Appendix 7

Evaluation of volume of
A
nn

Iu

Denition 12.21 on page 466 of the determinant of an n n matrix, A, is not the kind that permits immediate evaluation of det ( A ) . In order to accomplish this task, we have to introduce some ways to transform the matrix A, which dont affect its determinant, or affect it in a particularly simple way; after sufciently many such transformations, yield a simple form whose determinant is immediately recognizable.

The ideas behind the type of transformations we are seeking are based on a. the same type of matrix row operations used for matrix inversion (see Theorem 10.12 on page 339 and Example 10.7 on page 339) namely, adding a multiple of one row to another, and interchanging rows; b. showing that the volume (Jordan content) V ( M Iu ) = V ( I u ) = 1
n n

where I u is the n-dimensional unit interval (see Denition 12.1 on page 438, where a and b are n-dimensional vectors, a consisting of all 0s and b of all 1s) M is the n n matrix obtained by carrying out the matrix operation of adding a multiple of one row to another, or obtained by interchanging two rows, on the n n identity matrix (Denition 10.6 on page 334) and M I u is the image of I u under M (see Denition 12.20 on page 462); c. extending this result to show that V ( M A ) = V ( A ) whenever A is an n-dimensional set with well-dened Jordan content (Denition 12.7 on page 442). Before embarking on this proof, you should be curious about why the result is believable. That is, why would you expect the volume of the image of the unit cube, V ( M I u ) to be unchanged when the matrix M is of the above form ? Well, recall that this image in two dimensions is just the area of the indicated parallelogram, whose sides are the columns of M.

1999 by CRC Press LLC

Column 1

Projection of Column 1 on Column 2

Figure A7-1

Column 2

Subtracting this multiple of Column 2 from Column 1 changes the shape of the parallelogram converting it to a rectangle, but, as you can see, doesnt change the area; it just shifts the triangular piece to the other side. Algebraically, its because the total area is the sum of the areas formed by Column 2 with each of the legs of the triangle whose legs add up, vectorially, to Column 2. But since the parallelogram formed by the Projection and Column 2 has 0 area (due to the linear dependence of this projection and Column 2) only the component of Column 1 orthogonal to Column 2 contributes to the area. Although this may seem to represent only a special case, a little thought should convince you that the idea extends not only to all two-dimensional cases, but to any number of dimensions. Well work backwards , rst showing that if step b above has been proved, then step c follows. To do this, we approximate A closely by a nite union (see Footnote 1 on page 453) of intervals B , I j k
k=1 m q

j=1

where the (interior intervals) I j lie completely within the set A, and the (boundary intervals) B k , include both points in A and points not in A (see Figure 12-3 on page 443 and the surrounding discussion). Because A is assumed to have a well-dened Jordan content (Denition 12.7 on page 442) the total volume of the boundary intervals can be made small, by keeping the partition norm (Denition 12.2 on page 439) which generated these intervals, sufciently small. Now letting the symbol denote set inclusion,1 because I B Ij A j k
j=1 k=1 m m q

A7.1

j=1

it follows that

1. That is, A B means that each element of the set A is also an element of the set B.

1999 by CRC Press LLC

j=1

M ( A ) M I j B k . I j
j=1 k=1

A7.2

By the invertibility of M, which is easy to justify,1 M, considered as a function is one-one (Denition 12.18 on page 460). So, because the intervals I j and B k are non-overlapping, the images, M ( I j ) and M ( B k ) , also are nonoverlapping. Hence, using the fact that V ( ) denotes an integral, from Theorem 12.15 on page 453, we see that V and V M Similarly V M

j=1

I j=

j=1

V (I j) and V

j=1

I j

k =1 m

Bk =

V (I j) +V (Bk)
j=1 k =1 m

A7.3

j=1

I j = V

j=1

M(I j) =

V (M(I j)) =V (I j).

j=1 j=1 m q

A7.4

j=1

I j

k=1

B k =

j=1

V ( I j ) + V ( B k ).
k=1

A7.5

Applying the volume operator, V, to the elements expression A7.1 above yields V
j=1

V ( A ) V I j B k I j
j=1 k=1

A7.6

while applying V to the elements of expressions A7.2 yields V M V (M( A)) V M I B . I j j k

j=1 k=1 m m q

A7.7

j=1

Now substituting from equations A7.4 and A7.5 into inequalities yields

j=1

V (I j) V ( A)

j=1 m

V (I j) +

k=1

V ( Bk )
q

A7.8

and

j=1

V ( I j ) V ( M ( A ) ) V ( I j ) + V ( Bk ) .
j=1 k=1

A7.9

1. Since the operation is reversed by subtracting the same multiple of the rst row being considered from the second one.

1999 by CRC Press LLC

Now, comparing inequalities A7.8 and A7.9, and using the fact that by choosing sufciently small the norm of the partition giving rise to these sets makes V ( B k ) arbitrarily small, shows that V (M( A)) = V ( A),
n n

A7.10

provided that we know V ( M I u ) = V ( I u ) = 1. Well save this step for last. If we have shown that V ( M ( A ) ) = V ( A ) , for M an elementary row addition operation matrix, and A any n-dimensional set with well-dened Jordan content, we can now go to work and reduce the matrix C = DH ( w ) to diagonal form by a series of successive row operations of the type mentioned at the start of this appendix (you might want to look at Example 10.7 on page 339, but omitting multiplication of a row by a scalar, which may be needed to generate the identity from the original matrix, but is not needed to convert to diagonal form). In the diagonal form, only the diagonal elements are nonzero. Now note that each such operation can be carried out by multiplying on the left by the corresponding elementary row operation matrix e.g., to add 2 row 3 to row 1 in the matrix a b c B= d e f g h i multiply B on the left by 102 M = 010 . 001

A7.11

A7.12

Exercises
1. Verify that multiplying MB from equations A7.11 and A7.12 does have the effect of adding 2 row 3 to row 1. 2. Generalize the result of the previous problem to handle all such situations. So we have shown that we can nd elementary row operation matrices, say M 1 ,..., M r , such that M 1 M 2 ... M r C = D , A7.13

where D is a diagonal matrix, whose i -th diagonal element is d ii . The volume, V ( D ( I u ) ) , is easily seen to be given by
n

1999 by CRC Press LLC

V (D( I u )) =
n

i=1

d ii .

A7.14

So there remains only the rst step, to show that if M is a matrix obtained from the identity matrix via an elementary row operation, then V ( M Iu ) = V ( I u ) = 1 .
n n

We will leave it to the reader to verify this if M represents a row interchange. As for showing this result when M corresponds to adding a multiple of one row to another, this is where we must perform an integration. For simplicity, we will only consider the situation of adding a multiple of row 2, c row 2, to row 1. This really loses no generality, since row interchange can handle any other cases. So we want to prove that V 1 0 0 . 0 0 c 1 0 . . 0 0 0 1 . . . 0 . 0 . . . . . . . . 0 0 0 0 I = 1. u . n . 1

A7.15

This is where we need to represent volume (Jordan content) as an integral. Letting m j denote the j-th column of M, we have Mx = where m 2 = c e 1 + e 2 and m k = e k for k 2 , with e j being the standard n-dimensional j-th unit vector (Denition 8.2 on page 263) and a matrix times a vector given in Denition 10.1 on page 328. Hence, using the set notation introduced in Denition 12.16 on page 454, together with the denition of the image of a set under a function (Denition 12.20 on page 462) we nd M(I u )= {y in E n : y = (x1+cx2) e1+
n

j=

m j x j,

x je j
j=2

for 0 xi 1, all i = 1,..., n}.

To determine V ( M ( I u n ) ) we rst write M ( I u n ) in sequential form (again see Denition 12.16); here we let y 1 be the last coordinate determine. In this sequential form we nd

1999 by CRC Press LLC

M(I u ) = {(y1,y2,...,yn) in E n : 0 y2 1; 0 y3 1;...;0 yn 1;cy2 y1 1},

n

because the limits for y 2 ,..., y n , set rst, are evident. But for each choice of y 2 and x 1 , y 1 = cy 2 + x 1 so as x 1 varies from 0 to 1 in the nonsequential form of M ( I u ) just preceding the sequential form above, for n chosen y 2 , y 1 varies from cy 2 to cy 2 + 1. Then since V ( M ( I un ) ) =

M(Iu )
n

(the right side being a multiple integral (see Denition 12.3 on page 440) where 1 is the constant function of n variables whose value is 1) from Theorem 12.17 on page 455, on evaluation of a multiple integral as an iterated integral, we have V ( M ( I un ) ) =

0 ... 0 c y2

c y2 + 1

1 dy 1 dy 2 ... dy n = 1,

establishing the asserted result. Thus we have shown the following result.

Theorem A7.1: Volume of image of unit interval in E n

The volume of the image (Denition 12.20 on page 462) of the unit ndimensional interval (Denition 12.1 on page 438) A ( I u ) under an n n n matrix, A is the magnitude (absolute value) of the determinant of A, as dened in Denition 12.21 on page 466. Symbolically V ( A ( I un ) ) =

= det A .
n

A(Iu )

The next result is needed for connection of the geometric meaning of determinants with the suitability of the associated matrix for coordinate changes.

Theorem A7.2: Determinants and linear dependence

If A is a real n n matrix (Denition 10.1 on page 328) with determinant det ( A ) (Denition 12.21 on page 466) then the columns of A are linearly dependent (Denition 9.7 on page 297) if and only if det ( A ) = 0.

1999 by CRC Press LLC

Proof: It isnt hard to see that the row operations applied to a matrix to convert it to the diagonal form needed for simple evaluation of its determinant do not alter the linear dependence/independence status of the set of its column vectors, (or its row vectors). But in the diagonal form, it is easy to see that the columns of this nal matrix are linearly dependent if and only if at least one of the diagonal elements, d ii , is 0, and this occurs if and only if the determinant, n
i=1

d ii

= 0,

proving the desired result.

1999 by CRC Press LLC