MA 106: Linear Algebra

Lecture 17

Prof. B.V. Limaye

IIT Dharwad

Wednesday, 14 February 2018

B.V. Limaye, IITDH MA 106: Lec-17

Real Quadratic Forms

Let n ∈ N. A real n-ary quadratic form Q is a

homogeneous polynomial of degree 2 in n variables. Thus
∑
n ∑
n
Q(x1 , . . . , xn ) := αjk xk xj , where αjk ∈ R.
j=1 k=1

Examples
Let a, b, c, a′ , b ′ , c ′ ∈ R.
n = 1 : Q(x) := a x 2 (unary quadratic form)
n = 2 : Q(x, y ) := a x 2 + b y 2 + a′ xy (binary quadratic form)
n = 3 : Q(x, y , z) := a x 2 + b y 2 + c z 2 + a′ xy + b ′ yz + c ′ zx
(ternary quadratic form)

B.V. Limaye, IITDH MA 106: Lec-17

Given an n-ary quadratic form Q as above, define

Aα := [αjk ] ∈ Rn×n .
[ ]T
Then for x := x1 · · · xn ,
∑n 
α1k xk
[ ] k=1
.. 
x T Aα x = x1 · · · xn  . 
∑n
k=1 αnk xk
n (∑
∑ n )
= αjk xk xj = Q(x).
j=1 k=1

In general, the n×n matrix Aα := [αjk ] defined above will not

be symmetric. We shall show that there is a unique n×n real
symmetric matrix A such that Q(x1 , . . . , xn ) = xT A x for all
x ∈ Rn×1. For the uniqueness part, we need the following.
B.V. Limaye, IITDH MA 106: Lec-17
Lemma
Let A be a real symmetric matrix. Then for all x, y ∈ Rn×1 ,

4⟨A x, y⟩ = ⟨A(x + y), (x + y)⟩ − ⟨A(x − y), (x − y)⟩.

Consequently, if ⟨A x, x⟩ = 0 for all x ∈ Rn×1 , then A = O.

Proof. Let x, y ∈ Rn×1 . Since A is symmetric,

⟨Ay, x⟩ = (Ay)T x = yT AT x = yT A x = ⟨y, A x⟩ = ⟨A x, y⟩.
Hence
⟨A(x + y), (x + y)⟩ = ⟨A x, x⟩ + ⟨A x, y⟩ + ⟨Ay, x⟩ + ⟨Ay, y⟩
= ⟨A x, x⟩ + 2⟨A x, y⟩ + ⟨Ay, y⟩
Similarly,
⟨A(x − y), (x − y)⟩ = ⟨A x, x⟩ − 2⟨A x, y⟩ + ⟨Ay, y⟩
B.V. Limaye, IITDH MA 106: Lec-17
Subtraction gives

⟨A(x + y), (x + y)⟩ − ⟨A(x − y), x − y⟩ = 4⟨A x, y⟩,

as desired.
Now suppose ⟨A x, x⟩ = 0 for all x ∈ Rn×1 . Then the above
equality shows that ⟨A x, y⟩ = 0 for all x, y ∈ Rn×1 .
Let A := [ajk ], and let e1 , . . . , en be the basic column vectors
in Rn×1 . Then

ajk = ⟨A ek , ej ⟩ = 0 for all j, k = 1, . . . , n.

Hence A = O.

B.V. Limaye, IITDH MA 106: Lec-17

Proposition
Let Q be a real n-ary quadratic form. Then there is a unique
n×n real symmetric matrix A such that
[ ]T
Q(x1 , . . . , xn ) = xT A x for all x := x1 · · · xn ∈ Rn×1 .

Proof. ∑ ∑
Let Q(x1 , . . . , xn ) := nj=1 nk=1 αjk xk xj , where αjk ∈ R.
Define Aα := [αjk ] ∈ Rn×n . Then xT Aα x = Q(x1 , . . . , xn ) for
[ ]T
all x := x1 · · · xn ∈ Rn×1 as we have already seen.
Now define A := (Aα + ATα )/2. Then A is real symmetric, and
1 1
xT A x = xT Aα x + xT ATα x = xT Aα x = Q(x),
2 2
for all x ∈ Rn×1 , since xT ATα x = (xT ATα x)T = xT Aα x ∈ R.
To prove uniqueness, let B be a real symmetric matrix such
[ ]T
that Q(x1 , . . . , xn ) = xT B x for all x := x1 · · · xn ∈ Rn×1 .
B.V. Limaye, IITDH MA 106: Lec-17
Let C := A − B. Then C is a real symmetric matrix and

⟨C x, x⟩ = xT C x = xT A x − xT B x = 0 for all x ∈ Rn×1 .

By the preceding lemma, C = O, that is, B = A.

∑ ∑
Let Q(x1 , . . . , xn ) := nj=1 nk=1 αjk xj xk be a real quadratic
form. The unique real symmetric matrix A satisfying
[ ]T
Q(x1 , . . . , xn ) = xT A x for all x := x1 · · · xn ∈ Rn×1 is
called the matrix associated with the quadratic form Q,
and we write Q := QA . In fact, A := [ajk ], where

1
ajk := (αjk + αkj ) for j, k = 1, . . . , n.
2
The function QA : Rn×1 → R is defined by
[ ]T
QA (x1 , . . . , xn ) := xT A x for x := x1 · · · xn ∈ Rn×1 .

B.V. Limaye, IITDH MA 106: Lec-17

Thus to every real quadratic form, we associate a real
symmetric matrix. Conversely, every real symmetric matrix
defines a real quadratic form as above.
Example
Let Q(x1 , x2 ) := 3x12 + 4x1 x2 + 6x2 x1 + 2x22 for (x1 , x2 ) ∈ R2 .
[ ]T
Then for x = x1 x2 ∈ R2×1 ,

Q(x1 , x2 ) = 3x12 + 5x1 x2 + 5x2 x1 + 2x22

[ ][ ]
[ ] 3 5 x1
= x1 x2
5 2 x2
= QA (x),
[ ]
3 5
where A := is symmetric.
5 2

B.V. Limaye, IITDH MA 106: Lec-17

A real n-ary quadratic form Q is said to be a diagonal
quadratic form if there are λ1 , . . . , λn ∈ R such that
Q(x1 , . . . , xn ) = λ1 x12 + · · · + λn xn2 .
It is clear that a quadratic form Q is diagonal if and only if Q
is associated with a diagonal matrix D, that is, Q = QD .
Using the spectral theorem for real symmetric matrices, we
show that every quadratic form can be orthogonally
transformed to a diagonal quadratic form.
Theorem
Let Q be a real quadratic form and A ∈ Rn×n be the
symmetric matrix associated with Q. If C is an orthogonal
matrix such that the matrix D := CT A C is diagonal, then
Q(x) = QD (y), where y := CT x for all x ∈ Rn×1 .

Proof. Let x ∈ Rn×1 and y := CT x = C−1 x. Then x = Cy and

QA (x) = xT A x = (Cy)T A(Cy) = yT (CT AC)y = yT Dy = QD (y).
B.V. Limaye, IITDH MA 106: Lec-17
To diagonalise a real n-ary quadratic form Q, we first write
down the (real symmetric) matrix A ∈ Rn×n associated with
Q. We then find an orthonormal basis u1 , . . . , un consisting of
eigenvectors of A corresponding to its eigenvalues λ1 , . . . , λn
counted according to their algebraic multiplicities. If we let
[ ]
C := u1 · · · un and D := diag(λ1 , . . . , λn ),
 
uT1
 
Then Q(x) = λ1 y12 + · · · + λn yn2 , where y := CT x =  ...  x.
uTn
Example
Let us transform the quadratic form
Q(x) = x12 + x22 +x32 + 4x1 x2 +4x2 x3 − 4x3 x1 to a diagonal
1 2 −2
form. Then A := 2 1 2  is the associated matrix.

−2 2 1
B.V. Limaye, IITDH MA 106: Lec-17
In Lecture
[ √16, we√have seen
]T that [ √ √ √ ]T
u1 := 1/ 2 1/ 2 0 , u2 := −1/ 6 1/ 6 2/ 6
[ √ √ √ ]T
and u3 := 1/ 3 −1/ 3 1/ 3 are eigenvectors of A
corresponding to the eigenvalues 3, 3 and −3 respectively, and
they form an orthonormal basis for R3×1 . Hence let
 √ √ √ 
1/√2 −1/√ 6 1/ √3
C := 1/ 2 1/√6 −1/√ 3 and D := diag(3, 3, −3).
0 2/ 6 1/ 3
Then CT A C = D, and so Q(x) = 3(y12 + y22 − y32 ), where
   √ √  
y1 1/ √2 1/√2 0√ x1
y = y2  := CT x = −1/√ 6 1/ √6 2/√6 x2 ,
y3 1/ 3 −1/ 3 1/ 3 x3
√ √
that is, y1 = (x1 + x2√
)/ 2, y2 = (−x1 + x2 + 2x3 )/ 6 and
y3 = (x1 − x2 + x3 )/ 3.
B.V. Limaye, IITDH MA 106: Lec-17
Conic Sections
A conic section is the locus in R2 of an equation
a x 2 + b y 2 + c xy + a′ x + b ′ y + c ′ = 0,
where a, b, c, a′ , b ′ , c ′ ∈ R. We assume WLOG that not both a
and b are negative. It can be proved that this equation
represents one of the following:
(i) the empty set
(ii) a single point
(iii) one or two straight lines
(iv) an ellipse
(v) an hyperbola
(vi) a parabola.
Terms of the second degree on the LHS of the equation give
Q(x, y ) := a x 2 + b y 2 + c xy .
It is a binary quadratic form. It determines the type of the conic.
B.V. Limaye, IITDH MA 106: Lec-17
The (real symmetric) matrix associated with Q is
[ ]
a c/2
A := .
c/2 b
Hence the equation of the given conic section becomes
[ ][ ] [ ]
[ ] a c/2 x [ ′ ′] x
x y + a b + c ′ = 0,
c/2 b y y
that is,
[ ] [ ]T [ ][ ]T
x y A x y + a′ b ′ x y + c ′ = 0.
Let C := [u1 , u2 ] be an orthogonal matrix whose columns u1
and u2 are eigenvectors of A with corresponding eigenvalues
λ1 and λ2 , and let D := diag(λ1 , λ2 ).
[ ] [ ]
x u
We use the change of variables =C to transform the
y v
quadratic form Q(x, y ) to a diagonal form as follows.
B.V. Limaye, IITDH MA 106: Lec-17
 [ ] [ ]T
Q(x, y ) = x y A x y
[ ] [ ]T
= u v CT A C u v
[ ] [ ]T
= u v D u v
= λ1 u 2 + λ2 v 2 = QD (u, v ).
The orthonormal basis {u1 , u2 } determines a new set of
coordinate axes, and we obtain
[ ] [ ]
[ ] u [ ′ ′] u
u v diag(λ1 , λ2 ) + a b C + c′
v v
[ ]
[ ′ ′] [ ] u
2 2
= λ1 u + λ2 v + a b u1 u2 + c ′ = 0.
v
If the conic so determined is not degenerate, that is, if it does
not reduce to an empty set, a point, or line(s), then the signs
of λ1 and λ2 determine the type of the conic section as follows.
B.V. Limaye, IITDH MA 106: Lec-17
The equation represents

1. ellipse if λ1 λ2 > 0, that is, both λ1 and λ2 are positive,

2. hyperbola if λ1 λ2 < 0, that is, one of λ1 , λ2 is positive and

the other negative,

3. parabola if λ1 λ2 = 0, that is, one of λ1 , λ2 is zero.

B.V. Limaye, IITDH MA 106: Lec-17

Examples
1. Consider the conic section given by
2x 2 + 4xy + 5y 2 + 4x + 13y − 1/4 = 0, and the binary
quadratic form Q(x, y ) := 2x 2 + 4xy + 5y 2 . [ ]
2 2
Then the associated symmetric matrix A := has
2 5
eigenvalues λ1 := 1 and λ2 := √ 6, and the corresponding√
[ ]T [ ]T
eigenvectors u1 := 2 −1 / 5 and u2 := 1 2 / 5
form an orthonormal basis for R2×1 . Hence let
[ ]
1 2 1
C := √ and D := diag(1, 6).
5 −1 2

Then CT AC = D. So Q(x, y ) = QD (u, v ) = u 2 + 6v 2 , where

[ ] [ ] [ ][ ]
u T x 1 2 −1 x
:= C =√ , that is,
v y 5 1 2 y
√ √
u = (2x − y )/ 5 and v = (x + 2y )/ 5.
B.V. Limaye, IITDH MA 106: Lec-17
 [ ] [ ]
x u √
Since := C , substituting x = (2u + v )/ 5 and
y v
√
y = (−u + 2v )/ 5 in the given equation of the conic section,
we obtain
√ √ 1
u 2 + 6v 2 − 5u + 6 5v − = 0.
4
Completing the squares, we see that
( √ )2 ( √ )2
5 5
u− +6 v + = 9.
2 2

This
√ is an√equation of an ellipse with its centre at
( 5/2,− 5/2) in the uv -plane.
The u-axis and v -axis are determined by the eigenvectors u1
and u2 , as indicated in the following figure.

B.V. Limaye, IITDH MA 106: Lec-17

 V

B.V. Limaye, IITDH MA 106: Lec-17

 2. Consider the conic section given by
2x 2 − 4xy − y 2 − 4x + 10y − 13 = 0, and the binary quadratic
form Q(x, y ) := 2x 2 − 4xy − y 2 . [ ]
2 −2
The associated symmetric matrix A := has
−2 −1
eigenvalues λ1 = 3, λ2 = −2, and the corresponding √
[ ]T √ [ ]T
eigenvectors u1 := 2 −1 / 5 and u2 := 1 2 / 5
form an orthonormal basis for R2×1 . Hence let
[ ]
1 2 1
C := √ and D := diag(3, −2).
5 −1 2

Then CT A C = D. Hence Q(x, y ) = QD (u, v ) = 3u 2 − 2v 2 ,

[ ] [ ] [ ][ ]
u T x 1 2 −1 x
where := C =√ ,
v y 5 1 2 y
√ √
that is, u = (2x − y )/ 5 and v = (x + 2y )/ 5.

B.V. Limaye, IITDH MA 106: Lec-17

 √ √
Thus substituting x = (2u + v )/ 5 and y = (−u + 2v )/ 5
in the given equation of the conic section, we obtain
4 10
3u 2 − 2v 2 − √ (2u + v ) + √ (−u + 2v ) − 13 = 0,
5 5
that is,
1
3u 2 − 2v 2 − √ (18u − 16v ) − 13 = 0.
5
Completing the squares, we see that
√ √
(u − 3/ 5)2 (v − 4/ 5)2
− = 1.
4 6
equation of a hyperbola with its centre
This√is an √
(3/ 5, 4/ 5) in the uv -plane.

B.V. Limaye, IITDH MA 106: Lec-17

The u-axis and v -axis are determined by the eigenvectors u1
and u2 , as indicated in the following figure:

B.V. Limaye, IITDH MA 106: Lec-17

 3. Consider the conic section given by
9x 2 + 24xy + 16y 2 − 20x + 15y = 0, and the binary quadratic
form Q(x, y ) := 9x 2 + 24xy
[ + 16y 2
] . Then the associated
9 12
symmetric matrix A := has eigenvalues λ1 := 25 and
12 16
[ ]T
λ2 := 0, and the corresponding eigenvectors u1 := 3 4 /5
[ ]T
and u2 := −4 3[ /5 form ] an orthonormal basis for R .
2×1

1 3 −4
Hence let C := and D := diag(25, 0). Then
5 4 3
CT A C = D. Thus Q(x, y ) = QD (u, v ) = 25u 2 , where
[ ] [ ] [ ][ ]
u T x 1 3 4 x
:= C = ,
v y 5 −4 3 y
that is, u = (3x + 4y )/5 and v = (−4x + 3y )/5. Substituting
x = (3u − 4v )/5 and y = (4u + 3v )/5 in the equation of the
conic, we obtain u 2 + 25v = 0, an equation of a parabola with
its vertex at (0, 0) in the uv - plane, as shown below.
B.V. Limaye, IITDH MA 106: Lec-17
 U
V

B.V. Limaye, IITDH MA 106: Lec-17

Quadric Surfaces
A quadric surface is the locus in R3 of an equation
a x 2 +b y 2 +c z 2 +a′ xy +b ′ yz +c ′ zx +a′′ x +b ′′ y +c ′′ z +d = 0,
where a, b, c, a′ , b ′ , c ′ , a′′ , b ′′ , c ′′ , d ∈ R. We assume WLOG
that not all three a, b and c are negative.
Terms of the second degree on the LHS of the equation give
Q(x, y , z) := a x 2 + b y 2 + c z 2 + a′ xy + b ′ yz + c ′ zx.
It is a ternary quadratic form. It determines the type of the
quadric surface. The real symmetric matrix associated with
the quadratic form Q is
 
a a′ /2 c ′ /2
A := a′ /2 b b ′ /2 .
c ′ /2 b ′ /2 c
B.V. Limaye, IITDH MA 106: Lec-17
Hence the equation of the given quadric surface becomes
 ′ ′
   
[ ] ′a a /2 c ′ /2 x [ ′′ ′′ ′′ ] x
x y z a /2 b b /2 y + a b c y +d = 0,
c ′ /2 b ′ /2 c z z

that is,
[ ] [ ]T [ ][ ]T
x y z A x y z + a′′ b ′′ c ′′ x y z + d = 0.
[ ]
Let C := u1 u2 u3 be an orthogonal matrix whose
columns u1 , u2 , u3 are eigenvectors of A with corresponding
eigenvalues λ1 , λ2 , λ3 , and let D := diag(λ1 , λ2 , λ3 )..
   
x u
We use the change of variables y  = C  v  to transform
z w
the quadratic form Q(x, y , z) to a diagonal form as follows.

B.V. Limaye, IITDH MA 106: Lec-17

 [ ] [ ]T
Q(x, y , z) = x y z A x y z
[ ] [ ]T
= u v w CT A C u v w
[ ] [ ]T
= u v w D u v w
= λ1 u 2 + λ2 v 2 + λ23 w 2 = QD (u, v , w ).

The orthonormal basis {u1 , u2 , u3 } determines a new set of

coordinate axes, so that the locus of the original equation is
given by
   
[ ] u [ ′′ ′′ ′′ ] u
 
u v w diag(λ1 , λ2 , λ3 ) v + a b c C v  + d = 0.
w w

Leaving aside the degenerate cases, the primary cases are:

B.V. Limaye, IITDH MA 106: Lec-17

Equation Surface Signs of eigenvalues of A

x2 y2 z2
+ + =1 ellipsoid all three positive
a2 b2 c2
x2 y2 z
+ − =0 elliptic paraboloid two positive, one zero
a2 b2 c2
x2 y2 z
+ − =0 elliptic cone two positive, one negative
a2 b2 c 22
x2 y2 z
+ − =1 1-sheeted hyperboloid two positive, one negative
a2 b2 c 22
x2 y2 z
− − =1 2-sheeted hyperboloid one positive, two negative
a2 b2 c2
x2 y2 z
− − =0 hyperbolic paraboloid one positive, one negative,
a2 b2 c
one zero.

See the file surfaces.pdf kept on the moodle for pictures of

various quadric surfaces mentioned above.
B.V. Limaye, IITDH MA 106: Lec-17
Example Consider the quadric surface given by

x 2 + y 2 + z 2 + 4xy + 4yz − 4zx − 27 = 0,

and the associated ternary quadratic form

Q(x, y , z) := x 2 + y 2 + z 2 + 4xy + 4yz − 4zx.
We have
 already transformed
 the associated symmetric matrix
1 2 −2
A :=  2 1 2  to a diagonal form, and have obtained
−2 2 1
Q(x, y , z) = QD (u, v , w ) = 3(u 2 + v 2 − w 2 ) (with x1 , x2 , x3
and y1 , y2 , y3 in place of x, y , z and u, v , w ),
where D := diag(3,
√ 3, −3) and √ √
u = (x + y )/ 2, v = (−x + y + 2z)/ 6, w = (x − y + z)/ 3.
Under this change of coordinates, the quadric surface reduces
to u 2 + v 2 − w 2 = 9.
B.V. Limaye, IITDH MA 106: Lec-17
This is an equation of a hyperboloid of one sheet in the
uv -plane, as shown in the following figure.

B.V. Limaye, IITDH MA 106: Lec-17