LECTURE NOTES IN ANALYSIS (2011)

Sergiu Klainerman
Department of Mathematics, Princeton University, Princeton NJ 08544
E-mail address:

[email protected]

Part 1

INTRODUCTION TO PDE

1. The world of PDE

To start with partial differential equations, just like ordinary differential or integral
equations, are functional equations. That means that the unknown, or unknowns,
we are trying to determine are functions. In the case of partial differential equations (PDE) these functions are to be determined from equations which involve, in
addition to the usual operations of addition and multiplication, partial derivatives
of the functions. Below are the most basic examples,
(Laplace equation)
u = 0
2

(1)

where u = x
2 u + y 2 u + z 2 u. The other two examples described in the section
of fundamental mathematical definitions are

(Heat Equation)
t u + ku = 0,

(2)

t2 u + c2 u = 0.

(3)

(Wave equation)

In both cases one is asked to find a function u, depending on the variables t, x, y, z,

which verifies the corresponding equations. Observe that both (2) and (3) involve the symbol which has the same meaning as in the first equation, that is
2
2
2
2
2
2
u = ( x
2 + y 2 + z 2 )u = x2 u + y 2 u + z 2 u. In both (2) and (3) k > 0 and
c are fixed constants (representing the rate of diffusion for the first and the speed
of light in the second). It suffices to study to solve the equations for the special
cases k = 1 and c = 1. Indeed if u(t, x, y, z) is a solution of (3), for example, then
v(t, x, y, z) = u(t, x/c, y/c, z/c) verifies the same equation with c = 1. Both equations are called evolution equations, simply because they are supposed to describe
the change relative to the time parameter t of a particular physical object. Observe
that (1) can be interpreted as a particular case of both (3) and (2). Indeed solutions
u = u(t, x, y, z) of either (3) or (2) which are independent of t, i.e. t u = 0, verify
(1).
Here are some further examples of important PDEs:
dinger equation)
(Schro
it u + ku = 0

(4)

with u : R R3 C. The equation describes the quantum evolution of a massive

~
, where ~ > 0 is Plancks constant and m is the mass of the
particle, k = 2m
particle. As with the heat equation, one can normalize k = 1 by a simple change of
variables. Though the equation is formally very similar to the heat equation, it has
very different qualitative behavior. It is important to keep in mind when studying

1.

THE WORLD OF PDE

PDEs that small changes in the form of an equation can lead to very different
properties of solutions.
(Klein-Gordon equation)
t2 u + c2 u

mc2
~

2
u=0

(5)

This is the relativistic counterpart to the Schrodinger equation, the parameter m

has the physical interpretation of mass and mc2 has the physical interpretation of
rest energy (reflecting Einsteins famous equation E = mc2 ). One can normalize
2
the constants c and mc
~ to make them both equal 1 by applying a suitable change
of variables to both time and space.
Observe that all three PDE mentioned above satisfy the following simple property
called the principle of superposition: If u1 , u2 are solutions of an equation so is
any linear combination of them 1 u1 + 2 u2 where 1 and 2 are arbitrary real
numbers. Such equations are called linear. The following equation in the unknown
u = u(x, y), is manifestly not linear:
(Minimal surfaces
x

x u
(1 + |x

u|2

+ |y

1
u|2 ) 2

+ y

y u
(1 + |x

u|2

+ |y

1
u|2 ) 2

Here x and y are short hand notations for the partial derivatives

= 0.

and

(6)

y .

The equations we have encountered so far can be written in the form P[u] = 0,
where P is a differential operator applied to u. A differential operator is simply a
rule which takes functions u, defined in Rn or an open subset of it, into functions
P[u] by performing the following operations:
u
We can take partial derivatives i u = x
i relative to the variables x =
1
2
n
n
(x , x , . . . x ) of R . One allows also higher partial derivatives of u such
2
2
u
2
as the mixed second partials i j u = xi x
j or i = x2 .
i
The associated differential operators for (2) is P = t + and that
of (3) is t2 +
Can add and multiply u and its partial derivatives between themselves as
well as with given functions of the variables x. Composition with given
functions may also appear.

In the case of the equation (1) the associated differential operator is P = =

P3
12 + 22 + 32 = i,j=1 eij i j where eij is the diagonal 3 3 matrix with entries
(1, 1, 1) corresponding to the euclidean scalar product of vectors X, Y in R3 ,
< X, Y >= X1 Y1 + X2 Y2 + X3 Y3 =

3
X
i,j=1

eij Xi Xj .

(7)

The associated differential operators for (2), (3) and (4) are, resp. P = t + ,
P = t2 + and P = it + with variables are t, x1 , x2 , x3 R1+3 . In the
particular case of the wave equation (3) it pays to denote the variable t by x0 . The
wave operator can then be written in the form,
 = 02 + 12 + 22 + 32 =

3
X

(8)

,=0

where m is the diagonal 4 4 matrix with entries (1, 1, 1, 1), corresponding to

the Minkowski scalar product in R1+3 . This latter scalar product is defined, for 4
vectors X = (X0 , X1 , X2 , X3 ) and Y = (Y0 , Y1 , Y2 , Y3 ) by,
m(X, Y ) =

3
X

m X Y = X0 Y0 + X1 Y1 + X2 Y2 + X4 Y4

(9)

,=0

The differential operator  is called DAlembertian after the name of the French
mathematician who has first introduced it in connection to the equation of a vibrating string.
Observe that the differential operators associated to the equations (1)(4) are all
linear i.e.
P[u + v] = P[u] + P[v],
for any functions u, v and real, or complex, numbers , . The following is another
simple example of a linear differential operator
P[u] = a1 (x)1 u + a2 (x)2 u

(10)

where x = (x1 , x2 ) and a1 , a2 are given functions of x. They are called the coefficients of the linear operator. An equation of the form
P[u] = f,

(11)

corresponding to a linear differential operator P and a given function f = f (x), is

called linear-inhomogeneous . Any solution u of such an equation can be expressed
in the form u = u0 + v where u0 is a special solution of (11) and v solution to the
homogeneous equation
P[v] = 0.

(12)

In the case of the equation (6) the differential operator P can be written, relative
to the variables x1 and x2 , in the form,


2
X
1

u
,
P[u] =
i
i
1
(1 + |u|2 ) 2
i=1
where |u|2 = (1 u)2 + (2 u)2 . Clearly P[u] is not linear in this case. We call
it a nonlinear operator; the corresponding equation (6) is said to be a nonlinear
equation. An important property of both linear and nonlinear differential operators
is locality. This means that whenever we apply P to a function u, which vanishes
in some open set D, the resulting function P[u] also vanish in D.
Observe also that our equations (1)-(4) are also translation invariant. This means,
in the case (1) for example, that whenever the function u = u(x) is a solution so

1.

THE WORLD OF PDE

is the function uc (x) := u(Tc x) where Tc is the translation Tc (x) = x + c. On

the other hand the equation P[u] = 0, corresponding to the operator P defined by
(10) is not, unless the coefficients a1 , a2 are constant. Clearly the set of invertible
transformations1 T : Rn Rn which map any solution u = u(x), of P[u] = 0, to
another solution uT (x) := u(T x) form a group, called the invariance group of the
equation. The composition of two symmetries is again a symmetry, as is the inverse
of a symmetry, and so it is natural to view a collection of symmetries as forming a
group (which is typically a finite or infinite-dimensional Lie group).
The Laplace equation (1) is invariant not only with respect to translations but also
rotations, i.e linear transformations O : R3 R3 which preserve the euclidean
scalar product (7) in the sense that < OX, OY >=< X, Y > for all vectors X, Y
R3 . Similarly the wave equation (3) and Klein-Gordon equation (5) are invariant
under Lorentz transformations, i.e. linear transformations L : R1+3 R1+3 which
preserve the Minkowski scalar product (9), i.e. m(LX, LY ) = m(X, Y ). Our
other evolution equations (2) and (4) are clearly invariant under rotations of the
space variables x = (x1 , x2 , x3 ) R3 , keeping t fixed. They are also Galilean
invariant, which means, in the particular case of the Schrodinger equation (278),
2
that whenever u = u(t, x) is a solution so is uv (t, x) = ei(xv) eit|v| u(t, x vt) for
any vector v R3 .
So far we have tacitly assumed that our equations take place in the whole space
R3 for the Laplace equation, R4 for the Heat, Wave and Schrodinger equations and
R2 for the minimal surface equation. In reality, one is often restricted to a domain
of the corresponding space. Thus, for example, the equation (1) is usually studied
on a bounded open domain of R3 subject to a specified boundary condition. Here
is a typical example.
Example. The Dirichlet problem on an open domain of D R3 consists of
of D, twice continuously
finding a continuous functions u defined on the closure D
differentiable in D, such that u = 0 in D and the restriction of u to D, the
boundary of D, is prescribed to be a continuous function u0 . More precisely we
require that,
u|D = u0

(13)

One can impose the same boundary condition for solutions of (6), with D a bounded
open domain of R2 . A solution u = u(x, y) of (6) in D, verifying the boundary
condition (13), solves the Plateau problem of finding minimal surfaces in R3 which
pass through a given curve. One can show that the surface given by the graph
u = {(x, y, u(x, y))/(x, y) D R2 } has minimum area among all other graph
surfaces v verifying the same boundary condition, v|D = u0 .
Natural boundary conditions can also be imposed for the evolution equations (2)
(4). The simplest one is to prescribe the values of u on the hyperplane t = 0. In
the case of the heat and Schr
odinger equation we set,
u|t=0 = u0
1The transformations are often linear maps.

while in the case of the wave equation, which involves a second derivative in t, we
impose two conditions
u|t=0 = u0 and t u|t=0 = u1

(14)

where u0 , u1 are functions of the coordinates (x, y, z), called initial conditions. To
solve the initial value problem in both cases means to find solutions of the equations
for t > 0 which verify the corresponding initial conditions at t = 0. In addition
one may restrict the variables (x, y, z) to an open domain of D R3 . More to the
point one may try to solve a boundary value problem in a domain [0, ) D with a
boundary condition, such as (13), on [0, ) D and an initial condition at t = 0.
The choice of boundary condition and initial conditions, for a given PDE, is very
important. Finding which are the good boundary and initial conditions is an important aspect of the general theory of PDE which we shall address in section 2.
For equations of physical interest these appear naturally from the context in which
they are derived. For example, in the case of a vibrating string, which is described
by solutions of the one dimensional wave equation t2 u x2 u = 0 in the domain
(a, b) R, the initial conditions u = u0 , t u = u1 at t = t0 , amount to specifying
the original position and velocity of the string. On the other hand the boundary
condition u(a) = u(b) = 0 simply mean that the two ends of the of the string are
fixed.
So far we have only considered equations in one unknown. In reality many of
the equations of interest appear as systems of partial differential equations. The
following important example, contains two unknown functions u1 = u1 (x1 , x2 ), u2 =
u2 (x1 , x2 ) which verify,
(Cauchy-Riemann)
1 u2 2 u1 = 0,

1 u1 + 2 u2 = 0

(15)

It was first observed by Cauchy that u = u1 + iu2 , as a function of z = x1 + ix2 , is a

complex analytic function if and only if (15) is satisfied. Setting also = 1 + i2 ,
observe that (15) is equivalent to
= 0.
u

(16)

Equation (15) can also be written in the form P[u] = 0 by introducing u = (u1 , u2 )
as a column vector and P[u] the differential operator,

  
2 1
u1
P[u] =

1 2
u2
The system of equations (15) contains two equations and two unknowns. This is
the standard situation of a determined system. A system is called over-determined
if it contains more equations than unknowns and underdetermined if it contains
fewer equations than unknowns. For example the system of two equations and
one unknown x u(x, y) = f, y u(x, y) = g is clearly overdetermined. A necessary
condition for a solution to exist is y f = x g, condition which can be interpreted

1.

THE WORLD OF PDE

as requiring that the one-form w = f (x, y)dx + g(x, y)dy is exact, i.e. its exterior derivative d is identically zero. Overdetermined systems, such as De Rham
complexes, play a very important role in geometry.
Observe that (15) is a linear system. Observe also that the operator P has the
following remarkable property.


u1
P 2 [u] = P[P[u]] =
u2
In other words P 2 = I, with I the identity operator I[u] = u, and therefore
P can be viewed as a a square root of . One can define a similar type of square
root for the DAlembertian . To achieve this we need 4 4 complex matrices
0 , 1 , 3 , 4 which satisfy the property
+ = 2m I

(17)

with I the unit 4 4 matrix and m

as in (8). Using the matrices we can
introduce the Dirac operator acting on u = (u1 , u2 , u3 , u4 ) defined from R1+3 with
values in C4 by,
Du = i u

(18)

Using (17) we easily check that, D u = u. Thus the Dirac operator D can be
viewed as a square root of the DAlembertian . It leads to the following fundamental equation introduced by Dirac as the equation of free, massive, relativistic,
particle such as the electron:
(Dirac Equation)
Du = ku

(19)

Partial differential equations are ubiquitous throughout Mathematics and Science.

They provide the basic mathematical framework for some of the most important
physical theories, such as Elasticity, Hydrodynamics, Electromagnetism, General
Relativity and Non-relativistic Quantum Mechanics. The more modern relativistic
quantum field theories lead, in principle, to equations in infinite number of unknowns, which lie beyond the scope of partial differential equations. Yet, even in
that case, the basic equations preserve the locality property of PDE. Moreover the
starting point of a quantum field theory is always a classical field theory, described
by systems of PDEs. This is the case, for example, of the Standard Model of weak
and strong interactions, based on a Yang -Mills-Higgs field theory. If we also include
the ordinary differential equations of Classical Mechanics, which can be viewed as
one dimensional PDE, we see that, essentially, all of Physics is described by differential equations. Other examples of partial differential equations underlining some
of our most basic physical theories, are the Maxwell, Einstein, Euler and Navier
Stokes equations. Note that each equation are at the heart of an entire field of
Physics, i.e. Electrodynamics, General Relativity and Hydrodynamics,
An important feature of the main PDEs appearing in Physics is their apparent
universality. Thus, for example, the wave equation, first introduced by Dalembert
to describe the motion of a vibrating string was later found to be connected to
the propagation of sound and electromagnetic waves. The heat equation, first

10

introduced by Fourier to describe heat propagation, also comes in many other

situations in which dissipative effects play an important role. The same thing can
be said about the Laplace, Schrodinger and many other basic equations.
It is even more surprising that equations, originally introduced to describe specific
physical phenomena, also play a fundamental role in areas of mathematics, which
are considered pure, such as Complex Analysis, Differential Geometry, Topology
and Algebraic Geometry. Complex Analysis, for example, which studies the properties of holomorphic functions, can be regarded as the study of solutions to the
Cauchy-Riemann equations (15) in a domain of R2 . Hodge theory, which plays
a fundamental role in topology and algebraic geometry, is based on studying the
space of solutions to a class of linear systems of partial differential equations on
manifolds which generalize the Cauchy-Riemann equations. The Atiyah-Singer index theorem is formulated in terms of a special classes of linear PDE on manifolds,
related to the euclidean version of the Dirac operator (18).2
Important problems in geometry can be reduced to finding solutions to specific
partial differential equations, typically nonlinear. We have already seen such an
example in the case of the Plateau problem of finding surfaces of minimal total
area which pass through a given curve. The well known uniformization theorem
provides another excellent example.
To state the uniformization theorem, we need to recall the definition of a compact
Riemann surface S. This is a 2 -dimensional, compact manifold endowed with a
smooth, positive definite metric g. The Gauss curvature K = K(g) is an important
invariant of the surface which can be calculated explicitely at every point p
S in terms of the components gab relative to a local system of coordinates x =
(x1 , x2 ) near p. The calculation involves first and second partial derivatives of the
components gab relative to x1 and x2 . The remarkable fact is that the final value of
K does not depend on the particular system of coordinates in which one makes the
calculation. Moreover in the particular case when S is the standard sphere in R3 ,
given by the equation |x|2 = a2 , the Gauss curvature is equal to the expected value,
corresponding to our intuition of curvature, that is K = a2 . Another remarkable
property of the Gauss curvature is that its total integral along S does not depend on
the metric g but only on the topological properties of S. More precisely, according
to the Gauss-Bonnet formula, we have
Z
(S) = (2)1
Kdag ,
S
1
2
with dap
g denoting the area element of the metric g. In coordinates x , x we have
dag = |g|dx1 dx2 with |g| the determinant of the matrix (gab )a,b=1,2 . The number
(S) is one of the integers 2, 0, 2, . . . 2k . . . , called the Euler characteristic of
S, and has simple topological interpretation. Thus any surface which can be continuously deformed to the standard sphere has (S) = 2 while any surface which
can be continuously deformed to a torus has (S) = 0. We can now state the
uniformization theorem:
2This is the operator obtained when we change the minkowski metric m to the euclidean one
e in (17).

1.

THE WORLD OF PDE

11

Theorem 1.1. Let S be a 2-dimensional, compact, Riemann surface with metric g,

Gauss curvature K = K(g) and Euler characteristic (S). There exists a conformal
transformation of the metric g, i.e. g = 2 g, for some smooth non-vanishing
of the new metric g is identical equal
function , such that the Gauss curvature K
to 1, 0 or 1 according to whether (S) > 0, (S) = 0 or (S) < 0.
To prove this very important geometric result, which leads to the complete classification of all compact surfaces according to their Euler characteristic, we are led
to a nonlinear partial differential equation on S. Indeed assume that (S) = 2 and
of the metric g = e2u g to be exactly 1.
therefore we want the Gauss curvature K

It is easy to relate K, by a simple calculation, to the Gauss curvature K of the

original metric g . This leads to the following equation in u,
S u + e2u = K

(20)

where S , called the Laplace-Beltrami operator of S, is a straightforward adaptation of the Laplace operator, see (1), to the surface S. Thus the proof of the
uniformization theorem reduces to solve equation (20), i.e. for a given surface S
with Gauss curvature K, find a real valued function u which verifies (20).
We give below a precise definition of the operator S relative to a system of local coordinates x = (x1 , x2 ) on an open coordinate chart D S. Denote by
G(x) = (gab (x))a,b=1,2 the 2 2 matrix whose entries are the components of our
Riemannian metric on D. Let G1 (x) denote the matrix inverse to G(x) and denote
its components by (g ab (x))a,b=1,2 . Thus, for all x D,
X
gac (x)g cb (x) = ab
c

with ab the usual Kronecker symbol. We also set, as before, |g(x)| = det(G(x))
and define,
X
p
1
b ( |g(x)| g ab (x) a u(x))
S u(x) = p
|g(x)| a,b=1,2
Typically we suppress the explicit dependence on x in the above formula. It is also
very convenient to use Einsteins summation convention over repeated indices, and
thus write,
p
1
S u = p b ( |g| g ab a u)
(21)
|g|
As a third example we consider the Ricci flow equation on a compact n dimensional manifold M , which is described in one of the articles of the Compendium. In
the particular case of three dimensions the equation has been recently used, decisively, to provide the first proof of Thurstons geometrization conjecture, including
the well known Poincare conjecture. The geometrization conjecture, described in
the topology section of the Compendium, is the precise analogous, in three space
dimensions, of the 2-dimensional uniformization theorem mentioned above. The
Ricci flow is defined, in arbitrary local coordinates x = (x1 , x2 , x3 ) on M , by the
equation:

12

(Ricci Flow)
t gij = Rij (g)

(22)

Here gij = gij (t) is a family of Riemannian metrics depending smoothly on the
parameter t and Rij (g) denotes the Ricci curvature of the metric gij . This is simply
a three dimensional generalization of the Gauss curvature we have encountered
in the uniformization theorem. In a given system of coordinates Rij (g) can be
calculated in terms of the metric coefficients gij and their first and second partial
derivatives. Since both gij and Rij are symmetric relative to i, j = 1, 2, 3 we can
interpret (22) as a non-linear system of six equations with six unknowns. On a
closer look it turns out that (22) is related to the heat equation (2). Indeed, by
a straightforward calculation relative to a particular system of coordinates x =
(x1 , x2 , x2 ) called harmonic, it can be shown that the Ricci flow (22) takes the form
t gij g gij = Nij (g, g)

(23)

where each Nij , i, j = 1, 2, 3, are functions of the components gij and their first
partial derivatives with respect to the coordinates x and g is, again, a differential
operator very similar to the Laplacian in R3 , see (??). More precisely, if G1 =
(g ab )a,b=1,2,3 denotes the matrix inverse to G = (gab )a,b=1,2,3 we can write, using
the summation convention,
g = g ab a b =

3
X

g ab a b .

a,b=1

In a small neighborhood of a point p M we can choose the harmonic coordinate

xa such that g ab (p) = ab with ab denoting the usual Kronecker symbol. Thus,
near p, g looks indeed like = ab a b .
The Ricci flow3 allows one to deform an arbitrary Riemannian metric on M to a a
simple metric of constant sectional curvature. The idea is to start with a metric g
and look for solutions g(t) of (22) which verify the initial condition g(0) = g. One
hopes that as t the metric g(t) will converge to a metric of constant curvature.
Intuitively one can see this working out the same way heat gets evenly distributed
in space, as t increases, according to the heat equation (2). Indeed since (22) is
similar to (2) we expect the variations in the curvature of g(t) to become smaller
and smaller as the metric evolves according to (22). The type of metric we get in
the limit as t will allow us to determine the topological character of M . The
flow, however, can develop singularities before we achieve that goal. To overcome
this major technical difficulty one needs to make a detailed qualitative analysis of
the behavior of solutions to (22), task which requires just about all the advances
made in geometric PDE in the last hundred years.
As we have seen above the choice of harmonic coordinates allows us to write the
Ricci flow as a system of nonlinear heat equations (23). This fact is quite typical
to geometric equations. It is useful at this point to discuss another important
example, that of the Einstein equations in vacuum. An introduction to this equation
3In reality one needs to change the equation (22) slightly to make sure that the total volume
of of M , calculated with respect to the metric g(t), stays constant.

1.

THE WORLD OF PDE

13

and short discussion of its importance in General Relativity can be found (see
compendium article). Solutions to the Einstein vacuum equations are given by
Ricci flat spacetimes, that is Lorentzian manifolds (M, g) with M a four dimensional
manifold and g a Lorentz metric on it, for which the corresponding Ricci curvature
vanishes identically.
(Einstein-vacuum)
Ric(g) = 0.

(24)

The Ricci curvature of a Lorentz metric, Ric(g), can be defined in exactly the same
way as in the Riemannian case. Thus relative to a coordinate system x , with
= 0, 1, 2, 3, the Ricci curvature, denoted by R , can be expressed in terms of
the first and second partial derivatives of the metric coefficients g . As before,
we denote by g the components of the inverse metric. Moreover, by picking
a specified system of coordinates, called wave coordinates4, we can express the
Einstein-vacuum equations (24) in the form of a system of equations related to the
wave equation (3), in the same way the Ricci flow system (23) was related to the
heat equation (2). More precisely,
g g = N (g, g)

(25)

where, as in the case of the Ricci flow, the terms N (g, g) are expressions, which
can be calculated explicitely, depending on the metric g , its inverse g and the
first derivatives of g relative to the coordinates x . This is a system of 10 equations with respect to the ten unknown components of the metric (g ),=0,1,2,3 .
The differential operator,
X
g =
g
,

appearing on the left hand side is very similar to the wave operator  = m =
02 + which we have encountered before in (8). Indeed, in a neighborhood of a
point p M we can pick our wave coordinates x in such a way that g (p) = m .
Thus, locally, g looks like  = m and we can thus interpret (25) as a nonlinear
system of wave equations.
The two last examples illustrate the importance of choosing good coordinates for
equations which are defined in terms of geometric quantities, such as the Ricci
curvature. To solve such equations and find interesting properties of the solutions,
it is often very important to pick up a well adapted system of coordinates. In the
case of gauge field theories, such as Yang-Mills equations, the role of coordinates is
replaced by gauge transformations.
Finally we need to note that PDE arise not only in Physics and Geometry but also
in many fields of applied science. In engineering, for example, one often wants to
impose auxiliary conditions on solutions of a PDE, corresponding to a part of a
physical system which we can directly influence, such as the portion of the string
of a violin in direct contact with the bow, in order to control their behavior, i.e.
4they are the exact analogue of the harmonic coordinates discussed above.

14

obtain a beautiful sound. The mathematical theory dealing with this issue is called
Control Theory.
Often, when dealing with complex physical systems, when we cannot possible have
complete information about the state of the system at any given time, one makes
various randomness assumptions about various factors which influence it. This
leads to a very important class of equations called stochastic differential equations.
To give a simple example consider the N N system of the ordinary differential
equation,
dx
= f (x(t))
(26)
dt
Here f is a given function f : RN RN . A solution x(t) is a vector valued function
x : [0, ) RN . Given an initial data x(0) = x0 we can precisely determine
the position x(t) and velocity dx
dt of the solution at any given time t. In applied
situations, because of various factors which are hard to take into account, the state
of the solution may not be so neatly determined. It is thus reasonable to modify
the equation to take into account random effects which influence the system. One
then looks at en equation of the form,
dx
dW
= f (x(t)) + B(x(t))
(t)
(27)
dt
dt
where B(x) is a N M dimensional matrix and W (t) denotes the brownian motion
in RM . Similar modifications, which take randomness into account, can be made for
partial differential equations. A particularly interesting example of a PDE, which
is derived from a stochastic process, related to the price of stock options in finance,
is the well known Black- Scholes equation. The real price of a stock option u(s, t)
at time t and value s, verifies the PDE,
2 2 2
s s u ru = 0,
s > 0, t [0, T ],
(28)
2
subject to the terminal condition at expiration time T , u = max(0, (s p)) and
boundary condition u(0, t) = 0, t [0, T ]. Here p is the strike price of the option.
Observe that this equation is in fact a (time-reversed) variant of the heat equation (2), thus illustrating the point made above that a single class of mathematical
equations can arise in several completely different applications (in this case, thermodynamics and mathematical finance).
t u + rss u +

CHAPTER 1

Equations Derived by the Variational Principle

1. Basic Notions
In this section we will discuss some basic examples of linear and nonlinear equations
which arise variationaly from a relativistic Lagrangian. The fundamental objects
of a relativistic field theory are:
Space-time (M, g) which consists of an n + 1 dimensional manifold M
and a Lorentz metric g; i.e . a nondegenerate quadratic form with signature (1, 1, . . . , 1) defined on the tangent space at each point of M. We
denote the coordinates of a point in M by x , = 0, 1, . . . , n.
Throughout most of this chapter the space-time will in fact be the
simplest possible example - namely, the Minkowski space-time in which
the manifold is Rn+1 and the metric is given by

2
2
ds2 = m dx dx = dt2 + dx1 + + (dxn )
(29)
with t = x0 , m = diag(1, 1, . . . , 1). Recall that any system of coordinates for which the metric has the form (29) is called inertial. Any two
inertial coordinate systems are related by Lorentz transformations.
Collection of fields = (1) , (2) , . . . , (p) which can be scalars, tensors, or some other geometric objects1 such as spinors, defined on M.
Lagrangian density L which is a scalar function on M depending only
on the tensorfields and the metric2 g.
We then define the corresponding action S to be the integral,
Z
S = S[, g : U] =
L[]dvg
U

where U is any relatively compact set of M. Here dvg denotes the volume element
generated by the metric g. More precisely, relative to a local system of coordinates
x , we have

dvg = gdx0 dx1 dxn = gdx

with g the determinant of the matrix (g ).
By a compact variation of a field we mean a smooth one-parameter family of
fields (s) defined for s (, ) such that,
1For simplicity we restrict ourselves to covariant tensors.
2as well as its inverse g1
15

16

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

(1) At s = 0, (0) = .
(2) At all points p M \ U we have (s) = .
Given such a variation we denote := :=

d(s)
ds

. Thus, for small s,

s=0

(s) = + s + O(s2 )
A field is said to be stationary with respect to S if, for any compact variation
((s) , U) of , we have

d

S(s)  = 0
ds
s=0
where,
S(s) = S[(s) , g; U]
We write this in short hand notation as
S
=0

Action Principle, also called the Variational Principle, states that an acceptable
solution of a physical system must be stationary with respect to a given Lagrangian
density called the Lagrangian of the system. The action principle allows us to derive
partial differential equations for the fields called the Euler-Lagrange equations.
Here are some simple examples:
1. Scalar Field Equations :
One starts with the Lagrangian density
1
L[] = g V ()
2
where is a complex scalar function defined on (M, g) and V () a given real
function of .
Given a compact variation ((s) , U) of , we set S(s) = S[(s) , g; U]. Integration
by parts gives,
Z

d

V 0 ()]
gdx
S(s) 
=
[g
ds
s=0
ZU
g V 0 ()]dvg ]
=
[
U

where g is the DAlembertian,

g =




1
g g
g

.
In view of the action principle and the arbitrariness of we infer that must satisfy
the following Euler-Lagrange equation
g V 0 () = 0,

(30)

1. BASIC NOTIONS

17

Equation (30) is called the scalar wave equation with potential V ().
CONFORMAL PROPERTIES 2. Wave Maps :

The wave map equations will be defined in the context of a space-time (M, g), a
Riemannian manifold N with metric h, and a mapping
: M N.
We recall that if X is a vectorfield on M then X is the vectorfield on N defined
by X(f ) = X(f ). If is a 1-form on N its pull-back is the 1-form on M
defined by (X) = ( X), where X is an arbitrary vectorfield on M. Similarly
the pull-back of the metric h is the symmetric 2-covariant tensor on M defined by
the formula ( h)(X, Y ) = h( X, Y ). In local coordinates x on M and y a on
N , if a denotes the components of relative to y a , we have,

a b
hab ((p)) = h , i
x x
x x
where < , > denotes the Riemannian scalar product on N .
( h) (p) =

Consider the following Lagrangian density involving the map ,

1
L = Trg ( h)
2
where Trg ( h) denotes the trace relative to g of h. In local coordinates,
1
a b
L[] = g hab () .
2
x x
By definition wave maps are the stationary points of the corresponding action.
Thus by a a straightforward calculation,

d

0 =
S(s)  = I1 + I2
(31)
ds Z
s=0

1
hab () c
I1 =
g
a b gdx
c
2 U

I2 =
g hab () a b gdx
U

After integrating by parts, relabelling and using the symmetry in b, c, we can rewrite
I2 in the form,


Z
hab
c
b
I2 =
a hab ()g b + g

dvg
(32)

c
U




Z
1
hab
hac
b
c
=
a hab ()g b + g
+

dvg

2
c
b
U
Also, relabelling indices
1
I1 =
2

Z
U

hbc a
b c dvg .
a

18

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

Therefore,
0

= I1 + I2



Z
hac
hbc
hab
a
b
b
c 1

dvg
=
hab g + g
2 c
b
a
U



Z
hsc
hbc
hsb
a
d
b
c 1 ds

dvg
=
had g + g
h had
2
c
b
s
ZU


=
a had g d + b c g dbc dvg
U

where

dbc

= 12 hds

hsb
c

hsc
b

hbc
s

are the Christoffel symbols corresponding

to the Riemannian metric h. The arbitrariness of yields the following equation
for wave maps,
g a + abc g b c = 0

Example:

(33)

Let N be a two dimensional Riemannian manifold endowed with a

metric h of the form,

ds2 = dr2 + f (r)2 d2
Let be a wave map from M to N with components 1 , 2 , relative to the r,
0
(r)
.
coordinates. Then, 111 = 211 = 112 = 222 = 0 and 122 = f 0 (r)f (r), 212 = ff (r)
Therefore,
g 1
g 2

= f 0 (r)f (r)g 2 2
f 0 (r)
g 1 2
=
f (r)

The equations of wave maps can be given a simpler formulation when N is a submanifold of the Euclidean space Rm . In this case, the metric h is the Euclidean
metric3 so the first term in (31) vanishes.
Z


d

S(s) 
=
g h , idvg
ds
x x
s=0
Z U
=
< , > dvg
U

where h , i is the Euclidean scalar product and  the DAlembertian operator on

M. Thus the Euler-Lagrange equations take the form,
((p))T = 0

(34)

where T here means the projection onto the tangent space of N at (p).
In the special case when N Rm is a hypersurface, we can rewrite (34) in a more
concrete form. Let be the unit normal on N and k the second fundamental
form k(X, Y ) = hDX , Y i, with DX the standard covariant derivative of Euclidean
space. The hypersurface N is defined (locally) as the level set of some real valued
3Use the standard coordinates of the ambient Euclidean space.

1. BASIC NOTIONS

19

f . Differentiating the equation f ((x)) = 0 with respect to local coordinates x on

M yields 0 = < (), > along M. Hence,
0

= < (), >=< , > +g < (), >

= < 2, > +g < (E ) , (E ) >
i

Where (E ) = x
y i is the pushforward of E =
tangent to N . Therefore,

x .

In particular, (E ) is

< , >= k( (E ), (E ))

(35)

Thus the equation for wave maps becomes,

 = k( (E ), (E ))
In the case when N is the standard sphere Sm1 Rm , k(X, Y ) = < X, Y > and
the equation for wave maps becomes, in standard coordinates x in Rm , a () = a ,
a = a g <

,
>
x x

3. Maxwell equations:
An electromagnetic field F is an exact two form on a four dimensional manifold
M. That is, F is an antisymmetric tensor of rank two such that
F = dA

(36)

where A is a one-form on M called a gauge potential or connection 1-form. Note

that A is not uniquely defined - indeed if is an arbitrary scalar function then the
transformation
A A = A + d

(37)

yields another gauge potential A for F . This degree of arbitrariness is called gauge
freedom, and the transformations (37) are called gauge transformations.
The Lagrangian density for electromagnetic fields is
L[F ]

1
F F .
4

Any compact variation (F(s) , U) of F can be written in terms of a compact variation

(A(s) , U) of a gauge potential A, so that F(s) = dA(s) . Write

d

F =
F(s)  ,
ds
s=0


d

A =
A(s) 
ds
s=0

20

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

so that relative to a coordinate system x we have F = A A and therefore

F = A A . The action principle gives
Z

1
d

S(s)  =
F F dvg
0 =
ds
2 M
s=0
Z
1
=
( A A )F dvg
2 U


Z
Z

gF
dvg
=
A F dvg =
A
g
U
U
Note that the second factor in the integrand is just D F where D is the covariant
derivative on M corresponding to g. Hence the Euler-Lagrange equations take the
form
D F = 0.

(38)

Together, (36) and (38) constitute the Maxwell equations.

Exercise.

Given a vector field X on M, show

1
gX
D X =
g

We can write the Maxwell equations in a more symmetric form by using the Hodge
dual of F ,
1
?
F = F
2
and by noticing that (38) is equivalent to d ?F = 0. The Maxwell equations then
take the form
dF = 0,

d ?F = 0

(39)

D ?F = 0

(40)

or, equivalently,
D F = 0,

Note that since Lorentz transformations commute with both the Hodge dual and
exterior differentiation, the Lorentz invariance of the Maxwell equations is explicit
in (39).
Definition.

Given X an arbitrary vector field, we can define the contractions

E = (iX F )

X F

H = (iX ?F )

X ?F

called, respectively, the electric and magnetic components of F . Note that both
these one-forms are perpendicular to X.
d
We specialize to the case when M is the Minkowski space and X = dxd 0 = dt
. As
d
remarked, E, H are perpendicular to dt , so E0 = H0 = 0. The spatial components
are by definition

Ei

F0i

Hi

F0i =

1
1
0ijk F jk = ijk F jk
2
2

1. BASIC NOTIONS

21

We now use (39) to derive equations for E and H from above, which imply
D ?F = 0

(41)

and (38), respectively. Setting = 0 in both equations of (40) we derive,

i Ei

0,

i Hi = 0
k

Setting = i and observing that Fij =ijk H ,

(42)

Fij = ijk E we write

0 Ei + j Fij

= 0 Ei + ijk j H k = t Ei + ( H)i

t Hi ijk j Ek = t Hi ( E)i

Therefore,
t E + H

(43)

t H E

(44)

Alongside (43) and (44) we can assign data at time t = 0,

Ei (0, x)

(0)

Ei ,

(0)

Hi (0, x) = Hi

Exercise.
Show that the equations (42) are preserved by the time evolution of
the system (43)-(44). In other words if E (0) , H (0) satisfy (42) then they are satisfied
by E, H for all times t R.

4. Yang-Mills equations :
The Lagrangians of all classical field theories exhibit the symmetries of the spacetime. In addition to these space-time symmetries a Lagrangian can have symmetries
called internal symmetries of the field. A simple example is the complex scalar
Lagrangian,
1
L = m V (||)
2
where is a complex valued scalar defined on the Minkowski space-time Rn+1 ,
its complex conjugate. We note that L is invariant under the transformations
ei with a fixed real number. It is natural to ask whether the Lagrangian
can be modified to allow more general, local phase transformations of the form
(x) ei(x) (x). It is easy to see that under such transformations, the Lagrangian fails to be invariant, due to the term m . To obtain an invariant Lagrangian one replaces the derivatives by the covariant derivatives
(A)
D , + iA depending on a gauge potential A . We can now easily check
that the new Lagrangian
1
(A) V (||)
L = m D(A) D

2
is invariant relative to the local transformations,
(x ) ei(x) (x ) ,
called gauge transformations.

A A ,

22

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

Remark that the gauge transformations introduced above fit well with the definition
of the electromagnetic field F . Indeed setting F = dA we notice that F is invariant.
This allows us to consider a more general Lagrangian which includes F ,
1
1
L = F F m , , V (||)
4
2
called the Maxwell-Klein-Gordon Lagrangian.
The Yang-Mills Lagrangian is a natural generalization of the Maxwell-Klein-Gordon
Lagrangian to the case when the group SU (1), corresponding to the phase transformations of the complex scalar , is replaced by a more general Lie group G. In
this case the role of the gauge potential or connection 1-form is taken by a G valued
one form A = A dx defined on M. Here G is the Lie algebra of the Lie group G.
Let [ , ] its Lie bracket and < , > its Killing scalar product. Typically the Lie
group G is one of the classical groups of matrices, i.e. a subroup of either Mat(n, R)
or Mat(n, C). We pause briefly to recall some facts about the relavent Lie groups
and their Lie algebras.
(1) The orthogonal groups O(p, q). These are the groups of linear transformations of Ren which preserve a given nondegenerate symmetric bilinear
form of signature p, q, p + q = n. We denote by Rnp,q the corresponding
space. The case p = 0 is that of the Euclidean case, the group is then
simply denoted by O(n). The case p = 1, q = n is that of the Minkowski
space-time Rn+1 , the group O(1, n) is the Lorentz group. In general let
Q be the diagonal matrix whose first p diagonal elements are 1 and the
remaining ones are +1. Then,
O(p, q)

= {L Mat(n, R)|LT QL = Q}
= {L Mat(n, R)|LM LT = M }

Note that for L O(p, q), det(L) = 1.

Recall that the special orthogonal groups SO(p, q) are defined by
SO(p, q) = {L O(p, q)| det L = 1}.
They correspond to all orientation preserving isometries of Rnp,q . Both
O(p, q) and SO(p, q) have as Lie algebra4
so(p, q) = {A Mat(n, R)|AQ + QAT = 0}.
and that dimR O(p, q) = dimR SO(p, q) = n(n 1)/2. The Lie bracket for
so(p, q) is the usual Lie bracket of matrices, i.e. [A, B] = AB BA and
we have the Jacobi identity
[A, [B, C]] + [C, [A, B]] + [B, [C, A]] = 0

(45)

and its Killing scalar product < A, B >= Tr(AB T ) (where Tr is the
usual trace for matrices) enjoys the compatibility condition
< A, [B, C] >= < [A, B], C >

(46)

4Recall that the Lie algebra of a Lie group G is simply the tangent space to G at the origin.

1. BASIC NOTIONS

23

(2) The unitary groups U(p, q). These are the complex analogues of the
orthogonal groups. They are the groups of all linear transformations of
Cn which preserve a given nondegenerate hermitian bilinear form. Denote
by Cnp,q the corresponding space. Then, with the matrix Q as above,
U(p, q) = {U Mat(n, C) | U QU = Q}
and,
SU(p, q) = {U U(p, q) | det U = 1},
The corresponding Lie algebras are,
u(p, q)
su(p, q)

{A Mat(n, C) | AQ + QA = 0},

= {A U(p, q) | trM A = 0},

where the trace trQ A = Qij Aij . The Lie bracket is again the usual one for
matrices. The Killing scalar product is given by < A, B >= Tr(AB ).
Remark also that dimR U(p, q) = n2 , dimR SU(p, q) = n2 1.
In the Yang-Mills theory one is interested in compact Lie groups with a positive
definite Killing form. This is the case for the groups O(n), SO(n), U (n), SU (n).
In a given system of coordinates the connection 1-form A has the form, A dx and
we define the (gauge) covariant derivative of a G-valued tensor by
D(A)
= D + [A , ]

(47)

where D is the covariant derivative on M. Observe that (47) is invariant under the
following gauge transformations, for a given G-valued gauge potential A and a Gvalued tensor ,


= U 1 U,
A = U 1 A U + D U 1 U
(48)
with U G.
Proposition 1.1.



D(A) = U 1 D(A)
U
]
A
= D
Proof :

This just requires some patience. First we will show

D U 1 U = U 1 D + [, U (D U 1 )] U

Indeed
D U 1 U

=
=
=

D U 1 U + U 1 (D ) U + U 1 (D U )


U 1 (D U ) U 1 + D + (D U ) U 1 U


U 1 D + [, (D U ) U 1 ] U

24

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

as desired. Hence

D(A) = D + [A , ]





= U 1 D + [, U D U 1 ] + U 1 A U + D U 1 U, U 1 U




= U 1 D + [, (D U ) U 1 ] + [A , ] + U D U 1 , U
^
(A)
= U 1 (D + [A , ]) U = D

As in Riemmanian geometry, commuting two (gauge) covariant derivatives produces

a fundamental object called the curvature, here denoted by F
D D D D = [F , ]

(49)

where the components F of the curvature can be deduced by the following

straightforward computation:
D D

D (D ) + [A , D ]

D (D + [A , ]) + [A , D + [A , ]]

= D D + [D A , ] + [A , D ] + [A , D ] + [A , [A , ]]
So that
(D D D D )

[D A D A , ]
+ [A , [A , ]] [A , [A , ]]
{z
}
|
[[A ,A ],]

Therefore,
F

D A D A + [A , A ]

(50)

We leave it to the reader to show that the curvature tensor F is invariant under
gauge transformations. That is,
]

(A)
F

U 1 F (A) U

= F (A)

and that F satisfies the Bianchi identity

D F + D F + D F = 0

(51)

We are finally ready to present the generalization of the Maxwell theory provided
by the Yang-Mills Lagrangian:
L[A]

1
(A)
< F , F (A) >G
4

(52)

1. BASIC NOTIONS

25

We derive the Euler-Lagrange equations just as in the Maxwell theory,

Z

d
1

0 =
S(s)  =
< F , F >G dvg
ds
2 U
s=0
Z
1
< D A D A + [A , A ] + [A , A ], F >G dvg
=
2 U
Z
=
< D A , F > + < [A , A ], F >G dvg
U
Z


=
< A , D F >G + < A , A , F >G dvg
U

which implies
D F = 0

(53)

Together, (51) and (53) form the Yang-Mills equations.

Note that the equations are invariant under the group of gauge transformations.
A solution of the Yang-Mills equations, then, is an equivalence class of gaugeequivalent potentials A whose curvature F satisfies (53).
In our later treatment of Yang-Mills, we will almost always specify a representative
of a solutions equivalence class by imposing additional constraints - called gauge
conditions - on A. There are three standard ways of doing this, each yielding its
own rendition of the Yang-Mills equations with its own faults and advantages:
Coulomb Gauge is defined by,
i Ai (t, x) = 0 (t, x) Rn+1

(54)

To simplify notation, first write (53) in terms of the current J.



D F = J = A , F,

(55)

When = 0 (54) allows us to write (55) as

J0

= i F0i = i (0 Ai i A0 + [A0 , Ai ]) A0 + i [A0 , Ai ]

giving us for the time component of A:

A0 = 2 [i A0 , Ai ] + [A0 , t Ai ] + [Ai , [A0 , Aj ]]

(56)

When = i, (55) reads

Ji

= t + j Fij = t (i A0 + [Ai , A0 ]) + j (i Aj j Ai + [Ai , Aj ])

and after simplifying,
2Ai

t i A0 2 [Aj , j Ai ] + [Aj , i Aj ] + [t Ai , Aj ]
+2 [A0 , t Ai ] [A0 , i A0 ] [Aj , [Aj , Ai ]] + [A0 , [A0 , Ai ]]

(57)

Lorentz Gauge is specified by,

A (t, x) = 0 (t, x) R3+1

(58)

26

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

Appealing in its symmetric treatment of the time and space components

of A, , the Lorentz gauge also allows (53) to be written as a system of
wave equations:
D F

= D ( A A + [A , A ])


= 2A + [A , A ] + [A , A ] [A , A ] + A , [A , A ]

The system can be written schematically in the form

2

+ 3

Again, it is not at all clear that one can transform an arbitrary solution
into the Lorentz gauge. In addition, we will have a hard time finding good
estimates for this purely hyperbolic system of nonlinear wave equations.
Temporal Gauge is specified by the condition A0 = 0.

5. The Einstein Field Equations:

According to the general relativistic variational principle the space-time metric g
is itself stationary relative to an action,
Z
S=

Ldvg .
U

Here U is a relatively compact domain of (M, g) and L, the Lagrangian, is assumed

to be a scalar function on M whose dependence on the metric should involve no
more than two derivatives5. It is also assumed to depend on the matterfields =
(1) , (2) , . . . (p) present in our space-time.
In fact we write,
S = SG + SM
with,
Z
SG

LG dvg
U

Z
SM

LM dvg
U

denoting, respectively, the actions for the gravitational field and matter. The matter Lagrangian LM depends only on the matterfields , assumed to be covariant
tensorfields, and the inverse of the space-time metric g which appears in the
contraction of the tensorfields in order to produce the scalar LM . It may also
depend on additional positive definite metrics which are not to be varied 6.
5In fact we only require that the corrsponding Euler-Lagrange equations should involve no
more than two derivatives of the metric.
6This is the case of the metric h in the case of wave maps or the Killing scalar product in
the case of the Yang-Mills equations.

1. BASIC NOTIONS

27

Now the only possible candidate for the gravitational Lagrangian LG , which should
be a scalar invariant of the metric with the property that the corresponding EulerLagrange equations involve at most two derivatives of the metric, is given7 by the
scalar curvature R. Therefore we set,
LG = R.
d
Consider now a compact variation (g(s) , U) of the metric g. Let g = ds
g |s=0 .
2

Thus for small s, g (s) = g + sg + O(s ). Also, g (s) = g sg + O(s2 )

where g = g g g . Then,

Z
Z

d


SG (s)  =
Rdvg +
Rdv g
ds
s=0
U
U
Now,
1
dv g = g g dvg
2

Indeed, relative to a coordinate system, dvg = gdx0 dx1 . . . dxn Thus, the above
equality follows from,
g = gg g ,
with g the determinant
of g . On the other hand, writing R = g R and using


d
= g R + g R
. Therefore,
g(s)  = g , we calculate, R
the formula ds
s=0
Z
Z

d
1

dvg
SG (s)  = (R g R)g dvg +
g R
(59)
ds
2
s=0
U
U
we make use of the following Lemma,
To calculate R
Lemma 1.2. Let g (s) be a family of space-time metrics with g(0) = g and
d
d

ds g(0) = g. Set also, ds R (s)|s=0 = R . Then,

= D D
R

where is the tensor,

1

(D g + D g D g )
= g
2
Proof : Since both sides of the identity are tensors it suffices to prove the formula
at a point p relative to a particular system of coordinates for which the Christoffel
symbols vanish at p. Relative to such a coordinate system the Ricci tensor has

the form R =
.
can be written as a space-time diReturning to (59) we find that since g R
vergence of a tensor compactly supported in U the corresponding integral vanishes
identically. We therefore infer that,
Z

d

SG (s)  =
E g dvg
(60)
ds
s=0
U
7up to an additive constant

28

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

where E = R 12 g R. We now consider the variation of the action integral

SM with respect to the metric. As remarked before LM depends on the metric g
through its inverse g . Therefore if we denote SM (s) = SM [, g(s) ; U] we have,
writing dv g = 12 g g dvg ,


d

SM (s) 
ds
s=0

Definition.

Z
LM

g
dv
+
LM dv g
g

U g
U
Z
1
LM
= ( g LM )g dvg
2
U g

The symmetric tensor,



1
LM
g LM
T =
g
2

is called the energy-momentum tensor of the action SM .

With this definition we write,
Z

d

SM (s)  =
T g dvg
ds
s=0
U

(61)

Finally, combining 60 with 61, we derive for the total action S,

Z

d

S(s)  = (E T )g dvg
ds
s=0
U
Since g is an arbitrary symmetric 2-tensor compactly supported in U we derive
the Einstein field equation,
E = T
Recall that the Einstein tensor E satisfies the twice contracted Bianchi identity,
D E = 0
This implies that the energy-momentum tensor T is also divergenceless,

D T = 0

(62)

which is the concise, space-time expression for the law of conservation of energymomentum of the matter-fields.

2. The energy-momentum tensor

The conservation law (62) is a fundamental property of a matterfield. We now turn
to a more direct derivation.

2. THE ENERGY-MOMENTUM TENSOR

29

We consider an arbitrary Lagrangian field theory with stationary solution . Let

s be the one-parameter group of local diffeomorphisms generated by a given vectorfield X. We shall use the flow to vary the fields according to
gs

(s ) g

(s ) .

From the invariance of the action integral under diffeomorphisms,

S(s) = S[s , gs ; M] = SM [, g; M].
So that
Z
Z

d
S

0 =
S(s)  =
dvg +
T g dvg
(63)
ds
s=0
M
M
The first term is clearly zero, being a stationary solution. In the second term,
which represents variations with respect to the metric, we have


d

(gs )  = LX g = D X + D X
ds
s=0

Therefore
Z
0

T LX g dvg = 2

=
M

T D X dvg = 2
M

D T X dvg

As X was arbitrary, we conclude

D T = 0.

(64)

This is again the law of conservation of energy-momentum.

We list below the energy-momentum tensors of the field theories discussed before.
We leave it to the reader to carry out the calculations using the definition.
(1) The energy-momentum for the scalar field equation is,


1
1

T =
, , g (g , , + 2V ())
2
2
(2) The energy-momentum for wave maps is given by,


1
1

T =
< , , , > g (g < , , , >)
2
2
where < , > denotes the Riemannian inner product on the target manifold.
(3) The energy-momentum tensor for the Maxwell equations is,
1
T = F F g (F F )
4
(4) The energy-momentum tensor for the Yang-Mills equations is,
1
T =< F , F > g (< F , F >)
4
An acceptable notion of the energy-momentum tensor T must satisfy the following
properties in addition of the conservation law (64),

30

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

(1) T is symmetric
(2) T satisfies the positive energy condition that is, T(X, Y ) 0 , for any
future directed time-like vectors X, Y .
The symmetry property is automatic in our construction. The following proposition
asserts that the energy-momentum tensors of the field theories described above
satisfy the positive energy condition.
Proposition 2.1. The energy-momentum tensor of the scalar wave equation satisfies the positive energy condition if V is positive. The energy- momentum tensors
for the wave maps, Maxwell equations and Yang-Mills satisfy the positive energy
condition.

Proof : To prove the positivity conditions consider two vectors X, Y , at some

point p M, which are both causal future oriented. The plane spanned by X, Y
intersects the null cone at p along two null directions8. Let L, L be the two future
directed null vectors corresponding to the two complementary null directions and
normalized by the condition
< L, L >= 2
i.e. they form a null pair. Since the vectorfields X, Y are linear combinations
with positive coefficients of L, L, the proposition will follow from showing that
T(L, L) 0, T(L, L) 0 and T(L, L) 0. To show this we consider a frame at
p formed by the vectorfields E(n+1) = L, E(n) = L and E(1) , . . . , E(n1) with the
properties,
< E(i) , E(n) >=< E(i) , E(n+1) >= 0
and
< E(i) , E(j) >= ij
for all i, j = 1, . . . , n 1. A frame with these properties is called a null frame.
(1) We now calculate, in the case of the wave equation,
1
T(L, L) =
E()2
2
1
T(L, L) =
L()2 .
2
which are clearly non-negative. Now,
1
T(L, L) = [L()L() + (g , , + 2V ())]
2
and we aim to express g , , relative to our null frame. To do this,
observe that relative to the null frame the only nonvanishing components
of the metric g are,
gn(n+1) = 2

gii = 1 i = 1, . . . , n 1

and those of the inverse metric g are

1
g n(n+1) =
, g ii = 1 i = 1, . . . , n 1
2
8If X, Y are linearly dependent any plane passing through their common direction will do.

2. THE ENERGY-MOMENTUM TENSOR

31

Therefore,
g , , = L()L() + |
/ |2
where
|
/ |2 = (E(1) ())2 + (E(2) ())2 + . . . E(n1) ()2 .
Therefore,
1
|
/ |2 + V ().
2
according to the same calculation.
1
< E(), E() >
2
1
< E(), E() >
2
n1
1X
< E(i) (), E(i) () > .
2 i=1

T(L, L) =
(2) For wave maps we have,
T (E, E)

T (E, E)

T (E, E)

The positivity of T is then a consequence of the Riemannian metric h on

the target manifold N .
(3) To show positivity for the energy momentum tensor of the Maxwell equations in 3 + 1 dimensions we first write the tensor in the more symmetric
form
1
T = (F F + ?F ?F )
(65)
2
where ?F is the Hodge dual of F , i.e. ?F = 12 F .
Exercise. Check formula (65).
We introduce the following null decomposition of F at every point
p M,
A = FA4
,
A = FA3
1
1
,
= ?F34 .
= F34
2
2
which completely determines the tensor F . Here the indices A = 1, 2
correspond to the directions E1 , E2 tangent to the sphere while the indices
3, 4 correspond to E3 = L and E4 = L. We then calculate that for ?F ,
?

FA4 = ?A =
?

F34 = 2

FA3 = ?A

??

F34 = 2

where A =AB B . Here AB is the volume form on the unit sphere,

hence AB = 12 AB34 , i.e. 11 =22 = 0, 12 = 21 = 1. With this
notation we calculate,

2 
1X
1
T (E(4) , E(4) ) =
F4A F4A + ?F4A ?F4A
2
4
A=1

2
1X
(A A + ?A ?A )
2
A=1

2
X
A=1

A A = ||2 0.

32

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

Similarly,
T (E(3) , E(3) ) =

2
X

A A = ||2 0

A=1

and in the same vein we find

T (E, E) = 2 + 2 0
which proves our assertion.
(4) The positivity of the energy-momentum tensor of the Yang- Mills equations is proved in precisely the same manner as for the Maxwell equations,
using the positivity of the Killing scalar product < , >G .

Another important property which the energy momentum tensor of a field theory
may satisfy is the trace free condition, that is
g T = 0.
It turns out that this condition is satisfied by all field theories which are conformally
invariant.
Definition. A field theory is said to be conformally invariant if the corresponding
action integral is invariant under conformal transformations of the metric
= g
g g
a positive smooth function on the space-time.
Proposition 2.2. The energy momentum tensor T of a conformally invariant field
theory is traceless.
Proof : Consider an arbitrary smooth function f compactly supported in U M.
Consider the following variation of a given metric g,
g (s) = esf g .
Let S(s) = SU [, g(s)]. In view of the covariance of S we have S(s) = S(0). Hence,
Z
d
0=
S(s)|s=0 =
T g dvg
ds
U
where
g =
Hence,

R
U



d
g (s)
= f g .
ds
s=0

(T g ) f dvg = 0 and since f is arbitrary we infer that,

trT = g T 0.

3. CONSERVATION LAWS

33

We can easily check that the Maxwell and the Yang-Mills equations are conformally
invariant in 3 1-dimensions. The wave maps field theory is conformally invariant
in dimension 1 + 1, i.e. if the space-time M is two-dimensional9.

R
Remark: The action integral of the Maxwell equations, S = U F F dvg is
conformally invariant in any dimension provided that we also scale the electro = 2 g then dvg = n+1 dvg and if we also set
magnetic field F . Indeed if g
n3

F = 2 F we get
Z
F , g
] =
g
dvg
S[
F F g
Z
F F g g dvg
=
=

S[F, g].

We finish this section with a simple observation concerning conformal field theories
in 1+1 dimensions. We specialize in fact to the Minkowski space R1+1 and consider
the local conservation law, T = 0. Setting = 0, 1 we derive
0 T00 + 1 T01 = 0,

0 T01 + 1 T11 = 0

(66)

Since the energy-momentum tensor is trace-free, we get T00 = T11 := A. Set

T01 = T10 = B. Therefore (66) implies that both A and B satisfy the linear
homogeneous wave equation;
2A

= 0 =

2B.

(67)

Using this observation it is is easy to prove that smooth initial data remain smooth
for all time.
For example, wave maps are conformally invariant in dimension 1 + 1. In this case
A = T00 =

1
(< t , t > + < x , x >) ,
2

Given data in C0 (R), (67) implies that the derivatives of remain smooth for all
positive times. This proves global existence.

3. Conservation Laws
The energy-momentum tensor of a field theory is intimately connected with conservations laws. This connection is seen through Noethers principle,

Noethers Principle: To any one-parameter group of transformations preserving

the action there corresponds a conservation law.
9Similarly for the linear scalar wave equation

34

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

We illustrate this fundamental principle as follows: Let S = S[, g] be the action

integral of the fields . Let t be a 1-parameter group of isometries of M, i.e.,
(t ) g = g. Then
S[(t ) , g]

S[(t ) , (t ) g]

= S[, g].
Thus the action is preserved under (t ) . In view of Noethers Principle we
ought to find a conservation law for the corresponding Euler-Lagrange equations10.
We derive these laws using the Killing vectorfield X which generates t .
We begin with a general calculation involving the energy-momentum tensor T of
and an arbitrary vectorfield X. P the one-form obtained by contracting T with X.
P = T X
Since T is symmetric and divergence-free
D P
where

(X)

1
(D T ) X + T D X = T
2

(X)

is the deformation tensor of X.

(X)

Notation.

(LX g) = D X + D X

We denote the backward light cone with vertex p = (t, x

) Rn+1 by

N (t, x
) = {(t, x) 0 t t; |x x
| = t t}.

The restriction of this set to some time interval [t1 , t2 ], t1 t2 t, will be written
N[t1 ,t2 ] (t, x
). These null hypersurfaces are null boundaries of,

| t t}
J 1 (t, x
) = {(t, x)  0 t t; |x x


J
(t, x
) = {(t, x) t2 t t1 ; |x x
| t t}
[t2 ,t1 ]

We shall denote by St = St (t, x

) and Bt = Bt (t, x
) the intersection of the time slice

t with N , respectively J .
At each point q = (t, x) along N (p) , we define the null pair (E+ , E ) of future
oriented null vectors
L = E+

= t +

xi x
i
i ,
|x x
|

L = E = t

xi x
i
i
|x x
|

Observe that both L, L are null and < L, L >= 2.

The following is a simple consequence of Stokes theorem, in the following form.
10 The same argument holds for conformal isometries acting on a conformally invariant field
theory. We therefore also expect conservation laws in such a setting.

3. CONSERVATION LAWS

35

Proposition 3.1. Let P be a one-form satisfying P = F. Then11, for all

t1 t2 t,
Z
Z
Z
Z
hP, t i +
hP, E i =
hP, t i
F dtdx (68)

N[t

Bt2

1 ,t2 ]

(p)

J[t
(p)
,t ]

Bt1

where,
Z

t2

hP, E i =

N[t

1 ,t2 ]

(p)

Z
hP, E i dat .

dt
t1

St

Applying this proposition to Stokes theorem to (68 ) we get

Theorem 3.2. Let T be the energy-momentum tensor associated to a field theory
and X an arbitrary vector field. Then

Z
T(t , X) +

Bt2

N[t

1 ,t2 ]

T(E , X)

(p)

T(t , X)

(69)

B t1

J[t
(p)
1 ,t2 ]

T (X) dtdx

In the particular case when X is Killing, its deformation tensor vanishes identically. Thus,
Corollary 3.3. If X is a killing vectorfield,
Z
Z
Z
T(t , X) +
T(L, X) =
Bt2

N[t

1 ,t2 ]

(p)

T(t , X)

(70)

Bt1

Moreover (70) remains valid if T is traceless and X is conformal Killing.

The identity (70) is usually applied to time-like future-oriented Killing vectorfields
X in which case the positive energy condition for T insures that all integrands in
(70) will be positive. We know that (see appendix 4.2, up to a Lorentz transformation the only Killing, future oriented timelike vectorfield is a constant multiple of
t . Choosing X = t (70) becomes,
Z
Z
Z
T(t , t ) +
T(E , t ) =
T(t , t )
(71)
Bt2

N[t

1 ,t2 ]

(p)

Bt1

In the case of a conformal field theory we can pick X to be the future timelike,
conformal Killing vectorfield X = K0 = (t2 + |x|2 )t + 2txi i . Thus,
Z
Z
Z
T(t , K0 ) +
T(L, K0 ) =
T(t , K0 )
(72)
Bt2

N[t

1 ,t2 ]

(p)

B t1

In (71) the term T(t , t ) is called energy density while T(E , t ) is called energy
flux density. The corresponding integrals are called energy contained in Bt1 , and
11The brackets h, i in (68) denote inner product with respect to the Minkowski metric.

36

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

Bt2 and, respectively, flux of energy through N . The coresponding terms in (72)
are called conformal energy densities, fluxes etc.
Equation (71) can be used to derive the following fundamental properties of relativistic field theories.
(1) Finite propagation speed
(2) Uniqueness of the Cauchy problem
R
The first property follows from the fact that, if Bt T(t , t ) is zero at
1
R
R
time t = t1 then both integrals Bt T(t , t ) and N
T(E , t ) must vanish
Proof :

[t1 ,t2 ]

also. In view of the positivity properties of the T it follows that the corresponding
integrands must also vanish. Taking into account the specific form of T, in a
particular theory, one can then show that the fields do also vanish in the domain
of influence of the ball Bt1 . Conversely, if the initial data for the fields vanish in
the complement of Bt1 , the the fields are identically zero in the complement of the
domain of influence of of Bt1 .
The proof of the second property follows immediately from the first for a linear
field theory. For a nonlinear theory one has to work a little more.
Exercise 1. Formulate an initial value problem for each of the field theories we
have encountered so far, scalar wave equation (SWE), Wave Maps (WM), Maxwell
equations (ME) and Yang-Mills (YM). Proof uniqueness of solutions to the initial
value problem, for smooth solutions.
The following is another important consequence of (71) and (72). To state the
results we introduce the following quantities,
Z
E(t) =
T (t , t ) (t, x)dx
(73)
n
ZR
Ec (t) =
T (K0 , t ) (t, x)dx
(74)
Rn

Theorem 3.4 (Global Energy). For an arbitrary field theory, if E(0) < , then
E(t) = E(0)

(75)

Moreover, for a conformal field theory, if Ec (0) < ,

Ec (t) = Ec (0)

(76)

Proof : Follows easily by applying (71) and (72) to past causal domains J (p)
with p = (t, 0) between t1 = 0 and t2 = t and letting t +.
Exercise 2. Consider the Lagrangian,
1
L = m V (||)
2

3. CONSERVATION LAWS

37

where is a complex valued scalar defined on the Minkowski space-time Rn+1 ,

its complex conjugate. As noted before L is invariant under the continuous group
of transformations ei with R. According to Noethers theorem the
corresponding Euler-Lagrange equation should have a conservation law. Can you
derive it ?

3.5. Energy dispersion in the conformal invariant case. In this section

we shall make use of the global conformal energy identity (76) to show how energy
dissipates for a filed theories in Minkowski space. Consider a conformal field theory
defined on all of Rn+1 . At each point of Rn+1 , with t 0, define the standard null
frame where
L = E+

= t + r

L = E

t r .

Observe that the conformal Killing vectorfield K0 = (t2 + r2 )t + 2rtr can be

expressed in the form,
K0 =


1
(t + r)2 E+ + (t r)2 E .
2

Thus,
Z
Ec (t)

=
Rn

1
1
(t + r)2 T++ + (t r)2 T + (t + r)2 + (t r)2 T+ dx.
4
4
|
{z
}
2(t2 +r 2 )

Z
=
Rn

Z
Ec (0)

1
1
1
(t + r)2 T++ + (t2 + r2 )T+ + (t r)2 T dx
4
2
4
Z
T(t , K0 )(0, x)dx =
|x|2 T(t , t )dx

Rn

(77)

Rn

According to (76) we have Ec (t) = Ec (0). Assuming that Ec (0) =

is finite we conclude that,
Z
Ec (0)
T++ (t, )dx .
t2
n
ZR
Ec (0)
T+ (t, )dx .
.
t2
n
R

R
Rn

|x|2 T(t , t )dx

The remaining term in (77) contains the factor (t r)2 which is constant along
outgoing null directions r = t + c. Hence for any 0 <  < 1
Z
T = O(t2 )
|x|>(1+)t
Z
T = O(t2 ).
|x|<(1)t

We conclude that most of the energy of a conformal field is carried by the T

component and propagates near the light cone.

38

1. EQUATIONS DERIVED BY THE VARIATIONAL PRINCIPLE

3.6. The case of  = 0. The wave equation  = 0 is only conformal

invariant in dimension n = 1. However we can still derive useful conservation laws
corresponding to conformal Killing vectorfields in any dimension.
Lemma 3.7. Let T = T [] the corresponding energy momentum tensor to a
solution of  = 0. Let X be a conformal Killing vectorfield, i.e. = (X) =
LX m = m, and tr = m . It is easy to check that  = 0; in fact, in the
particular case of X = K0 , = 4(n + 1)t. Let
n1
n1
P = T X +
tr
(tr)2 .
4(n + 1)
8(n + 1)
We have,
P = 0.

Now consider the null pair L = t + r , L = t r as in the previous section. We

easily check,

Q(L, L)

= L()2

(78)

Q(L, L)

(79)

(80)

Q(L, L)
P

= |
/ |

= L()

where |
/ | = A |eA ()| with (eA )A=1,... ,n1 an orthonormal frame spanning the
orthogonal complement of L, L.

CHAPTER 2

General Equations

It is tempting to define PDE as the subject which is concerned with all partial
differential equations, just as Algebraic Geometry, say, deals with all polynomial
equations. According to this view, the goal of the subject is to find a general theory
of all, or very general classes of PDEs. Though this point of view is quite out of
fashion, it has nevertheless important merits which I hope to illustrate below. To
see the full power of the general theory we need to, at least, write down general
equations, yet will make sure to explain the main ideas in simplified cases. We
consider equations, or systems of equations, in Rd with respect to the variables

x = (x1 , x2 , . . . xd ). As before we denote by i = x

i the partial derivatives relative

to the coordinate xi and by = 1 1 2 2 d d the mixed partial derivatives

corresponding to a multi-index = (1 , 2 , . . . d ) Nd . We denote by k the
vector of all partial derivatives with || = 1 + + d = k. Finally we denote
by k u = (u, u, . . . k u) the set of all partial derivatives of order less or equal to
k. In most interesting examples k is one or two.
Example. To make these notations more transparent consider the case of R2 and

2
coordinates x1 , x2 . For the multi-index = (2, 0) we have u = x
1 x1 u = 1 u

while for = (1, 1) we have u = x1 x2 u = 1 2 u. Also

u,
u,
u) = (12 u, 1 2 u, 22 u)
x1 x1 x1 x2 x2 x2
and 2 u = (u, 1 u, 2 u, 12 u, 1 2 u, 22 u).
2u = (

With this notation the Laplace operator in Rd has the form = 12 + 22 + . . . d2

while the DAlembertian in the Minkowski space Rd+1 has the form  = t2 +12 +
. . . + d2 . To make sense of an equation in which there appear partial derivatives
of order up to k we need to work with functions which are k-time differentiable
at every point. It is convenient to work with the class C k of functions which are
continuous and whose all partial derivatives u of order || k are continuous.
Definition.

A general partial differential equation in Rd of order k is of the form,

F x, k u(x) = 0,
(81)

where F is a specified function. We also consider N N systems1 in which case

F and u are column N -vectors. A function u of class C k is said to be a classical2
solution (81) if it verifies the equation as all points x in a specified domain of Rd .
1 That is determined systems of N equations for N unknowns.
2We call it classical to distinguish from generalized solutions to be discussed in the following

sections.
39

40

2. GENERAL EQUATIONS

Consider first the one dimensional situation d = 1 in which case (81) becomes an
ordinary differential equation (ODE), or system of ODE. To simplify further take
k = 1 and N = 1, that is the case of an ordinary differential equation of order k = 1.
Then (81) is simply, F (x, u(x), x u(x)) = 0 where F is a given function of the three
variables x, u and p = x u such as, for example, F (x, u, p) = x p + u3 sinx. To
solve the equation (81) in this case is to find a function a C 1 function u(x) such
that
x x u(x) + u3 = sinx.

(82)

Now consider the case of a second order ODE, i.e. d = N = 1 and k = 2. Then
(81) becomes, F (x, u(x), x u(x), x2 u(x)) = 0, where F now depends on the four
variables x, u, p = x u, q = x2 u. As an example take F = q 2 +V 0 (u), for some given
function V = V (u), in which case (81) becomes the nonlinear harmonic oscillator
equation,
x2 u(x) + V 0 (u(x)) = 0

(83)

Passing to a system of ODE, with d = 1, k = 1 and N = 2 we will need a

vector function F = (F1 , F2 ) with both F1 and F2 depending on the five variables
x, u1 , u2 , p1 = x u1 , p2 = x u2 . Then (81) becomes,


F1 x, u1 (x), u2 (x), x u1 (x), x u2 (x) = 0


F2 x, u1 (x), u2 (x), x u1 (x), x u2 (x) = 0
The case of PDE gets a bit more complicated because of the large number of
variables involved in the definition of F . Thus for first order (k = 1) scalar equations
(N=1) in two space dimensions ( d = 2) we need functions F depending on the two
spatial variables x1 , x2 as well as u, p1 = 1 u and p2 = 2 u. For a given function
of five variables F = F (x, u, p), a general first order PDE in two space dimensions
takes the form,
F (x, u(x), 1 u(x), 2 u(x)) = 0.

(84)

As a particular example take F = p21 + p22 1. The corresponding equation is,

(1 u(x))2 + (2 u(x))2 = 1

(85)

which plays an important role in geometric optics. A classical solution of the

equation is a C 1 function u = u(x1 , x2 ) which verifies (85) at all points of a domain
D R2 . A similar example is that given by the Eikonal equation (149).
Remark 1. We have excluded from our definition over-determined (i.e. the number of equations exceeds that of unknowns) or underdetermined systems (i.e. the
number of equations is less than that of unknowns) despite their obvious interest
to Geometry and Physics. The Einstein vacuum equations (24), for example, look
underdetermined at first glance. They become determined once we fix a particular
coordinate condition, such as the wave coordinate condition alluded to in the introduction. Gauge field theories, such as Yang-Mills, have a similar structure. On
the other hand the equations defined by the De Rham complex on an open set of
Rn form an overdetermined system. For example, given a one form = i dxi, the
system df = is overdetermined and can only be solved, locally, if the exterior
derivative of vanishes i.e. d = 0.

2. GENERAL EQUATIONS

41

Remark 2. All higher order scalar equations or systems can in fact be re-expressed
as first order systems, i.e. k = 1, by simply introducing all higher order derivatives
of u as unknowns together with the obvious compatibility relations between partial
derivatives. As an example consider equation (83) and set v = x u. We can then
rewrite the equation as a first order system with N = 2, namely x v + V 0 (u) =
0, x u v = 0.
An equation, or system, is called quasi-linear if it is linear with respect to the
highest order derivatives. A quasilinear system of order one (k = 1) in Rd can be
written in the form,
d
X

Ai (x, u(x))i u = F (x, u(x))

(86)

i=1

Here u and F are column N vectors and the coefficients A1 , A2 , . . . Ad are N N

matrix valued functions.
The minimal surface equation is an example of a second order (k = 2) quasilinear
scalar equation (N = 1) in two space dimensions. Indeed, using the coordinates
x1 , x2 , instead of x, y, we can manipulate (6) with the help of Leibnitz formula and
rewrite in the form,
X
hij (u)i j u = 0,
(87)
i,j=1,2

with h11 (u) = 1 + (2 u)2 , h22 (u) = 1 + (1 u)2 , h12 (u) = h21 (u) = 1 u 2 u,
which is manifestly a second order quasi-linear equation.
In the particular case when the top order coefficients of a quasilinear equation,
i.e. those corresponding to the highest order derivatives, depend only on the space
variables x Rd , the equation, or system, is called semi-linear. For example,
equation (20) derived in connection to the uniformization theorem, is semi-linear.
A linear equation, or system, of order k can be written in the form,
X
A (x) u(x) = F (x).

(88)

||k

Observe that the differential operator on the left hand side is indeed linear in the
sense discussed in our introduction. If in addition the coefficients A are constant in
x, the system is called linear with constant coefficients. The five basic equations (1)
(5) discussed in the introduction are all linear with constant coefficients. Typically,
these are the only equations which can be solved explicitly.
We thus have our first useful, indeed very useful, classification of PDEs into fully
nonlinear, quasi-linear, semi-linear and linear. A fully nonlinear equation is nonlinear relative to the highest derivatives. The typical example is the Monge Ampere
equation. For simplicity consider the case of functions of 2 variables u(x1 , x2 )
in R2 with hessian 2 u = (i j u)i,j=1,2 . Clearly the determinant det( 2 u) =
(12 u) (22 u) (1 2 u)2 , is quadratic with respect to the second derivatives of

42

2. GENERAL EQUATIONS

u. Thus the Monge -Amp`ere equation,

det( 2 u) = f (x, u, u),

(89)

with f a given function defined on R2 RR2 , is fully nonlinear. This equation plays
an important role in Geometry, in relation to the isometric embedding problem as
well as to the problem of finding surfaces in R2 with prescribed Gauss curvature. A
variant of the Monge Amp`ere equation, for complex valued functions, plays a central
role in complex geometry in connection to Calabi -Yau manifolds. Calabi-Yau
manifolds, on the other hand, are central mathematical objects in String Theory.
Remark. Most of the basic equations of Physics, such as the Einstein equations,
are quasilinear. Fully nonlinear equations appear however in connection to the
theory of characteristics of linear PDE, which we discuss at length below, or in
geometry.
1. First order scalar equations
It does not make sense to give a systematic treatment of this classical topic since
there are many PDE books which do an excellent job, such as [E] or [J]. In what
follows I will only attempt to give the main ideas behind the theory. It turns out
that scalar (N = 1) first order (k = 1) PDE in d space dimensions can be reduced
to systems of first order ODE.
As a simple illustration of this important fact consider the following equation in
two space dimensions,
a1 (x1 , x2 )1 u(x1 , x2 ) + a2 (x1 , x2 )2 u(x1 , x2 ) = f (x1 , x2 )
1

(90)
2

where a , a , f are given real functions in the variables x = (x , x ) R . We

associate to (90) the first order 2 2 system
dx2
dx1
(s) = a1 (x1 (s), x2 (s)),
= a2 (x1 (s), x2 (s))
(91)
ds
ds
To simplify matters we assume f = 0. Observe that any solution u = u(x1 , x2 ) of
(90), with f = 0, is constant along any solution x(s) = (x1 (s), x2 (s)), i.e.


d
u x1 (s), x2 (s) = 0.
ds
Thus, in principle, the knowledge of solutions to (91), which are called characteristic
curves for (90), allows us to find all solutions to (90). I say in principle because,
in general, the nonlinear system (91) is not so easy to solve. Yet ODE are simpler
to deal with and the fundamental theorem of ODE, which we will discuss later in
this section, allows us to solve (91), at least locally for a small interval in s. The
constancy of u along characteristic curves allows us to obtain, even when we cannot
find explicit solutions, important qualitative information. For example, suppose
that the coefficients a1 , a2 are smooth (or real analytic) and that the initial data is
smooth (or real analytic) everywhere on H except at some point x0 H where it
is discontinuous. Then, clearly, as long as the trajectories of (91) are well defined
and distinct, the solution u remains smooth (or real analytic) at all points except
along the characteristic curve which initiates at x0 , i.e. along the solution to (91)

1. FIRST ORDER SCALAR EQUATIONS

43

which verifies the initial condition x(0) = x0 . The discontinuity at x0 propagates

precisely along . We see here the simplest manifestation of a general principle,
which we shall state later, that singularities of solutions to PDE propagate along
characteristics.
One can generalize equation (90) to allow the coefficients a1 , a2 and f to depend
not only on x = (x1 , x2 ) but also on u,
a1 (x, u(x))1 u(x) + a2 (x, u(x))2 u(x) = f (x, u(x))

(92)

The associated characteristic system becomes,

dx1
(s) = a1 x(s), u(s, x(s)) ,
ds
Then, as in the previous case,

dx2
= a2 x(s), u(s, x(s))
ds

(93)

d
u(x1 (s), x2 (s)) = f (x1 (s), x2 (s), u(x1 (s), x2 (s)))
(94)
ds
Unlike the previous case however (93) is undetermined; we need now to consider
the enlarged ODE system (93)-(94). where the unknowns are x1 (s), x2 (s), u(s) =
u(x1 (s), x2 (s)). As a special example of (92) consider the scalar equation in two
space dimensions,
t u + ux u = 0,
1

u(0, x) = u0 (x)
2

(95)
1

called the Burger equation. Since a = 1, a = u we can set x (s) = s, x (s) = x(s)
in (93) and thus derive its characteristic equation in the form,
dx
(s) = u(s, x(s)).
(96)
ds
Observe that, for any given solution u of (95) and any characteristic curve (s, x(s))
d
we have ds
u(s, x(s)) = 0. Thus, in principle, the knowledge of solutions to (96)
would allow us to determine the solutions to (95). This, however, seems circular
since u itself appears in (96). To see how this difficulty can be circumvented consider
the initial value problem for (95), i.e. look for solutions u which verify u(0, x) =
u0 (x). Consider an associated characteristic curve x(s) such that, initially, x(0) =
x0 . Then, since u is constant along the curve, we must have u(s, x(s)) = u0 (x0 ).
Hence, going back to (96), we infer that dx
ds = u0 (x0 ) and thus x(s) = x0 + su0 (x0 ).
We thus deduce that,
u(s, x0 + su0 (x0 )) = u0 (x0 )

(97)

which gives us, implicitly, the form of the solution u. We see once more, from (97),
that if the initial data is smooth (or real analytic) everywhere except at a point x0 ,
of the line t = 0, then the corresponding solution is also smooth (or real analytic)
everywhere, in a small neighborhood V of x0 , except along the characteristic curve
which initiates at x0 . The smallness of V is necessary here because new singularities
can form in the large. Observe indeed that u has to be constant along the lines
x + su0 (x) whose slopes depend on u0 (x). At a point when these lines cross, we
would obtain different values of u which is impossible unless u becomes singular
at that point. In fact one can show that the first derivative ux becomes infinite at
the first singular point, i.e. the singular point with the smallest value of |t|. This
blow-up phenomenon occur for any smooth, non-constant, initial data u0 .

44

2. GENERAL EQUATIONS

Remark. There is an important difference between the linear equation (90) and
quasi-linear equation (92). The characteristics of the first depend only on the
coefficients a1 (x), a2 (x) while the characteristics of the second depend, explicitely,
on a particular solution u of the equation. In both cases, singularities can only
propagate along the characteristic curves of the equation. For nonlinear equations,
however, new singularities can form in the large, independent of the smoothness of
the data.
The above procedure extends to fully nonlinear scalar equations in Rd of the form,
t u + H(x, u) = 0,

u(0, x) = u0 (x)

(98)

with H = H(x, p) a given function of the variables x = (x1 , x2 , . . . , xd ) and p =

(p1 , p2 , . . . pd ), called the Hamiltonian of the system, and u = (1 u, 2 u, . . . , d u).
We associate to (98) the ODE system, with i = 1, 2 . . . , d,

dxi
=
H(x(t), p(t)),
dt
pi

dpi

= i H(x(t), p(t)).
dt
x

(99)

The equation (98) is called a Hamilton-Jacobi equation while (99) is known as a

Hamiltonian system of ODE. The relationship between them is a little more involved
than in the previous cases discussed above. To simplify the calculations below we
assume d = 1, so that H = H(x, p) is only a function of two variables. Let u be a
solution of (98). Differentiating (98) in x and applying the chain rule we derive,
t x u + p H(x, x u)x2 u = x H(x, x u)

(100)

dx
dt

Now take x(t) a solution of the equation

= p H(x(t), x u(x(t)) and set p(t) :=
x u(t, x(t)). Then, by using first the chain rule and then equation (100) we derive,
dp
dt

= x t u(t, x(t)) + x2 u(t, x(t))p H(x(t), p(t))

= x H(x(t), x u(t, x(t))) = x H(x(t), p(t))

Hence x(t), p(t) verify the Hamilton equation

dx
= p H(x(t), p(t)),
dt

dp
= x H(x(t), p(t)).
dt

d
On the other hand, dt
u(t, x(t)) = t u(t, x(t)) + x u(t, x(t))p H(x(t), p(t)), and,
using equation (98), t u(t, x(t)) = H(x(t), x u(t, x(t)) = H(x(t), p(t)). Thus,

d
u(t, x(t)) = H(x(t), p(t)) + p(t)p H(x(t), p(t)),
dt
from which we see, in principle, how to construct u based only on the knowledge
of the solutions x(t), p(t), called the bicharacteristic curves of the nonlinear PDE.
Once more singularities can only propagate along bichararcteristics. As in the case
of the Burger equation singularities will occur, for essentially, all smooth data;
thus a classical, i.e. continuously differentiable, solution can only be constructed
locally in time. Both Hamilton-Jacobi equation and hamiltonian systems play a
fundamental role in Classical Mechanics as well as in the theory of propagation
of singularities in linear PDE. The deep connection between hamiltonian systems
and first oder Hamilton-Jacobi equations have played an important role in the
introduction of the Schr
odinger equation in quantum mechanics.

2.

INITIAL VALUE PROBLEM FOR ODE

45

2. Initial Value Problem for ODE

To go further with our general presentation we need to discuss the initial value
problem. For simplicity let us start with a first order ODE
x u(x) = f (x, u(x))

(101)

subject to the initial condition

u(x0 ) = u0

(102)

The reader may assume, for simplicity, that (101) is a scalar equation and that f
is a nice function of x and u, such as f (x, u) = u3 u + 1 + sinx. Observe that the
knowledge of the initial data u0 allows us to determine x u(x0 ). Differentiating the
equation (101) with respect to x and applying the chain rule, we derive,
x2 u(x) = x f (x, u(x)) + u f (x, u(x))x u(x) = cos x + 3u2 (x)x u(x) x u(x)
Hence, x2 u(x0 ) = x f (x0 , u0 ) + u f (x0 , u0 )x u0 and since x u(x0 ) has already
been determined we infer that x2 u(x0 ) can be explicitely calculated from the initial data u0 . The calculation also involves the function f as well as its first
partial derivatives. Taking higher derivatives of the equation (101) we can recursively determine x3 u(x0 ), as well as all other higher derivatives of u at x0 .
One can
u(x) with the help of the Taylor series
Pthan,1 ink principle, determine
1 2
x u(x0 )(xx0 )k = u(x0 )+x u(x0 )(xx0 )+ 2!
x (x0 )(xx0 )2 +. . . .
u(x) = k0 k!
We say in principle because there is no guarantee that the series converge. There is
however a very important theorem, called the Cauchy-Kowalewski theorem, which
asserts that, if the function f is real analytic, which is certainly the case for our
f (x, u) = u3 u + 1 + sinx, then there exists a neighborhood J of x0 where the
Taylor series converge to a real analytic solution u of the equation. One can the
easily show that the solution such obtained is the unique solution to (101) subject
to the initial condition (102).
The same result may not hold true if we consider a more general equation of the
form,
a(x, u(x))x u = f (x, u(x)),

u(x0 ) = u0

(103)

Indeed the recursive argument outlined above breaks down in the case of the scalar
equation (xx0 )x u = f (x, u) for the simple reason that we cannot even determine
x u(x0 ) from the initial condition u(x0 ) = u0 . A similar problem occurs for the
equation (u u0 )x u = f (x, u). An obvious condition which allows us to extend
our previous recursive argument to (103) is that a(x0 , u0 ) 6= 0. Otherwise we say
that the initial value problem (103) is characteristic. If both a and f are also
real analytic the Cauchy-Kowalewski theorem applies and we obtain a unique, real
analytic, solution of (103) in a small neighborhood of x0 . In the case of a N N
system,
A(x, u(x))x u = F (x, u(x)),

u(x0 ) = u0

(104)

A = A(x, u) is N N matrix and the non-characteristic condition becomes

det A(x0 , u0 ) 6= 0.

(105)

46

2. GENERAL EQUATIONS

It turns out, and this is extremely important, that while the non-degeneracy condition (105) is essential to obtain a unique solution of the equation, the analyticity
condition is not at all important, in the case of ODE. It can be replaced by a simple
local Lipschitz condition for A and F , i.e. it suffices to assume, for example, that
only their first partial derivatives exist and that they are merely locally bounded.
This is always the case if the first derivatives of A, F are continuous.
The following local existence and uniqueness (LEU) theorem is called the fundamental theorem of ODE.
Theorem[Fundamental theorem for ODE] If the matrix A(x0 , u0 ) is invertible
and if A, F are continuous and have locally bounded first derivatives then there
exists a time interval x0 J R and a unique solution3 u defined on J verifying
the initial conditions u(x0 ) = u0 .
Proof The proof of the theorem is based on the Picard iteration method. The idea
is to construct a sequence of approximate solutions u(n) (x) which converge to the
desired solution. Without loss of generality we can assume A to be the identity
matrix4. One starts by setting u(0) (x) = u0 and then defines recursively,
x u(n) (x) = F (x, u(n1) (x)),

u(n1) (x0 ) = u0

(106)

Observe that at every stage we only need to solve a very simple linear problem,
which makes Picard iteration easy to implement numerically. As we shall see below,
variations of this method are also used for solving nonlinear PDE.
...... To fill in the proof.....
Remark. The local existence theorem is sharp, in general. Indeed we have seen
that the invertibility condition for A(x0 , u0 ) is necessary. Also, in general, the
interval of existence J may not be extended to the whole real line. As an example
consider the nonlinear equation x u = u2 with initial data u = u0 at x = 0, for
u0
which the solution u = 1xu
becomes infinite in finite time, i.e. it blows-up.
0
Once the LEU result is established one can define the main goals of the mathematical theory of ODE to be:
(1) Find criteria for global existence. In case of blow-up describe the limiting
behavior.
(2) In case of global existence describe the asymptotic behavior of solutions
and family of solutions.
Though is impossible to develop a general theory, answering both goals (in practice
one is forced to restrict to special classes of equations motivated by applications),
the general LEU theorem mentioned above gives a powerful unifying theme. It
3Since we are not assuming analyticity for A, F the solution may not be analytic, but it has
continuous first derivatives.
4since A is invertible we can multiply both sides of the equation by the inverse matrix A1

3.

INITIAL VALUE PROBLEM FOR PDE

47

would be very helpful, really wonderful, if a similar situation were to hold for
general PDE.

3. Initial value problem for PDE

By analogy to the one dimensional situation it is natural to consider, instead of
points, hyper-surfaces H Rd on which to specify initial conditions for u. For a
general equation of order k, i.e. involving k derivatives, we would need to specify
the values of u and its first k 1 normal derivatives5 to H. For example in the case
of the second order wave equation (3) we need to specify the initial data for u and
t u. along the hypersurface t = 0. Without getting into details at this point we
can give the following general definition.
Definition. We say that an initial value problem, for a k-order quasilinear system, in which we specify, as data, the first k 1 normal derivatives of a solution u
along H, is non-characteristic at a point x0 of H, if we can formally determine all
other higher partial derivatives of u at x0 , uniquely, in terms of the data.
To understand the definition, which may seem too general at this point, consider
the much simpler case k = 1, N = 1. In this case we only need to specify the
restriction u|H = u0 of u to H. Our initial value problem takes the form,
d
X

ai (x, u(x))i u(x) = f (x, u(x)),

u|H = u0

(107)

i=1

with ai , f real valued functions of x Rd and u R. To simplify further take

d = 2, i.e. we have the equation in x = (x1 , x2 ),
a1 (x, u(x))1 u(x) + a2 (x, u(x))2 u(x) = f (x, u(x))

(108)

we have encountered earlier in (92). Consider a curve H in R , parametrized by

1
dx2
x1 = x1 (s), x2 = x2 (s) whose tangent vector V (s) = ( dx
ds , ds ) is non-degenerate,
1
dx2 2 1/2
2
i.e. |V (s)| = (| dx
6= 0. It has a well defined unit normal N (s) =
ds | + | ds | )
(n1 (s), n2 (s)), which verifies the conditions,
N (s) V (s) = 0,
1

N (s) N (s) = 1

Observe that the coefficients a , a in (108) can be completely determined, along

H, from the knowledge of the initial condition u0 = u0 (s). Consider the
first
derivatives (1 u, 2 u) evaluated along H, i.e. U (s) = 1 u(x(s)), 2 u(x(s) . At
every point along H our equation reads,
A(s) U (s) = f (s),
(109)




where A(s) = a1 (x(s), u0 (s)), a2 (x(s), u0 (s) and f (s) = f x(s), u0 (s) are completely determined by the data u0 (s). Differentiating u(x(s)) = u0 (s) with respect
to s we infer that,
d
U (s) V (s) = U0 (s),
U0 (s) =
u0 (s).
ds
5These are derivatives in the direction of the normal to H.

48

2. GENERAL EQUATIONS

To fully determine U (s) it remains to determine its projection on the normal vector
N (s), i.e. U (s) N (s). Indeed, since V (x) and N (x) span R2 , at all points x =
(x1 (s), x2 (s)) along our curve, we have
U (s) = (U V )(s)

V (s)
+ (U N )(s)N (s)
|V (s)|2

(110)

Therefore, from the equation (109),

f (s) = A(s) U (s) = (U (s) V (s))

A(s) V (s)
+ (U (s) N (s)) A(s) N (s)
|V (s)|2

from which we can determine U (s) N (s) provided that,

A(s) N (s) 6= 0.

(111)

If, on the other hand, A(s) N (s) = 0 then, since V (s) N (s) = 0, we infer that the
dx
vectors A(s) and V (s) = dx
ds must be proportional, i.e. ds = (s)A(s). One can
then reparametrize the curve H, i.e. introduce another parameter s0 = s0 (s) with
ds0
ds = (s), such that relative to the new parameter we have = 1. This leads to
the equation,


dx1
= a1 x(s), u(x(s) ,
ds

dx2
= a2 x(s), u(x(s))
ds

which is precisely the characteristic system (93). Thus,

Along a characteristic curve, the equation (108) is degenerate, that is we cannot
determine the first order derivatives of u uniquely in terms of the data u0 . On the
other hand the non-degenerate condition,
A(s0 ) N (s0 ) 6= 0,

i.e.

a1 (x0 , u(x0 ))n1 (x0 ) + a2 (x0 , u(x0 ))n2 (x0 ) 6= 0

(112)

at some point x0 = x(s0 ) H, allows us to determine all higher derivatives of u at

x0 , uniquely in terms of the data u0 .
Indeed, if the condition A(s0 ) N (s0 ) 6= 0 is satisfied at x0 = x(s0 ) H we have
seen already how to determine the first derivatives 1 u, 2 u at that point. Once we
have these it is not difficult to determine all higher derivatives of u. For example,
observe, by differentiating equation (108) with respect to x1 , that the function
v = 1 u verifies an equation of the form,
a1 (x, u(x))1 v(x) + a2 (x, u(x))2 v = g(x, u(x), v(x))
with a function g which can be easily determined from the coefficients a and f . We
can now proceed as before and determine the first derivatives of v i.e. 12 u, 2 1 u.
Thus, recursively, we can determine all partial derivatives of u of any order.
We can easily extend the discussion above to the higher dimensional case (107).
Given a hypersurface H in Rd , with unit normal N = (n1 , n2 , . . . nd ), we find that
H is non-characteristic at x0 for the initial value problem (107) if,
n
X
i=1

ai (x0 , u0 (x0 ))ni (x0 ) 6= 0

(113)

3.

INITIAL VALUE PROBLEM FOR PDE

49

With a little more work we can extend our discussion to general higher order quasilinear equations, or systems and get a simple, sufficient condition, for a Cauchy
problem to be non-characteristic. Particularly important for us are second order
(k = 2) scalar equations (N = 1). To keep things simple consider the case of a
second order, semi-linear equation in Rd ,
d
X

aij (x)i j u = f (x, u(x), u(x))

(114)

i,j=1

and a hypersurface H in Rd defined by the equation (x) = 0 with non-vanishing

, or in
gradient . Define the unit normal at a point x0 H to be N = ||
i
components ni = ||
. As initial conditions for (114) we prescribe u and its normal
derivative N u(x) = n1 (x)1 u(x) + n2 (x)2 u(x) + . . . nd (x)d u(x) on H,

xH

u(x) = u0 (x), N u(x) = u1 (x),

(115)

We need to find a condition on H such that we can determine all higher derivatives
of a solution u, at x0 H, from the initial data u0 , u1 . We can proceed exactly in
the same manner as before, and find that all second order derivatives of u can be
determined at a point x0 H, provided that,
d
X

aij (x0 )ni (x0 )nj (x0 ) 6= 0

(116)

i,j=1

It is indeed easy to see that the only second order derivative of u, which is not
automatically determined from u0 , u1 , is of the form N 2 u(x0 ) = N (N (u))(x0 ).
This latter can be determined from the equation (114), provided that (116) is
verified. One does this by decomposing all partial derivatives of u into tangential
and normal components, as we have done in (110). One can then show, recursively,
that all higher derivatives of u can also be determined. Thus, (116) is exactly the
non-characteristic condition we were looking for.
Pd
If, on the other hand, i,j=1 aij (x)ni (x)nj (x) = 0 at all points we call H a characteristic hypersurface for the equation (114). Since ni = |
we find that H is
i |
characteristic if and only if,
d
X

aij (x)i (x)j (x) = 0

(117)

i,j=1

Remark Observe that only the left hand side6 of (114), called 7, is relevant in
determining the characteristic surfaces of the equation.
Example 1.

Assume that the coefficients a of (114) verify,

d
X

aij (x)i j > 0,

Rd ,

x Rd

(118)

i,j=1

Then no surface in Rd can be characteristic. This is the case, in particular, for

the equation u = f . Consider also the minimal surfaces equation written in the
6containing second order derivatives
7principal part

50

2. GENERAL EQUATIONS

form (87). It is easy to check that, the quadratic form associated to the symmetric
matrix hij (u) is positive definite independent of u. Indeed,


hij (u)i j = (1 + |u|2 )1/2 ||2 (1 + |u|2 )1 ( u)2 > 0
Thus, even though (87) is not linear, we see that all surfaces in R2 are noncharacteristic.
Example 2. Consider the wave equation u = f in R1+d . All hypersurfaces of
the form (t, x) = 0 for which,
(t )2 =

d
X
(i )2 ,

(119)

i=1

are characteristic. This is the same eikonal equation which has appeared before in
(149). Observe that it splits into two Hamilton-Jacobi equations, see (98),
t =

d
X

(i )2 )1/2

(120)

i=1

The bicharacteristic curves of the associated Hamiltonians are called bicharacteristic curves of the wave equation. As particular solutions of (429) we find,
+ (t, x) = (t t0 ) + |x x0 | and (t, x) = (t t0 ) |x x0 | whose level surfaces
= 0 correspond to forward and backward light cones with vertex at p = (t0 , x0 ).
These represent, physically, the union of all light rays emanating from a point source
at p. The light rays are given by the equation (t t0 ) = (x x0 ), for R3
with || = 1, and are precisely the (t, x) components of the bicharacteristic curves
of the Hamilton-Jacobi equations (120).
More general, consider the linear wave equation,
g = 0.

(121)

where g is the inverse of a general Lorentz metric8 g . The characteristic surfaces

of (121) are given by the eikonal equation,
g = 0.

(122)

They are also called null hypersurfaces for the metric g.

Remark. In the particular case when g01 = . . . = g0n = 0, g00 < 0 and gij positive
definite, (121) takes the form,
X
a00 (t, x)t2 u +
aij (t, x)i j u = 0,
(123)
i,j
1
where with a00 = g00
and aij gjk = ki . Thus the characteristic equation has the
00
2
form a (t, x)(t ) + aij (x)i j = 0 or,
X

1/2
t = (a00 )1
aij (x)i j
.
(124)
i,j
8 i.e. g

= g and the associated quadratic form, g (p)X X is non-degenerate at every
point p and has signature (1, 1, . . . , 1).

3.

INITIAL VALUE PROBLEM FOR PDE

51

Thus, through any point p R1+n pass two distinct characteristic surfaces. The
same is true for the general case.
The bicharacteristics of the corresponding hamiltonian systems are called bicharacteristic curves of (123).
Remark. In the case of the first order scalar equations (90) we have seen how the
knowledge of characteristics can be used to find, implicitly, the general solutions.
We have shown, in particular, that singularities propagate only along characteristics. In the case of second order equations the situation is more complicated. the
characteristics are typically9 not sufficient to solve the equations, but they continue to provide important information, such as propagation of singularities. For
example, in the case of the wave equation u = 0 with smooth initial data u0 , u1
everywhere except at a point p = (t0 , x0 ), the solution u has singularities present
at all points of the light cone (t t0 )2 + |x x0 |2 = 0 with vertex at p. A more
refined version of this fact shows that the singularities propagate along bicharacteristics. The general principle here is that singularities propagate along characteristic
hypersurfaces of a PDE. Since this is a very important principle it pays to give it
a more precise formulation which extends to general boundary conditions, such as
the Dirichlet condition for (??).
Propagation of singularities10. If the boundary conditions, or the coefficients
of a linear PDE with smooth (or real analytic) coefficients are singular at some
point p, and smooth ( or real analytic) away from p in some small neighborhood
V , then a solution of the equation may only be singular in V along a characteristic
hypersurface passing through p. If there are no such characteristic hypersurfaces,
any solution of the equation must be smooth ( or real analytic) in V \ {p}.
Remark 1. The principle as stated is far too general, it can be proved only if
specific assumptions are made on the symbol of the operator. It should be viewed
however as something one might expect for a reasonable equation.
Remark 2. The principle can be extended, under specific minimum regularity
assumptions on solutions, to the nonlinear case. It is however invalid in the large.
Indeed, as we have shown in in the case of the Burger equation, solutions to nonlinear evolution equations, can develop new singularities independent of the smoothness of the initial conditions. Global versions of the principle can be formulated
for linear equations, based on the bicharacteristics of the equation, see remark 3
below.
Remark 3. According to the principle it follows that any solution of the equation
u = f , verifying the boundary condition u|D = u0 , with a boundary value u0
which is merely continuous, has to be smooth everywhere in the interior of D
9 Characteristics enter however in the explicit form of the fundamental solution for the
standard wave equation. This was made particularly obvious in the derivation starting with the
ansatz (148). The also play a major role to construct approximate solutions for wave equations
with variable coefficients, such as (123)
10A more precise version of the principle relates propagation of singularities to bicharacteristics curves.

52

2. GENERAL EQUATIONS

provided that f itself is smooth there. Moreover the solution is real analytic, if f
is real analytic.
Remark 4. More precise versions of this principle, which plays a fundamental role
in the general theory, can be given for linear equations. In the case of the general
wave equation (123), for example, one can show that singularities propagate along
bicharacteristics. These are the bicharacteristic curves associated to the HamiltonJacobi equation (124).

4. Cauchy-Kowalevsky Theorem
In the case of ODE we have seen that a non-characteristic initial value problem
admits always local in time solutions. Is there also a higher dimensional analogue
of this fact ? The answer is yes provided that we restrict ourselves to an extension
of the Cauchy -Kowalewsky theorem. More precisely one can consider general
quasilinear equations, or systems, with real analytic coefficients, real analytic hypersurfaces H, and real analytic initial data on H.
Theorem[Cauchy-Kowalevsky (CK)] If all the real analyticity conditions made
above are satisfied and if H is non-characteristic at x0 11, there exists locally, in a
neighborhood of x0 , a unique real analytic solution u(x) verifying the system and
the corresponding initial conditions.
The CK theorem validates thePmost straightforward attempts to find solutions

by formal expansions u(x) =

C (x x0 ) with constants C which can be
determined recursively, by simply algebraic formulas, from the equation and initial
conditions on H, using only the non-characteristic condition and the analyticity
assumptions. Indeed the theorem insures that the naive expansion obtained in this
way converges in a small neighborhood of x0 H.
Proof See [E] or [J]
In the special case of linear equations (88) an important companion theorem, due
to Holmgren, asserts that the analytic solution given by the CK theorem is unique
in the class of all smooth solutions and smooth non-characteristic hypersurfaces H.
Theorem 4.1 (Holmgren uniqueness theorem). Consider the initial value problem
for a linear equations of the type (88) with analytic coefficients. If the the hypersurface H is also analytic and non-characteristic at x0 H, then the corresponding
Cauchy problem is unique in the class of smooth solutions, in a small neighborhood
of x0 .
Proof See [J]
11In the case of second order equations of type (114) this is precisely condition (116).

4.

CAUCHY-KOWALEVSKY THEOREM

53

Remark. The remarkable thing about Holmgrens theorem is that it proves uniqueness even in cases where existence of solutions cannot be guaranteed. Thus, as we
shall see below, the Cauchy problem for the wave equation with data on the hyperplane x1 = 0 does not, in general, have solutions, yet Holmgrens theorem asserts
that if a solution exists it must be unique.
At first glance it may seem that the CK theorem is a perfect analogue of the fundamental theorem for ODEs. It turns out, however, that the analyticity conditions
required by the CK theorem are much too restrictive and thus the apparent generality of the result is misleading. A first limitation becomes immediately obvious when
we consider the wave equation u = 0 whose fundamental feature of finite speed
of propagation12 is impossible to make sense in the class of real analytic solutions.
A related problem, first pointed out by Hadamard, concerns the impossibility of
solving the Cauchy problem, in many important cases, for arbitrary smooth, non
analytic, data. Consider, for example, the Laplace equation u = 0 in Rd . As we
have established above, any hyper-surface H is non-characteristic, yet the Cauchy
problem u|H = u0 , N (u)|H = u1 , for arbitrary smooth initial conditions u0 , u1 may
admit no local solutions, in a neighborhood of any point of H. Indeed take H to
be the hyperplane x1 = 0 and assume that the Cauchy problem can be solved,
for a given, non analytic, smooth data in an domain which includes a closed ball
B centered at the origin. The corresponding solution can also be interpreted as
the solution to the Dirichlet problem in B, with the values of u prescribed on the
boundary B. But this, according to our heuristic principle13, must be real analytic
everywhere in the interior of B, contradicting our initial data assumptions.
On the other hand the Cauchy problem, for the wave equation u = 0 in Rd+1 ,
has a unique solution for any smooth initial data u0 , u1 , prescribed on a space-like
hyper-surface, that is a hypersurface (t, x) = 0 whose normal vector, at every
point p = (t0 , x0 ), is directed inside the interior of the future or past directed light
cone passing through that point. Formally this means,
|t (p)| >

d
X

|i (p)|2

1/2

(125)

i=1

The condition is clearly satisfied by the hypersurfaces of t = t0 , but any other

hypersurface close to it is also spacelike. On the other hand the IVP is ill posed,
i.e. not well posed, for a time-like hypersurface, i.e a hypersurface for which,
|t (p)| <

d
X

|i (p)|2

1/2

(126)

i=1

In this case we cannot, for general non real analytic initial conditions, find a solution
of the IVP. An example of a time-like hypersurface is given by the hyperplane
x1 = 0.
12Roughly this means that if a solution u is compactly supported at some value of t it must
be compactly supported at all later times. Analytic functions cannot be compactly supported
without vanishing identically.
13which can be easily made rigorous in this case

54

2. GENERAL EQUATIONS

Definition. A given problem for a PDE is said to be well posed if both existence
and uniqueness of solutions can be established for arbitrary data which belong to a
specified large space of functions, which includes the class of smooth functions14.
Moreover the solutions must depend continuously on the data.
The continuous dependence on the data is very important. Indeed solutions to the
IVP for a PDE would be of little use if very small changes of the initial conditions
will result, instantaneously, in very large changes in the corresponding solutions.
It is only in the class of smooth solutions that the theory of PDE becomes really
interesting, relevant and challenging. It means that we have to give up hope for
a all encompassing result and look instead for special classes of equations which
have common features, or really just on special important equations. It is in that
sense that the generality of the CK theorem is really an illusion. The true study
of partial differential equations only begins when we give up on analyticity. In the
next chapter we will analyze in detail the main analytic properties of the simplest
equations such as Cauchy-Riemann, Laplace, Heat and Wave equations using their
corresponding fundamental solutions. To do this we need fist to recall the theory
of distributions.

14Here we are necessarily vague. A precise space can be specified in each specific case.

CHAPTER 3

Distribution Theory

This is a very short summary of distribution theory, for more exposure to the
subject I suggest F.G. Friedlander and M. Joshis excellent book Introduction to
the Theory of Distributions, [Fr-Io]. Hormanders first volume of The Analysis of
Linear Partial Differential Operators, [Ho], in Springer can also be useful.
Notation. Throughout these notes we use the notation A . B to mean a cB
where c is a numerical constant, independent of A, B. When Rn is a set, we
may write (x ) to denote the indicator function of the set . For instance,
(5 |x| < 7) is a function equal to 1 for 5 x < 7 and 0 otherwise.

1. Introduction to Distribution Theory

A short description of the theory of distributions contains an unavoidable oxymoron:
It is an enabling theory which allows us to differentiate functions which are in no
way differentiable and manipulate them as if there were no problems whatsoever.
Its main application is to the theory of partial differential equations.
We begin by recounting how the notion of a fundamental solution in partial differential equations was born through classical electromagnetism. When charge does
not move, any charge distribution : R3 R gives rise to an electric field which
(up to a conventional sign and physical constant) is the gradient of a potential
function V : R3 R. The classical physical law relating V to the charge density
is Poissons equation
V =

(127)

P3
where = i=1 i2 denotes the Laplacian, and on physical grounds we may require
V (or at least its derivative) to vanish at infinity so that distant interactions are
weak.
As with any other field theory, the physical theory cannot be valid and complete
unless there exists a unique solution to the equation (127) (for reasonable data )
which depends continuously, in some sense, on the data. In addition to resolving
these issues, we seek at least a qualitative understanding of the behavior of the
solution. In the present case, thanks to a huge amount of symmetry, we will even
be able to derive an explicit formula, but forP
the heuristic analysis involved, it will
3
only be important that the operator = i=1 i2 is linear and commutes with
55

56

3. DISTRIBUTION THEORY

translations (i.e. it is a linear differential operator with constant coefficients). In

fact, the Laplacian is invariant under both translations and rotations in the sense
that (f T ) = (f ) T for all smooth functions whenever T is a rigid motion of
the Euclidean space.
The idea is to solve (127) first in the special case where the charge density (x)
is a unit charge completely concentrated at the point y R3 (we formally write
(x) = y (x) where the Dirac delta function corresponds to the density function of
a unit point mass at y). We will discuss the meaning of the Dirac delta function
later, but for the moment let us accept the formal definition of it as an operator
whose action on continuous functions f is to produce the value of f at the point y
i.e. y is the measure,
y (f ) = f (y).
Thus, we look for a solution Vy (x) to the equation Vy (x) = y (x) (which is currently meaningless since y is not a function). By linearity of , we can then obtain
the general solution as a superposition of solutions from the point contributions

Z
V (x) =

Vy (x)(y)dy

(128)

Formally, we can even manage to solve the equation Vy (x) = y (x) for any fixed
y R3 . In view of the translation invariance of , we may assume that y = 0.
Since is rotationally invariant (see Exercise 1) and so is 0 , then any solution
V0 (x) = V0 (|x|) should also be rotationally invariant if solutions are to be unique.
We call V0 (x) a fundamental solution for the Laplace operator. Then, postulating
the existence and spherical symmetry of V0 (x), we obtain (using the divergence
theorem)

Z
1=

0 (x)dx
|x|R

Z
=

V0 (x)dx
|x|R

dV
(|x|)d(x)
|x|=R dr
dV
= 4R2
(R)
dr
=

1
We choose the only fundamental solution decaying at infinity, namely V0 (x) = 1
4 R .
1
Therefore, translating back to y , we find Vy (x) = 1
4 |xy| . One can see by direct
computation that Vy (x) = 0 away from y, and one can even prove that (128) does
indeed solve (127) for (say) smooth, compactly supported densities . Furthermore,
by taking the gradient of (128), one obtains the experimentally refutable conclusion

1.

INTRODUCTION TO DISTRIBUTION THEORY

57

that the electric field decays asymptotically as

total charge
(distance)2
far away from the charge source.
1
Definition. We call Vy (x) = 1
4 |xy| a fundamental solution solution for in
R3 . More generally, we define a fundamental solution for a linear operator L in
Rn (i.e. acting on functions in Rn ) is a generalized function Vy (x) such that
L(Vy ) = y .

Given a fundamental solution for L we can find solutions for the equation Lu = f ,
for any smooth, compactly supported f by setting, formally (never mind, for the
moment, that the integration may make no sense),
Z
u(x) = Vy (x)f (y)dy
Exercise 1. Show, informally, that if L commutes with translations in the sense
that (Lf )( + y) = L(f ( + y)) for all translations x 7 x + y then the fundamental
solution also commutes with translations, in the sense that Vy (x) = V (x y) with
V verifying L(V ) = 0 .
Once a fundamental solution Vy of an operator L has been found, we need to make
sense of it as a generalized function as well as of the formal integration above. This is
precisely what the theory of distributions accomplishes. Distribution theory allows
us to make heuristic calculations rigorous and, even more importantly, enables us
to deal with singular objects as if they were regular functions. There are, of course,
limits to this new freedom which a good theory should spell out.
Exercise 2.
It is not difficult to show that, for R C0 (R3 ), the potential
R
(y)dy
1
1
R3
(y)dy behaves near infinity like 1
+ o(|x|2 ) away
V (x) = 4 R3 |xy|
4
|x|
from the support of . One way toR prove this asymptotic and understand the error
1 d
1
1
1
is to Taylor expand |xsy|
= |x|
+ 0 ds
|xsy| ds (the idea being that the parameter
y is relatively small).
When the
R charge distribution is centered at the origin (that is, the vector-valued
integral R3 y(y)dy = 0), show the more precise result (with explicit remainder)
that, as |x| ,
R
1 R3 (y)dy
V (x) =
+ O(|x|3 )
4
|x|
It may help keep computations
R 1 simple00 to apply the precise, first order Taylor expan1
0
.
sion (1) = (0) + (0) + 0 (1 s) (s)ds to the auxiliary function (s) = |xsy|
Also, a convenient way to differentiate the absolute value function is to observe
that |x sy|2 =< x sy, x sy > where <, > denotes the Euclidean inner product.
R
Remark:
IfR the total
is not 0, then one can find a center of
R charge (y)dy
R
charge yc = ydy/ dy so that (y yc )(y)dy = 0. In this situation, we

58

3. DISTRIBUTION THEORY

could Taylor expand about y = yc to see that the associated potential behaves
asymptotically as though it were centered at yc :
R
1 R3 (y)dy
V (x)
.
4 |x yc |
R
Notice, however, that when the charge cancels out in the sense that (y)dy = 0,
C
the associated potential function V decays more rapidly at infinity as |x|
2 . This
phenomenon of increased decay for localized, oscillatory data is not only physically
important for explaining why electric forces are weak over distances when charge
cancels, but it is also important in analysis where a similar cancelation arises in
many other naturally occurring situations. We will see this sort of cancelation being
used in a critical way later in the notes.
Exercise 3. The reasoning in the previous section can be extended to solve
for the potential inside of a bounded region whose boundary is grounded. That is,
consider the problem V (x) = (x) for x in a bounded domain with V = 0 on
the boundary.
In principle, how could you construct a general solution of the form
R
V (x) = K(x, y)(y)dy ? Where does linearity come in?
Exercise 4. Suppose that a unit of negative charge has been distributed uniformly
over the sphere of radius R1 in R3 , and that a unit of positive charge has been
distributed uniformly on the sphere of radius R2 . Find the electrostatic potential
function V associated to this charge configuration .
Exercise 5. a. Use the informal argument from the introduction to find the
fundamental solution Kn (x) of in Rn for every n 2; i.e. solve Kn (x) = 0 (x)
with an explicit formula for Kn .
b. Discuss the behavior as |x| of the corresponding solution
Z
V (x) = Kn (x y)(y)dy
for compactly supported. Namely, as |x| , what is the main term and how
large is the error?

2. Test Functions. Distributions

We start with some standard notation. We denote vectors in Rn by x = (x1 , . . . , xn )
and set x = (x1 , . . . , xn ), x + y = (x1 + y1 , . . . , xn + yn ). We denote by x y
1
the standard scalar product and by |x| = (x x) 2 the Euclidean length of x. Given
a function f : C we denote by supp(f ) the closure in of the set where
f (x) 6= 0. We denote by C k () the set of complex valued functions on which
are k times continuously differentiable and by C0k () the subset of those which are
also compactly supported. We also denote by C () = kN C k () the space of
infinitely differentiable functions, and by C0 () the subset of those which also have
compact support. The latter plays a particularly important role in the theory of
distributions; it is called the space of test functions on .

2. TEST FUNCTIONS. DISTRIBUTIONS

59

f
Let Rn and f C (). We denote by i f the partial derivative x
, i =
i
1, . . . , n. For derivatives of higher order we use the standard multi-index notation.
A multi-index is an n-tuple = (1 , . . . , n ) of nonnegative integers with length
|| = 1 + +n . Set + = (1 +1 , . . . , n +n ). We denote by ! the product
of factorials 1 ! n !. Now set f = 11 nn f . Clearly + f = f .
Given two smooth functions u, v we have the Leibniz formula,

(u v) =

X
+=

!
u v.
!!

Taylors formula, around the origin, for a smooth function f : Rn C can be

written as follows,
f (x) =

X 1
f (0)x + O(|x|k+1 )
!

as

x 0.

||k

n
1
Here x denotes the monomial x = x
1 xn .

We start by explaining a general method (often called mollification) which can

be used to approximate rougher functions by smooth ones. Essentially, one takes
the function f to be approximated, and replaces f by its average after randomly
translating f according to some smooth probability measure with small support.
It is intuitively clear and easy to prove that the randomly perturbed function is
a smooth approximation to the original (imagine a sharply formed sandpile after
a small earthquake), and in order to get a better approximation one shrinks the
support of to 0. The technical implementation of this method appears as follows.
Proposition 2.1. Let f C0k (Rn ), 0 k < . Let C0 (Rn ) be a smooth
contained in the unit ball B(0, 1) = {|x| 1} and
Rfunction with support supp() n
(x)dx = 1. We set  (x) =  (x/) and let
f (x) = f  (x) = n

Z
f (y)(

xy
)dy =


Z
f (x z)(z)dz.

We have:
(1) The functions f are in C0 (Rn ) and supp(f ) supp(f ) + B(0, ).
(2) We have f f uniformly as  0.

Proof : The first part of the proposition follows immediately from the definition
since the statement about supports is immediate and, by integration by parts, we
can transfer all derivatives of f on the smooth part of the integrand  . To prove
the second statement we simply write,

f (x) f (x) =

Z 


f (x z) f (x) (z)dz.

60

3. DISTRIBUTION THEORY

Therefore, for || k,
| f (x) f (x)|

| f (x z) f (x)||(z)|dz
|(z)|dz sup | f (x z) f (x)|
|z|1

sup | f (x z) f (x)|

|z|1

The proof follows now easily in view of the uniform continuity of the functions f .

As a corollary of the Proposition, one can easily check that the space of test functions C0 () is dense in the spaces C k () as well as Lp (),
R 1 p < . Of course,
one must first exhibit at least one such C0 (Rn ) with dx = 1. Some multiple
1

of the bump function (x) = e 1|x|2 (|x| < 1) will do. Another way to construct
an example is by starting with any C 1 bump function and taking advantage of the
smoothing effects of random translations (as in the above proposition) but keeping the support under control to obtain a smooth bump function as a limit of an
iterative process.
Definition 2.2. A distribution u D0 () is a linear functional u : C0 () C
verifying the following property:
For any compact set K there exists an integer N and a constant C = CK,N
such that for all C0 (), with supp() K we have
X
| < u, > | C
sup | |.
||N

If the same integer N can be used in the above definition for every K, then the
smallest such N is called the order of the distribution. For example, the Riesz
Representation theorem (characterizing the dual of C(X) for compact Hausdorff
spaces) guarantess that distributions of order 0 are Borel measures.
Equivalently, a distribution u is a linear functional u : C0 () C which is continuous with respect to some topology defined on C0 (). This topology turns out
to be a rather unorthodox one (non-metrizable1, locally convex) but never mind all
this; we can go quite far without worrying in the least about the precise definition.
All we need to know is that in this topology a sequence j converges to 0 in C0 ()
if all the supports of j are included in a compact subset of and, for each multiindex , j 0 in the uniform norm. With this definition in mind we have the
following very useful characterization of distributions:
1This topology can be constructed as an inductive limit topology of Fr
echet spaces CK , where
K is compact and CK is the space of all smooth functions supported in K, endowed with
a Fr
echet space structure by the seminorms 7 supK | | for all multi-indices . We do not,
however, need the precise definition.

2. TEST FUNCTIONS. DISTRIBUTIONS

61

Proposition 2.3. A linear form u : C0 () C is a distribution in D0 () iff

limj u(j ) = 0 for every sequence of test functions j which converges to 0, in
C0 (), as j .
Proof :

This proof can be found in Friedlander, section 1.3, Theorem 1.3.2.

Example 1:

Any locally integrable function f L1loc () defines a distribution,

Z
< f, >= f ,
C0 ().

We can thus identify L1loc () as a subspace of D0 (). This is true in particular for
the space C () L1loc ().
R
One often uses the formal notation < u, >= u(x)(x)dx even when u D0 ()
is not a locally integrable function, and even when is not technically a test function. This notation can be conceptually simpler, but keep in mind that this is in no
way a genuine Lebesgue integral. One can, however, typically interpret this formal
integration as a limit of classical integrals.
Example 2:
by

The Dirac measure with mass 1 supported at x0 Rn is defined

< x0 , >= (x0 ).

Remark: We shall also often denote the action of a distribution u on a test

function by u() instead of < u, >. Thus x0 () = (x0 ).
Definition 2.4. A sequence of distributions uj D0 () is said to converge, weakly,
to a distribution u D0 () if, uj () u() for all C0 ().
For example the sequenceR um = eimx converges weakly to 0 in D0 (R) as m .
Also if f L1 (Rn ), with Rn f (x)dx = 1, the family of functions f (x) = n f (x)
converges weakly to 0 as .
We will be able to show that any distribution is the weak limit of a sequence
un C0 (). Due to this fact, many operations defined initially for functions
extend by continuity in a unique, natural way to D0 ().
Exercise. Given a compact set K Rn and a positive distance > 0, construct
a smooth function : Rn R whose support is contained in {x | |y x|
for some y K} within a distance of K, and such that (y) = 1 for y in some
neighborhood of K. Hint: start with a rough cutoff and leave some wiggle room.

2.5. Operations with distributions. The advantage of working with the

space of distributions is that while this space is much larger than the space of
smooth functions most important operations on test functions can be carried over
to distributions.

62

3. DISTRIBUTION THEORY

1. Multiplication by smooth functions:

Given u D0 () and f C ()
we define,
< f u, >=< u, f >,
C0 ()
in order to be consistent with the identity when u is a function
Z
Z
(f u)dx = u(f )dx.
It is easily verified that multiplication by a smooth function is a continuous endomorphism of the space of distributions.
2. Convolution with a test-function: Consider, u D0 (Rn ), C0 (Rn ).
Generalizing the convolution of two functions in a natural way, we define
u(x) =< uy , (x y) >,
the subscript specifying that u is understood to be acting on functions of the variable
y. Observe that the definition coincides with the usual one if u is a locally integrable
function, u L1loc (Rn ), for which
Z
u (x) = u(y)(x y)dy.
Remark: The convolution of a distribution and a test function is not merely
another distribution. Rather, observe that for every distribution u D0 (Rn ) and
C0 (Rn ) we have that u C (Rn ) is in fact a smooth function. For example,
if ek denotes a standard unit vector, then we can differentiate in the direction ek
as follows:
u(x + hek ) u(x)
= h1 < uy , (x + hek y) (x y) >
h
Z 1
= < uy ,
k (x + thek y)dt > .
0

Now, since x K is restricted to some compact setRK Rn , then for every sequence
1
hi 0, the associated sequence of functions y 7 0 k (x + thi ek y)dt, together
with all its derivatives, converge uniformly toward k (x y) and its corresponding
derivatives. Moreover they are all compactly supported with supports contained in
some compact set K 0 . Therefore,
u(x + hek ) u(x)
= uk (x).
lim
h0
h
and thus u has continuous partial derivatives. We can continue in this manner
and conclude that in fact u C (Rn ).
3. Differentiation of distributions: For every distribution u D0 () we define
< u, >= (1)|| < u, > .
We make this definition to be consistent with the integration by parts formula for
functions
Z
Z
i u(x)(x)dx = u(x)(i (x))dx, C0 (Rn ),
which may be proven, for example, be considering difference quotients.

3. EXAMPLES OF DISTRIBUTIONS ON THE REAL LINE

63

Again, it is easily verified that we have thus defined a continuous endomorphism

of the space of distributions. Of course, the operations above are the only possible
extensions of the usual operations on smooth functions. The minus sign can be
viewed dually as a differentiation of the measure = (x)dx. If we temporarily let
v denote the operation translate by v, then
he1
h0
h

1 = lim

has a density function 1 (x) and the limit can be taken in the topology of C0 .
In this dual sense, we have 1 u = limh0 u(x+heh1 )u(x) in the weak topology, which
often enables us to differentiate under the integral sign provided we interpret all
integrals in the distribution-theoretic sense.
We can now define the action of a general linear partial differential operator on
distributions. Indeed let,
X
P (x, ) =
a ,
a C (),
||m

be such an operator. Then,

< P (x, )u, >=< u, P (x, ) >,
where P (x, ) is the formal adjoint operator,
X
P (x, ) v =
(1)|| (a v).
||m
0

Observe that if uj D () converges weakly to u D0 () then P (x, )uj converges

weakly to P (x, )u.
Exercise. Show that for all u D0 () there exists a sequence uj C0 () such
that uj u as j in the sense of distributions( weak convergence). Thus
C0 () is dense in D0 (), with respect to the weak topology of the latter.
3. Examples of distributions on the real line

1.) The simplest nontrivial distribution is the Dirac delta function 0 = 0 (x),
defined by < 0 (x), >= (0). We will sometimes write (x) without a subscript
to indicate the point mass at the origin on R.
2.) Another simple example is the Heaviside function H(x) equal to 1 for x 0
and zero for x < 0. Or, using the standard identification between locally integrable
functions and distributions,
Z
< H(x), >=
(x)dx.
0
0

Observe that H (x) = (x) and that H(x) =

tion function of 0 .

Rx

(t)dt is the cumulative distribu-

64

3. DISTRIBUTION THEORY

3.) A more elaborate example is pv( x1 ), or simply x1 , called the principal value
distribution,
Z 
Z


1
1
1
(x)dx +
(x)dx .
< , >= lim
0
x
x
x

Observe that log |x| is locally integrable and thus a distribution by the standard
d
identification. One can show easily that dx
log |x| = pv( x1 ). Note that pv( x1 ) is an
odd distribution (it is orthogonal to even test functions), and is of order 1 even
though it is of order 0 away from the origin. In fact, decomposing = ev + odd
into even and odd parts, we have
1
< pv( ), >=
x

odd
dx =
x

Z Z

0 (tx)dt dx

1
. We also remark that the function |x|
(x 6= 0), in contrast, does not admit an
extension as a distribution to the whole line.

Exercise 1.
Show that the distribution t d (t) on the line is equal to (t),
dt
which is a nonzero distribution. This may seem counterintuitive since either t or
0 (t) seems to vanish at every point.
Exercise 2. Let, for z C with 0 < arg(z) < , log z = log |z| + iarg(z). We can
+
regard x log z = log(x + iy) as a family of distributions

depending on y R .
For x 6= 0 we have limy0+ log z = log |x| + i 1 H(x) . Show that as y 0 in
1
R+ , x log z converges weakly to a distribution x+i0
and,
1
= x1 i0 (x).
x + i0

(129)

Exercise 3. If is open and connected, u D0 (), and all derivatives i u = 0

in the sense of distributions, then u is a constant.
Exercise 4. Any non-negative distribution (i.e. < u, > 0 when 0) is in
fact a Borel measure. By the Riesz representation theorem for measures, it suffices
to prove that for every compact set K Rn there is a constant C = CK such that
|u()| CK max ||
for all with support in K.
Exercise 5. (using the preceding exercises) If u : (, b) R is a nondecreasing
function which (for simplicity)
vanishes at , there exists a unique Borel measure
Rx
0 so that u(x) = 1d(t) for Lebesgue almost every x (, b]. In terms
of , when is u continuous? Absolutely continuous?
Remark: The classical result that monotonic functions are almost everywhere
differentiable can be derived from the above exercise in combination with some
basic measure theory and the Lebesgue differentiation theorem of section (2.6).

3. EXAMPLES OF DISTRIBUTIONS ON THE REAL LINE

65

Exercise 6. Characterize convex functions f : (a, b) R. Namely, show that the

following are equivalent:
(1) f (x + (1 )y) f (x) + (1 )f (y) for all [0, 1] and x, y (a, b)
(2) f is continuous; f 0 (in the distribution-theoretic sense) is a non-decreasing
function, and is therefore locally bounded
(3) f is continuous; f 00 is a non-negative distribution, and is therefore a finite
measure when restricted to any bounded set.
Hint: The class of convex functions remains invariant under the operation of random translation, therefore mollification may help.
2
We now define an important family of distributions z+ , with z C, by analytic
continuation. We will see this family again later while studying the fundamental
solution to the wave equation, and again in our study of restriction theorems for
the Fourier transform.
First recall the definition of the Gamma function,
Definition 3.1. For Re(z) > 0 we define
Z
(z) =
et tz1 dt

(130)

as well as the Beta function,

Z

B(a, b) =

sa1 (1 s)b1 ds

(131)

Clearly (a + 1) = a(a) and (1) = 1. Thus (n + 1) = n!. Recall that the

following identity holds:
B(a, b) =

(a) (b)
(a + b)

(132)

We also record for future applications,

(a)(1 a) = B(a, 1 a) =

sin(a)

(133)

In particular (1/2) = 1/2 .

Exercise.
3.4.

Prove formulas (132) and (133). For help see Hormander, [?] section

Definition 3.2. For Re(a) > 0, we denote by ja () the locally integrable function
which is identically zero for < 0 and
1 a1
ja () =

,
> 0.
(134)
(a)
The following proposition is well known,

66

3. DISTRIBUTION THEORY

Proposition 3.3. For all a, b, Re(a), Re(b) > 0,

ja jb = ja+b
Proof :

We have,
ja jb ()

=
=
=

1
1
(a) (b)

a1 ( )b1 d

Z 1
1
1 a+b1
sa1 (1 s)b1 ds

(a) (b)
0
B(a, b) a+b1
1

a+b1 = ja+b ()
=
(a) (b)
(a + b)

Proposition 3.4. There exists a family of distribution ja , defined for all a C,

which coincide with the functions ja for Re(a) > 0, such that, ja jb = ja+b ,
d
d ja () = ja1 () and j0 = 0 , the Dirac delta function at the origin. Moreover
for all positive integers m, jm (x) = xm 0 (x).
d
Proof : The proof is based on the observation that d
ja () = ja1 (). Thus, for
a test function ,
Z
Z
ja1 ()()d = ja ()0 ()d
R

Based on this observation we define, for every a C such that Re(a) + m > 0 as
distribution
Z
< ja , >= (1)m
ja+m () (m) ()d
0

In particular,
Z
< j0 , >=

j1 () ()d =
0

0 ()d = (0)

Hence j0 = 0 . It is also easy to see that ja jb = ja+b for all a, b C.

Remark: In applications one often sees the family of distributions a+ = ja+1 .
a+b+1
b1
d b
Clearly a+ b+ = +
, d
+ () = +
(), and 1
+ = 0 . Observe also that
k

k+ () = k! ( > 0) for positive integers k, and more generally a+ is homogeneous

of degree a, i.e. , a+ (t) = ta a+ (), for any positive constant t. This homogeneity
clearly makes sense for Re(a) > 1 when a+ is a function. Can you also make
sense of it for all a C ?
3.5. Support of a distribution. The support of a distribution can be easily
derived as follows:
Definition 3.6. For u D0 (), we define the complement of the support of u,
\supp(u) = {x | Vx 3 x open, such that < u, >= 0 C0 (Vx )}.

3. EXAMPLES OF DISTRIBUTIONS ON THE REAL LINE

67

Lemma 3.7. If u D0 () and is a test function with supp() \supp(u), then

< u, >= u() = 0.
Proof : This follows easily by a partition of unity argument. The argument can
be found in Friedlander, section 1.4.
The above lemma may be used to show that any distribution u D0 () of compact
support extends to test functions in C (Rn ) by taking an arbitrary cutoff
C0 (Rn ) equal to 1 on the support of u and defining u() = u(). In fact, the
following proposition shows that we may regard a compactly supported distribution
as an element of the dual to C N (Rn ) for some N .
Proposition 3.8. A distribution u D0 (Rn ) has compact support K Rn iff there
exists N N such that , C0 (Rn ) we have
X
|u()| CU sup
| (x)|,
xU

||N

where U is an arbitrary open neighborhood of K.

Proof : This is seen by using a cutoff function which is identically 1 on the support
of the distribution.
Remark: If we endow C (Rn ) with the Frechet topology induced by the family of
seminorms given by supKi | |, with Nn and Ki running over a countable
collection of compact sets exhausting Rn , then the space of compactly supported
distributions can be identified with C (Rn ) , i.e. the space dual to C (Rn ).
We have the following useful fact (essentially dual to Taylor expansion) concerning
the structure of distributions supported at one point. We will find this result useful
at various parts of the notes, although its application can essentially always be
replaced by repeating some variant of its proof. We will not give all the details,
but the main ideas are present.
Proposition
3.9. Let u D0 (Rn ) and assume that supp(u) {0}. Then we have
P
u = ||N a (0 ), for some integer N , complex numbers a and 0 the Dirac
measure in Rn supported at 0.
Proof Let u be a distribution supported at the origin, C0 (Rn ) be a test
function. It is not true in general that < u, > depends only on the value of at
0, but it is true that < u, > depends only on the restriction of to any small
neighborhood of 0. So let C0 (Rn ) be a cutoff function which is equal to 1 on
a neighborhood of the origin, and let (x) = ( x ) for > 0 be a cutoff function
with an even smaller support. Then < u, >=< u, >. We wish to prove that
< u, > depends only on the first N derivatives of , where
X
| < u, > | C||||C N = C
max | |, C0 (Rn )
||N

68

3. DISTRIBUTION THEORY

By Taylor expansion, one reduces to the case where all derivatives a = 0 for
|| N . More precisely, we can write
Z 1
d
(x) = (0) +
(tx)dt
dt
0
XZ 1
xi i (tx)dt
= (0) +
= (0) +

(1 t)

(0)x +
0

d2
(tx)dt
dt2

and continues integrating by parts until one has written as a multinomial with
coefficients corresponding to derivatives of at 0 plus a sum of terms of the form
x , where is a multi-index with | | > N and the functions are smooth (but
obviously not all of compact support).
Expanding in this way, we need to show that < u, x >= 0. Here we will
use the estimate | < u, > | C||||C N , so we will have to estimate derivatives of
the type ( x ) with order || N . Observe that
x
x
(( )) = || ( )( )

This type of scaling with derivatives is consistent with dimensional analysis, if we

view and x to have the same units and view each differentiation i to have
the reciprocal units. It is also a computation that comes up extremely often in
analysis as we shall see later on. We take advantage of this scaling by absorbing
the monomial factor into the cutoff ( x )x = ( x ) . We then obtain
| < u, x > | = | < u, > |
X
C
max | (
)|
||N

One could expand these derivatives using the Leibniz formula

X !
(
) =

!!
+=

and generate a tremendous number of terms, but to find the exact formula for these
products may not be useful even though it might be worthwhile to go through the
details at least once. In practice, however, one avoids details (such as the exact
values of constants) which are not at the heart of the matter by understanding
what kinds of terms will occur, and in particular one isolates the worst terms. In
this case, the worst terms occur when a derivative falls upon the cutoff = ( x ),
which is becoming increasingly sharp as 0. For such terms, each derivative
generates a factor of 1 . However, at most N derivatives can hit this cutoff, and
so we have, for some number C 0 independent of (although potentially dependent
on and ),
| < u, x > | C 0 N
which tends to 0 as 0 since > N .

4. PULL BACK OF DISTRIBUTIONS

69

Now that we have introduced the notion of support, it is important to observe that
the convolution of two distributions cannot be defined in general, but only when
certain conditions on the support of the distributions are satisfied. We note in
particular the fact that if u1 , u2 D0 (Rn ) one of which is compactly supported,
then the convolution u1 u2 can be defined. Indeed, assuming u2 to be compactly
supported, we simply define, ( )
(u1 u2 ) = u1 (u2 ),

C0 (Rn ).

Here, supp(u2 ) {x + y : x supp(u2 ), y supp()}, which is a compact set.

This definition extends the classical convolution for functions.

4. Pull back of distributions

Let Rn and 0 Rm be open sets, u D0 (0 ) be a distribution on 0 , and
f : 0 . We can sometimes define the pull-back f u D0 (). There may be
obstructions; for example, if f is a constant function, then u f makes sense only
for continuous functions u, and more generally if f maps some set with positive
measure into a set of zero measure, then u f does not even make sense for u Lp ,
which are only well-defined as functions up to a set of measure zero. To ensure
that none of the obstructions mentioned above occur, we assume that f is smooth
and that its derivative matrix Df has full rank at every point so that, at least, any
open set maps onto another open set.
To consistently define the pull-back of u by f , when possible, we use duality by
regarding the pull-back of a function as the operation adjoint to the push-forward
of a measure and set
< f u, >=< u, f# >
where f# is (the density function of) the pushforward of the finite, complex measure dx by the map f .
We shall later prove that f# is a smooth function and hence that pull-back is welldefined. It is then obvious that f u is continuous in the distribution u with respect
to weak limits. An immediate consequence of this continuity is that the chain rule
for smooth functions u and f generalizes to the case where u is a distribution. For
example,
u(f (x)) = u0 (f (x))f (x)
as distributions.
The above definition is certainly consistent with the formalism of measure theory.
However, it is not immediately clear that the pushforward measure has a density
function which is a valid test function, nor is it clear how to compute using this
definition, so let us first discuss a few concrete examples explicitly.

70

3. DISTRIBUTION THEORY

Example 1: When f : 0 is a C diffeomorphism with inverse g, then we

apply the familiar change of variables formula for y = f (x),
Z
Z
u(f (x))(x)dx = u(f (x))(g(f (x)))dx
Z
= u(y)(g(y)) | det Dg(y)|dy
where Dg is the derivative matrix of g. This calculation motivates the definition of
pullback for such diffeomorphisms
< f u, >=< u, |detDg(y)| g >,

C0 ()

In this case we see that the change of variables formula is equivalent to the definition
of pullback.
Example 2: If f : R has a nonvanishing gradient, then we can explicitly
1
obtain the pullback of the delta function t , namely f (t ) = |f
| d. Here, d
denotes the canonical surface measure on the embedded hypersurface f 1 (t) =
{f (x) = t} Rn and f denotes the gradient of f .
In other words, we can compute the value at t of the pushforward measures density
function
Z
d(x)
,
f# (t) =
(x)
|f |
f 1 (t)
and therefore compute < f u, >=< u, f# > not only for a -function, but also
for arbitrary distributions u D0 (R). In this way, the pullback formula may be
written informally as a sort of decomposition
Z
u(f (x)) = u(t)(f (x) t)dt
R
which can be formally derived from the identity u(y) = u(t)(y t)dt.
As a sample application of this formula, one can see that the derivative of the
volume of the ball of radius R is the surface area of the sphere of radius R from
x
the fact that the gradient of |x| = |x|
has norm 1 and from the differentiation
Z
Z
d
H(r |x|)dx = (r |x|)dx.
dr
This formula is clear geometrically: when one compares the volume of a ball of
radius r to a slightly larger ball of radius r + , the change in volume is essentially
 times the surface area.
Since the pullback of a delta-function will be very important for us, let us give
a proof of this formula. Once we have proven this formula, we have built up the
theory enough to carry out the details of the previous calculation in full. One would
take difference quotients in the variable r of the distribution H(r ), and these
difference quotients are essentially supported on a thickened sphere. We then use
the trivial observation that the pullback of a distribution < f u, >=< u, f# >
is continuous in u with respect to weak limits.

4. PULL BACK OF DISTRIBUTIONS

71

Now let us prove the pullback formula for a -function. The geometric picture
in the proof is basically a generalization of the special case f (x) = |x| considered
above.
Proof of the Pullback Formula for m = 1
By taking a partition of unity to decompose if necessary, we may assume that f
may be completed to a coordinate system on the support of , since this is always
possible on a small neighborhood of any point in the support of by the nonvanishing of |f |, and since finitely many such neighborhoods suffice. We consider the
measure = (y)dy and let (t) : R C be the distribution function defined by
Z
(t) = f# (, t] =

(y) (f (y) t)dy.

We now wish to show that

1
(t) = lim
0 
0

Z
(t < f (y) t + ) (y)dy

exists at every point t R, from which it will follow2 that is absolutely continuous and that 0 (t) = f# (t) given by the formula is in fact the correct density
function. For simplicity of notation, let us suppose  > 0.
We now verify by change of coordinates that, very close to a point y0 f 1 (t) the
thickened hypersurface {y | t < f (y) t+} can be parameterized to have height

1
|f |(y0 )| + o() and width df (t) (y0 ), which is at least intuitively plausible
from a picture of the generic situation (for example, in the case of the preceding
example f (x) = |x|).
We may assume (without loss of generality) that
smooth function h(x1 , . . . , xn ) satisfying
f (h(x), x0 ) = x1

f
y 1 (y0 )

6= 0 and consider the

(135)

for all x = (x1 , x2 , . . . , xn ) = (x1 , x0 ) in a neighborhood of (t = f (y 0 ), y01 , . . . , y0n )

containing the support of . We then essentially use f in place of y 1 as a coordinate by making the coordinate transformation (y 1 , y 2 , . . . , y n ) = (h(x), x2 , . . . , xn ),
obtaining:

2The Fundamental Theorem of Calculus applies when is continuous and classically differentiable at every point t.

72

3. DISTRIBUTION THEORY

Z 


1
(t < f (y) t + ) (y) |dy 1 dy 2 . . . dy n |



Z 
1
= lim
(t < x1 t + ) (h(x), x0 ) |dh dx2 . . . dxn |



Z 
1
h
= lim
(t < x1 t + ) (h(x), x0 ) | 1 (x)||dx1 dx2 . . . dxn |


x

Z 


1
f
= lim
(t < x1 t + ) (/| 1 |) h(x1 , x0 ), x0 dx1 . . . dxn


y
Z
f
= (/| 1 |) (h(t, x0 ), x0 ) dx0
y

lim

To compute the Jacobian of the transformation in the second line, we have used
the shorthand of differential forms, which quickly encapsulates the fact that the
volume of an n-dimensional parallelopiped remains unchanged when one vertex is
translated within the span of others through the identity
h 1
dx ) dx2 . . . dxn
x1
n
X
h i
+(
dx ) dx2 . . . dxn
i
x
2

dh dx2 . . . dxn = (

h 1
dx . . . dxn
x1

h
and x
1 is computed through implicit differentiation of equation (135) which defines
h implicitly. It is now clear that 0 = f# is a smooth function of t.

The equation (135) also shows that, for t fixed, the function t (x0 ) = h(t, x0 )
parameterizes the hypersurface f 1 (t) as the graph of the function x1 = t (x0 )
when x0 varies. We now wish to interpret the integral over f 1 (t) in terms of the
surface measure
!1/2
n
X
t 2
0
df 1 (t) (x ) = 1 +
( j)
dx0
x
2
so we compute the surface density by implicitly differentiating f (t (x0 ), x0 ) = t to
obtain
f
f t

+ j = 0 j = 2, . . . , n
y 1 xj
y
Hence, we see that
|f | =
Substituting into

s
X  f 2
y i

f
= | 1|
y

n
X
t
1+
( j )2
x
2

!1/2

f
0
0
(/| y
1 |) t (x )dx gives the formula in Example 2 above.

4. PULL BACK OF DISTRIBUTIONS

73

The proof above could have been simplified by employing the same change of variables (y 1 , . . . , y n ) = (h(x), x2 , . . . , xn ) to show directly that
Z
Z
u(f (y))(y)dy = u(t)f# (t)dt
for all u C (R), or equivalently by using a smooth approximate delta-function in
place of the sharp approximate delta-function 1 (0 < t ). We have alternatively
chosen the above, lengthier proof for its intuitive, geometric appeal and also to
demonstrate the use of distribution functions for computing f# . A briefer and
more general proof is given in the Appendix along with some computational tools.
Exercise 1. Let S : Rn Rn be the dilation map S (x) = x. We say that a
distribution u D0 (Rn ) is homogeneous of degree a if, S u = a u. Show that the
definition coincides with the usual one if u is a function. Show that, in Rn , 0 is
homogeneous of degree n.
Exercise 2. Show that any distribution in Rn which is both homogeneous of
degree n and also supported at the origin is a constant multiple of 0 .
The examples above are special cases of a more general formula. We can compute
f# when f : Rm is a smooth map whose derivative is everywhere surjective
by the following explicit formula:

Z
(f# )(y) =

(x)
f 1 (y)

d(x)
||f ||(x)

(136)

Here d denotes the induced measure on the codimension m submanifold f 1 (y)

, f = df 1 . . . df m is the pullback of the volume form = dy 1 . . . dy m on
Rm , and || || denotes the norm induced by the pointwise inner product on m-forms.
|f |

The measure d can also be written ||f

|| where f is the Hodge dual of f .
These notions are all reviewed further in the Appendix (??), where the general
formula (136) is proven in a different manner than the m = 1 case proven above.
The proof consists of decomposing the volume form
dx1 . . . dxn =

1
(f f )
||f ||2

and then integrating first over the level sets of f .

There are more general conditions under which the operation of pullback f u is
possible when the singularities of the distribution u are understood in a more precise
manner.
Applications
Our first application of these operations will be to prove Gausss divergence theorem, which involves expressing the integral of some derivative of a function over
the interior of a set in terms of a boundary integral of . One can express the

74

3. DISTRIBUTION THEORY

integral of any derivative over as a limit of integrals of difference quotients of ,

but dually one can take the function fixed and take adjoint difference quotients
of the characteristic function of ; thus, the divergence theorem turns out to be
equivalent to the differentiation of the characteristic function of an open set.
We will now compute the gradient of the characteristic function of a domain
with a smooth boundary, but let us first do it in words. One can picture the
graph of as an -shaped table. If we perturb slightly to can obtain a smooth
approximation
, it is clear that the gradient of
points inside of and normal
to the boundary (in the direction of maximal increase), and furthermore the
gradient remains supported essentially on the boundary of . It is therefore no
surprise that we obtain
Proposition 4.1 (Gausss divergence theorem). For any test function C0 (Rn ),
Z
Z
Z

(x)dx = (x)(x)dx =
(x)~nd (x),

where ~n denotes the inward, unit normal vector.

P
Proof By taking a partition of unity, we may decompose =
where
each is a smooth function supported within a small region V so that V
= {x V : f (x) > 0} is an upper contour set of some defining function
f = f with nonvanishing gradient on the boundary. In V we have the equality
(as distributions), = H f , i.e.
Z
Z
(x)(x)dx = H(f (x))(x)dx
for any test function supported in V . Therefore, by the chain rule,
(x) = (f (x))f (x) =

f
(x)d (x) = ~nd (x)
|fa |

which proves the proposition.

We could have slightly modified the proof to allow for far less stringent regularity
conditions on f by first approximating the characteristic function of and taking
a limit (for example by letting a family of pre-Heaviside functions H converge to
the Heaviside function). For example, when the boundary can be expressed locally
as a graph of a Lipschitz function, then the normal vector is well-defined almost
everywhere, and we obtain in this way the same formula for a larger variety of sets
such as polygons, cubes, etc. The details can be found in Hormanders book [?].
Extending these ideas, we can outline a quick proof of the more general Stokes
theorem. The proof goes essentially as follows: for an oriented k-dimensional manifold Y (which might be embedded inside a higher dimensional manifold X), a k 1
form on X, and a test function C0 (X), we have
Z
Z
Z
0=
d() =
d +
d
Y

4. PULL BACK OF DISTRIBUTIONS

75

which is clear when has support in a coordinate patch . If M Y is an open

subset of Y with smooth boundary M , we take to approach the characteristic
function of M (so that d has support on M ), and in the limit we obtain Stokes
theorem:
Z
Z
d =

The extension of distribution theory to the setting of manifolds is mostly straightforward and is outlined in H
ormander.
The Appendix on integration over submanifolds included at the end of the notes
may help for some of the following Exercises.
Exercise 3. Show that if f , g are two smooth functions on Rn with non-vanishing
differential everywhere, then for all a, b Rn Rn :
Z
0 (f (a) x)0 (g(b) x)dx = 0 (f (a) g(b)).
Hint: Both sides are to be interpreted as distributions onRRn Rn . One could
re-write the definition of pull-back in the form u(g(b)) = u(x)0 (g(b) x)dx.
Approximating with approximate -functions, we can extend to the case u(x) =
0 (f (a) x). Alternatively, use the obvious special case where f (a) = a and
g(b) = b are both the identity map and pull back for general f and g.
Exercise 4.

A point is drawn at random from the punctured square

S = {(x, y) : max |x|, |y| < 1, }\{(0, 0)}

What is the probability density of the random variable x y?

Exercise 5. Let f : R3 R2 be the map (, ) = f (x, y, z) = (x2 + y 2 + z 2 , z),
which is nonsingular away from the line x = y = 0. For C0 (R3 \{x = y = 0}),
show directly (e.g. by computing the distribution function) that the pushforward
measure f# is given by
Z
0

(, , )d ( > 2 )

|d d|
2

in the coordinates (, , ) on R3 where is the polar angle in the x, y plane. Check

that the formula (136) gives the same result.
Exercise 6. Show that, if 0 is the Dirac delta function on R, then when viewed
as a distribution on (t, x) R R3 \{(0, 0)} , we have
Z
dx
< 0 (t2 |x|2 ), >=
((|x|, x) + (|x|, x))
2|x|
R3
Why does this formula make sense as
on all of R3+1 , even though
Pa3 distribution
2
2
i
the derivative d(t |x| ) = 2tdt i=1 2xi dx vanishes at the origin? Why is
this the only possible definition extending 0 (t2 |x|2 ) to all of R3+1 which remains
homogeneous of degree 2?

76

3. DISTRIBUTION THEORY

4.2. Other topics we have not discussed.

We have not included a section on multiplying distributions because one cannot
define, in general, a meaningful, associative, product of distributions which continuously extends the usual multiplication of functions. (Try to produce an example of
three distributions on the real line whose product, if defined, could not be associative, or an example of two distributions whose product could not be commutative.)
There is no difficulty in multiplying together distributions whose singularities are
disjoint one simply uses a smooth partition of unity to localize in space. When the
singularities occur at the same points the matter is more subtle. The multiplication
is still possible when the singularities do not collide in a sense made precise by
the notion of a wavefront set, which measures inside the cotangent bundle of Rn the
position and direction of the singularities of a distribution in this situation one
must try to localize in both space and frequency to formalize the multiplication.
It is possible to show, for example, that 0 (x1 , x2 ) = (x1 ) (x2 ) D0 (R2 ). We
will allude to the concept of the wavefront set at various points in the notes, but we
will remain vague about what this set actually is; for a rigorous discussion of the
wavefront set and the propagation of singularities, see ( ). For our purposes, it
will be enough to remember that singularities have both location and directions.
To give a more explicit hint of the most general setting: given two distributions
u(x) and v(x) D0 (X), one defines a tensor product u(x1 )v(x2 ) D0 (X X)
in the obvious way, and multiplication, when possible, is the pull-back u(x)v(x)
of the tensor product by the diagonal embedding X , X X. Note, however,
that the diagonal embedding does not satisfy the conditions we assumed when we
originally defined pullback of distributions in that its derivative is not surjective.
It is sometimes possible to define pullback even in such circumstances, but again
some conditions involving the map and the wavefront set of the distribution to be
pulled back must be met, as it is clear, for example, that not all distributions can be
restricted to lower dimensional subsets. We will confront this issue later on during
some of the calculations involving fundamental solutions and again when we study
trace theorems (which allow us to make sense of boundary values when dealing
with certain generalized functions in PDE) and restriction theorems for the Fourier
transform.

CHAPTER 4

Fundamental solutions and the basic linear PDEs

In the Introduction, we introduced the basic concept of a fundamental solution in

the particular case of the Laplacian in R3 , and used this solution to deduce some
basic facts about Poissons equation for compactly supported data. We now begin
to study several of the basic linear partial differential equations from the same angle
but in greater depth.
Given a linear partial differential operator with constant coefficients
X
P () =
a
||k

with a C, we say that a distribution E is a fundamental solution if it verifies

P ()E = 0 . If this is the case, then we can always find solution of the equation
P ()u = f , where f D0 (Rn ) is a compactly supported distribution, by setting
u = Ef . This follows easily from the observation that 0 u = u for any u D0 (Rn )
together with the following proposition (which ultimately stems from the fact that
all translations commute in Rn ).
Proposition 0.3. Assume u, v D0 (Rn ) one of which is compactly supported.
Then,
P ()(uv) = P ()uv = uP ()v.
One can prove the following general result.
Theorem 0.4 (Ehrenpreis, Malgrange). Any linear partial differential operator
P () on Rn with constant coefficients has a fundamental solution.
The proof, which involves elementary Fourier and functional analysis, is actually
rather peripheral to these notes (although a midterm exercise demonstrates that
there could be obstructions to a similar theorem in more general settings). Ultimately a fundamental solution has to be quite explicit to be useful in deriving
interesting properties of the underlying equations.
We treat instead specific examples of important, translation-invariant linear differential operators which have special, useful invariance properties. From the simplicity and special symmetries of these operators, we are able to derive explicit
formulas for the fundamental solutions. This allows us to derive important qualitative properties of the corresponding equations (existence and regularity of solutions,
continuous dependence on data, etc.). These qualitative properties, however, will
77

78

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

often have nothing to do with the various symmetries, and clearly remain true if
one were to perturb the operator slightly, or even substantially. Therefore keep
in mind that, though explicit formulas are very useful, we will ultimately need to
develop more robust techniques to understand properties of PDEs.

1. Cauchy-Riemann equations
The operator1

= 12 ( x
+ i y
) is fundamental to complex analysis, which studies

solutions f : C C to the linear partial differential equation f

z = 0. Such
functions are called holomorphic and taken together form an algebra over C, in
addition to having many other fascinating properties. The pair of partial differential
equations relating the real and imaginary parts of f are known as the CauchyRiemann equations. We will assume the reader is already familiar with the subject
of complex analysis, and proceed to develop basic facts in the subject through the
application of distribution theory. For this section, we will use the formalism of
dz
differential forms and in particular denote by dz2i
= dx dy the volume form on
f
z.
C. Given a function f (z, z) we write, in complex notation, df = f
z dz + z d
According to our general definition, a fundamental solution for z is a distribution
K in R2 such that z K = 0 . Unlike the Laplace operator, the z operator does
not commute with rotations u(z) 7 u
= u(eit z) as an operator. In fact, we have

it u it
= e z (e z) from the calculation
z u
u it
u it
(e z)d(eit z) +
(e z)d(eit z)
z
z
u
u
= eit (eit z)dz + eit (eit z)d
z
z
z

d[u(eit z)] =

However, if one has any fundamental solution u

z = 0 , then by this computation,
one can construct another fundamental solution eit u(eit z) since the -function is
rotationally invariant. By averaging over the group of rotations, we can assume
without loss of generality that these two fundamental solutions are the same so
that K(eit z) = eit K(z), which motivates us to seek a fundamental solution of the
form K(rei ) = g(r)ei .
Since 0 is homogeneous of degree 2 and z lowers the degree of homogeneity by
1, we would suspect that g(r) = rc for some constant c C, so that the fundamental solution would be homogeneous of degree 1. Thus, we are led to guess that
K(z) = zc for some constant c. Indeed, since z1 is locally integrable, it defines a distribution everywhere in R2 with z z1 supported at the origin. Moreover, since z z1
is homogeneous of degree 2, we deduce from the characterization of distributions
supported on a point that it must be a constant multiple of 0 ; i.e. z z1 = C0 for
some constant C C (possibly 0).

f
z

z z
1An easy way to remember this definition is to write f (x, y) = f ( z+
, 2iz ). Note also that
2

1
(
2 x

i y
) and df =

f
dz
z

f
d
z.
z

1.

CAUCHY-RIEMANN EQUATIONS

79

We may determine the constant by applying the distribution to any test function
we wish, and we will choose our test function to be the characteristic function of
the unit disk D = {|z| 1}. Technically, doing so leaves the realm of distribution
theory that we have covered, but we will have no problem justifying our computations: the point is that one factor is at least continuous wherever the other is a bit
singular, which allows one to pass from smooth approximations. In the notation of
real variables, it is possible to evaluate
Z
Z
1

1
H(1 |z|)
dxdy =
H(1 |x + iy|)dxdy
z z
x + iy z
by integrating by parts. By applying the product rule to
2

|z| =
(z z)
z
z
z
, and hence
we obtain that z |z| = 2|z|
Z
Z
1

H(1 |x + iy|)dxdy =
(1 |z|)dxdy
x + iy z
2|z|
=

Equivalently, we can apply Stokes theorem2 to compute that

Z
D

 


Z
1 d
z dz
1 dz
=
d
z
2i
z 2i
ZD
1 dz
=
D z 2i
Z 2
iei d
=
ei
2i
0
=

giving the proposition:

Proposition 1.1. Let K(z) =

1 1
z,

then

K
z

= 0

Having found a fundamental solution, we will immediately obtain a representation

formula for holomorphic functions. We first note that a variation of the above

calculation allows us to compute

z when is an open set with Lipschitz boundary
1
3
and with z replaced by an arbitrary test function shows that
Z
Z

d
z dz
dz
(z)
=
(z)
C0 (C)
z
2i
2i

2 One must take care that the parameterization 7 ei gives the correct orientation of
the circle. A naive replication of the following calculation with the clockwise parameterization
7 ei would have led to a sign error.
3We proved the divergence theorem assuming the boundary was smooth, however a very
slight variation of the proof works for Lipschitz boundary. See H
ormander, for instance.

80

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

For example, this follows from integrating the identity

d( dz) = ( + )d
z dz
z
z
One of the easiest ways to check that such integral identities involving distributions
are valid is by allowing the singular distribution to be approximated by smooth
functions  . In this case, z is a measure since is Lipschitz and is a
continuous function, so one can already see that the computation is valid.
1
In the special case when f
z = 0 in and f is (say) C in a neighborhood of the
can pass from smooth approximations n f to conclude that
Rclosure of , we
4
f
(z)dz
=
0
.

Now, applying

to the product



1
1

= z0 +
z
(z z0 )
(z z0 ) z

gives a compactly supported distribution supported only at z0 and the boundary

of .
dz
, giving the identity:
We understand this equality by integrating against dz2i

Theorem 1.2. Let C be a bounded domain with Lipschitz boundary , z0

and C (C), then
1
2i

(z)
1
dz = (z0 ) +
z z0
2i

1
d
z dz
z z z0

We have only stated the theorem for smooth functions, but the theorem holds much
more generally by approximation. In particular, we can pass from smooth to f
when f is holomorphic in and C 1 in a neighborhood of the closure of , and by
doing so we obtain as a corollary
Corollary 1.3 ( Cauchy Integral Formula). Let f , and z0 as above, then
Z
1
f (z)
dz = f (z0 )
2i z z0
Remark: Some care must be taken when applying the Cauchy Integral formula
and calculating the integral over the boundary. For one thing, the assumption that
f remains well-behaved at the boundary is essential for the passing from smooth
approximations as the example of z1 on the unit disk with the origin removed
illustrates. In thise case, the Cauchy Integral
cannot apply to (say) points
R Formula
1
z0 very close to 0 the boundary integral D z1 zz
dz
clearly has size not much
0
R
larger than the arclength |z|=1 1|dz| = 2. The other important issue which our
4More directly, one can calculate that d(f dz) = df dz = f dz dz + f d
z dz = 0 and
z
z

integrate over .

1.

CAUCHY-RIEMANN EQUATIONS

81

use of Stokes theorem subsumes is that the orientation of the boundary must be
taken into account if.
By analyzing the Cauchy Integral Formula one can show that holomorphic functions
(under the above conditions) possess a convergent power series expansion about any
interior point of , and in particular are smooth. Even more usefully, one can make
this smoothness quantitative by deducing estimates of the form || f ||L (K) .
||f ||L () for compact sets K contained in . The same estimates also indicate
how the solution to f
z = 0 varies continuously upon its boundary values (when the
solution exists). This analyticity is one example of a more general phenomenon: the
regularity of a fundamental solution away from the origin corresponds to regularity
of solutions to the PDE. It would be nice in general, however, to achieve a regularity
result such as this one (perhaps not as strong) without relying upon the explicit
formulas. We will revisit holomorphic functions shortly.
Exercise 1. We say that u D0 () is a weak solution to the Cauchy-Riemann
equations if u
z = 0 in the distribution theoretic sense. Prove that a continuous
function which is a weak solution is in fact a classical holomorphic function (and
hence analytic). (Hint: the class of holomorphic functions is closed under translation and linear combination, so it may be useful to consider a mollification of u.
Then use the a-priori estimates.)
Exercise 2.

From Exercise 1, deduce:

Theorem 1.4 ( Schwartz Reflection Principle). : Let be an open subregion of C

intersecting the real line in an interval I. If f is continuous on and holomorphic
on \I, then f is holomorphic on .
Hint: f (x + iy) = lim0 (|y| > ) f (x + iy) in the weak sense.
Exercise 3.

Also prove:

Theorem 1.5 ( Moreras theorem). : Let f be a continuous function in the open

disk D. Then f is a holomorphic function in D if and only if for any right triangle
interior T with boundary T contained in D one has
Z
Z
f
T
d
z dz
f (z)dz = 0
z
T
Note, the leftmost integral makes no sense classically, as no regularity assumptions
about f
z have been made. (Hint: this can be proven either by using the linearity in
T to pass to general test functions, or by taking advantage of translation invariance
in the assumptions)
Exercise 4.
The Cauchy Integral Formula immediately implies an estimate
of the form |f (z0 )| C maxz |f (z)| for some positive constant C = C(z0 )
independent of f . However, essentially because z0 : f f (z0 ) is also a ring
homomorphism, we are able to choose C = 1 independent of z0 and thereby prove

82

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

Theorem 1.6 ( Maximum Modulus Principle). If is open with compact closure

, and f is a holomorhic function in a neighborhood of , then
max |f (z)| = max |f (z)|
z

One argument in this spirit is due to Landau.

On the other hand, the estimate |f (z0 )| C maxz |f (z)| alone implies (by the
Hahn Banach theorem) that the linear functional z0 : f f (z0 ) defined initially
for continuous boundary values of holomorphic functions extends to a continuous
linear functional on C() and is therefore represented by a measureR(i.e. for f holomorphic in and continuous up to the boundary, we have f (z0 ) = f (z)dz0 (z)
for some finite measure z0 on the boundary which is not necessarily unique a
priori). In fact, we have already calculated this measure by proving the Cauchy
Integral Formula. On the other hand, the Maximum Modulus Principle can be
proven without making use of the Cauchy Integral Formula (the reader is asked to
provide such a proof in Exercise 8 of the following section on the Laplace operator).
One sees, therefore, that a-priori estimates and the existence of integral representation formulas come hand in hand in expressing the uniqueness of solutions and
continuous dependence on data. However, notice:
Exercise 5. If arbitrary boundary data f C() could be realized by a holomorphic function u, u
z = 0 in then the continuous functional z0 : f 7 u(z0 ) for
z0 would be a continuous ring homomorphism defined on the algebra C()
according to the Maximum Modulus Principle above. However, any continuous,
linear functional on C() extending z0 as above cannot be a ring homomorphism
5
. Hence, there exist continuous functions which cannot be realized as boundary
values of holomorphic functions.
Therefore, we cannot conclude from a priori estimates (like the Maximum Modulus
Principle) or representation formulas alone that solutions exist. Understanding
the obstructions to the existence of solutions is an interesting problem in PDE, and
usually involves understanding the underlying geometry or topology of the equation
of interest. For example, in the present case, a holomorphic function cannot map
a closed curve to another curve with a reversed orientation.

2. Laplace Operator
.P As we have seen in the introduction, the Laplace operator (or Laplacian) =
n
2
n
i=1 i on R is one of the simplest and most important linear differential operators. Solutions to u = 0 are called harmonic functions. In two dimensions,
is related to the study of holomorphic functions (for example, from the identity

= 4 z z
and our preceding regularity results, we see that real and imaginary
parts of holomorphic functions are harmonic). The operator is also often denoted
5In fact, all such continuous ring homomorphisms on the algebra C() are point masses on
the boundary itself.

2.

LAPLACE OPERATOR

83

= (1 , 2 , . . . , n ) (1 , 2 , . . . , n ); this notation makes clear the rotational

symmetry of the operator, and also the integration by parts identity
Z
Z
u vdx = u vdx u, v C0 (Rn )
which can be taken as an alternative definition of the operator6.
During the exercises in the introduction, we used spherical symmetry to find the
fundamental solution to the Laplace equation. We can now verify rigorously that
the fundamental solution we derived formally is a true fundamental solution.

1 2n
Proposition 2.1. Define, for all n 3, Kn (x) = (2 n)n
|x|
while, for
n = 2, K2 (x) = (2)1 log |x|. Here n denotes the area of the unit sphere Sn1 .
Then, for all n 2,
Kn = 0 .
Proof : By a direct calculation, Kn = (r2 + (n1)
r r )Kn vanishes away from
the origin and therefore can be expressed as a sum of derivatives of 0 . Therefore,
Kn is a distribution supported at the origin in Rn and homogeneous of degree n,
implying it is a constant multiple of 0 . To determine the constant, we may use
any test function, and (with the same considerations as in the Cauchy-Riemann
equations) we choose the characteristic function of the unit ball H(1 |x|). By
abuse of notation, let us write Kn (x) = Kn (|x|).

Z
Kn (x)H(1 |x|)dx Kn (x) H(1 |x|)dx
Z
x
= (1 |x|)Kn (|x|)
dx
|x|
Z
dKn
x
x
(1 |x|)
(|x|)

dx
dr
|x| |x|
Z
dKn
=
(1)d
|x|=1 dr
=1

With the fundamental solution Kn (x) in hand, we can solve the inhomogeneous
equation
V = , C0 (Rn )
R
with the formula V (x) = Kn (x) = Kn (x y)(y)dy. This solution is also often
denoted by 1 .
6This latter definition generalizes to the Laplace-Beltrami operator on Riemannian manifolds,
where the gradient, dot product, and volume form must be taken with respect to the metric. This
geometric point of view gives us another way of seeing the rotational invariance of the standard
Laplacian.

84

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

With some basic knowledge of differential geometry, we can give another proof. In
polar coordinates x = r, r > 0, || = 1, takes the form,
= r2 +

n1
r + r2 Sn1 ,
r

where Sn1 is the Laplace -Beltrami operator on the unit sphere Sn1 .
Exercise 1. Show that the Laplace-Beltrami operator on a Riemannian manifold
with metric g 7 is given, in local coordinates xi , by
p
1
g = p i g ij |g|j ).
|g|
Here g ij are the components of the inverse metric g 1 relative to the coordinates
xi . The volume element dSg on M is given, in local coordinates, by dSg =
p
|g|dx1 dx2 . . . dxn . Observe that, on compact manifold M ,
Z
Z
g u vdSg =
u g vdSg .
M

Exercise 2. Calculate the Laplace-Beltrami operator for the unit sphere Sn1
and check the polar decomposition formula for . For the particular case n = 3,
relative to the coordinates x1 = r cos 1 , x2 = r sin 1 cos 2 , x3 = r sin 1 sin 2 ,
1 [0, ), 2 [0, 2) show that,
S2 = 21 + cotan1 1 +

1
22 .
sin2 1

Moreover the area element dS takes the form, dS = r2 sin 1 d1 d2 .

Proof (geometric derivation): For a smooth function (x) = (r), in polar coordinates r = |x|, Sn1 unit sphere in Rn , we have


n1
r + r2 Sn1
r


= r(n1) r (rn1 r + r2 Sn1

r2 +

We now pass to polar coordinates x = r so that the volume element may be

written dx = rn1 drdS . Integrating by parts on the Riemannian manifold Sn1 ,
we calculate that
< Kn , >

= < Kn , >
Z
Z
Z


n1
=
Kn (r)r r
r drdS +
||=1

||=1

(2 n)n

1

Z
||=1

Z
=
0

Kn (r)Sn1 drdS

rn+2 r rn1 r drdS + 0

rn+1

rn1 r dr =

r (r)dr
= (0)

7 Where u is defined implicitly by du(X) = g( u, X), and is defined by the identity

gR
g
g
g uvdSg = g(g u, g v)dSg for u, v C0 (M ),

2.

LAPLACE OPERATOR

85

where in the above calculation we define (r)

to be the average of over |x| = r.
We infer that, for n 3, Kn = 0 as desired. The case n = 2 can be treated in
the same manner.
Remark : Observe that, up to a constant, the expression of Kn (x) can also be
easily guessed by looking for spherically symmetric solutions K = K(|x|). Indeed,
0
the equation K = 0 reduces to the ODE, K 00 (r) + n1
r K (r) = 0.

Having found a fundamental solution, we can immediately deduce a representation

formula as before. Let C0 be a smooth test function, let a bounded, open
set with Lipschitz boundary, and let x .
We have
Z
(y)(x y)dy

(x) =
Z
=

Kn (x y)dy

Our strategy is to integrate by parts, allowing at most one derivative to hit the
characteristic function. We recall from our discussion of pullbacks of distributions
that = ~nd where ~n is the interior unit normal and d is the surface measure
on the boundary. In contrast to a classical integration by parts, one proceeds as
though there are no boundary terms since the product Kn (xy) has compact
support. For a function f with a continuous first derivative at the boundary ,
we let f
denote the outward unit normal.
Z
(x) = ( ) Kn (x y)dy
Z
Z
= Kn (x y)dy Kn (x y)dy
Z
Z

=
Kn (x y)d(y) ((( Kn ) Kn )

Z
Z
Z

=
Kn (x y) d(y) +
Kn (x y)dy
Kn (x y)d(y)

We thus derive the representation formula,
Z
(x) =
Kn (x y)dy
(137)
Z
Z

+
Kn (x y)d(y)
Kn (x y) d(y)

In particular (by approximation), if u is harmonic within (and, say, C 2 in a

neighborhood of )
Proposition 2.2.
Z
u(y) =


Kn
u
u(x)
(x y)
(x)Kn (x y) d(x)

(138)

86

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

With our representation formula in hand, we can repeat much of the same analysis
that had been remarked for the Cauchy-Riemann equations. We find that, thanks
to the real analyticity of the fundamental solution, harmonic functions as above are
in fact real analytic, with quantitative a priori estimates on derivatives in terms of
the boundary values of u and u
. We can also use these estimates to show that
continuous functions satisfying u = 0 are actually classical solutions. But to proceed with the analysis of harmonic functions from this formula may be misleading
because the interior values of a harmonic function are uniquely determined by its
boundary values alone, and therefore the normal derivative cannot be prescribed
arbitrarily.
Indeed, the Maximum Principle for harmonic functions which we now state implies that harmonic functions in the interior of a domain are determined by their
boundary values alone.
Theorem 2.3 ( Maximum Principle). If u : R is C 2 on a connected, open
set and u 0 in , then u cannot obtain an interior maximum unless u is a
constant. In particular, when is bounded,
sup u(x) = sup u(x)

(139)

and as a consequence, sup |u(x)| = sup |u(x)| when u is harmonic.

Proof : The first statement is called the strong maximum principle for C 2 , subharmonic functions (functions satisfying u 0 in the classical sense); the theorem
implies that a subharmonic function in a domain remains in the interior strictly
below any harmonic function with everywhere greater boundary values, hence the
term subharmonic. The strong maximum principle will be an easy consequence
of a theorem to be proved later (the mean value inequality for subharmonic functions), although we will leave the proof for the reader. At the moment, however,
we can at least prove the weak maximum principle (that is, inequality (139) ).
Indeed, when u > 0 is strictly positive, the (strong) maximum principle is obvious
because the function u is at any point strictly convex in at least one direction. By an
approximation (e.g. replacing u by, say, u + x21 ), we can obtain the weak maximum
principle when we only assume u 0 in a bounded domain and that u extends
continuously to .
It is clear that a strong minimum principle for superharmonic functions8 holds
upon replacing u with u, which in particular implies the last equality stated in
the theorem.
The concepts of superharmonic and subharmonic functions described in the proof
are useful even for the analysis of harmonic functions themselves because they are
much easier to construct explicitly (with exponentials, polynomials, etc.) and can
be used to bound harmonic functions according to the maximum principle above.
8 Superharmonic means u 0.

2.

LAPLACE OPERATOR

87

By the same reasoning that followed the discussion of the maximum modulus principle for holomorphic functions,R there must be a representation formula for harmonic
functions of the form u(y) = u(x)dy (x) for some finite measure y depending
on y . We can obtain such a representation formula as follows: if a harmonic function y (x) can be found which coincides with the fundamental solution
Kn (x y) on the boundary of , then the function G(x, y) = Kn (x y) y (x)
satisfies
x G(x, y) = y (x) in G(x, y) = 0 on
There can be only one such function by the maximum principle. This function
G(x, y) above is called the Greens function for , and wasR introduced formally in
the exercises in the Introduction. By computing u(y) = (x)u(x)G(x, y)dx
as in our previous representation formula (this time the boundary condition for
G cancels a boundary term: G(x, y) = ( G(x, y))) we obtain our desired
representation formula:
Proposition 2.4. If u is harmonic in and C 2 in a neighborhood of and G(x, y)
is as above, then

Z 
G
u(y) =
u(x)
(x, y) d(x)
(140)

Note: we have not proven that the function defined by the right hand side of the
formula is defined for arbitrary domains, nor that it defines a harmonic function,
nor even that it realizes the boundary values in the integrand as y tends towards
the boundary. When the boundary is sufficiently nice (say, Lipschitz), all of these
things can be proven and arbitrary continuous boundary values can be achieved by
harmonic functions (in contrast to the Cauchy-Riemann equations).
The probability measure G
(x, y)d(x) appearing in (140) describes the probability
distribution of the first contact with the boundary of a random walk beginning at
the point y. Thus, the value of a harmonic function at the point y may be considered
the expected value which the boundary data obtains at the first contact point of a
random walk beginning at y. From this interpretation some features of harmonic
functions (the maximum principle and mean value property below, for example)
are obvious, but we will not explore this interpretation here.
Example: For the half-space xn > 0 in Rn , one can obtain an explicit formula for
the Greens function G(x, y) = y , G(x1 , . . . , xn1 , 0) = 0 by placing a negative
point source at the point y = (y 1 , . . . , y n1 , y n ), and defining G(x, y) = K(x
y) K(x y ). Then G(x, y) = 0 on xn = 0 since such points are equidistant from
both y and y and the fundamental solution depends only on Euclidean distance.
The same method can also be used to construct a Greens function for a ball |y| 1.
In this case, one uses the conformal reflection y y = |y|y 2 , which fixes the sphere
|y| = 1. The Greens function then takes the form G(x, y) = K(x y) K(|x|(x
y )). Many more examples can be obtained in two dimensions using holomorphic
functions.
Among the main results of our analysis up to this point (the maximum principle,
some of the various a priori estimates which can be deduced from Greens formula,

88

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

and the existence of solutions on Rn for compactly supported data), many hold
for other operators closely analogous to the Laplacian. For example, by changing variables, we see that when u is a harmonic function and v((x)) = u(x) for
a
of Rn , then v will satisfy an equation of the form L[v] =
Pdiffeomorphism
n
ij
i
ij
i,j=1 a (y)i j v + b (y)i v = 0 where the smooth functions a (y) are the coefij
ficients of a symmetric, positive definite matrix a ((x)) = D(x)(D)t (x) and
the first order terms depend on second derivatives of .
Pn
Pn
More general operators of the form L = i,j=1 aij (x)i j + i=1 bi i where the
matrices (aij (x)) symmetric and positive definite are called elliptic, and it is no
surprise that they share many properties in common with the Laplacian, but they
generally do not necessarily possess the same amount of symmetry 9 as the Laplace
operator does, and therefore they require more robust methods to analyze successfully. However, there are also methods, for extending and transporting results and
estimates for the Laplace operator to more general elliptic operators.
The following theorem embodies the rotational and translational symmetry of the
Laplace operator, and in fact characterizes harmonic functions as well as the Laplace
operator itself. Therefore, it can be used to prove results for the Laplace operator
and harmonic functions which are beyond the reach of other methods, and therefore
its applications are also limited to these purposes. The theorem shows how the
Laplacian controls the change in spherical averages of varying radius.
Theorem 2.5. [Mean Value Property] When u is harmonic in the ball of radius
R > R about x, u(x) is equal to its average over the sphere of radius R centered
at the point x
Z
1
1
u(y)d(y)
u(x) =
nn R(n1) |yx|=R
with the = replaced by when u 0. In fact, for all u C (Rn ), 0 < R1 <
R2 < R ,
Z
Z
1
1
u(x)d(x)
u(x)d(x)
(n1)
(n1)
|x|=R2
|x|=R1
R2
R1
#
Z R2 "Z
=
u(y)dy (n1) d
R1

|y|

Proof We prove the last formula, since the first identity of the theorem is an
immediate consequence (by letting the inner radius tend to 0). In fact, the latter
formula shows that spherical averages increase with the radius when u 0.
9When one refers to the symmetries of a partial differential operator, one often has in

mind a collection of vector fields which commute with the operator, or their flows which leave
the operator invariant. The symmetries of on Rn are the symmetries of the underlying Euclidean geometry: the infinitesimal translations x1 , . . . , xn together with the infinitesimal rotations. The corresponding flows generate the group of rigid motions of Euclidean space. In
this diffeomorphism-invariant sense, a coordinate change of the Laplacian as above has the same
amount of symmetry. We will see later that such symmetries can be very helpful when analyzing
a differential operator. In any case, it is obvious that the condition of ellipticity alone does not
imply the existence of such operator-preserving flows.

2.

LAPLACE OPERATOR

89

Although the formula has been stated in integral

R form, we prove a 1differential
R
1
u(x)d(x) = (n1) (
version using the auxiliary function ( ) = (n1)
|x|=
|x|)u(x)dx. We denote r = |x|.

d
=
d
=
=
=
=

(n 1)

( r)u(x)dx

(n1)

Z 
1

(n 1)

( r)u(x)dx
r
r
(n1)

Z 
1

(n 1)

( )H( r)u(x)dx
r
r
r
(n1)
Z
1
H( r)u(x)dx
(n1)
Z
1
u(x)dx
(n1) |x|

In the second equality, we used the fact that ( r) is a distribution of order 0,

and that (n1)
and (n1)
coincide to 0th order on its support, as well as the fact
r

that any distribution pulled back by the map (, x) r remains fixed by any
vector field in the null space of d dr (hence, ( + r )( r) = 0). In the fourth
(n1)
2
line, we recognized that the operator r
2 +
r
r coincides with the Laplacian
when applied to spherically symmetric functions.
Integrating in from R1 to R2 gives the desired formula.
Remark 2.6. What we have essentially computed is that
Z
Z
u(x)d(x) = u1 dx
where is the measure in Exercise 5 of the Introduction.
A special case of the above formula has important applications to complex analysis.
When f is a nonzero holomorphic function in a disk D, we have the identity
X
1
k (z)
log |f (z)| =
2

where k runs over the finite collection zeros of f counted with multiplicity the
measure on the right hand side is known in algebraic geometry as the zero divisor of
f . One can see this identity locally near a zero j by writing f (z) = eg(z) (z j )n
for some function g holomorphic in a neighborhood of j , and by recalling that real
parts of holomorphic functions are harmonic and that the fundamental solution for
1
the Laplacian in two dimensions is given by 2
log |z|. This calculation implies in
particular that log |f (z)| is a subharmonic function (a fact which can be used in
combination with the maximum principle to give strong estimates on holomorphic
1
functions). Applying the general formula in the Theorem (2.5) to u = 2
log |f (z)|,
we obtain

90

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

Proposition 2.7 (Jensens formula).

Let f be a function which is holomorphic
in a neighborhood of the closed ball of radius R centered at zero and whose (discrete)
collection of zeros {j } satisfy 0 < |j | =
6 R. Then we have,
 
Z 2
X
R
1
log |f (Rei )|d log |f (0)| =
log  
2R 0
j
|j |<R

Proof Beginning with the general formula in the Theorem (2.5) we let u
1
2 log |f (z)| by the standard mollifier construction, and notice that our assumptions
on f (z) are enough to guarantee that the integral formula in the theorem is still
valid for R2 = R and 0 < R1 < min |j |. Letting R1 tend to 0 gives the left hand
side of Jensens formula.
We now calculate the right hand side of the general formula explicitly with Fubinis
theorem (or integration by parts):

#
Z R Z X
Z R "Z
1

log |f (z)|dy 1 d =
|j | (t)dt 1 d
0
0 j
0
|y| 2
"Z
#
Z RX
R
1
=
|j | (t)
d dt
0

 
X
R
=
log  
j
|j |<R

The theorems in Exercises 1 - 5 are basic theorems in the study of the Laplace
equation (along with their generalizations to other elliptic PDE), and their proofs
may be found in some form in either Evans or Gilbarg and Trudinger should the
reader wish to consult a reference.
Exercise 1.[ Hopf Lemma] states that a harmonic function on a bounded, open
set u : R must satisfy u
(x0 ) > 0 at a boundary point x0 where the boundary
is smooth and u(x0 ) > u(x) for x \{x0 }. Prove this fact. One approach is
to design an appropriate superharmonic perturbation of u close to x0 and use the
weak maximum principle to bound u below the superharmonic perturbation.
Exercise 2. Prove the strong maximum principle for subharmonic functions. (By
now you may be able to see more than one proof)
Try to obtain this result also for C 2 subsolutions to an elliptic equation in other
words, supposing
Lu =

n
X
i,j=1

aij (x)i j u +

n
X

bi (x)i u 0

i=1

where the matrices aij (x)i j ||2 are uniformly positive definite.

2.

LAPLACE OPERATOR

91

Exercise 3. Prove the following properties

of the Greens
function of an open set
R
R
with nice boundary : G(x, y) = x (z)G(z, y)dz = y (z)G(z, x)dz = G(y, x),
G(x, y) < 0 for x \{y} and G
(x0 ) > 0 at any boundary point x0 , and
R G
(x,
y)d(x)
=
1.

How would you interpret any of these facts either physically or probabilistically?
Exercise 4.[ Harnacks Inequality] Prove that, If K is compact, then there is
a constant C depending on K such that for all non-negative harmonic functions u
in
sup u C inf u
K

Theorem 2.8 (Liouvilles Theorem). A bounded harmonic function on all of Rn

is a constant.

Exercise 5.

Prove Liouvilles theorem.

Exercise 6. The co-area

formula, written distribution theoretically (and a bit
R
vaguely) as h(f (x)) = h(t)0 (f (x) t)dt, allows us to decompose a general density h(f (x)) into surface measures on the level sets of f , which can then be analyzed
individually. For example, when f (x) = |x|, the co-area formula reduces to integration in polar coordinates. This formula is equivalent to the definition of pull-back of
a distribution and is proven in the Appendix; use it to prove the following identity:
Let u C0 (Rn ), and (x) = (|x|) be a spherically symmetric density. Then if
 n ( x ) we have the formula

u  (x) = (n1)


"Z

#
u(x y)dy t
(t)dt

|y|t

Exercise 7. Prove (at least the n 3 case of) the following

Proposition 2.9. For any f C0 (Rn ), n 3 the equation u = f has a unique
smooth solution which vanishes
at infinity, i.e. tends to zero as |x| . The
R
solution is represented by Rn Kn (x y)f (y)dy. For n = 2 the same equation has
a smooth solution u(x) with lim|x| |u(x)|
|x| = 0 and |u(x)| 0 as |x| . The
R 1
solution is represented by R2 2 log |x y|f (y)dy , and is unique (in this class) up
to an additive constant.

Exercise 8. Prove the Maximum Modulus Principle for holomorphic functions.

We will return to the study of the Laplace equation (and some of its generalizations)
in Chapter 3.

92

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

3. DAlembertian operator
Recall that the Dalembertian  = t2 + is the simplest differential operator in
R1+n invariant under translations and Lorentz transformations, i.e. the Poincare.
group. The easiest way to see is to write  = m , with m the Minkowski
metric. Since m is invariant under the Poincare group so is . Thus it makes
sense to look for a fundamental solution of the form (t, x) = f () where10 =
t2 |x|2 = m x x is invariant under Lorentz transformations. Also, because
the distribution 0 on Rn+1 is homogeneous of degree (n + 1) and applying 
lowers the degree of homogeneity by 2, we conclude that f must be homogeneous
of degree n1
2 . Therefore, a good candidate for a fundamental solution must have
n1
the form E = cn (t2 |x|2 ) 2 , for some constant cn , in the region t > |x|. We are
therefore led to look for a distribution E+ , homogeneous of degree n + 1, which
n1
coincides with cn (t2 |x|2 ) 2 in the region t > |x|. This may seem difficult at
n1
first, due to the high degree of the singularity of (t2 |x|2 ) 2 along |x| = |t|,
until we realize that we can make use of the homogeneous family of distributions
ja defined by proposition ??. We need to choose in fact a = n1
2 + 1 and take
E+ to be proportional toj n1 +1 (t2 x2 ), understood as the pull back f (j n1 ),
2
2
with f = t2 |x|2 . It is more convenient in the context to change notation a little
bit and write,
a+ := ja+1
Thus,
n1
2

E = +
n1
2

Note that the expression +

(t2 |x|2 ).

(t2 |x|2 ) is not exactly rigorous, since the gradient

n1

of t2 |x|2 vanishes at the origin, and hence + 2 (t2 |x|2 ) defines a distribution
only on Rn+1 {0}. A rigorous formulation requires a bit more care, but the
particular degree of homogeneity of the distribution basically allows for a unique
extension to the whole space, See Exercise 3 of section (4) for the n = 3 case. Now
the distribution we have produced has the right properties except for the fact that
it supported in the entire region |x| |t|. For deterministic physical reasons we
prefer a distribution supported only in the future region |x| t. This defines our
(n+1)

n1

(t, x) = cn H(t)+ 2 (t2

candidate for a forward fundamental solution E+
2
|x| ) with H(t) the Heavyside function supported on t 0 and cn a normalizing
constant to be determined in the verification. Using the chain rule it is easy to


n1
(n+1)
show that, E+
= m + 2 (f ) must vanish outside the origin. By
(n+1)
the usual homogeneity considerations we deduce that E+
is proportional to
the function at the origin. It thus only remains to determine the normalizing
factor cn . We have the following result,
(n+1)

defined by
Theorem 3.1. The distribution E+
1n
1
n1
(n+1)
E+
(t, x) = 2 H(t)+ 2 (t2 |x|2 )
2

(141)

10 Recall that we denote t = x0 and we use the summation convention w.r.t the indices
, = 0, 1, . . . , n.

3. DALEMBERTIAN OPERATOR

93

is the unique fundamental solution of the wave equations supported in the forward
region |x| t.
We shall prove this theorem later for the moment a few remarks are in order. First,
observe a fundamental difference between the cases when n > 1 is odd and the
(n+1)
is supported only on the
cases when n is even. Indeed in the former case E+
boundary of the region |x| < t, i.e the future light cone |x| = t while in the latter
(n+1)
case E+
is supported in the entire forward region |x| t. More precisely, using
s1
the chain rule and the fact that d s+ () = +
(), we can write the fundamental
d
solution a bit more explicitly away from the origin: in dimensions n = 3 + 2k, the
fundamental solution looks like a derivative of a measure supported on the forward
light cone
k

k

1
1
r (t2 |x|2 ) = cn H(t)
t (t2 |x|2 )
cn H(t)
2r
2t
while in n = 2 + 2k dimensions, it is of the form

k

k
1
1
1
1
p
p
r
t
cn H(t)
(|x| t) = cn H(t)
(|x| t)
2
2
2
2r
2t
t |x|
t |x|2
(the above distributions being equal since 1t t + 1r r is in the null space of d(t2 r2 )).
In the most important particular case, when n = 3, we have,
(1+3)

E+

1
1
H(t)(t2 |x|2 ) =
(t |x|)
2
4

(142)

1
H(t)(t2 |x|2 )1/2 =
2 1/2

(143)

Also, for n = 2,
(1+2)

E+

It is important to observe that the knowledge of the fundamental solution in odd

dimensions allows one to determine it for even dimensions. This is called the method
(2k+1+1)
to a test functions which
of descent. This can be done by simply applying E+
are independent of one of the spatial variables. Never mind that this test function
does not have compact support, it will work because the fundamental solution has
compact support in x for any t. As an example the reader is invited to deduce
(148) from (142). As another simple remark, observe that though E+ is compactly
supported in x for every fixed t. Thus E+ can be applied to any smooth functions
whose compact support in t.
Exercise 1. Deduce the fundamental solution for dimension n = 1. Show in
fact11 that the general solution to the Cauchy problem takes on the form 1 (t +
x) + 2 (t x).
The fundamental solution allows us to solve the general Cauchy problem,
 = f,

(0, x) = f (x), t (0, x) = g(x)

11In 1 + 1 dimensions, the wave operator factors into  = ( ) ( + )

t
x
t
x

(144)

94

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

To see how to do this consider a point p = (t0 , x0 ) with t0 > 0 and observe that,
for any test function we have in the upper half space, D+ = {(t, x) t 0},
Z
Z
(t0 , x0 ) =
(t, x)p (t, x)dtdx =
+ (m )Ep (t, x)dtdx
R1+n

D+

where + is the characteristic function of D+ and Ep (t, x) = E+ (t to , x x0 ).

Therefore, integrating by parts,
Z
Z
(t0 , x0 ) =
m Ep
m Ep
1+n
1+n
ZR
ZR

=
m Ep +
m Ep
R1+n
R1+n
Z
+
m b Ep
R1+n

i.e.,
Z
Z
Ep
Ep +
Ep
1+n
1+n
R
R
D
Z
Z
Z +
 +
t t Ep
t t Ep
R1+n
R1+n
D
Z +
Z
Z
 +
(t) t Ep
(t)t Ep
D
R1+n
R1+n
Z +
Z
Z
Ep
(t) t0 Ep
(t)t Ep
D
R1+n
R1+n
Z +
Z
Z


Ep t0
f (x)Ep (0, x)dx
g(x)Ep (0, x)dx
Z

(t0 , x0 )

=
=
=
=
=

Rn

D+

Rn

The final formula takes the form,

Z
(t0 , x0 )

t0

=
0

ZR

t0
Rn

E+ (t0 t, x0 x)(t, x)dtdx

(145)
Z


E+ (t0 , x x0 )f (x)dx
E+ (t0 , x x0 )g(x)dx
Rn

Leaving aside the issue of uniqueness, which we shall treat separately later on, we
deduce the following.
Theorem 3.2. [Kirchoff-Hadamard] The initial value problem  = F , (0, x) =
f (x), t (0, x) = g(x) has a unique solution for arbitrary smooth functions f, g, F ,
given by formula (145).
Exercise 2. Compare formula (145) with (137) for the Laplacian. Explain what
may go wrong if we try to prove a result for the Laplace equation similar to that
of theorem 3.2 above.

3. DALEMBERTIAN OPERATOR

95

Exercise 3. Show that in the particular case of dimension 1 + 3 formula (145)

takes the more familiar Kirchoff formula form,
(t, x)

t (4t)1

f (y)da(y) + (4t)1

|xy|=t

Z
+
0

1
ds
ts

Z
g(y)da(y)
|xy|=t

Z
(s, y)da(y)

(146)

|xy|=ts

The traditional way to derive the Kirchoff formula (146) is to first prove it in the
homogeneous case, i.e.  = 0. In fact it suffices to prove it for the case f = 0
and arbitrary g using the beautiful method of spherical means, see [J] for a clean
derivation. Once the homogeneous case is treated one can derive the general formula
using the Duhamel principle. This goes as follows: Let W (t)g denote the solution
(t, ) of the homogeneous problem with data f = 0 and arbitrary g. Think of it
as an family of operators, parametrized by t, which take smooth functions in Rn to
smooth functions in Rn . We then have to verify that the solution of the equation
 = F is given by the formula
Z

W (t s)F (s, )ds

(t, x) =

(147)

Exercise 4. Prove the claim.

Exercise 5. What happens if we replace in the formulation of the Cauchy problem
the hypersurface t = 0 with a more general hypersurface 0 given by t = h(x) ?.
Show a similar formula with that in (145) can be deduced if the hypersurface is
space-like, i.e. |h(x)| < 1. What happens if the surface becomes time-like, i.e.
|h(x0 )| > 1 at some point (t0 , x0 ) 0 . Show that the Cauchy problem with
prescribed initial values and normal derivatives on the light cone t = |x| does not,
in general, admit a solution in the spatial interior of the cone. What happens when
you try to derive a representation formula for data on the light cone with H(t |x|)
replacing H(t)?
Exercise 6. Suppose is compactly supported and that is smooth outside of
a compact set K. Let u+ be the solution E+ to 2u = in Rn+1 . Show
that u is smooth outside of the set of light cones emanating from K given by
{k + (|x|, x) : k K, x Rn }.
Proof of theorem 3.1 in R1+3 . First remark that we can write  in terms of
spherical coordinates as follows,
2
 = t2 + = t2 + r2 + r + r2 Sn1
r
we have to check that
(1+3)

E+

(t, x)

1
1
= 1 H(t)0 (t2 |x|2 ) = r1 (t r)
2
4

96

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

with r = |x|. Thus, since 
= r1 (t + r )(t r )(r) + S2 , we have with

(t, r) = (t r ) r(t, r) ,
Z Z Z
1
(t r)(t + r )dtdrdS
< E+ ,  > =
4 S2 0
Z
Z0


1
d
=
dS
(r, r)dr
4 S2
dr
0
= (0, 0) = (0)
Thus, E+ = 0 as desired.
In what follows we give yet another derivation of the fundamental solution for the
wave equation in the special case R1+3 . This is the so called geometrics optics
derivation. We look for solutions of  = 0 of the form,
E

A(u)

(148)

for given real functions A and u to be determined. Here (u) is simply the pull
back of 0 by u as discussed in subsection 4, Example 2. A simple calculation leads
to,




m A(u) = m A (u) + (2 A u + u) 0 (u) + u u 00 (u)
To cancel the coefficient of 00 (u) we need to chose u such that,
m u u = 0.

(149)

This is the famous Eikonal equation in Minkowski space. A simple family of solutions is given by u(t, x) = t t0 |x x0 | for a given point (t0 , x0 ), whose level
hypersurfaces are simply backward light cones with vertex at (t0 , x0 ). For our purposes we choose u = t |x|. Next, to cancel the coefficient of 0 (u), we need to
choose A such that12,
2 A u + u = 0.
One can easily check that the choice A = |x|1 will do. Finally it only remains to
calculate the term containing (u), i.e.,
(A)(u) = (|x|1 )(u) = 40 (x)(u) = 0 (t, x)
where the first 0 (x) is the delta function in R3 while the final 0 is the desired
1+3
1 1
delta function in R1+3 . Hence E+
= 4
|x| (t |x|) as desired.
Exercise 7 Justify that last step involving products of distributions.
Uniqueness of the fundamental solution E+ . It suffices to prove uniqueness of
solutions to the general Cauchy problem in theorem (3.2).
Exercise 8.

Verify the above statement.

We start with the simple calculation involving the energy momentum tensor. To
calculate efficiently it helps to remember that we are using the summation convection with respect to the space-time indices , = 0, 1, . . . , n. We will also be using
12It turns out that the equation below can be interpreted as a transport equation along the
generators of the backward null cone t |x| = 0.

3. DALEMBERTIAN OPERATOR

97

the standard geometric convention of raising, or lowering, the indices relative to

the metric. Thus, if U is vector (so called contravariant) we define the covariant
vector U = m U . Similarly, if V is a covariant vector, we define V = m V .
Proposition 3.3. Let


1
1
Q [] = m m = m
2
2

(150)

the so called energy momentum tensor of . Then13

Q =  .

(151)

In particular, if  = 0, Q0 = 0. Now consider a point p(t0 , x0 ) R1+n and the

solution u(t, x) = t t0 |x x0 | to (149) introduced above. Let also t1 < tt < t0
and consider the distribution H(u)H(t t1 )H(t2 t), with H(t) the Heaviside
function, and perform an integration by parts, as we have done many times before,
to derive the identity,
Z
0 =
H(u)H(t t1 )H(t2 t) Q0
R1+n
Z
=
H 0 (u)H(t t1 )H(t2 t) uQ0
R1+n
Z
+
H(u)H(t t1 )H 0 (t2 t)Q00
1+n
ZR

H(u)H 0 (t t1 )H(t2 t)Q00

R1+n

This identity can be rewritten in the form,

Z
Z
Z
Q00 +
Q0 L =
D(t2 )

N (t1 ,t2 )

Q00

(152)

D(t1 )

where D(t1 ), D(t2 ) are t-sections through the solid light cone |x x0 | t t0 ,
N (t1 , t2 ) represents the portion of the light cone |x x0 | = t t0 between the
sections t = t1 and t = t2 and
L = u = m u.
It is easy to check that,
Q00 =

n
X
1
1
(|t |2 +
|i |2 ) = (|t |2 + ||2 ).
2
2
1=1

and that we have,

Exercise 9. Show that
Q0 L 0.
We thus deduce
13 In view of the raising and lowering of indices convention, Q
Q
. This
= m
works well, since m = 0.

98

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

Theorem 3.4 (Energy inequality). For every solution of the wave equation  = 0,
in a neighborhood of the solid region bounded by the surfaces t t0 |x x0 | = 0,
t = t1 and t = t2 , we have,
Z
D(t2 )

1
(|t |2 + ||2 )
2

Z
D(t1 )

1
(|t |2 + ||2 )
2

In particular any smooth solution which vanishes at D(t1 ) must also vanish at
D(t2 ).

Exercise 10. Deduce from the energy inequality the finite propagation speed for
the Cauchy problem, as we have deduced earlier from the explicit solution. Note
that strong Huygens principle in odd spatial dimensions n = 3+2k discussed earlier
is a very special phenomenon related to the precise form of the wave operator 2.
However the phenomenon of finite speed of propagation exhibited by the wave
equation is a more robust feature shared by many related equations for which an
energy inequality as above holds true.
Finally, we check below the validity of our forward fundamental solution in all
dimensions.
Proof [Theorem 3.1 all n] We prove the formula (modulo the absolute value of the
constant cn . )
Z Z

+

(n1)
2

(t2 |x|2 )H(1 |t|)dxdt

Z Z
(n1)

=
t2 + 2 (t2 |x|2 ) H(1 |t|)dxdt
Z Z
(n1)

=
t + 2 (t2 r2 )(1 t)dxdt
Z Z
(n1)

t + 2 (t2 r2 )(1 + t)dxdt

Z Z
(n1)

=2
t + 2 (t2 r2 )(1 t)dxdt
Z Z
(n1)
t

=2
( )r + 2 (t2 r2 )(1 t)dxdt
r

Taking the definition of (1t) the above becomes an integral over the hypersurface
t = 1 which we put in polar coordinates (writing the volume form rn1 drdn1
with dn1 the surface measure on the unit sphere in Rn ). We remark here that,
as we will see later on, a distribution does not quite have to be a continuous function in order to have a meaningful restriction to a lower dimensional submanifold.

4. HEAT OPERATOR H.

99

Dropping the unimportant factor of 2, we proceed

Z Z
(n1)
t

( )r + 2 (t2 r2 )(1 t)dxdt

r
Z Z
(n1)
1

=
r + 2 (1 r2 )rn1 drdn1
r
Z
(n1)

= |S n1 |
rn2 r + 2 (1 r2 )dr
0

Our proposition has therefore reduced to showing that the number

Z
(n1)

rn2 r + 2 (1 r2 )dr
0

is positive. Despite not having defined this number14, we prove its positivity by
induction, separating into cases based on the parity n. For n = 1 + 2k, we integrate
by parts and notice the boundary term vanishes to find

rn2 r +

(n1)
2

(1 r2 )dr =

r2k (

r k+1
)
H(1 r2 )dr
2r

Ik

r k
r r2k1 H(r) (
) H(1 r2 )dr
2r
Z
r k
) H(1 r2 )dr
= 2(k+1) (2k 1)
r2(k1) (
2r
0
(k+1)

=2

= 2(k+1) (2k 1)Ik1

Where the last integral should be positive by induction on k provided we can
calculate
Z
Z
r
I1 =
r2 (
)H(1 r2 )dr =
(1 r2 )dr
2r
0
0
Z
1
(1 r)dr = 1/2 > 0
=
1
+
r
0
This proves the proposition in odd spatial dimensions once cn > 0 has been chosen
appropriately. The case n = 2 + 2k is similar.

4. Heat Operator H.
We consider the heat operator H = t acting on functions defined on RRn =
R
. It makes sense to look for spherically symmetric solutions to Hu = 0: that
is to say, functions u(t, x) = u(t, |x|) = u(t, r). It is possible to find in this way
2
n
a class of locally integrable solutions Ec (t, x) = cH(t) t 2 e|x| /4t , with H(t) the
Heaviside function (although it is easier to proceed via the Fourier transform).
n+1

14Observe, however, that (by the product rule) r n2 H(r) has a decent amount of regularity,

and its only singularity is located away from r = 1.

100

4. FUNDAMENTAL SOLUTIONS AND THE BASIC LINEAR PDES

Indeed H(Ec ) = 0 for all (t, x) 6= (0, 0). We show below that, in the whole space,
H(Ec ) is proportional to 0 and that we can determine the constant c = cn =
n
2n 2 such that the corresponding E = Ec is a fundamental solution of H, i.e.
H(E) = 0 .
We could very easily reason by considering the parabolic scaling (t, x) (2 t, x),
that HEc = C0 for some constant C (possibly 0). To determine the constant, we
could use any test function, and it would be simple to take H(1 t). More or less,
this is exactly how we will proceed.
Let C0 (Rn+1 ),
< H(E), >

=
=
=
=

< E, H >= E(t, x)(t + )(t, x)dxdt

Z Z
lim+
E(t, x)(t + )(t, x)dxdt
0

Rn
Z Z
Z
lim
(t )E(t, x)(t, x)dxdt + lim
E(, x)(, x)dx
0+ 
0+ Rn
Rn
Z
Z
2
lim+
E(, x)(, x)dx = cn lim+ n/2
e|x| /4 (, x)dx
0

0

Rn

Rn

We now perform the change of variables x = 21/2 y,

Z
Z
2
< H(E), > = 2n cn lim+
(, 21/2 y)e|y| dy = 2n cn (0, 0)
0

Rn

e|y| dy

Rn

= (0, 0)
R
2
Modulo the fact that Rn e|y| dy = n/2 (which will be shown later), we have
proven that
2

E(t, x) = (4t)n/2 H(t) e|x|

/4t

(153)

is a fundamental solution for H. Notice that, for any fixed t > 0, E(t, x) has
support on all of Rn , implying that the heat equation (in contrast to the wave
equation) exhibits infinite speed of propagation. This phenomenon is related
P to
the parabolic scaling of the heat operator, which, in constrast to t2 + i i2 ,
endows time and space with different units. Also notice that E(t, x) is smooth
for t > 0; this fact will lead to instantaneous smoothing for the initial value problem
H = 0 on (0, ) Rn , (0, x) = 0 (x).
Exercise 1. Derive a representation formula for the initial value problem. Why
is it impossible to solve the heat equation backwards in time for arbitrary initial
data?
Exercise 2. Show that the above representation formula for the Cauchy problem
does indeed give a classical solution for sufficiently smooth data. Check that the
correct boundary value is obtained. (This is in contrast to the situation with the
Cauchy-Riemann equations).
Exercise 3. Write down a maximum principle for C 2 solutions to the H = 0 in
the interior of (0, T ] , for open and bounded.

4. HEAT OPERATOR H.

101

Exercise 4. Let : [0, T ] Rn be a solution to the heat equation with decays

rapidly at spatial infinity for fixed time. Show that the energy
Z
e(t) =
|(t, x)|2 dx
Rn

decreases in time. Deduce a uniqueness theorem for the heat equation.

Exercise 5. One often denotes by et 0 the restriction to a hypersurface of
fixed time t > 0 of the (canonical) solution to H = 0 on (0, ) Rn , (0, x) =
0 (x). Show that e(t+s) = et es as time-evolution operators on rapidly decaying,
smooth functions.
4.1. Schr
odinger operator S.
The Schrodinger operator, S = it +
has a fundamental solution which looks, superficially, exactly like that of the Heat
operator,
2

E(t, x) = (4it)n/2 H(t) ei|x|

/4t

(154)
2

Yet, of course, the presence of i in the exponential factor ei|x|

difference.

/4t

makes a world of

Exercise 1.
Show that (for the appropriately chosen branch cut of log) the
locally integrable function E is indeed a fundamental solution for S.
Exercise 2. Similarly to Exercise 5 for the heat equation, one denotes the timeevolution operator for the Schrodinger equation by eit . Show that eit is a unitary
operator in the sense that the quantity
Z
|(t, x)|2 dx
Rn

remains constant in time.

Part 2

Basic Tools of Analysis

CHAPTER 5

Fourier transform

1. Basic properties.
Recall that if f L1 (Rn ), then the Fourier transform F(f ) = f is defined as the
continuous function
Z

f () = f (x)eix dx
(155)
b n ), we have the inversion formula
In case that f L1 (R
f (x) = (2)n

f()eix d,

(156)

whose proof we shall indicate later. To distinguish between the two conceptually,
we refer to the Rn on which f lives as the physical space and the set of points
b n and
on which f lives as frequency space. We denote the frequency space by R
d
endow it with the normalized measure (2)n .
The inversion formula supplies us with a valuable heuristic understanding of what
the Fourier transform does. We see that f (x) can be written as some kind of linear
d describes the
combination of plane waves (x 7 eix ) and the measure f() (2)
n
distribution of f over the space of frequencies. If we view the plane waves eix
as eigenvectors of the translation operators on Rn , we can consider the Fourier
transform an attempt to simultaneously diagonalize translations. Similarly, if we

: j = 1 . . . n},
view the plane waves eix as the eigenvectors of the operators {i x
j
2
which are self-adjoint with respect to the L inner product (when restricted to the
appropriate domain), then we see that differentiation has also been diagonalized by
the Fourier transform.
With these heuristics in mind, we can begin to see how the Fourier transform
might be useful for analysis. For example, if f is concentrated nearby a frequency
b n , we expect f to behave in some ways like the plane waves nearby eix0 .
0 R
For instance, f may admit a bounded, complex-analytic extension into part of
Cn . We also expect that xj f (x) ij0 f (x) so that differentiation becomes a much
easier operation to study. Indeed, when we encounter Littlewood Paley theory later
on, the main idea will be to decompose general functions into frequency localized
components, analyze these components separately, and then reassemble.
105

106

5. FOURIER TRANSFORM

Another important principle regarding the Fourier transform is the duality between
smoothness and decay in physical and frequency space. Intuitively, a function f
whose graph has sudden jumps or spikes in physical space must be composed of
arbitrarily large frequencies, whereas when f is compactly supported, f must be
globally tame. Similarly, when f is very smooth, f can decay at infinity thanks
to interference ( cancelation ) among nearby plane waves in the inversion formula.
There are not one but many formal manifestations of these basic principles all over
Fourier analysis, so let us keep them in mind as we proceed to develop the theory.
The inversion formula takes particularly concrete form in the case of the Gaussian
2
function G(x) = e|x| /2 .
Lemma 1.1. The following calculation holds true for functions of one variable and
a, b R, b > 0,
Z
2

1/2 a2 /4b
e
(157)
eiax ebx =
b

Thus in Rn , for t > 0

Z
Rn

eixy ety =

n/2 |x|2 /4t

e
t

(158)

In particular
F(G)() = (2)n/2 G()

Proof : Make the change of variables in the complex domain, z = b1/2 x

and denote by the contour Im(z) = 2ba1/2 ,
Z
Z
Z
2
2
ea /4b
ea /4b x2
iax bx2
z 2
e e
dx =
e dz =
e
dx
b1/2
b1/2

(159)

a
i,
2b1/2

by
contour deformation argument.
to calculate the integral J =
R a standard
R Now
x2
|x|2
1/2
2
e
e
dx
=

,
we
observe
that
J
=
dx = by passing to polar

R2
coordinates and from this follows (157) . Formula (158) now follows immediately.

We can give another proof of the above identity after reviewing some of the fundamental properties of the Fourier transform.
Proposition 1.2. The Fourier transform is linear and verifies the following simple
properties.
Fourier transform takes translations in physical space Tx0 f (x) = f (xx0 )
into modulations in frequency space F(Tx0 f )() = eix0 f().
Fourier transform takes modulations in physical space M0 f (x) = eix0 f (x)
into translation in frequency space F(M0 f )() = f( 0 ).
Fourier transform takes conjugation in physical space into conjugation and
reflection in frequency, i.e. F(f)() = f().

1. BASIC PROPERTIES.

107

Fourier transform takes convolution in physical space into multiplication

in frequency space, fd
g = f g.
Fourier transform takes partial derivatives in physical space into multiplication in frequency space, F(xj f )() = ij f().
Fourier transform takes multiplication by xj in physical space into the
partial derivative j in frequency space, F(xj f )() = ij f().
We also have the simple self duality relation,
Z
Z
f (x)
g (x)dx = f()g()d
Fourier transform takes scaling in physical space S f (x) = f (x) into
a dual scaling in Fourier space, F(S f )() = n f(/). Observe that
S (f ) preserves size, i.e. kS f kL = kf kL while the dual scaling S f =
n f (x/) preserves mass, that is kS f kL1 = kf kL1 .
Proof Almost all of the above properties reduce to simple identities about exponentials when we specialize to the case where f and g are point masses in the
physical or frequency space1 (and hence are plane waves in the dual space) the
identities themselves may even be regarded as continuous, (bi)linear extensions of
these special cases.
Using these properties, we can give another proof of (157). Thanks to the scaling
identity, it suffices to consider b = 21 and compute the Fourier transform of G(x) =
2
ex /2 . Taking the Fourier transform of the identity, dG
dx = xG(x) and applying
b
the properties above, we see that G satisfies the same differential equation in ,
2 /2
b
and is therefore is of the form G(0)e
. Since we have already shown that
R x2 /2
b
G(0)
= e
= 1/2 , this completes the second proof of (157).

Let G,x0 ,0 (x) = eix0 G((xx0 )/ ) be a translated, modulated, rescaled Gaussian. Then,
Z

n/2 i(0 )x0

F(G,x0 ,0 )() = e
ei y(0 ) G(y)dy

= ()n/2 G ( 0 )
We can interpret this resultas saying that G,x0 ,0 is localized at spatial position x0 ,
with spatial
spread x , and at frequency position 0 with frequency spread
= 1/ . Observe that x 1, so our ability to localize simultaneously in
both physical and frequency space in this way seems to be limited. Surprisingly, this
construction is in some sense the best we can do, and it is our first encounter with
the uncertainty principle of Fourier analysis, which, in its various manifestations,
states that there is a bound on how well one can simultaneously localize in both
frequency and physical space.

1The Fourier transform of a finite, Borel measure is the continuous function

() =
eix d(x)..

108

5. FOURIER TRANSFORM

We now prove our first important manifestation of the duality between smoothness
and decay2.
Proposition 1.3 (Riemann Lebesgue). Given an arbitrary f L1 (Rn ) we have,
kfkL . kf kL1 . Moreover, f() 0 as || .
Proof : Only the last statement requires an argument. Observe that if f
C0 (Rn ), then we can use integration by parts to conclude that f decays rapidly.
Indeed for any multi-index , || = k N,
Z
Z

k
ix
k
f () = i
x e
eix x f (x)dx
f (x)dx = (i)
Z
| f()| .
|x f (x)dx| C
for some constant C . Thus, |f()| . (1 + ||)k which proves the statement in
this case. For general f L1 (Rn ), given  > 0, we can choose g C0 such that
||f g||L1 2 . From the preceding, we know that |
g ()| 2 if || > M = M
sufficiently large and therefore,
sup |f()| kf gkL1 (Rn ) + sup |
g ()| 
||>M

||>M

2. The Schwartz Space and the Inversion Formula

Many of the operations on smooth functions extend naturally to distributions (by
duality with C0 ), and we would like to see how the Fourier transform extends to
distributions. The only possible extension would have to be consistent with the
formula
<u
, >=< u, >
but this formula does not make sense for distributions u D0 (Rn ) and test functions
b n ) because one cannot guarantee that is also compactly supported. In
C0 (R
fact, as a manifestation of the uncertainty principle, both and cannot simultaneously be compactly supported unless = 0. Thus, if we desire a symmetric
theory generalizing the Fourier transform, we are lead to consider a new family of
test functions (and corresponding distributions) which behaves well with respect to
Fourier duality.
Definition 2.1. A function C (Rn ) is said to be rapidly decreasing if for all
multi indices , we have
sup |x (x)| < .
xRn
2We remark that this proposition is not valid for measures which do not have a density
function in L1 for example, the Fourier transform of a point mass is a plane wave, which has
absolutely no decay. Thus, some amount of smoothness is necessary.

2. THE SCHWARTZ SPACE AND THE INVERSION FORMULA

109

This so-called Schwarz space S(Rn ) of rapidly decreasing functions is endowed in

the usual way with a natural Frechet topology. A sequence of functions j converges
to zero in this topology if, for all multi-indices , , x j converges uniformly to
zero. Note that S(Rn ) contains the compactly supported functions C0 (Rn ). Since
C0 (Rn ) is dense in the Lp (Rn ) spaces, for 1 p < , S(Rn ) is also dense in the
Lp spaces. It is also easy to check that C0 (Rn ) is dense in S(Rn ).
We have the following important fact, which is the reason for considering the
Schwarz space in our context:
b n)
Proposition 2.2. The Fourier transform is an isomorphism of S(Rn ) onto S(R
with inverse given by the inversion formula (156). Moreover we have the Plancherel
identity, for all f, g S(Rn ),
Z
Z
n
(160)
(f, g)L2 =
f (x)
g (x)dx = (2)
fgd = (f, g)L2 (Rbn )
Rn

In particular we have the Parseval identity kf kL2 (Rn ) = kF(f )kL2 (Rbn ) .
| and that (x) decays faster than

Proof : Observe that | ()|

= |x\
||n1
b n ). Let Rf (x) =
|x|
. Thus we easily infer that F maps S(Rn ) into S(R
2
f (x) and define T = RF . Observe that T commutes with partial derivatives j
and multiplications by xj . Indeed, for all j = 1, . . . n,

T (j f ) = j (T f ),

T (xj f ) = xj (T f )

(161)

The inversion formula follows from the lemma.

Lemma 2.3. A linear operator T : S(Rn ) S(Rn ) which verifies (161) must be
of the form T = c for some constant c.
Proof : From the commuting property (161), we see that T is linear over the
algebra of polynomial functions. As a consequence of this linearity, we can show
that the value T (x0 ) depends only on the value of (x0 ) at the point x0 . For
example, in one dimension, if vanishes at the point x0 , then we may write
Z 1
(x) = (x x0 )
0 (x0 + t(x x0 ))dt = (x x0 )
0

with S(R). Applying T to this identity, we see that T (x0 ) = 0 as well, and
we may therefore write T = f for some function f possibly depending on .
But f = f does not depend on . If is any other Schwartz function, the linear
combination (x0 ) (x0 ) vanishes at the point x0 , and applying T we conclude
by the same property that f (x0 ) = f (x0 ) at any point x0 at which and are
simultaneously nonzero. It is clear that the function f must be smooth for T to
map S(Rn ) into itself, and in order for T to commute with differentiation, f must
be a constant.
To determine the constants we only have to remark that, in view of lemma 1.1 we

2
have T (G) = (2)n/2 G = (2)n G. Hence the constant c = (2)n which ends
the proof of the inversion formula, and the proposition, for Schwartz functions.

110

5. FOURIER TRANSFORM

The Plancherel and Parseval identities are immediate consequences of the inversion
formula.
Corollary 2.4. The following properties hold for all functions in S(Rn ):.
Z
Z

dx
=
d
Z
Z

dx = (2)
dx
d

(2)n

)()d

(
=

bn
The last convolution being taken with respect to the measure on R
We only completely worked out the proof of the inversion formula for one dimension,
although the same proof requires only a miniscule generalization of the Taylor
expansion to work for general n. The general case can also be deduced from the case
n = 1 as follows: the inversion formula is true for tensor products f1 (x1 ) fn (xn )
and linear combinations thereof, the delta-function is a tensor product 0 (x) =
(x1 ) (xn ), and an arbitrary
function may be written as a linear combination
R
of delta functions f (x) = f (t)(x t)dt.
Exercise 2.5. Make the above argument into a rigorous, self-contained proof of
the inversion formula for Rn by using approximate delta-functions.
It is worthwhile to explore the relationship of the above proof of the inversion formula via the Lemma 2.3 with other proofs of the formula. Just as a linear operator
between vector spaces of finite dimension can be studied via a matrix representation, we can study the operator T in terms of its Kernel K the distribution on
Rn Rn such that
Z
T (x) =

(x0 )K(x, x0 )dx0

In asserting that T (x0 ) depends only on (x0 ), we had proven that applying T was
the same as multiplying by some function; in terms of the kernel, we had established
that
Z
T (x) = (x0 )f (x)(x x0 )dx0
In order for T to commute with differentiation which is not so different from
commuting with translation we concluded that f (x) was a constant.
But we can see directly that an equivalent formulation of the inversion formula is
the distribution-theoretic identity
Z
0
d
= (x x0 )
(162)
ei(xx )
(2)n
n
c
R
which is really the special case of the inversion formula for a -function. Viewing the
integral on the left hand side as an inner product, the above identity can be regarded

2. THE SCHWARTZ SPACE AND THE INVERSION FORMULA

111

as a statement that the plane waves eix are in some sense orthonormal as
x varies. Thus, in writing
Z
Z
d
f (t)(x t)dt = f (x) = f()ei
2
we might regard the Fourier transform as analogous to a change of orthonormal
basis from -functions in physical space to plane waves so that the Plancharel
and Parseval identities should follow immediately. For example, in one dimension,
d
with distinct
distinct plane waves are eigenfunctions for the self-adjoint operator i d
eigenvalues, and therefore should be orthogonal as a matter of general principle
this argument can be made rigorous to show the above distribution vanishes away
from x x0 = 0, as was essentially done in the previous proof through Taylor
expansion and linearity over the polynomial ring. Let us mention several other
ways to establish this identity and hence prove the inversion formula.3
It suffices to show that

Z
1
eix d = 0 (x)
(2)n
as a distribution in the variable x on Rn this translation invariance corresponds
d
to T commuting with dx
in the previous proof. By viewing the above distribution
as a tensor product, it would suffice to consider the case n = 1, but let us refrain
from doing so. Recall that every distribution supported at the origin is a finite
linear combination of derivatives of (x), and hence the function itself is, up to a
constant, the only distribution homogeneous of degree n supported at 0 these
facts are easily established by Taylor expansion. As the integral on the left hand
side is clearly homogeneous of degree n in x, we will have proven the identity up
to a constant if we can show that
Z
eix d
(163)
is supported at the origin in this precise sense, a plane wave 7 eix is zero on
average.
Heuristically, let us outline a few ways to perform this calculation. Pretend that
the integral 163 is a classical integral and that x 6= 0 is fixed. If we view the plane
waves as eigenfunctions of differential operators, we may integrate in by parts
using the identity
Z
Z
1
eix d =
eix d,
|x|2
or alternatively we can rotate to the case x = |x|(1, 0, . . . , 0) and integrate by
1 eix
parts using the identity i|x|
1 . If we would rather view the plane waves as
eigenfunctions of translation operators, we may show the integral is zero for x 6= 0
by translating in the variable
Z
Z
Z
0
0
eix d = ei( )x d = ei x eix d
3As remarked, the plane waves may be viewed as eigenfunctions of the commuting family of
d
self-adjoint operators f 7 i dx
i f . Dually, the delta functions, which are similarly orthonormal
R
in the sense that 0 (y x)0 (y 0 x)dx = (y y 0 ), can be viewed as eigenfunctions for the
commuting family of operators f 7 xi f .

112

5. FOURIER TRANSFORM

by some appropriate frequency 0 depending on x. Of course these are all heuristic

lines of attack which treat the integral as though it were a classical one, and we
cannot treat x as a fixed point, but we can make these arguments rigorous by fixing
a test function localized around x and produce a complete argument akin to the
following.
n
Proof Let C
0 (R ) be supported away from the origin. Let be a smooth,
rapidly decreasing function with (0) = 1.

Z Z
Z Z
ix
ix
(x)e dx ()d
(x)e dxd = lim+
0
Z

Z
ix
= lim+ (x)
e ()d dx
0

This limit is zero by the dominated convergence theorem (the d integral is a rapidly
decreasing function of x and (x) 0).
Without assuming anything about the support of , the above proof would have
established the Inversion Formula directly with the constant had we chosen a
(such as a Gaussian) whose Fourier transform was understood. Indeed, if we know
the Inversion Formula for a Gaussian, the inversion formula is true for rescalings
and translates of Gaussians. As a limiting case, the Inversion Formula holds for
any
R function, and hence for an arbitrary function by the decomposition f (x) =
f (t)(x t)dt.
Exercise 2.6. Create a self-contained, direct proof of the Inversion Formula from
the case of a Gaussian.
In the case n = 1, there is Ralso a more complex-analytic way to evaluate the

distribution-theoretic integral eix d, which not only determines the constant4,

but directly relates the 2 in the inversion formula to the 2 in the Cauchy Integral formula (the circumference of a circle). Basically, one uses complex-analytic
1
extensions of the plane waves and the formula (129) for x+i0
in order to decompose

eix d =

eix d +

eix d

0
i(xiy)

= lim
e
d +
y0+


1
1
=i
+
x i0 x + i0
= 2(x)

ei(x+iy) d

4Incidentally, this argument also determines the integral of a Gaussian indirectly.

3. EXTENSION OF THE FOURIER TRANSFORM

113

3. Extension of the Fourier Transform

As a corollary to the Parseval and Plancherel formulas we can extend our definition
of the Fourier Transform to L2 (Rn ) functions by a simple density argument. Indeed
for any u L2 we can choose a sequence of S(Rn ) L1 functions uj converging to
u in the L2 norm. By Plancherel, kF(uj ) F(uk )kL2 . kuj uk kL2 . Hence the
sequence F(uj ) forms a Cauchy sequence in L2 and therefore converges to a limit
which we may call u
. Clearly, this definition does not depend on the particular
sequence. Moreover one can easily check that the Parseval identity extends to all
L2 functions. Since F is therefore an isometry onto its image, its image must be
closed, but then the image must be all of L2 since the image also contains the
Schwartz functions. Thus we have proved,
Theorem 3.1. The Fourier transform is an isometry of the Hilbert spaces L2 (Rn )
b n ).
and L2 (R
We can extend the Fourier transform even further to a special class of distributions
defined on Rn .
Definition. We define a tempered distribution to be an element in the dual space
of the Schwarz space.
Example. While ex is not a tempered distribution on R because it grows too
d
sin(ex ) is an example. Here we make the
quickly, the function ex cos(ex ) = dx
usual identification of a function with a distribution.
Note that the tempered distributions embed continuously into the space of ordinary
distributions defined earlier. In analogy with the properties of ordinary distributions, for every tempered distribution u, there exists a natural number N and a
constant C = C, such that
X
| < u, > | C
sup |x |, S(Rn )
||, ||N

We can now define the Fourier transform of a tempered distribution; namely,

<u
, >=< u, > .
One easily checks that this defines a tempered distribution u
for every tempered u.
Moreover, all the properties of the Fourier transform, which have been verified for
Schwartz functions in S(Rn ) can be easily extended to all tempered distributions.
In particular, since all Lp spaces are included in S 0 (Rn ) we have a definition of
Fourier transform for all such spaces. Observe that, in the case of L1 this definition
coincides with the definition given in (155).
The following simple, and very useful, formulas for the Fourier transform of the
Dirac measure 0 now make sense:
F(0 ) = 1,

F(1) = (2)n 0

(164)

114

5. FOURIER TRANSFORM

Observe also that if we denote by sign(x) the one dimensional tempered distribution
x
given by the locally integrable function |x|
we have,
d
sign()
= 2ipv()

(165)

d
Indeed sign0 (x) = 20 . Hence, i sign()
= 2. Therefore, for any rapidly decreasing
, we have
Z
Z
c

i sign(x)x(x)dx
= 2(0)
= 2 (x)dx
d
Also, observe that sign(x)
is an odd distribution so that whenever (x) = (x)
d >= 0. Now given a general test function
is an even test function, then < sign,
, write = 21 ((x) + (x)) + 12 ((x) (x)) = ev + odd . Hence, from the
preceding, we infer that


d >=< sign,
d 1 odd >= 2i < pv( 1 ), >
< sign,

as desired.
This fact may also be observed more directly by evaluating the distribution-theoretic
integral
Z
sign(x)eix dx

along the same lines as the complex-analytic proof of the Fourier Inversion Formula
outlined in the previous section.
Exercise 1. Show that the only harmonic functions which are tempered distributions are polynomials.
Exercise
2. Let f (x) = e|x| L1 (R). Compute f() (and hence f(0) =
R
f (x)dx = 2) using the fact that f satisfies a simple, second order differential
equation. Comment on the precise amounts of regularity and decay of f and f
and how they can be anticipated from the physical space representation. Note that
f continues meromorphically into the complex plane by considering correlations
against complex plane waves x eizx , z C, you can anticipate the location of
the poles from the form of f in physical space.
Exercise 3. Suppose that u is a tempered distribution which is invariant under
translation by a subgroup S of Rn for instance u could be periodic or a function
of less than n of the variables. Why can we assume S is closed? Show that the
Fourier transform u
is supported on the annihilator subgroup S of plane waves
which are invariant under S.
S = { | eix = 1 for all x S}

4. Uncertainty principle and localization

On the real line let the operators X, D defined by,
Xf (t) = tf (t),

Df (t) = if 0 (t)

4. UNCERTAINTY PRINCIPLE AND LOCALIZATION

115

Observe that,
[D, X]f = DXf XDf = if
This lack of commutation is responsible for the following:
Proposition 4.1 (Heisenberg uncertainty principle). The following inequality holds,
kXf kL2 kDf kL2
Proof :

1
kf k2L2
2

Observe, using the commutator relation above,

0 k(aX + ibD)f k2L2 = a2 kXf k2L2 + b2 kDf k2L2 abkf k2L2

Now, miniize the right hand side by choosing a = kDf kL2 and b = kXf kL2 .
The uncertainty principle, which can informally be described as5 x 1/2,
places a limit on how accurately we can localize a function, or any other relevant
object, simultaneously in both space and frequency. Let us investigate these localizations in more detail.

4.2. Physical space localization.

If we want to localize a function f to a
domain D Rn we may simply multiply f by the characteristic function D . The
problem with this localization is that the resulting function D f is not smooth even
if f is. To correct for this we choose D C0 (D) in such a way that D is not too
different from D . In the particular case when D is a ball B(x0 , R) centered at x0
we can choose D to be 1 on the ball B(x0 , R) and zero outside the ball B(x0 , 2R).
This leads to the following bounds for the derivatives of D ,
| D | . R|| .
In general given a domain D to which we can associate a length scale R ( such as
its diameter or distance from a fixed point in its interior), we can find a function
D C0 (D) such that,
| D | . R|| ,

(166)

for all multi-indices Nn .

A general remark: derivative estimates of the form (166) are very common in
analysis and almost always arise when the function obeying the estimates comes
from a rescaled version of another function, whose derivatives are simply bounded.
That is why the particular exponent which appears is consistent with dimensional
analysis.
5This notation comes from quantum mechanics, where one normalizes kf k
L2 = 1 and in|f()|2

terprets both |f (x)|2 and 2 as probability densities over states in position and momentum
space respectively. In this setting the Heisenberg uncertainty principle gives a lower bound for
the product of the standard deviations of position and momentum.

116

5. FOURIER TRANSFORM

4.3. Frequency space localization. . Just like before we can localize a

function to a domain D Rn in frequency space by F 1 (D f). Once more, it
often pays to use a smoother version of cut-off, thus we set,

Pd
D f () = D f ().
PD is an example of a Fourier multiplier operator, that is an operator of the type:

Td
m f () = m()f ().

(167)

with m = m() a given function called the symbol of the operator. Clearly,
Z
Tm f (x) = f K(x) = f (x y)K(y)dy
(168)
where K, the kernel of T , is the inverse Fourer transform of m,
Z
n
K(x) = (2)
eix m()d
Clearly any linear differential operator P () is a multiplier with symbol P (i).
To compare the action, in physical space, between rough and smooth cut-off operators it suffices to look at the corresponding kernels K. Let I = [1, 1] R and
I the rough cut-off (while ignoring the 2 constants). The corresponding kernel
Z 1
sinx
K(x) =
eix d = 2
x
1
decays very slowly as |x| . Because of this the operator
Z
sin(x y)
F 1 (I f)(x) = 2
f (y)dy
(x y)
has very poor localization properties. Indeed, the operator spreads around to the
whole R any function supported in some set J R. This situation corresponds to
a perfect localization in frequency space and a very bad one in physical space. The
exact opposite situation occurs when we do the rough cut-off localization I f in
physical space. On the other hand, when we use a smooth cut-off I in frequency
space, then the frequency cutoff operator PI f = F 1 (I f) is of the form f K f
where the kernel
Z
K(x) =
eix I ()d
R

is rapidly decreasing. In this case, we can prove that

Lemma 4.4. Let I = [1, 1], I a smooth cut-off on I and PI f = F 1 (I f). Then,
if f is any L2 function supported on a set D R,

j
|PI (f )(x)| . Cj kf kL2 1 + dist(x, D)
for all j N .
Thus PI spreads the support of any function f by a distance O(1) plus a rapidly
decreasing tail.

4. UNCERTAINTY PRINCIPLE AND LOCALIZATION

117

Exercise. Show that there exists no non-trivial function such that both and
F() are compactly supported.
The above discussion can be easily extended to higher dimensions. In particular
we can get a qualitative description of functions in Rn whose Fourier support is
restricted to a ball BR = B(0, R) centered at the origin. Let R be a smooth
cut-off for BR . More precisely we take it of the form
R () = (/R)
a smooth cut-off for B1 , i.e. is smooth, identically equal to 1 on B1 and
supported, say, in B2 . It is easy to check the estimate for any multi-index ,
sup | R ()| c R|| ,

with a constant c depending only on the fixed and its derivatives.

If f is a function whose Fourier support is restricted to BR then f = R f. Hence,
Z
f (x) =
f (y)KR (x y)dy
(169)
Rn

where KR (x) = F

(R ).

Lemma 4.5. The kernel KR (x) verifies the estimates,

|x KR (x)| CN, R|| Rn (1 + |x|R)N ,

(170)

for all R > 0, any N N and multi-index N , with a constant CN, which
depends only on N , , dimension n and choice of the fixed test function .
Proof Indeed, integrating by parts,
Z
Z
1
ix
ix
(e )R ()d
KR (x) =
e R ()d =
ix
n
n
R
ZR
1
ix
=
e R ()d
Rn ix
Thus, for any , || = N , denoting by |BR | = cn Rn the volume of BR ,
Z
|x KR (x)|
| R ()| c RN |BR | cn c RN +n
n

Rn
N

Hence, |KR (x)| CN R (|x|R) , for a constant CN which depends on N , n and

the fixed . On the other hand, for |x| R1 , |KR (x)| . Rn . Hence, for every
N N,
|KR (x)| . CN Rn (1 + |x|R)N .
It is easy to check also that each derivative of KR costs us a factor of R, proving
(170).
Now back to (169) we have
Z
| f (x)| = |
f (y) KR (x y)dy| .
Rn

R||+n

|f (y)|(1 + R|x y|)N dy

Rn

R||+n kf kL1

118

5. FOURIER TRANSFORM

Also, by H
olders inequality with

1
p

1
p0

= 1,
0

| f (x)| . kf kLp k KR kLp0 . R|| Rn Rn/p kf kLp

. R||+n/p kf kLp
We have just proved the following version (Lp L version) of the very important
Bernstein inequality,
Proposition 4.6. Assue that f is an Lp function which has its fourier transform
supported in the ball BR = B(0, R). Then f has infinitely many derivatives bounded
in L and we have,
k f kL (Rn ) . Rn/p+|| kf kLp .
Remark. Observe that the proposition could have been proved by reducing it
to the particular case of R = 1. More precisely assume that the result is true
for R = 1 and consider a function f whose Fourier transform is supported in BR .
Let g(x) = Rn f (R1 x) and observe that, supp g() = supp f(R) B1 and
therefore we have, k gkL (Rn ) . kgkL1 = Rn Rn/p kf kLp . Thus, k f kL (Rn ) .
Rn/p+|| kf kLp .
As we will very often see during these notes, dimensional analysis can be used to
rapidly figure out the exponent which arises above. For example, if we regard the
spatial variables as having a scale L, so that the volume element dx has scale Ln ,
then the frequency variables have units R L1 . We see that k f kL has a scale

1/p
R
L|| and kf kLp =
|f (x)|p dx
has scale Ln/p thus the power of R appearing
makes the estimate consistent with dimensional analysis.
Qualitatively, the Berenstein estimate embodies some basic intuition regarding Lp
norms and the Fourier Transform. In a sense, the higher Lp norms of a function such
as the L norm control a functions ability to blow up in a localized region of space,
whereas lower Lp norms control growth at infinity. The Berenstein estimate says
that if a frequency-localized function does not grow at infinity (i.e. has a bounded
Lp norm, 1 p < ), then due to the absense of high-frequency components, the
function is prohibited from having localized singularities as well (i.e. the L norm
is bounded).
A main reason we are interested in such estimates is not that we often run into
functions with compactly supported Fourier transforms, but rather that we often
decompose more general functions into a sum of parts which are frequency localized.
We shall return to this idea in our study of Littlewood-Paley theory.

5. Applications to PDE
Consider the initial value problems for our basic PDEs in R Rn , written in the
form

5. APPLICATIONS TO PDE

119

t = ,

)0, x) = f (x)

(171)

t = i,

(0, x) = f (x)

(172)

t2 = ,
t2 = ,

(0, x) = f (x),
(0, x) = f (x),

t (0, x) = g(x)
t (0, x) = g(x)

(173)
(174)

In each of these cases we can write down solutions using the Fourier transform
method. More precisely we can take the Fourier transform of each equation, set
Z

(t, ) = eix (t, x)dx

and solve the resulting differential equation in t. Once this is done we obtain our
solution simply using the inverse Fourier transform, i.e.
Z
) d
(t, x) =
eix (t,
(2)n
In the case of the heat equation (171) we derive,
Z +
2
d
(t, x) = (2)1
eix et|| u
0 ()
n
(2)

while in the case of the Schr

odinger equation,
Z +
2
d
1
(t, x) = (2)
eix eit|| u
0 ()
n
(2)

(175)

(176)

Exercise 1. Show how to relate the formulas (175) and (176) to the physical space
formulas (153) and (154).
In the particular case of the wave equation (173) we derive,
Z

d
sin t||
(t, x) =
eix cos t||f() +
g()
||
(2)n
Rn

(177)

Exercise 2. Derive a formula similar to (177) for the Laplace equation (174).
Show, using these formulas that (173) has solutions for all f, g S(Rn ) while (174)
does not. Show however that if we only prescribe (0, x) = f (this is the Dirichlet
problem for the Laplacian t2 + in Rn+1 ), then the problem has a unique solution
, which decays to zero as |t| + |x| , for all functions f S(Rn ).
Exercise 3. Show, in the special case of dimension 1 + 3, how to pass from formula
(177) to the Kirchoff formula (146)
Z
Z


1
1
(t, x) = t (4t)
f (y)da(y) + (4t)
g(y)da(y)(178)
|xy|=t

|xy|=t

which is consistent with the formulas derived in the previous chapter, based on the
explicit calculation of the fundamental solution.

120

5. FOURIER TRANSFORM

It is interesting to make a comparison between the Fourier based formula (177) and
the Kirchoff formula (178). Observe that it is quite easy, using Parseval, to derive
the global energy identity from (177),
Z
Z


(|t |2 + ||2 ) =
|f |2 + |g|2 dx
Rn

Rn

while obtaining such an identity from (178) seems not at all obvious, in fact quite
implausible. On the other hand (178) is perfect for giving us domain of influence
information. Indeed we read immediately from the formula that if the data f, g is
supported in ball Ba = {|x x0 | a} than (t, x) is supported in the ball Ba+|t|
for any time t. This fact, on the other hand, does not at all seem transparent6
in the Fourier based formula7 (177). The fact that different representations of
solutions have different, even opposite, strengths and weaknesses has important
consequences for constructing parametrices, i.e. approximate solutions, for more
complicated, linear variable coefficient or nonlinear wave equations. There are two
type of possible constructions, those in physical space, which mimic the physical
space formula (178) or those in Fourier space, which mimic formula (177). The first
are called Kirchoff-Sobolev, or Hadamard parametrices while the second are called
Lax parametrices, or, more generally, Fourier integral operators.

6 Support information can ve derived by Paley-Wiener type results.

7Support information can be extracted however from (177) using the Paley-Wiener method,

see [?] vol. 1.

CHAPTER 6

Basic Functional Inequalities

1. Basic interpolation theory

1.1. Introduction. Consider the Fourier transform as a linear operator F :
b n ). According to the Plancherel identity we have kF(f )kL2 kf kL2 .
S(Rn ) S(R
On the other hand, we have kF(f )kL kf kL1 . Can we get other bounds of
the type kF(f )kLq . kf kLp ? It turns out that such estimates can be easily
established by interpolating between the two estimates mentioned above. Complex
interpolation allows us to conclude an Lp to Lq estimate for any values of p and
q such that p1 + q 1 = 1 and q 2. This is known as the Young-Hausdorff
inequality. Interpolation theory is particularly useful for linear multiplier operators
of the form

Td
m f () = m()f ()
with bounded multipler m. In view of Parsevals identity it is very easy to check the
L2 L2 estimate, kTm f kL2 . kf kL2 . To obtain additional
R estimates we typically use
the integral representation (168) Tm f (x) = f K(x) = f (x y)K(y)dy where K
is the inverse Fourier transform of m. If, for example, we can establish that K L1
than we easily deduce that kTm f kL1 . kf kL1 , since kf KkL1 kf kL1 kKkL1 .
We thus have both L1 L1 and L2 L2 estimates for Tm . and it is tempting to
conclude we might have an Lp Lp estimate for all 1 p 2. Such an estimate
is indeed true and follows by interpolation. On the other hand, if we can establish
that K L then kf KkL . kf kL1 and thus we can prove, by interpolation, the
same Lp Lq estimate as in the Hausdorff-Young inequality.
1.2. Review of Lp spaces. Given a measurable subset Rn the space
L (), 1 p < , consists in all measurables functions f : C with finite Lp
norm,
Z
1/p
p
kf kLp =
|f (x)| dx
< .
p

The space L () consists of all measurable functions, bounded almost everywhere,

that is,
 Z

kf kL = inf : (|f (x)| > )dx = 0 = ess sup |f | < .

For all values of 1 p the spaces Lp () are Banach spaces. The theory of
Lp spaces generalizes when we replace the Lebesgue measure dx with a general,
121

122

6. BASIC FUNCTIONAL INEQUALITIES

positive measure

The following is called Holders inequality

kf gkLp kf kLq kgkLr ,

(179)

whenever 1/p = 1/q + 1/r. The relationship between the exponents is necessary
so that both sides are homogeneous of degree p1 in the measure. In particular, for
p = 1,
kf gkL1 kf kLq kgkLq0
where q 0 verifying

1
q0

= 1

1
q

is the exponent dual to q. This inequality implies

0

q
that we can identify each element g
R L with the bounded, linear functional on
q
the Banach space L given by f 7 f (x)g(x)dx. For all 1 q < the space
0
Lq () is dual to Lq () in the sense that the above identification is an isometry (in
particular, every bounded linear functional on Lq arises this way for a unique g),
while the dual of L () includes L1 (), but is vastly larger. Often taking the role
of L is the space C0 (Rn ) of continuous functions vanishing at infinity (since they
constitute the closure of C0 in the L norm), whose dual space is the set of finite,
Borel measures on Rn .

The different Lp norms measure different aspects of the size of a function. An

estimate of a higher Lp norm such as kf kL . 1 guarantees that |f | does not
become too large locally, whereas an estimate of a lower Lp norm such as kf kL1 . 1
controls the behavior of f at infinity. The space L2 () is especially important
because of its self-duality and its Hilbert space structure given by the inner product
Z
< f, g >=
f gdx

Exercise.

Show that C0 () is dense in Lp () for all 1 p < .

Given a measurable function f and a positive number , denote by (f, ) the

distribution function of f defined by
(f, ) = |{x : |f (x)| > }|
For 1 p < we have Chebyschevs inequality
(f, ) p kf kpLp

(180)
p

which quantitatively expresses the fact that the upper contour sets of an L function
have finite measure. It is helpful (at least as a mnemonic) to note that both sides
have the same units since f and have the same units.
Proof
Z 

Z
(f, ) =

(|f (x)| > )d(x)

|f (x)|p
p

(|f (x)| > )d(x) p kf kpLp

1There are some complications, however, when the whole space cannot be written as a countable union of finite -measure subsets. This will not concern us in these notes.

1. BASIC INTERPOLATION THEORY

123

We can
the Lp norm of f in terms of its distribution function. Indeed, the inteR write
p
gral |f | is the measure of the region bounded by the graph {(, x) : 0 < < |f (x)|p },
hence
Z
Z Z
Z
p
p
|f (x)| dx =
(|f (x)| > )ddx = p
p1 (f, )d
0
0
(181)
where the last integral is obtained from the substitution = p .
A measurable function f : C is said to be simple if its range consists of a
PN
finite number of points in C, that is f =
i=1 ai Ai for ai C and Ai
measurable. In this section we denote by S() the set of all simple functions in .
Recall that S() is dense in Lp () for all 1 p . The proof typically involves
approximating a fixed f (x) with linear combinations of characteristic functions
(f (x) E ), and letting the collection {E } tend towards a fine and complete
partition of C.
Exercise. Let f (x, y) be a measurable function on 1 2 Rn1 Rn2 . Prove
the following version of the Minkowskis inequality,
Z
Z
k
f (x, y)dykLpx (1 )
kf (x, y)kLpx (1 ) dy
2

for 1 p both by using duality and without doing so.

1.3. Three lines lemma. The method of analytic interpolation, for linear
operators acting on Lp spaces, is based on a variant of the maximum modulus
theorem for a strip-like domain called the three lines lemma. Consider the striplike domain,
D = {z C : 0 < Re(z) < 1} .
We will denote by ABC the set of bounded continuous functions on the closure of
D which are analytic on D.
Lemma 1.4 (Three lines lemma). Let f ABC such that
|f (0 + ib)| M0 ,

|f (1 + ib)| M1 ,

for all b R. Then for all 0 < a < 1 and b R,

|f (a + ib)| M01a M1a .
Remark. Recall that log |f (z)| is a subharmonic function when f is holomorphic
(and nontrivial). The bounds on the two strips guarantee that log |f (z)| obtains
lesser values than the harmonic function (a + bi) = (1 a) log M0 + a log M1 on
the boundary of D, and the conclusion of the Three Lines Lemma asserts that
log |f (a + bi)| (a + bi) within the domain D, however we cannot simply apply
the weak maximum principle, since it is not quite valid for unbounded domains (as
iz
the example z eie on D shows).
Proof First, by replacing f with

f (z)
M01z M1z

if necessary, we can assume that M0 =

M1 = 1 and it suffices to show |f (z)| 1 throughout. When f decays to 0

124

6. BASIC FUNCTIONAL INEQUALITIES

as |=z| , then one can simply apply the usual maximum modulus principle
to a sufficiently large subset of D to conclude |f (z)| 1 throughout. If this is
not the case, then (because we have assumed already that |f (z)| does not grow
substantially as |=z| ) we can apply the same argument to the approximation
F (z) = e(1z)z f (z) (which does decay for large |=z|) and conclude
|f (z)| = lim |F (z)| 1


throughout D.

1.5. Stein-Riesz-Thorin interpolation.

Definition 1.6. We say that a family of linear operators Tz , indexed by z D, is
an analytic family of operators if,
(1) Tz maps simple functions into measurable functions;
R
(2) For any pair of simple functions f, g S(), the map z 7 g(x)Tz f (x)dx
belongs to ABC .
Remark 1.7. The reason for choosing simple functions as test functions in the
previous definition is because they are easy to manipulate and they make a dense
set in Lp for every p [1, ).
Theorem 1.8. Let Tz be an analytic family of operators and assume there are
positive constants M0 , M1 such that, for every b R,
kTib f kLq0 (d) M0 kf kLp0 (d) ,

kT1+ib f kLq1 (d) M1 kf kLp1 (d) ,

with 1 q0 , p0 , q1 , p1 . Then, for z = a + ib D, Tz extends to a bounded

operator from Lp (d) to Lq (d) and
kTz f kLq (d) M01a M1a kf kLp (d) ,
where

1
1a
a
=
+ ,
p
p0
p1

1a
1
a
=
+ .
q
q0
q1

Proof : By changing the measures and themselves by a scalar multiple, we

can assume that M0 = M1 = 1. Adopting a bilinear formulation we have to prove
that
Z



 g(x)Tz f (x)dx 1
(182)


for every pair of simple functions f, g with kf kLp = kgkLq0 = 1. Fix such a pair
f, g and consider the related (analytic) families of simple functions
p

fz (x) = |f (x)| p(z) 1 f (x),

q0

gz (x) = |g(x)| q0 (z)

g(x),

with the exponents,

1
1z
z
=
+ ,
p(z)
p0
p1
We can easily check that
|fib | |f |p/p0 ,

|f1+ib | |f |p/p1 ,

1
1z
z
=
+ 0.
q 0 (z)
q00
q1
0

|gib | |g|q /q0 ,

|g1+ib | |g|q /q1 .

1. BASIC INTERPOLATION THEORY

125

Here we use the convention that 1/ = 0, and in particular if p0 = p1 = then

p = p(z) = and fz f , similarly q00 = q10 = then q 0 = q 0 (z) = and gz g.
It is immediate to verify that kfz k Re(p(z)) = kf kLp = 1 and kgz k Re(q0 (z)) =
L
L
kgkLq0 = 1.
Now consider the map defined on D,
Z
h(z) = gz (x)Tz fz (x)dx.
It is not difficult to see from our construction and the linearity and analyticity
properties of Tz , that h ABC . By hypothesis (and Holder) we have that |h(ib)| 1
and |h(1 + ib)| 1 for every b R. It follows from the three-lines lemma that
|h(z)| 1 and in particular 182.

1.9. Young inequality. We often need to estimate integral operators of the

form2.
Z
T f (x) = k(x, y)f (y)dy,
(183)
The simplest result of this type is given by Youngs theorem below.
Theorem 1.10 (Young). Let k(x, y) be a measurable function and assume that for
some 1 r we have
sup kk(x, )kLr 1,
x

sup kk( , y)kLr 1.

y

Then, for 1 p r and

1+

1
1 1
= + ,
q
r
p

(184)

we have

Proof :

kT f kLq kf kLp .

(185)

kT f kL kf kLr0 .

(186)

By H
older inequality,

On the other hand the dual operator T has the same form as T ,
Z
T g(y) = k(x, y)g(x)dx,
and hence,
kT gkL kgkLr0 ,
which by duality gives the other endpoint
kT f kLr kf kL1 .

(187)

Now, we can use Theorem 1.8, with Tz T , to interpolate between (186) and (187)
and obtain (185).
2In fact, the Schwartz Kernel theorem states that every continuous linear map from
C0 (1 ) D0 (2 ) is of the form (183) for some distribution k(x, y) D0 (1 2 )

126

6. BASIC FUNCTIONAL INEQUALITIES

As an immediate consequence, when k is translation invariant, k(x, y) = k(x y),

we obtain the well known estimate for convolutions:
kk f kLq kkkLr kf kLp ,

(188)

whenever the exponents 1 p, q, r satisfy (184). Note that this relationship

between the exponents is necessary so that both sides will have the same degree of
homogeneity in the measure.
r 1 fails and similarly for kkkL Lr ,
Exercise.
More generally, when kkkL
x Ly
y
x
one can reduce to the hypotheses of Theorem (1.10) by changing the measures in
the x and y variables. By doing so, what more general Young inequality do you
obtain?

Exercise. Prove, using complex interpolation, the Hausdorff-Young inequality for

the Fourier transform F,
kF(f )kLq . kf kLp ,

for all

q 2,

1/q + 1/p = 1.

1.11. Marcinkiewicz interpolation. A slightly weaker condition than Lp

integrability for a function f is the so called weak-Lp property.
Definition 1.12. For 1 p < , we say that f belongs to weak-Lp if (f, ) .
p , for every > 0. If p = we let weak-L coincide with L .
By Chebyschevs inequality (180), any function in Lp is also in weak-Lp . The following is the simplest example of real interpolation. It applies to sublinear operators,
that is,
|T (f + g)(x)| . |T f (x)| + |T g(x)| ,
Theorem 1.13. Consider a sublinear operator T mapping measurable functions on
X to measurable functions on Y . Assume that T maps Lpi (X) into weak-Lpi (Y ),
with bound
(T f, ) . pi kf kpLipi ,
for i = 1, 2 and 1 p1 < p2 . Then, for any p, p1 < p < p2 , T maps Lp (X)
into Lp (Y ), with the bound
kT f kLp . kf kLp ,
.
Proof : Given f Lp (X) and > 0 we write f = f + f , where f (x) =
f (x) (|f (x)| > ) and f (x) = f (x) (|f (x)| ) are cutoffs of f . In particular
f Lp1 and f Lp2 by H
olders inequality.
Consider first the case p2 < . By our assumptions on T we have
(T f, 2) . (T f , ) + (T f , ) . p1 kf kpL1p1 + p2 kf kpL2p2 .

(189)

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

127

Using the distributional characterization of kT f kLp and Fubinis theorem, we infer

that
Z
Z
Z
p
|T f (x)| dx .
|f (x)|p1 pp1 1 ddx +
|f (x)|p2 pp2 1 ddx.
|f (x)|

0<<|f (x)|

R |f (x)|

But 0
pp1 1 d ' |f (x)|pp1 , since p p1 1 > 1, and
|f (x)|pp2 , since p p2 1 < 1, and the conclusion follows.

R
|f (x)|

pp2 1 d '

In the case of p2 = the proof is actually simpler. We only have to observe that
|T f (x)|  implies |T f (x)|  , since |T f (x)| . kf kL . Hence we can
replace (189) by
(T f, C) . (T f , ) . p1 kf kpL1p1 ,
where C is some positive constant, and the proof proceeds as before.

2. Maximal function, fractional integration and applications

2.1. Maximal Function. A function f which is in Lp (Rn ), for some 1 p
, may possess very bad regularity properties. Given > 0, the set of points x
where |f (x)| > may merely be any measurable set (with finite measure if p < ).
It is often desirable to replace f with a positive function which has (almost) the
same integrability properties of f but better local regularity. This is achieved by
considering maximal averages of f .
Definition 2.2. Given a measurable function on Rn we define its maximal function
by
Z
1
Mf (x) = sup
|f (y)|dy.
xB |B| B
Here the supremum is taken over all possible euclidean balls B containing x (not
only those centered at x).
Remark 2.3. It follows immediately from the definition that Mf is lower semicontinuous. Indeed, for every 0, the sets E = {x Rn : Mf (x) > } are always
open: if x E then there exists a ball B containing x such that
Z
1
|f (y)|dy >
(190)
|B| B
and this also means that Mf (y) > for every y B, hence B E .
By the triangle inequality we also see that f 7 Mf is a subadditive operator,
M(f + g)(x) Mf (x) + Mg(x).

(191)

The averaging process may improve local regularity, but, because of the supremum,
it is not clear whether Mf preserves the integrability properties of f . If f is
essentially bounded, then Mf is bounded and
kMf kL kf kL .

(192)

128

6. BASIC FUNCTIONAL INEQUALITIES

But, if f is an integrable function, it doesnt follow that Mf is integrable. Take

for example f = B L1 , the characteristic function of a ball, then Mf (x) &
(1 + |x|)n which barely fails to be in L1 . Fortunately, the maximal function still
retains most of the information about the integrability properties of f .
Theorem 2.4. If f L1 then Mf is weakly in L1 , in the sense that for > 0 we
have
1
|E | = (Mf (x), ) . kf kL1
(193)

If f Lp with 1 < p then Mf Lp and we have

kMf kLp . kf kLp .

(194)

Proof : The second part of the statement follows from the first and the L boundedness of the maximal operator by Marcinkiewicz interpolation, Theorem 1.13.
Hence, we only need to prove (193).
Let f L1 and fix > 0. By the discussion in Remark 2.3 we can find a family
of balls B = {B}, such that E = BB B and each ball B satisfies (190). If these
balls were all disjoint then it would be easy to conclude, since in that case
Z
Z
X
1
1X
|f (y)|dy
|f (y)|dy.
|E |
|B| <

Rn
B
B

BB

In general these balls are not disjoint and we have to be more careful.
Let K be a compact subset of E , then it is possibile to select a finite subfamily B 0
of balls in B that cover K. Using the covering lemma proved below3, Lemma 2.5,
we can select among the balls in B 0 another finite subfamily B 00 made of disjoint
balls (which may no longer cover K) such that
X
|B 0 B0 B 0 | .
|B 00 |.
B 00 B00

Then, proceeding as above, we find

1
kf kL1 ,

and taking the supremum over all possible compact sets K we finally obtain (193).
|K| .

Lemma 2.5. Let B1 , . . . , BN be a finite collection of balls in Rn , then it is possible

to select a subcollection Bj1 , . . . , BjM , M N , of disjoint balls such that
M
 N
 X
j=1 Bj  .
|Bjk |.
k=1

Proof : We can assume that the balls Bj = B(xj , rj ) are labeled so that the radii
are in nonincreasing order, r1 r2 rN .
3This is sometimes known as the Vitali Covering Lemma

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

129

Take j1 = 1, so that Bj1 is the ball with largest radius. Then by induction, define
jk+1 to be the minimum index among those of the balls Bj which dont intersect
with the previously chosen balls Bj1 , . . . , Bjk ; if there are no such balls then stop
at step k.
With this construction we have that each ball Bj intersects one of the chosen balls
Bjk with rj rjk , hence Bj B(xjk , 3rjk ). This implies that
M
X

 N
 
n

j=1 Bj  M

3
|Bjk |.
B(x
,
3r
)
jk
jk
k=1
k=1

2.6. Lebesgue differentiation theorem. If a function f is continuous then,

clearly,
Z
1
f (y)dy = f (x).
(195)
lim
r0 |B(x, r)| B(x,r)
As an application of Theorem 2.4 we can show that this property continue to hold
for locally integrable functions.
Corollary 2.7 (Lebesgues differentiation theorem). If f L1loc (Rn ) then (195)
holds for almost every x.
Proof :

Since the statement is local we can assume that f L1 .

R
Let Ar be the averaging operator defined by Ar f (x) = |B(x, r)|1 B(x,r) f (y)dy.
The proof consist of two steps. First we prove that Ar f f in L1 as r 0, and
then it will be enough to show that limr0 Ar f (x) exists almost everywhere.
For the first step, given  > 0, using the density of C0 in L1 , we can always find a
compactly supported continuous function g which approximates f in L1 and have
kAr f Ar gkL1 kf gkL1 <  uniformly in r. Then by the uniform continuity of
g, we know that Ar g g in L1 as r 0, hence there exists an r such that
kAr f f kL1 kAr f Ar gkL1 + kAr g gkL1 + kf gkL1 3,
for r < r .
For the second step, we define the oscillation of an L1 function f by
f (x) = lim sup Ar f (x) lim inf Ar f (x).
r0

r0

The oscillation is a subadditive operator, (f +g) f +g and is bounded by the

maximal function operator, f 2Mf , moreover the oscillation of a continuous
function vanishes. If g is a continuous function which appoximate f in L1 then we
have that
f (f g) + g = (f g) 2M(f g).

130

6. BASIC FUNCTIONAL INEQUALITIES

We can apply now the weak-L1 property of the maximal function, and for any
positive we find that
1
|{x : f (x) > }| |{x : M(f g)(x) > /2}| . kf gkL1 .

Since kf gkL1 can be arbitrarily small, we infer that set of points where the
oscillation of f is positive is of measure zero.

2.8. Fractional integration. Let T be an integral operator acting on functions defined over Rn with kernel k as in (183). If the only information that we
have on k(x, y) is a decay estimate of the type
|k(x, y)| . |x y| ,
for some > 0, then Youngs inequality, Theorem 1.10, does not allow us to recover
a good control on T f , since the function | x | fails, barely, to be in Ln/ . However,
the convolution has smoothing properties that imply some positive results which
are contained in the following important theorem, originally proved by Hardy and
Littlewood for n = 1 and then extended by Sobolev to n > 1.
Theorem 2.9 (Hardy-Littlewood-Sobolev inequality). Let 0 < < n and 1 < p <
q < such that
1 1

(196)
1 = ,
n
p q
then
k| | f kLq (Rn ) . kf kLp (Rn ) .
Proof :

(197)

We can split the convolution with the singular kernel into two parts:
Z
Z
f (x y)
f (x y)
dy
+
dy,
I f (x) = | | f (x) =

|y|
|y|
|y|<R
|y|R

where the radius R is a positive constant to be chosen later We estimate the first
term simply by H
olders inequality,
Z

!1/p0
Z

n
0
f (x y) 


dy  kf kLp
|y|p dy
. R p0 kf kLp ,

 |y|R |y|

|y|R
where we need the integrability condition p0 > n, which by (196) is equivalent to
q < .
For the second part we perform a dyadic decomposition around the singularity and
get an estimate in terms of the maximal function,
Z

Z

|f (x y)|
f (x y)  X

dy

dy .



|y|
 |y|<R |y|

|y|
k
k1
R 2
k=0 2
Z

X
1
.
|f (x y)|dy .
(2k R) |y|2k R
.

k=0

(2k R)n Mf (x) ' Rn Mf (x),

k=0

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

131

where we need < n for the convergence of the last geometric series.
At this point we have found that for every x Rn and every R > 0,


n
| | f (x) . R p0 kf kLp + Rn Mf (x),
with constants independent of R and x. We optimize this inequality choosing, for
each x, a radius R = R(x) such that the two terms on the right hand side are equal,
n

R p0

kf kLp = Rn Mf (x),

i.e.,

R(x) =

kf kLp
Mf (x)

p/n
,

and since (n )p/n = 1 p/q, we have

1 p

|I f (x)| . kf kLp q Mf (x) q .

Then take the Lq norm on both sides,
1 p

kI f kLq . kf kLp q kMf kLq p .

If p > 1 we can conclude using the estimates for the maximal function (194).
Remark. The Hardy-Littlewood-Sobolev inequality has an equivalent bilinear
formulation, which reads
ZZ
f (x)g(y)
dxdy . kf kLp1 kgkLp2 ,
|x y|
for 0 < < n and 1 < p1 , p2 < such that
1
1

+ 0 = .
p01
p2
n
It is important to understand that the relation among the exponents can be quickly
derived from scaling arguments. If we assign a length scale to L to the variable x,
the expression
Z

k| | f kLq (Rn ) = k |x y| f (y)dykLqx (Rn )

n

has the units L Ln L q , whereas kf kLp has the units L p . The exponents , q, p
must relate in such a way that the exponents of both quantities match up. Indeed, if
they did not, then one could deduce the failure of the estimate (197) by considering
an arbitrary, nontrivial f and rescaling it to derive a contradiction.
Remark. In our proof of (197) we have not fully used the power of the Maximal
function (for example, by only considering balls centered at x). In fact, the same
estimate holds upon replacing the kernel |xy| with any kernel k(x, y) sharing the
same distribution function. A proof along these lines requires one to build up the
machinery of Lorentz spaces along with a more general form of the Marcinkiewicz
interpolation theorem. For this we refer to (***)

132

6. BASIC FUNCTIONAL INEQUALITIES

Using the Hardy-Littlewood-Sobolev inequality, we now show that it is possible to

give a very short proof of the Sobolev inequality,
kf kLq . kf kLp ,
for n/q = n/p 1, in the non sharp regime p > 1. As with the Hardy-Littlewood
Sobolev inequality, the exponents are easily deduced by considering the length
scaling of both sides. The Sobolev inequality quantifies the intuitive fact that a
function can only blow up somewhere if its derivatives blow up even worse.
Assume f C0 (Rn ). For every unit vector we have
Z
d
f (x + r)dr
f (x) =
dr
0

Proof.

To consider all possible directions in which f could grow, we integrate over the
whole sphere, and recalling that the volume element in Rn in polar coordinates is
dy = rn1 drd , we find that
Z


|f (y)|
|f (x)| .
dy = | |1n |f | (x).
n1
|x y|
We take the Lq norm and use (197) to get
kf kLq . k | |1n |f | kLq . k f kLp ,
whenever p > 1 and
1

1 1
n1
= .
n
p q

2.10. Sobolev Inequalities. In the previous section we have seen how to

1
1
estimate the Lq (Rn ) norm of a function in terms of an Lp norm, 1 n1
n = p q,
p > 1, of the gradient of f . We shall now prove a stronger version of this.
Theorem 2.11 (Galgliardo-Nirenberg-Sobolev). The inequality
f C0 (Rn ),

(198)

(1 p < q < ).

(199)

f C0 (Rn ),

(200)

kf kLq (Rn ) . k m f kLp (Rn ) ,

holds for
1
1 m
=
> 0,
q
p
n

m N,

While for q = , we have

kf kL (Rn ) .

m
X

k k f kLp (Rn ) ,

k=0

when

n
p

m < 0.

Remark. We dont need to remember the precise condition (199); it can be deduced by a simple dimensional analysis. Since the estimate is homogeneous, it has
to be invariant under dilations, and (199) simply says that both sides in (198) have
the same scaling. Also the condition np m < 0 is a comparison of the scalings of
the two sides of (200) which excludes a very localized and spiky counterexample.

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

133

Remark. The following non-sharp version of estimate (198) also holds for all
1 p < q < and 1/p m/n < 1/q,
X
kf kLq (Rn ) .
k f kLp f C0 (Rn ),
(201)
||m

Exercise. Show by an example that the inequality (200) can fail to be true for
p = n/m. Prove (201) for m = 1, using the results of theorem 2.11.
Exercise. Show by a scaling argument that if the inequality (201) holds true for
1/p = 1/q + m/n < 0 then the homogeneous inequality (198) is also true.
Proof [Proof of (198)]: We obtain the cases with m > 1 by repeated iterations of
the case m = 1. Hence, we can assume m = 1 and, by (199),
np
n
q=
< .
1 p < n,
n1
np
Once we have the estimate for p = 1 and q = n/(n 1), then we get the cases
with p > 1 and q > n/(n 1) by simply applying Holder inequality. Indeed, let
q = n/(n 1), for some > 1, then
n
kf kLq = k|f | kL n1
. k|f |1 f kL1 k|f |1 kLp0 kf kLp ,

and we just have to check that

( 1)p0 =

n1
n q
1 n1

1
= q.
1q

But this essentially needs no verification by the scaling of the inequality, the
exponents must work out.
It only remains to prove the special case m = 1, p = 1, q = n/(n 1). Following
Nirenberg, [?], one can show the stronger result that for f C0 (Rn ) we have
n
kf kL n1
.
(Rn )

n
Y

1/n

kj f kL1 (Rn ) .

(202)

j=1

When n = 1, this comes easily from writing

Z x
f (x) =
f 0 (y)dy.

When n = 2, we do the same with respect toeach variable and then multiply and
integrate:
ZZ
ZZ Z
Z
|f (x1 , x2 )|2 dx1 dx2
|1 f (y1 , x2 )|dy1 |2 f (x1 , y2 )|dy2 dx1 dx2
= k1 f kL1 k2 f kL1 .
When n 3 things become more tricky and, to separate the variables, we have to
make a repeated use of H
older inequality. Let just look at the case n = 3. To ease

134

6. BASIC FUNCTIONAL INEQUALITIES

the notation set fj = j f and

3
2

Z

|f (x)|

(x)dxj =

(
xj ). We start with

 12 Z
 21 Z
 21
|f1 (, x2 , x3 )|
|f2 (x1 , , x3 )|
|f3 (x1 , x2 , )|
.

Then integrate with respect to x1 . The first factor on the right hand side doesnt
depend on x1 , while we use Holder to separate the second from the third,
Z
 12 Z
 12 Z
 12
Z
3
2
|f (, x2 , x3 )|
|f1 (, x2 , x3 )|
|f2 (, , x3 )|
|f3 (, x2 , )|
.
1

1,2

1,3

Proceed similarly with the integration with respect tox2 ,

 12 Z
 12 Z
Z
Z
3
|f1 (, , x3 )|
|f2 (, , x3 )|
|f (, , x3 )| 2
1,2

1,2

1,2

and finally do the same with x3 ,

Z
 12 Z
Z
3
2
|f ()|
|f1 ()|
1,2,3

1,2,3

 21
|f3 ()|
,

1,2,3

 21 Z
|f2 ()|

1,2,3

 12
|f3 ()|

1,2,3

When n > 3 the procedure is exacly the same.

Proof [Proof of (200)]: It clearly suffices to look at the case m = 1, since the
cases m > 1 will follow from it applying (198). Assume thus m = 1 and p > n, we
want to prove that
|f (0)| . kf kLp + kDf kLp .
Suppose first that f has support contained in the unit ball B = {|x| < 1}, then
Z 1
d
f (0) =
f (r)dr, Sn1 .
(203)
dr
0
Integrate with respect to and then apply Holder,
1/p0
Z
Z
|f (x)|
dx
p
|f (0)| .
dx
.
kf
k
. kf kLp ,
L
n1
(n1)p0
B |x|
B |x|

(204)

where the integrability condition needed here is (n 1)p0 < n, which is precisely
p > n.
In general, fix a cutoff function C0 with support in B and (0) = 1, then in
view of the above, |f (0)| = |(0)f (0)| . k(f )kLp . kf kLp + kf kLp .

2.12. Classical Sobolev spaces. The Sobolev inequalities of theorem (2.11)

lead us to the introduction of Sobolev spaces.
Definition 2.13. Let be an open subset of Rn . Fix 1 p and let s N
be a non-negative integer. The space W s,p () consists of all locally integrable, real
(or complex) valued functions u on such that for all multiindex with || s

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

135

the weak4 derivatives u belong to Lp (). These spaces come equiped with the
norms,
X

1/p
kukW s,p () =
k ukpLp ()
,
for 1 p <
||s

kukW s, ()

k ukL ()

||s

We also denote by W0k,p () the closure of C0 () in W k,p ().

In the particular case p = 2 we write H s () = W s,2 (). Clearly H 0 () = L2 ().
We also write H0s () = W0s,2 (). These spaces are especially important because of
their Hilbert space structure.
In the particular case p = we work with the smaller space C s () W s, (),
the set of functions which are s times continuously differentiable and have bounded
k kW s, norm.
Exercise. Show that for each s N and 1 p the spaces W s,p () are
Banach spaces.
There is a lot more to be said about Sobolev spaces in domains Rn . For
instance, We refer the reader to Evans ([E], ch. 5). For the time being we specialize
to the case = Rn .
Exercise. Show that the spaces W k,p (Rn ) and W0k,p (Rn ) coincide. That means
that C0 is dense in W k,p (Rn ).
The Sobolev inequalities proved in the previous subsection can be interpreted as
embedding theorems. Indeed (198) and (201) can be interpreted as saying that
the Sobolev space W m,p (Rn ) is included in the Lebesgue space Lq (Rn ) as long as
1
m
1
p n q.
Proposition 2.14. The following inclusions are continuous
W m,p (Rn ) Lq (Rn ),

if

1 m
1

.
p
n
q

Moreover, for q = , W m,p (Rn ) embeds into the space of bounded continuous
functions on Rn provided that m > n/p.
Proof :

Follows from theorem 2.11 and the density of C0 (Rn ) in W m,p (Rn ).

2.15. H
older spaces. Together with Sobolev spaces Holder spaces play a very
important role in Analysis, especially in connection to elliptic equations. Before
introducing these spaces we recall the definitions of the spaces C m () of m times
4That is derivatives in the sense of distributions.

136

6. BASIC FUNCTIONAL INEQUALITIES

continuously differentiable functions u : R on an open domain for which the

W m, norm is bounded,
X
kukC m () =
k u(x)kL () < .
||m

Definition 2.16. Let be an open domain in Rn . We say that a function u :

R is H
older continuous with exponent 0 < 1 if,
|u(x) u(y)|
< .
|x y|
x6=y

[u]C 0, () = sup

(205)

The H
older space C k, () consists of all functions u C k () for which the norm,
X
kukC k, () = kukC k () +
[ u]C 0, () .
(206)
||=k

is finite.
Exercise 2.17.

The space C k, () is a Banach space.

Exercise 2.18.
Show that C 0,1 ((a, b)), the space of Lipschitz functions on an
interval, consists exactly of those distributions whose derivative belongs to L .
Exercise 2.19.
Let f (x) = (a x b) be the characteristic function of an
interval. Show that the seven-fold convolution f f is in the Holder class C 5,1 .
The following stronger version of the Sobolev embedding in L is important in
elliptic theory. As usual, the relationship between the exponents involved can be
deduced from dimensional analysis.
Theorem 2.20 (Morreys inequality). Assume n < p . Then, for all u
C0 (Rn ),
kukC 0, (Rn ) . kukW 1,p (Rn )

(207)

provided that = 1 n/p.

Proof : In one dimension, this is an easy application of the fundamental theorem
of calculus and H
olders inequality. For the general case, see [E], section 5.6.2.

2.21. Fractional H s - Sobolev spaces. Consider the Sobolev space

H s (Rn ) = {u L2 : u L2 ,

|| s}.

Proposition 2.22. The Sobolev space H s (Rn ) coincides with the set of all tempered
distributions u S 0 (RN ) for which u
is locally integrable and,
Z
kuk2H s =
(1 + ||2 )s |
u()|2 <
(208)
RN

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

137

Proof : Follows easily from the Parseval identity, the density of C0 in each space,

and the fact that d

j f () = ij f ().
Observe that the equivalent definition of proposition 2.22 makes sense not only for
positive integers but for all real numbers s. We can thus talk about Sobolev spaces
H s for all real values of s. We shall also make use of the following homogeneous
Sobolev norm, for all s 0,
Z
kuk2H s =
||2s |
u()|2 <
(209)
RN

Question. Why does kukH s have units L 2 s if we consider the physical space
variable to have the unit L?
Exercise. For s (0, 1) the space H s (Rn ) coincides with the space of locally
integrable functions such that,
1/2
Z Z
|u(x) u(x + y)|2
2
dxdy + kukL2 (Rn )
<
(210)
|y|n+2s
Rn Rn
Exercise. Prove that, for s > n/2 the Sobolev space H s (Rn ) embeds in the space
of bounded continuous functions.
2.23. A Trace Theorem. In order to make sense of boundary values of generalized functions for partial differential equations, it is important to prove that the
operation of restriction, which obviously makes sense for continuous functions, continues to make sense even when the function is not continuous. Such theorems are
called trace theorems. Consider, for simplicity, the case of the hyperplane xn = 0
in Rn and define the trace operator,
T f (x1 , . . . , xn ) = f (x1 , . . . , xn1 , 0).

(211)

Clearly the operator makes sense for any continuous functions f , in particular for
any test function, in Rn .
Theorem 2.24. The following estimate holds true, uniformly for any test function
f C0 (Rn ), n 2 and any s > 1/2.
kT f k

H s 2 (Rn1 )

. kf kH s (Rn )

(212)

Therefore T extends as a linear map T : H s (Rn ) H s1/2 (Rn1 ).

Remark. Observe that the result is dimensionally sharp, which is somewhat surprising if one compares it to the usual embedding of the Sobolev spaces H s (R) in
L (R), in which case we know that the sharp case, s = 1/2, is false. In fact the
above trace theorem is also false for the case s = 1/2.
Proof : Take f smooth and g(x0 ) = f (x0 , 0). Let f be the Fourier transform of f in
xn only, and f, g be the Fourier transforms of f and g in Rn and Rn1 , respectively.

138

6. BASIC FUNCTIONAL INEQUALITIES

I.e.
Z

f(x0 , n ) =

f (x0 , xn )eixn n dxn

By applying Fourier inversion (with xn = 0) and then the Fourier transform, we

get
Z
1
g(x0 ) = f (x0 , 0) =
f(x0 , m )dn
2
Z
1
g( 0 ) =
f( 0 , m )dn
2
We can then see, using our knowledge of fractional H s spaces and Cauchy-Schwartz:
Z
|
g ( 0 )|2 (1 + | 0 |2 )s1/2 d 0
kgkH s1/2 (Rn1 .
Rn1

Z
.

2

f()dn  (1 + | 0 |2 )s1/2 d 0

Z

0 2 s1/2
2
2 s

(1 + | | )
|f ()| (1 + || ) dn J( 0 )d 0

Z



n1

Z
.

Rn1

with,
J( 0 ) =

(1 + ||2 )s dn

And since s > 1/2, we have

Z
Z
0
2 s
J( ) =
(1 + || ) dn =
(1 + | 0 |2 + |n |2 )s dn

Z
0 2 s+1/2
= (1 + | | )
(1 + y 2 )s dy

Plugging this into our above estimate for kgkH s1/2 proves the result.
Similar results hold for traces to higher co-dimension hypersurfaces. Here is such
a result, which can be proved by elementary means.
Proposition 2.25. Consider the trace operator T in R3 which takes continuous
functions f (t, x1 , x2 ) to T f (t) = f (t, 0, 0). We have, for any test function f ,
kt (T f )kL2 (R) . k 2 f kL2 (R3 )
2

(213)
3

Thus, T extends, as a bounded linear operator, to H (R ) with values in H (R).

Proof
Z
R

|t f (t, 0, 0)|2 dt

Z


dx2 1 2 t f (t, x)t f (t, x)dt
0
0
Z
Z
Z R


= 2
dx1
dx2
1 2 t f (t, x)t f (t, x)dt
Z0
Z0
ZR


+ 2
dx1
dx2
1 t f (t, x)2 t f (t, x)dt
dx1

2. MAXIMAL FUNCTION, FRACTIONAL INTEGRATION AND APPLICATIONS

Clearly,
Z




dx

dx

139




1 t f (t, x)2 t f (t, x)dt  . k 2 f k2L2 (R3 )

On the other hand, integrating by parts in t,

Z
Z
1 2 t f (t, x)t f (t, x)dt = 1 2 f (t, x)t2 f (t, x)dt
R

Hence,
Z




dx

dx




1 2 t f (t, x)t f (t, x)dt  .

k 2 f kL2 (R3 ) .

as desired.
Exercise. Prove the same result using Fourier transform and extend it to all dimensions and general H s spaces. Exercise Extend the result to bounded intervals
in t.
2.26. Extensions. To extend results which hold true for functions in Rn to
domains in Rn we need to extend the functions in a controlled manner. I will restrict
the discussion to the case of the half space Rn+ = {x Rn /xn 0}. Consider the
Sobolev space W 1,p (Rn+ ). We want to prove the following.
Proposition 2.27. There exists a bounded linear operator E : W 1,p (Rn+ ) W 1,p (Rn ),
such that for any continuous f ,
Ef |Rn+ = f
and,
kEf kW 1,p (Rn ) . kf kW 1,p (Rn+ )
Proof It suffices to prove the result for functions f C 1 (Rn+ ). Given such a function we define, using its higher order reflection, its extension barf which coincided
with f in Rn+ and, for all xn < 0,
1
f(x0 , xn ) = 3f (x0 , xn ) + 4f (x0 , xn )
2
Observe first that f is also C 1 . Indeed f is continuous across xn = 0 and so are its
derivatives with respect to the variables x0 = (x1 , . . . , xn1 ). On the other hand,
for xn < 0,
1
n f(x0 , xn ) = 3n f (x0 , xn ) 2n f (x0 , xn )
2
Hence, letting xn tend to zero with xn < 0
(n f) (x0 , 0) = n f (x0 , 0).
Using these calculations we immediately derive the desired estimate.

140

6. BASIC FUNCTIONAL INEQUALITIES

Exercise Extend the result to the W s,p spaces, with s N. What about fractional
H s spaces ?.

3. Littlewood-Paley theory
In its simplest manifestation Littlewood-Paley theory is a systematic and very useful method to understand various properties of functions
f , defined on Rn , by
P
decomposing them in infinite dyadic sums f =
kZ fk , with frequency localized components fk , i.e. fbk () = 0 for all values of outside the dyadic annulus
2k1 || 2k+1 . Such a decomposition can be easily achieved by choosing a test
function () in Fourier space, supported in 21 || 2, and such that, for all
6= 0,
X
(2k ) = 1.
(214)
kZ

Indeed choose () to be a real radial bump function supported in || 2 which

equals 1 on the ball || 1. Then the function () = () (2) verifies the
desired properties.
We now define
k
Pd
k f () = (/2 )f ()

(215)

Pk f = fk = mk f

(216)

or, in physical space,

where mk (x) = 2nk m(2k x) and m(x) the inverse Fourier transform of . Clearly,
from (214)
X
f=
Pk f
(217)
kZ

as desired. Denoting the map multiplication by 2k by 2k , we also have the

important scaling identity
(Pk f ) 2k = P0 (f 2k )

(218)

Observe that the Fourier transform of Pk f is supported in the dyadic interval

2k1 || 2k+1 and therefore,
Pk0 Pk f = 0,

k, k 0 Z,

|k k 0 | > 2.

Therefore,
Pk f =

X
k0 Z

Pk0 (Pk f ) =

Pk 0 Pk f

|kk0 |1

Thus,P
since Pk1 , Pk , Pk+1 do not differ much between themselves we can write
Pk = |kk0 |1 Pk0 Pk Pk2 . It is for this reason that the cut-off operators Pk are
called, improperly, LP projections.
P
Denote PJ = kJ Pk for all intervals J Z. We write, in particular, Pk =
P(,k] and P<k = Pk1 . Clearly, Pk = Pk P<k .

3. LITTLEWOOD-PALEY THEORY

141

The following properties of these LP projections lie at the heart of the classical LP
theory:
Theorem 3.1. The LP projections verify the following properties:
LP 1. Almost Orthogonality.
The operators Pk are selfadjoint and verify
Pk1 Pk2 = 0 for all pairs of integers such that |k1 k2 | 2. In particular,
X
kf k2L2
kPk f k2L2
(219)
k

LP 2.

Lp -boundedness:

For any 1 p , and any interval J Z,

kPJ f kLp . kf kLp

(220)

LP 3.
Finite band property. We can write any partial derivative Pk f in the
form Pk f = 2k Pk f and the symbol of Pk is a cut-off operator5 which verifies
property LP2. In particular, for any 1 p
kPk f kLp

. 2k kf kLp

(221)

. kf kLp

(222)

2 kPk f kLp
LP 4.
Bernstein inequalities.
inequalities,
kPk f kLq
kP0 f kLq

For any 1 p q we have the Bernstein

. 2kn(1/p1/q) kf kLp ,

k Z

. kf kLp .

(223)
(224)

In particular,
kPk f kL . 2kn/p kf kLp .
LP5.

Commutator estimates

Consider the commutator

[Pk , f ] g = Pk (f g) f Pk g
with f, g

C0 (Rn ).

We have,
k [Pk , f ] gkLp . 2k kf kL kgkLp .

LP6. Square function inequalities.

(Pk f )kZ . The quantity

Let Sf be the vector valued function Sf =

Sf (x) = |Sf (x)| =

|Pk f (x)|2

1/2

(225)

kZ

is known as the Littlewood-Paley square function. For every 1 < p < there exists
constant(s), depending on p, such that for all f C0
kf kLp . kSf kLp . kf kLp

(226)

5Associated with a slightly different test function

which remains supported in 12 || 2,

but may fail to satisfy (214).

142

6. BASIC FUNCTIONAL INEQUALITIES

Proof : Only the proof of LP6 is not straightforward and we postpone it until
next section. The proof of LP1 is immediate. Indeed we only have to check (219).
Clearly,
X
X
kf k2L2 = k
Pk f k2L2 =
< Pk f, Pk0 f >L2
|kk0 |1

kPk f kL2 kPk0 f kL2

|kk0 |1

kPk f k2L2

P
To show that k kPk f k2L2 . kf k2L2 we only need to use Parsevals identity together
with the definition of the projections Pk .
It suffices to prove LP2 for intervals of the form J = (, k] Z, that is to prove
k
\
Lp boundedness for Pk . If () = () (2) then P
k f = (/2 )f (). Thus
Pk f = m
k f,
where m
k (x) = 2nk m(2
k x) and m(x)

is the inverse Fourier transform of . Observe

that km
k kL1 = kmk
L1 . 1. Thus, using the convolution inequality (188),
kPk f kLp km
k kL1 kf kLp . kf kLp
as desired.
To prove LP3 we write i (Pk f ) = 2k (i m)k f where (i m)k (x) = 2nk i m(2k x).
Clearly k(i m)k kL1 = ki mkL1 . 1. Hence,
ki (Pk f )kLp . 2k kf kLp
Pn
which establishes (221). To prove (222) we write f() = j=1
2k Pd
k f () =

n
X

Hence,

2k

j=1

where j () =

j [
i||2 xj f ().

j
i||2 ().

X
j
(/2k )[
2k j (/2k )[
xj f () =
xj f ()
2
i||
j=1

Hence, in physical space,

2k Pk f =

n
X

(j m)k j f

j=1

with (j m)k (x) = 2nk j m(2k x) and j m the inverse Fourier transform of j . Thus,
as before,
2k kPk f kLp .

n
X

kj f kLp = kf kLp

j=1

as desired.
Property LP4 is an immediate consequence of the physical space representation
(216) and the convolution inequality (188).
kPk f kLq

= kmk f kLq . kmk kLr kf kLp

3. LITTLEWOOD-PALEY THEORY

143

where 1 + q 1 = r1 + p1 . Now,
Z

1/r
nk
kmk kLr = 2
= 2nk 2nk/r kmkLr . 2nk(11/r) . 2nk(1/p1/q)
|m(2k x)|r dx
Rn

It only remains to prove LP5. In view of (216) we can write,

Z


Pk (f g)(x) f (x)Pk g(x) =
mk (x y) f (y) f (x) g(y)dy
Rn

On the other hand,

Z


d
f (x + s(y x))ds
0 ds
. |x y|kf kL


|f (y) f (x)| . 

Hence,
|Pk (f g)(x) f (x)Pk g(x)|

. 2k kf kL

Z
|m
k (x y)||g(y)|dy
Rn

where m
k (x) = 2nk m(2
k x) and m(x)

= |x|m(x). Thus,
kPk (f g) f Pk gkLp . 2k kf kL kgkLp
We leave the proof of property LP6 for the next section.
Remark. It could have simplified matters in the preceding proof to prove properties LP2-4 only in the case k = 0, and deduce the more general estimates from
the scaling identity (218). In particular, note that the Bernstein inequality is simply the statement that lower Lp norms control higher Lp norms when f is localized
in frequency space (as opposed to the other way around, which occurs when f is
localized in physical space). This accords with our intuition for Lp norms: while
a frequency localized function may be too large at in physical space to be integrable, one need not worry about sudden jumps or spikes where the function blows
up locally, and hence only the former phenomenon needs to be controlled.
Definition. We say that a Fourier multiplier operator Pk is similar to a standard
LP projection Pk if its symbol
k is a bump function adapted to the dyadic region
|| 2k . More precisely we can write
k () = (
2k ) for some bump function

1 k
k
supported in the region c 2 . || c2 for some fixed c > 0.
Remark. Observe that the inequality kPk f kLp . kf kLp holds for every other
operator Pk similar to Pk . The same holds true for the properties LP3, LP4 and
LP5.
Remark: We have the following pointwise relation of the operator Pk with the
maximal function:
|Pk f | . Mf (x)
Indeed we have, as before,
Pk f = m
k f,

(227)

144

6. BASIC FUNCTIONAL INEQUALITIES

where m
k (x) = 2nk m(2
k x) and m(x)

S(Rn ). Therefore,
Z
Z


nk
k
nk

|Pk f | . 2
|f (y)|m
2 (x y) |dy . 2
|f (y)|(1 + 2k |x y|)n1 dy
Z
. 2nk
|f (y)|(1 + 2k |x y|)n1 dy
+

2nk

B(x,2k )

XZ
j=0

.
.

2nk

|f (y)|(1 + 2k |x y|)n1 dy

2j 2k |xy|2j+1

Z
|f (y)|dy +
B(x,2k )

Mf (x) +

2(n+1)j

. Mf (x) + 2n

2(n+1)j 2nk 2n(j+1k)

|f (y)|dy

|xy|2j+1k

j0

j>0

1
|B(x, 2k+j+1 )|

Z
|f (y)|dy
B(x,2k+j+1 )

2j Mf (x) . Mf (x)

j>0

as desired.
Properties LP3-LP4 go a long way to explain why LP theory is such a useful tool
for partial differential equations. The finite band property allows us to replace
derivatives of the dyadic components fk by multiplication with 2k . The Lp L
Bernstein inequality is a dyadic remedy for the failure of the embedding of the
n
Sobolev space W p ,p (Rn ) to L (Rn ). Indeed, in view of the finite band property,
the Bernstein inequality does actually imply the desired Sobolev inequality for each
LP component
fk , the failure of the Sobolev inequality for f is due to the summation
P
f = k fk .
In what follows we give a few applications of LP -calculus.
3.2. Interpolation inequalities.
The following inequality holds true for
arbitrary functions in C0 (Rn ) and any integers 0 i m:
1i/m

k i f kLp . kf kLp

i/m

k m f kLp

(228)

To prove it we decompose f = Pk f + P>k f = fk + f>k . Now, using LP2-LP4,

for any fixed value of k Z,
k i f kLp

k i fk kLp + k i f>k kLp

2ki kf kLp + 2k(im) k m f kLp

Thus,
k i f kLp

i kf kLp + im k m f kLp

for any 2Z . To finish the proof we would like to choose such that the two
terms on the right hand side are equal to each other, i.e.,
 m
1/m
k f kLp
0 =
kf kLp

3. LITTLEWOOD-PALEY THEORY

145

since we are restricted to 2Z we choose the dyadic number 2Z such that,

0 2 Hence,
2
mi m
1i/m
i/m
k i f kLp i0 kf kLp +
k f kLp . kf kLp
k m f kLp .
0
In general when an estimate for functions on Rn fails to be dimensionally consistent
(in that the scalings of the two sides are not the same), such an estimate can be
amplified into one which appears even stronger (or proven false).
Exercise. Assuming the inequality
k i f kLp . kf kLp + k m f kLp
deduce the estimate (228) by considering the rescalings f f .
3.3. Non-sharp Sobolev inequalities.
We shall prove the following slightly
improved version of the inequality (201), for functions f C0 (Rn ) and exponents
1 p < q < with 1/p m/n < 1/q,
kf kLq (Rn ) . kf kLp + k m f kLp
P
P
We decompose f = P0 f + kN Pk f = f<0 + k>0 fk . Thus, using LP4 and then
LP3,
X
X
kf kLq kf<0 kLq +
kfk kLq . kf kLp +
2kn(1/p1/q) kf kLp
k>0

. kf kLp +
. kf k

Lp

k>0

kn(m/n)

k>0
m

kf kLp . kf kLp +

2kn k m f kLp

k>0

+ k f k

Lp

3.4. Spaces of functions. The Littlewood-Paley theory can be used both to

give alternative descriptions of Sobolev spaces and introduce new, more refined,
spaces of functions. We first remark that, in view of the almost orthogonality
property LP1,
X
X
kf k2L2 = k
Pk f k2L2 .
kPk f k2L2
kZ

kPk f k2L2

kZ

. kf kL2

kZ

We can thus give an LP description of the homogeneous Sobolev norms k kH s (Rn )

X
kf k2H s
22ks kPk f k2L2
(229)
kZ
+

For k Z , define operator k = Pk if k > 0, and 0 = P0 . Also for the H s

norms,

X
kf k2H s
22ks kk f k2L2
(230)
k=0

The Littlewood- Paley decompositions can be used to define new spaces of functions
such as Besov spaces.

146

6. BASIC FUNCTIONAL INEQUALITIES

Definition:
norm:

s
The Besov space Bp,q
(Rn ) is the closure of C0 (Rn ) relative to the

s
kf kBp,q
=(

2ksq kk f kqLp ) q

(231)

k=0

The corresponding homogeneous Besov norm is defined by,

X
1
kf kB s = (
2sqk kPk f kqLp ) q ,

(232)

p,q

kZ

s
One similarly define Triebel space Fp,q
by reversing the Lp norm and lq norm in
s
s
(231). Thus, for example, the H norm is equivalent with the Besov norm B2,2
.
s
s
s
Observe that, H B2,1 . One reason why the larger space B2,1 is useful is because
of the following

kf kL . kf kB n/2

(233)

2,1

which follows from the Bernstein inequality LP4. (233) will play a key role in
s
the following section. Another reason to use the Besov norms B2,1
will become
transparent in the next section where we discuss product estimates.

3.5. Product estimates. The LP calculus is particularly useful for nonlinear

estimates. Let f, g be two functions on Rn . Consider,
X
Pk (f g) =
Pk (Pk0 f Pk00 g)
(234)
k0 ,k00 Z
0

Now, since Pk0 f has Fourier support in the set D0 = 2k 1 || 2k +1 and

00
00
Pk00 f has Fourier support in D = 2k 1 || 2k +1 it follows that Pk0 f Pk00 g
0
00
has Fourier support in D + D . We only get a nonzero contribution in the sum
(234) if D0 + D00 intersects 2k1 || 2k+1 . Therefore, writing fk = Pk f and
f<k = P<k f , and fJ = PJ f for any interval J Z we derive,
Lemma 3.6. Given functions f, g we have the following decomposition:
Pk (f g)
HHk (f, g )

= HHk (f, g ) + LLk (f, g ) + LHk (f, g ) + HLk (f, g ) (235)

X


=
Pk fk0 Pk00 g
k0 ,k00 >k+5,|k0 k00 |3

= Pk f[k5,k+5] g[k5,k+5]


LHk (f, g ) = Pk fk5 g[k3,k+3]


HLk (f, g ) = Pk f[k3,k+3] gk5
LLk (f, g )

The term HHk (f, g ) corresponds to high-high interactions. More precisely, each
term in the sum defining HHk (f, g ) has frequency 2m for some 2m >> 2k . We
shall write schematically,
X


HHk (f, g ) = Pk
fm gm
(236)
m>k

3. LITTLEWOOD-PALEY THEORY

147

The term LLk (f, g ) consists of a finite number of terms which can be typically
ignored. Indeed they can be treated, in any estimates, like either a finite number
of HH terms or a finite number of LH and HL terms. We write, schematically,
LLk (f, g ) = 0

(237)

Finally the LHk and HLk terms consist of low high, respectively high-low, interactions. We shall write schematically,


LHk (f, g ) = Pk f<k gk
(238)


HLk (f, g ) = Pk fk g<k
(239)
Remark. In the correct expression of LHk given by (235) the terms of the form
fk5 gk00 , k 00 [k 3, k + 3], have Fourier supports in the dyadic region 2k .
Thus Pk can be safely ignored and we can write,
LHk (f, g ) f<k gk .
We have thus established, the famous trichotomy formula,
Pk (f g) = LHk (f, g) + HLk (f, g) + HHk (f, g)

(240)

which is the basis of paradifferential calculus. In practice whenever we apply formula

(240) we have to recall that formulas (237)(239) are only appproximate; the correct
definitions are given by (235). However in any estimates we can safely ignore the
additional terms as they are estimated precisely in the same way as the terms we
keep.
We shall now make use of the trichotomy formula to prove a product estimate.
Theorem 3.7. The following estimate holds true for all s > 0.
kf gkH s . kf kL kgkH s + kgkL kf kH s

(241)

kf gkH s . kf kH s kgkH s

(242)

Thus for all s > n/2,

In what follows we give a somewhat simple proof of theorem (3.7) which is very
instructive. The proof6 shows that it is sometimes better not to rely on the full
decomposition (235) but rather using decompositions sparingly whenever needed.
Indeed, we write,
X
X
X
kf gk2H s .
22ks kPk (f g)k2L2 .
22ks kPk (f<k g)k2L2 +
22ks kPk (fk g)k2L2
k

Now,
X

22ks kPk (fk g)k2L2

kgk2L

kgk2L

XX

kgk2L

X X

kgk2L kf k2H s

22ks kfk k2L2

k
0

22(kk )s k2k s fk0 k2L2

k k0 k

k0

0
0
22(kk )s k2k s fk0 k2L2

kk0

6 I thank Igor Rodnianski for pointing the argument to me.

148

6. BASIC FUNCTIONAL INEQUALITIES

P
To estimate k 22ks kPk (fk g)k2L2 we shall decompose further, proceeding as in the
P
decomposition (235). But first observe that the term k 22ks kPk (f[k3,k] g)k2L2 can
P
be treated precisely as k 22ks kPk (f>k g)k2L2 . Indeed we might as well estimated
P 2ks
kPk (f>k3 g)k2L2 instead. Now,
k2
X
X
X
Pk (fk3 gk0 ) =
Pk (fk3 g) =
Pk (fk3 gk0 ) +
Pk (fk3 gk0 )
k0

k0 <k2

k2k0 k+2

Pk (fk3 gk0 )

k0 >k+2

Observe that the first and last term are zero, therefore,
X
Pk (fk3 g) =
Pk (fk3 gk0 ) Pk (fk3 gk ).
k2k0 k+2

Often, for simplicity, we simply write,

Pk (f<k g) f<k gk

(243)

Of course this formula is not quite right, but is morally right. Now,
X
X
22ks kPk (f<k g)k2L2 =
22ks kf<k gk k2L2
k

. kf k2L

22ks kgk k2L2 = kf k2L kgk2H s

as desired.
Remark.

In view of (243) we have the following partial decomposition formula,

Pk (f g) = f<k gk + Pk (fk g) = LHk (f, g) + Pk (fk g)

(244)

Contrast this with the full trichotomy decomposition (240).

Similar estimates, easier to prove, hold in Besov spaces. Indeed, for every s > 0 we
have,
s
s
s
kf gkB2,1
. kf kL kgkB2,1
+ kgkL kf kB2,1

Exercise.

(245)

Prove estimate (245).

4. Wentes Inequality
In this section we prove Wentes inequality as an application of Littlewood-Paley
theory. In what follows given two functions f, g in R2 we consider the bilinear
expression (df dg) = x f y g y f x g, where denotes the trivial Hodge duality
in R2 . By abuse of language we drop the dual sign below and write simply df dg
.
Theorem 4.1. On R2 , assume f , g H 1 (R2 ), u = (df dg). Then u L is
in fact continuous.

4. WENTES INEQUALITY

149

Remark. In fact df dg If is replaced by ordinary multiplication, then the best

we can get is df dg L1 . This is obviously not enough to obtain that u L .
It turns out however that df dg has special structure which allows us to derive
the desired estimate. In the above theorem, we refer to the canonical solution
u = 1 (df dg) obtained through the canonical solution.
Proof : It is easy to see from finite band property that is a isometric operator
s2
s
from B p,1
to B p,1
. In fact we shall work with p = 2, In view of the Sobolev
1
inequality (233), it suffices to show that df dg B 2,1
(R2 ). Using the trichotomy
formula and the fact that the LP projections Pk commute with d we write,
I

= df dg = LHk + HLk + HHk

LHk

= dP<k f dPk g

HLk

= dPk dP<k g
X
= Pk (
(dPm f dPm g)

HHk

mk

By symmetry we only need to deal with LH and HH. The LH term is trivial to
estimate, without using the special structure of the wedge product. Using the
Bernstein inequality we write,
X
2k kLHk kL2 . 2k
kdPl f kL kdPk (g)kL2
l<k

2lk kDPl f kL2 kDPk f kL2

l<k

The proof now follows with the following discrete version of the Young inequality.
Lemma 4.2. Let f (k) l1 (Z) and g(k), h(k) l2 (Z). Then,
X
f (k l)g(l)h(k) kf kl1 kgkL2 khkl2 .
k,l

Using the lemma, we derive,

X
2k kLHk kL2 .

kDPl f k2L2

1/2 X
(
kDPk f k2L2 )1/2

. kDf kL2 kDgkL2

We now consider HHk . It is here that we need to use the special structure of the
wedge product. In fact we shall simply use the identity, df dg = d(f dg). Thus,
X
HHk =
Pk (dPm f dPm g)
mk

dPk (Pm f dPm g)

mk

Thus, using the finite band property and Bernstein inequality,

kHHk kL2

22k kPm f dPm gkL1

22k kPm f kL2 kDPm gkL2

22km kDPm f kL2 kDPm gkL2

150

6. BASIC FUNCTIONAL INEQUALITIES

Therefore,
2k kHHk kL2

. 2km kDPm f kL2 kDPm gkL2

Thus, again, using the discrete Young inequality of the lemma above,
X
2k kLHk kL2 . kDf kL2 kDgkL2
k

as desired.

5. A Sharp Trace Theorem

In this section, we provide another application of LP theory: a stronger version of
the the Trace Theorem, in Besov spaces, see [Kl-Rodn3]
For simplicity, let I = [0, 1] and consider I R2 . We will use the mixed norm
notation:
Z 1
 q1
q
q
p
kf kLt Lx =
kf (t, )kLpx (R2 ) dt
0

 p1

Z
kf k

q
Lp
x Lt

=
R2

kf (, x)kpLq (I) dx
t

with the obvious modifications if p = or q = .

We will get the following trace-like estimate:
Z
1
k |t f |2 dtkB2,1
. kf k2H 2 (IR2 )

(246)

We observe that
1
0
kgkB2,1
. kgkB2,1
+ kgkL2

Thus, (246) follows from the sharp bilinear trace theorem below.
Theorem 5.1. For any smooth, scalar functions g,h on I R2 , we have
Z
0
k t g hdtkB2,1
. kgkH 1 (IR2 ) khkH 1 (IR2 )
I

Proof Immediately we see:

Z
Z
X
0
k t g hdtkB2,1
=
kPk
I

X
k0

Z
kPk
0

Z
t g hdtkL2x + kP<0

k0

t g hdtkL2x

t g hdtkL2x

(247)

5. A SHARP TRACE THEOREM

151

P
P
We will then decompose g and h with respect to x; g = k Pk g = k gk , h =
R1
P
P
k Pk h =
k hk . Then we can decompose Pk 0 (t g h) = Ak + Bk + Ck + Dk ,
where
Z 1
A k = Pk
(t g)<k hk
0
1

(t g)k h<k

Bk = P k
0
1

(t g)<k h<k

Ck = Pk
0

(t g)k hk

Dk = Pk
0

As in the Trichotomy Formula, Ck is essentially zero (with the exception of finitely

many terms which can be subsumed in Ak , Bk , or Dk ).
We now briefly sketch how to estimate each of Ak , Bk , Dk , leaving
R 1 the details to be
filled in. Note that Pk trivially commutes with the integrals 0 dt and any partial
derivatives t .
To estimate Ak , note that we can write (using LP2):
X Z 1
kAk kL2x .
k(t g)k0 hk00 kL2x dt
k0 <kk00

We can then use Bernstein inequality LP4 and property LP3 on h to pull out the
0
00
0
00
0
00
power 2k k . Writing 2k k . 2(k k)/2+(kk )/2 , using LP1, and summing over
k, we can then get:
X
2 . kt gkL L2 khkL L2
kAk kL
x
x
t Lx
t
t
k0

R1

(t g)k hk , write
Z 1
X
1
2
Dk = Dk + Dk =
Pk
(t g)k0 hk00 +

To estimate Dk = Pk

kk0 k00

X
kk0 k00

(t g)k00 hk0

Pk
0

Dk1 can be estimated straightforwardly, without integration by parts. Use LP4 and
LP3 to write
0
kDk1 kL2x . 2kk kt gkL2t L2x khkL2t L2x
Then sum over k and use LP1 to get:
X
kDk1 kL2x . kt gkL2t L2x khkL2t L2x
k0

To estimate Dk2 we use integration by parts to transfer the t from the highfrequency gk00 to the low-frequency hk0 . After integrating by parts we treat the

152

6. BASIC FUNCTIONAL INEQUALITIES

result exactly as Dk1 . Thus, we need only estimate the boundary terms: kIk (1)
2 , where
Ik (0)kL2x . kIk kL
t Lx
X
Ik =
Pk (gk00 hk0 )
kk0 <k00

We use the following lemma to do so:

Lemma 5.2. For any k,k,k we have
0

kPk (gk0 hk00 )k . 2 4 (|k k|+|k

00

k|)

kgk0 kkhk00 k

Using this lemma, we integrate by parts and bound Dk2 just as Dk1 plus the boundary
term, and eventually get:
X
kDk2 kL2x . kgkH 1 khkH 1
k

R1
P
Now we estimate Bk by similarly decomposing to Bk = k0 <kk00 Pk 0 (t g)k00 hk0 .
As above,
P we integrate by parts and use the lemma to estimate the boundary terms
Jk = k0 <kk00 Pk (gk00 ) hk0 ). It is then not hard to manipulate and sum over k to
get
X
kBk kL2x . kgkH 1 khkH 1
k

Combining all the estimates for Ak , Bk , and Dk completes the proof of the theorem.
It only remains to prove the above Lemma which helped us estimate the boundary
terms. Without going into all the details, this is done by considering the three
cases:
k 0 k 00 k, k 0 k > k 00 , k > k 0 k 00
We note that the third (low-low) case is impossible. The other two cases are
bounded using LP3 and the the following (simple) calculus inequality:
1

2
2
2 . kt f k 2 2 kf k 2 2 + kf kL2 L2
kf kL
L L
L L
t Lx
t x
t

(248)

2 using (248) and LP3 yields the estimate in the

Estimating kPk (gk0 hk00 )kL
t Lx
lemma.

Exercise.

Fill in the missing steps in the proof of the above theorem.

6. Calderon-Zygmund theory
The following L2 identity
n
X
i,j=1

ki j uk2L2 = kuk2L2 .

6. CALDERON-ZYGMUND THEORY

153

for any u C0 (Rn ) can be easily established by integration by parts, see below in
(252). Thus,
k 2 ukL2 . kukL2

(249)

It is natural to ask whether such estimate still holds true for other Lp norms. It
turns out that the problem can be reduced to that of study the Lp boundedness
properties for a very important class of linear operators called Calderon-Zygmund.
Definition 6.1. A linear operator T acting on L2 (Rn ) is called a Calderon-Zygmund
operator if:
(1) T is bounded from L2 to L2 .
(2) There exists a measurable kernel k such that for every f L2 with compact support and for x 6 suppf , we have
Z
T f (x) =
k(x y)f (y)dy,
Rn

where the integral converges absolutely for all x in the complement of

suppf .
(3) There exists constants C > 1 and A > 0 such that
Z
|k(x y) k(x)| dx A,
(250)
|x|C|y|

uniformly in y. For simplicity one can take C = 2.

Proposition 6.2. Assume that the kernel k(x) verifies, for all x 6= 0,
|k(x)| . |x|n ,

|k(x)| . |x|n1

(251)

Then k verifies the cancellation condition (250).

Exercise.

Prove the proposition.

R
Example 1. Hilbert transform Hf (x) = eix sign f()d. By Plancherel it is
easy to check that H is a bounded linear operator on L2 . On the other hand we
know that the inverse Fourier transform of sign is proportional to the principal
value distribution pv(1/x). Hence, if x 6 suppf ,
Z +
1
f (y)dy.
Hf (x) = c
x

It is easy to check that the kernel k(x) =

1
x

verifies condition 3 above.

Example 2. Consider the equation u = f in Rn , n 3, for f , smooth, compactly supported. Recall, see (??), that any solution u, vanishing at7 , can be
represented in the form, u = Kn f where Kn (x) = cn |x|2n . Thus, if x 6 suppf ,
it makes sense to differentiate under the integral sign and derive,
Z
i j u = i j Kn f =
i j Kn (x y)f (y)dy.
Rn
7In the case of n = 2 any solution whose first derivatives vanish at .

154

6. BASIC FUNCTIONAL INEQUALITIES

It is easy to check that the kernel

R k(x) = i j Kn (x) verifies condition 3. To show
that the operators Rij f (x) = Rn i j Kn (x y)f (y)dy are Calderon-Zygmund
operators, it only remains to check the L2 -boundedness property. This follows
easily from the equation u = f . Indeed u = Kn f is the unique solution of
the equation vanishing at . Moreover |u(x)| . |x|2n , |u(x)| . |x|1n and
Rij f = i j u(x). Thus we can integrate by parts in the expression,
Z
Z
n Z
X
2
|f (x)| dx =
u(x)u(x)dx =
|i j u(x)|2 dx
Rn

Rn
n
X

i,j=1

Rn

i,j=1

|Rij f (x)|2 dx

(252)

Rn

Hence for each pair 1 i, j n,

kRij f kL2 kf kL2 .
Thus the operators Rij are Calderon-Zygmund. We shall write schematically Rij =
i j ()1 .
Theorem 6.3. Calderon-Zygmund operators are bounded from L1 into weak-L1 .
As a consequence we derive,
Corollary 6.4. Calderon-Zygmund operators are bounded from Lp into Lp , for
any 1 < p < . They are not bounded, in general, for p = 1 and p = .
Proof : The boundedness over Lp for 1 < p < 2 follows from the weak-L1 and
the L2 boundedness by Marcinkiewicz interpolation. The cases p > 2 follow by
duality from the fact that the dual of a Calderon-Zygmund operator, with kernel
k(x), is again a Calderon-Zygmund operator, with kernel k(x). More precisely, if
f, g have disjoint supports,
Z
Z Z
Z
T f (x)g(x)dx =
k(x y)f (y)g(x)dx =
f (y) T g(y)dy
Rn

Rn

Rn

Rn

where
T g(y) =

Z
k(y + x)g(x)dx,

y 6 suppg.

Rn

On the other hand kT f kL2 = kT f kL2 . kf kL2 . Hence T is indeed a CZ operator.

0
Now, using the duality between Lp and Lp , 1/p + 1/p0 = 1 and the fact that T is
0
Lp bounded for p0 2,
Z
Z
f (x) T g(x)dx|
kT f kLp =
sup |
T f (x)g(x)dx| = sup |
kgkLp0 1

sup
kgkLp0 1

Rn

kgkLp0 1

Rn

kf kLp kT gkLp0 . kf kLp .

We shall prove the main theorem 6.3 in the next two subsections.

6. CALDERON-ZYGMUND THEORY

155

6.5. Calderon-Zygmund decompositions.

During our study of the Laplace operator in the Introduction, we found
R that for f
C0 , 1 (f ) would decay rapidly away from the support of f provided f (x)dx = 0.
This fact is physically important: it explains why we must have our hands in
contact with an item in order to move it, even though the same electromagnetic
force is well-known to move objects at much greater distances when there is a
concentration of positive or negative charge. We also find that a related special
behavior with respect to oscillation is quite important to the analysis of CZOs (of
which 1 is not an example, but its close relatives the Riesz potentials are). We
therefore devote the following section to a way of decomposing a general function
into one part which is bounded and other parts which oscillate and are physically
localized, and this decomposition will allow us to prove theorem 6.3.
Definition 6.6. We define a dyadic cube in Rn to be a cube Q of the form
Q = [2k a1 , 2k (a1 + 1)] [2k an , 2k (an + 1)],
where k, a1 , . . . , an Z. We then say that size (Q) = 2k . If Q is a dyadic cubes then
its parent is the only dyadic cube Q such that Q Q and size (Q ) = 2 size (Q)
and we say that Q is a child of Q .
Lemma 6.7 (Whitney decomposition). Any proper open set in Rn can be covered
by a family Q = {Q} of disjoint dyadic cubes
= QQ Q,
where each cube Q Q satisfies the property
size (Q) dist (Q, ).

(253)

Proof : For each x denote by Qx the largest dyadic cube containing x with
the property: dist (Qx , ) > size (Qx ). If Q denotes the parent of Qx then
dist (Q , ) size (Q ). By the triangular inequality it follows that

n + 2 size (Qx ).
dist (Qx , ) n size (Qx ) + dist (Q , )
Hence, Qx verifies (253). If y Qx then, by the maximality property of Qx and
Qy , we necessarily have Qy = Qx . Hence, the family Q = {Qx }x is formed of
disjoint cubes and covers .
Proposition 6.8 (Calderon-Zygmund decomposition). Let f L1 (Rn ) and > 0.
Then it is possible to findP
a countable family of disjoint dyadic cubes Q = {Q} and
a decomposition f = g + QQ bQ , such that:
kgkL . ,

(254a)

supp bQ Q,
Z
bQ (x)dx = 0,

(254b)

kbQ kL1 . |Q|,

X
1
|Q| . kf kL1 .

(254d)

(254c)

(254e)

156

6. BASIC FUNCTIONAL INEQUALITIES

Remark Note that in the above , g, bQ and f all have the same units, so that
these estimates on the sizes and supports of g and bQ are the only ones possible
that are still dimensionally correct.
Proof : Let Q be the Whitney decomposition of the open set
R = {Mf (x) > }
as indicated in Lemma (6.7). For each Q, define fQ = |Q|1 Q f (x)dx. Let
(
f (x), if x 6 ,
g(x) =
fQ ,
if x Q,
and bQ (x) = Q (x)(f (x) fQ ) with
P Q the characteristic function of the cube
Q. Of course we have f = g + Q bQ . The important property, which follows
from (253), is that each cube Q is contained inside a ball B which is not entirely
contained in and with |Q| |B|. Let x B \ , we have
Z
Z
1
1
|f (y)| dy .
|f (y)| dy Mf (x) .
(255)
|fQ |
|Q| Q
|B| B
We check now that this decomposition has the desired properties. For almost every
x outside , by Lebesgues differentiation theorem, Corollary 2.7, we have |g(x)|
Mf (x) . When x it follows from (255) that g(x) . . Hence (254a) is
satisfied. Properties (254b) and (254c) are immediate consequences of the definition
of hQ . Property (254d) is implied by (255). Finally, (254e) is nothing but the weak
L1 property for Mf proved in Theorem 2.4.

1
P 6.9. Proof of Theorem 6.3. Consider f L and > 0. Let f = g +
Q bQ = g + b be the Calderon-Zygmund decomposition of f according to Theorem 6.8. Since

{|T f (x)| > } {|T g(x)| > /2} ({|T b(x)| > /2})
and in view of (254e) it is enough to prove separately that
1
|{|T g(x)| > /2}| . kf kL1 ,
(256)

1
|{|T b(x)| > /2}| . kf kL1
(257)

Estimate (256) follows from Chebyschevs inequality, the boundedness of T on L2

and the uniform bound on g,
1
1
1
|{|T g(x)| > /2}| . 2 kT gk2L2 . 2 kgk2L2 . kgkL1

X
X
1
1
1
kf kL1 +
kbQ kL1 . kf kL1 +
|Q| . kf kL1 .

It remains to derive (257). Since the family Q is countable we denote its members
j to be the cube with
by Qj , j N. For each Qj let y(j) be its center and take Q
1/2
the same center but with the sides expanded by 2n , such that for all x in the
j ,
complement of Q
|x y(j) | 2 max |y y(j) |
yQj

6. CALDERON-ZYGMUND THEORY

157

j and F its complement. We denote bj = bQ . Since

Let = j Q
j
write, for x F ,
Z


T (bj )(x) =
k(x y) k(x y(j) ) bj (y)dy,

bj dy = 0 we

Qj

or, since the cubes Qj are disjoint,

Z


T (bj )(x) =
k(x y) k(x y(j) ) b(y)dy,
Qj

Thus, in view of (250),

Z
XZ
XZ
|T (b)(x)|dx
|T (b)(x)|dx .
F

XZ

j
xRn \Q

j y(j) }
xRn \{Q

|k(x (y yj) ) k(x)|

Z
|b(y)|

|k(x (y yj) ) k(x)|

|x|2|(yyj) |

yQj

XZ
j

|k(x y) k(x y(j) )|

|b(y)|

XZ

|k(x y) k(x y(j) )| |b(y)|,

yQj

yQj

|bj (y)|

XZ
j

j
xRn \Q

yQj

|b(y)| . kf kL1

yQj

Therefore,
|{x F : |T b(x)| > /2}|| . 1 kf kL1
j is also
On the other hand, the measure of the complement of F , i.e. = Q
controlled by,
X
X
j | .
||
|Q
Qj . 1 kf kL1 .
j

Hence,
|{x Rn : |T b(x)| > /2}|| . 1 kf kL1
as desired.
6.10. Michlin-H
ormander theorem. An important class of CZ operators
can be defined by means of Fourier multiplier operators. Recall that these are
defined by Fourier transform,
Tcf () = m()fb(),

(258)

where m is a bounded function, called the multiplier. We can view these operators
as convolution operators, T f = k f , where b
k = m. It is natural to ask when a
Fourier multiplier operator gives rise to a CZ operator. Since we know that a CZO
will grant extra decay to a localized function of mean zero, we would expect that
the multiplier m should be fairly away from the origin. This is precisely the content
of the following theorem.

158

6. BASIC FUNCTIONAL INEQUALITIES

Theorem 6.11. Let l > n/2. Suppose m is a Fourier multiplier of class C l on

b n \ 0, such that
R


bn \ 0
 m() . |||| ,
R
for every multiindex with || l. Then the operator defined by (258) is a
Calderon-Zygmund operator.
Proof : Consider the same dyadic partition of unity as that used in the LP projections,
X
1=
()
for Rn \ 0,
2Z

generated by C0 with supp {1/2 || 2}, and () = (/).

P
Decompose m into dyadic pieces, m = m , where m = m. Since | m()| .
|||| and all derivatives of () are bounded,
X
X




 m ()
 ||  .
|| || || .
(259)
||+||||
||+||||
Let k be the inverse Fourier transform of m . Since m has compact support k
is a smooth function. Moreover, for any integer N we have8
|k (x)| . |x|N k N m kL1 . |x|N nN .
P
Now take N > n and sum over 2Z . Observe that k converges to a well
defined measurable function k on Rn \0, and it easy to see that k satisfies property 2
of Definition 6.1.
The boundedness of T on L2 follows immediately from the boundedness of m on
b n.
R
For 0 j l, by Plancherels theorem and
Z
X Z
2
|x|2j |k (x)| dx '

(259) we obtain


 m ()2 d . n2j .

||=j

Let R > 0, using the case j = 0 we find that

Z
1/2
Z
2
|k (x)| dx .
|k (x)| dx
Rn/2 . (R)n/2 ,

(260)

|x|R

while using the case j = l we find that

Z

Z
|k (x)| dx .

|x|2l |k (x)| dx

1/2

|x|R

Z
|x|>R

dx
|x|2l

!1/2
. (R)n/2l .
(261)

8Recall that, by integration by parts, we have F 1 f (x) |x|N k N f k ,

L1

6. CALDERON-ZYGMUND THEORY

159

If we choose R = 1/, summing (260) and (261) we obtain kk kL1 . 1 uniformly

in . We can apply the same procedure to k , which has symbol m m , to
prove that kk kL1 . . Hence,
Z
Z Z |y|
|k (x y) k (x)| dx
|k (x ty/|y|)| dtdx (262)
|x||y|

= |y| kk kL1 . |y|,

but also, by (261),
Z
Z
|k (x y) k (x)| dx 2
|x||y|

(263)

n/2l

|k (x)| dx . (|y|)

|x||y|

(264)

We sum over using (262) when |y| 1 and (264) when |y| > 1, and obtain9
Z
X
X
|k(x y) k(x)| dx . |y|
+ |y|n/2l
n/2l . 1.
|x||y|

|y|1

>|y|1

as desired.
Exercise.
Let C0 (C) and let f be the solution to the inhomogeneous
Cauchy-Riemann equations f
z = which decays at infinity. Show that for 1 < p <
we have the estimate
||f ||Lp . ||||Lp
6.12. Square function estimates. We recall property LP6 for the square


P
2 1/2
function, Sf =
,
k |Pk f |
Theorem 6.13 (Littlewood-Paley). We have,
kf kLp . kSf kLp . kf kLp

(265)

for all 1 < p < .

We give two proofs of this estimate.
Proof [first proof]: First we show using duality arguments that the first inequality
in (265) follows from the second one. Indeed using Plancherels theorem, the fact
that Pk Pk0 = 0 unless k k 0 , and Cauchy-Schwartz inequality we obtain
Z
Z X
f (x)g(x)dx '
Pk f (x)Pk0 g(x)dx
kk0

Z
.

!1/2
X

|Pk f (x)|

!1/2
X

|Pk0 g(x)|

dx

k0

kSf kLp kSgkLp0 . kSf kLp kgkLp0 .

The left inequality in (265) now follows by taking the sup over all g with kgkLp0 = 1.
9 Here we used the following summation properties, in dyadic notation, for geometric series,

' L and

' L for > 0.

160

6. BASIC FUNCTIONAL INEQUALITIES

To prove the right inequality in (265) we need to introduce the Rademacher functions rk (t) defined on R as follows: for every k 0, k Z and t R set
rk (t) = r0 (2k t), where r0 (t) is the periodic function, r0 (t + 1) = r0 (t), such that
r0 (t) = 1 for 0 t < 1/2 and r0 (t) = 1 for 1/2 t < 1. These Rademacher
functions form an orthonormal sequence in L2 [0, 1] and they form a sequence of
independent identically distributed random variables. The basic property that we
need is that the Lp norm of a linear combination of Rademacher function is equivalent to the l2 norm of its coefficients.
P 2
Lemma 6.14. Given a sequence of real numbers {ak } satisfying
k=0 ak < ,
define

X
F (t) =
ak rk (t).
k=0

P
Then F L ([0, 1]) with kF kL2 = ( k=0 a2k )1/2 . In addition, F Lp ([0, 1]) for
1 < p < , and there exist constants Ap so that
2

A1
p kF kLp kF kL2 Ap kF kLp .
For a proof of this lemma see Stein, [?, Appendix D].
Define the operator Tt so that
Tt f =

rk (t)Pk f

k=0

P
Clearly Tt is the Fourier multiplier operator with symbol mt () = k rk (t)(2k ),
where is the smooth cut-off function used to define the LP projections. For 6= 0,
at most three of the terms in the sum defining mt () can be non-zero. We can then
easily verify that mt verifies the condition of Thm. 6.11. That is, that
| mt ()| C |||| ,
with constants C independent of t. Thus, by Calderon-Zygmund theory (specifically Corollary 6.4), we have:
kTt f kLp . kf kLp
And so,
Z
0

1/p
kTt f kpLp dt
. kf kLp

In addition, we can use Lemma 6.14 to see that:

Z 1
Z 1Z X
p
kTt f kLp dt =
|
rk (t)(Pk f )(x)|p dxdt
0

Z
&
R

!p/2
X

|(Pk f )(x)|

And so combining our results we get:

kSf kLp . kf kLp

dx

6. CALDERON-ZYGMUND THEORY

161

(Note that this argument proves the theorem only in the one-dimensional case,
n = 1. It can, however, be extended to Rn as in Stein, Singular Integrals, Ch. IV,
Section 5.)

Proof [second proof]:

We recall the definition for the vector-valued function,

Sf (x) = Pk f (x) kZ .

Clearly, if f S(Rn ), for every x Rn , Sf (x) l2 and Sf (x) = |Sf (x)| denotes
the l2 norm of Sf (x). We claim that
Z
Sf (x) = K(x y)f (y)dy
is a an l2 -valued Calderon-Zygmund operator with the l2 -valued kernel defined by,


K(x) = Kk (x) kZ ,
Kk (x) = 2nk (2
k x)




P
P
2 1/2
2 1/2
Denote |K(x)| =
, |K(x)| =
. We easily
k |Kk (x)|
k |Kk (x)|
check that the l2 valued version of the condition (251) is verified,
|K(x)| . |x|n

|K(x)| . |x|(n+1) ,

for x 6= 0.

(266)

On the other hand,

kSf kL2 := kSf kL2 . kf kL2 .
2

Thus S is indeed an l valued C-Z operator and therefore, in view of a straightforward extension of Theorem 6.3 and its corollary, we infer that,
kSf kLp := k|Sf |kLp = kSf kLp . kf kLp
In view of the beginning of the first proof of our theorem we infer that also,
kf kLp . kSf kLp .

P
P
2 1/2
Remark that, according to theorem 6.13, | k Pk f |
. A more
k |Pk f |
general principle asserts that if a sequence of functions f1 , f2 , . . . fk . . . oscillate at


P
P
2 1/2
different rates, that is any two phases are different, then | k fk |
.
k |fk |
The following version of the property LP6, and theorem 6.13, also holds true for
LP projections Pk Pk . More precisely,
X

1/2
k
|Pk f |2
kLp . kf kLp ,
1 < p < .
(267)
k

This can be proved in the same manner as the inequality kSf kLp . kf kLp by
= (Pk f )kZ , and proceeding exactly as in
introducing the l2 valued operator, Sf
the second proof of theorem 6.13. Given an l2 valued vector function g = (gk )kZ
observe that
Z
Z
Z X
X

(x)dx =
f (x)
Pk gk (x)dx
< Sf, g >=
Sf (x) g
Pk f (x)gk (x)dx =
Rn

Rn

Rn

162

6. BASIC FUNCTIONAL INEQUALITIES

Thus,
g =
S

Pk gk

(268)

gk p0 . kgk p0 , for
and therefore the estimate dual to (267) has the form, kS
L
L
0
1/p + 1/p = 1. In other words,
k

Pk gk kLp . k

|gk |2

1/2

kLp ,

1 < p < .

(269)

The following is an easy consequence of theorem 6.13.

Corollary 6.15. For 2 p < we have
X
kf k2Lp .
kPk f k2Lp .

(270)

kZ

For 1 < p 2 we have

X

kPk f k2Lp . kf k2Lp .

(271)

kZ

P
2
Proof : Recall that Sf (x)2 =
kZ |Pk f | . If p/2 1, in view of LP6 and
Minkowski inequality, we have
X
X
X
kf k2Lp . kSf k2Lp = k
|Pk f |2 kLp/2
k |Pk f |2 kLp/2 =
kPk f k2Lp .
k

If p/2 1, we make use instead of the reverse Minkowski inequality,

X
X
X
kf k2Lp & k
|Pk f |2 kLp/2
k |Pk f |2 kLp/2 =
kPk f k2Lp .
k

The reverse Minkowski inequality we have used here states that for 0 < q 1 and
a sequence of positive functions (fk )kZ
X
X
k
|fk |kLq
kfk kLq .
(272)
k

We briefly sketch a proof of (272); it can be found in many books (e.g. Garling, Inequalities or DiBenedetto, Real Analysis, from which we take this particular
proof).
One way is to first prove a reverseR Holder inequality: For 0 < p < 1, q < 0,
1
1
p
q
|f g| kf kLp kgkLq . This can be easily shown
p + q = 1, f L , g L , we have
R
1/p
p
|f g|
by writing kf kLp =
and applying the usual Holder inequality with the
|g|p
exponents p = 1/p > 1 and q = 1/(1 p) > 1.

6. CALDERON-ZYGMUND THEORY

163

With this in hand, the reverse Minkowski inequality in two terms (k|f | + |g|kLq
kf kLq + kgkLq for 0 < q 1) follows (writing q10 = 1 1q ):
Z
k|f | + |g|kqLq =
(|f | + |g|)q1 (|f | + |g|)
Z

(|f | + |g|)

(q1)q 0

1/q0
(kf kLq + kgkLq )

q1
k|f | + |g|kL
q (kf kLq + kgkLq )

6.16. W s,p - Sobolev spaces. We recall that we have defined the W s,p norm
of a function by,
s
X
kf kW s,p =
k j f kLp
j=0

We claim the following

Lemma 6.17. For any j 0, 1 < p < we have,
X

1/2
k j f kLp k
|2jk Pk f |2
kLp
k

Proof :

We first write,
k j f kLp . k

j Pk f kLp

As in the proof of the property LP5, we can express j Pk f = 2jk Pk Pk f for some
Pk similar to Pk . Hence, using the estimate (269)
X
X

1/2
k j f kLp . k
2jk Pk Pk f kLp . k
|2jk Pk f |2
kLp .
k

On the other hand, we can also write 2 Pk f = Pk j f for some other similar LP
projection. Then, in view of (267),
X
X

1/2

1/2
k
|2jk Pk f |2
kLp . k
|Pk j f |2
kLp . k j f kLp
jk

Using the lemma we can now find an equivalent definition using LP projections:
Proposition 6.18. For any 1 < p < and any s N we have,
X
kf kW s,p k
(1 + 2k )s Pk f kLp .

(273)

Moreover, for the homogeneous W s,p norm kf kW s,p = k s f kLp ,

X
kf kW s,p k
2ks Pk f kLp .
k

(274)

164

6. BASIC FUNCTIONAL INEQUALITIES

Observe that the expressions on the right hand side of (273) and (274) make sense
s,p spaces
for every value s R. We can thus extend the definitions of W s,p , and W
to all real values s.
Additional characterizations of the homogeneous Sobolev norms k kW s,p can be
given using the following,
Proposition 6.19. For 2 p < and any s we have,
!1/p
!1/2
X
X
p
2kps kPk f kLp
. kf kW s,p .
22ks kPk f k2Lp
.
k

For 1 < p 2 and s R we have

!1/2
X
2ks
2
2 kPk f kLp
. kf kW s,p .
k

Proof :

(275)

!1/p
X

kps

kPk f kpLp

(276)

If p/2 1, by Theorem 6.13 and Minkowski inequality we have

X
X
X
kf k2Lp . k
|Pk f |2 kLp/2
k|Pk f |2 kLp/2 =
kPk f k2Lp .
k

If p/2 1, by Theorem 6.13 and the reverse Minkowski inequality we have

X
X
X
kf k2Lp & k
|Pk f |2 kLp/2
k|Pk f |2 kLp/2 =
kPk f k2Lp .
k

The remaining details should be clear to fill in.

7. Problems
Problem 1.[Distributions in R]
Let f (z) be a an analytic function in the domain D+ = {z C : 0 < =(z) < }
such that |f (z)| . |=(z)|N for all z D. Show that there exists a distribution
f+ = f ( + i0) such that for every C0 (Rn ),
Z
lim
f (x + iy)(x)dx = < f+ , >,
y0,y>0

Similarly, for analytic functions defined on D = {z C/ )  < =(z) < 0} we can

define a distribution f = f ( i0),
Z
lim
f (x + iy)(x)dx = < f , >
y0,y<0

This defines, in particular when f =

(x i0)1 . Prove the formulas,

1
z

1
x+iy ,

(x + i0)1 (x i0)1

the distributions (x + i0)1 and

= 2i0 (x).

Show also that,

(x + i0)1 = x1 i0 (x)

7. PROBLEMS

where

1
x

165

is the principal value distribution defined in the text.

Problem 2.[Fundamental solutions] Consider the operator Lu = u + u in R3 .

Find all solutions of Lu = 0 with spherical symmetry. Show that
cos |x|
K(x) =
4|x|
is a fundamental solution for L.
Problem 3.[Initial value problem] Consider the initial value problems for the
following, four evolution equations in R Rn ,

t u = u,

u(0, x) = f (x)

(277)

t u = iu,

u(0, x) = f (x)

(278)

t2 u = u,
t2 u = u,

u(0, x) = f (x),
u(0, x) = f (x),

t u(0, x) = g(x)
t u(0, x) = g(x)

(279)
(280)

In each of these cases write down solutions using the Fourier transform method. In
other words take the Fourier transform of each equation, set
Z
u
(t, ) = eix u(t, x)dx
and solve the resulting differential equation in t. Compare the results for the last
two equations. Show that (279) has solutions for all f, g S(Rn ) while (280)
does not. Show however that if we only prescribe u(0, x) = f (this is the Dirichlet
problem for the Laplacian t2 + in Rn+1 ), then the problem has a unique solution
u, which decays to zero as |t| + |x| , for all functions f S(Rn ). In all cases
express10 the resulting solutions as integral operators applied to the initial data(in
physical space).
Problem 4.[Extension operator] Let H be the half space xn > 0 in Rn and
1 p . Show that there exists an extension operator, that is a bounded linear
operator E : W 1,p (H) W 1,p (Rn ) such that for all u W 1,p (H) we have Eu = u
a.e. in H and
kEukW 1,p (Rn ) . kukW 1,p (H) .
Extend the result to any s N. Can you extend the result to arbitrary domains
U Rn ? What about domains with smooth boundaries ?
Problem 5.[Distributions and Fourier Analysis on the Circle] A smooth function
on the circle R/Z is a smooth function on R which is 1-periodic
f (x + k) = f (x),

kZ

\ = Z corresponding to the
The circle has a discrete space of frequencies m (R/Z)
functions x 7 e2imx . The discreteness of the frequency space is intimately related
10 You will have to perform the inverse Fourier tarnsform, u(t, x) = F 1 u
(t, ). For the wave
equation this is more difficult, in general, but you can do it for dimension n = 3.

166

6. BASIC FUNCTIONAL INEQUALITIES

to the compactness of the circle. A Schwartz function on the circle is just a smooth
function; a Schwartz function on Z is one which decays faster than any polynomial
at infinity.
a. We define the Fourier transform of a periodic function f(m) =
Prove the Fourier inversion formula
X
f (x) =
f(m)e2imx

R1
0

f (x)e2imx .

mZ

for smooth functions on the circle. Deduce the Plancharel formula < f, g >=<
f, g >.
b. We define a distribution u on the circle to be an element of the dual of C k (R/Z)
for some k, i.e. < u, > C||||C k for some k, C and all C (R/Z). The circle
has a smooth structure, so it is possible to formulate the notion of a fundamental
solution for a differential operator (the group structure on the circle allows convolution to make sense as well) however it is not always possible to find such a
d
. In other
solution. Show that there is no fundamental solution u to the operator dx
du
words, there is no distribution u for which dx = (x) in the sense that
du
d
, > < u,
>= (0), C (R/Z)
dx
dx
There are many ways to prove this. Can you see this in both physical and frequency
d
space? What if we replace the vector field dx
by another nonvanishing vector field
d

D = dx for some nonvanishing, smooth function C (R/Z)?

<

Problem 6.[Trace theorems] Let Rn1 be a hyperplane in Rn , for example xn =

0. For any f S(Rn ) let Rf denote the restriction of f to Rn1 .
i.

Prove that, for any s > 12 ,

kRf kL2 (Rn1 ) . kf kH s (Rn )

(281)

ii. Show that the result is not true for s 1/2. Show however that the following
sharp trace theorem holds for all s > 0,
kRf kH s (Rn1 ) . kf kH s+1/2 (Rn )

(282)

iii. Show that f is a function with Fourier support in the ball || . 2k for some
integer k then, for all 1 p and s > 1/p,
kf kLp (Rn1 ) . 2k/p kf kW s,p (Rn )
Can you deduce from here a trace result, in Lp norms, generalizing that of (281) ?
What about (282) ?
iv. Let H be the half space xn > 0. According to the above considerations we
can talk about the trace of a function in W 1,p (H) to the hyperplane xn = 0( Prove
this !). Show that a function f W 1,p (H) belongs11 to W01,p (H) if and only if its
trace to xn = 0 is zero.
11recall that W 1,p (H) is the closure of C (H) in W 1,p (H)
0
0

8. RESTRICTION THEOREMS

167

Problem 7.[Littlewood-Paley] Consider the spaces = C 0, (Rn ) with norm

kf k = kf kL (Rn ) + sup

x6=yRn

i.

|u(x) u(y)|
|x y|

Show, using the Littlewood-Paley projections Pk , that

kf k kP0 f kL + sup 2k kPk f kLp
k>0

ii.

Define the Zygmund class of functions with norm,

kf k = kf kL +

sup
xRn , 0h1

|f (x + h) + f (x h) 2f (x)|
h

Show that
kf k kP0 f kL + sup 2k kPk kLp .
k>0

iii.

Prove the product estimate in Besov spaces B s = H s,1 , s > 0.

kf gkB s . kf kL kgkB s + kgkL kf kB s

Problem 8. Read on your own the section on Calderon-Zygmund operators. Indicate how the theory can be extended to operators valued in a given Hilbert space,
such as l2 .

8. Restriction Theorems
It is well known that when f L1 (Rn ) then its Fourier transform f is a bounded
and continuous function, thus the restriction of f to any hypersurface is perfectly
well defined. On the other hand, if f L2 (Rn ) then f may be any function in L2 ,
hence defined only almost everywhere and completely arbitrary on sets of measure
zero like hypersurfaces.
Can one make sense of the restriction of f to a smooth hypersurface S when f
belongs to some Lp with 1 < p < 2? This is a basic question in modern Fourier
analysis, which, as we shall see, turns out to be intimately tied to regularity properties of solutions to wave equations.
If we take S to be a hyperplane, we immediately see that the answer is negative.
Indeed, let f (x1 , x0 ) = u(x1 )v(x0 ), f(1 , 0 ) = u
(1 )
v ( 0 ), with x1 , 1 R and
0 0
n1

x , R
.R The restriction of f to the hyperplane 1 = 0 is well defined only
when u
(0) = u(x)dx
R is well defined. For any p > 1 it is always possible to find
u Lp (R) such that udx doesnt make sense. We deduce that the restriction of
the Fourier transform on hyperplanes cannot be defined when p > 1.
The answer is different if we consider hypersurfaces which have non vanishing curvature. For simplicity we consider the model case of the sphere.

168

6. BASIC FUNCTIONAL INEQUALITIES

8.1. The Stein-Tomas theorem. The following type of result was first proved
by Stein [], then extended by Tomas [] and given its final form again by Stein [].
Theorem 8.2 (Stein-Tomas). Let S = Sn1 be the standard unit sphere in Rn and
d its standard volume element. Let f Lp (Rn ) with
1p

2(n + 1)
.
n+3



Then Rf = f L2 (S) and
S

kRf kL2 (S) . kf kLp (Rn ) .

This theorem has an equivalent dual formulation. Define
the Stein operator to be



the dual of the Fourier restriction operator Rf = f  ,

S

Sg(x) = R g(x) =

eix g()d ' (gd) (x),

where now g is a function defined on the sphere.

Theorem 8.3. Let f L2 (S) and
2(n + 1)
p .
n1
Then Sf Lp (Rn ) and
kSf kLp (Rn ) . kf kL2 (S) .

(283)

Remark 8.4. It suffices to prove Theorem 8.3 for p = p = 2(n+1)/(n1). Indeed

for p > p , by Sobolev inequality we have
kSf kLp . kDs Sf kLp
for s = n(1/p 1/p) > 0, where (Ds u) () = ||s u
(). But here
Ds Sf = S(| |s f ) = Sf.
Thus, if we can prove the theorem when p = p then
kSf kLp . kSf kLp . kf kL2 (S)
Remark 8.5. The result remains true if we replace d by d = d, with
C0 (Rn ), since the theorem implies
k(f d) kLp . kf kL2 (S) . kf kL2 (S) .
Moreover, using a partition of unity, it suffices to prove Theorem 8.3 just for Sf =
(f d) , with d = d and C0 (Rn ) supported in a small neighborhood of
a point on the sphere. Though obvious, it is a very important fact that we can
localize the restriction estimate as we shall see in the future.

8. RESTRICTION THEOREMS

169

8.6. Knapp counterexample. The result of theorem 8.3 is false for any
p < p in virtue of the following counterexample ([?]).
Define, for some small > 0, the region in phase space


D = Rn : |1 1| < 2 , | 0 | < .
Let now f = SD be the characteristic function of the cap S D, then
kf kL2 (S) = |S D|1/2 (n1)/2 .
We can write
Sf (x) = eix1

ei(x,) d ,

SD

with phase (x, ) = x1 (1 1) + x0 0 . It then possible to fix a region in physical

space,
o
n

R = x Rn : |x1 | < 2 , |x0 | < 1 ,

6
6
such that for x R and D we have |(x, )| /3, hence, when x R,
Z
1
ix1
|Sf (x)| Re(e
Sf (x)) =
cos((x, ))d |S D|.
2
SD
This implies that
n1
n+1
kSf kLp
& |S D|1/2 |R|1/p 2 p .
kf kL2
For small values of , an estimate like (283) will necessarily require
which is possible only if p p = 2(n + 1)/(n 1).

n1
2

n+1
p 0,

This example suggests that there is some sort of parabolic scaling property in the
structure of the operator S which comes from the nonvanishing curvature of the
sphere.
8.7. The importance of curvature. The restriction theorem and its dual
counterpart remain true if we replace the standard sphere Sn1 by a compact
hypersurface H Rn with non-vanishing Gauss curvature. The importance of
non-vanishing Gauss curvature is illustrated by the following result.
Lemma 8.8. Let H Rn be a compact hypersurface with non-vanishing Gauss
curvature (i.e. with all its principal curvatures different from zero) and volume
element d. Then, for any smooth function , we have,
|(d) (x)| . (1 + |x|)

n1
2

(284)

If exactly one principal curvature vanishes then we have instead,

|(d) (x)| . (1 + |x|)

n2
2

Proof The general proof is based on the method of stationary phase, see Steins
Harmonic Analysis book. For the particular case of the standard sphere H = Sn1
and odd n the proof can be done by a direct computation in polar coordinates.
Exercise

Prove the lemma for S2 R3 .

170

6. BASIC FUNCTIONAL INEQUALITIES

Remark 8.9. Another interesting observation links these restiction theorems to

partial differential equations. Indeed if u = d f , then u is a solution of the
linear elliptic equation
u + u = 0,
as we can be easily seen taking the Fourier transform,
F(u + u)() ' (1 ||2 )(1 ||)f() = 0.
where is the Dirac distribution.
8.10. T T principle. The following simple functional analysis result plays
an important role in restriction and Strichartz type estimates. Let B be a Banach
space and denote by B 0 its dual. Let H be an Hilbert space with inner product
denoted by h, i. Consider a linear operator T : H B 0 . Since we can identify
H with its dual, we can consider T to be the adjoint of the operator T : B H
defined by
hh, T (x)i = T h(x).
Actually, T is the adjoint of T when B is reflexive, but for our purposes we shall
keep calling T the adjoint of T .
The T T principle states that the boundedness of T is equivalent to the boundedness of T T . More precisely we have:
Proposition 8.11. The following statements are equivalent:
(i)
(ii)
(iii)
(iv)

T : H B 0 is bounded and kT k = M ;
T : B H is bounded and kT k = M ;
T T : B B 0 is bounded and kT T k = M 2 ;
the bilinear form (x, y) 7 hT x, T yi is bounded on B B with norm M 2 .

The proof is a standard exercise in functional analysis.

8.12. T T formulation of the restriction theorem. The T T formulation for the Stein operator corresponds to a convolution with the (inverse) Fourier
transform of the measure on the sphere. Formally, we have,
Z
Z Z

ix
SS f (x) = SRf (x) = e f ()d =
ei(xy) d f (y)dy = d f (x).
S

Rn

We are thus led to the following equivalent form of the restriction theorem,
kd f kLp (Rn ) . kf kLp0 (Rn ) ,

(285)

for p p .
One can give three distinct proofs of Theorem 8.3. We shall sketch the first proof
based on analytic interpolation. This is essentially the original proof of Stein and
Tomas. The second proof, based on introducing a time parameter and treating
Sf as an evolution operator allows us to regard the restriction theorem as part
of a more general framework which includes Strichartz estimates for various linear

8. RESTRICTION THEOREMS

171

PDE with constant coefficients. Finally the third approach, which only applies for
specific exponents, will allow us to to connect with bilinear estimates.
8.13. First proof: analytic interpolation. According to Remark 8.12 and
Remark 8.4 it suffices to prove that U f = d f verifies
kU f kLp (Rn ) . kf kLp0 (Rn ) ,

(286)

where p = 2(n + 1)/(n 1) and p0 = 2(n + 1)/(n + 3).

0

In general, to obtain Lp Lp estimates directly is usually very complicated and

we dont know any direct proof except in cases where p is a nice exponent like
p = 4, 6 (which happens only for n = 2 or n = 3). We would feel more comfortable
with L2 L2 type estimates, where Plancherels theorem is a powerful tool, or with
L1 L type estimates, since pointwise decay estimates of oscillatory integrals can
be obtained from stationary phase methods. This suggests to use some interpolation
theory for Lp spaces. But, an L2 L2 estimate for the operator U is ruled out by
the Knapp counterexample and a L L1 one is too trivial and doesnt answer to
our question. It is here that the Stein interpolation theorem, Thm. 1.8, shows its
0
power, since it allows us to obtain the Lp Lp estimate for U from L2 L2 and
L L1 estimates for other (reasonable) operators different from U .
We will accomplish this by constructing a family of convolution operators Uz f =

z f , with z being distributions depending analytically in z. The parameter z will

essentially reflect the degree of homogeneity of the distribution z . For this reason
it is natural to place our target at z = 1, requiring U1 = U or 1 = d, since
d can be written as the pullback of a delta distribution (which is homogeneous of
degree 1) on the sphere: d ' (1 ||)d.
An L2 L2 estimate for Uz will follow if z coincides with a bounded function,
indeed, by Plancherels theorem, we have
kUz f kL2 ' k(Uz f ) kL2 ' kz fkL2 . kz kL kf kL2 .

(287)

To have z () bounded we must require that z () is essentially homogeneous of

degree 0, hence when z lies on the line Re(z) = 0.
An L1 L estimate for Uz will follow instead when
z coincides with a bounded
function, since we directly have
kUz f kL . k
z kL kf kL1 .

(288)

To obtain (286) from the analytic interpolation of (287) and (288), we would like
the latter to happen on the line Re(z) = a, where a is chosen so that
1 = a + (1 )0,

1
=
+
,
p

1
1
= +
,
0
p
1
2

and this happens precisely when Re(z) = a = (n + 1)/2.

This argument leads to the precise version of the Stein analytic interpolation theorem that we are going to use.

172

6. BASIC FUNCTIONAL INEQUALITIES

Proposition 8.14. Let Uz be an analytic family of linear operators such that:

(i) U1 = U ;
(ii) kUz f kL2 . kf kL2 , uniformly on the line Re(z) = 0;
(iii) kUz f kL . kf kL1 , uniformly on the line Re(z) = (n + 1)/2.
Then it follows that
kU f kLp . kf kLp0 .
The above discussion showed that, when we write Uz f as the convolution
z f,
then the hypothesis of the proposition are fulfilled whenever z is an analytic family
of distribution such that
(i) 1 = d;
(ii) z () coincides with a bounded function, with a uniform bound on the
line Re(z) = 0;
(iii)
z (x) coincides with a bounded function, with a uniform bound on the
line Re(z) = (n + 1)/2.
It thus remains to define the distributions z and verify these properties.
Inspired by the identity = 1
+ and d ' (1 ||), we define our family of
distributions as
2

z () = ez z+ (1 ||)(||),

(289)

where C0 (R) is a cut-off function supported in a small neighborhood of 1, say

[1/2, 3/2], and (1) = 1.
We recall that the homogeneous distributions z+ , when Re(z) > 1, coincide with
the functions:
(
tz /(z + 1) if t 0,
z
+ (t) =
0
if t < 0,
R
where the Gamma function is defined by (z + 1) = 0 tz et dt. From the identity
(z + 1) = z(z), it follows that
d z
z1
(t) = +
(t).
(290)
dt +
z
Using this formula, + can be analytically continued for all z C by performing
repeated integrations by parts. To do this we first observe that for Re(z) > 1 and
C0 we have
Z
Z
Z
z+1
z
0
m
+ (t)(t)dt = + (t) (t)dt = . . . = (1)
z+m
(t)(m) (t)dt.
+
R
Thus integrating by parts sufficiently many times we can make sense of z+ dt
when Re(z) > 1 m for any m, and hence for all z. To see that 1
+ = it takes
just an integration by parts, indeed
Z
Z
1
1
0 (t)dt = (0).
+ dt =
(1) 0

8. RESTRICTION THEOREMS

173

For more information about z+ and distribution theory one can consult the books
by Gelfand and Shilov [Ge-S] or Hormander [?].
2

The factor ez in the definition of z is chosen in order to garantee a uniform

2
boundedness of our operators for large =(z), indeed ez decreases exponentially as
=(z) , uniformly on the strip (n + 1)/2 Re(z) 0. This permits to allow
the various constants in the following inequalities to have a polynomial growth in
terms of b = =(z).
Clearly 1 ' (1 ||)(||) ' d. This verifies (i).
Condition (ii) is immediately verified, since z
+ is always a bounded function
when Re(z) = 0. Condition (iii) will follow from stationary phase arguments,
more generally we have:
Proposition 8.15.
Re(z)1 n1
2

|
z (x)| . (1 + |x|)

(291)

8.16. Second proof: evolution operators approach. In this section we

make the following assumption on f :
f C (S),

suppf {1 > 1/2} .

(292)

With this assumption we can relabel x1 = t as a time parameter and rewrite Sf as

Z

p
d 0
0
it 1| 0 |2 ix0 0
0 |2 , 0 ) p
Sf (t, x ) =
e
e
f
(
1

1 | 0 |2
| 0 |< 3/2
Z

0 0
0 2
= eit 1| | eix (| 0 |)g( 0 )d 0 .
p
p
with C0 supported in | 0 | < 1 and g( 0 ) = f ( 1 | 0 |2 , 0 )/ 1 | 0 |2 . Observe that
Z
Z
|f ()|2
0 2
0
|g( )| d =
d ' kf k2L2 (S)
2
S |1 |
by the assumption on the support of f .
Theorem 8.17. Let C0 (Rn1 ) be supported in the unit ball { Rn1 : || <
1} and consider the operator
Z

2
T g(t, x) =
eit 1|| eix ()g()d, t R, x Rn1 .
Rn1

Let q, r be Lebesgue exponents verifing the conditions:

2
0 min {1, (r)} ,
q


2
, (r) 6= (1, 1),
q

(293)
(294)

where (r) = (n 1)(1/2 1/r). Then the following estimate holds true for all
g C0 (Rn1 ),
kT gkLqt Lrx (RRn1 ) . kgkL2 (Rn1 ) .

(295)

174

6. BASIC FUNCTIONAL INEQUALITIES

where we use the mixed norm notation defined in section 7.

n+1
By Remark 8.5, Theorem 8.3 follows from the special case q = r = 2 n1
.

Remark 8.18. We can run again the Knapp example to prove the necessity of
condition (293), when q 2. Indeed let D Rn1 be the disk defined by || ,
for sufficiently small > 0, and take g = D to be the characteristic function of D.
We write,
Z

2
T g(t, x) = eit
eit( 1|| 1) eix ()d
D

and observe that for |t| 2 and |x| 1 we have, with a p

fixed constant c > 0,
|T g(t, x)| c. Indeed this follows easily from | and | 1 ||2 1| . 2 .
Therefore, if (295) holds true, we must have, for all sufficiently small > 0,
2

c q

n1
r

. kT gkLqt Lrx . kD kL2 .

n1
2

from which (293), q 2 follows.

Remark 8.19. The end-point restriction (294) can be removed when n 6= 3, due to
a well known result by Keel and Tao [K-T] (Endpoint Strichartz Inequalities).
The other restriction q 2, implicit in (293) will be discussed in the next chapter.
We start by calculating T and T T .
ZZ
< T F, g > =< F, T g >=
F T gdtdx =
ZZ
Z

2
=
F (t, x) eit 1|| eix ()g()ddtdx =
Z Z

Z

it 1||2 ix
= g()()
e
e
F (t, x)dtdx d.
Hence
T F () = ()
and

2
eit 1|| eix F (t, x)dtdx,

2
eit 1|| eix ()T F ()d
ZZ

2
=
ei(ts) 1|| eix |()|2 F (s, )dds,

T T F (t, x) =

where F (s, ) =

ZZ

eix F (s, x)dx. If we introduce the family of operators

Z

2
U (t)f (x) = eit 1|| eix |()|2 f()d,

we can write T T as a convolution operator,

Z
T T F (t, ) = U (t s)F (s, )ds.

(296)

By Proposition 8.11, to show that T is a bounded operator from Lqt Lrx (Rn ) to
0
0
L2 (Rn1 ) it suffices to prove that T T is a bounded operator from Lqt Lrx (Rn ) to
Lqt Lrx (Rn ).

8. RESTRICTION THEOREMS

175

We shall first prove an estimate for U (t).

Proposition 8.20. Let 2 r and (r) = (n 1)(1/2 1/r). Then U (t)
verifies the estimate
kU (t)f kLr (Rn1 ) . (1 + |t|)(r) kf kLr0 (Rn1 ) .

(297)

Proof Once we have proved the two extreme cases r = 2 and r = ,

kU (t)f kL2 (Rn1 ) . kf kL2 (Rn1 )

(298)

kU (t)f kL (Rn1 ) . (1 + |t|)(n1)/2 kf kL1 (Rn1 )

(299)

then the estimate follows from the standard Riesz interpolation theorem.
We obtain (298) immediately using Plancherel formula, since

2
(U (t)f ) () ' eit 1|| |()|2 f().
To prove (299) we write
Z
U (t)f (x) =
where

Kt (x y)f (y)dy,

2
eix eit 1|| |()|2 d
ZZ
p
'
eix eit (1 2 ||2 ) 1 ||2 |()|2 d d
ZZ


'
ei(t,x)(,) (1 |(, )|)1 (, )d d,
1 (, ) = |()|2 ,
Z

Kt (x) =

= (1 dn1 ) (t, x).

Hence Kt is just the Fourier transform of a measure supported on the sphere Sn1 ,
for which we have the decay estimate
|Kt (x)| . (1 + |t| + |x|)(n1)/2 ,
which implies (299).
We next apply Proposition 8.20 to (296),
Z
1
kT T F (t, )kLrx .
(300)
kF (s, )kLrx0 ds.
(1 + |t s|)(r)
Finally, we are in a position to apply the Hardy-Littlewood-Sobolev inequality and,
if 0 < (r) < 1, we obtain
kT T F kLqt Lrx . kF kLq0 Lr0 ,
t

when (r) + 1 + 1/q = 1/q , hence (r) = 2/q. Therefore we proved Theorem
8.17 in the case 0 < (r) = 2/q < 1.
On the other hand if q = 2 and (r) > 1 we have from (300),
kT T F kL2t Lrx . kF kL2t Lrx0 ,
by an application of the standard Hausdorff-Young inequality.

176

6. BASIC FUNCTIONAL INEQUALITIES

Finally, if 2/q < 1 and (r) > 2/q the result follows from the case (r) = 2/q using
Sobolev inequalities.
8.21. Third proof: bilinear forms (n = 2 and n = 3). We present now
another method to prove the restriction theorem for the sphere that works for the
special cases n = 2, p = 6 or n = 3, p = 4. The idea is that when p is an even integer,
the restriction theorem can be viewed as an L2 estimate for a multilinear form,
which, through the Fourier transform, has a convolution structure that provides
some smoothing effects. The proofs given below are at the root of the so called
bilinear trilinear estimates, which play a fundamental role in the modern theory of
nonlinear wave and dispersive equations.
Let us see the case n = 3 first. We consider the Stein operator Sf = (f d) , and
use the fact that (Sf Sf ) ' (f d) (f d). Let B(f, g) = Sf Sg, then an L4
estimate for Sf corresponds to an L2 estimate for B(f, f ). We have
Z

B(f, g)() ' (f d) (gd)() =

(1 | |)(1 ||)f ( )g()d,
R3

and applying Cauchy-Schwarz with respect to the measure (1 | |)(1 ||)d

we find


g)()|2 B(1,
1)()B
|f |2 , |g|2 ().
|B(f,
Integrating with respect to , we obtain
kB(f, g)k2L2 (R3 ) . Akf k2L2 (S2 ) kgk2L2 (S2 ) ,

(301)

with
1)()| = sup
A = sup |B(1,

Z
(1 | |)(1 ||)d.

(302)

Thus, to prove the theorem in this case it suffices to check that A is finite. It is
1)(). For any dimension
useful to carry out the explicit calculation of A() = B(1,
n 2 we have:
Lemma 8.22.
Z
(1 | |)(1 ||)d '

A() =
Rn

n3
1
4 ||2 +2 .
||

(303)

Proof
Z
A() =

Z
(1 | |)(1 ||)d '

(1 | |2 )d =

||=1

1
=
(|| 2 )d '
||
||=1
2

||=1


||

d .
2
||

Because of the rotational symmetry, we may assume that = (||, 0, . . . , 0), so that

Z 
1
||
A() '

cos (sin )n2 d =

|| 0
2


 n3
Z 1 
2
n3
1
||
1
||2
=

u (1 u2 ) 2 du =
1
,
|| 1
2
||
4

8. RESTRICTION THEOREMS

177

when ||/2 [1, 1].

When n = 3, A() ' 1/|| is singular only at = 0, but we can avoid this difficulty
by assuming that f and g are supported in a small neighborhood of a point in
S 2 (recall that without loss of generality we can localize the estimate on a small
cap on the sphere). Then the supremum in (302) can be taken over just all
supp(f ) + supp(g), which is a set bounded away from 0. Hence we may restrict
to || C > 0 in (302) and the singularity disappears leaving A < .
From the L2 estimate (301) of the bilinear form B(f, g), it follows the L4 estimate
for the Stein operator Sf :
kSf k2L4 (R3 ) = kB(f, f )kL2 ' A1/2 kf k2L2 (S2 ) ,
with the assumption that f is supported in a small cap on the sphere.
In the case n = 2 what we want is an L6 estimate for Sf . Since 6 = 3 2 we
can try to repeat the same calculation using this time a trilinear form, T (f, g, h) =
Sf Sg Sh, and the fact that kSf k3L6 = kT (f, f, f )kL2 . We have
T(f, g, h)() ' (f d) (gd) (hd)() =
ZZ
=
(1 | |)(1 ||)(1 ||)f ( )g()h()dd,
R2 R2

and applying Cauchy-Schwarz with respect to the measure (1||)(1||)(1

||)dd we find


|T(f, g, h)()|2 T(1, 1, 1)()T |f |2 , |g|2 , |h|2 ().
Integrating with respect to , we obtain
kT (f, g, h)k2L2 (R2 ) . Akf k2L2 (S1 ) kgk2L2 (S1 ) khk2L2 (S1 ) ,

(304)

with
A = sup |T(1, 1, 1)()| = sup

ZZ
(1 | |)(1 ||)(1 ||)dd.

(305)

The convolution structure allows us to restrict to the set suppf + suppg+ supph,
and, if we make the hypothesis of f, g, h supported in a small cap of the sphere, we
can assume 1 || 3. Using Lemma 8.22 we can evaluate T (1, 1, 1) and show
that A is bounded,
Z
T (1, 1, 1)() = B(1, 1)( ) (1 ||)d
Z
Z
(1 ||)
d

d
=
'
S1
2 )1/2
(4

|
(3

+ ||2 )1/2
||<2
||<2
Z 1
Z 1
da
da
'

' 1,
2
1/2
2
1/2
1/2
(1 a )
(1 a)1/2
a() (3 || + 2||a)
a() (a a())
2

2
where a() = 3||
2|| . From the L estimate (304) of the trilinear form T (f, g, h),
it follows the L6 estimate for the Stein operator Sf :

kSf k3L6 (R2 ) = kT (f, f, f )kL2 ' A1/2 kf k3L2 (S2 ) .

178

6. BASIC FUNCTIONAL INEQUALITIES

We can also try to repeat the bilinear argument for n = 2. As before, for B(f, g) =
Sf Sg we have


g)()|2 B(1,
1)()B
|f |2 , |g|2 ().
|B(f,
1),
Integrate with respect to , and use Lemma 8.22 to evaluate B(1,
ZZ
(1 | |)(1 ||)
|f ( )|2 |g()|2 dd.
kB(f, g)k2L2 (R2 ) .
||(4 ||2 )1/2
Change variable, = , and observe that when || = || = 1 we have
|| = | + | ' (1 + )1/2 ,
(4 ||2 )1/2 = (4 | + |2 )1/2 ' (1 )1/2 ,
hence
kB(f, g)k2L2 (R2 ) .

|f ()|2 |g()|2

ZZ
S1 S1

(1 ( )2 )

1/2

d d .

(306)

This is an interesting formula. Observe that if the supports of f and g on S1 are

projectionally disjoint, i.e. dont contain points in the same direction, then the
quantity 1 ( )2 is bounded below by a positive constant and in this case we
obtain the bilinear restriction estimate
kB(f, g)kL2 (R2 ) . kf kL2 (S1 ) kgkL2 (S1 ) .
We can consider also other types of bilinear forms which have a special structure that cancel the singularity in the denominator. Take for example Q(f, g) =
1 Sf 2 Sg 2 Sf 1 Sg, then taking the Fourier transform and proceeding as before
we see that
ZZ
|1 2 2 1 |2
2
kQ(f, g)kL2 (R2 ) .
|f ()|2 |g()|2 d d
1/2
S1 S1 (1 ( )2 )
.

kf k2L2 (S1 ) kgk2L2 (S1 ) ,

since we have the identity |1 2 2 1 |2 = 1 ( )2 1.

Part 3

Analysis of the Wave Equation in

Minkowski Space

CHAPTER 7

Decay estimates
Consider the standard wave equation in Minkowski space Rn+1
 = 0.

(307)

The canonical, inertial, coordinates in Rn+1 are denoted by x , = 0, 1, . . . , n relative to which the Minkowski metric takes the diagonal form m = diag(1, 1, . . . , 1).
We have x0 = t and x = (x1 , . . . , xn ) denote the spatial coordinates. We make use
of the standard summation convention over repeted indices and those concerning
raising and lowering the indices of vectors and tensors. In particular, if x = m x ,
we have x0 = t and xi = xi , i = 1, . . . , n. We denote by t0 the spacelike
P hyperplanes t = t0 . The wave operator is defined by  = m = t2 + i i2 . We
study the initial value problem,
(0, x) = f (x),

t (0, x) = g(x)

(308)

For convenience we denote [0] = (f, D1 g) with D1 the pseudodifferential operator with symbol ||1 . Let,
Z
X
E[](t) =
(|t |2 +
|i |2 )dx
(309)
t

be the total energy of at time t. The conservation law for the energy is,
E[](t) = E[](0)

(310)

As a consequence we have the energy inequalities, for all s 0,

k(t)kH s (Rn ) k (0) kH s (Rn ) .
The energy identity can be proved both by the geometric techniques discussed in
the previous sections, involving only integration by parts, or by the Fourier method,
using Plancherel formula together with the Fourier representation formula,
Z


sin t||
n
g() d
(311)
(t, x) = (2)
eix cos t||f () +
||
Remark 0.23. The standard Sobolev embedding H s (Rn ) L (Rn ), for s >
n

bounds of solutions to (307) without using the explicit

2 allows us to get L
representation. This procedure generalizes to nonlinear equations and plays an
important role in the proof of the local existence theorem.
Proposition 0.24 (Dispersive inequality). The solutions to (307), (308) verify,
k(t)kL c|t|

n1
2

181

k[0] k

n+1

B1,12

(312)

182

7. DECAY ESTIMATES
n+1

with B1,12 the Besov space slightly larger than W n+1

2 , 1. More precisely,
X n+1
kf k n+1

2k 2 kPk f kL1 (Rn )

2
B1,1

kZ

Exercise. Show that the inequality (312) follows from its frequency localized
version. In other words show that it suffices to prove the following inequality,
k(t)kL c|t|

n1
2

k[0] kL1

for initial data f, g whose Fourier transforms are localized to

(313)
1
2

|| 2.

Proof The standard proof of (312) is based on the method of stationary phase
applied to the representation (311). In odd dimensions one can prove a related form
of the dispersive estimate using the spherical means representation of solutions.
This is particularly easy to do for n = 3. We shall later discuss a derivation of
(312) which avoids any representation formulas.
Remark 0.25. The dispersive inequality provides two types of information. The
first concerns the precise decay rate of k(t)kL as t while the second provides information about the regularity properties of k(t)kL for t > 0. As far as
improved regularity is concerned the estimate (312) gains, for t > 0, n1
2 derivatives
when compared to the Sobolev embedding L (Rn ) W 1,n (Rn ).
In many applications, especially to nonlinear equations, (312) is not very useful.
A more effective procedure to derive the asymptotic properties of solutions of the
wave equation is based on generalized energy estimates, obtained by the commuting
vectorfields method, together with global Sobolev inequalities. In what follows
we review the commuting vectorfields method for deriving the above decay rate
estimate. The idea is to use the energy identity (310) together with the vectorfields
which commute with the wave operator  and and a global version of the classical
Sobolev inequalities We refer the reader to [?] and [?] for details.
The Minkowski space-time Rn+1 is equipped, see appendix 4.2, with a family of
Killing and conformal Killing vector fields, the translations T = , Lorentz rotations L = x x , scaling S = tt + xi i and the inverted translations
K = 2x S+ < x, x > . Recall that x , denote the standard variables x0 = t,
x1 , . . . , xn , and x = m x . The Killing vector fields T and L commute with
 while S preserves the space of solutions in the sense that  = 0 implies S = 0
as [, S] = 2. One can split the operators L into the angular rotation operators
(ij)
O = xi j xj i and the boosts (i) L = xi t + ti , for i, j, k = 1, . . . , n. Recall
the energy expression in (309). Based on the commutation properties described
above we define the following generalized energies

Ek [] =

X
Xi1 ,..,Xij

E[Xi1 Xi2 ...Xij ]

(314)

7. DECAY ESTIMATES

183

with the sum taken over 0 j k and over all Killing vector fields T, L as well
as the scaling vector field S. The crucial point of the commuting vectorfield method
is that the quantities Ek , k 1 are conserved by solutions to (307). Therefore, if,


X Z
2k
k+1
2
k
2
(1 + |x|)
|
f (x)| + | g(x)| dx Cs <
(315)
0ks

then for all t, Es [](t) Cs . The desired decay estimates of solutions to (307) can
now be derived from the following global version of the Sobolev inequalities ( see
[?], [?]):
Theorem 0.26 (Global Sobolev). Let be an arbitrary function in Rn+1 such that
Es [] is finite for some integer s > n2 . Then,
|(t, x)| . (1 + t + |x|)

n1
2

(1 + |t |x||) 2 sup Es [](t0 )

0t0 t

for all t > 0. Therefore if the data f, g in (307) satisfy 315, with s >
all t 0,
1
|(t, x)| .
n1
1
2
(1 + t + |x|)
(1 + |t |x||) 2

(316)
n
2,

then for
(317)

Remark 0.27. Clearly this estimate, whose proof is purely geometric1, implies the
decay properties given by the dispersive inequality (312). In fact it provides more
information outside the wave zone |x| t which fit very well with the expected
propagation properties of the linear equation  = 0. On the other hand, as (316)
is really a global version of the Sobolev inequality, it seems that the estimates of
the Proposition 0.26 have no bearing on the improved regularity features of (312).
This is however not quite true as we shall see, later.
Proof We only sketch the main ideas of the proof below. Consider the canonical
null pair L = t r ,, an associated null frame e1 , . . . en1 , en = L , en+1 = L+
as well as the angular vectorfields, Ai = i xri r . Clearly,
X
X
|Ai | . |
/ | .
|Ai |.
i

where |
/ |2 =

Pn1

2
i=1 |ei ()| . Also,
X
X
|r | +
|Ai | . || . |r | +
|Ai |
i

We can also easily check the following simple algebraic identities,

1
(t + r)L+ + (t r)L
= S
2
X xi


1
(t + r)L+ (t r)L
=
Li
2
|x|
i
xi X xj
tAi = Li
Lj
|x| j |x|
tOij

= xi Lj xj Li

1 In particular it does not require any explicit representation of solutions

184

7. DECAY ESTIMATES

From the first two identities we easily derive,

1
|(t, x)|
t
1
 |(t, x)|
|L (t, x)| . 
t |x|
|L+ (t, x)| .

(318)

with || = |S| + |L|.

|
/ (t, x)| .

1
|(t, x)|.
t

(319)

Clearly, we also have,

1
 |(t, x)|
|(t, x)| . 
t |x|
or, more generally,

1
N
| N (t, x)| . 
 | (t, x)|
t |x|N
where |N | =

(320)

|1 . . . N | with 1 , . . . , N any of the vectorfields S, L1 , . . . Ln .

Combining the above inequalities with the definition of our norms we derive

tkE+ (t)kL2

. k(t)kL2 (Rn )

tk
/ (t)kL2

. k(t)kL2 (Rn )

kuE (t)k

. k(t)kL2 (Rn )

L2

where u = |t |x|| It remains to derive sup-norm estimates from the L2 estimates

above.
Proposition 0.28. Let  = 0 with initial data verifying the assumptions above.
Then, for all t 0, s > n2 ,
1
n1
2
ks (t)kL2 (Rn )
t
1
n1
2
.
ks (t)kL2 (Rn )
t

k(t)kL .
k(1 + |u|)k k (t)kL
Also,

1
n+1
2
ks+1 (t)kL2 (Rn )
t
1
n+1
2
.
ks+1 (t)kL2 (Rn )
t
1
n1
2
.
ks+1 (t)kL2 (Rn )
t

kE+ (t)kL .
k
/ (t)kL
k(1 + |u|)E (t)kL

(321)
(322)

7. DECAY ESTIMATES

185

The proof is based on the following Lemma

Lemma 0.29. Let u(x) be a smooth, compactly supported function on Rn , n 2.
We have,


1
n1
/)
r ukL1
(323)
|u(x)| C n1 kr ukL1 + k(r
|x|

Proposition 0.30 ( see[?]). The commuting vectorfields method implies the dispersive inequality (312).

Proof Without loss of generality we may assume that t = g = 0 and that the
Fourier transform of f = (0) is supported in the shell 2 || 2 for some
2N . By a simple scaling argument we may in fact assume = 1. For such
initial conditions, with Fourier supports restricted to 1/2 || 2, it suffices to
prove,
k(t)kL . (1 + |t|)

n1
2

kf kL1 (Rn )

the Fourier transform of relative to the space variables x, is also supported

Since ,
in the same shell it suffices to prove the estimates for instead of .
Next we cover Rn by an union of discs DI centered at points I Zn with integer
coordinates such that each DI intersects at most a finite number cn of discs DJ
with cn depending only on the dimension n. Consider a smooth partition of unity
(I )IZn with supp I DI and each I positive. Clearly we can arrange to have,
for all k,
X

|k I (x)| Ck,n

(324)

IZn

uniformly in x Rn , with a constant Cn,k depending only on n and k. Now

set, fI = I f , and I thePcorresponding
P solution to (307) with data I (0) =
fI , t I (0) = 0. Clearly f = I fI , = I I . It suffices to prove that for all I,
kk I (t)kL . (1 + t)

n1
2

n+k+1
X

kDj fI kL1

(325)

j=0

Indeed if (325) holds true we easily infer that,

kk (t)kL

X
I

kk I (t)kL . (1 + t)

n1
2

n+k+1
X X
j=0

kDj fI kL1

186

7. DECAY ESTIMATES

In view of (324) we have,

X
X
X X Z
kDj fI kL1 =
kDj (I f )kL1 .
I

0ij

X Z
X

Rn

0ij

|Di I (x)||Dji f (x)|dx

Rn


|Di I (x)| )Dji f (x)|dx

ci,n kD

ji

f kL1 (Rn ) . kf kL1 (Rn )

0ij

Hence,
kk (t)kL

(1 + t)

n1
2

kf kL1 (Rn )

as desired.
It therefore remains to check (325). Without loss of generality, by performing a
space translation, we may assume that I = 0. Applying the proposition 0.26 to
= 0 we derive, for s the first integer strictly larger than n2 ,
k(t)kL

c(1 + t)

n1
2

n1
2

c(1 + t)

Es [0 ](t)
Es [0 ](0).

Since the support of 0 is included in in the ball of radius 1 centered at the origin
we have,
sX
+1
Es [0 ](0) Cn
kDj f0 kL2 .
j=0
n

Finally, according to the standard Sobolev inequality in Rn , kf kL2 ck 2 f kL1 ,

we conclude with,
k(t)kL c(1 + t)

n1
2

n+2
X
j=0

as desired.

kDj f0 kL1

CHAPTER 8

Strichartz Inequalities

Strichartz inequalities are an important tool in the study of linear and nonlinear
wave equations. They are intimately tied to restriction theorems. In this chapter
we shall only consider the case of the standard linear wave equation. Similar inequalities hold true however for linear dispersive equations such as the Schrodinger,
linear KdV etc.
0.30.1. Homogeneous wave equation. Consider solutions u = u(t, x), t R, x
Rn to the equation
u =
u(0, x)

F,

(326)

f (x),

t u(0, x) = g(x),

(327)

with  the wave operator 2 = t2 u + . Clearly, a solution to eqrefeq:genwave

can be written as a superposition between a solution to the homogeneous wave
equation,
2u = 0,

(328)

verifying the initial condition (327) at time t = 0, and a solution to the purely
inhomogeneous wave equation
2u = F,

(329)

with zero initial data

u(0, x) = 0,

t u(x, 0) = 0.

We denote by W (t)h
the fundamental solution of the homogeneous problem (328),

i.e. u(t, x) = W (t)h (x) is the unique solution of (328) which verifies the initial
conditions
u(0, x) = 0,

t u(0, x) = h(x)

By Duhamels principle any solution of the inhomogeneous equation can itself be

written as a superposition of solutions to the homogeneous equation according to
the formula,
Z t
u(t) =
W (t t0 )F (t0 )dt0 ,
(330)
0

Before stating the main result of this section we make the following definition.
187

188

8. STRICHARTZ INEQUALITIES

Definition 0.31. We say that the pair of real numbers (q, r) is an admissible wave
pair if they satisfy the conditions
q
2

q
(q, r, n) 6=

2,

(n 1)

1 1

2 r


,

(2, , 3).

We are now ready to state the following.

Theorem 0.32. Suppose that n 2 and (q, r) is a wave admissible pair
r < .

with

(1) Assume the dimensional condition, 1q + nr = n2 . Then, if u verifies the

homogeneous equation (328) with initial conditions (327),
kukLqt Lrx + kukL H x + kt ukL H x1
t

. kf kH + kgkH 1

(331)

(2) Assume the dimensional condition,2 1q + nr = n2 = q10 + rn0 2, with q 0

dual to q and r0 dual to r. Then, if u verifies the purely inhomogeneous
problem (0.30.1) with zero initial conditions, then on a finite time interval
[0, T ]:
kukLq ([0,T ];Lr ) + kukC([0,T ];H ) + kt ukC([0,T ];H 1 )

. kF kLq0 ([0,T ];Lr0 )(332)

(3) We also have the following more general version of (332) for admissible
pairs (q1 , r1 ), (q2 , r2 ) with r1 , r2 < verifying the dimensional condition,
1
n
1
n
n
+
= = 0 + 0 2
q1
r1
2
q2
r2
Then,
kukLq1 ([0,T ];Lr1 ) + kukC([0,T ];H ) + kt ukC([0,T ];H 1 )

kF kLq20 ([0,T ];Lr20 )

(333)

Remark 0.33. For n 4, the region of admissable exponents corresponds to a

quadrilateral OEP Q in the plane (1/q, 1/r) with vertices O = (1/, 1/), E =
n3
(1/, 1/2), P = (1/2, 2(n1)
) and Q = (1/2, 1/). When n = 3 the point P
coincides with Q and the region reduces to the triangle OEQ. When n = 2 we have
a smaller triangle OEQ2 where Q2 = (1/4, 1/).
For n = 3, the boundary of the triangular region is allowed except for the endpoint
P . For n 4, the boundary of the quadrilateral region is entirely allowed, as we
will note below.
The interesting cases are the ones on the segment EP and the ones on P Q close
to P , since all the others can be deduced from these using Sobolev embeddings.
The point E corresponds to the energy estimates. There are counterexamples that
1The case when r = can also be included provided that we modify the spaces on the left
of the estimates below to appropriate Besov spaces.
2Thus, in fact, = 1/2.

8. STRICHARTZ INEQUALITIES

189

Figure 1. Admissable exponents for n 4

exclude the point P when n = 3, while the inclusion of P in higher dimensions were
recently obtained by Keel and Tao [K-T].
n1
n1
The standard Strichartz estimate3 corresponds to the point S = ( 2(n+1)
, 2(n+1)
).

Remark 0.34. We remark that in even though the end-point case n = 3, q = , r =

2 is forbidden, the estimates holds in the spherically symmetric case. Indeed let
be a solution of the homogeneous wave equation  = 0 in R3+1 subject to the
initial conditions
(0, x) = 0, t (0, x) = f (x)
and assume that f is spherically symmetric i.e. f (x) = f (|x|). Then,
Z
k(t, )k2L (R3 ) dt ckf k2L2 .
(334)
0

The proof is an immediate consequence of the HardyLittlewood maximal theorem4

in view of the fact that, for spherically symmetric f ,
Z |x|+t
c
(x, t) =

 f () d.
|x| |x|t
Remark 0.35. We give an elementary example below to illustrate how the end point
result n = 3, q = , r = 2 fails in the general case due to possible concentrations
along null rays. We show below that there exists a sequence of functions fn in
C0 (R3 ), with kfn kL2 = 1 such that for the corresponding solutions n ,
Z
|n (t, t, 0, 0)|2 dt n.
(335)
0
3i.e. the one actually proved by Strichartz.
4This is obviously so in the region r t while for r t the argument is elementary.

190

8. STRICHARTZ INEQUALITIES

R
assume by contradiction that in fact, J := 0 (t, t, 0, 0)(t) dt < C for all f
C0 (R3 ) with kf kL2 = 1 and some S(R), 6 0. In view of the formula (see
section on the fundamental solution of  in R3+1 ),
Z
1
(t, x) = (4) t
f (x + t) d
||=1

we find that,
1

|y|1 f1 (y1 + |y|, y2 , y3 )(|y|) dy

J = (4)

R3

or, changing the variables z = y + (|y|, 0, 0)

Z
 |z|2 
1
1
f (z)
dz < c.
J = (4)
z1
2z1
z1 >0

Since f is an arbitrary

C0 (R3 )

function, kf kL2 = 1, we must have that,

1  |z|2 
z
z1
2z1

is in L2 (R3+ ) which is false whenever 6 0. In fact,

Z
Z
Z
Z
1
1 2  |z|2 
sin
2
2
(|y|)dy = 2
d
dz =
()
2
3
z
2z
(y
+
|y|)|y|
1
+
cos
3
1
1
R+ 1
R
0
0
diverges logarithmically if 6 0.
0.36. Fourier representation of solutions. We can solve the homogeneous
problem (328) by the Fourier method. To recall, If we apply the Fourier transform
with respect to the space variables, the initial value problem (328), (327) becomes
a Cauchy problem for an ordinary differential equation:
2u
b + ||2 u
b = 0, u
b(0, ) = f(), t u
b(0, ) = g(),
t

which can be solved explicitly:

u
b(t, ) = cos (t||) f() + sin (t||)

g()
||

Thus the fundamental solution W (t)h, defined above, takes the form,
Z
sin(t||)
W (t)h(x) =
eix
h()d.
||
n
R

(336)

(337)

By Duhamel principle, see (330), the general solution of the inhomogeneou equation
u = F can be expressed in the form,
Z t
u(t) = t W (t)f + W (t)g +
W (t s)F (s)ds.
(338)
0

let D = ()
Observe that,

1/2

be the operator whose symbol in Fourier space is given by ||.

Z

(DW (t))f (x)

(W (t)Df )(x) =

eix sin t||)f()d.

Rn

Since sin t|| and cos t|| are bounded the operators t W (t) and DW (t) map H s (Rn )
in itself. In particular, solutions u of (328), (327) preserves the (Sobolev) regularity

8. STRICHARTZ INEQUALITIES

191

of the initial data f and g. More precisely, If f, D1 g H s for some s R, then

u(t), D1 t u(t) H s uniformly for t R. We can also write,
kukL H x + kt ukL H x1
t

. kf kH + kgkH 1

which provides the easy part of estimate5 (331). Therefore to prove (331) it suffices
to prove,
kukLqt Lrx

. kf kH + kgkH 1

(339)

for and wave admissible pair (q, r).

We also remark that,
Z
t W (t)h(x)

eix cos(t||)h()d

=
Rn

and,
D1 W (t)h(x)

eix

=
Rn

cos(t||)
h()d
||

We can rewrite (336) as

where f =

1
2

u
b(t, ) = eit|| f+ () + eit|| f (),


f D1 g . It follows that u = u+ + u where
Z

u
=
ei(xt||) f ()d

Observe that to prove (339) it suffices to prove,

ku+ kLqt Lrx

. kf + kH

(340)

and a similar estimate for f .

0.37. Energy estimates. We will derive a simple L2 estimate for general
solutions of u = F by integration by parts. It all follows from the simple algebraic
identity:
1
t (|t u|2 + |u|2 ) + i (t ui u) = t u F
(341)
2
P
n
2
where |u|2 =
i=1 (i u) and i = xi . Integrating with respect to x, and
assuming that u and its derivatives vanish6 at infinity we derive,
Z
Z


t
(t u)2 + |u|2 dx = 2
t u F dx
Rn

Rn

Thus integrating in t,
kt u(t)k2L2 + ku(t)k2L2 kt u(0)k2L2 + ku(0)k2L2 + 2

Z tZ
t u F dxds
0

Rn

5Another derivation, based on energy identities, is given in the next subsection.

6This can easily be justified by the finite propagation speed property of solutions to the wave

equation

192

8. STRICHARTZ INEQUALITIES

which we rewrite, with |u|2 = |t u|2 + |u|2 ,

Z tZ
ku(t)k2L2 = ku(0)k2L2 + 2

t u F dxds.

(342)

Rn

In particular, applying H
older,
ku(t)k2L2

ku(0)k2L2

kt u(s)kL2 kF (s)kL2 ds

+2
0

from which we derive the inhomogeneous energy estimate,

Z T
sup ku(t)kL2 . ku(0)kL2 +
kF (s)kL2 ds
t[0,T ]

(343)

Now let Ds be the operator Ds = ()s/2 whose symbol in Fourier space is given
by ||s . Since Ds commutes with  we easily derive,
Z
s
2
s
2
kD u(t)kL2 = kD u(0)kL2 + 2
t Ds u Ds F dx
Rn

We can write, using Plancherel with respect to the x variables,

Z
Z
s
s
t D u D F dx =
t D2s u F dx,
Rn

Rn

Therefore, by H
older, in the slab DT = [0, T ] Rn ,
sup kDs u(t)k2L2 kDs u(0)k2L2 + 2kD2s t ukLqt Lrx (DT ) kF kLq0 Lr0 (D
t

t[0,T ]

T)

Choosing s = 1 we infer that,

sup kD1/2 u(t)k2L2 kD1/2 u(0)k2L2 + 2kD1 t ukLqt Lrx (DT ) kF kLq0 Lr0 (D
t

t[0,T ]

T)

We apply this energy estimate to solution of the inhomogeneous problem (0.30.1)

with zero initial conditions. We also assume that the dimensional condition 1q + nr =
n
1
n
1
2 = q 0 + r 0 2 is verified. That implies = 2 . We thus have,
sup kD1/2 u(t)k2L2 2kD1 t ukLqt Lrx (DT ) kF kLq0 Lr0 (D
t

t[0,T ]

T)

Assume for a moment that we can prove the estimate,

kD1 t ukLqt Lrx (DT ) . kF kLq0 Lr0 (D

T)

(344)

Then,
sup kD1/2 u(t)kL2 . kF kLq0 Lr0 (D
t

t[0,T ]

T)

which is equivalent to,

sup
t[0,T ]

ku(t)kH ) + kt ukH 1 )

kF kLq0 ([0,T ];Lr0 )

thus proving half of estimate (332). Therefore the inhomogeneous estimate (332)
reduces to proving,
kukLq ([0,T ];Lr ) + kD1 t ukLq ([0,T ];Lr )

kF kLq0 ([0,T ];Lr0 )

(345)

8. STRICHARTZ INEQUALITIES

193

0.38. Homogenous Case. In this section we prove estimate (340) and thus
complete the proof for the homogeneous Strichartz estimate of theorem 0.32. Using
the space-time Fourier transform, i.e. Fourier transform with respect to both t and
x,
u
e+ (, ) = ( ||)f+ (), u
e (, ) = ( + ||)f (),
(346)
These are the components of u
e living on the forward null cone C+ = { = ||} and
on the backward null cone C = { = ||}, respectively. Thus we can interpret
(340) from the point of view of a restriction theorem for the half light cones C+ or
C . We next show that it suffices to prove (340) for the case when f+ is included
in fixed dyadic piece. More precisely, dropping the label + it suffices to show that,
q
ku+
k kLt Lrx

. 2k kfk+ kL2

(347)

+
+
+
+
where u+ = k2Z u+
and Pk the standard LP projeck , uk = Pk u , fk = Pk f
tions with respect to the spatial variables x.

To show that (348) implies (340) is highly nontrivial7 as we need to rely on corollary
6.15 adapted to the mixed norms Lqt Lrx with both q and r larger than 2. Thus,
X
X
2
ku+ k2Lq Lr .
ku+
22k kfk+ k2L2 . kf + kH
k kLq Lr .
t

kZ

kZ

Finally we observe, using a simple scaling argument, that (348) follows from,
q
ku+
0 kLt Lrx

kf0+ kL2

(348)

We now define the truncated cone operator C to be the operator

Z
Cf (t, x) = eit|| eix ()f()d,

(349)

where is a cut-off function supported in 1.2 || 2, such as the one used in

the definition of the LP projections, see (214). The operator C can be viewed as
the adjoint of the restriction of the Fourier transform to a truncated cone,
F () = () F
e(||, )
[
C

Estimate (348) is an immediate consequence of the following theorem.

Theorem 0.39. Let (q, r), (q1 , r1 ), (q2 , r2 ) be admissable pairs of exponents. Then
we have the estimates
kCf kLqt Lrx . kf kL2 ,

(350)

and also
kCC F kLqt 1 Lrx1 . kF k

q0

r0

Lt 2 Lx2

(351)

Composing C with C we derive,

Z
CC F (t, x) ' ei[(ts)||+(xy)] |()|2 F (s, y)dsdyd,
7Without using corollary 6.15 we would only derive a weaker estimate with the Besov norm

B 2,1
replacing H norm on the right.

194

8. STRICHARTZ INEQUALITIES

which can be rewritten as the convolution

Z
CC F (t, ) = U (t s)F (s, )ds,

(352)

with the evolution operator

Z
U (t)f (x) =

ei(t||+x) |()|2 f()d.

(353)

(Observe that U is essentially the same operator as C !) By the T T principle, we

know that the estimate (350) is equivalent to the following estimate for CC ,
kCC F kLqt Lrx . kF kLq0 Lr0 .

(354)

which is also equivalent to the polarized form (351). Thus, to prove the theorem it
suffices to prove (354). As in the second proof of the restriction theorem presented
in the previous section to prove (354) we need to prove the following properties for
the evolution operators U (t).
Proposition 0.40. Let () be a fixed C0 (Rn ) function supported in 1/2 || 2
and,
Z
U (t)f (x) = ei(t||+x) ()f()d.
(355)
Then,
kU (t)f kL2
kU (t)f kL

. Ckf kL2

(356)
n1
2

. (1 + |t|)

kf kL1

(357)

from which, interpolating, for all 2 r ,

kU (t)f kLr

. (1 + |t|)

n1
2
2 (1 r )

kf kLr0

(358)

Moreover, if in addition, = is supported in a cube of size , then (357) can

be strengthened to
kU (t)f kL

(1 + |t|)

n1
2

kf kL1

(359)

Proof We prove directly the stronger version (359). We only need to check (??).
We write,
Z
U (t)f = Kt f,
Kt (x) = ei(x+t||) ()d
It suffices to show that,
|Kt (x)| .

1
(1 + |t| + |x|)

In the regions |x| < |t|/2 and |x| 2|t| we integrate by parts k times with respect to
P xj +t j
the operator L = i j |x+t ||2 j , such that L(ei(x+t||) ) = ei(x+t||) . We also
||

make use of the straightforward estimate, | ()| . || to derive, |Kt (x)| .

(1 + |t|)k nk or, choosing k = n1
2 ,
|Kt (x)| . (1 + |t|)

n1
2

n+1
2

8. STRICHARTZ INEQUALITIES

195

On the other hand, in the region |t| |x|, we write, with (||) vanishing on the
support of h ,
Z 1+2
Z
Kt (x) =
eit ()
eix h ()d()
||=

12

We now need to rely on the following estimate,

Z
n1
sup |
eix h()d()| . (1 + |x|) 2
1/22

(360)

||=

which follows easily from the decay of the Fourier transform of measures supported
on Sn1 discussed in the previous section, see lemma 8.8. Therefore, for |t| |x|,
|Kt (x)| . (1 + |x|)

n1
2

. (1 + |t|)

n1
2

as desired.
We are now ready to prove (354) by following the same argument as in the second
proof of the restriction theorem. Indeed, in view of (352) and (358) we derive,
Z +
kCC F kLrx (t) .
(1 + |t s|)(r) kF (s)kLrx0 ds
(361)

n1
2 (1

2
r ).

where (r) =

We are now precisely in the same situation as in the

second proof of the restriction theorem, see the argument following formula (300).
If 0 < (r) < 1 we can apply the Hardy-Littlewood-Sobolev inequality to obtain
kCC F kLqt Lrx . kF kLq0 Lr0 ,
t

when (r) + 1 + 1/q = 1/q , hence (r) = 2/q. This proves (348), and thus
theorem 0.39, in the case 0 < (r) = 2/q < 1. If q = 2 and (r) > 1 we have from
(361),
kCC F kL2t Lrx . kF kL2t Lrx0 ,
by an application of the standard Hausdorff-Young inequality.
Finally, if 2/q < 1 and (r) > 2/q the result follows from the case (r) = 2/q using
Sobolev inequalities. Due to the fact that one of the principal curvatures of the
light cone vanishes, the Strichartz estimates for the wave equation is not as strong
as it could be. Using the improved dispersive estimate (359) we can however derive
a stronger statement ,which is very useful in applications.
Proposition 0.41. Let 0 < < 1. Let f be an L2 function with Fourier transform
supported in a cube of size at a distance 1 from the origin. Let (q, r) be an
admissable pair of exponents for the Strichartz estimates. Then
1
1
(362)
kCf k q r . ( 2 r ) kf k 2 .
L

Lt Lx

The proof is based on the improved dispersive estimate (359). Interpolating it with
(356) we derive,
kU (t)f kLr

1 r (1 + |t|)

n1
2
2 (1 r )

The proof the continues exactly as above to derive,

2

kCC F kLqt Lrx . 1 r kF kLq0 Lr0 ,

t

kf kLr0

196

8. STRICHARTZ INEQUALITIES

1
1
and therefore, by the T T argument, kCf kLqt Lrx . ( 2 r ) kf kL2 , as desired. As a
straightforward corollary to the proposition we derive:

Theorem 0.42. Consider a general solution of u = 0 with data f, g supported, in

Fourier space, on a cube of size situated in a dyadic shell of size , with much
larger than , say 8. Then,


2
(363)
kukLqt Lrx . 1 r kf kH + kf kH 1
Proof The proof follows easily by a scaling argument from the proposition above.

Finally we state below another result, which follows easily from the decay estimate
(357).
Theorem 0.43. Let u be a free wave, i.e. solution of the homogeneous equation
u = 0, with initial data (f, g). Then,
 n+1

n1 X
n1
ku(t)kL . |t| 2
2 kf kL1 + 2 kg kL1
2Z

n1
2

|t|


kf kB n+1/2 + kgkB n1/2 .
1,1

1,1

n1

The uniform decay rate |t| 2 , for large t, plays a very important role in the study
of nonlinear perturbations of the standard wave equation.
0.44. Inhomogeneous Strichartz estimates. We have already reduced the
inhomogeneous Strichartz estimate (332) of theorem 0.32 to estimate (345). Proceeding as in the case of the homogeneous estimates we can now reduce (345) to
the case when the spatial Fourier transform of F is supported in the unit dyadic
ring 1/2 || 2. Moreover, decomposing u as before in the parts it suffices to
prove the estimates separately for u+ and u . Therefore we need to prove,
ku+ |Lq ([0,T ];Lr ) + kD1 t u+ kLq ([0,T ];Lr )

. kF kLq0 ([0,T ];Lr0 )

(364)

We have,
Z
u+ (t, )

U (t s)F (s, )ds

=
0

D1 t u+ (t, )

Z
=

t D1 U (t s)F (s, )ds

Since, in view of the dyadic restriction, t D1 U (t) U (t) it suffices to prove

the estimate for ku+ |Lq ([0,T ];Lr ) . Clearly, u+ differs from CC F in (352) only by
the restriction of the interval of integration to [0, t]. In view of this fact we write
u+ = (CC )R F . We are thus led to the following theorem, from which (364) and
thus (332).
Theorem 0.45. Let U (t) defined as in (355) and let
Z t

(CC )R F (t, ) =
U (t s)F (s, )ds
0

8. STRICHARTZ INEQUALITIES

197

Then, for all admissible pairs (q1 , r1 ), (q2 , r2 ),

k(CC )R F kLqt 1 Lrx1 ([0,T ]Rn )

. kF k

q0

r0

Lt 2 Lx2 ([0,T ]Rn )

(365)

Proof The proof is straightforward in the case (q1 , r1 ) = (q2 , r2 ) = (q, r). Indeed
in this case we can simply repeat the proof of estimate (354) and just take into
account the limits of integration. We have also treated the case when q1 = ,
r1 = 2, see the subsection on energy estimates. The other non-diagonal case cases
are a little more difficult and will be treated in the more general abstract setting
discuss later in this section. The proof we have given covers however the most
interesting case of estimate (332). We have thus given complete proofs for the first
two parts of theorem 0.32

0.46. Necessity of the admissibility conditions. To understand what is

the optimal range of exponents q and r we consider the analog of the Knapp counterexample in the context of the truncated cone operator C defined in (349).
For some small > 0, let
D = { Rn : |1 1| < 1/2, | 0 | < } ,
and consider f = D . We have
Cf (t, x) = ei(t+x1 )

ei[t(||1 )+(t+x1 )(1 1)+x ] d,

and observe that

| 0 |2
. 2 .
|| + 1
We can then choose a region of space-time R defined by
|| 1 =

|t| . 2 ,

|t + x1 | . 1,

|x0 | . 1 ,

such that, when (t, x) R and D, then the oscillatory factor inside the last
integral can be treated as a constant. Hence, |Cf (t, x)| & |D| for (t, x) R and we
have
kCf kLqt Lrx
|D|kR kLqt Lrx
n1
n1
2
&
2 q r .
1/2
kf kL2
|D|
In the limit 0, an estimate of the form (354) will necessarily imply that q and
r satisfy the condition


2
1 1
(n 1)

.
(366)
q
2 r
The other restriction on the range for q, i.e. q 2 is a consequence of the invariance of the operator CC under time translations. Indeed for translation invariant
operators we have the following general result due to Hormander, [?].
Proposition 0.47. Let T : Lp (Rn ) Lq (Rn ) be a (non trivial) linear operator
which commutes with translations, in the sense that (T f ) y = T (f y ), where
y (x) = x + y, for x, y Rn . If T is bounded from Lp to Lq then we necessarily
have q p.

198

8. STRICHARTZ INEQUALITIES

The proof is based on the following lemma.

Lemma 0.48. Let f Lp (Rn ), then
lim kf + f y kLp = 21/p kf kLp .

|y|

Proof For every R > 0 consider the decomposition f = gR + hR , where gR (x) =

f (x) if |x| < R and 0 if |x| R, and hR (x) = 0 if |x| < R and f (x) if |x| R.
Then
lim kgR kLp = kf kLp ,
lim khR kLp = 0.
R

For R = |y|/2 we have

f + f y = gR + gR y + hR + hR y .
The functions gR and gR y have disjoint supports, so that
kgR + gR y kpLp = kgR kpLp + kgR y kpLp = 2kgR kpLp ,
while
lim khR + hR y kLp lim 2khR kLp = 0,

|y|

|y|

hence
lim kf + f y kLp = lim 21/p kgR kLp = 21/p kf kLp .

|y|

|y|

Proof [Proof of Proposition 0.47] Let C > 0 be the optimal constant for the
estimate
kT f kLq Ckf kLp , f Lp .
Then by linearity and the translation invariance,
kT f + (T f ) y kLq Ckf + f y kLp .
When |y| , applying the lemma we obtain
21/q kT f kLq C21/p kf kLp ,
1

f Lp .

The optimality of C implies that 2 p q 1, hence q p.

The proposition generalizes easily to vector valued Lp spaces and if we consider
0
0
CC as an operator from Lq (R; Lxr ) to Lq (R; Lrx ), then we must have q q 0 ,
which is the condition q 2.
0.49. A general, abstract framework. It turns out that the method of
proving Strichartz estimates described above applies to many other equations, such
as Schr
odinger, KdV etc. It thus pays to have a general framework which applies
to all these cases.
Let (X, d) be a measure space and H a Hilbert space. Consider a family (U (t))tR
of operators U (t) : H L2 (X), which describes the evolution of some system with
data in H. We assume that this evolution satisfies the following two properties:

8. STRICHARTZ INEQUALITIES

199

for all t R and f H we have the energy estimate:

kU (t)f kL2 (X) . kf kH ;

(367)

for all t 6= s and g L1 (X) we have the dispersive inequality:

kU (t)U (s)gkL (X) . |t s|0 kgkL1 (X) ,

(368)

for some 0 > 0.

Interpolating between (367) and (368) we obtain the estimate
kU (t)U (s)gkLr (X) . |t s|(r) kgkLr0 (X) ,

(369)

for r 2, where


2
(r) = 0 1
.
r
Theorem 0.50. If the evolution operator U (t) satisfies (367) and (368), then the
estimates
kU (t)f kLqt LrX . kf kH ,
hold for all q, r 2 verifing:
2
= (r),
q

(370)

(q, r, 0 ) 6= (2, , 1).

(371)

Remark 0.51. This form of the Strichartz inequalities applies to linear dispersive
equations such as Schr
odinger.
Proof If we consider the operator T : H Lqt LrX defined by T f (t, x) = (U (t)f )(x)
0
0
then it is easy
to verify that the dual of T is the operator T : Lqt LrX H given
R
by T F = U (s)F (s, )ds. By the T T method, (370) is then equivalent to the
estimate
Z
k U (t)U (s)F (s)dskLqt LrX . kF kLq0 Lr0 .
(372)
t

By duality and symmetry considerations, this is in turn equivalent to

|B(F, G)| . kF kLq0 Lr0 kGkLq0 Lr0 ,
t

(373)

where B(F, G) is the bilinear form

ZZ
B(F, G) =
hU (t)F (t), U (s)G(s)idtds.

(374)

s<t

From the bilinear version of (369) we have that

ZZ
kF (t)kLr0 kG(s)kLr0
dsdt.
|B(F, G)| .
|t s|(r)

(375)

If (r) < 1, we can apply the Hardy-Littlewood-Sobolev inequality and obtain

(373). This concludes the proof for the cases q = 2/(r) > 2.
The endpoint case, corresponding to (r) = 2/q = 1, is allowed when r < . Its
proof will be described in the next section.

200

8. STRICHARTZ INEQUALITIES

Remark 0.52. If we strengthen the dispersive condition (368) to

kU (t)U (s)gkL (X) . (1 + |t s|)0 kgkL1 (X) ,

(376)

then (375) can be improved to

ZZ
|B(F, G)| .

kF (t)kLr0 kG(s)kLr0
dsdt.
(1 + |t s|)(r)

(377)

Now we can obtain (373) from Youngs inequality when 2/q = 1/p and (1 +
|t|)(r) Lp (R), i.e. (r)p > 1. Hence, (376) allows us to extend the Strichartz
estimates (370) in Theorem 0.50 to the range
2
(r), (q, r, 0 ) 6= (2, , 1).
q
This case applies to the linear wave equations.

(378)

Remark 0.53. We observe that there is a natural scaling associated to the objects
in this abstract formulation. More precisely, the estimates (370) in Theorem 0.50
are invariant under the change of scale defined by
U (t) U (t/),

U (s) U (s/),

d 0 d,

hf, giH 0 hf, giH .

(379)

We can also consider the endpoint case.

20
, 0 > 1.
0 1
This, in fact, is more difficult than the previous non-endpoint case, and requires
a two-parameter estimate which is better than the one-parameter family given by
the interpolation (369). This proof is presented in the previously mentioned paper
by Keel and Tao, Endpoint Strichartz Estimates. We omit it here.
q = 2,

r=

0.54. Inhomogeneous estimates. Saying that an operator T maps the Hilbert

0
0
space H into Lqt LrX , is equivalent to saying that its dual T maps Lqt LrX into H,
0
0
and is also equivalent to saying that the T T operator maps Lqt LrX into Lqt LrX . If
the pair (q, r) is allowed to vary in a set E of admissable exponents, we can view
T T as a composition of two operators associated with different pairs of exponents.
It follows that T T actually satisfies a larger set of mapping properties, since it
0
r0
maps Lqt LX
into Lqt LrX , for any couple of pairs (q, r), (
q , r) E.
The operator T f (t) = U (t)f defined in the previous subsection can be viewed as the
solution of some homogenous, translation invariant, linear evolution equation. The
solution of the corresponding inhomogenoues problem, using Duhamels principle,
would be represented by the retarded operator
Z
RF (t) =
U (t)U (s)F (s)ds.
s<t

Observe that operator R looks very similar to the T T operator, which is given by
Z

T T F (t) = U (t)U (s)F (s)ds.

8. STRICHARTZ INEQUALITIES

201

The restriction s < t in the definition of R, however, destroys the composition

structure of T T . Fortunately, all the mapping properties of T T , which we have
derived above, can be transfered to R.
0

Theorem 0.55. The operator R maps Lqt LrX

into Lqt LrX , for any couple of pairs
(q, r), (
q , r) for which the Strichartz estimate 370 holds.

Proof First of all observe that in the proof of theorem 0.50 we have actually
proved the diagonal case (q, r)
q , r). Indeed, the bilinear form defined in (374)
RR= (
can be written as B(F, G) =
R(F ) Gdxdt and (373) is the dual formulation of
the mapping property for R.
The non diagonal cases with 1q + 1q < 1 follow from the mapping properties of T T
by using a general argument about integral operators due to Christ and Kiselev
(see [] and []) which we summarize in Proposition 0.57 below.
It remains to consider the cases with q = q = 2 and r 6= r, under the assumption
that the evolution U (t) satisfies the stronger dispersive inequality (376) with 0 > 1.
Since, we have already proved the case r = r, by interpolation it is enough to
0
, r = , and show that
consider the extreme case: r = r = 2
0 1
|B(F, G)| . kF k

r0

L2t LX

kGkL2t L1X .

This estimate follows by decomposing B(F, G) into dyadic pieces, B =

where
ZZ
B (F, G) =
hU (t)F (t), U (s)G(s)idtds,

2Z

B ,

(380)

/2|ts|2

The desired conclusion follows immediately from the lemma below.

Lemma 0.56. Let B (F, G) be the bilinear form defined in (380). Then, there exists
an > 0 such that


|B (F, G)| . min , 1 kF k 2 r0 kGkL2t L1X .
Lt LX

Proof We may assume that F and G are supported on disjoint time intervals of
length O() separated by a distance O(). Then B (F, G) = hT F, T GiH . We
use the energy estimate to bound kT F kH and the Strichartz estimate with q = 2
and r = to bound kT GkH , so that
|B (F, G)| . kF kL1t L2X kGkL2t L1X .
We then apply Holder inequality and use the assumption on the support of F to
obtain
|B (F, G)| . 1/2 kF kL2t L2X kGkL2t L1X .
We can also write B (F, G) =
dispersive inequality,

RRR

F (t) U (t)U (s)G(s)dxdsdt and make use of the

|B (F, G)| . (1 + )0 kF kL1t L1X kGkL1t L1X .

202

8. STRICHARTZ INEQUALITIES

Again, we apply Holder inequality and use the assumption on the support of F and
G to obtain

kF kL2t L1X kGkL2t L1X .

|B (F, G)| .
(1 + )0
Hence, B is bounded on L2t L2X L2t L1X with constant 1/2 and on L2t L1X L2t L1X

p
with constant (1+)
spaces we obtain that B
0 . By standard interpolation of L
r0

is bounded on L2t LX L2t L1X with constant C , where

1

1
1

,
= +
C = /2
,
(1 + )0
r0
2
1

r =

20
.
0 1

Simplyfing the expression we find that

0 +1



20
. min , 1 ,
C =
1+
with

= min

0 + 1
0 + 1
,1
20
20


=

0 1
1
=
> 0.
20
r

0.56.1. Integral operators with restricted kernel. In this subsection we give a self
contained exposition of the results of Christ-Kisselev mentioned above. Consider
an integral operator with a measurable kernel K(s, t),
Z
T f (t) =
K(s, t)f (s)ds,
R

and its restricted version associated with the kernel K(s, t)(s < t),
Z
Rf (t) =
K(s, t)f (s)ds.
s<t

If T maps Lp into Lq and 1 p < q then we have that R also maps Lp into
Lq . An equivalent formulation of this fact is given in the following proposition.
Proposition 0.57. Let K(s, t) be a measurable function on R R. Let B(f, g) be
the bilinear form with kernel K,
ZZ
B(f, g) =
K(s, t)f (s)g(t)dsdt,
e g) the bilinear form with kernel restricted to the region s < t,
and B(f,
ZZ
e g) =
B(f,
K(s, t)f (s)g(t)dsdt.
s<t

Let p, q 1, with the condition

1 1
+ > 1.
p q
If B is bounded on Lp Lq ,
|B(f, g)| . kf kLp kgkLq ,

(381)

8. STRICHARTZ INEQUALITIES

203

e is also bounded on Lp Lq ,
then B


e

B(f, g) . kf kLp kgkLq .
Remark 0.58. There are cases for which equality in condition (381) is not allowed.
Consider for the example the case of the Hilbert transform, which corresponds to
1
the kernel K(s, t) = st
, with p = q = 2.
Proof Let f Lp and g Lq with kf kLp = kgkLq = 1.
R
Define F (t) = s<t |f (s)|p ds. F is a continuous non-decreasing function which
maps [, +] onto [0, 1]. In particular, the inverse image of an interval of the
type I = [a, b] [0, 1] will be an interval of the same type, F 1 (I) = [A, B], with
RB
F (A) = a, F (B) = b, and A |f (s)|p ds = F (B) F (A) = b a. Hence,
kf kLp (F 1 (I)) = |I|1/p .

(382)



Consider now a Whitney decomposition of the set = (x, y) R2 : x < y into
disjoint dyadic squares, as in Lemma 6.7, = Q Q, where each square Q = I J
has the property
dist (I, J) |I| = |J| = ,

(383)

for some dyadic value of . If we look only at those squares needed to cover the
triangle [0, 1]2 , then 1/2.
Observe that s < t implies that either F (s) < F (t) or f 0 almost everywhere on
the interval [s, t]. Hence, we can write
ZZ
X
e g) =
B(f,
K(s, t)f (s)g(t)dsdt =
B(F 1 (I) f, F 1 (J) g).
F (s)<F (t)

Q
p

Using the boundedness of B on L L we obtain


 X
e

kf kLp (F 1 (I)) kgkLq (F 1 (J)) .
B(f, g) .
Q

Now we use (382), (383) and the fact that, for each given dyadic interval J, the
number of intervals I for which I J is one of the squares in the decomposition of
is bounded by a universal constant. Hence,


X 1 X
e

kgkLq (F 1 (J)) .
p
B(f, g) .
1/2

|J|=

Next, we apply H
olders inequality to the summation over the dyadic intervals J of
length and since there are 1 of them in [0, 1] we have


X 1 1
X 1 1
e

p q0 kgkLq =
p + q 1 . 1.
B(f, g) .
1/2

1/2

CHAPTER 9

Bilinear Estimates

1. Bilinear proofs of some Strichartz estimates

Consider the homogeneous wave equation u = 0 in R1+3 . The Strichartz estimate
(331) with q = r = 4 and = 1/2. Takes the form,
kukL4 (R1+3 )
+

Writing u = u + u

. kf kH 1/2 + kgkH 1/2

it suffices to prove,
ku+ kL4 (R1+3 )

kf + kH 1/2

(384)

where
Z
u+ (t, x) =

eix+t|| f()d

Clearly,
ku+ k2L4 (R1+3 )

f
+u
+k 2
ku+ u+ kL2 = kuf
L

Now, recalling (346), and dropping the index +,

Z Z
u
u
(, ) =
( | |)f( )( ||)f()dd
Z
=
(( || | |)f()f( )d
Clearly, (384) follows from the following:
Theorem 1.1. The bilinear operator,
Z
F ( ) G()
B(F, G) =
( || | |)
d.
| |1/2 ||1/2
verifies the estimate,
kB(F, G)kL2 (R1+3 ) . kF kL2 (R3 ) kGkL2 (R1+3 )
Proof By Cauchy-Schwartz,
Z
. J(, ) ( || | |)|F ( )|2 |G()|2 d
Z
1
1
d
J(, ) =
( || | |)
| | ||

|B(F, G)(, )|2

205

(385)

206

9. BILINEAR ESTIMATES

It suffices to show that J is uniformly bounded. Indeed, if that is the case,

Z Z
kB(F, G)kL2 (R1+3 ) . sup J(, )
( || | |)F ( )|2 |G()|2 dd d
,

. sup J(, )kF k2L2 kGk2L2

,

Therefore the bilinear estimate is an immediate consequence of the uniform boundedness of J. This follows from the following more general lemma below.
Lemma 1.2. Let F be an arbitrary function of two variables and JF the integral
Z
JF (, ) =
( || | |)F (||, | |)
Rn

Then,
JF (, )

JF+ (, )

( 2 ||2 )

n3
2

F
1

( 2 ||2 )

n3
2

F
1

n3
+ s|| + s||
2
,
( x2 ||2 )(1 |x|2 ) 2 dx,
2
2
(386)

n3
+ s|| + s||
2
,
( x2 ||2 )(1 |x|2 ) 2 dx
2
2
(387)

Proof : Observe that in the case = the measure ( ||| |) is supported

on the ellipsoid of revolution with foci at 0 and , E(, ) = { Rn : || + | | = } ,.
In this case || . In the = + the measure ( ||+| |) is supported in the
hyperboloid of revolution with foci at 0 and , H(, ) = { Rn : || | | = } ,
which is an unbounded hypersurface with infinite volume. In this case ||2 2 .
In the sense of distributions, we have the identity
( || | |)

( ||)2 | |2

2( ||)

2( ||) ( ||)2 | |2

2( ||) 2 ||2 2 + 2 cos )

2( ||) 2 ||2 2 + 2a||)

with a the cosine of the angle between and . Thus, for fixed and we must
have, on the support of the measure,
a

2 ||2 2
2||

(388)

Observe that in the ellipsoidal case a can take any values in the interval [1, 1] and
2 ||2
||
thus, since = 2(
+||
a||) , we have
2
2 . On the other hand, in the
2
2
hyperboloidal case when || > , we must also have the restriction,

a.
||
and thus, =

2 +||2
2( +a||)

+|
2 .

2.

IMPROVED BILINEAR STRICHARTZ

207

n3

Thus, since d = n1 ddS = (1 a2 ) 2 n1 ddS0 ,

Z +||
2

2 ||2 2
 n3
1

2
F (, )( )n2 1
d
JF =
|| ||
2||
2
n3 Z +||
2
 + ||


( 2 ||2 ) 2
||
 n3
2
=
F (, )( )

n2
||
||
2
2
2
At last we perform the change of variables x = 2
|| to derive the desired formula
(386). The proof for (387) follows in the same manner.

2. Improved Bilinear Strichartz

Consider two solutions of the homogeneous wave equations, u = v = 0. For
simplicity, and without loss of generality, we assume that u, v verify the reduced
initial data at t = 0,
u(0, x) = f (x), v(0, x) = g(x), t u(0, x) = t v(0, x) = 0.
We consider estimates of the form,
kDb (uv)kLq/2 Lr/2
t

kf kH a kgkH a

with (q, r) an acceptable pair. By dimensional analysis and recalling the exponent
= n(( 21 1r )) 1q in (331), we must have,
2a

1

1 1
= b + 2
b + 2 n( )
2 r
q

(389)

We decompose the product u v by the trichotomy formula,

X
X
X
uv =
u v +
v u +
P (u v )
<

<

(u v)LH + (u v)HL + (u v)HH

Here , 2 , u = P u and P the usual LP projections. Now,

X
X
kDb (uv)LH kLq/2 Lr/2
b ku v kLq/2 Lr/2
b ku kLqt Lrx kv kLqt Lrx
Z

in view of the Strichartz estimates of the previous section

ku kLqt Lrx . (a) kf kH a = b/2 kf kH a
kv kLqt Lrx . (a) kgk kH a = b/2 kg kH
and therefore, for b > 0,
kDb (uv)LH kLq/2 Lr/2
t

X
b
kf kH a kg kH a

. kf kH a kgkH a

208

9. BILINEAR ESTIMATES

By symmetry,
kDb (uv)LH kLq/2 Lr/2
t

kf kH a kgkH a

It thus only remains to estimate the high-high term k(u v)HH kLq/2 Lr/2 . This
x
t
requires a more subtle argument based on theorem ??. We write,
X
kDb (u v)HH kLq/2 Lr/2 .
b kP (u v )kLq/2 Lr/2
x

If we use the standard Strichartz estimate, i.e.,

kP (u v )kLq/2 Lr/2
x

. ku kLqt Lrx kv kLqt Lrx 2(a) kf kH a kgkH a

= b kf kH a kgkH a

(390)

we would derive,
kDb (u v)HH kLq/2 Lr/2 .
t

b b kf kH a kgkH a

which diverges. We need to replace (390) by a stronger estimate which takes into
account the presence of P in front of u v . To achieve this, we need first to exploit
some orthogonality properties. We decompose the
, g , in Fourier space,
P the data fP
into pieces supported on cubes of size , f = Q fQ , g = Q gQ and denote by
uQ , vQ the corresponding solutions.
Clearly P
the decomposition commutes with the
P
wave operator . Thus, u Q uQ , v Q vQ and
X


P (u v )
P uQ1 vQ2
Q1 ,Q2

Observe that P (uQ1 uQ2 ) 6= 0 only if Q1 + Q2 intersects the region of frequencies

of size where P is supported. For each cube Q1 , of size , there are only a finite
number (which depends only on n) of cubes Q2 for which this happens. Morally,
by enlarging the cubes if necessary we may assume that Q2 = Q1 and thus,
X
P (u v )
uQ vQ .
Q

Hence,
kP (u v )kLq/2 Lr/2 .
t

kuQ vQ kLqt Lrx .

kuQ kLqt Lrx kvQ kLqt Lrx .

We are now in a position to apply theorem 0.42. Thus,

12 r1
kuQ kLqt Lrx .
kfQ kH

and similarly for vQ . Hence,

1 r2 X
kP (u v )kLq/2 Lr/2 .
kfQ kH kgQ kH
x
t

.
.
.

1 r2
kf kH kg kH

2

1 r 22a

kf kH a kg kH a

2


1 r b
kf kH a kg kH a

3. BILINEAR ESTIMATES FOR NULL FORMS.

209

and, consequently,
kDb (u v)HH kLq/2 Lr/2
t

X
1 2 b
r
kf kH a kg kH a

<

. kf kH a kgkH a
provided that b < 1
[?].

2
r.

We have just proved the following bilinear estimate, see

Theorem 2.1. The following estimate 1 holds for solutions u = v = 0, any

admissible pair (q, r) and any 0 b < 1 2r ,
kDb (u v)Lq/2 Lr/2
t

. ku[0]kH a kv[0]kH a

(391)

provided that the dimensional condition,

b
a = + ,
2

1 1
1
= n( )
2 r
q

(392)

3. Bilinear estimates for null forms.

In this subsection we discuss the simplest bilinear estimates for null quadratic forms,
see [?], [?], [?] and [?].
Definition 3.1. Let u, v be two smooth solutions of  =P
v = 0 on Rn+1 . The
n
standard null quadratic forms are Q0 (u, v) = t u t v + i=1 i ui v, as well as
Qij (u, v) = i uj v i vj u, and Q0i (u, v) = i ut v i vt u for i, j = 1, . . . , n.
Theorem 3.2. For any null form Q and any solutions to  = v = 0 on Rn+1 ,
n 2, we have,
kQ(u, v)kL2 (Rn+1 )

ku[0]kH 1 (Rn ) kv[0]k

n+1
2

(Rn )

(393)

Remark 3.3. Without loss of generality, it suffices to consider the reduced initial
value problems
u(0, x) = f (x), v(0, x) = g(x), t u(0, x) = t v(0, x) = 0

(394)

In what follows we show how to deduce the estimate (3.2) from a more general form
of bilinear estimates presented in the next section.

Definition 3.4. Let D , D+

and D
be the operators in Rn+1 defined by the
multipliers with symbols, respectively



|| , (| | + ||) , | | || .

Observe that we can write, for any smooth functions u, v,

2Q0 (u, v) = (uv) uv uu
1Here ku[0]k
a = ku(0)kH
a + kt u(0)kH
a
H

210

9. BILINEAR ESTIMATES

Thus, if u = v = 0, using Plancherel,

kQ0 (u, v)kL2 (Rn+1 )

1
1
k(uv)kL2 (Rn+1 ) = (2)n k( 2 ||2 )f
uvkL2 (Rn+1 )
2
2
. kD+ D (uv)kL2 (Rn+1 )

Therefore,
kQ0 (u, v)kL2 (Rn+1 )

kD+ D (uv)kL2 (Rn+1 )

(395)

Thus, in the case of the null form Q0 , theorem 3.2 reduces to,
kD+ D (uv)kL2 (Rn+1 ) . ku[0]kH 1 (Rn ) kv[0]k

n+1
2

(Rn )

(396)

which is a special case of theorem ??.

Below we show that similar estimates hold true for the other null forms, Qij , Q0i .
Remark 3.5. Given a solution u of u = 0 with initial data u(0, x) = f (x),
t u(0, x) = 0 we denote by u0 the solution of the same equation with data u0 (0, x) =
f 0 (x), t u0 (0, x) = 0 where f 0 = F 1 (|f|). Observe, of course, that kf 0 kH a =
kkf kH a and thus, from the point of view of the L2 type estimates we are considering
u and u0 are indistinguishable.
Proposition 3.6. Let u, v be smooth solutions of the homogeneous wave equation
with initial . The following estimates hold true:
kQij (u, v)kL2 (Rn+1 )
kQ0i (u, v)kL2 (Rn+1 )

1/2

. kD1/2 D (D1/2 u0 D1/2 v 0 )kL2 (Rn+1 )

.

1/2 1/2
kD+ D (D1/2 u0

1/2 0

v )kL2 (Rn+1 )

(397)
(398)

Proof : We first decompose, as before, u = u+ + u , v = v + + v We write, in

Fourier variables,
Z
Qij^
(u+ , v )(, ) =
qij (, )( || | |)f()
g ( )d
where qij (, ) = i ( )j j ( )i = ( )ij We now rely on the following
simple lemma.
Lemma 3.7. The following inequalities hold true,
| | . ||1/2 ||1/2 | + |1/2 (|| + || | + |)1/2


| | . ||1/2 ||1/2 | + |1/2 (| + | || ||)1/2
We have indeed,
4| |2

4(|||| )(|||| + )

((|| + || | + |)((|| + || + | + |)




(| + | || ||)(| + | + || ||)

from which the lemma immediately follows.

(399)
(400)

3. BILINEAR ESTIMATES FOR NULL FORMS.

211

Therefore, in both cases, |Qij^

(u+ , v )(, )| can be bounded by the expression,
Z
|qij (, )| ( || | |)|f()| |
g ( )|d
Z
1/2

g ()|d
. | | || ||1/2 ( || | |) ||1/2 | |1/2 |f()| |
1/2

= D1/2 D (D1/2 u0 D1/2 v 0 )

as desired.

According to proposition 3.6, theorem 3.2 reduces, for Q = Qij , resp. Q = Q0i , to
the statements,
1/2

kD1/2 D (u v)kL2 (Rn+1 )

. ku[0]kH 1/2 ku[0]kH n/2

1/2 1/2
kD+ D (u

. ku[0]kH 1/2 ku[0]kH n/2

v)kL2 (Rn+1 )

which are particular cases of theorem ??.

APPENDIX A

Basic Concepts in Riemannian and Lorentzian

Geometry

In what follows we give a short overview of the basic notions in Riemannian and
Lorentzian geometry. For a more detailed review we refer to [Pet], for Riemannian
geometry, and [Car], [Ha-E], [Wa] for Lorentzian geometry.

1. Introduction
A pseudo-riemannian manifold 1, or simply a spacetime, consist of a pair (M, g)
where M is an orientable p + q = n-dimensional manifold and g is a pseudoriemannian metric defined on it, that is a smooth, a non degenerate, 2-covariant
symmetric tensor field of signature (p, q). This means that at each point p M one
can choose a basis of p + q vectors, {e() }, belonging to the tangent space T Mp ,
such that
g(e() , e() )

(401)

for all , = 0, 1, ..., n , where is the diagonal matrix with 1 in the first p entries
and +1 in the last q entries. If X is an arbitrary vector at p expressed, in terms of
the basis {e() }, as X = X e() , we have
g(X, X) = (X 1 )2 . . . (X p )2 + (X p+1 )2 + .... + (X p+q )2

(402)

The case when p = 0 and q = n corresponds to Riemannian manifolds of dimension

n. The other case of interest for us is p = 1, q = n which corresponds to a Lorentzian
manifolds of dimension n + 1. The primary example of Riemannian manifold is the
Euclidean space Rn . Any other Riemannian manifold looks, locally, like Rn . Similarly, the primary example of a Lorentzian manifold is the Minkowski spacetime,
the spacetime of Special Relativity. It plays the same role, in Lorentzian geometry,
as the Euclidean space in Riemannian geometry. In this case the manifold M is
diffeomorphic to Rn+1 and there exists globally defined systems of coordinates, x ,
relative to which the metric takes the diagonal form 1, 1, ..., 1. All such systems
are related through Lorentz transformations and are called inertial. We shall denote
the Minkowski spacetime of dimension n + 1 by (Rn+1, m).
1We assume that our reader is already familiar with the basics concepts of differential geometry such as manifolds, tensor fields, covariant, Lie and exterior differentiation. For a short
introduction to these concepts see Chapter 2 of Hawking and Ellis, The large scale structure of
space-time, [?]
213

214

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

Relative to a given coordinate system x , the components of a pseudo-riemannian

metric take the form
g = g( , )

where = x are the associated coordinate vectorfields. We denote by g the

components of the inverse metric g 1 relative to the same coordinates x, and by |g|
the determinant of thepmatrix g
p. The volume element dvM of M isR expressed, in
local coordinates, by |g|dx = |g|dx1 . . . dxn . Thus the integral M f dvM of a
p
R
function f , supported in coordinate chart U M is defined by U f (x) |g(x)|dx.
The integral on M of an arbitrary function f is defined by making a partition of
unity subordinated to a covering of M by coordinate charts. One can easily check
that the definition is independent of the particular system of local coordinates.
In view of (402) we see that a Lorentzian metric divides the vectors in the tangent
space T Mp at each p, into timelike, null or spacelike according to whether the
quadratic form
(X, X) = g X X

(403)

is, respectively, negative, zero or positive. The set of null vectors Np forms a double
cone, called the null cone of the corresponding point p. The set of timelike vectors
Ip forms the interior of this cone. The vectors in the union of Ip and Np are called
causal. The set Sp of spacelike vectors is the complement of Ip Np .
A frame e() verifying (401) is said to be orthonormal. In the case of Lorentzian
manifolds it makes sense to consider, in addition to orthonormal frames, null frames.
These are collections of vectorfields2 e consisting of two null vectors en+1 , en and
orthonormal spacelike vectors (ea )a=1,... ,n1 which verify,

g(en , en ) = g(en+1 , en+1 ) = 0 , g(en , en+1 ) = 2

g(en , ea ) = g(en+1 , ea ) = 0 , g(ea , eb ) = ab
One-forms A = A dx are sections of the cotangent bundle of M. We denote
by A(X) the natural pairing between A and a vectorfield X. We can raise the
indices of A by A = g A . A0 = A defines a vectorfield on M and we have,
A(X) = g(A0 , X). Covariant tensors A of order k are k-multilinear forms on T M.
Notation: We will use the following notational conventions: We shall use boldface characters to denote important tensors such as the metric g, and the Riemann
curvature tensor R. Their components relative to arbitrary frames will also be denoted by boldface characters. Thus, given a frame {e() } we write g = g(e , e ),
R = R(e , e , e , e ) and, for an arbitrary tensor T ,
T... T (e , e , e , e , ...)
In the case of a Riemannian manifold we often use latin letters i, j, k, l, . . . to
denote indices of coordinates x1 , x2 , . . . , xn or tensors. For a Lorentzian manifold
2We write e instead of e

() to simplify the notation, whenever there can be no confusion.

2. VARIOUS NOTIONS OF DIFFERENTIATION

215

we use greek letters , , , . . . to denote indices 0, 1, . . . , n. For a general pseudoriemannian metric of signature s we shall also use greek indices.
We will review the following topics below:
1.)

Lie brackets of vectorfields. Frobenius theorem

2.)

Lie derivative of a tensorfield

3.)

Multilinear forms and exterior differentiation

4.)

Connections and covariant derivatives

5.)

Pseudo-riemannian metrics. Riemannian and Lorentzian geometry.

6.)

Levi-Civita connection associated to a pseudo-riemannian metric.

7.)

Parallel transport, geodesics, exponential map, completeness

8.)
Curvature tensor of a pseudo-riemannian manifold. Symmetries. First and
second Bianchi identities.
9.)
Isometries and conformal isometries. Killing and conformal Killing vectorfields.

2. Various notions of differentiation

We recall here the three fundamental operators of the differential geometry on a
Riemann or Lorentz manifold: the exterior derivative, the Lie derivative, and the
Levi-Civita connection with its associated covariant derivative.
2.1. The exterior derivative. Given a scalar function f its differential df is
the 1-form defined by
df (X) = X(f )
for any vector field X. This definition can be extended for all differential forms on
M, i.e. k (M), in the following way:
i) d is a linear operator defined from the space of all k-forms to that of k + 1-forms
on M. Thus for all k-forms A,B and real numbers ,
d(A + B) = dA + dB
ii)

For any k-form A and arbitrary form B

d(A B) = dA B + (1)k A dB

iii)

For any form A,

d2 A = 0 .

216

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

We recall that, if is a smooth map defined from M to another manifold M0 , then

d( A) = (dA) .
Finally if A is a one form and X, Y arbitrary vector fields, we have the equation


1
dA(X, Y ) =
X(A(Y )) Y (A(X)) A([X, Y ])
2
where [X, Y ] is the commutator X(Y ) Y (X). This can be easily generalised to
arbitrary k forms, see Vol.I, Chapter 7, Theorem 13 in [Sp].
2.2. Lie derivative. Consider an arbitrary vector field X. In local coordinates x , the flow of X is given by the system of differential equations
dx
= X (x1 (t), ..., xp+q (t)) .
dt
The corresponding curves, x (t), are the integral curves of X. For each point p M
there exists an open neighborhood U, a small  > 0 and a family of diffeomorphism
t : U M, |t| , obtained by taking each point in U to a parameter distance
t, along the integral curves of X. We use these diffeomorphisms to construct, for
any given tensor T at p, the family of tensors (t ) T at t (p) .
The Lie derivative LX T of a tensor field T , with respect to X, is:
1
LX T |p lim (T |p (t ) T |p ) .
t0 t
It has the following properties:
i)

LX linearly maps (p, q)-tensor fields into tensor fields of the same type.

ii) LX commutes with contractions.

iii)

For any tensor fields S, T ,

LX (S T ) = LX S T + S LX T .

If X is a vector field we easily check that

LX Y = [X, Y ]




d
((t ) Y )i 
and expressing (t ) Y )i  =
by writing (LX Y )i = dt
t=0

xi (t (q)) j 
xj (q) Y 

,
q

where q = t (p). (See [?], Hawking and Ellis, section 2.4 for details.)
If A is a k-form we have, as a consequence of the commutation formula of the
exterior derivative with the pull-back ,
d(LX A) = LX (dA) .
For a given k-covariant tensorfield T we have,
LX T (Y1 , . . . , Yk ) = XT (Y1 , . . . , Yk )

k
X
i=1

T (Y1 , . . . , LX Yi , . . . , Yk )

2. VARIOUS NOTIONS OF DIFFERENTIATION

217

We remark that the Lie bracket of two coordinate vector fields vanishes,



,
= 0.
x x
The converse is also true, namely, [Sp], Vol.I, Chapter 5,
Proposition 2.3. If X(0) , ...., X(k) are linearly independent vector fields in a neighbourhood of a point p and the Lie bracket of any two of them is zero then there exists
a coordinate system x , around p such that X() = x for each = 0, ..., k .
Proof When k = 2 the proposition is equivalent to the fact that if if two linearly
independent vector fields X, Y commute , i.e. [X, Y ] = 0, then their flows also
commute in the following sense. Starting with any point p, if we go a parameter
distance t along the integral curve of X initiating at p then a parameter distance
s along the integral curve of Y , we arrive at the same point as if one first distance
s along the integral curve of Y and then distance t along the integral curve of X.
To prove the assertion we first consider coordinates t, x relative to which X = t .
The coefficient of Y , relative to the same coordinates, are t independent. We then
perform another change of coordinates x = x(y) such that Y = A(y)t + y1 .
Finally we can make another change of coordinates of the form s = a(y) + t, with
a the primitive of A relative to y 1 . In the new coordinates s, y we must have
X = s , Y = y1 as desired. The general case can be proved in the same manner.

The above proposition is the main step in the proof of Frobenius Theorem. To state
the theorem we recall the definition of a k-distribution in M. This is an arbitrary
smooth assignment of a k-dimensional plane p at every point in a domain U of
M. The distribution is said to be involute if, for any vector fields X, Y on U with
X|p , Y |p p , for any p U, we have [X, Y ]|p p . This is clearly the case for
integrable distributions3. Indeed if X|p , Y |p T Np for all p N , then X, Y are
tangent to N and so is also their commutator [X, Y ]. The Frobenius Theorem
establishes that the converse is also true4, that is being in involution is also a
sufficient condition for the distribution to be integrable,
Theorem 2.4. (Frobenius Theorem) A necessary and sufficient condition for a
distribution (p )pU to be integrable is that it is involute.
Proof If is involute we can find linearly independent vector fields X1 , . . . , Xk
spanning at every point and such that the Lie bracket of any two of them is a
linear combination of X1 , . . . , Xk . One can then redefine X10 as a linear combination
of the other vector fields such that it commutes with X2 , . . . , Xk . One can then
proceed by induction to redefine the other vectorfields such that they all commute.
The proof then follows from the previous proposition.
3Recall that a distribution on U is said to be integrable if through every point p U there
passes a unique submanifold N , of dimension k, such that p = T Np .
4For a proof see Spivak, [Sp], Vol.I, Chapter 6.

218

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

2.5. The connection and the covariant derivative. A connection D is a

rule which assigns to each vectorfield X a differential operator DX . This operator
maps vector fields Y into vector fields DX Y in such a way that, with , R and
f, g scalar functions on M,
a) Df X+gY Z = f DX Z + gDY Z
b) DX (Y + Z) = DX Y + DX Z

(404)

c) DX f Y = X(f )Y + f DX Y
Therefore, at a point p,
DY Y; () e()

(405)

where the () are the one-forms of the dual basis respect to the orthonormal frame
e() . Observe that Y; = () (De() Y ). On the other side, from c),
D(f Y ) = df Y + f DY
so that
DY = D(Y e() ) = dY e() + Y De()
and finally, using df () = e() (f )() (),


DY = e() (Y ) + Y () (De() e() ) () e()

(406)

Therefore

Y; = e() (Y ) +
Y

and the connection is, therefore, determined by its connection coefficients,

()

(De() e() )
=

(407)

which, in a coordinate basis, are the usual Christoffel symbols and have the expression

= dx (D )
x x
Finally



DX Y = X(Y ) +
e()
(408)
X Y
In the particular case of a coordinate frame we have



DX Y = X
+ X Y
x
x
A connection is said to be a Levi-civita connection if Dg = 0. That is, for any
three vector fields X, Y, Z,
Z(g(X, Y )) = g(DZ X, Y ) + g(X, DZ Y )

(409)

A very simple and basic result of differential geometry asserts that for any given
(pseudo-riemannian) metric there exists a unique affine connection associated to it.

2. VARIOUS NOTIONS OF DIFFERENTIATION

219

Proposition 2.6. There exists a unique connection on M, called the Levi-Civita

connection, which satisfies D g = 0. The connection is torsion free, that is,
DX Y DY X = [X, Y ] .
Moreover, relative to a system of coordinates, x , the Christoffel symbol of the
connection is given by the standard formula
=
Exercise:

1
g ( g + g g ) .
2

Prove the proposition yourself, without looking in a book.

So far we have only defined the covariant derivative of a a vector field. We can
easily extend the definition to one forms A = A dxa by the requirement that,
X(A(Y )) = DX A(Y ) + A(DX Y ),
for all vectorfields X, Y . Given a k-covariant tensor field T we define its covariant
derivative DX T by the rule,
DX T (Y1 , . . . , Yk ) = XT (Y1 , . . . , Yk )

k
X

T (Y1 , . . . , DX Yi , . . . , Yk )

i=1

We can talk about DT as a covariant tensor of rank k + 1 defined by,

DT (X, Y1 , . . . , Yk ) = DX T (Y1 , . . . , Yk ).
Given a frame e we denote by T1 ... ,k ; = DT (e , ea1 , . . . , ek ) the components
of DT relative to the frame. By repeated covariant differentiation we can define
D2 T, . . . Dm T. Relative to a frame e we write,
D1 . . . Dm T1 ...k = T1 ...k ;1 ...m = Dm T (e1 . . . , em , e1 , ..., ek ).
The fact that the Levi-Civita connection is torsion free allows us to connect covariant differentiation to the Lie derivative. Thus, if T is a k-covariant tensor we have,
in a coordinate basis,
(LX T )1 ...k = X T1 ...k ; + X ; 1 T2 ...k + .... + X ; k T1 ...k1 .
The covariant derivative is also connected to the exterior derivative according to
the following simple formula. If A is a k-form, we have5 A[1 ...k ;] = A[1 ...k ,]
and
X
dA =
A1 ...k ; dx dx1 dx2 .... dxk .

be the
Given a smooth curve x : [0, 1] M, parametrized by t, let T = t
x
corresponding tangent vector field along the curve. A vector field X, defined on
the curve, is said to be parallelly transported along it if DT X = 0. If the curve
has the parametric equations x = x (t), relative to a system of coordinates, then
5[ ... ; ] indicates the antisymmetrization with respect to all indices (i.e. 1 (alternating
1
k
k!
sum of the tensor over all permutations of the indices)) and , indicates the ordinary derivative

with respect to x .

220

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

T = dx
= X (x(t)) satisfy the ordinary differential
dt and the components X
system of equations

D
dX
dx
X
+ (x(t))
X =0.
dt
dt
dt
The curve is said to be geodesic if, at every point of the curve, DT T is tangent
to the curve, DT T = T . In this case one can reparametrize
the curve such that,

satisfies DS S = 0 .
relative to the new parameter s, the tangent vector S = s
x
Such a parameter is called an affine parameter. The affine parameter is defined
up to a transformation s = as0 +b for a, b constants. Relative to an affine parameter
s and arbitrary coordinates x the geodesic curves satisfy the equations

d2 x
dx dx
+

=0.

ds2
ds ds
A geodesic curve parametrized by an affine parameter is simply called a geodesic.
In Lorentzian geometry timelike geodesics correspond to world lines of particles
freely falling in the gravitational field represented by the connection coefficients. In
this case the affine parameter s is called the proper time of the particle.

Given a point p M and a vector X in the tangent space Tp M, let x(t) be the
unique geodesic starting at p with velocity X. We define the exponential map:
expp : Tp M M .
This map may not be defined for all X Tp M. The theorem of existence and
uniqueness for systems of ordinary differential equations implies that the exponential map is defined in a neighbourhood of the origin in Tp M. If the exponential
map is defined for all Tp M, for every point p the manifold M is said geodesically
complete. In general if the connection is a C r connection6 there exists an open
neighbourhood U0 of the origin in Tp M and an open neighbourhood of the point
p in M, Vp , such that the map expp is a C r diffeomorphism of U0 onto Vp . The
neighbourhood Vp is called a normal neighbourhood of p.

3. Riemann curvature tensor, Ricci tensor, Bianchi identities

Riemann curvature tensor, Ricci tensor, Bianchi identities
In the flat spacetime if we parallel transport a vector along any closed curve we
obtain the vector we have started with. This fails in general because the second
covariant derivatives of a vector field do not commute. This lack of commutation
is measured by the Riemann curvature tensor,
R(X, Y )Z = DX (DY Z) DY (DX Z) D[X,Y ] Z
or written in components relative to an arbitrary frame,


R = () (D D D D )e()

(410)

(411)

6A C r connection is such that if Y is a C r+1 vector field then DY is a C r vector field.

3.

RIEMANN CURVATURE TENSOR, RICCI TENSOR, BIANCHI IDENTITIES

221

Relative to a coordinate system x and written in terms of the g components,

the Riemann components have the expression

R =

+
(412)

x
x
The fundamental property of the curvature tensor, first proved by Riemann, states
that if R vanishes identically in a neighbourhood of a point p one can find families
of local coordinates such that, in a neighbourhood of p, g = 7.
The trace of the curvature tensor, relative to the metric g, is a symmetric tensor
called the Ricci tensor,
R = g R
The scalar curvature is the trace of the Ricci tensor
R = g R .
The Riemann curvature tensor of an arbitrary spacetime (M, g) has the following
symmetry properties,
R = R = R = R
R + R + R = 0

(413)

The second identity in (413) is called the first Bianchi identity.

It also satisfies the second Bianchi identities, which we refer to here as the Bianchi
equations and, in a generic frame, have the form:
D[ R] = 0

(414)

The traceless part of the curvature tensor, C is called the Weyl tensor, and has the
following expression in an arbitrary frame,
1
(g R + g R g R g R )
C = R
n1
1
(g g g g )R
(415)
+
n(n 1)
Observe that C verifies all the symmetry properties of the Riemann tensor:
C = C = C = C
C + C + C = 0
and, in addition,

(416)

C = 0 .

are conformal if g
= 2 g for some non zero
We say that two metrics g and g
differentiable function . Then the following theorem holds (see Hawking- Ellis,
[?], chapter 2, section 2.6):
the Weyl tensor relative to g
= 2 g, C
and C the Weyl tensor
Theorem 3.1. Let g
relative to g. Then
= C .
C

Thus C is conformally invariant.

7For a thorough discussion and proof of this fact, see [Sp], Vol. II.

222

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

3.2. Isometries and conformal isometries, Killing and conformal Killing

vector fields. Definition. A diffeomorphism : U M M is said to be a
conformal isometry if, at every point p, g = 2 g, that is,
( g)(X, Y )|p = g( X, Y )|(p) = 2 g(X, Y )|p
with 6= 0. If = 1, is called an isometry of M.
Definition. A vector field K which generates a one parameter group of isometries (respectively, conformal isometries) is called a Killing (respectively, conformal
Killing) vector field.
Let K be such a vector field and t the corresponding one parameter group. Since
the (t ) are conformal isometries, we infer that LK g must be proportional to the
metric g. Moreover LK g = 0 if K is a Killing vector field.
Definition. Given an arbitrary vector field X we denote
tensor of X defined by the formula

(X)

the deformation

(X)

= (LX g) = D X + D X .

The tensor (X) measures, in a precise sense, how much the diffeomorphism generated by X differs from an isometry or a conformal isometry. The following Proposition holds, (see Hawking-Ellis, citeHawkEll, chapter 2, section 2.6):
Proposition 3.3. The vector field X is Killing if and only if
conformal Killing if and only if (X) is proportional to g.

(X)

= 0. It is

Remark: One can choose local coordinates such that X = x . It then immediately follows that, relative to these coordinates the metric g is independent of the
component x .
Proposition 3.4.
On any pseudo-riemannian spacetime M, of dimension n =
p + q, there can be no more than 21 (p + q)(p + q + 1) linearly independent Killing
vector fields.
Proof: Proposition 3.4 is an easy consequence of the following relation, valid for
an arbitrary vector field X, obtained by a straightforward computation and the use
of the symmetries of R.
D D X = R X + (X)

(417)

where
(X)

1
(D + D D )
2

(418)

and (X) is the X deformation tensor.

If X is a Killing vector field equation (417) becomes
D (D X ) = R X

(419)

3.

RIEMANN CURVATURE TENSOR, RICCI TENSOR, BIANCHI IDENTITIES

223

and this implies, in view of the theorem of existence and uniqueness for ordinary
differential equations, that any Killing vector field is completely determined by the
1
2 (n + 1)(n + 2) values of X and DX at a given point. Indeed let p, q be two points
connected by a curve x(t) with tangent vector T . Let L D X , Observe that
along x(t), X, L verify the system of differential equations
D
D
X =T L ,
L = R(, , X, T )
dt
dt
therefore the values of X, L along the curve are uniquely determined by their values
at p.
The n-dimensional Riemannian manifold which possesses the maximum number of
Killing vector fields is the Euclidean space Rn . Simmilarily the Minkowski spacetime Rn+1 is the Lorentzian manifold with the maximum numbers of Killing vectorfields.
3.5. The volume form, Hodge duality and divergence theorem. Let
(M, g) be an orientable pseudo-riemannian manifold M of dimension n and signature s. Let e() be an arbitrary, positively oriented, frame on M e() the corresponding dual frame. We can associate it with the n-form,
e(1) e(2) . . . e(n)
()
the corresponding
If e0() = M e() denotes another basis and e0 () = (M 1 )
e
dual basis of 1-forms, then,

e0 (1) e0 (2) . . . e0 (n) = det(M 1 )e(1) e(2) . . . e(n)

On the other hand if we denote by g and g0 the determinants of the matrices
g(e() , e() ), g(e0() , e0() ) then, g0 = (det M )2 g. Hence, the n from,
= |g|1/2 e(1) . . . e(n)

(420)

is independent of changes of basis which preserves the orientation. Thus, if M is

orientable, we can define a uniques n form called the volume form of M. Relative
to a positively oriented orthonormal frame e() we have 1 ...n = 0 whenever two
indices coincide and equal to the signature of the permutation 1 2 . . . n otherwise.
The following multiplication property of are very useful in calculations,
1 ...n 1 ...n = (1)s det(ji )i.j=1,... ,n

(421)

where s is the signature of the metric. Also, by repeated contractions,

1 2 ...n 1 2 ...n

(1)s det(ji )i.j=2,... ,n

1 2 3 ...n 1 2 3 ...n

2(1)s det(ji )i.j=3,... ,n

...

=
=

...
n!(1)s

1 ...n

1 ...n

Remark that the covariant derivative of vanishes. Indeed this follows easily by
differentiating the formula 1 ...n 1 ...n = n!(1)s from which it follows that,
1 ...n 1 ...n ; = 0 and therefore, using (421), D 1 ...n =1 ...n ; = 0.

224

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

Hodge duality is a naturally defined operator from k (M) to nk (M). In a fixed,

positively oriented frame, if we denote by A0 ...k1 the coefficients of A k (M)
we have,
1
A1 ...nk =
1 ...nk 1 ...k A1 ...k
k!
It is easy to check that, for k-form A we have8,
(A) = (1)s (1)k(nk) A
k

We define the divergence operator : (M)

k1

(422)
(M) by the formula,

A = d A

(423)

Given a continuous, compactly supported function, f on M we define its integral

according to the formula,
Z
Z
f dvg := f
M

If f is supported in a (positively oriented) coordinate chart we have,

Z
Z
f dvg = f (x1 , . . . , xn )|g|1/2 dx1 . . . dxn .
M

We can thus introduce the scalar product for functions on M,

Z
< f1 , f2 >=
f1 f2 dvg

(424)

Given two (compactly supported) forms A k (M), B nk (M) we define,

Z
AB
M

with there integration understood as the usual integration of n-forms on M. Also

given A, B k (M) we define, with respect to a fixed frame e1 , e2
g(A, B) =

1 1 ...k
A
B1 ...k
k!

and,
Z
< A, B >:=

g(A, B)

(425)

With these definitions we can easily check the following formula for A k (M), B
nk (M)
Z
< ?A, B > =
AB
M

Observe also that the divergence operator is dual to d in the following sense (for
A k (M), B k1 (M)),
< AB >=< A, dB >
8 In particular, if g is a Lorentz manifold of dimension 4 we have (A) = (1)k+1 )A.

3.

RIEMANN CURVATURE TENSOR, RICCI TENSOR, BIANCHI IDENTITIES

225

If X is a (compactly supported ) vectorfield on M we define its Hodge dual X to

be the Hodge dual of its associated 1 f orm, i.e.
(X)1 ...n1 =1 ...n1 X
We easily check that, relative to an arbitrary positively oriented coordinate frame,
1
i.e. X = (k1)!
1 ...n1 X dx1 . . . dxn1 , we have
d(X) = (1)n1 (D X )dvg
and hence,
X =: DivX = D X
Combining this fact with the Stokes theorem we deduce the following.
Proposition 3.6. Let X be a vector field in M with ?X its Hodge dual. Let U be a
domain in M with smooth boundary oriented with the positive orientation induced
by that of M. Then,
Z
Z
?
DivX dvg = (1)n1
X.
U

We apply this proposition to a Lorentzian manifold of dimension n + 1. Given a

space-like hypersurface with induced Riemannian metric g we denote by dvg the
corresponding volume element of g, relative to the induced positive orientation. We
now apply the divergence theorem to a lens shaped domain U in M bounded by
two space-like hypersurfaces 0 , 1 .
Corollary 3.7. Let U be a lens shaped domain bounded by two space-like hypersurfaces 0 , 1 . Let T denote the future oriented unit normal on both sides of the
boundary and X an arbitrary smooth vectorfield on M. Then,
Z
Z
Z
(X, T )dvg
(X, T )dvg =
DivXdvg .
1

Proof Consider first the 0 part of the boundary and define, locally around every
point p 0 , a system of coordinates x0 , x1 , . . . xn such that,
(1) x0 = 0 in 0 and x0 < 0 in U.
(2) x1 , x2 , . . . , xn is positively oriented in 0 .
(3) g00 = 1, g0i = 0 on 0 near p.
Relative to these coordinates we have dvg =12...n dx1 . . .dxn where 12lodtsn =01...n .
Therefore, on 0 ,
( ?X)i1 ...in

i1 ...in X =i1 ...i0 X 0 = (1)n X 0 dvg

226

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

3.8. Laplace-Beltrami operator. The scalar Laplace-Beltrami operator on

a pseudo-riemannian manifold M is defined by,
M u(x) = g D D u
where u is a scalar function on M. Or, in local coordinates,
p
1
M u(x) = p
(g |g(x)| )u(x)
|g(x)|

(426)

(427)

The Laplace-Beltrami operator is called DAlembertian in the particular case of a

Lorentzian manifold, and is then denoted by M . On any pseudo-riemannian manifold, M is symmetric relative to the following scalar product for scalar functions
u, v:
Z
(u, v)M = u(x)v(x)dvM
Indeed the following identities are easily established by integration by parts, for
any two smooth, compactly supported9 functions u, v,
Z
(u, v)M =
u v dvM = (u, v)M
(428)
M

= g ij i uj v. In the particular case when u = v we derive, (u, v)M =

Rwhere uv
2
|u|
,
with
|u|2 = u u. Thus, = M is symmetric for functions
M
u C0 (M). It is positive definite if the manifold M is Riemannian. This is not
the case for Lorentzian manifolds: M is non positive definite.
4. Minkowski space
4.1. Basic definitions. The n+1 dimensional Minkowski space, which we denote by Rn+1 , consists of the manifold Rn+1 together with a Lorentz metric m and
a distinguished system of coordinates x , = 0, 1, . . . n, called inertial, relative
to which the metric has the diagonal form m = diag(1, 1, . . . , 1). Two inertial systems of coordinates are connected to each other by translations or Lorentz
transformations. We use standard geometric conventions of lowering and raising
indices relative to m, and its inverse m1 = m, as well as the usual summation
convention over repeated indices. The coordinate vectorfields x are denoted by
, an arbitrary vectorfield is denoted by X = X with X = X (x0 , . . . , xn ).
Observe that by lowering indices relative to m we get X0 = X 0 and Xi = X i
for all i = 1, . . . , n. We denote by D the flat covariant derivative of Rn+1 , that is
D = for an arbitrary 1- form w = dx . We also split the spacetime coordinates x into the time component x0 = t and space components x = xi , . . . xn .
Note that t0 = t and xi = xi for i = 1, . . . , n.
A vector X is said to be timelike, null or spacelike according to whether m(X, X)
is < 0, = 0 or > 0. Accordingly a smooth curve x (s) is said to be timelike,

null or spacelike if its tangent vector dx

ds is timelike, null or spacelike at every
one of its points. A causal curve may be timelike or null. Similarly a hypersurface u(x0 , . . . xn ) = 0 is said to be spacelike, null or timelike if its normal
9This is automatically satisfied if the manifold M is compact.

4. MINKOWSKI SPACE

227

N = m u is, respectively, timelike, null or spacelike. The metric induced by

m on a spacelike hypersurface is necessarily positive definite, that is Riemannian.
A function t(x0 , x1 , . . . , xn ) is said to be a time function if its level hypersurfaces
t = t are spacelike. On a null hypersurface the induced metric is degenerate relative
to the normal direction, i.e. m(N, N ) = 0. In particular function u = u(x0 , . . . xn )
whose level surfaces u = u are null must verify the Eikonal equation
m u u = 0

(429)

Equation (429) can also be written in the form DN N = 0. We call N a null geodesic
generator of the level hypersurfaces of u.
A causal curve can be either timelike and null at any of its points. The canonical
time orientation of Rn+1 is given by the vectorfield T0 = 0 . A timelike vector X is
said to be future oriented if m(X, T0 ) < 0 and past oriented if m(X, T0 ) > 0. The
causal future J + (S) of a set S consists of all points in Rn+1 which can be connected
to S by a future directed causal curve. The causal past J (S) is defined in the
same way. Thus, for a point p = (t, x), J + (p) = {(t t0 , x)/|x x0 | t t0 }.
Given a smooth domain D, its future set J + (D) may, in general, have a nonsmooth
boundary, due to caustics.
We consider conservative domains J + (D1 ) J (D2 ) with D1 1 , D2 2 ,
spacelike hypersurfaces. The domain is regular if both D1 , D2 are regular and
its non- spacelike boundaries N1 (J + (D1 )) \ D1 and N2 (J (D2 )) \ D2
are smooth. In the particular case, when D1 = 1 and D = D2 2 , we obtain
J + (1 )J (D), called domain of dependence of D relative to 1 , consisting of all
points in the causal past of D 2 , to the future of 1 . Similarily J + (D)J (2 ),
with D 1 is called the domain of dependence of influence of D relative to
2 . Particularly useful examples are given in terms of a time function t with
1 = {(t, x)/t(t, x) = t1 }, 2 = {(t, x)/t(t, x) = t1 } two, nonintersecting, level
hypersurfaces, 2 lying in the future of 1 .
A pair of null vectorfields L, L form a null pair if m(L, L) = 2. A null pair
en = L, en+1 = L together with vectorfields e1 , . . . en1 such that m(L, ea ) =
m(L, ea ) = 0 and m(ea , eb ) = ab , for all a, b = 1, . . . , n 1, is called a null frame.
The null pair,
L = t + r ,

L = t r ,

(430)

with r = |x| and r = xi /ri , is called canonical. Simmilarly a null frame

e1 , . . . en+1 with en = L, en+1 = L is called a canonical null frame. In that case
e1 , . . . , en1 form, at any point, an orthonormal basis for the the sphere St,r , of constant t and r, passing through that point. Observe also that L is the null geodesic
generator associated to u = t r while L the null geodesic of u = t + r.

4.2. Conformal Killing vectorfields. Let x be an inertial coordinate system of Minkowski space Rn+1 . The following are all the isometries and conformal
isometries of Rn+1 .

228

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

1. Translations: for any given vector a = (a0 , a1 , ...., an ) Rn+1 ,

x x + a
2. Lorentz rotations: Given any = O(1, n),
x x
3. Scalings: Given any real number 6= 0,
x x
4. Inversion: Consider the transformation x I(x ), where
I(x ) =

x
(x, x)

defined for all points x Rn+1 such that (x, x) 6= 0.

The first two sets of transformations are isometries of Rn+1 , the group generated
by them is called the Poincar`e group. The last two type of transformations are
conformal isometries. the group generated by all the above transformations is called
the Conformal group. In fact the Liouville theorem, whose infinitesimal version will
be proved later on, states that it is the group of all the conformal isometries of Rn+1 .
We next list the Killing and conformal Killing vector fields which generate the above
transformations.
i. The generators of translations in the x directions, = 0, 1, ..., n:
T =

ii. The generators of the Lorentz rotations in the (, ) plane:

L = x x
iii. The generators of the scaling transformations:
S = x
iv. The generators of the inverted translations
K = 2x x

10

(x x )
x
x

10Observe that the vector fields K can be obtained applying I to the vector fields T .

4. MINKOWSKI SPACE

229

We also list below the commutator relations between these vector fields,
[L , L ] = L L + L L
[L , T ] = T T
[T , T ] = 0
[T , S] = T
[T , K ] = 2( S + L )
[L , S] = [K , K ] = 0
[L , K ] = K K

(431)

Denoting P(1, n) the Lie algebra generated by the vector fields T , L and K(1, n)
the Lie algebra generated by all the vector fields T , L , S, K we state the following version of the Liouville theorem,
Theorem 4.3. The following statements hold true.
1) P(1, n) is the Lie algebra of all Killing vector fields in Rn+1 .
2) If n > 1, K(1, n) is the Lie algebra of all conformal Killing vector fields in Rn+1 .
3) If n = 1, the set of all conformal Killing vector fields in R1+1 is given by the
following expression
f (x0 + x1 )(0 + 1 ) + g(x0 x1 )(0 1 )
where f, g are arbitrary smooth functions of one variable.
Proof: The proof for part 1 of the theorem follows immediately, as a particular
case, from Proposition (3.4). From (417) as R = 0 and X is Killing we have
D D X = 0 .
Therefore, there exist constants a , b such that X = a x + b . Since X is
Killing D X = D X which implies a = a . Consequently X can be
written as a linear combination, with real coefficients, of the vector fields T , L .
Let now X be a conformal Killing vector field. There exists a function such that
(X)

(432)

From (417) and (418) it follows that

1
(, + , , )
2
Taking the trace with respect to , , on both sides of (433) we infer that
D D X =

n1

2
n+1
D X =

(433)

X =

(434)

and applying D to the first equation,  to the second one and subtracting we
obtain
 = 0

(435)

230

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

Applying D to the first equation of (434) and using (435) we obtain

n1
(n 1)D D =
(D D + D D ) = (D X + D X )
2
= () = 0
(436)
Hence for n 6= 1, D D = 0. This implies that must be a linear function of x .
We can therefore find a linear combination, with constant coefficients, cS + d K
such that the deformation tensor of X (cS + d K ) must be zero. This is the
case because (S) = 2 and (K ) = 4x . Therefore X (cS + d K ) is Killing
which, in view of the first part of the theorem, proves the result.
Part 3 can be easily derived by solving (432). Indeed posing X = a0 + b1 , we
obtain 2D0 X0 = , 2D1 X1 = and D0 X1 + D1 X0 = 0. Hence a, b verify the
system
b
b
a
a
=
,
=
.
0
1
0
x
x
x
x1
2b
2a
Hence the one form adx0 + bdx1 is exact, adx0 + bdx1 = d, and x
0 2 = x1 2 , that
is  = 0. In conclusion




1

1
+

(
+

)
+
(0 1 )
X=
0
1
2 x0
x1
2 x0
x1
which proves the result.
Remark.

Expresse relative to the canonical null pair,

T0 = 21 (L + L),

S = 21 (u L + u L),

K0 = 21 (u2 L + u2 L).
(437)

Both T0 = t and K0 = (t +|x| )t +2tx i are causal. This makes them important
in deriving energy estimates. Observe that S is causal only in J + (0) J (0).
4.4. Null hypersurfaces. Null hypersurfaces are particularly important as
they correspond to the propagation fronts of solutions to the wave or Maxwell
equation in Minkowski space11. The simplest way to describe the geometry of a
null hypersurfaces is to start with a codimension one hypersurface S0 0 , where
0 is a fixed spacelike hypersurface of Mn+1 . At every point p S0 there are
precisely two null directions ortogonal to the tangent space Tp (S0 ). Let L denote
a smooth null vectorfield orthogonal to S0 and consider the congruence of null
geodesics12 generated by the integral curves of L. As long as these null geodesics
do not intersect the congruence forms a smooth null hypersurface N . We can also
extend L, by parallel transport, to all points of N . Clearly DL L = 0, m(L, L) = 0,
moreover m(L, X) = 0 for every vector X tangent to N . Observe also that L is
uniquely defined up to multiplication by a conformal factor depending only on S0 .
Define, for all vectorfields X, Y tangent to N ,
(X, Y ) = m(X, Y ),

(X, Y ) = m(DX L, Y )

(438)

They are both symmetric tensors, called, respectively, the first and second null fundamental forms of N . Observe that is uniquely defined up to the same conformal
11Or more generally on a Lorentz spacetime.
12 These are in fact straight lines in Minkowski space.

4. MINKOWSKI SPACE

231

factor associated to L. Clearly (L, X) = (L, X) = 0 for all X tangent to N ,

therefore they both depend, at a fixed p N , only on a fixed hyperplane transversal to Lp . Define s, called affine parameter, by the condition L(s) = 1, s = 0 on S0 .
Its level surfaces defines the geodesic foliation of N . Given coordinates w = ( a ),
a = 1, . . . n 1 on S0 we can parametrize points on Ss by the flow x (s, ) defined

by dx
ds = L with x (0, ) the point on S0 of coordinates w. Let,
ab = (

,
),
a
a

ab = (

,
)
a
b

denote the components of and relative to these coordinates. One can easily
d
check that ds
ab = 2ab . The volume element of Ss is given by
p
daSs = ||d 1 . . . dwn1
d
d
with the determinant of the metric . Observe that ds
log || = ab ds
ab = 2tr,
ab
with tr = ab the expansion coefficient of the null hypersurface. Thus,

p
dp
|| = tr ||.
ds
The rate of change of the total volume |Ss | is given by the following formula,
Z
d
|Ss | =
trdaSs .
(439)
ds
Ss
We also remark that verifies the following Ricatti type equation,
d
+ 2 = 0
ds

(440)

which can be explicitely integrated. Thus one can verify that tr(s, 0 ) may become
at a finite value of s > 0 if tr(0, 0 ) < 0 at some point of S0 . This occurence
corresponds to the formation of a caustic.
An arbitrary foliation Sv on N can be parametrized by v(s, ) with (s, ) the
geodesic coordinates defined above. We call = dv
ds the null lapse function of the
foliation and denote by 0 and 0 the restiction of , to Sv . If X is a vectorfield
tangent to the geodesic foliation Ss then X 0 = X 1 X(v)L is tangent to Sv .
Thus, if X, Y are tangent to Ss then (X, Y ) = (X 0 , Y 0 ) and (X 0 , Y 0 ) = (X, Y ).
Relative to the coordinates (v, ) we have
0
ab
= ab ,

0ab = ab .

To define the volume element on a null hypersurface N we choose an arbitrary

foliation v with null lapse function dv
ds = and induced metric and set
daN = 1 daSv dv

(441)

where daSv denotes the area element of Sv induced by . The definition does not
depend on the particular foliation.

232

A. BASIC CONCEPTS IN RIEMANNIAN AND LORENTZIAN GEOMETRY

4.5. Energy momentum tensor. An energy momentum tensor in Rn+1 is

a symmetric two tensor Q verifying the positive energy condition,
Q(X, Y ) 0,
for all X, Y causal, future oriented. We say that Q is divergenceless if,
D Q = 0

(442)

Given an arbitrary vectorfield X,

D (Q X ) = Q D X =

1
Q
2

(X)

where (X) = LX m denotes the deformation tensor of X. Recall that (X) =

X + X . In the particular case when X is a Killing vectorfield, that is
(X)
= 0, we derive
D (Q X ) = 0.

(443)

The same identity holds if X is conformal Killing and Q is traceless, that is

m Q = 0.
A typical conservation law is obtained when we integrate the latter identity, and
apply Stokes theorem, on a regular conservative spacetime domain( see section 4.1)
J + (D1 ) J (D2 ) with smooth spacelike boundaries Di i and null boundaries
Ni , i = 1, 2. We denote by T1 , T2 the future unit normals to the spacelike hypersurfaces 1 , 2 and chose the null normals L1 , L2 such that m(Li , Ti ) = 1 along
the boundaries Di i , i = 1, 2. For simplicity we denote both timelike normals
by T and both null normals by L whenever there is no possibility of confusion.
Proposition 4.6. Assume that Q is a divergenceless energy momentum tensor
and X a Killing vectorfield in a neighborhood of the regular conservative domain
J (D1 , D2 ) as above. Then,
Z
Z
Z
Z
Q(X, T )
(444)
Q(X, L) +
Q(X, T ) =
Q(X, L) +
N2

N1

D2

D1

The integrals are taken with respect to the area elements daN along the null hypersurfaces N1 , N2 and the area elements of the Riemannian metrics induced by m on
1 , 2 . Observe that all integrands are positive if X is causal. The identity (444)
remains valied if X is conformal Killing and Q is traceless.
Proof :

Let P = Q X . According to eqrefeq:cons-law1 we have D P = 0.

The result simplifies for domains of dependence J + (1 )J (D 2 ), or influence

J + (D 1 ) J (2 ), with 2 in the future of 1 . We normalize L by the
condition m(L, T ) = 1 on D 2 where T denotes the unit normal to 1 , 2 .
Corollary 4.7. If Q is divergenceless, X is Killing and D 2 ,
Z
Z
Z
Q(X, L) +
Q(X, T ) =
Q(X, T )
N

D2

J (D)1

(445)

4. MINKOWSKI SPACE

Similarily, if D 1 ,
Z
Z
Q(X, L) +
N

D1

233

Z
Q(X, T ) =

Q(X, T )
J + (D)2

The identity remains true if X is conformal Killing and Q is traceless.

(446)

Bibliography

[Car] S. M. Carroll, Lecture Note on General Relativity, arXiv:gr-qc/9712019v1, Dec 1997

[E] L.C. Evans, Partial Differential Equations, Providence: American Mathematical Society,
1998
[Fr-Io] F.G. Friedlander and M. Ioshi An introduction to the theory of distributions, Cambridge
University Press, second edition, 1998.
[G-T] D. Gilbarg and N. Trudinger, Elliptic partial differential equations of second order, Springer
Verlag, second edition,1983.
[Ge-S] I.M. Gelfand and G.E. Shilov Generalized Functions Volume I: Properties and Operations.
Academic Press, 1964.
[Ho] L. Hormander, The Analysis of linear partial differential operators vols 1-4, Springer Verlag
1983, 1985.
[Ha-E] S.W. Hawking, G.F.R. Ellis, The Large Scale Structure of Space-Time Cambridge University Press, 1973.
[J] F. John, Partial Differential Equations, Providence: American Mathematical Society, fourth
edition, 1982.
[K-T] Markus Keel and Terrence Tao, Endpoint Strichartz Estimates, Am. J. Math. 120 (1998),
955 980.
[Kl-Rodn2] S. Klainerman and I. Rodnianski, Causal geometry of Einstein-vacuum spacetimes
with finite curvature flux, Invent. Math. 159 (2005), 437529.
[Kl-Rodn2] S. Klainerman and I. Rodnianski, A geometric approach to the Littlewood-Paley theory, Geom. Funct. Anal. 16 (2006), 126163.
[Kl-Rodn3] S. Klainerman and I. Rodnianski, Sharp trace theorems for null hypersurfaces on
Einstein metrics with finite curvature flux, Geom. Funct. Anal. 16 (2006), 164229.
[Pet] P/ Petersen, Rimeannain Geometry, GTM 171, Springer, 1998.
[Sp] M. Sivak, A Comprehensive introduction to Differential Geometry, vols I-V, Wilmington:
Publish or Perish.
[T] M.E. Taylor, Partial Differential Equations, Springer Verlag, texts in applied mathematics,
vol 23, 1996.
[Wa] , R.M. Wald, General Relativity, The university of Chicago Press, 1984

235