Use of the Wigner-Ville distribution in

interpreting and identifying ULF waves in

triaxial magnetic records
P. J. Chi
1
and C. T. Russell
1
Received 13 April 2007; revised 2 August 2007; accepted 8 October 2007; published 31 January 2008.
[1] Magnetospheric ultra-low-frequency (ULF) waves (f = 1 mHz to 1 Hz) exhibit highly
time-dependent characteristics due to the dynamic properties of these waves and, for
observations in space, the spacecraft motion. These time-dependent features
may not be properly resolved by conventional Fourier techniques. In this study we
examine how the Wigner-Ville distribution (WVD) can be used to analyze ULF waves.
We find that this approach has unique advantages over the conventional Fourier
spectrograms and wavelet scalograms. In particular, for Pc1 wave packets, field line/cavity
mode resonances in the Pc 34 band, and Pi2 pulsations, the start and end times
of each wave packet can be well identified and the frequency better defined. In addition,
we demonstrate that the Wigner-Ville distribution can be used to calculate the
polarization of wave signals in triaxial magnetic field data in a way analogous to Fourier
analysis. Motivated by the large amount of ULF wave observations, we have also
developed a WVD-based algorithm to identify ULF waves as a way to facilitate the rapid
processing of the data collected by satellite missions and the vast network of
ground magnetometers.
Citation: Chi, P. J., and C. T. Russell (2008), Use of the Wigner-Ville distribution in interpreting and identifying ULF waves in
triaxial magnetic records, J. Geophys. Res., 113, A01218, doi:10.1029/2007JA012469.
1. Introduction
[2] The ultra-low-frequency (ULF) waves (f = 1 mHz to
1 Hz) in the magnetosphere are intimately related to many
magnetospheric processes, ranging from the interaction of
particles, the normal-mode resonances of magnetospheric
field lines, to the onset of substorms [see, e.g., Takahashi et
al., 2006]. One of the fundamental characteristics of the
ULF waves, as clearly indicated by the classification into
the subgroups of continuous pulsations (Pc) and irregular
pulsations (Pi) [Jacobs et al., 1964], is that the wave
frequency and phase can be quite changeable, creating
complications and challenges in wave analysis. For satellite
observations the motion of the spacecraft can also introduce
an artificial phase drift that produces frequency shifts.
[3] Fourier analysis has been the dominant spectral tech-
nique in ULF research, as its formulation closely matches
those used in wave physics. A few other techniques have
also been used in ULF analysis. For example, complex
demodulation was occasionally used due to its power in
distinguishing time-dependent changes in wave phase
[Beamish et al., 1979]. Since the 1990s, as the wavelet
technique became popular, there has gradually been more
use of wavelet in ULF research [e.g., Rees et al., 2003]. A
pertinent question then is, Are the existing techniques
sufficient for a thorough investigation of ULF waves or
can any other techniques not presently in our arsenal
provide advantages in analyzing ULF waves?
[4] In this study we demonstrate that the use of the
Wigner-Ville distribution (WVD) in time-frequency analy-
sis can better characterize the time-frequency properties of
some types of ULF waves than Fourier spectrograms and
wavelet scalograms. In addition, the WVD can be used to
calculate the polarization of wave signals in triaxial mag-
netic field data in a way analogous to Fourier analysis. Also
motivated by the formidable task of investigating the
enormous amount of ULF wave events collected by satel-
lites, we have developed an algorithm based on the WVD
together with statistical tests to automatically identify ULF
waves. This procedure can significantly facilitate statistical
studies of ULF waves as many future multisatellite missions
will dramatically increase the amount of data that can be
analyzed.
[5] In the following we start with a brief description of
the WVD and its use in time-frequency analysis, followed
by the examples of several often-studied classes of ULF
waves, including Pc 1 wave packets, Pi 2 pulsations, and the
field line/cavity mode resonances in the Pc 34 band.
Section 4 lays out the procedure of the automatic identifi-
cation of wave events based on WVD, and section 5
describes how to use WVD in polarization analysis. Dis-
JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 113, A01218, doi:10.1029/2007JA012469, 2008
Click
Here
for
Full
Article
1
Institute of Geophysics and Planetary Physics, University of
California, Los Angeles, Los Angeles, California, USA.
Copyright 2008 by the American Geophysical Union.
0148-0227/08/2007JA012469$09.00
A01218 1 of 11
cussed at the end are some considerations of the choice of
time-frequency methods.
2. Wigner-Ville Distribution
[6] The Wigner-Ville distribution W
s
of a time series
signal s(t) is defined as
W
s
t; w

1
1
s t
t
2

s* t
t
2

e
iwt
dt: 1
[7] This distribution was first introduced by Eugene
Wigner in his calculation of the quantum corrections of
classical statistical mechanics [Wigner, 1932]. It was inde-
pendently derived again by J. Ville in 1948 as a quadratic
representation of the local time-frequency energy of a
signal. The time series function s(t) in equation (1) can be
either real or complex. The only type of complex signal
considered in this paper is the analytic signal, which is
defined as (s(t) + iH[s(t)]) where H denotes the Hilbert
transform [e.g., Cohen, 1994]. The analytic signal has the
same spectrum (or, to be precise, the spectrum multiplied by
two) as that of the original real signal at positive frequencies
but keeps the spectrum at negative frequencies zero.
[8] A significant characteristic of WVD, which essentially
correlates the signal with a time- and frequency-translated
version of itself, is that it does not contain a windowing
function as those in the Fourier and wavelet frameworks.
This unique feature frees WVD from the smearing effect
due to the windowing function, and, as a result, the WVD
provides the representation that has the highest possible
resolution in the time-frequency plane. For example, if a
wave packet is nonzero only for t 2 (t
1
, t
2
), it can be shown
that the corresponding W
s
has to be 0 for t outside this exact
time interval.
[9] There is, however, a well-known disadvantage of
using WVD. When there are multiple components in a
signal, the WVD can become difficult to interpret. Suppose
that the signal s(t) is composed by the sum of two signals,
s
1
(t) and s
2
(t), the WVD becomes
W
s
t; w W
s1
t; w W
s2
t; w 2<W
s1;s2
t; w ; 2
in which W
s
1
,s
2
(t, w)

1
1
s
1
(t +
t
2
) s*
2
(t
t
2
) e
iwt
dt is
called the cross Wigner-Ville function. Representing the
interference between the two signals, the term2<W
s
1
,s
2
(t, w)
has significant nonzero values located between the auto terms
in the time-frequency plane. This cross-term interference
makes it difficult at times to interpret signal properties from
WVD.
[10] One of the special properties of the interference term
is that it is highly oscillatory in the time-frequency plane
compared to the auto terms that represent the true signals
[e.g., Cohen, 1994]. This property inspired the idea that a
smoothing function, or a kernel, can be used to suppress
the interference pattern without much of an effect on the
desired signal. In this approach, the kernel is placed in the
Wigner-Ville function as

W
s
t; w

1
1

1
1
f t
0
; w
0
W
s
t t
0
; w w
0
dt
0
dw
0
; 3
in which f(t, w) is a two-dimensional (2-D) low-pass filter
that can suppress the cross-term interference in the time-
frequency plane. In the last 2 decades, significant progress
has been made in understanding the properties of the
smoothed version of the WVD. There are also a wide range
of kernels constructed, each carrying its special advantages
in suppressing the interference term. In this study we use the
popular kernel, Choi-William function, that has proven
capable of suppressing the interference effectively [Choi
and Williams, 1989].
[11] In this study we also converted real-valued signals to
analytic signals before calculating the WVD. The reason is
that real-valued signals have spectral power at negative
frequencies, which can create significant cross-term inter-
ference in the vicinity of w = 0. Because the analytic signal
is a halfband function, the resulting WVD effectively avoids
all cross-term interference associated with the negative
frequency components [Qian, 2002].
[12] An example of the comparison among the Wigner-
Ville distribution, the short-time Fourier transform (STFT),
and the wavelet scalogram is demonstrated in Figure 1. The
sample time series consists of a wave packet with a constant
frequency, a sharp pulse, and a wave packet with decreasing
frequency (Figure 1a). Figures 1b, 1c, and 1d show the
corresponding STFT, wavelet scalogram (where the Morlet
wavelet is used as the basis), and the magnitude of WVD.
For the STFT, there is a well-known tradeoff between time
and frequency resolution. A Fourier window that can
properly address low-frequency waves will smear the rep-
resentation for sharp pulses. The wavelet scalogram has
smarter architecture in which the window size is inversely
proportional to the wave frequency of interest. Nevertheless,
the wavelet representation still smears the sharp pulse in the
time domain due to the wavelet basis involved. In contrast,
the WVD has a better defined representation of both wave
packets and a very narrow line for the sharp pulse. The
tradeoff is the interference that cannot be removed com-
pletely even when the smoothing function is used.
3. Time-Frequency Analysis of ULF Waves
[13] In the following, we examine the Wigner-Ville
representations for three classes of ULF waves and compare
them with the conventional Fourier spectrograms and wave-
let scalograms. Each class of these ULF waves has been a
well-recognized research subject because of its specific
implications in magnetospheric physics. The detailed
time-dependent characteristics of these waves, however,
may not have been studied extensively or still await
definitive physical explanation.
3.1. Pc 1 Wave Packets
[14] The Pc 1 waves (f = 0.52 Hz) are essentially
generated by the ion cyclotron instability in the magneto-
sphere. One major objective of Pc 1 research is the waves
interaction with ring current particles and its impact on ring
current decay. These waves may occur in the form of
separate bursts, gradually developing into a series of pulsa-
tions that appear as pearls in time series plots. They may
also occur in the form of consecutive groups of pulsations
with varying frequency. A review of Pc 1 waves can be
found, for example, in the paper by Fraser et al. [2006].
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
2 of 11
A01218
[15] Figure 2a shows an example of Pc 1 waves observed
by the Polar satellite. Three wave packets, each containing
fine structure, can be seen in the 3-min interval presented.
The magnetosphere was in the recovery phase of a magnetic
storm that started 3 d earlier, and Polar was located at L
3.5 and 15 magnetic latitude, traveling approximately
along the field line and toward the equator. The
corresponding STFT, wavelet scalogram, and WVD
(Figures 2b2d) all identify the three wave packets. A hint
of frequency splitting in the third wave packet is seen in the
wavelet scalogram but is more clearly evident in the STFT
and WVD plots. We have experimented with several dif-
ferent Fourier windows for STFT and with the numbers of
analyzed frequencies for wavelets, but, as demonstrated by
comparing Figures 2b2d, WVD provides the best defini-
tion of the detailed time-frequency structures of these wave
packets.
[16] The basic properties of these waves are consistent
with the generation via the ion cyclotron instability in the
equatorial region. The carrier frequency of all the wave
packets is 1.8 Hz, which was below the equatorial value of
f
He
+, 3.9 Hz, and implied the possibility of O
+
participating
in the process. The relatively steady wave frequency is a
clear indication that the wave source was far from the
observation location because the ambient field measured
by the satellite fell off by 12% during the 3 min of
observation. Instead, the steady frequency is consistent with
the field mapped to the equator that was almost unchanged.
Additional interesting features of this Pc 1 event reside in
the periodicity of wave packets, the drift of the carrier
frequency, and frequency splitting. The repetitive wave
packets have a periodicity of approximately one minute,
but Polar did not observe any other concurrent perturbations
at a similar periodicity, which is a possible mechanism
of wave modulation [e.g., Coroniti and Kennel, 1970;
Rasinkangas and Mursula, 1998]. Also mysterious is the
slight frequency drift, increasing from 1.62 Hz to 1.68 Hz,
in the second wave packet and the frequency splitting in the
third wave packet. Because the Polar satellite essentially
stayed on the same field line during the Pc 1 event, these
detailed characteristics are more likely associated with the
wave source at the equator. In summary, this Pc 1 example
suggests that the source of EMIC waves in the equatorial
region can be bursty, and the WVD is a powerful tool to
pinpoint the detailed evolution of time-frequency features of
Pc 1 wave packets.
3.2. Irregular Pulsation Pi 2
[17] Pi 2 pulsations are associated with the onset of
magnetospheric substorms or the bursty bulk flows in the
magnetotail (see an overview by Olson [1999] and the
references therein). They are excited by the movement of
plasma due to reconnection, and the signal can propagate
via the MHD waves. The signatures of Pi 2 pulsations are
mostly seen in the records of ground magnetometers located
in the nightside, but occasionally they can be found in the
dayside as well [e.g., Sutcliffe and Yumoto, 1991]. Although
the detailed dynamics of these waves are being actively
studied, the frequency of Pi 2 could be the consequence of
several different processes, such as surface waves at the
plasmaspheric boundary [Lester and Orr, 1983], field line
resonance [Takahashi et al., 1988], cavity mode resonance
Figure 1. (a) A signal containing a wave packet with constant carrier frequency, a sharp pulse, and a
chirp; (b), (c), and (d) the short-time Fourier transform (STFT), the wavelet spectrogram, and the Wigner-
Ville distribution (WVD) of the signal.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
3 of 11
A01218
[e.g., Sutcliffe and Yumoto, 1991], and the periodicity of
bursty bulk flows [Kepko et al., 2001]. These waves attract
substantial interest because they are associated with the
explosive phenomena in the nightside magnetosphere, and
their timing and other properties may reveal the properties
of substorms and the magnetotail.
[18] The Pi 2, the second type of impulsive pulsations,
was so designated because of its broadband and short-
duration nature. Its small number of cycles imposes diffi-
culty on traditional Fourier analysis as it is not capable of
identifying the time-dependent features within a short
period of time. Beamish et al. [1979] have attempted to
circumvent this problem by performing complex demodu-
lation analysis on the instantaneous values of amplitude and
phase of Pi 2. This technique, however, assumes that the
signal is confined in a narrow frequency band and can focus
on only the amplitude modulation but not the possible
frequency drift.
[19] To assess if WVD is useful for analyzing the time-
frequency relation in Pi 2 pulsations, we compare the three
techniques using the Pi 2 event observed by several ground
magnetometer stations in North America at 0850 UT on
28 January 2003. This Pi 2 event occurred at the onset of a
substorm, after which the AL index dropped to a minimum
of 300 nT at approximately 0920 UT. The magnetic
field in the H-component observed at Cambridge, MN
(CAM, L = 3.20), and Island Lake (ISLL, L = 5.23), are
shown in Figures 3a and 3e. The trend in B
H
for Pi 2
pulsations is often in a form that is more complicated than
the first-order and second-order polynomial fits, causing a
further complication in spectral analysis. To alleviate this
problem, we use the first difference of B
H
, which preserves
the frequency content of the signal as
d
dt
(A exp
iwt
) = iw A
exp
iwt
, for suppressing the low-frequency modulations
without affecting the time-dependent frequency of Pi 2.
Both STFT spectrogram (Figures 3b and 3f) and wavelet
scalogram (Figures 3c and 3g) show that a dominant Pi 2
frequency at 14.5 mHz was seen at both stations. The
wavelet scalogram shows a lower frequency at 5.7 mHz,
which is not seen in other time-frequency representations
and could be related to the duration of the wave packet.
Both STFT and WVD plots show the slightly dropping tone
in the Pi 2 wave packet, and the WVD provides the best
frequency determination.
[20] The WVD results of this Pi 2 event demonstrate the
similarities and differences in frequency observed by the
two stations. Both stations saw an identical frequency at
14.5 mHz initially, and then the higher-latitude station ISLL
experienced a stronger decline in frequency to 9.7 mHz,
compared to the end frequency 11.9 mHz at CAM. It should
be noted that the two stations were located on different sides
of the plasmapause, evidenced by the opposite polarization
presented in the time series plots (Figures 3a and 3e). The
initial frequency at 15 mHz could be dictated by the surface
waves on the plasmaspheric boundary, but the wave period
was then modified by the local field line resonance after 2
3 cycles when the wave energy had sufficient time to
bounce between the two ionospheres. At the lower-latitude
station CAM, on the other hand, a cavity resonance could be
excited and result in the weak oscillations seen after t =
600 s. It is clear that, to study the detailed evolution of Pi 2
frequency, one needs to address the change in wave period
Figure 2. (a) Structured Pc 1 waves observed by the Polar satellite on 7 May 1998. The start time is
2121 UT. (b), (c), and (d) the STFT spectrogram, the wavelet spectrogram, and the WVD for the time
series in Figure 2a.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
4 of 11
A01218
from one cycle to the next, and the WVD is a powerful tool
for such time-frequency analysis.
3.3. Pc 35 Waves: Field Line Resonance and
Phase Skipping
[21] Evidenced by the success of the field line resonance
observations in the last 3 decades, the traditional Fourier
method is sufficient in handling most analysis need. There
are additional benefits, however, that WVD could bring in
through its superior resolution in both time and frequency.
Figure 4 presents the first derivative of the azimuthal
magnetic field (B
f
), the corresponding STFT, wavelet
scalogram, and WVD plot for a Pc 34 event observed
by the ISEE-1 satellite on 24 November 1977. This event is
identical to one of the examples shown by Kivelson et al.
[1997] for demonstrating the observational evidence of the
putative cavity mode resonance. The ISEE-1 satellite was
located at 5 magnetic latitude in the morning sector,
traveling inbound from 10.5 R
E
to 8.6 R
E
in the distance to
the Earth. Shown in all the three time-frequency plots
(Figures 4b4d) are the intermittent wave activities at
frequencies between 7 and 41 mHz recorded during the
80-min interval. The wave power was confined in several
field line resonance harmonics, exhibiting a trend of in-
creasing frequency as the satellite was moving closer to the
Earth.
[22] Although the STFT, wavelet scalogram, and WVD
show consistent results with regard to the frequencies of
field-line and cavity-mode resonances, it is evident that the
Figure 3. Pi 2 waves observed at (left) Cambridge, Minnesota (CAM, L = 3.20) and (right) Island Lake
(ISLL, L = 5.23) on 28 January 2003, and their STFT, wavelet, and Wigner-Ville representations. The
start time is 0850 UT.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
5 of 11
A01218
WVD best preserves the narrow peaks of the wave spec-
trum. Not only can this improved precision provided by
WVD help compare with the modeling of cavity mode
resonance, it also means a better determination of the
plasma mass density inferred from the satellite and ground
observations of field line resonance harmonics [Denton et
al., 2004; Chi et al., 2000, 2005]. The WVD plot also
reveals many broadband (i.e., vertical) power enhancements
that are not found in either STFT or wavelet scalogram.
These broadband structures are associated with large values
in dB
f
/dt or, equivalently, sudden changes in wave phases.
A consequence of impulsive energy input, phase skipping is
a common feature of dayside Pc 34 waves [Chi and
Russell, 1998] but not suited for Fourier or wavelet analysis.
Figure 4 demonstrates that without the need of any window-
ing function, the WVD has a unique advantage to properly
exhibit phase skips rather than broaden the spectral peaks
for neighboring wave cycles.
4. Wave Identification
[23] After decades of operation, satellites and ground
magnetometers have generated a vast amount of data.
However, these data have not been fully explored for
ULF analysis. Identifying the ULF events in a few years
of data recorded by a single satellite can already be a
formidable task if it is done without a computer algorithm
for wave detection. This issue will likely be more serious
Figure 4. Field line resonance in the Pc 35 band observed by the ISEE-1 satellite on 24 November
1977, showing (a) time series; (b), (c), and (d) STFT, wavelet, and WVD representations. The start time is
2101:19 UT.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
6 of 11
A01218
when more multisatellite missions are in operation and the
amount of data increase dramatically. The need to investi-
gate a greater amount of data more efficiently motivated us
to develop a WVD-based algorithm to identify wave events
automatically.
[24] We propose that the selection of wave events in the
WVD can be performed in a very similar way to the
selection in Fourier and wavelet spectrum. For Fourier-
based analysis, one popular selection approach is based on
the comparison with background noise, which can be
modeled as either white noise or red noise. Because the
Fourier spectrum follows the chi-square distribution, we can
calculate the level of background noise for any assigned
confidence interval (such as 99%). Any spectral peak above
such confidence interval can be considered as a wave event
since the probability that it is generated by noise is ex-
tremely small. Torrence and Compo [1998] show that
wavelet-based analysis can follow the same procedure, but
no study to our knowledge has demonstrated whether the
same approach is still valid for WVD-based analysis. In
fact, the WVD has an important frequency marginal condi-
tion that

1
1
W
s
t; w dt jS w j
2
4
where S(w) denotes the Fourier transform of the signal s(t).
Further integration with respect to the frequency,
1
2p

1
1
jS(w)j
2
dw =

1
1
js(t)j
2
dt, yields the total energy of the
signal. In other words, the WVD can be considered as the
signals time frequency density function. When the interval
of time integration is long enough to cover the wave packets
of interest, we can use the finite integral

DT
0
W
s
(t, w)dt to
approximate the spectrum as shown in equation (4). We
have also confirmed, through Monte Carlo analysis, that the
time-integrated WVD follows the chi-square distribution,
and the details are presented in Appendix A.
[25] To illustrate an example of wave detection, we
applied the above method to the Pc 1 wave packets in the
same event presented in section 3.1. For ion cyclotron
waves in triaxial records, the wave power of interest is the
polarized power that is defined as the total power multiplied
by the degree of polarization [e.g., Means, 1972]. More
details on the wave polarization expressed by WVD are
described in the next section and Appendix B. Figure 5a
shows the WVD representation of polarized power and the
selected wave events, based on a confidence interval at
99%, that are denoted by black contour lines. All three wave
packets visualized in Figure 2, including the frequency
splitting in the third wave packet, are successfully selected.
5. Polarization Analysis
[26] For any spectral analysis method, it is desirable to
establish a formulation for polarization analysis to examine
the wave properties frommultichannel data, such as the triaxial
magnetometer data recorded by satellites. The Fourier-based
algorithm for polarization analysis has been widely used for
4 decades [e.g., Fowler et al., 1967; Means, 1972], and it is
especially useful for studying the polarization properties of the
Pc 1 waves. The polarization analysis can also be based on
wavelets, and such approach has been applied to the studies of
seismic waves [e.g., Lilly and Park, 1995].
[27] Here we demonstrate for the first time that the
Wigner-Ville distribution can be used to calculate the
polarization properties of waves. The polarization is often
expressed by the coherence matrix:
J
s
x
t s
x
* t h i s
x
t s
y
* t

s
x
t s
z
* t h i
s
y
t s
x
* t

s
y
t s
y
* t

s
y
t s
z
* t

s
z
t s
x
* t h i s
z
t s
y
* t

s
z
t s
z
* t h i

5
in which the brackets indicate average values within the
coherence time [see, e.g., Means, 1972]. If we define a
matrix J
W
based on the cross WVD as
J
W
t; w
W
xx
t; w W
xy
t; w W
xz
t; w
W
yx
t; w W
yy
t; w W
yz
t; w
W
zx
t; w W
zy
t; w W
zz
t; w

; 6
it can be shown that, for a monochromatic wave, J
W
can be
an alternative coherence matrix to calculate polarization
parameters. A simple proof for the two-dimensional case is
provided in Appendix B.
[28] Using the same Pc 1 example, Figures 5b and 5c
show the details of the propagation angle (to the back-
ground B) and the ellipticity for the same event. It is shown
that all three Pc 1 wave packets propagated in directions
roughly parallel or antiparallel to the background magnetic
field. The ellipticity for the first two wave packets is
basically linear. The third wave packet, however, is more
left-handed polarized. Although it is difficult to provide a
definitive explanation from single-spacecraft measurements,
the difference in ellipticity for the third wavelet is consistent
with the conventional model of ion cyclotron waves in a
plasma with multiple ion species [Perraut et al., 1984]:
When a wave generated near the equator propagates to
higher latitudes, all the critical frequencies increase due to
the increasing magnetic field, and the originally left-handed
polarized wave becomes linearly polarized when the wave
frequency matches the crossover frequency. In this Pc 1
event, the position of the Polar satellite could be in the
region where the wave frequency was close to the crossover
frequency, and the satellite moved closer to the equator
where the wave was more left-handed.
6. Concluding Remarks
[29] In this study we present how the Wigner-Ville
distribution can be used to study ULF waves. We have
demonstrated that the method can either better determine the
time dependence of wave frequency or reveal previously
little explored properties of ULF signals. Compared to
STFT and wavelets, the WVD is most powerful when it is
applied to the types of pulsations that contain rapid varia-
tions in signal amplitude or phase, such as Pi 2, Pc 1
packets, and the phase skips in Pc 34 waves. The WVD
still satisfies the Heisenberg-Gabor uncertainty principle,
but it is free from any reduction in resolution due to the use
of a windowing function. For completeness, we also pro-
vided the formulation for wave detection and polarization
analysis based on the WVD.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
7 of 11
A01218
[30] One caution in using the WVD is that the WVD is
more susceptible to trends than STFT or wavelets. The
WVD results can be seriously affected if the trend is not
properly removed beforehand. In most cases, the often-used
approaches in trend estimation such as a linear fit or a low-
order polynomial fit are sufficient. For some types of ULF
waves, such as Pi 2 pulsations, the trend can be the fast
varying magnetic field related to substorm onsets. In order
to maintain the consistency with the Wigner-Ville method, it
is preferable not to remove the trend by Fourier-based
filtering. One possibility is to decompose the time series
into multiple subsets, each at a different timescale, through
the empirical mode decomposition (EMD) [e.g., Huang et
al., 1998], and then calculate the WVD for the subsets
associated with Pi 2 timescales. Although EMD is an
appealing method and does not rely on any functional basis,
there has been no theoretical foundation for it to interpret
the meaning of each subset. In this study we have demon-
strated that differencing the time series is a useful technique
in preparing Pi 2 events for WVD analysis, and it also
provides an additional benefit to better reveal the phase
skipping in the Pc 34 waves.
[31] Another technical issue with using the WVD tech-
nique is its higher computation costs. Our tests show that,
for a 512-point time series, the STFT and WVD calculations
take 0.075 s and 1.5 s on a PC with a 2.0 GHz CPU.
Although WVD needs a computation time 20 times longer,
it is still an affordable method for wave detection even on
personal computers.
[32] Last, we would like to address an often asked ques-
tion: What is the best time frequency/timescale method to
use? As seen in our presentations, each of STFT, wavelet,
and WVD has its own strengths and weaknesses, and one
should not expect that any of these methods can clearly
express all the features in a wave signal. For analyzing waves
with known properties, the question is often reduced to
Figure 5. WVD representation of (a) polarized wave power, (b) angle of propagation, and (c) ellipticity.
Contour lines in black or white denote the wave events selected for a confidence interval at 99%.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
8 of 11
A01218
identifying the simplest but sufficient method. For time
series with unknown characteristics, we argue that one
should generate all the three, and possibly other, time
frequency/timescale representations to make the best possi-
ble interpretation of wave signals. This is a feasible approach
now considering that the needed computer algorithms and
fast CPUs are readily available. We hope that this study can
be a start of such practice for the analysis of ULF waves and
other types of wave signals.
Appendix A: Relationship Between WVD and
Chi-Square Distribution
[33] A commonly used method to identify wave events is
through the comparison between the wave spectrum and the
background noise. If a spectral peak is much stronger than
the background noise, it can be considered as a true wave
event because the probability for such hypothesis to be false
is minuscule. One necessary piece of knowledge to make
such a comparison is the probability distribution of spec-
trum. It is known that the Fourier spectrum of a time series
follows the chi-square distribution [Jenkins and Watts,
1968], and the wavelet spectrum also has the same property
[Torrence and Compo, 1998]. This Section demonstrates
that the time-integrated Wigner-Ville distribution also fol-
lows the chi-square distribution.
[34] The background noise can be modeled as white noise
x
n
z
n
; A1
where z
n
follows the Gaussian distribution N(0, s
2
), or as
red noise that follows the first-order autoregressive process,
or AR(1),
x
n
ax
n1
z
n
: A2
[35] When a = 0, the equation is reduced to white noise.
The spectral density function for the AR(1) process is
known to be
P f
s
2
1 a
2
2acos 2pf
: A3
[36] Because of the frequency marginal condition of the
Wigner-Ville distribution (see equation (4)), the integration
of W(t, w) with respect to time is expected to follow the
same spectrum and the associated significance levels. To
verify this argument, we performed a Monte Carlo analysis
on 10,000 Wigner-Ville distributions derived from time
series that follow equations (7) and (8). Owing to the use
of analytic signals, the WVD values were divided by two for
comparison. The mean spectrum and the confidence inter-
vals at 80%, 90%, and 95%, for both the time-integrated
WVD the chi-square distribution, are shown in Figure A1.
The comparison confirms that the Wigner-Ville spectrum
closely follow the chi-square distribution.
Appendix B: Polarization Parameters Expressed
by Cross Wigner-Ville Functions
[37] It is shown here that the coherence matrix as shown
in equation (5) can be expressed by cross Wigner-Ville
functions. If we consider two orthogonal wave fields, H
x
and H
y
, of frequency w,
H
x
a
1
t e
i wtf
1

H
y
a
2
t e
i wtf
2

the coherence matrix J can be expressed as
J
H
x
t H
x
* t h i H
x
t H
y
* t

H
y
t H
x
* t

H
y
t H
y
* t

a
2
1
t e
i f
1
f
2

a
1
t a
2
t
e
i f
1
f
2

a
2
t a
1
t a
2
2
t

: B1
Figure A1. (top) Spectral density function for white noise
at the mean value and three different confidence intervals at
80%, 90%, and 95%. Lines and symbol represent the
theoretical values calculated from the chi-square distribu-
tion and the Monte Carlo results from the Wigner-Ville
distributions of 10
4
time series samples. (bottom) The same
as the upper plot except for the red noise following an
AR(1) process with a = 0.7.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
9 of 11
A01218
The quantity a
i
t a
j
t is the average value over the
coherence time. The corresponding expression of J using
the cross Wigner-Ville distribution can be written as
J
W
t; w
W
xx
t; w W
xy
t; w
W
yx
t; w W
yy
t; w

; B2
where
W
xx

a
1
t
t
2

a
1
t
t
2

dt;
W
xy
e
i f
1
f
2

a
1
t
t
2

a
2
t
t
2

dt;
W
yx
e
i f
1
f
2

a
2
t
t
2

a
1
t
t
2

dt;
W
yy

a
2
t
t
2

a
2
t
t
2

dt:
[38] To compare the values of J and J
W
, we can consider a
simple case where the signal has constant amplitude over a
finite time such that a
i
(t) = a
i
for
T
2
t
T
2
and a
i
(t) = 0
otherwise. The elements in J are simply a
i
t a
j
t = a
i
a
j
.
One can also show that, for the integrals in J
W
,

1
1
a
i
(t +
t
2
)
a
j
(t
t
2
) dt = a
i
a
j
(2T 4jtj) for
T
2
t
T
2
. For each t 2
[
T
2
,
T
2
], J
W
is equal to J multiplied by a constant, and
therefore polarization parameters derived from J
W
remain
the same. This result also indicates that the matrix expressed
by the cross Wigner-Ville functions (as equation (11)) can
be used to calculate the properties of wave polarization.
Several commonly used parameters are as follows:
B1. Propagation Angle
[39] Following Means [1972], we can use the Wigner-
Ville coherence matrix (in three dimensions) to obtain the
components of the wave normal vector k = [k
x
, k
y
, k
z
]
k
x

J
W;yz
K
; k
y

J
W;xz
K
; k
z

J
W;xy
K
; B3
in which K = (J
W,yz
2
+ J
W,xz
2
+ J
W,yz
2
)
1/2
. The direction of the
wave propagation with respect to B is
q
kB
cos
1
^
k
^
B

: B4
B2. Polarized Power
[40] Most polarization parameters are more conveniently
calculated in the coordinate system where ^z
0
=
^
k, ^y
0
= ^z
0

^
B,
and ^x
0
= ^y
0
^z
0
. The coherence matrix in this coordinate
system can be obtained through the similarity transformation
J
0
W
R
1
J
W
R; B5
which R is the rotation matrix that connects the original
coordinates (^x, ^y, ^z) and the new coordinates (^x
0
, ^y
0
, ^z
0
). If
the 2 2 submatrix of J
0
W
that contains elements for x
0
and
y
0
is denoted as J
0
W
[x
0
, y
0
; x
0
, y
0
], the degree of polarization h
is
h 1
4jJ
0
W
x
0
; y
0
; x
0
; y
0
j
J
0
W;x
0
x
0 J
0
W;y
0
y
0

2

1=2
B6
and the polarized power is the total power multiplied by the
degree of polarization
hTr J
W
: B7
B3. Ellipticity
[41] The auxiliary angle b = tan
1
(a
y
0/a
x
0) can be
expressed as
sin 2b
2Im J
0
W;x
0
y
0

J
0
W;x
0
x
0 J
0
W;y
0
y
0

2
4jJ
0
W
x
0
; y
0
; x
0
; y
0
j

1=2
; B8
and the ellipticity is
tan b: B9
For right-handed waves,
^
k
^
B > 0 and this quantity remains
the same. For left-handed waves, however,
^
k
^
B < 0 and the
ellipticity is multiplied by 1.
[42] Acknowledgments. This work was supported in part by NASA
NAG5-12953 and the Polar project. The magnetometer from the Island
Lake station is provided by the Canadian Space Science Data Portal. We
thank the authors of the tftb toolbox for developing the computer programs
for time-frequency calculations.
[43] Amitava Bhattacharjee thanks Lars Blomberg and Fouad Sahraoui
for their assistance in evaluating this paper.
References
Beamish, D., H. W. Hanson, and D. C. Webb (1979), Complex demodula-
tion applied to pi2 geomagnetic pulsations, Geophys. J. R. Astron. Soc.,
58, 471493.
Chi, P. J., and C. T. Russell (1998), Phase skipping and poynting flux of
continuous pulsations, J. Geophys. Res., 103(A12), 29,47929,491.
Chi, P., C. Russell, S. Musman, W. Peterson, G. Le, V. Angelopoulos,
G. Reeves, M. Moldwin, and F. Chun (2000), Plasmaspheric depletion
and refilling associated with the September 25, 1998 magnetic storm
observed by ground magnetometers at L = 2, Geophys. Res. Lett.,
27(5), 633636.
Chi, P. J., C. T. Russell, J. C. Foster, M. B. Moldwin, M. J. Engebretson,
and I. R. Mann (2005), Density enhancement in plasmasphere-ionosphere
plasma during the 2003 Halloween superstorm: Observations along the
330th magnetic meridian in North America, Geophys. Res. Lett., 32,
L03S07, doi:10.1029/2004GL021722.
Choi, H.-I., and W. J. Williams (1989), Improved time-frequency represen-
tation of multicomponent signals using exponential kernels, IEEE Trans.
Acoust. Speech Signal Proc., 37(6), 862871.
Cohen, L. (1994), Time Frequency Analysis: Theory and Applications,
Prentice Hall, Upper Saddle River, N. J.
Coroniti, F. V., and C. F. Kennel (1970), Electron precipitation pulsations,
J. Geophys. Res., 75, 12791289.
Denton, R. E., K. Takahashi, R. R. Anderson, and M. P. Wuest (2004),
Magnetospheric toroidal Alfven wave harmonics and the field line dis-
tribution of mass density, J. Geophys. Res., 109, A06202, doi:10.1029/
2003JA010201.
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
10 of 11
A01218
Fowler, R. A., B. J. Kotick, and R. D. Elliott (1967), Polarization analysis
of natural and artificially induced geomagnetic micropulsations, J. Geo-
phys. Res., 72, 28712883.
Fraser, B. J., T. M. Lotoaniu, and H. J. Singer (2006), Electromagnetic ion
cyclotron waves in the magnetosphere, in Magnetospheric ULF Waves:
Synthesis and New Directions, Geophys. Monogr. Ser., vol. 169, edited by
K. Takahashi et al., pp. 195212, AGU, Washington, D. C.
Huang, N. E., Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-C.
Yen, C. C. Tung, and H. H. Liu (1998), The empirical model decomposi-
tion and the hilbert spectrum for nonlinear and non-stationary time series
analysis, Proc. R. Soc. London, Ser. A, 454, 903995.
Jacobs, J. A., Y. Kato, S. Matsushita, and V. A. Troitskaya (1964), Classi-
fication of geomagnetic micropulsations, J. Geophys. Res., 69, 180181.
Jenkins, G. M., and D. G. Watts (1968), Spectral Analysis and its Applica-
tions, Holden-Day, Boca Raton, Fla.
Kepko, L., M. G. Kivelson, and K. Yumoto (2001), Flow bursts, braking,
and Pi2 pulsations, J. Geophys. Res., 106(A2), 19031915.
Kivelson, M. G., M. Cao, R. L. McPherron, and R. J. Walker (1997), A
possible signature of magnetic cavity mode oscillations in isee spacecraft
observations, J. Geomagn. Geoelectr., 49, 10791098.
Lester, M., and D. Orr (1983), Correlations between ground observations of
Pi 2 geomagnetic pulsations and satellite plasma density observations,
Planet. Space Sci., 31, 143160, doi:10.1016/0032-0633(83)90050-8.
Lilly, J. M., and J. Park (1995), Multiwavelet spectral and polarization
analyses of seismic records, Geophys. J. Int., 122, 10011021.
Means, J. D. (1972), Use of the three-dimensional covariance matrix in
analyzing the polarization properties of plane waves, J. Geophys. Res.,
77(28), 55515559.
Olson, J. V. (1999), Pi2 pulsations and substorms: A review, J. Geophys.
Res., 104(A8), 17,49917,520.
Perraut, S., R. Gendrin, A. Roux, and C. de Villedary (1984), Ion cyclotron
waves: Direct comparison between ground-based measurements and ob-
servations in the source region, J. Geophys. Res., 89(A1), 195202.
Qian, S. (2002), Introduction to Time-Frequency and Wavelet Transforms,
Prentice Hall, Upper Saddle River, N. J.
Rasinkangas, R., and K. Mursula (1998), Modulation of magnetospheric
EMIC waves by Pc 3 pulsations of upstream origin, Geophys. Res. Lett.,
25(6), 869872.
Rees, J. M., C. R. Wilford, R. J. Moffett, Z. C. Dent, and I. R. Mann (2003),
Analysis of magnetometer data using wavelet transforms, Phil. Trans. R.
Soc. London, Ser. A, 361, 133137.
Sutcliffe, P. R., and K. Yumoto (1991), On the cavity mode nature of low-
latitude Pi2 pulsations, J. Geophys. Res., 96(A2), 15431551.
Takahashi, K., S. Kokubun, T. Sakurai, R. W. McEntire, and R. E. Lopez
(1988), AMPTE/CCE observations of substorm-associated standing alf-
ven waves in the midnight sector, Geophys. Res. Lett., 15(11), 1287
1290.
Takahashi, K., P. J. Chi, R. E. Denton, and R. L. Lysak (Eds.) (2006),
Magnetospheric ULF Waves: Synthesis and New Directions, Geophys.
Monogr. Ser., vol. 169, AGU, Washington, D. C.
Torrence, C., and G. P. Compo (1998), A practical guide to wavelet ana-
lysis, Bull. Am. Meteorol. Soc., 79(1), 6178.
Wigner, E. (1932), On the quantum correction for thermodynamic equili-
brium, Phys. Rev., 40, 749759.

P. J. Chi and C. T. Russell, Institute of Geophysics and Planetary Physics,

University of California, Los Angeles, CA 90095-1567, USA. (pchi@igpp.
ucla.edu)
A01218 CHI AND RUSSELL: WIGNER-VILLE DISTRIBUTION
11 of 11
A01218