Académique Documents
Professionnel Documents
Culture Documents
Opening remarks
The Fourier series representation for discrete-time signals has some similarities with that
of continuous-time signals. Nevertheless, certain di↵erences exist:
I Discrete-time signals are unique over the frequency range f 2 [ 0.5, 0.5) or
]! 2 [ ⇡, ⇡) (or any interval of this length).
I The period of ?a discrete-time signal is expressed in samples.
Discrete-time signals
ej2⇡kn/N , k = 0, 1, · · · , N 1
N
X1
x[k] = cn ej2⇡kn/N (22)
n=0
N 1
1 X j2⇡kn/N
cn = x[k]e (23)
N k=0
Parseval’s result for discrete-time signals provides the decomposition of power in the
frequency domain,
N 1 NX1
1 X 2
Pxx = |x[k]| = |cn |2 (24)
N k=0 n=0
Thus, we have the line spectrum in frequency domain, as in the continuous-time case.
I The term |cn |2 denotes therefore the power associated with the nth frequency
component
I The di↵erence between the results in the c.t. and d.t. case is only in the restriction
on the number of basis functions in the expansion.
Remarks
I The Fourier coefficients {cn } enjoy the conjugate symmetry property
I The Fourier coefficients {cn } are periodic with the same period as x[k]
I The power spectrum of a discrete-time periodic signal is also, therefore, periodic,
3
1X j2⇡kn/4 1 j2⇡n/4
cn = x[k]e = (1 + e ) n = 0, 1, 2, 3
4 k=0 4
1 1 1
c0 = ; c1 = (1 j); c2 = 0; c3 = (1 + j)
2 4 4
Observe that c1 = c?3 .
N 1
1 X j2⇡ln/N
Pxx [n] = xx [l]e (28a)
N l=0
N
X1
xx [l] = Pxx [n]ej2⇡ln/N (28b)
l=0
N
X1 N 1 N
X1
Discrete- j2⇡kn/N 1 X 2
x[k] = cn e Pxx = |x[k]| = |cn |2
Time N
n=0 k=0 n=0
N
X1
Fourier 1
cn , x[k]e j2⇡kn/N
x[k] is periodic with fundamental period N
Series N
k=0
Opening remarks
I The discrete-time aperiodic signal is treated in the same way as the continuous-time
case, i.e., as an extension of the DTFS to the case of periodic signal as N ! 1.
I Consequently, the frequency axis is a continuum.
I The synthesis equation is now an integral, but still restricted to f 2 [ 1/2, 1/2) or
! 2 [ ⇡, ⇡).
DTFT
Remarks
I The DTFT is unique only in the interval [0, 1) cycles/ sample or [0, 2⇡) rad/sample.
I The DTFT is periodic, i.e., X(f + 1) = X(f ) or X(! + 2⇡) = X(!) (Sampling
in time introduces periodicity in frequency)
P
I Further, the DTFT is also the z-transform of x[k], X(z) = 1k= 1 x[k]z
k
,
evaluated on the unit circle z = ej!
Existence conditions
I The signal should be absolutely convergent, i.e., it should have a finite 1-norm
1
X
|x[k]| < 1 (31)
k= 1
I A weaker requirement is that the signal should have a finite 2-norm, in which case
the signal is guaranteed to only converge in a sum-squared error sense.
I Essentially signals that exist forever in time, e.g., step, ramp and exponentially
growing signals, do not have a Fourier transform.
I On the other hand, all finite-length, bounded-amplitude signals always have
a Fourier transform.
Arun K. Tangirala (IIT Madras) Applied Time-Series Analysis 73
Fourier Transforms for Deterministic Processes References
Energy conservation
Energy is preserved under this transformation once again due to Parseval’s relation:
1
X Z 1/2 Z ⇡
2 1
2
Exx = |x[k]| = |X(f )| df = |X(!)|2 d! (32)
k= 1 1/2 2⇡ ⇡
2 |X(!)|2
Sxx (f ) = |X(f )| ; Sxx (!) = (33)
2⇡
qualifies to be a density function, specifically as the energy spectral density of x[k].
Given that X(f ) is periodic (for real-valued signals), the spectral density of a
discrete-time (real-valued) signal is also periodic with the same period.
1 2
0.9 1.8
0.8 1.6
0.6 1.2
Amplitude
0.5 1
0.4 0.8
0.3 0.6
0.2 0.4
0.1 0.2
0 0
−10 −5 0 5 10 −0.4 −0.2 0 0.2 0.4 0.6
Time Frequency (cycles/sample)
1
X L 1
X j2⇡f L
j2⇡f k j2⇡f k 1 e
X(f ) = x[k]e = Ae =A j2⇡f
k= 1 k=0
1 e
1 cos(2⇡f L)
Sxx (f ) = A2
1 cos 2⇡f
0.5 50
0.4 40
0.3 30
0.2 20
0.1 10
0 0
−10 −5 0 5 10 15 20 −0.6 −0.4 −0.2 0 0.2 0.4 0.6
Time Frequency (cycles/sample)
1
X
j2⇡lf
Sxx (f ) = xx [l]e (36a)
l= 1
Z 1/2
xx [l] = Sxx (f )ej2⇡f l df (36b)
1/2
When x2 [k] and x1 [k] are the output and input of a linear time-invariant system
respectively, i.e.,
n=1
X
1
x2 [k] = G(q )x1 [k] = g[n]x1 [k n] = g1 [k] ? x1 [k] (38)
n= 1
j2⇡f j2⇡f
Sx2 x1 (f ) = G1 (e )Sx1 x1 (f ); Sx2 x2 (f ) = |G1 (e )|2 Sx1 x1 (f ) (39)
Summary
It is useful to summarize our observations on the spectral characteristics of di↵erent
classes of signals.
Properties of DTFT
Linearity property
1. Linearity:
F F
If x1 [k] ! X1 (!) and x2 [k] ! X2 (!) then
F
a1 x1 [k] + a2 x2 [k] ! a1 X1 (f ) + a2 X2 (f )
Shift property
2. Time shifting:
F
If x1 [k] ! X1 (!) then
F
x1 [k D] ! e j2⇡f D X1 (f )
Time reversal
3. Time reversal:
F F
If x[k] ! X(!), then x[ k] ! X( f ) = X ? (f )
If a signal is folded in time, then its power spectrum remains unchanged; however,
the phase spectrum undergoes a sign reversal.
Dual: The dual is contained in the statement above.
Scaling property
4. Scaling:
F F
If x[k] ! X(!) (or x(t) !✓X(F
◆ )),
k F t F
then x ! X(sf ) (or x ! X(sF ))
s s
1
If X(F ) has a center frequency Fc , then scaling the signal x(t) by a factor
s
Fc
results in shifting the center frequency (of the scaled signal) to
s
Note: For real-valued functions, it is more appropriate to refer to |X(F )|,
Convolution
5. Convolution Theorem: Convolution in time-domain transforms into a product in
the frequency domain.
Theorem
F F
If x1 [k] ! X1 (!) and x2 [k] ! X2 (!) and
1
X
x[k] = (x1 ? x2 )[k] = x1 [n]x2 [k n]
n= 1
Product
Z 1/2
F
x[k] = x1 [k]x2 [k] ! X1 ( )X2 (f )d
1/2
Correlation theorem
7. Correlation Theorem (Wiener-Khinchin theorem for deterministic signals)
Deterministic Fourier
x[k] M
ag
X(f )
ni
tu
de
-s
x[k] x[ k]
DF
T
X(f )X (f )
x1 x2 [l] is the cross-energy spectral density |XN (fn )|2
N lim
1
X F
j2⇡f l xx [l] Sxx (f )
F{ x1 x2 [l]} = x1 x2 [l]e = Sx1 x2 (f ) = 2⇡Sx1 x2 (!) Wiener-Khinchin
Theorem
Spectral
ACVF
Density
l=1
Opening remarks
The transform derives its name from the fact that it is now discrete in both time and
frequency.
Main result
For signal x[k] of length Nl , its DTFT X(f ) is perfectly recoverable from its sampled
version X(fn ) if and only if the frequency axis is sampled uniformly at Nl points in
[ 1/2, 1/2), i.e., i↵
1 2⇡
4f = or 4! = (42)
Nl Nl
N-point DFT
The resulting DFT is known as the N -point DFT with N = Nl . The associated analysis
and synthesis equations are given by
N
X1
j 2⇡
X[n] , X(fn ) = x[k]e N
nk
n = 0, 1, · · · , N 1 (43a)
k=0
N
X1
1 2⇡
x[k] = X[n]ej N kn k = 0, 1, · · · , N 1 (43b)
N n=0
Unitary DFT
p
It is also a common practice to use a factor 1/ N on both (43a) and (43b) to achieve
symmetry of expressions.
N 1
1 X j2⇡fn k n
X[n] = p x[k]e fn = , n = 0, 1, · · · , N 1 (44a)
N k=0 N
N 1
1 X
x[k] = p X[n]ej2⇡fn k k = 0, 1, · · · , N 1 (44b)
N n=0
The resulting transforms are known as unitary transforms since they are norm-preserving,
i.e., ||x[k]||22 = ||X[n]||22 .
Arun K. Tangirala (IIT Madras) Applied Time-Series Analysis 101
Fourier Transforms for Deterministic Processes References
N
X1 ✓ ◆ ✓ ◆
2⇡n 2⇡n
X(f ) = X P 2⇡f N Nl (45)
n=0
N N
sin(⇡f N ) j⇡f (N 1)
where P (f ) = e
N sin(⇡f )
I Equation (45) has very close similarities to that for a continuous-time signal x(t)
from its samples x[k] (Proakis and Manolakis, 2005).
I Further, the condition N Nl is similar to the requirement for avoiding aliasing.
Arun K. Tangirala (IIT Madras) Applied Time-Series Analysis 102
Fourier Transforms for Deterministic Processes References
1
X
Now, define xp [k] = x[k lN ], with period Np = N .
l= 1
Arun K. Tangirala (IIT Madras) Applied Time-Series Analysis 103
Fourier Transforms for Deterministic Processes References
N
X1
j2⇡nk/N
N cn = xp [k]e (47)
k=0
The N -point DFT X[n] of a sequence xN = {x[0], x[1], · · · , x[N 1]} is equivalent to
the coefficient cn of the DTFS of the periodic extension of xN . Mathematically,
N 1
1 X j 2⇡ kn
X[n] = N cn , cn = x[k]e N (48)
N k=0
Putting together · · ·
I The basis blocks are cos(2⇡ Nk n) and sin(2⇡ Nk n) characterized by the index n
I The quantity n denotes the number of cycles completed by each basis block
for the duration of N samples
I DFT inherits all the properties of DTFT with the convolution property replaced by
circular convolution.
DFT: Summary
Definition
The N-point DFT and IDFT are given by
N
X1 N 1
j2⇡kn/N 1 X
X[n] = x[k]e ; x[k] = X[n]ej2⇡kn/N
k=0
N n=0
N
X1 N 1
kn 1 X
X[n] = x[k]WN ; x[k] = X[n]WNkn
k=0
N n=0
Points to remember
R: fft
Arun K. Tangirala (IIT Madras) Applied Time-Series Analysis 108
Fourier Transforms for Deterministic Processes References
Thus, strictly speaking we have neither densities. Instead DFT always implies a
power spectrum (line spectrum) regardless of the nature of underlying signal!
|X[n]|2
Pxx (fn ) = |cn |2 = (49)
N2
A heuristic power spectral density (power per unit cyclic frequency), known as the peri-
odogram, introduced by Schuster, (1897), for the finite-length sequence is used,
Routines in R
Bibliography I
Antoniou, A. (2006). Digital Signal Processing: Signals Systems and Filters. USA: McGraw-Hill.
Bloomfield, P. (2000). Fourier Analysis of Time Series: An Introduction. 2nd edition. New York, USA:
John Wiley & Sons, Inc.
Cohen, L. (1994). Time Frequency Analysis: Theory and Applications. Upper Saddle River, New Jersey,
USA: Prentice Hall.
Hamilton, J. D. (1994). Time Series Analysis. Princeton, NJ, USA: Princeton University Press.
Lighthill, M. (1958). Introduction to Fourier Analysis and Generalized Functions. Cambridge, UK: Cam-
bridge University Press.
Priestley, M. B. (1981). Spectral Analysis and Time Series. London, UK: Academic Press.
Proakis, J. and D. Manolakis (2005). Digital Signal Processing - Principles, Algorithms and Applications.
New Jersey, USA: Prentice Hall.
Bibliography II
Schuster, A. (1897). “On lunar and solar periodicities of earthquakes”. In: Proceedings of the Royal
Society 61, pp. 455–465.
Shumway, R. and D. Sto↵er (2006). Time Series Analysis and its Applications. New York, USA: Springer-
Verlag.
Smith, S. W. (1997). Scientist and Engineer’s Guide to Digital Signal Processing. San Diego, CA, USA:
California Technical publishing.
Tangirala, A. K. (2014). Principles of System Identification: Theory and Practice. Boca Raton, FL, USA:
CRC Press, Taylor & Francis Group.