Indian Statistical Institute, Kolkata

Second semester 2023-24

M.Math I year
Subject : Functional Analysis
Lecture 2

DefinitionP0.1. Let {vn }∞

P
n=1 be a sequence of points in V . Then the seriesP n vn is sum-
∞ ∞
mable if { m m
P
n=1 vn }m=1 converges, and n vn is absolutely summable if { n=1 ||vn ||}m=1
converges.
This is basically analogous to the definitions of convergence and absolute convergence for
series for real numbers, and we have a similar result as well:
∞
Proposition 0.2. If n vn is absolutely summable, then the sequence of partial sums { m
P P
n=1 vn }m=1
is Cauchy.
This proof is left to us as an exercise (it’s the same proof as when V = R), and we should
note that the theorem is that we have a Cauchy sequence, not necessarily that it is summable
(like in the real-valued case). And that’s because we need completeness, and that leads to
our next result:
Theorem 0.3. A normed vector space V is a Banach space if and only if every absolutely
summable series is summable.
This is sometimes an easier property to verify than going through the Cauchy business –
in particular, it’ll be useful in integration theory later on.
Proof. We need to prove both directions. For the forward direction, suppose that V is Ba-
nach. Then V is complete, so any absolutely summable series is Cauchy and thus convergent
in V (that is, summable).
For the opposite direction, suppose that every absolutely summable series is summable.
Then for any Cauchy sequence {vn }, let’s first show that we can find a convergent subse-
quence. (This will imply that the whole sequence converges by a triangle-inequality metric
space argument.)
To construct this subsequence, we basically “speed up the Cauchy-ness of {vn }.” We know
that for all k ∈ N, there exists Nk ∈ N such that for all n, m ≥ Nk , we have
||vn − vm || < 2−k .
(We’re choosing 2−k because it’s summable.) So now we define
nk = N1 + · · · + Nk ,
so n1 < n2 < n3 < · · · is an increasing sequence of integers, and for all k, nk ≥ Nk . And
now we claim that {vnk } converge: after all,
||vnk+1 − vnk || < 2−k
(because of how we choose nk and nk+1 ), and therefore the series
X
(vnk +1 − vnk )
k∈N
must be summable (it’s absolutely summable because k∈N 2−k = 1, and we assumed that
P
all absolutely summable sequences are summable). Thus the sequence of partial sums
m
X
(vnk+1 − vnk ) = vnm+1 − vn1
k=1

converges in V , and adding vn1 to every term does not change convergence. Thus the
sequence {vnm+1 }∞
m=1 converges, and we’ve found our convergent subsequence (meaning that
the whole sequence indeed converges). This proves that V is Banach. 
Now that we’ve appropriately characterized our vector spaces, we want to find the analog
of matrices from linear algebra, which will lead us to operators and functionals. Here’s
a particular example to keep in mind (because it motivates a lot of the machinery that we’ll
be using):
Example 0.4. Let K : [0, 1] × [0, 1] → C be a continuous function. Then for any function
f ∈ C([0, 1]), we can define
Z 1
T f (x) = K(x, y)f (y)dy.
0
The map T is basically the inverse operators of differential operators, but we’ll see that later
on.
We can check that T f ∈ C([0, 1]) (it’s also continuous), and for any λ1 , λ2 ∈ C and
f1 , f2 ∈ C([0, 1]), we have
T (λ1 f1 + λ2 f2 ) = λ1 T f1 + λ2 T f2
(linearity). We’ve already proven that C([0, 1]) is a Banach space, so T here is going to be
an example of a linear operator.
Definition 0.5. Let V and W be two vector spaces. A map T : V → W is linear if for all
λ1 , λ2 ∈ K and v1 , v2 ∈ V ,
T (λ1 v1 + λ2 v2 ) = λ1 T v1 + λ2 T v2 .
(We’ll often use the phrase linear operator instead of “linear map” or “linear trans-
formation.”)
We’ll be particularly curious about linear operators that are continuous: recall that a map
T : V → W (not necessarily linear) is continuous on V if for all v ∈ V and all sequences
{vn } converging to v, we have T vn → T v. (Equivalently, we can use the topological notion
of continuity and say that for all open sets U ⊂ W , the inverse image
T −1 (U ) = {v ∈ V : T v ∈ U }
is open in V .) For linear maps, there’s a way of characterizing whether a function is con-
tinuous on a normed space – in finite-dimensional vector spaces, all linear transfor-
mations are continuous, but this is not always true when we have a map between two
Banach spaces.
Theorem 0.6. Let V, W be two normed vector spaces. A linear operator T : V → W is
continuous if and only if there exists C > 0 such that for all v ∈ V , ||T v||W ≤ C||v||V .
2
 In this case, we say that T is a bounded linear operator, but that doesn’t mean the
image of T is bounded – the only such linear map is the zero map! Instead, we’re saying
that bounded subsets of V are always sent to bounded subsets of W .
Proof. First, suppose that such a C > 0 exists (such that ||T v||W ≤ C||v||V for all v ∈ V ):
we will prove continuity by showing that T vn → T v for all {vn } → v. Start with a convergent
subsequence vn → v: then
||T vn − T v||W = ||T (vn − v)||W
(by linearity of T ), and now by our assumption, this can be bounded as
≤ C||vn − v||V .
Since ||vn − v||V → 0, the squeeze theorem tells us that ||T vn − T v||W → 0 (since the norm
is always nonnegative), and thus T vn → T v.
For the other direction, suppose that T is continuous. This time we’ll describe continuity
with the topological characterization: the inverse of every open set in W is an open set in
V , so in particular, the set
T −1 (BW (0, 1)) = {v ∈ V : T v ∈ BW (0, 1)}
is an open set in V . Since 0 is contained in BW (0, 1), and T (0) = 0, we must have 0 ∈
T −1 (BW (0, 1)), and (by openness) we can find a ball of some radius r > 0 so that BV (0, r) is
contained inside T −1 (BW (0, 1)). This means that the image of BV (0, r) is contained inside
BW (0, 1).
Now, we claim we can take C = 2r . To show this, for any v ∈ V − {0} (the case v = 0
r
automatically satisfies the inequality), we have the vector 2||v|| V
v, which has length 2r < r.
This means that
 
r r
v ∈ BV (0, r) =⇒ T v ∈ BW (0, 1)
2||v||V 2||v||V
(because BV (0, r) is all sent within BW (0, 1) under T ), and thus
 
r 2
T v < 1 =⇒ ||T (v)||W ≤ ||v||V
2||v||V W r
by taking scalars out of T and using homogeneity of the norm, and we’re done. 
The “boundedness property” above will become tedious to write down, so we won’t use
the subscripts from now on. (But we should be able to track which space we’re thinking
about just by thinking about domains and codomains of our operators.)
Example 0.7. The linear operator T : C([0, 1]) → C([0, 1]) in 0.4 is indeed a bounded linear
operator (and thus continuous).
We should be able to check that T is linear in f easily (because constants come out of the
integral). To check that it is bounded, recall that we’re using the C∞ norm, so if we have a
function f ∈ C([0, 1]),
||f ||∞ = sup |f (x)|
x∈[0,1]
3
(and this supremum value will actually be attained somewhere, but that’s not important).
We can then estimate the norm of T f by noting that for all x ∈ [0, 1],
Z 1 Z 1
T f (x) = K(x, y)f (y)dy ≤ |K(x, y)| |f (y)|dy
0 0
by the triangle inequality, and now we can bound f and K by their supremum (over [0, 1]
and [0, 1] × [0, 1], respectively) to get
Z 1 Z 1
≤ |K(x, y)| ||f ||∞ dy ≤ ||K(x, y)|| ||f ||∞ dy = ||K(x, y)|| ||f ||∞ .
0 0
Since this bound holds for all x, it holds for the supremum also, and thus
||T f ||x ≤ ||K||∞ ||f ||∞
and we can use C = ||K||∞ to show boundedness (and thus continuity). We will often refer
to K as a kernel.
Definition 0.8. Let V and W be two normed spaces. The set of bounded linear operators
from V to W is denoted B(V, W ).
We can check that B(V, W ) is a vector space – the sum of two linear operators is a linear
operator, and so on. Furthermore, we can put a norm on this space:
Definition 0.9. The operator norm of an operator T ∈ B(V, W ) is defined by
||T || = sup ||T v||.
||v||=1,v∈V

This is indeed a finite number, because being bounded implies that

||T v|| ≤ C||v|| = C
whenever ||v|| = 1, and the operator norm is the smallest such C possible.
Theorem 0.10. The operator norm is a norm, which means B(V, W ) is a normed space.
Proof. First, we show definiteness. The zero operator indeed has norm 0 (because ||T v|| = 0
for all v). On the other  hand,
0
 suppose that T v = 0 for all ||v|| = 1. Then rescaling tells us
that 0 = T v 0 = ||v 0 ||T ||vv0 || = 0 for all v 0 6= 0, so T is indeed the zero operator.
Next, we can show homogeneity, which follows from the homogeneity of the norm on W .
We have
||λT || = sup ||λT v|| = sup |λ|||T v||,
||v||=1 ||v||=1
and now we can pull the nonnegative constant |λ| out of the supremum to get
= |λ| sup ||T v|| = |λ|||T || .
||v||=1

Finally, the triangle inequality also follows from the triangle inequality on W : if S, T ∈
B(V, W ), and we have some element v ∈ V with ||v|| = 1, then
||(S + T )v|| = ||Sv + T v|| ≤ ||Sv|| + ||T v|| ≤ ||S|| + ||T ||.
So taking the supremum of the left-hand side over all unit-length v gives us ||S + T || ≤
||S|| + ||T ||, and we’re done. 
4
 For example, if we return to the operator T from 0.4, we notice that for any f of unit
length, we have
||T f ||∞ ≤ ||K||∞ .
Therefore, ||T || ≤ ||K||. And in general, now that we’ve defined the operator norm, it gives
us a bound of the form
 
v
T ≤ ||T || =⇒ ||T v|| ≤ ||T || ||v||
||v||
for all v ∈ V (not just those with unit length).
Since we have a normed vector space, it’s natural to ask for completeness, which we get
in the following way:
Theorem 0.11. If V is a normed vector space and W is a Banach space, then B(V, W ) is
a Banach space.
Proof. We’ll use the characterization given in 0.3. Suppose that {Tn } is a sequence of
bounded linear operators in B(V, W ) such that
X
C= ||Tn || < ∞.
n

(In other words,

P we have an absolutely summable series of linear operators.) Then we need
to show that n Tn is summable, and we’ll do this in a similar way to how we showed that
the space C∞ (X) was Banach: we’ll come up with a bounded linear operator and show that
we have convergence in the operator norm.
Our candidate will be obtained as follows: for any v ∈ V and m ∈ N, we know that
m
X m
X X
||Tn v|| ≤ ||Tn || ||v|| ≤ ||v|| ||Tn || = C||v||.
n=1 n=1 n

Thus, the sequence of partial sums of nonnegative real numbers m

P
n=1 ||T
Pn v|| is bounded and
thus convergent. Since Tn v ∈ W for each n, we’ve shown
P that a series n Tn v is absolutely
summable in W , and thus (because W is Banach) n Tn v is summable as well. So we can
define the “sum of the Tn s,” T : V → W , by defining
m
X
T v = lim Tn v
m→∞
n=1

(because this limit does indeed exist). We now need to show that this candidate is a bounded
linear operator.
Linearity follows because for all λ1 , λ2 ∈ K and v1 , v2 ∈ V ,
m
X
T (λ1 v1 + λ2 v2 ) = lim Tn (λ1 v1 + λ2 v2 ),
m→∞
n=1

and now because each Tn is linear, this is

m
X m
X
= lim λ1 T v1 + λ2 T v2 .
m→∞
n=1 n=1
5
Now each of the sums converge as we want, since the sum of the limits is the limit of the
sums:
= λ1 T v1 + λ2 T v2 .
(The proof that the sum of two convergent sequences also converges to the sum of the limits
is the same as it is in R, except that we replace absolute values with norms.)
Next, to prove that this linear operator T is bounded, consider any v ∈ V . Then
m
X
||T v|| = lim Tn v ,
m→∞
n=1

and limits and norms interchange, so this is also

m
X m
X
= lim Tn v ≤ lim ||Tn v||
m→∞ m→∞
n=1 n=1

by the triangle inequality. But now this is bounded by

X m
≤ ||Tn || ||v|| = C||v||,
n=1

where C is finite by assumption (because we have an absolutely summable series). So we’ve

P linear operator in B(V, W ).
verified that T is a bounded
It remains to show that mn=1 Tn actually converges to T in the operator norm (as m → ∞).
If we consider some v ∈ V with ||v|| = 1, then
m m 0 m m0
X X X X
Tv − Tn v = lim Tn v − Tn v , = lim Tn v ,
m0 →∞ m0 →∞
n=1 n=1 n=1 n=m+1

and now we can bring the norm inside the limit and then use the triangle inequality to get
m 0 m 0
X X
≤ lim
0
||Tn v|| ≤ lim
0
||Tn ||
m →∞ m →∞
n=m+1 n=m+1

(because v has unit length). And now this is a series of nonnegative real numbers
∞
X
= ||Tn ||,
n=m+1

and thus we note that (taking the supremum over all unit-length v)
m
X ∞
X
T− Tn ≤ ||Tn || → 0
n=1 n=m+1

because we have the tail of a convergent series of real numbers. So indeed we have convergence
in the operator norm as desired. 
Definition 0.12. Let V be a normed vector space (over K). Then V 0 = B(V, K) is called
the dual space of V , and because K = R, C are both complete, V 0 is then a Banach
space by 0.11. An element of the dual space B(V, K) is called a functional.
6
 We can actually identify the dual space for all of the `p spaces: it turns out that
0
(`p )0 = `p ,
where p, p0 satisfy the relation p1 + p10 = 1. So the dual of `1 is `∞ , and the dual of `2 is itself
(this is the only `p space for which this is true), but the dual of `∞ is not actually `1 . (Life
would be a lot easier if this were true, and this headache will come up in the Lp spaces as
well.)