Outline of the week

1 Abstract vector spaces

2 Bases and dimensions
3 Inner product spaces
4 Gram-Schmidt orthogonalization
5 Linear transformation and matrices
6 Change of bases and similarity
7 Symmetric maps and spectral theorem

1 / 58
Abstract vector spaces

After studying vector spaces which were contained in Rn or even

Cn , we now state the very important:
Definition (Abstract vector space)
A non-empty set V with two algebraic operations namely,
1 VECTOR ADDITION:

+ : V × V −→ V , (a, b) 7→ a + b

and
2 SCALAR MULTIPLICATION:

(·) : R(or C) × V −→ V , (λ, a) 7→ λa

is called a vector space if the following eight axioms hold.

2 / 58
Axioms contd.

∀ a, b, c ∈ V and λ ∈ R(or C)

Definition (continuation)
A1 a + b = b + a. [Commutativity]
A2 (a + b) + c = a + (b + c). [Associativity]
A3 ∃! 0 ∈ V s.t. a + 0 = a. [Additive identity]
A4 ∃! − a s.t. a + (−a) = 0. [Additive inverse]

M1 λ(a + b) = λa + λb. [L distributivity(scal. mult. over vect. add.)]

M2 (λ + µ)a = λa + µa [R distributivity(scal. mult. over scal. add.)]
M3 λ(µa) = (λµ)a [mixed associativity]
M4 and 1 · a = a [normalization]

∃! is read as ”there exists a unique”.

3 / 58
Some natural consequences

The following natural properties can be logically deduced form the

above axioms.
0 · a = 0 = λ · 0.
(−1)a = −a.
Remark: The uniqueness assertions in A3 (about 0) and A4
(about −a) is also derivable if we chose to omit it from the
axioms. i.e. ∃! is implied by ∃.

4 / 58
Examples

Example (Vector spaces)

1 Rn and Cn .
2 N (A) = {x ∈ Rn |Ax = 0}, A ∈ Mm×n (R).
3 Mm×n (R) and Mm×n (C).
4 R[x] -the set of all the polynomials in x with real coefficients.
Similarly C[x].
5 R[x]3 -the set of all the polynomials in x with real coefficients
of degree ≤ 3. Similarly C[x]3 or R[x]d etc..
6 Solutions of the equation y 000 + µ2 y 0 = 0 or of the equation
y 00 + p(x)y 0 + q(x)y = 0 (over some interval I ). (These are
null spaces of obvious ”linear maps”.)
 
y1 (t)
7 The set of vector functions y(t) = satisfying ÿ = Ay,
y2 (t)
where A ∈ M2 (R).

5 / 58
Non-examples

Example (Non vector spaces)

1 All m × n real matrices with entries ≥ 0.
2 All real (or complex) coefficient quintic polynomials.
3 Solutions of xy 0 + y = 3x 2 , x > 0.
4 Solutions of y 0 + y 2 = 0. (much worse than previous)

etc..

6 / 58
Linear combinations

Definition (Linear combinations)

Given v1 , v2 , ..., vk ∈ V a linear combination is a vector

c1 v1 + c2 v2 + · · · + ck vk

for any choice of scalars c1 , c2 , ..., ck .

Definition (Linear dependence)

A set of vectors v1 , v2 , ..., vk ∈ V is called linearly dependent If
scalars c1 , c2 , ..., ck , at least one NON-zero, can be found such that

c1 v1 + c2 v2 + · · · + ck vk = 0.

7 / 58
Example

Contrapositive of lin.Dependence is lin.INdependence

Example
1 In R[x] the set {1, x, x 2 , x 5 } is a linearly independent set.
2 On the other hand the set {1, x, 1 + x 2 , 1 − x 2 } ⊂ R[x] is a
linearly dependent set since

(−2)1 + 0x + 1(1 + x 2 ) + 1(1 − x 2 ) = 0.

c1 = −2, c2 = 0, c3 = 1 = c4 are not all zeroes.

8 / 58
Wronskian

Definition
(Wronskian) Let f1 , f2 , ..., fn be functions over some interval (a, b).
Their Wronskian is another function on (a, b) defined by a
determinant involving the given functions and their derivatives
upto the order n − 1.
 
 f1
 0 f2 ··· fn 
 f1
def 
f20 ··· fn0 
Wf1 ,f2 ,...,fn (x) =  .. .. ..  .
 . . . 
 (n−1) (n−1) (n−1) 
1
f
2 f ··· fn

9 / 58
Wronskian

Example
Prove that if c1 f1 + c2 f2 + · · · + cn fn = 0 holds over the interval
(a, b) for some constants c1 , c2 , ..., cn and Wf1 ,f2 ,...,fn (x0 ) 6= 0 at
some x0 , then c1 = c2 = · · · = cn = 0. In other words,
nonvanishing of Wf1 ,f2 ,...,fn at a single point establishes linear
independence of f1 , f2 , ..., fn on (a, b).

Caution:The converse is false in general.

Wf1 ,f2 ,...,fn ≡ 0 =⇒
6 f1 , f2 , ..., fn linear dependence on (a, b).

However, the converse is true for ”good” functions like

polynomials and power series.

10 / 58
Solution

Differentiating c1 f1 + c2 f2 + · · · + cn fn = 0 (n − 1) times we
get a system of equations
    
f1 f2 ··· fn c1 0
 f10 f 2
0
· · · f n
0 
 c 2  0
 ..

.. ..  

.  =  .  , x ∈ (a, b).
 . . .   ..   .. 
(n−1) (n−1) (n−1)
f1 f2 · · · fn cn 0

c1 0
   
c2  0
If Wf1 ,f2 ,...,fn (x0 ) 6= 0 for some x0 ∈ (a, b), then 
 ...  =  ...  as
  

cn 0
required.

11 / 58
Example

Example
Show that the set {1, x, 1 + x 2 , 1 − x 2 } ⊂ R[x] is a linearly
dependent set.
Solution:
The Wronskian is
 
1 x 1 + x 2 1 − x 2 
 
0 1
 2x −2x 
 = 0.
0 0 2 2 

0 0 0 0 

Therefore, linearly dependent , since the converse of the

criterion is true for polynomial functions.

12 / 58
Example

Example
Show that the set {1, x, x 2 , x 5 } ⊂ R[x] is a linearly
independent set.
Solution:
The Wronskian is
 
1
 x x 2 x 5 
0 1 2x 5x 4 
 = 120x 2 ≡
6 0.
0 0 2 20x 3 
0 0 60x 2 

0

Therefore, by Wroskian criterion, the set is linearly

independent.

13 / 58
Linear span, basis

Definition (Linear span)

Given a (finite) set S of vectors in a vector space V , its linear span
L(S) is the set of ALL linear combinations of elements of S.

Definition (Basis)
If a set of vectors S in a vector space V is such that
S is linearly independent and
Every vector in V is a (finite) linear combination of vectors
from S.
then the set S is called a basis of S.

Definition (Dimension)
The cardinality of any basis of V is called the dimension of V .

14 / 58
Examples

Example
{i, j, k} is the standard basis of R3 .

{e1 , e2 , ..., en } is the standard basis of Rn .

A basis of R[x] is

S = {1, x, x 2 , ..., x n , ...}.

Hence it has infinite dimensions.

A basis of R[x]5 is

S = {1, x, x 2 , x 3 , x 4 , x 5 }

which makes it 6-dimnesional.

15 / 58
More examples

Example
{Ejk , 1 ≤ j ≤ m, 1 ≤ k ≤ n} ⊂ Mm×n (R)} is a basis of
Mm×n (R).

{cos µx, sin µx} is a basis of the solution space of the

differential equation y 00 + µ2 y = 0, µ > 0s.
(Apply Wronskian criterion)
n  o
Let V = p(x) ∈ R[x]3 p(1) = 0 . Then V is a vector

space with a bases

B = {1 − x, 1 − x 2 , 1 − x 3 }.

It has 3 dimensions. 
Compare with {[a, b, c, d] ∈ R4 a + b + c + d = 0}.

16 / 58
Inner product spaces

Definition (Real inner product)

A real vector space V is called an inner product space if to each
pair of vectors v, w ∈ V is associated a real number hv, wi such
that
1 hλv + µw, xi = λhv, xi + µhw, xi. (Linearity in the 1st slot)
2 hv, wi = hw, vi (Symmetry)
3 hv, vi ≥ 0 and equality only if v = 0. (Positive definiteness)

(1) and (2) imply linearity in the second variable

i.e. hx, λv + µwi = λhx, vi + µhx, wi.

Definition (Length/Norm)
p
For a vector v ∈ (V , h, i), the non-negative square root hv, vi is
called the length or norm of v. It is usually denoted kvk.

17 / 58
Inner product spaces

Definition (Unitary or Hermitian inner-product)

A complex vector space V is called a unitary inner-product space if
to each pair of vectors v, w ∈ V is associated a complex number
hv, wi if
1 hλv + µw, xi = λhv, xi + µhw, xi. (Linearity in the 1st slot)
2 hv, wi = hw, vi. ((Hermitian symmetry)
3 hv, vi ≥ 0 and equality only if v = 0. (Positive definiteness)

(1) and (2) imply conjugate-linearity in the second variable

i.e. hx, λv + µwi = λ̄ hx, vi + µ̄ hx, wi.

Definition (Length)
p
For a vector v ∈ (V , h, i), the non-negative square root hv, vi is
called the length of v. It is also usually denoted kvk.

18 / 58
Cauchy-Schwartz inequality
Theorem (Cauchy-Schwartz inequality)
For v, w ∈ V , we have |hv, wi| ≤ kvkkwk.
Equality holds if and only if {v, w} is a linearly dependent set.

Corollary (Triangle inequality)

Let (V , h, i) be any (unitary) inner-product space and v, w be any
two vectors. Then kv ± wk ≤ kvk + kwk.

Proof:

kv ± wk2 = hv ± w, v ± wi
bi-lin.
= kvk2 + kwk2 ± [hv, wi + hw, vi]
≤ kvk2 + kwk2 + |hv, wi| + |hw, vi|
C-S
≤ kvk2 + kwk2 + 2kvkkwk = (kvk + kwk)2 .

19 / 58
Proof of C-S inequality, real case

If either v or w = 0, there is lin. dependence and equality.

So let A = hv, vi > 0, B = hv, wi, C = hw, wi > 0.

At 2 − 2Bt + C = hw − tv, w − tvi ≥ 0 ∀t ∈ R.

By the theory quadratic equations, the discriminant

B 2 − AC ≤ 0 ⇐⇒ |hv, wi| ≤ kvkkwk ,

with equality ⇐⇒ ∃t0 s.t. w = t0 v (linear dependence). This

means that
|hv, wi| ≤ kvkkwk
with equality if and only if {v, w} is a l.d. set. [1.0]

20 / 58
Proof of C-S inequality, complex case
In the unitary case, working as above and using
hv, wi + hw, vi = 2Rehv, wi,
we will get  
Rehv, wi ≤ kvkkwk.
Let
hv, wi = |hv, wi|e ıθ ( polar form in C)
and replace w by w0 = e ıθ w. This yields,
Rehv, w0 i ≤ kvkkw0 k = kvkkwk.
 

But
Rehv, w0 i = Re e −ıθ hv, wi 
  

| {z }
=|hv,wi|≥0
= |hv, wi|

21 / 58
Angle between two vectors and orthogonality

Let v.w ∈ V \ {0} be two non zero vectors in an i.p. space V .

The angle between v and w is defined abstractly as

hv, wi
θ = cos−1 .
kvkkwk

Re hv, wi
In the unitary case the angle is θ = cos−1 .
kvkkwk
In either case θ ∈ [0, π].

Definition (Orthogonality)
Two vectors v, w ∈ V are said to be orthogonal if hv, wi = 0.

In short we express this by writing v ⊥ w. Naturally, if both are

π
non zero then the angle between them is θ = .
2
22 / 58
Pythagoras theorem, parallelogram law

Theorem (Pythagoras)
Let v ⊥ w be any two mutually orthogonal vectors in an i.p. space
V , then
kv + wk2 = kvk2 + kwk2 .

Theorem (Parallelogram law)

Let v, w be any two vectors in an i.p. space V , then
 
kv + wk2 + kv − wk2 = 2 kvk2 + kwk2 .

Proofs:
Directly from axioms and ”R-bilinearity”. [1.5]

23 / 58
Examples

Example (1)
V = Rn , treated as the set of n × 1 columns. Then
X
hv, wi := vT w = vj w j
j

is the standard inner product of Rn .

In the case of Cn , we take hv, wi := w∗ v = j vj w̄j P
P
.
Caution: For row vectors we take hv, wi := vw∗ = j vj w̄j .

Example (2)
Let P be a symmetric n × n matrix with all eigenvalues positive. In Rn ,
define
hx, yiP = xT Py.
Then h, iP is the most general inner product in Rn .

24 / 58
Examples

Example (3(Lorentz contraction))

In R3 , define
hx, yiβ = x1 y1 + x2 y2 + β 2 x3 y3 ,
p
where β = 1 − v 2 /c 2 .

This describes the geometry of an inertial frame moving with

velocity ±v k as seen by a ’stationary’ observer.

Example (4)
In the space Mn (R), define hX , Y i = trX T Y . This defines an
inner product in the vector space of n × n matrices. It is analogous
to the standard inner product in Rn .

25 / 58
Function spaces

Example (4)

Let V = C [a, b] = {f : [a, b] −→ Rf is continuous} and define
Z b
hf , g i = f (t)g (t)dt.
a

Then C [a, b] becomes an i.p. space.

26 / 58
Orthogonal and orthonormal sets
Definition (Orthogonal/orthonormal sets)
A subset S of an i.p. space V is said to be an orthogonal set if
hv, wi = 0 ∀ v =
6 w ∈ S.

If in addition kvk = 1 ∀ v ∈ S, then S is called an orthonormal set.

Theorem
If S is an orthogonal set of non-zero vectors in V , then S is
linearly independent.

Proof:
Exercise.
Example
{i, j, k} is an orthonormal set in R3 .
{i, j, β −1 k} is an orthonormal set in R3 w.r.t. h, iβ .
27 / 58
Orthogonal sets, examples

Example (2)
Z π
Let V = C [−π, π] and define as before hf , g i = f (t)g (t)dt.
−π
Then the infinite set

S = {1, cos x, sin x, cos 2x, sin 2x, ...ad inf }

is an orthogonal set in C [−π, π].

It follows that
 
1 cos x sin x cos 2x sin 2x
S = √ , √ , √ , √ , √ , ...ad inf
2π π π π π

is an orthonormal set in C [−π, π].

28 / 58
Orthonormal bases

Definition (Orthonormal basis)

A basis of an inner-product space V which is an orthonormal set is
called an orthonormal basis of V .
Let B = {u1 , u2 , ..., un } be an (ordered)
X o.n. basis of V . Let v be
any element of V . Can write v = cj uj .
j
Recall that in general it is difficult to compute cj , but
Xin this case
2
(o.n. basis), we just have cj = huj , vi. Also kvk = |cj |2 just
j
like in Rn or Cn .

29 / 58
Gram-Schmidt process
Q. How to get orthonormal sets and bases?
A. By Gram-Schmidt process .
Start with any spanning set {v1 , v2 , v3 , ...} -finite or infinite.
Recursively, produce an orthogonal set as follows (same as before
in Rn or Cn ):

w1 = v 1
hv2 , w1 iw1
w2 = v 2 − (= v2 if w1 = 0.)
kw1 k2
hv3 , w1 iw1 hv3 , w2 iw2
w3 = v3 − − (similar remark)
kw1 k2 kw2 k2
..
.
k−1
X hvk , wj iwj
wk = vk − etc.
j=1
kwj k2
wj 6=0

30 / 58
G-S process

The following facts can be verified recursively (by induction):

{w1 , w2 , ..., } is an orthogonal set.

LS{v1 , ..., vk } = LS{w1 , ..., wk } ∀k = 1, 2, 3, ...

Finally, zero vectors, if any, can be dropped from the

orthogonal set obtained.

31 / 58
An example
Example
Consider the linearly independent set {1, x, x 2 , ...ad inf } ⊂ V = C [−1, 1].
The (real) space V has the inner-product defined by
Z 1
hf , g i = f (t)g (t)dt.
−1

The G-S process yields, upto scalar multiples, the famous

Legendre polynomials.

w0 = v0 = 1
hx, 1i
w1 = x− 1=x
k1k2
hx 2 , 1i hx 2 , xi 1
w2 = x2 − 2
.1 − x = x2 −
k1k kxk2 3
3
hx , xi 3
w3 = x3 − 0 − − 0 = x3 − x
kxk2 5
6 3
w4 = x4 − x2 + etc.
7 35 32 / 58
Bessel’s inequality

Theorem (Bessel’s inequality)

Let {u1 , ..., uk } be an o.n. set in V and v ∈ V . Then
k
X
2
kvk ≥ |hv, uj i|2 .
j=1

The above inequality becomes equality if and only if

v ∈ LS{u1 , ..., uk }.

Proof:
Define
k
X
v0 = hv, uj iuj .
j=1

33 / 58
Bessel’s inequality
Observe that

v0 ∈ LS{u1 , ..., uk } and ⊥ (v − v0 ).

Now,

kvk2 = kv0 + (v − v0 )k2

Pythagoras
= kv0 k2 + k(v − v0 )k2
k
X
2
≥ kv0 k = |hv, uj i|2 .
j=1

Equality holds

⇐⇒ k(v − v0 )k = 0 ⇐⇒ v = v0 ⇐⇒

v ∈ LS{u1 , ..., uk }.

34 / 58
G-S process; an example in C [0, π]

Example
Orthogonalize the ordered set in [0, π]:

{cos 2x, sin 2x, cos x, sin x}

Z π
relative to the inner product hf , g i = f (t)g (t)dt.
0

Solution:

w1 = v1 = cos 2x
0
w2 = sin 2x − cos 2x = sin 2x
π/2
0 4/3 8
w3 = cos x − cos 2x − sin 2x = cos x − sin 2x
π/2 π/2 3π

35 / 58
G-S process; example in C [0, π] contd.

−2/3 0 0
w4 = sin x − cos 2x − sin 2x − w3
π/2 π/2 π/2
4
= sin x + cos 2x.
3π
The orthogonalized set is
    
8 4
cos 2x, sin 2x, cos x − sin 2x , sin x + cos 2x
3π 3π
 
π π π 32 π 8
Squares of lengths are , , − , − .
2 2 2 9π 2 9π

36 / 58
Linear maps

Definition (Linear map or transformation)

Let V , W be two (abstract) vector spaces. A map
T : V −→ W is called a linear map or transformation if
T (v + w) = T v + T w
and T (λv) = λT v.

Example
A ∈ Mm×n (R) can be viewed as a linear map A : Rn −→ Rm
via the matrix multiplication v 7→ Av. Here v is treated as
n × 1 column which maps to m × 1 column Av.

37 / 58
More examples

d2
(a) 2
+ µ2 : R[x] −→ R[x]
dx
(b) τc : R[x]m −→ R[x]m where


τc p(x) = p(x − c), c ∈ R.

(c) µx : R[x]m −→ R[x] defined by

µx p(x) = xp(x).

(d) I : C ([0, 1]) −→ C ([0, 1]) defined by

Z x
I (f ) (x) =


f (t)dt, 0 ≤ x ≤ 1.
0

38 / 58
Significance of linear maps

Levels of complexity of maps:

1 Constant maps: Easiest to describe. Need to know at any
one point only. (Constant maps can not tell ANYTHING
about the input by looking at the output.)
2 Linear maps: Easiest among non constant maps. It is
enough to sample a linear map on a basis to determine it
completely.
3 Non-linear maps: Hardest to deal with. Often the local
behaviour is studied by linear approximation. Hence the
importance of linear maps.

16:05 [2.0]

39 / 58
Matrix of a linear transformation (or map)

If T : V −→ W is a linear map and B ⊂ V is a basis, then

T  determines T completely.

B
In other words, the values or images

{T v ∈ W }v∈B

determine T on ALL of V .

Conversely, we can assign arbitrary values T v to the vectors

v ∈ B and create a linear map T : V −→ W .

40 / 58
Matrix of a linear transformation (or map)

Now suppose B1 = {v1 , v2 , ..., vn } is a basis of V and

B2 = {w1 , w2 , ..., wm } is that of W . Assume that

T : V −→ W

is a linear map, then we get a unique m × n array of scalars on writing

T v1 = t11 w1 + t21 w2 + · · · + tm1 wm

T v2 = t12 w1 + t22 w2 + · · · + tm2 wm
.. .. ..
. . .
T vn = t1n w1 + t2n w2 + · · · + tmn wm

The above array completely determines T

This gives us a (unique) matrix representation [tjk ] of T w.r.t. to
the ordered bases B1 and B2 of V and W respectively.

41 / 58
Matrix of a linear transformation contd.

Conversely if A = [ajk ] is any m × n matrix of scalars, then we can

construct a unique linear map TA : V −→ W by setting
 
linear
X X X X
TA ( c k vk ) = ck TA vk := ck  ajk wj 
k k k j
X
= bj wj , say.
j

This can also be en-coded as

    
a11 a12 · · · a1n c1 b1
 a21 a22 · · · a2n  c2   b2 
   
..   ..  =  ..  .
 
 .. ..
 . . .  .   . 
am1 am2 · · · amn cn bm
To break the code (recover TA : V −→ W ) must have the ”keys”
B1 and B2 .
42 / 58
Example

Example
Write down the matrix of the linear map

T p(x) = xp 0 (x − 1) + p 00 (x)

from R[x]3 to itself w.r.t. the ordered basis B = {1, x, x 2 , x 3 } on

both the sides.
Solution:
1 T (1) = 0 = 0.1 + 0.x + 0.x 2 + 0.x 3 .
2 T (x) = x = 0.1 + 1.x + 0.x 2 + 0.x 3 .
3 T (x 2 ) = x(2x − 2) + 2 = 2.1 − 2.x + 2.x 2 + 0.x 3 .
4 T (x 3 ) = 3x(x − 1)2 + 6x = 0.1 + 9.x − 6.x 2 + 3.x 3 .

43 / 58
Example contd.

T p(x) = xp 0 (x − 1) + p 00 (x) implies

T (1) = 0.1 + 0.x + 0.x 2 + 0.x 3

T (x) = 0.1 + 1.x + 0.x 2 + 0.x 3
T (x 2 ) = 2.1 − 2.x + 2.x 2 + 0.x 3
T (x 3 ) = 0.1 + 9.x − 6.x 2 + 3.x 3

Therefore the 4 × 4 matrix of T w.r.t. the basis B is

 
0 0 2 0
0 1 −2 9 
[T ] = 
0 0 2 −6


0 0 0 3

44 / 58
The change of basis

Theorem
Let T : V −→ W be a linear map. Let B, B 0 be two bases of V
and C, C 0 be those of W . Let
1 [T ] be the matrix of T relative to B, C and [T ]0 be the matrix
of T relative to B 0 , C 0 .
2 Let P be the matrix wich gives B 0 in terms of B and Q be the
matrix which gives C 0 in terms of C.
Then
[T ]P = Q[T ]0 .

45 / 58
The change of basis, proof

Proof: Let B = {v1 , v2 , ..., vn } and B 0 = {v10 , v20 , ..., vn0 }.

Let P = [pk` ]. (B 0 in terms of B)
Similarly, let C = {w1 , w2 , ..., wm } and C 0 = {w10 , w20 , ..., wm 0
}.
0
Let Q = [qij ]. (C in terms of C)
Also let [T ] = [tjk ] and [T ]0 = [tjk0 ].
X X X
OTO1 H, T v`0 = T pk` vk = pk` T vk = pk` tjk wj .
k k j,k
X X
OTO2 H, T v`0 = 0
ti` wi0 = 0
ti` qji wj .
i i,j

Hence due to linear independence of C, we find on equating the

coefficients of wj , X X
0
pk` tjk = ti` qji .
k i

, while the RHS = Q[T ]0 ∴ [T ]P = Q[T ]0 .

LHS = [T ]P j` j`
,

46 / 58
Special case of interest

The case of V = W (self map) is most interesting.

We use the same basis ON BOTH SIDES in the case of a self-map.

Theorem (Change of basis)

Let T : V −→ V be a linear map of a vector space V .
Suppose B and B 0 are two bases of V .
Let [T ] and [T ]0 be matrices of T w.r.t. B and B 0 respectively.
Then [T ] ∼ [T ]0 .
More precisely, if B 0 is expressed by the matrix P w.r.t. B then
P −1 [T ]P = [T ]0 .

It is understood that we used B on BOTH SIDES to write [T ].

Likewise, B 0 for [T ]0 .

47 / 58
The null space and the range

Let T : V −→ W be a linear transformation.

Definition (Null space)

The set {v ∈ V T v = 0} is called the null space of T . It is also
called the kernel of T .
It is a vector subspace of V and usually denoted as N (T ).

Definition (Range space)


The set {T v ∈ W v ∈ V } is known as the range space of T .

The range of T is denoted by R(T ) and it is a vector subspace of

W.

48 / 58
Rank-nullity theorem

Definition (Rank and nullity)

The dimension of N (T ) is called the nullity of T and is denoted
null(T ).
The dimension of R(T ) is called the rank of T and is denoted
rank(T ).

Theorem (Rank-nullity theorem)

For any linear transformation T : V −→ W between two vector spaces
we have
rank(T ) + null(T ) = dimV .

Theorem
Let [T ] be a matrix of T w.r.t. to any choices of bases in V and W .
then rank([T ]) = rank(T ) and null([T ]) = null(T ).

49 / 58
Solvabilty of T x = b

T ’transmits’ V into W . null(T ) is the number of dimensions

which disappear (transmission losses) and rank(T ) is the number
of dimensions received. Naturally the sum rank(T ) + null(T ) is
the total number of dimensions transmitted i.e. dimV .
Solvabilty of T x = b:
Given a linear map T : V −→ W and b ∈ W , we can solve
T x = b for the unknown x ∈ V if and only if b ∈ R(T ).
This geometrically obvious statement is equivalent to
rank(A) = rank(A+ ) for the solvability of the linear systems.

If p ∈ V is any particular
 solution i.e. T p = b, then the set
p + N (T ) = {p + v v ∈ N (T )} is the full set of solutions of

T x = p.
This is identical to the description of the solution set of the linear
systems.

50 / 58
Eigenvalue problem
We restrict to the real vector spaces for simplicity.
Let V be a real vector space. Let T : V −→ V be a linear map.
(It is important that W = V .)
Eigenvalue Problem (EVP):

Find a scalar λ and a vector v(6= 0) ∈ V such that T v = λv.

Definition (Eigenvalue, eigenvector)

1 The scalar λ in the above problem is called an eigenvalue of T .
2 The vector v in the above problem is called an eigenvector of T
corresponding to the eigenvalue λ.

Diagonalization Problem: Given T : V −→ V , a linear map, does

there exist a T -eigenbasis of V ?
If yes, then
the matrix [T ] of T w.r.t. to such a basis will be diagonal.
51 / 58
Eigenvalue problem, example

Example
Let T : R[x]3 −→ R[x]3 be defined by

T p(x) = xp 0 (x − 1) + p 00 (x)

Find a T -eigenbasis of R[x]3 if it exists.

Solution: The matrix of T w.r.t. the basis B = {1, x, x 2 , x 3 } is

 
0 0 2 0
0 1 −2 9 
[T ] = 
0 0 2 −6


0 0 0 3

The eigenvalues are 0, 1, 2, 3, all simple. Hence T is diagonalizable.

By inspection: for 0 and 1, the eigenvectors are p0 = 1 and p1 = x
respectively.
52 / 58
For the eigenvalue λ = 2 and eigenvector c0 + c1 x + c2 x 2 + c3 x 3
say,
 
−2 0 2 0
 0 −1 −2 9 
[T ] − 2I =  
0 0 0 −6
0 0 0 1
      
-2 0 2 0 -2 0 2 0 c0 0
 0 1 2 −9  0 1 2 −9 c1  0
      
7→   =⇒  = .
0 0 1  c2  0

 0 0 0 1   0
0 0 0 0 0 0 0 0 c3 0

Hence an eigenvector is p2 = 1 − 2x + x 2 . (c2 = 1 chosen)

53 / 58
Example contd.
Similarly, an eigenvector for λ = 3 is p3 = −4 + 10.5x − 6x 2 + x 3 . The
matrix of eigenvectors is
   
1 0 1 −4 1 0 −1 −2
0 1 −2 10.5 −1
0 1 2 1.5
P= , P =  
0 0 1 −6  0 0 1 6
0 0 0 1 0 0 0 1

Verify that P −1 [T ]P = diag{0, 1, 2, 3}.

  
0 0 0 0 1 0 1 −4
P −1 [T ]P = 
0 1 2 1.5  0 1 −2 10.5
  
 = diag{0, 1, 2, 3}.
0 0 2 12  0 0 1 −6 
0 0 0 3 0 0 0 1
| {z }
=P −1 [T ]

d
Exercise: Show that D = : R[x]3 −→ R[x]3 is not
dx
diagonalizable.
54 / 58
Symmetric maps and spectral theorem

First we define symmetric linear maps which are analogues of

symmetric matrices.
Definition (Symmetric maps)
Let (V , h, i) be a real vector space with an inner-product. A linear
T : V −→ V is called symmetric or self-adjoint if

hT v, wi = hv, T wi ∀ v, w ∈ V .

In concrete terms we can take an orthonormal basis and write the

matrix [T ] for T . This matrix must be symmetric for T to be a
symmetric linear map.

55 / 58
Symmetric maps and spectral theorem

Theorem (Spectral theorem)

Let T be a self-adjoint (≡ symmetric) linear map of a finite
dimensional real inner-product space (V , h, i). Then,
1 All the eigenvalues of T are real.
2 If {v1 , v2 , ..., vk } are eigenvectors corresponding to the
distinct eigenvalues {λ1 , λ2 , ..., λk } respectively, then the
former is an orthogonal set in V .
3 There exists an orthonormal basis of V consisting of
eigenvectors of T . (Spectral theorem)

56 / 58
Legendre polynomials again

Example
Consider the polynomials of degree n defined by
d
pn (x) = D n (x 2 − 1)n , n = 0, 1, 2, ... where D ≡ .
dx
Show using integration by parts n times that
Z 1
pn (x)pm (x) dx = 0 if m < n (Orthogonality).
−1

Let L : R[x] −→ R[x] be defined by

0
Lf (x) = (1 − x 2 )f 0 (x) .


Prove that
Z 1 Z 1



Lf (x) g (x) dx = f (x) Lg (x) dx. (L is self-adjoint)
−1 −1

57 / 58
Legendre polynomials again
Example
Finally prove that

Lpn (x) = −n(n + 1)pn (x). (pn are eigenfunctions of L)

Thus Legendre polynomials are eigenvectors of a self-adjoint

R1
L : R[x] −→ R[x] with the inner-product hf , g i = −1 f (x)g (x) dx, hence
they are automatically mutually orthogonal, since eigenvalues are distinct.

Legendre’s polynomials form an orthogonal basis of R[x].

Legendre polynomials appear in the physical phenomena involving

spherical symmetries. For example
(i) Propogation of heat or waves along a spherical surface like
earthquakes
(ii) Quantum states (eigenvectors) of a hydrogen atom and their
corresponding energies (eigenvalues).

(iii) Upto scalar multiplication by 2n
n , pn (x) is the degree n polynomial
wn obtained by orthogonalizing {1, x, x 2 , ...ad inf } earlier. [END]
58 / 58