Fis - Fourier Transform
Fis - Fourier Transform
Fis - Fourier Transform
by
Campus da FCT da UNL, Quinta da Torre, 2825 114 Monte da Caparica, Portugal
Tel. +351 1 2948520, Fax. +351 1 2957786, E-mail: mdo@uninova.pt
Abstract
In this paper, we study the aperiodic comb signal from the point of view of the Fourier Transform. The comb is
very important in the theory of ideal sampling. The knowledge of its properties is crucial for the establishment of suitable
interpolation schemes. Here, we present sufficient conditions so that the Fourier Transform of an aperiodic comb is an
aperiodic comb. We use this result to propose: 1) an alternative approach to the definition of an almost periodic signal
and its anharmonic Fourier series; 2) a generalisation of the Shannon-Whittakker sampling/reconstruction for the
irregular sampling case. Application of this theory to Pulse Duration Modulation and Pulse Position Modulation is also
presented.
1
Number of pages: 20
Number of figures: 3
2
1. INTRODUCTION
The comb signal is one of the most important entities in Signal Processing, because of its connections with
Fourier Series (FS) and ideal sampling [1]. The usual comb is a periodic repetition of the Dirac's delta (generalised)
function [2,3]. As it is well known, its FT is also a periodic comb [4]. In this paper we will study the Fourier Transform
(FT) of the general aperiodic comb signal and formulate conditions to guarantee that its FT is a aperiodic comb, too. To
see the importance of this subject, let us consider the following practical situation. One of the objectives in
electrocardiogram (ECG) processing is the study of the variability of the cardiac frequency. This is usually done from the
so-called RR intervals that are the time intervals between peaks of consecutive cardiac beats. These values constitute a
time series. We can model the excitation of the heart as a pulse frequency modulation signal. With this, the RR interval
signal can be considered as being proportional to the modulating signal. Therefore, we have a signal that is sampled at a
non-uniform spacing: the beat peak positions. Let dn (n=1, 2, , L) be a sequence of RR intervals. Taking 0 as the time
origin reference, we define a set of sampling instants:
tn = t n-1 + d n t0=0, n=1, 2, , L (1.1)
and a signal, v(t)
v(t n) = d n (1.2)
that is proportional to the modulating signal. In the available commercial systems, the signal v(t) is treated as if it was
obtained by uniform sampling. To analyse the error we are making, we took a signal d n obtained from an ECG signal and
constructed the sequence of instants, t n, through (1.1). We sampled a sinusoid at those instants and at a uniform spacing
nT,
figure 1
where the sampling interval, T, is the mean value of dn. For these two signals, we computed their FT by using the FFT.
The results are shown in figure 1 (top and middle pictures). As it is clear, the use of FFT to compute the FT of the non-
uniformly sampled signal is incorrect. To avoid the problem, we assumed that the signal v(t) was ideally sampled by a
non-periodic comb:
+
p(t) = (t - tn) (1.3)
-
We used this expression to obtain the spectrum shown at the bottom picture in figure 1 that shows a good agreement
with the picture in the middle. This fact means that the available approaches to studying the variability of cardiac
frequency are intrinsically wrong. These considerations served as motivation for the study we present in this paper.
3
The problem of non-uniform sampling has received increasing attention due to practical applications in real life. The
theory of frames [5] has being the most important tool for dealing with the problem. Here we adopt a more general point
of view. We intend to state conditions generalising some of current results on ideal sampling and reconstruction. In
section 2, we present the main result of this paper: under stated conditions, the Fourier Transform of an aperiodic comb
is an aperiodic comb [6]. We precise such (sufficient) conditions. The proof is in appendix A. That result has interesting
implications in some well known fields, as: almost periodic functions and non-uniform sampling. These subjects are
treated in section 3. In section 4 we present two applications to communication theory: the pulse duration modulation
(PDM) and pulse position modulation (PPM) [7]. Most of mathematical base of the theory is in Distribution Theory. We
present in appendix B a brief overview of the Axiomatic Theory of Distributions [8,9].
In the following, we represent the sets of integer and real numbers by Z and R, respectively. The Dirac's symbol will
always be represented by (t).
2. THE FT OF A COMB
Consider a set of instants tn (n=-, ...,0, ..., +) assumed to form a, as fast as n, increasing sequence such that
t= (for theorems 2.1 and 2.2, we only need to assume that the sequence increases faster than |n| ).
Definition 2.1
A comb is a distribution, c(t), defined by:
+
c(t) = (t - tn) (2.1)
-
Theorem 2.1
+
The series (t - tn) is convergent.
-
t t0
r(t) = (2.3)
0 t<0
It is not hard to see that s(t) is a continuous function. In fact, for every finite t, s(t) is a finite sum of continuous
functions. This means that both the series in (2.2) are uniformly convergent. According, to definition b.3 in appendix B
[8,9], the series (2.1) defining the comb is convergent, because c(t) is the second derivative of s(t), that is a continuous
function.
The Fourier Transform [see appendix B] of an aperiodic comb is an anharmonic [10] Fourier Series (FS)
+
-jt n
C() = e (2.4)
-
4
Theorem 2.2
+
-jt n
The series e is convergent.
-
+
1 1
As |P()| < 2
that is a convergent series, since its terms converge to zero faster than (t (n=-, ...,0, ..., +) is
|n| n
- t n
n0
assumed to form a, as fast as n, increasing sequence), the series in (2.5) converges uniformly. P() is defined by a
uniformly convergent series and so it is a continuous function. Its second derivative is C() and then the series
+
-jt n
e is convergent.
-
As it is well known, the FT of a periodic comb is a periodic comb. We are looking for a generalisation of this result to the
non-periodic case. According, to what we just wrote above, we are looking for conditions that ensure that the series (2.4)
represents a comb . Let us introduce some generality by using the comb
+
C() =2 Cn ( - n) (2.6)
-
where n is a, as fast as n, increasing sequence, for now. Its FT-1 is the anharmonic FS
+
jnt
c(t)= Cn e (2.7)
-
*
We assume that the sequence Cn is bounded (but not necessarily convergent!) and that, to ensure realness of c(t), Cn =
+
c(t)= 1 + 2. |Cn| cos ( n t+n ) n =arg(Cn ) (2.8)
1
Again, the conditions we are looking for must ensure that (2.7) represents the comb (2.1). If these conditions exist, (2.7)
will be a Fourier (nonharmonic) series associated to (2.1) and (2.4) a Fourier series associated to (2.6).
+ +
jnt
(t - tn) ~ Cn e (2.9)
- -
and
+ +
-jt n
2 Cn ( - n) ~ e (2.10)
- -
According to the above theorems, all these four series are convergent.
1 2
To go further introduce a sequence of intervals In and In (n=-, ...,0, ..., +) defined by:
5
1
In = ]n-1/4, n+1/4] nZ (2.11)
and
2
In = ]n+1/4, n+3/4] nZ (2.12)
Definition 2.2
We define an almost linear sequence (ALS)[10], TF, with uniform density equal to |F| [5], as a strictly increasing
sequence satisfying:
1
TF= tn: Ft nIn nZ , FR (2.13)
or :
2
TF= tn: Ft nIn nZ , FR (2.14)
{see figure 2).
In other words, to form an ALS with density F we only have to pick one point and only one in each interval of a
This is a more precise definition than relatively dense and has a wider generality than the similar definition given by
Davis [10]2. It is clear that, According to this definition, all the t n can be written as:
tn = n.T + n T=1/F nZ (2.15)
in the type 1 case and
tn = (n + 1/2).T + n T=1/F nZ (2.16)
in the type 2 case, with
0|n|< 1/(4F) = T/4 (2.17)
If all the n are zero, the t n sequence will be uniform t n=nT=n/F. In the following we will assume that the sequence of n has
a zero mean value (if the n sequence were not zero mean valued, t n could be written as t n= nT+0 +n, corresponding to a
sliding of the whole sequence. So, we do not loose generality with this assumption). In figure 2 we show how the
intervals are defined for different values of F (f1 > f2)
Figure 2
From (2.15) or (2.16), we conclude that the minimum distance between consecutive elements in an ALS is equal to
1 3
, while the maximum distance is . This means that every strictly increasing sequence with minimum and maximum
2F 2F
distances equal, respectively, to m and M with M 3m is an ALS. We compute T from a given sequence, tn, by
adjusting to it a stair function:
n = t n nT
6
and forcing n to have zero mean. As nT is an odd function, we force both sequences (corresponding to n>0 and n<0) to
be approximated by a stair function. Consider the expressions:
N N N N
lim lim 2 lim lim 2
N (tn nT) = 0 T=
N N(N+1) tn and
N (-t-n nT) = 0 T= -
N N(N+1) t-n
1 1 1 1
The average of the two is
N
lim 2
N N(N+1) n -n
T= (t - t ) (2.18)
1
Definition 2.3
We define an almost periodic comb as a comb defined on an Almost Linear Sequence.
Theorem 2.3
If t n is an Almost Linear Sequence, there is an Almost Linear Odd Sequence k = --k such that:
+ +
FT[ (t - tn)] = 2 Cn ( - n) (2.19)
- -
where the Cn coefficients form a bounded sequence and are given by:
N
1 lim 1 -j t
Ck =
T N 2N+1 e k n (2.20)
n=-N
7
Figure 3
As it is clear the previous procedure can be used to compute other frequencies (almost harmonic frequencies) and we
can also conclude that the set of these frequencies is numerable.
3. CONSEQUENCES
Let xb(t) a signal with FT Xb() and tn an ALS . Let us assume that Xb() is a bounded function. We define an almost
periodic function, x(t), as the generalised function resulting from convoluting xb(t) with c(t) given by (2.1). Attending to
the properties of the function, we can write:
x(t) = xb (t - tn) (3.1)
n=-
By the use of eq. (2.19) we conclude that x(t) is represented by the anharmonic FS [11,12]:
+
jnt
x(t) = Xn e (3.2)
-
with
Xn = Cn. Xb(n) (3.3)
However, using (2.1) and using the definition of FT, we obtain:
+
Xn = <c(t)e -jn t
>. x ().e-jn d
b
-
+ +
=
lim 1
-j s
c(t) xb(s-t).e n dsdt
2
- -
or, by changing the integration orders and noting that the convolution of c(t) with xb(t) is x(t):
lim 1 -j s
2
Xn = x(s).e n ds (3.4)
-
that is the usual way of computing the Fourier coefficients associated with an almost periodic function [11,12]. Therefore,
an almost periodic function is an almost periodic repetition of a basic function.
It is important to remark that the reduction of (3.2) to a trigonometric polynomial is obtained only for signals, xb(t), which
are band-limited or have a FT with nulls at the n for n>N0 Z+ (since Xn = 0).
The dual problem leads to the ideal sampling. Defining a signal xp(t) by:
xp(t) = x(t).c(t) (3.5)
we obtain after using eq. (2.19)
Xp() = Cn X(-n) (3.6)
n=-
generalising the well-known result. Similarly, insert (2.1) in (3.5) and transforming, we obtain, immediately:
8
-jtn
Xp() = x(tn) e (3.7)
n=-
that shows that Xp() is an almost periodic function. If x(t) is a band-limited signal, -BL, we can recover x(t) from Xp()
by ideal low-pass filtering provided that 12. So, let W1. As, from (2.20) and (a.9), C0=/2, the filter must have a
gain equal to 2/ and its output is easily obtained from (3.7):
W
x(tn)
1
j(ttn )
x(t) = e d
n=- -W
sin[W(ttn )]
= x(tn)
(3.8)
n=- (ttn )
2
that is the generalisation of the Shannon-Whittaker cardinal series. In particular, putting t=kT and tn= (n+n)T and
noting that T.=2, we obtain:
sin[.(kn-n )]
x(kT) = x(tn)
(kn-n )
(3.9)
n=-
where
WT 2W
= = 1 (3.10)
The equation (3.9) allows us to obtain the values of a signal on a uniform time spacing from an irregular one. If one puts
t=t k= (k+k)T (kZ) and t n = nT in (3.8), and notes that the sinc function is even, we have:
sin[.(kn-n )]
x(tk) = x(nT)
(kn-n )
(3.11)
n=-
A comparison of equations (3.9) and (3.11) leads us to conclude, that under the described circumstances the sinc
functions appearing in those equations are orthogonal.
9
with |x(t)| 1 and 0<<T/4. Without loosing generality, x(t) is assumed to have a null mean.
It is not a simple task to obtain directly the expression of the modulated signal, sPDM(t). This is a sequence of
rectangular width-variant pulses. However, it is very easy to write its derivative:
+ +
'
s
PDM
(t) = (t - nT) - (t - tn) (4.3)
- -
easily obtained from (2.9). The Fourier coefficients are computed from (2.20). Computing the primitives of both series in
(4.4), we obtain the expression of the PDM signal:
2
1 j T nt Cn jnt
s
PDM
(t) = S0 + j2n
e - jn
e (4.5)
n0 n0
Cn jnt Cn jn jnt
s
PPM
(t) =
T
+ jn
e - jn
e .e (4.7)
n0 n0
or
s
PPM
(t) =
T
+
Cn
jn
[ 1 -ej ].ej t
n n
(4.8)
n0
5. CONCLUSIONS
In this paper, we studied the aperiodic comb signal and its Fourier Transform. We showed how we can guarantee
that the FT of an aperiodic comb is an aperiodic comb. We used this result to propose an alternative approach to the
definition of an almost periodic signal and computed its anharmonic Fourier series. The original first goal of this theory
was the generalisation of the Shannon-Whittakker sampling/reconstruction for the irregular sampling case was also
presented. Based on the previous results we presented an application to Pulse Communication by proposing exact
expressions for Pulse Duration Modulation and Pulse Position Modulation signals.
10
6. APPENDIX A PROOF OF THEOREM 2.3
Before going into the proof, we introduce the long-term average (LTA)[11] or mean value [12],<g>, of a generalised
function, g(t), by:
lim 1
<g> = 2 g(t) dt (a.1)
-
+
A more rigorous statement would write <g > =
lim 1
g(t) dt and the convergence would be uniform in .
2
-+
However, as we work in the field of the generalised functions, this subject is not important. With (a.1), it is easy to show
that:
0 if n k
<e jn t -j t
e k > = (a.2)
1 if n = k
For proof of theorem 2.3, assume that (2.19) is valid. Computing the inverse FT of both members, we obtain:
+ +
jnt
(t - tn) = Cn e (a.3)
- -
-jk t
Multiplying both members of (a.3) by e and computing the LTA, we obtain:
>< >
+
< jnt -j t
Cn e
-jk t
(t - tn)e = e (a.4)
- n=-
Using (a.2 ) the right hand side in (a.4) is Ck, that, by using (a.1) leads to
lim 1 + -j t
2
Ck = (t - tn)e k dt (a.5)
- -
that is the formula to compute the Fourier coefficients [11,12]. If the frequencies k k=1, 2, ..., are multiples of a given
frequency 0, we obtain the ordinary (harmonic) FS. Now, write =(N+1/2)T. Performing the computation, we obtain:
N
1 lim 1 -j t
Ck =
T N 2N+1 e k n (a.6)
n=-N
1
The sequence Cn is bounded which guarantees the convergence of the series in (2.9) and (2.10). In fact |Cn| . We are
T
going to give an interpretation to equation (a.6).
j k t n j k t
a) The signal e is the result of sampling a cisoid e in a time spacing that is an ALS. Equation (a.4)
represents the average of those values.
b) When performing the summation in (a.6) we are adding unitary vectors in the complex plane. The resulting
vector is, in general, of finite length, even if we extend the summation to . This does not happen if the
vectors are collinear, or in a more general situation, if all their extremities lye in the same complex half plane
11
(right or left). If the vectors are collinear, the resulting vector will have a length equal to the sum of the
lengths of all the vectors; otherwise, the resulting vector will have a length inferior to the sum of the lengths.
Assuming that tn has the form (2.15) or (2.16) with n having zero mean, the length of the resulting vector will
be equal to the sum of the real parts of the vectors and so:
N
1 lim 1
Ck =
T N 2N+1 cos(k t n ) (a.7)
n=-N
where ktn = 2f ktn and f ktn satisfies (2.13) or (2.14) for all n,kZ. As the 1st member in (2.19) is given by (2.4) and using the
LTA definition we can write:
W
+
lim 1 jt
1=
W2W 2 Cn ( - n)e k d (a.8)
-W -
that is valid for every nZ. Inserting (a.6) into (a.9), we obtain after some manipulation:
N M
2 lim 1 -jk ( t m - t n )
1=
T N,M (2N+1)(2M+1) e nZ (a.10)
m=-N k=-M
for nZ. According to the considerations we did before concerning the interpretation of the summations as vectors, we
N M
-jk ( t m - t n )
conclude that in the e only the terms corresponding to tm = tn, contribute to the limit. This is not
m=-Nk=-M
N M -jk ( t m - t n ) N M -jk ( t m - t n )
e |e | = (2N+1)(2M+1)
m=-Nk=-M m=-Nk=-M
But the terms corresponding to t m = t n, contribute with the value 1 independently of k. So,
N M
1 = (2N+1)(2M+1)
m=-Nk=-M
This is a consequence of the fact that tm -tn = (m-n)T+m -n where m -n ]-1/2,1/2[ (m,nZ, mn) has zero mean. This
means that the corresponding vectors are likely to have their extremities almost uniformly distributed on the unit circle,
so adding to zero. It is easy, now, to obtain the important formula:
2
=1 (a.11)
T
which is the generalisation of the Nyquist result obtained in the uniform case. Consider (a.9) again, multiply both sides
-jm t n
by e , compute the average value and use (a.10)
12
1 lim 1 N jm t n 1 lim 1 N N
j(k - m ) t n
TN2N+1 e = 2
T N (2N+1)2 Ck e
which states orthogonality between two different sinusoids sampled at the same points and is the discrete-time version
of (a.2). As k is an ALS sequence
k = k + k kZ, |k|</4, and 0=0 (a.13)
k - m = (k-m) + k - m with |k - m | < /2, ensuring that, with t n an ALS, the extremities of vectors in (a.8) are likely to
be at any point on the unit circle, leading to a null average of the vectors. Equation (a.13) fixes only the format of the k
sequence; it does not mean that all the corresponding "almost harmonics" exist, because the respective coefficients may
be zero. For example, consider the "extreme" sequence tn = nTT/4, with the signal being selected randomly or
irregularly. In this case, the first almost harmonic has a null coefficient.
13
7. APPENDIX B - ON THE AXIOMATIC THEORY OF DISTRIBUTIONS
7.1 Motivation
In the following, we will present a brief overview on the axiomatic theory of distributions. Although this theory
was developed by Prof. J. Sebastio and Silva [8] in the fifties, it remains almost unknown in the engineering community.
However, in our opinion, it is the most intuitive and direct of the approaches to distributions. Davis [13] proposed an
approach that we consider to be a particular case of the theory we are going to present. The idea underlying the
axiomatic theory of distributions is very simple: enlarge the class of functions in order to make possible to differentiate
indefinitely any continuous function.
p
To begin let I be an interval in R, C(I) the set of continuous functions on I and C (I) the set of functions p times
d p
continuously derivable in the usual sense. We represent by D the operator . So, a function g C (I) can be
dt
represented by:
p
f=D g (b.1)
where f C(I). The right inverse operator of D ( primitivation operator) is represented by S and satisfies:
p
S f=g (b.2)
with
f() d
Sf = (b.3)
I
As the primitive of a given function is defined aside a constant, the set:
1 1
C (I) ={ Sf + K: fC(I), K constant }is contained in C(I): C(I) C (I) . Defining the powers of the operator S by the
n 0
recursion: S f=S(Sn-1) f, with S f=f and denoting by Pn-1 the set of (n-1)th degree polynomials, we will have
C (I) = S f + p : f C(I) and pPn-1
n n
(b.4)
n
where C (I) is the set of the n times continuously differentiable functions in R. Putting, now, C (I) as the intersection
n
of all the C (I) (n=1,..., ) , we conclude that:
1 n-1 n n+1
C(I)C (I) ... C (I) C (I) C (I) ... C (I)
n+1 n
For every n, the derivation can be viewed as an application of C in C . Its generalisation consists
essentially in prolonging the previous sequence to the left in order to obtain:
C(I) ... Cn+1 (I) Cn (I) Cn-1 (I) ... C1 (I) C(I)
We Represent by C(I) the reunion of all the Cn (I) (n=1,..., ) say, the set of all the "functions" that result from the
14
Definition b.1
The elements C(I) such that:
p
=D f p N0, fC(I) (b.5)
Thus, C(I) is a nonempty set, since it contains, at least, the continuous functions.
Axiom b.2
For each fC(I), there is an element Df C(I) such that if f has derivative, f C(I) in the usual sense, then Df
each pair (f,n) [2]. To the least natural number in these conditions, we call degree of the distribution. So, zero degree GF
are the continuous functions; the Heavisides unit step has degree 1 and Diracs impulse has degree 2.
Axiom b.3
For each C(I) there is a n N0 and a function f C(I) such that
= Dn f (b.6)
We conclude that any distribution is defined by sets of pairs (f,n). For example, the Diracs delta function can be defined
by the pairs (r,2) and (u,1), where r and u are the ramp and unit step functions.
Axiom b.4
If n N0 and f and g C(I), then we have Dn f = Dn g if and only if f-g has the form f-g = Pn-1 (t), where Pn-1
15
Consider the simple case of two continuous functions defined in R. In this case, Df=f and Dg=g, Df=Dg being
equivalent to (f-g)=0, or, f-g=constant in R. Axiom 4 corresponds to accept and generalise this fact. As Dn f=Dn+m(Smf)
and Dmg = Dn+m(Sn g ), we conclude that Dn f = Dmg, if and only if Smf - Sn g = Pn+m-1 .
To finish these considerations relative to the axiomatic it is convenient to say that we would show the it must be
compatible and categorical. We do not do it, because it is beyond the objectives of this work.
Having this axiomatic in hands, it is not difficult, now, to show that, with suitable definitions of sum of
distributions and multiplication by a constant, the space of distributions is a linear vectorial space. This means that,
particularly, the sum of two GF is a GF and the addition enjoys the usual properties. The same does not happen with
multiplication. In fact, it is not possible to define product of two GF in order to guarantee that it enjoys the usual
properties (namely, to be associative, to satisfy the product derivative rule and to coincide with usual product when both
factors are continuous functions. We do not go further {see [9]}.
Let fn be a sequence of GF. We will say that fn converges to f in C(I) if and only if there exist pN, F and FnC(I) such
that we have DpFn =fn, DpF = f and that Fn converges uniformly to F in I.
With this definition, we are able to define sum of a series of distributions.
Definition b.3
Let fn a sequence of distributions in I. We say that the series fn is convergent in C if and only if the sequence
1
k
of partial sums g k = fn is convergent.
1
+ 1
If the series is bilateral: fn , it is convergent, if the series fn and fn are both convergent. The sum of the
- 1 -
16
An integral exists or is convergent if and only if both limits exist. It can be proved that if a GF is integrable in R, then (t)
= O(t -1) when t . On the other hand, if there exist < -1 such that (t) = O(t ) when t then is integrable in R [9].
Similarly to the usual procedure, we define Fourier Transform of a generalised function (t) as being the function
() such that:
() =
(t) e-jt dt (b.9)
R
if the integral exists.
Definition b.4
The integral
(t,)d is said to be convergent in R if and only if there exists a primitive of relatively to convergent
in R when .
17
8. REFERENCES
[1] Jerri, A. J. The Shannon Sampling Theorem - Its Various Extensions and Applications: A Tutorial Review,
Proceedings of the IEEE, Vol. 65, No. 11, November 1977, pp. 1565-1596.
[2] Schwartz, L., "Thorie des Distributions", Herman & Cie, 1951.
[3] Lighthill, M. J. - "Introduction to Fourier Analysis and Generalised Functions", Cambridge University Press,
1964.
[4] Bracewell, R. The Fourier Transform and Its Applications, McGraw-Hill Book Company 1965.
[5] Duffin, R. J. And Schaeffer, A. C., A Class of Nonharmonic Fourier Series, Trans. American Math. Soc. Vol.
72, 1952, pp. 341-366.
[6] Ortigueira, M.D., "What about the FT of the comb signal?", Proceedings of SAMPTA-97, Aveiro, Portugal, 1997,
pp. 28-33.
[7] Carls on, A.B., " Communication Systems," McGraw-Hill International Editios, 1983.
[8] Silva, J. S. ,"The Axiomatic Theory of the Distributions", Complete Works, INIC, Lisbon, 1989
[9] Ferreira, J. C., Introduo Teoria das Distribuies, Fundao Calouste Gulbenkian, 1990.
There is an English version: Introduction to the Theory of distributions, Pitman Monographs and Surveys in
Pure and Applied Mathematics, July 1997.
[10] Davis, A. M., Almost Periodic Extension of Band-Limited Functions and its Application to Nonuniform
Sampling, IEEE Trans. On Circuits and Systems, Vol. CAS-33, No. 10, October 1986, pp. 933-938
[11] Levitan, B. M. And Zhikov, V. V.,Almost Periodic Functions and Differential Equations, Cambridge University
Press, 1982.
[12] Cordoneanu, C. Almost Periodic Functions, Wiley Interscience Publishers, 1968.
[13] Davis, A.M.,K Generalised Functions Proceedings of the IEEE International Conference on Circuits and
Systems, Seattle, Washington, USA, 1995, 1652-1649.
18
1000
500
0
0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26
2000
1000
0
0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26
2000
1000
0
0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26
Figure 1 - FFT of a non-uniformly sampled sinusoid (top), FFT of a uniformly sampled sinusoid (middle) and the FT of
f2
17/4
15/4 f1
13/4
11/4
9/4
7/4
5/4
3/4
Figure 2 - definition of (2.11) intervals for two different values of the density
13/4
11/4
9/4
7/4
5/4
3/4
t
t1 t2 t3
Figure 3 - sequence of intervals defined by (2.23) to compute the almost harmonics from the set of instants.
19
1 1 Also with INESC, R. Alves Redol, 9, 2, 1000 Lisbon, Portugal
2
We tried, but failed to prove the conjecture: "every ALS with density F is a sub-sequence of all the ALS with density
kF, k any integer". In this case the proof of theorem 2.3 would be simpler.
3
The even sequence would correspond to a type 2 ALS, that is not suitable here, since 0=0 (see figure 3).
(t)
4
a generalised function is said to be tempered or of polynomial type if and only if R such that is bounded
t
when t .
20