Notes on distribution theory

H.Narayanan

1 Why Distributions?
Consider an ordinary linear differential equation with initial conditions set
to zero
p(D)x = u,
where x(·) and u(·) are real functions over the real line.

Figure 1:

Suppose we construct the functions u1 , . . . , un . . . with un agreeing with

u over the interval [n∆t, (n + 1)∆t] and elsewhere zero.
It is clear that ∞
X
u(·) = un (·)
n=0

Let, as stated earlier, x(0) be zero.

Then the response due to u(·) must be the sum of the responses due to
1
the uj (·). Let p̃τ represent the pulse of width ∆t and a constant height ∆t
in the interval [τ, τ + ∆t].
Let us approximate the functions un (·) by ũn (·) = (u(n∆t) · ∆t)p̃n∆t (·),
essentially making the function constant over [n∆t, (n + 1)∆t].

1
 If ∆t is small we expect
∞
X
u(·) ≈ ũn (·)
n=0

and the response to be approximately the sum of the responses due to the
ũn (·). As ∆t → 0, one expects the approximation to become exact.
SupposeP∞the response due to p̃τ (·) is h̃τ (·). Then the response due to u(·)
would be n=0 h̃τ (·)u(τ ) · ∆t, (approximately) where τ = n∆t, i.e.,
∞
X
y(t) ≈ (h̃τ (t))u(τ ) · ∆t.
n=0

So we expect Z ∞
y(t) = hτ (t)u(τ )dτ,
0

where hτ (t) = lim∆t→0 h̃τ (t). If we assume the system is causal i.e., the
response to an input which is non-zero only for t ≥ t0 is also non-zero only
for t ≥ t0 , then the above integral would have limits from 0 to t. We remind
the reader that h̃τ (·) is the response due to the pulse p̃τ (·) from τ to τ +
1
∆t, of width ∆t and height ∆t . We may think of hτ (·) as the response due
1
to the infinite pulse p̂τ (·) with width ∆t and height ∆t and ∆t → 0.
Of course there exists no such function. But it is nevertheless very conve-
nient to work with this generalized function. Distribution theory will justify
the use of this construct.

2 Definition of distributions: general and tem-

pered
We begin by generalizing the notion of a function over the real line. The
generalization goes through routinely to the functions over Rn i.e., instead of
f (t) we could have f (t1 , t2 , . . . , tn ) where t1 , t2 , . . . , tn are real numbers. We
first build a domain that is ’richer’ than the R. This is the space D of test
functions. A test function φ(·) is defined over the real line and takes complex
values and satisfies the following :

• φ(·) vanishes outside a finite interval and

dk φ
• dtk
exists for every positive integer k (in other words, φ(·) is infinitely
differentiable).

2
 Example of test function:
( 1
e x2 −1 if |x| ≤ 1
φ(x) =
0 if |x| ≥ 0.

It is clear that this function is infinitely differentiable at |x| =

6 1.
Let us examine what happens at x = 1− (i.e., limx→1,x<1). We have
 
d 21 −2x 1
e x −1 = e x2 −1
dx (x2 − 1)2

Now y k e−y → 0 as y → ∞ for all k. Hence

d  x21−1 
lim e = 0.
x→1− dx
The same idea can be used for all orders of derivatives and it can be proved
that
dk  21 
e x −1 = 0 at x = 1−
dxk
Thus the given function is indeed an example of a test function. Observe
that we can shift this function and change its width quite easily. Also its
integral moves from 0 to another constant smoothly from −1 to 1. Therefore
one can build test functions with flat tops (or bottoms) of whatever height
and widthe we please.
A second class of more general functions, namely the class S1 of functions
of rapid decay, is useful in our study. A function of rapid decay φ(·) is defined
over the real line and takes complex values and satisfies the following :

• φ(·) is infinitely differentiable,

• φ(·) together with all its derivatives vanishes at |x| = ∞ faster than
the reciprocal of any polynomial. Thus for each pair of nonnegative
integers k and l,  
 k dl φ 

lim x  = 0.
|x|→∞  dxl 

(We may think of φ as being similar to e−x in its power to kill polyno-
mials).

Clearly every test function is a function of rapid decay since it vanishes

outside a finite interval.
We say a sequence φn of test functions is a ’null sequence’ in D iff

3
 • all φn vanish outside a finite interval,

• φn and all its derivatives approach 0 uniformly in this interval.

Basically the null sequence φn is analogous to a sequence of numbers

tending to zero and we need this idea to talk about small perturbations of a
test function φ̂ by studying φ̂ + φn .
We define a null sequence φm of functions of rapid decay in a similar
manner for each pair of nonnegative integers k and l,
 
 k d l φm 
lim max x   = 0.
m→∞ −∞<x<∞  dxl 

Observe that even after being multiplied by a polynomial of any degree

l
the function ddxφm
l , must be bounded in the interval −∞ < x < ∞ and further

this maximum value must tend to zero as m → ∞.

We are now in a position to define a distribution. A distribution is a
continuous linear functional on the space of test functions, i.e., a distribution
q

• takes a complex value on each test function φ,

usually denoted as < q, φ >

• is linear i.e. < q, α1 φ1 + α2 φ2 >= α1 < q, φ1 > +α2 < q, φ2 >,

• is continuous on the space of test functions i.e. limn→∞ < q, φn >= 0,

whenever φn is a null sequence in D.

A special class of distributions called ‘tempered distributions’ or ‘distri-

butions of slow growth’ are of interest to us. These are linear continuous on
the larger class S1 of functions of rapid decay i.e., limn→∞ < q, φn >= 0, for
every null sequence of rapidly decaying functions, whenever q is a tempered
distribution. Henceforth we use [q, φ] in place of < q, φ >.
Distributions are generalizations of ordinary functions on the real line.
Let f be locally integrable over the real line i.e.,
Z T2
|f (x)| dx
T1

is finite whenever T2 − T1 is finite.

Define Z ∞
[qf , φ] ≡ f (x)φ(x) dx.
−∞

4
Then qf is linear continuous on D. To see this, first note that
Z ∞ Z T2
f (x)φ(x) dx = f (x)φ(x) dx,
−∞ T1

where φ vanishes outside [T1 , T2 ].

Now since φ is differentiable everywhere it must be continuous in [T1 , T2 ]
and therefore has a maximum value in [T1 , T2 ] say M.
So
Z ∞  Z T2
 
 f (x)φ(x) dx. ≤ |f (x)||φ(x)| dx.

−∞ T1
Z T2
≤M |f (x)| dx.
T1
R∞
Hence −∞ f (x)φ(x) dx exists.
The linearity is clear since
Z ∞
f (x)(α1 φ1 (x) + α2 φ2 (x)) dx =
−∞
Z ∞ Z ∞
α1 f (x)φ1 (x) dx + α2 f (x)φ2 (x) dx.
−∞ −∞

To prove continuity we need to show limm→∞ [qf , φm ] → 0 when φm is a

null sequence in D.
We have Z ∞ Z T2
f (x)φm (x) dx = f (x)φm (x) dx
−∞ T1

for some finite T1 , T2 , by the definition of null sequence.

Let Mm = maxT1 ≤x≤T2 φm (x). Hence
Z ∞  Z T2
 
 
f (x)φm (x) dx. ≤ Mm |f (x)| dx

−∞ T1

RT
Since f is locally integrable, T12 |f (x)| dx is finite. Hence, since limm→∞ Mm =
R∞
0, we must have limm→∞ −∞ f (x)φm (x) dx = 0, as needed.
We claim that the action [qf , φ] is indeed a generalization of the value of
the function f . Indeed,we can recover the value of f at points where it is
continuous by making f act on suitable φn and taking limits. Choose φn as a
nonnegative test function with nonnegative values in [a − n1 , a + n1 ] and zero
outside.

5
 R a+ n1 R a+ n1
Further let 1
a− n
φn (x) dx = 1. Observe that 1
a− n
f (x)φn (x) dx gives a
1 1
‘weighted average’ of f in the interval [a − n
n → ∞, provided f
, a + n ]. As
is continuous at a, the integral will therefore converge to f (a).
R ∞ When the context is clear we will denote the distribution qf ([qf , φ] ≡
−∞
f (x)φ(x) dx) also by f .
A distribution qf that arises from a locally integrable function through
the definition Z ∞
[qf , φ] = f (x)φ(x)dx
−∞

is said to be regular. Not all distributions arise in this manner. These latter
distributions are said to be singular. The R ∞ convention even in the case of
singular distributions is to write [q, φ] = −∞ q(x)φ(x)dx.

3 δ function and δ-sequences

For our purpose, the most important singular distribution is the delta [Dirac
delta] functional δ(x) defined by

[δ, φ] = φ(0),

where φ is a function of rapid decay. (In fact δ functional is linear and

continuous even on the space of functionsR continuous at 0). One cannot
∞
expect any function f to have the property −∞ f (x)φ(x)dx = φ(0). However
a sequence of functions, in the limit, can have such property.
Let fn be a sequence of functions continuous at the origin with the fol-
lowing properties.
R∞
1. −∞ fn (x)dx = 1 for each n.
Rb
2. limn→∞ a fn (x)dx = 0, a, b both positive or both negative and there-
Rb
fore limn→∞ a fn (x)dx = 1, if a < 0 < b.

We call such a sequence fn , a delta sequence. Clearly we must have, for every
rapidly decaying φ,
Z ∞ Z −ǫ
fn (x)φ(x)dx = fn (x)φ(x)dx
−∞ −∞
Z ǫ Z ∞
+ fn (x)φ(x)dx + fn (x)φ(x)dx
−ǫ ǫ

6
where ǫ > 0. We then have
Z ∞ Z −ǫ Z ǫ Z ∞ 
lim fn (x)φ(x)dx = lim + +
n→∞ −∞ n→∞ −∞ −ǫ ǫ
Z ǫ
= lim fn (x)φ(x)dx, for every ǫ > 0.
n→∞ −ǫ

Since φ is continuous at φ(0) the right side must be equal to φ(0) (taking the
limit ǫ → 0, ǫ > 0).
So if we define Z ∞
[q, φ] ≡ lim fn (x)φ(x)dx,
n→∞ −∞

where f has the above properties, we will have q = δ.

It is easy to see by the definition that q is a linear functional. Indeed

[q, α1 φ1 + α2 φ2 ] = α1 φ1 (0) + α2 φ2 (0).

To see that q is continuous on the space of rapidly decaying functions we

need to show that limn→∞ [q, φn ] = 0 where {φn } is a null sequence of rapidly
decaying functions. This is clear since limn→∞ φn (0) = 0, by the definition
of a null sequence in the space of rapidly decaying functions.
We often write, whenever fn is a delta sequence

lim fn (x) = δ(x).

n→∞

It is to be interpreted as
Z ∞
[δ, φ] = lim fn (x)φ(x) dx.
n→∞ −∞

We now give a couple of examples of delta sequences which are available in

the literature.

1.
1 n
sn (x) ≡ .
π 1 + n2 x2
These functions are continuous at the origin. It is clear that
Z b
n
lim dx = 0
n→∞ a 1 + n2 x2

whenever a, b are both positive or negative.

7
 d 1 d
We remind the reader that dx
arctan(x) = 1+x2
and therefore dx
arctan(kx) =
k
1+k 2 x2
. Hence
Z ∞
n
dx = [arctan(nx)]∞
−∞
−∞ 1 + n2 x2
= π.
R∞
Clearly therefore, −∞
sn (x) dx = 1. This means that sn is a delta
sequence.

2.
sin(nx)
sn (x) ≡
πx
is a delta sequence.
We claim that

(a) Z ∞
sin(kx)
dx = 1
−∞ πx
This result can be proved by using contour integration. Build a
closed contour C moving from −R1 to −R2 along the real axis,
moving from −R2 to +R2 along a semicircle in the upper half
plane, from R2 to +R1 along the real axis and close the contour
by moving from R1 toR −R1 along a semicircle in the upper half
jkz
plane. Now integrate C eπz dz, letting R1 tend to ∞ and R2 to
zero. Since within the contour there is no pole of the integrand, the
contour integral will equal zero. The larger semicircle integration
can be shown to become zero while the smaller one yields −j.
Now
Z ∞ Z R2 Z R1
sin kx sin kx sin kx
dx = lim dx + dx.
−∞ πx R1 →∞,R2 →0 −R
1
πx R2 πx

The RHS is the imaginary part of the portion of the above contour
integral along the real axis, which by the above argument equals
+j. The result follows.
We can also prove this using Fourier transform ideas as follows.
Let

pk (t) =1 0 ≤ |k|
0 otherwise.

8
 Fourier transform of this function is
Z ∞ Z T
−jωt
e pT (t)dt = e−jωt dt
−∞ −T
T
e−jωt 
=
−jω −T
ejωT − e−jωT
=
jω
2 sin(ωT )
=
ω

Therefore F[ sinπxkx ] = πpk π(−ω) = pk (ω). Observe that

Z ∞ Z ∞ 
sin(xt) sin(xt) −jωt 
dt = e dt
−∞ πt −∞ πt ω=0
= pk (ω)|ω=0 = 1.
(b) We have,
Z b Z kb
sin kx sin υ
dx = dυ, taking υ = kx.
a x ka υ
Let us consider the case when a, b are both positive or both neg-
ative. Using integration by parts as k → ∞, the above integral,
 Z kb 
cos υ kb sin(υ)
= lim −  + dυ
k→∞ υ ka ka υ2
Z kb
1
≤ 0 + lim dυ
k→∞ ka υ 2
kb
1 
≤ lim −  = 0
k→∞ υ ka
We thus see that the sequence sn is a δ-sequence.

4 Operations on Distributions
R∞
The properties of the integral −∞ f (x)φ(x)dx are used to define various
notions related to distributions. In this regard, we will consider the notions
of value over intervals, translations, scaling,derivative and Fourier transform.
We remind the reader that the support of a function is the closure of the set
of points on which it takes non zero values.

9
4.1 Equality in an interval:
R∞
If two functions f1 , f2 have the same value in an interval [a, b] clearly −∞ f1 (x)φ(x)dx =
Rb
a 2
f (x)φ(x)dx, whenever the support of φ is contained in [a, b]. By analogy
we define distributions q1 , q2 to be equal over the interval [a, b] provided
[q1 , φ] = [q2 , φ], whenever the support of φ is contained in [a, b].

4.2 Translation:
Let fa (x) ≡ f (x − a). We say that fa is a translation of f by a. Clearly
Z ∞ Z ∞
fa (x)φ(x)dx = f (x − a)φ(x)dx
−∞ −∞
Z ∞
= f (y)φ(y + a)dy
−∞

where y = x − a.
Thus the action of fa on φ is the same as the action of f on φ−a . In the
case of distributions we are thus motivated to define the translation of the
distribution q by a as follows:-

[qa , φ] ≡ [q, φ−a ].

4.3 Linear Combination:

If we scale a function f by a to yield g, i.e. if g(x) = af (x), we would have
Z ∞ Z ∞
g(x)φ(x)dx = f (x)(aφ(x))dx.
−∞ −∞

We therefore define, for a distribution q

[aq, φ] = [q, aφ] = a[q, φ].

Similarly the sum of the distributions q1 and q2 is defined to be

[q1 + q2 , φ] = [q1 , φ] + [q2 , φ].

10
4.4 Scaling the Domain:
Next suppose we define g by scaling the domain of f , i.e. g(x) ≡ f (ax). In
this case
Z ∞ Z ∞
g(x)φ(x)dx = f (ax)φ(x)dx
−∞ −∞
Z
1 ∞ y
= f (y)φ( )dy, a > 0
a −∞ a
Z ∞
1 y
=− f (y)φ( )dy, a < 0.
a −∞ a

Hence Z ∞ Z ∞
1 y
g(x)φ(x)dx = f (y)φ( )dy.
−∞ |a| −∞ a
We therefore define for a distribution q (writing it as q(x)),
1 x
[q(ax), φ] ≡ [q(x), φ( )].
|a| a

4.5 Differentiation:
Distributions were conceived to handle differential equations in a convenient
manner, particularly impulse response and its derivatives. The entire theory
has been built around the idea that distributions should be differentiable to
all orders. Observe that for a differentiable function f ,
Z ∞ Z ∞
˙
f (x)φ(x)dx = − f (x)φ̇(x)dx + f (x)φ(x)]∞
−∞ .
−∞ −∞

Now let f be a function of ‘slow growth’, i.e., some polynomial grows

faster than f (x) as x → ∞, i.e.,
1
lim f (x) = 0
x→∞ (1 + x2 )p

for some positive integer p. In the above equation, in the RHS f (x)φ(x)]∞
−∞ =
0, since φ(x) either has finite support when it belongs to D or, when it
is rapidly decaying, has the property that limx→∞ p(x)φ(x) = 0 for every
polynomial p(x).
We thus have
Z ∞ Z ∞
˙
f (x)φ(x)dx = − f (x)φ̇(x)dx
−∞ −∞

11
whenever f is a function of slow growth and φ, a rapidly decaying function.
dq
This motivates us to define dx , for a distribution as follows:
  h i
dq
, φ ≡ − q, φ̇ .
dx

4.6 Multiplication by a function f :

If q is a distribution, by the preceding development, we would like to define

[f q, φ] ≡ [q, f φ].

However f φ would not always be a test function or a rapidly decaying func-

tion when φ is one. In general, therefore, we require f to be infinitely dif-
ferentiable for the above definition to work in the case of test functions.
For rapidly decaying functions we need additionally that f grow slower than
some polynomial, i.e. f (x) ≤ cxp , |x| ≥ x0 for some c, p. In the case of special
distributions, this rule can be relaxed. For instance

[f δ, φ] ≡ [δ, f φ] ≡ f (0)φ(0).

Here f δ is clearly defined, provided f is continuous at 0. Similarly we see

 k   k   k 
d δ d δ k d (f φ)
f k,φ ≡ , f φ ≡ (−1) δ,
dx dxk dxk
which is defined if f has continuous k th derivatives.

4.7 Fourier transform of distributions:

Suppose f and φ are both Fourier transformable, we have
Z ∞ Z ∞ Z ∞Z ∞
−jyx
f (y) φ(x)e dxdy = f (y)φ(x)e−jyx dxdy
−∞ −∞
Z−∞
∞
−∞
Z ∞
= φ(x) f (y)e−jyxdydx.
−∞ −∞
R∞
b
Let us denote the Fourier transform −∞ f (x)e−jyx dx by f(y). We there-
fore have Z ∞ Z ∞
b
f (t)φ(t)dt = fb(t)φ(t)dt.
−∞ −∞
This motivates our definition of the Fourier transform qb of a distribution
q:
b
q , φ] ≡ [q, φ],
[b

12
whenever φ is a rapidly decaying function. (It is shown below that whenever
φ is a rapidly decaying function so will φb be.)
The reader will note that by making the functions on which distributions
act very ’well behaved’ we are able to define all the above operations on dis-
tributions. In particular because rapidly decaying functions are differentiable
to any order, distributions also become ’differentiable’ to any order. Because
rapidly decaying functions are Fourier transformable, tempered distributions
become ’Fourier transformable’.

4.8 Some properties of rapidly decaying functions:

If φ is rapidly decaying then the following holds
1. xk φ is rapidly decaying.
2. φ is bounded (since φ is differentiable in (-∞,∞) and lim φ(x) = 0).
|x|→∞
Z ∞
3. |φ(x)|dx exists.
−∞
Z ∞ Z ∞
1+x2
Proof: | φ(x) |dx = 1+x2
| φ(x) |dx
−∞ −∞
Z ∞
2 1
≤ (maxx∈(−∞,∞) (1 + x )φ(x))| 2
dx|.
−∞ 1 + x
Since the RHS is finite the result follows.
dφ
4. dx
is rapidly decaying.
Proof: We need to only show that lim | xk dφdx
|=0
|x|→∞
Z ∞ Z ∞
k dφ
We have x dx dx = - k xk−1 φ dx + [xk φ]∞
−∞
−∞ −∞

Of the two terms in the RHS the first integral clearly exists,since xk−1
φ is rapidly decaying and the second is zero since φ is rapidly decaying.
Z ∞
Thus xk dφ
dx
dx exists for all k. But this can only happen if lim
−∞ |x|→∞
| xk dφ
dx
| = 0.
b
5. If φ is a rapidly decaying function so is φ.
Proof: We need to show
 
 dp φb 
 
lim xk p  −→ 0
x→∞  dx 

13
 for every k and p.
Consider  p Z  
d ∞ 
 e −jxt
φ(t)dt xk  .
 dxp
−∞

This expression is equal to

Z ∞  
 
 p −jxt
(−jt) e φ(t)dt xk 

−∞
Z ∞ 
 
=  p
(−jt) (−jx) e k −jxt
φ(t)dt
Z−∞ 
 ∞ d k 
=  p
(−jt) φ(t) k e −jxt
dt
−∞ dt

Integrating by parts this becomes

 Z ∞ 
 d k 
= (−1) k
e−jxt
[(−jt) p
φ(t)]dt

−∞ dtk
Z ∞ k 
 d p 
≤  k
(t φ(t))dt 

−∞ dt

dk
Since φ is rapidly decaying so is tp φ(t) and therefore so is dtk
(tp φ(t)).
So the integral exists for all p.
b
dp φ
This means | xk dxp
| is bounded for all x and all k but this can clearly
pb
happen only if for each k, lim | xk ddxφp | = 0, which proves that φb is a
x→∞
rapidly decaying function.

4.9 Duality for rapidly decaying functions and tempered dis-

tributions
Suppose f (·) is continuous in (−∞, ∞) and its Fourier transform exists. Let
Z R
1
fR (x) = fb(ω)ejωx dω,
2π −R

where fb(ω) is the Fourier transform of f (x). We will show that

lim fR (x) = f (x),

R→∞

bb
or equivalently f(x) = 2π(f (−x)).

14
 We have
Z R Z ∞ 
1 jωx −jωt
fR (x) = e e f (t)dt dω.
2π −R −∞
Z " R #
∞
1 ejω(x−t)
= f (t)dt.
2π −∞ j(x − t) −R
Z ∞
1 2 sin R(t − x)
= f (t)dt.
2π −∞ t−x
Z
1 ∞ sin Rz
= f (x + z)dz.
π −∞ z

sin Rz
We have already seen that πz
is a δ−sequence with R → ∞. Hence
Z ∞
lim fR (x) = δ(z)f (x + z)dz = f (x).
R→∞ −∞
R∞ sin Rz
[Note that RHS involving δ is just short form for limR→∞ −∞ πz
f (x+z)dz].
In particular we see that if φ is a rapidly decaying function
Z ∞
1 b jωx
φ(ω)e dx = φ(x).
2π −∞

An easy consequence of the duality for rapidly decaying functions is the

duality for tempered distributions,i.e., b
qb(x) = 2πq(−x). To see this note that
h i h i  
b b
qb(x), φ(x) = qb, φb = q, φb = [q(x), 2πφ(−x)] = [2πq(−x), φ(x)] .

4.10 Fourier Transform of derivatives

d
Let f, g be absolutely integrable in (−∞, ∞). Suppose dx f (x) = g(x). We
know that b b
g (x) = jxf (x). We now show that this relation is valid even if
d b
f, g are tempered distributions. We first remind the reader that dx φ(x) =
b
−jxφ(x), when φ(x) is rapidly decaying and
h i h therefore Fourier
i h transformable.i
We have [b b d b d b
g (x), φ(x)] = g(x), φ(x) , i.e., dx f (x), φ(x) = f (x), dx φ(x)
h i h i h i
d
= − f (x), −jxφ(x) = fb(x), jxφ(x) = jxfb(x), φ(x) , as required.
Similarly, or by invoking duality, we can prove that d b
f (x) \(x),
= −jxf
dx
for any tempered distribution f.

15
4.11 Convergent sequences of distributions:
For distributions we say lim qn = q, iff for each φ lim [qn ,φ] = [q, φ].
n→∞ n→∞
Suppose a sequence { qn } has the property that lim [qn ,φ] exists for
n→∞
each φ then we can define a functional q on the space of φ by

[q, φ] ≡ lim [qn , φ].

n→∞

The linearity of q is clear. Continuity involves showing that lim [q, φn ] =

n→∞
0, whenever φn is a null sequence. We will skip this sophisticated proof. For
practical purposes, the most important such convergent sequences are those
that arise from locally integrable functions. For the specific cases of interest,
we will give alternative proofs of the fact that the limit is a distribution.
Let fn converge to f uniformly over every finite interval. We then have

lim [fn , φ] = [f, φ]

n→∞

for every test function. We prove this claim as follows:-

Z ∞ Z b
lim fn φdx = lim fn φdx
n→∞ −∞ n→∞ a

(where [a, b] contains the support of φ).

Since fn converge uniformly to f over [a, b], for each ǫ we can choose Nǫ
such that
|f (x) − fn (x)| ≤ ǫ for x ∈ [a, b], whenever n ≥ Nǫ .
Hence
Z b Z b
| (f − fn )φdx| ≤ |f − fn |φdx
a a
Z b
≤ |(maxa≤x≤b φ(x))| |f − fn |dx
a
≤ |(maxa≤x≤b φ(x))ǫ(b − a)|, n ≥ Nǫ .

The claim follows.

16
5 Some special tempered distributions
5.1 Unit step function
A function that is very commonly encountered in studying the solution of
differential equations is the unit step function
1(x) ≡ 0 , x < 0
≡ 1 , x ≥ 0.

This function is locally integrable. Clearly, the action of 1(·) on a rapidly

decaying function φ is given by
Z ∞ Z ∞
[1(x), φ] ≡ 1(x)φ(x)dx = φ(x)dx.
−∞ 0
We have already shown that this integral exists for a rapidly decaying func-
tion φ. Thus 1(·) is a tempered distribution. In this distributional sense,
 
d
1(x), φ(x) = −[1(x), φ̇(x)]
dx
Z ∞
= − φ̇(x)dx
0
= [φ(x)]0∞ = φ(0).

Thus the distributional derivative of 1(x) is δ(x).

We will call a sequence {rn } of functions a 1-sequence if
(a) rn (·) is locally
R ∞ integrable R∞
(b) limn→∞ −∞ rn (x)φ(x)dx = 0 φ(x)dx for every rapidly decaying
function φ(x).
Suppose the functions rn are all differentiable over (−∞, ∞). We then have
Z ∞ Z ∞
lim r˙n (x)φ(x)dx = − rn (x)φ̇(x)dx
n→∞ −∞ −∞
Z ∞
= φ̇(x)dx
0
= φ(0).

It follows therefore that

{r˙n (x)} is a δ-sequence.
d
This gives another interpretation to the expression dx
1(x) = δ(x).

17
 1
5.2 The tempered distribution x
In signal processing applications, the tempered distribution x1 plays an im-
portant role. It is related to the unit step function through the Fourier
transform. Convolution of a rapidly decaying function by this distribution is
called the Hilbert transform and is of use in studying some kinds of modula-
tion of signals.
The tempered distribution x1 is defined by, for φ ∈ S1 ,
  Z −ǫ Z ∞
1 φ(x) φ(x)
, φ ≡ lim dx + dx.
x ǫ→0,ǫ>0 −∞ x ǫ x

Since φ is continuous at 0, the limit in the above expression exists. To see

this, define ψ(x) ≡ φ(x) − φ(−x). The above expression on the RHS reduces
to Z ∞
ψ(x)
lim dx.
ǫ→0,ǫ>0 ǫ x
Observe that ψ(0) = 0 and further, since φ is rapidly decaying, we need only
examine the convergence of the above integral for some positive b in place of
∞.
We therefore need to show that the limit exists in the following expression.
Z b
ψ(x)
limǫ→0,ǫ>0 dx.
ǫ x

In the interval [ǫ, b], we can bound ψ(x) between ψ(0) + k1 x and ψ(0) + k2 x
Rb
for some k1 , k2 . Noting that ψ(0) = 0, we see that the integral ǫ ψ(x)
x
dx lies
between k1 (b − ǫ) and k2 (b − ǫ) and therefore, as ǫ tends to zero, the integral
R b ψ(x)
ǫ x
dx converges.
 
Further, x1 , φ is linear
 in φ and by using the above argument, it can be
1
seen that limn→∞ x , ψn = 0, whenever {ψn } is a null sequence of rapidly 
decaying functions (in the limit ψ(0), k1 and k2 go to zero). Thus x1 , φ is
continuous in φ. It follows that x1 is a tempered distribution.

1
5.3 Fourier transform of 1(x) and x
In what follows we use 1 for the constant function on the real line which
takes value 1 on all real numbers. By 1(·) or 1(x) , we mean the unit step
function which takes value 0 for x < 0 and 1 for x > 1. h i h i
b φ = δ, φb =
Let us first compute the Fourier transform of δ. We have δ,
b = [1, φ] We therefore conclude δb = 1.
φ(0)

18
 We have d d
dx
1(x) = δb = 1. But in Section4.10 we saw that dx cd
q(x) = jxb
q (x),
b
for any tempered distribution q. So jx1(x) = 1. We will show that this means
b
1(x) = jx1
+ cδ(x) for some constant c and later show that the constant must
be π.
1
First observe that the tempered distribution jx satisfies jxq(x) = 1. We
have
  Z −ǫ Z ∞ 
1 jx jx
, jxφ = lim φdx + φdx
jx ǫ→0,ǫ>0 −∞ jx ǫ jx
Z ∞ Z ∞
= φ(x)dx = 1φ(x)dx.
−∞ −∞

Consider the equation jxq(x) = 0. ( b 1(x) − jx1

satisfies this equation.)
We have [jxq(x), φ(x)] = 0 i.e., [q(x), jxφ(x)] = 0.
We first show that a test function ψ(x) has the form xφ(x) iff ψ(0) = 0.
Clearly xφ(x) at x = 0 has value 0. Suppose ψ(x) = 0. Consider ψ(x) x
. Define

ψ(x)
φ(x) = , x 6= 0
x
ψ(x)
= ψ̇(0) = lim , x = 0.
x→0 x

It is clear that φ(x) is a test function and ψ(x) = xφ(x).

We will next show that q = cδ for some constant c.
We have [xq(x), φ(x)] = 0, i.e., [q(x), xφ(x)] = 0.
Write φ(x) as φ(0)φ0 (x) + [φ(x) −φ(0)φ0(x)], where φ0 (x) is any test function
with φ0 (x) = 1.
Now, [q, φ] = [q(x), φ(0)φ0 (x)] + [q(x), (φ(x) − φ(0)φ0 (x))].
But φ(x) − φ(0)φ0(x) = 0 at x = 0 since φ0 (0) = 1.
Hence φ(x) − φ(0)φ0 (x) has the form xρ(x) for some test function ρ(x).
Hence [q(x), (φ(x) − φ(0)φ0(x))] = [q(x), xρ(x)] = 0.
Hence [q, φ] = [q(x), φ(0)φ0(x)] = φ(0)[q(x), φ0 (x)].
So [q, φ] = φ(0)c, where c = [q(x), φ0 (x)]. So q = cδ.
It is thus clear that b 1
1(x) = jx + cδ(x).
We next evaluate the constant c. Consider

1(x) + 1(−x) = 1.

By duality we know that b 1 = 2πδ, i.e., (1(x)\+ 1(−x)) = 2πδ,

i.e., b
1(x) + b \
1(−x) = 2πδ(x). But q(−x) = qb(−x). So b 1(x) + b
1(−x) = jx1
+
1
−jx
+ cδ(x) + cδ(−x) = 2cδ(x). It follows that c = π. Thus b 1
1(x) = jx + πδ(x).

19
 b1 + πδ
By duality, we must have jx c = 2π1(−x) = 2π−2π1(x). Since πδ c = π,
b1 = π − 2π1(x), i.e., b1 = −jπsgn(x), where sgn(x) is −1 for
it follows that jx x
negative x and 1 for positive x.

6 Multiplication Rule for Distributional Deriva-

tive
As we saw before, under certain conditions, f q would be defined when f is
a function and q is a distribution. For instance, if f has derivatives of all
orders f q is always defined.

[f q, φ] ≡ [q, f φ].

In special cases, the harsh conditions on f can be relaxed. When q is say

the δ-function, f q is defined if f is continuous at the origin. Let us examine
d df dq
if the usual multiplication rule dx (f q) = q dx + f dx works in the case of
distributions when f is a function whose derivatives of all orders exist, and
q, a distribution. We have
     
d dφ dφ
(f q), φ ≡ − f q, = − q, f
dx dx dx
 
d(f φ) df
= − q, −φ
dx dx
   
d(f φ) df
= − q, + q, φ
dx dx
   
dq df
= , f φ + q, φ
dx dx
   
dq df
= f ,φ + q ,φ
dx dx

Therefore it follows that

d(f q) dq df
=f +q
dx dx dx
as is the case with functions.
Example:
Consider the function in Figure 2 below

20
 Figure 2:

g(x) = x(1(x − 1) − 1(x − 2))

dg
= 1(1(x − 1) − 1(x − 2)) + x(δ(x − 1) − δ(x − 2))
dx
= 1(x − 1) − 1(x − 2) + δ(x − 1) − 2δ(x − 2).

(Observe that [f (x)δ(x − a), φ(x)] = [δ(x − a), f (x)φ(x)]

= [δ(x), f (x − a)φ(x − a)] = [f (x − a)δ(x), φ(x − a)]
= [f (a)δ(x), φ(x − a)] = [f (a)δ(x − a), φ(x)].)

Figure 3:

Consider the function f in Figure 3. This function is piecewise continuous

being made up of polynomials in the open intervals (0, 1), (1, 2), (2, 3), (3, 4).
The function has derivatives of all orders (of course the third derivative
is the zero function). However, from the first or second derivative functions

21
(g = f˙, h = f),
¨ we cannot recapture f (t) back again if we work with ordinary
functions and use the ordinary notion of derivatives. On the other hand, if
f (·) is treated as a distribution and the derivatives are taken in the distri-
butional sense, f (·) can be fully recovered. Let us denote the distributional
derivative of f by dfdt
.

df d 2 
= t (1(t) − 1(t − 1)) + t(1(t − 1) − 1(t − 2))
dt dt
+ [2(1(t − 2) − 1(t − 3)) − (t − 4)(1(t − 3) − 1(t − 4))]
= 2t(1(t) − 1(t − 1)) + 1(1(t − 1) − 1(t − 2))
+0(1(t − 2) − 1(t − 3)) − 1(1(t − 3) − 1(t − 4))
−δ(t − 1) + δ(t − 1) − 2δ(t − 2) + 2δ(t − 2) − 2δ(t − 3) + δ(t − 3)
= g(t) − δ(t − 3)
2
df dg
similarly, = − δ̇(t − 3)
dt2 dt
= h(t) − δ(t − 1) − δ(t − 2) − δ(t − 3) + δ(t − 4) − δ̇(t − 3)

Observe that from the distributional derivative of any order of the function,
we can recover the original function - the δs and the δ̇s do the bookkeeping
for us.

7 Convolution of Distributions
7.1 q1 ∗ q2 when q2 has finite support
In order to define the notion of convolution for distributions we first examine
the concept in the case of absolutely integrable functions. Let f1 , f2 be
absolutely integrable, i.e.,
Z ∞
|fi (x)| dx, i = 1, 2 ,
−∞

be finite. We define the convolution f1 ∗ f2 by

Z ∞
f1 ∗ f2 (y) ≡ f1 (x)f2 (y − x) dx
−∞
Z ∞
= f1 (y − x)f2 (x) dx
−∞
= f2 ∗ f1 (y)

22
 R∞
We see that −∞
|f1 ∗ f2 (y)| dy exists, since
Z ∞ Z ∞ Z ∞ 
 
|f1 ∗ f2 (y)| dy =  f (x)f (y − x) dx  dy
 1 2 
−∞
Z−∞
∞
−∞
 Z ∞ 
≤ |f1 (x)| |f2 (y − x)| dy dx
−∞ −∞
Z ∞  Z ∞ 
≤ |f1 (x)| dx |f2 (z)| dz , taking z = (y − x).
−∞ −∞

Let us examine the distribution qf1 ∗f2 defined by

Z ∞
[qf1 ∗f2 , φ] ≡ f1 ∗ f2 (y)φ(y) dy.
−∞

We can write the above integral as

Z ∞ Z ∞  Z ∞ Z ∞ 
f1 (x)f2 (y − x) dy φ(y) dy = f1 (x) f2 (y − x)φ(y) dy dx.
−∞ −∞ −∞ −∞

This has the form

Z ∞
f1 (x)ψ(x) dx,
−∞
Z ∞
where ψ(x) ≡ f2 (y − x)φ(y) dy
Z−∞
∞
= f2 (z)φ(x + z) dz (taking z = y − x).
−∞

In order to make this appear as the action of a distribution on a test function

we would require ψ(x) to be a test function. It is clear that it is infinitely
R∞
differentiable since dψ
dx
= f (z) d(φ(x+z))
−∞ 2 dx
dz and φ(x + z) is infinitely dif-
ferentiable. But if f2 doesn’t have finite support then ψ(x) can not have
finite support. The above discussion brings out the difficulties in defining
convolution for distributions and suggests we attempt to q1 ∗ q2 only when
q2 has finite support or when q1 , q2 have some other special properties.
When q1 , q2 are distributions and q2 has finite support [a, b] ( ie [q2 , φ] = 0
whenever support of φ doesn’t intersect [a, b]), we define q1 ∗ q2 as follows.

[q1 ∗ q2 (x), φ(x)] = [q1 (x), [q2 (z), φ(x + z)]] .

For this notion to be well defined we need to verify that ψ(x) = [q2 (z), φ(x +

23
z)] is a test function. First we observe that

dψ d[q2 (z), φ(x + z)]

=
dx dx
[[q2 (z), φ(x + z + △x)] − [q2 (z), φ(x + z)]]
= lim
△x→0 △x
 
φ(x + z + △x) − φ(x + z)
= lim q2 (z),
△x→0 △x
  
φ(x + z + △x) − φ(x + z)
= q2 (z), lim
△x→0 △x
(by the continuity of q2 )
 
dφ(x + z)
= q2 (z),
dx

Since φ is infinitely differentiable it will follow that ψ will also be. Next let q2
have support [−T1 , T2 ] and let φ have support [−T3 , T4 ]; T1 , T2 , T3 , T4 being
positive. Suppose x ∈ / [−T1 − T4 , T2 + T3 ]. We have ψ(x) = [q2 (z), φ(x + z)].
Now φ(x + z) has support (in terms of variable z) [−T3 + x, T4 + x]. We
must have [−T3 + x, T4 + x] ∩ [−T1 , T2 ] 6= ∅, in order that ψ(x + z) is nonzero.
If
x > T2 + T3
or if
x < −T1 − T4
the above intersection is null, i.e., the support of φ(x + z) and q2 (z) do not
intersect so that [q2 (z), φ(x + z)] = 0. Thus ψ(x) has finite support and is
infinitely differentiable and is therefore a test function. Hence,

[q1 (x), [q2 (z), φ(x + z)]]

is well defined.
Linearity of q1 ∗ q2 is clear, since,

[q1 ∗ q2 (x), (αφ1 + βφ2 )(x)] = [q1 (x), [q2 (z), (αφ1 + βφ2 )(x + z)]]
= [q1 (x), α[q2 (z), φ1 (x + z)] + β[q2 (z), φ2 (x + z)]]
= α[q1 ∗ q2 (x), φ1 (x)] + β[q1 ∗ q2 (x), φ2 (x)].

Continuity of q1 ∗ q2 can be shown as follows. Let φn be a null sequence of

test functions.

lim [q1 ∗ q2 (x), φn (x)] = lim [q1 (x), [q2 (z), φn (x + z)].
n→∞ n→∞

24
Since q2 is continuous and has finite support, it is clear that ψn (x) ≡ [q2 (z), φn (x+
z)] is a null sequence of test functions so that the limit is zero as required for
continuity.
We next examine the derivative of the convolution of distributions. We
will show that
d(q1 ∗ q2 ) dq2
= q1 ∗ .
dx dx
We have
 
d(q1 ∗ q2 )
, φ(x) = −[q1 ∗ q2 , φ̇(x)]
dx
to avoid notational confusion we denote φ̇(x) by ψ(x).
= −[q1 (x), [q2 (z), ψ(x + z)]]
  
dq2
= − q1 (x), − , φ(x + z)
dz
  
dq2
= q1 (x), , φ(x + z)
dz
dq2
Let us denote the distribution dz
by q3 (z). It is defined by

[q3 (z), φ(z)] = −[q2 (z), φ̇(z)].

Thus,
 
d(q1 ∗ q2 )
, φ(x) = [q1 (x), [q3 (z), φ(x + z)]]
dx
= [q1 ∗ q3 (x), φ(x)].
1 ∗q2 )
Hence, d(qdx = q1 ∗ dq
dx
2
.
The way convolution has been defined for distributions does not make the
operation naturally commutative. In special cases, however, this would be
true. For instance, if q1 is δ or its derivative of some order and q2 is of finite
support or regular, it can verified that the operation is indeed commutative.
So, φ ∗ δ = δ ∗ φ = φ.

7.2 Convolution of a distribution with φ ∈ S1 .

The special case of a distribution with a rapidly decaying function is of
importance in signal processing- the Hilbert transform, for instance, is of
this kind.
We will show that q ∗ φ is a tempered distribution when φ ∈ S1 and q is
tempered. We need the following preliminary lemma.

25
 Lemma: Let φ, ψ ∈ S1 . Then φ ∗ ψ ∈ S1 .
Proof: We have φ, b ψb ∈ S1 and therefore by the definition of rapidly
b ψb ∈ S1 . But we know that φ[
decaying functions φ. b ψb and φ ∗ ψ is
∗ ψ = φ.
continuous in (−∞, ∞).
By duality we know that,
d
b ψ)(x)
b
(φ. = 2πφ ∗ ψ(−x)
Thus φ ∗ ψ(−x) ∈ S1 and therefore φ ∗ ψ(x) ∈ S1 .
QED

Theorem: If q is a tempered distribution and φ ∈ S1 , q ∗ φ is a tempered

distribution.
Proof: We have
[q ∗ φ, ψ] ≡ [q(x), [φ(z), ψ(x + z)]]
when ψ ∈ S1 . We have,
[φ(z), ψ(x + z)] = [φ(y − x), ψ(y)]
= [φ̃(x − y), ψ(y)]
= ψ ∗ φ̃(x), denoting φ(−t) by φ̃(t)
Since ψ, φ̃ ∈ S1 , ψ ∗ φ̃(x) ∈ S1 . Hence, [q(x), ψ ∗ φ̃(x)] and therefore
[q(x), [φ(z), ψ(x + z)]] is well defined. Hence, [q ∗ φ, ψ] is well defined for
all ψ ∈ S1 . Linearity of q ∗ φ over S1 is clear. We need to verify continuity,
i.e., that,
lim [q ∗ φ, ψn ] = 0
n→∞
whenever ψn is a null sequence in S1 .
Clearly this would follow if {ψn ∗ φ̃} is a null sequence in S1 , whenever
φ ∈ S1 and {ψn } is a null sequence in S1 . We need to verify that,
 
 dl (ψ ∗ φ̃(x) 
 n 
lim max xk =0
n→∞ −∞<x<∞  dxl 

for every pair of integers k, l.

Now,
dl (ψn ∗ φ̃)(x) dl φ̃(x)
xk = ψn ∗
dxl dxl
dl φ̃(x)
Since dxl
∈ S1 whenever φ ∈ S1 , it is adequate to verify that,
 
lim max xk (ψn ∗ φ(x)) = 0
n→∞ −∞<x<∞

26
for every φ ∈ S1 .
We have,
Z 
 k   ∞ 
x (ψn ∗ φ(x)) =  |x ψn (x − τ ) ∗ φ(τ ) dτ 
k

−∞
Z ∞
≤ Mn φ(τ ) dτ,
−∞
 
where Mn = max xk ψn (x − τ )
−∞<τ <∞

Since limn→∞ Mn = 0, we have

 
lim max xk (ψn ∗ φ(x)) = 0.
n→∞ −∞<x<∞

Thus, ψn ∗ φ̃(x) is a null sequence whenever ψn ∈ S1 and φ ∈ S1 and the

continuity of q ∗ φ on S1 follows. Thus q ∗ φ is a tempered distribution as
required.
QED

7.3 Convolution of a distribution with δ and its derivatives

Let us first examine the convolution of δ with φ ∈ S1 . We have
[δ ∗ φ(x), ψ(x)] = [δ(x), [φ(z), ψ(x + z)] = [φ(z), ψ(z)]. So, δ ∗ φ = φ.
On the other hand, [φ ∗ δ(x), ψ(x)] = [φ(x), [δ(z), ψ(x + z)]] = [φ(x), ψ(x)].
So, φ ∗ δ = δ ∗ φ = φ.
Next, dφdx
= d(φ∗δ)
dx
= φ ∗ dxdδ
.
dδ
On the other hand, consider the distribution dx ∗ φ. We have,
dδ dδ dδ
[ dx ∗ φ, ψ] = [ dx , [φ(z), ψ(x + z)]] = [ dx , φ̃ ∗ ψ(x)], where φ̃(x) ≡ φ(−x). Now
˜ ∗ ψ] = [δ ∗ dφ , ψ].
[ dδ , φ̃ ∗ ψ(x)] = −[δ, d(φ̃∗ψ(x)) ] = [δ, dφ
dx dx dx dx
Thus, dxdδ
∗ φ = φ ∗ dxdδ
= dφ
dx
.
By induction it will follow that the convolution of the k th derivative of δ
with φ ∈ S1 will yield the k th derivative of φ and the order of convolution is
immaterial.
The convolution of a distribution q with δ yields the same distribution.
We have, [q ∗ q2 , φ] ≡ [q(x), [q2 (x), φ(x + z)]].
So, [q ∗ δ, φ] = [q(x), [δ(z), φ(x + z)]] = [q(x), φ(x)].
1 ∗q2 )
Next we have d(qdx = q1 ∗ dq
dx
2
, when q2 has finite support. Hence if q2 is the
k derivative of δ, q ∗ q2 would be the k th derivative of q ∗ δ = q.
th

27
7.4 Fourier transform of convolution of distributions
We define Fourier transform of convolution of distributions only in the special
case where one of them is tempered and other a rapidly decaying function.
Let q be tempered and let φ ∈ S1 . We have seen that q ∗ φ is tempered and
therefore has a Fourier transform. We verify below that qb b (We
∗φ = qb.φ.
have denoted φ(−x) by φ̃.)
h i
∗φ), ψ] ≡ q ∗ φ, ψ
[(qb b
h h ii
= q(x), φ(z), ψ(zb + x)
h   i
= q(x), φ̃ ∗ ψb (x)
" \  #
1 b
= q(x), 2π φ.ψ (x)
2π
h   i
b
= qb(x), φ.ψ (x)
h i
b
= qb(x), φ(x).ψ(x)
h i
b
= qb(x).φ(x), ψ(x) .

By using duality, or directly, we can verify, when q is tempered and φ is

rapidly decaying, that
c = 1 qb ∗ φ.
q.φ b
2π
We remind the reader that the tempered distribution x1 is defined by,for
φ ∈ S1 ,   Z −ǫ Z ∞
1 φ(x) φ(x)
, φ ≡ lim dx + dx.
x ǫ→0,ǫ>0 −∞ x ǫ x
b
We have shown earlier that x1 = −jπsgn(x).
The distribution x1 ∗ φ where φ ∈ S1 is tempered and has the Fourier
b
transform −jπsgn(x)φ(x).
1
In signal processing literature πt ∗ φ(t) is called the Hilbert transform of
φ(t) and has important applications in the study of modulation. We give an
illustration below.

7.5 An application of the Hilbert transform

In the discussion to follow, we follow the convention of signal processing
literature and take signals to be lower case functions of ‘t′ and their Fourier

28
transform to be the corresponding upper case functions of ‘jω ′ . Thus b
a(t) is
denoted A(jω).
Consider an amplitude modulated wave a(t) cos(ω0 t) = α(t). We assume
that A(jω) is zero for |ω| ≥ ω0 and is infinitely differentiable. As a conse-
quence, we have that a(t) and α(t) are rapidly decaying. Given α(t), the
problem is to recover the signal a(t). For simplicity let us take a(t) to be
real. We have
 
δ(jω − jω0 ) + δ(jω + jω0 ) 1
b(jω) = A(jω)∗
α = (A(jω−jω0 )+A(jω+jω0)).
2 2

Hence,

A(jω − jω0 ) = 2b
α(jω).1(jω)

and therefore

α(jω).1(jω)) ∗ δ(jω + jω0 )] .

A(jω) = [(2b

We then have
  
−1 j
a(t) = F α(jω).1(jω)) ∗ δ(jω + jω0 )] = (α(t) ∗
[(2b + δ(t) e−jω0 t.
πt
 
= α(t) + jαH (t) e−jω0 t ,

where
1
αH (t) =
∗ α(t).
πt
Since we assumed a(t) to be real, we must have

a(t) = α(t) cos(ω0 t) + αH (t) sin(ω0 t).

αH (t) is the Hilbert transform of α(t).

1
As we noted before α(t) is a rapidly decaying function. πt is a tempered
1
distribution. πt ∗ α(t) is therefore a tempered distribution. In the present
case it turns out to be a regular distribution.
Z ∞
1 1
∗ α(t) = α(t − τ )dτ.
πt −∞ πτ
Z −ǫ Z ∞ 
1 1
lim α(t − τ )dτ + α(t − τ )dτ .
ǫ→0 −∞ πt ǫ πt

29
8 Summary of properties of Fourier Trans-
forms for tempered distributions
For notational convenience, we use the convention of signal processing liter-
ature: a tempered distribution is written as q(t) and its Fourier Transform
is written as qb(jω).
1. Linearity

αq\
1 + βq2 = αqb1 + β qb2 .

2. Time shifting:

\
q(t − to ) = e−jωto qb(jω).

3. Frequency shifting:

\
ejω \
0 t q(t) = q(jω − jω0 ).

4. Time scaling:

\
[ = 1 q( jω ).
q(at)
|a| |a|
5. Time differentiation:

c
dq
= jωb
q(jω).
dt
6. Frequency Differentiation:

d = j db
tq(t)
q
.
dω
7. Convolution:

q\
∗ f (t) = qb.fb,
where q is tempered and f rapidly decaying.
8. Multiplication:

\ 1
q.f (t) = qb ∗ fb(jω),
2π
where q is tempered and f rapidly decaying.

30
 9. Duality:

b
qb = 2π q̃,
˜ ≡ q(−t).
where q(t)
b and proceed by
q , φ] ≡ [q, φ]
The proofs are routine with the starting point [b
using the corresponding property for Fourier transformable functions.

9 Periodic Distributions
In the discussion to follow, we follow the convention of signal processing
literature and take signals to be lower case functions of ‘t′ and their Fourier
transform to be the corresponding upper case functions of ‘jω ′ . Thus b a(t) is
denoted A(jω).
A distribution q is said to be periodic with period T iff

[q, φ] = [q, φT ].

In signal processing applications it is quite common to encounter the situation

described in Figure 4 over one period.
We wish to show that we can obtain the Fourier series expansion (a
distributional equation that has the same form as the usual Fourier series
expansion) of this generalized function by the usual process valid for the
functions satisfying Dirichlet conditions:
∞
X
q(t) = cn ejω0 nt
n=−∞

where RT
0
q(t)e−jω0 nt dt
cn = .
< ejω0 nt , ejω0 nt >
The integral on the numerator is to be interpreted appropriately for the
singular distributions which are the constituent parts of q(t). The technique
that we describe is valid for periodic distributions whose ‘average value’ over
one period is zero. Essentially this means that q(t) must be composed of
regular distributions, δ functions and derivatives. In this case the action
[q, 1] would be defined over one period and this can be subtracted out before
we seek an expansion.

31
 Figure 4:

While we usually treat the above construct as a function it is clearly a

distribution q which is the sum of f, δ(t − 3T
4
) and 2δ̇(t − T4 ). This may be
regarded as periodic with period T in which case we would be working with
∞
X X∞
3T . T
fp + δ(t − − nT ) + 2 δ (t − − nT ).
n=−∞
4 n=−∞
4

fp being a periodic function agreeing with f over the period [0, T ].

Its action on a rapidly decaying function φ is given by
Z ∞ X∞ X∞
3T . T
[q, φ] = f (t)φ(t)dt + φ( + nT ) − 2 φ ( + nT ).
−∞ n=−∞
4 n=−∞
4

Let us examine whether we can obtain a Fourier series expansion of this

periodic distribution. By such an expansion we mean that we should be able
to write the following distributional equation
∞
X 2π
q= cn ejω0 nt + c0 , ω0 = .
n=−∞n6=0
T

We will now assume that f in [0, T ] is made up of polynomials over some

subintervals say [0, T1 ], [T1 , T2 ]...[Tk−1 , T ].
Clearly there exist functions f2 (t), gδ (t), gδ1 (t) such that
(a.) f2 (t) is made up of polynomials over [0, T1 ], . . . , [Tk−1 , T ],
d2 f2 (t)
(b.) dt2
= f (t) in (0, T1 ), . . . , (Tk−1, T ),
d2 gδ (t) 3T
(c.) dt2
= δ(t − 4
),
d2 gδ1 (t)
(d.) dt2
= +2δ(t − T4 ).

32
 The Fourier
P∞series expansion of m(t) = f2 (t) + gδ (t) + gδ1 (t) over [0, T ] is
jω0 nt
say m(t) = n=−∞ an e . The second derivative of the periodic function
mp (t) would differ from q(t) by a constant.
k
Let p(t) be such that d dtp(t)
k = m(t). then the Fourier series expansion of
p(t) would be
X∞
an
p(t) = k
ejω0 nt + b0 . (**)
n=−∞
(jω 0 n)
n6=0

Since m(t) is made up of polynomials, for k sufficiently large (actually 2)

we can take the series to be absolutely convergent. It would follow that in
the equation (**) the series on the right converges to p(t) uniformly. Hence
(**) is an equation valid distributionally. Differentiating term by term, we
get,
X∞
dk+2 p(t) an (jω0 n)k+2 jω0 nt
= e (*)
dtk+2 n=−∞
(jω 0 n) k

n6=0

to be a valid distributional equation. But the left side would be the distribu-
tion q (within a constant) that we began with. Thus ** is a distributionally
valid Fourier expansion for q minus a constant.
We will now show that the coefficients could have directly been obtained
as RT
< q(t), ejω0 nt > q(t)e−jω0 nt dt
cn = = 0
T T
as though q(t) is an ordinary function satisfying Dirichlet conditions. In the
present case this term would be
Z T Z T Z T
−jω0 nt 3T −jω0nt . T
f (t)e dt + δ(t − )e dt + 2 δ (t − )e−jω0 nt dt
0 0 4 0 4
where we interpret the second and third terms above as
Z ∞ Z ∞
3T −jω0 nt . T −jω nt − 3 (jω nT ) − 1 (jω nT )
δ(t− )e dt+2 δ (t− )e 0 dt = e 4 0 +2(jω0 n)e 4 0
−∞ 4 −∞ 4

Consider the Fourier series expansion for m(t)

∞
X
m(t) = an ejω0 nt .
n=−∞

33
Hence R
< m(t), ejω0 nt > m(t)e−jω0 nt dt
an = =
< ejω0 nt , ejω0 nt > T
We saw thatPthis was also valid distributionally. The Fourier series ex-
pansion q(t) = ∞ n=−∞ cn e
jω0 nt
could be obtained by taking cn = (jω0 n)2 an .
But this is exactly the same as
Z T 2
2 d
(−1) 2
(e−jω0nt )dt.
0 dt

If m(t) had a second derivative in the ordinary function sense by integrating

by parts we would get (using m(0) = m(T ), ????) the above to be equal to
R T d2 m(t) −jω nt
0 dt2
e 0 dt.
2 RT
When d dtm(t)
2 = q, the meaning of 0 q(t)e−jω0nt dt would be the same as
RT 2
(−1)2 0 dtd 2 (e−jω0 nt )dt
The above discussion may be summarized as:
If a periodic distribution q is composed of polynomials, delta functions
and its derivatives, the Fourier series coefficient can be obtained directly as

[q, e−jω0 nt ]
cn =
T
just as though q(t) is a regular function and the integral is over [0, T ].

Let us consider an important special case of periodic distributions.

∞
X
q(t) = δ(t − nT ).
n=−∞

In the interval (0− , T− ) we have thePsingle delta function δ(t). The Fourier
series expansion is therefore q(t) = ∞ n=−∞ cn e
jω0 nt
where

[q(t), e−jω0nt ]
cn =
< ejω0 nt , ejω0nt >
R T−
0−
δ(t)e−jω0 nt dt
=
T
1
=
T
1
P∞
Thus q(t) = T n=−∞ ejω0 nt .

34
 Distributionally
∞
X ∞
1 X jω0 nt
δ(t − nT ) = e
n=−∞
T n=−∞

Consider the infinite series of distributions on the right. We know that if

∞
X ∞
X
qi = q then q̂ = q̂i .
i=−∞ i=−∞

The Fourier transform of ejω0 nt is 2πδ(ω − ω0 ) (by duality). Hence

∞
2π X
q̂ = δ(ω − ω0 ).
T i=−∞
P∞
Thus the Fourier transform
P of the train of impulses n=−∞ δ(t − nT ) is the
train of impulses ω0 ∞n=−∞ δ(ω − ω0 )

10 Infinite series of tempered distributions

Our primary interest in infinite series of distributions is in dealing with
Fourier transforms of such series. Therefore we confine ourselves only to
tempered distributions.
P Pn
We say ∞ i=1 qi = q iff the sequence of partial sums sn = i=1 qi converges
to q,i.e., " n #
X
lim qi , φ = [q, φ], φ ∈ S1 .
n→∞
i=1
P∞
We sayPn i=−∞ qi = q when the same thing happens to the partial sum
sn = i=−n qi P
If a sequence ni=−n fi of locally integrable functions converges to another
such function f uniformly within (−∞, ∞), we know that
" n #
X
lim fi , φ = [f, φ] , φ ∈ S1 .
n→+∞
i=−n

We thus have the distributional equation

∞
X
fi = f.
i=−∞

35
 Distributional equations involving infinite series have a very convenient
property (which is not shared in general by uniformly convergent series of
functions ) viz. the equations remain valid even if we differentiate both sides
term by term.
The proof is easy to see

∞
X
Suppose qi = q
i=−∞
" ∞
# " ∞
#
X dqi X
Then ,φ = − qi , φ̇
i=−∞
dx i=−∞
h i  
dq
− q, φ̇ = ,φ
dx

This fact is of great importance in signal processing. Suppose the periodic

function f is expanded into Fourier series as
∞
X
f (t) = cn ejw0nt
n=−∞

When f satisfies Dirichlet conditions we expect pointwise convergence at

t provided f is continuous at t . Sometimes however, the series on the right
converges uniformly to f in (−∞, ∞). More often the following situation
occurs.
We consider the sequence,
∞
X cn
g(t) = k
ejω0 nt , n 6= 0.
n=−∞
(jω 0 n)
P
For k sufficiently large, the series ∞ |cn |
n=−∞ (ω0 n)k e
jω0 nt
, n 6= 0, would often
P∞ cn
be convergent (i.e. n=−∞ (jω0 n)k , n 6= 0, is absolutely convergent).
P
In such a case, ∞ cn
n=−∞ (jω0 n)k e
jω0 nt
, n 6= 0, converges to g(t) uniformly in
the interval (−∞, ∞). We then have the distributional equation
∞
X cn
g(t) = k
ejω0 nt , n 6= 0.
n=−∞
(jω 0 n)

dk g
We know that the distribution dtk
is then obtained by differentiating the
right side term by term k times.

36
 We therefore have the distributional equation
∞
X
f1 (t) = cn ejω0 nt , n 6= 0,
n=−∞

k
where f1 = ddtkg , even though regarded as functions the RHS does not converge
uniformly over (−∞, ∞) to f1 (t). On the other hand, the series
∞
X
cn ejω0 nt , n 6= 0,
n=−∞

converges pointwise (at points of continuity) in the interval (−∞, ∞) to

f (t) − c0 . The function f (t) − c0 and the distribution f1 (t) are therefore
equal distributionally. Therefore f (t) has the Fourier series expansion
∞
X
f (t) = cn ejω0 nt ,
n=−∞

that is also distributionally valid. Once this happens we can take derivatives
on either side any number of times and get valid distributional equations that
will have the form of a Fourier series expansion. This situation occurs for
instance when 0 < T1 < ..Tm = T and f equals some polynomial in [Ti , Tj ]
and the function is periodic with period T .
These polynomials could be different in different subintervals and f could
be discontinuous at the Ti .

Figure 5:

In particular, consider the function f (t) = t, t ∈ [0, T ]. Let fp be the

periodic function which agrees with f over [0, T ] (see Fig 6).
We can show that

37
 Figure 6:

∞
X 1 1
fp (t) = ejω0 nt + .(∗ ∗ ∗)
n=−∞
−jω0 n 2

We have,
∞
X cn
gp (t) = k
ejω0 nt , − ∞ < t < ∞
n=−∞
(jω0 n)

with the right side converging uniformly to the left side over (−∞, ∞). We
can obtain equation (∗ ∗ ∗) from the latter equation by differentiating term
by term k times and adding the term c0 . Therefore equation (∗ ∗ ∗) is valid
distributionally.
Therefore, we have the distributional equation,
X∞ X∞
dfp
=1− δ(t − nT ) = − ejω0 nt , n 6= 0.
dt n=−∞ n=−∞

Thus ∞ ∞ ∞
X X X
δ(t − nT ) = 1 + ejω0 nt , n 6= 0 = ejω0 nt .
n=−∞ n=−∞ n=−∞

38