This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

VIPUL NAIK

Abstract. This is a somewhat modiﬁed version of the notes I had prepared for a lecture on inequalities

that formed part of a training camp organized by the Association of Mathematics Teachers of India for

preparation for the Indian National Mathematical Olympiad (INMO) for students from Tamil Nadu.

1. Basic idea of inequalities

1.1. What we need to prove. An “inequation” is an expression of the form:

F ≥ 0

where F is an expression in terms of certain variables. An “inequality”’ is an inequation that is

satisﬁed for all values of the variables (within a certain range).

For instance:

x

2

−x + 1 ≥ 0

and

x

2

−x −1 ≥ 0

are both inequations. Among these, the ﬁrst inequation is true for all real x, while the second

inequation is true for all values of x within a certain range.

Thus, when we talk of an inequality, we have the following in mind:

• The underlying inequation

• The range of values over which the inequality is true

A strict inequation is an inequation of the form:

F > 0

where F is an expression in terms of the variables.

Given any inequation F ≥ 0 we can consider the corresponding strict inequation F > 0.

Thus, when studying an inequality, we are interested in:

• The underlying inequation

• The range of values over which the inequality is true

• The values for which exact equality holds

Some other points to note:

• Any inequation of the form F ≥ G where F and G are both expressions can be written in the

standard form as F −G ≥ 0. The original inequation is true for precisely those values for which

the standard form is true. The equality conditions are also the same.

• An inequation of the form F ≤ G can be expressed as G−F ≥ 0. Again, the original inequation

is true for precisely those values for which the standard form is true. The equality conditions

are also the same.

c Vipul Naik, B.Sc. (Hons) Math and C.S., Chennai Mathematical Institute.

1

1.2. No square is negative. This basic inequality states:

x

2

≥ 0

The range is all x ∈ R and equality holds iﬀ x = 0.

This can be generalized to something of the form:

(f(x

1

, x

2

, . . . , x

n

))

2

+ (g(x

1

, x

2

, . . . , x

n

))

2

≥ 0

The range is all x ∈ R and equality holds iﬀ f(x

1

, x

2

, . . . , x

n

) = g(x

1

, x

2

, . . . , x

n

) = 0.

Problem 1. Prove that x

4

−x

2

y

2

+ y

4

≥ 0 for all real x and y, equality holding iﬀ x = y = 0.

Proof. We use:

x

4

−x

2

y

2

+ y

4

= (x

2

−y

2

)

2

+ (xy)

2

Thus, (x

2

−y

2

) plays th role of f above and xy plays the role of g

Clearly then, the left-hand-side is nonnegative, and is 0 if and only if x

2

= y

2

and xy = 0, thus forcing

x = y = 0.

We can extend the idea to sums of more than two squares:

Problem 2. Prove that a

2

+ b

2

+ c

2

+ ab + bc + ca ≥ 0 with equality holding only if a = b = c = 0.

Proof. The left-hand-side can be expressed as 1/2(a

2

+ b

2

+ c

2

+ (a + b + c)

2

). So it is nonnegative and

can be zero only if a = b = c = 0.

Alternatively, the left hand side can also be written as 1/2((a +b)

2

+(b +c)

2

+(c +a)

2

) and is hence

nonnegative, taking the value 0 if and only if a = b = c = 0

Another problem (for which I’m not writing the solution here):

Problem 3. Prove that a

2

+ b

2

+ c

2

−(ab + bc + ca) ≥ 0 with equality holding only if a = b = c.

It turns out that one of the solution techniques for the previous problem can be applied to this one.

1.3. Manipulating about the inequality symbol. The following results are typically used for ma-

nipulating inequalities:

• We can add two inequalities. The greater side gets added to the greater side, the smaller side

to the smaller side. If either inequality is strict, the resultant inequality is again strict. More

generally, the set of values for which the resultant inequality becomes equality is the intersection

of the corresponding sets for each inequality.

• We can multiply both sides of an inequality by a positive number. In general, however, we cannot

multiply two inequalities.

2. Mean inequalities

2.1. Deﬁnition of means. A mean is a good notion of average for a collection of numbers. A mean of

n numbers is thus typically a function from n-tuples of reals to reals, such that:

• If all the members of the tuple are equal, the mean should be equal to all of them. That is, if

a = a

1

= a

2

= . . . a

n

then the mean of a

1

, a

2

, . . . , a

n

is a.

• The mean is a symmetric function of all the elements of the tuple, that is, if the elements are

permuted, the value of the mean remains unchanged. That is, the mean of a

1

, a

2

, . . . , a

n

is the

same as the mean of a

σ(1)

, a

σ(2)

, . . . , a

σ(n)

.

2

• The mean of a collection of positive numbers should be between the smallest number and the

largest number. That is, if a

1

≤ a

2

≤ . . . ≤ a

n

, the mean lies between a

1

and a

n

.

• The mean is an increasing function in each of the arguments. That is, if a

i

≤ a

i

, then the mean of

a

1

, a

2

, . . . , a

i−1

, a

i

, a

i+1

, . . . , a

n

is less than or equal to the mean of a

1

, a

2

, . . . , a

i−1

, a

i

, a

i+1

, . . . , a

n

.

We now deﬁne some typical notions of mean:

Deﬁnition. (1) The arithmetic mean(deﬁned) of n real numbers a

1

, a

2

, a

3

, . . . , a

n

is deﬁned as:

a

1

+ a

2

+ . . . a

n

n

The arithmetic mean is a well-deﬁned notion for any collection of real numbers (positive, negative

or zero).

(2) The geometric mean(deﬁned) of n positive real numbers a

1

, a

2

, a

3

, . . . , a

n

is deﬁned as

(a

1

a

2

. . . a

n

)

1/n

The geometric mean is deﬁned only for positive numbers.

(3) The quadratic mean(deﬁned) or the root-mean-square of n real numbers a

1

, a

2

, a

3

, . . . , a

n

is

deﬁned as:

a

2

1

+ a

2

2

+ . . . + a

2

n

n

(4) The harmonic mean(deﬁned) of n nonzero real numbers a

1

, a

2

, a

3

, . . . , a

n

is deﬁned as:

a

−1

1

+ a

−1

2

+ . . . + a

−1

n

n

−1

For two positive reals a and b, these boil down to the formulas:

Name of the mean Value

Arithmetic mean

(a+b)

2

Geometric mean

√

ab

Quadratic mean

a

2

+b

2

2

Harmonic mean

2ab

a+b

2.2. Inequalities for two variables.

Claim. For positive reals a and b, Q.M. ≥ A.M. ≥ G.M. ≥ H.M.

Proof. We prove Q.M. ≥ A.M. The remaining proofs follow along similar lines:

What we would like to show is that, for all reals a and b:

a

2

+ b

2

2

≥

a + b

2

Since the left side is nonnegative, it suﬃces to show that the square of the left side is greater than or

equal to the square of the right side. That is, we need to show that:

a

2

+ b

2

2

≥

(a + b)

2

4

But the latter rearranges to (a −b)

2

≥ 0. This tells us that the inequality is valid for all real a and b

with equality holding iﬀ a = b.

3

Let’s look at the pattern. The Q.M. is essentially obtained by taking the arithmetic mean of squares

and then taking squareroot. The A.M. is obtained by taking the arithmetic mean of ﬁrst powers and

then taking the ﬁrst root. The H.M. is obtained by taking the arithmetic mean of inverses and then

taking the inverse. This suggests a general deﬁnition:

M

r

(a, b) =

a

r

+ b

r

2

1/r

Then the quadratic mean is M

2

, the arithmetic mean is M

1

, and the harmonic mean is M

−1

.

By this deﬁnition, M

0

does not make sense. But it turns out that, through a suitable limit argument,

we can take M

0

as the geometric mean. In that case, we have:

M

2

≥ M

1

≥ M

0

≥ M

−1

We also know that:

2 ≥ 1 ≥ 0 ≥ −1

Does this suggest something?

2.3. The mean inequalities: an explanation. Let a and b be positive reals. What can we say about

the behaviour of M

r

(a, b) as r varies from −∞ to ∞. It turns out that as r → −∞, M

r

approaches

min{a, b}, and as r → ∞, M

r

→ max{a, b}. Thus, as r steadily increases, M

r

(a, b) steadily goes from

the minimum to the maximum.

The explanation for this can be sought by viewing the r as a kind of weighting of a and b. The greater

the value of r, the greater the dominance of the bigger term, and hence, the greater the mean is to the

bigger term.

2.4. The mean inequalities for many variables. The same phenomena which we observe for two

variables also generalize to more than two variables. We deﬁne:

M

r

(a

1

, a

2

, . . . , a

n

) =

a

r

1

+ a

r

2

+ . . . a

r

n

n

1/r

Again, as r → −∞, M

r

approaches the minimum of the a

n

s, and as r → ∞, M

r

approaches the

maximum of the a

n

s.

3. Cauchy-Schwarz inequality

3.1. Statement. Let (a

1

, a

2

, . . . , a

n

) and (b

1

, b

2

, . . . , b

n

) be two n-tuples of real numbers. Then:

(

¸

i

a

2

i

)(

¸

i

b

2

i

) ≥ (

¸

i

(a

i

b

i

))

2

With equality holding if and only if one of the tuples is zero or if b

i

= λa

i

for some ﬁxed λ independent

of i (that is, the tuple of b

i

s is a scalar multiple of the tuple of a

i

s).

3.2. Vector interpretation. The vector interpretation of Cauchy Schwarz inequality looks at both

a = (a

1

, ma

2

, . . . , a

n

) and b = (b

1

, b

2

, . . . , b

n

) as vectors in R

n

. Then, the left-hand-side is:

|a|

2

|b|

2

where |a| denotes the magnitude or length of the vector a

The right-hand-side is the square of the dot product of the vectors, which is the same as:

(a.b)

2

= |a|

2

|b|

2

cos

2

θ

where θ is the angle between the vectors. Since cos

2

θ ≤ 1 and quality holds if and only if a and b are

collinear, we get a geometric proof of Cauchy-Schwarz inequality.

4

3.3. A trigonometric problem. Consider the following problem:

Problem 4. Maximize

a cos θ + b sin θ

as a function of θ where a and b are ﬁxed reals (and not both zero).

The idea is to view this as a dot product of vectors (a, b) and (cos θ, sin θ). We have:

(a

2

+ b

2

)(cos

2

θ + sin

2

θ) ≥ (a cos θ + b sin θ)

2

Since cos

2

θ + sin

2

θ = 1, we obtain:

(a cos θ + b sin θ) ≤

a

2

+ b

2

A necessary and suﬃcient condition for the magnitude of the left-hand side to be

√

a

2

+ b

2

is that

a/ cos θ = b/ sin θ, giving tan θ = b/a. Among the two possible values for the pair (cos θ, sin θ) we must

pick the one making a cos θ + b sin θ positive.

3.4. A geometric problem. Consider the following problem:

Problem 5. Let A and B be two points in a plane at distance 1. Find the maximum length of a path

from A to B, comprising at most n line segments, with the property that at every stage, the distance

from B is reducing.

The answer is

√

n.

Proof. The idea of the proof is to use induction on n. Let f(n) denote the maximum value for a given n.

We observe that any such optimal path is memoryless in the following sense:

Suppose γ is a path from A to B comprising at most n line segments, and suppose that the ﬁrst line

segment of γ ends at a point P. Now, the part from P to B must be composed of (n −1) line segments

with the property that at every stage, the distance from B is reducing.

Now, whatever path we choose, we could replace it by a path of maximum length from P to B

comprising (n − 1) line segments and with the property that distance from B is reducing. Since the

original thing was longest, we conclude that the part from P to B must also be the longest one.

Now what is the longest possible path of (n − 1) line segments from P to B? Since lengths scale, it

is the length PB times the value f(n −1). We thus get:

length ofγ = AP + PBf(n −1)

Thus the maximum of the possible lengths of γ is the maximum over all P of the above expression.

Now, from the fact that along the path AP, the distance from P is steadily reducing, we obtain that

the angle ∠APB is either obtuse or right. Thus, in particular, for any given length AP, we have:

PB ≤

1 −AP

2

If equality does not hold, we could replace P by another point Q such that AQ = AP and such that

∠AQB = π/2. Then, QB would be greater than PB, and hence, the length of the longest path would

increase. Hence, we conclude that equality does indeed hold for the longest path, viz ∠APB = π/2.

Let θ be ∠BAP. Then AP = cos θ and PB = sin θ. We thus get:

length of γ = max

θ

cos θ + f(n −1) sin θ

Thus, applying the result of the previous problem:

f(n) =

1 + (f(n −1))

2

Since f(1) = 1 (clearly) we get f(n) =

√

n.

5

4. Rearrangement and Chebyshev inequality

4.1. Rearrangement inequality: statement. Let (a

1

, a

2

, . . . , a

n

) and (b

1

, b

2

, . . . , b

n

) be two n-tuples

of real numbers such that a

1

≥ a

2

≥ . . . ≥ a

n

and b

1

≥ b

2

≥ . . . b

n

. Let σ be a permutation of the

numbers 1, 2, . . . , n. Then:

¸

i

a

i

b

i

≥

¸

i

a

i

b

σ(i)

In other words, the sum of pairwise products is maximum if we pair the largest with the largest, the

second largest with the second largest, and so on.

Equality holds if and only if, for each i, a

i

= a

σ(i)

or b

i

= b

σ(i)

.

Further:

¸

i

a

i

b

σ(i)

≥

¸

i

a

i

b

n+1−i

In other words, the sum of pairwise products is minimum if we pair the largest with the smallest, the

second largest with the second smallest, and so on.

4.2. Idea behind the inequality. Think of it as a resource allocation problem. For instance, suppose

a thief has 3 bags and 3 kinds of coins (gold, silver, copper) to pack in them, and she must pack a

diﬀerent kind of coin in each bag. Assume further that the coins are available in unlimited quantities.

Then, in order to maximize her loot, she will put the gold coins in the biggest bag, the silver coins in

the second biggest bag, and the copper coins in the third biggest bag.

The idea is: send the most to the best. Such an allocation principle is often called a greedy allocation

principle.

The Rearrangement inequality is best proved for two elements, and then extended by induction. Let

a

1

≥ a

2

and b

1

≥ b

2

. Then we have:

(a

1

−a

2

)(b

1

−b

2

) ≥ 0

Manipulating this gives us:

a

1

b

1

+ a

2

b

2

≥ a

1

b

2

+ a

2

b

1

The rearrangement inequality thus illustrates the general statement the principles of optimization and

equality are often at crossroads.

To use this to prove the result globally, we start with the expression

¸

i

a

i

b

σ(i)

and locate indices i, j

for which i < j but σ(i) > σ(j). We then change the permutation to one sending i to σ(j) and j to σ(i)

(and having the same eﬀect as σ on the others). This local change increases the value of the expression

and hence it is clearly not the optimum value.

Note here that equality holds only if a

i

= a

j

or b

σ(i)

= b

σ(j)

.

4.3. An application of rearrangement. Consider the following problem I had mentioned earlier:

Prove that a

2

+ b

2

+ c

2

−(ab + bc + ca) ≥ 0 with equality holding only if a = b = c.

This problem can also be solved using the rearrangement inequality. First observe that since the

expression is symmetric in a, b and c, we can assume without loss of generality that a ≥ b ≥ c.

Consider the triple (a, b, c). This is an ordered triple with the property that the elements are in non-

increasing order. Then (b, c, a) is a permutation of this expression. Thus, by rearrangement inequality:

aa + bb + cc ≥ ab + bc + ca

Which gives us what we want.

Also note that in this case, equality holds if and only if a = b = c.

4.4. Chebyshev inequality. Chebyshev inequality says that sending the most to the best is better

than giving the average to the average. More formally, if (a

1

, a

2

, . . . , a

n

) and (b

1

, b

2

, . . . , b

n

) are two

n-tuples of decreasing reals:

¸

i

a

i

b

i

≥

¸

i

a

i

¸

i

b

i

n

Where equality holds iﬀ either all the a

i

s are equal or all the b

i

s are equal.

6

4.5. Fundamental diﬀerence between Chebyshev and Cauchy-Schwarz. Both the Chebyshev

and the Cauchy-Schwarz inequalities are similar in the following sense:

• They are both true for all reals

• They both provide bounds of

¸

i

a

i

b

i

But they are diﬀerent in the following ways:

• In Chebyshev, it is important to order the a

i

s and b

i

s in descending order, whereas Cauchy-

Schwarz is applicable for any ordering

• Chebyshev gives a bound in terms of

¸

i

a

i

and

¸

i

b

i

while Cauchy-Schwarz gives a bound in

terms of the sums of their squares.

• Chebyshev provides a lower bound on

¸

i

a

i

b

i

while Cauchy-Schwarz provides an upper bound

• The equality case is diﬀerent in both. In Chebyshev, equality holds if all the elements in one of

the tuples are equal. In Cauchy-Schwarz, equality holds if the two tuples are scalar multiples of

one another.

A word of caution, though, when deciding whether to apply Chebyshev or Cauchy-Schwarz. Just

because the inequality seems to require a lower bound on

¸

i

F

i

G

i

, does not mean that Chebyshev is the

one to be used. In fact, we could still use Cauchy-Schwarz by taking a

i

= F

i

G

i

and b

i

to be 1/F

i

.

5. Nesbitt’s inequality

5.1. Statement of the inequality.

Problem 6 (Nesbitt’s inequality). For positive a, b and c, prove that:

a

b + c

+

b

c + a

+

c

a + b

≥

3

2

with equality holding if and only if a = b = c.

5.2. Applying Cauchy-Schwarz (direct application fails). To apply Cauchy-Schwarz we need to

put the terms

a

b+c

and its analogues on the left side, which means we should view each of them as a

square. Their squareroots are

a

b+c

and its analogues. Thus, one tuple is:

a

b + c

,

b

c + a

,

c

a + b

**We would like the other tuple to be something that cancels the denominator. A natural choice is
**

√

b + c,

√

c + a,

√

a + b

**. Unfortunately, this fails to yield the answer, because the expression that we
**

get is upper-bounded, rather than lower-bounded, in the case of equality.

5.3. Applying Chebyshev. Consider the tuples (a, b, c) and ((b + c)

−1

, (c + a)

−1

, (a + b)

−1

). We ﬁrst

need to determine whether they are arranged in the same order. Assume without loss of generality that

a ≥ b ≥ c. Then b + c ≤ c + a ≤ a + b, and taking inverses, we obtain that the second tuple also has its

coordinates in descending order.

We are thus in a position to apply Chebyshev’s and obtain that the give expression is at least:

(a + b + c)((b + c)

−1

+ (c + a)

−1

+ (a + b)

−1

)

3

Now using A.M.-H.M. inequality for the quantities (b + c), (c + a) and (a + b), we get the required

result.

5.4. A short proof. Another way of proving the result is to add and subtract 3, thus writing it as:

(a + b + c)

1

b + c

+

1

c + a

+

1

a + b

**And now apply the A.M.-H.M. inequality.
**

7

6. A past IMO problem

6.1. The problem statement.

Problem 7 (IMO 1995). Prove that if a, b and c are positive reals such that abc = 1, then:

1

a

3

(b + c)

+

1

b

3

(c + a)

+

1

c

3

(a + b)

≥

3

2

The ﬁrst trick is to put x = 1/a, y = 1/b and z = 1/c. The left-hand side becomes:

x

2

y + z

+

y

2

z + x

+

z

2

x + y

6.2. Cauchy-Schwarz. After this point, the ﬁrst possibility to consider is Cauchy-Schwarz. Since we

want to lower-bound the sum here, we must view

x

2

y+z

and its analogues as squares of a tuple. The other

tuple is obtained by cancelling denominators from third tuple. We thus have tuples:

x

√

y + z

,

y

√

z + x

,

z

√

x + y

and

√

y + z,

√

z + x,

√

x + y

We apply Cauchy-Schwarz to these tuples, and then use A.M.-G.M. inequality and the fact that

xyz = 1.

If we keep track of the inequality constraints at each step, we obtain that equality holds if and only

if x = y = z = 1, and hence a = b = c = 1.

8

Index

arithmetic mean, 2

Cauchy-Schwarz inequality, 3

Chebyshev inequality, 5

geometric mean, 2

harmonic mean, 2

mean

arithmetic, 2

geometric, 2

harmonic, 2

quadratic, 2

quadratic mean, 2

rearrangement inequality, 4

9

x2 . 2 . xn ))2 + (g(x1 . More generally. 1. if the elements are permuted. the mean should be equal to all of them.2. . an is a. if a = a1 = a2 = . . . Prove that x4 − x2 y 2 + y 4 ≥ 0 for all real x and y. the smaller side to the smaller side. . That is. . . 2. So it is nonnegative and can be zero only if a = b = c = 0. Proof. • We can multiply both sides of an inequality by a positive number. x2 . an then the mean of a1 . the set of values for which the resultant inequality becomes equality is the intersection of the corresponding sets for each inequality. taking the value 0 if and only if a = b = c = 0 Another problem (for which I’m not writing the solution here): Problem 3. Deﬁnition of means. we cannot multiply two inequalities. The left-hand-side can be expressed as 1/2(a2 + b2 + c2 + (a + b + c)2 ). Mean inequalities 2.1. . . . . . thus forcing x = y = 0. (x2 − y 2 ) plays th role of f above and xy plays the role of g Clearly then. x2 . If either inequality is strict. Manipulating about the inequality symbol. A mean is a good notion of average for a collection of numbers. . . The following results are typically used for manipulating inequalities: • We can add two inequalities. This basic inequality states: x2 ≥ 0 The range is all x ∈ R and equality holds iﬀ x = 0. the left-hand-side is nonnegative. such that: • If all the members of the tuple are equal. . . Problem 1. . .3.1. equality holding iﬀ x = y = 0. aσ(n) . In general. that is. the resultant inequality is again strict. Proof. . . the value of the mean remains unchanged. . a2 . xn ))2 ≥ 0 The range is all x ∈ R and equality holds iﬀ f (x1 . x2 . . an is the same as the mean of aσ(1) . Prove that a2 + b2 + c2 − (ab + bc + ca) ≥ 0 with equality holding only if a = b = c. . A mean of n numbers is thus typically a function from n-tuples of reals to reals. No square is negative. a2 . . It turns out that one of the solution techniques for the previous problem can be applied to this one. . . the mean of a1 . the left hand side can also be written as 1/2((a + b)2 + (b + c)2 + (c + a)2 ) and is hence nonnegative. Prove that a2 + b2 + c2 + ab + bc + ca ≥ 0 with equality holding only if a = b = c = 0. . . . and is 0 if and only if x2 = y 2 and xy = 0. Alternatively. This can be generalized to something of the form: (f (x1 . . xn ) = g(x1 . . xn ) = 0. That is. • The mean is a symmetric function of all the elements of the tuple. We use: x4 − x2 y 2 + y 4 = (x2 − y 2 )2 + (xy)2 Thus. aσ(2) . We can extend the idea to sums of more than two squares: Problem 2. The greater side gets added to the greater side. however.

a2 . an )1/n The geometric mean is deﬁned only for positive numbers. ai . . negative or zero). . . . . a2 . an . . . . ≤ an . an is deﬁned as (a1 a2 . then the mean of a1 . We prove Q. a3 . . a2 . . . .M. (1) The arithmetic mean(deﬁned) of n real numbers a1 .M. . . if ai ≤ ai . . a3 . . a3 . Q. ai . . Proof. . . That is. + a−1 n 1 2 n −1 (4) The harmonic mean(deﬁned) For two positive reals a and b. (a+b) 2 √ ab a2 +b2 2 2ab a+b Claim. . We now deﬁne some typical notions of mean: Deﬁnition. . ai−1 . .M. a2 . + a2 n 1 2 n of n nonzero real numbers a1 . That is. . (3) The quadratic mean(deﬁned) or the root-mean-square of n real numbers a1 .• The mean of a collection of positive numbers should be between the smallest number and the largest number. • The mean is an increasing function in each of the arguments. . a2 .2. (2) The geometric mean(deﬁned) of n positive real numbers a1 . . For positive reals a and b. an is deﬁned as: a−1 + a−1 + . . . . ai−1 . ≥ H.M. ≥ G. these boil down to the formulas: Name of the mean Value Arithmetic mean Geometric mean Quadratic mean Harmonic mean 2. . we need to show that: a2 + b2 (a + b)2 ≥ 2 4 But the latter rearranges to (a − b)2 ≥ 0. a2 . . . ai+1 . ≥ A. the mean lies between a1 and an . . . a3 . . for all reals a and b: a2 + b2 a+b ≥ 2 2 Since the left side is nonnegative. . . . The remaining proofs follow along similar lines: What we would like to show is that. 3 . ai+1 . . if a1 ≤ a2 ≤ . That is. an is less than or equal to the mean of a1 . an n The arithmetic mean is a well-deﬁned notion for any collection of real numbers (positive. .M. it suﬃces to show that the square of the left side is greater than or equal to the square of the right side. . This tells us that the inequality is valid for all real a and b with equality holding iﬀ a = b. an is deﬁned as: a2 + a2 + . . Inequalities for two variables. an is deﬁned as: a1 + a2 + .M. ≥ A. .

b) steadily goes from the minimum to the maximum. This suggests a general deﬁnition: ar + br 2 1/r Mr (a. a2 . Cauchy-Schwarz inequality 3. Let a and b be positive reals. the tuple of bi s is a scalar multiple of the tuple of ai s). Statement.4. Vector interpretation. . and hence. The explanation for this can be sought by viewing the r as a kind of weighting of a and b. We deﬁne: ar + ar + . Let (a1 . the arithmetic mean is M1 . through a suitable limit argument. . which is the same as: (a. . . Since cos2 θ ≤ 1 and quality holds if and only if a and b are collinear. an ) and (b1 . bn ) as vectors in Rn . Mr → max{a. The mean inequalities: an explanation. But it turns out that. Then: ( i a2 )( i i b2 ) ≥ ( i i (ai bi ))2 With equality holding if and only if one of the tuples is zero or if bi = λai for some ﬁxed λ independent of i (that is. is obtained by taking the arithmetic mean of ﬁrst powers and then taking the ﬁrst root.b)2 = |a| |b| cos2 θ where θ is the angle between the vectors. The same phenomena which we observe for two variables also generalize to more than two variables. Mr approaches min{a. . and as r → ∞. b}. . is obtained by taking the arithmetic mean of inverses and then taking the inverse. b2 . b2 . ar n 1 2 n 1/r Mr (a1 . a2 .M. Mr approaches the minimum of the an s. . b}. 3. . . an ) = Again. . Thus. . It turns out that as r → −∞. The mean inequalities for many variables. . .Let’s look at the pattern. The vector interpretation of Cauchy Schwarz inequality looks at both a = (a1 . Mr approaches the maximum of the an s.2. .M. The A. The Q. . we can take M0 as the geometric mean. The greater the value of r. . is essentially obtained by taking the arithmetic mean of squares and then taking squareroot. . 3. M0 does not make sense. The H. What can we say about the behaviour of Mr (a. b) as r varies from −∞ to ∞.M. . . we have: M2 ≥ M1 ≥ M0 ≥ M−1 We also know that: 2 ≥ 1 ≥ 0 ≥ −1 Does this suggest something? 2.1. . the greater the mean is to the bigger term.3. as r → −∞. and as r → ∞. By this deﬁnition. bn ) be two n-tuples of real numbers. . 4 2 2 . Mr (a. b) = Then the quadratic mean is M2 . the greater the dominance of the bigger term. we get a geometric proof of Cauchy-Schwarz inequality. . ma2 . and the harmonic mean is M−1 . 2. an ) and b = (b1 . Then. the left-hand-side is: |a| |b| 2 2 where |a| denotes the magnitude or length of the vector a The right-hand-side is the square of the dot product of the vectors. as r steadily increases. In that case.

and suppose that the ﬁrst line segment of γ ends at a point P . with the property that at every stage. We observe that any such optimal path is memoryless in the following sense: Suppose γ is a path from A to B comprising at most n line segments. in particular. we could replace it by a path of maximum length from P to B comprising (n − 1) line segments and with the property that distance from B is reducing.3. Now. Among the two possible values for the pair (cos θ.4. Consider the following problem: Problem 4. comprising at most n line segments. the distance from B is reducing. the length of the longest path would increase. we obtain: √ A necessary and suﬃcient condition for the magnitude of the left-hand side to be a2 + b2 is that a/ cos θ = b/ sin θ. QB would be greater than P B. 1 + (f (n − 1))2 5 . The idea of the proof is to use induction on n. Then AP = cos θ and P B = sin θ. 3. We thus get: length of γ = max cos θ + f (n − 1) sin θ θ Thus. Now. the distance from B is reducing. whatever path we choose. Now what is the longest possible path of (n − 1) line segments from P to B? Since lengths scale. Let θ be ∠BAP . Let A and B be two points in a plane at distance 1. Let f (n) denote the maximum value for a given n. sin θ). giving tan θ = b/a. we conclude that equality does indeed hold for the longest path. for any given length AP . √ The answer is n. Find the maximum length of a path from A to B. it is the length P B times the value f (n − 1). The idea is to view this as a dot product of vectors (a. we obtain that the angle ∠AP B is either obtuse or right. Maximize a cos θ + b sin θ as a function of θ where a and b are ﬁxed reals (and not both zero). the distance from P is steadily reducing. viz ∠AP B = π/2.3. A trigonometric problem. we could replace P by another point Q such that AQ = AP and such that ∠AQB = π/2. We have: (a2 + b2 )(cos2 θ + sin2 θ) ≥ (a cos θ + b sin θ)2 Since cos2 θ + sin2 θ = 1. Thus. Then. we have: PB ≤ 1 − AP 2 If equality does not hold. Consider the following problem: (a cos θ + b sin θ) ≤ a2 + b2 Problem 5. from the fact that along the path AP . applying the result of the previous problem: f (n) = √ Since f (1) = 1 (clearly) we get f (n) = n. Now. sin θ) we must pick the one making a cos θ + b sin θ positive. b) and (cos θ. A geometric problem. Hence. We thus get: length ofγ = AP + P Bf (n − 1) Thus the maximum of the possible lengths of γ is the maximum over all P of the above expression. the part from P to B must be composed of (n − 1) line segments with the property that at every stage. we conclude that the part from P to B must also be the longest one. Proof. and hence. Since the original thing was longest.

For instance. Then: ai bi ≥ i i ai bσ(i) In other words. an ) and (b1 . . a2 . we start with the expression i ai bσ(i) and locate indices i. . . Consider the triple (a. bn . the second largest with the second smallest. bn ) are two n-tuples of decreasing reals: ai bi ≥ i 6 i ai n i bi Where equality holds iﬀ either all the ai s are equal or all the bi s are equal. . . she will put the gold coins in the biggest bag. Such an allocation principle is often called a greedy allocation principle. Thus. j for which i < j but σ(i) > σ(j). suppose a thief has 3 bags and 3 kinds of coins (gold. . To use this to prove the result globally. . . Note here that equality holds only if ai = aj or bσ(i) = bσ(j) . copper) to pack in them. The Rearrangement inequality is best proved for two elements. . This local change increases the value of the expression and hence it is clearly not the optimum value. and then extended by induction. We then change the permutation to one sending i to σ(j) and j to σ(i) (and having the same eﬀect as σ on the others). ≥ an and b1 ≥ b2 ≥ . 4. Let a1 ≥ a2 and b1 ≥ b2 . 2. 4. ai = aσ(i) or bi = bσ(i) . . Equality holds if and only if. Then (b.4. Rearrangement and Chebyshev inequality 4. equality holds if and only if a = b = c. . . 4. silver. b2 . . and so on. Then we have: (a1 − a2 )(b1 − b2 ) ≥ 0 Manipulating this gives us: a1 b1 + a2 b2 ≥ a1 b2 + a2 b1 The rearrangement inequality thus illustrates the general statement the principles of optimization and equality are often at crossroads. c). c. an ) and (b1 . . b2 .1. the sum of pairwise products is maximum if we pair the largest with the largest. for each i. . Chebyshev inequality. a2 . Consider the following problem I had mentioned earlier: Prove that a2 + b2 + c2 − (ab + bc + ca) ≥ 0 with equality holding only if a = b = c. b.3. Let (a1 . . .2. and she must pack a diﬀerent kind of coin in each bag. More formally. in order to maximize her loot. An application of rearrangement.4. we can assume without loss of generality that a ≥ b ≥ c. . b and c. First observe that since the expression is symmetric in a. Rearrangement inequality: statement. Assume further that the coins are available in unlimited quantities. The idea is: send the most to the best. a) is a permutation of this expression. This problem can also be solved using the rearrangement inequality. . Then. bn ) be two n-tuples of real numbers such that a1 ≥ a2 ≥ . . by rearrangement inequality: aa + bb + cc ≥ ab + bc + ca Which gives us what we want. Also note that in this case. the sum of pairwise products is minimum if we pair the largest with the smallest. and the copper coins in the third biggest bag. n. and so on. Think of it as a resource allocation problem. . if (a1 . Let σ be a permutation of the numbers 1. . the silver coins in the second biggest bag. . Further: ai bσ(i) ≥ i i ai bn+1−i In other words. Chebyshev inequality says that sending the most to the best is better than giving the average to the average. the second largest with the second largest. . This is an ordered triple with the property that the elements are in nonincreasing order. . Idea behind the inequality.

inequality for the quantities (b + c). A word of caution.4. prove that: b c 3 a + + ≥ b+c c+a a+b 2 with equality holding if and only if a = b = c. A natural choice is √ b + c. whereas CauchySchwarz is applicable for any ordering • Chebyshev gives a bound in terms of i ai and i bi while Cauchy-Schwarz gives a bound in terms of the sums of their squares.4. it is important to order the ai s and bi s in descending order. b+c b .M. c+a c a+b √ √We would like the other tuple to be something that cancels the denominator. (c + a)−1 . Their squareroots are a b+c and its analogues. 5. one tuple is: a . In Chebyshev. To apply Cauchy-Schwarz we need to a put the terms b+c and its analogues on the left side. Statement of the inequality. Thus. a + b . thus writing it as: (a + b + c) And now apply the A.-H. • Chebyshev provides a lower bound on i ai bi while Cauchy-Schwarz provides an upper bound • The equality case is diﬀerent in both. though. (c + a) and (a + b). 5. Unfortunately. 7 1 1 1 + + b+c c+a a+b .M. does not mean that Chebyshev is the one to be used. we obtain that the second tuple also has its coordinates in descending order. We ﬁrst need to determine whether they are arranged in the same order.1.5. in the case of equality. b and c. equality holds if all the elements in one of the tuples are equal. this fails to yield the answer. We are thus in a position to apply Chebyshev’s and obtain that the give expression is at least: (a + b + c)((b + c)−1 + (c + a)−1 + (a + b)−1 ) 3 Now using A. Fundamental diﬀerence between Chebyshev and Cauchy-Schwarz. Applying Cauchy-Schwarz (direct application fails). Just because the inequality seems to require a lower bound on i Fi Gi . inequality. In Cauchy-Schwarz. In fact. c + a. we could still use Cauchy-Schwarz by taking ai = Fi Gi and bi to be 1/Fi . Nesbitt’s inequality 5. and taking inverses. which means we should view each of them as a square. Then b + c ≤ c + a ≤ a + b.2. b. because the expression that we get is upper-bounded. Applying Chebyshev. For positive a. rather than lower-bounded.M. 5. Consider the tuples (a.M. Problem 6 (Nesbitt’s inequality). Assume without loss of generality that a ≥ b ≥ c. equality holds if the two tuples are scalar multiples of one another. when deciding whether to apply Chebyshev or Cauchy-Schwarz. (a + b)−1 ). Both the Chebyshev and the Cauchy-Schwarz inequalities are similar in the following sense: • They are both true for all reals • They both provide bounds of i ai bi But they are diﬀerent in the following ways: • In Chebyshev.3. Another way of proving the result is to add and subtract 3. we get the required result.-H. A short proof. c) and ((b + c)−1 . 5.

M. The other tuple is obtained by cancelling denominators from third tuple.2. The left-hand side becomes: x2 y2 z2 + + y+z z+x x+y 6. Cauchy-Schwarz.√ .M. After this point. Prove that if a.6.√ y+z x+y z+x 8 . the ﬁrst possibility to consider is Cauchy-Schwarz.1. inequality and the fact that xyz = 1. z + x. b and c are positive reals such that abc = 1. Problem 7 (IMO 1995). y = 1/b and z = 1/c. √ y z x . x + y We apply Cauchy-Schwarz to these tuples. we must view y+z and its analogues as squares of a tuple. and then use A. and hence a = b = c = 1. We thus have tuples: √ and √ √ y + z. Since we x2 want to lower-bound the sum here. The problem statement. If we keep track of the inequality constraints at each step. A past IMO problem 6. we obtain that equality holds if and only if x = y = z = 1.-G. then: a3 (b 1 1 3 1 + 3 + 3 ≥ + c) b (c + a) c (a + b) 2 The ﬁrst trick is to put x = 1/a.

4 9 . 2 quadratic. 2 harmonic mean. 2 harmonic. 2 mean arithmetic. 2 geometric. 5 geometric mean. 3 Chebyshev inequality.Index arithmetic mean. 2 Cauchy-Schwarz inequality. 2 quadratic mean. 2 rearrangement inequality.

- The0ry of Number
- Proof by Bygone Experience
- Rearrangement Inequality
- 05putnam5.Ps
- 24 Trang Tiep Tuyen Hay
- Math assignment
- Concept 4
- Module 2 - Math
- fasc5a
- Reading 1_ Week 1_FN-553-E-1027295
- Math Proof
- IPaMV
- 12027-0130670227_ismSec6
- PS 1 2007 Solutions
- Coherent-States Anna 2012 05
- Calculus to Second Exam
- Zeta_2
- IntegrationTRIGsub Tony
- 6678_01R_msc_20130815
- Pantellas
- Edexcel Mathematics Core 3
- Calculus-Triple integrals
- h4s
- Table
- Fourier Series
- Calculus one and several variables 10E Salas solutions manual ch19
- geotransformation
- Goldstein 13-14-2
- Assg4 (1)
- Practice Paper C3

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd