Bilinear and quadratic forms

Zdeněk Dvořák
April 28, 2015

1 Bilinear forms
Definition 1. Let V be a vector space over a field F. A function b : V×V →
F is a bilinear form if

b(u + v, w) = b(u, w) + b(v, w) b(u, v + w) = b(u, v) + b(u, w)

b(αu, v) = αb(u, v) b(u, αv) = αb(u, v)

for all u, v, w ∈ V and α ∈ F.

The bilinear form b is symmetric if b(u, v) = b(v, u) for all u, v ∈ V.

Remark: b(o, v) = b(v, o) = 0.

Examples:

• b(x, y) = hx, yi in Rn is bilinear and symmetric for any scalar product.

• b((x1 , y1 ), (x2 , y2 )) = x1 x2 + 2x1 y2 + 3y1 x2 + 4y1 y2 is bilinear, but not

symmetric.

Definition 2. Let C = v1 , . . . , vn be a basis of V and let b be a bilinear form

on V. The matrix of b with respect to C is
 
b(v1 , v1 ) b(v1 , v2 ) . . . b(v1 , vn )
 b(v2 , v1 ) b(v2 , v2 ) . . . b(v2 , vn ) 
[b]C =  .
 ... 
b(vn , v1 ) b(vn , v2 ) . . . b(vn , vn )

Lemma 1. Let C = v1 , . . . , vn be a basis of V and let b be a bilinear form

on V. For any x, y ∈ V, we have

b(x, y) = [x]C [b]C [y]TC .

1
Proof. Let [x]C = (α1 , . . . , αn ) and [y]C = (β1 , . . . , βn ). We have
b(x, y) = b(α1 v1 + . . . + αn vn , β1 v1 + . . . + βn vn )
Xn X n
= αi βj b(vi , vj )
i=1 j=1
Xn X n


= αi βj [b]C i,j
i=1 j=1

= [x]C [b]C [y]TC .

Remark: [b]C is the only matrix with this property. A bilinear form b is
symmetric if and only if [b]C is a symmetric matrix.
Corollary 2. Let V be a vector space over a field F. Let C = v1 , . . . , vn be a
basis of V. For every n × n matrix M over F, there exists a unique bilinear
form b : V × V → F such that b(vi , vj ) = Mi,j for 1 ≤ i, j ≤ n.
Proof. Define b(x, y) = [x]C M [y]TC and observe that b is bilinear. No other
bilinear form with this property exists, since any bilinear form satisfying
the assumptions has matrix M , which by Lemma 1 uniquely determines the
values of the bilinear form.
Example 1. Thebilinear  form b((x1 , y1 ), (x2 , y2 )) = x1 x2 + 2x1 y2 + 3y1 x2 +
1 2
4y1 y2 has matrix with respect to the standard basis;
3 4
  
1 2 x2
b((x1 , y1 ), (x2 , y2 )) = (x1 , y1 ) .
3 4 y2
Lemma 3. Let B = v1 , . . . , vn and C be two bases of V and let b be a bilinear
form on V. Let S = [id]B,C . Then
[b]B = S T [b]C S.
Proof. We have
S T [b]C S = eTi S T [b]C Sej


i,j
= [vi ]B S T [b]C S[vj ]TB
= [vi ]C [b]C [vj ]TC
= b(vi , vj ) = ([b]B )i,j .

2
2 Quadratic forms
Definition 3. A function f : V → F is a quadratic form if there exists a
bilinear form b : V × V → F such that f (x) = b(x, x) for every x ∈ V.

Example 2.
f ((x, y)) = x2 +5xy+4y 2 is a quadratic form, since f ((x, y)) = b((x, y), (x, y))
for the bilinear form b((x1 , y1 ), (x2 , y2 )) = x1 x2 + 2x1 y2 + 3y1 x2 + 4y1 y2 .
Also, f ((x, y)) = b1 ((x, y), (x, y)) for the symmetric bilinear form b((x1 , y1 ), (x2 , y2 )) =
x1 x2 + 25 (x1 y2 + y1 x2 ) + 4y1 y2 .

Lemma 4. Let V be a vector space over a field F whose characteristic is

not 2. For every quadratic form f , there exists a unique symmetric bilinear
form b such that f (x) = b(x, x) for every x ∈ V.

Proof. Since f is quadratic, there exists a bilinear form b0 such that f (x) =
b0 (x, x) for every x ∈ V. Let b(x, y) = 21 (b0 (x, y) + b0 (y, x)). Then b is a
symmetric bilinear form and b(x, x) = b0 (x, x) for every x ∈ V. Hence, b is
a symmetric bilinear form such that f (x) = b(x, x) for every x ∈ V.
To show that b is unique, it suffices to note that

b(x + y, x + y) − b(x, x) − b(y, y) f (x + y) − f (x) − f (y)

b(x, y) = =
2 2
whenever b is a symmetric bilinear form b satisfying f (x) = b(x, x) for every
x ∈ V.
Hence, if f is a quadratic form on V and C is a basis of V, then

f (x) = [x]C A[x]TC

for a unique symmetric matrix A. We write [f ]C = A. Also, by Lemma 3, if

B is another basis of V and S = [id]B,C , then

[f ]B = S T [f ]C S.

Lemma 5. Let f : V → F be a quadratic form. Let B1 = v1 , . . . , vn and

B2 = w1 , . . . , wn be two bases of V such that both [f ]B1 and [f ]B2 are diagonal
matrices. Then [f ]B1 and [f ]B2 have the same number of positive entries.

Proof. For i ∈ {1, 2}, let ai be the number of positive entries of [f ]Bi , and
suppose for a contradiction that a1 > a2 . Let I = {i : f (vi ) > 0}, J = {i :
f (wi ) ≤ 0}. Let U1 = span({vi : i ∈ I}) and U2 = span({wi : i ∈ J}). Note

3
that dim(U1 ) + dim(U2 ) = a1 + (dim(V) − a2 ) > dim(V), and thus there
exists a non-zero vector v ∈ U1 ∩ U2 . However,
X
f (v) = [v]B1 [f ]B1 [v]TB1 = ([v]B1 )2i f (vi ) > 0
i∈I
X
f (v) = [v]B2 [f ]B2 [v]TB2 = ([v]B2 )2i f (wi ) ≤ 0
i∈J

This is a contradiction.
For integers a, b, and c, let D(a, b, c) be the diagonal matrix with

+1
 for i = 1, . . . , a,
Di,i = −1 for i = a + 1, . . . , a + b, .

0 for i = a + b + 1, . . . , a + b + c.


Theorem 6 (Sylvester’s law of inertia). If f is a quadratic form on a vector

space V over real numbers of finite dimension, then there exist unique integers
a, b, and c and a basis B of V such that [f ]B = D(a, b, c).

Proof. Let A be a matrix of f with respect to any basis C of V. We need to

find a regular matrix S (which will serve as the transition matrix from basis
B to C) such that S T AS = D(a, b, c) for some a, b, and c. Perform Gaussian
elimination, applying the same operations to the rows and columns of A,
until we obtain a diagonal matrix with only +1, −1, or 0 on the diagonal.
For the uniqueness, suppose that [f ]B1 = D(a1 , b1 , c1 ) and [f ]B2 = D(a2 , b2 , c2 )
for some bases B1 and B2 . Let S = [id]B1 ,B2 ; hence, D(a1 , b1 , c1 ) = S T D(a2 , b2 , c2 )S.
Since S is regular, we have rank(D(a1 , b1 , c1 )) = rank(D(a2 , b2 , c2 )), and thus
c1 = c2 . Also, by Lemma 5, we have a1 = a2 , and thus b1 = b2 .

Definition 4. We say that a quadratic form f on a vector space V of finite

dimension has signature (a, b, c) if there exists a basis B such that [f ]B =
D(a, b, c).
 
−1 1 −3 −1
1 7 −1 5 
Example 3. Let A =  −3 −1 −1 1 . By performing simultaneous


−1 5 1 7

4
row and column operations, we have
   
−1 1 −3 −1 −1 0 −3 −1
0 8 −4 4  → 0
 8 −4 4 
A→ −3 −1 −1 1  −3 −4 −1 1 


−1 5 1 7 −1 4 1 7
   
−1 0 −3 −1 −1 0 0 −1
0 8 −4 4  → 0
 8 −4 4 
→ 0 −4 8

4   0 −4 8 4
−1 4 1 7 −1 4 4 7
   
−1 0 0 −1 −1 0 0 0
0 8 −4 4  → 0
 8 −4 4
→ 0 −4 8  0 −4 8 4

4
0 4 4 8 0 4 4 8
   
−1 0 0 0 −1 0 0 0
0 4 −2 2 0 2 −2 2
→ 0 −4 8 4 →  0 −2 8 4
  

0 4 4 8 0 2 4 8
   
−1 0 0 0 −1 0 0 0
 0 2 −2 2  0 2 0 2
→ 0 0 6 6 →  0 0 6 6
  

0 2 4 8 0 2 6 8
   
−1 0 0 0 −1 0 0 0
 0 2 0 2  0 2 0 0
→ 0 0 6 6 →  0 0 6 6
  

0 0 6 6 0 0 6 6
   
−1 0 0 0 −1 0 0 0
 0 2 0 0
 →  0 2 0 0
 
→ 0 0 6 6  0 0 6 0
0 0 0 0 0 0 0 0
−1 √0
   
0 0 −1 0 0 0
0 2 √0 0  →  0 1 0 0
 
→0 0 6 0   0 0 1 0
0 0 0 0 0 0 0 0
   
0 0 1 0 1 0 0 0
 0 1 0 0
 → 0 1 0 0 .
 
→−1 0 0 0 0 0 −1 0
0 0 0 0 0 0 0 0
Hence, the signature of A is (2, 1, 1).

5
Corollary 7. Let A be a real symmetric n×n matrix and let f (x) = xT Ax for
x ∈ Rn . If f has signature (a, b, c), then the sum of algebraic multiplicities
of the positive eigenvalues of A is a, and the sum of algebraic multiplicities
of the negative eigenvalues of A is b.
Proof. Recall that since A is real and symmetric, we have A = Q−1 DQ =
QT DQ for a diagonal matrix D and an orthogonal matrix Q, where A and D
have the same eigenvalues, equal to the diagonal entries of D. Equivalently,
we have [f ]B1 = D, where B1 is the basis formed by the columns of QT .
Since the signature of A is (a, b, c), there exists a basis B2 such that [f ]B2 =
D(a, b, c). The claim follows by Lemma 5.

3 Quadrics and conics

Definition 5. For any n × n symmetric real matrix A, a real row vector
b and a real number γ, the set {x ∈ Rn : xT Ax + bx + γ = 0} is called a
quadric. If n = 2, it is called a conic.
Example 4. The set {(x, y, z) ∈ R3 : x2 − y 2 − z + 1 = 0}:

6
 Consider a quadric C = {x ∈ Rn : xT Ax + bx + γ = 0}. We have
A = QT DQ for an orthogonal matrix Q and a diagonal matrix D. Let
b0 = bQT and C 0 = {y ∈ Rn : y T Dy + b0 y + γ = 0}. Observe that x ∈ C
if and only if Qx ∈ C 0 , and thus thus the sets C 0 and C only differ by the
isometry described by Q. Let (p, n, z) be the signature of A; without loss
of generality, the first p entries of D are positive, the next n entries are
negative and the last z are zeros. Furthermore, since C 0 is also equal to
{y ∈ Rn : y T (−D)y − b0 y − γ = 0}, we can assume that p ≥ n.
For any vector d, let bd = 2dT D + b0 , γd = dT Dd + b0 d + γ and Cd = {v ∈
Rn : v T Dv + bd v + γd }. Note that v ∈ Cd if and only if v + d ∈ C 0 , and
thus Cd is obtained from C 0 by shifting it by the vector d (another isometry).
Choose the first p + n coordinates of d so that the first p + n coordinates of
the vector 2dT D are equal to the first p + n coordinates of −b0 (the remaining
coordinates of 2dT D are always 0). Furthermore, if z > 0 and at least one of
the last z entries of b0 is not 0, we can choose the last z coordinates of d so
that γd is 0.
Thus, we get the following.

Lemma 8. Every quadric is up to isometry equal to a quadric {x ∈ Rn :

xT Ax + bx + γ = 0} satisfying the following conditions:

• A is diagonal with the first p entries positive, the next n entries negative
and the last z equal to 0 for some p ≥ n,

• the first p + n entries of b are equal to 0, and

• either b = o or γ = 0.

Example 5. Classification of conics:

p = 2: α1 x21 + α2 x22 = γ for α1 , α2 > 0.

• empty if γ < 0,
• the point (0, 0) if γ = 0,
p p
• an ellipse with axes γ/α1 and γ/α2 if γ > 0.

7
 y
3

2.5

1.5

0.5

x
−3 −2.5 −2 −1.5 −1 −0.5 0.5 1 1.5 2 2.5 3

−0.5

−1

−1.5

−2

−2.5

−3

p = n = 1: α1 x21 − α2 x22 = γ for α1 , α2 > 0.

p
• Two intersecting lines |x1 | = α2 /α1 |x2 | if γ = 0.
• Hyperbola if γ 6= 0.
y
3

2.5

1.5

0.5

x
−3 −2.5 −2 −1.5 −1 −0.5 0.5 1 1.5 2 2.5 3

−0.5

−1

−1.5

−2

−2.5

−3

p = z = 1, b = o: αx21 = γ for α > 0.

• Empty, a line, or two parallel lines depending on γ.

p = z = 1, b 6= o, γ = 0: αx21 = βx2 for α, β 6= 0.

• A parabola.

8
 y
3

2.5

1.5

0.5

x
−3 −2.5 −2 −1.5 −1 −0.5 0.5 1 1.5 2 2.5 3

−0.5

z = 2, b = o: γ = 0.

• Empty or R2 depending on γ.

z = 2, b 6= o, γ = 0: β1 x1 + β2 x2 = 0.

• A line.

Similarly, we can classify quadrics in higher dimensions, based on their

signatures.