MA110:

Lecture 11

Saurav Bhaumik
Department of Mathematics
IIT Bombay

Spring 2025

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Similarity of Square Matrices
Definition
Let A, B ∈ Kn×n . We say that A is similar to B (over K) if
there is an invertible P ∈ Kn×n such that B = P−1 AP, that is,
AP = PB. In this case, we write A ∼ B.
One can easily check (i) A ∼ A, (ii) if A ∼ B then B ∼ A,
and (iii) if A ∼ B and B ∼ C, then A ∼ C.
Examples:    
4 −3 3 1
(i) Let A := . It is easily seen that P := is
2 −1   2 1
1 −1
invertible and . P−1 = . Hence A is similar to
−2 3
        
1 −1 4 −3 3 1 1 −1 6 1 2 0
B := = = ,
−2 3 2 −1 2 1 −2 3 4 1 0 1
which is a diagonal matrix.
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
More Examples and a Characterization of Similarity

(ii) Let A ∈ Kn×n . Then A ∼ I ⇐⇒ A = I.

(iii) Let A ∈ Kn×n , and let E be n×n an elementary matrix.
Then B := EAE−1 is similar to A. Note: EA is obtained from
A by an elementary row operation on A, and B is obtained
from EA by the ‘reverse column operation’ on EA.

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Diagonalizability

Definition
A matrix A ∈ Kn×n is called diagonalizable (over K) if A is
similar to a diagonal matrix (over K).

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Examples
   
4 −3 3 1
(i) Let A := . Let P := , so that
 2 −1 2 1
−1 1 −1
P := . Then A is similar to the matrix
−2 3
        
1 −1 4 −3 3 1 1 −1 6 1 2 0
B := = = ,
−2 3 2 −1 2 1 −2 3 4 1 0 1

which is a diagonal matrix.

(ii) Let A ∈ Kn×n . Then A ∼ I ⇐⇒ A = I.
(iii) Let A ∈ Kn×n , and let E be n×n an elementary matrix.
Then B := EAE−1 is similar to A. Note: EA is obtained from
A by an elementary row operation on A, and B is obtained
from EA by the ‘reverse column operation’ on EA.

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Similarity of matrices has the following characterisation.
Proposition
Let A, B ∈ Kn×n . Then A ∼ B if and only if there is an
ordered basis E for Kn×1 such that B is the matrix of the
linear transformation TA : Kn×1 → Kn×1 with respect to E .
In fact, B = P−1 AP if and only if the columns of P form an
ordered basis, say E , for Kn×1 and B = MEE (TA ).

Proof. Let B := [bjk ]. Now A ∼ B ⇐⇒ there is an invertible

matrix P such that AP = PB. Let xk be the k-th column of
P. Then E := (x1 , . . . , xn ) is an ordered basis for Kn×1 .
Conversely, if E = (x1 , . . . , xn ) is an ordered basis then the
matrix P = [x1 , . . . , xn ] is invertible. Now AP = PB implies:
 
b11 · · · b1n
A x1 · · · xn = x1 · · · xn  ... .. ..  .
   
. . 
bn1 · · · bnn
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
The kth column of LHS is Axk and the kth column of RHS is
the linear combination of x1 , . . . , xn with coefficients from the
k column of B. Thus Axk = b1k x1 + · · · + bnk xn for
k = 1, . . . , n. This means the kth column of MEE (TA ) is the
 T
kth column b1k · · · bnk of B, k = 1, . . . , n, that is,
B = MEE (TA ).
Conversely, if B = MEE (TA ) then again
Axk = TA (xk ) = b1k x1 + · · · + bnk xn for k = 1, . . . , n, so
AP = PB where P = [x1 , . . . , xn ].

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Proposition
A matrix A ∈ Kn×n is diagonalizable if and only if there is a
basis for Kn×1 consisting of eigenvectors of A.
In fact, X−1 AX = D, where X := x1 · · · xn and
 

D := diag(λ1 , . . . , λn ) ⇐⇒ {x1 , . . . , xn } is a basis for Kn×1

and Axk = λk xk for k = 1, . . . , n.

Proof. A is diagonalizable ⇐⇒ there is an invertible matrix

X and a diagonal matrix D such that AX = XD. This is the
case if and only if there is a basis {x1 , . . . , xn } for Kn×1 and
there are λ1 , . . . , λn ∈ K such that
   
A x1 · · · xn = x1 · · · xn diag(λ1 , . . . , λn ),

The LHS is just [Ax1 , · · · , Axn ] and the RHS is just

[λx1 , · · · , λn xn ]. Thus the proposition follows.

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

If A = XDX−1 , then for any m ∈ N,

Am = (XDX−1 ) · · · (XDX−1 ) = XDm X−1 = X diag(λm m −1

1 , . . . , λn )X .

Example
We have seen that
     −1
4 −3 3 1 2 0 3 1
A := =
2 −1 2 1 0 1 2 1
 −1  
3 1 1 −1
and = . Hence
2 1 −2 3
     m  
m 3 1 2m 0 1 −1 2 3 1 1 −1
A = = m
2 1 0 1 −2 3 2 2 1 −2 3
 m 
2 3 − 2 −2m 3 + 3
= for m ∈ N.
2m 2 − 2 −2m 2 + 3

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Similarity and Eigenvalues
Recall that λ ∈ K is an eigenvalue of A ∈ Kn×n if Ax = λx
for some x ∈ Kn×1 with x 6= 0. It turns out that similar
matrices have the same eigenvalues. In fact, more is true.
Proposition
Let A, A0 ∈ Kn×n be similar. Then pA (t) = pA0 (t). In
particular, λ ∈ K is an eigenvalue of A if and only if λ is an
eigenvalue of A0 . Consequently, the algebraic multiplicity of λ
as an eigenvalue of A is equal to the algebraic multiplicity of λ
as an eigenvalue of A0 . Furthermore, the geometric
multiplicity of λ as an eigenvalue of A is equal to the
geometric multiplicity of λ as an eigenvalue of A0 .
Proof: SInce A ∼ A0 , there is an invertible P ∈ Kn×n such
that A0 = P−1 AP. Writing I = P−1 P, we see that
pA0 (t) = det(A0 −tI) = det(P−1 AP−tI) = det(A−tI) = pA (t).
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
Proof Contd. It remains to prove the assertion about
geometric multiplicity. Let λ be an eigenvalue of A and x be
an eigenvector of A corresponding to λ. Then x 6= 0 and
Ax = λx. Since A0 = P−1 AP, we see that x0 := P−1 x satisfies
A0 x0 = λx0 . Also x0 6= 0 since P is invertible. Thus x0 is an
eigevector of A0 corresponding to λ. Also, it is easy to check
that {x1 , . . . , xg } is a basis for N (A − λI) if and only if
{P−1 x1 , . . . , P−1 xg } is a basis for N (A0 − λI). Hence the
geometric multiplicity of λ as an eigenvalue of A is equal to
the geometric multiplicity of λ as an eigenvalue of A0 .

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Example

   
λ 1 λ−t 1
A := ⇒ pA (t) = det = (λ − t)2 .
0 λ 0 λ−t
Hence λ is the only eigenvalue of A, and its algebraic
multiplicity
 is 2. But its geometric multiplicity is 1 since
0 1
A−λI = =⇒ rank(A−λI) = 1 =⇒ nullity(A−λI) = 1.
0 0
Note that in the above example, if A were similar to a
diagonal matrix D, then we must have D = diag(λ, λ), since
eigenvalues and their algebraic multiplicities of A and D have
to be the same. But the geometric multiplicity of λ as an
eigenvalue of D is 2. This shows that A is not diagonalizable.

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Example
   
3 0 0 3−t 0 0
A := −2 4 2 ⇒ pA (t) = det  −2 4 − t 2 .
−2 1 5 −2 1 5−t
Computing the determinant, we find pA (t) = (3 − t)2 (6 − t).
Hence 3 is an eigenvalue of A of algebraic multiplicity 2, and 6
is an eigenvalue
 of A of algebraic multiplicity 1. Also,
0 0 0
A − 3I = −2 1 2 ⇒ rank(A − 3I) = 1.
−2 1 2
 T  T
So nullity(A − 3I) = 2. In fact, 1 0 1 , 1 2 0 is
a basis of the eigenspace of A corresponding to eigenvalue 3,
and so its geometric multiplicity is equal to 2.

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Relating geometric and algebraic multiplicities

Proposition
Let A ∈ Kn×n , and let λ be an eigenvalue of A. Then the
geometric multiplicity of λ is less than or equal to its algebraic
multiplicity.

Proof. Let g be the geometric multiplicity of λ. Let

(v1 , . . . , vg ) be an ordered basis of the eigenspace of λ; extend
n×1
it to an ordered  basis (v1 ,. . . , vg , vg +1 , . . . , vn ) of K .
Define P := v1 · · · vn . Then P is invertible since its n
columns v1 , . . . , vn are linearly independent. Consider
A0 := P−1 AP. Since Avj = λvj and Pej = vj for j = 1, . . . , g ,
we see that the j th column of A0 is given by

A0 ej = P−1 APej = P−1 Avj = λP−1 vj = λej .

Saurav Bhaumik, IIT Bombay MA110: Lecture 11

Hence
λ ··· 0
 
 ... . . . ... C 
0
A =
 
 0 ··· λ


O D
where C ∈ Kg ×(n−g ) , O ∈ K(n−g )×g and D ∈ K(n−g )×(n−g ) .
Expanding by the first column, we see that
det(A0 − tI) = (λ − t)g q(t),
where q(t) is a polynomial of degree n − g . Thus
pA (t) = p A0 (t) = det(A0 − tI) = (λ − t)g q(t).
Thus (λ − t)g divides the characteristic polynomial pA (t) of
A. Since the algebraic multiplicity of λ is the largest natural
number m such that (λ − t)m divides pA (t), we obtain
g ≤ m.
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
Eigenvectors corresponding to distinct eigenvalues
Our next result is about the linear independence of
eigenvectors corresponding to distinct eigenvalues of a matrix.
Lemma
Let A ∈ Kn×n , and let λ1 , . . . , λk be distinct eigenvalues of A.
Let x1 , . . . , xk belong to the eigenspaces of A corresponding to
λ1 , . . . , λk respectively. Then

x1 + · · · + xk = 0 ⇐⇒ x1 = · · · = xk = 0.

In particular, if x1 , . . . , xk are eigenvectors of A corresponding

to λ1 , . . . , λk respectively, then the set {x1 , . . . , xk } is linearly
independent.
Proof. We use induction on the number k of distinct
eigenvalues of A. Clearly, the result holds for k = 1.
Let k ≥ 2 and assume that the result holds for k − 1.
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
Suppose x := x1 + · · · + xk−1 + xk = 0. Then Ax = 0, that is,
λ1 x1 + · · · + λk−1 xk−1 + λk xk = 0. Also, multiplying the first
equation by λk , we obtain λk x1 + · · · + λk xk−1 + λk xk = 0.
Subtraction gives (λ1 − λk )x1 + · · · + (λk−1 − λk )xk−1 = 0.
By the induction hypothesis,
(λ1 − λk )x1 = · · · = (λk−1 − λk )xk−1 = 0. Since
λ1 6= λk , . . . , λk−1 6= λk , we obtain x1 = · · · = xk−1 = 0, and
so xk = 0 as well.
Now let x1 , . . . , xk be eigenvectors. If α1 x1 + · · · + αk xk = 0,
then α1 x1 = · · · = αk xk = 0. But x1 6= 0, . . . , xk 6= 0, so that
α1 = · · · = αk = 0. Thus {x1 , . . . , xk } is linearly independent.
Theorem
Let A ∈ Kn×n , and let λ1 , . . . , λk be the distinct eigenvalues
of A. Let gj be the geometric multiplicity of λj for
j = 1, . . . , k. Then g1 + · · · + gk ≤ n. Further, A is
diagonalizable if and only if g1 + · · · + gk = n.
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
Proof. Let pA (t) = kj=1 (λj − t)nj . Then nj is the algebraic
Q
multiplicity of λj . We have already seen that gj ≤ nj . Thus,
k
X k
X
gj ≤ nj = n.
j=1 j=1

If A is diagonalizable, that is, similar to a diagonal matrix D,

then the distinct eigenvalues of D are exactly λ1 , . . . , λk and
for each j = 1, . . . , k, the geometric and algebraic
multiplicities of λj are same for A and D. However as D is a
diagonal matrix, the algebraic and geometric multiplicities for
each
Pn eigenvalue are same, hence gj = nj for each j, thus
j=1 gj = n.
Conversely, suppose nj=1 gj = n. Let Vj denote the
P

eigenspace N (A − λj I) of Kn×1 , and let Ej be a basis for Vj

consisting of gj eigenvectors of A corresponding to λj for
j = 1, . . . , k.
Saurav Bhaumik, IIT Bombay MA110: Lecture 11
Proof contd.

We claim that the set E := E1 ∪ · · · ∪ Ek is linearly

independent. Let x be a linear combination of elements of E .
Collate the elements of Ej for each j = 1, . . . , k and write
x = x1 + · · · + xk , where xj ∈ Vj for j = 1, . . . , k. Suppose
x = 0. Then xj = 0 for j = 1, . . . , k by the previous lemma.
For j ∈ {1, . . . , k}, xj is a linear combination of elements of
the set Ej , and since the set Ej is linearly independent, every
coefficient in this linear combination must be 0. Since this
holds for each j = 1, . . . , k, we see that every coefficient in the
linear combination x of elements of E must P be 0.
Now Ej has gj many elements, so E has j gj = n many
elements. As the dimension of Kn is n, E must be a basis. As
every element of E is an eigenvector of A, by results already
proved, A is diagonalizable.

Saurav Bhaumik, IIT Bombay MA110: Lecture 11