THE METHOD OF

LAGRANGE MULTIPLIERS

William F. Trench
Andrew G. Cowles Distinguished Professor Emeritus
Department of Mathematics
Trinity University
San Antonio, Texas, USA
[email protected]

This is a supplement to the authors Introduction to Real Analysis. It has been

judged to meet the evaluation criteria set by the Editorial Board of the American
Institute of Mathematics in connection with the Institutes Open Textbook Initiative.
It may be copied, modified, redistributed, translated, and built upon subject to the
Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. A
complete instructors solution manual is available by email to [email protected],
subject to verification of the requestors faculty status.

THE METHOD OF LAGRANGE MULTIPLIERS

William F. Trench

Foreword

This is a revised and extended version of Section 6.5 of my Advanced Calculus (Harper
& Row, 1978). It is a supplement to my textbook Introduction to Real Analysis, which
is referenced via hypertext links.

Introduction

To avoid repetition, it is to be understood throughout that f and g1 , g2 ,. . . , gm are

continuously differentiable on an open set D in Rn .
Suppose that m < n and
g1 .X/ D g2 .X/ D


D gm .X/ D 0

(1)

on a nonempty subset D1 of D. If X0 2 D1 and there is a neighborhood N of X0 such

that
f .X/  f .X0 /
(2)
for every X in N \D1 , then X0 is a local maximum point of f subject to the constraints
(1). However, we will usually say subject to rather than subject to the constraint(s).
If (2) is replaced by
f .X/  f .X0 /;
(3)
then maximum is replaced by minimum. A local maximum or minimum of f
subject to (1) is also called a local extreme point of f subject to (1). More briefly, we
also speak of constrained local maximum, minimum, or extreme points. If (2) or (3)
holds for all X in D1 , we omit local.
Recall that X0 D .x10 ; x20 ; : : : ; xn0 / is a critical point of a differentiable function
L D L.x1 ; x2; : : : ; xn / if
Lxi .x10 ; x20 ; : : : ; xn0/ D 0;

1  i  n:

Therefore, every local extreme point of L is a critical point of L; however, a critical

point of L is not necessarily a local extreme point of L (pp. 334-5).
Suppose that the system (1) of simultaneous equations can be solved for x1 , . . . ,
xm in terms of the xmC1 , . . . , xn ; thus,
xj D hj .xmC1 ; : : : ; xn/;

1  j  m:

(4)

Then a constrained extreme value of f is an unconstrained extreme value of

f .h1 .xmC1 ; : : : ; xn /; : : : ; hm.xmC1 ; : : : ; xn /; xmC1 ; : : : ; xn /:
2

(5)

However, it may be difficult or impossible to find explicit formulas for h1 , h2 , . . . , hm ,

and, even if it is possible, the composite function (5) is almost always complicated.
Fortunately, there is a better way to to find constrained extrema, which also requires
the solvability assumption, but does not require an explicit formula as indicated in (4).
It is based on the following theorem. Since the proof is complicated, we consider two
special cases first.
Theorem 1 Suppose that n > m: If X0 is a local extreme point of f subject to
g1 .X/ D g2 .X/ D


D gm .X/ D 0
and

@g1 .X0 / @g1 .X0 /

@g1 .X0 /

@x
@x
@x
r1
r2
rm

@g2 .X0 / @g2 .X0 /

@gm .X0 /

(6)
0

@x
@x
@x
r1
r2
rm

::
::
:
:
::
::

:
:

@gm .X0 / @gm .X0 /

@g
.X
/
m
0

@xr1
@xr2
@xrm
for at least one choice of r1 < r2 <


< rm in f1; 2; : : : ; ng; then there are constants
1 ; 2 ; . . . ; m such that X0 is a critical point of
f

1 g1

2 g2

m gm I

that is;
@f .X0 /
@xi

1

@g1 .X0 /
@xi

2

@g2 .X0 /
@xi

m

@gm .X0 /
D 0;
@xi

1  i  n.
The following implementation of this theorem is the method of Lagrange
multipliers.
(a) Find the critical points of
f

1 g1

2 g2

m gm ;

treating 1 , 2 , . . . m as unspecified constants.

(b) Find 1 , 2 , . . . , m so that the critical points obtained in (a) satisfy the constraints.
(c) Determine which of the critical points are constrained extreme points of f . This
can usually be done by physical or intuitive arguments.
If a and b1 , b2 , . . . , bm are nonzero constants and c is an arbitrary constant, then the
local extreme points of f subject to g1 D g2 D


D gm D 0 are the same as the local
extreme points of af c subject to b1 g1 D b2 g2 D


D bm gm D 0. Therefore, we
can replace f 1 g1 2 g2


m gm by af 1 b1 g1 2 b2 g2


m bm gm c
to simplify computations. (Usually, the c indicates dropping additive constants.)
We will denote the final form by L (for Lagrangian).
3

Extrema subject to one constraint

Here is Theorem 1 with m D 1.

Theorem 2 Suppose that n > 1: If X0 is a local extreme point of f subject to g.X/ D
0 and gxr .X0 / 0 for some r 2 f1; 2; : : : ; ng; then there is a constant  such that
fxi .X0 /

gxi .X0 / D 0;

1  i  nI thus; X0 is a critical point of f

Proof

(7)

g:

For notational convenience, let r D 1 and denote

U D .x2 ; x3 ; : : : xn / and U0 D .x20 ; x30 ; : : : xn0 /:

Since gx1 .X0 / 0, the Implicit Function Theorem (Corollary 6.4.2, p. 423) implies
that there is a unique continuously differentiable function h D h.U/; defined on a
neighborhood N  Rn 1 of U0 ; such that .h.U/; U/ 2 D for all U 2 N , h.U0 / D x10 ,
and
g.h.U/; U/ D 0; U 2 N:
(8)
Now define
D

fx1 .X0 /
;
gx1 .X0 /

(9)

which is permissible, since gx1 .X0 / 0. This implies (7) with i D 1. If i > 1,
differentiating (8) with respect to xi yields
@g.h.U/; U/ @h.U/
@g.h.U/; U/
C
D 0;
@xi
@x1
@xi

U 2 N:

(10)

Also,
@f .h.U/; U/
@f .h.U/; U/ @h.U/
@f .h.U/; U//
D
C
;
@xi
@xi
@x1
@xi

U 2 N:

(11)

Since .h.U0 /; U0 / D X0 , (10) implies that

@g.X0 /
@g.X0 / @h.U0 /
C
D 0:
@xi
@x1
@xi

(12)

If X0 is a local extreme point of f subject to g.X/ D 0, then U0 is an unconstrained

local extreme point of f .h.U/; U/; therefore, (11) implies that
@f .X0 /
@f .X0 / @h.U0 /
C
D 0:
@xi
@x1
@xi
Since a linear homogeneous system


 

a b
u
0
D
c d
v
0
4

(13)

has a nontrivial solution if and only if

b
D 0;
d

(Theorem 6.1.15, p. 376), (12) and (13) imply that

@f .X0 / @f .X0 /

@f .X0 /

@xi

@xi
@x
1

D 0; so
@g.X / @g.X /

@f .X /
0
0
0

@xi
@x1
@x1

@g.X0 /
@xi

D 0;
@g.X0 /

@x1

since the determinants of a matrix and its transpose are equal. Therefore, the system
2 @f .X /
0
6 @xi
6
6
4 @f .X /
0
@x1

@g.X0 / 3
 

7
@xi
0
7 u
D
7
0
@g.X / 5 v
0

@x1

has a nontrivial solution (Theorem 6.1.15, p. 376). Since gx1 .X0 / 0, u must be
nonzero in a nontrivial solution. Hence, we may assume that u D 1, so
@f .X0 /
6 @xi
6
6
4 @f .X /
0
@x1
2

3
@g.X0 /
 

7
@xi
0
7 1
D
:
7
0
@g.X0 / 5 v
@x1

(14)

In particular,
@f .X0 /
@g.X0 /
Cv
D 0; so
@x1
@x1

vD

fx1 .X0 /
:
gx1 .X0 /

Now (9) implies that v D , and (14) becomes

2 @f .X /
0
6 @xi
6
6
4 @f .X /
0
@x1

@g.X0 / 3
7
@xi
7
7
@g.X0 / 5
@x1

1


0
0

Computing the topmost entry of the vector on the left yields (7).
Example 1 Find the point .x0 ; y0/ on the line
ax C by D d
closest to a given point .x1 ; y1 /.

p
Solution We must minimize .x x1 /2 C .y y1 /2 subject to the constraint. This
is equivalent to minimizing .x x1 /2 C .y y1 /2 subject to the constraint, which is
simpler. For, this we could let
x1 /2 C .y

L D .x
however,
LD

.x

y1 /2

x1 /2 C .y
2

.ax C by

y1 /2

d /I

.ax C by/

is better. Since
Lx D x

x1

a

Ly D y

and

y1

b;

.x0 ; y0/ D .x1 C a; y1 C b/, where we must choose  so that ax0 C by0 D d .
Therefore,
ax0 C by0 D ax1 C by1 C .a2 C b 2 / D d;
so
D

ax1 by1 /a
.d
; and y0 D y1 C
2
2
a Cb
The distance from .x1 ; y1 / to the line is
x0 D x1 C

.d

ax1 by1
;
a2 C b 2

.x0

x1 /2 C .y0

y1 /2 D

jd

ax1 by1 /b
:
a2 C b 2

ax1 by1 j
p
:
a2 C b 2

Example 2 Find the extreme values of f .x; y/ D 2x C y subject to

x 2 C y 2 D 4:
Solution Let

 2
.x C y 2 /I
2

L D 2x C y
then
Lx D 2

y;
p
so .x0 ; y0 / D p
.2=; 1=/. Since
C
4,  D 5=2. Hence, the constrained
p
maximum is 2 p
5, attained
at
.4=
5;
2=
5/,
and the constrained minimum is 2 5,
p
attained at . 4= 5; 2= 5/.
x02
p

x and Ly D 1
2
yp
0 D

Example 3 Find the point in the plane

3x C 4y C D 1

(15)

closest to . 1; 1; 1/.
Solution We must minimize
f .x; y; / D .x C 1/2 C .y
6

1/2 C .

1/2

subject to (15). Let

LD

.x C 1/2 C .y

1/2 C .

1/2

.3x C 4y C /I

then
Lx D x C 1

3;

Ly D y

4; and L D

;

so
x0 D 1 C 3;

y0 D 1 C 4;

0 D 1 C :

From (15),
3. 1 C 3/ C 4.1 C 4/ C .1 C /
so  D 1=26 and
.x0 ; y0 ; 0 / D

1 D 1 C 26 D 0;


29 22 25
; ;
:
26 26 26

The distance from .x0 ; y0 ; 0 / to . 1; 1; 1/ is

p

.x0 C 1/2 C .y0

1
1/2 D p :
26

1/2 C .0

Example 4 Assume that n  2 and xi  0, 1  i  n.

n
n
X
X
(a) Find the extreme values of
xi subject to
xi2 D 1.
i D1

(b) Find the minimum value of

i D1

n
X

xi2

subject to

i D1

Solution (a) Let

LD
then
Lxi D 1
Hence,

n
X
i D1

n
X
i D1

n
X

xi D 1.

X 2
xi I
2

xi

i D1

i D1

xi ; so xi 0 D

1
;


1  i  n:

p
xi20 D n=2 , so  D n and



1
1
1
.x10 ; x20; : : : ; xn0 / D p ; p ; : : : ; p
:
n
n
n
p
p
Therefore, the constrained maximum is n and the constrained minimum is
n.
Solution (b) Let

1X 2
LD
xi
2
i D1

n
X
i D1

xi I

then

Hence,

n
X
i D1

Lxi D xi

; so xi 0 D ;

1  i  n:

xi 0 D n D 1, so xi 0 D  D 1=n and the constrained minimum is

n
X
i D1

xi20 D

1
n

There is no constrained maximum. (Why?)

Example 5 Show that
x 1=p y 1=q 
if

x
y
C ;
p
q

1
1
C D 1;
p
q

x; y  0;

p > 0; and q > 0:

(16)

Solution We first find the maximum of

f .x; y/ D x 1=p y 1=q
subject to

y
x
C D ;
p
q

x  0;

y  0;

(17)

where  is a fixed but arbitrary positive number. Since f is continuous, it must assume
a maximum at some point .x0 ; y0 / on the line segment (17), and .x0 ; y0 / cannot be an
endpoint of the segment, since f .p; 0/ D f .0; q/ D 0. Therefore, .x0 ; y0 / is in the
open first quadrant.
Let


x
y
1=p 1=q
LDx y

C
:
p
q
Then
Lx D

1
f .x; y/
px


1
and Ly D
f .x; y/
p
qy


D 0;
q

so x0 D y0 D f .x0 ; y0 /=. Now(16) and (17) imply that x0 D y0 D . Therefore,

f .x; y/  f .; / D  1=p  1=q D  D

x
y
C :
p
q

This can be generalized (Exercise 53). It can also be used to generalize Schwarzs
inequality (Exercise 54).

Constrained Extrema of Quadratic Forms

In this section it is convenient to write

2

x1
x2
::
:

6
6
XD6
4

7
7
7:
5

xn

An eigenvalue of a square matrix A D aij ni;j D1 is a number  such that the system
AX D X;
or, equivalently,
.A

I/X D 0;

has a solution X 0. Such a solution is called an eigenvector of A. You probably

know from linear algebra that  is an eigenvalue of A if and only if
I/ D 0:

det.A

Henceforth we assume that A is symmetric .aij D aj i ; 1  i; j  n/. In this case,

det.A

I/ D . 1/n .

where 1 ; 2 ; : : : ; n are real numbers.

The function
Q.X/ D

1 /.

n
X

2 /


.

n /;

aij xi xj

i;j D1

is a quadratic form. To find its maximum or minimum subject to

n
X
i D1

the Lagrangian
L D Q.X/
Then
Lxi D 2
so

n
X

j D1

n
X

j D1

aij xj

n
X

xi2 :

i D1

2xi D 0;

aij xj 0 D xi 0 ;

xi2 D 1, we form

1  i  n;

1  i  n:

Therefore, X0 is a constrained critical point of Q subject to

n
X
i D1

xi2 D 1 if and only

if AX0 D X0 for some ; that is, if and only if  is an eigenvalue and X0 is an
9

associated unit eigenvector of A. If AX0 D X0 and

Q.X0 /

n
X
i

xi20 D 1, then

n
n
n
X
X
X
@
aij xj 0 A xi 0 D
.xi 0 /xi 0
i D1

D 

n
X
i D1

j D1

i D1

xi20 D I

therefore, the largest and smallest eigenvalues of A are the maximum and minimum
n
X
values of Q subject to
xi2 D 1.
i D1

Example 6 Find the maximum and minimum values

Q.X/ D x 2 C y 2 C 22

2xy C 4x C 4y

subject to the constraint

x 2 C y 2 C 2 D 1:

(18)

Solution The matrix of Q is

2

AD4

and
det.A

I/

1
1
2

3
1 2
1 2 5
2 2

1 
1
2

D
1 1 
2
2
2
2 
D
. C 2/. 2/. 4/;

so
1 D 4;

2 D 2;

3 D 2

are the eigenvalues of A. Hence, 1 D 4 and 3 D 2 are the maximum and minimum
values of Q subject to (18).
To find the points .x1 ; y1; 1 / where Q attains its constrained maximum, we first
find an eigenvector of A corresponding to 1 D 4. To do this, we find a nontrivial
solution of the system
2
3 2
32
3 2 3
x1
3
1
2
x1
0
3
2 5 4 y1 5 D 4 0 5 :
.A 4I/ 4 y1 5 D 4 1
1
2
2
2
1
0

10

3
1
All such solutions are multiples of 4 1 5 : Normalizing this to satisfy (18) yields
2
2
3
2 3
x1
1
1
X1 D p 4 y1 5 D 4 1 5 :
6
1
1

To find the points .x3 ; y3 ; 3 / where Q attains its constrained minimum, we first
find an eigenvector of A corresponding to 3 D 2. To do this, we find a nontrivial
solution of the system
2
3 2
32
3 2
3
x3
3
1
2
x3
0
.A C 2I/ 4 y3 5 D 4 1
3
2 5 4 y3 5 D 4 0 5 :
3
2
2
4
3
0
2
3
1
All such solutions are multiples of 4 1 5 : Normalizing this to satisfy (18) yields
1
2
3
2
3
x2
1
1
X3 D 4 y2 5 D p 4 1 5 :
3
2
1
As for the eigenvalue 2 D 2, we leave it you to verify that the only unit vectors
that satisfy AX2 D 2X2 are
2
3
1
1
X2 D p 4 1 5 :
2
1
For more on this subject, see Theorem 4.

Extrema subject to two constraints

Here is Theorem 1 with m D 2.

Theorem 3 Suppose that n > 2: If X0 is a local extreme point of f subject to g1 .X/ D
g2 .X/ D 0 and

@g1 .X0 / @g1 .X0 /

@xr
@xs

(19)

0
@g .X / @g .X /
2
0
2
0

@xr
@xs
for some r and s in f1; 2; : : : ; ng; then there are constants  and  such that
@f .X0 /
@xi

@g1 .X0 /
@xi

1  i  n.
11

@g2.X0 /
D 0;
@xi

(20)

Proof

For notational convenience, let r D 1 and s D 2. Denote

U D .x3 ; x4 ; : : : xn / and U0 D .x30 ; x30 ; : : : xn0 /:

Since

@g1 .X0 / @g1.X0 /

@x1
@x2

(21)

0;
@g .X / @g .X /
2
0
2
0

@x1
@x2
the Implicit Function Theorem (Theorem 6.4.1, p. 420) implies that there are unique
continuously differentiable functions
h1 D h1 .x3 ; x4; : : : ; xn / and h2 D h1 .x3 ; x4 ; : : : ; xn /;

defined on a neighborhood N  Rn 2 of U0 ; such that .h1 .U/; h2 .U/; U/ 2 D for all

U 2 N , h1 .U0 / D x10 , h2 .U0 / D x20 , and
g1 .h1 .U/; h2 .U/; U/ D g2 .h1 .U/; h2 .U/; U/ D 0;
From (21), the system
2 @g .X /
1
0
6
@x
1
6
6
4 @g .X /
2
0
@x1

U 2 N:

@g1 .X0 / 3
 

7
@x2
fx1 .X0 /
7 
D
7
fx2 .X0 /
@g .X / 5 
2

(22)

(23)

@x2

has a unique solution (Theorem 6.1.13, p. 373). This implies (20) with i D 1 and
i D 2. If 3  i  n, then differentiating (22) with respect to xi and recalling that
.h1 .U0 /; h2 .U0 /; U0 / D X0 yields
@g1 .X0 /
@g1 .X0 / @h1.U0 /
@g1.X0 / @h2 .U0 /
C
C
D0
@xi
@x1
@xi
@x2
@xi

and

@g2.X0 /
@g2 .X0 / @h1 .U0 /
@g2 .X0 / @h2.U0 /
C
C
D 0:
@xi
@x1
@xi
@x2
@xi
If X0 is a local extreme point of f subject to g1 .X/ D g2 .X/ D 0, then U0 is an
unconstrained local extreme point of f .h1 .U/; h2 .U/; U/; therefore,
@f .X0 /
@f .X0 / @h1 .U0 /
@f .X0 / @h2.U0 /
C
C
D 0:
@xi
@x1
@xi
@x2
@xi
The last three equations imply that

@f .X0 /
@f .X0 /

@x
@x1
i

@g .X / @g .X /

1
0
1
0

@xi
@x1

@g .X / @g .X /
2
0
2
0

@xi
@x1
12

@f .X0 /
@x2

@g1 .X0 /
D 0;

@x2

@g2 .X0 /

@x2


@f .X0 /

@xi

@f .X /

@x1

@f .X /
0

@x2

@g1 .X0 /
@xi
@g1 .X0 /
@x1
@g1 .X0 /
@x2

@g2 .X0 /
@xi

@g2 .X0 /

D 0:
@x1

@g2 .X0 /

@x2

Therefore, there are constants c1 , c2, c3 , not all zero, such that
2
3
@f .X0 / @g1.X0 / @g2 .X0 /
6 @xi
7
@xi
@xi
7
6
7
6
3 2
3
6 @f .X / @g .X / @g .X / 7 2
c1
0
6
0
1
0
2
0 7
74
6
7 c2 5 D 4 0 5 :
6 @x1
@x1
@x1
7
6
0
7 c3
6
6 @f .X / @g .X / @g .X / 7
1
0
2
0 7
0
6
5
4
@x2
@x2
@x2

If c1 D 0, then

2 @g .X /
1
0
6
@x1
6
6
4 @g .X /
2
0
@x1

@g1.X0 / 3
 

7
@x2
0
7 c2
D
;
7
0
@g .X / 5 c3
2

@x2

so (19) implies that c2 D c3 D 0; hence, we may assume that c1

solution of (24). Therefore,
2
3
@f .X0 / @g1.X0 / @g2.X0 /
7
6 @xi
@xi
@xi
7
6
7
6
3 2
6 @f .X / @g .X / @g .X / 7 2
1
7
6
1
0
2
0
0
74
6
6 @x1
7 c2 5 D 4
@x1
@x1
6
7
6
7 c3
6 @f .X / @g .X / @g .X / 7
0
1
0
2
0
6
7
4
5
@x2
@x2
@x2

D 1 in a nontrivial

3
0
0 5;
0

which implies that

@g1.X0 /
6
@x1
6
6
4 @g .X /
2
0
@x1
2

(24)

3
@g1 .X0 /
7
@x2
7
7
@g2 .X0 / 5
@x2
13

c2
c3

fx1 .X0 /
fx2 .X0 /

(25)

Since (23) has only one solution, this implies that c2

becomes
2
3
@f .X0 / @g1.X0 / @g2 .X0 /
6 @xi
7
@xi
@xi
6
7
6
7
6 @f .X / @g .X / @g .X / 7 2
6
1
0
2
0 7
0
6
74
6 @x1
7
@x1
@x1
6
7
6
7
6 @f .X / @g .X / @g .X / 7
0
1
0
2
0 7
6
4
5
@x2
@x2
@x2

 and c2 D

, so (25)

3 2
3
1
0
 5 D 4 0 5:

0

Computing the topmost entry of the vector on the left yields (20).
Example 7 Minimize
f .x; y; ; w/ D x 2 C y 2 C 2 C w 2
subject to
x C y C C w D 10 and x

y C C 3w D 6:

(26)

Solution Let
LD

x 2 C y 2 C 2 C w 2
2

.x C y C C w/

.x

y C C 3w/I

then
Lx
Ly
L
Lw

D x
D y

 
C

D  
D w  3;

so
x0 D  C ;

y0 D 

;

0 D  C ;

w0 D  C 3:

This and (26) imply that

. C / C . / C . C / C . C 3/
. C / . / C . C / C .3 C 9/
Therefore,
4 C 4 D 10

4 C 12 D

6;

so  D 3 and  D 1=2. Now (27) implies that



5 7 53
.x0 ; y0 ; 0 ; w0/ D
; ;
:
2 2 22
14

D 10
D
6:

(27)

Since f .x; y; ; w/ is the square of the distance from .x; y; ; w/ to the origin, it attains
a minimum value (but not a maximum value) subject to the constraints; hence the
constrained minimum value is


5 7 5 3
; ; ;
D 27:
f
2 2 2 2
Example 8 The distance between two curves in R2 is the minimum value of
p
.x1 x2 /2 C .y1 y2 /2 ;

where .x1 ; y1 / is on one curve and .x2 ; y2 / is on the other. Find the distance between
the ellipse
x 2 C 2y 2 D 1
and the line
x C y D 4:

(28)

Solution We must minimize

d 2 D .x1

x2 /2 C .y1

y2 /2

subject to
x12 C 2y12 D 1 and x2 C y2 D 4:
Let
LD

.x1

x2 /2 C .y1

y2 /2
2

.x12 C 2y12 /

.x2 C y2 /I

then
Lx1
Ly1
Lx2
Ly2

D x1
D y1
D x2
D y2

x2
y2
x1

x1
2y1


y1

;

so
x10

x20

y10
x20
y20

y20
x10
y10

D x10

(i)

D 2y10 (ii)
D 
(iii)
D :
(iv)

From (i) and (iii),  D x10 ; from (ii) and (iv),  D 2y10 . Since the curves do
2
2
not intersect,  0, so x10 D 2y10 . Since x10
C 2y10
D 1 and .x0 ; y0/ is in the first
quadrant,


2
1
.x10 ; y10 / D p ; p :
(29)
6
6
15

Now (iii), (iv), and (28) yield the simultaneous system

1
x20 y20 D x10 y10 D p ; x20 C y20 D 4;
6
so


1
1
.x20 ; y20/ D 2 C p ; 2
p :
2 6
2 6
From this and (29), the distance between the curves is
"


 #1=2

p
2 2
1
1 2
1
p
p
C 2
p
D 2 2
2C p
2 6
6
2 6
6


3
p :
2 6

Proof of Theorem 1

Proof

Denote

For notational convenience, let r` D `, 1  `  m, so (6) becomes

@g1 .X0 /
@g1 .X0 / @g1 .X0 /

@x1
@x2
@xm

@g .X / @g .X /
@g2 .X0 /
2
0
2
0

0
@x1
@x2
@xm

::
::
::
::

:
:
:
:

@g .X / @g .X /

@g
.X
/
m
0
m
0
m
0

@x1
@x2
@xm

(30)

U D .xmC1 ; xmC2 ; : : : xn / and U0 D .xmC1;0 ; xmC2;0 ; : : : xn0 /:

From (30), the Implicit Function Theorem implies that there are unique continuously
differentiable functions h` D h` .U/, 1  `  m, defined on a neighborhood N of U0 ,
such that
.h1 .U/; h2 .U/; : : : ; hm .U/; U/ 2 D; for all U 2 N;
.h1 .U0 /; h2 .U0 /; : : : ; hm .U0 /; U0 / D X0 ;
(31)

and

g` .h1 .U/; h2 .U/; : : : ; hm.U/; U/ D 0; U 2 N; 1  `  m:

Again from (30), the system
2
3
@g1 .X0 / @g1 .X0 /
@g1 .X0 /




6
7
@x1
@x2
@xm
6
72
3 2
6
7
1
fx1 .X0 /
6 @g .X / @g .X /
7
@g2 .X0 / 7 6
6
2
0
2
0
7
6





6
7 6 2 7 6 fx2 .X0 /
6
76 : 7 D 6
@x1
@x2
@xm
::
6
74 : 5 4
::
::
::
::
:
:
6
7
:
:
:
:
6
7

f
6 @g .X / @g .X /
7
m
x
m .X0 /
@gm .X0 / 7
m
0
m
0
6




4
5
@x1
@x2
@xm
16

(32)

3
7
7
7
5

(33)

has a unique solution. This implies that

@f .X0 /
@xi

1

@g1 .X0 /
@xi

2

@g2 .X0 /
@xi

m

@gm .X0 /
D0
@xi

(34)

for 1  i  m.
If m C 1  i  n, differentiating (32) with respect to xi and recalling (31) yields
m

@g`.X0 / X @g`.X0 / @hj .X0 /

C
D 0;
@xi
@xj
@xi
j D1

1  `  m:

If X0 is local extreme point f subject to g1 .X/ D g2 .X/ D


D gm .X/ D 0, then U0

is an unconstrained local extreme point of f .h1 .U/; h2 .U/; : : : hm .U/; U/; therefore,
m

@f .X0 / X @f .X0 / @hj .X0 /

C
D 0:
@xi
@xj
@xi
j D1

The last two equations imply that

@f .X0 /
@f .X0 /

@x
@x1
i

@g1 .X0 / @g1 .X0 /

@xi
@x1

@g .X / @g .X /
2
0
2
0

@xi
@x1

::
::

:
:

@gm .X0 / @gm .X0 /

@xi
@x1

so

@f .X0 /
@x2

@g1 .X0 /
@x2

@g2 .X0 /
@x2
::
:
@gm .X0 /
@x2

::

@f .X0 /
@xi

@g1.X0 /
@xi

@g2.X0 /
@xi

:::

@f .X0 /
@x1

@g1.X0 /
@x1

@g2.X0 /
@x1

:::

@f .X0 /
@x2
::
:
@f .X0 /
@xm

@g1.X0 /
@x2
::
:
@g1.X0 /
@xm

@g2.X0 /
@x2
::
:
@g2.X0 /
@xm

17

:::
::

:::

@f .X0 /
@xm

@g1 .X0 /

@xm

@g2 .X0 / D 0;

@xm

::

@gm .X0 /

@xm

@gm .X0 /
@xi

@gm .X0 /

@x1

D 0:
@gm .X0 /

@x2

::

@gm .X0 /

@xm

Therefore, there are constant c0 , c1 , . . . cm , not all zero, such that

2
3
@f .X0 / @g1 .X0 / @g2 .X0 /
@gm .X0 /
:
:
:
6 @xi
7
@xi
@xi
@xi
6
7
6
72
6
7
c0
@gm .X0 / 7
6 @f .X0 / @g1 .X0 / @g2 .X0 /
6
76
:::
c1
6 @x1
7
@x
@x
@x
1
1
1
6
76
6
76
6 c3
6 @f .X / @g .X / @g .X /
6 :
@gm .X0 / 7
1
0
2
0
0
6
7
:::
6
7 4 ::
@x2
@x2
@x2
6 @x2
7
6
7 cm
::
::
::
::
::
6
7
:
:
:
:
:
6
7
4 @f .X0 / @g1 .X0 / @g2 .X0 /
@gm .X0 / 5
:::
@xm
@xm
@xm
@xm
If c0 D 0, then
2
@g1 .X0 /
6
@x1
6
6
6 @g .X /
6
2
0
6
6
@x1
6
::
6
:
6
6 @g .X /
m
0
6
4
@x1

@g1 .X0 /
@x2
@g2 .X0 /
@x2
::
:
@gm .X0 /
@x2

::

3
@g1 .X0 /
7
@xm
72
7
@g2 .X0 / 7
76
76
76
@xm
74
::
7
:
7
@gm .X0 / 7
7
5
@xm

and (30) implies that c1 D c2 D


D cm D

a nontrivial solution of (35). Therefore,
2
@f .X0 / @g1 .X0 / @g2 .X0 /
:::
6 @xi
@xi
@xi
6
6
6
6 @f .X0 / @g1 .X0 / @g2 .X0 /
6
:::
6 @x1
@x1
@x1
6
6
6 @f .X / @g .X / @g .X /
0
1
0
2
0
6
:::
6
6 @x2
@x2
@x2
6
::
::
::
::
6
:
:
:
:
6
4 @f .X0 / @g1 .X0 / @g2 .X0 /
:::
@xm
@xm
@xm

18

7 6
7 6
7 6
7D6
7 6
5 4

c1
c2
::
:
cm

7 6
7 6
7D6
5 4

0
0
::
:
0

0
0
0
::
:
0

7
7
7
7 : (35)
7
5

3
7
7
7
5

0; hence, we may assume that c0 D 1 in

3
@gm .X0 /
7
@xi
7
72
7
@gm .X0 / 7
76
76
@x1
76
76
@gm .X0 / 7
76
74
@x2
7
7
::
7
:
7
@gm .X0 / 5
@xm

1
c1
c2
::
:
cm

7 6
7 6
7 6
7D6
7 6
5 4

0
0
0
::
:
0

7
7
7
7 ; (36)
7
5

which implies that

2
@g1 .X0 / @g1 .X0 /
6
@x1
@x2
6
6
6 @g .X / @g .X /
6
2
0
2
0
6
6
@x1
@x2
6
::
::
6
:
:
6
6 @g .X / @g .X /
m
0
m
0
6
4
@x1
@x2

::

3
@g1 .X0 /
7
@xm
72
7
@g2 .X0 / 7
76
76
76
@xm
74
::
7
:
7
@gm .X0 / 7
7
5
@xm

Since (33) has only one solution, this

becomes
2
@f .X0 / @g1 .X0 / @g2 .X0 /
6 @xi
@xi
@xi
6
6
6
6 @f .X0 / @g1 .X0 / @g2 .X0 /
6
6 @x1
@x1
@x1
6
6
6 @f .X / @g .X / @g .X /
0
1
0
2
0
6
6
6 @x2
@x2
@x2
6
::
::
::
6
:
:
:
6
4 @f .X0 / @g1 .X0 / @g2 .X0 /
@xm
@xm
@xm

c1
c2
::
:
cm

implies that cj D
:::

:::

:::
::

:::

7 6
7 6
7D6
5 4

fx1 .X0 /
fx2 .X0 /
::
:
fxm .X0 /

3
7
7
7
5

j , 1  j  n, so (36)

3
@gm.X0 /
7
@xi
7
72
7
@gm.X0 / 7
76
76
@x1
76
76
@gm.X0 / 7
76
74
@x2
7
7
::
7
:
7
@gm.X0 / 5

1
1
2
::
:
m

7 6
7 6
7 6
7D6
7 6
5 4

0
0
0
::
:
0

7
7
7
7:
7
5

@xm

Computing the topmost entry of the vector on the left yields yields (34), which completes the proof.
Example 9 Minimize

n
X

xi2 subject to

i D1

n
X
i D1

where

ar i xi D cr ;

n
X
i D1

Solution

ar i asi D

Let

1X 2
LD
xi
2
i D1

Then

Lxi D xi

m
X

1  r  m;
1 if r D s;
0 if r s:

m
X

s

sD1

s asi ;

sD1

19

n
X

asi xi :

i D1

1  i  n;

(37)

(38)

so
xi 0 D
and

m
X

s asi

ar i xi 0 D
Now (38) implies that
n
X
i D1

1  i  n;

sD1

ar i xi 0 D

m
X

(39)

s ar i asi :

sD1

m
X

s

sD1

n
X
i D1

ar i asi D r :

From this and (37), r D cr , 1  r  m, and (39) implies that

xi 0 D
Therefore,
xi20 D
and (38) implies that
n
X

xi20

i D1

m
X

cs asi ;

1  i  n:

sD1

m
X

cr cs ar i asi ;

r;sD1

m
X

cr cs

r;sD1

n
X
i D1

1  i  n;

ar i asi D

m
X

cr2 :

r D1

The next theorem provides further information on the relationship between the
eigenvalues of a symmetric matrix and constrained extrema of its quadratic form. It
can be proved by successive applications of Theorem 1; however, we omit the proof.
Theorem 4 Suppose that A D ar s nr;sD1 2 Rnn is symmetric and let
Q.x/ D

n
X

ar s xr xs :

r;sD1

Suppose also that

2

minimizes Q subject to

Pn

i D1

6
6
x1 D 6
4

x11
x21
::
:
xn1

3
7
7
7
5

xi2 . For 2  r  n, suppose that

2
3
x1r
6 x2r 7
6
7
xr D 6 : 7 ;
4 :: 5
xnr
20

minimizes Q subject to
n
X
i D1

xi2 D 1 and

Denote
r D

n
X

n
X
i D1

xi s xi D 0;

aij xi r xj r ;

i;j D1

1sr

1:

1  r  n:

Then
1  2 


 n and Axr D r xr ;

21

1  r  n:

Exercises
1.

Find the point on the plane 2x C 3y C D 7 closest to .1; 2; 3/.

2.

Find the extreme values of f .x; y/ D 2x C y subject to x 2 C y 2 D 5.

3.

Suppose that a; b > 0 and a 2 C b 2 D 1. Find the extreme values of

f .x; y/ D x C y subject to ax 2 C by 2 D 1.

4.

Find the points on the circle x 2 C y 2 D 320 closest to and farthest from .2; 4/.

5.

Find the extreme values of

f .x; y; / D 2x C 3y C

subject to x 2 C 2y 2 C 32 D 1:

6.

Find the maximum value of f .x; y/ D xy on the line ax C by D 1, where

a; b > 0.

7.

A rectangle has perimeter p. Find its largest possible area.

8.

A rectangle has area A. Find its smallest possible perimeter.

9.

A closed rectangular box has surface area A. Find it largest possible volume.

10.

The sides and bottom of a rectangular box have total area A. Find its largest
possible volume.

11.

A rectangular box with no top has volume V . Find its smallest possible surface
area.

12.

Maximize f .x; y; / D xy subject to

x
y

C C D 1;
a
b
c
where a, b, c > 0.

13.

Two vertices of a triangle are . a; 0/ and .a; 0/, and the third is on the ellipse
x2
y2
C
D 1:
a2
b2
Find its largest possible area.

14.

Show that the triangle with the greatest possible area for a given perimeter is
equilateral, given that the area of a triangle with sides x, y, and perimeter s is
p
A D s.s x/.s y/.s /:
22

15.

A box with sides parallel to the coordinate planes has its vertices on the ellipsoid
x2
y2
2
C 2 C 2 D 1:
2
a
b
c
Find its largest possible volume.

16.

Derive a formula for the distance from .x1 ; y1; 1 / to the plane
ax C by C c D :

17.

Let Xi D .xi ; yi ; i /, 1  i  n. Find the point in the plane

ax C by C c D 
for which
plane.

18.
19.

Pn

i D1

jX

Xi j2 is a minimum. Assume that none of the Xi are in the

Find the extreme values of f .X/ D

n
X

ci /2 subject to

i D1

n
X
i D1

xi2 D 1.

Find the extreme values of

f .x; y; / D 2xy C 2x C 2y

20.

.xi

subject to x 2 C y 2 C 2 D 1:

Find the extreme values of

f .x; y; / D 3x 2 C 2y 2 C 32 C 2x subject to x 2 C y 2 C 2 D 1:

21.

Find the extreme values of

f .x; y/ D x 2 C 8xy C 4y 2

subject to x 2 C 2xy C 4y 2 D 1:

22.

Find the extreme value of f .x; y/ D C xy subject to .ax C by/2 D 1.

Assume that ab 0.

23.

Find the extreme values of f .x; y; / D x C y 2 C 2 subject to

4x 2 C 9y 2

24.

362 D 36:

Find the extreme values of f .x; y; ; w/ D .x C /.y C w/ subject to

x 2 C y 2 C 2 C w 2 D 1:

25.

Find the extreme values of f .x; y; ; w/ D .x C /.y C w/ subject to

x 2 C y 2 D 1 and 2 C w 2 D 1:
23

26.

Find the extreme values of f .x; y; ; w/ D .x C /.y C w/ subject to

x 2 C 2 D 1 and y 2 C w 2 D 1:

27.

Find the distance between the circle x 2 C y 2 D 1 the hyperbola xy D 1.

28.

Minimize f .x; y; x/ D

x2
y2
2
C
C
subject to ax C by C c D d and x,
2
2
2

y, > 0.
29.

Find the distance from .c1 ; c2; : : : ; cn / to the plane

a1 x1 C a2 x2 C


C an xn D d:

30.

Find the maximum value of f .X/ D

n
X

ai xi2 subject to

i D1

i D1

p; q > 0 and ai , bi xi > 0, 1  i  n.

31.

Find the extreme value of f .X/ D

q>0 and ai , bi , xi > 0, 1  i  n.

32.

n
X

p
ai xi

n
X

subject to

n
X
i D1

i D1

bi xi4 D 1, where

bi xi D 1, where p,

Find the minimum value of

f .x; y; ; w/ D x 2 C 2y 2 C 2 C w 2
subject to
x C y C C 3w
x C y C 2 C w

33.

D
D

1
2:

Find the minimum value of

f .x; y; / D

x2
y2
2
C 2 C 2
2
a
b
c

subject to p1 x C p2 y C p3 D d , assuming that at least one of p1 , p2 , p3 is

nonzero.
34.

Find the extreme values of f .x; y; / D p1 x C p2 y C p3 subject to

x2
y2
2
C
C
D 1;
a2
b2
c2
assuming that at least one of p1 , p2 , p3 is nonzero.

35.

Find the distance from . 1; 2; 3/ to the intersection of the planes

x C 2y 3 D 4 and 2x y C 2 D 5.
24

36.

Find the extreme values of f .x; y; / D 2x C y C 2 subject to x 2 C y 2 D 4

and x C D 2.

37.

Find the distance between the parabola y D 1 C x 2 and the line x C y D

38.

Find the distance between the ellipsoid

1.

3x 2 C 9y 2 C 62 D 10
and the plane
3x C 3y C 6 D 70:
39.

Show that the extreme values of f .x; y; / D xy C y C x subject to

y2
2
x2
C 2 C 2 D1
2
a
b
c
are the largest and smallest eigenvalues of the matrix
2
3
0 a2 a2
4 b2 0 b2 5 :
c2 c2 0

40.

Show that the extreme values of f .x; y; / D xy C 2y C 2x subject to

x2
y2
2
C 2 C 2 D1
2
a
b
c
are the largest and smallest eigenvalues of the matrix
2
3
0
a2 =2 a2
4 b 2=2
0
b2 5 :
2
2
c
c
0

41.

Find the extreme values of x.y C / subject to

x2
y2
2
C 2 C 2 D 1:
2
a
b
c

42.

Let a, b, c, p, q, r , , , and be positive constants. Find the maximum value

of f .x; y; / D x y subject to
ax p C by q C cr D 1 and x; y; > 0:

43.

Find the extreme values of

f .x; y; ; w/ D xw

subject to x 2 C 2y 2 D 4 and
25

22 C w 2 D 9:

44.

Let a, b, c,and d be positive. Find the extreme values of

f .x; y; ; w/ D xw

subject to
ax 2 C by 2 D 1;

c2 C dw 2 D 1;

if (a) ad bc; (b) ad D bc:

45.

Minimize f .x; y; / D x 2 C y 2 C 2 subject to

a1 x C a2 y C a3 D c and b1 x C b2 y C b3 D d:
Assume that
a12 C a22 C a32 0; and b12 C b22 C b32 0:

; ; > 0;

Formulate and apply a required additional assumption.

46.

Minimize f .X; Y/ D

n
X
i D1

n
X
i D1

where

n
X
i D1

47.

i /2 subject to

.xi

ai xi D c and

ai2 D

n
X
i D1

n
X
i D1

bi x i D d ;
n
X

bi2 D 1 and

i D1

ai bi D 0:

Find .x10;x20 ; : : : ; xn0 / to minimize

Q.X/ D

n
X

xi2

i D1

subject to
n
X
i D1

Prove explicitly that if

xi D 1 and

n
X

j D1

yi D 1;

n
X
i D1

n
X
i D1

ixi D 0:

iyi D 0

and yi xi 0 for some i 2 f1; 2; : : : ; ng, then

n
X

yi2 >

i D1

n
X
i D1

26

xi20:

48.

Let p1 , p2 , . . . , pn and s be positive numbers. Maximize

x1 /p1 .s

f .X/ D .s

x2/p2


.s

xn /pn

subject to x1 C x2 C


C xn D s.
49.

Maximize f .X/ D x1p1 x2p2


xnpn subject to xi > 0, 1  i  n, and

n
X
xi

i

i D1

D S;

where p1 , p2 ,. . . , pn , 1 , 2 , . . . , n , and V are given positive numbers.

50.

Maximize
f .X/ D
subject to xi > 0, 1  i  n, and
p

n
X
xi
i D1

i

x1 1 x2 2


xnpn D V;
where p1 , p2 ,. . . , pn , 1 , 2 , . . . , n , and S are given positive numbers.
51.

Suppose that 1, 2 , . . . n are positive and at least one of a1 , a2 , . . . , an is

nonzero. Let .c1 ; c2 ; : : : ; cn/ be given. Minimize
Q.X/ D

n
X
.xi
i D1

ci /2
i

subject to
a1 x1 C a2 x2 C


C an xn D d:
52.

Schwarzs inequality says that .a1 ; a2 ; : : : ; an / and .x1 ; x2 ; : : : ; xn/ are arbitrary n-tuples of real numbers, then
ja1 x1 C a2 x2 C


C an xn j  .a12 C a22 C


C an2 /1=2 .x12 C x22 C


C xn2/1=2 :
Prove this by finding the extreme values of f .X/ D

53.

n
X
i D1

ai xi subject to

n
X
i D1

xi2 D  2 .

Let x1 , x2 , . . . , xm , r1 , r2 , . . . , rm be positive and

r1 C r2 C


C rm D r:
Show that

r1 x 1 C r2 x 2 C


rm x m
;
r
and give necessary and P
sufficient conditions for equality. (Hint: Maximize
rm
x1r1 x2r2


xm
subject to m
j D1 rj xj D  > 0, x1 > 0, x2 > 0, . . . , xm > 0.)
rm
x1r1 x2r2


xm

1=r

27

54.

Let A D aij be an m  n matrix. Suppose that p1 , p2 , . . . , pm > 0 and

m
X
1
D 1;
pj

j D1

and define
i D

n
X

j D1

jaij jpi ;

1  i  m:

Use Exercise 53 to show that

1=p1 1=p2
1=pm

a
a






a
:
2



m
ij 2j
mj  1

j D1

(With m D 2 this is Hlders inequality, which reduces to Schwarzs inequality

if p1 D p2 D 2.)

55.

Let c0 , c1 , . . . , cm be given constants and n  m C 1. Show that the minimum

value of
n
X
Q.X/ D
xr2
r D0

subject to

n
X

r D0

is attained when

xr r s D c s ;

xr D
where

m
X
`D0

m
X

0  s  m;

s r s ;

0  r  n;

sD0

sC` ` D cs and s D

n
X

r s;

r D0

0  s  m:

Show that if fxr gnrD0 satisfies the constraints and xr xr 0 for some r , then
n
X

xr2 >

r D0

56.

n
X

xr20:

r D0

Suppose that n > 2k. Show that the minimum value of f .W/ D
subject to the constraint
n
X

wi P .r

iD n

28

i / D P .r /

n
X

iD n

wi2,

whenever r is an integer and P is a polynomial of degree  2k, is attained with

wi 0 D

2k
X

r i r ;

1  i  n;

r D0

where
2k
X

r D0

r r Cs D

Show that if fwi gniD

n
X
1 if s D 0;
and s D
j s:
0 if 1  s  2k;
jD n

satisfies the constraint and wi wi 0 for some i , then

n
X

iD n

57.

n
X

wi2 >

wi20:

iD n

Suppose that n  k. Show that the minimum value of f

constraint

n
X

wi P .r

n
X

wi2, subject to the

i D0

i / D P .r C 1/

i D0

whenever r is an integer and P is a polynomial of degree  k, is attained with

wi 0 D

k
X

r i r ;

0  i  n;

r D0

where
k
X

r D0

r Cs r D . 1/s ;

Show that if

0  s  k;
n
X

ui P .r

and ` D

n
X
i D0

i `;

0  `  2k:

i / D P .r C 1/

i D0

whenever r is an integer and P is a polynomial of degree  k, and ui wi 0

for some i , then
n
n
X
X
u2i >
wi20 :
i D0

58.

i D0

Minimize
f .X/ D
subject to
n
X
i D1

n
X
.xi
i D1

ai r xi D dr ;
29

ci /2
i

1r m

Assume that m > 1, 1 , 2 , . . . m > 0, and

(
n
X
1 if r D s;
i ai r ai s D
0 if r s:
i D1

30

8
1.

Answers to selected exercises

15
7

2 25
;
7 7

p
p
3. 1= ab, 1= ab

2. 5

p
4. .8; 16/ is closest, . 8; 16/ is farthest. 5. 53=6 6. 1=4ab 7. p 2 =4
p
p
p
8. 4 A 9. A3=2 =6 6 10. A3=2=6 3 11. 3.2V /2=3 12. abc=27
p
13. ab 15. 8abc=3 3

18. .1

/2 and .1 C /2 , where  D @

n
X

j D1

26. 2
jd
n
X

31.

i D1

11=2

cj2 A

19.

p
22. jj=4jabj 23.
5, 73=16
2
p
d
27. 2 1 28.
.a/2 C .b 2 / C .c /2

2=3, 2

21.

29.

a1 c1 a2 c2


an cn /ai j
q
a12 C a22 C


an2
!1 p=q
aiq=.q

p/ p=.p q/
bi

30.

24. 1

n
X
a2
i

i D1

1, 2

bi

20. 2, 4
25. 2

!1=2

is a constrained maximum if p < q, a constrained

minimum if p > q
d2
34. .p12 a2 C p22 b 2 C p32 c 2 /1=2
p12 a2 C p22 b 2 C p32 c 2
p
p
p
p
35.
693=45 36. 2, 6 37. 7=4 2 38. 10 6=3 41. jcj a2 C b 2=2


p
p
p

3

42.
C C
43. 3 44. (a) 1= bc (b) 1= ad D 1= bc
pqr p
q
r
!2
!2
n
n
X
X
46. c
ai i C d
bi i
47. xi 0 D .4n C 2 6i /=n.n 1/
32. 689=845

33.

i D1

.n 1/s
P

iP

i D1

p1 1 p2 2


pn n

p1 Cp2 C


Cpn
S
49.
.p1 1 /p1 .p2 2 /p2


.pn n /pn
p1 C p2 C


C pn

 p Cp 1C


Cp
n
1
2
V
50. .p1 C p2 C


C pn /
p
p
p
n
1
2
.1 p1 / .2 p2 /


.n pn /
!
!
!1=2 n
!1=2
n
n
n
X
X
X
X
2
2
2
2
51. d
ai ci / =
ai i
52.
ai
xi 0
48.

i D1

58.

m
X

r D1

dr

i D1

n
X
i D1

ai r ci

i D1

!2

31

i D1