by Dr.

Yi Zhang
Institute of Mathematical Sciences
Fudan University
Shanghai 200433
P.R.China
zhangyi math@fudan.edu.cn
Lecture given at Fudan University, 2006.
http://homepage.fudan.edu.cn/yizhang/teaching/analysis-exercise.pdf
The author holds the copyright of this lecture notes. Any person(s) intending to copy a part or whole of the materials
in the notes in a proposed publication must seek copyright release from the author.
1
Analysis i

Dedicate to Katie
,
,

ii Lecture Notes at Fudan

Contents
References ii
1. Problems on Sets,Sequences and Limits 1
1.1. Elementary Technique 1
1.2. Applications of the Stolz theorem 11
2. Inequality 14
3. Orders Estimate of Innitesimal 20
3.1. Notations and Examples 20
3.2. Exercises and homework 23
4. Application of Innitesimal: Wallis Formula and Stirling Formula 27
5. Topic on the Gamma functions 30
5.1. functions 30
5.2. Exercises 34
6. Working Technique in function theory 36
6.1. Iteration technique 36
6.2. Exercises and Homework 39
7. Applications of Dierential 41
8. Treasures in Calculus 42
8.1. Eulers identication (2) = 1 +
1
2
2
+
1
n
2
+

2
6
42
8.2. Irrationality of , log 2, (2), (3). 43
8.3. The properties of Tchebychevs functions 46
8.4. Mertens Theorem and Selbergs Inequality 53
8.5. An Elementary Proof of The Prime Number Theorem by Selberg and Erdos 58
8.6. Gausss Proof of The Fundamental Theorem of Algebra 59
9. Advanced Techniques in Analysis 60
9.1. Preliminary Results in Analysis 60
9.2. Scaling Technique and Schauder Estimate 72
9.3. Campanatos characterization of L
2
functions to be Holder continuous 79
10. Advanced Topics in Analysis 82
10.1. Topic on Riemann Zeta Function(To be Continuous) 82
10.2. Elementary on Nevanlinna Theory(To be Continuous) 83
10.3. Elementary on p-adic Series(To be Continuous) 84
References
[F] .M. : , VOL I., II. , III. Higher Education Press.
[H] G.H. Hardy : A Course of Pure Mathematics (Tenth Edition) Combridge University Press 2002
[HUA] : , 1979.
[J] J.Jost: Partial Dierential Equations (GTM 214), Spinger 2002.
[P] Problems Selection in The William Lowell Putnam Mathematical Competitions.
[P-Y] , : .
Analysis iii

Analysis 1
1. Problems on Sets,Sequences and Limits
1.1. Elementary Technique.
Exercise 1.1. Show that the set of all irrational number of R is uncountable.
Proof. Sucient to show that the set of all irrational number in [0, 1] is uncountable. Otherwise, the
set of real numbers in [0, 1] is countable, ie.,
[0, 1] = x
1
, x
2
, .
Cover each x
i
, i = 1, with the corresponding interval
I
i
:= [x
i

1
2
(
1
3
)
i
, x
i
+
1
2
(
1
3
)
i
],
Then,there holds
[0, 1] = x
1
, x
2
,

_
i=1
I
i
,
and so
1 = l([0, 1])

i=1
l(I
i
) =

i=1
l([x
i

1
2
(
1
3
)
i
, x
i
+
1
2
(
1
3
)
i
]) =

i=1
(
1
3
)
i
=
1
2
.
It is a contradiction.

Example 1.2. Let , R

+
Q and 1/ + 1/ = 1. Let
A = [n] : n N, B = [n] : n N
be two strictly increasing sequences of positive integers.
Show that
A
_
B = N; A

B = .
Proof. We have > 1, > 1. W.L.O.G., we assume < 2. Then 1 A B.
1. If a B ,= , then there exist two integers m, n such that
[m] = [n] = q N.
Hence there holds that q < m +n < q + 1. A contradiction.
2. If there is a positive integer p A B, then there exist two integers m, n such that
[m] < p < [(m + 1)], and [n] < p < [(n + 1)].
Therefore
m < p < p + 1 [(m + 1)] < (m + 1), and n < p < +1 [(n + 1)] < (n + 1).
Since 1/ + 1/ = 1, we have
m +n < p < p + 1 < m +n + 2.
A contradiction.

Example 1.3. Let f : Z Z be a bounded function. Assume for any integers m, n there is a
relation:
f(m +n) +f(mn) = 2f(m)f(n).
To show all such function f.
2 Lecture Notes at Fudan
Hint. a) f(0) 0, 1, and f(0) = 0 f 0.
b) Now, we assume f(0) = 1. Then, f(n) = f(n)n Z, and f(2m) = 2f(m)
2
1, m Z. If
there is an integer k such that [f(k)[ > 1, then
[f(2k)[ = [2f(k)
2
1[ = (f(k)
2
1) +f(k)
2
> [f(k)[ > 1,
and so we have an increasing sequence of integer
[f(k)[ < [f(2k)[ < [f(4k)[ < < [f(2
l
k)[ < .
But it contradict that f is a bounded function with all values in Z. Therefore
f(Z) 1, 0, 1.
c) Let f(1) = cos with ,

2
, 0. We can show that
f(n) = cos(n), n Z
by induction and the formula
2 cos cos = cos( +) + cos( ).

Example 1.4. Dene the function G : N 0 Z by

_
G(0) = 0 ;
G(n) = n G(G(n 1)), n N.
To show that
G(n) = [

5 1
2
(n + 1)].
Proof. 1. Actually, we have
1 G(n) n and G(n 1) G(n) n 1.
At rst G(1) = 1, G(2) = 1.By induction, we assume
1 G(k) k and G(k 1) G(k) 1 k n 1.
Then, we have
1 G(n 1) n 1, and so G(G(n 1)) G(n 1) n 1,
1 = n (n 1) G(n) = n G(G(n)) n 1 < n,
G(n) G(n 1) = 1 [G(G(n 1)) G(G(n 2))] 0 since 1 G(n 2) G(n 1) n 1.
2. By induction, we can show
G(n + 1) G(n) = 1 or 0 n.
3. Dene F(n) = [(n + 1)] where =

51
2
. Dene
S(n) = F(n) +F(F(n 1)).
We are going to show that S(n) = n, and so F(n) = G(n).
Let K = [n], and so n = K + with 0 < < 1. Then
F(n) = [(n + 1)] = [K + +] = K + [ +], F(F(n 1)) = F([n]) = [(K + 1)].
Since
2
= 1 , we have
(K + 1) = + (n ) = (1 ) +n
2
= (1 ) +n K ,
S(n) = F(n) +F(F(n 1)) = n + [ +] + [(1 ) ].
Analysis 3
Let T = (1 ) . Then
1 < < T < < 1, and so [T] =
_
0, T > 0;
1, T < 0.
On the other hand, we have
T = (1 ) = (1 +)(

1 +
) = (1 +)(
2
) = (1 +)(1 ( +)),
and so
[T] =
_
0, [ +] = 0;
1, [ +] = 1.
Therefore,
[ +] + [T] = 0, i.e., S(n) = n.

Example 1.5. Let a

n
is strictly increasing sequence of positive integers, i.e.,a
n+1
> a
n
, and
a
n
N. Dene
b
n
= [a
1
, a
2
, , a
n
] the least common multiple of a
1
, a
n
.
To show that

n=1
1
b
n
< .
Proof. Let
d(n) = #p : p N p[n.
Then d(n) < 2

n, and so we have
n d(b
n
) < 2
_
b
n
.
By the previous exercise, we obtain the result.

Example 1.6. To show

lim
n
n

k=1
(
3
_
1 +
k
n
2
1) =
1
6
.
Proof. Let

k
=
3
_
1 +
k
n
2
1.
Then 0 <
k
< 1 and
k
<
k
3n
2
by
1 +
k
n
2
= (1 +
k
)
3
> 1 + 3
k
.
ON the other hand,
1 +
k
n
2
= (1 +
k
)
3
= 1 + 3
2
k
+ 3
k
+
3
k
1 + 4
2
k
+ 3
k
1 + 4(
k
3n
2
)
2
+ 3
k
,
and so
n

k=1
k
3n
2

n

k=1

k

n

k=1
4k
2
27n
4
+
k
3n
2
.
4 Lecture Notes at Fudan
Since
n

k=1
k =
n(n + 1)
6
and
n

k=1
k
2
=
n(n + 1)(2n + 1)
6
,
we have
lim
n
(
n(n + 1)
6n
2

2n(n + 1)(2n + 1)
3 27n
4
) lim
n
n

k=1

k
lim
n
n(n + 1)
6n
2
.

Remark. Let
g(x) =
n

i=1
i(1 +x)
i
.
Then,
g(x) (1 +x)g(x) =
n

i=1
(1 +x)
i
n(1 +x)
n+1
,
and so
g(x) =
(1 +x) (1 +x)
n+1
x
2
+
n(1 +x)
n+1
3x
.
Comparing the coecient of the term x of the polynomial g(x), we have:
n

i=1
i
2
= C
3
n+1
+nC
2
n+1
=
n(n + 1)(2n + 1)
6
.
Example 1.7. Let a
n
, b
n
be two sequence satisfying
(1)
(a
n
+b
n
)b
n
,= 0, n = 1, 2, ,
(2)
[

n=1
a
n
b
n
[ = A < ,

n=1
(
a
n
b
n
)
2
= B < .
Then the sum

n=1
a
n
a
n
+b
n
<
is convergent.
Proof. By the condition (2), for any R with 0 < <
1
2
, there exist an integer N > 0, for any
n, l N, we have
[
a
n
b
n
[ , [

i=l
a
i
b
i
[ < , and

i=l
(
a
i
b
i
)
2
< .
and so
(1.7.1)
a
n
a
n
+b
n
=
a
n
b
n
1 +
a
n
b
n
=
a
n
b
n
(1
a
n
b
n
)
1 (
a
n
b
n
)
2
= (
a
n
b
n
(
a
n
b
n
)
2
)(1 +

k=1
(
a
n
b
n
)
2k
), n N
(1.7.2)
n
=

k=1
(
a
n
b
n
)
2k
=
(
a
n
b
n
)
2
1 (
a
n
b
n
)
2
2(
a
n
b
n
)
2
2
2
(n N),
Analysis 5
(1.7.3) [

i=l
(
a
i
b
i
)
3
[

i=l
[
a
i
b
i
[
3

i=l
(
a
i
b
i
)
2

i=l
(
a
i
b
i
)
4

i=l
(
a
i
b
i
)
2
(l N).
On the other hand, for l Nwe have
[

n=l
a
n
a
n
+b
n
[ = [

n=l
(
a
n
b
n
(
a
n
b
n
)
2
)(1 +

k=1
(
a
n
b
n
)
2k
)[ 2 +[

n=l
(
a
n
b
n
(
a
n
b
n
)
2
)

k=1
(
a
n
b
n
)
2k
[.
But by 1.7.2 and 1.7.3 we have the following
[

n=N
(
a
n
b
n
(
a
n
b
n
)
2
)

k=1
(
a
n
b
n
)
2k
[ [

n=l
a
n
b
n

n
[ +[

n=l
(
a
n
b
n
)
2

n
[

n=l
[
a
n
b
n
[
3
1 (
a
n
b
n
)
2
+ 2
3
.

n=l
[
a
n
b
n
[
3
1 (
a
n
b
n
)
2
2

n=l
[
a
n
b
n
[
3
2.
At all, we obtain
[

n=l
a
n
a
n
+b
n
[ < 6, l N,
and so by Cauchy principal, it is convergent.

Another clever proof by a student. It is sucient to consider the convergence of

n=1
(
a
n
a
n
+b
n

a
n
b
n
).
We have

n=1
(
a
n
b
n

a
n
a
n
+b
n
) =

n=1
a
2
n
(a
n
+b
n
)b
n
=

n=1
(
a
n
b
n
)
2
b
n
a
n
+b
n
.
Since
[
a
n
b
n
[
1
2
, (n >> 0),
and b
n
,= 0, n N, we have
0 <
b
n
a
n
+b
n
=
1
1 +
a
n
b
n
2, (n >> 0),
and so
b
n
a
n
+b
n
has a positive upper bounded uniformly for n N. Therefore the sum

n=1
(
a
n
a
n
+b
n

a
n
b
n
)
is convergent.

Example 1.8. Let a

n
be a sequence satisfying
a
n+1
(2 a
n
) = 1 n N.
To shown that
lim
n
a
n
exists, and
lim
n
a
n
= 1.
6 Lecture Notes at Fudan
Hint. Let b
n
= 1 a
n
, n = 1, 2 . We have a new sequence b
n
with the relation
(1 +b
n
)(1 b
n+1
) = 1, i.e., b
n+1
=
b
n
1 +b
n
.
We nd b
n
= 0 b
n+1
= 0. If one b
m
= 0 then b
n
= 0, n m and lim
n
a
n
= 1.
No we assume b
n
,= 0, n = 1, 2, . Then we have
1
b
n+1

1
b
n+1
= 1,
and so
1
b
n+1
= n +
1
b
1
, i.e., b
n+1
=
b
1
nb
1
+ 1
0 (n ).

Example 1.9. Let a

n
be a sequence satisfying
(1)

n=1
a
n
= 1;
(2) for all n N,
0 < a
n

k=n+1
a
k
.
Then, for any x (0, 1), there exists a subsequence a
n
k
of a
n
such that

k=1
a
n
k
= x.
Proof. Dene the n
1
, n
2
by induction.
1. Let
n
1
= minn[ a
n
< x.
It is well-dened since by the condition (1) lim
n
a
n
= 0. Then by the condition (2) we have
0 < a
n
1
1

k=n
1
a
k
,
and so

k=n
1
a
k
x.
2. If

k=n
1
a
k
> x
then we dene
n
2
= minn > n
1
[ a
n
1
+a
n
< x..
We must show that
a
n
1
+

k=n
2
a
k
x.
Analysis 7
Otherwise, we have n
2
> n
1
+ 1 since by step 1, and then by the condition (2) we have
a
n
1
+a
n
2
1
a
n
1
+

k=n
2
a
k
< x.
A contradiction to the denition of n
2
.
3. If
a
n
1
+

k=n
2
a
k
> x
then we dene
n
3
= minn > n
1
[ a
n
1
+a
n
2
+a
n
< x..
and we have
a
n
1
+a
n
2
+

k=n
3
a
k
x.
4. By induction, assume we have n
1
, n
2
, , n
k
with
(1.9.1)
k1

j=1
a
n
j
+

i=n
k
a
i
x.
If
k1

j=1
a
n
j
+

i=n
k
a
i
= x,
then the proof will be nished. Otherwise, we dene
n
k+1
= minn > n
k
[
k

i=1
a
n
i
+a
n
< x.
Only necessarily to show that
(1.9.2)
k

j=1
a
n
j
+

i=n
k+1
x.
If 1.9.2 is false, then n
k+1
> n
k
+ 1 since 1.9.1. By the condition (2) there must hold
a
n
1
+a
n
2
+ +a
n
k
+a
n
k+1
1

k

j=1
a
n
j
+

k=n
k+1
a
k
< x.
It is a contradiction to the denition of n
k+1
.

Example 1.10. 1. Let f(x) =

x
2
+c. Compute the iteration
f
(n)
(x) = f f f
. .
n
.
Solute.Let
(x) = x
2
, g(x) = x +c.
Then we have :
f =
1
g and f
(n)
(x) =
1
g
(n)
,
where
1
is the inverse function of .
g
(n)
(x) = x +nc, and so f
(n)
(x) =

x
2
+nc.
8 Lecture Notes at Fudan

2. Let
f(x) =
x
3

x
3
+c
.
Compute the iteration
f
(n)
(x) = f f f
. .
n
.
Solute.Let
(x) = x
3
, g(x) =
x
x +c
.
Then we have :
f =
1
g and f
(n)
(x) =
1
g
(n)
,
where
1
is the inverse function of . By induction, we have:
g
(n)
(x) =
x
(

n1
i=o
c
i
)x +c
n
.
Then,
f
(n)
(x) =
x
3
_
(

n1
i=o
c
i
)x
3
+c
n

Exercise 1.11. a) Let

f(x) = x
3
+ 6x
2
+ 12x + 6.
Compute the iteration
f
(n)
(x) = f f f
. .
n
.
b) Let
f(x) =
x + 4
2x 1
.
Compute the iteration
f
(n)
(x) = f f f
. .
n
.
c) Let
f(x) =
2x
1 x
2
.
Compute the iteration
f
(n)
(x) = f f f
. .
n
.
Analysis 9
Exercise 1.12. The sequence a
0
, a
1
, a
n
is dened by
a
0
=
1
2
, a
k
= a
k1
+
1
n
a
2
k1
.
To show that
1
1
n
< a
n
< 1.
Hint. By induction to show for any 1 k n, we have
n + 1
2n k + 2
< a
k
<
n
2n k
,
and let k = n, we get
n + 1
n + 2
< a
n
< 1.

Exercise 1.13. Let a 1, a

2
, , a
n
be a sequence of dierent positive integers. To show that
n

k=1
a
k
k
2

n

k=1
1
k
.
Hint. Let S
k
= a
1
+ +a
k
. Then
S
k

k

i=1
i
k(k + 1)
2
,
and we have:
n

k=1
a
k
k
2
=
n

k=1
S
k
[
1
k
2

1
(k + 1)
2
] +
S
n
n
2

n

k=1
k(k + 1)
2
[
1
k
2

1
(k + 1)
2
] +
n + 1
2n
.

Exercise 1.14. Let a

1
a
2
, a
n
and b
1
b
2
, b
n
. To show
(
n

k=1
a
k
)(
n

k=1
b
k
) n
n

k=1
a
k
b
k
.
Proof. Let
D

k=1
a
k
b
k
(

k=1
a
k
)(

k=1
b
k
).
We have D
1
= 0 and
D
+1
D

k=1
(a
k
a
k+1
)(b
k
b
k+1
) 0.

Exercise 1.15. 1. To show that

log(n + 1) <
n

k=1
1
k
< log n + 1.
10 Lecture Notes at Fudan
2. Consider

N
() =
N

k=1
1
k

, R.
To show that
lim
N

N
() < , (1, );
and
lim
N

N
() = , (0, 1).
Exercise 1.16. Show the sum
S = C
2
n
x +C
5
n
x
2
+C
8
n
x
3
+ + +C
n1
n
x
n
3
,
where n is a positive integer with 3[n.
Hint. Let be the primitive root of y
3
= 1, i.e.,
2
+ + 1 = 0. Then, we have

2(k+1)
+
k+1
+ 1 =
_
3, 3[k + 1;
0, others.
Let a
3

x,
3

x,
2 3

x. Consider the equality

a(1 +a)
n
=
n

k=0
C
k
n
a
k+1
,
Let a goes through the three values in the set a
3

x,
3

x,
2 3

x, we then sum the three equalities

and get
S =
3

2
i=0

i
(1 +
3

x
i
)
n
3
.

Exercise 1.17. Let the sequence x

n
satisfy the following condition:
_
x
1
= 815, ;
x
n+1
= x
2
2, n N.
Show the limit
lim
n
x
n+1
x
1
x
2
x
n
.
Hint. We have
x
2
n+1
4 = (x
2
1
4)(x
1
x
n
)
2
and x
n
> 2
n
for all n N.

Exercise 1.18. If the following is true for all n N,

[Bn] = [A[An]] + 1,
show that
B = A
2
.
Analysis 11
Hint.
lim
n
[n]
n
= .

1.2. Applications of the Stolz theorem.

Theorem (Stolz theorem). Let y
n
is strictly increasing sequence, i.e.,y
n+1
> y
n
, and y
n
. If
lim
n
x
n
x
n+1
y
n
y
n+1
= a, a [, +],
Then
lim
n
x
n
y
n
= a.
Example 1.19. Let a
n
, b
n
be two sequences satisfying that
b
n+1
= a
n
+ 918a
n+1
n N.
Then, we have:
lim
n
b
n
exists lim
n
a
n
exists.
Proof. We only prove the part. Let b = lim
n
b
n
. We dene two new sequences

n
=
b
919
a
n
,
n
=
b b
n
918
, n = 1, 2, .
Let =
1
918
. then the sequences
n
,
n
satisfy

n+1
+
n
=
n+1
.
By induction, we have

n+1
=
n+1

i=1

n+1i
+
n+1

0
=

n+1
i=1

i
(
1

)
i
+
0
(
1

)
n+1
,
and so
[
n+1
[

n+1
i=1
[
i
[[
1

[
i
+[
1
[
[
1

[
n+1
.
By the Stolz theorem, we have
lim
n

n+1
i=1
[
i
[[
1

[
i
+[
1
[
[
1

[
n
= lim
n
[
n+1
[[
1

[
n+1
[
1

[
n
([
1

[ 1)
= 0.

1.20. Exercises.
1. Let p
1
, p
2
, , p
n
, be positive numbers. If
lim
n
p
n
= p (0, ),
then
lim
n
n

p
1
p
2
p
n
= p.
12 Lecture Notes at Fudan
2. Let a
0
, a
1
, a
2
, , a
n
, be positive numbers. If
lim
n
a
n+1
a
n
= p < ,
then
lim
n
n

a
n
= p.
3. To show for any k N
lim
n

n
i=1
i
k
n
k+1
=
1
k + 1
.
Proof.By Stolz theorem, we compute
lim
n

n+1
i=1
i
k

n
i=1
i
k
(n + 1)
k+1
n
k+1
= lim
n
(n + 1)
k
(n + 1)
k+1
n
k+1
= lim
n
(n+1)
k
n
k
(n+1)
k+1
n
k+1
n
k
.
On the other hand,
(n + 1)
k
n
k
= 1 +O(
1
n
),
(n + 1)
k+1
n
k+1
n
k
= k + 1 +O(
1
n
).

4. To show for any k N

lim
n
n(

n
i=1
i
k
n
k+1

1
k + 1
) =
1
2
.
Proof.
n(

n
i=1
i
k
n
k+1

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

n
i=1
i
k
) n
k+1
(k + 1)n
k
.
By Stolz theorem, it is sucient to compute that
lim
n
(k + 1)(n + 1)
k
(n + 1)
k+1
+n
k+1
(k + 1)((n + 1)
k
n
k
)
On the other hand
(k + 1)(n + 1)
k
(n + 1)
k+1
+n
k+1
= (k + 1)
k

i=0
C
i
k
n
i

i=0
C
i
k+1
n
i
=
k1

i=0
((k + 1)C
i
k
C
i
k+1
)n
i
So
(k + 1)(n + 1)
k
(n + 1)
k+1
+n
k+1
(k + 1)((n + 1)
k
n
k
)
=
1
2
k(k + 1) +O(
1
n
)
k(k + 1) +O(
1
n
)

5. Let x
1
(0, 1) and x
n
be a sequence with x
n+1
= x
n
(1 x
n
)n N. Then we have:
lim
n
nx
n
= 1.
Analysis 13
Proof.Since x
2
= x
1
(1 x
1
) (1/2)
2
< 1, by induction we have x
n
(0, 1)n N. On the other
hand,
x
n+1
x
n
= 1 x
n
< 1,
so x
n
is a strictly decreasing sequence with a nite bound, and then
lim
n
x
n
= A [0, 1).
Moveover, by A = A(1 A), we have A = 0. By Stolz theorem,
lim
n
nx
n
= lim
n
n
1
x
n
= lim
n
n (n 1)
1
x
n

1
x
n1
= lim
n
(1 x
n
) = 1.

6. let a
k
be a sequence satisfying that
lim
n
A
n
= lim
n
n

k=1
a
k
= A < .
Let p
k
be a strictly increasing sequence of positive numbers with p
k
+ (k ). Show
that
lim
n
p
1
a
1
+p
2
a
2
+ +p
n
a
n
p
n
= 0.
Proof.Assume A
0
= 0, we have:
p
1
a
1
+p
2
a
2
+ +p
n
a
n
p
n
=

n
k=1
p
k
(A
k
A
k1
)
p
k
=

n1
k=1
A
k
(p
k
p
k+1
) +A
n
p
n
p
n
By STolz theorem, we have
lim
n

n1
k=1
A
k
(p
k
p
k+1
)
p
n
= lim
n
A
n
(p
n
p
n+1
)
p
n+1
p
n
= A

14 Lecture Notes at Fudan

2. Inequality
Theorem 2.1 (Holder Inequality). Let a
k
, b
k
0(k = 1, , n), q, p > 0 with
1
p
+
1
q
= 1. Then, we
have
n

k=1
a
k
b
k
(
n

k=1
a
p
k
)
1
p
(
n

k=1
b
q
k
)
1
q
.
Moreover, the inequality becomes to be an equality if and only if there exists a t R
+
such that
a
p
k
= tb
q
k
(k = 1, , n).
The proof of the Holder inequality depends heavily on the following lemma.
Lemma 2.2. Let A, B 0. then for any [0, 1], we have:
A

B
1
A + (1 )B,
and the inequality becomes to be an equality if and only if A = B.
Proof. W.L.O.G. we assume A B > 0, ,= 0. Then we have
(
A
B
)

=
_ A
B
1
x
1
d x + 1 <
_ A
B
1
1 d x + 1 = (
A
B
1) + 1 = (
A
B
) + (1 ).

Exercise 2.3 (Young Inequality). Let A, B 0. and q, p > 0 with

1
p
+
1
q
= 1. Then, we have
AB < A
p
+

q
p
B
q
> 0.
Exercise 2.4 (Minkowski Inequality). Let a
k
, b
k
0(k = 1, , n), p > 1. Then, we have

k=1
(a
k
+b
k
)
p

1
p

n

k=1
a
p
k

1
p
+
n

k=1
b
p
k

1
p
.
Moreover, the inequality becomes to be an equality if and only if there exists a t R such that
a
k
= tb
k
(k = 1, , n).
Theorem 2.5 (Arithmetic Mean-Geometry Mean Inequality). Let a
1
, , a
n
be positive real
numbers and a
1
, , a
n
be positive real numbers with p
1
+ +p
n
= 1. Then we have:
G
n
=
n

i=1
a
p
i
i

n

i=1
p
i
a
i
= A
n
,
moreover the inequality becomes to be an equality if and only if a
1
= a
2
= = a
n
.
Proof. W.L.O.G. we assume a
1
a
2
a
n
. Then there exists an integer k [1, n1] such that
a
k
G
n
a
k+1
.
Analysis 15
Consider the following formula
A
n
G
n
1 =
n

i=1
p
i
(
a
i
G
n
G
n
)
n

i=1
p
i
(
log a
i
log G
n
G
n
)
=
k

i=1
p
i
(
G
n
a
i
G
n
+ (log G
n
log a
i
)) +
n

i=k+1
p
i
(
a
i
G
n
G
n
(log a
i
log G
n
))
=
k

i=1
p
i
_
G
n
a
i
(
1
G
n

1
t
)dt +
n

i=k+1
p
i
_
a
i
G
n
(
1
G
n

1
t
)dt
=
k

i=1
+p
i
_
G
n
a
i
(
1
t

1
G
n
) +
n

i=k+1
p
i
_
a
i
G
n
(
1
G
n

1
t
)
0.
Moreover the inequality becomes to be an equality if and only if
a
1
= a
2
= = a
n
= G
n
.

2.6. Exercises.
1. Let a
k
, b
k
0(k = 1, , n), with
1
p
+
1
q
= 1 with 0 < p < 1 Assume that
(a
p
k
)
n
k=1
,= t(b
q
k
)
n
k=1
, t R.
To show that
n

k=1
a
k
b
k
> (
n

k=1
a
p
k
)
1
p
(
n

k=1
b
q
k
)
1
q
.
Hint.We have
0 <
1
(
1
p
)
< 1, 0 <
1
(
q
p
)
= 1 k < 1,
and
1
(
1
p
)
+
1
(
q
p
)
= 1.
Then, we use the Holder inequality to the following
n

i=1
a
p
i
=
n

i=1
(a
i
b
i
)
p
b
p
i
=
n

i=1
(a
i
b
i
)
1
1/p
b
1
q/p
i
.

2. Let a
k
0(k = 1, , n), p > 0. We dene
M
p
(a) =

n
k=1
a
p
k
n

1
p
.
To show that for 0 < r < s, we always have
M
r
(a) < M
s
(a)
except for a
1
= a
2
= = a
n
.
16 Lecture Notes at Fudan
3. To show
n

k=1
_
1
k
sin(
k
n
)

1 + log n
sin

2n
.
4. To show
n

k=1
1
k
2
_
arctan
1
k
2
+k + 1

_

3
.
Example 2.7 (Hilbert Inequality). Assume

n=1
a
2
n
. Then we have:

m=1

n=1
a
m
a
n
m +n

n=1
a
2
n
.
Proof. Consider
N

m=1
N

n=1
a
m
a
n
m +n
=
N

m=1
N

n=1
(
m
n
)
1
4
a
m

m +n
(
n
m
)
1
4
a
n

m +n
.
Thus, by the Holder inequality, we have:
N

m=1
N

n=1
a
m
a
n
m +n

N

m=1
N

n=1
(
m
n
)
1
2
a
2
m
m +n

1
2

N

m=1
N

n=1
(
n
m
)
1
2
a
2
n
m +n

1
2
=
N

m=1
N

n=1
(
m
n
)
1
2
a
2
m
m +n
,
N

m=1
N

n=1
a
m
a
n
m +n

N

m=1
a
2
m
N

n=1
(
m
n
)
1
2
1
m +n
.
On the other hand, we have:
(
m
n
)
1
2
1
m +n
=
1
m
(
n
m
)
1
2
(1 +
n
m
)

_ n
m
n1
m
dx
(1 +x)

x
.
N

n=1
(
m
n
)
1
2
1
m +n

N

n=1
_ n
m
n1
m
dx
(1 +x)

x

_ N
m
0
dx
(1 +x)

x

_

0
dx
(1 +x)

x
= .
At all, taking limit n, m , we have

m=1

n=1
a
m
a
n
m +n

m=1
a
2
m
=

m=1
a
2
m
.

Example 2.8 (Hardy Inequality). Let p, q > 0 with

1
p
+
1
q
= 1. Assume

n=1
a
p
n
and

m=1
b
q
n
Then we have:

m=1

n=1
a
n
b
m
m +n


sin(

p
)
(

n=1
a
p
n
)
1
p
(

m=1
b
q
m
)
1
q
.
Hint. Consider the following arguments.
a)
N

m=1
N

n=1
a
m
a
n
m +n
=
N

m=1
N

n=1
(
m
n
)
1
pq
a
m

m +n
(
n
m
)
1
pq
a
n

m +n
.
Analysis 17
b)
(
m
n
)
1
q
1
m +n
=
1
m
(
n
m
)
1
q
(1 +
n
m
)

_ n
m
n1
m
dx
(1 +x)x
1
q
.

Example 2.9 (Hardy-Landan inequality). Let p and a

k
, k = 1, , n be positive numbers. To
show for p > 1, we have:
n

k=1
(
a
1
+ +a
k
k
)
p
< (
p
p 1
)
p
n

k=1
a
p
k
.
Hint. Let
A
k
=
a
1
+ +a
k
k
.
Then, we have:
a)
A
p
k

p
p 1
A
p1
k
a
k
= A
p
k

p
p 1
A
p1
k
(A
k
A
k1
) = A
p
k
(1
kp
p 1
) +
(k 1)p
p 1
A
p1
k
A
k1
A
p
k

p
p 1
A
p1
k
a
k
A
p
k
(1
kp
p 1
) +
k 1
p 1
[(p 1)A
p
k
+A
p
k1
] =
1
p 1
[(k 1)A
p
k1
kA
p
k
].
b) Thus, we have
n

k=1
A
p
k

p
p 1
A
p1
k
a
k
<
p
p 1
n

k=1
A
p1
k
a
k
,
again use Holder inequality.

Example 2.10 (Carleman inequality). Let a

k
, k = N be positive numbers. If

k=1
a
k
< ,
then

k=1
(a
1
a
2
a
k
)
1/k
e

k=1
a
k
.
Another proof directly by elementary technique. We only show that
n

k=1
(a
1
a
2
a
k
)
1/k
e
n

k=1
a
k
.
Let b
j
= j(1 +
1
j
)
j
. Then, we have
b
j
j
e and b
1
b
2
b
k
= (k + 1)
k
,
and so
(a
1
a
2
a
k
)
1/k
=
1
k + 1
k
_
(a
1
b
1
) (a
k
b
k
)

k
i=1
a
i
b
i
k(k + 1)
=
k

i=1
[
1
k

1
k + 1
]a
i
b
i
.
18 Lecture Notes at Fudan
Therefore,
n

k=1
(a
1
a
2
a
k
)
1/k

k=1
k

i=1
[
1
k

1
k + 1
]a
i
b
i
=
n

i=1
n

k=i
[
1
k

1
k + 1
]a
i
b
i
=
n

i=1
a
i
b
i
n

k=i
[
1
k

1
k + 1
],
and so
n

k=1
(a
1
a
2
a
k
)
1/k

i=1
a
i
b
i
[
1
i

1
n
]
n

i=1
a
i
b
i
1
i
=
n

i=1
a
i
b
i
i
e
n

i=1
a
i
.

2.11. Homework.
1. Let A
k
0, t
k
> 0, k = 1, , n. Denote T
l
=

l
k=1
t
k
. Show that
(
n

k=1
A
t
k
k
)
1
T
n

1
T
n
n

k=1
t
k
A
k
.
Hint.Consider the following argument. Let
L
m
=
m

k=1
A
t
k
k
.
Then,we have:
(L
m
)
1
T
m
= (A
t
m
m
)
1
T
m
= (L
1
T
m1
m1
)
T
m1
T
m
A
t
m
T
m
m

T
m1
T
m
L
1
T
m1
m1
+
t
m
T
m
A
m
.
By induction, we prove the statement.

2. To show the Carleman inequality by using the Hardy-Landan inequality.

Proof.By the Hardy-Landan inequality, for all p > 1, we have
n

k=1
(a
1
a
2
a
k
)
1/k

k=1
(
a
1
p
1
+a
1
p
2
+ +a
1
p
k
k
)
p
(
p
p 1
)
p
n

k=1
a
k
.
Since
lim
p+
(
p
p 1
)
p
= e,
we get
n

k=1
(a
1
a
2
a
k
)
1/k
e
n

k=1
a
k
.

2.12 (Appendix). Computing the integral

_

0
dx
(1 +x)x
1
q
with residue theory.
Since
1
q
+
1
p
= 1. Then
_

0
dx
(1 +x)x
1
q
=
_

0
x
1
p
1
1 +x
dx.
Analysis 19
Let f(z) =
1
(1+z)
. f(z) is a holomorphic at C 0. We have
2

1Res
z=1
(z
1
p
1
f(z)) =
_
R

x
1
p
1
f(x)dx +
_

R
e
(
1
p
1)2

1
x
1
p
1
f(x)dx
+
_
|z|=
z
1
p
1
f(z)dz
_
|z|=R
z
1
p
1
f(z)dz
Res
z=1
(z
1
p
1
f(z)) = e
(
1
p
1) log z
[
z=1
= e
(
1
p
1)(log |1|+

1 arg(1))
= e
(
1
p
1)

1
= e
1
p

1
On the other hand,
[f(z)[ 1([z[ << 1), [f(z)[
1
[z[
([z[ >> 1).
Therefore, we have
[
_
|z|=
z
1
p
1
f(z)dz[
1
p
1
2 0( 0),
[
_
|z|=R
z
1
p
1
f(z)dz[
[R[
1
p
1
R
2R 0(R ),
_
R

x
1
p
1
f(x)dx +
_

R
e
(
1
p
1)2

1
x
1
p
1
f(x)dx = (1 e
(
1
p
1)2

1
)
_
R

x
1
p
1
f(x)dx
At all, we have
(1 e
(
1
p
1)2

1
)
_

0
x
1
p
1
f(x)dx = 2

1e
1
p

1
,
and so
_

0
dx
(1 +x)x
1
q
=
2

1e
1
p

1
(1 e
2
p

1
)
=
2

1
e
1
p

1
e

1
p

1
=

sin

p
.
20 Lecture Notes at Fudan
3. Orders Estimate of Infinitesimal
3.1. Notations and Examples.
3.1. Notations.
1. If
lim
xx
o
f(x)
g(x)
= 0,
we denote f(x) = o(g(x)).It is obvious that o(h) +o(g) = o([h[ +[g[).
In particular, let a
n
be a sequence, if
lim
n
a
n
= 0
we denote a
n
= o(1), n .
2. If
lim
xx
o
f(x)
g(x)
= 1,
we say f(x) and g(x) is equivalent for x x
0
, and we denote
f(x) g(x), x x
0
3. Let g(x) > 0, if there is a constant A > 0, such that
[f(x)[ Ag(x), x (a, b),
then we denote f(x) = O(g(x)), x (a, b).
4. Here we have some well-know elementary functions: For n even, we denote n!! = n(n 2)(n
4) 4 2; For n odd, we denote n!! = n(n 2)(n 4) 3 1.
a)
sin x =

k=0
(1)
k
x
2k+1
(2k + 1)!
(x R), and cos x =

k=0
(1)
k
x
2k
(2k)!
(x R),
b)
arctan x =

k=0
(1)
k
x
2k+1
2k + 1
([x[ < 1),
c)
log(1 +x) =

k=1
(1)
k1
x
k
k
([x[ < 1),
d)
(1 +x)

k=0
_

k
_
x
k
([x[ < 1),
e)
e
x
=

k=0
x
k
k!
(x R).
Analysis 21
f)
arctan x =

k=
(1)
n
x
2n+1
2n + 1
(x R).
g)
arcsin x = x +
1
3
1
2!
x
3
+
1
5
3!!
4!!
x
5
+
1
2n + 1
(2n 1)!!
(2n)!!
x
2n+1
+ (x R).
Theorem. In the neighborhood of x
0
, if f
(n)
(x) exists and [f
(n)
(x
0
)[ M then
f(x) =
n1

k=0
f
(k)
(x
0
)
k!
(x x
0
)
k
+O([x x
0
[
n
).
For examples, we have:
a) sin x = x
x
3
3!
+O([x[
5
) (x R), and cos x = 1
x
2
2
+O([x[
4
) (x R),
b) log(1 +x) = x
x
2
2
+O([x[
3
) ([x[ < 1),
c) (1 +x)

= 1 +x +O(x
2
) R, ([x[ < 1),
d) e
x
= 1 +x +

n
k=1
x
k
k!
+O([x[
n+1
) (x R),
e) arctan x = x
x
3
3
+
x
5
5
+O(x
7
)(x R),
f) arcsin x = x +
1
3
1
2!
x
3
+
1
5
3!!
4!!
x
5
+ +O(x
7
) (x R).
Example 3.2. Let > 0 and A be any two constants. Then for any > 0 we have :
1. x
A
= o((1 +)
x
), x ;
2. log x
A
= o(x

), x ;
3. (f(x))
A
= o(e
f(x)
) for any increasing function f(x) with
lim
x
f(x) = .
Proof. Let n = [x] be the integer part of x and m = [A] + 1. If x then n . We have :
(1 +)
x
(1 +)
n
C
m+1
n

m+1

m+1
(m + 1)!
(
n
2
)
m+1
,
if n 2m + 1. Thus, if x , we have
(1 +)
x
x
A


m+1
2
m+1
(m + 1)!
n
m+1
(1 +n)
m
.
Let x = y, we get (1).
Let = e 1, x = log y, we get (2).
(3) is obvious.

Example 3.3. Let a

n
, b
n
be two sequences with a
n
= o(b
n
) n and b
n
> 0. If

n=1
b
n
= ,
then we have:
N

n=1
a
n
= o(
N

n=1
b
n
), N .
22 Lecture Notes at Fudan
Proof. For any > 0 there exists M > 0 such that for any n > M we always have [a
n
[ < b
n
.
Thus, for any N > M, we have:
[
N

n=1
a
n
[ [
M

n=1
a
n
[ +
N

n=M+1
b
n

M

n=1
[a
n
[ +
N

n=M+1
b
n
.
Since

n=1
b
n
= , there exists N
0
> M such that if N > N
0

n=1
b
n
>
M

n=1
[a
n
[.
So
[
N

n=1
a
n
[ 2
N

n=1
b
n
.

Example 3.4. Let a

k
be a sequences with a
k
> 0 and

k=1
a
k
= L < , if

k=n
a
n
= O(a
n
),
then we have:
n

k=1
1
a
k
= O(
1
a
n
).
Proof. Let
A
n
=

k=n
a
n
,
then a
n
= A
n
A
n+1
< A
n
and A
n
= O(a
n
). Thus, there exists a nite constant A > 1 such that
A
n
Aa
n
. We therefore have:
A
n
A
n
A
n+1
A,
A
n+1

A 1
A
A
n
(
A 1
A
)
l+1
A
nl
.
Thus,
n

k=1
1
a
k

A
n
a
n
n

k=1
A
A
k

A
a
n
n

k=1
A
n
A
k

A
a
n
n1

l=0
(
A 1
A
)
l
<
A
a
n

l=0
(
A 1
A
)
l
=
A
2
a
n
.

Example 3.5 (Stolz theorem). Let y

n
is strictly increasing sequence, i.e.,y
n+1
> y
n
, and y
n
.
If
lim
n
x
n
x
n+1
y
n
y
n+1
= a, a [, +],
Then
lim
n
x
n
y
n
= a.
Analysis 23
Proof. In case of a = +, there is a N N such that x
n+1
x
n
> y
n+1
y
n
for n > N and
x
n
+, and
lim
n
y
n
y
n+1
x
n
x
n+1
= 0.
In case of a = , let t
n
= x
n
. Thus these two cases are reduced to the case nite constant a.
Now we assume that a is a nite constant. Then we have
x
k
x
k1
= a(y
k
y
k1
) +o(y
k
y
k1
),
sum with k, we get
n

k=2
(x
k
x
k1
) =
n

k=2
a(y
k
y
k1
) +
n

k=2
o(y
k
y
k1
).
Since y
n+1
> y
n
n N and y
n
we have
n

k=2
o(y
k
y
k1
) = o(
n

k=2
y
k
y
k1
) = o(y
n
y
1
) = o(y
n
).
Thus,
x
n
x
1
= a(y
n
y
1
) +o(y
n
).

3.2. Exercises and homework.

3.6. Exercises.
1. To show
lim
n
n(
n

n 1)
log n
= 1.
Proof.
n

n = exp(
log n
n
) = 1 +
log n
n
+O(
log
2
n
n
2
),
n(
n

n 1) = log n +O(
log
2
n
n
).

Another proof given by a student.Let t =

n

n 1. Then
t 0
+
(n ) and n = (1 +t)
n
,
and so
n(
n

n 1)
log n
=
nt
nlog(1 +t)
=
t
log(1 +t)
=
1
log(1 +
1
1/t
)
1/t
log e = 1 (n ).

2. To prove for x > 0,

lim
n
n
2
(
n

x
n+1

x) = log x.
24 Lecture Notes at Fudan
Proof.
n
2
(
n

x
n+1

x) = n
2
x
1/n
(1 x
1/(n(n+1))
) = n
2
x
1/n
(1 exp(
log x
n(n + 1)
)).
exp(
log x
n(n + 1)
) = 1
log x
n(n + 1)
+O(
1
n
4
),
so we have:
n
2
(
n

x
n+1

x) = n
2
x
1/n
(
log x
n(n + 1)
+O(
1
n
4
)) = x
1/n
log x +O(
1
n
2
).

3. To prove
lim
n
cos
n
(
x

n
) = e

x
2
2
.
Proof.
(cos(
x

n
))
n
= (1
x
2
2n
+O(
1
n
2
))
n
= exp(nlog(1
x
2
2n
+O(
1
n
2
))),
Since
log(1
x
2
2n
+O(
1
n
2
)) =
x
2
2n
+O(
1
n
2
),
we have:
(cos(
x

n
))
n
= e

x
2
2
(1 +O(
1
n
)).

Another proof given by a student.

(cos(
x

n
))
n
= (1 sin
2
(
x

n
))
n
2
((1
x
2
n
)
n
x
2
)

x
2
2
e

x
2
2
, (n ).

4. To prove for 0 < < 2, we have

lim
x
_
x +
_
x +

x =
1
2
.
Proof.
_
x +
_
x +

x(1 +
_
1
x
+x

2
2
)
1/2
=

x(1 +
1
2

x
_
1 +x

2
1
) +O(
1

x
),
thus
_
x +
_
x +

x =
1
2
(1 +
1
2
x

2
1
+O(x
2
)) +O(
1

x
).

5. To show
lim
x0
(
x
tan x
)
1
x
2
= e
1
3
.
Analysis 25
Proof.
x
tan x
=
x cos x
sin x
= x
1
x
2
2
+O(x
4
)
x
x
3
6
+O(x
5
)
= (1
x
2
2
+O(x
4
))(1 +
x
2
6
+O(x
4
)) = 1
x
2
3
+O(x
4
),
(
x
tan x
)
1
x
2
= exp(
log 1
x
2
3
+O(x
4
)
x
2
) = exp(
x
2
3
+O(x
4
)
x
2
) = e
1
3
+O(x
2
)
.

3.7. Homework.
1. Euler constant.To show that
lim
n
n

k=1
1
k
log n = < .
Proof.Let
a
n
=
n1

k=1
1
k
log n,
b
n
=
n

k=1
1
k
log n.
Then b
n
a
n
=
1
n
and
a
n+1
a
n
=
1
n
(log(n + 1) log n) =
_
n+1
n
(
1
n

1
t
)dt > 0,
b
n+1
b
n
=
1
n + 1
(log(n + 1) log n) =
_
n+1
n
(
1
n + 1

1
t
)dt < 0.
Thus
lim
n
a
n
= lim
n
b
n
= < .

2. In the textbook, e is dened as

e := lim
n
(1 +
1
n
)
n
.
Show that
e =

k=0
1
k!
.
Proof.Let
a
n
=
n

k=0
1
k!
,
b
n
= (1 +
1
n
)
n
.
then
b
n
= 1 + 1 +
1
2!
(1
1
n
) +
1
3!
(1
1
n
)(1
2
n
) + +
1
n!
(1
1
n
)(1
2
n
) (1
n 1
n
) a
n
On the other hand, given any k N, for n > k we have
b
n
> 1 + 1 +
1
2!
(1
1
n
) +
1
3!
(1
1
n
)(1
2
n
) + +
1
k!
(1
1
n
)(1
2
n
) (1
k
n
).
26 Lecture Notes at Fudan
Thus,
e lim
n
1 + 1 +
1
2!
(1
1
n
) +
1
3!
(1
1
n
)(1
2
n
) + +
1
k!
(1
1
n
)(1
2
n
) (1
k
n
) = a
k
.
Therefore,
b
k
a
k
e, k N, and so lim
k
a
k
= e.

3. To show that for (1, )

k=n
1
k

= O(
1
n
1
),
n

k=1
1
k

= C +O(
1
n
1
),
where C is a constant.
Proof.Since
_
k+1
k
dx
x

<
1
k

<
_
k
k1
dx
x

,
So

k=n
1
k

<
_

n1
dx
x

=
1
1
_

n1
dx
1
= O(
1
n
1
).

Analysis 27
4. Application of Infinitesimal: Wallis Formula and Stirling Formula
4.1. (Wallis Formula).
1. Let n N. If n is even, we denote
n!! = n(n 2)(n 4) 4 2.
If n is odd, we denote
n!! = n(n 2)(n 4) 3 1.
Show the integral
J
m
=
_
2
0
sin
m
xd x =
_

2
(m1)!!
m!!
m is even ;
(m1)!!
m!!
, m is odd.
Proof.By partial integral, we have
J
m
=
_
2
0
sin
m1
x cos xd x
= sin
m1
x cos x[

2
0
(m1)
_
2
0
sin
m2
x cos
2
xd x
= (m1)J
m2
(m1)J
m
,
and so,
J
m
=
m1
m
J
m1
.

2. To show the Wallis Formula

2
= lim
n
1
2n + 1
[
(2n)!!
(2n 1)!!
]
2
.
Proof.From
_
2
0
sin
2n+1
xd x
_
2
0
sin
2n
xd x
_
2
0
sin
2n1
xd x,
we have
(2n)!!
(2n + 1)!!

(2n 1)!!
(2n)!!


2

(2n 2)!!
(2n 1)!!
(4.1.1)
1
2n + 1
[
(2n)!!
(2n 1)!!
]
2

2

1
2n
[
(2n)!!
(2n 1)!!
]
2
But by the above inequality, we have
1
2n
[
(2n)!!
(2n 1)!!
]
2

1
2n + 1
[
(2n)!!
(2n 1)!!
]
2
=
1
2n(2n + 1)
[
(2n)!!
(2n 1)!!
]
2

4n
0 (n ).
Therefore,
lim
n
1
2n + 1
[
(2n)!!
(2n 1)!!
]
2
= lim
n
1
2n
[
(2n)!!
(2n 1)!!
]
2
=

2
.

4.2. (Stirling Formula).

28 Lecture Notes at Fudan
1. To show
n

k=1
1
k
= log n + +O(
1
n
),
where is the EuLer constant.
Proof.Since
log(1 +
1
k
) =
1
k
+C
k
where C
k
= O(
1
k
2
). we have
n

k=1
1
k
=
n

k=1
log(1 +
1
k
) +
n

k=1
C
k
=
n

k=1
log(1 +
1
k
) +

k=1
C
k

k=n+1
C
k
= log n + +O(
1
n
).

2. To show that
(4.2.1) log n! = nlog n n +
1
2
log n +C +O(
1
n
),
where C is a constant, i.e.,
n! Ae
n
n
n+
1
2
.
Proof.
log n! = log n
n1

k=1
= log n+
_
n
1
log tdt
n1

k=1
_
k+1
k
log
t
k
dt = (n+1) log nn+1
n1

k=1
_
1
0
log
t +k
k
dt
log n! = (n + 1) log n n + 1
n1

k=1
_
1
0
log(1 +
t
k
)dt = (n + 1) log n n + 1
n1

k=1
_
1
0
(
t
k
+O(
t
2
k
2
))dt
log n! = (n + 1) log n n + 1
1
2
(
n1

k=1
(
1
k
+O(
1
k
2
)))
Since
n

k=1
1
k
= log n + +O(
1
n
),
where is the Euler constant, we have:
log n! = nlog n n +
1
2
log n +C +O(
1
n
),
i.e.,
n! = Ae
n
n
n+
1
2
(1 +O(
1
n
)).

3. To show
1
n +
1
2
< log(1 +
1
n
) <
1
2
(
1
n + 1
+
1
n
).
Analysis 29
Proof.The right inequality is clear by the picture of area and the convex of picture of
y = f(x) = 1/x.
The left inequality is equivalent to the following inequality :
(n +
1
2
) log(1 +
1
n
) 1 > 0.
If n = 1, it is obvious. If n 2, we consider the following formula:
log(1 +x) =

k=1
(1)
k1
x
k
k
([x[ < 1).
Thus,for n 0 we have
(n +
1
2
) log(1 +
1
n
) 1 = (n +
1
2
)(
1
n

1
2n
2
+
1
3n
3
+O(
1
n
4
)) 1 =
1
12n
2
+O(
1
n
3
) > 0.
The equality is true for all n, but we only need the inequality for n 0 to prove the Stirling
formula.
4. Use the above inequality to give another proof of the following formula:
n! Ae
n
n
n+
1
2
(n ).
Proof.Let
a
n
=
n!
e
n
n
n+
1
2
.
Then we have:
log
a
n
a
n+1
= (n +
1
2
) log(1 +
1
n
) 1.
From the the above inequality, we have
0 < log
a
n
a
n+1
<
1
4
(
1
n

1
n + 1
),
thus a
n
is a strictly decreasing sequence and b
n
is a strictly decreasing sequence where b
n
=
a
n
e

1
4n
. Moreover,
0 < a
n
b
n
= a
n
(1 e

1
4n
) a
1
(1 e

1
4n
) 0 (n ).
Therefore, there exists a nite constant A such that
lim
n
a
n
= lim
n
b
n
= A.

5. Use the Wallis Formula to show that the constant A is exact

2.
Proof.Let n N. If n is even, we denote
n!! = n(n 2)(n 4) 4 2.
If n is odd, we denote
n!! = n(n 2)(n 4) 3 1.
The Wallis Formula says that

2
= lim
n
1
2n + 1
[
(2n)!!
(2n 1)!!
]
2
.
30 Lecture Notes at Fudan
Thus

2
= lim
n
1
2n + 1

[(2n)!!]
2
(2n)!

2
= lim
n
1
2n + 1

(2
n
n!)
2
(2n)!

2
= lim
n
1
2n + 1

2
2n
(n!)
2
(2n)!

2
.
on the other hand, we have
n! Ae
n
n
n+
1
2
and (2n)! Ae
2n
(2n)
2n+
1
2
(n )
Thus

2
= lim
n
1
2n + 1

2
2n
A
2
n
2n+1
e
2n
A(2n)
2n+1/2
e
2n

2
= lim
n
1
2n + 1
(
_
n
2
A)
2
=
A
2
4
.
At all, we obtain
A =

2,
and so
(4.2.2) n! =

2n(
n
e
)
n
(1 +O(
1
n
))

5. Topic on the Gamma functions

5.1. functions.
5.1. Denition and basic properties. Refer to the textbook. The Gamma function is dened by
(5.1.1) (x) =
_

0
t
x1
e
t
dt, x > 0.
The Gamma function has the following properties:
1. (x) C

((0, )), moreover

(n)
(x) =
_

0
e
t
t
z1
(log t)
n
dt.
2.
(x + 1) = x(x),
so that
lim
x0+
(s) = +
and for any n N,
(n) = n!.
3. For any , > 0,
(5.1.2)
()()
( +)
=
_

0
t
1
(1 +t)
+
dt =
_
1
0
y
1
(1 y)
1
dy.
so that
(5.1.3) (x)(1 x) =

sin(x)
, 0 < x < 1
and
(
1
2
) =

.
Analysis 31
Proof.Put y = xt in the formula 5.1.1, there is
()() =
_

0
x
1
e
x
dx
_

0
y
1
e
y
dy
=
_

0
x
1
e
x
dx
_

0
x

t
1
e
xt
dt
=
_

0
x
+1
e
x
dx
_

0
t
1
e
xt
dt
=
_

0
t
1
dt
_

0
x
+1
e
x(1+t)
dx.
Put back y = x(1 +t) in the above formula, there holds
()() =
_

0
t
1
dt
_

0
x
+1
e
x(1+t)
dx
=
_

0
t
1
dt
_

0
y
+1
e
y
(1 +t)
+
dy
= ( +)
_

0
t
1
(1 +t)
+
dt.
Put t = tan
2
, we obtain
_

0
t
1
(1 +t)
+
dt = 2
_
2
0
(sin )
21
(cos )
21
d =
_
1
0
y
1
(1 y)
1
dy.

4. Legendre formula
(5.1.4) 2
2x1
(x)(x +
1
2
) = (
1
2
)(2x).
Proof.Put = = x into the formula 5.1.2, there holds
((x))
2
(2x)
=
_
1
0
y
x1
(1 y)
x1
dy = 2
_
1/2
0
y
x1
(1 y)
x1
dy.
Let y = (1

t)/2, we the obtain

2
_
1/2
0
y
x1
(1 y)
x1
dy =
1
2
_
1
0
(
1 t
4
)
x1
t
1
2
dt
= 2
12x
_
1
0
(1 t)
x1
t
1
2
dt
= 2
12x
(x)(
1
2
)
(x +
1
2
)
.

Lemma 5.2. Let a be any constant, then there holds

(5.2.1)
(x)
(x +a)
= x
a
+O(x
a1
).
32 Lecture Notes at Fudan
Proof. If a < 1 then there is a k N with a +k > 1, and so
(x)
(x +a)
=
(x)
(x +a +k)
k

i=1
(x +a +i).
Thus we reduce the question to the case a > 1. Now we suppose that a > 1. From 5.1.2, there is
(x)(a)
(x +a)
=
_
1
0
y
x1
(1 y)
a1
dy =
_

0
(1 e
t
)
a1
e
xt
dt = I
1
+I
2
.
I
1
=
_ 1

x
0
(1 e
t
)
a1
e
xt
dt
=
_ 1

x
0
(t +O(t
2
))
a1
e
xt
dt
=
_ 1

x
0
(t)
a1
(1 +O(t))e
xt
dt
=
_ 1

x
0
(t)
a1
e
xt
dt +O(
_ 1

x
0
(t)
a
e
xt
dt)
= x
a
_

x
0
(t)
a1
e
t
dt +O(x
a1
_

x
0
(t)
a1
e
t
dt)
= x
a
(a) +O(x
a1
).
I
2
=
_

1

x
(1 e
t
)
a1
e
xt
dt
= O(
_

1

x
e
xt
dt)
= O(x
a1
)
At all , we obtain
(x)(a)
(x +a)
= x
a
(a) +O(x
a1
),
i.e.,
(x)
(x +a)
= x
a
+O(x
a1
).

Theorem 5.3 (Stirlings Formula).

(5.3.1) (x) = x
x
1
2
e
x

2(1 +O(
1
x
)), x > 0.
In particular, if x = n N, there is the Stirling formula:
n! =

2n(
n
e
)
n
(1 +O(
1
n
)).
Proof. The formula 9.16.3 says:
log (n) = nlog n n +
1
2
log n +C +O(
1
n
),
Analysis 33
where C is a constant. Let x = n +a where n N and 0 < a < 1. The formula 5.2.1 says that
log (x) = log (n +a) = log (n) +a log n +O(
1
n
).
The formula 9.16.3 says:
log (n) = nlog n n +
1
2
log n +C +O(
1
n
),
where C is a constant, thus we obtain
log (x) = (n
1
2
) log n n +C +a log n +O(
1
n
)
= (x a
1
2
) log(x a)n x +a +C +a log(x a) +O(
1
x
)
= (x
1
2
) log x x +C +O(
1
x
)
On the other hand, the Legendre formula says
(2x 1) log 2 + log (x) + log (x +
1
2
) = log (
1
2
) + log (2x).
We then have
(5.3.2)
(2x
1
2
) log 2x2x+C+log (
1
2
)+O(
1
x
) = (2x1) log 2+(x
1
2
) log x+x log(x+
1
2
)2x
1
2
+2C+O(
1
x
)
Therefore,
C = log

2.

Theorem 5.4 (Euler Product Formula).

(5.4.1) (s) =
1
s

n=1
(1 +
1
n
)
s
(1 +
s
n
)
1
, s > 0
Proof. 1. First, we show that
(5.4.2) (s) =
_

0
t
s1
e
t
dt = lim
n
_
n
0
t
s1
(1
t
n
)
n
dt
Let
_
n
0
t
s1
(1
t
n
)
n
dt =
_
n
1
3
0
t
s1
(1
t
n
)
n
dt +
_
n
n
1
3
t
s1
(1
t
n
)
n
dt = I
1
+I
2
.
34 Lecture Notes at Fudan
I
1
=
_
n
1
3
0
t
s1
(1
t
n
)
n
dt
=
_
n
1
3
0
t
s1
e
nlog n(1
t
n
)
dt
=
_
n
1
3
0
t
s1
e
t+o(n
1
3 )
dt
=
_
n
1
3
0
t
s1
e
t
dt +O(n
1
3
)
= (s) +O(n
1
3
).
I
2
=
_
n
n
1
3
t
s1
(1
t
n
)
n
dt = I
1
+I
2
.
= O(
_
n
n
1
3
t
s1
(1
n
1
3
n
)
n
dt)
= O(
n
s
s
e
n
1
3
).
2. We then show that
(5.4.3)
_
n
0
t
s1
(1
t
n
)
n
dt ==
(1 +
s
n
)
1
s
n1

k=1
(1 +
1
k
)
s
(1 +
s
k
)
1
.
_
n
0
t
s1
(1
t
n
)
n
dt = n
s
_
1
0
(1 t)
n
t
s1
dt =
n
s
s
_
1
0
(1 t)
n
dt
s
= n
s
n
s
_
1
0
(1 t)
n1
t
s
dt.
By induction, we obtain:
_
n
0
t
s1
(1
t
n
)
n
dt = n
s
n(n 1) 1
s(s + 1) (s +n 1)
_
1
0
t
s+n1
dt = n
s
n(n 1) 1
s(s + 1) (s +n)
,
so
_
n
0
t
s1
(1
t
n
)
n
dt =
(1 +
s
n
)
1
s
n1

k=1
(1 +
1
k
)
s
(1 +
s
k
)
1
.

To be Continuous .....
5.2. Exercises.
Exercise 5.5. To show that for n = 1, 2, 3, , there holds
(xe
2(xn)
)
n
= O(e
(x
2
+x)
), x > 0
Proof. Consider the function
f
n
(x) = xe
2(xn)
)
n
e
(x
2
+x)
.
Then
f

n
(x) = e
2n(xn)x
2
x
(nx
n1
+x
n
(2x + 2n 1)),
and f

n
(n) = 0,f
n
(x) gets the maximal value at x = n,
f
n
(n) = e
nlog nn
2
n
= O(1).

Analysis 35
Exercise 5.6. For any > 0, to show that
log x = o(x

).
Exercise 5.7. Dene
Lix = lim
0
(
_
1
0
+
_
x
1
)
dt
log t
.
To show that
Lix
x
log x
.
Exercise 5.8. To show that
_
1
0
(log
1
t
)
n
dt = n!.
Exercise 5.9. Denote be the Euler constant. For any x > 0, show that

nN,nx
1
n
= log x + +O(
1
x
).
Exercise 5.10. To show that

(1) = lim
n
(
_
1
0
1 y
n
1 y
dy log n) = lim
n

i=1
1
i
log n = .
36 Lecture Notes at Fudan
6. Working Technique in function theory
6.1. Iteration technique.
Example 6.1 (De Giorgi-Nash iteration). Let (t) be a nonnegative function decreasing on [k
0
, ].
Assume that satises
(6.1.1) (h) (
M
h k
)

[(k)]

, h > k k
0
,
where M, > 0, > 1. Then
(k
0
+d) = 0
where
d = M[(k
0
)]
1

1
.
Proof. Dene sequence k
n
by the following rule
k
n
= k
0
+d
d
2
n
, n = 0, 1, 2, .
Then (k
n
) is a decreasing sequence. From the rule 6.2.1, we have
(k
n+1
)
M

2
(n+1)
d

[(k
n
)]

, n = 0, 1, 2, .
If we can show that
(6.1.2) (k
n
)
(k
0
)
r
n
for some r > 1, n = 0, 1, 2, ,
where r is under-determined. Then, we have
0 (k
0
+d) (k
n
)
(k
0
)
r
n
, n = 0, 1, 2, .
Take the limit, we then obtain the result.
Now, we choose the number r an prove the inequality 6.1.2 by using induction : if the inequality
is true for n, then, we have
(k
n+1
)
M

2
(n+1)
d

[(k
n
)]

(k
0
)
r
n+1
M

2
(n+1)
d

r
n(1)1
[(k
0
)]
1
.
Choose r 2
/(1)
, we have
M

2
(n+1)
d

r
n(1)1
[(k
0
)]
1

2
/(1)
d

[(k
0
)]
1
1.

Example 6.2. Let (t) be a nonnegative function increasing on [0, R

0
]. Assume that for some
, [0, 1) and (0, 1], K 0, satises
(6.2.1) (R) (R) +KR

, R (0, R
0
].
Then, we have
(6.2.2) (R) C(
R
R
0
)

[(R
0
) +KR

0
], R (0, R
0
],
where = (, , ) (0, ) and C = C(, , ) > 0 are constant.
Analysis 37
Proof. Assume that

> 1. Let

R
i
=
i

R
0
, i = 0, 1, 2, . From 6.2.1, we have
(

R
i+1
)

R
i
) +K

i
,
and by iteration, we have
(

R
i
)
i
(

R
0
) +
i1

m=0
K
m

im1
,

i
(

R
0
) +K

(i1)
i1

m=0
(

)
m
=
i
(

R
0
) +K

(i1)
(

)
i
1

1

i
(

R
0
) +K

(i1)
(

)
i

1
=
i
[(

R
0
) +C
1
K

0
],
where
C
1
=

1
.
Since
i =
log(

R
i

R
0
)
log
,
let = log / log (o, ), we have
(

R
i
) (

R
i

R
0
)

[(

R
0
) +C
1
K

0
], i = 0, 1, 2 .,
When

R
0
(R
0
, R
0
], let C
2
= max(C
1
, 1), we have
(

R
i
)
C
1

R
i
R
0
)

[(R
0
) +KR

0
], i = 0, 1, 2 .,
On the other hand, for any R [0, R
0
], there are i N,

R
0
(R
0
, R
0
] such that R =

R
i
. Let
C =
C
1

, we then have the inequality 6.2.2.

Example 6.3. Let 0 T

0
< T
1
, (t) be a nonnegative bounded function on [T
0
, T
1
]. Assume that
for some t, s with 0 T
0
t < s T
1
, satises
(6.3.1) (t) (s) +
A
(s t)

+B,
where , A, B, are non-negative constant and < 1. Then, we have
(6.3.2) () C(
A
(R )

+B), T
0
< R T
1
,
where C = C(, ) > 0 is a constant.
Proof. Let T
0
< R T
1
. Dene
t
0
= , t
i+1
= t
i
+ (1 )
i
(R ) (i = 0, 1, 2, ),
38 Lecture Notes at Fudan
where (0, 1) is under-determined. By the relation 6.3.2, we have
(t
i
) (t
i+1
) +
A
((1 )
i
(R ))

+B, i = 0, 1, .
by induction, for any k 1 we have
(t
0
)
k
(t
k
) + (
A
(1 )

(R )

+B)
k1

i=0
(

)
i
.
Choose with

< 1, and let k , we then obtain the inequality 6.3.2.

Example 6.4. Let 0 T

0
< T
1
, (t) be a nonnegative function increasing on [0, R
0
]. Assume that
for some constants , with 0 < < , satises
(6.4.1) () C(

R
)

[(R) +BR

], 0 < < R R
0
.
where C is a constant. Then, we have
(6.4.2) () C(

R
)

[(R) +BR

], 0 < < R R
0
,
where C = C(A, , ) > 0 is a constant.
Proof. Let =
1
2
( +). Choose (0, 1) such that A

< 1. Then, we have

(R) A

(R) +BR

= A

(R) +BR

(R) +BR

,
and
(
k+1
R) =
(k+1)
(R) +B(
k
+
(k1)+
+ +
k
)R

=
(k+1)
(R) +B
k
(
k()
+
(k1)()
+ + 1)R

=
(k+1)
(R) +B

k
(1
(k+1)()
1
()
R

C
1

k
[(R) +BR

],
Where C
1
1 is a constant independent on k. Thus,
(6.4.3) (
k+1
R) C
1

(k1)
[(R) +BR

], k 0.
for any with 0 < < R R
0
, we choose any positive integer k such that

k+1
R <
k
R,
then we have
() (
k
R) C
1

(k1)
[(R) +BR

]
C
1

2
(

R
)

[(R) +BR

]
= C(

R
)

[(R) +BR

],
where C = C
1

2
.

Analysis 39
6.2. Exercises and Homework.
Exercise 6.5. Let f : R R be a continuous function. dene
f
(n)
(x) = f f f
. .
n
.
Assume that there exist a, C R with C > o such that for all n N,
[f
(n)
(a)[ < C.
Then, the function f has a xed point, i.e., x
0
R, f(x
0
) = x
0
.
Proof. Let g(x) = f(x) x. We can assume
g(x) > 0, x R or g(x) > 0, x R.
Otherwise, by the Cauchy mean value theorem we will have ar least a zero point of g.
a) If
g(x) > 0, x R
we will get a bounded increasing x
n
dened by
x
0
= a, x
n+1
= f(x
n
), n = 1, 2 .
Since A = lim
n
x
n
exists, by f is continuous we then have
A = lim
n
x
n+1
= lim
n
f(x
n
) = f( lim
n
x
n
) = f(A).
b) If
g(x) < 0, x R
we will get a bounded decreasing y
n
dened by
y
0
= a, y
n+1
= f(y
n
), n = 1, 2 .
Similarly, B = lim
n
y
n
exists and is a xed point of f.

Exercise 6.6. Let f : R R be a continuous function satisfying

f(2x
2
1) = 2xf(x), xR.
To show that
f(x) 0, x [1, 1].
Proof. We have f(x) = f(x), x ,= 0. By the continuity of f, f(0) = f(0), and so f(0) = 0.
Now, for any , we have
f(cos 2) = 2 cos f(cos ),
and so
f(cos ) = 0 f(cos 2) = 0.
Again the continuity of f, for any , we have
f(cos 2)
sin 2
=
f(cos )
sin
.
On the other hand,
f(cos 2) = f(2 sin
2
1) = 2 sin f(sin ),
then
cos f(cos ) = sin f(sin ),
40 Lecture Notes at Fudan
and so
f(sin )
cos
=
f(cos )
sin
.
Therefore, for any t = sin [1, 1], we have
f(sin )
cos
=
f(sin

2
)
cos

2
= =
f(sin

2
n
)
cos

2
n

f(0)
1
= 0 (n ).

Exercise 6.7. Let f(x), g(x) be functions dened on (a, +). Suppose that f(x), g(x) satisfy the
following conditions :
a) f(x), g(x) are bounded on any nite interval (a, b);
b)
g(x + 1) > g(x), x (a, +);
c)
lim
x+
g(x) = +.
If
lim
x+
f(x + 1) f(x)
g(x + 1) g(x)
= l,
where l [, +]. Then, we have
lim
x+
f(x)
g(x)
= l.
Hint. Refer to the proof of the Stolz theorem.

Exercise 6.8. (1) Show that the set of all irrational number of R is uncountable.
(2) Let f : R R be a function satisfying
f(Q) RQ, and f(RQ) Q.
To show that f is not a continuous function, and show an example.
Analysis 41
7. Applications of Differential
Example 7.1 (Liouville). Suppose R is an algebraic integer of degree d > 0, i.e, is a root of
a polynomial with integer coecients
f(x) = a
0
x
d
+a
1
x
d1
+ +a
d
, a
i
Z,
which is irreducible in Q[x]. Then there exists a constant c = C() > 0 such that for each pair of
integers (p, q), there holds
[
p
q
[
c
[q[
d
.
Proof. For a given pair (p, q), if [
p
q
[ > 1 then let c = 1. Thus,We suppose [
p
q
[ 1. We have:
f(
p
q
) = f(
p
q
) f() = f

()(
p
q
), for some [[[ 1, [[ + 1].
Let M be the maximal value of [f

(x)[ on [[[ 1, [[ + 1], and let c = 1/M.

Since f(
p
q
) ,= 0 for any national number
p
q
(we omit the proof here), so
[
p
q
[
[f

()[[
p
q
[
M
= c[f(
p
q
)[ = c
[

d
i=0
a
i
p
di
q
i
[
[q[
d
c
1
[q[
d
.

Exercise 7.2. For any n N, dene

P
n
= 1 +
x
1!
+
x
2
2!
+ +
x
n
n!
,
and denote F
n
= P
n+1
(x)P
n
(x). To show that F
n
(x) has a unique root in R.
Hint. Suciently to prove that if n is even then P
n
(x) has no root in R, and if n is odd then P
n
(x)
has a unique root in R.
To be continuous......
42 Lecture Notes at Fudan
8. Treasures in Calculus
8.1. Eulers identication (2) = 1 +
1
2
2
+
1
n
2
+

2
6
. At rst, there holds
_
1
0
_
1
0
1
1 xy
dxdy =
_
1
0
_
1
0

i0
x
i
y
i
dxdy
=

i0
_
1
0
_
1
0
x
i
y
i
dxdy
=

i0
_
1
0
x
i
dx
_
1
0
y
i
dy
=

i0
1
(i + 1)
2
= (2).
On the other hand, replacing x
2
with X and y
2
with Y, there holds
_
1
0
_
1
0
(
1
1 xy

1
1 +xy
)dxdy =
_
1
0
_
1
0
(
2xy
1 x
2
y
2
)dxdy
=
1
2
_
1
0
_
1
0
1
1 XY
dXdY
=
(2)
2
and let x = sin / cos , y = sin / cos , we obtain
_
1
0
_
1
0
(
1
1 xy
+
1
1 +xy
)dxdy = 2
_
1
0
_
1
0
(
1
1 x
2
y
2
)dxdy
= 2
_
2
0
d
_
2

0
1
1 (
sin
cos
)
2
(
sin
cos
)
2
(1 tan
2
tan
2
)d
=

2
4
Therefore, we obtain
2(2) =
_
1
0
_
1
0
(
1
1 xy

1
1 +xy
)dxdy +int
1
0
_
1
0
(
1
1 xy
+
1
1 +xy
)dxdy
=
(2)
2
+

2
4
,
and so
(8.0.1) (2) = 1 +
1
2
2
+
1
n
2
+

2
6
Analysis 43
8.2. Irrationality of , log 2, (2), (3).
8.1. Legendre polynomial
1. For any n N, the Legendre polynomial is dene by
(8.1.1) P
n
(x) =
1
n!
d
n
dx
n
(x
n
(1 x)
n
).
It is obvious all coecients of P
n
are integers, i.e.,
P
n
=
n

j=0
p
n,j
x
j
, p
n,j
Z.
For examples, P
0
(x) = 0, P
1
(x) = 1 2x, P
2
(x) = 2 12x 12x
2
.
2. The following is an important property of Legendre polynomials:
(8.1.2)
_
1
0
P
n
(x)f(x)dx = (1)
n
_
1
0
1
n!
x
n
(1 x)
n
d
n
f(x)
dx
n
dx.
3. A technique to show a number is irrational by using Legendre polynomials.
Suppose there is a family of integrals such that (j N)
_
1
0
x
j
f(x) = R
j
+S
j
,
where R
j
, S
j
Q and f(x) is an under-determined function.
Assume is rational, then
_
1
0
P
n
(x)f(x) =
A
n
B
n
with A
n
, B
n
Z, n N.
On the other hand, if A
n
,= 0n >> 0 and
[B
n
(1)
n
_
1
0
1
n!
x
n
(1 x)
n
d
n
f(x)
dx
n
dx[ 0 (n )
then there is a coniction:
1 [A
n
[ 0 (n ),
and so is irrational.
8.2. is irrational.
1. It is obvious for any j N,
_
1
0
x
j
sin(x)dx = h(
1

),
where h(x) is a polynomial in Z[x] with degree at most j.
2. Thus, if = a/b is rational then for any Legendre polynomial P
n
(x), there holds
_
1
0
P
n
(x) sin(x)dx =
A
n
a
n
with A
n
Z.
44 Lecture Notes at Fudan
3. On the other hand,by stirling formula, there is
[A
n
[ = [a
n
_
1
0
P
n
(x) sin(x)dx[
= [a
n
_
1
0
1
n!
x
n
(1 x)
n
d
n
dx
n
(sin(x))dx[
= [a
n
_
1
0
1
n!
x
n
(1 x)
n

n
dx[
[
1
n!
(
a
4
)
n
[ 0(n ).
8.3. log 2 is irrational. In this case, we choose f(x) = 1/(1 +x).
1. Since x
j
= x
j1
(x + 1) x
j2
(x + 1) + (x + 1) 1 for j N, it is obvious that
_
1
0
x
j
1 +x
dx =
1
j

1
j 1
+ 1 log 2 = C
j
log 2,
where C
j
Q.
2. Thus, if log 2 = a/b is rational then for any Legendre polynomial P
n
(x), there holds
_
1
0
P
n
(x)
1
x + 1
dx =
A
n
bd
n
with A
n
Z,
and d
n
= LCM(1, 2, , n).
3. On the other hand,
[A
n
[ = [bd
n
_
1
0
P
n
(x)
1
(1 +x)
dx[
= [bd
n
_
1
0
1
n!
x
n
(1 x)
n
[
d
n
dx
n
1
x + 1
]dx[
= [bd
n
_
1
0
(
x(1 x)
1 +x
)
n
1
1 +x
dx[.
Since
max
x[0,1]
(
x(1 x)
1 +x
) = 3 2

2,
and d
n
3
n
,
[A
n
[ [b[(3(3 2

2))
n
0.
8.4. (2) is irrational. In this case, we choose f(x) =
_
1
0
(1y)
n
1xy
dy.
1. Consider the integral family(j N)
_
1
0
x
j
[
_
1
0
(1 y)
n
1 xy
dy]dx =

r,sN
p
j;r,s
_
1
0
_
1
0
x
r
y
s
1 xy
dydx, p
j;r,s
Z,
By Euclieans division method, for any pair (r, s) N
2
, we have:
_
1
0
_
1
0
x
r
y
s
1 xy
dydx =

p,qN
A
r,s;p,q
_
1
0
_
1
0
x
p
y
q
dydx +

pN
B
r,s;p
_
1
0
_
1
0
x
p
1 xy
dydx
+

qN
C
r,s;q
_
1
0
_
1
0
y
q
1 xy
dydx +D
r,s
_
1
0
_
1
0
1
1 xy
dydx.
Analysis 45
Therefore, we obtain
_
1
0
_
1
0
x
r
y
s
1 xy
dydx =
_
k
r,s
d
2
n+1
, r ,= s
(2)

r
i=1
1
i
2
, r = s,
where each k
r,s
is an integer and d
n
= LCM(1, , n).
2. Suppose (2) = a/b is rational, then or any Legendre polynomial P
n
(x),
[
_
1
0
P
n
(x)f(x)dx[ =
[A
n
[
bd
2
n+1
, A
n
Z.
3. On the other hand,
[A
n
[ = [bd
2
n+1
_
1
0
P
n
(x)f(x)dx[
= [bd
2
n+1
_
1
0
1
n!
x
n
(1 x)
n
[
d
n
dx
n
f(x)]dx[
= [bd
2
n+1
_
1
0
_
1
0
(
x(1 x)y(1 y)
1 +xy
)
n
1
1 xy
dydx[.
Since
d
n+1
3
n+1
,
and
max
(x,y)[0,1][0,1]
(
x(1 x)y(1 y)
1 xy
) = 3 2

2 = (
1 +

5
2
)
5
,
then
[A
n
[ [9b(2)[(9(
1 +

5
2
)
5
)
n
0,
since
9(
1 +

5
2
)
5
=
45

5 99
2
< 1.
Exercise 8.5. To show that
(3) =

i=1
1
i
3
is irrational.
Hint. Consider the function
f(x) =
_
1
0
P
n
(y)
1 xy
log xydy,
and the family of integrals(j N)
_
1
0
x
j
f(x)dx.
Show that if (3) = a/b is rational then
[
_
1
0
P
n
(x)f(x)[ =
[A
n
[
[bd
3
n+1
[
.

46 Lecture Notes at Fudan

8.3. The properties of Tchebychevs functions. Denote P be the set of all prime numbers.
8.6. Tchebychevs functions
1. Dene
(x) =

pP,px
1,
and
(x) =

pP,px
log p, x > 0.
Example. For any n N, there holds
(8.6.1)

pn,pP
[
n
p
] log p = nlog n +O(n).
Proof.At rst,
n! =

pn,pP
p

p
,
where

p
=

i=1
[
n
p
i
].
Then,we obtain
log n! =

pn,pP

p
log p =

pn,pP
[
n
p
] log p +

pn,pP

i=2
[
n
p
i
] log p.
But

pn,pP

i=2
[
n
p
i
] log p n

pn,pP

i=2
1
p
i
log p
n

pn,pP
log p
p
2

p
p 1
n

k=2
log k
k
2

k
k 1
2n

k=2
log k
k
2
= O(n).
On the other hand, the Stirling formula says that
log n! = nlog n +O(n).
Comparing the above two formula, we obtain 8.6.1.
Theorem. Let x 2, there is important relations between the functions (x) and (x):
(x) = (x) log x
_
x
2
(t)
t
dt (8.6.2)
(x) =
(x)
log x
+
_
x
2
(t)
t(log t)
2
dt (8.6.3)
The proof is dependent of the following lemma heavily :
Analysis 47
Lemma (Abel identity). Let a(n) : N C be a arithmetic function, its sum function is
A(x) =

nx
a(n), x 1.
Let f(t) is a dierential function on [y, x] where y [1, x]. Then, there holds
(8.6.4)

y<nx
a(n)f(n) = A(x)f(x) A(y)f(y)
_
x
y
A(t)f

(t)dt
Proof.We leave it as an exercise.

2. Dene Mangoldt function on N

(n) =
_
log p, n = p
m
, m N for some p P;
0, others.
Dene
(x) =

nN,nx
(n), x > 0.
Example. There holds (x) (x), moreover
(8.6.5) (x) = (x) +O(

x log x), (x ).
Proof.It is obvious that there holds
(8.6.6) (x) = (x) +

i=2
(x
1
i
).
We note
0 (x
1
i
) x
1
i
log(x
1
i
),
if i 2; and
(x
1
i
) = 0
if i > [
log x
log 2
] = M. Thus
(x) (x) (x) +
M

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

i=2
x
1
i
log(x
1
i
) (x) +Mx
1
2
log(x
1
2
).

3. Moreover, we have
(8.6.7) lim
x
(x)
x(log x)
1
= lim
x
(x)
x
= lim
x
(x)
x
.
Proof.i. At rst, from 8.6.6we obtain
(x) =

pP,px
[
log x
log p
] log p,
and so
(x) (x)

pP,px
log x
log p
log p = (x) log x
Thus, we obtain
lim
x
(x)
x
lim
x
(x)
x
= lim
x
(x)
x(log x)
1
.
48 Lecture Notes at Fudan
ii. On the other hand, for any with 0 < < 1, and x > 1, there holds
(x)

pP,x

<px
log x ((x) (x

)) log x

((x) x

) log x.
Thus
lim
x
(x)
x
lim
x
(x)
x(log x)
1
.
Since it holds for any 0 < < 1, there is
lim
x
(x)
x
lim
x
(x)
x(log x)
1
.

Similarly, we obtain
(8.6.8) lim
x
(x)
x(log x)
1
= lim
x
(x)
x
= lim
x
(x)
x
.
Corollary. There holds
(8.6.9) (x)
(x)
log x

(x)
log x
, (x ).
4. Dene Moblius function : N = 1, 0, 1 by
(n) =
_
_
_
1, n = 1;
(1)
r
, n = q
1
q
2
q
r
, q
1
< q
2
< < q
r
;
0, others.
Thus, if (n
1
, n
2
) = 1, then (n
1
n
2
) = (n
1
)(n
2
), and for any n N, there holds
(8.6.10)

k|n
(k) = [
1
n
].
Example. For any x 1, there holds
(8.6.11) [

dx
(d)
d
[ 1.
Proof.
1 =

nx
[
1
n
] =

nx

d|n
(d) =

dlx;d1,l1
(d) =

dx
(d)

1lx/d
1 =

dx
(d)[
x
d
],
thus we obtain
x

dx
(d)
d
= 1

dx
(d)
x
d
.
Since

2dx
(d)
x
d
<

2dx
1 [x] 1 = x x 1,
we have
[1

dx
(d)
x
d
[ = [1 x

2dx
(d)
x
d
[ x.

Analysis 49
Lemma (Mobius Transform). Let f(n), g(n) be arithmetic functions, i.e., from N to C. Then the
follow two formulas are equivalent:
g(n) =

d|n
f(d); (8.6.12)
f(n) =

d|n
(d)g(
n
d
). (8.6.13)
Corollary. For any n N, there holds

d|n
(d) = log n.
Proof.Let n = q

1
1
q

2
2
q

r
r
. Then, by induction on the number r, there is

d|n
(d) =

1
=0

2
=0

r
=0
(q

1
1
q

2
2
q

r
r
)
=

1
=1

2
=0

r
=0
(q

1
1
q

2
2
q

r
r
) +

2
=0

r
=0
(q

1
1
q

2
2
q

r
r
)
=

1
=1
(q

1
1
) +

d|q

2
2
q

r
r
(d)
=
1
log q
1
+ log q

2
2
q

r
r
= log n.

Example. [Tchebychevs identity]

(8.6.14)

nx
(
x
n
) = log[x]!.
Proof.
log([x]!) =

pP,p[x]
(

i=1
[
[x]
p
i
]) log p =

pP,p[x]
(

i=1
[
x
p
i
]) log p
=

pP,p[x]

i=1
(p
i
)[
x
p
i
] =

pP,p
i
[x]
(p
i
)[
x
p
i
]
=

dx
(d)[
x
d
] =

dx
(
x
d
).

By the exercise 8.6.16, we obtain

Corollary.
(8.6.15)

nx
(
x
n
) = x log x x +O(log x), x 1.
50 Lecture Notes at Fudan
Example (Tchebychevs inequality). For x 2, there holds
(8.6.16) (
log 2
4
)x (x) (4 log 2)x.
Proof.Let n N. we obtain
log(2n)! 2 log n! =

k2n
(
2n
k
) 2

kn
(
n
k
)
=

k2n
(
2n
k
) 2

kn
(
2n
2k
)
=

k2n
(1)
k1
(
2n
k
),
then
(8.6.17) (2n) (n) log
(2n)!
(n!)
2
(2n).
It is obvious that
2
n

(2n)!
(n!)
2
4
n
, n 2,
thus
(2n) nlog 2, (2n) (n) 2nlog 2.
Then, for x 2, there is m N with 2
m1
x 2
m
we obtain
(x) (2[
x
2
]) [
x
2
] log 2 > (
log 2
4
)x,
and
(x) (2
m
) =
m1

i=0
(2
i+1
) (2
i
) (
m1

i=0
2
i+1
) log 2 < 2
m+1
log 2 < (4 log 2)x.

Exercise 8.7. 1. To show that for any n N, there holds

log n

d|n
(d) = 0,
and so
(n) =

d|n
(d) log
n
d
=

d|n
(d) log d,
i.e., The Mobius transform of (n) log n is (n). Therefore, there is
(8.7.1) (n) log n =

d|n
(
n
d
), n N.
2. Generalized Mobius Transform Let F(x), G(x) be two functions dened over x 1. Then
the follow two formulas are equivalent:
G(x) =

nx
F(
x
n
); (8.7.2)
F(x) =

nx
(n)G(
x
n
). (8.7.3)
Analysis 51
3. To show the Tchebychevs identity 4

nx
(
x
n
) = log[x]!.
by the generalized Mobius transform 8.7.2.
4. To show that for any x 1, there holds
(8.7.4) log[x]! = x log x x +O(log(x)).
Proof.Using
_
x
1
log tdt = x log x x + 1,
we obtain
log[x]! =

n[x]1
log n + log x

n[x]1
_
n+1
n
log tdt + log x

_
x
1
log tdt + log x
= x log x x + 1 + log x,
and
log[x]! =

2n[x]
log n

2n[x]
_
n
n1
log tdt
=
_
x
1
log tdt
_
x
[x]
log tdt

_
x
1
log tdt log x
= x log x x + 1 log x,

5. To show that
(x) = O(x), x 1.
Proof.At rst, by 8.6.1 we have

pn,pP
[
n
p
] log p = nlog n +O(n),
thus

pP,p2n
[
2n
p
] log p 2[
n
p
] log p = O(n).
52 Lecture Notes at Fudan
Since
[
2n
p
] 2[
n
p
] =
_
1, n < p 2n
0, p n,
there is
(2n) (n) =

n<p2n
log p = O(n),
and so
(2x) (x) = ([2x]) ([x])
= ([2x]) (2[x]) +(2[x]) ([x])
= O(log[2x]) +O([x])
= O([x]).
Dene k(x) = [log x/log2] + 1, then for x 1, we obtain
(x) =

i=1
k(x)
((
x
2
i1
) (
x
2
i
)) = O(1)

i=1
x
2
i
= O(x).

Analysis 53
8.4. Mertens Theorem and Selbergs Inequality.
8.8. Mertens Theorems
1. By 8.7.2, the generalized mobius transform of (x) is
(8.8.1) T(x) =

nx
(
x
n
) =

m,nN,mnx
(m) =

mx
(m)[
x
m
].
2. For x 1, there holds
(8.8.2)

nx
(n)
n
= log x +O(1).
Proof.By 8.8.1, we obtain

mx
(m)[
x
m
] = T(x) = x log x +O(x),
and so the Tchebychev inequality say
0

nx
(n)(
x
n
[
x
n
])

nx
(n) = (x) < 4x.

3. Also, we have
(8.8.3)

px,pP
log p
n
= log x +O(1), x 1.
Proof.At rst, we have

nx
(n)
n

px,pP
log p
p
=

m2

pP,px
1/m
log p
p
m
<

pP
log p(

i=2
1
p
i
)
=

pP
log p
p(p 1)
<

n=2
log n
n(n 1)
= O(1).
Then,by 8.8.3we obtain the 8.8.3.

4.
Lemma. Let
n
be an increasing sequence with
n
(n ) and a
n
be any sequence.
Let b(x) be a function on R such that it is continuous on any nite interval [
1
, ]. Then, there
holds
(8.8.4)

n
x
a
n
b(
n
) = S(x)b(x)
_
x

1
S(x)b

(x)dt,
54 Lecture Notes at Fudan
where
S(x) =

n
x
a
n
.
Proof.
S(x)b(x)

n
x
a
n
b(
n
)
=

n
x
a
n
(b(x) b(
n
))
=

n
x
_
x

n
a
n
b

(t)dt
=
_
x

1
(

n
t
a
n
)b

(t)dt
=
_
x

1
S(x)b

(x)dt.

Theorem. For x 2, there is

(8.8.5)

pP,px
1
p
= log log x +A
1
+O(
1
log x
), x 2
where A
1
is a constant.
Proof.Dene a sequence
n
by

1
= 2,
n
= p
n
P(n 2);
and dene sequence a
n
by
a
n
=
log p
n
log p
n
, p
n
P.
From 8.8.3, we obtain
S(x) =

n
x
a
n
= log x +O(1).
Let b(x) =
1
log x
and r(x) = S(x) logx From8.8.4, we have

pP,px
1
p
=
S(x)
log x
+
_
x
2
S(t)
t(log t)
2
dt
=
log x +O(1)
log x
+
_
x
2
log t +O(1)
t(log t)
2
dt
= log log x +O(
1
log x
) +A +O(1)
_

x
dt
t(log t)
2
= log log x +O(
1
log x
) +A,
where
A = 1 log log 2 +
_

2
r(t)dt
t(log t)
2
is a constant.
Analysis 55
8.9. Selbergs Asymptotic Formula
1.
Lemma 8.10. Let F(x), G(x) be two functions dened over x 1 with F(1) = G(1). The the
following two relations are equivalent :
(8.10.1) G(x) =

1nx
F(
x
n
)
and
(8.10.2)

1nx
(n)G(
x
n
) log
x
n
= F(x) log x +

1nx
F(
x
n
)(n)
Proof.If the relation
G(x) =

1nx
F(
x
n
)
holds then we obtain

1nx
(n)G(
x
n
) log
x
n
=

nx
(n) log
x
n

mx/n
F(
x
mn
)
=

kx
F(
x
k
)

n|k
(n) log
x
n
= log x

nx
F(
x
n
)

k|n
(k)

nx
F(
x
n
)

k|n
(k) log k
= F(x) log x +

1nx
F(
x
n
)(n).
Assume the relation

1nx
(n)G(
x
n
) log
x
n
= F(x) log x +

1nx
F(
x
n
)(n)
holds.Let
k(n) =

1nx
(n)G(
x
n
) log
x
n
,
then by the generalized Mobius transform 8.7.2 we obtain
G(x) log x =

1nx
k(
x
n
),
thus,
G(x) log x =

1nx
F(
x
n
) log
x
n
+

1m
x
n
F(
x
nm
)(m)
= log x

1nx
F(
x
n
)

1nx
F(
x
n
) log n +

1kx
F(
x
k
)

m|k
(m)
= log x

1nx
F(
x
n
)

1nx
F(
x
n
) log n +

1kx
F(
x
k
) log k
= log x

1nx
F(
x
n
)

56 Lecture Notes at Fudan

2. Denote
1
(x) = x 1, where is the Euler constant. Let
T(x) =

nx
(
x
n
), T
1
(x) =

nx

1
(
x
n
).
Since

nN,nx
1
n
= log x + +O(
1
x
), x 1,
Then
T
1
(x) =

nx
(
x
n
1) = x log x x +O(1).
Moreover, since there is
T(x) = [n]! = x log x x +O(log(x)) x 1,
we obtain
(8.10.3) T(x) T
1
(x) =

nx
(
x
n
)
1
(
x
n
) = O(log x),
and for any l Z
0
, we obtain
(8.10.4)

nx
(n)T(
x
n
) T
1
(
x
n
)(log
x
n
)
l
<<

nx
(log
x
n
)
l+1
<<

nx
(
x
n
)
1
2
<< x.
Actually we again obtain the Tchebychev inequality while l = 0.
Theorem 1 (Selbergs Inequality).
(8.10.5) ((x) x) log x +

nx
(n)((
x
n
)
x
n
) = O(x).
Proof.Denote
F(x) = (x)
1
(x) = (x) (x 1), and G(x) = T(x) T
1
(x).
From 8.10.2 and 8.10.4, we obtain
((x) (x 1)) log x +

1nx
(n)(
x
n
1))
=

1nx
(n)(T(
x
n
) T
(
x
n
)) log
x
n
<<

nx
(log
x
n
)
2
<< x.

3. Selbergs inequality says that

(x) log x +

nx
(n)(
x
n
) = log x +

nx
(n)
x
n
+O(x),
then withith Merterns theorem,we immediately obtain
Analysis 57
Corollary (Selbergs Asymptotic Formula).
(8.10.6) (x) log x +

nx
(n)(
x
n
) = 2x log x +O(x).
(8.10.7)

px,pP
log
2
p +

p,qP,pqx
log p log q = 2x log x +O(x).
Proof.Only 8.10.7 is left to prove. Since
(x) = (x) +O(

x log x), (x ),
the formula 8.10.6 can be formulated into
(8.10.8) (x) log x +

px,pP
(
x
n
) log p = 2x log x +O(x).
On the other hand,
(x)

px,pP
log
2
p =

px,pP
log p log
x
p
=

px,pP
log p(

n
x
p
1
n
+O(1))
=

nx
1
n

n
x
p
log p +O((x))
= O(x

nx
1
n
2
) +O(x)
= O(x).

58 Lecture Notes at Fudan

8.5. An Elementary Proof of The Prime Number Theorem by Selberg and Erdos.
Analysis 59
8.6. Gausss Proof of The Fundamental Theorem of Algebra.
Theorem. Let
f(x) = x
n
+a
n1
x
n1
+a
n2
x
n2
+ +a
1
x +a
0
be a polynomial with a
i
Ci. Then, there exists a root C of f(x), i.e., f() = 0.
Proof. Let x = re

1
= r(cos + sin ). Then f(re

1
) = P(r, ) +

1Q(r, ) with
P(r, ) = r
n
cos(n) + , Q(r, ) = r
n
sin(n) + .
We know that
F(r, ) = (P
2
+Q
2
)(r, ) = r
2n
+
is a function of (r, ). If there is a pair (r
0
,
0
) satisfying F(r
0
,
0
) = 0 then r
0
e

1
0
is a root of the
polynomial f(x). Suppose that F(r, ) has no zero point, we show there must be a coniction: We
can dene a dierentiable function of (r, ) by U = arctan
P
Q
since F(r, ) has no zero point. Thus,
there is
U
r
=
1
P
2
+Q
2
(Q
P
r
P
Q
r
),
U

=
1
P
2
+Q
2
(Q
P

P
Q

),
and so

2
U
r
=
H
P
2
+Q
2
,
where H is a continuous function of (r, ).
Denote
I
1
=
_
R
0
(
_
2
0

2
U
r
d)dr, and I
2
=
_
2
0
(
_
R
0

2
U
r
dr)d.
Since

2
U
r
is a continuous function of (r, ), there holds I
1
= I
2
.
On the other hand, we have the following calculations
i. I
1
= 0 is given by
_
2
0

2
U
r
d =
U
r
[
2
0
= 0.
ii.
P

= nr
n
sin(n) + , and
Q

= nr
n
cos(n) + ,
and so,
Q
P

P
Q

= nr
2n
+ ,
U

=
1
P
2
+Q
2
(Q
P

P
Q

) =
nr
2n
+
r
2n
+
.
Regarding Q
P

P
Q

and P
2
+Q
2
as polynomials of r, then all coecients in both polynomials
are bounded function of . Therefore
lim
r
U

= n (uniformly for all ),

_
R
0

2
U
r
dr =
U

[
R
0
n (uniformly for all ),
and so
lim
R
I
2
= 2n.
It is a coniction to that I
1
= I
2
. Thus P
2
+Q
2
has at least one zero point.

60 Lecture Notes at Fudan

9. Advanced Techniques in Analysis
9.1. Preliminary Results in Analysis.
9.1. Some useful techniques.
1. Unit decomposition. Let K R
n
is a compact set, U
1
, , U
N
is open covering of K. Then
there exist functions
1
C

0
(U
1
), ,
N
C

0
(U
N
), such that
i.
0
i
(x) 1, x U
i
(i = 1, , N);
ii.
N

i
(x) 1, x K.
2. Local C
k
-attening Let R
n
be a bounded domain. We call the boundary is C
k
, if for
any x
0
there is a neighborhood U R
n
of x
0
and a bijective function : U B
1
(0) with
(x
0
) = 0 such that
i. ,
1
are C
k
.
ii.
(U ) = B
+
1
(0) = y B
1
(0) : y
n
> 0,
iii.
(U ) = B
+
1
(0) = y B
1
(0) : y
n
= 0.
3. Lipschitz attening An open subset R
n
is said to be Lipschitz if for every the boundary
is C
k
, if if for any x
0
there exist a neighborhood U R
n
of x
0
and a bijective function

x
0 : U B
1
(0) with (x
0
) = 0 such that
i. ,
1
are Lipschitz.
ii.
(U ) = B
+
1
(0) = y B
1
(0) : y
n
> 0,
iii.
(U ) = B
+
1
(0) = y B
1
(0) : y
n
= 0.
Corollary. Let R
n
be a bounded Lipschitz domain.There exists a continuous linear extension
operator
E : W
1,2
() W
1,2
(R
n
)
such that
[[E(u)[[
W
1,2
(R
n
)
C()[[u[[
W
1,2
()
, u W
1,2
().
Moreover, it is possible to arrange such an extension E such that
x R
n
: E(u) ,= 0 x R
n
: dist(x, ) 1.
9.2. Friedrichs smoothing molliers.
1. Consider a function j C

0
(R
n
) having support in B
1
(0) = x R
n
: [x[ < 1 with the properties
j(x) 0,

_
B
1
(0)
j(x)dx = 1.
Analysis 61
Example. Denote
j(x) =
_
1
A
exp(
1
|x|
2
1
), [x[ < 1;
0, [x[ 1;
where
A =
_
B
1
(0)
exp(
1
[x[
2
1
)dx.
2. For any > 0, we dene the molliers
j

(x) =
1

n
j(
x

),
then j

has compact support in B

(0) = x R
n
: [x[ < with the properties
j

(x) 0,

_
B

(0)
j

(x)dx = 1.
Remark. The following are useful remarks:
a) For any > 0, there is a with 0 < < 1 such that
0 <
_
R
n
\B
R
(0)
j(x) , R < 1,
and
0 <
_
R
n
\B
R
(0)
j

(x) , R < 1.
b) If j(x) has additional property j(x) = j([x[), then we have
n
n

n
_

0
j(
r

)r
n1
dr = 1,
where
n
is the volume of unit ball in R
n
, actually

n
=
2
n
2
n(
n
2
)
.
3. We can use the molliers j

to smooth a function u L
1
loc
()( R
n
). Let

:= x : dist(x, ) >
and for any x

, denote
J

u(x) = (u j

)(x) :=
_

u(y)j

(x y)dy,
J

u is evidently smooth on

. In fact, by virtue of the usual dierential under the integral lemma,

for any x

,
D

x
J

u(x) =
_

u(y)D

x
j

(x y)dy.
We then call that J

u is the smoothing (mollifying)of u, and J

the smoothing operator, j

the smoothing (mollifying) kernel.

Remark. Mollifying kernels were indroduced into PDE theory by K.O.Friedrichs.Therefore, they
are often called Friedrichs molliers.
Proposition 9.3. Let u be a function on R
n
with
Suppu R
n
.
We will put u(y) = 0 for y R
n
.(We shall always use that convention in the sequel.) Then,
we have the following properties:
62 Lecture Notes at Fudan
a) If u L
1
loc
(), then J

u C

).
b) If dist(Suppu, ) > , then J

u C

0
().
c)
J

u
a.e.
u on ,
and moreover for all compact sets K ,
i. J

u u on K(on for u C
0
()) uniformly if u is continuous.
ii.
J

u
L
1
(K)
u.
Proof.Now let u C
0
(). At rst, Let
=
dist(K, )
2
.
Then, for any given > 0, there exists with > > 0 such that
B

(x) K

, x K, where K

is a compact set in denoted by

K

= x [ dist(x, K) ;
if 0 < < then
sup
zB
1
(0)
[u(x +z) u(x)[ < , x K,
since u is uniformly continuous on the compact set K

.
If < then
J

u(x) = (u j

)(x) =
_

u(y)
1

n
j(
x y

)dy =
_
B
1
(0)
j(z)u(x +z)dz.
Therefore, for all x K, if < then
[J

u(x) u(x)[
_
B
1
(0)
j(z)[u(x +z) u(x)[dz sup
zB
1
(0)
[u(x +z) u(x)[
_
B
1
(0)
j(z)dz .

d) If u L
p
()(1 p < ), then J

u L
p
(), and
lim
0
+
[[J

u u[[
L
p
()
= 0.
Moreover, if u L
P
loc
() then for all compact sets K , we have
J

u
L
p
(K)
u.
Proof.Denote

= x R
n
[ dist(x, ) .
At rst,
J

u(x) = (u j

)(x) =
_

u(y)
1

n
j(
x y

)dy =
_
B
1
(0)
j(z)u(x +z)dz,
then for q > 0 with 1/q + 1/q = 1, we obtain
[J

u(x)[
p
(
_
B
1
(0)
j(z)dz)
p
q
(
_
B
1
(0)
j(z)[u(x +z)[
p
dz).
Analysis 63
If < , it follows that
(9.3.1)
_

[J

u(x)[
p
dx
_

_
B
1
(0)
j(z)[u(x +z)[
p
dzdx =
_
B
1
(0)
j(z)(
_

[u(x +z)[
p
dx)dz
_

[u(x)[
p
dx.
For a given > 0, we choose w C
0
(

) with
[[u w[[
L
p
(

)
<

3
.
On the other hand, for suciently small ,
[[w J

w[[
L
p
(

)
<

3
.
Applying 9.3.1 to u w, we obtain
[[J

u J

w[[
L
p
()
[[u w[[
L
p
(

)
<

3
,
and hence
[[J

u u[[
L
p
()
[[J

u J

w[[
L
p
()
+[[J

w w[[
L
p
()
+[[u w[[
L
p
()
.

Remark 9.4. In the proof, we did not use the smoothness of the kernel j at all. Thus, some
results holds for other kernels, and in particular for
(x) =
_
1

n
, [x[ 1;
0, [x[ 1;
The corresponding convolution is
u

=
1

n
_

(
x y

)u(y)dy =
1
[B

(x)[
_
B

(x)
u(y)dy.
We obtain the the following result:
Corollary. If u L
p
()(1 p < ), then u

L
p
(), and
lim
0
+
[[
1
[B

(x)[
_
B

(x)
u(y)dy u[[
L
p
()
= 0.
Moreover,
1
[B

(x)[
_
B

(x)
u(y)dy
a.e.
u on ,
and for all compact sets K ,
i.
1
|B

(x)|
_
B

(x)
u(y)dy u on K(on for u C
0
()) uniformly if u is continuous.
ii.
1
[B

(x)[
_
B

(x)
u(y)dy
L
1
(K)
u.
9.5. Cut-o functions.
Let be a bounded domain in R
n
with smooth boundary, and

(i.e,

). Let d :=
1
4
dist(

, , ) then d > 0. Let

:= x : dist(x,

) < d,
then dist(

, ) = 3d.
64 Lecture Notes at Fudan
1. Let

be characteristic function of

, we consider the smoothing function of

(x) := J
d
(

).
Then, we have
C

0
(); and

0 1 with (x) 1 on

[[
C()
d
,
where C() is a constant only dependent on . We then call the cut-o function.
2. In application, we always consider the case
= B
R
(x
0
) := x R
n
: [x x
0
[ < R.
Let 0 < < R and (x) be the cut-o function related to B

(x
0
). We have moreover property of
(x) :
[D
k
(x)[
C
[R [
k
,
[D
k
]

C
[R [
k+
,
where the constant C() absolutely does not depend on R and .
9.6. Applications of smoothing kernels and cut-o functions.
1. Mean-value formula. Let u be a continuous function on a domain R
n
such that
(9.6.1) u(x) = [B(x, r)[
1
_
B(x,r)
u(y)dS(y)
for any ball B(x, r) . Then, u is a harmonic function.
Proof.Let j(x) be the smoothing kernel having the additional property j(x) = j([x[), then we
have
n
n

n
_

0
j(
r

)r
n1
dr = 1,
Let j

(x) :=
1

n
j(
x

), > 0. For any sucient small > 0, denote

:= x : dist(x, ) >
and then for any x

, denote
J

u(x) = (u j

)(x) :=
_

u(y)j

(x y)dy.
Analysis 65
Then J

u(x) C

), and on

, we have
J

u(x) =
1

n
_
R
n
u(y)j

(x y)dy
=
1

n
_
R
n
j(
x y

)u(y)dy
=
1

n
_

0
j(
r

)
_
B(x,r)
u(y)dS(y)dr
=
1

n
_

0
j(
r

)[B(x, r)[u(x)dr
= u(x)
n
n

n
_

0
j(
r

)r
n1
dr = u(x).
Hence u C

(). On the other hand,for any x , denote

(r) := [B(0, 1)[
1
_
B(0,1)
u(x +rz)dS(z).
Then
(r) = [B(x, r)[
1
_
B(x,r)
u(y)dS(y) u(x),
and so
0
(r)
r
= [B(0, 1)[
1
_
B(0,1)
Du(x +rz) zdS(z)
= [B(x, r)[
1
_
B(x,r)
Du(y)
y x
r
dS(y)
= [B(x, r)[
1
_
B(x,r)
u(y)
n
dS(y)
=
r
n
[B(x, r)[
1
_
B(x,r)
u(y)dy.
Since u is smooth, let r 0, we then obtain
u(x) = 0, x .

Remark. A continuous function on is harmonic if and only if u satises 9.6.1.

2. Let u be a harmonic function on a domain R
n
and

, d = dist(

, ). Then u C

()
and there holds for every multi-index
sup

[D

u[ C(, d)[[u[[
L
1
()
.
Proof.Let j
d
(x) be the mollier smooth function related to having additional property j
d
(x) =
j
d
([x[). Then, for every xed y

, we have
u(y) =
_
R
n
j
d
(x y)u(x)dx,
66 Lecture Notes at Fudan
and hence
[D

y
u(y)[
_
B
d
(y)
[D

y
j
d
(x y)[[u(x)[dx sup

[D

j
d
(x)[
_

[u(x)[dx.
Since
[D

j
d
(x)[ = C
n,
d

,
we complete the proof.
3. Weyl Lemma.
Denition. Let u W
1,2
(), where is an open set in R
n
, we say that u is weakly harmonic
on if
_

Du D = 0, C

0
().
Theorem 9.7 (Weyl Lemma). Suppose that u W
1,2
() is weakly harmonic. Then the L
2
class
of u has a C

representative which is harmonic.

Proof.For any > 0, let j

(x) be the mollier smooth function. then J

u is the smoothing of u
dened on

:= x : dist(x, ) > ,
and

.
i. Since

x
j

(x y) = (1)
2

y
j

(x y) =
y
j

(x y),
for any x

, we have

x
J

u(x) =
_

u(y)
x
j

(x y)dy
=
_

u(y)
y
j

(x y)dy
=
n

i=1
_

D
y
i
u(y)D
y
i
[j

(x y)]dy
= 0,
i.e., J

u is C

and harmonic on

.
ii. For any K , we can nd a
0
such that if 0 < <
0
, then K

. Let d = dist(K, ).
We have for any x K, any 0 <
1
,
2
<
0
[D

x
(J

1
u(x) J

2
u(x))[ C
n,
d

[J

1
u(y) J

2
u(y)[dy = C
n,
d

[[J

1
u J

2
u[[
L
1
()
.
Denote
u = lim
0
+
J

u,
then by Arzela-Ascoli lemma u is in C

() and is harmonic.

4. Application.
Lemma 9.8 (Rellich Compactness Lemma). Suppose that is a bounded Lipschitz domain in
R
n
and u
k
is a sequence of W
1,2
() with sup
k
[[u
k
[[
W
1,2
()
< . Then there is a subsequence u
k
i
and u W
1,2
() such that
i.
u
k
i
weakly
u in W
1,2
();
Analysis 67
ii.
u
k
i
strongly
u in L
2
();
( The same in fact is true for L
p
() with p < 2

:=
2n
n2
)
iii.
_

[Du[
2
lim
i
inf
_

[Du
k
i
[
2
Theorem 9.9 (Poincare inequality). Suppose that is a bounded and connected Lipschitz domain
in R
n
. Then there exists a constant C

dependent only on the domain such that for every

u W
1,2
() there holds
(9.9.1)
_

[u [
2
C

[Du[
2
,
where = [[
1
_

u.
Proof.a) Suppose the assertion is false: Then for each k N, there exist function u
k
W
1,2
()
such that 9.9.1 fails for C

= k:
_

[Du
k
[
2

1
k
_

[u
k

k
[
2

1
k
inf
R
_

[u
k
[
2
,
where
k
= [[
1
_

u
k
. Dene
v
k
:=
u
k

k
[[u
k

k
[[
L
2
()
,
we nd
[[v
k
[[
L
2
()
= 1 and [[Dv
k
[[
1

k
,
and hence v
k
is a bounded sequence in W
1,2
().
b) By Rellich Compactness Lemma, there exists v W
1,2
() and a subsequence v
k
i
of v
k
such
that

v
k
i
strongly
v in L
2
(),
and so
_

v = 0;

v
k
i
weakly
v in W
1,2
(),
and so
Dv
k
i
weakly
Dv in L
2
();

[Dv[
2
lim
i
inf
_

[Dv
k
i
[
2
and so
_

[Dv[
2
lim
i
inf
1

k
= 0.
68 Lecture Notes at Fudan
c) Hence Dv = 0 a.e on . Since is connected and
D(J

v) = J

(Dv) = 0,
J

v is a constant on any connected component

. Hence v = 0 a.e. on by
_

v = 0, but
it contradicts to that [[v[[
L
2
()
= 1.

Remark. By using the special case for = B

1
(0) and by changing scale x Rx, the Poincare
inequality is : For any u W
1,2
(B
R
(x
0
)),
(9.9.2) R
n
_
B
R
(x
0
)
[u [
2
CR
2n
_
B
R
(x
0
)
[Du[
2
,
where = [B
R
(x
0
)[
1
_
B
R
(x
0
)
u and C is a constant depending only on n.
9.10. More properties on harmonic functions. Let u be a harmonic function on an open set
R
n
.
1. Harnark inequality.
Proposition. Let B
R
(x) be a ball.For any ball B
r
(x) B
R
(x), there holds
(9.10.1) sup
yB
r
(x)
u(y) (
R +r
R r
)
n
inf
yB
r
(x)
u(y).
Moreover, for any connected compact subset V subset, we have
(9.10.2) sup
yV
u(y) C inf
yV
u(y),
where the constant C depends only on n and dist(V, ).
2.
Proposition. For any ball B
r
(x) and any multi-index with k = [[, there holds
(9.10.3) [D

u(x)[
C
k
r
n+k
_
B
r
(x)
[u(y)[dy,
where
C
0
=
1

n
, C
k
=
(n +k)
n+k
(n + 1)
k

n
(n + 1)
n+1
(k = 1, 2, ).
Proof.At rst, we have
[u
x
i
(x)[
n + 1

n
r
n+1
_
B
r
(x)
[u(y)[dy.
Since D

u is still a harmonic function, we have

[D

u(x)[
n + 1

n
(tr)
n+1
_
B
tr
(x)
[D

u(y)[dy,
where is a multi-index with [[ = k 1 such that there is i 1, 2, , n with D

u = (D

u)
x
i
.
Since y B
tr
(x) implies that B
(1t)r
(y) B
r
(x) , by induction on k there holds
[D

u(y)[
C
k1
((1 t)r)
n+k1
_
B
(1t)r
(y)
[u(z)[dz
C
k1
((1 t)r)
n+k1
_
B
r
(x)
[u(z)[dz,
Hence
[D

u(x)[
n + 1

n
(tr)
n+1

n
(tr)
n
max
yB
tr
(x)
[D

u(y)[
(n + 1)C
k1
r
n+k
t(1 k)
n+k1
_
B
r
(x)
[u(z)[dz.
Analysis 69
Denote
C
k
:= max
t[0,1]
(n + 1)C
k1
t(1 k)
n+k1
,
then
C
k
=
(n + 1)C
k1
(n +k)
n+k
(n +k 1)
n+k1
=
(n +k)
n+k
(n + 1)
k

n
(n + 1)
n+1
.

3. As a corollary, by Harnark inequality,we have:

Corollary. Let u be a harmonic function on the total R
n
. Suppose u has upper bound (or lower
bound), then u is a constant.
4.
Proposition. A u harmonic function on is an analytic function.
Proof.Fix any point x
0
, we only need to check the convergent radios of the following the
Taylor series of u at x
0
. First the Taylor-Maclaurin formula says
u(x) =
N1

k=0

||=k
D

u(x
0
)
!
(x x
0
)

+R
N
(x),
where
R
N
(x) =

k=N

||=k
D

u(x
0
)
!
(x x
0
)

||=N
D

u(x
0
+(x x
0
))
!
(x x
0
)

, (0, 1).
Let r =
1
4
dist(x
0
, ), and denote
A =
1

n
(n + 1)
n+1
r
n
_
B
2r
(x
0
)
[u(y)[dy.
Then, for any x B
2r
(x
0
) (so that B
r
(x) B
2r
(x
0
) ), the formula 9.10.3 says that for
N >> o,
[R
N
(x)[

||=N
[D

u(x
0
+(x x
0
))[
!
[x x
0
[

||=N
1
!
(n +N)
n+N
[
(n + 1)[x x
0
[
r
]
N
= 2
n
e
n
A(

||=N
)(
N!
!
)e
N
N
n1/2
[
(n + 1)[x x
0
[
r
]
N
2
n
e
n
A[
en(n + 1)[x x
0
[
r
]
N
N
n1/2
.
Here we use the combination formula

||=N
1
!
= n
N
,
and the Stirling formula (refer to the formula 4.2.2):
lim
N
N
N+1/2
N!e
N
=
1

2
< 1,
70 Lecture Notes at Fudan
and so if N >> 0
(n +N)
n+N
= (1 +n/N)
N
(N +n)
n
N
N
e
n
(2N)
n
N
N
2
n
e
n
N!e
N
N
n1/2
.
Hence if
x B
r
2n(n+1)e
(x
0
)
then
[R
N
(x)[ 2
n
e
n
A
N
n1/2
2
N
,
and so
[R
N
(x)[
uniformly
0 (N ).

5. The following harmonic approximation (or blow up) lemma will be of fundamental importance.
Proposition (Harmonic Approximation Lemma). Let f W
1,2
(B
1
(0)). If for each > 0 there
is = (n, ) > 0 such that f satises
i.
_
B
1
(0)
[Df[
2
< 1,
and
ii.
[
_
B
1
(0)
Df D[ < sup
B
1
(0)
[D[, C

0
(B
1
(0)).
Then there is a harmonic function u on B
1
(0) such that
_
B
1
(0)
[Du[
2
< 1,and
_
B
1
(0)
[u f[
2

2
.
Proof.If this fails for some , then there is a sequence f
k
W
1,2
(B
1
(0)) with

_
B
1
(0)
[Df
k
[
2
< 1,
[
_
B
1
(0)
Df
k
D[ <
1
k
sup
B
1
(0)
[D[, C

0
(B
1
(0)),
and such that
(9.10.4)
_
B
1
(0)
[u f
k
[
2
>
2
,
for every harmonic u on B
1
(0) with
_
B
1
(0)
[Du[
2
< 1.
Notice that since the same holds with

f
k
= f
k

k
for any choice of constants
k
, we
can assume without loss of generality that
_
B
1
(0)
f
k
= 0 for each k. But then by the Poincare
inequality(c.f.9.9.1) we conclude
sup lim
k
_
B
1
(0)
([f
k
[
2
+[Df
k
[
2
) < ,
and hence by the Rellich Compactness Theorem we have a subsequence f
k
and an f W
1,2
(B
1
(0))
such that
f
k

strongly
f in L
2
(B
1
(0)); and Df
k

weakly
Df in L
2
(B
1
(0)).
But we then conclude
_
B
1
(0)
Df D = 0 C

0
(B
1
(0)),
Analysis 71
so that f is a weakly harmonic function on B
1
(0), and the Weyls lemma guarantees that f is a
smooth harmonic function on B
1
(0). Moreover, since by Lebsegue lemma, we have
_
B
1
(0)
[Df[
2
inf lim
k

[Df
k
[
2
1,
hence that f
k

strongly
f in L
2
(B
1
(0)) contradicts to the relation 9.10.4.
72 Lecture Notes at Fudan
9.2. Scaling Technique and Schauder Estimate.
9.11. Notions. Let R
n
be an open domain and (0, 1]. Let u be a function dened on .
1. u is said to be Holder continuous with exponent on if there is a constant C such that
(9.11.1) [u(x) u(y)[ C[x y[

, x, y
2. Let u be a Holder continuous function with exponent on . The Holder coecient of u is
dened by
[u]
;
:= sup
x=y,x,y
[u(x) u(y)[
[x y[

.
we dene the set of all bounded Holder continuous functions with exponent on by
C
0,
() = u[ u is a function dened on with [u]
;
< .
We call u Lipschitz continuous if u is a Holder continuous function with exponent 1.
Remark. []
;
is only a semi-norm in the space C
0,
(). It is not a norm since it is zero for any
constant function.But by the Arzela-Ascoli theorem, the space C
0,
() becames a Banach space
in the norm
[u[
0,;
:= [u[
0;
+ [u]
;
.
where [u[
0;
= [[u[[
L

()
.
3. u is called locally Holder continuous with exponent on if it is Holder continuous on
each

, i.e., u C
0,
(

), where

means the closure of

in is a compact set of
. We denote the space of all locally Holder continuous functions with exponent on be C
0,
().
4. Scaling. Let u C
0,
(). For any R > 0, we dene the scaled function
u(x) = R

u(Rx) x

where the domain

R
n
is dened by

:= R
1
y [ y .
Then, u C
0,
(

), and
[u]
;
= [ u]
;

.
5. Oscillation. For any real function u : R. The oscillation is dened as
osc

u := sup
x
u(x) inf
x
u(x).
Corollary 9.12. Let u : B
R
(x
0
) R be a real value function. If for every y B
R/2
(x
0
), and every

R
2
, osc
B
R(y)
u < and there is a xed (0, 1) such that
(9.12.1) osc
B

(y)
u <
1
2
osc
B

(y)
u,
Analysis 73
then u C
0,
(B
R/2
(x
0
)) with =
log 2
log
, and moreover
[u]
;B
R/2
(x
0
)
C

osc
B
R
(x
0
)
u.
Proof. By induction we get from 9.12.1 that for every y B
R/2
(x
0
), and every
R
2
, the following
estimates
(9.12.2) osc
B

(y)
u <
1
2
k
osc
B

(y)
u, k = 1, 2.
are valid. Let x, y B
R/2
(x
0
).
a) If r := [x y[
R
2
, then we have
[u(x) u(y)[ (
2[x y[
R
)

osc
B
R
(x
0
)
u,
and so
[u(x) u(y)[
[x y[

osc
B
R
(x
0
)
u.
b) If r := [x y[
R
2
, we can choose an integer k with

k+1

2r
R
<
k
.
Using 9.12.2 with = R/2, we have
[u(x) u(y)[ osc
B

k R
2
u 2
k
osc
B
R
2
(y)
u 2
k
osc
B
R
(x
0
)
u,
and so
[u(x) u(y)[
k
osc
B
R
(x
0
)
u

(
2[x y[
R
)

osc
B
R
(x
0
)
u.
Since

=
1
2
, we have
[u(x) u(y)[
[x y[

2
+1
R

osc
B
R
(x
0
)
u.
Let C

= 2
+1
. The claim is then established.

9.13. Notions.
Let R
n
be an open domain and (0, 1].. Let = (
1
, ,
n
) be the multiple index with

1
, ,
n
non-negative integers.
Denote [[ :=
1
+ +
n
.
Let u be a function dened on .
1. If u C
k
(), denote
D

u =

||u

1
x
1

n
x
n
.
The space C
k
() is a Banach Space with the norm dened by
[u[
k;
=

||k
sup

[D

u[.
74 Lecture Notes at Fudan
2. Recall
C
0,
() = u[ u is a function dened on with [u]
;
< ,
and it is a Banach space with the norm
[u[
0,;
:= [u[
0;
+ [u]
;
.
We dene
C
k,
() = u[ u is a function dened on with D

u C
0,
() for any mult-index with [[ k.
The space C
k,
() will be a Banach space if it is denoted the following two norms:
[u[
k,;
=

||k
[D

u[
;
,
and
[u[
k,0;
= [u[
k;
=

||k
[D

u[
0;
.
Remark. The following two are not norms but semi-norms:
[u]
k,;
=

||=k
[D

u]
;
,
and
[u]
k,0;
= [u]
k;
=

||=k
[D

u[
0;
.
Lemma 9.14. Let u be a harmonic function over R
n
. If there are constants C, q > 0 such that
[u[
0;B
R
CR
q
R 1.
Then u is a polynomial.
Proof. At rst, it is well known that
[B
R
[
[B
R
[
=
n
R
.
Denote
n
be the volume of B
1
.
For any x R
n
, any harmonic function u, we have
u(x) =
1

n
_
B

u(x +y)dy, > 0.

and
u(x) =
1

n1
n
n
_
B

u(x +y)dS(y), > 0.

Since D
i
u is still a harmonic function on R
n
, we have
D
i
u(x) =
1
R
n

n
_
B
R
D
i
u(x +y)dy
=
1
R
n

n
_
B
R
u(x +y) cos(

n , y)dS(y)
=
1
R
n
_
|w|=1
u(x +Rw)w
i
dw,
where

n
i=1
(w
i
)
2
= 1. By induction, we have
Analysis 75
D
i
1
D
i
2
D
i
k
u(x)) =
1
R
n
_
|w
1
|=1
D
i
2
D
i
k
u(x +Rw
1
)w
i
1
1
dw
1
=
1
R
k

k
n
_
|w
1
|=1
dw
1

_
w
k
=1
u(x +Rw
1
+ Rw
k
)w
i
1
1
w
i
k
1
dw
k
,
and so for all x R
n
, k 1, we have
[D
i
1
D
i
2
D
i
k
u(x)[
C
R
k

k
n
([x[ +kR)
q
(n
n
)
k
.
Fix R
0
such that [x[ R
0
, Let R R
0
, if k > q we then have
[D
i
1
D
i
2
D
i
k
u(x)[
Cn
k
R
kq
(
R
0
R
+k)
q
0(R ).
Therefore, D
k
u(x) 0 x B
R
0
if k > q, and then u has to be a polynomial of degree less than
q + 1.

Corollary 9.15. Let u C

2
(R
n
+
) with u(x

, 0) = 0, and u is a harmonic function over R

n
+
. If there
are constants C, q > 0 such that
[u[
0;B
R
R
n
+
CR
q
R 1.
Then u is a polynomial.
Proof. Dene
u(x

, x
n
) =
_
u(x

, x
n
), x
n
0;
u(x

, x
n
), x
n
< 0.
Then, u C
0
(R
n
) and
D
2
ij
u C
0
(R
n
), i +j 2n 1.
On the other hand, we have
_
u = 0, x
n
> 0;
u = u(x

x
n
) = 0, x
n
< 0.
Thus, we have
u
x
n
x
n
(x

, +0) = u
x
n
x
n
(x

, +0) =
n1

i=1
u
x
i
x
i
(x

, 0),
u
x
n
x
n
(x

, 0) =
n1

i=1
u
x
i
x
i
(x

, 0) = u
x
n
x
n
(x

, +0);
and so u C
2
(R
n
) and it is harmonic on R
n
. Since
[ u[
0,B
R
= [u[
0;B
R
R
n
+
CR
q
R 1,
by the above lemma u is a polynomial, and then so is u.

Theorem 9.16. Let (0, 1). Then there is a constant C = C(n, ) such that for any u C
2,
(R
n
),
we have
(9.16.1) [u]
2,
C[u]

.
76 Lecture Notes at Fudan
Proof. 1. Assume the estimate 9.16.1 is not true for all u C
2,
(R
n
), then we have a series u
k
, k =
1, 2, , such that
1 = [u
k
]
2,
k[u
k
]

, k = 1, 2, ,
i.e., we have
(9.16.2) [u
k
]

<
1
k
, k = 1, 2, .
2. By the denition, we have
[u]
2,
=

||=2
[D

u]

||=2
sup
xR
n
,eS
n1
,h>0
[D

u
k
(x +he) D

u
k
(x)[
h

,
it implies that for any k N, We can choose x
k
R
n
, h
k
> 0 and

([[ = 2), e

S
n1
with

, e

independent of k, satisfying
[D

u
k
(x
k
+h
k
e

) D

u
k
(x
k
)[
h

1
2n
3
.
(actually, we can choose e

= e
i
k
such that i
k
,= , where [

[ =

.)
3. Using scaling technique: Let x = x
k
+h
k
y and denote v
k
(y) = h
2
k
u
k
(x
k
+h
k
y). Then, we have
D
2
y
v
k
(y) == h

k
D
2
x
u
k
(x
k
+h
k
y),
and so we have: for k N, there is
(9.16.3)
[v
k
]
2,
=

||=2
[D

v
k
]

||=2
sup
y,y

R
n
[D

u
k
(x
k
+h
k
y) D

u
k
(x
k
+h
k
y

)[
h

[y y

||=2
[D

u
k
]

= [u
k
]
2,
= 1;
(9.16.4) [v
k
]

= [u
k
]

1
k
, k N;
(9.16.5) [D

v
k
(e

) D

v
k
(0)[
1
2n
3
, k N.
4. Denote
v
k
(y) = v
k
(y) v
k
(0)

1in
D
y
i
v
k
(0)y
i

1
2

1i,jn
D
y
i
y
j
v
k
(0)y
i
y
j
.
Then, we have
(9.16.6) v
k
(0) = 0, D v
k
(0) = 0, D
2
v
k
(0) = 0, k N;
(9.16.7) [ v
k
]

1
k
, k N;
(9.16.8) [D
2
v
k
(y)[ [ v
k
]
2,
[y[

= [y[

, [D v
k
(y)[ [y[
+1
, [ v
k
(y)[ [y[
+2
, k N;
(9.16.9) [D

v
k
(e

)[
1
2n
3
, k N.
Then, we obtain that for R > 0, k N, v
k
are bounded uniformly and is equai-continuous on
B
R
, and so on B
R
with any R > 0, we have
v
k
=v

C
2
(B
R
), v

C
2,
(B
R
);
and
[v

]
;B
R
2, v
k
=v

Analysis 77
5. From 9.16.7, we have [v

0, and so v

is a constant.Moreove, by 9.16.8 and let y 0,

we have v

0, i.e., v

is a harmonic function.Using the formula 9.16.8 and the above lemma

9.14,v

is polynomial of degree at least 4. Thus,by 9.16.7,D

2
v

(0) = 0 and so
D

(te

) = a
i
t
i
+ +a
1
t, with i 2 and a
i
,= 0.
i. If i 1, we have a contradiction from
[D

(te

)[
t

[v

]
2,
2 (t > 0).
ii. If i = 0, then D

(te

) 0. But, by the formula 9.16.9 we have

[D

(e

)[
1
2n
3
.
It is a contradiction.

Theorem 9.17. Let (0, 1). Then there is a constant C = C(n, ) such that for any u C
2,
(R
n
+
)
with u[
R
n
+
= 0, we have
(9.17.1) [u]
2,;R
n
+
C[u]
;R
n
+
.
Proof. 1. Assume the estimate 9.17.1 is not true for all u C
2,
(R
n
+
), then we have a series u
k
, k =
1, 2, , such that
1 = [u
k
]
2,;R
n
+
k[u
k
]
;R
n
+
, k = 1, 2, ,
i.e., we have
(9.17.2) [u
k
]
;R
n
+
<
1
k
, k = 1, 2, .
By the denition, we have
[u]
2,;R
n
+
=

||=2
[D

u]
;R
n
+
=

||=2
sup
xR
n
+
,eS
n1
,h>0
[D

u
k
(x +he) D

u
k
(x)[
h

,
it implies that for any k N, Similarly as in the proof of the theorem 9.16,we can choose
x
k
R
n
+
, h
k
> 0 and

([[ = 2), e

S
n1
with

, e

independent of k, satisfying
[D

u
k
(x
k
+h
k
e

) D

u
k
(x
k
)[
h

1
2n
3
.
2. Using scaling technique: Let t
k
= x
k;n
h
1
k
where x
k
= (x
k;1
, , x
k;n
). We consider the following
two cases:
a) In case of t
k
+(h
k
+0). Denote x = x
k
+ yh
k
, then y = h
1
k
x x
k
h
1
k
and y
n
> t
k
with t
k
. Denote v
k
(y) = h
2
k
u
k
(x
k
+h
k
y). Then,
D
2
y
v
k
(y) == h

k
D
2
x
u
k
(x
k
+h
k
y),
also, we have: for k N,
(9.17.3)
[v
k
]
2,
=

||=2
[D

v
k
]

||=2
sup
y,y

R
n
[D

u
k
(x
k
+h
k
y) D

u
k
(x
k
+h
k
y

)[
h

[y y

||=2
[D

u
k
]

= [u
k
]
2,
= 1,
and so
(9.17.4) [v
k
]

= [u
k
]

1
k
, k N,
78 Lecture Notes at Fudan
(9.17.5) [D

v
k
(e

) D

v
k
(0)[
1
2n
3
, k N.
Similar as in the proof of the theorem 9.16, we get a harmonic function v

(y) dened on whole

R
n
with
[v

]
2,
2, [v

(y)[ [y[
2+
, y R
n
and [D

(e

) D

(0)[
1
2n
3
.
We then have a contradiction.
b) In case of t
k
t

< +(h
k
+0). Writing x
k
= (x

k
, x
k;n
), we denote v
k
(y) = h
2
k
u
k
((x

k
, 0)+
h
k
y). Then,
v
k
(y)[
y
n
=0
= 0, [v
k
]
2,
= 1, [v
k
]

1
k
, k N.
and
[D

v
k
((0, t
k
) +e

) D

v
k
(0, t
k
)[
1
2n
3
.
Denote
v
k
(y) = v
k
(y) v
k
(0)

1in
D
y
i
v
k
(0)y
i

1
2

1i,jn
D
y
i
y
j
v
k
(0)y
i
y
j
.
Then, we have
(9.17.6) v
k
(0) = 0, D v
k
(0) = 0, D
2
v
k
(0) = 0, k N;
(9.17.7) [ v
k
]

1
k
, k N;
(9.17.8) [D
2
v
k
(y)[ [ v
k
]
2,
[y[

= [y[

, [D v
k
(y)[ [y[
+1
, [ v
k
(y)[ [y[
+2
, k N;
(9.17.9) [D

v
k
((0, t
k
) +e

) D

v
k
((0, t
k
))[
1
2n
3
, k N.
Similar as in the proof of the theorem 9.16, we obtain v

C
2,
(R
n
+
) satisfying
(9.17.10) [v

]
2,;R
n
+
2, [v

(y)[ [y[
2+
, y R
n
+
and [D

((0, t

) +e

) D

(0, t

)[
1
2n
3
,
(9.17.11)
_
v

= 0, in R
n
+
v

[
R
n
+
= 0.
Thus, from 9.17.10,9.17.11 and the corollary 9.15, v

is a polynomial on R
n
+
. Then it contradicts
to that
[D

((0, t

) +e

) D

(0, t

)[
1
2n
3
.

Analysis 79
9.3. Campanatos characterization of L
2
functions to be Holder continuous.
Theorem 9.18 (Campanato Lemma). Suppose u L
2
(B
2R
(x
0
)), (0, 1], > 0 are constants,
and
(9.18.1) inf
R

n
_
B

(y)
[u [
2

2
(

R
)
2
for every ball B

(y) such that y B

R
(x
0
) and R. Then there is a Holder continuous representa-
tion u for the L
2
-class of u with
[ u(x) u(y)[ C
n,
(
[x y[
R
)

. x, y B
R
(x
0
),
where C
n,
depends only on n and .
Remark. Actually,
inf
R
_
B

(y)
[u [
2
=
_
B

(y)
[u
y,
[
2
,
where

y,
=
1

n
_
B

(y)
u.
Proof. a) First by the inequality 9.18.1, we note that
(

2
)
n
_
B
/2
(y)
[u
y,
[
2
2
n

n
_
B

(y)
[u
y,
[
2
2
n

2
(

R
)
2
,
where

y,
=
1

n
_
B

(y)
u.
On the other hand,we have
(

2
)
n
_
B
/2
(y)
[u
y,/2
[
2
= inf
R
(

2
)
n
_
B
/2
(y)
[u [
2
(

2
)
n
_
B
/2
(y)
[u
y,
[
2
2
n

2
(

R
)
2
.
Since
[
y,

y,/2
[
2
2([u
y,
[
2
+[u
y,/2
[
2
),
we conclude that
[
y,/2

y,
[ 2
N

1
2
n
(

R
)

provided that < R and y B

R
(x
0
), where
n
is the volume of the unit ball in R
n
.
b) For any = 0, 1, 2, , we can choose = 2

R and obtain
(9.18.2) [
y,2
1

y,2

[ 2
N

1
2
n
2

.
Denote
s
k
=
k

=0
(
y,2
1

y,2

).
We nd

=0
[
y,2
1

y,2

[ < ,
hence lim
k
s
k
exists and so does lim
k

y,2

y
:= lim
k

y,2

= lim
k
s
k
+
y,
< .
80 Lecture Notes at Fudan
Moreover, denote u :=
y
,using the Lebesgue Lemma, we have
u(y) = u(y), a.e. on B
2R
(x
0
),
and
[
y,2

uy[

j=
[
y,2
j1

y,2
j

[ C2

,
where the constant C = C(n, ).
c) Using
[u u(y)[
2
2([u
y,2

[
2
+[ u(y)
y,2

[
2
),
we obtain

n
_
B

(y)
[u u(y)[
2
dx C
2
(

R
)
2
, = 2

R = 0, 1, 2, .
On the other hand for any (0, R/2] there is an integer
0
1 such that 2
nu1
R < 2

R.
Replacing 2
n+2
C with C, we conclude that
(9.18.3)
n
_
B

(y)
[u u(y)[
2
dx C
2
(

R
)
2
, (0, R/2]
d) Now take any pair y, z B
R
(x
0
) with [y z[ R/4 and let = [y z[. Since
B
/2
(
y +z
2
) B

(y) B

(z),
and
[ u(y) u(z)[
2
2([u u(y)[
2
+[u u(z)[
2
),
the inequality 9.18.3 gives
[ u(y) u(z)[
2
[B
/2
(
y +z
2
)[
1
_
B
/2
(
y+z
2
)
2([u u(y)[
2
+[u u(z)[
2
)dx
2C
n
_
B

(y)
[u u(y)[
2
dx + 2C
n
_
B

(z)
[u u()[
2
dx
4C
2
(

R
)
2
.
Hence,
(9.18.4) [ u(y) u(z)[ 2C(

R
)

= 2
1+
C(
[y z[
R
)

, y, z B
R
(x
0
) with [y z[ R/4.
e) If y, z B
R
(x
0
), we can pick points z
1
= y, , z
8
= z on the segment jointing y, z such that
[z
i
z
i+1
[ R/4. Applying 9.18.4, we obtain
[ u(y) u(z)[ C(n, )
[y z[

, y, z B
R
(x
0
).
We complete the proof.

Theorem 9.19 (Morreys Lemma). Let B

R
(x
0
) R
n
be an open ball. Suppose u W
1,2
(B
R
(x
0
)),
(0, 1], > 0 are constants, and

2n
_
B
|
Rho(y)
[Du[
2
dx
2
(

R
)
2
, y B
R
(x
0
), (0,
R
2
].
Analysis 81
Then u C
0,
(B
R
(x
0
)) ( there is a Holder continuous representation for the L
2
-class of u ), and in
fact
[u(x) u(y)[ C(
[x y[
R
)

, x, y B
R
(x
0
),
where C depends only on n.
Proof. The Poincare inequality (c.f. the remark 9.9.2) gives
_
B

(y)
[u
y,
[
2
dx C
2n
_
B

(y)
[Du[
2
dx
2
(

R
)
2
,
for each y BR
2
(x
0
) and each (0,
R
2
], where

y,
=
1

n
_
B

(y)
u.
Using the Campanato Lemma (c.f. the theorem 9.18.1), we then have the required result.
82 Lecture Notes at Fudan
10. Advanced Topics in Analysis
10.1. Topic on Riemann Zeta Function(To be Continuous). To be Continuous .....
Analysis 83
10.2. Elementary on Nevanlinna Theory(To be Continuous). To be Continuous .....
84 Lecture Notes at Fudan
10.3. Elementary on p-adic Series(To be Continuous). To be Continuous .....