Scribd
Upload
22 views

Convex 1

This document discusses convex functions and their properties. Some key points: - A function φ is convex if φ(tx + (1-t)y) ≤ tφ(x) + (1-t)φ(y) for all x, y in its domain and t between 0 and 1. - Holder's inequality states that for convex φ, the sum of φ over a weighted sum is less than or equal to the sum of the weights times the individual φ values. - Jensen's inequality states that for convex φ, applying φ to the average of a function f is less than or equal to the average of φ applied to f.

Uploaded by

Sebastian Ilinca
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
22 views

Convex 1

This document discusses convex functions and their properties. Some key points: - A function φ is convex if φ(tx + (1-t)y) ≤ tφ(x) + (1-t)φ(y) for all x, y in its domain and t between 0 and 1. - Holder's inequality states that for convex φ, the sum of φ over a weighted sum is less than or equal to the sum of the weights times the individual φ values. - Jensen's inequality states that for convex φ, applying φ to the average of a function f is less than or equal to the average of φ applied to f.

Uploaded by

Sebastian Ilinca
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

CONVEX FUNCTIONS GIVE INEQUALITIES

PROF. MARIA GIRARDI

1. Basics
Notation 1.1. Throughout this presentation:
(a, b) = I R where a, b R {}
: I R is a function
p (1, ) and its conjugate exponent p0 (1, ) is defined by
Definition 1.2. : I R is convex
x, y I , t (0, 1)

1
p

1
p0

= 1.

(tx + (1 t) y) t(x) + (1 t) (y) .

(1.1)

Picture.

|
|
|
| |

tx + (1 t) y

Date: 7 June 2001.


1

2. Holders Inequality
Lemma 2.1. Let : I R be convex. Then:
xi I ,

ti (0, 1) ,

n
X

ti = 1

n
X

i=1

xi I ,

ti (0, 1)

ti xi

i=1

n
X

ti (xi )

i=1

Pn

 Pn
t (xi )
ti xi
i=1
i=1
Pn i

.
Pn
i=1 ti
i=1 ti

(2.1)

P
Proof. (2.1): Use (1.1) and induction. (2.2): Let ti = [ nj=1 tj ]1 ti and apply (2.1).

(2.2)


Recall. Geometric-Arithmetic Mean Inequality: GM AM


!1/n
n
n
Y
1X
xi .
xi 0 , n N
=
xi

n
i=1
i=1
Using convex functions, we can generalize the GM-AM inequality.
Proposition 2.2. Generalized GM-AM inequality:
xi 0 ,

ti (0, 1) ,

n
X

ti = 1

n
Y

i=1

xtii

n
X

ti xi .

(2.3)

i=1

i=1

Proof. Let : (0, ) R be (x) = ln x. Then


is convex

n
X

ln

!
t i xi

i=1

n
X

ln

xtii

i=1

ln

n
X
i=1
n
X

ti ( ln xi )
!

ti xi

(2.4)

i=1

Now exponentiate both sides of (2.4).

An immediate corollary follows now.


Corollary 2.3. Youngs Inequality:
0

xi 0

1<p<

=
2

xp1
xp2
x1 x2
+ 0 .
p
p

(2.5)

Proof. Apply (2.3) to


x1 x2 (xp1 )1/p

xp2

1/p0

.


Recall. For a sequence {xi }ni=1 = {x1 , . . . , xn } from R:


" n
#1/p
X
k{xi }ni=1 k`p =
|xi |p
.
i=1

Theorem 2.4. Holders Inequality in `p : For sequences {xi }ni=1 and {yi }ni=1 from R:
k{xi yi }ni=1 k`1 k{xi }ni=1 k`p k{yi }ni=1 k`p0
that is
n
X

"
|xi yi |

i=1

n
X

#1/p
|xi |p

"

i=1

n
X

#1/p0
p0

|yi |

(2.6)

i=1

When p = 2 = p0 , note that (2.6) is just the Cauchy-Schwarz Inequality.


Proof. WLOG: Neither k{xi }i k`p nor k{yi }i k`p0 is 0. WLOG: k{xi }i k`p = 1 = k{yi }i k`p0
for if not then:
x
ei :=

xi
= k{e
xi }i k`p = 1
k{xj }j k`p

yei :=

yi
k{yj }j k`

= k{e
yi }i k`p0 = 1 .

p0

By Youngs Inequality
n
X
i=1

#
"
0
n
X
0
|yi |p
1
1
|xi |p
+
=
k{xi }i kp`p + 0 k{yi }i kp`p0 = 1 .
|xi yi |
0
p
p
p
p
i=1


Here are two exercises of Generalized Holders Inequalities in `p .


P
Exercise. For sequences {xji }ni=1 from R and kj=1 p1j = 1
(
)n
k

k
Y
Y


j n
j


{x }i=1
xi
.
i


`pj
j=1

j=1
i=1
`1

Exercise. For sequences {xi }ni=1 and {yi }ni=1 from R and
k{xi }ni=1 k`p

k{xi yi }ni=1 k`p

Hint:

1
p1 /p3

1
p2 /p3

1
p1

1
p2

1
p3

k{yi }ni=1 k`p .


2

=1.

Recall. For a nice function f : I R,


Z
kf kLp =

1/p

|f (x)| dx

Theorem 2.5. Holders Inequality in Lp : For nice functions f, g : I R,


kf gkL1 kf kLp kgkLp0
that is,
Z

1/p

|f (x)|

|f (x)g(x)| dx

Z

1/p0

|g(x)|

p0

Proof. WLOG: Neither kf kLp nor kgkLp0 is 0 or . WLOG: kf kLp = 1 = kgkLp0 for if
not so:
fe :=


f

= fe = 1
kf kLp
Lp

ge :=

g
= ke
g kLp0 = 1 .
kgkLp0

By Youngs Inequality
#
0
Z
Z "
0
|g(x)|p
1
1
|f (x)|p
|f (x)g(x)| dx
+
=
kf kpLp + 0 kgkpLp0 = 1 .
0
p
p
p
p
I
I

Here are two exercises of Generalized Holders Inequalities in Lp .
P
Exercise. For nice functions fj : I R and kj=1 p1j = 1


k
k
Y

Y

kfj kLp .
fj
j


j=1

j=1

L1

Exercise. For nice functions f, g : I R and


kf gkLp

1
p1

kf kLp

1
p2

1
p3

kgkLp .
2

3. Jensens Inequality
Lemma 3.1. Let : I R be convex. Let x, y, z I with x < y < z. Then
(z) (x)
(z) (y)
(y) (x)

.
yx
zx
zy

(3.1)

Picture.

| |||

|
a

Proof. Key idea:


y = tx + (1 t)z =

zy
yx
x+
z
zx
zx

(y) t(x) + (1 t)(z) .

Think of what this says in the picture; the needed algebra then follows easily.
5

Proposition 3.2. Let : I R be convex and x0 I. Then


(1) is continuous at x0
(2) 0 (x0 ) := limxl x0

(x0 )(xl )
x0 xl

exists

(3) 0+ (x0 ) := limxr x+0

(xr )(x0 )
xr x0

exists

(4) 0 (x0 ) 0+ (x0 ) .


Proof. (1): Find: a < a1 < a2 < x0 < b2 < b1 < b. If x [a2 , b2 ] \ {x0 }, then

|
|
[ |
| |
] |
|

a1

a2

x0

b2

b1

by (3.1)
A :=

(a2 ) (a1 )
(b1 ) (b2 )
(x0 ) (x)

:= B
a2 a1
x0 x
b1 b2

and so


(x0 ) (x)
max (|A| , |B|) .



x0 x
(2)(4): Consider

|
|
|
|
|
| |

xl

x
el

x0

x
er

xr

By (3.1)
(x0 ) (xl )
(x0 ) (e
xl )
(e
xr ) (x0 )
(xr ) (x0 )

x0 xl
x0 x
el
x
er x0
x r x0
and so
lim

xl %x0

(x0 ) (xl )
% 0 (x0 ) 0+ (x0 ) .
x 0 xl

lim

xr &x0

(xr ) (x0 )
.
xr x0

(3.2)


Observation 3.3. Let : I R be convex and x0 I. Let


0 (x0 ) m 0+ (x0 ) .
Consider the line
l(x) = m(x x0 ) + (x0 )
through the point (x0 , (x0 )). Then
l(x) (x)

x I .

(3.3)

A line y = l(x) through the point (x0 , (x0 )) that satisfies (3.3) is called a supporting line
of y = (x) at x0 . Draw yourself a picture to see the choice of terminolgy here. Thus
Observation 3.3 says that convex functions always have supporting lines.
Picture.

|
|
|
| |

xl

x0

xr

Proof. By (3.2)
(x0 ) (xl )
(xr ) (x0 )
m
.
x0 x l
xr x0
Thus
m(x x0 ) (x) (x0 )

x I \ {x0 } .


Theorem 3.4. Jensens Inequality: Let : R R be convex and f : I R be integrable.


If I = (0, 1), then
Z


Z
f (x) dx
(f (x)) dx

(3.4)

If I = (a, b) has finite length, then


R
R

f
(x)
dx
(f (x)) dx
I
I
.
ba
ba

(3.5)

Remark. Theorem 3.4 may be thought of as a continuous version of Lemma 2.1.


Proof. (3.4): Let

R
I

f (x) dx = x0 and 0 (x0 ) m 0+ (x0 ). Then by Observation 3.3


m(x x0 ) + (x0 ) (x)

x R .

Thus
m(f (x) x0 ) + (x0 ) (f (x))

x I .

(3.6)

Now integrate both sides of (3.6) over I := (0, 1).


(3.5) follows applying (3.4) to
g(x) := f (a + x (b a)) : I R .


You might also like