Linear Algebra

May 26, 2020

Eigenvalues and Eigenvectors

Question
Let's ask this question:
Ÿ given a linear transformation T ∶ V Ñ V where V is an n
dimensional vector space over F . Does there exists an ordered
basis B  pv1 , ..., vn q such that the matrix rT sB is a diagonal
matrix:  
α1 0 ⋯ 0
 
rT sB   0 α2 ⋯ 0 
⋮ ⋱ ⋮
(1)
0 0 ⋯ αn
Ÿ Obviously if this occurs then we should have Tvi  αvi for
1 ¤ i ¤ n.
Ÿ In other words vi 's are eigenvectors and αi 's are eigenalues:
Eigenvalues and Eigenvectors

Ÿ Denition 1.
Let V be a vector space over a eld F and let T ∶ V Ñ V be a
linear transformation. A scalar α P F is called an eigenvalue of T if
there exists a nonzero vector v P V such that
Tv  αv
The vector v is named an eigenvector of T .
In other words eigenvalues are those elements α P F for which
ker pT
αI q is a non-trivial subspace of V . And each nonzero
element of pT
αI q is named an eigenvector of T corresponding to
the eigenvalue α. The subspace ker pT
αI q is also called an
eigenspace of T corresponding to the eigenvalue α P F .

( Here I denotes the identity transformation on V .)

Eigenvalues and Eigenvectors

Ÿ It is obvious that if v is an eigenvector of T corresponding to

the eigenvalue α P F then any scalar multiple rv where
0  r P F is also an eigenvector of T . In other words all the
non-zero vectors in the one dimensional subspace v ¡
generated by v are also eigenvectors. This is why we can talk
also about eigen direction corrsponding to the eigenvalue α.
Ÿ If we are given an eigenvalue α for T in order to compute the
corresponding eigenvector we should resolve the linear system
of equations
pT
αI qX  0
Eigenvalues and Eigenvectors

Ÿ Denition 2 (Diagonalizability).
A linear transformation T ∶ V Ñ V dened on an n dimensional
vector space is called diagonizable if there exists an ordered basis
B  pv1 , ..., vn q with respect to which the matrix representation of
T denoted by rT sB has the diagonal form (1)
Eigenvalues and Eigenvectors

Diagonalizability
Ÿ Proposition 1.
A linear transformation T ∶ V Ñ V on an n dimensional vector
space is diagonizable i there exists a basis B  pv1 , ..., vn q all of
whose members are eigenvectors of T .
Proof:
Ÿ The proof should be trivial from the way we introduced rT sB .
Eigenvalues and Eigenvectors
Characteristic Polynomial
Ÿ In order to answer the question in page 1 we rst try to
compute eigenvalues and corresponding eigenvectors.
Ÿ Assume that α is an eigenvalue of the linear transformation T .
Then according to denition (1), T
αI should be singular.
Ÿ We have already seen that this is equivalent to saying that
detpT
αI q  0 (2)
Ÿ In other words in order to compute eigenvalues we may resolve
the equation:

detpT
xI q  0 (3)
in terms of the unknown x .
Eigenvalues and Eigenvectors
Characteristic Polynomial

Ÿ Assume that A  raij snn represents the matrix of T in some

ordered basis B . Then we can rewrite the equation (3) in this
coordinates:
 
x
a11
a12 ⋯
a1n

a21 x
a22 ⋯
a2n 
det 
 ⋮ 
0
⋱ ⋮

an 1
an2 ⋯ x
ann

Ÿ As can be seen from this expression detpxI
T q is a monic

polynomial of degree n in terms of x .

( We have already seen that this equation is independent of

the choice of the basis B .)
Eigenvalues and Eigenvectors
Characteristic Polynomial

Ÿ Denition 3.
The monic polynomial p px q ∶ detpxI
T q is named the
characteristic polynomial of T .
Eigenvalues and Eigenvectors
Example
Ÿ Let T ∶ R2 Ñ R2 be a self-adjoint linear transformation (w.r.t.
standard inner product). We know that the matrix of A  rT s
in the standard basis B  tp1, 0q, p0, 1qu is a symmetric matrix:

A
a b
b c

Ÿ Then we the characteristic polynomial of T is given by


x
a
b
p px q  detpxI
Aq  det  x 2
pa c qx ac
b 2

b x
c
Ÿ This polynomial has the following roots:
a
λ  pa cq  pa
c q2
2
4b 2
Eigenvalues and Eigenvectors
Example

Ÿ Thus in the case where a  c or b  0 there exists two

distincte eigenvalues λ  λ
.
Ÿ This means that there exists also two distinct engen-directions
Spantv u  Spantv
u:

Av λ v Av
 λ
v

Ÿ This means that A and tus T is diagonizable.

Eigenvalues and Eigenvectors
Example

Ÿ Let T ∶ R2 Ñ R2 be the rotation with angle θ around the

origin.
Ÿ In the standard basis tp1, 0q, p0, 1qu the matrix of T has the
following representation:

cos θ
sin θ
sin θ cos θ

Ÿ if θ  kπ for any integer k then T can not have any

eigenvector since no straight line remains invariant under the
application of T .
Ÿ And for θ  kπ , k P Z, T is not diagonalizable. (Why?)
Eigenvalues and Eigenvectors
Example

Ÿ This can also be seen by doing direct computation since we

have

x
cos θ
det
sin θ

sin θ x
cos θ  x 2
2 cos θx 1

Ÿ Thus the characteristic polynomial p px q  x 2
2 cos θx 1

has no root in R unless cos θ  1.
Eigenvalues and Eigenvectors
Diagonizability and Characteristic Polynomial

Ÿ Theorem 1.
if T ∶ V Ñ V is diagonaizable then its characteristic polynomial
p px q can split into linear factors:

p px q  px
α1 qb1 ...px
αk qbk

where α1 , ..., αk are scalars and b1 , ..., bk are positive integers with
°k

i 1 bi n
Eigenvalues and Eigenvectors

Ÿ Proof: In fact α1 , ..., αk are distinct entries on the diagonal of

rT sB in (1) and b1, ..., bk respectively represent the the
number of times they are repeated.
Ÿ Example: If
α 0 0 0 0
1

 0 α1 0 0 0
rT sB  
 0 0 α2 0 0
 0 0 0 α2 0
0 0 0 0 α2

then
detpxI
T q  px
α1q2px
α2q3
Eigenvalues and Eigenvectors
Diagonizability:
Ÿ In fact if we reorder the elements of the basis B in such a way
that the eigenvectors having identical eigenvalues lie next to
each other and if we denote the new ordered basis by B 1 then
the matrix of T with respect to B 1 is written as
 
α1 I1 0 ⋯ 0
 0 ⋯ 0 
rT sB 1   ⋮
α2 I2
⋱ ⋮

 (4)
0 0 ⋯ αk Ik

where Ij is a bj  bj identity matrix.

Ÿ From this representation we nd that the eigen space
Wj  ker pT
αj I q, 1¤j ¤k
has dimension bj . (Where bj is the number of repetition of αj
on the diagonal).
Eigenvalues and Eigenvectors

Diagonizability:
Ÿ In fact if
B 1  pu1 , .., ub1 , ..., ub1 ... bj
1 1 , ..., ub1 b2 ...bj , ..., un q
Then Wj  ub1 ... bj
1 1 , ..., ub1 ... bj ¡, 1 ¤ j ¤ k and we
have
 W1 ` ... ` Wj ` ... ` Wk
V
and also the restriction T |W ∶ Wj Ñ Wj is nothing but αj IdW
where IdW ∶ Wj Ñ Wj is the identity matrix.
j j

j
Eigenvalues and Eigenvectors
Diagonizability:
Ÿ To sum up we get to the following theorem

Theorem 2.
A linear transformation T ∶ V Ñ V from an n dimensional vector
space to itself is diagonizable i V can be decomposed into a direct
sum
V  W1 ` ... ` Wk
where Wj € V for 1 ¤ j ¤ k is a bj dimensional subspace and
T |W ∶ Wj Ñ Wj is given by
j

T |W  αj IdW , 1 ¤ j ¤ k
j j

Moreover if T is diagonizable then its characteristic polynomial

splits into linear factors as follows:
p px q  Πkj1 px
αj qbj
Eigenvalues and Eigenvectors
Example

Ÿ Consider the following 2  2 matrix:


1 1
A
0 1
Ÿ The characteristic polynomial of A consists of

x
1

1  px
1q2
p px q  det
0 x
1

Ÿ This polynomial has a double root x  1 (of multiplicity 2)

and thus A has only one eigenvalue.
Eigenvalues and Eigenvectors
Example
Ÿ In order to compute the corresponding eigenvectors we have to
resolve the following linear system of equation
pαI
AqX  0
whereα is an eigenvalue of A: in our example α  1, and
X  x1
x2
Ÿ We get to the following linear system of equations
#

x2  0
00
Ÿ Therefore we have only one eigen direction corresponding to
the eigenvalue α  1 which consists of the one dimensional
subspace Spantp1, 0qu.
Eigenvalues and Eigenvectors
Example

Ÿ From the above computation

 and proposition 1 we deduce
1 1
that the matrix A  is not diagonizable.
0 1
Ÿ In other words the fact that the characteristic polynomial splits
into linear factors is not sucient for a linear transformation to
be diagonizable.
Ÿ Yet in the next theorem we see that if the roots of the
characteristic polynomial have multiplicity one then
diagonizability holds.
Eigenvalues and Eigenvectors
Diagonizability:

Ÿ Theorem 3.
Let V be an n-dimensional vector space over a eld F . Let
T ∶ V Ñ V be a linear transformation. If T has n distinct
eigenvalues then T is diagonizable.
Eigenvalues and Eigenvectors
Diagonizability:

Proof:
Ÿ Let α1 , ..., αn denote the eigenvalues of T . We know that for
1 ¤ i, j ¤ n if i  j then αi  αj , Let vi for 1 ¤ i ¤ n be the
eigenvector corresponding to the eigenvalue αi :
Tvi  αi vi
Ÿ In order to prove the theorem it suces to show that v1 , ..., vn
are linearly independent and thus form a basis for V .
Eigenvalues and Eigenvectors
Diagonizability:
Proof:
Ÿ If v1 , ..., vn are linearly dependent then we can nd 2 ¤ j ¤ n
such that v1 , ..., vj
1 are linearly independent but
v1 , ..., vj
1 , vj are linearly dependent.
Ÿ Thus vj can be written as a linear combination of v1 , ..., vj
1 :

¸
vj  λi vi , λi P F, 1¤i ¤j
1 (5)
1¤i ¤j
1

so we get ¸
Tvj  λi Tvi
1¤i ¤j
1

or equivalently ¸
αj vj  αi λi vi
1¤i ¤j
1
Eigenvalues and Eigenvectors
Diagonizability:

Proof:
Ÿ so using (5) we obtain
¸ ¸
αj λi vi  αi λi vi
1¤i ¤j
1 1¤i ¤j
1

Ÿ since at leas one of the vectors λi vi for 1 ¤ i ¤ j
1 is non

zero and since v1 , ..., vj
1 are linearly independent the above
equation implies that
αj  αi
for some i ¤ j
1 which is a contradiction.
Ÿ Hence v1 , ..., vn should be linearly independent and the
theorem is proved.
The theory of a single linear transformation
Main Question and Jordan Canonical Form:
Ÿ From now on we try to understand as clearly as possible the
behavior of a general linear transformation on nite
dimensional vector spaces.
Ÿ Consider a nite dimensional vector space V over a eld F
and let T ∶ V Ñ V be a linear transformation.
Ÿ What we are looking for is to decompose the vector space V
as a direct sum
V  W1 ` W2 ` ... ` Wk
such that each of the subspaces Wi € V for i  1, ..., k are
invariant under the application of T :
T pW i q € W i , 1¤i ¤k
and moreover the restrictions T |Wi ∶ Wi Ñ Wi are as simple as
possible.
The theory of a single linear transformation
Minimal Polynomial:

Ÿ To do so we will see that polynomials play a crucial role.

Ÿ if we assume that dimF V  n then we know that the space of
linear transformations LpV , V q is a vector space of dimension
n2 over F .
This means that n2 1 elements I , T , T 2 , ..., T n of LpV , V q
2
Ÿ
are linearly dependent where I denotes the identity
transformation.
Ÿ It follows that there exists scalars a0 , a1 , .., an2 such that
¸
n2
ai T i 0

i 0
The theory of a single linear transformation
Minimal Polynomial:

Ÿ This means that there exists a polynomial q px q P F rx s given

°2
by q px q  ni 0 ai x i for which we have
q pT q  0

Ÿ Denition 4.
A polynomial q px q such that q pT q  0 is called an annihilating
polynomial for T .
Ÿ Consider the subset AT € F rx s of all annihilating polynomials
for T :
AT ∶ tq P F rx s|q pT q  0u
The theory of a single linear transformation
Minimal Polynomial:

Ÿ Denition 5.
A ring R is a mathematical system consisting of a nonempty set
R  ta, b, ...u together with two operations, addition and
multiplication, each of which assignes to a pair of elements a and b
in R other elements of R , denoted by a b in the case of addition
and ab in the case of multiplication, such that the following
conditions hold for all a, b, c P R
The theory of a single linear transformation
Minimal Polynomial:

1. pR, q is a commutative group.

2. multiplication is associative: pabqc  apbc q.
3. There exists an identity element 1 P R for multiplication i.e
1a  a1  a
4. apb c q  ab ac and pa b qc  ac bc
5. If the commutative law for multiplication holds (ab  ba for all
a, b P R ) Then R is called a commutative ring.
The theory of a single linear transformation
Minimal Polynomial:

Ÿ Example: Z with natural and  is a commutative ring.

Ÿ Example: F rx s the space of polynomials on a eld of scalars F
with and  in the space of polynomials is a commutative
ring.
Ÿ Example: LpV , V q the space of linear transformations on a
vector space V with and the composition of operators  as
multiplication is a (noncommutative) ring.
The theory of a single linear transformation

Minimal Polynomial:
Ÿ Denition 6.
Given a commutative ring R a subset I € R of R is named an ideal
if
1. pI , q is a group
2. for all a P I and for all r P R we have ra P I

Ÿ Example: For example the set of integers which are a multiple

of 3 form an ideal in Z.
Ÿ We will soon characterize the ideals of the F rx s.
The theory of a single linear transformation
Minimal Polynomial:

Ÿ AT is an ideal of F rx s which by denition means it has the

following two properties:
1. pAT , q is a subgroup of F rx s. This is equivalent to say that
AT is closed under addition and subtraction.
2. AT is closed under multiplication by arbitrary elements of F rx s:

@r P F rx s, @p P AT Ôñ rp P AT
Ÿ The second property for AT comes from the following lemma
The theory of a single linear transformation
Minimal Polynomial:

Ÿ Lemma 1.
If p1 , p2 P F rx s are two polynomials with coecients in F then we
have
pp1p2qpT q  p2pT q  p1pT q  p1pT q  p2pT q
Ÿ The proof is easy and left as exercise.
The theory of a single linear transformation
Minimal Polynomial:
Ÿ Let 0  m P AT be a nonzero polynomial of least degree in AT

Ÿ Theorem 4.
Every polynomial p P AT is a multiple of m i.e. for every p P AT
there exists q P F rx s such that
p  qm

Proof:
Ÿ This follows from the so called division algorithem in the space of
polynomials by which we know that for every pair of polynomials
p1 , p2 P F rx s with p1  0 there exists a unique polynomial q P F rx s
named quotient such that
p2  qp1 r

where r P F rx s named the remainder is either zero r  0 or it

satises deg pr q deg pp1 q.
The theory of a single linear transformation
Minimal Polynomial:

Proof:
Ÿ In order to prove the theorem we divide p by m to obtain the
corresponding quotient q and remainder r such that

p  mq r
with deg pr q deg pmq
Ÿ But since AT is an ideal and p, m P AT we have
r  p
mq P AT .
Ÿ Since m P AT was the nonzero polynomial of least degree we
should have r  0. This proves the theorem.
The theory of a single linear transformation

Ÿ In fact we have just proved the following theorem

Ÿ Theorem 5.
Every ideal I of the ring of polynomials F rx s is generated by a
single element i.e. there exists a P I such that
I  a ¡ tra|r P F rx su

Ÿ Corollary 1.
If m1 P AT is another nonzero polynomial of least degree then there
exists a scaler α P F such that m1  αm.
The theory of a single linear transformation

Ÿ Corollary 2.
Given a linear transformaton T P LpV , V q then there exists a
unique monic polynomial m P F rx s of least degree such that
m pT q  0

Ÿ We recall that a monic polynomial is a polynomial whose

leading coecient (the nonzero coecient of highest degree)
is equal to 1. So m has the following form
mpx q  b0 b1 x ... bk
1 x k
1 xk
The theory of a single linear transformation

Ÿ Denition 7.
Given a linear transformation T P LpV , V q the monic polynomial
mT of least degree annihilating T is called the minimal polynomial
for T .
The theory of a single linear transformation
Reminder from polynomials

Ÿ An important property of the polynomial ring F rx s where F is

an arbitrary eld is the unique factorization property.
Ÿ Denition 8.
A polynomial p P F rx s is called reducible if there exists
non-constant polynomials q1 , q2 P F rx s such that
p  q1 q2 ,

Since both q1 and q2 are non-constant we should have

1 deg q1 , deg q2 deg p .
p is called irreducible or prime if it is not reducible.
The theory of a single linear transformation
Reminder from polynomials

Ÿ Denition 9.
Here the notation r |s ("read r devides s " or "s is a multiple of r " )
for two polynomials r , s P F rx s means that there exists t P F rx s
such that
s  rt
or in other words the remainder of the division of s to r is zero.
The theory of a single linear transformation
Reminder from polynomials

Ÿ Denition 10.
Given two polynomials a, b P F rx s a common divisor of a and b
consists of a polynomial c P F rx s such that c |a and c |b.
The theory of a single linear transformation
Reminder from polynomials

Ÿ Denition 11.
Two distinct polynomials a and b are relatively prime if they do not
have any common divisor of degree greater than zero. Or in other
words their only common divisors are constant polynomials.
The theory of a single linear transformation
Reminder from polynomials

Ÿ Denition 12.
An element d P F rx s is called a greatest common divisor of
r1 , ..., rk P F rx s if d |ri , for 1 ¤ i ¤ k , and if d 1 is such that d 1 |ri ,
1 ¤ i ¤ k , then d'|d.
The theory of a single linear transformation
Reminder from polynomials

Ÿ Now consider polynomials f1 , ..., fk P F rx s. Then we set

ķ
I ∶ f1 , ..., fk ¡ t ai fi |ai P F rx s1 ¤ i ¤ k u
i 1 
Ÿ It is not dicult to see that I is an ideal of F rx s.
Ÿ By theorem in page 9 we know that there exists d P F rx s
which generates°I . In other words there exists h1 , ..., hk P F rx s
such that d  ki1 hi fi and I  d ¡.
Ÿ Since f1 , ..., fk P I this means that d is a divisor of f1 , ..., fk
The theory of a single linear transformation
Reminder from polynomials

Ÿ Theorem 6. °
There exists g1 , ..., gk P F rx s such that d  ki1 gi fi and d is the
greatest common divisor of f1 , ..., fk
Proof:
Ÿ The existence of gi 's follow from the fact that d P I
Ÿ If d 1 is a common divisor of f1 , ..., fk we should have
°
d 1 |d  ki1 gi fi .
Ÿ This proves d is the greatest common divisor of f1 , .., fk
The theory of a single linear transformation
Reminder from polynomials

Ÿ Thus far we have proved the following

Ÿ Theorem 7.
Given polynomials f1 , ..., fk P F rx s their greatest common divisor
always exists. If we denote this greatest common divisor by d then
there exists polynomials g1 , ..., gk P F rx s such that
d  g1f1 ... gk fk
The theory of a single linear transformation
Reminder from polynomials

Ÿ Theorem 8.
Given a polynomial p P F rx s there exists a factorization
p px q  p1 px qa1 ...pk px qak

where p1 , ..., pk P F rx s are distinct irreducible polynomials and

a1 , ..., ak are positive integers. Moreover the decomposition is
unique in the sense that if
p px q  q1 px qb1 ....ql px qbl

is another factorization where q1 , ..., ql are distinct irreducible

polynomials then k  l , pi  θi qi and ai  bi for i  1, ..., k , where
θi 's are scalars belonging to F .
The theory of a single linear transformation
Reminder from polynomials

Proof
Ÿ Proof: The proof of existence of a factorization into irreducible
polynomials can be done by induction and is left as exercise.
Ÿ The proof of uniqueness is an immediate consequence of the
following lemma and this also is left as exercise.
The theory of a single linear transformation
Reminder from polynomials
Ÿ Lemma 2.
Let a, b P F rx s be two arbitrary polynomials and let p P F rx s be an
irreducible polynomial. Assume that p |ab. Then either p |a or p |b
Ÿ Proof: Suppose p does not divide a. Then a and p are
relatively prime.So we have
au pv 1
for some u, v P F rx s.
Ÿ Then
abu pvb  b
and since p |ab, we have p |b.
The theory of a single linear transformation
Minimal Polynomials Toy Example:

Ÿ Now let's return to our initial problem of characterization of

the behavior of linear transformations.
Ÿ For simplicity we rst assume that mT the minimal polynomial
of T is written as
mT  pq
where p and q are polynomials which are relatively prime, by
which we mean their only common divisors are scalars.
The theory of a single linear transformation
Minimal Polynomials Toy Example:

Ÿ Theorem 9.
With the above hypothesis we have
Impp pT qq  Ker pq pT qq, Ker pp pT qq  Impq pT qq

moreover if we set W1 ∶ ker pp pT qq and W2 ∶ ker pq pT qq then

W1 and W2 are invariant subspaces of V i.e T pW1 q € W1 and
T pW2 q € W2 and we have

V  W1 ` W2
The theory of a single linear transformation
Minimal Polynomials Toy Example:

Proof:
Ÿ We rst show that
Impp pT qq X ker pp pT qq  Impq pT qq X ker pq pT qq  t0u
Ÿ To see this we note that due to the property pp, q q  1 (this is
the notation for being relatively prime) according to theorem 7
we can nd polynomials a, b P F rx s such that
ap bq  1

or
a p T qp pT q b pT qq pT q  Id (6)
The theory of a single linear transformation
Minimal Polynomials Toy Example:

Proof:
Ÿ On the other hand we know that mT pT q  p pT qq pT q  0 so

ImpP pT qq € ker pq pT qq, Impq pT qq € ker pp pT qq (7)

Ÿ Thus if v P Impp pT qq X ker pp pT qq then we have
p pT qv  q pT qv  0. But this contradicts (6)
Ÿ So we obtain

V  ker pppT qq ` ImpppT qq, (8)

(due to the property dim ker pp pT qq dim Impp pT qq  n)
The theory of a single linear transformation
Minimal Polynomials Toy Example:

Proof:
Ÿ Also from (6) we know that

Impp pT qq Impq pT qq  V (9)

Ÿ Comparing (7), (8) and (9) we can conclude that
Impq pT qq  ker pp pT qq. Similarly we can derive that
Impp pT qq  ker pq pT qq.
Ÿ Thus due to (8) the assertion of theorem 8 is proved i.e.
V  ker pppT qq ` kerqpT q
The theory of a single linear transformation
Minimal Polynomials Toy Example:
Second proof of theorem 9:
Ÿ We rst note that

Impq pT qq € ker pp pT qq and Impp pT qq € ker pq pT qq

This can be seen simply by taking an arbitrary vector v P V
and see that
p pT qpq pT qpv qq  q pT qpp pT qv q  0

Ÿ Also since p and q are relatively prime we can nd polynomials

a, b P F rx s such that

ap bq  1

which implies that

a p T qp pT q b pT qq pT q  Id
The theory of a single linear transformation
Minimal Polynomials Toy Example:
Second proof of theorem 9:
Ÿ From which we conclude that ker pp pT qq X ker pq pT qq  t0u.
Ÿ On the other hand since we have

dim Impp pT qq dim ker pp pT qq  dim Impq pT qq dim ker pq pT qq

 dim V
Ÿ Thus if dim ker pq pT qq ¡ dim Impp pT qq it would follow that
dim ker pq pT qq dim ker pp pT qq ¡ dim V

and thus

ker pq pT qq X ker ppT qq  t0u

which is a contradiction
The theory of a single linear transformation
Minimal Polynomials Toy Example:

Second proof of theorem 9:

Ÿ We deduce that

dim ker pq pT qq  dim Impp pT qq

which means ker pq pT qq  Impp pT qq similarly we have

ker pp pT qq  Impq pT qq.
Ÿ Since ker pp pT qq X ker pq pT qq  t0u we conclude
V  ker pppT qq ` ImpppT qq  ker pqpT qq ` ImpqpT qq
Ÿ The invariance of ker pp pT qq and Impp pT qq with respect to T
comes from
T  p p T q  p pT q  T
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Ÿ Consider a linear transformation T ∶ V Ñ V where V is an n

dimensional vector space over a eld F .
Ÿ Let mT P F rx s denote the minimal polynomial of T . Consider
the factorization of mT into irreducible polynomials
mT px q  p1 px qa1 ...pk px qak

Ÿ We recall that here p1 , ..., pk are distinct irreducible

polynomials in F rx s.
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Ÿ Theorem 10.
With the hypothesis of previous slide if we set Wi ∶ ker ppi pT qai q,
1 ¤ i ¤ k then we have
V  W1 ` ... ` Wk
Moreover the subspaces W1 , ..., Wk are invariant under T i.e.
T pW i q € W i , 1¤i ¤k
and the minimal polynomial mi P F rx s of T |Wi ∶ Wi Ñ Wi is equal
to
pi px qai
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Proof:
Ÿ The proof is easily done by induction and using the theorem 9.
Ÿ Note rst that if we set p ∶ pp1 qa1 and q ∶ p a2 ...p ak then p
2 k
and q are relatively prime.
Ÿ So if we set W1 ∶ ker pp q and W̃2 ∶ ker pq q then by theorem
9 we have
V  W1 ` W̃2
Ÿ Furthermore we know that W1 and W̃2 are invariant under T .
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:
Proof:
Ÿ If m1 and m̃2 denote the minimal polynomials of
T |W1 ∶ W1 Ñ W1 and T |W̃2 ∶ W̃2 Ñ W̃2 respectively, then we
know by denition of W1 and W̃2 that
m1 |p1a1 m̃2 |pp2a2 ...pkak q

and so m1 and m̃2 are relatively prime and

m1 m̃2 |mT (10)
Ÿ On the other hand due to the direct decomposition
V  W1 ` W̃2 it is obvious that m1 m̃2 is an annihilator of T .
Thus
mT |m1 m̃2 (11)
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Proof:
Ÿ By comparing (10) and (11) we conclude that mT  m1m̃2
and more precisely
m1  p1a1 m̃2  p2a2 ...pkak

Ÿ Continuing inductively the proof of the theorem can be

completed.
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Ÿ Let's assume the hypothesis of theorem 10 i.e. a linear

transformaton T ∶ V Ñ V with minimal polynomial
mT  p1a1 ...pkak and a decomposition V  W1 ` ... ` Wk .
where Wj  ker ppjaj pT qq for j  1, ..., k .
°k
Ÿ Let q1 , ..., qk P F rx s be such that 
mT
i 1 qi p ai  1. This is
i
feasible due to theorem 7.
Ÿ We set
Πj ∶ V Ñ V, Πj  qj pT q maT pT q, 1¤j ¤k
pj j

Ÿ Then we have the following theorem

The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Ÿ Theorem 11.
The operators Π1 , ..., Πk are nothing but the projections on
W1 , ..., Wk respectively with respect to the decomposition
V  W1 ` ... ` Wk . More precisely if v P V is decomposed as
v  v1 ... vk for vj P Wj , 1 ¤ j ¤ k then we have

Πj pv q  vj , 1¤j ¤k
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Proof:
Ÿ If i  j then the pia |qj mp
i T
aj . By denition of Wi 's this means
that Πj |W  0 for i  j .
j

i
°
Ÿ On the other hand since we have kj1 Πj  Id thus
Πj |Wj  IdWj .
Ÿ The combination of the above two observations proves
theorem 11.
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

Ÿ Corollary 3.
The projection operators Πj ∶ V Ñ Wj can be expressed as
polynomials in terms of T .
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:
The case of algebraically closed elds
Ÿ As a consequence of the recent theorem we nd that the
problem of understanding the behavior of an arbitrary linear
transformation T reduces to that of characterizing the
behavior of T |Wi 's.
Ÿ In other word we restrict ourselves to the cases where the
minimal polynomial mT has the form p px qa P F rx s where p px q
is an irreducible polynomial in F rx s and a is a positive integer.
Ÿ In the case where F  C we know by fundamental theorem of
algebra that the only irreducible polynomial upto multiplication
by scalars are polynomials of degree 1:
p px q  x
ξ

where ξ P C is a xed complex number.

The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

The case of algebraically closed elds

Ÿ As a special case let V be a vector space over C. Let
T ∶ V Ñ V be a linear transformation.
Ÿ In this case according to fundamental theorem of algebra the
minimal polynomial mT of T splits as
mT px q  px
α1 qa1 ...px
αk qak

If we set Wi ∶ ker ppT
αi Id qai q, where Id ∶ V Ñ V denotes

the identity map, then we know by previous theorem that
V  W1 ` ... ` Wk
The theory of a single linear transformation
Minimal Polynomials and Decomposition into Invariant Subspaces:

generalized eigenvectors
Ÿ Denition 13.
A nonzero vector v P V is a generalized eigenvector of rank m of T
corresponding to the eigenvalue α if
pT
αpId qqm v  0
In other words given an eigenvalue α P F we dene the generalized
eigenspace of V of rank m to be equal to ker pT
αpId qqm and the
nonzero elements of ker pT
αpId qqm are called generalized
eigenvectors.
Ÿ Note that for any positive integer m the subspace ker pT
αpId qqm
can be non trivial only if α is an eigenvector of T . (Prove this!)
The theory of a single linear transformation
Nilpotent Maps

Ÿ As can be seen from theorem 10 and our recent discussion in

the case where F is algebraically closed and in particular in the
case where F  C, the problem of characterization of arbitrary
linear operators T ∶ V Ñ V reduces to the characterization of
T |Wj where Wj  ker pT
αj I qaj for some eigenvalues αj and
some positive integers aj .
Ÿ This means that pT
αj I q|Wj is a nilpotent operator on Wj ,
1 ¤ j ¤ k.
Ÿ Thus if we can characterize nilpotent operators then we will be
able to understand the behavior of T |Wj .
Ÿ This is because the dierence between pT
αj I q|Wj and T |Wj
is just αj I which has a trivial behavior.
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ There is a close relation between nilpotent maps and triangular

matrices.
Ÿ Denition 14.
A square n  n matrix A is called upper triangular if all the entries
below the main diagonal are zero
 
a11 a12 ⋯ a1n
 0 a22 ⋯ a2n 
A
 

⋮ ⋱ ⋮
0 0 ⋯ ann
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Now assume that A is an upper triangular matrix such that all

the entries on the main diagonal vanish: a11  ...  ann  0.
Ÿ Then we claim that A is a nilpotent matrix with index at most
n:
An  0
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ In fact a simple computation shows that if we multiply an

upper triangular matrix A with vanishing main diagonal by
itself the result will have zero entries both on its main and on
its super diagonal.
Ÿ Similarly Aj should have zero entries on its i -th superdiagonal
for i ¤ j
1.
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Let's write this argument more rigorously. Consider an n  n

matrix A  raij s such that
aij 0 if j ¤ i
Ÿ Let Ak  rbijk s for k ¥ 2 and 1 ¤ i, j ¤ n. Then we have
bijk  0 if , j ¤ i k
1
Ÿ Prove the above claim.
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Corollary 4.
Any upper triangular matrix with vanishing main diagonal:
 
0 a12 ⋯ a1n
 0 0 ⋯ a2n 
rT sB   ⋮ ⋱ ⋮

 (12)
0 0 ⋯ 0
is nilpotent.
We will see that the inverse of this is also true in other words
any nilpotent linear operator in appropriate basis has a
representation like (12).
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Let T ∶ V Ñ V be a linear transformation on an n-dimensional

vector space V .
Ÿ Let the matrix of T with respect to an ordered basis
B  pv1 , ..., vn q has the following form:
 
0 a12 ⋯ a1n
 0 0 ⋯ a2n 
rT sB   ⋮ ⋱ ⋮

 (13)
0 0 ⋯ 0
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ If we set Wj  v1 , ..., vj ¡ for j  1, ..., n, and W0  t0u

then we have
1.
t0u  W0 € W1 € ... € Wn  V
2. and Tvj P Wj
1 for 1 ¤ j ¤ n thus

T pWj q € Wj
1 for 1 ¤ j ¤ n
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Theorem 12.
There exists a basis B for the linear application T ∶ V Ñ V with
respect to which the matrix representation rT sB has the upper
triangular form (13) if and only if there exists nested subspaces
t0u  W0 ˆ W1 ˆ ... ˆ Wn
1 ˆ Wn  V such that
T pW j q € W j
1 for 1 ¤ j ¤ n

Proof:
Ÿ The proof is easy and left as exercise.
The theory of a single linear transformation
Nilpotent maps and triangular matrices
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Theorem 13.
Let V be a vector space dened over an algebraically closed eld F .
Let T ∶ V Ñ V be a linear transformation. Then there exists an
ordered basis B for V with respect to which the matrix
representation rT sB has the following form:
 A1 0

 ⋱ 

 Aj 
  (14)

 ⋱
0 Ak
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Where Aj 1 ¤ j ¤ k is a bj  bj upper triangular matrix like

 
αj  ⋯ 
 0 αj ⋯  
Aj  
 
 (15)
⋮ ⋱ ⋮
0 0 ⋯ αj
such that Aj
αj Ij is a nilpotent bj  bj matrix of index aj :
pAj
αj Ij qaj  0
Ÿ It is obvious that
aj ¤ bj (16)
Eigenvalues and Eigenvectors
Diagonizability and minimal polynomials:

Ÿ Theorem 14.
A linear transformation T P LpV , V q is diagonizable i the minimal
polynomial of T has the form
mT px q  px
α1 q...px
αk q

where α1 , ..., αk are distinct elements of F .

Eigenvalues and Eigenvectors
Diagonizability and minimal polynomials::
Proof:
Ÿ If T is diagonizable we know from theorem 2 that there exists
a decomposition of V  `kj1 Wj into a direct sum of
subspaces W1 , ..., Wk such that
T |Wj  αj IdW j

for some distinct scalers α1 , ..., αk P F .

Ÿ In other words

pT
αj Id q|W  0j
1¤j ¤k
thus
pT
α1Id q  ...  pT
αk Id q  0
Eigenvalues and Eigenvectors
Diagonizability and minimal polynomials::
Proof:
Ÿ This means that
mT px q|px
α1 q...px
αk q

Ÿ On the other hand if each of the prime factors x
αj

1 ¤ j ¤ k is deleted the obtained polynomial
Π k p x
α q
m px q  i px1
αj q i does not annihilate T .
Ÿ in fact if vj P Wj then Tvj  αi vj so pT
αi I qvj  pαj
αi qvj .
Ÿ so m pT qvj  Πi j pαj
αi qvj  0
Ÿ This shows that

mT px q  px
α1 q...px
αk q
Eigenvalues and Eigenvectors

Diagonizability and minimal polynomials::

Proof:
Ÿ Conversely assume that the minimal polynomial mT of T has
the form
mT px q  px
α1 q...px
αk q
where α1 , ..., αk are distinct scalars.
Ÿ Then by theorem 10 we know that V  W1 ` ... ` Wk where
Wj  ker pT
αj I q, 1 ¤ j ¤ k

Ÿ Thus by theorem 2, T is diagonizable.

Eigenvalues and Eigenvectors

Diagonizability and minimal polynomials::

Ÿ Theorem 15.
Let T1 , ..., Tk be a set of diagonizable linear transformations on V
such that Ti Tj  Tj Ti for 1 ¤ i, j ¤ k . Then there exists a basis of
V such that the basis B for V with respect to which all the
transformations rT1 s, ..., rTk s are diagonal matrices.
Eigenvalues and Eigenvectors
Diagonizability and minimal polynomials::
Proof:
Ÿ We proceed the proof by induction on k .
Ÿ Let α1 , ..., αr denote the distinct eigenvalues of T1 and let
Wj  ker pT
αj I q 1 ¤ j ¤ k be the eigenspace corresponding
to eigenvalue αj . We know from theorem 2 that
V  W1 ` ... ` Wk .
Ÿ The fact that Tj 's commute with each other implies that the
subspaces W1 , ..., Wr are invariant with respect to all the
operators T1 , ..., Tk . This is because if vj P Wj then we have
T1 Ti vj  Ti T1vj  Ti pαj vj q  αj Ti vj
which means that Ti vj is also an eigenvector of T1 with
eigenvalue αj thus it belongs to Wj
Eigenvalues and Eigenvectors
Diagonizability and minimal polynomials::
Proof:
Ÿ Now we consider the operators S2 , .., Sk ∶ W1 Ñ W1 dened by
Si ∶ Ti |W1 ,2 ¤ i ¤ k
Ÿ We claim that S2 , ..., Sk are diagonizable. This is because for
each 2 ¤ i ¤ k the minimal polynomial mSi of Si divides the
minimal polynomial of mTi of Ti :
mSi |mTi 2¤i ¤k
Ÿ Now we can apply theorem 14 and the hypothesis on
diagonizability of Ti 's to deduce that Si 's are also diagonizable.
Ÿ Since obviously S2 , ..., Sk commute with each other by
induction hypothesis we can infer that S2 , ..., Sk are
simultaneously diagonizable.
Ÿ Repeating this for all the subspaces W1 , ..., Wr completes the
proof.
Ÿ
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Theorem 16 (Cayly Hamilton Theorem For

Algebraically Closed Fields).
Let T be a linear transformation on a nite dimensional vector
spaceV over an algebraically closed eld F , then the minimal
polynomial mT divides the characteristic polynomial p px q:
mT |p

In particular P pT q  0.
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Proof: The proof is more or less obvious from theorem 13.

Due to relation (14) we have
detpxI
T q  detpxI1
A1q  ...  pxIk
Ak q
Ÿ Also from (15) we have
detpxIj

Aj q  px
αj qb j

Ÿ The inequality (16) shows also that aj ¤ bj for 1 ¤ j ¤ k .

Ÿ The proof is complete by recalling that
mT px q  Πkj1 pxj
α j qa j
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Ÿ Theorem 17 (Jordan Decomposition).

Let T P LpV , V q, where V is a nite dimensional vector space over
an algebraically closed eld. Then there exists linear transformation
D and N on V such that
a) T  D N
b) D is diagonizable and N is nilpotent.
c) There exist polynomials f px q and g px q P F rx s such that
D  f pT q and N  g pT q
The transformations D and N are uniquely determined in the sense
that if D 1 and N 1 are diagonizable and nilpotent transformations,
respectively, such that T  D 1 N 1 and D 1 N 1  N 1 D 1 , then
D  D 1 and N  N 1 .
The theory of a single linear transformation
Nilpotent maps and triangular matrices
Proof:
Ÿ We have already seen that if mT P F px q denotes the minimal
polynomial of T then since F is algebraically closed we have
the decomposition
mT px q  px
α1 qa1 ...px
αk qak

and if we set Wj ∶ ker pT
αj I qaj 1 ¤ j ¤ k then

V  W1 ` ... ` Wk and according to corollary 3 the projection
maps
Πj ∶ V Ñ W j , m 1 ¤ j ¤ k
are polynomials in terms of T . So there exists fi px q P F rx s
such that

Πj  fj pT q, 1¤j ¤k (17)
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Proof:
Ÿ Now if we dene D; V ÑV as
ķ
D ∶ αj Πj

j 1

then D is diagonizable since D |Wj  αj IdWj and more over

V  `kj1 Wj .
°
Ÿ Also if we wet f px q  kj1 αj fj px q then due to (17) we have
D  f pT q
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Proof:
Ÿ We set N ∶ T
D . Then obviously N  g pT q for
g px q  x
f px q.
Ÿ Also if a ∶ max1¤j ¤k taj u then we claim that N a  0.
Ÿ This is because we know that

N |Wj  pT
αj Id q|W j

thus N |aWj j  0 by denition of Wj . In particular since aj ¤a

for all 1 ¤ j ¤ k we have

pN |W qa  0 @1 ¤ j ¤ k
j

Ÿ Since V  `kj1 Wj we conclude that N a  0

The theory of a single linear transformation
Nilpotent maps and triangular matrices

Proof:
Ÿ In order to prove uniqueness part assume that T  D 1 N1
1 1
and D and N satisfy the hypothesis of the last part of the
theorem.
Ÿ Since D 1 N 1  N 1 D 1 we have TD 1  D 1 T and TN 1  N 1 T .
Ÿ Since D and N are polynomials in terms of T we can deduce
that DD 1  D”D and NN 1  N 1 N .
Ÿ Thus from T  D 1 N 1  D N it follows that

D1
D  N
N1
The theory of a single linear transformation
Nilpotent maps and triangular matrices
Proof:
Ÿ Due to relation
¸ 
pN
N 1 q s  n N j N 1 s
j
j
which comes from commutativity of N and N 1 we can deduce
that if N a  0 and N 1 b  0 then pN
N 1 qa b  0
Ÿ on the other hand due to theorem 15 D and D 1 are
simultaneously diagonizable we can nd a basis B with respect
to which both D and D 1 and thus D
D 1 is diagonizable
 β1 0

 ⋱ 

 βj 
  (18)

 ⋱
0 βk
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Proof:
Ÿ This means that D
D 1 can be nilpotent i D  D 1 and this
completes the uniqueness part of the theorem.
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Examples:
Ÿ Exercise. Prove that the minimal polynomial and the
characteristic polynomial of a linear transformation T ∶ V Ñ V
have the same roots. In particular if p px q  Πki1 px
αi qbi
can be decomposed into linear factors then
mT px q  Πki1 px
αi qai where 1 ¤ ai ¤ bi for all 1 ¤ i ¤ k
The theory of a single linear transformation
Nilpotent maps and triangular matrices
Examples:
Ÿ Let V be a three dimensional vector space over C with basis
tv1, v2, v3u. And let T P LpV , V q be dened by the equations:
Tv1 
v1 2v3
Tv2  3v1 2v2 v3
Tv3 
v3

Ÿ Thus the matrix of T with respect to pv1 , v2 , v3 q is given by



1 3 0
0 2 0
2 1
1
Ÿ Find the distinct eigenvalues of T .
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Examples:
Ÿ To this end we can proceed as follows. The characteristic
polynomial p px q of T is computed as follows
RRRx 1
3 0 RRRR
RRR R
p px q  detpxI
Aq  RRR 0 x
2 0 RRRR  px 1 q2 p x
2 q
RRR
RR
2
1 x 1RRRRR
Ÿ Thus the distinct eigenvalues of T are t
1, 2u. Also the
minimal polynomial of T is either
p1 x qp2
x q or p1 x q2 p 2
x q
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Examples:
Ÿ Now we consider A I and A
2I
 
0 3 0
3 3 0
A 
I 0 3 0 , A
2I 
 0 0 0
2 1 0 2 1
3
Ÿ It is not dicult to see that rank pA I q  rank pA
2I q  2
and thus dim ker pA I q  dim ker pA
2I q  1
Ÿ This means that T is not diagonizable and so we have
mT px q  p1 x q 2 p2
x q
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Examples:
Ÿ In order to nd the Jordan decomposition we need to nd
ker pT I q2 and ker pT
2q.
Ÿ It is an exercise to see that ker pT I q2  Spantv1 , v2 u and
ker pT
2I q  Spantpv1 v2 v3 qu.
Ÿ We set

W1  Spac tv1 , v2 u and W2  Spantv1 v2 v3 u

Ÿ According to our previous discussion our desired matrix

representation for T is obtained through taking appropriate
basis for W1 and W2 .
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Examples:
Ÿ More precisely we look for a basis tw1 , w2 u for W1 such that

pT I qw 1  0 and pT I qpw2 q  Spantw1 u

Ÿ we know that Tv1 
v1 2v3 and Tv3 
v3 . Thus

pT I qv1  2v3 and pT I qv3  0.
Ÿ Thus if we set w1  v3 and w2  v1 we get
pT I qw1  0 and pT I qw2  2w1
Ÿ If we set w3  v1 v2 v3 we also have
Tw3  2w3 .
The theory of a single linear transformation
Nilpotent maps and triangular matrices
Examples:
Ÿ Consequently if we set B 1  pw1 , w2 , w3 q then



1 2 0
rT sB 1   0
1 0
0 0 2
Ÿ We can also see easily that the matrix of change of
coordinates P consists of

0 1 1
0 0 1
1 0 1
and we have
rT sB 1  P
1AP
The theory of a single linear transformation
Nilpotent maps and triangular matrices

Examples:
Ÿ In this example we have
 

1 0 0 0 2 0
D   0
1 0 and N  0 0 0
0 0 2 0 0 0
The theory of a single linear transformation
Jordan Canonical Form

Ÿ Consider a linear transformation T ∶ V Ñ V . Let α P F be a

scalar such that T
αI is a nilpotent linear transformation of
index k i.e.
pT
αI qk  0
and k is the smallest positive integer with this property.
Ÿ Let S ∶ V Ñ V be the operator S  T
αI and we set
Vj  tv P V |S j pv q  0u  ker pS j q, 1 ¤ j ¤ k
Ÿ Obviously we have V1  kerpS q and Vk  V and
t0u  V0 € V1 € V2 € ... € Vk
1 € Vk  V
The theory of a single linear transformation
Jordan Canonical Form

Ÿ In fact we have
V1  S
1 pt0uq, V2  S
1 pV1 q, ..., Vi  S
1pVi
1q for 1 ¤ i ¤ k
The theory of a single linear transformation
Jordan Canonical Form

Ÿ We have Vk
1  ker pS k
1 q. Let Wk
1 be a complement of

Vk
1 in V :
V  Vk
1 ` Wk
1
Ÿ Then for every v P Wk
1 we have S k
1 pv q  0
Ÿ Thus
S pWk
1 q X Vk
2  t0u
where Vk
2  ker pS k
2 q
Ÿ Let Wk
2 be a complement of Vk
2 in Vk
1 containing
S pW k
1 q :

Vk
1  Vk
2 ` Wk
2 , S p W k
1 q € W k
2
The theory of a single linear transformation
Jordan Canonical Form

Ÿ Continuing inductively we assume that the subspaces

Wk
1 , ..., Wi € V are constructed such that

Vj  Vj
1 ` Wj
1 i 1¤j ¤k
where Vj  ker pS j q and satisfying

S pWj
1 q € Wj
2 , for i 2¤j ¤k

Ÿ Then S pWi q € Vi is such that S pWi q X Vi
1  t0u
Ÿ We then construct a complement Wi
1 for Vi
1 in Vi
comtaining S pWi q:
Vi  Vi
1 ` Wi
1 S pWi q € Wi
1
The theory of a single linear transformation
Jordan Canonical Form

Ÿ Consquently we obtain
S pWk
1 q ↪ Wk
2 , S pWk
2 q ↪ Wk
3 , ..., S pW2 q ↪ W1

Ÿ Also S pW1 q ↪ V1 and V1  ker pS q. We also set W0 ∶ V1 .

Jordan Canonical Form
The theory of a single linear transformation
Jordan Canonical Form

Ÿ Assume that dim Wj  dj for 0 ¤ j ¤ k
1.

Ÿ Let Bj  pvj 1 , vj 2 , ..., vjdj q be an ordered basis for Wj for
0 ¤ j ¤ k
1 such that
Svji  vj
1i for 1 ¤ j ¤ k
1, 1 ¤ i ¤ dj
Ÿ As an exercise prove that such bases exists.
The theory of a single linear transformation
Jordan Canonical Form

Ÿ We know that B0 , ..., Bk
1 are disjoint and their union

B  Ykj 
11 Bj forms a basis for V .
Ÿ Consider chains of maximal lenght of the form
pv , Sv , S 2v , ..., S l v q in B .
Ÿ for example pv01 , v21 , ..., vk
11 q is one such chain.
Ÿ In fact we have dk
1 chains of length k and we have pd2
d1 q
chains of length k
1...and we have d0
d1
...
dk
1
chains of lenth 1. (Why?)
The theory of a single linear transformation

Ÿ The partition of B into these chains gives rise to a

decomposition of V into a direct sum
V  E1 ` ... ` Er
where each Ej , 1 ¤ j ¤ r corresponds to the subspace
generated by the above mentioned chains.
Ÿ For example E1 , v01 , v21 , ..., vk
1,1 ¡ etc.
The theory of a single linear transformation
Jordan Canonical Form

Ÿ It is obvious that the subspaces E1 , ..., Er are invariant under

the application of S . And thus the matrix of S with respect to
the basis B has the following block form
 C1 0

 ⋱ 

 Cj 
  (19)

 ⋱
0 Cr
The theory of a single linear transformation

Ÿ where  0 1 0 ⋯ 0


 0 0 1 ⋯ 0 

Cj  
 ⋮ ⋮ ⋱ ⋱ ⋮ 
 (20)
 0 0 ⋯ ⋱ 1 
0 0 ⋯ 0
The theory of a single linear transformation

Ÿ
Ÿ  αj 1 0 ⋯ 0


 0 αj 1 ⋯ 0 

Aj  
 ⋮ ⋮ ⋱ ⋱ ⋮ 
 (21)
 0 0 ⋯ ⋱ 1 
0 0 ⋯ αj
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Let V be a real n dimensional vector space. We dene a

complex vector space V b C as follows:
V b C ∶ tu iv |u, v P Vu
Ÿ The addition and scalar multiplication are dened by
pu iv q pu1 iv 1q ∶ pu u1q i pv v 1q

for u iv , u 1 iv 1 P V b C

pa ibqpu iv q  pau
bv q i pbu av q

for pa ib q P C and u iv P V b C
The theory of a single linear transformation
Complexication of a real vector space

Ÿ It is not dicult to see that V b C with the above operations

is a vector space over C. This is called complexication of the
vector space V .
Ÿ Now if T ∶ V Ñ V is an R-linear transformation we can dene
Tc ∶ V bCÑV bC
by
Tc pu iv q ∶ Tu iTv
we have
Tc ppa ib qpu iv qq  T pau
bv q iT pav bu q  pa ib qTc pu iv q

Tc ppu iv q pu 1 iv 1 qq  Tc pu iv q Tc pu 1 iv 1 q
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Thus Tc is a C-linear transformation on V b C

Ÿ Example If V  Rn then V b C is nothing but Cn . Also if A is
an n  n matrix with real coecients then the complexication
of A is represented by the same matrix A.
The theory of a single linear transformation
Complexication of a real vector space

Ÿ If W̃ € V b C is a complex subspace of V b C then we dene

¯
W̃  tu iv |u iv P W u  tu
iv |u iv P Wu
Ÿ ¯ is also a complex subspace of V b C. To see this we take a
W̃
vector u iv P W̃ and we note that
pa ib qpu
iv q  pa
ib qpu iv q

and pa
ibqpu iv q P W̃ . Also it is not dicult to see that W̃

¯
is closed under addition.
Ÿ
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Several cases migh occur:

¯
W̃  W̃ or
¯
W̃ X W̃  t0u or t0u ˆ W̃¯ X W̃ ˆ W̃
Ÿ If W € V is a real subspace of V then W b C is a complex
subspace of V b C. and we have
W bCW bC
Ÿ This is not dicult to verify but we can also prove the
converse.
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Lemma 3.
If W̃ € V b C is a complex subspace such that W̃ ¯  W̃ then there
exists a subspace W € V such that W̃  W b C.
Ÿ Proof: In fact for every u iv P W̃ we have u P W̃ . This is
because u
iv P W̃ .
Ÿ Also
i pu iv q  v
iu P W̃ thus we also have v P W̃ .
Ÿ Therefore if we set W ∶ W̃ X V then W̃ € W b C and thus
we get
W̃ W bC
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Now let u1 , ..., uk , v1 , ..., vk be 2k linearly independent vectors

in V .
Ÿ Set
W̃ ∶ Spantu1 iv1 , ..., uk bC ivk u € V
by which we mean W̃ is the complex subspace of V b C:
ķ
W̃  t p aj ibj qpuj ivj q|aj , bj P R for 1 ¤ j ¤ k u

j 1
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Lemma 4.
W0 ∶ W̃ X W̃¯  t0u
Ÿ Proof: By denition we have W̄0  W0 so according to
previous lemma there exists a a subspace V0 € V such that
W0  V0 b C.
Ÿ Now if the complex numbers a1 ib1 , ..., ak ibk P C are such
that
ķ
v  paj ibj qpuj ivj q P V0

j 1

Ÿ Or equivalently
¸ ?
rpaj uj
bj vj q
1 p aj v j bj uj qs P V0
The theory of a single linear transformation
Complexication of a real vector space

Ÿ Then we should have

¸
paj vj bj uj q  0

Since the family of vectors tu1 , ..., uk , v1 , ..., vk u is linearly

independent this implies that
aj  bj  0 @1 ¤ j ¤ k
Ÿ Thus v  0 and this completes the proof.
The theory of a single linear transformation
Complexication of a real vector space
Ÿ A converse for lemma 4 also holds in the sense that if
W̃ € V b C is a complex vector subspace satisfying

X W̃¯  t0u
W̃
and if tu1 iv1 , ..., uk ivk u constitute a basis for W̃ then the
family of vectors tu1 , ..., uk , v1 , ..., vk u are linearly independent.
Ÿ this is because due to the identity
ķ ¸ ?
paj ibj qpuj ivj q  rpaj uj
bj vj q

1paj vj bj uj qs

j 1

if tu1 , ..., uk , v1 , ..., vk u are linearly dependent we can nd

a1 ib1 , ..., ak ibk such that
ķ
w  paj ibj qpuj ivj q P V

j 1
The theory of a single linear transformation
Complexication of a real vector space

Ÿ It follows that wP W̃ X W̃¯

The theory of a single linear transformation
Orthogonal Transformations

Ÿ Now consider an inner product vector space pV , x, yq. Let

T ∶ V Ñ V be an orthogonal transformation.
Ÿ Let Tc ∶ V b C Ñ V b C be the corresponding
complexication of T .
Ÿ Let α  a ib P C be a complex number with b  0. and
assume that α is an eigenvalue of Tc .
Ÿ Since the characteristic polynomial of Tc and T are identical
we know that ᾱ is also an eigenvalue of Tc .
The theory of a single linear transformation
Orthogonal Transformations

Ÿ
Ÿ Also since α  ᾱ the eigen spaces W1  ker pT
αI q and
W̄1  ker pT
ᾱI q satisfy

W1 X W̄1  t0u
The theory of a single linear transformation
Orthogonal Transformations

Ÿ Note that
Tc pu iv q  pa ib qpu iv q ðñ Tc pu
iv q  pa
ib qpu
iv q
which implies that ker pT
ᾱI q  W̄1 .
Ÿ Now take an eigenvector u iv P W1. Then tu, v u is linearly
independent and we have
Tc p u iv q  pa ib qpu iv q

which is equivalent to
Tu  au
bv , Tv  bu av

Ÿ Thus the 2-dimensional real subspace Spantu, v u € V is

invariant under the application of T .
The theory of a single linear transformation
Orthogonal Transformations

Ÿ Since T is an orthogonal transformation if we set

W2  Spantu, v u then the orthogonal complement W K is also
invariant under T .
Ÿ For real eigenvalues λ P R of Tc Due to orthogonality one can
see that λ  1. If v P V denote the eigenvector associated
to λ and V1 ∶ Spantv u then we can again decompose V like
V  V1 ` V1K
where V1K is the orthogonal complement of V1 . We also know
that V1K is invariant under V .
Ÿ So by induction we get to the following theorem:
The theory of a single linear transformation
Orthogonal Transformations

Ÿ Theorem 18.
Let pV , x, yq be an inner product vector space and let T ∶ V Ñ V be
an orthogonal transformation. Then there exists a decomposition
V  W1 ` ... ` Wk ` V1 ` ... ` Vl
into subspaces
dim Wi  2, 1¤i ¤k dim Vj 1 1¤j ¤l
Each of the vector spaces Wi 's and Vj 's are invariant under the
application of T and they are mutually orthogonal subspaces of V .
The theory of a single linear transformation
Orthogonal Transformations
Ÿ Corollary 5.
With the hypothesis of the previous theorem there exists an
orthonormal base B for V such that the matrix A  rT sB in this
basis has the form
 C1 0

 ⋱ 

 Cj 
  (22)

 ⋱
0 Cr

where C1 , ...Ck are 2  2 rotation matrices


Cj  cos θj
sin θj 1¤j ¤k
sin θj cos θj
The theory of a single linear transformation
Orthogonal Transformations

Ÿ And Ci  1 for k+1¤ i ¤ k l n

The theory of a single linear transformation
Self-Adjoint Transformations

Ÿ Let pV x, yq be an inner product vector space and let

T ∶ V Ñ V be a self-adjoint linear transformation which by
denition satises:
xTu, v y  xu, Tv y
Ÿ Theorem 19.
With the above hypothesis T is diagonizable in an orthogonal basis.
This completes the proof of the theorem.
The theory of a single linear transformation
Self-Adjoint Transformations

Ÿ Proof: Let Tc ∶ V b C Ñ V b C be the complexication of T

as dened before.
Ÿ Let p px q be the characteristic polynomial of T which is also
the characteristic polynomial of Tc .
Ÿ We rst claim that all the roots of p are real. Consequently
both T and Tc have n real eigenvalues (taking into account
the multiplicities of these eigenvalues.)
The theory of a single linear transformation
Self-Adjoint Transformations

Ÿ To see this we assume that α  a ib is an eigenvalue of Tc .

Let also u iv P V b C be the eigenvector associated to α.
Ÿ so we have:
Tc pu iv q  pa ib qpu iv q
Ÿ By denition of Tc this is equivalent to say that
Tu  au
bv Tv  bu av (23)
Ÿ This relation means that the subspace W  Spantu, v u is
invariant under the application of T :
T pW q € W
The theory of a single linear transformation
Self-Adjoint Transformations

Ÿ If dim W  1 then there exists a real λ P R such that u  λv .

Ÿ In this case by (23) we nd that
T pλv q  paλ
b qv T pv q  pa bλqv

Ÿ It is not dicult to see from the above equations that b  0.

So α P R
Ÿ If dim W  2 then we use xTu, v y  xu, Tv y to obtain
axu, v y
b xv , v y  b xu, u y axu, v y

We can again conclude that b  0.

The theory of a single linear transformation
Self-Adjoint Transformations

Ÿ Consequently all the eigenvalues of T are real. Let now α P R

be an eigenvalue of T and let v1 P V denote the associated
eigenvector
Tv1  αv1
Ÿ Now let W0  Spantv1 uK . Then W0 is invariant under the
application of T .
Ÿ This is because if w P W0 then
xTw , v1y  xw , Tv1y  αxv1, w y  0 for all w P W0
Ÿ Thus Tw P W0
The theory of a single linear transformation
Self-Adjoint Transformations

Ÿ T |W0 ∶ W0 Ñ W0 is also self adjoint and by repeating the

above argument we can obtain an orthogonal basis of
eigenvalues for T like B  tv1 , ..., vn u.
The theory of a single linear transformation
Polar decomposition theorem

Ÿ Theorem 20 ( Polar decomposition theorem).

Consider Rn with its standard inner product x, y. Let T ∶ Rn Ñ Rn
be an invertible linear transformation. Then there exists an
orthogonal transformation U ∶ Rn Ñ Rn and a self-adjoint operator
S ∶ Rn Ñ Rn such that
T  SU
The theory of a single linear transformation
Polar decomposition theorem

Ÿ Proof: We rst not that if such a decomposition exists then

we can write
TT t  SUU t S t  S 2
Here we are using the orthogonality of U and self-adjointness
of S
?
Ÿ Thus in some sense S  TT t . This has sense since TT t is a
positive denite symmetric matrix.
Ÿ By our last theorem TT t is diagonizable in an orthonormal
basis which means that there exists an orthogonal matrix P
such that
 P t DP
TT t
where D is a dianal matrix like D  diag pλ1 , ..., λn q where
λ1 , ..., λn are the eigenvalues of TT t .
The theory of a single linear transformation
Polar decomposition theorem

Ÿ Also since TT t is non-singular and positive denite λi 's are

positive.
Ÿ Thus if we set
a a
S ∶ P t diag p λ1 , ..., λn qP

then we have S 2  TT t .
Ÿ Now if we set U ∶ S
1 T then we can write
UU t  pS
1T qpT t pS
1qt q  S
1pS 2qS
1  I
Ÿ This means that U is orthogonal.