page 1 of Section 9.

CHAPTER 9 QUADRATIC FORMS

SECTION 9.1 THE MATRIX OF A QUADRATIC FORM

quadratic forms and their matrix notation
If q = a1 x2 + a2 y2 + a3 z2 + a4 xy + a5 xz + a6 yz
then q is called a quadratic form (in variables x,y,z). There is a q value (a
scalar) at every point. (To a physicist, q is probably the energy of a system with
ingredients x,y,z.)
The matrix for q is
 a1
1
a
1
a 
 2 4 2 5

A = 1
a
2 4
a2
1
a
2 6 
1 1 
 a
2 5 2 6
a a3 
It's the symmetric matrix A with this connection to q:
 x 
(1) q = [ x y z ]
A y
 z 
or equivalently
Û Û Û  x 
(2) q = x T A x where x =  y 
 z 
example 1
2 2 2 2 1
If q = 3x1 + 6x2 - 7x3 + πx 4 - √ 2 x1 x2 + 8x1 x3 + πx3 x4 + 10x2 x4 - 2 x1 x4
then the matrix for q is
 3 -
√ 2/2 4 -1/4 
=  -√ 2/2 6 0 5 
A
 4 
 -1/4 0 -7 π/2

5 π/2 π

warning
In example 1, do not write q = A. The quadratic form does not equal a matrix (q is
 x 
ÛT Û Û  1
a scalar quantity, not a matrix). What q does equal is x Ax where x =  » 
 x4 
 
example 2
 2 3 5 
If q has matrix  3 7 -2  then q = 2x2 + 7y2 + 12 z2 + 6xy + 10xz - 4yz.
 5 -2 1/2 
 page 2 of Section 9.1

basis changing rule for the matrix of a quadratic form

Suppose q is a quadratic form in variables x,y,z with (old) matrix A. Let

u = u1i + u2j + u3k

v = v1i + v2j + v3k
w = w1i + w2j + w3k
be a new basis for R3. Let P be the usual basis—changing matrix:

 u1 v1 w1 
 
P =  u2 v2 w2 

 u v3 w3 
 3 
Then the new matrix for q in the new coord system with variables X,Y,Z and
basis vectors u,v,w is given by

new = PT old P

example 3
Suppose
(3) q = x2 + 6xy + y2.

Let
u = i + 2j
v = -i + 2j

Find the matrix for q w.r.t. basis u,v and express q in terms of new coordinates X
and Y.
solution method 1 (using an algebraic substitution)
Let
 1 -1 
P = 
 2 2 

The old coordinates x,y and new coordinates X,Y are related by

[ xy ] = P [ XY ]
so
x = X - Y
y = 2X + 2Y

Substitute for x and y in (3):

q = (X-Y)2 + 6(X-Y)(2X+2Y) + (2X+2Y)2.

Multiply out and collect terms to get q in terms of X and Y:

q = 17X2 + 6XY - 7Y2.

 17 3 
The matrix for q w.r.t. basis u,v is   .
 3 -7

method 2 (using the basis changing rule for the matrix of a quadratic form)
The matrix for q (w.r.t. the standard coord system) is A = [ 13 3
1 ].
By the basis changing rule above, the new matrix for q is
 page 3 of Section 9.1

 1 2   1 3   1 -1   17 3 
PT AP =    3   2  =  3 
 -1 2 1 2 -7

Use the matrix to express q in terms of X and Y:

q = 17X2 + 6XY - 7Y2

proof of the basis changing rule

 x   x  T  x 
q = [x y z] A  yz  = y
 z  A y
 z 

  X  T X
=  P  Y   A P  Y  by the basis changing rule

  Z   Z  for coords of a vector
 X 
= [ X Y Z ]
PTAP  Y 
 Z  by T rule

So PTAP satisfies (1) but using new coordinates.

Furthermore PTAP is symmetric because

(PTAP)T = PTATPTT = PTAP (since AT = A)

So PTAP is the matrix of q w.r.t. the new basis.

warning
1. The matrix of a quadratic form must be symmetric.
2. If A is symmetric then PTAP is also symmetric.
If yours isn't, check your arithmetic.

congruent matrices (versus similar matrices)

If A is symmetric and P is invertible (so that its cols are independent and can
serve as basis vectors) then PTAP and A are called congruent.
The matrix PTAP represents the same quadratic form as A, but w.r.t. a new basis
consisting of the cols of P.

Remember from Section 6.2 that P-1 AP and A are called similar (whether or not A is
symmetric).
The matrix P-1AP represents the same linear transformation as A but w.r.t. a new
basis consisting of the cols of P.

If the matrix P is orthogonal, which happens when the new basis is orthonormal,
then PT = P-1 and congruence is the same as similarity.

mathematical catechism
question 1 What is a quadratic form.
answer A quadratic form say in variables x1, x2, x3, x4 is an expression of the form
2 2 2 2
a1x1 + a2x2 + a3x3 + a4x4 + cross terms like b1x1x2, b2x1x3 etc.
question 2 What is the matrix of a quadratic form.
answer The matrix of a quadratic form in variables x1,..., xn is the symmetric
 page 4 of Section 9.1

 x1 
 
matrix A such that q = [x1 ... xn] A  » 

 xn 
 

question 3 What are congruent matrices.

answer Matrices A and B are congruent if A is symmetric and there exists an
invertible matrix P such that B = PTAP (this automatically makes B symmetric too).

question 4 What are similar matrices.

answer Matrices A and B are similar if there exists a matrix P such that
B = P-1AP.

PROBLEMS FOR SECTION 9.1

1. Let q = x2 + 3y2 + 8z2 - 3xy - 4yz.

Find the matrix A for q and write q in terms of A using matrix notation.

 2 3 4 
2. Write out the quadratic form which has matrix  3 6 7 .
 4 7 9 
 page 5 of Section 9.1

2
3. Suppose q = x2 + 3x4 x5. Find the matrix for q.
 1 2 3 4 
 2 5 

4. Find the quadratic form with matrix  3 0
0 0 .
0 -3 
 
 4 0 -3 6 
5. Let
q = x2 + 4xy + 3y2
Û Û Û
u = 3i + j
Û Û Û
v = 2i - j.

(a) Find the new formula for q w.r.t. the basis u,v using the basis changing rule
for a quadratic form.
(b) Find the new formula for q w.r.t. the basis u,v again using an algebraic
substitution.
(c) Suppose a point has coords X=1, Y=2 w.r.t. basis u,v.
Find the value of q at the point two ways, using its X,Y coordinates and then
again using its x,y coordinates.

6. Start with q = x2 + 3xy - 5y2 and make the change of variable

X = x - y
Y = x + y

(a) Find q in terms of X and Y just with algebra.

(b) What new basis is involved when you use variables X and Y.
(c) Find q in terms of X and Y again using the basis changing rule for q.
(d) Find the (old) matrix for q. What is the connection between q and the old matrix
(write an equation beginning "q = ").
(e) Find the new matrix for q. What is the connection between q and the new matrix
(write an equation beginning "q = ").
(f) What's the connection between the old matrix for q and the new matrix for q.

7. Start with q = x2 + 4xy - y2 and make the change of variable

x = 2X - Y
y = X + 3Y

(a) Find q in terms of X and Y just with algebra.

(b) What new basis is involved when you use coordinates X and Y.
(c) Find q in terms of X and Y again using the basis changing rule for q.

8. Suppose the X—axis is the same as the x—axis and the Y—axis is found by rotating
the y—axis clockwise 45o. Use the basis changing rule for a quadratic form to find
(a) the new formula for q = x2 + y2.
(b) the old formula for q = XY.
(c) the equation of the circle x2 + y2 = 1 in the new system.

9. In the usual x,y coord system, q is 2x2 + 3xy + 4y2.

Switch to a new X,Y coord system which has the same axes as before but new scales.
If the old scale was the inch, on the new X-axis use the foot and on the new Y-axis
use a half—inch. Find q in the new coord system.
 page 1 of Section 9.2

SECTION 9.2 DIAGONALIZING A QUADRATIC FORM

diagonalizing q
Start with a quadratic form q, in say 3 variables, with matrix A.
Diagonalizing q means finding a new X,Y,Z coord system in which the formula for q
has no cross terms, i.e., is of the form aX2 + bY2 + cZ2.
Equivalently, diagonalizing q means finding an invertible matrix P so that the
T
P AP, the new matrix for q, is diagonal.

first method for diagonalizing q: using eigenvalues

Suppose q is a quadratic form in variables x,y,z with matrix A. Since A is

symmetric it has a complete set of orthonormal eigenvectors with corresponding real
eigenvalues ¬1, ¬2, ¬3. Then w.r.t. a basis for R3 of orthonormal eigenvectors of
A,
q = ¬1 X2 + ¬2 Y2 + ¬3 Z2

In other words, the new matrix for q w.r.t. the new basis is

 ¬1 0 0 
 
Ò =  0 ¬2 0 

 0 0 ¬3 
 

proof
Let u,v,w be the orthonormal eigenvectors and let P be the matrix with cols u,v,w.
Then
new matrix for q = PT AP basis changing rule for q
-1
= P AP P is an orthog matrix since its cols are orthonormal
= Ò by (1) in Section 8.3

example 1
Let
(1) q = 5x2 + 8y2 + 5z2 - 4xy + 8xz + 4yz

The matrix for q is

 5 -2 4 
A =  -2 8 2 
 4 2 5 

It turns out that

¬=0 is an eigenvalue of A with corresponding eigenvector u = (2,1,-2);
¬=9 is a 2—fold eigenvalue with corresponding eigenvectors v = (0,2,1), w = (1,-2,0).
The Gram Schmidt process on v,w produces
2 4
v1 = (0,2,1), w1 = (1, - 5 , 5 )
So far, u, v1, w1 are orthog eigenvectors. Normalize them to get
2 1 2 2 1 1 2 4
r = (3 , 3 , - 3 ), s = (0, , ), t = ( , - , )
5 √
√ 5 45
√ 45 √
√ 45
In the new coord system with basis vectors r,s,t,

(2) q = 9Y2 + 9Z2.

The substitution that turns (1) into (2) is

 page 2 of Section 9.2

 x   X 
 y = P  Y
 z   Z


where P = [r s ]
t

2 1
x = 3 X + Z
√45
1 2 2
y = 3 x + Y - Z
5
√ 45
√
2 1 4
z = - 3 X + Y + Z
5
√ 45
√
As a check, if you substitute for x,y,z in (1) you should get the new version of q
in (2):
2 1 1 2 2 2 1 4
q = 5( 3 X + Z)2 + 8( 3 x + Y - Z)2 + 5(- 3 X + Y + Z)2
45
√ 45
√ 45
√ 5
√ 45
√
2 1 1 2 2
- 4( 3 X + Z)( 3 x + Y - Z)
45
√ 45
√ 45
√
2 1 2 1 4
+ 8( 3 X + Z)(- 3 X + Y + Z)
√45 5
√ 45
√
1 2 2 2 1 4
+ 4( 3 X + Y - Z)(- 3 X + Y + Z)
45
√ 45
√ 5
√ 45
√
= 9Y2 + 9Z2

warning
You need orthonormal eigenvectors to get q to be ¬1X2 + ¬2Y2 + ¬3Z2. In example 1,
if you use the original eigenvectors u,v,w (v and w are not orthog) as a basis, you
won't get a diagonal q. And if you use eigenectors u, v1, w1 (orthogonal but not
normalized) you will get a diagonal q but not ¬1X2 + ¬2Y2 + ¬3Z2.

example 2
Diagonalize q = 3x2 + 3y2 + 2xy.
Find the new basis vectors and the change of variable that produces the diagonal q.
solution The matrix for q is
A =
3 1
1 3 [ ]
The eigenvalues are 2,4 with respective eigenvectors r = (1,-1), s = (1,1). They are
orthog since A is symmetric. Normalize them to get

1 1 1 1
u = ( , - ), v = ( , ).
2
√ √ 2  2 √ 2
√
In the new coord system with basis u,v,

q = 2X2 + 4Y2.

The change of variable is

[ xy ]= P[ XY ] where P = [u
1 1
]
v

x = X + Y
2
√ 2
√
1 1
y = - X + Y
2
√ 2
√
warning
Don't forget to use orthonormal eigenvectors if you claim that q = ¬1X2 + ¬2Y2.
Remember to normalize!
 page 3 of Section 9.2

example 2 continued
Identify the graph of
3x2 + 3y2 + 2xy = 2

(temporarily obscured because of the cross term 2xy).

solution In the new X,Y cord system with basis u,v, the equation is

2X2 + 4Y2 = 2

so the graph is an ellipse (Fig 1).

y-axis
Y-axis
((-1,0))•
• ((0, 1
2
))
x-axis
1
((0, - 2
)) •

• ((1,0))

X-axis
FIG 1

warning
Example 2 involved the quadratic form q = 3x2 + 3y2 + 2xy. The ellipse in Fig 1 is
not the graph of q. It's the graph of the equation q = 2 (the ellipse is actually one
of the level sets of q).

row/col operations on a matrix

A row/col operation is a pair of matching row and col ops such as

R6 = 3R1 + R6
C6 = 3C1 + C6

Note that row/col operations preserve symmetry; i.e., if you row/col operate on a
symmetric matrix, the result is also symmetric.

row/col operation rule

Let A be the matrix of a quadratic form q w.r.t. the standard basis i,j,k.
Do some row/col ops on A to get matrix B.
Do just the col ops to I to get a new matrix P.
Then B is the matrix of q w.r.t. a new basis consisting of the cols of P.

In other words, B = PTAP.

proof
Each col op corresponds to right multiplication by an elementary matrix (±1.3).
Each row op corresponds to left mult by an elem matrix.
When the operations are matching, the elem matrices turn out to be transposes of
one another.
So row/col operations on A (say there are three of them) and the col ops on I
amount to the scheme in Fig 2 where
 page 4 of Section 9.2

I turns into P = E1E2E3

T T T
A turns into B = E3E2E1 AE1E2E3 = (E1E2E3)T A (E1E2E3) = PTAP
B is of the form PTAP where P is invertible (elementary matrices are invertible
and the product of invertibles is invertible), so B is the matrix for q in the new
coord system whose basis vectors are the cols of P. QED
A I

row/col ops col ops

T T T
B = E3E2E1 AE1E2E3 P = IE1E2E3 = E1E2E3
FIG 2

second method for diagonalizing q: using row/col ops

Suppose A is the matrix of a quadratic form q (say in variables x,y,z)

Then you can do row/col ops to A to get a diagonal matrix
 a 0 0 
 0 b 0 
 0 0 c 

Do just the col ops to I to get a new matrix P.

Then with the cols of P as the new basis vectors,
q = aX2 + bY2 + cZ2

example 3
Let q = 2x2 + 5y2 + 6z2 + 8xy + 4xz + 2yz.

Let A be the matrix for q. To diagonalize q using row/col ops, do the following
row/col ops to A and do just the col ops to I.

 2 4 2   1 0 0 
start with A =  4 5 1  start with I =  0 1 0 
 2 1 6   0 0 1 
R2 = -2R1 + R2 2 4 2 2 0 2 1 -2 0
C2 = -2C1 + C2 0 -3 -3 0 -3 -3 0 1 0
2 1 6 2 -3 6 0 0 1

R3 = R1 + R3 2 0 2 2 0 0 1 -2 -1
C3 = C1 + C3 0 -3 -3 0 -3 -3 0 1 0
0 -3 4 0 -3 4 0 0 1

R3 = -R2 + R3 2 0 0 2 0 0 1 -2 1
C3 = -C2 + C3 0 -3 -3 0 -3 0 0 1 -1
0 0 7 0 0 7 0 0 1

After every step, you get a new symmetric matrix in the A column and a new set of
basis col vectors in the I col. Each of the symmetric matrices represents q w.r.t.
the new basis. For instance, after the first row/col op you know (but it isn't very
 page 5 of Section 9.2

useful) that
q = 2X2 - 3Y2 + 6Z2 + 4XZ - 6YZ

in the new coord system with basis vectors (1,0,0), (-2,1,0), (0,0,1).
At the end you know (very useful) that
q = 2X2 - 3Y2 + 7Z2

in the coord system with basis vectors

u = (1,0,0), v = (-2,1,0), w = (1,-1,1).

As a check, let P = [u v w]. Then

 1 0 0   2 4 2   1 -2 1   2 0 0 
PTAP =  -2 1 0   4 5 1   0 1 -1  =  0 -3 0 
 1 -1 1   2 1 6   0 0 1   0 0 7 

so q does come out to be 2X2 - 3Y2 + 7Z2 in the coord system with basis u,v,w.

warning
In example 3, the diagonal coeffs 2,-3,7 are not eigenvalues of A and the cols of P
are not eigenvectors. This was a different method of diagonalizing.

diagonalization process continued

You can continue the diagonalization in example 3 so that the coeffs are –1's.
Do some more row/col ops:
divide R1 by √ 2; divide C1 by √ 2
divide R2 by √ 3; divide C2 by √ 3
divide R3 by √ 7; divide C3 by √ 7
Do the col ops to P, where the earlier col ops left off. All in all, this turns A
into
1 0 0
0 -1 0
0 0 1
and turns I into
1/√ 2 -2/√ 3 1/
√7
0 1/
√3 -1/√7
0 0 1/
√7
So
q = X2 - Y2 + Z2

in the coord system with basis vectors

1 2 1 1 1 1
( , 0, 0 ), (- , , 0 ), ( , - , )
2
√ 3 √
√ 3 7
√ √ 7 √7

footnote This continuation in example 3 can also be accomplished

algebraically by going from the latest X,Y,Z coord system to a new X1, Y1,
1 1 1
Z1 coord system by substituting X = X1, Y = Y1, Z = Z . You get
2
√ 3
√ √ 7 1
 page 6 of Section 9.2

1 1 1 2 2 2
q = 2( X1)2 - 3( Y1)2 + 7( Z1)2 = X1 - Y1 + Z1.
2
√ 3
√ 7
√
This substitution amounts to changing the scale on the X,Y,Z axes, i.e., to
getting new basis vectors by multiplying the latest basis vectors by 1/√2,
1/
√ 3 and 1/√ 7 respectively (Section 2.4).

a generalization of the row/col operation rule for getting a new matrix for q
old rule

Start here with the matrix of q Start with I, i.e., with [i j k]

w.r.t.the usual basis i,j,k

Row/col operate on the left side. Col op on the right side.

You end up on the left side with a new matrix for q w.r.t. a new basis consisting of
the cols on the right side.

more general rule

Start here with the matrix of q Start here with [u v w]

w.r.t. basis u,v,w

Row/col operate on the left side. Col op on the right side.

You end up on the left side with a new matrix for q w.r.t. a new basis consisting of
the cols on the right side.

(This works whether or not you have reached a diagonal matrix on the left side
although that's usually what you're aiming for.)

The more general version says you don't have to start at the "beginning", with the
matrix for q w.r.t. the usual basis i,j,k.

example 4 (mixing two methods).

Suppose the matrix (2 ≈ 2) for q has eigenvalues ¬ = 2, -3 with corresponding
eigenvectors (2,1), (-1,2).
Diagonalize q so that the diagonal coeffs are –1's and find the new basis.

solution
2 1 1 2
q = 2X2 - 3Y2 w.r.t. basis u = ( , ), v = (- , ).
5 √
√ 5 5 √
√ 5
Now pick up where the eigenvalue method left off and use row/col ops to make q
diagonal with diagonal coeffs –1's.
solution
matrix for q w.r.t. basis u,v [u v]

2 0 2/√ 5 -1/√ 5

0 -3 -1/√ 5 2/√ 5

Do these row /col ops on the left side and the col ops on the right side:

1 1
R1 = R , C = C
2 1 1 2 1
√
1 1
R2 = R2, C2 = C
3
√ 3 2
√
 page 7 of Section 9.2

You get

1 0 √
2/5 -1/√ 5
0 -1 1/√10 2/√15

So q = X2 -Y2 w.r.t. basis (√

2/5, 1/√10 ), (-1/√15, 2/√15 ).

graphing a conic in a NON-orthonormal coord system

In a non—orthonormal coord system, an ellipse can still be distinguished from a
hyperbola in the usual way, by the signs of the coeffs in the equation.
But the usual way of distinguishing circles from non-circular ellipses doesn't
work. For example, 2X2 + 6Y2 = 1 can be a circle and X2 + Y2 = 1 can be a
non—circular ellipse.

Furthermore, in a non—orthonormal system, the usual rules for finding vertices,

foci, major and minor axes etc. no longer hold.
For example, in an orthonormal coord system, the graph of

X2 Y2
+ = 1
4 9
is a (non—circular) ellipse with vertices (0, –3); the Y—axis is the major axis, the
X—axis is the minor axis. But in a non—orthonormal system all you can conclude is
that it is either a circle or an ellipse. Even if it is an ellipse, the points
(0, –3) are not necessarily vertices, the X—axis and Y—axis are not necessarily the
axes of the ellipse (Fig 2).
So to distinguish a hyperbola from an ellipse/circle, any diagonalization method
is OK. But to graph accurately and locate vertices and axes of the conic and
distinguish circles from non—circular ellipses, you have to use the eigenvector
method to get q diagonal in a new orthonormal coord system.
y

vertex Y
•
• ((0,3))
X

x
((0,-3))•

•
vertex

X2 Y2
possible graph of + = 1
4 9
FIG 2

sign distribution law

Given a quadratic form q in n variables.
All diagonal versions of q have the same coefficient sign distribution, i.e.,
the same number of positive, negative and zero coeffs. This sign distribution is the
same as the sign distribution of the eigenvalues of the matrix of q since one of the
diagonal forms of q has the eigenvalues as the coeffs.
(proof omitted)
For the q in example 3 which turned out to be 2X2 - 3Y2 + 7Z2 in another coord
system, every other diagonal form of q will have two positive and one negative coeff
. In particular, if you find the eigenvalues of the matrix for q, two will be
positive and one will be negative (but not necessarily 2,-3 and 7 specifically).
 page 8 of Section 9.2

example 4
Suppose q is a function of four variables and in some coord system q is
2 2 2
2X1 + 5X3 - 7X4. Then every diagonalization of q has two positive coeffs, one
negative coeff and one zero coeff. And the matrix for q has two positive
eigenvalues, one negative eigenvalue (not necessarily 2,5 and -7 though) and one
zero eigenvalue.

third method for diagonalizing q: completing the square

Look at the quadratic form

(3) q = 3x2 + y2 + 4z2 + 3xy + 6xz

To diagonalize q, first collect the x terms like this:

q = 3 [ x2 + (y + 2z)x ] + y2 + 4z2
The coeff of x in the bracket is y+2z. Take half the coeff, square it and add it
on to complete the square.

q = 3
[ x2 + (y+2z)x + (y+2z
2
)
2
] + y2 + 4z2 - 3 (
y+2z 2
2
)
2
[ 2 ] + y2 - 3yz + z2
= 3 x +
y+2z 1
4

Now collect the y terms and complete the square again:

2
[
q = 3 x +
y+2z
2 ]1
[
+ 4 y2 - 12yz + z2 ]
The coeff of y in the second brackets is 12z. Take half the coeff, square it and
add it on to complete the square.
2
[
q = 3 x +
y+2z
2 ] [
+ 4
1
y2 - 12zy + 36z2
1
]
+ z2 - 4 36z2

2 2
[
= 3 x +
y+2z
2 ] [1
+ 4 y - 6z - 8z2]
Now let
y+2z
(4) X = x + , Y = y - 6z, Z = z.
2

Then
1
(5) q = 3X2 + 4 Y2 -8Z2

If you want the basis vectors corresponding to the new coords X,Y,Z, write (4) in
matrix form:

 X   1 1/2 1   x 
 Y  =  0 1 -6   y 
 Z   0 0 1   z 

P-1

Invert to get
 page 9 of Section 9.2

 2 -1/2 -4 
(6) P =  0 1 6 
 0 0 1 

The new basis vector are u = (2,0,0), v = (-1/2, 1, 0),(-4,6,1).

PROBLEMS FOR SECTION 9.2

1. (a) Let q = x2 + 4xy - 2y2. Use the eigenvalue method here.

(i) Diagonalize q and find the basis for the new coord system in which q is
diagonal.
(ii) What change of variable produced the diagonal q.
(iii) Find the old matrix for q and find the connection between q and the old
matrix.
(iv) Find the new matrix for q and find the connection between q and the new
matrix.
(v) What's the connection between the old matrix and the new matrix.
(b) Sketch the graph of x2 + 4xy - 2y2 = 2 and identify the major axis and
vertices.

2. (a) Sketch the graph of 2x2 + 2y2 + 3z2 + 4xy+ 2xz + 2yz = 9 by switching to a
new orthonormal coord system.
The eigenvalues you need come out to be 2,5,0 with corresponding eigenvectors
u = (-1,-1,2), v = (1,1,1), w = (1,-1,0).
(b) Why did you need eigenvalues in part (a). Wouldn't row/col ops or completing the
square also have worked.
(c) What change of variables did you use in part (a). Express x,y,z in terms of X,Y,Z
and vice versa.
(d) Repeat part (a) but change the 9 on the righthand side to -9.

3. Let q = x2 + 4xy - 2y2.

(a) Diagonalize q using row/col ops.
(b) Find the change of variable that produces the diagonal q in part (a) and check
that it really works.
(c) Check that the new matrix for q really is PT old P.

 1 1 2 
4. Let q have matrix A =  1 3 5 .
 2 5 9 
(a) Diagonalize q using row/col ops and make the diagonal coeffs –1's.
(b) What change of variable produced the diagonal q in part (a).

 2 0 2 
5. Let A =  0 1 3 .
 2 3 4 
Without computing the eigenvalues, find their sign distribution (how many positives
and how many negatives).

6. Suppose q has matrix A and the eigenvalues of A are 1/3 and -4 with corresponding
eigenvectors u = (-4,3), v = (3,4).
Diagonalize q so that the coeffs are –1's, and find the new basis vectors.

7. Diagonalize q by completing the square and find the new basis vectors.
(a) q = x2 + 4xy - 2y2
(b) q = 2x2 + 3xy + y2
(c) q = 3x2 - 6y2 + z2 + 6xy + 18xz
(d) q = 3x2 + y2 + 4z2 + 2xy + 5xz + 6yz
 page 1 of Section 9.3

SECTION 9.3 DEFINITENESS

leading principal minors

The pm's of a square matrix are the northwest subdeterminants.
For example, the pm's of
 1 2 3 4 
 5 
 6 7 8 
 9 10 11 12 
 
 13 14 15 16 
are
 1 2 3 4 
 1 2 3   
|1|, | 1 2
5 6 |,  5 6 7
 9 10 11
 ,





5
9
13
6
10
14
7
11
15
8
12
16




positive definiteness
Let q be a quadratic form with matrix A (necessarily symmetric).
The following are equivalent (either all happen or none happen).
When they do happen then q and A are called positive definite.
Û
(1) q is positive at every point (x,y,z) except at the origin 0 (where every q has
value 0.)
(2) In every diagonal version of q the coeffs are positive.
In particular if you use row/col ops on A to get a diagonal matrix, the diagonal
entries are positive.
(3) The eigenvalues of A are positive.
(4) The pm's of A are positive.

positive semi (non-negative ) definiteness

Let q be a quadratic form with matrix A (necessarily symmetric).
The following are equivalent (either all happen or none happen).
When they do happen then q and A are called non—negative definite or positive
semi—definite.
(1) q ≥ 0 at every point.
(2) In any diagonal version of q, the coeffs are ≥ 0.
In particular if you use row/col ops on A to get a diagonal matrix, the diagonal
entries are ≥ 0.
(3) The eigenvalues of A are ≥ 0.
(4) The pm's of A are ≥ 0.

The categories positive definite and positive semi—definite overlap: If A is

positive definite then it is also positive semi—definite.

If in addition to (1)-(4) above, q actually is 0 somewhere besides the origin or in

some diagonal version of q one of the coeffs actually is 0 or one of the eigenvalues
of A actually is 0 or one of the pm's actually is 0 then q and A are positive semi—
definite but not positive definite.

negative definiteness
Let q be a quadratic form with matrix A (necessarily symmetric)
The following are equivalent (either all happen or none happen).
When they do happen then q and A are called negative definite.
Û
(1) q is negative at every point (x,y,z) except at the origin 0 (where q = 0).
(2) In every diagonal version of q, the coeffs are negative.
In particular if you use row/col ops on A to get a diagonal matrix, the diagonal
entries are negative.

(3) The eigenvalues of A are negative.

(4) The pm's of A starting with the most northwesterly have the signs
- + - + - + ...

negative-semi (non-positive) definiteness

 page 2 of Section 9.3

Let q be a quadratic form matrix with A (necessarily symmetric)

The following are equivalent (either all happen or none happen).
When they do happen then q and A are called non—positive definite or negative
semi—definite.
(1) q ≤ 0 at every point.
(2) In any diagonal version of q, the coeffs are ≤ 0.
In particular if you use row/col ops on A to get a diagonal matrix, the diagonal
entries are ≤ 0.

(3) The eigenvalues of A are ≤ 0.

(4) The pm's of A starting with the most northwesterly are ≤ 0, ≥ 0, ≤ 0, ≥ 0, etc.

The categories negative definite and negative semi—definite overlap: If A is

negative definite then it is also negative semi—definite.

If in addition to (1)-(4) above, q actually is 0 somewhere besides the origin or in

some diagonal version of q one of the coeffs actually is 0 or one of the eigenvalues
of A actually is 0 or one of the pm's actually is 0 then q and A are negative semi—
definite but not negative definite.

indefiniteness
If q and A are none of the above then they are called indefinite.

Only symmetric matrices are assigned a definiteness. So if a matrix is called

positive, positive semi, neg, negative semi or indefinite you may assume that the
matrix is symmetric to begin with.

For each kind of definiteness, the equivalence of (1)-(3) follows from the
preceding sections.
The proof that (4) belongs on the list with (1)-(3) is too messy.

example 1
Let q = x2 - 3y2 + 7xy. Find the definiteness of q.

method 1 If x = 0, y = 1 then q =-3, negative. If y=0 and x = 1 then q = 1,

positive. So by inspection, q is indefinite. (But it's not always this easy to find
the definiteness by inspection.)

method 2 The matrix for q is

[ 1
7/2 -3
7/2
]
The pm's are 1, -61/4 which makes q indefinite.
7
method 3 Start with the matrix for q and diagonalize it by adding - 2 row1 to
7
row2 and adding - 2 col1 to col2 to get

[ 10 0
-61/4 ]
61
In another coord system (I don't care about the basis vectors), q = X2 - 4 Y2. There
is one positive and one negative coeff. So q is indefinite.

-8 – 4√65
method 4 The matrix for q has eigenvalues . One is positive and one is
8
negative. So q is indefinite.
 page 3 of Section 9.3

PROBLEMS FOR SECTION 9.3

1. Let
 2 1 1 
A =  1 1 0 
 1 0 1 

Find the definiteness of A with several methods, for practice.

2. Find the definiteness. (a) [ 21 1

2 ] (b) [ 12 2
1 ]
3. True or False. If q is a quadratic form with all positive coeffs (e.g.,
q = ax2 + bxy + cy2 where a,b,c are positive) then q is positive definite.

4. Show that if A is negative definite then -A is positive definite

(a) using (1) on the lists
(b) using (4) on the lists

5. Suppose A is positive definite. What about 3A.

6. Suppose q has matrix A, say 3 ≈ 3. What can you conclude about the definiteness
of A, the eigenvalues of A, and |A| if
(a) in some new coord system, q = -2X2 - 3Y2 - 5Z2
(b) in some new coord system, q = X2 - 2Y2 + Z2

7. Let B be an arbitrary matrix.

(a) Show that BTB is symmetric.
(b) Let q be the quadratic form with matrix BTB. Write q in matrix notation.
(c) Show that BTB is positive semi—definite by continuing from part (b) with matrix
algebra until you can see that q ≥ 0.
Suggestion: Use the fact that for column vectors u and v, u…v = uTv (Section 2.1).

8. Suppose A is symmetric. True or False.

(a) If A is positive definite then |A| is positive
(b) If |A| is positive then A is positive definite.

9. Let q = 2x2 + 3y2 + bxy. For what values of b will q be

(a) positive definite
(b) positive semi—definite
(c) positive semi—definite but not positive definite (i.e., only positive
semi—definite)
(d) indefinite
(e) negative definite

10. (a) Show that if A is positive definite then A is invertible.

(b) Is the converse true?
(c) Suppose A is positive definite (so that by part (a), it is invertible). Show
that the inverse must also be positive definite by thinking further about the
eigenvalues of A-1.

 a b c 
11. Let A =  b d e  . True or False.
 c e f 
(a) If A is positive definite then the diagonal entries a,d,f are positive.

(b) If the diagonal entries a,d,f are positive then A is positive definite.
SUMMARY OF BASIS CHANGING AND DIAGONALIZING
Read this before the 3rd exam and then again before the final exam.

basis changing for points, linear transformations, quadratic forms

Given a new coord system (i.e., given new basis vectors).
Let P be the basis changing matrix, with the new basis vectors as its cols.
Here's how to get the new coords of a point, the new matrix for a linear
transformation T and the new matrix for a quadratic form q.

 coords
new 
  coords
old 

 of  = P-1  of 
 vector
the   vector
a 

new Matrix for T = P-1 oldMatrix P (the old and new matrices are called similar)

new Matrix for q = PT oldMatrix P (the old and new matrices are called congruent)

warning
The rule involves PT when you get a new matrix for a quadratic form.
The rule involves P-1 when you get a new matrix for a transformation.
(If the new coord system is orthonormal then PT = P-1.)

diagonalizing a matrix
Diagonalizing M means finding a diagonal matrix Ò and an invertible matrix P so
that P-1MP = Ò.
Equivalently, diagonalizing M means finding a new basis so that the transformation
which has matrix M w.r.t. the standard basis will have a diagonal matrix Ò w.r.t.
the new basis.
There is only one method. The columns of P must be eigenvectors of M and the
diagonal entries of Ò are the corresponding eigenvalues of M.
Not all matrices can be diagonalized (only the ones that have a complete set of
eigenvectors).
Hermitian matrices are not only diagonalizable but you can do it with a real
Ò and a unitary P provided you take the trouble to choose orthogonal normalized
eigenvectors.

diagonalizing a quadratic form

Let q have matrix A.
Diagonalizing q means switching to a new coord system in which q has only square
terms.
Equivalently, diagonalizing q means finding an invertible P so that PTAP is
diagonal.
Every q can be diagonalized (in many ways).
One diagonalizing method is to use as a basis (i.e., as the cols of P) a complete
set of orthonormal eigenvectors of A in which case

q = ¬1X2 + ¬2Y2 + ¬3Z2.

This is the method to use if you want to graph an equation of degree 2, in two or
three variables, accurately.
Other methods use row/col ops and completing the square and do not involve
eigenvalues or eigenvectors or orthogonalizing or normalizing. These are the methods
to use if all you care about is the definiteness of q (e.g., whether the graph of an
equation of degree 2, in two variables, is an ellipse or hyperbola).
 warning
To diagonalize q, you don't have to use eigenvectors at all. And the coeffs of the
square terms in the new coord system don't have to be eigenvalues.
But if you want to turn q into ¬1X2 + ¬2Y2 + ¬3Z2, you must not only use
eigenvectors as the new basis then you must use orthonormal eigenvectors.
To diagonalize a matrix M you must use eigenvectors but they don't have to be
orthogonal or unit length. It's just convenient if they are.
 page 1 of review problems for chapter 9

REVIEW PROBLEMS FOR CHAPTER 9

1. Let q = 2x2 - 4xy + 5y2

1
(a) Find the new formula for q in the coord system with basis u = 2i, v = 3 j
twice.
(i) Find the connection between X,Y and x,y and then use algebra
(ii) Use the basis changing rule

(b) Diagonalize q using three different methods.

(c) Sketch the graph of 2x2 - 4xy + 5y2 = 7 and identify some of its significant
features (in the old coord system).

2. Let
 2 -1 1 
A =  -1 2 -1 
 1 -1 2 

(a) Use two methods to show that A is positive definite.

(b) Find an invertible P so that PTAP = I.

3. Suppose a quadratic form in variables x1, x2, x3 has matrix A.

 x1 
Û   Û
Let x =  x2  . Write the equation relating q,A, and x .
x 
 3

4. Show that if A is congruent to B then |AB| ≥ 0.

5. Let A be symmetric.
If row/col ops turn A into B and the col ops turn I into M, how are A,B,M related.

6. Suppose q = x2 - y2 + 3xy in the standard coord system, and q = -2X2 + 3Y2 in

another coord system with basis vectors u,v. Find the old matrix for q, the new
matrix for q and write an equation relating the two matrices.