OPERATOR THEORY

Finite dimensional spectral theory:

Introduction: Let 𝐻 be a Hilbert space which is finite dimensional. Let 𝑇 be an
operator that maps 𝐻 into itself we may enquire the existence of 𝑥 ∈ 𝐻, (𝑥 ≠ 0)
which satisfies the equation
𝑇𝑥 = 𝜆𝑥, . . . (1)
where 𝜆 is a scalar.
An element 𝑥 ∈ 𝐻, 𝑥 ≠ 0 that satisfies the equation (1) is called an
eigenvector of 𝑇 and a scalar 𝜆 for which the equation (1) is satisfied by some non-
zero element 𝑥 is called an eigen value of 𝑇. If 𝜆 is an eigen value of 𝑇, then (1) may
be satisfied by more than any one values of 𝑥, each such 𝑥 ≠ 0 is an eigenvector of
𝑇 corresponding to the eigen value 𝜆. On the other hand, it is clear that to each
eigenvector of 𝑇 there corresponds exactly one eigen value.
If 𝜆 is an eigen value of 𝑇, let 𝑀 denote the set of all eigenvectors of 𝑇
corresponding to 𝜆. Although the zero element is not an eigenvector of 𝑇 by
definition, we assume that 𝑀 contains also the zero element. Therefore, every
element of 𝑀 satisfies the equation
𝑇𝑥 = 𝜆𝑥 𝑜𝑟 𝑇 − 𝜆𝐼 𝑥 = 0.
If 𝑇 is a continuous linear operator, then 𝑇 − 𝜆𝐼 is also so and we can now
verify that 𝑀 is closed subspace. This subspace is called eigen space of 𝑇
corresponding to the eigen value 𝜆. The set of eigen values of 𝑇 is called its
spectrum and is denoted by 𝜎 𝑇 .
Existence Theorem: Let 𝑋 be 𝑛-dimensional normed linear space and 𝑇 be a linear
operator such that 𝑇: 𝑋 → 𝑋. We know that 𝑇 can be represented by matrices (that
depend on the choice of the basis of 𝑋) and since here 𝑇 maps 𝑋 into itself, these
are square matrices. Therefore, analogous to (1) we consider the equation
𝑇𝑥 = 𝜆𝑥 . . . (2)

where 𝑇 = 𝛼𝑗𝑘 is an 𝑛-rowed square matrix, 𝑥 is a column matrix of unknowns

𝜉1 , 𝜉2 , . . . , 𝜉𝑛 and 𝜆 is a scalar. In this case, we similarly introduce the concepts of
eigen values, eigenvectors and eigen space.
3 2
Example: Let 𝑇 = .
1 2
It may be verified by direct calculation that
𝜉1 2 𝜉 1
= 𝑎𝑛𝑑 1 =
𝜉2 1 𝜉2 −1
are eigen vector of 𝑇 corresponding to the eigen values 𝜆1 = 4, 𝜆2 = 1.

1
 We may ask naturally how we arrived at the above values. The answer to this
question is provided in the following theorem.
Theorem: An 𝑛- rowed square matrix 𝑇 = 𝛼𝑗𝑘 has at least one eigen value and
almost 𝑛 numerically different eigenvalues. These are determinable from a
polynomial equation.
Proof: We consider the equation (2)
𝑇𝑥 = 𝜆𝑥

where 𝑇 = 𝛼𝑗𝑘 is an 𝑛-rowed square matrix, 𝑥 is a column matrix of unknowns

𝜉1 , 𝜉2 , . . . , 𝜉𝑛 and 𝜆 is a scalar. This equation can be written in the form
𝑇 − 𝜆𝐼 𝑥 = 0 ... 3
where 𝐼 is the 𝑛-rowed unit matrix.
The equation (3) is a homogeneous system of 𝑛 linear equations in 𝑛
unknowns 𝜉1 , 𝜉2 , . . . , 𝜉𝑛 . In order that (3) has a non-zero solution, we must have
𝑑𝑒𝑡 𝑇 − 𝜆𝐼 = 0
which is equivalent to
𝛼11 − 𝜆 𝛼12 ⋯ 𝛼11 − 𝜆
⋯ 𝛼2𝑛
𝑑𝑒𝑡 𝑇 − 𝜆𝐼 = 𝛼⋯21 𝛼22 − 𝜆
⋯ =0 . . . (4)
⋯ ⋯
𝛼𝑛1 𝛼𝑛2 ⋯ 𝛼𝑛𝑛
The equation (4) is called the characteristic equation of 𝑇 and determinant 𝑑𝑒𝑡 𝑇 −
𝜆𝐼 is called the characteristic determinant of 𝑇. If we develop the determinant
𝑑𝑒𝑡 𝑇 − 𝜆𝐼 , then we obtain from (2), a polynomial in 𝜆 of degree 𝑛. Since this
equation has at most 𝑛 numerically different roots, the proof of the theorem is
complete.
If we apply the above theorem to the earlier example, then the characteristic
equation of 𝑇 is
3−𝜆 2
= 0, 𝑖. 𝑒. , 𝜆2 − 5𝜆 + 4 = 0.
1 2−𝜆
So, the spectrum of 𝑇, 𝜎 𝑇 = 1,4 .
It is clear that the representation of a linear operator by a matrix depends
essentially on the selection of a basis. Therefore, for the selection of different bases,
we obtain different matrices representing the same linear operator. Now a question
naturally may arise. Whether different matrices representing a linear operator
corresponding to different bases have the same eigen values. This is answered in
the following theorem.
Theorem: Suppose that 𝑇 is a linear operator such that 𝑇: 𝑋 → 𝑋 where 𝑋 is a finite
dimensional normed linear space. Then all matrices representing 𝑇 corresponding to
various bases for 𝑋 have the same eigen values.

2
Proof: Let 𝑋 be 𝑛-dimensional. Let 𝑒 = 𝑒1 , 𝑒2 , . . . , 𝑒𝑛 and 𝑓 = 𝑓1 , 𝑓2 , . . . , 𝑓𝑛 be two
bases for 𝑋 written as row matrices. Then every 𝑓𝑖 can be written as a linear
combination of 𝑒𝑖 . So, let
𝑛

𝑓𝑖 = 𝛼𝑖𝑗 𝑒𝑗 , 𝑖 = 1,2, . . . , 𝑛
𝑗 =1

Then 𝑓 = 𝑓1 , 𝑓2 , . . . , 𝑓𝑛
𝑛 𝑛 𝑛

= 𝛼1𝑗 𝑒𝑗 , 𝛼2𝑗 𝑒𝑗 , . . . , 𝛼𝑛𝑗 𝑒𝑗

𝑗 =1 𝑗 =1 𝑗 =1

𝛼11 𝛼12 ⋯ 𝛼11

𝛼21 𝛼22 ⋯ 𝛼2𝑛
= 𝑒1 , 𝑒2 , . . . , 𝑒𝑛 ⋯ ⋯ ⋯ ⋯ = 𝑒𝐶 . . . (5)
𝛼𝑛1 𝛼𝑛2 ⋯ 𝛼𝑛𝑛
where 𝐶 is a non-singular 𝑛-rowed square matrix.
We can represent every 𝑥 ∈ 𝑋 as a linear combination of elements with
respect to the bases 𝑒 and 𝑓. So, let
𝑛

𝑥= 𝜉𝑗 𝑒𝑗 = 𝑒𝑥1
𝑗 =1

and also
𝑛

𝑥= 𝜂𝑘 𝑓𝑘 = 𝑓𝑥2
𝑘 =1

where 𝑥1 = 𝜉1 , 𝜉2 , . . . , 𝜉𝑛 and 𝑥2 = 𝜂1 , 𝜂2 , . . . , 𝜂𝑛 are column matrices.

So,
𝑛 𝑛

𝑥 = 𝑒𝑥1 = 𝜉𝑗 𝑒𝑗 = 𝑓𝑥2 = 𝜂𝑘 𝑓𝑘 . . . (6)

𝑗 =1 𝑘 =1

From (5) and (6), we obtain that

𝑒𝑥1 = 𝑓𝑥2 = 𝑒𝐶𝑥2 and therefore 𝑥1 = 𝐶𝑥2 ---(7)
Similarly, if 𝑇𝑥 = 𝑦 = 𝑒𝑦1 = 𝑓𝑦2 , then we shall have
𝑦1 = 𝐶𝑦2 ... 8

Now 𝑇 𝑒𝑗 , 𝑗 = 1,2, … , 𝑛 can be expressed as a linear combination of elements of 𝑒.

So, let
𝑛

𝑇 𝑒𝑗 = 𝜂𝑖𝑗 𝑒𝑖 , 𝑗 = 1,2, … , 𝑛.
𝑖=1

3
Therefore, from (6),
𝑛

𝑒𝑦1 = 𝑇𝑥 = 𝑇 𝜉𝑗 𝑒𝑗
𝑗 =1

= 𝜉𝑗 𝑇 𝑒𝑗 = 𝜉1 𝑇 𝑒1 + 𝜉2 𝑇 𝑒2 +. . . +𝜉𝑛 𝑇 𝑒𝑛
𝑗 =1
𝑛 𝑛 𝑛

= 𝜉1 𝜂𝑖1 𝑒𝑖 + 𝜉1 𝜂𝑖2 𝑒𝑖 +. . . +𝜉1 𝜂𝑖𝑛 𝑒𝑖 = 𝑒𝑇1 𝑥1 ,

𝑖=1 𝑖=1 𝑖=1

where𝑇1 is the matrix which represents 𝑇 with respect to the basis 𝑒.

Therefore, 𝑦1 = 𝑇1 𝑥1 . Similarly, if 𝑇2 denotes the matrix which represents 𝑇
with respect to the basis 𝑓, then 𝑦2 = 𝑇2 𝑥2 .
So, 𝑦1 = 𝑇1 𝑥1 and 𝑦2 = 𝑇2 𝑥2 ---(9)
From (7), (8) and (9), we get that
𝐶𝑇2 𝑥2 = 𝐶𝑦2 = 𝑦1 = 𝑇1 𝑥1 = 𝑇1 𝐶𝑥2 whence, operating by 𝐶 −1 , we obtain that
𝑇2 = 𝐶 −1 𝑇1 𝐶 . . . (10)
With the help of this, we verify that the characteristic determinants of 𝑇1 and 𝑇2
are equal and from this the equality of eigenvalues of 𝑇1 and 𝑇2 follows from the
earlier theorem as already proved. Using the fact 𝑑𝑒𝑡 𝐶 𝑑𝑒𝑡 𝐶 −1 = 1, we obtain that
𝑑𝑒𝑡 𝑇2 − 𝜆𝐼 = 𝑑𝑒𝑡 𝐶 −1 𝑇1 𝐶 − 𝜆𝐶 −1 𝐼𝐶 , 𝑓𝑟𝑜𝑚 10
= 𝑑𝑒𝑡 𝐶 −1 𝑇1 − 𝜆𝐼 𝐶 = 𝑑𝑒𝑡 𝐶 −1 𝑑𝑒𝑡 𝑇1 − 𝜆𝐼 𝑑𝑒𝑡𝐶 = 𝑑𝑒𝑡 𝑇1 − 𝜆𝐼
This proves the theorem.
If 𝑇 is a linear operator such that 𝑇: 𝑋 → 𝑋 where 𝑋 is a finite dimensional
normed linear space and 𝑇1 is the corresponding matrix with respect to a basis then
we identify the eigenvalues of the matrix 𝑇1 with that of the operator 𝑇. Therefore the
above two theorems give the following theorems.
Theorem (Existence theorem): A linear operator on a finite dimensional normed
(complex) linear space 𝑋 ≠ 0 has at least one eigen value.

Spectral Theorem
Let 𝐻 be a finite dimensional Hilbert space and 𝑇 be a continuous linear operator
such that 𝑇: 𝐻 → 𝐻. If 𝐻 has no non-zero element, then 𝑇 has no eigen value by
definition and so there is nothing to be considered. [We therefore suppose that
𝐻 ≠ 0 . Let 𝜆1 , 𝜆2 , . . . , 𝜆𝑚 be all the distinct eigenvalues of 𝑇 and let 𝑀1 , 𝑀2 , . . . , 𝑀𝑚 be
their corresponding eigenspaces. Let 𝑃1 , 𝑃2 , . . . , 𝑃𝑚 be the projections on these eigen
spaces.
The following statements are equivalent to one another:

4
I. The eigenspaces 𝑀𝑖 ’s are pairwise orthogonal and span 𝐻, i.e.,
𝐻 = 𝑀1 ⊕ 𝑀2 ⊕. . .⊕ 𝑀𝑚 .
II. The projection 𝑃𝑖 ’s are pairwise orthogonal and
𝑚 𝑚

𝐼= 𝑃𝑖 𝑎𝑛𝑑 𝑇 = 𝜆𝑖 𝑃𝑖 .
𝑖=1 𝑖=1
III. 𝑇 is normal.
Uniqueness of Spectral Resolution:
We have seen that if 𝑇 is normal then 𝑇 can be expressed like
𝑇 = 𝜆1 𝑃1 + 𝜆2 𝑃2 +. . . +𝜆𝑚 𝑃𝑚 . . . (𝐴)
where 𝑃𝑖 ’s are pairwise orthogonal. This expression for 𝑇 is called the spectral
resolution of 𝑇. This expression for 𝑇 is unique except for notations and order of
terms.
Theorem: Suppose that 𝑃1 , 𝑃2 , . . . , 𝑃𝑛 are the projections on closed subspaces
𝑀1 , 𝑀2 , . . . , 𝑀𝑛 respectively of 𝐻. Then 𝑃 = 𝑃1 + 𝑃2 +. . . +𝑃𝑛 is a projection if and
only if the 𝑃𝑖 ’s are the pairwise orthogonal. If this condition is satisfied then 𝑃 is
the projection on 𝑀 = 𝑀1 + 𝑀2 +. . . +𝑀𝑛 .
Proof: The operator𝑃 is clearly self-adjoint. Therefore, 𝑃 is a projection if and
only if 𝑃2 = 𝑃. Suppose that 𝑃𝑖 ’s are pairwise orthogonal, then
𝑛
2
𝑃 = 𝑃12 + 𝑃22 +. . . +𝑃𝑛2 + 𝑃𝑖 𝑃𝑗 , 𝑖 ≠ 𝑗
𝑖,𝑗 =1

= 𝑃12 + 𝑃22 + ⋯ + 𝑃𝑛2 = 𝑃1 + 𝑃2 + ⋯ + 𝑃𝑛 = 𝑃

So, 𝑃 is a projection operator.
Suppose now that 𝑃 is a projection operator. Let 𝑥 be an element in the range
of 𝑃𝑖 , i.e., 𝑥 ∈ 𝑀𝑖 , so we have 𝑃𝑖 𝑥 = 𝑥. Then
𝑛 𝑛 𝑛 𝑛
2 2 2
𝑥 = 𝑃𝑖 𝑥 ≤ 𝑃𝑗 𝑥 = 𝑃𝑗 𝑥, 𝑃𝑗 𝑥 = 𝑥, 𝑃𝑗 ∗ 𝑃𝑗 𝑥 = 𝑥, 𝑃𝑗 𝑥
𝑗 =1 𝑗 =1 𝑗 =1 𝑗 =1
𝑛
2 2
= 𝑥, 𝑃𝑗 𝑥 = 𝑥, 𝑃𝑥 ≤ 𝑥 𝑃 ≤ 𝑥 , 𝑏𝑒𝑐𝑎𝑢𝑠𝑒 𝑃
𝑗 =1
2
=1 ∵𝑃 =𝑝
So all the steps are equal and therefore
𝑛
2 2
𝑃𝑗 𝑥 = 𝑃𝑖 𝑥
𝑗 =1

and so 𝑃𝑖 𝑥 = 0 for 𝑗 ≠ 𝐼, i.e., 𝑃𝑗 𝑥 = 𝜃 for 𝑗 ≠ 𝑖.

5
This indicates that the range 𝑀𝑖 of 𝑃𝑖 is contained in 𝑁𝑗 , the null space of 𝑃𝑗 i.e.,
𝑀𝑖 ⊂ 𝑁𝑗 for 𝑖 ≠ 𝑗.

On the other hand, if 𝑦 ∈ 𝑀𝑗 and 𝑗 ≠ 𝑖 and 𝑥 ∈ 𝑁𝑗 then 𝑦, 𝑥 = 𝑃𝑗 𝑦, 𝑥 =

𝑦, 𝑃𝑗 ∗ 𝑥 = 𝑦, 𝑃𝑗 𝑥 = 𝑦, 𝜃 = 0.[∵ 𝑦 ∈ 𝑀𝑗 , ∴ 𝑃𝑗 𝑦 = 𝑦]

So 𝑥 ⊥ 𝑦, i.e., 𝑥 ∈ 𝑀𝑗⊥ and therefore 𝑁𝑗 ⊂ 𝑀𝑗⊥ . Hence from above, we have

𝑀𝑖 ⊂ 𝑀𝑗⊥ for 𝑗 ≠ 𝑖. This means that 𝑀𝑖 ⊥ 𝑀𝑗 , 𝑖 ≠ 𝑗 and by an earlier theorem, the
𝑃𝑖 ’s are pairwise orthogonal.
Now, we prove the last statement we have already seen that if 𝑥 ∈ 𝑀𝑖 then
𝑃𝑗 𝑥 = 𝜃 for 𝑗 ≠ 𝑖. Let 𝑥 ∈ 𝑀𝑖 . Then 𝑃𝑥 = 𝑃1 𝑥 + 𝑃2 𝑥 + ⋯ + 𝑃𝑛 𝑥 = 𝑃𝑖 𝑥 = 𝑥 and so
each 𝑀𝑖 is contained in the range of 𝑃. Clearly, the range of 𝑃 is a subspace and
so 𝑀 = 𝑀1 + 𝑀2 +. . . +𝑀𝑛 is contained in the range of 𝑃.
Now, let 𝑥 be an element in the range 𝑃. The 𝑥 = 𝑃1 𝑥 1 + 𝑃2 𝑥 1 +. . . +𝑃𝑛 𝑥 1 is
evidently in 𝑀, i.e., the range of 𝑃 is contained in 𝑀. So, 𝑀 is the same as the
range of 𝑃. Therefore 𝑃 is the projection on 𝑀. This proves the theorem.
Theorem: Suppose that 𝐴 is a self-adjoint operator, 𝐴: 𝐻 → 𝐻. Then
𝐴 = sup 𝐴𝑥, 𝑥 : 𝑥 = 1 .
Proof: Let us write
𝑆𝐴 = sup 𝐴𝑥, 𝑥 : 𝑥 = 1
If 𝑥 = 1, then
2
𝐴𝑥, 𝑥 ≤ 𝐴𝑥 𝑥 ≤ 𝐴 𝑥 = 𝐴
and so 𝑆𝐴 ≤ 𝐴 … (1)
On the other hand if 𝑦 = 1 then clearly
2 2
𝐴𝑦, 𝑦 ≤ sup 𝐴𝑦, 𝑦 . 𝑦 : 𝑦 = 1 = 𝑆𝐴 𝑦
𝑦
If 𝑦 ≠ 0 and 𝑦 ≠ 𝜃, let 𝑦′ = , then 𝑦′ = 1 and
𝑦

𝐴𝑦′, 𝑦′ ≤ sup 𝐴𝑥, 𝑥 : 𝑥 = 1 = 𝑆𝐴

1
𝑖. 𝑒., 2
𝐴𝑦, 𝑦 ≤ 𝑆𝐴
𝑦
2
𝑖. 𝑒., 𝐴𝑦, 𝑦 ≤ 𝑆𝐴 𝑦 … (2)
The inequality (2) clearly holds if 𝑦 = 𝜃 also.
If 𝑧 ∈ 𝐻 and 𝑧 ≠ 𝜃, we put
1 2
𝐴𝑧 1
𝜆= 𝑎𝑛𝑑 𝑢 = 𝐴𝑧.
𝑧 𝜆

We obtain that

6
 2
𝐴𝑧 = 𝐴𝑧, 𝐴𝑧 = 𝐴𝜆𝑧, 𝑢
1
= 𝐴 𝜆𝑧 + 𝑢 , 𝜆𝑧 + 𝑢 , 𝐴 𝜆𝑧 − 𝑢 , 𝜆𝑧 − 𝑢 , 𝑏𝑒𝑐𝑎𝑢𝑠𝑒 𝐴 𝑖𝑠 𝑠𝑒𝑙𝑓 − 𝑎𝑑𝑗𝑜𝑖𝑛𝑡
4
1 2 2
≤ 𝑆𝐴 𝜆𝑧 + 𝑢 + 𝜆𝑧 − 𝑢 ,
4
2
because 𝐴𝑦, 𝑦 ≤ 𝑆𝐴 𝑦 and − 𝐴𝑦, 𝑦 ≤ 𝑆𝐴 𝑦 2 ,
1 2 2
= 𝑆 𝜆𝑧 + 𝑢 , 𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙𝑜𝑔𝑟𝑎𝑚 𝑙𝑎𝑤
2 𝐴
1 1
= 𝑆𝐴 𝜆2 𝑧 2
+ 𝐴𝑧 2
2 𝜆2
1 𝐴𝑧 2
𝑧 2
= 𝑆𝐴 𝑧 + 𝐴𝑧
2 𝑧 𝜆2 𝐴𝑧
1
= 𝑆𝐴 𝐴𝑧 𝑧 + 𝑧 𝐴𝑧 = 𝑆𝐴 𝑧 𝐴𝑧 .
2
So, 𝐴𝑧 ≤ 𝑆𝐴 𝑧 and consequently
𝐴 ≤ 𝑆𝐴 . . . (3)
From (1) and (3), we obtain that
sup 𝐴𝑥, 𝑥 : 𝑥 = 1 = 𝑆𝐴 = 𝐴
This proves the theorem.
1
[ 𝐴𝜆𝑧, 𝑢 = 𝐴𝜆𝑧, 𝜆 𝐴𝑧 = 𝐴𝑧, 𝐴𝑧

So 𝐴𝑧, 𝐴𝑧 = 𝐴𝜆𝑧, 𝑢 = 𝜆𝑧, 𝐴 ∗ 𝑢 = 𝜆𝑧, 𝐴𝑢 [∵ 𝐴 = 𝐴 ∗]]

Closed Linear Operators: Suppose that 𝑋 and 𝑌 are sets and 𝐴 is an operator from
some subset 𝐷 of 𝑋 into 𝑌. Then the collection of order pairs 𝐺𝐴 = { 𝑥, 𝐴 𝑥 , 𝑥 ∈ 𝐷} is
called the graph of the operator 𝐴.
The Cartesian product of two linear spaces 𝑋 and 𝑌 consisting of the ordered
pairs [𝑥, 𝑦] where 𝑥 ∈ 𝑋 and 𝑦 ∈ 𝑌 and denoted by 𝑋 × 𝑌 as
𝑥1 , 𝑦1 + 𝑥2 , 𝑦2 = [𝑥1 + 𝑥2 , 𝑦1 + 𝑦2 ]
and scalar multiplication
𝛼 𝑥, 𝑦 = [𝛼𝑥, 𝛼𝑦]
So the graph 𝐺𝐴 is considered as a subset of 𝑋 × 𝑌 and if 𝐷 is a subspace
then it may be verified that 𝐺𝐴 is a subspace of 𝑋 × 𝑌 provided 𝐴 is a linear operator.
If 𝑋 and 𝑌 are normed linear spaces, then we can also introduce a norm in
𝑋 × 𝑌. If [𝑥, 𝑦] ∈ 𝑋 × 𝑌 then we define
[𝑥, 𝑦] = 𝑥 + 𝑦

7
 We can now verify that with this definition of norm, 𝑋 × 𝑌 becomes a normed
linear space. We introduce the following definition:
Definition: Let, 𝑋 and 𝑌 be normed linear spaces and 𝐷 be a subspace of 𝑋. Then
the linear operator 𝐴: 𝐷 → 𝑌 is called closed if the relations 𝑥𝑛 ∈ 𝐷, 𝑥𝑛 → 𝑥, 𝐴𝑥𝑛 → 𝑦
imply that 𝑥 ∈ 𝐷 and 𝐴𝑥 = 𝑦.
We can now distinguish between a closed operator and a continuous linear
operator. If the linear operator 𝐴 is continuous and if lim⁡𝑥𝑛 exists then lim⁡𝐴𝑥𝑛 also
exists; whereas if the operator 𝐴 is only closed then the existence of lim⁡𝑥𝑛 does not
necessarily guarantee the existence of lim⁡𝐴𝑥𝑛 .
Theorem: The linear operator 𝐴 is closed if and only if its graph is a closed
subspace.
Proof: We suppose that 𝐴: 𝐷 → 𝑌, 𝐷 ⊂ 𝑋 is a closed operator. We show 𝐺𝐴 is closed
let 𝑥𝑛 , 𝐴𝑥𝑛 → [𝑥, 𝑦] as 𝑛 → ∞,
or, 𝑥𝑛 , 𝐴𝑥𝑛 − [𝑥, 𝑦] → 0 as 𝑛 → ∞
or, 𝑥𝑛 − 𝑥, 𝐴𝑥𝑛 − 𝑦 → 0 as 𝑛 → ∞
or 𝑥𝑛 − 𝑥 + 𝐴𝑥𝑛 − 𝑦 → 0 as 𝑛 → ∞
This implies that 𝑥𝑛 → 𝑥 and 𝐴𝑥𝑛 → 𝑦. Since 𝐴 is closed, we must have then
𝑥 ∈ 𝐷 and 𝑦 = 𝐴𝑥. Therefore 𝑥, 𝑦 = [𝑥, 𝐴𝑥] ∈ 𝐺𝐴 and thus 𝐺𝐴 is thus closed.
Conversely, suppose that 𝐺𝐴 is closed and that 𝑥𝑛 → 𝑥, 𝑥𝑛 ∈ 𝐷 and 𝐴𝑥𝑛 → 𝑦.
So, 𝑥𝑛 , 𝐴𝑥𝑛 − [𝑥, 𝑦]
= 𝑥𝑛 − 𝑥 + 𝐴𝑥𝑛 − 𝑦 → 0 as 𝑛 → ∞
i.e., 𝑥𝑛 , 𝐴𝑥𝑛 → [𝑥, 𝑦]
Since, 𝐺𝐴 is closed, it follows that [𝑥, 𝑦] ∈ 𝐺𝐴 . By the definition of 𝐺𝐴 , it follows that
𝑥 ∈ 𝐷 and 𝑦 = 𝐴𝑥. Therefore, 𝐴 is closed and this proves the theorem.
Theorem: Let 𝑋 and 𝑌 be normed linear spaces and 𝐷 ⊂ 𝑋 be a closed subspace of
𝑋. If 𝐴 is a bounded linear operator such that 𝐴: 𝐷 → 𝑌, then 𝐴 is closed.
Proof: Let 𝑥𝑛 ∈ 𝐷, 𝑥𝑛 → 𝑥 and 𝐴𝑥𝑛 → 𝑦, then since 𝐷 is closed, 𝑥 ∈ 𝐷. Continuity of 𝐴
gives that 𝐴𝑥𝑛 → 𝐴𝑥 and so 𝑦 = 𝐴𝑥 and 𝐴 becomes closed. This proves the theorem.
Now since the space 𝑋 may be considered as a closed subspace,
we have the following corollary.
Corollary: If 𝐴 is a continuous linear operator such that 𝐴: 𝑋 → 𝑌, then 𝐴 is closed.
Theorem: Suppose that 𝐴 is a closed linear operator such that 𝐴: 𝐷 → 𝑌, 𝐷 ⊂ 𝑋.
Then if 𝐴−1 exists, it is a closed linear operator.
Proof: If 𝐴−1 exists, it is a linear operator by an earlier theorem (Let 𝐴: 𝐸 → 𝐹 be a
linear operator. If 𝐴−1 exists, then 𝐴−1 is linear)

8
Since 𝐴 is closed, 𝐺𝐴 = { 𝑥, 𝐴𝑥 : 𝑥 ∈ 𝐷} is a closed subspace.
Consider 𝑅(𝐴), the range of 𝐴. Since 𝐴−1 exists, there is a unique 𝑥 ∈ 𝐷 such that
𝑦 = 𝐴𝑥 i.e., 𝑥 = 𝐴−1 𝑦. Therefore 𝐺𝐴 can be written as
𝐺𝐴 = { 𝐴−1 𝑦, 𝑦 : 𝑦 ∈ 𝑅 𝐴 }
If we consider the mapping 𝑋 × 𝑌 → 𝑌 × 𝑋, i.e., 𝑥, 𝑦 → [𝑦, 𝑥] then
this mapping is an isometry. Because, clearly an isometric mapping maps a closed
set into a closed set, it follows that { 𝑦, 𝐴−1 𝑦 : 𝑦 ∈ 𝑅 𝐴 } ---(i) is closed.
Since 𝐴−1 is linear, this set is also a subspace. Hence the set (i) is a closed
subspace. But this set is the graph of 𝐴−1 . So by an earlier theorem, 𝐴−1 is closed.
This proves the theorem.
We now present below two fundamental theorems- open mapping
theorem and closed graph theorem. For this purpose, we introduce some definition:
Definition: Let 𝑋 and 𝑌 be metric spaces and 𝑓 be mapping of 𝑋 into 𝑌. Then 𝑓 is
called an open mapping if whenever 𝐺 is open in 𝑋, 𝑓 𝐺 = {𝑓 𝑥 : 𝑥 ∈ 𝐺} is open in
𝑌.
Definition: Let 𝑋 and 𝑌 be metric spaces and 𝑓 be mapping of 𝑋 onto 𝑌. Then 𝑓 is
called a homeomorphism if 𝑓 and 𝑓 −1 are each continuous functions.
Theorem (Open Mapping Theorem): If 𝐵 and 𝐵′ are Banach spaces and if 𝑇 is a
continuous linear operator of 𝐵 onto 𝐵′, then 𝑇 is an open mapping.
Open mapping theorem has various applications. These
applications generally use the following theorem which is an immediate
consequence of the open mapping theorem.
Theorem: If 𝑇 is a one-to-one continuous linear operator of one Banach space onto
another, then 𝑇 is a homeomorphism.
The above theorem has an immediate application in the following
basic theorem which is known as closed graph theorem.
Theorem (Closed Graph Theorem): If 𝐵 and 𝐵′ are Banach spaces and if 𝑇 is a
linear operator of 𝐵 into 𝐵′ then 𝑇 is continuous if its graph is closed (i.e., 𝑇 is
closed).
Proof: We denote by 𝐵1 the linear space 𝐵 renormed by
𝑥 1 = 𝑥 + 𝑇(𝑥)
We verify that with this norm, 𝐵1 is a Banach space. 𝐵1 is already a
normed linear space. We show that 𝐵1 is complete. Let {𝑥𝑛 } be a Cauchy sequence
in 𝐵1 . Then
𝑥𝑛 − 𝑥 𝑚 1 = 𝑥𝑛 − 𝑥𝑚 + 𝑇(𝑥𝑛 ) − 𝑇(𝑥𝑚 )

9
 From this it follows that {𝑥𝑛 } and {𝑇(𝑥𝑛 )} are both Cauchy
sequences in 𝐵 and 𝐵′ respectively. Since 𝐵 and 𝐵′ both are complete, there are
elements 𝑥 and 𝑦 in 𝐵 and 𝐵′ respectively such that
𝑥𝑛 → 𝑥 𝑎𝑛𝑑 𝑇(𝑥𝑛 ) → 𝑦
Since 𝑇 is closed, it follows that 𝑦 = 𝑇(𝑥).
So, 𝑥𝑛 − 𝑥 1 = 𝑥𝑛 − 𝑥 + 𝑇 𝑥𝑛 − 𝑇 𝑥 = 𝑥𝑛 − 𝑥 + 𝑇 𝑥𝑛 − 𝑦 → 0 as 𝑛 → ∞.
Consequently, 𝐵1 is complete and so is a Banach space.
Now, 𝑇𝑥 ≤ 𝑥 + 𝑇 𝑥 = 𝑥 1

shows that 𝑇 is bounded and consequently continuous as an operator from 𝐵1 onto 𝐵

is clearly continuous, because
𝑥 ≤ 𝑥 + 𝑇 𝑥 = 𝑥 1 𝑎𝑛𝑑 𝑥𝑛 − 𝑥 ≤ 𝑥𝑛 − 𝑥 1 .
Therefore, if 𝑥𝑛 → 𝑥 in 𝐵1 then 𝑥𝑛 → 𝑥 in 𝐵. By an earlier theorem, the identity
mapping from 𝐵1 onto 𝐵 is theorefore a homeomorphism.
We now show that 𝑇 is continuous from 𝐵 into 𝐵′. The operator 𝑇 is
already continuous from 𝐵1 into 𝐵′. So, if 𝐺 is open in 𝐵′, then 𝑇 −1 𝐺 is open in 𝐵1
and since the identity mapping from 𝐵1 onto 𝐵 is a homeomorphism (so open also), it
follows that 𝑇 −1 𝐺 is open in 𝐵 also which proves the continuity of 𝑇. This proves
the theorem.
Theorem: If 𝑇 is a positive operator on a Hilbert space 𝐻 then 𝐼 + 𝑇 is non-singular.
Proof: We know that any transformation is non-singular ⟺ It is one-one and onto.
To show that (𝐼 + 𝑇) is one-one
Let (𝐼 + 𝑇)𝑥 = (𝐼 + 𝑇)𝑦
⇒ (𝐼 + 𝑇)(𝑥 − 𝑦) = 0
⇒ (𝐼 + 𝑇)𝑧 = 0 𝑤𝑕𝑒𝑟𝑒 𝑧 = 𝑥 − 𝑦
⇒ 𝐼𝑧 + 𝑇𝑧 = 0 ⇒ 𝑇𝑧 = −𝑧
2
⇒ (𝑇𝑧, 𝑧) = (−𝑧, 𝑧) = −(𝑧, 𝑧) = − 𝑧
2
But L.H.S (𝑇𝑧, 𝑧) is ≥ 0 as 𝑇 is a positive operator where as R.H.S − 𝑧 is negative.
Above is possible only when each side is zero.
2
∴ 𝑧 =0
Consequently 𝑧 = 0 or 𝑥 − 𝑦 = 0 or 𝑥 = 𝑦.
Hence, 𝐼 + 𝑇 is one-one.
To show that (𝐼 + 𝑇) is onto

10
Since 𝐼 + 𝑇 is one-one and for showing it to be onto we have to establish that range
of 𝐼 + 𝑇 denoted by 𝑅(𝐼 + 𝑇) (say) is 𝐻 itself. For this we shall first establish that
𝑅(𝐼 + 𝑇) is a closed linear subspace of 𝐻.
Let 𝑥, 𝑦 ∈ 𝐻 so that 𝛼𝑥 + 𝛽𝑦 ∈ 𝐻 where 𝛼, 𝛽 are scalars. (𝐼 + 𝑇)𝑥,
(𝐼 + 𝑇)𝑦 both belong to 𝑅(𝐼 + 𝑇).
Now, 𝛼(𝐼 + 𝑇)𝑥 + 𝛽(𝐼 + 𝑇)𝑦
= (𝐼 + 𝑇)𝛼𝑥 + (𝐼 + 𝑇)𝛽𝑦
= (𝐼 + 𝑇)(𝛼𝑥 + 𝛽𝑦)
= (𝐼 + 𝑇)𝑧 𝑤𝑕𝑒𝑟𝑒 𝑧 ∈ 𝐻
∈ 𝑅(𝐼 + 𝑇)
∴ 𝑅(𝐼 + 𝑇) is a subspace of 𝐻.
Now we shall show that 𝑅(𝐼 + 𝑇) is a closed linear subspace, i.e.,
every limit point of it belongs to it or it is sufficient to show that it is complete, i.e.,
every Cauchy sequence in it is convergent.
Let 𝐼 + 𝑇 𝑥𝑛 be a Cauchy sequence in 𝑅(𝐼 + 𝑇) so that for every
𝜖 > 0∃𝑛0 such that ∀𝑚, 𝑛 > 𝑛0 ,
(𝐼 + 𝑇)𝑥𝑛 − (𝐼 + 𝑇)𝑥𝑚 < 𝜖
𝑜𝑟, (𝐼 + 𝑇)(𝑥𝑛 − 𝑥𝑚 ) < 𝜖
𝑜𝑟, (𝑥𝑛 − 𝑥𝑚 ) + 𝑇(𝑥𝑛 − 𝑥𝑚 ) < 𝜖
2
𝑜𝑟, (𝑥𝑛 − 𝑥𝑚 ) + 𝑇(𝑥𝑛 − 𝑥𝑚 ) < 𝜖2 . . . (1)
But we know that
2
𝑥+𝑦 = 𝑥 + 𝑦, 𝑥 + 𝑦
2 2
= 𝑥 + 𝑦 + (𝑥, 𝑦) + 𝑦, 𝑥
2 2
= 𝑥 + 𝑦 + (𝑥, 𝑦) + 𝑥, 𝑦
∴ By (1),
2 2
𝑥𝑛 − 𝑥𝑚 + 𝑇 𝑥𝑛 − 𝑥𝑚 + 𝑇 𝑥𝑛 − 𝑥𝑚 , 𝑥𝑛 − 𝑥𝑚 + 𝑇 𝑥 𝑛 − 𝑥𝑚 , 𝑥𝑛 − 𝑥𝑚
< 𝜖2
But when 𝑇 is self-adjoint then (𝑇𝑥, 𝑥) is real so that (𝑇𝑥, 𝑥) = (𝑇𝑥, 𝑥) and as 𝑇 is
positive, (𝑇𝑥, 𝑥) ≥ 0. [(𝑇𝑥, 𝑥) = (𝑥, 𝑇 ∗ 𝑥) = (𝑥, 𝑇𝑥) = (𝑇𝑥, 𝑥).
2 2
𝑥𝑛 − 𝑥𝑚 + 𝑇 𝑥𝑛 − 𝑥𝑚 + 2 𝑇 𝑥𝑛 − 𝑥𝑚 , 𝑥𝑛 − 𝑥 𝑚 < 𝜖2

All the terms in the L.H.S are positive.

2
∴ 𝑥 𝑛 − 𝑥𝑚 < 𝜖 2 ⇒ 𝑥𝑛 − 𝑥𝑚 < 𝜖.

11
⇒ 𝑥𝑛 is a Cauchy sequence in 𝐻 which is complete.
Therefore 𝑥𝑛 is a convergent sequence and let 𝑥𝑛 → 𝑥 in 𝐻.
So, (𝐼 + 𝑇)𝑥 ∈ 𝑅(𝐼 + 𝑇) is the limit of (𝐼 + 𝑇)𝑥𝑛 .
Thus, 𝑅(𝐼 + 𝑇) is a closed linear subspace of 𝐻.
We have to prove that 𝑅(𝐼 + 𝑇) is 𝐻 itself. Let us suppose that 𝑅(𝐼 + 𝑇)
is not equal to 𝐻 then it is a proper closed linear subspace of 𝐻 and in that case by
an earlier theorem, (Statement: If 𝑀 is a proper closed subspace of a Hilbert space
𝐻 then ∃𝑦0 (≠ 0) ∈ 𝐻 such that 𝑦0 ⊥ 𝑀) there exists a non-zero element 𝑦0 in 𝐻 such
that 𝑦0 ⊥ 𝑅(𝐼 + 𝑇).
Now,
𝑦0 ∈ 𝐻 ⇒ (𝐼 + 𝑇)𝑦0 ∈ 𝑅(𝐼 + 𝑇)
∴ 𝑦0 ⊥ (𝐼 + 𝑇)𝑦0 𝑖. 𝑒., 𝑦0 ⊥ (𝑦0 + 𝑇𝑦0 )
𝑖. 𝑒. , 𝑦0 + 𝑇𝑦0 , 𝑦0 = 0
𝑖. 𝑒., (𝑦0 , 𝑦0 ) + (𝑇𝑦0 , 𝑦0 ) = 0
2
𝑖. 𝑒., (𝑇𝑦0 , 𝑦0 ) = − 𝑦0
2
As 𝑇 is positive, therefore, (𝑇𝑦0 , 𝑦0 ) ≥ 0 and so − 𝑦0 ≥ 0. Above is possible only
when 𝑦0 = 0, i.e., 𝑦0 = 𝜃.
But this will be a contradiction as 𝑦0 is a non-zero element in 𝐻 such
that 𝑦0 ⊥ 𝑅(𝐼 + 𝑇). Hence our supposition that 𝑅(𝐼 + 𝑇) is not equal to 𝐻 is wrong.
∴ 𝑅(𝐼 + 𝑇) = 𝐻
𝑖. 𝑒., 𝐼 + 𝑇 𝑖𝑠 𝑜𝑛𝑡𝑜.
Since 𝐼 + 𝑇 is both one-one onto it follows that it is non-singular.
Corollary: We have already proved that both 𝑇𝑇 ∗ and 𝑇 ∗ 𝑇 are positive operators
and hence by the above theorem, we can say that both 𝐼 + 𝑇𝑇 ∗ and 𝐼 + 𝑇 ∗ 𝑇 are
non-singular.
Theorem:𝐸 is a projection on a closed linear subspace 𝑀 of 𝐻 ⇔ 𝐼 − 𝐸 is a
projection on 𝑀⊥ .
Proof: Let 𝐸 be a projection on 𝐻 then 𝐸 2 = 𝐸 and 𝐸 ∗= 𝐸 . . . (1)
Now, (𝐼 − 𝐸) ∗= 𝐼 ∗ −𝐸 ∗= 𝐼 − 𝐸, 𝑖. 𝑒, 𝐼 − 𝐸 is self-adjoint.
2
𝐼−𝐸 = (𝐼 − 𝐸)(𝐼 − 𝐸) = 𝐼 − 𝐸 − 𝐸 + 𝐸 2 = 𝐼 − 𝐸 − 𝐸 + 𝐸 = 𝐼 − 𝐸
Hence, 𝐼 − 𝐸 is also a projection on 𝐻.
Now, if 𝑀 be the range of 𝐸 we will show that the range of 𝐼 − 𝐸 is 𝑀⊥ .
Let 𝑁 be the range of 𝐼 − 𝐸 then

12
 𝑥 ∈ 𝑁, ⇒ (𝐼 − 𝐸)𝑥 = 𝑥 ⇒ 𝑥 − 𝐸𝑥 = 𝑥 ⇒ 𝐸𝑥 = 0 ⇒ 𝑥 ∈ 𝑁𝑢𝑙𝑙 𝑠𝑝𝑎𝑐𝑒 𝑜𝑓 𝐸 ⇒ 𝑥 ∈ 𝑀⊥
∴ 𝑁 ⊆ 𝑀⊥ . . . (2)
Again let 𝑥 ∈ 𝑀⊥ which is null space of 𝐸 so that 𝐸𝑥 = 0, which implies that
𝑥 − 𝐸𝑥 = 𝑥 ⇒ 𝐼 − 𝐸 𝑥 = 𝑥 ⇒ 𝑥 ∈ 𝑅𝑎𝑛𝑔𝑒𝑜𝑓 𝐼 − 𝐸 , 𝑖. 𝑒. , 𝑥 ∈ 𝑁
∴ 𝑀⊥ ⊆ 𝑁 … (3)
Hence from (2) and (3), we get that 𝑁 = 𝑀⊥ ,
𝑖. 𝑒., 𝑟𝑎𝑛𝑔𝑒 𝑜𝑓 𝐼 − 𝐸 𝑖𝑠 𝑀⊥
Conversely, let 𝐼 − 𝐸 be a projection on 𝑀⊥ then by the first part proved above
𝐼 − (𝐼 − 𝐸) is a projection on 𝑀⊥ ⊥
𝑖. 𝑒., 𝐸 is a projection on 𝑀 (Since, 𝑀⊥ ⊥
= 𝑀, 𝑀 being a closed linear subspace of
𝐻)
This proves the theorem.

********************************

13