4.

5 Computational Examples 257

is a linear transformation with no invariant subspace of dimension less than

n. Thus Corollary 4.59 depends on the special nature of the real numbers.

At this point we would like to reemphasize the remark prior to Propo-

sition 4.3.

(4.61) Remark. Our entire approach in this section has been to analyze a
linear transformation T by analyzing the structure of the F [X]-module VT .
Since T1 and T2 are similar precisely when VT1 and VT2 are isomorphic, each
and every invariant we have derived in this section—characteristic and min-
imal polynomials, rational, Jordan, and generalized Jordan canonical forms,
(generalized) eigenvalues and their algebraic and geometric multiplicities,
etc.—is the same for similar linear transformations T1 and T2 .

4.5 Computational Examples

This section will be devoted to several numerical examples to illustrate

the general canonical form theory developed in Section 4.4. In Chapter
5, further techniques will be presented using an analysis of equivalence of
matrices over a PID. Throughout this section it will be assumed that the
reader is familiar with the procedure for solving a system AX = B of linear
equations over a field F by the process of row reduction of the matrix [A B].
This procedure is a standard topic in elementary linear algebra courses; it
will be developed in the more general context of matrices with entries in a
PID in Section 5.2.

(5.1) Example. Construct, up to similarity, all linear transformations T :

F 6 → F 6 with minimal polynomial
mT (X) = (X − 5)2 (X − 6)2 .

Solution. If mT (X) = (X − 5)2 (X − 6)2 , then we must have

cT (X) = (X − 5)i (X − 6)j
where i + j = 6 = dim(F 6 ) and 2 ≤ i, 2 ≤ j. This is because mT (X)
divides cT (X) (Corollary 4.10 (2)), and mT (X) and cT (X) have the same
irreducible factors (Corollary 4.10 (3)).
Suppose i = 2 and j = 4. Then 5 has algebraic multiplicity 2, and 2
is the size of the largest Jordan block with eigenvalue 5, so there is exactly
one such block. Also 6 has algebraic multiplicity 4, and 2 is the size of its
largest Jordan block, so either it has two blocks of size 2, or one block of
size 2 and two of size 1. Thus the possibilities for the Jordan canonical form
of T are
258 Chapter 4. Linear Algebra
 
5 1 0 0 0 0
0 5 0 0 0 0
 
0 0 6 1 0 0
J5,2 ⊕ J6,2 ⊕ J6,2 = 
0 0 0 6 0 0
 
0 0 0 0 6 1
0 0 0 0 0 6
and
 
5 1 0 0 0 0
0 5 0 0 0 0
 
0 0 6 1 0 0
J5,1 ⊕ J6,2 ⊕ J6,1 ⊕ J6,1 = .
0 0 0 6 0 0
 
0 0 0 0 6 0
0 0 0 0 0 6

If i = 4 and j = 2, then there is a similar analysis, with the roles of

5 and 6 reversed. The possibilities for the Jordan canonical form of T are
then
J6,1 ⊕ J5,2 ⊕ J5,2 or J6,2 ⊕ J5,2 ⊕ J5,1 ⊕ J5,1 .

If i = 3 and j = 3, then for each eigenvalue we must have one block of

size 2 and one of size 1. This gives the single possibility
 
5 1 0 0 0 0
0 5 0 0 0 0
 
0 0 5 0 0 0
J5,2 ⊕ J5,1 ⊕ J6,2 ⊕ J6,2 = 
0 0 0 6 1 0
 
0 0 0 0 6 0
0 0 0 0 0 6

for the Jordan canonical form of T . u

t

(5.2) Example. Construct, up to similarity, all linear transformations T with

characteristic polynomial cT (X) = (X −8)7 , minimal polynomial mT (X) =
(X − 8)3 , and 8 an eigenvalue of geometric multiplicity 4.
Solution. Since deg(cT (X)) = 7, T is a linear transformation of a 7-
dimensional vector space V , which we may take to be F 7 , since we are
working up to similarity, and 8 is the only eigenvalue of T . Thus the Jordan
blocks in the Jordan canonical form of T are J8,r . Since mT (X) has degree
3, the largest Jordan block has size 3, so the block sizes must be 3, 3, 1, or 3,
2, 2, or 3, 2, 1, 1, or 3, 1, 1, 1, 1. Since the geometric multiplicity of 8 is equal
to 4, there must be 4 Jordan blocks (Corollary 4.48 (1)), and the only ways
to partition 7 into 4 parts are 7 = 4 + 1 + 1 + 1 = 3 + 2 + 1 + 1 = 2 + 2 + 2 + 1,
these being the possible block sizes (Corollary 4.48 (2)). Comparing these
last two lists, there is only only coincidence; thus there is only one possi-
bility for T :
 4.5 Computational Examples 259
 
8 1 0 0 0 0 0
0 8 1 0 0 0 0
 
0 0 8 0 0 0 0
 
J8,3 ⊕ J8,2 ⊕ J8,1 ⊕ J8,1 = 0 0 0 8 1 0 0.
 
0 0 0 0 8 0 0
 
0 0 0 0 0 8 0
0 0 0 0 0 0 8
u
t

In contrast to the first two examples, the problems which generally

arise are not to construct matrices in Jordan canonical form, but rather
to find the Jordan canonical form of a given linear transformation (or ma-
trix). In this, Proposition 4.23 is our starting point, telling us how to find
the characteristic polynomial. (We do not wish to address at this point the
problem of practical methods for computing the characteristic polynomial,
but simply assume that this is done. Chapter 5 will contain further in-
formation, which will be of computational interest in computing canonical
forms.)

(5.3) Example. Let V = F 3 and let T : V → V be the linear transformation

with matrix  
1 −1 1
[T ]C = A =  −2 1 2 
−2 −1 4
where C is the standard basis on F 3 . Find out “everything” about T .
Solution. A calculation of the characteristic polynomial cT (X) shows that

cT (X) = det(XI3 − A)
= X 3 − 6X 2 + 11X − 6
= (X − 1)(X − 2)(X − 3).

Since this is a product of distinct linear factors, it follows that T is diago-

nalizable (Corollary 4.32), mT (X) = cT (X), and the Jordan canonical form
of T is J = diag(1, 2, 3). Then the eigenvalues of T are 1, 2, and 3, each of
which has algebraic and geometric multiplicity 1.
Since mT (X) = cT (X), it follows that VT is a cyclic F [X]-module and
the rational canonical form of T is
 
0 0 6
C(mT (X)) = R =  1 0 −11  .
0 1 6

We now compute the eigenspaces of T . This is a straightforward matter

of solving systems of linear equations. The details are left to the reader. If
Vλ = Ker(T − λ1V ), then we find
260 Chapter 4. Linear Algebra
* + * + * +
0 1 1
V1 =  1  , V2 =  0  , V3 =  1  .
1 1 1
Thus, if      
½ 0 1 1 ¾
B = 1, 0, 1 ,
1 1 1
then [T ]B = J. Also, if  
0 1 1
P = 1 0 1
1 1 1
is the matrix whose columns are the vectors of B, then P −1 AP = J =
diag(1, 2, 3). Finally, if we let
       
0 1 1 2
v = 1 + 0 + 1 = 2,
1 1 1 3
then

Ann(v) = h(X − 1)(X − 2)(X − 3)i = hmT (X)i = Ann(VT ).

Thus, v is a cyclic vector for the F [X]-module VT . If we let

     
½ 2 3 5 ¾
B = v, T v, T v} = { 2 , 4 , 10 
0 2     
3 6 14

be the basis of V given by Corollary 4.16, then [T ]B0 = R, and if Q is the

matrix whose columns are the column vectors of B0 , then Q−1 AQ = R. u t

(5.4) Example. Let V = F 3 and let T : V → V be the linear transformation

with matrix with respect to the standard basis C
 
6 −2 −2
[T ]C = A =  6 −1 −3  .
4 −2 0
Find out “almost everything” about T .
Solution. We compute

cT (X) = det(XI3 − A)
= X 3 − 5X 2 + 8X − 4
= (X − 1)(X − 2)2 .

We see at this point that we have two possibilities:

 4.5 Computational Examples 261

(a) νalg (1) = νgeom (1) = 1 and νalg (2) = νgeom (2) = 2.
In this case T is diagonalizable and the Jordan canonical form is J =
diag(1, 2, 2). Also mT (X) = (X − 1)(X − 2), so the F [X]-module VT has
invariant factors f1 (X) = X − 1 and f2 (X) = (X − 1)(X − 2), and VT has
rank 2 as an F [X]-module. Thus, T has the rational canonical form
 
1 0 0
R = C(f1 (X)) ⊕ C(f2 (X)) =  0 0 −2  .
0 1 3

(b) νalg (1) = νgeom (1) = 1 and νalg (2) = 2, νgeom (2) = 1.
In this case T is not diagonalizable, and the eigenvalue 2 has a single
Jordan block, so its Jordan canonical form is
 
1 0 0
J = J1,1 ⊕ J2,2 =  0 2 1  .
0 0 2

Also, mT (X) = (X −1)(X −2)2 = cT (X) so that VT is a cyclic F [X]-module

(Lemma 4.11 (3)) and T has the rational canonical form
 
0 0 4
R = C(mT (X)) =  1 0 −8  .
0 1 5

To decide between these two alternatives, we may proceed in either of

two ways:
(1) Compute νgeom (2). If it is 2, we are in case (a), and if it is (1), we are
in case (b).
(2) Compute (T − 1)(T − 2) = (T − 1V )(T − 2 · 1V ). If it is 0 we are in
case (a), if it is not, we are in case (b).
Since we are in any event interested in (generalized) eigenvectors, we
choose the first method. We find
*   +
1 1
Ker(T − 2 · 1V ) =  2  ,  0  = hv2 , v3 i,
0 2

so we are in case (a) here. Also,

* +
2
Ker(T − 1V ) =  3  = hv1 i,
2

so in the basis B = {v1 , v2 , v3 } (note the order), [T ]B = J. u

t
262 Chapter 4. Linear Algebra

(5.5) Example. Let V = F 3 and let V → V be the linear transformation

with the matrix  
5 −5 7
[T ]C = A =  −4 7 −6 
−5 7 −7
in the standard basis C. Find out “almost everything” about T .
Solution. We compute that cT (X) = (X − 1)(X − 2)2 , exactly the same
as in Example 5.4. Thus, we have the same two possibilities and we will
proceed in the same way. Now we find that
* +
1
Ker(T − 2 · 1V ) =  2  = hv2 i,
1
so in this example we are in case (b). To find the basis for the Jordan
canonical form, we find a vector v2 ∈ F 3 with (T − 2)(v3 ) = v2 . By solving
a system of linear equations we find that we may take
 
−3
v3 =  −2  .
0
Also, * +
−3
Ker(T − 1V ) =  −1  = hv1 i.
1
Thus, if B = {v1 , v2 , v3 } (note the order), then B is a basis of F 3 and
[T ]B = J.
As a practical matter, there is a second method for finding a suitable
basis of F 3 . Note that dim(Ker(T − 2 · 1V )2 ) = 2. Pick any vector v30 , which
t
is in Ker(T − 2 · 1V )2 but not in Ker(T − 2 · 1V ), say v30 = [ 2 0 −1 ] .
t
Then let v20 = (T − 2)(v30 ) = [ −1 −2 −1 ] . If B 0 = {v1 , v20 , v30 }, then we
also have [T ]B0 = J. u
t

(5.6) Example. Let V be a vector space over F of dimension 8, let B =

{vi }8i=1 be a basis of V , and suppose that T ∈ EndF (V ) is a linear trans-
formation such that
[T ]B = A
= J2,2 ⊕ J2,2 ⊕ J2,1 ⊕ J1,2 ⊕ J1,1
 
2 1 0 0 0 0 0 0
0 2 0 0 0 0 0 0
 
0 0 2 1 0 0 0 0
 
0 0 0 2 0 0 0 0
= .
0 0 0 0 2 0 0 0
 
0 0 0 0 0 1 1 0
 
0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 1
 4.5 Computational Examples 263

That is, [T ]B is already in Jordan canonical form. Compute µ(VT ), the

cyclic decomposition of VT , and the rational canonical form of T .
Solution. Since νgeom (2) = 3 and νgeom (1) = 2, it follows from Corollary
4.48 (4) that µ(VT ) = 3. Moreover, the elementary divisors of T are (X−2)2 ,
(X − 2)2 , (X − 2), (X − 1)2 , and (X − 1), so the invariant factors are (see
the proof of Theorem 3.7.15):
f3 (X) = (X − 2)2 (X − 1)2
f2 (X) = (X − 2)2 (X − 1)
f1 (X) = (X − 2).
Since
Ann(v2 ) = Ann(v4 ) = h(X − 2)2 i
Ann(v5 ) = h(X − 2)i
Ann(v7 ) = h(X − 1)2 i
Ann(v8 ) = h(X − 1)i,
it follows (by Lemma 3.7.18) that
Ann(w3 ) = hf3 (X)i
Ann(w2 ) = hf2 (X)i
Ann(w1 ) = hf1 (X)i
where w3 = v2 + v7 , w2 = v4 + v8 , and w1 = v5 . Thus
VT ∼
= Rw1 ⊕ Rw2 ⊕ Rw3
and the rational canonical form R of T is the matrix
R = C(f1 (X)) ⊕ C(f2 (X)) ⊕ C(f3 (X)).
Moreover, Q−1 AQ = R if Q ∈ M8 (F ) is the invertible matrix
 
0 0 0 0 0 1 4 12
0 0 0 0 1 2 4 8 
 
0 0 1 4 0 0 0 0 
 
C 0 1 2 4 0 0 0 0 
Q = PB =  
1 0 0 0 0 0 0 0 
 
0 0 0 0 0 1 2 3 
 
0 0 0 0 1 1 1 1
0 1 1 1 0 0 0 0
where
C = {w1 , w2 , T w2 , T 2 w2 , w3 , T w3 , T 2 w3 , T 3 w3 }.
u
t
264 Chapter 4. Linear Algebra

(5.7) Remark. We have given information about the rational canonical form
in the above examples in order to fully illustrate the situation. However, as
we have remarked, it is the Jordan canonical form that is really of interest.
In particular, we note that while the methods of the current section produce
the Jordan canonical form via computation of generalized eigenvectors, and
then the rational canonical form is computed (via Lemma 3.7.18) from the
Jordan canonical form (e.g., Examples 5.3 and 5.6), it is the other direction
that is of more interest. This is because the rational canonical form of T can
be computed via elementary row and column operations from the matrix
XIn − [T ]B . The Jordan canonical form can then be computed from this
information. This approach to computations will be considered in Chapter
5.

We will conclude this survey of computations by sketching the algo-

rithm that has been used in the previous examples to compute a basis of V
in which a given linear transformation T ∈ EndF (V ) is in Jordan canonical
form. The Jordan canonical form can be computed solely by solving sys-
tems of homogeneous linear equations, provided the eigenvalues of T are
available. It should, however, be emphasized again that the computation of
eigenvalues involves solving a polynomial equation.
Suppose that T ∈ EndF (V ) has mT (X) = (X − λ)r q(X) where q(λ) 6=
0. Then the primary decomposition theorem allows one to write

VT ∼
= W1 ⊕ · · · ⊕ Wt ⊕ Ve

where Wi = F [X]wi is cyclic with Ann(wi ) = h(X − λ)ri i and

W = W1 ⊕ · · · ⊕ Wt

is the generalized eigenspace of the eigenvalue λ. By rearranging the order

of the ri if necessary we may assume that

r = r1 ≥ r2 ≥ · · · ≥ rt

(and of course r1 + · · · + rt = νalg (λ).) We will let

r1 = r̃1 > r̃2 > · · · > r̃s = rt

denote the distinct ri . It is the vectors wi that are crucial in determining the
Jordan blocks corresponding to the eigenvalue λ in the Jordan canonical
form of T . We wish to see how these vectors can be picked out of the
generalized eigenspace W corresponding to λ. First observe that

B = {w1 , Sw1 , . . . , S r1 −1 w1 , w2 , . . . , S r2 −1 w2 , . . . , S rt −1 wt },

where S = T − λ1V , is a basis of W . This is just the basis (in a different

order) produced in Theorem 4.38. It is suggestive to write this basis in the
following table:
 4.5 Computational Examples 265

Table 5.1. Jordan basis of T

w1 ··· wk1
Sw1 ··· Swk1
.. ..
. . wk1 +1 ··· wk2
Swk1 +1 ··· Swk2
.. .. .. ..
. . . . ··· wt
.. .. .. .. ..
. . . . .
S r1 −1 w1 ··· S r1 −1 wk1 S r̃2 −1 wk1 +1 ··· S r̃2 −1 wk2 ··· S rt −1 wt

(Of course, W is an F [X]-module of rank t, and as we observe from

Table 5.1, t = dimF Ker(T − λ1V ), as we expect from Corollary 4.48 (4).)
Note that the linear span of each column is the cyclic submodule Wi = Rwi
while the last k rows form a basis of Ker(T − λ1V )k . As a concrete example
of this scheme, suppose that T : F 13 → F 13 is multiplication by the matrix

A = J2,4 ⊕ J2,4 ⊕ J2,2 ⊕ J2,2 ⊕ J2,1 .

Using the standard basis on F 13 the above table becomes

Table 5.2. Jordan basis of T : F 13 → F 13

e4 e8
e3 e7
e2 e6 e10 e12
e1 e5 e9 e11 e13

Thus, the bottom row is a basis of Ker(T − 2 · 1V ), the bottom two rows are
a basis of Ker(T − 2 · 1V )2 , etc. These tables suggest that the top vectors
of each column are vectors that are in Ker(T − λ1V )k for some k, but
they are not also (T − λ1V )w for some other generalized eigenvector w.
This observation can then be formalized into a computational scheme to
produce the generators wi of the primary cyclic submodules Wi .
We first introduce some language to describe entries in Table 5.1. It
is easiest to do this specifically in the example considered in Table 5.2, as
otherwise the notation is quite cumbersome, but the generalization is clear.
We will number the rows of Table 5.1 from the bottom up. We then say (in
the specific case of Table 5.2) that {e13 } is the tail of row 1, {e10 , e12 } is
266 Chapter 4. Linear Algebra

the tail of row 2, the tail of row 3 is empty, and {e4 , e8 } is the tail of row
4. Thus, the tail of any row consists of those vectors in the row that are at
the top of some column.
Clearly, in order to determine the basis B of W it suffices to find the
tails of each row. We do this as follows:
(i)
For i = 0, 1, 2, . . . , let Vλ = Ker(S i ) (where, as above, we let S =
T − λ1V ). Then
(0) (1) (2) (r)
{0} = Vλ ⊆ Vλ ⊆ Vλ ⊆ · · · ⊆ Vλ = W.
(i)
Let di = dim(Vλ ) so that

0 = d0 < d1 ≤ d2 ≤ · · · ≤ dr = νalg (λ).

Note that r is the smallest value of i for which di = νalg (λ), where r is as
above, i.e., where mT (X) = (X − λ)r q(X) with q(λ) 6= 0. This observation
is useful for determining r without first computing mT (X) explicitly. In
particular, we will have r = 1, i.e., d1 = νalg (λ), if and only if the generalized
eigenspace of T corresponding to λ is the eigenspace of T corresponding to
(r) (r+1) (r)
λ. Note that in any case Vλ = W , so Vλ = Vλ , and dr+1 = dr .
Observe that for 1 ≤ i ≤ r, row i of Table 5.1 has length di − di−1 ,
and hence, its tail has length

(di − di−1 ) − (di+1 − di ).

We make the following observation:

(i+1) (i) (i+1)
Let V λ be any complement of Vλ in Vλ , i.e., any subspace of
(i+1)
Vλ such that
(i+1) (i) (i+1)
Vλ = Vλ ⊕ V λ .
(i+1) (i−1) (i)
Then S(V λ ) and Vλ are both subspace of Vλ and

(i+1) (i−1)
(5.8) Claim. S(V λ ) ∩ Vλ = {0}.

(i+1) (i−1)
To see this, let v ∈ S(V λ ) ∩ Vλ . Then v = S(v) for some v ∈
(i+1) i−1
V λ , and also S (v) = 0. But

0 = S i−1 (v) = S i−1 (S(v)) = S i (v)

(i+1)
implies that v = 0, by the definition of V λ .

Now for our algorithm. We begin at the top, with i = r, and work
(i) (i−1) (i)
down. By Claim 5.8, we may choose a complement V λ of Vλ in Vλ ,
(i+1) (i)
which contains the subspace S(V λ ). Let V λ be any complement of
 4.5 Computational Examples 267

(i+1) (i) (i)

S(V λ ) in V λ . Then a basis for V λ gives the tail of row i of Table 5.1.
(r+1) (r) (r)
(Note that V λ = {0}, so at the first stage of the process V λ = V λ
(r−1) (r)
is any complement of Vλ in Vλ = W . Also, at the last stage of the
(0) (1) (1) (1) (2)
process Vλ = {0}, so V λ = Vλ and V λ is any complement of S(V λ )
(1)
in Vλ , the eigenspace of T corresponding to λ.)

(5.9) Example. We will illustrate this algorithm with the linear transforma-
tion T : F 13 → F 13 whose Jordan basis is presented in Table 5.2. Of course,
this transformation is already in Jordan canonical form, so the purpose is
(i) (i) (i)
just to illustrate how the various subspace Vλ , V λ , and V λ relate to the
basis B = {e1 , . . . , e13 }. Since there is only one eigenvalue, for simplicity,
(i)
let V2 = Vi , with a similar convention for the other spaces. Then
V5 = V4 = F 13
V4 = he1 , . . . , e13 i
V3 = he1 , e2 , e3 , e5 , e6 , e7 , e9 , e10 , e11 , e12 , e13 i
V2 = he1 , e2 , e5 , e6 , e9 , e10 , e11 , e12 , e13 i
V1 = he1 , e5 , e9 , e11 , e13 i
V0 = {0},
while, for the complementary spaces we may take
V 5 = {0}
V 4 = he4 , e8 i
V 3 = he3 , e7 i
V 2 = he2 , e6 , e10 , e12 i
V 1 = he1 , e5 , e9 , e11 , e13 i.
Since V4 = F 13 , we conclude that there are 4 rows in the Jordan table of
T , and since V 4 = V 4 , we conclude that the tail of row 4 is {e4 , e8 }. Since
S(e4 ) = e3 and S(e8 ) = e7 , we see that S(V 4 ) = V 3 and hence the tail
of row 3 is empty. Now S(e3 ) = e2 and S(e7 ) = e6 so that we may take
V 2 = he10 , e12 i, giving {e10 , e12 } as the tail of row 2. Also, S(e2 ) = e1 ,
S(e6 ) = e5 , S(e10 ) = e9 , and S(e12 ) = e11 , so we conclude that V 1 = he13 i,
i.e., the tail of row 1 is {e13 }.

Examples 5.3 and 5.4 illustrate simple cases of the algorithm described
above for producing the Jordan basis. We will present one more example,
which is (slightly) more complicated.

(5.10) Example. Let V = F 4 and let T : V → V be the linear transformation

with the matrix
268 Chapter 4. Linear Algebra
 
−1 1 1 1
 −3 3 1 1
[T ]C = A =  
−8 2 5 3
2 0 −1 1
in the standard basis C. Find the Jordan canonical form of T and a basis
B of V such that [T ]B is in Jordan canonical form.
Solution. We compute that

cT (X) = cA (X) = (X − 2)4 .

Thus, there is only one eigenvalue, namely, 2, and νalg (2) = 4. Now find the
eigenspace of 2, i.e.,
   
* 1 1 +
(1) 1 1
V2 = Ker(T − 2 · 1V ) =   ,   ,
2 0
0 2
so d1 = 2. We also find that
(2)
V2 = Ker((T − 2 · 1V )2 ) = V,

so d2 = 4. Hence, we choose {w1 , w2 } ⊆ V so that

(1)
V2 ⊕ hw1 , w2 i = V

in order to obtain a Jordan basis of V .

Take    
1 0
0 1
w1 =   and w2 =   .
0 0
0 0
Then compute

−3
 −3 
v1 = (T − 2 · 1V )(w1 ) =  
−8
2
and
 
1
1
v3 = (T − 2 · 1V )(w2 ) =   .
2
0
Setting

v1 = (T − 2 · 1V )(w1 ), v2 = w1 ,
v3 = (T − 2 · 1V )(w2 ), v4 = w2 ,
 4.6 Inner Product Spaces and Normal Linear Transformations 269

we obtain a basis B = {v1 , v2 , v3 , v4 } such that [T ]B = J. The table

corresponding to Table 5.1 is

v2 v4
v1 v3

We also note that VT ∼

= Rv2 ⊕ Rv4 so that µ(VT ) = 2. u
t

4.6 Inner Product Spaces and Normal

Linear Transformations

In this section we will study vector spaces possessing an additional struc-

ture, that of an inner product, as well as endomorphisms of these spaces
which are “normal,” normality being a property defined in terms of an inner
product.

(6.1) Definition. Let V be a vector space over F = R or C. An inner

product on V is a function ( : ) : V × V → F such that for all u, v, w ∈ V
and a ∈ F ,
(1) (u + v : w) = (u : w) + (v : w),
(2) (au : v) = a(u : v),
(3) (u : v) = (v : u), and
(4) (u : u) > 0 if u 6= 0.
A vector space V , together with an inner product ( : ) is called an inner
product space.

The bar in Definition 6.1 (3) refers to conjugation in the field F . Of

course, if F = R, conjugation in F is trivial, but it is convenient to handle
the real and complex cases simultaneously.

(6.2) Examples.
(1) The standard inner product on F n is defined by
n
X
(u : v) = uj v j .
j=1

t
(2) If A ∈ Mm,n (F ), then let A∗ = A where A denotes the matrix ob-
tained from A by conjugating all the entries. A∗ is called the Hermitian
transpose of A. If we define

(A : B) = Tr(AB ∗ )