LINEARLY CONSTRAINED QUADRATIC FORMS

AND BORDERED MATRICES

In economics, there are many situations where the variables are subject to a constraint. One example is
the following. Assume the wealth (total value) is xed as w > 0 and there are n goods with prices xed as
p
i
> 0. Assume the commodity bundles (x
1
, . . . , x
n
) are restricted to those that satisfy p
1
x
1
+ + p
n
x
n
=
w. Somethings, we then want to maximize another quantify, such as utility, U(x
1
, . . . , x
n
).
When we study multidimensional calculus we will consider the method of Lagrange multipliers. In this
handout, we give the linear algebra necessary to state a second derivative test for the method of Lagrange
multipliers. In this context, we need to consider maximizing a quadratic form Q(x) = x
T
Ax on a linear
subspace given by Bx = 0. Here, B can have k rows, which corresponds to k linear constraints.
1. Quadratic Forms with Two Variables and One Linear constraint
Before considering the general case, we consider the case of two variable and one linear constraint;
Q(x, y) = a
1,1
x
2
+ 2a
1,2
xy + a
2,2
y
2
on the points satisfying the linear constraint b
1
x + b
2
y = 0 with
(b
1
, b
2
) = 0. The points on the linear constraint can be parameterized by (x, y) = t (b
2
, b
1
). Then,
Q(t b
2
, -t b
1
) = a
1,1
(t b
2
)
2
+2a
1,2
(t b
2
)(-t b
1
) +a
2,2
(-t b
1
)
2
=
_
a
1,1
b
2
2
+2a
1,2
b
2
b
1
a
2,2
b
2
1
_
t
2
= det
_
_
0 b
1
b
2
b
1
a
1,1
a
1,2
b
2
a
1,2
a
2,2
_
_
t
2
.
Therefore, Q(x, y) is positive denite on b
1
x +b
2
y = 0, if and only if
det
_
_
0 b
1
b
2
b
1
a
1,1
a
1,2
b
2
a
1,2
a
2,2
_
_
> 0.
2. General Theorem
In this section, we discuss the general case of the quadratic form Q(x) = x
T
Ax on the null space B(x) =
0. The matrix A is n n, so there are n variables. We assume that B is k n and has rank k with k < n. We
assume that the variables have been ordered so that det(B
1
) = 0, where B is partitioned so that B
1
contains
the rst k columns and B
2
contains the last n k columns of B,
B
1
=
_
_
_
b
1,1
. . . b
1,k
.
.
.
.
.
.
b
k,1
. . . b
k,k
_
_
_
and B
2
=
_
_
_
b
1,k+1
. . . b
1,n
.
.
.
.
.
.
b
k,k+1
. . . b
k,n
_
_
_
,
The condition that determines whether Q is denite on the null space of B is the (n + k) (n + k)
symmetric matrix called the bordered Hessian that combines the matrix A with B and B
T
on the borders:
H =
_
0 B
B
T
A
_
=
_
_
_
_
_
_
_
_
_
0 . . . 0 b
1,1
. . . b
1,n
.
.
.
.
.
.
.
.
.
.
.
.
0 . . . 0 b
k,1
. . . b
k,n
b
1,1
. . . b
k,1
a
1,1
. . . a
1,n
.
.
.
.
.
.
.
.
.
.
.
.
b
1,n
. . . b
k,n
a
n,1
. . . a
n,n
_
_
_
_
_
_
_
_
_
.
1
2 BORDERED MATRICES
Since the dimension of the null space is n k, we should need to check the signs of determinants of n k
submatrices. Using the fact that det(B
1
) = 0, we consider the j j submatrices of H, formed from H by
deleting the last n +k j rows and columns:
H
j
=
_
_
_
_
_
_
_
_
_
0 . . . 0 b
1,1
. . . b
1, j k
.
.
.
.
.
.
.
.
.
.
.
.
0 . . . 0 b
k,1
. . . b
k, j k
b
1,1
. . . b
k,1
a
1,1
. . . a
1, j k
.
.
.
.
.
.
.
.
.
.
.
.
b
1, j k
. . . b
k, j k
a
j k,1
. . . a
j k, j k
_
_
_
_
_
_
_
_
_
.
Theorem 1. Let A be n n symmetric matrix and B = [B
1
B
2
] be k n with rank k and det(B
1
) = 0. Let
the bordered Hessians H
j
be dened as above, and Nul(B) = {x : B(x) = 0} be the null space of B.
a The quadratic form Q(x) = x
T
Ax is positive denite on the null space Nul(B) iff
(-1)
k
det(H
j
) > 0 for 2k +1 j n +k.
b The quadratic form Q(x) = x
T
Ax is negative denite on the null space Nul(B) iff
(-1)
j k
det(H
j
) > 0 for 2k +1 j n +k. (Note that the signs of det(H
j
) alternate with j .)
Remark 1. (i) When we considered one linear constraint k = 1 and n = 2, we required that - det(H) > 0
which is compatible with the general result, since 2k +1 = 3 = n +k and we only need to consider the sign
of det(H
n+k
) = det(H). (ii) For n k > 1, we need to consider the sign of more than one determinant.
For cases with more than one constraint, k > 1 and the border has more than one extra row and column.
Proof.
Since det(B
1
) = 0, we can solve for the null space of B in terms of the last n k variables:
0 = B
1
_
_
_
x
1
.
.
.
x
k
_
_
_
+B
2
_
_
_
x
k+1
.
.
.
x
n
_
_
_
_
_
_
x
1
.
.
.
x
k
_
_
_
= B
1
1
B
2
_
_
_
x
k+1
.
.
.
x
nk
_
_
_
= J
_
_
_
x
k+1
.
.
.
x
nk
_
_
_
,
where J = -B
1
1
B
2
. Partitioning A =
_
A
11
A
12
A
T
12
A
22
_
into blocks, where A
11
is k k, A
12
is k (n k), and
A
22
is (n k) (n k), the quadratic form on the null space has the following symmetric matrix E:
E =
_
J
T
I
_
_
A
11
A
12
A
T
12
A
22
_ _
J
I
_
=
_
J
T
I
_
_
A
11
J +A
12
A
T
12
J +A
22
_
= J
T
A
11
J +J
T
A
12
+A
T
12
J +A
22
.
On the other hand, we can perform a (non-orthogonal) change of basis of the n +k-dimensional space on
which the quadratic form H is dened:
_
_
I
k
0 0
0 I
k
0
0 J
T
I
nk
_
_
_
_
0
k
B
1
B
2
B
T
1
A
11
A
12
B
T
2
A
T
12
A
22
_
_
_
_
I
k
0 0
0 I
k
J
0 0 I
nk
_
_
=
_
_
0 B
1
0
B
T
1
A
11
C
12
0 C
T
12
E
_
_
Here the matrix E induces the quadratic form on the null space as we showed above. Since the determinant
of the change of basis matrix is one, this change of basis preserves the determinant of H, and also the
determinants of H
j
for 2k +1 j n +k.
BORDERED MATRICES 3
By using k row interchanges
det(H) = det
_
_
0 B
1
0
B
T
1
A
11
C
12
0 C
T
12
E
_
_
= (-1)
k
det
_
_
B
T
1
A
11
C
12
0 B
1
0
0 C
T
12
E
_
_
= (-1)
k
det(B
T
1
) det
_
B
1
0
C
T
12
E
_
= (-1)
k
det(B
1
)
2
det(E).
This calculation carries over to all the H
j
, (-1)
k
det(H
j
) = det(B
1
)
2
det(E
j 2k
). Therefore, we can use the
signs of the determinants of the H
j
for 2k + 1 j n + k to check the signs of the determinants of the
principal submatrices E
j 2k
with size ranging from 1 to n k.
The quadratic form Q for A is positive denite on the null space iff the quadratic form for E is positive
denite iff
(-1)
k
det(H
j
) = det(B
1
)
2
det(E
j 2k
) > 0 for 2k +1 j n +k.
For the negative denite case, the quadratic form Q for A is negative denite on the null space iff the
quadratic form for E is negative denite iff
(-1)
j k
det(H
j
) = (-1)
j 2k
det(B
1
)
2
det(E
j 2k
) > 0 for 2k +1 j n +k.

3. EXAMPLES
Example 1. Consider the quadratic form
Q(x
1
, x
2
, x
3
, x
4
) = x
2
1
+ x
2
2
x
2
3
+ x
2
4
+4x
2
x
3
2x
1
x
4
and linear constraints
x
1
+ x
2
+ x
3
= 0 and x
1
9x
3
+ x
4
= 0.
The matrices are
A =
_
_
_
_
1 0 0 -1
0 1 2 0
0 2 -1 0
-1 0 0 1
_
_
_
_
and B =
_
1 1 1 0
1 0 -9 1
_
Notice that the 2 2 matrix formed from the rst two columns of B is invertible:
det(B
1
) = det
_
1 1
1 0
_
= -1 = 0.
We need to form the bordered matrices for 2k +1 = 4 +1 = 5 j n +k = 6,
H
6
=
_
0 B
B
T
A
_
=
_
_
_
_
_
_
_
_
0 0 1 1 1 0
0 0 1 0 -9 1
1 1 1 0 0 -1
1 0 0 1 2 0
1 -9 0 2 -1 0
0 1 -1 0 0 1
_
_
_
_
_
_
_
_
and H
5
=
_
_
_
_
_
_
0 0 1 1 1
0 0 1 0 -9
1 1 1 0 0
1 0 0 1 2
1 -9 0 2 -1
_
_
_
_
_
_
.
Since (-1)
k
= (-1)
2
= +1, det(H
6
) = 24 > 0, and det(H
5
) = 140 > 0, Q is positive denite on the linear
constraint set.
4 BORDERED MATRICES
Problems
1. Let Q be as in the example. Use only the one linear constraint,
x
1
+ x
2
+ x
3
= 0.
Is Q positive denite on the null space of this one equation?
2. Let Q be as in the example. Use the three linear constraints,
x
1
+ x
2
+ x
3
= 0
x
1
9x
3
+ x
4
= 0
x
2
+ x
4
= 0.
Is Q positive denite on the null space of these three equations?
3. Let
A =
_
_
1 1 1
1 1 -1
-1 1 1
_
_
and B =
_
1 -1 -1
_
.
Is the quadratic form for A positive denite on the null space of for B?
REFERENCES
[1] C. Hassell and E. Rees, The index of a constrained critical point, The Mathematical Monthly, October 1993, pp. 772778.