Professional Documents
Culture Documents
Multiplication Thoery PDF
Multiplication Thoery PDF
a r t i c l e i n f o a b s t r a c t
Article history: In the theory of signal processing, signals are usually classified either by determining whether their
Available online 26 May 2016 time domain is discrete or continuous, or by determining whether they are periodic. However, no
comprehensive definitions of multiplication and convolution exist that are consistent with the theories
Keywords:
behind all classes, although some important theorems in signal processing involve multiplication and
Theory of signal processing
Multiplication of distributions
convolution. In order to unite the theories behind these classifications, we will consider tempered
Convolution of distributions distributions. In this paper, we propose an approach to the multiplication and convolution of distributions
Tempered distributions that is appropriate to signal processing theory, and prove a well-known theorem regarding the impulse
LTI systems response of continuous linear time-invariant systems of tempered distributions in the context of this new
approach.
© 2016 Elsevier Inc. All rights reserved.
discrete-time periodic signals, and discrete-time Fourier transforms there is in fact no function that satisfies both conditions. Distribu-
for discrete-time non-periodic signals. Corresponding to the the- tion theory provides a mathematically precise approach for dealing
ory of Fourier transforms, there are four different definitions of with δ(t ), such that the distributional Fourier transform gives us
a convolution. However, while the relationships between Fourier the four Fourier transforms mentioned above [1].
transforms and convolutions are similar in each of the four cases, Even if we consider discrete-time signals as distributions, mul-
four separate theories of signal processing are presented in stan- tiplications between continuous-time signals and discrete-time sig-
dard textbooks. nals are not preserved for existing definitions of multiplication for
distributions, because the product δ · δ is not considered to be a
distribution. Moreover, most conventional convolutions of distribu-
E-mail address: akitad@mail.sci.hokudai.ac.jp. tions are only valid for distributions with a specific period, or only
http://dx.doi.org/10.1016/j.dsp.2016.05.008
1051-2004/© 2016 Elsevier Inc. All rights reserved.
2 D. Akita / Digital Signal Processing 56 (2016) 1–14
for distributions that have no period. In this paper, we propose a If T f has a value for any ϕ ∈ S and continuity on S , then T f is a
multiplication of distributions that preserves the multiplication of tempered distribution. Then, we also use the expression f to refer
signals and admits a definition of a convolution of distributions to T f as a distribution.
that does not depend on whether the distributions have a period In fact, there exist some sufficient conditions for a function f
or not. To this end, the theory of distributions and the results of on R to be a tempered distribution T f . The following theorem
previous studies on the multiplication and convolution of distri- states one of these conditions [3].
butions are introduced in Section 2. Then, a multiplication and a
convolution of distributions appropriate to signal processing theory Theorem 1. Let f be a complex-valued, Lebesgue-measurable function
are defined in Sections 3 and 4, respectively. on R. If there exist C > 0 and a non-negative integer K such that
2.1. Conditions for signals to be tempered distributions for almost all t, then T f ∈ S .
Schwartz [2] defined a distribution as a continuous linear func- We let Q M denote the space of functions satisfying the condi-
tional, so that δ is defined as a functional δ : ϕ (t ) → ϕ (0) corre- tions of Theorem 1.
sponding to the heuristic equation Before introducing the condition for a discrete-time signal to be
a tempered distribution, we justify expression (1). As mentioned in
∞
the first part of this subsection, δ is defined as a tempered distri-
δ(t )ϕ (t )dt = ϕ (0). (4) bution, so that for any ϕ ∈ S ,
−∞
δ, ϕ = ϕ (0). (12)
However, we have to consider that the domain of the functional is
a restricted function space. The translation of distributions τh is an operator parametrized with
A rapidly decreasing function is an infinitely differentiable com- h ∈ R, such that for any F ∈ S ,
plex-valued function ϕ (t ) satisfying
τh F , ϕ = F , ϕ (t + h) . (13)
k (n)
sup (1 + |t |) ϕ (t ) < ∞, (5) We may employ the notation δ(t − h) for τh δ . In addition, we de-
t ∈R
n+k≤ N fine the operator reflection of distributions, denoted by ˇ, as
for any non-negative integer N. We express the space of rapidly
decreasing functions as S , and define convergence in S as follows. F̌ , ϕ (t ) = F , ϕ (−t ) . (14)
A sequence of a rapidly decreasing functions {ϕ j } is said to con-
Furthermore, we define the convergence of tempered distributions
verge to ϕ if for any non-negative integer N,
as follows. A sequence of tempered distributions { F N } is said to
lim
(n)
sup (1 + |t |)k ϕ j (t ) − ϕ (n) (t ) = 0. (6) converge to F ∈ S if
j →∞ t ∈R
n+k≤ N ∀ϕ ∈ S lim F N , ϕ = F , ϕ . (15)
N →∞
A tempered distribution F is a continuous linear functional of
rapidly decreasing functions, where the value of F at ϕ is de- In
∞particular, for a sequence Nof tempered distributions { F N },
scribed as F , ϕ . The continuity of a functional F in S is defined k=−∞ F k denotes lim N →∞ k=− N F k . Using this definition, a
as follows. A functional F is continuous if for any sequence {ϕ j } discrete-time signal f is regarded as a tempered distribution with
which converges to 0 in S , sampling time T s > 0 if
lim F , ϕ j = 0. (7) ∞
N
j →∞
f , ϕ = f [k]δ(t − kT s ), ϕ = lim f [k]ϕ (kT s )
N →∞
If necessary, we employ the notation F (t ), ϕ (t ) to express the k=−∞ k=− N
variable of ϕ explicitly. Moreover, we denote the space of a tem- (16)
pered distributions by S . Then, the Fourier transform of a tempered
distribution F is defined as follows: has a value for any ϕ ∈ S , and if f , ϕ j → 0 whenever ϕ j → 0
in S .
F [ F ], ϕ = F , F [ϕ ] , (8) Just as arbitrary continuous-time signals cannot be regarded
where ϕ ∈ S , F ∈ S , and
F [ϕ ] on the right-hand side is the clas- as tempered distributions, not all discrete-time signals are tem-
sical Fourier transform of the rapidly decreasing function pered distributions. However, in contrast to the case of continuous-
time signals, there exists a necessary and sufficient condition for a
∞ discrete-time signal to be regarded as a tempered distribution. To
F [ϕ ](ω) = ϕ (t ) exp(−i ωt )dt . (9) show this, we introduce two theorems from [4].
−∞
Theorem 2. Let f [n] be a discrete-time signal. If there exist real numbers
This Fourier transform is a one-to-one mapping of S onto it-
M and K such that
self [3]. Here, and throughout this article, the Lebesgue integral
is used. | f [n]| ≤ M |n| K (17)
Now, we describe how signals can be regarded as distributions. N
For a continuous-time signal f , consider a functional T f such that for all non-zero n ∈ Z, then for any ω0 > 0, lim N →∞ k=− N f [k] ×
for a given ϕ ∈ S , exp(ikω0 t ) converges to a tempered distribution.
∞
Theorem 3. Let F ∈ S . If there exists P > 0 such that
T f ,ϕ = f (t )ϕ (t )dt . (10)
−∞
τP F = F , (18)
D. Akita / Digital Signal Processing 56 (2016) 1–14 3
√
then there exists a discrete-time signal f [n] such that almost periodic function f (t ) = sin t + sin 2t, despite being non-
∞
periodic, is not usually discussed in textbooks because its Fourier
2π ikt transform as function
F= f [k] exp . (19)
P ∞
k=−∞
Moreover, there exist real numbers M and K such that F [ f ](ω) = f (t ) exp(−i ωt )dt (25)
−∞
| f [n]| ≤ M |n| K (20)
cannot be defined. However, as tempered distribution
√ f (√
t ) has a
for all non-zero n ∈ Z. Fourier transform (δ(t − 1) − δ(t + 1) + δ(t − 2) − δ(t + 2))/2i.
Recognizing signals as tempered distributions enables us to treat
Further, we will employ the following additional theorem mathematically troublesome problems of this sort.
from [3]. To conclude this subsection, we define some sets relating to
distributions. A tempered distribution F is said to vanish on an
Theorem 4. Let { F N } be a sequence of tempered distributions. Then, open set if
{ F N } converges to the tempered distribution F if and only if {F [ F N ]}
converges to F [ F ]. ∀ϕ ∈ S , supp ϕ ⊂ ⇒ F , ϕ = 0, (26)
Corollary 1. Let f [
n] be a discrete-time signal, and let T s be a positive 2.2. Preservation of multiplication and integration of convolution
∞
real number. Then, k=−∞ f [k]δ(t − kT s ) is a tempered distribution if
and only if there exist real numbers M and K such that Let T be a mapping from Q M to S such that
Proof. Let and for T s > 0, let T T s be a mapping from O M to S such that
N
∞
TN = f [k]δ(t − kT s ) (22)
∀ f ∈ OM , ϕ ∈ S TT s [ f ], ϕ = f [k]ϕ (kT s ). (29)
k=− N
k=−∞
and
We let T [Q M ] = {T [ f ]| f ∈ Q M } and T T s [O M ] = {T T s [ f ]| f ∈ O M }.
N Then, continuous-time signals in Q M and discrete-time signals in
S N = F [T N ] = f [k] exp(ikT s t ). (23) OM can be regarded as tempered distributions given by T and T T s
k=− N respectively, according to the subsection 2.1. Moreover, the Fourier
transforms of signals correspond to the Fourier transforms of their
Then, { T N } converges to a tempered distribution if and only if { S N }
representations as tempered distributions.
converges, from Theorem 4. Here, if there exist such M and K ,
However, while both Q M and O M constitute rings, the usual
then Theorem 2 implies that { S N } converges. Furthermore, if { S N }
multiplication of distributions holds only for very restricted dis-
converges to a tempered distribution, then there also must exist
tributions. For example, in many textbooks on distributions, we
such M and K according to Theorem 3, as
can find the following definition of multiplication for a tempered
∞
∞
distribution and a slowly increasing function, where a slowly in-
τ2π /T s f [k] exp(ikT s t ) = f [k] exp(ikT s (t − 2π / T s )) creasing function is an infinitely differentiable function f ∈ Q M . Let
k=−∞ k=−∞ O M denote the space of a slowly increasing function. Then, the
∞
product of F ∈ S and f ∈ O M is defined as
= f [k] exp(ikT s t ). (24)
f F , ϕ = F , f ϕ (30)
k=−∞
Thus, { T N } converges if and only if there exist such numbers M for any ϕ ∈ S , because f ϕ ∈ S . However, this definition of multi-
and K . 2 plication cannot be applied for f , g ∈ T T s [O M ].
Consider defining a multiplication in a subset of S containing
We let O M denote the space of discrete-time signals satisfying T [Q M ] and T T s [OM ] for any T s > 0, such that T and T T s are ring
the conditions in Corollary 1. homomorphisms. Then, we find that δ · δ = δ . To explain this, set
Assuming that signals that are classified into the four classes
1, n = 0
mentioned above can be expressed as tempered distributions in f [n] = (31)
this way, the Fourier transform defined by (8) gives us Fourier 0, else
series, the Fourier transform of a function, the discrete Fourier
transform, and the discrete-time Fourier transform, depending on so that T T s [ f ] = δ . Then, we see that
the classification [1]. Thus, the four classifications of signals and δ = TT s [ f ] = TT s [ f · f ] = TT s [ f ] · TT s [ f ] = δ · δ. (32)
their corresponding Fourier transforms may be integrated into a
single theory using tempered distributions. As will be described later, the existing definition of multiplication
In fact, tempered distributions include elements that are not for distributions considers restricted distributions, and the majority
covered by the four theories of signal processing. For example, the of studies do not define δ · δ as a distribution.
4 D. Akita / Digital Signal Processing 56 (2016) 1–14
Moreover, the convolution of distributions is only defined for various ways [15,16,18,19]. In some of these, δ 2 = δ . For exam-
a restricted subset of distributions [5–8], and usually the defini- ple, according to Vladimirov [19], δ 2 = C δ , where C is an arbitrary
tion of a convolution only holds for non-periodic distributions, as constant, so that δ 2 = δ if we let C = 1. Bagarello [16] proposed a
follows. For F ∈ S and ϕ ∈ S , multiplication of generalized functions with some parameters, and
we can choose these parameters such that δ 2 = δ . In this paper,
F ∗ ϕ (x) = F (t ), ϕ (x − t ) . (33) however, these theories of generalized functions are not employed.
For F , G ∈ S , if supp G is compact, A new multiplication is proposed, in which δ 2 = δ in the distribu-
tion sense. The reason for not using generalized functions is that
we can utilize the theory of the Fourier transform for tempered
F ∗ G , ϕ = F , Ǧ ∗ ϕ . (34)
distributions, which has been studied for many years. Moreover we
The proof that these convolutions have values in S can be found can avoid the complex theory of generalized functions. Although,
in addition to the theory of generalized functions, non-standard
in certain textbooks on distributions (e.g., [9,10]), and in fact, the
analysis has been used to justify δ 2 [20,21], our proposed defini-
convolution of F ∈ S and ϕ ∈ S given by equation (33) belongs
tion of multiplication does not require either of these theories.
to O M . Smith [4] provides a rigorous and concise review of peri-
In addition to the above literature, we should refer to Ober-
odic tempered distributions, and defines a convolution for periodic
guggenberger’s book [22], where many studies on multiplication
tempered distributions. However, this cannot be applied to non-
of distributions are reviewed. He describes three main approaches
periodic tempered distributions. It is preferable to have a theory
to defining multiplication. The first is called “regular intrinsic op-
with one simple definition of convolution for an integrated treat-
erations”, in which regular objects such as functions are treated
ment of signals, as with Fourier transforms, rather than separate
and the operations are defined classically. The second is “irregu-
definitions for the four cases mentioned above.
lar intrinsic operations”, in which a product for particular pairs of
These difficulties concerning multiplication and convolution are
singular distributions is defined. The last approach, “extrinsic prod-
frustrating, especially considering that there is a single important
ucts and algebras containing the distributions”, uses objects out-
general formula relating the Fourier transform to multiplication
side of distributions. Furthermore, he describes a list of methods
and convolution in signal processing theory,
relating to the second approach: “(a) the duality method; (b) the
F [ f ∗ g ] = F [ f ]F [ g ], (35) method of Fourier transform and convolution; (c) mollification and
passage to the limit; and (d) analytic and harmonic regularization
which appears in the theories of the respective signal classifica- and passage to the limit or even extraction of the finite part.” The
tions. If we define multiplication and convolution in a comprehen- approach in this paper is similar to (b), with some differences. In
sive way, this theorem can be described without the classification the method (b), a convolution of distributions is first defined, and
of signals. then a multiplication is derived by
In fact, the two problems are very closely related: finding inte-
grated definition of convolution means determining a multiplica- F G = F F −1 [ F ] ∗ F −1 [ G ] . (37)
tion that meets the criterion of formula (35), in which the right-
hand side denotes the multiplication of the spectrum. If the signal In contrast, we first define a multiplication of distributions and
is not periodic, then with some exceptions the spectrum is ex- then derive a convolution by equation (36). As it has been noted
pressed by a function on R, whereas if the signal does have a by Oberguggenberger [22] that it is not possible to simultaneously
period, then the spectrum is expressed by a function on Z [1]. have full generality, freedom in differential-algebraic manipula-
Therefore, if we want to calculate the convolution of periodic sig- tions, and coherence with classical definitions, the domain of the
nals such that equation (35) holds, we must justify the equation multiplication in this paper is not closed under differentiation, as
δ · δ = δ .1 That is, these problems are solved if there exists a mul- mentioned in subsection 4.1.
tiplication such that δ · δ = δ , and we can define convolution as
3. δ -Series distributions and multiplication
f ∗ g = F −1 [F [ f ]F [ g ]] , (36)
3.1. Definition
where F −1 denotes the inverse Fourier transform.
The difficulty of defining the multiplication of any two distribu- In this section, we will present our general definition of sig-
tions was considered by Schwartz [11], who showed that there is nal multiplication, in which the product of two discrete-time sig-
no vector space that allows multiplication, derivation, and the ex- nals of δ -series form, equation (1), is the δ -series whose coeffi-
istence of δ , so the expression δ · δ does not make sense. Therefore, cients are products of the corresponding coefficients in the two
although some studies have extended the domain of multiplica- series. In addition, in this general definition, the product of two
tion of distributions, none of these includes δ · δ (e.g., [6,12,13]). continuous-time signals or the product of a continuous-time signal
Thus, δ 2 has been excluded from the distribution space, but stud- and discrete-time signal will involve a limiting process, also based
ies concerning δ 2 have proceeded by using the limit of a sequence on δ -series. Before presenting this general definition of multiplica-
of functions [14]. As a result of such studies, an extended space tion, we first show that continuous-time signals may be expressed
that includes the space of distributions has begun to be consid- as a limit of δ -series.
ered, termed generalized functions,2 and various multiplications of Let f : R → C be a continuous-time signal, and let
the generalized functions have been proposed [15–17]. The expres- ∞
sion δ 2 , in terms of the generalized functions, has been justified in 1 k k
fN = f δ t− N . (38)
2N 2 N 2
k=−∞
1
In [4], where Smith reviewed periodic tempered distributions, he stated a con- From Corollary 1, we have f N ∈ S for all N ∈ N. Considering the
volution theorem for periodic tempered distributions. However, equation (35) in the limit of f N as N → ∞, we obtain the following theorem.
theorem was not for the Fourier transform of tempered distributions, but for the fi-
nite Fourier transform of periodic tempered distributions.
2
Although the space of generalized functions may or may not include distribu-
Theorem 5. Let f be a complex-valued Lebesgue-measurable right con-
tions, depending on its definition, we refer to it as a set that contains the space of tinuous function on R. If there exist C > 0 and a non-negative integer K
distributions. such that
D. Akita / Digital Signal Processing 56 (2016) 1–14 5
| f (t )| ≤ C (1 + |t |) K (39) which means that h(t ) is integrable over R. Hence, it follows from
Lebesgue’s dominated convergence theorem that
for almost all t, then the sequence { f N } defined by (38) converges to f .
∞ ∞
Proof. Let ϕ ∈ S and I N ,k = [(k − 1)/2N , k/2N ). We consider lim g N (t )dt = g (t )dt . (52)
N →∞
g (t ) = f (t )ϕ (t ), (40) −∞ −∞
∞
k k That is,
g N (t ) = f ϕ 1 I N ,k (t ), (41)
2N 2N
k=−∞ lim f N , ϕ = f , ϕ . 2 (53)
N →∞
where 1 I N ,k (t ) is an indicator function such that
Remark. The idea of a multiplication involving coefficients of δ is
1, t ∈ I N ,k
1 I N ,k (t ) = (42) not new. For example, as described in [4], we can define a map-
0. t∈
/ I N ,k ping F P called a finite Fourier transform from a set of tempered
distributions of period P to a set of sequences, such that the se-
Here,
quences correspond to Fourier coefficients. Moreover, there exists
∞
∞ a convolution ∗ P for tempered distributions of period P such that
1 k k
f N , ϕ = f ϕ = g N (t )dt . (43)
k=−∞
2N 2N 2N F P [ F ∗ P G ] = P F P [ F ]F P [G ]. (54)
−∞
First, we show that g N (t ) converges pointwise to g (t ). For an Here, f , g ∈ O M , F = F −1 [T
2π / P [ f ]], and G = F −1 [T
2π / P [ g ]]
arbitrary number t 0 , we can take a sequence {k1 , k2 , ...} such that with a period P . Therefore, F [ F ∗ P G ] denotes the product of the
coefficients of T2π / P [ f ] and T2π / P [ g ], although this is valid only
∀ N t 0 ∈ I N ,k N . (44) for a specific period.
Then, we see that
kN kN Thus, any function that satisfies the conditions of Theorem 5 is
g N (t 0 ) = f ϕ . (45) a tempered distribution, as the limit of a δ -series. Because, as se-
2N 2N
ries of δ , continuous-time signals attain the same form as discrete-
Noting that {k N /2 N } is a monotonically decreasing sequence con- time signals, it follows that we can define a new series of δ as a
verging to t 0 and f is right continuous, we have product of distributions by using the product of the coefficients of
δ . We now define these concepts in detail. In the following defini-
kN kN
lim g N (t 0 ) = lim f ϕ = f (t 0 )ϕ (t 0 ) = g (t 0 ). tion, aZ + c and Z/b + d denote {ak + c |k ∈ Z} and {k/b + d|k ∈ Z},
N →∞ N →∞ 2N 2N respectively, for a, b, c , d > 0.
(46)
Now, we will show that g N (t )dt → g (t )dt ( N → ∞) by us- Definition 1. A tempered distribution F is a δ -series distribution if
ing Lebesgue’s dominated convergence theorem.3 there exist f : R → C and a sequence of sets { B N } N ∈N satisfying
Given that the following conditions:
∞
h(t ) = sup | g (x)|1 I 1,k (t ), (47) (I) { B N } = {Z/2 N }, otherwise the sequence { B N } is independent
k −1
≤x< 2k of N and there exists a finite number of positive real numbers
k=−∞ 2
T 1 , ..., T n , p 1 , ..., pn such that
we have
∀t ∈ R, N ∈ N h(t ) ≥ g N (t ). (48)
n
BN ⊂ T jZ + p j. (55)
On the other hand, there exist C , K , and L such that j =1
Let A δ denote the set of δ -series distributions. If f and where t is the floor function. Then, T T s [ f ] is a δ -series distribu-
{ B N } satisfy the above conditions for F ∈ A δ , then we write tion ( g , { T s Z})δ .
F = ( f , { B N })δ . Moreover, we use superscripts for { B N } such as
f g fg
{ B N }, { B N }, and { B N } if we need to distinguish between different Example 5. Let f ∈ Q M be right continuous. Then, T [ f ] = ( f ,
{ B N }. The following theorem is easy to derive. {Z/2N })δ ∈ A δ according to Theorem 5.
Theorem 6. For any f and { B N } satisfying the conditions (I) and (II) of According to the definition of a δ -series distribution, { B N } in-
Definition 1, there exists a tempered distribution F such that dicates whether F is continuous or discrete. That is, if { B N } =
F = ( f , { B N })δ ∈ A δ . {Z/2N }, then F is described by a function, otherwise F is described
∞
by a summation of δ(t − h). We denote the closure of N =1 B N
by B ∞ . If { B N } = {Z/2 } then B ∞ = R, otherwise B 1 = B 2 = · · ·
N
Proof. We have already treated the case in which { B N } = {Z/2 N }, = B ∞ and the series contains no accumulation point of R accord-
so we assume that { B N } = {Z/2 N } and omit the “N” in B N and F N , ing to condition (I) of Definition 1. We use this fact in the proofs
because B N does not depend on N according to condition (I) of of some lemmas in the subsection 3.2.
Definition 1. Then, there exists a finite number of positive real For these δ -series distributions, a new multiplication of distri-
numbers T 1 , ..., T n , p 1 , ..., pn , according to that condition. There- butions is defined.
fore, given that ϕ ∈ S , we have
f g
∞
Definition 2. Given F = ( f , { B N })δ and G = ( g , { B N })δ , their prod-
n fg fg
| f (h)ϕ (h)| ≤ | f (kT j + p j )||ϕ (kT j + p j )|. (60) uct is defined as F G = ( f g , { B N })δ ,
where we choose { B N } satis-
j =1 k=−∞
fying condition (I) of Definition 1 and
h∈ B
fg f g
The last term is guaranteed to have a finite value because of Corol- B∞ = B∞ ∩ B∞. (65)
lary 1, so that F is a linear functional on S . The continuity of F
in S can also be derived from equation (60), by considering a se- In Definition 2, it is obvious that f g satisfies condition (II) of
fg
quence of rapidly decreasing functions {ϕ j } that converges to 0. Definition 1, and we can choose { B N } satisfying condition (I) of
Thus, it is shown that F = ( f , { B N })δ ∈ A δ . 2 Definition 1 and equation (65). Therefore, to demonstrate well-
definedness of the multiplication, we will illustrate later that
f g
Example 1. The Heaviside function F G is independent of the choice of f , g, { B N }, and { B N } for
f g
F = ( f , { B N })δ and G = ( g , { B N })δ . Deferring the proof of the
0, t < 0 well-definedness to the subsection 3.2, we present some exam-
u (t ) = (61) ples.
1, 0≤t
satisfies the condition (II) of Definition 1. Therefore, u is a δ -series Example 6. Let F = (t , {Z/2 N })δ and G = (sin t , {Z/2 N })δ . Here,
fg fg f g
distribution expressed by (u , {Z/2 N })δ . { B N } = {Z/2N }, because B ∞ must be B ∞ ∩ B ∞ = R ∩ R = R.
Hence, F G = (t sin t , {Z/2 N })δ .
∞
Example 2. Let F = k=−∞ kδ(t − k). Then, F ∈ A δ , because ∞
f Example 7. Let f ∈ Q M and ∞G = k=−∞ δ(t − kT s ) = (1, { T s Z})δ .
f (t ) = t and {B N }
= {Z} meet the conditions of Definition 1 for F .
Then, f G = ( f , { T s Z})δ = k=−∞ f (kT s )δ(t − kT s ).
Therefore, F is denoted by (t , {Z})δ .
g
Here, suppose that g (t ) = t cos2 (π t ) and { B N } = {Z/2}. Then,
g
g and { B N } also satisfy the conditions of Definition 1, so that Example 8. Let F = (t , {Z})δ and G = (sin t , {Z/2 N })δ . In this
fg f g fg
(t cos2 (π t ), {Z/2})δ ∈ A δ . Given that ϕ ∈ S , we have that case, B ∞ must be B ∞ ∩ B ∞ = Z ∩ R = Z. Note that B ∞ = R,
fg
{B N }
= {Z}. Thus, F G = (t sin t , {Z})δ .
∞
k kπ k Consider the additional expression (t cos2 (π t ), {Z/2})δ for F ,
(t cos2 (π t ), {Z/2})δ , ϕ = cos2 ϕ which is shown in Example 2. Then, F G = (t sin(t ) cos2 (π t ), {Z/2})δ .
2 2 2
k=−∞ Given that ϕ ∈ S , we see that
∞
= kϕ (k) = (t , {Z})δ , ϕ . (62) (t sin(t ) cos2 (π t ), {Z/2})δ , ϕ
k=−∞
∞
k k kπ k
= sin cos2 ϕ
Thus, (t cos2 (π t ), {Z/2})δ also denotes F , and so the expression of 2 2 2 2
k=−∞
F ∈ A δ is not unique.
∞
√ ∞ = k sin(k)ϕ (k) = (t sin t , {Z})δ , ϕ . (66)
Example 3. δ(t − 2) + − 2k − 1) is a δ -series distri-
k=−∞ δ(t k=−∞
√
bution expressed √ by (1, {{2} ∪ ( 2Z + 1)})δ . We can confirm that This is an example of the well-definedness.
{ B N } = {{2} ∪ ( 2Z + 1)} satisfies the condition (I), because
√ Example 9. Let F = (exp(2π it /3), {2Z})δ and G = (exp(−π it ),
B N ⊂ Z ∪ ( 2Z + 1). (63) fg
{3Z})δ and consider their product. Then, { B N } = {6Z}. Therefore,
Example 4. Let f [n] ∈ O M and F G = (exp(−π it /3), {6Z})δ = (1, {6Z})δ . (67)
t The next two examples show that T and T T s are ring homo-
g (t ) = f , (64)
Ts morphisms.
D. Akita / Digital Signal Processing 56 (2016) 1–14 7
This new multiplication does not contradict the multiplication Fig. 1. Graph of θ(t ).
f
f g
Lemma 3. Let F = ( f , { B N })δ , G = ( g , { B N })δ , and F G = ( f g , { B N })δ ,
fg In particular, f = g on R if { B N } = {Z/2 N }.
as in Definition 2. Then,
Proof. If F = 0, then supp F is an empty set. Therefore, we assume
Z ( F G ) ⊃ Z ( F ) ∪ Z (G ). (90) f g
that F = 0 and both B ∞ and B ∞ are not empty sets.
f
That is, (i) { B N } = {Z/2 N }
g
From Lemma 4, we have { B N } = {Z/2 N }. Hence, for any ϕ ∈ S,
supp F G ⊂ supp F ∩ supp G . (91)
∞
Proof. Suppose that Z ( F ) is not empty and let ϕ ∈ S such that f
( f , { B N })δ , ϕ − ( g , { B N })δ , ϕ =
g
( f (t ) − g (t )) ϕ (t )dt = 0.
f
supp ϕ ⊂ Z ( F ). According to Lemma 2, f (t ) = 0 on B ∞ ∩ supp ϕ . −∞
Then,
(99)
| F G , ϕ | ≤ lim J N | f (h) g (h)ϕ (h)| Let r (t ) = f (t ) − g (t ). Then, we can show that r (t ) = 0 for all t ∈ R.
N →∞
fg
h∈ B N We assume that there exists x such that r (x) = 0, and let
r (x) > 0 without loss of generality. Because r (x) is right continu-
≤ lim J N | f (h) g (h)ϕ (h)| ous, there exists ε > 0 such that
N →∞
f
h∈ B N ∀t x < t < x + ε ⇒ r (t ) > 0. (100)
= lim J N | f (h) g (h)ϕ (h)| = 0, (92) Then, given that
N →∞
f
h∈ B N ∩supp
ϕ t−x 1
ϕ (t ) = θ − , (101)
which means that Z ( F G ) ⊃ Z ( F ). Even if Z ( F ) is empty, Z ( F G ) ⊃ ε 2
Z ( F ). Z ( F G ) ⊃ Z (G ) is also shown in the same way, and it follows we have that
that Z ( F G ) ⊃ Z ( F ) ∪ Z (G ). 2
∞
Using the derived lemmas, we consider the arbitrariness of our r (t )ϕ (t )dt > 0. (102)
choice of B N . −∞
and
4. exp-Series distribution and convolution
supp F 2 G 2 ⊂ supp F ∩ supp G . (109)
4.1. Definition
Moreover, according to Lemma 5, f 1 = f 2 on supp F and g 1 = g 2
on supp G, so that f 1 g 1 = f 2 g 2 on supp F ∩ supp G.
In this subsection, we define the convolution corresponding to
Next, we will show that supp F 1 G 1 = supp F 2 G 2 . For this pur-
the multiplication defined in Section 3. First, we define a class of
pose, suppose that supp F 1 G 1 = supp F 2 G 2 , and derive a contra-
f g f g
distributions for convolution.
diction. Noting that neither of { B N1 1 } and { B N2 2 } are equal to
{Z/2N }, we can assume without loss of generality that there ex- Definition 3. Given that f and { B N } satisfy conditions (I) and (II)
ists x ∈ supp F 2 G 2 \ supp F 1 G 1 . Here, we choose ε > 0 such that of Definition 1, if F ∈ S is expressed as
(x − ε , x + ε ) ∩ (supp F 1 G 1 ∪ supp F 2 G 2 ) = {x}, (110) 1
F = lim J N f (h) exp(iht ) (119)
and set ϕ as N →∞ 2π
h∈ B N
t−x
ϕ (t ) = θ . (111) in terms of J N from Definition 1, then we call F an exp-series
2ε distribution and denote F by ( f , { B N })e .
Then, we find that
Let us call the set of exp-series distributions Ae . The elements
F 1 G 1 , ϕ = 0, (112) of Ae and those of A δ are in a one-to-one correspondence by the
Fourier transform, i.e., for any ( f , { B N })e ∈ Ae ,
F 2 G 2 , ϕ = f 2 (x) g 2 (x), (113)
which means that f 2 (x) and g 2 (x) are both not equal to 0. Mean- F [( f , { B N })e ] = ( f , { B N })δ . (120)
while, f 1 (x) = f 2 (x) and g 1 (x) = g 2 (x) are also not equal to 0, Therefore, Ae can be written as
f g
although F 1 G 1 , ϕ = 0. Thus, we have x ∈
/ B ∞1 1 , so that at least
f1 g1
one of B ∞ and B ∞ does not contain x. Let x ∈
f
/ B ∞1 , without loss of Ae = {F −1 [ F ]| F ∈ A δ }. (121)
f1 f2
generality. Because neither of { B N } and { B N } are equal to {Z/2 N },
there exists η for which Remark. Ae is not equal to A δ . For example, the derivative of δ ,
for which
f f
(x − η, x + η) ∩ B ∞1 ∪ B ∞2 = {x}. (114)
δ , ϕ = −ϕ (0), (122)
Then, given that is not in A δ . This can be derived from the fact stated in [2] that
for any tempered distribution T ∈ S that has support {0}, there
t−x
ψ(t ) = θ , (115) uniquely exist c 0 , c 1 , ... such that
2η
∞
we see that
T= c j δ ( j) , (123)
f
F , ψ = ( f 1 , { B N1 })δ , ψ = 0, (116) j =0
However, δ = (it , {Z/2 N })e ∈ Ae , because Now, let us calculate f ∗ g according to Definition 4:
it , {Z/2 N } = F −1 it , {Z/2 N } = F −1 [it] = δ . (124) f ∗ g = (F [ f ], {Z/2 N })e ∗ (F [ g ], {Z/2 N })e
e δ
Conversely, f (t ) = it is not in Ae , because F [it ] = −2π δ is not = (F [ f ]F [ g ], {Z/2 N })e = (F [ f ∗ g ], {Z/2 N })e . (133)
in A δ . We can determine whether or not a tempered distribution
F is in Ae by considering whether F [ F ] is in A δ . Thus, the convolution of Ae is consistent with that of L 1 .
Example 12. δ ∈ A δ belongs to Ae , because Example 15. δ is the identity element of this convolution. In fact,
for any F = ( f , { B N })e ∈ Ae ,
F [δ] = 1 = (1, {Z/2 N })δ ∈ A δ . (125)
That is to say, δ = (1, {Z/2 N })e . F ∗ δ = ( f , { B N })e ∗ (1, {Z/2 N })e = ( f , { B N })e = F . (134)
Example 13. Let f ∈ L 1 . Then, the Fourier transform of f is a Example 16. We can calculate the convolution of distributions
bounded continuous function [9]. Therefore, f = (F [ f ], even if their periods are √ different. For example, given f (t ) =
{Z/2N })e ∈ Ae . That is, L 1 ⊂ Ae . exp(it ) and g (t ) = exp(i 2t ),
√
Convolution is now defined for exp-series distributions: f ∗ g (t ) = (2π , {1})e ∗ (2π , { 2})e = (4π 2 , ∅)e = 0, (135)
f
Definition 4. Given F = ( f , { B N })e and G = ( g , { B N })e , the convo-
g where ∅ denotes an empty set.
fg
lution of F and G is defined as F ∗ G = ( f g , { B N })e , where we
fg
choose { B N } satisfying condition (I) of Definition 1 and Example 17. Let us confirm that the convolution of exp-series
distributions agrees with the convolution in formula (33). An
fg
B∞ = B∞ ∩ B∞.
f g
(126) arbitrary ϕ ∈ S is an exp-series distribution, because F [ϕ ] =
(F [ϕ ], {Z/2N })δ , so that ϕ = (F [ϕ ], {Z/2N })e . Therefore, we can
This convolution, in fact, is derived from calculate the convolution of F = ( f , { B N })e ∈ Ae and ϕ ∈ S .
First, suppose that { B N } = {Z/2 N }. Then, for any ψ ∈ S ,
F ∗ G = F −1 [F [ F ]F [G ]]. (127)
Therefore, the relationship in equation (35) holds, and we can see F ∗ ϕ , ψ = ( f F [ϕ ], {Z/2 N })e , ψ
that the convolution is well-defined, because the multiplication is = ( f F [ϕ ], {Z/2 N })δ , F −1 [ψ]
well-defined and there is a one-to-one correspondence between
A δ and Ae . It is interesting that the convolution derived from ∞
equation (127) is the product of the coefficients of complex ex- = f (ω)F [ϕ ](ω)F −1 [ψ](ω)dω
ponential functions, whereas the multiplication is the product of −∞
the coefficients of δ . We can find a convolution corresponding to
the one described in Smith’s paper [4] for tempered distributions ∞ ∞
1
F and G that have the same period P , where F and G are ex- = f (ω)F [ϕ ](ω) ψ(t ) exp(i ωt )dtdω
2π
pressed with the coefficients an and bn as −∞ −∞
∞
∞ ∞
F= an exp(2π t / P ), (128) 1
= f (ω) F [ϕ ](ω) exp(i ωt )dωψ(t )dt
k=−∞ 2π
−∞ −∞
∞
G= bn exp(2π t / P ). (129) ∞ ∞
k=−∞ = f (ω)F −1 [ϕ (t − x)](ω)dωψ(t )dt
Then, Smith’s convolution F ∗ G is given by −∞ −∞
∞
∞
F ∗G= P an bn exp(2π t / P ). (130) = ( f , {Z/2 N })δ , F −1 [ϕ (t − x)]ψ(t )dt
k=−∞ −∞
Thus, Smith’s convolution also calculates the coefficients of com- ∞
plex exponential functions, and agrees with our convolution. Fur- = ( f , {Z/2 N })e , ϕ (t − x)ψ(t )dt
thermore, our definition does not depend on the period of the
−∞
signal, nor on whether the signal is continuous or discrete.
∞
Example 14. Suppose that f , g ∈ L 1 . Then, the convolution of f = F (x), ϕ (t − x)ψ(t )dt = F (x), ϕ (t − x), ψ(t ).
and g is defined as −∞
∞ (136)
f ∗ g (t ) = f (x) g (t − x)dx, (131)
Next, let { B N } = {Z/2 N } and T 1 , ..., T n , p 1 , ..., pn be positive
−∞ real numbers such that
for which the following formula holds for the Fourier transform of
functions [9]:
n
B∞ ⊂ T jZ + p j. (137)
F [ f ∗ g ] = F [ f ]F [ g ]. (132) j =1
D. Akita / Digital Signal Processing 56 (2016) 1–14 11
∞
This definition of convolution enables us to derive an important
1
F ∗ ϕ , ψ = f (h)F [ϕ ](h) ψ(t ) exp(iht )dt theorem regarding continuous linear time-invariant (LTI) systems.
2π First, we define a continuous LTI system:
h∈ B ∞−∞
∞ (I) Continuity:
| f n |dμ < ∞, (144) For any converging sequence of tempered distributions { F N } →
n=1 E F ( N → ∞),
∞
then n=1 f n is finite for almost all x ∈ E and lim A ( F N ) = A ( F ). (152)
N →∞
∞
∞
f n dμ = f n dμ. (145)
(II) Linearity:
n=1 E For any α , β ∈ C and F , G ∈ S ,
E n=1
The proof can be found in textbooks on the Lebesgue integral (e.g., [23]). A (α F + β G ) = α A ( F ) + β A (G ). (153)
12 D. Akita / Digital Signal Processing 56 (2016) 1–14
(III) Time-invariance: Here, J is not bounded, because if there exists a positive number
For any h ∈ R and F ∈ S , r for which J ⊂ [−r , r ], then given
∞
so F , ϕ = j =1 ϕ (a j ) converges. For a sequence of rapidly de- Theorem 9. Let A be a continuous LTI system. Then, if A (δ) has a com-
creasing functions {ϕl } that converges to 0, we can take a decreas- pact support,
ing sequence { Ll } such that
A ( F ) = F ∗ A (δ) (186)
Ll
|ϕl (t )| < , (179) for any F ∈ S .
(1 + |t |)2
where Ll → 0 as l → ∞. Then, Furthermore, he gave an example of a continuous LTI system
A such that A (δ) does not have compact support, and such that
∞
equation (186) holds when applied to any rapidly decreasing func-
1
| F , ϕl | ≤ Ll → 0 (l → ∞), (180) tion belonging to the set
(1 + j )2 ⎧ ∞ ⎫
j =1
⎨ ⎬
so that F ∈ S . Although F [ A (F −1 [ F ])] must be in S too, we see χ0 = ϕ ∈ S ϕ (t )dt = 0 . (187)
⎩ ⎭
that −∞
⎡ ⎛∞ ⎞⎤
1 The following example, in contrast, presents a case in which
F [ A (F −1
[ F ])], ψ = F ⎣ A ⎝ exp(ia j t )⎠⎦ , ψ supp A (δ) is not compact and the codomain of A is S .
2π
j =1
⎡∞ ⎤ Example 19. Choose a rapidly decreasing function ψ and let the
1 mapping A : S → S be defined by
= F⎣ λ(a j ) exp(ia j t )⎦ , ψ
2π
j =1
K
∞
∞ A ( F ), ϕ = lim ψ(k)τ−k F , ϕ (188)
λ(a j ) K →∞
= λ(a j )δ(t − a j ), ψ = , (181) k=− K
|λ(a j )|
j =1 j =1 for all ϕ ∈ S . First, we will show that A is a continuous LTI system.
Given that F ∈ S and ϕ ∈ S , we have
which does not converge. Therefore, F [ A (F −1 [ F ])] does not have
any value at ψ , giving a contradiction. 2
K
A ( F ), ϕ = lim ψ(k) F , ϕ (t − k)
K →∞
k=− K
In the following, we present a theorem regarding the impulse
response A (δ), which is a well-known theorem in signal process-
K
N
signals and discrete-time signals are transformed to subsets of
(1 + |t |) K φ (n) − ψ(k)ϕ (n) (t − k) tempered distributions by ring homomorphisms. Moreover, this
k=− N proposed convolution is consistent with conventional convolutions
of periodic distributions and non-periodic distributions.
= (1 + |t |) K ψ(k)ϕ (n) (t − k) As a result of these new definitions of operations, we have ob-
|k|> N tained a proof of a well-known theorem regarding the impulse
response of continuous LTI systems of tempered distributions, un-
≤ (1 + |k|) K |ψ(k)|(1 + |t − k|) K |ϕ (n) (t − k)| der the assumption that the input of the system satisfies the con-
|k|> N ditions for the convolution. However, because A δ = Ae , not all
⎛ ⎞⎛ ⎞ elements of A δ are suitable for calculating the convolution. For
∞
C3 applications, it is more useful to have a convolution that is defined
≤⎝ |ψ(k)|(1 + |k|) K⎠⎝ ⎠ , (194) for A δ , because some noise filters are defined using a convolution.
(1 + |k| + 1/2)2
|k|> N k=−∞ This extension of the convolution will be the subject of further re-
∞ search.
k=−∞ |ψ(k)|(1 + |k|)
K
for all t. Noting that converges, we have
N Acknowledgments
lim sup(1 + |t |) K φ (n) − ψ(k)ϕ (n) (t − k) = 0, (195)
N →∞ t ∈R I would like to thank professor Tohsiyuki Nakagaki, associate
k=− N
professor Katsuhiko Sato, and assistant professor Shigeru Kuroda
N for their discussions regarding this theory.
which means that k=− N ψ(k)ϕ (t − k) converges to φ(t ) in S .
Therefore, given a sequence of tempered distributions {G N } that
converges to G, it follows that References
∞
[1] D. Kammler, A First Course in Fourier Analysis, 2nd ed., Cambridge University
lim A (G N ), ϕ = lim GN, ψ(k)ϕ (t − k) Press, 2008.
N →∞ N →∞
k=−∞ [2] L. Schwartz, Théorie des Distributions, Hermann, 1966.
[3] D. Smith, An introduction to distribution theory for signals analysis, Digit. Sig-
= lim G N , φ nal Process. 13 (2) (2003) 201–232.
N →∞
∞
[4] D. Smith, An introduction to distribution theory for signals analysis. Part III.
Periodic tempered distributions, Digit. Signal Process. 41 (2015) 1–21.
= G , φ = G , ψ(k)ϕ (t − k) = A (G ), ϕ , (196) [5] H. Youn, I. Richards, On the general definition of convolution for distributions,
J. Korean Math. Soc. 17 (1980) 13–37.
k=−∞
[6] H. Youn, I. Richards, The Theory of Distributions, Cambridge University Press,
which means A is continuous. The linearity and time-invariance 1995.
are obvious, giving that A is a continuous LTI system. [7] R. Shiraishi, On the definition of convolutions for distributions, J. Sci. Hiroshima
University, Ser. A 23 (1959) 19–32.
According to the proof of Theorem 8, it holds that [8] D. Smith, An introduction to distribution theory for signals analysis. Part II. The
∞
convolution, Digit. Signal Process. 16 (4) (2006) 419–444.
[9] H. Arai, Fourier kaisekigaku (Fourier Analysis), Asakura Shoten, 2003.
λ = F [ A (δ)] = ψ(k) exp(ikt ). (197) [10] W. Krabs, Mathematical Foundations of Signal Theory, Heldermann Verlag,
k=−∞ 1995.
[11] L. Schwartz, Sur l’impossibilité de la multiplication des distributions, C. R. Acad.
For example, let an input of A be sin t = (−π it , {−1, 1})e . Then, Sci. 239 (1954) 847–848.
we see that [12] R. Shiraishi, M. Itano, On the multiplicative products of distributions, J. Sci.
Hiroshima University, Ser. A 28 (2) (1964) 223–235.
A (δ) ∗ (sin t ) = (λ, {Z/2 N })e ∗ (−π it , {−1, 1})e [13] I. Richards, H. Youn, Localization and multiplication of distributions, J. Korean
Math. Soc. 37 (3) (2000) 371–389.
= (−π it λ, {−1, 1})e [14] J. Mikusiński, On the square of the Dirac delta-distribution, Bull. Acad. Pol. Sci.,
⎛ ⎞ Sér. Sci. Math. Astron. Phys. 14 (9) (1966) 511–513.
∞
1 [15] J. Colombeau, A multiplication of distributions, J. Math. Anal. Appl. 94 (1)
= ⎝π i ψ(k) exp(−ik)⎠ exp(−it ) (1983) 96–115.
2π [16] F. Bagarello, Multiplication of distributions in any dimension: applications to
k=−∞
⎛ ⎞ δ -function and its derivatives, J. Math. Anal. Appl. 337 (2) (2008) 1337–1344.
∞
[17] J. Colombeau, New Generalized Functions and Multiplication of Distributions,
1
+ ⎝−π i ψ(k) exp(ik)⎠ exp(it ) Elsevier Science Ltd., 1984.
2π [18] Y. Egorov, A contribution to the theory of generalized functions, Russ. Math.
k=−∞ Surv. 45 (5) (1990) 1–49.
∞
[19] V. Vladimirov, Obobshchennye funktsii i ikh primeneniya (Generalized Func-
exp(i (t + k)) − exp(−i (t + k))
= ψ(k) tions and Their Applications), Znanie (1990).
2i [20] Li Bang-He, Nonstandard analysis and multiplication of distributions, Sci. Sin.
k=−∞ 21 (1978) 561–585.
∞
[21] T. Nitta, T. Okada, Infinitesimal Fourier transformation for the space of func-
= ψ(k) sin(t + k) = A (sin t ). (198) tionals, in: Proceedings of the Conference on Differential & Difference Equa-
tions and Applications, 2006, pp. 871–883.
k=−∞ [22] M. Oberguggenberger, Multiplication of Distributions and Application to Partial
Differential Equations, Longman Sc & Tech, 1992.
5. Conclusion [23] S. Ito, Lebesgue sekibun nyuumon (Introduction to Lebesgue Integral), Shokabo,
1963.