Fourier

Fourier transform and distributions
with applications to the Schrodinger operator

800674S
Lecture Notes
2nd Edition
Valeriy Serov
University of Oulu
2007
Edited by Markus Harju

Contents
1 Introduction 1
2 Fourier transform in Schwartz space 3
3 Fourier transform in Lp (Rn ), 1 p 2 10
4 Tempered distributions 18
5 Convolutions in S and S 29
6 Sobolev spaces 34
7 Homogeneous distributions 44
8 Fundamental solutions of elliptic partial differential operators 55
9 Schrodinger operator 63
10 Estimates for Laplacian and Hamiltonian 79
11 Some inverse problems for the Schrodinger operator 93
References 103
i
1 Introduction
This text assumes that the reader is familiar with the following concepts:
1) Metric spaces and their completeness.
2) Lebesgue integral in a bounded domain Rn and in Rn .
3) Banach spaces (Lp , 1 p , C k ) and Hilbert spaces (L2 ): If 1 p < then

we set
Z 1/p
p p
L () := {f : C measurable : kf kLp () := |f (x)| dx < }

while
L () := {f : C measurable : kf kL () := ess sup |f (x)| < }.
x
Moreover
X
C k () := {f : C : kf kC k () := max | f (x)| < },
x ||k
where is the closure of . We say that f C () if f C k (1 ) for all k N

and for all bounded subsets 1 . The space C () is not a normed space.
The inner product in L2 () is denoted by
Z
(f, g)L2 () = f (x)g(x)dx.

Also in L2 (), the duality pairing is given by

Z
hf, giL2 () = f (x)g(x)dx.

4) Holders inequality: Let 1 p , u Lp and v Lp with
1 1
+ = 1.
p p
Then uv L1 and
Z Z 1 Z 1
p
p p p
|u(x)v(x)|dx |u(x)| dx |v(x)| dx ,

where the Holder conjugate exponent p of p is obtained via

p
p =
p1
with the understanding that p = if p = 1 and p = 1 if p = .
1
5) Lebesgues theorem about dominated convergence:
Let A Rn be measurable and let {fk } k=1 be a sequence of measurable functions
converging to f (x) point-wise in A. If there exists function g L1 (A) such that
|fk (x)| g(x) in A, then f L1 (A) and
Z Z
lim fk (x)dx = f (x)dx.
k A A
6) Fubinis theorem about the interchange of the order of integration:

Z Z Z Z Z
|f (x, y)|dxdy = dx |f (x, y)|dy = dy |f (x, y)|dx ,
XY X Y Y X
if one of the three integrals exists.
2
2 Fourier transform in Schwartz space
Consider
the Euclidean space Rn , n 1 with x = (x1 , . . . , xn ) Rn and with |x| =
P
x21 + + x2n and scalar product (x, y) = nj=1 xj yj . The open ball of radius > 0
centered at x Rn is denoted by
U (x) := {y Rn : |x y| < }.
Recall the Cauchy-Bunjakovsky inequality
|(x, y)| |x||y|.
Following L. Schwartz we call an n-tuple = (1 , . . . , n ), j N {0} N0 an

n-dimensional multi-index. Denote
|| = 1 + + n , ! = 1 ! n !
and
x = x1 1 xnn , 00 = 1, 0! = 1.
Moreover, multi-indices and can be ordered according to
if and only if j j for all j = 1, 2, . . . , n. Let us also introduce a shorthand notation

= 11 nn , j = .
xj
Definition. The Schwartz space S(Rn ) of rapidly decaying functions is defined as

S(Rn ) = {f C (Rn ) : |f |, := sup x f (x) < for any , Nn0 }.
xRn
The following properties of S = S(Rn ) are readily verified.

1) S is a linear space.
2) : S S for any 0.
3) x : S S for any 0.
4) If f S(Rn ) then |f (x)| cm (1 + |x|)m for any m N. The converse is not

true (see part 3) of Example 2.1).
5) It follows from part 4) that S(Rn ) Lp (Rn ) for any 1 p .

2
Example 2.1. 1) f (x) = ea|x| S for any a > 0.
2 )a
2) f (x) = ea(1+|x| S for any a > 0.
3
3) f (x) = e|x| 6 S.
4) C0 (Rn ) S(Rn ), where C0 (Rn ) = {f C (Rn ) : supp f is compact in Rn }

and supp f = {x Rn : f (x) 6= 0}.
The space S(Rn ) is not a normed space because |f |, is only a seminorm for 0
and > 0 i.e. the condition
|f |, = 0 if and only if f = 0
fails to hold for e.g. constant function f . But the space (S, ) is a metric space if the
metric is defined by
X |f g|,
(f, g) = 2|||| .
,0 1 + |f g|,
Exercise 1. Prove that is a metric, that is,
1) (f, g) 0 and (f, g) = 0 if and only if f = g.
2) (f, g) = (g, f ).
3) (g, h) (g, f ) + (f, h).
4) (In addition) |(f, h) (g, h)| (f, g).
Theorem 1 (Completeness). The space (S, ) is a complete metric space i.e. every
Cauchy sequence converges.
Proof. Let {fk }

k=1 , fk S, be a Cauchy sequence, that is, for any > 0 there exists
n0 () N such that
(fk , fm ) < , k, m n0 ().
It follows that
sup (fk fm ) <
xK
for any 0 and for any compact set K in Rn . It means that {fk }
k=1 is a Cauchy
sequence in the Banach space C (K). Hence there exists a function f C || (K) such
||
that
C || (K)
lim fk = f.
k
Thats why we may conclude that our function f C (Rn ). It only remains to prove
that f S. It is clear that
sup |x f | sup |x (fk f )| + sup |x fk |

xK xK xK
C (K) sup | (fk f )| + sup |x fk |.

xK xK
4
Taking k we obtain
sup |x f | lim sup |fk |, < .

xK k
The last inequality is valid since {fk }

k=1 is a Cauchy sequence, so that |fk |, is
bounded. The last inequality doesnt depend on K either. Thats why we may conclude
that |f |, < or f S.
S
Definition. We say that fk f as k if and only if
|fk f |, 0, k
for all , 0.
Exercise 2. Prove that C0 (Rn ) = S, that is, for any f S there exists {fk }
k=1 , fk
n S
C0 (R ), such that fk f, k .
Now we are in the position to define the Fourier transform in S(Rn ).
Definition. The Fourier transform F f () or fb() of the function f (x) S is defined
by Z
F f () fb() := (2)n/2 ei(x,) f (x)dx, Rn .
Rn
Remark. This integral is well-defined since

Z
b n/2
f () cm (2) (1 + |x|)m dx < ,
Rn
for m > n.
Next we prove the following properties of the Fourier transform:
1) F is a linear continuous map from S into S.
2) fb() = (i)||+||

x (x f (x)).
Indeed, we have Z
fb() = (2) n/2
(ix) ei(x,) f (x)dx
Rn
and hence Z
|x|||
b
f () cm (2)n/2 dx <
L (Rn ) Rn (1 + |x|)m
if we choose m > n + ||. At the same time we obtained the formula

fb() = (ix) f (x). (2.1)
Further, integration by parts gives us

Z
fb() = (i)|| (2)n/2

ei(x,) x f (x)dx.
Rn
5
Thats why we have the estimate
Z
b
f c |x f (x)| dx <
L (Rn ) Rn
since x f (x) S for any 0, if f (x) S. And also we have the formula

fb = (i)|| f .
x (2.2)
If we combine these two last estimates we may conclude that F : S S and F is

a continuous map since F maps every bounded set from S to a bounded set from S
again.
The formulas (2.1) and (2.2) show us that it is more convenient to use the following
notation:

Dj = ij = i , D = D11 Dnn .
xj
For this new derivative the formulas (2.1) and (2.2) can be rewritten as
D fb = (1)|| x
f ,
fb = D f .
Example 2.2. It is true that

1 2 1 2
F (e 2 |x| )() = e 2 || .
Proof. The definition gives us directly

Z
21 |x|2 n/2 1 2
F (e )() = (2) ei(x,) 2 |x| dx
Rn
Z
1 2 1 2 +2i(x,)||2 )
= (2)n/2 e 2 || e 2 (|x| dx
Rn
n Z
Y
1 2 1 2
= (2)n/2 e 2 || e 2 (t+ij ) dt.
j=1
z2
In order to calculate the last integral we consider the function f (z) = e 2 of the
complex variable z and the domain DR depicted in Figure 1. We consider the positive
direction of going around the boundary DR . It is clear that f (z) is a holomorphic
function in this domain and due to Cauchy theorem we have
I
z2
e 2 dz = 0.
DR
But
I 2
Z R 2
Z j Z R Z 0
z2 t2 21 (R+i )2 21 (t+ij )2 1 2
e dz = e dt + i e d + e dt + i e 2 (R+i ) d.
DR R 0 R j
If R then Z j 1 2
e 2 (R+i ) d 0.
0
6
DR
R t
R
Figure 1: Domain DR .
Hence Z Z t2
21 (t+ij )2
e dt = e 2 dt, j = 1, . . . , n.

Using Fubinis theorem and polar coordinates we can evaluate the last integral as
Z 2
2 Z Z 2 Z r2
t2 21 (t2 +s2 )
e dt = e dtds = d e 2 rdr
R2 0 0
Z
= 2 em dm = 2.
0
Thus Z
1 2
e 2 (t+ij ) dt = 2

and n
|x|2 n 1 2 Y 1 2
F (e 2 )() = (2) 2 e 2 || 2 = e 2 || .
j=1
Exercise 3. Let P (D) be a differential operator

X
P (D) = a D
||m

with constant coefficients. Prove that P (D)u = P ()ub.
Definition. Adopt the following notation for translation and dilation of a function
(h f )(x) := f (x h), ( f )(x) := f (x), 6= 0.
Exercise 4. Let f S(Rn ), h Rn and R, 6= 0. Prove that
1)
f () = ||
n
1 fb() and fb() = ||n
1 f ()

7
2)
h f () = e f () and h fb() = e
i(h,) b i(h,) f ().
Exercise 5. Let A be a real-valued nn-matrix such that A1 exists. Denote fA (x) :=

f (A1 x). Prove that
f b
A () = (f )A ()
if and only if A is an orthogonal matrix (a rotation), that is, AT = A1 .
Let us now consider f and g from S(Rn ). Then

Z Z Z
(F f, g)L2 = fb()g()d = (2)n/2 g() i(x,)
e f (x)dx d
R n Rn Rn
Z Z
= (2)n/2 f (x) ei(x,) g()d dx = (f, F g)L2 ,
Rn Rn
where F g(x) := F g(x).

Remark. Here F is the adjoint operator (in the sense of L2 ) which maps S into S
since F : S S. The inverse Fourier transform F 1 is defined as: F 1 := F .
In order to justify this definition we will prove the following theorem.
Theorem 2 (Fourier inversion formula). Let f be a function from S(Rn ). Then
F F f = f.
To this end we will prove first the following (somewhat technical) lemma.
R
Lemma 1. Let f0 (x) be a function from L1 (Rn ) with Rn f0 (x)dx = 1 and let f (x) be
a function from L (Rn ) which is continuous at {0}. Then
Z
x
lim n f0 f (x)dx = f (0).
0 Rn
Proof. Since
Z Z
x x
n f0 f (x)dx f (0) = n f0 (f (x) f (0))dx,
Rn Rn
then we may assume without loss of generality that f (0) = 0. Since f is continuous at
{0} then for any > 0 there exists > 0 such that

|f (x)| < ,
kf0 kL1
whenever |x| < . Note that Z

f0 dx kf0 kL1 .
n
R
8
Thats why we may conclude that
Z Z

n x

n
f0
x
n f f (x)dx dx

0
R kf0 kL1 |x|<
Z
n
f0
x
+ kf kL dx
|x|>
Z

kf0 kL1 + kf kL |f0 (y)| dy = + kf kL I .
kf0 kL1 |y|>
But I 0 as 0. This proves the lemma.

|x|2
Proof of theorem 2. Let us consider v(x) = e 2 . We know from Example 2.2 that
R n/2
Rn v(x)dx = (2) and F v = v. If we apply Lemma 1 with f0 = (2)n/2 v(x) and
n
f S(R ) then
Z
x
(2)n/2 f (0) = lim n v f (x)dx = limhf, n 1 viL2 = limhf, n 1 F viL2
0 Rn 0 0
Z
Ex.4
= limhf, F ( v)iL2 = limhF f, viL2 = hF f, 1i = F f ()ei(0,) d
0 0 Rn
after using the Lebesgue theorem of dominated convergence in the last step. This
proves that
Z
n/2
f (0) = (2) F f ()ei(0,) d = (F F f )(0).
Rn
The proof is now finished by
Z
f (x) = (x f )(0) = (F F (x f ))(0) = (2)n/2 F (x f )()ei(0,) d
Rn
Z
Ex.4
= (2)n/2 ei(x,) F f ()d = F F f (x).
Rn
Corollary. The Fourier transform is an isometry (in the sense of L2 ).

Proof. The fact that Fourier transform retains norm of f S follows from the following
Parseval equality
kF f k2L2 = (F f, F f )L2 = (f, F F f )L2 = (f, f )L2 = kf k2L2 .
Note that
(F f, g)L2 = (f, F g)L2
means that Z Z Z
fb()g()d = f (x)F g(x)dx = f (x)F (g)(x)dx.
Rn Rn Rn
It implies that Z Z
fb()g()d = f (x)gb(x)dx
Rn Rn
or
hF f, giL2 = hf, F giL2 .
9
3 Fourier transform in Lp(Rn), 1 p 2
Let us start with a preliminary proposition.
Proposition. Let X be a linear normed space and E X a subspace of X such that

E = X, that is, the closure E of E in the sense of the norm in X is equal to X. Let
Y be a Banach space. If T : E Y is continuous linear map, i.e. there exists M > 0
such that
kT ukY M kukX , u E,
then there exists a unique linear continuous map Tex : X Y such that Tex |E = T
and
kTex ukY M kukX , u X.
Exercise 6. Prove the previous proposition.
Lemma 1. Let 1 p < . Then

Lp
C0 (Rn ) = Lp (Rn ),
that is, C0 (Rn ) is dense in Lp (Rn ) in the sense of Lp -norm.
Proof. We will use the fact that the set of finite linear combinations of characteristic
functions of bounded measurable sets in Rn is dense in Lp (Rn ), 1 p < . This is a
well-known fact from functional analysis.
Let now A Rn be a bounded measurable set and let > 0. Then there exists
a closed set F and open set Q such that F A Q and (Q \ F ) < p (or only
(Q) < p if there is no closed set F A). Here is the Lebesgue measure in Rn .
Let now be a function from C0 (Rn ) such that supp Q, |F 1 and 0 1.
Then
Z Z
k A kpLp (Rn ) = p
|(x) A (x)| dx 1dx = (Q \ F ) < p
Rn Q\F
or
k A kLp (Rn ) < ,
where A denotes the characteristic function of A i.e.

1,xA
A (x) =
0, x
/ A.
Thus, we may conclude that C0 (Rn ) = Lp (Rn ) for 1 p < .

Remark. Lemma 1 does not hold for p = . Indeed, for a function f c0 6= 0 and for
any function C0 (Rn ) we have that
kf kL (Rn ) |c0 | > 0.
10
Hence we cannot approximate any function from L (Rn ) by functions from C0 (Rn ).
It means that
L
C0 (R ) 6= L (Rn ).
n
But the following result holds:

L
Exercise 7. Prove that S(Rn ) = C(Rn ), where
C(Rn ) := {f C(Rn ) : lim f (x) = 0}.

|x|
Now we are in the position to extend F from S L1 to L1 .
Theorem 1 (Riemann-Lebesgue lemma). Let F : S S be the Fourier transform in

Schwartz space S(Rn ). Then there exists a unique extension Fex as a map from L1 (Rn )
to C(Rn ) with norm kFex kL1 L = (2)n/2 .
Proof. We know that kF f kL (2)n/2 kf kL1 for f S. Now we apply the pre-
L1
liminary proposition to E = S, X = L1 and Y = L . Since S = L1 (it follows
L1
from C0 S and C0 = L1 ) for any f L1 (Rn ) there exists {fk } S such that
kfk f kL1 0 as k . In that case we can define
L
Fex f := lim F fk .
k
L
Since S = C (see Exercise 7) then Fex f C and kFex kL1 L (2)n/2 . On the
other hand
kF f kL |fb(0)| = (2)n/2 kf kL1
for f L1 and f 0. Hence kFex kL1 L = (2)n/2 .
Alternative proof. If f L1 (Rn ) then we can define Fourier transform F f () directly
by Z
n/2
F f () := (2) ei(x,) f (x)dx,
Rn
since Z Z
i(x,)
e
f (x)dx |f (x)|dx = kf kL1 .

Rn Rn
Also we have
Z
b
(2) n/2
f ( + h) fb() =
sup ei(,x) i(h,x)
(e 1)f (x)dx
L (Rn ) Rn Rn
Z Z
2
|f (x)|dx + |f (x)|dx := I1 + I2 .
|x|> |h| |x||h|
Here we have used the fact that |eiy 1| |y| for y R with |y| 1. It is easily seen
that I1 0 for |h| 0 and I2 0 for 0, since f L1 (Rn ).
11
It means that the Fourier transform fb() is continuous (even uniformly continuous)
on Rn . Moreover, we have
Z

2fb() = (2)n/2 i(x,)
e f (x) f x + 2 dx.
Rn ||
This equality follows from
Z Z

i , i(x,)
fb() = (2)n/2 e e i i(x,)
f (x)dx = (2) n/2
e ||2 e f (x)dx
Rn Rn
and Exercise 4. Thus,

2|fb()| (2) n/2
f (x) f x +

0
||2 L1 (Rn )
for || .
Theorem 2 (Plancherel). Let F : S S be the Fourier transform in S with kF f kL2 =
onto
kf kL2 . Then there exists a unique extension Fex of F to L2 -space, such that Fex : L2
L2 and kFex kL2 L2 = 1. Also the Parseval equality remains valid.
L2 L2
Proof. We know that S = L2 since C0 = L2 . Thus for any f L2 (Rn ) there exists
{fk } n
k=1 S(R ) such that kfk f kL2 (Rn ) 0 as k . By Parseval equality in S
we get
kF fk F fl kL2 = kfk fl kL2 0, k, l .
L2
Thus {F fk } 2 n 2
k=1 is a Cauchy sequence in L (R ) and, therefore, F fk g, where g L .
Thats why we may put Fex f := g. Also we have the Parseval equality
kFex f kL2 = lim kF fk kL2 = lim kfk kL2 = kf kL2

k k
which proves the statement about the operator norm.

Remark. In L2 -space we also have the Fourier inversion formula Fex

Fex f = f or
1
Fex Fex f = f .
Exercise 8. Prove that if f L2 (Rn ) then
L2
1) Fex f () = lim F fR (), where fR (x) = {x:|x|R} (x)f (x)
R+
L2
2) Fex f () = lim F (e|x| f ).
+0
Exercise 9. Let us assume that f L1 (Rn ) and F f () L1 (Rn ). Prove that

Z
n/2
f (x) = (2) ei(x,) F f ()d = F 1 F f (x).
Rn
It means that the Fourier inversion formula is valid.
12
Exercise 10. Let f1 and f2 belong to L2 (Rn ). Prove that
(f1 , f2 )L2 = (F f1 , F f2 )L2 .
Theorem 3 (Riesz - Torin interpolation theorem). Let T be a linear continuous map

from Lp1 (Rn ) to Lq1 (Rn ) with norm estimate M1 and from Lp2 (Rn ) to Lq2 (Rn ) with
norm estimate M2 . Then T is a linear continuous map from Lp (Rn ) to Lq (Rn ) with p
and q such that
1 1 1 1
= + , = + , 0 1,
p p1 p2 q q1 q2
with norm estimate M1 M21 .
Proof. Let F and G be two functions with the properties:
1) F, G 0,
2) kF kL1 = kGkL1 = 1.
Let us consider now the function (z) of complex variable z C given by

Z z
+ 1z
z
+ 1z
q q
(z) := M1z M21+z T (f0 F p1 p2
)(x)g0 G 1 2 (x)dx,
Rn
where q11 + q1 = 1, q12 + q1 = 1, |f0 | 1 and |g0 | 1. The two functions f0 and g0 will
1 2
be selected later. We assume also that 0 Re (z) 1.
Our aim is to prove the inequality
|hT f, giL2 | M1 M21 kf kLp kgkLq ,
where
1 1 1 1 1 1
= + , = + , + = 1.
p p1 p2 q q1 q2 q q
z
+ 1z
z
+ 1z

Since T is a linear continuous map and F p1 p2 , G q1 q2 are holomorphic functions
with respect to z (consider az = ez ln a , a > 0) we may conclude that (z) is a holomor-
phic function also.
1) Let us assume now that Re (z) = 0, i.e. z = iy. Then we have (iy) =
iy
iy
+ 1iy + 1iy
M1iy M21+iy hT (f0 F p1 p2 ), g0 G q1 q2 iL2 . Since |aix | = 1 for a, x R, a > 0, it

follows from Holder inequality and the assumptions on T that

iy 1iy iy
+ 1iy
f0 F p1 + p2
|(iy)| M21 M2 g0 G q1 q2

Lp2 q
L 2
1 1 1
1 q
= |f0 |F p2 |g0 |G q2 kF kLp21 kGk 2
= 1.
q L1
Lp2 L 2
13
2) Let us assume now that Re (z) = 1, i.e., z = 1 + iy. Then we have similarly that

1+iy iy 1+iy
+ iy
f0 F p1 + p2 q q
|(1 + iy)| M11 M1 g0 G 1

2

Lp1 q
L 1
1 1 1
1 q
= |f0 |F p1 |g0 |G q1 kF k p1
kGk 1
= 1.
q L1 L1
Lp1 L 1
If we apply now the Phragmen-Lindelevs theorem for the domain 0 < Re (z) < 1 we
obtain that |(z)| 1 for any z such that 0 < Re (z) < 1. Then |()| 1 also for
0 < < 1. But this is equivalent to the estimate
1 1
|hT (f0 F p ), g0 G q iL2 | M1 M21 , (3.1)
where p1 = p1 + 1 , 1 = q1 + 1
p2 q q2
and 1q + q1 = 1. In order to finish the proof of this

theorem let us choose (for arbitrary functions f Lp and g Lq with p and q as
above) the functions F, G, f0 and g0 as follows:

F = |f1 |p , G = |g1 |q , f0 = sgn f1 , g0 = sgn g1 ,
f g
where f1 = kf kLp
, g1 = kgk
and
Lq

1,
f >0
sgn f = 0, f =0

1, f < 0.
1 1
In that case f1 = f0 F p and g1 = g0 G q . Applying the estimate (3.1) we obtain

f g

T , M1 M21 ,
kf kLp kgkLq L2

which is equivalent to
|hT f, giL2 | M1 M21 kf kLp kgkLq .
It implies the desired final estimate
kT f kLq M1 M21 kf kLp .
Theorem 4 (Hausdorff-Young). Let F : S S be the Fourier transform in Schwartz

space. Then there exists a unique extension Fex as a linear continuous map

Fex : Lp (Rn ) Lp (Rn ),
1 1
where 1 p 2 and p
+ p
= 1. What is more, we have the norm estimate
kFex kLp Lp (2)n( p 2 ) .

1 1
It is called the Hausdorff-Young inequality.
14
Proof. We know from Theorems 1 and 2 that there exists a unique extension Fex of
the Fourier transform from S to S for spaces:
n
1) Fex : L1 (Rn ) L (Rn ) with norm estimate M1 = (2) 2
2) Fex : L2 (Rn ) L2 (Rn ) with norm estimate M2 = 1.
Applying now Theorem 3 we obtain that Fex : Lp Lq , where
1 1 1 1 1 1
= + = + , = + = .
p 1 2 2 2 q 2 2 2
It follows that
1 1
+ =1
p q
i.e. q = p and = p2 1. For to satisfy the condition 0 1 we get 1 p 2.
We may also conclude that
kFex kLp Lp ((2) 2 ) 11 = (2)n( p 2 ) .

n 1 1
Remark. In order to obtain Fex in Lp (Rn ), 1 p 2, constructively we can apply the

following procedure. Let us assume that f Lp (Rn ), 1 p 2, and {fk } n
k=1 S(R )
such that
kfk f kLp (Rn ) 0, k .
It follows from Hausdorff-Young inequality that
kF fk F fl kLp (Rn ) Cn kfk fl kLp (Rn ) .

It means that {F fk } p n
k=1 is a Cauchy sequence in L (R ). Thats why we can define

Lp
Fex f := lim F fk .
k
And we also have the Hausdorff-Young inequality
kFex f kLp = lim kF fk kLp lim Cn kfk kLp = Cn kf kLp .

k k
1
Example 3.1 (Fourier transform on the line). Let f2 (x) = (xi)2
, where > 0 is
fixed. It is clear that f2 L1 (R) and
Z
1 + eix dx
fc2 () = .
2 (x i)2
e iz
In order to calculate this integral we consider the function F (z) := (zi) 2 of complex
variable z C. It is easily seen that z = i, > 0 is a pole of order 2. We consider the

cases > 0 and < 0 separately, see Figure 2.
15
Im z Im z
+
DR
b
i b
i
R R Re z Re z
R R

DR
+
Figure 2: Domains DR and DR of integration.
1) > 0. By Cauchy theorem we have

I Z R Z

F (z)dz = 0 = F (z)dz + F (z)dz = I1 + I2 .
DR R
|z|=R
Im z<0
It follows that Z + eix dx

I1 , R ,
(x i)2
and
I2 0, R
due to Jordans lemma, since Im z < 0. Thats why we may conclude that
Z + eix dx
=0
(x i)2
for > 0.
2) < 0. In this case again Im z < 0. So we may apply Jordans lemma again and
obtain
I Z Z
R eix dx
F (z)dz = + F (z)dz = 2iRes F (z).
+
DR R (x i)2 z=i
|z|=R
Im z>0
Hence Z eix dx
= 2i((z i)2 F (z)) |z=i = 2e .
(x i)2
If we combine these two cases we obtain

1
() = 2H()e ,
(x i)2
16
where
t01,
H(t) =
0, t < 0
is the Heaviside function. Similarly we obtain

1
() = 2H()e .
(x + i)2
1
Example 3.2. Let f1 (x) = xi / L1 (R), but
, where > 0 is fixed. It is clear that f1
f1 Lp (R), 1 < p 2. Analogous to Example 3.1 we obtain:

1 i 2H()e , 6 0
=
() =
x + i i , 2
=0
and

1 i 2H()e , 6= 0
() =
x i i , 2
= 0.
Exercise 11. Find the Fourier transforms of the following functions on the line.

ex , x>0
a) f (x) =
0, x0
1
b) f (x) = e|x| and f (x) = 1+x2
1
c) f3 (x) = (xi)3
, > 0.
Exercise 12. Define the Laplace transform by

Z
L(p) := f (x)epx dx,
0
where |f (x)| M eax , x > 0, f (x) = 0, x < 0 and p = p1 + ip2 , p1 > a. Prove that

a) L(p) = 2 f (x)ep1 x (p2 ).
b) Apply the Fourier inversion formula to prove the Mellin formula

Z
1 p1 +i
f (x) = L(p)epx dp, p1 > a.
2i p1 i
17
4 Tempered distributions
In this chapter we will consider two types of distributions: Schwartz distributions
and tempered distributions. To that end we consider the space D := C0 (Rn ) of test
functions. It is clear that D is a linear space and D S. A notion of convergence is
given in
Definition. A sequence {k }
k=1 is a null-sequence in D if and only if
1) there exists a compact set K Rn , such that supp k K for any k and
2) for any 0 we have
sup |D k (x)| 0, k .
xK
D D D
We denote this fact by k 0. As usual, k D means that k 0.
Now we are in the position to define the Schwartz distribution space.
Definition. Functional T : D C is a Schwartz distribution if it is linear and

continuous, that is,
1) T (1 1 + 2 2 ) = 1 T (1 ) + 2 T (2 ) for any 1 , 2 D and 1 , 2 C
2) for any null-sequence k in D it holds that T (k ) 0 in C as k .
The linear space of Schwartz distributions is denoted by D . The action of T on is

denoted by T () = hT, i.
Example 4.1. Every locally integrable function f , that is, f L1loc (Rn ), defines a
Schwartz distribution by the formula
Z
hTf , i := f (x)(x)dx.
Rn
It is clear that Tf is a linear map. It remains to prove only that Tf is continuous map
on D. Let {k } k=1 be a null-sequence in D. Then
Z
|hTf , k i| sup |k (x)| |f (x)|dx 0, k
xK K
by the definition of null-sequence.

D
Example 4.2. If hT, i := (0), then T D . Indeed, T is linear and if k 0
then hT, k i = k (0) 0 for k . This distribution is called the -function and is
denoted by i.e.
h, i = (0), D.
18
Remark. A distribution T is regular if it can be written in the form
Z
hT, i = f (x)(x)dx
Rn
for some locally integrable function f . All other distributions are singular.
Exercise 13. Prove that is a singular distribution.
Definition. The functional T defined by
Z Z
(x) (x)
hT, i := lim dx p.v. dx
+0 |x|> x x
on D(R) is called the principal value of x1 . We denote it by T = p.v. x1 .

1
Remark. Note that x
/ L1loc (R) but we have the following

Exercise 14. Prove that
Z Z
1 (x) (x) (x) (0)
hp.v. , i = dx = p.v. dx.
x 0 x x
Example 4.3. Let be a hypersurface of dimension n 1 in Rn and let d stand for
an element of surface area on . Consider the functional
Z
hT, i = a(x)(x)d

on D, where a(x) is a locally integrable function over . We can interpret T in terms

of surface source. Indeed,
Z Z Z
h a()(x )d , i := a()h(x ), (x)id = a()()d .

It is easy to see that T is a singular distribution. This distribution is known as the

simple layer.
Definition. If T D and g C (Rn ) then we may define the product gT by
hgT, i := hT, gi, D.
This product is well-defined because g D.

f
If f is a locally integrable function whose derivative xj
is also locally integrable,
then
Z Z
f f
, = (x)dx = f dx = f, , D
xj Rn xj Rn xj xj
by integration by parts. This property is used to define the derivative of any distribu-
tion.
19
Definition. Let T be a distribution from D . For any multi-index we define the
derivative T by
h T, i := hT, (1)|| i, D.
It is easily seen that T D .
Example 4.4. Consider the Heaviside function H(x). Since H L1loc (R) then
Z

hH , i = hH, i = (x)dx = (0) = h, i.
0
Hence H = .
Example 4.5. Let us prove that in the sense of Schwartz distributions (log |x|) =
p.v. x1 . Indeed,
Z
h(log |x|) , i = hlog |x|, i = log(|x|) (x)dx

Z Z 0

= log(x) (x)dx log(x) (x)dx
0
Z Z

= log(x)( (x) + (x))dx = log(x)((x) (x)) dx
0 0
Z (x) (x) 1
= log(x) [(x) (x)]
0 + dx = hp.v. , i
0 x x
by integration by parts and Exercise 14.

1 1
p.v. = p.v. .
x x2
The following characterization of D is given without a proof: T D if and only
if for any compact K Rn there exists n0 (K) N0 such that
X
|hT, i| C0 sup |D |
||n0 xK
for any D with supp K.
Definition. Functional T : S C is a tempered distribution if
1) T is linear i.e. hT, + i = hT, i + hT, i for all , C and , S
2) T is a continuous on S, i.e. there exists n0 N0 and constant c0 > 0 such that

X
|hT, i| c0 ||,
||,||n0
for any S.
20
The space of tempered distributions is denoted by S . In addition, for Tk , T S the
S C
convergence Tk T means that hTk , i hT, i for all S.
Remark. Since D S the space of tempered distributions is more narrow than the
space of Schwartz distributions, S D . Later we will consider even more narrow
distribution space E which consists of continuous linear functionals on the (widest test
function) space E := C (Rn ). In short, D S E implies that
E S D .
It turns out that members of E have compact support and thats why they are called
distributions with compact support. But more on that later.
Example 4.6. Let us consider R1 .

2
1) It is clear that f (x) = e|x| is a Schwartz distribution, but not a tempered distri-
bution, because part 2) of the previous definition is not satisfied.
m
P
2) If f (x) = ak xk is a polynomial then f (x) S since
k=0

Z X m
k
|hTf , i| = a x (x)dx
k
R k=0
X m Z
|ak | (1 + |x|)1 (1 + |x|)1+ |x|k |(x)|dx
k=0 R
Xm Z
C |ak |||0,k+1+ (1 + |x|)1 dx,
k=0 R
so the condition 2) is satisfied e.g. Rfor = 1, n0 = m + 2. This polynomial is a

regular distribution since hTf , i = R f (x)(x)dx is well-defined.
Definition. Let T be a distribution from D . Then the support of T is defined by
supp T := Rn \ A,
where A = {x Rn : hT, i = 0 for any C with supp U (x)}.
1) if f is continuous then
supp Tf = supp f
2) supp( T ) supp T
3) supp = {0}.
21
Example 4.7. 1) The weighted Lebesgue spaces are defined as
Z 1
p
Lp (Rn ) := {f Lploc (Rn ) : kf kLp := p
(1 + |x|) |f (x)| dx p
< }
Rn
for 1 p < and
L n n
(R ) := {f Lloc (R ) : kf kL

:= ess sup(1 + |x|) |f (x)| < }.
xRn
If f L1 (Rn ) for some > 0 then Tf S . In fact,

Z

|hTf , i| = f dx kf kL1 kkL .
R n
R
It means that R f dx is well-defined in this case and
Z
hTf , i := f dx.
Rn
2) If f Lp , 1 p , then f S . Indeed,
n
Lp (Rn ) L1 (Rn ) for > ,
p
1 1
where p
+ p
= 1. This follows from Holders inequality
Z Z 1
p
(1 + |x|) |f (x)|dx (1 + |x|)p dx kf kLp .
R R

x
3) Let T S , and 0 (x) C0 (Rn ) with 0 (0) = 1. The product 0 k
T is
well-defined in S by

x x
0 T, := T, 0 .
k k

x
If we consider the sequence Tk := 0 k
T then
k S S
1) hTk , i hT, 0 ( xk )i hT, i (since 0 ( xk ) ) so that Tk T .
2) Tk has compact support as a tempered distribution. This fact follows from
the compactness of k = 0 ( xk ).
Now we are ready to prove more serious and more useful fact.
Theorem 1. Let T S . Then there exists Tk S such that
Z
hTk , i = Tk (x)(x)dx hT, i, k ,
Rn
S
where S. In short, S = S .
22
R
Proof. Let j(x) be a function from D C0 (Rn ) with Rn j(x)dx = 1 and j(x) = j(x).
Let jk (x) := k n j(kx). By Lemma 1 of Chapter 2 we have
Z
lim hjk , i = lim jk (x)(x)dx = (0)
k k Rn
S
for any S. That is, jk (x) (x).
The convolution of two integrable functions g and is defined by
Z
(g )(x) := g(x y)(y)dy.
Rn
If h and g are integrable functions and S then it follows from Fubinis theorem
that
Z Z Z Z
hh g, i = (x)dx h(x y)g(y)dy = g(y)dy h(x y)(x)dx
n n Rn Rn
ZR ZR
= g(y)dy Rh(y x)(x)dx = hg, Rh i,
Rn Rn
where Rh(z) := h(z) is the reflection of h.

f
0 (x) D with 0 (0) = 1. For any T S let us put Tk := jk Tk , where
Let now
f x f f
Tk = 0 k T . From above considerations we know that hjk Tk , i = hTk , Rjk i.
Let us prove that this Tk meets the requirements of this theorem. First of all,

x
hTk , i hjk Tfk , i = hTfk , Rjk i = h0 T, jk i
k
x
= hT, 0 (jk )i hT, i, k ,
k
because
S
x
a) 0 k
1 pointwise for k , since 0 (0) = 1 and 0 ( xk )
S
b) jk for k by Lemma 1 of Chapter 2:
Z Z
jk (x y)(y)dy = jk (z)(x z)dz (x).
Rn Rn
Finally jk (x) C0 (Rn ) implies that Tk C0 (Rn ) S also.

Definition. Let us assume that L : S S is a linear continuous map. The adjoint
map L : S S is defined by
hL T, i := hT, Li, T S .
Clearly, L is also a linear continuous map.

Corollary. Any linear continuous map (operator) L : S S admits a linear continu-
: S S .
ous extension L
23
S
Proof. If T S then by Theorem 1 there exists Tk S such that Tk T . Then
i,
hLTk , i = hTk , L i hT, L i := hLT, k .
Now we are in the position to formulate

Theorem 2 (Properties of tempered distributions). The following linear continuous
operators from S into S admit unique linear continuous extensions as maps from S
into S :
1) huT, i := hT, ui, u S;
2) h T, i := hT, (1)|| i;
3) hh T, i := hT, h i;
4) h T, i := hT, ||n 1 i, 6= 0;

5) hF T, i := hT, F i.
Proof. See the previous definition, Theorem 1 and its corollary.
Remark. Since hF 1 F T, i = hF T, F 1 i = hT, F F 1 i = hT, i we have that
F 1 F = F F 1 = I in S .
Example 4.8. 1) Since
Z Z
n n
hF 1, i h1, F i = (F )()d = (2) 2 (2) 2 ei(0,) F d
Rn Rn
n n n
= (2) 2 F 1 F (0) = (2) 2 (0) = (2) 2 h, i
for any S we have that

n
1b = (2) 2
in S .
n
2) b = (2) 2 1, since
Z
b i = h, F i = F (0) = (2) n2
h, ei(0,x) (x)dx = (2) 2 h1, i,
n
S.
Rn
n
Moreover, F 1 = (2) 2 1 in S .
x2 n 2
3) ea 2 = a 2 e 2a , Re a 0, a 6= 0. Indeed, for a > 0 we know that

x2 ( ax)2 n 2
F (ea 2 ) = F (e 2 ) = a 2 e 2a .
If a is such that Re a 0, a 6= 0, then we can use analytic continuation of these

formulas.
24
2 2
n
4) Consider (1 )u = f , where = x 2 + + x2 is the Laplacian in R and
1 n

u and f S . This equation can be solved in S using the Fourier transform.
Indeed, we get
(1 + ||2 )ub = fb
or
ub = (1 + ||2 )1 fb
or
u = F 1 ((1 + ||2 )1 F f ).
If f S then F f S and (1 + ||2 )1 F f S also and then u S exists. If
S
f S then by Theorem 1 there exists fk S such that fk f . Thats why we
may conclude that
S
u = lim uk ,
k
1 2 1
where uk = F ((1 + || ) F fk ).
Exercise 17. Let P (D) be an elliptic partial differential operator

X
P (D) = a D
||m
with constant coefficients and P () 6= 0 for 6= 0. Prove that if u S and P u = 0

then u is a polynomial.
Corollary. If u = 0 in S and |u| const then u const.

1) F (p.v. x1 ) = i 2
sgn

2) F (p.v. x12 ) = 2
||.
Definition. Introduce the tempered distributions

1 1
:= lim
x i0 0 x i
i.e.
1 1
, = lim , , S.
x i0 0 x i
In a similar fashion,
1 1
:= lim
(x i0)2 0 (x i)2
in S .
25
Example 4.9. We know from Example 3.2 that

1 i 2H()e , 6= 0
() =
x + i i 2 , =0
and

1 i 2H()e , 6= 0
() =
x i i , 2
= 0.
Hence

1
1 i 2H(), 6 0
=
= lim =
x + i0 0 x + i i , 2
=0
and

1
1 i 2H(), =6 0
= lim =
x i0 0 x i i , 2
= 0.
It follows from Exercise 18 that

1
1 1
+ = i 2 sgn = 2 i sgn = 2p.v.
x + i0 x i0 2 x
and thus
1 1 1
+ = 2p.v. .
x + i0 x i0 x
In a similar fashion,

1
1
= i 2 1 = i 2 2 b = 2ib
x i0 x + i0
and so
1 1
= 2i.
x i0 x + i0
Add and subtract to get finally
1 1 1 1
= p.v. i and = p.v. + i.
x + i0 x x i0 x
1)

1
1
= 2H() and = 2H()
(x + i0)2 (x i0)2
2)
1 1 1 1 1
2
+ 2
= 2p.v. 2 and 2
= 2i
(x + i0) (x i0) x (x i0) (x + i0)2
26
3)
1 1 1 1
2
= p.v. 2 + i and 2
= p.v. 2 i
(x + i0) x (x i0) x
4)
1
log |x| = p.v.
2 ||
5)

x = (2)n/2 i|| .

1)
c i 1
H() =
2 i0
2)
i 1
s
gn() = p.v. .
2

Example 4.10. Since
h
, i = h , i b = h, i||
b = (1)|| h, i i
b (i) i = h(2) 2 , (i) i = h(2) n2 (i) , i

= h,
n
we get

n
= (2) 2 (i) .
In particular, in dimension one,

1 k k

(k) = i , xck = 2ik (k) .
2
Consider the Cauchy-Riemann operators

1
:= +i
2 x y
and
1
:= i
2 x y
in R2 .
Let us prove the following facts about these operators:
1)
1
= = ,
4
27
2)
1 1
= in R2
z
The last fact means that
1 1
x + iy
Taking the Fourier transform of 2) gives us
is the fundamental solution of .
d1 b
() = ()
z
which is equivalent to
1
1 1
(i1 2 ) () = (2)1 1 =
2 z 2
or
1
1 1
= i () = , 6= 0.
z i1 2 1 + i2
Let us check that this indeed holds true. We have, by Example 3.2,
Z Z Z
1 1 ei(1 x+2 y) 1 i2 y ei1 x
() = dxdy = e dy dx
z 2 R2 x + iy 2 x + iy
Z Z i1 x Z Z i1 x
1 i2 y e 1 0 i2 y e
= e dy dx + e dy dx
2 0 x + iy 2 x + iy
Z
1 i2 y
= e 2(i 2H(1 )ey1 )dy
2 0
Z
1 0 i2 y
+ e 2(i 2H(1 )ey1 )dy
2
Z Z 0
= i H(1 ) ey(1 +i2 ) dy H(1 ) ey(1 +i2 ) dy .
0
For 1 > 0 we have

Z

y(1 +i2 ) ey(1 +i2 ) 1
i e dy = i = i

.
0 1 + i2 0 1 + i2
For 1 < 0 we have
Z 0
0
y(1 +i2 ) ey(1 +i2 ) 1
i e dy = i

= i .
1 + i2 1 + i2
Hence
i
1
() =
z 1 + i2
which proves 2). Part 1) is established with a simple calculation:
2 2 !
1 1 1
= i +i = + = = .
4 x y x y 4 x y 4
28
5 Convolutions in S and S
Consider first the direct product of distributions. Let T1 , T2 , . . . , Tn be one-dimensional
tempered distributions, Tj S (R), j = 1, 2, . . . , n. The product T1 (x1 ) Tn (xn ) can
be formally defined by
hT1 (x1 ) Tn (xn ), (x1 , . . . , xn )i = hT1 (x1 ) Tn1 (xn1 ), 1 (x1 , . . . , xn1 )i
= hT1 (x1 ) Tn2 (xn2 ), 2 (x1 , . . . , xn2 )i
= = hT1 (x1 ), n1 (x1 )i,
where
1 (x1 , . . . , xn1 ) := hTn (xn ), (x1 , . . . , xn )i S(Rn1 )
j (x1 , . . . , xnj ) := hTnj+1 , j1 (x1 , . . . , xnj+1 )i S(Rnj ).
In this sense it is clear that
(x1 , . . . , xn ) = (x1 ) (xn ).
But the product T1 (x)T2 (x), where x are the same, in general case does not exist, that
is, it is impossible to define such product. We remedy this by recalling
Definition. The convolution of the functions S and S is defined as
Z
( )(x) := (x y)(y)dy.
Rn
We can observe the following immediately.

1) The convolution is commutative for any n 1. If n 2, then
Z Z
( )(x) = (x y)(y)dy = (z)(x z)dz = ( )(x).
Rn Rn
If n = 1 then
Z Z
( )(x) = (x y)(y)dy = (z)(x z)dz

Z
= (x z)(z)dz = ( )(x).

2) It is also clear that the convolution is well-defined for and from S, and
moreover for any 0,
Z
x (
)(x) = ( )(x) = x (x y)(y)dy
Rn
Z Z
= (1)|| y (x y)(y)dy = (1)2|| (x y)y (y)dy
Rn Rn
= ( )(x),
where we integrated by parts and used the fact that xj (x y) = yj (x y).
29
We would like to prove that for and from S it follows that from S also. In
fact,
a) C (Rn ) since ( ) = and : S S.
b) decreases at the infinity faster than any inverse power:

Z Z Z

1
(x y)(y)dy c1 |(y)|dy + c2 |(y)|dy
Rn |y|
|x|
2
|x y|m |y|>
|x|
2
Z Z
c1
m
|(y)|dy + c2 |y|m |y|m |(y)|dy
|x| |y| 2
|x| |x|
|y|> 2

c1 c
m
+ 2m = c|x|m , m N.
|x| |x|
Next we collect some important inequalities involving the convolution.
1) Holders inequality implies that
k kL (Rn ) kkLp (Rn ) kkLp (Rn ) , (5.1)
where p1 + p1 = 1, 1 p . It means that the convolution is well-defined even

for Lp (Rn ) and Lp (Rn ). In particular,
k kL (Rn ) kkL1 (Rn ) kkL (Rn ) . (5.2)
2) It follows from Fubinis theorem that

Z Z
k kL1 dx |(x y)||(y)|dy
n Rn
ZR Z
= |(y)|dy |(x y)|dx = kkL1 kkL1 . (5.3)
Rn Rn
3) Interpolating (5.2) and (5.3) leads us to
k kLp kkL1 kkLp . (5.4)
4) Interpolating (5.1) and (5.3) leads us to
k kLs kkLr kkLq , (5.5)
where
1 1 1 1 1 1
= + , = + , = + .
s 1 r p 1 q p 1
It follows that
1
=1 .
s
30
Thus

1 1 1 1 1 1 1 1 1 1 1 1 1
= 1 + = + 1 = + = 1 + .
r p s s p s p p s p p s p
But
!
1 1 1 1 1 1 1 1
= = 1 +1= 1 1 + 1.
p q q q 1 s
Finally we have
" #
1 1 1 1 1 1 1 1
=1 1 =1 1 1 1 +1 =1 +11+
r p s s q 1 s
q s
or
1 1 1
1+ = + .
s r q
Now we are in the position to consider the Fourier transform of a convolution.
1) Let , S. Then S and F ( ) S. Moreover,

Z Z
n
F ( ) = (2) 2 ei(x,) dx (x y)(y)dy
n Rn
ZR Z
n
= (2) 2 (y)dy ei(x,) (x y)dx
Rn Rn
Z Z
n i(y,) n
= (2) 2 (y)e dy (z)ei(z,) dz = (2) 2 F F ,
Rn Rn
i.e.

= (2) 2 b .
n
Similarly,
n
F 1 ( ) = (2) 2 F 1 F 1 .
Hence
n
= (2) 2 F (F 1 F 1 )
| {z } | {z }
1 1
which implies that

n
F 1 F 1 = (2) 2 F (1 1 )
or

= (2) 2 b .
n
2) Let us assume that L1 and Lp , 1 p 2. Then (5.4) implies that

Lp , 1 p 2. Further, F ( ) belongs to Lp by Hausdorff-Young
inequality. Thus,
n
L p .
= (2) 2 b
31
R
Lemma 1. Let (x) be a function from L1 (Rn ) with Rn (x)dx = 1 and let (x) be a
function from L2 (Rn ). Let us set (x) := n x , > 0. Then
L2 (Rn )
lim = .
+0
Proof. By (5.4) we have that L2 (Rn ). Then

n
= (2) 2
c
in L2 . But
n
x
n n n 1 L
=
c
= 1 () = ()
b = ()
b (0),
b +0.

Note also that Z
n n
(0)
b = (2) 2 ei(0,x) (x)dx = (2) 2 .
Rn
Hence

n
L2
= (2) 2 ()
b () (), +0.
By Fourier inversion formula it follows that
L2
, +0.
Theorem 1. For any fixed function from S(Rn ) the map T has, as a linear
continuous map from S to S (with respect to T ), a unique linear continuous extension
as a map from S to S (with respect to T ) as follows:
h T, i := hT, R i,
where R(x) := (x). Moreover, this extension has the properties

1) T = (2) 2 b T
n
2) ( T ) = T = T .
Proof. Let us assume that , and T belong to the Schwartz space S(Rn ). Then we
have checked already the properties 1) and 2) above. But we can easily check that for
such functions the definition is also true. In fact,
Z Z Z
h T, i = ( T )(x)(x)dx = (x y)T (y)dy(x)dx
n Rn Rn
ZR Z
= T (y) (x y)(x)dxdy
Rn Rn
Z Z
= T (y)dy R(y x)(x)dx = hT, R i.
Rn Rn
S
For the case T S the statement of this theorem follows from the fact that S = S
(see Theorem 1 from Chapter 4).
32
Corollary. Since T = T for and T from S then we may define T as
follows (for T S )
hT , i := hT, R i.
Example 5.1. 1) It is true that = . Indeed,
h , i = h, R i = (R )(0)
Z Z
= (y x)(y)dy |x=0 = (y)(y)dy = h, i.
Rn Rn
Alternatively, we note that

n
= (2) 2 b b = 1 b = b
is equivalent to
=
in S .
2) Property 2) of Theorem 1 and part 1) of this example imply that
( ) = = .
3) Let us consider again the equation (1 )u = f for u and f L2 (or even from
S ). Then (1 + ||2 )ub = fb is still valid in L2 and ub = (1 + ||2 )1 fb or
Z
1 n 1
u(x) = F 1
2
fb = (2) 2 F 1 f = K(x y)f (y)dy,
1 + || 1 + ||2 Rn
where Z
1 ei(xy,)
K(x y) := d.
(2)n Rn 1 + ||2
This is the inverse Fourier transform of locally integrable function. This function
K is the free space Greens function of the operator 1 in Rn . We calculate
this integral precisely later.
1 n
R
2. Let j(x) be a function from L (R ) with Rn j(x)dx = 1. Set j (x) =
Lemma
n x
j , > 0. Then
kj f f kLp 0, +0
for any function f Lp (Rn ), 1 p < . In the case p = we can state only the fact
Z Z
(j f )gdx f gdx, +0
Rn Rn
for any g L1 (Rn ).

Exercise 21. Prove Lemma 2 and find a counterexample as to why the first part fails
for p = .
Remark. If j C0 (Rn ) or S(Rn ) then j f C0 (Rn ) or S(Rn ) also for any f
Lp (Rn ), 1 p < .
33
6 Sobolev spaces
Lemma 1. For any function f L2 (Rn ) the following statements are equivalent:
f
1) xj
(x) L2 (Rn ),
2) j fb() L2 (Rn ),
t f (x)
3) lim j t exists in L2 (Rn ). Here tj f (x) := f (x + tej ) f (x) with t R and
t0
ej = (0, . . . , 1, 0, . . . , 0).
L2 fk
4) There exists {fk }
k=1 , fk S, such that fk f and xj
has a limit in L2 (Rn ),
Proof. 1) 2): Since

D b
j f = j f
we have
b
j f = kDj f kL2
L2
by the Parseval equality.
2) 3): Let j fb be a function from L2 (Rn ). Then the equality
1 1 eitj 1
tj f () = (eitj 1)fb() = j fb()
t t tj
holds. But
eitj 1
i
tj
pointwise as t 0. Hence
1 L2
tj f ij fb, t0
t
i.e. (again due to Parseval equality)
1 t L2 f
f , t 0.
t j xj
The same arguments lead us to the statement that 3) 1).

L2 S
fk S f
4) 1): Let fk be a sequence from S such that fk f . Then fk f and xj
xj
fk L2
also. By the condition 4) we have that the limit lim x j
=g exists. Thats why
k
fk S f
we may conclude that xj
g. It means that g = x j
in S .
34
2) 4): Write fb() as the sum of two functions fb() = g() + h(), where
g() = fb(){|j |<1} , h() = fb(){|j |>1} .

L2
Let {gk } be a sequence in S such that, gk g and supp gk {|j | < 2}. Let
L2
{hk } be a sequence in S such that,

hk j h and supp hk {|j | > 21 }. If we
define the sequence fk (x) = F 1 gk + hjk (x), then
hk L2
fck () = gk + g() + h() = fb().
j
But

fk L2
= ij gk + ihk ij (g + h) = ij fb.
xj
It means that (by Fourier inversion formula or Parseval equality)
fk L2 1 f
F (ij fb) = .
xj xj
We have also the following generalization of Lemma 1 to multi-index .

Lemma 2. Let f be a function from L2 (Rn ) and let s N. Then the following
statements are equivalent:
1) D f L2 (Rn ), || s;
2) fb L2 (Rn ), || s;
f
3) lim hh exists in L2 (Rn ), || s. Here h f := (h11 hnn )f and h Rn with
h0
hj 6= 0 for all j = 1, 2, . . . , n.
L2
4) There exists fk S such that, fk f and D fk has a limit in L2 (Rn ) for
|| s.
Proof. Follows from Lemma 1 by induction on ||.
Definition. Let s > 0 be an integer. Then
X
H s (Rn ) := {f L2 (Rn ) : kD f kL2 < }
||s
is the (L2 -based) Sobolev space of order s with the norm

1/2
X
kf kH s (Rn ) := kD f k2L2 (Rn ) .
||s
35
Remark. It is easy to check that H s (Rn ), s N, can be characterized by
Z
H s (Rn )={f L2 (Rn ) : (1 + ||2 )s |fb()|2 d < }.
Rn
Proof. It follows from Parseval equality that

X X

2
X
b 2
kD f k2L2 = D f 2 = f 2
L L
||s ||s ||s
X Z Z X
= | |2 |fb()|2 d = | |2 |fb()|2 d.
||s Rn Rn ||s
But it is easily seen that there are positive constants c1 and c2 such that
X
c1 (1 + ||2 )s | |2 c2 (1 + ||2 )s
||s
or X
||2 (1 + ||2 )s .
||s
Therefore we may conclude, that

X X Z

kD f kL2 < kD
f k2L2 < (1 + ||2 )s |fb()|2 d < .
||s ||s Rn
This property for an integer s justifies the following definition.
Definition. Let s be a real number. Then

s
H s (Rn ) := {f S : (1 + ||2 ) 2 fb L2 (Rn )}
with the norm Z 1

2
kf kH s (Rn ) := (1 + || ) |fb()|2 d
2 s
.
Rn
Definition. Let s > 0 be an integer and 1 p . Then

X
Wps (Rn ) := {f Lp (Rn ) : kD f kLp (Rn ) < }
||s
is called the Sobolev space with norm

1/p
X
kf kWps (Rn ) := kD f kpLp (Rn ) .
||s
36
Exercise 22. Let s > 0 be an even integer and 1 p . Prove that
Z 1
s p
|f |Wps (Rn ) := |F 1
((1 + || ) fb)|p dx
2 2
Rn
is an equivalent norm in Wps (Rn ).
Definition. Let s > 0 be a real number and 1 p . Then

Z 1
s p
Wps (Rn ) := {f S : |F 1 ((1 + ||2 ) 2 fb)|p dx < }
Rn
with the norm Z 1

s p
kf kWps (Rn ) := |F 1
((1 + || ) fb)|p dx
2 2 .
Rn
Exercise 23. Let s R. Prove that
f H s (Rn )
if and only if
fb L2s (Rn ).
Proposition. Let us assume that 0 < s < 1. Then

Z Z Z Z
|f (x) f (y)|2
(1 + || )|fb()|2 d =
2s 2
|f (x)| dx + As dxdy, (6.1)
Rn Rn Rn Rn |x y|n+2s
where As is a positive constant depending on s and n.
Remark. Since 1 + ||2s (1 + ||2 )s , 0 < s < 1, then the right hand side of (6.1) is an
equivalent norm in H s (Rn ).
Proof. Denote by I the double integral appearing in the right hand side of (6.1). Then
Z Z
I = |f (y + z) f (y)|2 |z|n2s dydz
Rn Rn
Z Z Z Z
|ei(z,) 1|2
= |z|n2s dz |ei(z,) 1|2 |fb()|2 d = |fb()|2 d dz
Rn Rn Rn Rn |z|n+2s
by Parseval equality and Exercise 4. We claim that

Z
|ei(z,) 1|2
dz = ||2s A1
s .
Rn |z|n+2s
Indeed, if we consider the Householder reflection matrix
2vv T
A := I , v = ||e1 , Rn
|v|2
37
then AT = A1 = A and A = ||e1 = (||, 0, . . . , 0). It follows that
Z Z Z
|ei(z,) 1|2 |ei(Az,A) 1|2 |ei(y,A) 1|2
||2s dz = ||2s dz = ||2s
dy
Rn |z|n+2s Rn |z|n+2s Rn |y|n+2s
Z Z
|eiy1 || 1|2 |eiz1 1|2
= ||2s dy = dz := A1s .
Rn |y|n+2s Rn |z|n+2s
Therefore
Z Z Z Z Z
2
|f (x)| dx+As |f (x)f (y)| |xy| 2 n2s
dxdy = b 2
|f ()| d+ ||2s |fb()|2 d.
Rn Rn Rn Rn Rn
Remark. Note that As exists only for 0 < s < 1.

Z 2 X Z
kf k2H k+s (Rn ) (1 + || 2k+2s b 2 b
)|f ()| d f 2 +
||2s |D f |2 d
Rn L Rn
||=k
X Z Z
= kf k2L2 + As |D f (x) D f (y)|2 |x y|n2s dxdy.
||=k Rn Rn
Example 6.1. 1 / H s (Rn ) for any s. Indeed, sassume that 1 H s0 (Rn ) for some s0
0
(it is clear that s0 > 0). It means that (1 + ||2 ) 2 1b L2 (Rn ). It follows from this fact
n
that 1b L2loc (Rn ) and, further, 1b L1loc (Rn ). But 1b = (2) 2 , and we know that is
not a regular distribution.
Next we list some properties of H s (Rn ).
1) Since f H s (Rn ) if and only if fb() L2s (Rn ) and L2s (Rn ) is a separable Hilbert
space with the scalar product
Z
(f1 , g1 )
L2s (Rn ) = (1 + ||2 )s f1 g1 d
Rn
then H s (Rn ) is also a separable Hilbert space and the scalar product can be
defined by Z
(f, g)H s (Rn ) = (1 + ||2 )s fb gbd.
Rn
Let us define H s (Rn ) for negative s.

Definition. Let us set for any positive real number s that
H s (Rn ) := (H s (Rn )) ,
where (H s (Rn )) denotes the adjoint space of H s (Rn ) in the sense of Hilbert space
L2 (Rn ) with the norm defined by
|(f, g)L2 (Rn ) |
kf kH s (Rn ) := sup .
06=gH s (Rn ) kgkH s (Rn )
38
2) For < s < t < it follows that S H t (Rn ) H s (Rn ) S .
Example 6.2. H s (Rn ) if and only if s < n2 . Indeed, if we denote

1
hi := (1 + ||2 ) 2
n
then H s (Rn ) is equivalent to (2) 2 his L2 (Rn ) which in turn is equivalent to
s < n2 .
3) Let be a function from H s (Rn ) and a function from H s (Rn ). Then b

L2 (Rn ), so that
L2s (Rn ) and L1 (Rn ) by Holder inequality. Thats
b
s
why we may define (momentarily, and with slight abuse of notation)
Z
h, iL2 (Rn ) :=
b d
Rn
and get
|h, iL2 (Rn ) | kkH s (Rn ) kkH s (Rn ) .
n
For example, if is a function from H 2 +1+ (Rn ), > 0 and = xj
, then
Z Z
n
, = d
b = i(2) 2 j ()d
b
xj L2 (Rn ) Rn xj Rn
is well-defined, since b L2n +1+ (Rn ) and j L2 n 1 (Rn ).

2 2
4) Let X
P (D) = a D
||m
be a differential operator with constant coefficients. Then P (D) : H s (Rn )

H sm (Rn ) for any real s.
Proof.
Z
kP (D)f k2H sm (Rn ) =
(1 + ||2 )sm |P (D)f |2 d
Rn
Z
= (1 + ||2 )sm |P ()|2 |fb()|2 d
Rn
Z
c (1 + ||2 )sm (1 + ||2 )m |fb()|2 d = c kf k2H s (Rn ) .
Rn
There is a generalization of this result. Let P (x, D) be a differential operator

X
P (x, D) = a (x)D
||m
39
with variable coefficients such that |a (x)| c0 for all x Rn and || m.
Then
P (x, D) : H m (Rn ) L2 (Rn ).
Indeed,
X X
b
kP (x, D)f kL2 c0 kD f kL2 = c0 f
L2
||m ||m

c0 (1 + ||2 ) 2 fb
m
= c0 kf kH m .
L2
5)
Lemma 3. Let be a function from S and f a function from H s (Rn ), s R.

Then f H s (Rn ) and

2 2
|s|
kf kH s
c (1 + || ) b kf kH s .
L1
Proof. We know that

Z

f () = (2) 2
n
b )fb()d.
(
Rn
Hence Z
n his
his f = (2) 2 b )his fb()d.
(
Rn his
Let us prove that
|s|
his 2 2 his h i|s|
for any s R. Indeed,
1 1
hi = (1 + ||2 ) 2 (1 + ||2 ) 2 + | | = hi + | | hi(1 + | |).

Since 1 + | | 2h i we have
s
his 2 2 his h is
for s 0. Moreover, for s < 0 we have
his hi|s| |s|
s
= |s|
2 2 h i|s| .
hi hi
It now follows from (5.4) that

d
kf kH s = his f c |hi|s| |
b |his fb| c hi|s| b his fb
L2 L2 L1 L2
for any s R.
40
6) Let us consider now distributions with compact support in more detail than in
Chapter 4.
Definition. Denote E = C (Rn ). We say that T E if T is a linear functional
on E which is also continuous i.e. k 0 in E implies that hT, k i 0 in C.
Here k 0 in E means that
sup | k | 0, k
K
for any compact subset K Rn and for any multi-index .
It can be proved that T E if and only if there exist c0 > 0, R0 > 0 and n0 N0
such that X
|hT, i| c0 sup |D (x)|
||n0 |x|R0
for any C (Rn ). Moreover, members of E have compact support.

Assume that T E . Since (x) = ei(x,) C (Rn ) then hT, ei(x,) i is well-
defined and there exists c0 > 0, R0 > 0 and n0 N0 such that
X X n0
|hT, ei(x,) i| c0 sup |Dx ei(x,) | c0 | | (1 + ||2 ) 2 .
||n0 |x|R0 ||n0
If we now set
T() := (2)n/2 hT, ei(x,) i
then T is a usual function of . The same is true for
T() = (2)n/2 (1)|| hT, ei(x,) i
and hence T C (Rn ). On the other hand |hT, ei(x,) i| c0 hin0 implies that
|T()| c0 hin0 and hence T L2s (Rn ) for s < n0 n2 . So, by Exercise 23, we
may conclude that any T E belongs to H s (Rn ) for s < n0 n2 .
7)
Lemma 4. The closure of C0 (Rn ) in the norm of H s (Rn ) is H s (Rn ) for any
Hs
s R. In short, C0 (Rn ) = H s (Rn ).
Proof. Let f be an arbitrary function from H s (Rn ) and let fR be a new function
such that
fb(), || < R,
f b
R () = R ()f () =
0, || > R.
n
Then fR (x) = F 1 (R fb)(x) = (2) 2 (F 1 R f )(x). It follows from above con-
siderations that F 1 R C (Rn ) as an inverse Fourier transform of a compactly
supported function (but / C0 (Rn )) and
Z Z
kf fR k2H s = |fb() f 2 2s
R ()| hi d = |fb()|2 hi2s d 0, R
Rn ||>R
41
since f H s (Rn ). This was the first step.
R
The second step is as follows. Let j() C0 (|| < 1) with Rn j()d = 1.
Let us set jk () := k n j(k). We remember from Lemma 2 of Chapter 5 that
Lp
jk g g, 1 p < . Define the sequence vk := F 1 (jk f
ck = jk f
R ). Since v R
then supp vck UR+1 (0) and so vck C0 (Rn ). Hence vk S. Thats why
vk H s (Rn ) and
Z
kvk fR k2H s (Rn ) = hi2s |jk f 2
R fR | d
||<R+1
Z
CR |jk f 2
R fR | d 0, k .
||<R+1
/ C0 (Rn ) we take a function C0 (Rn ) with (0) = 1. Then

Since vk

x S
vk vk , A .
A
Hs
x x
This fact implies that A
vk vk as A . Setting fk (x) := A
vk (x)
C0 (Rn ) we get finally

x
kf fk kH s kf fR kH s + kfR vk kH s + vk vk 0
A Hs
if A, k and R are large enough.
Now we are in the position to formulate the main result concerning H s (Rn ).
Theorem 1 (Sobolev embedding theorem). Let f be a function from H s (Rn ) for
s > k + n2 , where k N0 . Then D f C(Rn ) for all such that || k. In short,
n
H s C k (Rn ), s>k+ .
2
Proof. Let f H s (Rn ) S . Then
D f = F 1 F (D f ) = F 1 ( fb()).
What is more,
Z Z Z
|||| s b
| fb()|d c

||||
|fb()|d = c hi |f ()|d
Rn Rn Rn his
Z !1/2 Z 1/2
||2||
c d hi |fb()|2 d
2s
c kf kH s (Rn )
Rn hi2s Rn
if and only if 2s 2|| > n or s > || + n/2.

It means that for such s and the function D f is a Fourier transform of some
function from L1 (Rn ). Due to Riemann-Lebesgue lemma we have that D f from
C(Rn ).
42
2 n q n
Lemma 5. Ls (R ) L (R ) if and only if q = 2 and s 0 or 1 q < 2 and
s > n 1q 12 .
Exercise 25. Prove Lemma 5.
Lemma 6 (Hormander).
1) F : H s (Rn ) Lq (Rn ) for 1 q < 2 and s >
n 1q 12 .

1 1
2) F : Lp (Rn ) H s (Rn ) for 2 n 2
p
.
3) F : L2 (Rn ) L2 (Rn ).
Proof. 1) See Lemma 5.
2) Let f be a function from Lp (Rn ) for 2 n p
2
then kk
b Lp c kkH s . So
|hfb, iL2 (Rn ) | c kf kp kkH s .

Thats why (by duality)
b
f c kf kLp
H s

1 1 1 1
for s > n p
2
=n 2
p
.

3) This is simply the Parseval equality fb = kf kL2 .
L2

1) [0,1] H s (R) if and only if s < 1/2.
2) [0,1][0,1] H s (R2 ) if and only if s < 1/2.

1
3) K(x) := F 1 1+||2
H s (R) if and only if s < 2 n/2.
4) Let f (x) = (x) log log |x|1 in R2 , where (x) C0 (|x| < 1/3). Prove that
f H 1 (R2 ) but f
/ L (R2 ).
Remark. This counterexample shows us that Sobolev embedding theorem is sharp.
Lemma 7. Let us assume that and f from H s (Rn ) for s > n2 . Then F (f ) L1 (Rn ).
Proof. Since f, H s (Rn ) then fb, b L2s (Rn ) for s > n2 . But this implies (see Lemma
5) that fb and b L1 (Rn ) and
n
F (f ) = (2) 2 b fb
also belongs to L1 (Rn ).
n
Remark. It is possible to prove that if , f H s (Rn ) for s > 2
then f H s (Rn )
with the same s.
Exercise 27. Prove that Wp1 (Rn ) Wp1 (Rn ) Wp1 (Rn ) if p > n.
43
7 Homogeneous distributions
We start this chapter with the Fourier transform of a radially symmetric function.
Lemma 1. Let f (x) be a radially symmetric function in Rn i.e. f (x) = f1 (|x|). Let
us assume also that f (x) L1 (Rn ). Then the Fourier transform fb() is also radial
and Z
n n
fb() = ||1 2 f1 (r)r 2 J n2 (r||)dr,
0 2
where J () is the Bessel function of order .
Proof. Let us take the Fourier transform

Z Z Z
n n
fb() = (2) 2 ei(x,) f1 (|x|)dx = (2) 2 f1 (r)rn1 dr ei||r(,) d,
Rn 0 Sn1
where x = r, = || and , Sn1 := {x Rn : |x| = 1}. It is known that

Z n1 Z
i||r(,) 2 2
e d = n1 ei||r cos (sin )n2 d,
Sn1 ( 2 ) 0
where is the gamma function. This fact implies that fb() is a radial function, since
the last integral depends only on ||. A property of Bessel functions is that
Z

i||r cos n2 n
1
n1 J n2 (r||)
e (sin ) d = 2 2 2
n2 . (7.1)
0 2 (r||) 2
Collecting these things we obtain

Z
2n n
fb() = || 2 f1 (r)r 2 J n2 (r||)dr.
0 2
Remark. If we put variable u = cos in the integral appearing in (7.1), then we obtain
Z Z 1
i||r cos n2
I= e (sin ) d = ei||ru ( 1 u2 )n3 du.
0 1
In particular, if n = 3 then (7.1) implies that

Z 1 sin(||r) J 21 (r||)
I= ei||ru du = 2 = 2 1
1 ||r (r||) 2
i.e.
2 sin(||r)
J 1 (r||) = .
2 (||r) 21
44
If n = 2 then Z 1 ei||ru
I= du = J0 (r||)
1 1 u2
i.e. Z
1 1 ei||ru
J0 (r||) = du.
1 1 u2
Remark. For further considerations we state the small and large argument asymptotics
of J for > 1 as

c |x| , |x| +0
J (|x|)
= 1
c cos(A |x| + B ), |x| +.
|x|
Exercise 28. Prove that fb(A) = fb() if A is a linear transformation in Rn with

A = A1 and f is radially symmetric.
Let us return again to the distribution (cf. Example 5.1)

1 1 1
K1 (x) := n F (x).
(2) 2 1 + ||2
Let us assume now that n = 1, 2, 3, 4. Then the last integral can be understood in the
classical sense. It follows from Lemma 1 that
n n
Z r 2 J n2 (r|x|)dr Z 2 J n2 ()d
g(|x|) = (2) n 1 n n 2n
K1 (x) = K 1 2 |x| 2 2
= (2) 2 |x| 2
.
0 1+ r2 0 2 + |x|2
It is not too difficult to prove that for |x| < 1 we have

1, n = 1,

1
|K1 (x)| c log |x| , n = 2,

|x|2n , n = 3, 4.
Exercise 29. Prove this fact.
Remark. A little later we will prove estimates for K1 (x) for any dimension and for all
x Rn .
There is one more important example. If we havethe equation (1 )u = f in
2 n n 1 1
L (R ) (or even in S), then formally u = (2) F 2
||2 1
f = K1 f, where
n
Z 2 J n2 ()d
n 2n
K1 (|x|) = (2) 2 |x| 2
.
0 2 |x|2
But there is a problem with the convergence of this integral near = |x|. Thats why
this integral must be regularized as
n
Z 2 J n2 ()d
lim 2
.
0 0 2 |x|2 i
Recall that
45
1) f (x) := f (x), 6= 0 and
2) h T, i := n hT, 1 i, > 0.

Definition. A tempered distribution T is said to be a homogeneous distribution of

degree m C if
T = m T
for any > 0. In other words,
h T, i = m hT, i
or
hT, i = nm hT, 1 i

for S. The space of all such distributions is denoted by Hm (Rn ).
Lemma 2. F : Hm (Rn ) Hmn (Rn ).
Proof. Let T Hm (Rn ) and S. Then
h T, i = n hT, 1 i = n hT,
1 i =
n
hT, n i
b

b = n h 1 T, i
= hT, i b = n m hT, i , i.
b = nm hT

Definition. Hm (Rn ) := {T Hm (Rn ) : T C (Rn \ {0})}.

1) if T Hm then D T Hm||

and x T Hm+||

2) F : Hm Hmn .
Exercise 31. Let (x) be a function from C (Rn ) with |D (x)| chxim|| for
all 0 and m R. Prove that ()
b C (Rn \ {0}) and (1 )b S, where
n
C0 (R ) and 1 in U (0).

Example 7.1. 1) Hn (Rn ). Indeed,
h , i = n h, 1 i = n 1 (0) = n (0) = n h, i.

But supp = {0}. It means that C (Rn \ {0}). Alternatively one could note
that
n
b = (2) 2 1 H0 (Rn )
and use Exercise 30 to conclude that
n

= F 1 ((2) 2 1) Hn (Rn ).
46

x
2) Let us assume that C (Sn1 ) and m > n. Set Tm (x) := |x|m |x|
,x
Rn \ {0}. Then Tm (x) L1loc (Rn ) and Tm Hm

(Rn ). Indeed,
Z Z
x m
h Tm , i = Tm (x)(x)dx = |x| (x)dx = m hTm , i.
Rn Rn |x|

x
Since |x|m and |x| are from C (Rn \ {0}) then Tm Hm
(Rn ). Moreover,
D Tm Hm||

(Rn ) and x Tm Hm+||

(Rn ) by Exercise 30.
R
3) Let now m = n in part 2) and in addition assume that Sn1 ()d = 0. Note
/ L1loc (Rn ). But we can define Tn as a distribution from S by
that Tn (x)
Z
hp.v.Tn , i := Tn (x)[(x) (0)(|x|)]dx,
Rn
where S(Rn ) and S(R) with (0) = 1. We assume that R

is fixed. But
it is clear that this definition does not depend on , because Sn1 ()d = 0.
Exercise 32. Prove that,

Z
hp.v.Tn , i = lim Tn (x)(x)dx,
+0 |x|
R
x
where Tn = |x|n |x|
, Sn1 ()d = 0.
Let us prove that:

1) p.v.Tn Hn and
H (Rn ) and moreover it is bounded.
2) p.v.Tn 0
Proof. Part 1) is clear. Part 2) follows from

Z
, i|
|hp.v.Tn b
= |hp.v.Tn , i| = Tn (x)[(x) (0)(|x|)]dx
b b
Rn
Z Z
n
(2) 2 |()|d Tn (x)[ei(x,) (|x|)]dx
Rn Rn
Z Z Z
n

1 ir(,)
= (2) 2 |()|d ()d [e (r)]dr
n R n1 S 0 r
c kkL1 (Rn ) .
L (Rn ) by duality.
Hence p.v.Tn
Part 2) implies that
3) p.v.Tn : L2 (Rn ) L2 (Rn ).
47
Indeed, if f L2 (Rn ) then
fb
F (p.v.Tn f ) = (2) 2 p.v.T
n
n
which implies that

kp.v.Tn f kL2 (Rn ) (2) 2 p.v.T
n
n kf kL2 .
L
Remark. Actually, it follows from the Calderon-Zigmund theory that
p.v.Tn : Lp (Rn ) Lp (Rn ), 1 < p < .
Now we want to consider more difficult case than previous one. Denote
Z
1
hp.v. , i := |x|n [(x) (0)(|x|)]dx, (7.2)
|x|n Rn
where
R
S and S with (0) = 1. But now we dont have the condition
Sn1 ()d = 0 as above. This is the reason why (7.2) must depend on the function
(|x|). We will try to choose an appropriate function . Applying the operator we
get
Z
1 1 x
h p.v. n , i = hp.v. n , n 1 i = |x|n n (0)(|x|) dx
|x| |x| Rn
Z
= n |y|n [(y) (0)(|y|)]dy
Rn
Z
= n |y|n [(y) (0)(|y|)]dy
Rn
Z
n
|y|n (0)[(|y|) (|y|)]dy
Rn
1
= hn p.v. , i + Rest,
|x|n
where
Z
Rest = n (0) |y|n [(|y|) (|y|)]dy
Rn
Z Z
n(r) (r)
= h, i dr d
0 r Sn1
Z
(r) (r)
= n n h, i dr,
0 r
n
2 2
where n = ( n
)
is the area of the unit sphere Sn1 . Let us denote the last integral by
2
G(), > 0. Then
Z Z

1 1 1
G () = (r)dr = (t)dt = (0) = .
0 0
48
Also we have that G(1) = 0. Thats why we may conclude that G() = ln . It
implies that
Rest = n n ln h, i
and so
1 1
p.v. n = n p.v. n
+ n n ln (x).
|x| |x|
Taking the Fourier transform we get

1 1 n
F p.v. n = n F (p.v. n
) + (2) 2 n n ln
|x| |x|
or
n 1 1 n
1 F p.v. n = n F (p.v. n
) + (2) 2 n n ln
|x| |x|
or
1 1 n
F p.v. n = F p.v. n () + (2) 2 n ln .
|x| |x|
Let us put now = ||. Then

1 n 1
F p.v. n () = (2) 2 n ln || + F p.v. n .
|x| |x| ||

Since p.v. |x|1n for such is a homogeneous distribution and radial then F p.v. |x|1n

is also a homogeneous distribution and radial. Thats why F p.v. |x|1n ||
depends

only on || = 1. So this term is a constant that depends on the choice of . We will
choose our function (|x|) so that this constant is zero. Then, finally

1 n
F p.v. n () = (2) 2 n ln ||.
|x|
Now let us consider Tm = |x|m , 0 < m < n. It is clear that |x|m L1loc (Rn ).
Thats why the situation is more simple. We have
Z

h|x|m , i = h|x|m , i
b = |x|m (x)dx.
b
Rn
Lemma 1 implies that

n
Z r 2 J n2 (r||) Z

|x|m = || 1 n
2 2
dr = || n+m n
2 m J n2 ()d.
0 rm 0 2
n1
Last integral converges if 2
< m < n. Thats why we may write that
m = C mn n1
|x| n,m || , < m < n.
2
49
In fact, this is true even for m such that 0 < Re (m) < n. It follows by analytic
continuation on m. In order to calculate the constant Cn,m let us apply this distribution
|x|2
to = e 2 . Since b = we get
|x|2 ||2
h|x|m , e 2 i = hCn,m ||mn , e 2 i.
The left hand side is

Z Z
|x|2 r2
|x|m e 2 dx = n rnm1 e 2 dr
Rn 0

n m
Z
nm2 nm nm2
1
= 2 2 n t 2 et dt = 2 2 n .
0 2
Using this the right hand side becomes

mn
||2 n(nm)2 m
Cn,m h|| ,e 2 i = Cn,m 2 2 n .
2
Thats why
m2 m nm2

n m
Cn,m 2 2 n =2 2 n
2 2
which gives us
nm
n
m

Cn,m = 2 2 2 .
m2
Finally, we have
nm
n
m

|x|m = 2 2 ||mn .
2
(7.3)
m
2
Definition. The Hilbert transform Hf of f S is defined by

1 1
Hf := p.v. f
x
i.e. Z
1 f (t)dt
Hf (x) = lim , x R.
0 |xt| xt
1) kHf kL2 (R) = kf kL2 (R) .
2) Hilbert transform has an extension to functions from L2 (R).
3) H 2 = I i.e. H 1 = H.
1
4) (Hf1 , Hf2 )L2 = (f1 , f2 )L2 for f1 Lp and f2 Lp , where p
+ p1 = 1, 1 0 where c does not depend on .

An n-dimensional analogue of the Hilbert transform is developed via
Definition. The functions
xj
Rj (x) := , x 6= 0, j = 1, 2, . . . , n
|x|n+1
are called the Riesz kernels.
xj
Remark. We can rewrite Rj (x) in the form Rj (x) = |x|n j (x), where j (x) = |x|
.
R
1) Sn1 j ()d = 0
2) Rj (x) = n Rj (x), > 0.

These properties imply that we may define Riesz transform by
Rj f = p.v.Rj f,

because in our previous notation Rj (x) = Tn Hn (Rn ) is a homogeneous distribu-
tion. Let us calculate the Fourier transform of the Riesz kernels. Due to homogeneity
it suffices to consider || = 1. We have,
Z
() () = (2) n2 ei(x,) xj
R j = p.v.R j dx
Rn |x|n+1
Z
n ei(x,) xj
= lim (2) 2 dx.
+0,+ <|x|< |x|n+1
Split
Z Z Z
ei(x,) xj ei(x,) xj ei(x,) xj
dx = dx + dx := I1 + I2 .
<|x|< |x|n+1 <|x|<1 |x|n+1 1<|x|< |x|n+1
For I1 we will use integration by parts:
Z
1
I1 = ei(x,) (|x|1n )dx
1n <|x|<1 xj
Z i(x,) Z Z
e ei(x,) xj ei(x,) xj
= cn ij dx + d d
<|x|<1 |x|n1 |x|=1 |x|n |x|= |x|n
Z Z
0 ei(x,)
cn ij n1
dx + ei(x,) xj d 0.
|x|<1 |x| |x|=1
51
But
Z Z Z
i(x,)
xj e d = i ei(x,) d = i cos(|| x1 )d
|x|=1 j |x|=1 j |x|=1
Z
ij
= x1 sin(|| x1 )d = ij C1 , || = 1,
|| |x|=1
where we have used the fact that a rotation maps to (||, 0, . . . , 0). Similarly we may
conclude that
Z Z
ei(x,)
dx = cos(||x1 )|x|1n dx = C2 , || = 1.
|x|<1 |x|n1 |x|<1
If we collect all of these things we obtain:

0
I1 Cn ij , || = 1.
For I2 we will use the following technique

Z Z
+ ei(x,) xj ei(x,)
I2 dx = i dx
|x|>1 |x|n+1 j |x|>1 |x|n+1
Z

= i |x|n1 cos(|| x1 )dx
j |x|>1
Z
ij x1 sin(||x1 )
= dx
|| |x|>1 |x|n+1
= ij const, || = 1.
Exercise 34. Prove the convergence of the last integral.
Collecting these integrals we obtain that

= i C
p.v.R j j n
H (Rn ). Thats why we may

for || = 1. But we know from Exercise 30 that p.v.R j 0
j
conclude that p.v.Rj () = iCn || . Further, we have
b n
j b
R j f = (2) 2 Rj f = iCn f
||
or
j b
Rj f = iCn F 1 f .
||
Corollary.
n
X
.
Rj Rj = Cn
j=1
52
Proof.
n
X n
X j j
F (Rj Rj f ) = iCn iCn fb() = (Cn )2 fb().
j=1 j=1 || ||
Remark. By Parseval equality we have

kRj f kL2 = R j b
C fb
j f = C f = C kf kL2
L2 || L2
L2
i.e.
Rj : L2 (Rn ) L2 (Rn ),
and it follows from Calderon-Zigmund theory that
Rj : Lp (Rn ) Lp (Rn ), 1 < p < .
Let us now introduce the Riesz potential by

1 1 1 b n 1
I f := F f () = (2) 2 F 1 f = I1 f,
|| ||
where, by (7.3),
1
I1 (x) = cn .
|x|n1
Thats why
Z
1 f (y)dy
I f (x) = cn .
Rn |x y|n1

It is straightforward to check that I
xj 1
= cn Rj and hence
1
I f = cn Rj f.
xj
We would like to prove that
I 1 : Ls (Rn ) Ws1 (Rn )
for some s and . Since Rj is a bounded map from L2 (Rn ) to L2 (Rn ) we may conclude
that
1
I : L2 (Rn ) L2 (Rn ). (7.4)
xj
Now let us assume for simplicity that n 3. Let us try to prove that
I 1 : L2 (Rn ) L2 (Rn ). (7.5)
Indeed, for f L2 (Rn )

I 1 f L2 (Rn )
53
if and only if
1 b
f L2 (Rn ).
||
Let us assume now that > 1.
Lemma
5 in Chapter 6 implies that L2 (Rn ) Lr (Rn ) for any 1 r < 2 and
1 1 2n
> n r 2 . But for > 1 we may find appropriate r such that r < n+2 . Thats why
2 n
we may conclude that for function f L (R ) with > 1 it follows from Hausdorff-
r
Young inequality that fb Lr (Rn ) for some r > n2 or |fb|2 L 2 (Rn ). This fact
2n
implies that for || < 1 we have
||1 fb() L2loc .
Indeed,
Z Z 2/r Z 1

r
d ( 2 ) <
r
2
2
|| |fb()|2 d |fb()|r d || 2
||<1 ||<1 ||<1

since r2 > n2n
and r2 < n2 . For || > 1 the function ||1 fb() belongs to L2 (Rn ).
This fact follows from the inequality ||1 |fb()| < |fb()| and from the positivity of
(see Lemma 5 in Chapter 6). This proves (7.5) for > 1.
If we collect (7.4) and (7.5) then we will obtain that
I 1 : L2 (Rn ) W21 (Rn ), > 1.
Let us consider now L n n

(R ) for > 1. If f L (R ) then |f (x)| C(1 + |x|)

and thus Z
(1 + |y|) dy
|I 1 f (x)| C < .
Rn |x y|n1
It means that
I 1 : L n n
(R ) L (R ).
Interpolating this with (7.5) we can get the following result:
I 1 : Ls (Rn ) Ls (Rn ), 2 s , > 1.
If we recall the fact that Rj : Ls (Rn ) Ls (Rn ) for any 1 < s < then we obtain
I 1 : Ls (Rn ) Ws1 (Rn ), 2 s < , > 1.
54
8 Fundamental solutions of elliptic partial differen-
tial operators
Let us consider a linear partial differential operator of order m in the form
X
L(x, D) = a (x)D , x Rn ,
||m
where = (1 , . . . , n ) is a multi-index, D = D11 Dnn and Dj = 1

i xj
.
In this chapter is a bounded domain in Rn or = Rn .
Definition. A fundamental solution for L in is a distribution E in x, which satisfies
Lx E(x|y) = (x y)
in D () with parameter y , i.e., hLx E, i = (y) for C0 ().

We understand that hLE, i is defined in distributional form
hLE, i = hE, L i,
where L is the formal adjoint operator for L given by

X
L f = (1)|| D (a (x)f (x)).
||m
In that case L must be in D() for from D(). This will be the case, for example,
for a (x) C ().
Any two fundamental solutions for L with the same parameter y differ by a solution
of the homogeneous equation Lu = 0. Unless boundary conditions are imposed, the
homogeneous equation will have many solutions and the fundamental solution will not
be uniquely determined. In most problems there are grounds of symmetry or causality
for selecting the particular fundamental solution for the appropriate physical behavior.
We also observe that if L has constant coefficients, we can find the fundamental
solution in the form E(x|y) = E(x y|0) := E(x y). This fact follows from the
properties of the Fourier transform:
X X
L
x E(x y) =

a E(x y) = = ei(,y) (x)
a ei(,y) E(x) = (x
y)
||m ||m
i.e.
Lx E(x y) = (x y).
Exercise 35. Let L be a differential operator with constant coefficients. Prove that
u = q E = E q solves the inhomogeneous equation
Lu = q
in D .
55
Remark. In many cases the fundamental solution is a function. Thats why we can
write u as an integral Z
u(x) = E(x y)q(y)dy.

Remark. In order for convolution product E q (or q E) to be well-defined we have

to assume that, for example, q vanishes outside a finite sphere.
Remark. If L does not have constant coefficients, we can no longer appeal to convolution
products; instead one can often show that
Z
u(x) = E(x|y)q(y)dy.

Definition. Denote by a0 (x, ) the main (or principal) symbol

X
a0 (x, ) = a (x) , Rn
||=m
of L(x, ). Assume that a (x) are smooth. Operator L(x, D) is said to be elliptic in
if for any x and Rn \{0} it follows that
a0 (x, ) 6= 0.
Exercise 36. Let a (x) be real for || = m. Prove that the previous definition is
equivalent to
1) m is even,
2) a0 (x, ) CK ||m , CK > 0, for any compact set K and for all Rn and
x K.
Let us consider the heat equation

u = u, t > 0, x Rn
t
u(x, 0) = f (x), x Rn
in S (Rn ). Take the Fourier transform with respect to x to obtain

u b(, t) = 2 u
b(, t), t>0
t
u b
b(, 0) = f ().
This initial value problem for an ordinary differential equation has the solution
2
ub(, t) = et|| fb().
Hence
2 n 2
u(x, t) = F 1 (et|| fb()) = (2) 2 F 1 (et|| ) f = P (, t) f,
56
where Z
n 2 1 |x|2
P (x, t) = (2) et|| ei(x,) d = n e
4t
.
Rn (4t) 2
This formula implies that
Z
1 |xy|2
u(x, t) = n e 4t f (y)dy.
(4t) 2 Rn
Definition. The function P (x, t) is the fundamental solution of the heat equation and
satisfies

t
P (x, t) = 0, t > 0
S

lim P (x, t) = (x).
t0
We can generalize this situation as follows. Let us consider an elliptic differential

operator X
L(D) = a D
||m
P
with constant coefficients. Assume that L() = ||m a > 0 for all Rn \{0}. If
we consider PL (x, t) as a solution of

t + L(D) PL (x, t) = 0, t>0
S

lim PL (x, t) = (x)
t0

then PL (x, t) is the fundamental solution of t
+ L(D) and can be calculated by
Z
PL (x, t) = (2)n etL() ei(x,) d.
Rn
Lemma 1. Let PL (x, t) be as above. Then the function

Z
S
F (x, ) := lim et PL (x, t)dt (8.1)
0
is a fundamental solution of the operator L(D) + I, > 0.

Proof. By definitions of F and PL we have
Z Z
hF (x, ), i = limh et PL (x, t)dt, i = lim et hPL , idt.
0 0
Therefore
Z
h(L(D) + )F, i = lim et h(L(D) + )PL , idt
0
Z Z
= lim et hL(D)PL , idt + et hPL , idt
0 0
Z
= lim et h PL , idt + hF, i
0 t
Z
t
= lim e hPL , i|
t
e hPL , idt + hF, i
0
= lim e hPL (, ), i = h, i.
0
57

Exercise 37. Let us define a fundamental solution (x, t) of t
+ L(D) as a solution
of
( + L)(x, t) = (x)(t)
t
(x, 0) = 0.
Prove that Z
F (x, ) := et (x, t)dt
0
is a fundamental solution of the operator L(D) + I, > 0.

P 2
Example 8.1. Assume that L(D) = nj=1 1i x j = . Then L() = ||2 and the
fundamental solution F (x, ) of the operator L(D) + = + has the form
Z Z
1 t x2 1 x2 n
F (x, ) = n e e 4t dt = n et 4t t 2 dt
0 (4t) 2 (4) 2 0
Z Z
1 n
1
n
4 1
( |x|)2 n r2 n
= n 2 e 2 d = n 2
1
e 4 2 d,
(4) 2 0 (4) 2 0

where r = |x|. From our previous considerations we know that

n/2 1 1
F (x, ) = (2) F 2
(x),
|| +
where F 1 is the inverse Fourier transform. The function

1 r
Z r2
K (r) = et 4t t1 dt
2 2 0
is called the Macdonald function of order . So, we have

1 n
n |x| 2
F (x, ) = (2) 2 K n2 1 ( |x|).

It is known that
i i (1)
K (r) =
e 2 H (ir), r > 0,
2
where H(1) is the Hankel function of first kind and order .
Next we want to obtain estimates for F (x, ) for x Rn , > 0 and n 1. Let us
R r2 n R1 R
consider the integral 0 e 4 2 d in two parts I1 + I2 = 0 + 1 .
1) If 0 < r < 1 then

Z 1 Z 1 Z
r2 n r2 n n
I1 = ey 4y y 2 dy e 4y y 2 dy = cn r2n r2
ez z 2 2 dz = cn r2n I1 .
0 0 4
58
n = 1: Since I1 cr1 , r +0 then |I1 | cn
n = 2: Since I1 c ln 1r , r +0 then |I1 | cn ln 1r
n 3: Since I1 c, r +0 then |I1 | cn r2n .
For I2 we can simply argue that

Z Z
r2 n r2 r2
I2 = ey 4y y 2 dy e 4 ey dy e 4 1, r +0.
1 1
2) If r > 1 then

Z 1 Z
2
r 2 e r4 ,
2
r4y n n n = 1, 2, 3, 4
I1 e y 2 dy = cn r2n r2
ez z 2
2
dz cn 2n r2
0 4 r e , n 5,
n
where 0 < < 14 . The last inequality follows from the fact that z 2 2 c ez for
n
2
2 > 0 and for any > 0 (z > 1).
Since Z r2
I2 ey 4y dy
1
r2
we perform the change of variable z := y + 4y
. Then z r and z +. Thus
Z Z

y r4y
2
z z
e dy = c e 1+ 2 dz
1 r z r2
Z Z
z zdz
= c e dz + c ez 2
r r z Z r2

= cer + c ez z 2 r2 + ez z 2 r2 dz
r
Z r
= c er + ez z 2 r2 dz cer
r
for any 0 < < 1.
If we collect all these estimates we obtain that

1) If |x| < 1 then

n=11,

1, n=1
n n
1 1 1 |x| 1
|F (x, )| cn 2 log |x| , n = 2 cn 2 e log ,
|x|
n=2

( |x|)2n , n 3 ( |x|)2n , n 3.

2) If |x| > 1 then
|F (x, )| cn e |x|
, n 1.
59
We will rewrite these estimates in more appropriate form for all > 0 and x Rn as

1 , n=1

|F (x, )| cn e |x|

1
1 + | log |x| |, n = 2

|x|2n , n 3.
Remark. It is not too difficult to observe that F (x, ) is positive.

Example 8.2. Recall from Chapter 7 that the solution of the equation (1 )u = f
can be written in the form

1 1
u(x) = K1 f = F 2
f,
|| 1
where n
Z 2 J n2 ()d
2n
K1 (|x|) = cn |x| lim 2
.
0 0 2 |x|2 i
Actually K1 is a fundamental solution of the operator 1 . Let us consider more
general operator for > 0 or even for C. The operator is called
the Helmholtz operator. Its fundamental solution En (x, ) satisfies
En En = (x).

We define with nonnegative imaginary part i.e. = + i, where 0 and
= 0 if and only if [0, +). We require that En is radially symmetric. Then, for
x 6= 0, En must solve the equation
(rn1 u ) + rn1 u = 0.
This equation can be reduced to one of Bessel type by making the substitution u =
n
wr1 2 . A straightforward calculation shows that

n 2 w
(rw ) 1 + rw = 0
2 r
or
w n 2 1
w + + 1 w=0
r 2 r2
or

v (r ) n 2 1
v (r ) + + 1 1 v(r ) = 0, w(r) = v(r ).
r 2 r2
This is the Bessel equation of order n2 1. Its two linearly independent solutions are the
Bessel functions J n2 1 and Y n2 1 of the first and second kind, respectively. Therefore
the general solution is of the form

w(r) = c0 J n2 1 ( r) + c1 Y n2 1 ( r).
60
For us it is convenient to write it in terms of Hankel functions of first and second kind
as
(1) (2)
w(r) = c0 H n 1 ( r) + c1 H n 1 ( r),
2 2
where
H(1) (z) = J (z) + iY (z), H(2) (z) = J (z) iY (z).
The corresponding general solution u is
n
h
(1)
(2)
i
u(r) = r1 2 c0 H n 1 ( r) + c1 H n 1 ( r) .
2 2
(2)

If
/ [0, +) then has positive imaginary part and the solution H n 1 ( r) is
(1)
2
exponentially large at z = +, whereas H n 1 ( r) is exponentially small. Hence we

2
take
n (1)
En (x, ) = c0 r1 2 H n 1 ( r).
2
Exercise 38. Prove that Z

En
lim d(x) = 1
0 |x|= r
or
En
lim rn1 n = 1,
r0 r
where n = |Sn1 | is the area (measure) of the unit sphere Sn1 .
For small values of r, we have the asymptotic expansions

n2
(1) i2 2 ( n2
2
) n2
H n2 (r) r 2 , n 6= 2
2
and
(1) 2i
H0 (r) log r.

It can be proved using Exercise 38 that
! n2
i 2
c0 = .
4 2
Thus for n 2 and

/ [0, +) we obtain
! n2
i 2
(1)

En (x, ) = H n2 ( |x|). (8.2)
4 2|x| 2
A direct calculation shows that for n = 1 we have

i
E1 (x, ) = ei |x|
2
61
for all 6= 0. The formula (8.2) is valid also for (0, +). This fact follows from
the definition:
! n2
i + i 2 (1)

En (x, ) = lim En (x, + i) = lim H n2 ( + i|x|)
0 4 0 2|x| 2
! 2 n2
i (1)

= H n2 ( |x|).
4 2|x| 2
Remark. We may conclude that

! n2
1 i 2
(1)

(2)n/2 F 1 2
= H n2 ( |x|),
|| i0 4 2|x| 2
for > 0. A direct calculation shows that

|x| ,
2
n=1
1
En (x, 0) = log 1 , n=2
2 2n |x|

|x|
(n2)n
, n 3.
62
9 Schrodinger operator
There are many inverse scattering problems which are connected with the reconstruc-
tion of the quantum mechanical potential in the Schrodinger operator (Hamiltonian)
H = + q(x). This operator is defined in Rn . Here and throughout we assume that
q is real-valued.
First of all we have to define H as a self-adjoint operator in L2 (Rn ). Our basic
assumption is that the potential q(x) belongs to Lp (Rn ) for n2 R (9.1)
with some 0 and R > 0 large enough. This parameter will be specified later,
depending on the situation. We would like to construct the self-adjoint extension of this
operator by Friedrichs method, because formally our operator is defined now only for
smooth functions, say for functions from C0 (Rn ). In order to construct such extension
let us consider the Hilbert space H1 which is defined as follows
Z
2 n 2 n
H1 = {f L (R ) : f (x) L (R ) and |q(x)||f (x)|2 dx < }.
Rn
The inner product in H1 is defined by

Z
(f, g)H1 = (f, g)L2 + q(x)f (x)g(x)dx + 0 (f, g)L2 ,
Rn
with 0 > 0 large enough and fixed.

n
Lemma 1. Assume that f W21 (Rn ) and q Lp (Rn ) for 2
 0 such that
|(qf, f )L2 | kf k2L2 (Rn ) + c kf k2L2 (Rn ) .
Proof. If p = then
Z
|(qf, f )L2 | |q(x)||f (x)|2 dx kqkL (Rn ) kf k2L2 (Rn )
Rn
kf k2L2 (Rn ) + kqkL (Rn ) kf k2L2 (Rn ) .
n
If 2
A
Z
|q(x)||f (x)|2 dx + A kf k2L2 (Rn ) .
|q(x)|>A
63
Let us consider the integral appearing in the last estimate. For n 3 it follows from
Holder inequality that
Z Z 2 Z n2
n n 2n n
2
|q(x)||f (x)| dx |q(x)| dx 2 |f (x)| n2 dx
|q(x)|>A |q(x)|>A |q(x)|>A
Z 2
n
A( 2 p) n
n 2
|q(x)| dxp
c1 kf k2W 1 (Rn )
|q(x)|>A 2
2p 2p
c1 A1 n kqk n
Lp (Rn ) kf k2W 1 (Rn ) .
2
For getting the last inequality we used the fact that n2 < p < and a well-known
2n
embedding: W21 (Rn ) L n2 (Rn ), n 3, with the norm estimate

kf k n2
2n
n
c1 kf kW 1 (Rn ) .
L (R ) 2
Collecting these estimates we obtain

2p 2p
|(qf, f )L2 | c1 A1 n kqkLnp (Rn ) kf k2W 1 (Rn ) + A kf k2L2 (Rn )
2p
2
2p

1 2p 2p
= c1 A n kqk n
Lp (Rn ) kf k2L2 (Rn ) + A + c1 A1 n kqkLnp (Rn ) kf k2L2 (Rn ) .
2p
The claim follows now from the last inequality since A1 n can be chosen sufficiently
small for n2 1.

Find the conditions on 1 and 2 which ensure the statement of Lemma 1.
Remark. Lemma 1 holds for any potential q Lp (Rn ) + L (Rn ) for p > n2 , n 2.
Using Lemma 1 we obtain
kf k2H1 = kf k2L2 (Rn ) + 0 kf k2L2 (Rn ) + (qf, f )L2

kf k2L2 (Rn ) + 0 kf k2L2 (Rn ) kf k2L2 (Rn ) c kf k2L2 (Rn )
= (1 ) kf k2L2 (Rn ) + (0 c ) kf k2L2 (Rn ) .
We choose here 0 < < 1 and 0 > c . On the other hand
kf k2H1 (1 + ) kf k2L2 (Rn ) + (0 + c ) kf k2L2 (Rn ) .
64
These two inequalities mean that the new Hilbert space H1 is equivalent to the space
W21 (Rn ) up to equivalent norms. Thus we may conclude that for any f H1 our
operator is well defined by
(f, (H + 0 )f )L2 (Rn ) = kf k2H1 .
Moreover, since H + 0 is positive then

2
kf k2H1 = (H + 0 ) 2 f
1
L2 (Rn )
and the following statements hold:

1
1) Domain of (H + 0 ) 2 is W21 (Rn )
2) D(H + 0 ) D(H) W21 (Rn )
3) D(H) = {f W21 (Rn ) : Hf L2 (Rn )}.
Remark.
1 1 1
(H+0 )f = (H+0 ) 2 (H+0 ) 2 f D(H) = {f W21 (Rn ) : g := (H+0 ) 2 f W21 (Rn )}.
Remark. Let us consider this extension procedure from another point of view. The
inequality
(f, (H + 0 )f )L2 (1 ) kf k2L2 (Rn ) + (0 c ) kf k2L2 (Rn )
allows us to conclude that
a) (f, (H + 0 )f )L2 c kf k2L2 (Rn ) and
b) (f, (H + 0 )f )L2 c kf k2W 1 (Rn )

2
for any f C0 (Rn ). It means that there exists (H + 0 )1 which is also defined for
g C0 (Rn ) and satisfies the inequality
1
a) k(H + 0 )1 gkL2 (Rn ) c
kgkL2 (Rn ) or even
b) k(H + 0 )1 gkW 1 (Rn ) 1

c
kgkW 1 (Rn ) , where W21 (Rn ) is the conjugate (adjoint)
2 2
space of W21 (Rn ).
L2 W 1
Since (H + 0 )1 is bounded operator and C0 (Rn ) = L2 (Rn ) and C0 (Rn ) = 2
1
W2 (Rn ) then we can extend (H + 0 )1 as the bounded operator onto L2 (Rn ) in
the first case and onto W21 (Rn ) in the second case. The extension for the differential
operator is H + 0 = ((H + 0 )1 )1 and D(H + 0 ) = R((H + 0 )1 ) in both cases.
It is also clear that H + 0 and (H + 0 )1 are self-adjoint operators.
65
Lemma 2. Let us assume that q Lp (Rn ) for 2 p if n = 2, 3 and q Lp (Rn )
for n2 < p if n 4. Then
W22 (Rn ) D(H).
Proof. Since H = +q and D(H) = {f W21 (Rn ) : Hf L2 (Rn )} then for required
embedding it is enough to show that for f W22 (Rn ) it follows that qf L2 (Rn ).
If p = then
Z
|qf |2 dx kqk2L (Rn ) kf k2L2 (Rn ) <
Rn
for any f W22 (Rn ), n 2.

For finite p let us consider first the case n = 2, 3. Since W22 (Rn ) C(Rn ) L (Rn )
(Sobolev embedding) then
Z Z Z
|qf |2 dx = |qf |2 dx + |qf |2 dx
Rn |q|<A |q|>A
Z Z
A 2 2
|f | dx + kf k2L (Rn ) |q|p |q|2p dx
|q|<A |q|>A
A 2
kf k2L2 (Rn ) +C kf k2W 2 (Rn ) A2p kqkpLp (Rn ) < .
2
In the case n 4 we have the embeddings:
n=4: f W22 (R4 ) Lp (R4 ), p < .

2n
n5: f W22 (Rn ) L n4 (Rn ).
Thats why applying the Holder inequality we obtain
n5:
Z Z Z
2 2
|qf | dx = |qf | dx + |qf |2 dx
Rn |q|<A |q|>A
Z 4 Z n4
n n 2n n
A 2
kf k2L2 (Rn ) + |q| dx 2 |f | n4 dx
|q|>A |q|>A
Z 4
n
(n 4
A 2
kf k2L2 (Rn ) + CA 2
p) n p
|q| dx kf k2W 2 (Rn ) < .
|q|>A 2
n=4:
Z Z 2 Z 2
p
2 p p p
|qf | dx |q| dx |f | dx <
R4 R4 R4
for 2 < p < and p < .
66
Exercise 41. Prove this lemma for q Lp (Rn ) + L (Rn ), n2 < p if n 4 and
for q L2 (Rn ) + L (Rn ) if n = 2, 3.
n
Remark. For n 5 we may consider q L 2 (Rn ).
Lemma 3. Let us assume that q Ln (Rn ), n 3. Then
D(H) = W22 (Rn ).
Proof. The embedding W22 (Rn ) D(H) was proved in Lemma 2. Let us now assume
that f D(H) i.e. f W21 (Rn ) and Hf L2 (Rn ). Note that for g := Hf L2 we
have the following representation
f = ( + 1)1 (q 1)f ( + 1)1 g
= ( + 1)1 (qf ) ( + 1)1 g ( + 1)1 f .
| {z } | {z }
W22 W22
Thats why it is enough to show that qf L2 (Rn ). We use the same arguments as in
Lemma 1 and Lemma 2. So it suffices to show that for any f W21 (Rn ) it follows that
2n
qf L2 (Rn ). From the embedding W21 (Rn ) L n2 (Rn ) for n 3 we have by Holder
inequality
Z Z Z
|q(x)|2 |f (x)|2 dx = |q(x)|2 |f (x)|2 dx + |q(x)|2 |f (x)|2 dx
Rn |q|<A |q|>A
Z 2 Z n2
n 2n n
A 2
kf k2L2 (Rn ) + n
|q| dx |f | n2 dx < .
|q|>A |q|>A
Thus lemma is proved.

n
Exercise 42. Describe the domain of H for the case 2
< p < n, n 3. Hint: Prove
that D(H) W22 (Rn ) + Ws2 (Rn ) with some s = s(p).
Let us present some mathematical background material concerning self-adjoint op-
erators in Hilbert spaces. If A : H H is a linear operator in a Hilbert space H with
D(A) = H then the set
(A) := { C : (A I)1 exists as a bounded operator from H to D(A)}
is called the resolvent set of A and its complement
(A) := C \ (A)
is called the spectrum of A. The operator-valued function
R := (A I)1 , (A),
is called the resolvent of A.
If A = A (or even if A is closed) then the resolvent set is open and the spectrum
is closed. Moreover, (A) 6= and (A) R in this case.
67
Definition. If A = A then
a) the point spectrum p (A) of A is the set of all eigenvalues of A, i.e.,
p (A) = { (A) : Ker (A I) 6= 0}.
It means that there exists a non-trivial f D(A) such that Af = f . The linear
subspace
{f D(A) : Af = f }
is called the eigenspace of A corresponding to .
b) the complement (A) \ p (A) is the continuous spectrum c (A) of A.
c) the discrete spectrum d (A) of A is defined as
d (A) := { p (A) : dim Ker (A I) < }
and must also be isolated in (A).
d) the set ess (A) := (A) \ d (A) is called the essential spectrum of A.
Remark. c (A) means that (A I)1 does exist but it is not bounded. It is
equivalent to R(A I) 6= H, i.e., there exists f H such that f
/ R(A I).
Theorem 1 (Spectral theorem of J. Neumann). Let us assume that A : H H and

D(A) = H. Then A = A if and only if there exists a spectral family {E }
= i.e.
an orthogonal projection E (E is a bounded and self-adjoint on H with E2 = E )
satisfying the conditions
1) E E for (that is E H E H)
2) E+0 = E in norm
3) E = 0 and E = I
4) E A = AE on D(A).
The domain D(A) can be defined (or described) as:

Z
D(A) = {f H : 2 d kE f k2 < }.

Moreover, if f D(A) then

Z Z
L2
Af = dE f, kAf k2L2 = 2 d kE f k2 .

Remark. The spectral family {E } is uniquely defined.
68
Remark. If F () is an arbitrary complex-valued function then the operator F (A) can
be defined by Z
F (A) = F ()dE

with domain Z
{f H : |F ()|2 d kE f k2 < }.

Exercise 43. Prove that for any f, g H the real-valued function V () := (E f, g)

is a function of bounded variation.
Let us return to our Schrodinger operator H = + q with q Lp (Rn ), n2 < p
, n 2. We proved that H = H and for f D(H) it follows that
(Hf, f )L2 c0 kf k2L2 (Rn ) , (9.2)
where
D(H) = {f W21 (Rn ) : Hf L2 (Rn )}.
Z Z
H= dE (Hf, f )L2 = d(E f, f ).
c0 c0
Our next problem is to investigate the spectrum of this Schrodinger operator.

Exercise 44. Let A : L2 (R) L2 (R) be such that Af (t) := tf (t), t R for f D(A).
Define D(A) and formulate the spectral theorem in this case.
Exercise 45. Let A be a self-adjoint operator in Hilbert space H. Assume that
z C \ R, that is, Im z 6= 0. Prove that the resolvent Rz = (A zI)1 is a bounded
operator in H with the norm estimate kRz kHH |Im1 z| .
Exercise 46. Let A be a self-adjoint operator in Hilbert space H with the spectral
ti
family {E }= and let f (t) = t+i , t R. Define the Cayley transform by UA :=
R i
+i dE . Prove that
1) kUA uk = kuk for any u H.
2) A = i(I + UA )(I UA )1 .
We will need also the following facts about the spectrum of a self-adjoint operator
in H.
1) A real number belongs to (A) if and only if there is a sequence {fm } D(A)
such that kfm k = 1 and k(A )fm k 0 as m .
2) The essential spectrum ess (A) is the union of
a) c (A);
69
b) the limiting points of p (A);
c) eigenvalues of infinite multiplicity.
3) A real number belongs to ess (A) if and only if there is a sequence {fm } D(A)
such that
a) kfm k = 1
b) fm 0 weakly
c) (A )fm 0 in norm.
4) If 0 (A) is not an isolated point of (A) then 0 ess (A). In other words,
if 0 (A) and 0 / ess (A) then 0 is isolated.
5) If (A) \ ess (A) then is an eigenvalue of A of finite multiplicity.
6) If A is a self-adjoint and K a compact operator then ess (A + K) = ess (A).
Let us consider now Laplacian H0 = in Rn , n 1. Since (f, f )L2 =

kf k2L2 (Rn ) 0 for any f W21 (Rn ), then H0 is a non-negative operator. Moreover,
H0 = H0 with domain D(H0 ) = W22 (Rn ) and this operator has the spectral represen-
tation Z
H0 f = dE f.
0
It follows that (H0 ) [0, +), but actually (H0 ) = [0, +) and even (H0 ) =
c (H0 ) = ess (H0 ) = [0, +). In order to understand this fact it is enough to observe
that for any [0, +) the homogenous equation (H0 )u = 0 has a solution

of the form u(x, k ) = ei( k ,x) , where ( k , k ) = and k Rn . These solutions

u(x, k ) are called generalized eigenfunctions, but u(x, k ) / L2 (Rn ). These solutions

are bounded and correspond to the continuous spectrum of H0 . Thats why u(x, k )
are not eigenfunctions, but generalized eigenfunctions. If we consider the solutions
of the equation (H0 )u = 0 for < 0, then these solutions will be exponentially
increasing at the infinity. It implies that < 0 does not belong to (H0 ).
For the spectral representation of H0 we have two forms:
1) The Neumann spectral representation

Z
f = dE f, f W22 (Rn ),
0
2) and Scattering theory representation

Z Z
f = F 1
(|| fb) = (2)n
2 2 i(,x)
|| e d ei(,y) f (y)dy.
Rn Rn
Exercise 47. Determine the connection between these two representations.
70
There are some important remarks about the resolvent (z)1 for z
/ [0, +).
A consequence of the spectral theorem is that
Z
1
( z) = ( z)1 dE , z C \ [0, +),
0
and for such z the operator ( z)1 is bounded operator in L2 (Rn ). Moreover,
/ [0, +) the operator ( z)1 as a operator-valued function is
with respect to z
a holomorphic function. This fact follows immediately from
Z
(( z)1 )z = ( z)2 dE = ( z)2 .
0
The last integral converges as well as the previous one (even better). Now we are in
the position to formulate a theorem about the spectrum of H = + q.
Theorem 2. Assume that q Lp (Rn ), n2 < p , n 2 and q(x) 0 as |x| +.

Then
1) c (H) (0, +);
2) p (H) [c0 , 0] is of finite multiplicity with only accumulation point at {0} with
c0 from (9.2).
In order to prove this theorem we will prove two lemmas.
Lemma 4. Assume that the potential q(x) satisfies the assumptions of Theorem 2.
Assume in addition that q(x) L2 (Rn ) for n = 2, 3. Then
( z)1 q : L2 (Rn ) L2 (Rn )
is a compact operator for z

/ [0, +).
Proof. Due to our assumptions on the potential q(x) it can be represented as the sum
q(x) = q1 (x) + q2 (x), where q1 Lp (|x| < R) with the same p and q2 0 as |x| .
We may assume (without loss of generality) that q2 is supported in {x Rn : |x| > R}
and that it is a continuous function. Let us consider first the case when n = 2, 3.
If f L2 (Rn ) then q1 f L1 (|x| < R) and ( z)1 (q1 f ) W12 (Rn ) (by Fourier
transform). By the embedding theorem for Sobolev spaces we have that
2 n
( z)1 (q1 f ) W12 (Rn ) W2 2
(Rn ), n = 2, 3
with the norm estimate

( z)1 (q1 f ) ( z)1 (q1 f ) 2 n
2
L2 (Rn ) W
2

c ( z) (q1 f )
1
c kq1 f kL1 (Rn )
W12
c kq1 kL2 (|x|<R) kf kL2 (|x|<R)
71
or

( z)1 q1 c kq1 kL2 ,
L2 (|x|<R)L2 (Rn )
where c may depend only on z.

In the case n 4 and q Lp (|x| < R), p > n2 , we may obtain by Holder inequality
that
2n
q1 f Ls (|x| < R), s > ,
n+4
for f L2 (Rn ) and, therefore, ( z)1 (q1 f ) Ws2 (Rn ). Again by embedding
theorem for Sobolev spaces we have
2n( 1s 21 )
( z)1 (q1 f ) W2 (Rn )
2n
for some s > n+4
, with the norm estimate

( z)1 q1 c kq1 kLp (|x|<R) .
L2 (|x|<R)L2 (Rn )
In order to prove that ( z)1 q1 is a compact operator we approximate it as

follows:
A := ( z)1 q1 , Aj := j (x)A,
where j (x) C0 (Rn ), |j (x)| C and
kA Aj kL2 L2 0, j .
The reason is that ( z)1 q1 is actually an integral operator with a kernel

Kz (x y) which tends to 0 when |x| uniformly with respect to |y| < R (note
that q1 is supported in |y| < R). Thats why we can approximate this kernel Kz by the
functions j C0 (Rn ). But Aj is a compact operator for each j = 1, 2, . . . because
the embedding
W2 (|x| < R) L2 (|x| < R)
is compact for positive . It implies that also A is compact operator.
Next we consider q2 . Since for f (x) L2 (Rn ) we know that (z)1 f W22 (Rn )
then we may conclude that q2 ( z)1 f L2 (|x| > R). Actually
q2 : W22 (Rn ) L2 (|x| > R)
is compact embedding. In order to establish this fact let us consider again j (x)
L
C0 (Rn ), |j (x)| c and j q2 as j . We can state this because C0 = C.
Thats why we required such behavior of q(x) at the infinity (q 0 as |x| +). If
we denote A := q2 ( z)1 and Aj := j ( z)1 then we obtain

kA Aj kL2 L2 sup |j q2 | ( z)1 c sup |j q2 | 0, j +.
x L2 L2 x
(9.3)
72
2
But we know that W2,comp L2comp is compact embedding. This implies (together with
(9.3)) that A is compact operator. Thus lemma is proved, because
( z)1 q2 = (q2 ( z)1 ) .
Lemma 5. Let Q be an open and connected set in C. Let A(z) be compact, operator
valued and holomorphic function in Q and in L2 (Rn ). If (I + A(z0 ))1 exists for some
z0 Q then (I + A(z))1 exists in all of Q except for finitely many points from Q with
only possible accumulation points on Q.
Proof. We will prove this lemma for the concrete operator A(z) := ( z)1 q(x).
Lemma 4 shows us that A(z) is compact operator for z / [0, +). The remarks about
Rz = ( z)1 show us that A(z) is a holomorphic function in C \ [0, +). Also
we can prove that (I + ( z)1 q)1 exists for any z C \ R or for real z < c0 ,
where + q c0 . Indeed, if z C with Im z 6= 0 then (I + ( z)1 q)u = 0
or ( z)u = qu or (u, u) + z(u, u) = (qu, u). It implies for z, Im z 6= 0, that
(u, u) = 0 if and only if u = 0. In the real case z < c0 we have that equality
(I + ( z)1 q)u = 0 implies
(( + q)u, u) z(u, u) = 0.
It follows that
(c0 z) kuk2L2 0
or u = 0. These remarks show us that in C \ [0, +) our operator I + ( z)1 q
may be non-invertible only on [c0 , 0).
Let us consider an open and connected set Q in C \ [0, +) such that [c0 , 0) Q,
see Figure 3. It is easily seen that there exists z0 Q such that (I + ( z0 )1 q)1
exists also. It is not so difficult to show that there exists > 0 such that (I + (
z)1 q)1 exists in U (z0 ). Indeed, let us choose > 0 such that
1
kA(z) A(z0 )kL2 L2 < (9.4)
k(I + A(z0 ))1 kL2 L2
for all z such that |z z0 | < . Then
(I + A(z))1 = (I + A(z0 ))1 (I + B)1 ,
where B := (A(z) A(z0 ))(I + A(z0 ))1 . But kBk < 1 due to (9.4) and then
(I + B)1 = I B + B 2 + + (1)n B n +
exists in the strong topology from L2 to L2 . Thats why we may conclude that I +A(z)
may be non-invertible only for finitely many points in Q. This fact follows from the
holomorphicity of A(z) with respect to z by analogy with the theorem about zeros of the
73
Im z
Q
c0 Re z
Figure 3: The set Q.
holomorphic function in complex analysis. Moreover, since A(z) is compact operator

then by Fredholms alternative Ker (I + A(z)) has finite dimension. Thats why we
may conclude that (I + ( z)1 q)1 does not exist only in the finite numbers of
the points (at most) on [c0 , ] for any > 0 and these points are finite multiplicity.
This finishes the proof.
Let us return to the proof of Theorem 2.
Proof of Theorem 2. Let be a positive number and + c0 > 0 (H c0 I). Let us
consider for such the second resolvent equation
(H + )1 = (H0 + )1 (H + )1 q (H0 + )1 , (9.5)
where H0 = and H = + q(x). It follows from Lemma 4 that q (H0 + )1

is compact operator in L2 (Rn ). It means that (H + )1 is a compact perturbation of
(H0 + )1 . Hence, by fact 6) above we have
ess ((H + )1 ) = ess ((H0 + )1 ).
But ess ((H0 + )1 ) = [0, 1 ] = c ((H0 + )1 ). Thats why we may conclude that
ess (H + ) = [, +].
Outside of this set we have only points of the discrete spectrum with one possible
accumulation point at . This statement is a simple corollary from Lemma 5. Moreover,
these points of discrete spectrum are situated on [ c0 , ) and they are of finite
multiplicity. Hence the discrete spectrum d (H) of H belongs to [c0 , 0) with only
74
one possible accumulation point at {0}. And (0, +) is continuous part of (H).
There is only one problem. As a matter of fact, Weyls theorem (fact 6)) states
that the operators H and H0 dont have the same spectrum but the same essential
spectrum. Thats why on (0, +) there can be eigenvalues of infinite multiplicity
(see the definition of ess ). In order to eliminate such possibility and to prove that
0 c (H) and d (H) is finite let us assume additionally that our potential q(x) has a
special behavior at the infinity:
|q(x)| c|x| , |x| +,
where > 2. In that case we can prove that on the interval [c0 , 0) the operator H has
at most finitely many points of discrete spectrum. And we prove also that 0 c (H).
Assume on the contrary that H has infinitely many points of discrete spectrum or
one of them has an infinite multiplicity. It means that in D(H) there exists the infinite
dimensional space of the functions {u} which satisfy the equation
( + q)u = u, c0 0.
It follows that
Z Z
2 + 2
(|u(x)| + q (x)|u(x)| )dx q (x)|u(x)|2 dx,
Rn Rn
where q + and q are the positive and negative parts of the potential q(x), respec-
tively. Let us Rconsider an infinite sequence of functions {u(x)} which are orthogonal
2
in the metric
R Rn q (x)|u(x)| dx. Thats why this sequenceR
is uniformly bounded in
the metric Rn (|u| + |q||u| )dx and hence, in the metric Rn (|u|2 + |u|2 )dx. But for
2 2
every eigenfunction u(x) of the operator H with eigenvalue [c0 , 0] the following
inequality holds (see [1]):
Z
|u(x)| c|| |u(y)|dy,
|xy|1
where c does not depend on x. It follows from this inequality that

a) = 0 is not an eigenvalue.
b) this orthogonal sequence is uniformly bounded in every fixed ball.

Lemma 6. Denote by U the set of functions u(x) D(H) which are uniformly bounded
in every fixed ball in Rn . Then U is a precompact set in the metric
Z
|q||u|2 dx
Rn
if it is a bounded set in the metric

Z
(|u|2 + |u|2 )dx.
Rn
75
Proof. Let {uk (x)}
k=1 U be an arbitrary sequence which is bounded in the second
metric. Then for u(x) := uk (x) um (x) we have for r large enough that
Z Z Z
2 |u(x)|2
|q(x)||u(x)| dx c dx + |q(x)||u(x)|2 dx
Rn |x|>r |x| |x|r,|q(x)|A
Z
+ |q(x)||u(x)|2 dx := I0 + I1 + I2 .
|x|r,|q(x)|>A
For n 3 (for n = 2 we need some changes) and > 2 we get

Z Z
2 2 2 2
I0 cr |x| |u(x)| dx cr |u(x)|2 dx, u W21 (Rn ).
|x|>r Rn
Due to the uniform boundedness of U in every ball we may conclude that

Z
I2 c |q(x)|dx 0
|x|r,|q(x)|>A
as A + uniformly on U with fixed r. Since the embedding W21 L2 for every

ball is compact the boundedness of the sequence in the second metric implies the
precompactness in L2 for every ball. Thats why we have
Z
I1 A |u(x)|2 dx 0, m, k
|x|r
with r and A fixed. After limiting processes, these inequalities for I0 , I1 and I2 show
that Z
|q(x)||u(x)|2 dx 0, m, k .
Rn
Thus lemma is proved.
Let us return to the proof R
of 1). By Lemma 6 we obtain that our sequence (which is
orthogonal in the metric q |u|2 dx) is a Cauchy sequence in the first metric. But this
fact contradicts its orthogonality. Thus 1) is proved.
2) Let us discuss (briefly) the situation with a positive eigenvalue at the continuous
spectrum. If we consider the homogeneous equation
[I + ( k 2 i0)1 q]f = 0, k 2 > 0,
in the space C(Rn ) then by Greens formula one can show (see, [2] or [3]) that the
n1
solution f (x) of this equation behaves at the infinity as o(|x| 2 ). Thats why may
conclude (T. Kato) that f (x) 0 outside some ball in Rn . By the unique continuation
principle for the Schrodinger operator (see, for example, [4]) it follows that f 0 in
the whole Rn .
Let us consider now the spectral representation of the Schrodinger operator H =
+q(x), with q(x) as in Theorem 2 with the behavior O(|x| ), > 2 at the infinity.
For any f D(H),
Z
n 2
Z
M
X
Hf (x) = (2) k u(x, k )d k f (y)u(y, k )dy + j fj uj (x),
Rn Rn j=1
76

where u(x, k ) are the solutions of the equation Hu = k 2 u, uj (x) are the ortonor-
mal eigenfunctions corresponding to the negative eigenvalues j , taking into account

multiplicity of j and fj = (f, uj )L2 (Rn ) . The functions u(x, k ) are called generalized
eigenfunctions. In the case when q 0 the generalized eigenfunctions have the form

u(x, k ) = ei(x, k ) . This fact follows by Fourier transform. Indeed,
( k 2 )u = 0
if and only if
(||2 k 2 )ub = 0
or X

ub = c () ( k ),

since
||2 = k 2
if and only if

k = 0.
Hence

X

u(x, k ) = c F 1 ( () ( k ))(x)

X
X

= c ei(x, k ) F 1 ( () ())(x) = c ei(x, k ) x .

But u(x, k ) must be bounded. Thats why u(x, k ) = c0 ei(x, k ) . We choose c0 = 1. If
we have the Schrodinger operator H = + q with q 6= 0, then it is natural to look

for the solutions of Hu = k 2 u of the form u(x, k ) = ei(x, k ) + usc (x, k ). Due to this
representation we have

( k 2 )(ei(x, k ) + usc ) = qu
or
( k 2 )usc = qu.
In order to find usc let us recall that from Chapter 8 we know the fundamental solution
of the operator k 2 . Thats why

Z
u(x, k) = e i(x, k )
G+
k (|x y|)q(y)u(y)dy,
Rn
where
n2
i |k| 2
(1)
G+
k (|x|) = H n2 (|k||x|)
4 2|x| 2
is the fundamental solution for the operator k 2 . This equation is called the
Lippmann-Schwinger integral equation.
77
In order to investigate this equation we will investigate the integral operator (
k 2 i0)1 in some weighted spaces. As a matter of fact, (k 2 i0)1 is not bounded
in L2 (Rn ) but it is bounded from L2 (Rn ) to L2 (Rn )) for > 12 with the norm estimate
c
k( k 2 i0)1 kL2 L2 .
|k|
This fact was proved by S. Agmon in [5]. We will prove this estimate for n 3 in
Chapter 10.
78
10 Estimates for Laplacian and Hamiltonian
Let us recall Agmons (2, 2)-estimate for Laplacian:
c
k( k 2 i0)1 kL2 L2 ,
|k|
where > 12 . In fact, this estimate allows us to consider the Hamiltonian with L loc -
potentials only (if we want to preserve (2, 2)-estimates). But we would like to consider
Hamiltonian with Lploc -potentials. Thats why we need to prove (p, q)-estimates. In
this case we follow A. Ruiz.
1 1
We have proved in Example 4.9 that the limit lim xi := xi0 exists in the sense
0
of tempered distributions and
1 1
= p.v. + i(x)
x i0 x
i.e. Z
1 (x)
h , i = lim dx + i(0).
x i0 +0 |x|> x
In Example 4.3 we have considered the simple layer
Z
hT, i := a()()d ,

where is a hypersurface of dimension n 1 in Rn and a() is a density. These

examples can be extended as follows. If H : Rn R and |H| =
6 0 at any point where
H() = 0 then we can define the distribution
1
(H() i0)1 := lim
+0 H() i
in S (R) and we can also prove that

1
(H() i0)1 = p.v. + i(H() = 0),
H()
where (H() = 0) is defined as follows:
Z
h(H), i= ()d , S(Rn ).
H()=0
The equality H() = 0 defines an n 1 dimensional hypersurface and is any (n 1)-

dH
form such that d |H| = d (in local coordinates).
1
(H) = (H)

for any function which does not vanish at any point where H() = 0.
79

1 H
Due to Exercise 48 we may conclude that (H) = |H| |H| if |H| =
6 0 for
H = 0.
Let us consider now H() := ||2 + k 2 , k > 0. Then H() = 0 or || = k is a
sphere and H() = 2 and |H()| = 2k at any point on this sphere. If we change
the variables then we obtain
Z Z
1
h(H), i= ()d = (k)d.
H()=0 2k Sn1
We know that ( k 2 i0)1 f can be represented as
Z
2
( k i0) f = 1
G+
k (|x y|)f (y)dy,
Rn
n2
|k| (1)
where G+
k (|x|) =
i
4 2|x|
2
H n2 (|k||x|). On the other hand we can write
2
Z
2 1 1 2 1 n fb()ei(x,) d
( k i0) f = F (F [( k i0) f ]) = (2) 2
Rn ||2 k 2 i0
Z
n 1
= (2) 2 p.v. fb()ei(x,) d
Rn ||2 k2
n Z
i(2) 2
+ (H)fb()ei(x,) d
2k Rn
Z Z
n fb()ei(x,) d i
= (2) 2 p.v. 2 2
+ n fb(k)eik(x,) d
Rn || k 2k(2) 2 Sn1
Z
n fb()ei(x,) d
= (2) 2 p.v.
Rn ||2 k 2
Z Z
i
+ f (y)dy eik(,xy) d.
2k(2)n Rn Sn1
Our aim is to prove the following result.

Theorem 1. Let k > 0 and n2 p1 p1 n+1 2
for n 3 and 1 > p1 p1 23 for n = 2,
where p1 + p1 = 1. Then there exists a constant C independent of k and f such that

1
2 1 n p1 2
k( k i0) f kLp (Rn ) Ck p
kf kLp (Rn ) .
instead of ( k 2 i0)1 .
Remark. In what follows we will use G k
Proof. First we prove that if the claim holds for k = 1 then it holds for any k > 0. So,
let us assume that
kG f k p n Ckf k p n .
1 L (R ) L (R )
n
Denote T f := f (x), > 0. It is clear that kT f kLp (Rn ) = p kf kLp (Rn ) . It is not so
= k 2 T G
difficult to show that G
k k 1 T 1 . Indeed, since
k
Z Z
f = (2)n ei(y,) f (x y)ddy
G k
Rn Rn ||2 k 2 i0
80
we get
Z Z
T 1 f = (2)n ei(y,) f ( xy
k
)ddy
G 1 2
.
k Rn Rn || 1 i0
It follows that
Z Z
T 1 f = (2)n ei(y,) f (x ky )ddy
Tk G 1 , (z := y/k, := k)
k Rn Rn ||2 1 i0
Z Z
ei(z,) k n f (x z)k n dzd
= (2)n ||2
Rn Rn
k2
1 i0
Z Z i(z,)
e f (x z)dzd
= (2)n k 2 .
Rn Rn ||2 k 2 i0
This proves that
f T G
k2G
k k 1 T 1 f.
k
We use it to get
f k p = k 2 kT G
2 pn T 1 f k p
kG k L k 1 T 1 f kLp = k k kG 1 L
k k
n
2 pn 2 pn 1 p
n 1
p1 2
Ck kT 1 f kLp = Ck kf kLp = Ck p
kf kLp .
k k
Thats why it is enough to prove this theorem for k = 1.

Lemma 1. Let (x) S(Rn ), 0 < < 1 and () = n
. Let us denote

1
P () := p.v. 2
().
|| 1
Then
c
|P ()| .

Proof. For P we have the following representation:
Z Z Z
( )
P = p.v. + + d = I1 + I2 + I3 .
1||1+ ||<1 ||>1+ ||2 1
1
The integrals
I2 and I3 can be easily bounded by kkL1 because || < 1 implies
1 1
that ||2 1 = 1||2 < and || > 1 + implies that ||2 1 = ||211 < 1 . By the
1 1
definition of p.v. we have

Z Z Z Z
( ) 1 1+ rn1
I1 = lim d = lim + (r) ddr.
+0 <|1|||< ||2 1 +0 1 1+ Sn1 r2 1
Replacing r with 2 r in the latter integral we obtain
Z 1 F (r, )
I1 = lim dr,
+0 1 r1
81
where
Z
rn1 (2 r)n1
F (r, ) = ( r) ( (2 r)) d.
Sn1 r+1 3r
If we observe that F (1, ) = 0 then we get by the mean value theorem (Lagrange
formulae) that
Z Z
1 F (r, ) 1 F (r, ) F (1, ) F

1 r 1
dr =

1
dr
( ) sup

(r, )

r1 1<r<1 r

F
sup
.
1<r<1 r
But
Z
F rn1
= ( r)d
r r+1 Sn1
Z Z
rn1 (2 r)n1
( ( r))d ( (2 r))d
r + 1 Sn1 3r Sn1
Z
(2 r)n1
( ( (2 r)))d = 1 + 2 + 3 + 4 .
3r Sn1
By the proof of Lemma 2 below we get |1 | c1 1 and |3 | c3 1 , where the

constants c1 and c3 depend on . The second integral 2 can be estimated as (see
Lemma 2)
n Z
1
X rn1
j ( r)d c2 2 .
j=1 r + 1 Sn1 x j
The same estimate holds for 4 . Thus, Lemma 1 is proved.

Lemma 2. Let us assume that f L (Sn1 ) and S(Rn ). Then
Z

( )f ()d C1 .
Sn1 L (Rn )
S
Proof. We can reduce the proof to compactly supported , since C0 =S. Let us take
P P 1
a C0 -partition of unity in Rn such that j=0 j () = 1 or even j=0 j = 1, where
0 is supported in || < 1 and j = (2j ) for j = 1, 2, 3, . . . , with supported in
the annulus 1/2 < || < 2. Thats why we may write
Z Z
X n
( )f ()d = j f ()d.
Sn1 j=0 S
n1

For j = 1, 2, 3, . . . , the function j

is supported in the annulus 2j1
| | 2j+1 . Since is rapidly decreasing we have that, in this annulus,

CM

,
(1 + 2j )M
82
for all M N. Hence
Z

n (2j+1 )n1 n j n1
1 (2 )

Sn1
(j ) f ()d

CM CM .
(1 + 2j )M (1 + 2j )M
Taking M large enough the sum in j converges to C1 . To end the proof of Lemma
2 notice that the term for j = 0 satisfes this inequality trivially.
Exercise 49. Prove that ()1 : L2 (R3 ) L2 (R3 ) for > 1.
f in the form
Let us return to the proof of Theorem 1. We can rewrite G 1
Z
f = Cp.v. fb()ei(x,) d
G 1 + I1 f,
Rn ||2 1
where Z
I1 f = C fb()ei(,x) d.
Sn1
P
Let us take a partition of unity j=0 j (x) = 1 such that supp 0 {|x| < 1} and
supp j {2j1 < |x| < 2j+1 }, where j = (2j x) with a fixed function S.
Denote j := j G+ +
1 and Kj f := j f , where G1 is the kernel of the integral operator
. Using the estimates of the Hankel function H (1) (|x|) for |x| < 2 we obtain
G 1 n2
2
|0 | C|x|2n , n3
and
|0 | C(|log |x|| + 1), n = 2.
1 1
Exercise 50 (Sobolev inequality). Let 0 < < n, 1 K, |x| > .
From Sobolev inequality for = 2 we may conclude that the operator K0 is bounded

from Lp (Rn ) Lp (Rn ) for the range n2 p1 p1 0 if n 3 and 1 > p1 p1 0 if
n = 2. From Lemma 1 and 2 with = 21j we can obtain that
kF (j )k = k(||2 1 i0)1 j k C 2j .
This inequality leads to

kKj kL2 L2 C 2j ,
83
because
fbk 2 k
kKj f kL2 = kF (j f )kL2 = Ck k kfbk 2 C 2j kf k 2 .
j L j L L L
On the other hand, due to the estimate of the fundamental solution at infinity we can
n1
obtain that |j (x)| C 2j 2 and
n1
kKj kL1 L C 2j 2 .
We have used here two facts:

(1)
H n2 (|x|)
C
1 , |x| > 1
2 |x| 2
and supp j (x) {x : 2j1 < |x| < 2j+1 }. Interpolating these estimates we obtain the
self-dual estimates:
kKj kLp Lp C(2j )2(1 p ) 2 ( p 1) .
1 n1 2
For convergence of this series we need the condition 2(1 p1 ) n12

( p2 1) < 0 or
1
p
p1 > n+1
2
. If we want to get the sharper inequality p1 p1 n+1
2
we have to use
Steins theorem about interpolation. Thus, Theorem 1 is proved.
It follows from Theorem 1 that if we consider the values of p from the interval
2n 2n + 2
p , n3
n+2 n+3
1 < p 6/5, n = 2,
then we have the self-dual estimate
k p p C
kG k L L .
1
2n p1
|k| p
But we would like to extend the estimates for G for 2n p 2, n 3, and

k n+2
1 < p 2, n = 2. In order to do so we use interpolation of the Agmons estimate and
the latter estimate for p = 2n+2
n+3
. This process leads to the estimate
k p p C
kG k L L ,
1(n1)( p1 21 )

|k|

2n+2 1 1 1
where n+3
 2
(n + 1) 2p
4
.
Theorem 2. Assume that the potential q(x) belongs to Lp (Rn ), n 2, with n2 1 n+1
2p
for n+1
2
< p +. Then for all k 6= 0
the limit
:= lim (H k 2 i)1
G q
+0
84
2p 2p
exists in the uniform operator topology from L p+1

(Rn ) to Lp1 n
(R ) with the norm
2 2
estimate
kG f k 2p C|k| kf k 2p
q
p1 p+1
L L
2 2
n n n+1
for large k with p and as above and with = 2 2
for 2
<p 2
and = 1 n1
2p
for n+1
2
< p .
c with the kernel
Proof. Let us prove first that the integral operator K
1
K(x, y) := |q| 2 (x)G+
k (|x y|)q 1 (y), 2
1
where q 1 (y) = |q(y)| sgn q(y) maps from L2 (Rn ) to L2 (Rn ) with the same norm es-
2
2
1
timate as in Theorem 2. Indeed, if f L2 (Rn ) and q Lp (Rn ) then |q| 2 L2p n
(R )
2
2p
1 p+1 n
and, therefore, f |q| L
2 (R ). Applying Theorem 1 we obtain
2
(|q| 2 f )k
kG
1
2p C|k| kf k 2p ,
k
p1
L L p+1

2 2
(q 1 f )
where is as in Theorem 2. Then, by Holders inequality we have that |q| 2 G
1
k
2
2 n
L (R ) as asserted.
. This operator satisfies the resolvent equation
Let us consider now the operator G q
=G
G G
qG
q k k q
which follows easily from (H k 2 )G = I. Denote by G and G the integral operators

q l r
1
having the kernels G+k (|x y|)q +
1 (y) and |q(x)| 2 G (|x y|), respectively. Then one can
k
2
show that
=G
G G (1 + K)
c 1 G
q k l r
2p 2p
: L p+1
c : L2 L2 , G
for large k. Since K : L2 L p1
L2 and G
r
2
l 2 then we may
conclude that Theorem 2 is proved.
There are some corollaries from these two theorems.
Corollary 1. Assume that the potential q(x) belongs to Lp (Rn ) L1 (Rn ) with p and
as in Theorem 2. Then for > 0 large enough there is a unique solution u of the
equation
( + q )u = 0
of the form

u(x, k ) = ei( k ,x) + usc (x, k ),

where k , x Rn , ( k , k ) = and
1 C 1
k|q| 2 usc kL2 (Rn ) kqkL1 (Rn )
2
2
with as in Theorem 2.
85
Proof. Let us rewrite the Lippmann-Schwinger equation
Z

u(x, k ) = ei( k ,x) G+
(|x y|)q(y)u(y, k )dy
Rn
in the form
Z
1
1
1

|q(x)| 2 u(x, k ) = |q(x)| 2 ei( k ,x) K(x, y)|q(y)| 2 u(y, k )dy
Rn
or
c
v = v0 Kv,
1
1

c is as above. Since v L2
where v(x) = |q(x)| 2 u(x, k ), v0 (x) = |q(x)| 2 ei( k ,x) and K 0
if and only if q L1 and Kc : L2 L2 with a good norm estimate, we may conclude
that
v = v0 + Kvc +K c2 v + . . .
0 0

for | k | large enough and
c 2 2 kv k 2
kv v0 kL2 CkKkL L 0 L
i.e.
1 C 1
k|q(x)| 2 usc kL2 kqkL1 .
2
/2
Corollary 2. Let v be the outgoing solution of the inhomogeneous Schrodinger equation
(H k 2 )v = f
i.e.
v = (H k 2 i0)1 f,
where f S(Rn ). Then the following representation holds:
(f q G
v(x) = G (f ))(x).
k q
Moreover, for |x| and fixed positive k,

n3 !
eik|x| k 2
1
v(x, k) = Cn n1 Af (k, ) + o n1 ,
|x| 2 |x| 2
x
where = |x|
and the function Af is defined by
Z
Af (k, ) :=

(f ))dy.
eik( ,y) (f (y) q(y)G q
Rn
86
Proof. The first representation follows immediately from the definition of G . Indeed,
q

since v = Gq f then Gk f = v + Gk qv or v = Gk f Gk qv = Gk (f q Gq f ).
In order to prove the asymptotic behavior for v let us assume that q and f have
compact support, say in the ball {x : |x| R}. We will use the following asymptotic
behavior of G+k (|x|) :
1) k|x| < 1 :
a) G+
k (|x|) C|x|
2n
, n 3,
b) G+
k (|x|) C log(k|x|), n = 2.
2) k|x| > 1 :
n3
G+
k (|x|) C
k 2
n1 eik|x| , n 2.
|x| 2
Since k is fixed, |y| R and |x| + we may assume that k|x y| > 1 for x large
enough. Thats why
n3 Z
eik|x| k 2
f )dy
v(x) = C n1 eik(|xy||x|) (f q G q
|x| 2 |y|R
Z !
1 f )dy = I + I .
+ o n1 (f q G q 1 2
|y|R |x y| 2
It is clear that for I2 the following is true

Z ! !
1 f (y))dy = o 1
I2 = o n1 (f (y) q(y)G q n1 ,
|x| 2 |y|R |x| 2
f is an integrable function. Next, let us note that

because f q G q

|x y|2 |x|2 y 2 2(x, y) x 1
|x y| |x| = = = ,y + O , |x| +.
|x y| + |x| |x y| + |x| |x| |x|
Thats why we can rewrite the integral appearing in I1 as follows:

Z Z
ik( |x|
x
,y )+O( |x|
1
) f )dy
f )dy
e (f q G q = eik( ,y) (f q G q
|y|R |y|R
Z
1
f )dy
+ O eik( ,y) (f q G q
|x| |y|R
Z
ik( ,y) f )dy + O 1
= e (f q G q ,
|y|R |x|
x
where = |x| Sn1 . Thus, Corollary 2 is proved when q and f have compact support.
The proof in the general case is much more difficult and is therefore omitted.
87
Remark. Hint for the general case: The integral over Rn might be divided in two parts:
|y| < |x| and |y| > |x| , where > 0 is chosen appropriately.
Lemma 3 (Optical lemma). For the function Af (k, ) the following equality holds:
Z Z
21
|Af (k, )| d = 2 n2 Im (f v)dx,
Sn1 C k Rn
where C is the constant from the asymptotic representation of v = (H k 2 i0)1 f .

Proof. Let be a smooth real-valued function on [0, +) such that 0 1 and
(r) = 1 for 0 r < 1 and (r) = 0 for r 2. We set m (r) = mr . Multiplying f
by vm (|x|), integrating over Rn and taking imaginary parts leads to
Z Z
Im f (x)m (|x|)v(x)dx = Im (v)m (|x|)v(x)dx.
Rn Rn
R
As m tends to infinity, the left-hand side converges to Im Rn f (x)v(x)dx. To get the
desired limit for the right-hand side, we integrate by parts and obtain
Z Z
x
Im (v)m (|x|)v(x)dx = Im vm (|x|)v(x)dx
R n R |x|
n
Z
= Im ( v ikv)v(x)m (|x|) + ikm (|x|)|v|2 dx
n
ZR
= Im ( v ikv)v(x)m (x)dx
Z Rn
+ k m (|x|)|v(x)|2 dx = I1 + I2 .
Rn
Since v = (H k 2 i0)1 f then using the asymptotical representation we may conclude

that v satisfies the Sommerfeld radiation condition

v 1
ikv = o n1 , r = |x|
r r 2
at the infinity. Thats why I1 0 as m . By Corollary 2 the second term I2 is
equal to
Z Z
k n3
k m (|x|)|v(x)|2 dx = k m (|x|) C 2
n1
|Af (k, )|2 dx
Rn R n |x|
Z
1
+ k m (|x|)o dx
Rn |x|n1
Z Z 2m n1
2 n2 2 r
= C k |Af (k, )| d n1
m (r)dr
m r
Sn1
Z Z 2m
1
+ k d rn1 o n1 m (r)dr
Sn1 m r
Z Z 2m
= C 2 k n2 |Af (k, )|2 d m (r)dr + om (1)
Sn1 m
Z
2 n2 2
= C k |Af (k, )| d [(2) (1)] + om (1)
Sn1
Z
= C 2 k n2 |Af (k, )|2 d + om (1).
Sn1
88
Letting m we obtain
Z Z
Im f (x)v(x)dx = C 2 k n2 |Af (k, )|2 d .
Rn Sn1
Thus, Lemma 3 is proved.
Exercise 51. Let n = 2 or n = 3. Assume that q Lp (Rn ) L1 (Rn ) with 1 < p

if n = 2 and 3 < p if n = 3. Prove that the generalized eigenfunctions u(x, k )

are uniformly bounded with respect to x Rn and | k | large enough.
We will obtain very important corollaries from Optical lemma. Let Aq (k) denote
the linear mapping that takes the inhomogeneity f to the corresponding scattering
amplitude
Aq (k) : f (x) Af (k, ).
Lemma 4. Let the potential q(x) satisfy the conditions from Theorem 2. Then Aq
2p
is a well-defined bounded operator from L p+1

(Rn ) to L2 (Sn1 ) with the operator norm
2
estimate
C
kAq k p+12p n2 ,
L L2 |k| 2 + 2
2
where p, and are as in Theorem 2.
Proof. By Lemma 3 and the definition of Aq f we have that

Z Z
1
kAq f k2L2 (Sn1 ) = |Af (k, )|2 d = Im (f v)dx
Sn1 C 2 |k|n2 Rn
1
kvk p1
2p kf k p+1
2p .
C 2 |k|n2 L (Rn ) L (Rn )
2 2
f, we obtain from Theorem 2 that

Further, since v = G q
C
kAq f k2L2 (Sn1 ) n2 |k| kf k2 2p .
|k| L p+1
(Rn )
2
Thus, Lemma 4 is proved.

Let us denote by A0 (k) the operator Aq (k) which corresponds to the potential q 0
i.e. Z

A0 f ( ) = eik( ,y) f (y)dy.
Rn
It is not so difficult to see that
Z
Aq f ( ) = Af (k, ) = f (y)u(y, k, )dy,
Rn
89
where u(, k, ) is the solution of Lippmann-Schwinger equation. Indeed, by Corollary
2 of Theorem 2 we have
Z
Af (k, ) =

(f ))dy = ((I q G
eik( ,y) (f (y) q(y)G )f, eik( ,y) ) 2 n
q q L (R )
Rn
Z
(q))eik( ,) ) 2 n =
= (f, (I G

(q))(eik( ,) )(y)dy
f (y)(I G
q L (R ) q
Rn
Z
= f (y)u(y, k, )dy,
Rn
is a self-adjoint operator.
since G q
Let us prove now that
(q))(eik( ,) )(y)
u(y, k, ) := (I G

q
is the solution of Lippmann-Schwinger equation. Indeed,
(H k 2 )u = (H k 2 )(eik( ,y) ) (H k 2 )G

(q) (eik( ,) )(y)
q
2 ik( ,y) ik( ,y)
= ( k )e + qe qeik( ,y) = 0,

= I. It means that this u(y, k, ) is the
since ( k 2 )eik( ,y) = 0 and (H k 2 )G q
2
solution of the equation (H k )u = 0.
Remark. Let us consider the Lippmann-Schwinger equation
Z
u(x, k, ) = eik(x,) G+
k (|x y|)q(y)u(y, k, )dy.
Rn
Then for fixed k > 0 and |x| the solution u(x, k, ) admits the asymptotical
representation
n3 !
ik(x,) eik|x| k 2
1
u(x, k, ) = e + Cn n1 A(k, , ) + o n1 ,
|x| 2 |x| 2
x
where = |x|
and the function A(k, , ) is called the scattering amplitude and has
the form Z

A(k, , ) = eik( ,y) q(y)u(y, k, )dy.
Rn
For k < 0 we set
A(k, , ) = A(k, , ), u(x, k, ) = u(x, k, ).
Proof. If (H k 2 )u = 0 and u = eik(,x) + usc (x, k, ) then usc (x, k, ) satisfies the
equation
(H k 2 )usc = qeik(,x) .
Thats why we may apply Corollary 2 of Theorem 2 with v := usc and f := qeik(,x)
and obtain n3 !
eik|x| k 2 1
usc (x, k, ) = Cn n1 Af (k, ) + o n1 ,
|x| 2 |x| 2
90
where
Z
Af (k, ) =

(qeik(,) )(y))dy
eik( ,y) (qeik(,y) + q G q
Rn
Z
=

(qeik(,) )dy.
eik( ,y) q(y)(eik(,y) G q
Rn
(qeik(,) )(y) is solution of the equation (H k 2 )u =

But we have proved that eik(,y) G q
0. Thats why we may conclude that
Z

Af (k, ) = eik( ,y) q(y)u(y, k, )dy := A(k, , ).
Rn
Thus, this remark is proved.

Now let 0 (k) and (k) be the operators defined for f L2 (Sn1 ) as
Z
1
(0 (k)f )(x) := |q(x)| 2 eik(x,) f ()d (10.1)
Sn1
and Z
1
((k)f )(x) := |q(x)| 2 u(x, k, )f ()d. (10.2)
Sn1
Lemma 5. The operators 0 (k) and (k) are bounded from L2 (Sn1 ) to L2 (Rn ) with
the norm estimates
C
k0 (k)k, k(k)k + n2 , k > 0,
k2 2
where is as in Theorem 2.
Proof. Let us prove that

1
(0 (k)f )(x) = |q(x)| 2 (A0 f )(x) (10.3)
and
1
((k)f )(x) = |q(x)| 2 (Aq f )(x), (10.4)
where A0 and Aq are the adjoint operators for A0 and Aq , respectively. Indeed, if
f L2 (Sn1 ) and g L2 (Rn ) then
Z Z Z
f ()(A0 g)()d = f ()d eik(,y) g(y)dy
Sn1 Sn1 Rn
Z Z
= g(y)dy eik(,y) f ()d
Rn Sn1
Z Z
ik(,y)
= e f ()d g(y)dy.
Rn Sn1
It means that Z
A0 f (y) = eik(,y) f ()d
Sn1
91
and (10.3) is immediate. Similarly for (10.4). Since (see Lemma 4)
C
kA0 k, kAq k 2p
+ n2
L p+1
L2 (Sn1 ) k 2 2
2
we have that
C
kA0 k, kAq k 2p
+ n2
.
p1
L2 (Sn1 )L n
(R ) k 2 2
2
The proof is finished by

1 1
k0 (k)f kL2 (Rn ) = k|q| 2 (A0 f )kL2 (Rn ) kqkL2 p (Rn ) kA0 f k 2p
p1
L n
(R )
2
C 1
n2 kqkLp (Rn ) kf kL2 (Sn1 ) ,
2
k2+ 2
where we have made use of Holder inequality in the first estimate. It is clear that the
same is true for (k).
92
11 Some inverse problems for the Schrodinger op-
erator
The classical inverse scattering problem is to reconstruct the potential q(x) from the
knowledge of the far field data (scattering amplitude) A(k, , ), when k, and are
restricted to some given set.
If q L1 (Rn ) then q(y)u(y, k, ) L1 (Rn ) uniformly with respect to Sn1 due
to
1
q(y)u(y, k, ) = q(y)(eik(,y) + usc (y, k, )) = q(y)eik(,y) + |q| 2 q 1 usc (y, k, )
| {z } |{z} | 2 {z }
L1 L2 L2
and Holders inequality. Thats why we may conclude that the scattering amplitude
A(k, , ) is well-defined and continuous. Also the following representation holds:
Z Z
ik( ,y) ik(,y)
A(k, , ) = e q(y)(e + usc )dy = eik( ,y) q(y)dy + R(k, , )
Rn Rn
= (2)n/2 (F q)(k( )) + R(k, , ),
where R(k, , ) 0 as k + uniformly with respect to and . This fact implies
that
A(k, , ) (2)n/2 (F q)(k( ))
or
q(x) (2)n/2 F 1 (A(k, , ))(x),
where the inverse Fourier transform must be understood in some special sense.
Let us introduce the cylinders M0 = R Sn1 and M = M0 Sn1 , and the
measures and on M0 and M , respectively, as
1
d (k, ) = |k|n1 dk| |2 d ,
4
1
d(k, , ) = n1 dd (k, )
|S |
n/2
where |Sn1 | = 2
( n )
is the area of the unit sphere Sn1 and d and d denote the
2
usual Lebesgue measures on Sn1 . We shall define the inverse Fourier transform on
M0 and M as Z
1 1
(FM 1 )(x) = n/2
eik( ,x) 1 (k, )d ,
0
(2) M0
Z
1 1 ik( ,x)
(FM 2 )(x) = e 2 (k, , )d.
(2)n/2 M
If we write = k( ) then k and are obtained by
|| b ,
b
k= b
, = 2(, ) b = . (11.1)
2(, ) ||
93
Exercise 52. Let u (k, ) be the coordinate mapping M0 Rn given as
u (k, ) = k( ),
where is considered as a fixed parameter. Prove that
1) the formulas (11.1) for k and hold
2) the following is true:

Z Z

u (k, )d (k, ) = (x)dx
M0 Rn
if S is even and
Z Z

u (k, )d(k, , ) = (x)dx
M Rn
if S.
3) in addition:
1
FM 0
( u ) = F 1
if S is even and
1
FM ( u ) = F 1
if S. Here F 1 is the usual inverse Fourier transform in Rn .
1) A(k, , ) = A(k, , )
2) A(k, , ) = A(k, , ).
n
The approximation q(x) (2) 2 F 1 (A(k, , )(x)) for all and and for suffi-
ciently large k allows us to introduce the following definitions.
Definition. The inverse Born approximations qB (x) and qB (x) of the potential q(x)
are defined by
Z
1
qB (x) n/2
= (2) 1
(FM A)(x) = eik( ,x) A(k, , )d ,
0
(2)n M0
and Z
n/2 1 1
qB (x) = (2) (FM A)(x) = eik( ,x) A(k, , )d.
(2)n M
Remark. The equalities from the latter definition must be understood in the sense of
distributions.
94
Theorem 1 (Uniqueness). Assume that the potential q(x) belongs Lploc (Rn ), n2 2, |x| at the infinity.
Then the knowledge of qB (x) with restricted to an (n 2) dimensional semisphere
determines q(x) uniquely.
Proof. It is not so difficult to check that if q(x) satisfies the conditions of present
theorem then q(x) will satisfy the conditions of Theorem 2 from Chapter 10:
n n+1
q Lp (Rn ), 1 .
2 2p
Now we can represent qB (x) in the form
Z
1
qB (x) = n
eik( ,x) A(k, , )d (k, )
(2) M0
Z Z
1
= n
d (k, ) eik( ,x) eik( ,y) q(y)u(y, k, )dy
(2) M0 R n
Z Z
1 ik( ,xy)
= d e q(y)eik(,y) u(y, k, )dy,
(2)n M0 R n
where u(y, k, ) is the solution of the Lippmann-Schwinger equation. Denoting
v(y, k, ) := eik(,y) u(y, k, )
and making the change of variables = k( ) we obtain

Z Z !
1 ||
qB (x) = d i(,xy)
e q(y)v y, b
, dy.
(2)n Rn Rn 2(, )
The usual Fourier transform of qB (x) is equal to

Z " ! #
c ||

qB () = q()
b + (2)n/2 ei(,y) q(y) v y, b
, 1 dy
R n
2(, )
and it implies that
Z !
||
c
n/2
|qB q|
b (2) |q(y)| v y, b
, 1 dy,

Rn 2(, )
where the function v(y, k, ) solves the equation
Z
v(x, k, ) = 1 eik(x,) G+
k (|x y|)e
ik(y,)
q(y)v(y, k, )dy
Rn
i.e.
f (qv), c
v =1G k
95
= eik(xy,) G+ . For k large enough we may obtain that
where G k k
c
f q)1 (1)
v = (I + G k
or
f (q), c
v =1G q (11.2)
c
f is the integral operator with the kernel G = eik(xy,) G and the integral
where G q q q
2
operator Gq with this kernel satisfies also the equation (H k )Gq = I. In order to
prove (11.2) we recall that
=G
G G qG

q k k q
and, therefore
c
f =Gf Gc
f qGf c c
G q k k q
or
c
f = (I + G
f q)1 G
f.c c
G q k k
The last equality implies that

c
f (q) = (I + G
f q)1 Gc c
f (q) = (v 1)
G q k k
because
c
f q)1 G c
f (q) = (v 1)
(I + G k k
is equivalent to
c
f (q) = (I + G c
f q)(v 1) = (v 1) (G
f q)(v) + (G
f q)(1) c c
G k k k k
f (q) = G
f (q).c c
= v + 1 1 + v + G k k
Thats why we may apply Theorem 2 of Chapter 10 and get
kv 1k p1
2p
f (q)k 2p C kqk 2p ,
c
= kG q
L (Rn ) p1
L
k L p+1

2 2 2
where , p and are as in that theorem. It remains to check only that the potential
2p
q Lploc (Rn ) with the special behavior at the infinity belongs to L p+1
(Rn ). But it is a
2
very simple exercise. Hence, the letter inequality leads to
!
b )|
|(,
c
2
|qB () q()|
b Ckqk 2p , 6= 0
L p+1
(Rn ) ||
2
with the same . If q1 and q2 are as q then

|qb1 () qb2 ()| = |qb1 () qcB + qcB qb2 ()| |qb1 () qcB | + |qcB qb2 ()|
! !
b )|
|(, b )|
|(,
2 2
Ckq1 k 2p + Ckq2 k 2p =0
L p+1
(Rn ) || L p+1
(Rn ) ||
2 2
b ) = 0. Thus, this theorem is proved because (,

if (, b ) = 0 precisely when runs
through an (n 2)-dimensional semisphere, see [6, 7].
96
Theorem 2 (Saitos formula). Under the same assumptions for q(x) as in Theorem
1,
Z Z Z
(2)n q(y)dy
lim k n1 eik( ,x) A(k, , )dd = ,
k+ Sn1 Sn1 Rn |x y|n1
where the limit holds in classical sense for n < p and in the sense of the distri-
butions for n2 < p n.
n
Proof. Let us consider only the case n < p . The proof for 2
< p n takes place
with some changes.
By definition of the scattering amplitude,
Z Z
n1
I := k A(k, , )eik( ,x) dd
Sn1 Sn1
Z Z Z

= k n1 q(y)dy eik( ,yx) dd
Rn
Z Z Sn1 ZSn1

+ k n1 q(y)dy eik( ,y) R(y, k, )eik( ,x) dd := I1 + I2 ,
Rn Sn1 Sn1
where R(y, k, ) is given by

Z
R(y, k, ) = G+
k (|y z|)q(z)u(z, k, )dz
Rn
and u(z, k, ) is the solution of the Lippmann-Schwinger equation. Since

Z Z Z 2
ik( ,yx) ik(,yx)
e dd = e d
Sn1 Sn1 Sn1
Z
n1 2
4 ik|yx| cos n2
= e (sin ) d
2 ( n1
2
) 0
J 2n2 (k|x y|)
n
= (2) 2
,
(k|x y|)n2
then I1 can be represented in the form
Z
q(y)
I1 = (2)n k J 2n2 (k|x y|)dy.
Rn |x y|n2 2
We consider two cases: k|x y| < 1 and k|x y| > 1. In the first case using Holders
inequality the integral I1 over {y : k|x y| < 1} can be estimated by
Z
|q(y)|(k|x y|)n2
|I1 | Ck dy
|xy|< k1 |x y|n2
Z 1 Z 1
p p
n1 p
Ck |q(y)| dy 1 dy
|xy|< k1 |xy|< k1
Z 1 Z 1
p p
n1 pn p n
1 p
= Ck k |q(y)| dy = Ck p |q(y)| dy 0
|xy|< k1 |xy|< k1
97
as k + since n 1}. The asymptotic behavior of the Bessel function
J () for large argument implies that
Z
q(y)
I1 = (2) k n
|xy|> k1 |x y|n2
"s #2
2 n 1
cos k|x y| + +O dy
k|x y| 4 4 (k|x y|)3/2
Z " #
n q(y) 2 cos2 (k|x y| n
4
+ 4 ) 1
= (2) k +O dy
|xy|> k1 |x y|n2 k|x y| (k|x y|)2
Z
(2)n q(y)dy
=
|xy|> k |x y|n1
1

n
Z
(2)n q(y)
+ cos 2k|x y| + dy
|xy|> k1 |x y|n1 2 2
Z
1 |q(y)|O(1)
+ dy
k |xy|> k1 |x y|n
(1) (2) (3)
= I1 + I1 + I1 .
It is clear that Z
(1) (2)n q(y)dy
lim I1 =
k+ Rn |x y|n1
and
(2)
lim I1 = 0.
k+
The latter fact follows from the following arguments. Since q belongs to Lp (Rn ) for
p > n and has the special behavior at the infinity then we may conclude that L1 -norm
q(y)
of the function |xy|n1 is uniformly bounded with respect to x. Hence it follows from
(2)
the Riemann-Lebesgue lemma that I1 approaches to zero uniformly with respect to
(3)
x as k +. For I1 we have the estimate
Z
(3) C |q(y)|dy
|I1 | .
k 1 Rn |x y|n
R q(y)dy
If we choose such that 1 > > np , then Rn |xy|n will be uniformly bounded with
(3)
respect to x. Therefore, I1 0 as k uniformly with respect to x. If we collect
all estimates we obtain that
Z
(2)n q(y)dy
lim I1 = .
k Rn |x y|n1
Our next task is to prove that I2 0 as k . Since
Z Z Z
n1
I2 = k q(y)dy eik( ,y) R(y, k, )eik( ,x) dd ,
Rn Sn1 Sn1
98
where Z
R(y, k, ) = G+
k (|y z|)q(z)u(z, k, )dz = Gk (qu),
Rn
(qeik(,z) ). Hence, I can be represented as
then one can check that R(y, k, ) = G q 2
Z Z Z
I2 = k n1
q(y)dy e ik( ,xy)
d Gq q(z) eik(,zx)
d
Rn Sn1 Sn1

Z J n2 (k|x y|) J n2 (k|x y|)
= k n1 (2)n q(y) 2
G q(z) 2
dy
n2 q n2
Rn (k|x y|) 2 (k|x y|) 2

Z J n2 (k|x y|) J n2 (k|x y|)
1
= (2)n k q 1 (y) 2 c
K |q(z)| 2 2
dy,
n2 q n2
Rn 2
(|x y|) 2 (|x y|) 2
c is the integral operator with the kernel

where K q
1
Kq (x, y) = |q(x)| 2 Gq (k, x, y)q 1 (y).
2
It follows from Theorem 2 of Chapter 10 that K c : L2 (Rn ) L2 (Rn ) with the norm
q
estimate
kKck 2 2 C,
q L L
k
where is as in that theorem. Thats why we can estimate I2 using Holders inequality
as
C
Z J 2n2 (k|x y|)
|I2 | k |q(y)| 2 dy.
k Rn |x y|n2
By the same arguments as in the proof for I1 we can obtain that
Z J 2n2 (k|x y|)
k |q(y)| 2
dy <
Rn |x y|n2
uniformly with respect to x. It implies that

C
|I2 | 0, k +.
k
Remark. This proof holds also for n = 2. In dimension n = 1 there is an analogous

result in which we replace the double integral in the left hand side by the sum of four
values of the integrand at = 1 and = 1.
Theorem 3. Let us assume that n 2. Under the same assumptions for q1 (x), q2 (x)
as in Theorem 2 let us assume that the corresponding scattering amplitudes Aq1 and
Aq2 coincide for some sequence kj and for all , Sn1 . Then q1 (x) = q2 (x)
in the sence of Lp for n < p and in the sense of distributions for n2 < p n.
99
Proof. Saitos formula shows that we only have to show that the homogeneous equation
Z
q(y)dy
(x) := =0
Rn |x y|n1
has only the trivial solution q(y) 0. Let us assume that n < p . Introduce the
space S0 (Rn ) of all functions from the Schwartz space which vanish in some neighbor-
hood of the origin. Due to the conditions for the potential q(x) we may conclude (as it
was before) that L (Rn ) and it defines a tempered distribution. Then for every
function S0 (Rn ) it follows that
i = C h||1 q(),
0 = h, b i = Cn hq(),
b ||1 i.
n
Since () S0 (Rn ) then ||1 S0 (Rn ) also. Hence, for every h S0 (Rn ) the
following equation holds
hq,
b hi = 0.
This means that the support of q()

b is at most at the origin and therefore q()
b can be
represented as X
q()
b = C D .
||m
Hence, q(x) is a polynomial. But due to the behavior at the infinity we must conclude
that q 0. This proves Theorem 3.
Let us return now to the Born approximation of q(x). A repeated use of the
Lippmann-Schwinger equation leads to the following representation for the scattering
amplitude A(k, , ):
m Z
X 1

A(k, , ) = cj (|q| 2 eik(x,) )(y)dy
eik( ,y) q 1 (y)K
Rn 2
j=0
Z
1
+ cm+1 (|q| 2 (u(x, k, ))(y)dy,
eik( ,y) q 1 (y)K
Rn 2
c is an inte-
where u(x, k, ) is the solution of the Lippmann-Schwinger equation and K
gral operator with the kernel
1
K(x, y) = |q(x)| 2 G+
k (|x y|)q 1 (y). 2
The equality for A can be reformulated in the sense of integral operators in L2 (Sn1 )
as m X
=
A cj (k) + (k) sgn q K
0 (k) sgn q K cm+1 (k),
0 0
j=0
where 0 and (k) are defined by (10.1) and (10.2) and 0 is the L2 -adjoint of 0 .
Using this equality and the definition of Borns potential qB (x) we obtain
m
X
qB (x) = 1
FM 0 (k) sgn q K cm+1 (k) ,
cj (k) + F 1 (k) sgn q K
0 M 0
j=0
100
where the inverse Fourier transform is applied to the kernels of the corresponding
integral operators. If we rewrite the latter formula as
m
X
qB (x) = qj (x) + qem+1 (x),
j=0
then the term qj has the form

Z Z
1 1
1 ik(z,) j ik( ,y)
qj (x) = FM |q(z)| e 2 dz sgn q(z)K (z, y, k)|q(y)| e
2 dy
Rn Rn
1 ,y)

= FM1 cj
0 sgn q K (|q| 2 eik(
)
Z
1 ik( ,x) cj (|q| 21 eik( ,y) )
= e d(k, , ) 0 sgn q K
(2)n M
and a similar formula holds for qem+1 with obvious changes.
In order to formulate the result about the reconstruction of singularities of the
unknown potential q(x) let us set A(k, , ) = 0 for |k| k0 , where k0 > 0 is arbitrarily
large.
Theorem 4. Assume that the potential q(x) satisfies all conditions of Theorem 1 and
also belongs to L1 (Rn ). Then qj (x) and qej (x) for any j 1 belong to the Sobolev space
H t (Rn ) for any t < (j + 12 ) 1 with as in Theorem 2 of Chapter 10.
Proof. Using the change of variables = k( ) we obtain
t
kqj k2H t (Rn ) = k(1 + ||2 ) 2 F qj ()k2L2 (Rn )
Z Z ! 2
||
= Cn (1 + || ) 2 t cj
0 sgn q K 0 b )
, 2(, b d d,
b ),

Rn Sn1 2(,
where in the brackets [] there is the kernel of the corresponding integral operator. The
last estimate can further be bounded by
Z Z
Cn (1 + || )
2 t
0 sgn q K cj (...)2 dd
0
n Sn1
ZR Z
= Cn (1 + |k( )|2 )t cj (k( , ))2 d
0 sgn q K 0
M Sn1
Z 0 Z Z 2
n1 2 t cj (|q| 21 eik( ,y) ) dd
Cn k (1 + k ) dk 0 (sgn q K
k Sn1 Sn1
Z 0 Z
C k n1 (1 + k 2 )t dk k k2 kKcj (|q| 21 eik( ,y) )k 1 d
n 0 L
k0 Sn1
Z Z
Cn k n1 (1 + k 2 )t dk c 2j kqk 1 d
k0 k2 kKk L
k0 Sn1
Z k n1+2t dk
Cn 1 .
k0 k 2(j+ 2 )+n2
But the latter integral converges if and only if t < (j + 12 ) 1. This proves the
theorem.
101
Theorem 5 (Reconstruction of singularities). Assume that the potential q(x) belongs to
Lploc (R3 ) for 3 3, |x|
at the infinity. Then
qB (x) q(x) q1 (x) H t (R3 ),
where t < 32 2p
5
and q1 (x) is continuous and bounded (more precisely, q1 (x) Wp1 (R3 )
for p > 3).
Proof. The statement about the first nonlinear term q1 (x) was proved by Lassi Pai-
varinta and Valery Serov, see [8]. It is also easy to check that q0 (x) is simply the
unknown potential q(x). Hence, we can write
qB q q1 = qe2
and so the claim follows from Theorem 4.

Remark. Let us assume that 3 < p < . Then the following embedding holds
t3( 21 p1 )
H t (R3 ) Wp (R3 ) = Wp (R3 ),

where = t 3 12 p1 and, therefore, < 2p 1
. It means that qe2 belongs to the
smoother space than the unknown potential q(x) and so we can reconstruct main
singularities of the potential q(x) by the Born approximation.
Remark. Actually it is possible to prove (see [9]) that
3
q2 (x) C 1 p (R3 )
for any 3 < p . Using Theorem 4 we obtain that qB q q1 q2 H t for any

7
t < 25 2p . But for 3 < p the following embedding holds: H t C(Rn ). It means
that for 3 < p we can reconstruct all singularities of the unknown potential
because q1 and q2 are continuous in this case.
Remark. It follows from this theorem that by Born approximation we can reconstruct
an arbitrary bounded domain D. In order to see this fact it is enough to consider
q(x) = D (x).
102
References
[1] Cycon H L, Froese R G, Kirsch W and Simon B, Schrodinger Operators, Academic
Press, New-York, 1989
[2] Ikebe I, Arch. Rational Mech. Anal., 5 (1960), 1-34
[3] Kato T, Comm. Pure Appl. Math., 12 (1959), 403-425
[4] Ruiz A, lectures at the summer school in Oulu, June, 2002
[5] Agmon S, Ann. Scuola Norm. Sup. Pisa Cl. Sci., 4 (1975), 151-218
[6] Saito Y, Osaka, J. Math.19 (1982), 527-547
[7] Paivarinta L and Somersalo E, SIAM J. Math. Anal., 22 (1991), 480-499
[8] Paivarinta L and Serov V, SIAM J. Math. Anal., 29 (1998), 697-711
[9] Paivarinta L and Serov V, Inverse Probl. Imaging, 1 (2007), 525-535
103

Fourier

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fourier

Uploaded by

Copyright:

Available Formats

Fourier transform and distributions

with applications to the Schrodinger operator

Edited by Markus Harju

2 Fourier transform in Schwartz space 3

3 Fourier transform in Lp (Rn ), 1 p 2 10

8 Fundamental solutions of elliptic partial differential operators 55

10 Estimates for Laplacian and Hamiltonian 79

11 Some inverse problems for the Schrodinger operator 93

2) Lebesgue integral in a bounded domain Rn and in Rn .

3) Banach spaces (Lp , 1 p , C k ) and Hilbert spaces (L2 ): If 1 p < then

where is the closure of . We say that f C () if f C k (1 ) for all k N

Also in L2 (), the duality pairing is given by

where the Holder conjugate exponent p of p is obtained via

6) Fubinis theorem about the interchange of the order of integration:

if one of the three integrals exists.

Recall the Cauchy-Bunjakovsky inequality

|(x, y)| |x||y|.

Following L. Schwartz we call an n-tuple = (1 , . . . , n ), j N {0} N0 an

if and only if j j for all j = 1, 2, . . . , n. Let us also introduce a shorthand notation

Definition. The Schwartz space S(Rn ) of rapidly decaying functions is defined as

The following properties of S = S(Rn ) are readily verified.

4) If f S(Rn ) then |f (x)| cm (1 + |x|)m for any m N. The converse is not

5) It follows from part 4) that S(Rn ) Lp (Rn ) for any 1 p .

4) C0 (Rn ) S(Rn ), where C0 (Rn ) = {f C (Rn ) : supp f is compact in Rn }

Exercise 1. Prove that is a metric, that is,

1) (f, g) 0 and (f, g) = 0 if and only if f = g.

3) (g, h) (g, f ) + (f, h).

4) (In addition) |(f, h) (g, h)| (f, g).

Proof. Let {fk }

sup |x f | sup |x (fk f )| + sup |x fk |

C (K) sup | (fk f )| + sup |x fk |.

sup |x f | lim sup |fk |, < .

The last inequality is valid since {fk }

Remark. This integral is well-defined since

Further, integration by parts gives us

If we combine these two last estimates we may conclude that F : S S and F is

Example 2.2. It is true that

Proof. The definition gives us directly

Exercise 3. Let P (D) be a differential operator

(h f )(x) := f (x h), ( f )(x) := f (x), 6= 0.

Exercise 4. Let f S(Rn ), h Rn and R, 6= 0. Prove that

Exercise 5. Let A be a real-valued nn-matrix such that A1 exists. Denote fA (x) :=

if and only if A is an orthogonal matrix (a rotation), that is, AT = A1 .

Let us now consider f and g from S(Rn ). Then

where F g(x) := F g(x).

Theorem 2 (Fourier inversion formula). Let f be a function from S(Rn ). Then

whenever |x| < . Note that Z

But I 0 as 0. This proves the lemma.

Corollary. The Fourier transform is an isometry (in the sense of L2 ).

Proposition. Let X be a linear normed space and E X a subspace of X such that

Exercise 6. Prove the previous proposition.

Lemma 1. Let 1 p < . Then

that is, C0 (Rn ) is dense in Lp (Rn ) in the sense of Lp -norm.

Thus, we may conclude that C0 (Rn ) = Lp (Rn ) for 1 p < .

kf kL (Rn ) |c0 | > 0.

But the following result holds:

C(Rn ) := {f C(Rn ) : lim f (x) = 0}.

Now we are in the position to extend F from S L1 to L1 .

Theorem 1 (Riemann-Lebesgue lemma). Let F : S S be the Fourier transform in

and Exercise 4. Thus,

kFex f kL2 = lim kF fk kL2 = lim kfk kL2 = kf kL2

which proves the statement about the operator norm.

Exercise 9. Let us assume that f L1 (Rn ) and F f () L1 (Rn ). Prove that

It means that the Fourier inversion formula is valid.