Local instantaneous frequency estimation

of multi-component signals
Mahmut O

ztu rk, Aydn Akan

*
Istanbul University, Department of Electrical and Electronics Engineering Avcilar, 34320 Istanbul, Turkey
Received 5 September 2006; received in revised form 1 February 2007; accepted 5 March 2007
Available online 5 June 2007
Abstract
We present a method for estimating the instantaneous frequency (IF) of multi-component signals. This method involves
the calculation of a timefrequency energy density of the signal, then obtaining a local IF estimate from this joint density.
Timefrequency energy density is calculated as a least squares optimal combination of multi-window Gabor based evolu-
tionary spectra. The optimal weights are obtained by minimizing an error criterion that is the dierence between a refer-
ence timefrequency distribution and the combination of evolutionary spectra. IF of the signal components is estimated
from the nal evolutionary spectrum at small timefrequency regions as the average of frequencies at each time. As such,
local IF information of a multi-component signal can be estimated in the timefrequency plane.
2007 Elsevier Ltd. All rights reserved.
Keywords: Instantaneous frequency; Timefrequency analysis; Evolutionary spectrum; Multi-component signals
1. Introduction
Instantaneous frequency (IF) of a signal, x(t), is dened as the derivative of the phase of its corresponding
analytic signal, x(t) = A(t)e
j/(t)
where A(t) is the amplitude function and /(t) is the phase [1]. Moreover, from a
joint timefrequency (TF) perspective, the IF of a signal is dened as the average of frequencies at a given time
(or time-conditional mean frequency) [2]:
hxi
t
xt
_
x
St; x
St
dx; 1
where S(t) = jx(t)j
2
is the density in time or time marginal of the TF energy density function S(t, x). Estimat-
ing the IF of a signal is an important problem in many signal processing applications such as communications,
radar, bioengineering, etc. [3,4]. For instance, in spread spectrum communication systems, jammers can be
eliminated by estimating their IF and removing them by a time-varying lter [5].
0045-7906/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compeleceng.2007.03.004
*
Corresponding author. Tel.: +90 212 473 70 70/17914; fax: +90 212 473 71 80.
E-mail addresses: mahmutoz@istanbul.edu.tr (M. O

ztu rk), akan@istanbul.edu.tr (A. Akan).

Available online at www.sciencedirect.com
Computers and Electrical Engineering 34 (2008) 281289
www.elsevier.com/locate/compeleceng
There have been many work in the literature on the estimation of IF of mono-component signals [1,68],
however, not so much contribution have been made for multi-component signals. Conventionally, the IF of a
mono-component signal may be obtained from its TF distribution function as the average of frequencies pres-
ent in the signal at a given time. For multi-component signals such a computation of IF does not have the
same signicance [1]. Furthermore, the usual denition of the IF being the derivative of the phase of the cor-
responding analytical signal fails (or do not approach our intuition) in the case of multi-component signals. In
our approach, we dene local IF of signal components in small TF regions. Hence, we are able to calculate IF
of multi-component signals in the TF domain. IF is estimated using a least squares multi-window evolutionary
spectrum as the TF energy density for the signal.
TF signal analysis is a helpful tool for analyzing the time-varying frequency content of a non-stationary
signal [2]. The WignerVille Spectrum (WVS) is dened as a time-dependent spectrum for non-stationary sto-
chastic process x(t) and given by [9]:
Pt; x EfW t; xg E
_
1
1
x t
s
2
_ _
x

t
s
2
_ _ _ _
e
jxs
ds
_ _
;
where W(t, x) denotes the Wigner Distribution (WD) and the above is the statistical average of the WDs of
the realizations of the process. When we have several observations of the non-stationary process x(t), we can
use an ensemble average of the individual WDs of these observations to estimate the WVS. However, this is
not the case in general; we are only given a single realization of the process. In that case, timefrequency dis-
tributions (TFDs) with a smoothing kernel function is used to estimate the WVS [2]. A good amount of re-
search has been done to design kernels with desired properties yielding unbiased and low variance WVS
estimates [9,11].
A new estimate of the WVS is proposed as the optimal average of multiple-window spectrograms of the
process in [4,12]. In this work, we use a timefrequency TF spectral estimate that is an optimal combination
of evolutionary spectra obtained by a multi-window Gabor expansion [10,13]. The optimal combination coef-
cients are obtained by minimizing the squared error between a reference TFD (which is taken to be the
WignerVille distribution of the signal) and the multi-window spectral estimate.
2. Evolutionary spectral analysis
WoldCramer representation [14], of a non-stationary random signal n(n) can be expressed as an innite
sum of sinusoids with random and time-dependent amplitudes and phases, or
nn
_
p
p
Cn; xe
jxn
dZx;
where Z(x) is considered a random process with orthogonal increments. This is a generalization of the spectral
representation of stationary processes. Priestleys evolutionary spectrum [14,15] of c(n) is given as the magni-
tude square of the evolutionary kernel C(n, x).
On the other hand, for a deterministic signal with a time-dependent spectrum, x(n), 0 6 n 6 N 1, we pre-
sented an analogous representation called the discrete evolutionary transform (DET) [10,16]:
xn

K1
k0
Xn; x
k
e
jx
k
n
; 2
where x
k
= 2pk/K, K is the number of frequency samples, and X(n, x
k
) is the DET kernel. The evolutionary
spectrum is obtained from this kernel as Sn; x
k

1
K
jXn; x
k
j
2
. The DET is obtained by expressing the kernel
X(n, x
k
) in terms of the signal. Thus, for the representation in (2) the DET that provides the evolutionary ker-
nel X(n, x
k
), 0 6 k 6 K 1, is given by
Xn; x
k

N1
0
xw
k
n; e
jx
k

; 3
where w
k
(n,) is, in general, a time and frequency dependent window.
282 M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289

The DET can be seen as a generalization of the short-time Fourier transform, where the windows are con-
stant. The windows w
k
(n, ) can be obtained from either the Gabor representation that uses non-orthogonal
bases, or the Malvar wavelet representation that uses orthogonal bases. Computation of the evolutionary
spectrum using DET and its application to IF estimation and jammer excision in communication systems
can be found in [1618]. The multiwindow Gabor expansion is given by [10]
xn

M1
m0

K1
k0
a
i;m;k
h
i
n mLe
jx
k
n
4

K1
k0
X
i
n; x
k
e
jx
k
n
; 5
where {a
i,m,k
} are the Gabor coecients, {h
i,m,k
} are the Gabor basis functions that are obtained by scaling,
translating and modulating a window function by a sinusoid as
h
i;m;k
n h
i
n mLe
jx
k
n
6
and the synthesis window h
i
(n) is obtained by scaling a unitenergy mother window g(n) as
h
i
n 2
i=2
g2
i
n; i 0; 1; . . . ; I 1:
The multi-window Gabor coecients are evaluated by
a
i;m;k

N1
n0
xnc

i
n mLe
jx
k
n
; 7
where the analysis window c
i
(n) is solved from the bi-orthogonality condition between h
i
(n) and c
i
(n) [10].
Hence, by comparing the representations of the signal in (4) and (5) we obtain the DET kernel as
X
i
n; x
k

M1
m0
a
i;m;k
h
i
n mL: 8
Substituting for the coecients {a
i,m,k
}, we obtain that
X
i
n; x
k

N1
0
xw
i
n; e
jx
k

; 9
where we dened a time-varying window for scale i as
w
i
n;

M1
m0
c

i
mLh
i
n mL:
Then the evolutionary spectrum of x(n) calculated using the window h
i
(n) is obtained by
S
i
n; x
k

1
K
jX
i
n; x
k
j
2
;
where the factor 1/K is used for proper energy normalization. We should mention that normalizing the w
i
(n, )
to unit energy, the total energy of the signal is preserved thus justifying the use of S
i
(n, x
k
) as a TF energy
density for x(n). Furthermore, S
i
(n, x
k
) is always non-negative and it approximates the marginal conditions
[2]; hence, as opposed to many other TFDs, it is interpretable as TF energy density function [10].
3. IF estimation
Estimation of IF is a complex and not well understood task [1,6]. For multi-component signals, time-con-
ditional mean frequency does not give a local information on signal components. The evolutionary spectrum
can be used to obtain a general denition of IF by considering the signal x(n) as a sum of analytic signals with
time-dependent magnitudes and phases, that is
xn

k
jXn; x
k
je
jWn;x
k

;
M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289 283
where W(n, x
k
) = Arg[X(n, x
k
)] + x
k
n. Computing W(n, x
k
) only where jX(n, x
k
)j is signicant, a general IF
function is dened as
IFn; x
k
Wn; x
k
Wn 1; x
k
: 10
This can be accomplished by determining the instantaneous phase at the peaks of the spectra. On the other
hand, as we will see, decomposing the signal into its components jXn; x
k
je
jWn;x
k

, these are analytic functions

that will also provide the IF.
Consider a multi-component signal. The estimation of the signal IF is complicated by the multi-component
nature of the signal. We need thus to separate the dierent components. The estimation is especially dicult at
places where there is overlap of the spectra of the signal components.
4. IF estimation by least squares evolutionary spectrum
Given a realization of a discrete-time, non-stationary process {x(n)}, we intend to obtain a high resolution,
non-negative evolutionary spectral estimate so that the IF of the signal can be estimated. We calculate a
weighted average combination of evolutionary spectra S
i
(n,x
k
) described above that is closest to a reference
TFD in a least squares sense. Given the signal x(n), we calculate evolutionary spectra S
i
(n,x
k
) for i = 0, 1, . . . ,
I 1 as
S
i
n; x
k

1
K

N1
0
xw
i
n; e
jx
k

2
: 11
Gaussian windows are used as h
i
(n), for their optimal concentration in the TF plane [16]. Then we estimate
the WVS of the process x(n) as a weighted average of the evolutionary spectra

P n; x
k

I1
i0
c
i
S
i
n; x
k
; 12
where the weights {c
i
} are obtained by minimizing the error function
e
i

N1
n0

K1
k0
P
R
n; x
k

I1
i0
c
i
S
i
n; x
k

2
13
and P
R
(n, x
k
) is a reference TFD which is taken here as WignerVille Distribution of the signal because of its
optimal TF resolution for a certain class of signals [2]. The reference TFD in this optimization can be chosen
as any one of the high resolution TFDs considering interference terms, positivity and marginal issues. By
using a matrix notation, the minimization problem in (13) can be rewritten as
min
c
i
kP
R
Sck
2
: 14
The solution of this least squares minimization problem is
c
o
S
T
S
1
S
T
P
R
;
where the superscript o denotes optimum. Then a WVS estimate is obtained as optimal weighted average
using fc
o
i
g in Eq. (12).
Finally, we mask or threshold our estimate

P n; x
k
to eliminate any possible negative values caused by any
negative c
o
i
coecient, and result in a non-negative time-varying spectrum, as proposed in [4], i.e.,

P n; x
k

P n; x
k
;

P n; x
k
P0;
0;

P n; x
k
< 0;
_
15
where

P n; x
k

denotes the positive-only part of the spectrum. The global IF of the signal can be calculated
form this TF spectrum as time-conditional mean frequency. However, for a multi-component signal the global
284 M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289

IF will demonstrate the combined average frequency at each time. On the other hand, we dene a local IF of
the signal that can be estimated from the TF density in P Q dimensional small TF regions:
IFn; x
k

k
x
k

P n; x
k

Sn
; p 1P 1 6 n 6 pP q 1Q 1 6 k 6 qQ; 16
where p = 1, 2, . . . , N/P and q = 1, 2, . . . , K/Q.
In our experiments, we obtain both TF density and local IF estimate of several multi-component signals.
5. Experimental results
To illustrate the performance of our proposed method, we rst consider combination of two-chirps. In
Fig. 1a, we show the least squares evolutionary spectral estimate

P n; x
k

of this signal which is calculated

by using I = 4 windows. Fig. 1b shows the global IF estimate obtained from this evolutionary spectrum.
Notice that the local IF information of signal components cannot be observed in the global IF. In Fig. 1c,
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Least Squares Evolutionary Spectrum
0 50 100 150 200 250
0.5
1
1.5
2
2.5
3
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Global IF Estimate
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Local IF Estimate
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Fig. 1. (a) Least squares evolutionary spectrum, (b) global IF estimate and (c) local IF of the signal.
M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289 285
we show the local IF estimate of this two-chirp signal, calculated in 8 8 TF regions using our proposed
method. We see that in the local IF estimate, there are blocking eects due to windowing in the TF domain.
Furthermore, the method fails to extract the local IF information, if the signal components are not resolved in
the evolutionary spectrum.
In the second example, we test the performance of the proposed method on a noisy signal. Consider the
combination of a sinusoidal FM signal and a sinusoidal signal that is added a white Gaussian noise with
10 dB and 5 dB signal-to-noise ratio (SNR). Fig. 2a shows the least squares evolutionary spectral estimate
of this signal with 10 dB additive noise. In Fig. 2b, we give the global IF estimate obtained from the least
squares evolutionary spectrum. In Fig. 2c, the local IF estimate of this signal is shown. Finally, we calculate
the local IF estimate with 5 dB noise. It can be seen from the gures that the local IF estimate can resolve the
signal components even with 5 dB additive noise.
As a nal example, we consider the combination of a sinusoidal FM and two-chirp signals. In Fig. 3a, we
show the least squares evolutionary spectral estimate of this three-component signal. Fig. 3b shows the global
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Least Squares Evolutionary Spectrum
50 100 150 200 250
0.5
1
1.5
2
2.5
3
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Global IF Estimation
Local IF Estimation
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Time [n]
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Local IF Estimation
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Time [n]
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Fig. 2. (a) Least squares evolutionary spectrum with 10 dB SNR, (b) global IF estimate, local IF estimate with (c) 10 dB SNR and (d) 5 dB
SNR.
286 M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289

IF estimate obtained from the least squares evolutionary spectrum. Fig. 3c shows the local IF estimate calcu-
lated in 4 4 timefrequency TF regions. As shown in the above examples, the proposed evolutionary spec-
trum based local IF estimation method reveals the local characteristics of multi-component signals and gives
more information than the classical global, time-conditional IF estimation method. In the literature, global IF
function is estimated using dierent TFD methods and similar results are obtained [6].
6. Conclusions
In this work, we present a new method for obtaining the IF of multi-component signals. Our method uses
the optimal combination in the least squares sense, of evolutionary spectra that are calculated by multi-win-
dow Gabor expansion. The optimal weights are obtained by minimizing the squared error between the
combination of evolutionary spectra and a reference timefrequency TF distribution. Examples show that
our method combines the advantages of multiple-window evolutionary spectral analysis and high resolution
TFDs, i.e., it provides non-negative and high resolution time-varying spectral estimates. Thus, the local IF
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Least Squares Combination of Evolutionary Spectra
50 100 150 200 250
0.5
1
1.5
2
2.5
3
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Time [n]
F
r
e
q
u
e
n
c
y

[
r
a
d
]
Local IF Estimate
50 100 150 200 250
0.5
1
1.5
2
2.5
3
Fig. 3. (a) Least squares evolutionary spectrum, (b) global IF estimate and (c) local IF estimate of the signal.
M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289 287
estimation obtained from this nal TF density can give the local IF of signal components. Meanwhile, the
classical global IF estimation method, gives a single IF function for all components of the signal. It is also
observed from our simulation results that the proposed method is capable of obtaining the local IF informa-
tion only in the TF regions where the signal components are resolved. Hence it is crucial to calculate a high
resolution TF distribution to use in the IF calculation procedure. We also observe from the simulation results
that the proposed local IF estimate has blocking eects due to windowing in the TF domain.
Acknowledgements
This work was partially supported by the Research Fund of The University of Istanbul. Project Nos: UDP-
770/07062006 and UDP-901/31017001.
References
[1] Cohen L, Lee C. Instantaneous frequency, its standard deviation and multicomponent signals. SPIE, Adv Algorithms Archit Signal
Process III 1988;975:186208.
[2] Cohen L. Timefrequency analysis. Englewood Clis, NJ: Prentice Hall; 1995.
[3] Mandel L. Interpretation of instantaneous frequencies. AJP 1974;42:8406.
[4] Cakrak F, Loughlin PJ. Multiwindow time-varying spectrum with instantaneous bandwidth and frequency constraints. IEEE Trans
Signal Process 2001;49(8):165666.
[5] Akan A, Cekic Y, Interference suppression in DSSS communication systems using instantaneous frequency estimation. In: The
6th IEEE international conference on electronics, circuits and systems ICECS99, September 58, 1999, Paphos, Cyprus. pp. 4614.
[6] Boashash B. Estimating and interpreting the instantaneous frequency of a signal Part 1: fundamentals. Proc IEEE
1992;80(4):52039.
[7] Loughlin PJ, Tacer B. Comments on the interpretation of instantaneous frequency. IEEE Signal Process Lett 1997;4(5):1235.
[8] Girolami G, Vakman D. Instantaneous frequency estimation and measurement: a quasi-local method. Measur Sci Technol
2002;13(6):90917.
[9] Martin W, Flandrin P. WignerVille spectral analysis of nonstationary processes. IEEE Trans ASSP 1985;1 33(6):146170.
[10] Akan A, Chaparro LF. Multi-window Gabor expansion for evolutionary spectral analysis. Signal Process 1997;63:24962.
[11] Jones DL, Baraniuk RG. An adaptive optimal-kernel timefrequency representation. IEEE Trans Signal Process
1995;43(10):236171.
[12] Bayram M, Baraniuk RG. Multiple window timefrequency analysis. In: IEEE SP international symposium on timefrequency and
timescale analysis, TFTS96, June 1996, Paris, France. pp. 1736.
[13] Akan A, Chaparro LF. Evolutionary chirp representation of non-stationary signals via Gabor transform. Signal Process
2001;81(11):242936.
[14] Priestley MB. Non-linear and non-stationary time series analysis. London: Academic Press; 1988.
[15] Melard G, Schutter AH. Contributions to evolutionary spectral theory. J Time Series Anal 1989;10:4163.
[16] Suleesathira R, Akan A, Chaparro LF. Discrete evolutionary transform for timefrequency signal analysis. J. Franklin Inst, Special
Issue on TimeFrequency Signal Analysis and its Applications 2000;337(4):34764.
[17] Chaparro LF, Suleesathira R, Akan A, Unsal RB, Instantaneous frequency estimation using discrete evolutionary transform for
Jammer excision. In: IEEE The 26th international conference on acoustic, speech and signal processing, ICASSP-2001, May 711,
2001, Salt Lake City, Utah.
[18] Akan A, Yalcin M, Chaparro LF. An iterative method for instantaneous frequency estimation. In: The 8th IEEE international
conference on electronics, circuits and systems, ICECS 2001, Malta, September 25, 2001. pp. 13358.
Mahmut

Ozturk was born in Istanbul, Turkey in 1977. He received the B.Sc. degree in 2000 and the M.Sc. degree
in 2003 from Istanbul University, Istanbul, Turkey both in Electrical Engineering. He is currently a Ph.D. student
at the same university. He is also a Research Assistant at the department of Electrical and Electronics Engi-
neering, Istanbul University. His research interests are digital signal processing, timefrequency signal analysis
and its applications.
288 M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289

Aydn Akan was born in Bursa, Turkey in 1967. He received the B.Sc. degree in 1988 from the University of
Uludag, Bursa, Turkey, the M.Sc. degree from the Technical University of Istanbul, Istanbul, Turkey in 1991, and
the Ph.D. degree from the University of Pittsburgh, Pittsburgh, PA, USA, in 1996 all in Electrical Engineering. He
has been with the Department of Electrical and Electronics Engineering, University of Istanbul since 1996 where
he currently holds a Professor position. His current research interests are digital signal processing, statistical signal
processing, timefrequency analysis methods and applications of timefrequency methods to communications and
bioengineering. Dr. Akan is a Senior Member of the IEEE Signal Processing Society.
M. O

ztu rk, A. Akan / Computers and Electrical Engineering 34 (2008) 281289 289