QUADRATIC FORMS AND DEFINITE MATRICES

1. D EFINITION AND CLASSIFICATION OF QUADRATIC FORMS

1.1. Definition of a quadratic form. Let A denote an n x n symmetric matrix with real entries and
let x denote an n x 1 column vector. Then Q = x’Ax is said to be a quadratic form. Note that
 
a11 · · · a1n  
 .. ..  x1
Q = x´Ax = (x1 ...xn)  .. 
xn
a
 P n1 · · · a 
nn
a1ixi
 .
. 
= (x1, x2, · · · , xn)  
P.
anixi
(1)
= a11x21 + a12x1x2 + ... + a1nx1 xn
+ a21x2 x1 + a22x22 + ... + a2nx2 xn
+ ...
+ ...
+ ...
2
+ aPn1 xn x1 + an2 xn x2 + ... + ann xn
= i ≤ j aij xi xj

For example, consider the matrix

 
1 2
A=
2 1
and the vector x. Q is given by

  
0 1 2 x1
Q = x Ax = [x1 x2 ]
2 1 x2
x1
= [x1 + 2 x2 2 x1 + x2 ]
x2
= x21 + 2 x1 x2 + 2 x1 x2 + x22
= x21 + 4 x1 x2 + x22

1.2. Classification of the quadratic form Q = x0Ax: A quadratic form is said to be:
a: negative definite: Q < 0 when x 6= 0
b: negative semidefinite: Q ≤ 0 for all x and Q = 0 for some x 6= 0
c: positive definite: Q > 0 when x 6= 0
d: positive semidefinite: Q ≥ 0 for all x and Q = 0 for some x 6= 0
e: indefinite: Q > 0 for some x and Q < 0 for some other x

Date: September 14, 2004.

1
2 QUADRATIC FORMS AND DEFINITE MATRICES

Consider as an example the 3x3 diagonal matrix D below and a general 3 element vector x.
 
1 0 0
D= 0 2 0 
0 0 4
The general quadratic form is given by
   
1 0 0 x1
Q = x0A x = [x1 x2 x3]  0 2 0   x2 
  0 0 4 x3
x1
= [x1 2 x2 4 x3 ]  x2 
x3
= x21 + 2 x22 + 4 x23

Note that for any real vector x 6= 0, that Q will be positive, because the square of any number
is positive, the coefficients of the squared terms are positive and the sum of positive numbers is
always positive. Also consider the following matrix.
 
−2 1 0
E =  1 −2 0 
0 0 −2
The general quadratic form is given by
   
−2 1 0 x1
Q = x0A x = [x1 x2 x3 ]  1 −2 0   x2 
0 0 −2  x3
x1
= [−2 x1 + x2 x1 − 2 x2 − 2 x3]  x2 
x3
= −2 x21 + x1 x2 + x1 x2 − 2 x22 − 2 x23
= −2 x21 + 2 x1 x2 − 2 x22 − 2 x23
= −2 [x21 − x1 x2] − 2 x22 − 2 x23
= −2 x21 − 2[x22 − x1 x2] − 2 x23

Note that independent of the value of x3 , this will be negative if x1 and x2 are of opposite sign
or equal to one another. Now consider the case where |x1| > |x2|. Write Q as

Q = − 2 x21 + 2 x1 x2 − 2 x22 − 2 x23

The first, third, and fourth terms are clearly negative. But with | x1| > | x2 |, | 2 x21 | > | 2 x1 x2 |
so that the first term is more negative than the second term is positive, and so the whole expression
is negative. Now consider the case where |x1| < |x2|. Write Q as

Q = − 2 x21 + 2 x1 x2 − 2 x22 − 2 x23

The first, third, and fourth terms are clearly negative. But with | x1 | < | x2 | , | 2 x22 | > | 2 x1 x2 |
so that the third term is more negative than the second term is positive, and so the whole expression
is negative. Thus this quadratic form is negative definite for any and all real values of x 6= 0.
 QUADRATIC FORMS AND DEFINITE MATRICES 3

1.3. Graphical analysis. When x has only two elements, we can graphically represent Q in 3 di-
mensions. A positive definite quadratic form will always be positive except at the point where x
= 0. This gives a nice graphical representation where the plane at x = 0 bounds the function from
below. Figure 1 shows a positive definite quadratic form.

F IGURE 1. Positive Definite Quadratic Form 3x21 + 3x22

10
x2 5
0
-5
10
-10
600

400
Q
200

0
-10
-5
0
x1 5
10
Similarly, a negative definite quadratic form is bounded above by the plane x = 0. Figure 2 shows
a negative definite quadratic form.
4 QUADRATIC FORMS AND DEFINITE MATRICES

F IGURE 2. Negative Definite Quadratic Form −2x21 − 2x22

10
x2 5
0
-5
10
-10
0
-100
Q
-200
-300
-400
-10
-5
0
x1 5
10

A positive semi-definite quadratic form is bounded below by the plane x = 0 but will touch the
plane at more than the single point (0,0), it will touch the plane along a line. Figure 3 shows a
positive semi-definite quadratic form.
A negative semi-definite quadratic form is bounded above by the plane x = 0 but will touch the
plane at more than the single point (0,0). It will touch the plane along a line. Figure 4 shows a
negative-definite quadratic form.
An indefinite quadratic form will not lie completely above or below the plane but will lie above
for some values of x and below for other values of x. Figure 5 shows an indefinite quadratic form.

1.4. Note on symmetry. The matrix associated with a quadratic form B need not be symmetric.
However, no loss of generality is obtained by assuming B is symmetric. We can always take definite
and semidefinite matrices to be symmetric since they are defined by a quadratic form. Specifically
consider a nonsymmetric matrix B and define A as 12 (B + B 0 ), A is now symmetric and x0 Ax =
x0Bx.

2. D EFINITE AND SEMIDEFINITE MATRICES

2.1. Definitions of definite and semi-definite matrices. Let A be a square matrix of order n and
let x be an n element vector. Then A is said to be positive semidefinite iff for all vectors x
 QUADRATIC FORMS AND DEFINITE MATRICES 5

F IGURE 3. Positive Semi-Definite Quadratic Form 2x21 + 4x1x2 + 2x22

x2
5 2.5
0
-2.5
-5

100

75
Q
50

25

0
-5 0 5

x1

F IGURE 4. Negative Semi-Definite Quadratic Form −2x21 + 4x1 x2 − 2x22

x2
5
2.5
0
-2.5
-5
0

-25

-50
Q

-75

-100

0 5
-5

x1

x0Ax ≥ 0 (2)
The matrix A is said to be positive definite if for non zero x

x0 Ax > 0 (3)
6 QUADRATIC FORMS AND DEFINITE MATRICES

F IGURE 5. Indefinite Quadratic Form −2x21 + 4x1x2 + 2x22

x2
5
2.5
0
-2.5
-5

50

0
Q

-50

-5 0 5

x1

Let A be a square matrix of order n. Then A is said to be negative (semi)definite iff -A is positive
(semi)definite.

2.2. Diagonal elements of positive definite matrices.

Theorem 1. Let A be a positive definite matrix of order m. Then

aii > 0, i = 1, 2, ..., m.

If A is only positive semidefinite then

aii ≥ 0, i = 1, 2, ..., m.

Proof. Let e·i be the m-element vector all of whose elements are zeros save the ith, which is unity.
For example if m = 5 and i = 2 then e. 2 = [0, 1, 0, 0, 0 ] If A is positive definite, because e·i is not
the null vector, we must have

e0·i Ae·i > 0, i = 1, 2, ..., m. (4)

But

e0·i Ae·i = aii, i = 1, 2, ..., m. (5)

If A is positive semidefinite but not positive definite then repeating the argument above we find

aii = e0·i Ae·i ≥ 0, i = 1, 2, ..., m. (6)

 QUADRATIC FORMS AND DEFINITE MATRICES 7

2.3. Factoring positive definite matrices (Cholesky factorization).

Theorem 2. Let A be a positive definite matrix of order n. Then there exists a lower triangular matrix T
such that

A = TT0 (7)

Proof. Define T as follows

 
t11 0 0 ··· 0
 t21 t22 0 ··· 0 
 
 t31 t32 t33 ··· 0 
T =   (8)
 .. .. .. .. .. 
 . . . . . 
tn1 tn2 tn3 ··· tnn
0
Now define T T

   
t11 0 0 ··· 0 t11 t21 t31 ··· tn1
 t21 t22 0 ··· 0   0 t22 t32 ··· tn2 
   
 t31 t32 t33 ··· 0   0 0 t33 ··· tn3 
T T´ =    
 .. .. .. .. ..   .. .. .. .. .. 
 . . . . .   . . . . . 
tn1 tn2 tn3 ··· tnn 0 0 0 ··· tnn
 
t211 t11t21 t11 t31 ··· t11tn1
 t21t11 t221 + t222 t21t31 + t22t32 ··· t21tn1 + t22tn2 
 
 t231 + t232 + t233 
=  t31t11 t31t21 + t32t22 ··· t31tn1 + t32tn2 + t33tn3  (9)
 .. .. .. .. .. 
 . . . . . 
tn1t11 tn1t21 + tn2t22 tn1 t31 + tn2 t32 + tn3 t33 ··· Σn 2
i=1 tni

Now define A = T T 0 and compare like elements

A =T T´ 
a11 a12 a13 ··· a1n
 a21 a22 a23 ··· a2n 
 
 
⇒  a31 a32 a33 ··· a3n  =
 .. .. .. .. .. 
 . . . . . 
 an12 an2 an3 · · · ann  (10)
t11 t11t21 t11t31 ··· t11tn1
 t21 t11 t 2 2
21 + t22 t21t31 + t22t32 ··· t21tn1 + t22tn2 
 
 t31 t11 t31t21 + t32t22 t231 + t232 + t233 ··· t31tn1 + t32tn2 + t33tn3 
 
 .. .. .. .. .. 
 . . . . . 
tn1 t11 tn1t21 + tn2t22 tn1t31 + tn2t32 + tn3t33 ··· Σn 2
i=1 tni

Solve the system now for each tij as functions of the aij . The system is obviously recursive
because we can solve first for t11, then t21, etc. A schematic algorithm is given below.
8 QUADRATIC FORMS AND DEFINITE MATRICES

√ a12 a13 a1n

t11 = ± a11 , t21 = , t31 = , · · · , tn1 =
t11 t11 t11
s
a122 a23 − t21 t31 a2n − t21 tn1
t22 = ± a22 − , t32 = , · · · , tn2 =
a11 t22 t22
s
q  2
2 2 a213 a23 − t21 t31
t33 = ± a33 − t31 − t32 = ± a33 − −
a11 t22
a34 − t31 t41 − t32 t42 a3n − t31 tn1 − t32 tn2
t43 = , · · · , tn3 = (11)
t33 t33
This matrix is not unique because the square roots involve two roots. The standard procedure is
to make the diagonal elements positive. Consider the following matrix as an example
 
4 2 0
F =  2 9 0 .
0 0 2
We can factor it into the following matrix T
 
2 0
√ 0
T =  1 2 2 √0 
0 0 2
0 0
and its transpose T . Then T T = F .
     
2 √0 0 2 1
√ 0 4 2 0
T T 0 =  1 2 2 √0   0 2 2 √0  =  2 9 0 
0 0 2 0 0 2 0 0 2
2.4. Characteristic roots of positive definite matrices.
Theorem 3. Let A be a symmetric matrix of order n. Let λi , i=1,...,n be its characteristic roots. Then if A is
positive definite, λi > 0, for all i.
Proof. Because A is symmetric, choose an orthonormal set of eigenvectors Q. Clearly Q0 AQ = Λ,
where Λ is a diagonal matrix with the eigenvalues of A on the diagonal. Now consider any one of
the rows of Q0. This is one of the eigenvectors of A. Denote it by qi0 . Then clearly

qi0 Aqi = λi > 0 (12)

2.5. Nonsingularity of positive definite matrices.
Theorem 4. Let A be a symmetric matrix of order n. If A is positive definite then r(A) = n.
Proof. Because A is symmetric, choose an orthonormal set of eigenvectors Q. Clearly Q0 AQ = Λ,
where Λ is a diagonal matrix with the eigenvalues of A on the diagonal. Because Q is orthogonal
its inverse is its transpose and we also obtain that AQ = QΛ. Now because A is positive definite,
all the characteristic roots on the diagonal of Λ are positive. Thus the inverse of Λ is just a matrix
with the reciprocal of each characteristic root on its diagonal. Thus Λ is invertible. Because QΛ
is the product of two invertible matrices, it is invertible. Thus AQ is invertible, and because Q is
invertible, this means A is invertible and of full rank. See Dhrymes [1, Proposition 2.61] or Horn
and Johnson [4].
 QUADRATIC FORMS AND DEFINITE MATRICES 9

Theorem 5. If A in the above theorem is merely positive semidefinite then r(A) < n.
Proof. Because A is positive semidefinite, we know that λi ≥ 0, i = 1,2, ..., n. The proof is based on
showing that at least one of the roots is zero. We can diagonalize A as

Q0AQ = Λ. (13)
Consequently, for any vector y,

y0 Q0AQy = Σn 2
i=1 λi yi . (14)
Now, if x is any nonnull vector, by the semidefiniteness of A we have

0 ≤ x0Ax = x0QQ0 AQQ0x = x0QΛQ0x = Σn 2

i=1 λi yi , (15)
where now we set

y = Q0 x. (16)
Because x is nonnull then y is also nonnull, because Q is orthogonal and thus non-singular.

If none of the λi is zero, 15 implies that for any nonnull x

x0Ax > 0, (17)

thus showing A to be positive definite. Consequently, at least one of the λi , i = 1, 2,...,n, must be
zero and there must exist at least one nonnull x such that

x0 Ax = Σn 2
i=1 λi yi = 0. (18)
But this shows that

r(A) < n. (19)

2.6. Factoring symmetric positive definite matrices.
Theorem 6. Let A be a symmetric matrix, of order m. Then A is positive definite if and only if there exists
a matrix S of dimension n x m and rank m (n ≥ m) such that

A = S 0 S.
It is positive semidefinite if and only if

r(S) < m.
Proof. (Dhrymes [1, Proposition 2.61] or Horn and Johnson [4]). If A is positive (semi)definite then,
as in the proof of theorem 56, we have the representation

A = QΛQ0.
Here Q is an orthonormal set of eigenvectors and Λ is a diagonal matrix with the eigenvalues of
A on the diagonal. Taking

S = Λ1/2Q0
we have
10 QUADRATIC FORMS AND DEFINITE MATRICES

A = S 0 S.
If A is positive definite, Λ is nonsingular and thus

r(S) = m.
If A is merely positive semidefinite then r(Λ) < m and hence

r(S) < m.
This proves the first part of the theorem.

On the other hand suppose

A = S0S (20)
and S is n x m matrix (n ≥ m) of rank m. Let x be any nonnull vector and note

x0 Ax = xS 0 Sx. (21)
The right side of the equation above is a sum of squares and thus is zero if and only if

Sx = 0. (22)
If the rank of S is m, equation 22 can be satisfied only with null x. Hence A is positive definite.
So for any x

x0 Ax = x0S 0 Sx ≥ 0,
and if S is of rank less than m, there exists at least one nonnull x such that

Sx = 0.
Consequently, there exists at least one nonnull x such that

x0 Ax = 0.
which shows that A is positive semidefinite but not positive definite.

2.7. Using naturally ordered principal minors to test for positive definiteness.

2.7.1. Definition of naturally ordered (leading) principal minors. The naturally ordered principle mi-
nors of a matrix A are defined as determinants of the matrices
 
a11 a12 ··· a1k
 a21 a22 ··· a2k 
 
 .. .. ..  k = 1, 2, ..., n. (23)
 . . . 
ak1 ak2 ··· akk
A principal minor is the minor of a principal submatrix of A where a principal submatrix is a
matrix formed from a square matrix A by taking a subset consisting of n rows and column elements
from the same numbered columns. The natural ordering considers only those principal minors that
fall along the main diagonal. Specifically for a matrix A, the naturally ordered principal minors are
 QUADRATIC FORMS AND DEFINITE MATRICES 11

 
   a11 a12 ··· a1n 
   a11   
 a   a12 a13   a21 a22 ··· a2n 
a12  
a11 ,  11  ,  a21
  a22 a23  , ···
  .. .. ..  (24)
a21 a22  a31   . . . 
a32 a33  
 an1 an2 ··· ann 
or schematically

F IGURE 6. Naturally Ordered Principle Minors of a Matrix

Theorem 7. Let A be a symmetric matrix of order m. Then A is positive definite iff its naturally ordered
principal minors are all positive.
For a proof, see Gantmacher [2, p. 306] or Hadley [3, p. 260-262].

As an example consider the matrix G1.

12 QUADRATIC FORMS AND DEFINITE MATRICES

 
4 2 0
G1 =  2 9 0 .
0 0 2
Element a11 = 4 > 0. Now consider the first naturally occurring principal 2x2 submatrix
 
 4 2 
 = 36 − 4 = 32 > 0.
 2 9 
Now consider the determinant of the entire matrix
 
 4 2 0 
 
 2 9 0  = (4)(9)(2) + (2) (0) (0) + (2) (0) (0) − (9) (0) (0) − (2) (2) (2) − (4) (0) (0) =
 
 0 0 2 
= 72 + 0 + 0 − 0 − 8 − 0 = 64 > 0
This matrix is then positive definite.

2.8. Characteristic roots of positive semi-definite matrices. Let A be a symmetric matrix of order
n. Let λi , i=1,...,n be its characteristic roots. If A is positive semi-definite then

λi ≥ 0 i = 1, 2, ..., n and at least one λi = 0.

2.9. Using principal minors to test for positive definiteness and positive semidefiniteness.

2.9.1. Definition of principal minors. A principal minor of order r is defined as the determinant of a
principal submatrix. A principal submatrix is defined as follows. If A is a matrix of order n, and
we wipe out r of the rows and the corresponding r columns as well, the resulting (n-r) x (n-r)
submatrix is called a principal submatrix of A. The determinant of this matrix is called a principal
minor of A. Another way to write a principal minor of order p is
 
 ai1 i1 , ai1 i2 , · · · ai1 ip 
   
 ai2 i1 , ai2 i2 , · · · ai2 ip 
i1 , i2 , i3 , ..., ip  
A =  ..  (25)
i1 , i2 , i3 , ..., ip  . 
 
 ai i , ai i ··· aip ip 
p 1 p 2

Note: A matrix of order m has

 
m
p
principal minors of order p where
 
m m!
= .
p p! (m − p)!
For example a 4x4 matrix has 1 principal minor of order 4, (the matrix itself), 4 principal minors
of order 3, 6 principle minors of order 2, and 4 principle minors of order 1 for a total of 15 principal
minors.
 QUADRATIC FORMS AND DEFINITE MATRICES 13

As an example, consider the following matrix of order 3 which has 1 principal minor of order 3,
3 principal minors of order 2,
 
3 3! (3) (2) 6
= = = = 3
2 2! (3 − 2)! (1) (2) (1) (1) 2
and 3 principle minors of order 1 for a total of 7 principal minors.
 
4 2 0
G = 2 9 0 .
0 0 2
 
 4 2 0 
 
Order 3  2 9 0 .

 0 0 2 
     
 4 2   4 0   9 0 
Order 2  


 0 2



 0 2


2 9

Order 1 4 2 9

Now consider the general 4x4 matrix A and some of its principal minors
 
a11 a12 a13 a14  
 a11 a12 a14 
 a21 a22 a23 a24 
 
A =  a31 a32 a33 a34 
 A 1 2 4 =  a21 a22 a24
124 


 a41 a42 a44 
a41 a42 a43 a44 (26)
   

 a a13 
 a a24 

A 13
=  11 A 24
=  22 A 3
= a33
13 a31 a33  24 a42 a44  3

2.9.2. A test for for positive definiteness and positive semidefiniteness using principle minors.
Theorem 8. A matrix A is positive semidefinite iff all the principal minors of A are non-negative.
For a proof see Gantmacher [2, p. 307].

As an example consider the matrix G2.

 
2 0 2
G2 =  0 4 4  .
2 4 6
The diagonal elements are all positive so the 1x1 test is passed. Now consider the principal 2x2
minors
     
 2 0   2 2   4 4 
  =8 − 0 = 8 > 0  = 12 − 4 = 8 > 0   = 24 − 16 = 8 > 0
 0 4   2 6   4 6 

These are all positive and so we pass the 2x2 test. Now consider the determinant of the entire
matrix
14 QUADRATIC FORMS AND DEFINITE MATRICES

 
 2 0 2 
 
 0 4 4  = (2)(4)(6) + (0) (4) (2) + (2) (4) (0) − (2) (4) (2) − (4) (4) (2) − (6) (0) (0) =
 
 2 4 6 
= 48 + 0 + 0 − 16 − 32 − 0 = 0
This determinant is zero and so the matrix is positive semidefinite but not positive definite.
2.9.3. Some characteristics of negative semidefinite matrices. The results on positive definite and posi-
tive semidefinite matrices have counterparts for negative definite and semidefinite matrices.
a: A negative semidefinite matrix is negative definite only if it is non-singular.
b: Let A be a negative definite matrix of order m. Then
aii < 0 i = 1, ..., m
c: If A is only negative semidefinite then

aii ≤ 0 i = 1, ..., m.
d: Let A be a symmetric matrix of order m. Then A is negative definite iff its naturally or-
dered (leading)principal minors alternate in sign starting with a negative number.
The naturally ordered principle minors of a matrix A are defined as determinants of
matrices
 
a11 a12 · · · a1k
 a21 a22 · · · a2k 
 
 .. .. ..  k = 1, 2, ..., m. (27)
 . . . 
ak1 ak2 · · · akk
As an example consider the matrix
 
−2 1 0
E =  1 −2 0 
0 0 −2
Element a11 = -2 < 0. Now consider the first naturally occurring principal 2x2 submatrix
 
 −2 1 
 
 1 −2  = 4 − 1 = 3 > 0.
Now consider the determinant of the entire matrix
 
 −2 1 0 
 
 1 −2 0  = (−2)(−2)(−2) + (1) (0) (0) + (1) (0) (0) − (−2) (0) (0) − (−2) (0) (0) − (−2) (1) (1)
 
 0 0 −2 
= −8 + 0 + 0 − 0 − 0 + 2 = −6 < 0
This matrix is then negative definite.
e: Let A be a symmetric matrix of order m. A is negative semidefinite iff the following in-
equalities hold:
 
i , i2 , · · · , ip
(−1)p A 1 ≥0 (28)
i 1 , i2 , · · · , ip
[1 ≤ i1 ≤ i2 ≤, · · · , ≤ ip ≤ m, p = 1, 2, ..., m]
 QUADRATIC FORMS AND DEFINITE MATRICES 15

Here A ( ) is the determinant of the submatrix of A with p rows and columns of A, i.e.,
it is a principle minor of A.
 
 ai1 i1 ai1 i2 ··· 
   
i1 , i2 , ..., ip  ai2 i1 ai2 i2 ··· 
 
A =  . ..  (29)
i1 , i2 , ..., ip  .. . 
 
 ai i aip i2 ··· 
p 1

For a 3x3 matrix this means that all the diagonal elements are non-positive, all 2x2 prin-
cipal minors are non-negative and the determinant of the matrix is non-positive. Consider
as an example the matrix G3
 
−2 1 −1
G3 =  1 −2 −1  .
−1 −1 −2
The diagonal elements are all negative so the 1x1 test is passed. Now consider the
principal 2x2 minors

     
 −2 1   −2 −1   
 = 4 − 1 = 3 > 0   = 4 − 1 = 3 > 0  −2 −1  = 4−1 = 3 > 0 .
 1 −2  −1 −2   −1 −2 

These are all positive and so we pass the 2x2 test. Now consider the determinant of the
entire matrix
 
 −2 1 −1 

|G3| =  1 −2 −1 
 −1 −1 −2 

= [(−2)(−2)(−2)] + [(1) (−1) (−1)] + [(1) (−1) (−1)]

− [(−1) (−2) (−1)] − [(−1) (−1) (−2)] − [(1) (1) (−2)]

= −8 + 1 + 1 − (−2) − (−2) − (−2) = 0

This determinant is zero and so the matrix is negative semidefinite but not negative
definite.
f: Characteristic roots of negative definite matrices.
Let A be a symmetric matrix of order m and let λi , i = 1,...,m be its real characteristic
roots. If A is negative definite then

λi < 0 i = 1, 2, ..., m
g: Characteristic roots of negative semi-definite matrices.
Let A be a symmetric matrix of order m and let λi , i = 1,...,m be its real characteristic
roots. If A is negative semi-definite then

λi ≤ 0 i = 1, 2, ..., m and at least one λi = 0.

16 QUADRATIC FORMS AND DEFINITE MATRICES

2.10. Example problems. Determine whether the following matrices are positive definite, positive
semidefinite,
 negative definite,
 negative semidefinite, or indefinite.
−2 0 −1
A =  0 −2 −1 
−2 −4 −3

 
−2 4 −1
B=  4 −2 −1 
−1 −1 −2

 
2 1 −1
C=  1 4 −2 
−1 −2 4

 
2 −1 3
D =  −1 5 3 
3 3 9

 
−2 1 −1
E=  1 −3 −2 
−1 −2 −5

3. S ECOND ORDER CONDITIONS FOR OPTIMIZATION PROBLEMS AND DEFINITENESS CONDITIONS

ON MATRICES

3.1. Restatement of second order conditions for optimization problems with 2 variables.

Theorem 9. Suppose that f(x1, x2) and its first and second partial derivatives are continuous throughout a
∂f ∂f
disk centered at (a, b) and that ∂x 1
(a, b) = ∂x 2
(a, b) = 0 . Then
2 2 2
h 2 i2
a: f has a local maximum at (a, b) if ∂∂xf2 (a, b) < 0 and ∂∂xf2 ∂∂xf2 − ∂x∂1 ∂x
f
2
> 0 at (a, b).
1 1 2
We can also write this as f11 < 0 and f11 f22 − f122 > 0 at (a, b).
h i2
∂ 2f ∂2f ∂2f ∂2f
b: f has a local minimum at (a, b) if ∂x21
(a, b) > 0 and ∂x21 ∂x22
− ∂x1 ∂x2 > 0 at (a, b).
We can also write this as f11 > 0 and f11 f22 − f122 > 0 at (a, b).
2 2
h 2 i2
c: f has a saddle point at (a, b) if ∂∂xf2 ∂∂xf2 − ∂x∂1 ∂x
f
2
< 0 at (a, b). We can also write this as
1 2
f11 f22 − f122 < 0 at (a, b).
2 2
h 2 i2
d: The test is inconclusive at (a, b) if ∂∂xf2 ∂∂xf2 − ∂x∂1 ∂x
f
2
= 0 at (a, b). If this case we must
1 2
find some other way to determine the behavior of f at (a, b).
2 2
h 2 i2
The expression ∂∂xf2 ∂∂xf2 − ∂x∂1 ∂x
f
2
is called the discriminant of f.
1 2

3.2. Expressing the second order conditions in terms of the definiteness of the Hessian of the
objective function.
 QUADRATIC FORMS AND DEFINITE MATRICES 17

3.2.1. Second order conditions for a local maximum. The Hessian of a function f is the nxn matrix of
second order partial derivatives, that is
 
 f11 f12 · · · f1n 
 
 f21 f22 · · · f2n 
 
|H| =  . .. .. ..  (30)
 .. . . . 
 
 fn1 fn2 · · · fnn 
We can write the discriminant condition as the determinant of the Hessian of the objective func-
tion f when there are just 2 variables in the function as
 2  
∂2f ∂2f ∂2f  f11 f12 
− =   = |H| (31)
∂x21 ∂x22 ∂x1 ∂x2 f21 f22 
The second order condition for a local maximum is then that
 
 f11 f12 
f11 < 0 and    > 0
f21 f22 
which is just the condition that H is negative definite.

3.2.2. Second order conditions for a local minimum. The second order condition for a local minimum
is that
 
 f f12 
f11 > 0 and  11 > 0
f21 f22 
which is just the condition that H is positive definite.

3.2.3. Extension of condition on Hessian to more than two variables. The second order conditions, for
local maxima and minima based on the sign of f11 and the discriminant written in terms of whether
the Hessian of the objective function is positive or negative, extend to problems involving objective
functions with more than 2 variables.

4. C ONVEXITY AND CONCAVITY AND DEFINITENESS CONDITIONS ON H ESSIAN MATRICES

4.1. Definition of concavity. Let S be a nonempty convex set in Rn.. The function f: S → R1 is said
to be concave on S if f(λx1 + (1-λ)x2) ≥ λf(x1 ) + (1-λ)(x2) for each x1 , x2 ∈ S and for each λ ∈ [0, 1].
The function f is said to be strictly concave if the above inequality holds as a strict inequality for
each distinct x1 x2 ∈ S and for each λ ∈ (0, 1).

4.2. Characterizations of concave functions.

a: The function f is continuous on the interior of S.
b: The function f is concave on S if and only if the set {(x, y): x ∈ S, y ≤ f(x)} is convex.
c: The set {x ∈ S, f(x) ≥ α } is convex for every real α. This set is called the hypograph of f. It
is a subset of R2. Thus concavity of f is equivalent to convexity of its hypograph.
d: A differentiable function f is concave on S if and only if

f ( x) ≤ f (x̄) + f´(x̄) (x − x̄) for each distinct x, x̄ ∈ S.

This implies that tangent line is above the graph.
e: A twice differentiable function f is concave iff the Hessian H(x) is negative semidefinite for
each x ∈ S.
18 QUADRATIC FORMS AND DEFINITE MATRICES

f: Let f be twice differentiable. Then if the Hessian H(x) is negative definite for each x ∈
S, f is strictly concave. Further if f is strictly concave, then the Hessian H(x) is negative
semidefinite for each x ∈ S.
g: Every local maximum of f over a convex set W ⊆ S is a global maximum.
h: If f 0 (x̄) = 0 for a concave function then, x̄ is the global maximum of f over S.

4.3. Definition of convexity. Let S be a nonempty convex set in Rn. . The function f: S → R1 is said
to be convex on S if f ( λ x1 + (1 − λ) x2) ≤ λ f (x1 ) + (1 − λ )f ( x2 ) for each x1 , x2 ∈ S and for
each λ ∈ [0, 1]. The function f is said to be strictly convex if the above inequality holds as a strict
inequality for each distinct x1 , x2 , ∈ S and for each λ ∈ (0, 1).

4.4. Characteristics of convex functions.

a: The function f is continuous on the interior of S.
b: The function f is convex on S if and only if the set {(x, y): x ∈ S, y ≥ f(x)} is convex.
c: The set {x ∈ S, f(x) ≤ α } is convex for every real α. This set is called the epigraph of f. It is
a subset of R2. Thus concavity of f is equivalent to convexity of its epigraph.
d: A differentiable function f is convex on S if and only if

f(x) ≥ f (x̄ ) + f 0 (x̄) (x − x̄) for each distinct x, x̄ ∈ S.

This implies that tangent line is below the graph.
e: A twice differentiable function f is convex iff the Hessian H(x) is positive semidefinite for
each x ∈ S.
f: Let f be twice differentiable. Then if the Hessian H(x) is positive definite for each x ∈
S, f is strictly concave. Further if f is strictly concave, then the Hessian H(x) is positive
semidefinite for each x ∈ S.
g: Every local minimum of f over a convex set W ⊆ S is a global minimum.
h: If f 0 (x̄) = 0 for a convex function then, x̄ is the global minimum of f over S.

5. L INEAR CONSTRAINTS AND BORDERED MATRICES

5.1. Definition of a quadratic form with linear constraints. Let the quadratic form be given by
 
a11 · · · a1n  
 .. ..  x1
Q = x´Ax = (x1...xn)  ..  (32)
xn
an1 · · · ann

with a set of m linear constraints represented by

x0 B = 0
   
b11 b12 ··· b1m 0
 b21 b22 ··· b2m   0  (33)
   
(x1 x2 · · · xn)  .. ..  =  .. 
 . .   . 
bn1 bn2 ··· bnm 0
 QUADRATIC FORMS AND DEFINITE MATRICES 19

5.2. Graphical analysis. Consider the indefinite matrix A given by

 
−2 2
A = (34)
2 2

The quadratic form is given by

 
 
0 −2 2x1
Q = x Ax = [x1 x2]
2 2
 x2 
x1
= [−2 x1 + 2 x2 2 x1 + 2 x2 ]
x2 (35)
= −2 x21 + 2 x1 x2 + 2 x1 x2 + 2 x22
= −2 x21 + √4 x1 x2 +√2 x22
= 4 x22 − ( 2 x1 − 2 x2 )2

The graph in 3 dimensions in contained in figure 7.

F IGURE 7. Indefinite Quadratic Form −2x21 + 4x1x2 − 2x22

20

0
Q

-20

-5 5
-2.5 2.5
0 0
2.5 -2.5
5 -5
x1 x2

where it is clear that Q takes both positive and negative values.  

1
If we restrict our attention to values of x1 and x2 where x1 = x2 or a matrix b = then
−1
the function will be positive for all values of x 6= 0 as is obvious from the last line of equation 35. To
see this more clearly, draw a vertical plane through the graph above along the x1 = x2 line in figure
8.
20 QUADRATIC FORMS AND DEFINITE MATRICES

F IGURE 8. Indefinite Quadratic Form with Restrictions

50

Q 0

-50

-4 4
-2 2
0 0
2 -2
4 -4
x1 x2

If we combine figure 8 with the plane divding the positive and negative orthants, the positive
definiteness of the quadratic from subject to the constraint is even more obvious as shown in figure
9.

F IGURE 9. Indefinite Quadratic Form with Restrictions

50

0 Q

-50

-5 5
-2.5 2.5
0 0
2.5 -2.5
5 -5
x1 x2

Now along the set of points where x1 = x2 , the function is always positive except where x1 = x2
= 0. So this function is positive definite subject to the constraint that x1 = x2 .

5.3. Definition of a bordered matrix with constraints. Define the bordered matrix HB as follows
 QUADRATIC FORMS AND DEFINITE MATRICES 21

 
a11 a12 ··· a1n | b11 b12 ··· b1m
 a21 a22 ··· a2n | b21 b22 ··· b2m 
 
 .. .. .. .. .. .. .. .. .. 
 . . . . . . . . . 
 
 an1 an2 ··· ann | bn1 bn2 ··· bnm 
 
 .. 
HB =  .  (36)
 
 b11 b21 ··· bn1 | 0 0 ··· 0 
 
 b12 b22 ··· bn2 | 0 0 ··· 0 
 
 . .. .. .. .. .. .. .. .. 
 .. . . . . . . . . 
b1m b2m ··· bnm | 0 0 ··· 0

On the right of the A matrix we append the columns of the B matrix. If there are three constraints,
then the matrix HB will have n + 3 columns or in general n+m columns. Below the A matrix we
append the transpose of the B matrix, one row at a time as we add constraints. So if m = 2, then HB
will have n+2 rows.

5.4. Definiteness of a quadratic form subject to linear constraints.

5.4.1. Constructing minors of HB . To determine the definiteness of the quadratic form in equation
32 subject to equation 33 construct the matrix HB in equation 36. The definiteness is checked by
analyzing the signs of the naturally order principal minors of HB starting with the minor that has
m+1 rows and columns of the matrix A along with the borders for those rows and columns. For
example, if m = 1, then the first minor we check is

 
a11 a12 | b11 

a21 a22 | b21 

 ..  (37)
 . 
 
 b11 b21 | 0 

Then we check

 
a11 a12 a13 | b11

a21 a22 a23 | b21

a31 a32 a33 | b31
 (38)
 .. 
 . 
 
 b11 b21 b31 | 0

and so forth. In general we are checking the signs of minors with p + m rows, where p goes from
m+1 to n. The minors we check can be written as below when there are m constraints.
22 QUADRATIC FORMS AND DEFINITE MATRICES

 
 a11 a12 ··· a1p | b11 b12 ··· b1m 

 a21 a22 ··· a2p | b21 b22 ··· b2m 

 .. .. .. .. .. .. .. .. .. 
 . . . . . . . . . 

 ap1 ap2 ··· app | bp1 bp2 ··· bpm 

 .. 
 .  (39)
 
 b11 b21 ··· bp1 | 0 0 ··· 0 

 b12 b22 ··· bp2 | 0 0 ··· 0 

 . .. .. .. .. .. .. .. .. 
 .. . . . . . . . . 

b b2m ··· bpm | 0 0 ··· 0 
1m

5.4.2. If determinant of HB has the same sign as (−1)p and if these last n-m leading principal minors
alternate in sign, then the quadratic form Q is negative definite on the constraint set x0B = 0. With one
constraint, m = 1, and so p starts at 2, so that the first minor is positive, the second negative and so forth.
5.4.3. If the determinant of HB and these last n-m leading principal minors all have the same sign as (−1)m ,
then Q is positive definite on the constraint set x0B = 0. With one constraint, m = 1, so that the first
minor is negative as are all subsequent ones.
5.4.4. If both of these conditions (b) and (c) are violated by nonzero leading principal minors, then Q is
indefinite on the constraint set x0 B = 0.
5.5. Definiteness of a quadratic form subject to one linear constraint. Construct the (n+1)x(n+1)
matrix HB as in 36 where the constraint equation is now given by

b1 x1 + b2 x2 + · · · + bn xn = 0 (40)
Suppose that b1 6= 0. Then we can show that the general conditions above reduce to the follow-
ing. If the last n leading principal minors of HB have the same sign, Q is positive definite on the
constraint set. If the last n leading principal minors alternate in sign, then Q is negative definite on
the constraint (Simon [5, Section 16.3] ).
 QUADRATIC FORMS AND DEFINITE MATRICES 23

R EFERENCES
[1] 3r d
Dhrymes, P.J. Mathematics for Econometrics - Edition. New York: Springer-Verlag, 2000.
[2] Gantmacher, F.R. The Theory of Matrices. Vol I. New York: Chelsea Publishing Company, 1977.
[3] Hadley, G. Linear Algebra. Reading: Addison Wesley Publishing Company, 1961.
[4] Horn, R.A. and C.R. Johnson. Matrix Analysis. Cambridge: Cambridge University Press, 1985.
[5] Simon, C.P., and L. Blume. Mathematics for Economists. New York: W.W. Norton and Company, 1994.