1

EEG oscillations and wavelet analysis

Christoph S. Herrmann1,2, Maren Grigutsch2 & Niko A. Busch1,2

1
Otto-von-Guericke University, Dept. Biological Psychology,
Magdeburg, Germany

2
Max-Planck-Institute of Cognitive Neuroscience,
Leipzig, Germany

Electroencephalographic recordings are analyzed in an

event-related fashion when we want to gain insights into
the relation of the electroencephalogram (EEG) and
experimental events. The standard analysis method is to
focus on event-related potentials (ERPs) by averaging.
However, another approach is to concentrate on event-
related oscillations (EROs). This chapter will introduce
the notion of EEG oscillations and a method suited to
analyze the temporal and spatial characteristics of EROs
at the same time, namely the wavelet analysis. At first
an introduction to oscillatory EEG activity will be
given, followed by details of the wavelet analysis. Some
general prerequisites of recording EROs will be reviewed
and finally, recently introduced wavelet-based methods
for studying dynamical interrelations between brain
signals will be discussed.

1. OSCILLATIONS IN THE EEG

Both EEG and ERP measures can be investigated in the

frequency domain, and it has been convincingly
demonstrated that assessing specific frequencies can
often yield insights into the functional cognitive
correlations of these signals (Başar et al., 1999).

Oscillations were the very beginning of EEG research

when the German neurophysiologist Berger (1929) first
observed the dominant oscillations of approximately 10 Hz
 2

recorded from the human scalp. Berger coined the term

alpha frequency for activity in this frequency range by
using the first letter of the Greek alphabet. Berger
dubbed the second type of rhythmic activity that he found
in the human EEG as beta, which is now considered to be
the frequency range of approximately 12-30 Hz. Following
this consecutive ordering, Adrian (1942) referred to
oscillations around 40 Hz (more general 30-80 Hz)
observed after odor stimulation in the hedgehog as gamma
waves. The slow oscillations below 4 Hz, which were
discovered next, had been coined as delta waves. Finally,
waves that cycle between 4 and 8 times per second (4-8
Hz) were named theta oscillations after the first letter
of their assumed region of origin, the thalamus.

Frequency: Name:
0-4 Hz Delta
4-8 Hz Theta
8-12 Hz Alpha
12-30 Hz Beta
30-80 Hz Gamma
Table 1. A list of well-established frequency bands and their names.

1.1 Evoked versus induced oscillations

Oscillations are characterized by their amplitude and

phase. The amplitude of an EEG oscillation is typically
between 0 and 10 µV. The (cyclic) phase ranges between 0
and 2π. At every point in time the amplitude and phase of
an oscillation can be determined. According to a
classification of different types of oscillatory activity
by Galambos (1992), there are spontaneous, induced, and
evoked rhythms, all of which are differentiated by their
degree of phase-locking to the stimulus (emitted rhythms
in response to omitted stimuli also have been observed,
but these will not be considered here). In this
framework, spontaneous activity is completely
uncorrelated with the occurrence of an experimental
condition. Induced activity is correlated with
experimental conditions but is not strictly phase-locked
to its onset. Evoked activity is strictly phase-locked to
 3

the onset of an experimental condition across trials,

i.e. it has the same phase in every stimulus repetition.

avg

Figure 1. If oscillations occur at the same latency after stimulus

onset and with the same phase relative to stimulus onset in multiple
trials (rows 1-4), they are considered evoked by the stimulus
(left). If latency or phase jitter relative to stimulus onset, the
oscillations are considered to be induced by the stimulus (right).
Evoked activity sums up in the average (bottom row), while induced
activity is nearly cancelled out.

Figure 1 (left) illustrates such evoked oscillations

which start at the same time after stimulation in every
trial and have identical phases. In this case, the
activity is called evoked, sums, and is visible in the
averaged ERP. However, evoked oscillations are only
visible in the ERP to the bear eye if they are of
sufficient amplitude. But, since high frequency
oscillations usually have lower amplitude than low-
frequency oscillations they are often not visible. Evoked
oscillations usually result from any kind of sensory
events, like auditory, visual or somatosensory
stimulation.
If oscillations occur after each stimulation but with
varying onset times and/or phase jitter, they are
considered as being induced by the stimulus rather than
evoked and are not visible in the averaged ERP. Figure 1
 4

(right) illustrates this outcome. Special methods have to

be applied to analyze this type of activity (see below).

1.2 Delta and theta oscillations

An ERP constitutes a mixture of multiple waves of

various frequencies. Using digital filters single
frequencies can be selectively shown while all others are
filtered out. E.g. a band-pass filter lets only certain
frequencies pass. When only theta frequencies are
admitted (4-8 Hz) only such slow oscillations will remain
in the event-related signal. This is illustrated in Fig.
2. Evoked delta and theta oscillations represent the slow
potentials in ERPs, i.e. P300, N400, P600 etc. (Basar-
Eroglu et al., 1992). Functional correlates of event-
related theta oscillations are described for working
memory functions (Tesche & Karhu, 2000; Jensen & Tesche,
2002). Event-related theta oscillations which relate to
memory performance have been shown to interact with
faster oscillations in the gamma frequency range (Fell et
al., 2003). Note that these event-related signals are not
identical to those which can be seen with the bare eye by
physicians in raw EEG and usually relate to deep sleep
(Steriade et al., 1993) or malfunctions (Gloor et al.,
1977).

Figure 2: EROs in the theta band resulting from applying a 4-8 Hz

band-pass filter to an ERP. An event-related theta oscillation
emerges after stimulation which then decays over time.
 5

1.3 Alpha oscillations

The term alpha oscillation usually refers to the ongoing

alpha rhythm. This rhythm with approximately 10 Hz can be
observed in routine EEG recordings without averaging (cf.
Fig. 3). Typically, the amplitude of the 10 Hz rhythm
increases and decreases over time which has been
described as waxing and waning. Some authors even
hypothesized that there exist several independent rhythms
in the alpha band with different functional properties
(e.g. Niedermeyer, 1997). However, this is not the type
of alpha activity which we want to discuss here. We are
interested in 10 Hz oscillations which occur in relation
to an experimental condition, i.e. evoked or induced 10
Hz oscillations.

Figure 3. Ten seconds of unfiltered, spontaneous EEG showing alpha

activity (8-12Hz).

Fig. 4 shows a burst of 10 Hz oscillations after visual

stimulation in an occipital electrode (Oz in the 10-20
system). In order to compute this evoked 10 Hz activity,
at first an ERP was averaged which was then bandpass-
filetered in the alpha frequency range. Note that the
burst of oscillatory activity seems to start before the
onset of stimulation. This is an artefact of the filter
algorithm. The filter uses time points of the past and
future to compute each time point of the filtered signal.
 6

Therefore, the activity ‘leaks’ into past and future

events around its real peak.

Figure 4. Short burst of 10 Hz oscillations evoked by visual

stimulation.

Such bursts of alpha activity are evoked by mostly any

sensory stimulation, i.e. visual, auditory, and
somatosensory. The topography of this evoked alpha
response is restricted to the primary sensory cortex
which was stimulated. Interestingly, this burst of alpha
activity is not due to an increase in amplitude. This
becomes clear when the total power of alpha activity
following a visual stimulus is computed. The total power
contains both evoked and induced activity and typically
decreases after visual stimulation. Thus, the amplitude
of alpha oscillations is reduced after stimulation while
the evoked alpha activity is enhanced. This phenomenon
has been called the alpha paradox (Klimesch et al.,
1998b) and is depicted schematically in Fig. 5. The first
two traces show band-pass filtered alpha activity and its
amplitude reduction after a visual stimulus. When these
traces are simply added up as in the case of an average
(third row) only those oscillations which are phase-
locked (evoked) will remain visible. Due to a so-called
phase-resetting the randomly distributed phase will be
reset to start from the same value after stimulation for
a short period of time (Brandt, 1997). This leads to the
short burst of evoked alpha activity, since oscillations
add up if they have identical phases across trials.
However, the behavior of the amplitude is only reflected
 7

in the total power measure (bottom row) which is

independent of the phase of the oscillations.

Figure 5. A resetting of the phase of alpha oscilations at the

time of stimulation leads to a short increase of evoked activity
(sum) despite the fact that the amplitude (power) decreases.

Alpha activity has been associated with a large number

of cognitive processes. The most important of them are
memory processes (Klimesch 1997; Klimesch et al, 1993),
attention (Klimesch et al., 1998a; Yordanova et al.,
2001) and visual awareness (Sewards & Sewards, 1999;
Strüber & Herrmann, 2002). For an overview see Basar et
al. (1997). While the generators of EEG alpha activity
are aasumed to reside in cortex, these generators are
probably driven by thalamic cells (Steriade et al., 1990;
Lopes Da Silva, 1991).

1.4 Beta oscillations

The frequency range from 12-30 Hz constitutes the beta

frequency band. These oscillations have been investigated
in the context of motor actions. Beta oscillations are
suppressed during motor action but increase (a so-called
rebound) approximately one second after movement with a
 8

topography close to the primary sensorimotor regions

which represent the involved body part (Neuper &
Pfurtscheller,2001). They are also observed during
imagined movements and can be elicited by median nerve
stimulation (Salmelin & Hari,1994). During somatosensory
stimulation beta activity is evoked together with gamma
and alpha activity (Chen & Herrmann, 2001). It has been
assumed that beta oscillations are induced by faster
gamma oscillations (Haenschel et al., 2000) and maybe
they in turn induce slower alpha oscillations which would
explain the presence of all three frequencies in one
experimental paradigm. In addition to motor and sensory
processes beta oscillations also have been associated
with cognitive processes like memory rehearsal (Tallon-
Baudry et al., 2001).

1.5 Gamma oscillations

In recent years technical improvements have revealed

oscillations of ever higher frequencies which are evoked
by sensory stimulation. These go up to about 600 Hz
(Curio, 1999) where the theoretical limit of EEG activity
is assumed due to the temporal width of single action
potentials in the brain which range between 1 and 2 ms.
Among high-frequency oscillations gamma waves (30-80 Hz,
cf. Fig. 6) have received a considerable amount of
attention due to their important correlates with higher
brain functions (Engel et al., 2001). It has even been
assumed that they might be a neural correlate of
consciousness (Llinas & Ribary, 1993).

Figure 6. Evoked 40 Hz oscillations show a burt shortly after visual

stimulation.
 9

The correlates of processes that are most frequently

associated with gamma oscillations are binding phenomena
(Müller et al., 1997; Tallon et al., 1995; Tallon-Baudry
et al., 1996), perceiving meaningful objects (Keil et
al., 1999; Tallon-Baudry et al., 1997), and attention
(Tiitinen et al., 1993; Müller et al., 1998; Herrmann et
al., 1999; Herrmann & Mecklinger, 2001; Debener et al.,
2003).
Reviews related to the functional relevance of gamma
oscillations can be found in Başar-Eroglu et al. (1996b),
Tallon-Baudry & Bertrand (1999), Müller et al. (2000) and
Herrmann & Knight (2001).

2. WAVELET ANALYSIS

2.1 FREQUENCY ANALYSIS METHODS

In principle, every signal can be decomposed into

sinusoidal oscillations of different frequencies. Such
decomposition is usually computed using the Fourier
transform to quantify the oscillations that constitute
the signal (Dumermuth, 1977).
Several methods exist to exclusively extract
oscillations of a specific frequency from ERP data. Among
the most popular are filtering, Fourier transformation,
and wavelet analysis.
Fig. 7 shows the results of those three methods to
extract frequency information from an ERP. Left panel:
filtering two ERPs with a band pass filter (35-45 Hz)
shows a clear burst of 40 Hz activity around 100
milliseconds. This oscillatory activity is enhanced for
the dotted as compared to the solid condition. Middle
panel: Fourier spectrum analyses of the two ERPs. An
increase of activity for the dotted condition can be
noticed around 40 Hz. However, it is unclear at what
point in time this difference between conditions occurs.
Right panel: the absolute values of the wavelet
coefficients of the ERP are shown for a 40 Hz wavelet.
The difference between conditions is very prominent and
can be observed at every point in time due to the lack of
oscillations in the signal. The wavelet transform can be
 10

thought of as the envelope of the bandpass-filtered ERP.

The wavelet transform is advantageous over the FFT, since
the time course of frequency information can be observed.
While this is also true for the filtered signal, the
wavelet transform yields directly the amplitude and the
phase of the signal oscillations in the analyzed
frequency band when a complex wavelet function is used.
The wavelet amplitude has only positive values and does
not bear the problem that oscillations might cancel out
when averaging across multiple subjects (negative values
shown in Fig. 7 result from a baseline correction making
the wavelet transform a relative measure with respect to
the pre-stimulus interval). A review of using wavelets
for EEG analysis is given by Samar et al. (1999).

Figure 7. Three possibilities to extract frequency information from

ERP data: two 35-45 Hz filtered ERPs (left), two FFT spectra of the
ERPs (middle) and the wavelet transforms of the ERPs (right). Note
that only the filtered signal and the wavelet transform still
represent changes over time. The FFT spectra show the whole
frequency range but no temporal information.

2.2 THE WAVELET TRANSFORM

To compute a wavelet transform, the original signal

time series, x(t), is convolved with a scaled and
translated version of a mother wavelet function, Ψ(t).
The convolution leads to a new signal of wavelet
coefficients,

t −b
W xΨ (b, a ) = AΨ ⋅ ∫ Ψ *   ⋅ x(t ) ⋅ dt ,
 a 
 11

where Ψ* denotes the complex conjugation of the wavelet

function, b is the translation parameter, a is the
wavelet’s scaling parameter, and AΨ denotes a (wavelet-
specific) normalization parameter. The wavelet
coefficients quantify the similarity between the original
signal and the wavelet function at a specific scale a and
target latency b. Hence, the wavelet coefficients depend
on the choice of the mother wavelet function.
The mother wavelet is constructed in such a way that it
has zero mean and is localized in both time and frequency
space. This is in contrast to the Fourier transform where
the harmonic basis functions have a well determined
frequency but extend over the whole time axis. Due to its
localization properties the wavelet transform allows to
follow the time-course of component structures in the
signal. This feature is of crucial importance when
analyzing non-stationary signals but has to be paid for
with a reduced frequency resolution.
Another important feature of the wavelet transform is
its zooming property. When the scaling parameter, a, is
varied from high to low values the wavelet function,
ψ([t-b]/a), will be compressed. The corresponding wavelet
transform zooms from coarser (i.e. low-frequency) to
finer (i.e. high-frequency) signal structures.

In the case of Morlet’s wavelets, also referred to as

Gabor wavelets, the mother wavelet function is given by
the formula

Ψ (t ) = e jω 0t ⋅ e − t
2
/2
,

where j denotes the imaginary unit, (-1)1/2, and ω0 is

2π times the frequency of the unshifted and uncompressed
mother wavelet (if less than 6 cycles of a wavelet are
used a correction term e-ω0/2 has to be subtracted from
ejω0t to guarantee that the wavelet still has a mean value
of zero). Morlet wavelets are complex functions. Both
their real and imaginary part consist of a harmonic
oscillation windowed in time by a Gaussian envelope. This
is schematically illustrated in Fig. 8.
 12

Figure 8. Multiplying a sinusoidal function (a) and an envelope function

(b) results in a wavelet (c).

Using sinusoidal wavelets like the Morlet wavelet is

ideally suited for detecting sinusoidal EEG activity
since the wavelet transform is similar to detecting
whether the used wavelet is contained in the signal.
Other wavelets which are more spiky can be used for
detecting transient phenomena in EEG like epileptic
spikes (Schiff et al., 1994).

In the frequency domain, Morlet wavelets also have a

Gaussian shape around their modulation frequency, i.e.
the wavelet scale can be directly interpreted in terms of
a well-defined center frequency (the terms scale and
frequency will be used synonymously here). Hence, the
scaled, unshifted wavelet can be written as a function of
frequency, f:

Ψ (t , f ) = e j 2π f t ⋅ e − t
2
/ 2σ t2
,

where the standard deviation σt of the Gaussian temporal

envelope is reciprocally related to the frequency
(σt∼1/f) in order to retain the wavelet’s scaling
properties. By this scaling one obtaines the same number
of significant wavelet cycles, nco=6σtf, at all
frequencies. The standard deviation in the frequency
domain is given by σf=(2πσt)-1. It grows proportionally to
the modulation frequency, i.e. σf/f is constant. This
implies that the Morlet wavelet transform has a different
time and frequency resolution at each scale. If the
number of significant cycles of the wavelet is kept
constant it varies in temporal width as a function of
frequency, since the same number of cycles spread over a
longer time interval for lower frequencies. Therefore, at
high frequencies the temporal resolution of a wavelet is
better than at low frequencies. However, the inverse is
 13

true for the frequency resolution of the wavelet

transform. At low frequencies the wide temporal extention
of the wavelet results in a good frequency resolution,
since many time points are considered for the analysis.
At high frequencies, where the small width leaves fewer
time points, the frequency resolution decreases. This is
illustrated in Fig. 9.

Figure 9. Three Morlet wavelets (leftpanel, only real part shown) with
different central frequencies and the corresponding frequency spectra (right
panel). A low-frequency wavelet of 10 Hz is very broad in the time domain
but has a good frequency resolution picking up only activity from adjacent
frequencies in a wavelet analysis (left peak in frequency spectrum). A
wavelet with a frequency of 40 Hz is more localized in time but has a lower
frequency resolution picking up frequencies from a wider range in a wavelet
analysis (right peak in frequency spectrum).

Besides this general trade-off between temporal and

frequency resolution wavelets also allow to adjust their
temporal and frequency width for any given center
frequency. By using a wavelet with more cycles (i.e.
larger nco) the frequency resolution increases, since the
frequency can be determined via more time points – of
course the temporal resolution decreases at the same
time. Using fewer cycles has the opposite effect.

Convolutions with Morlet wavelets can be computed for

multiple frequencies in order to yield a time-frequency
(TF, cf. Fig. 16) representation of the analyzed signal,
x(t). Because the Morlet wavelet function is complex, the
wavelet transform, Wx(t,f), is also a complex function,
 14

which can be divided into its real part, ℜ{Wx}, and its
imaginary part, ℑ{Wx}. Alternatively, using the polar
notation, Wx=|Wx|exp{jθx}, the wavelet coefficients can be
described by an amplitude, |Wx|=[ℜ{Wx}2+[ℑ{Wx}2]1/2, and a
phase angle, θx(t,f)=tan-1[ℑ{Wx}/ℜ{Wx}].
A wavelet function can be thought of as a finite
impulse response filter. In this context, the real part,
ℜ{Wx}, of the Morlet wavelet transform represents a
bandpass-filtered signal, xf(t), while the imaginary
part, ℑ{Wx}, yields a 90-degree phase shifted signal
(Hilbert transform). The amplitude, |Wx(t,f)| corresponds
to the envelope of the filtered signal, xf(t). It
quantifies the instantaneous oscillatory strength of the
signal with respect to the analyzed frequency band. Fig.
10 shows a time-frequency representation of the signals
depicted in Figure 1. The TF representation has been
obtained by gray-scale coding of the wavelet amplitudes.
Positions on the horizontal axis correspond to different
latencies, while different wavelet center frequencies
have been mapped to the vertical axis.

Figure 10. Multiple convolutions can be mapped in a time frequency

representation. This is shown for the evoked gamma activity (top) of the
example in Figure 1, the sum of evoked and induced gamma activity (middle)
and isolated induced gamma activity (bottom). The induced activity has been
estimated by the difference of total and evoked activity.

In analogy to the Fourier power spetrum, the wavelet

power spectrum is defined as |Wx(t,f)|2. It is a measure
for the signal energy (signal variance) contained in the
time-frequency bin covered by the transform, centered
 15

around time point t and frequency f. The Wavelet

functions can be normalized prior to the convolution to
have unit energy at all scales. In this case, the wavelet
power spectra of an analyzed signal are then directly
comparable to each other across all scales. For the
Morlet wavelet transform this normalization is achieved
with the factor Aψ=σt-1/2π-1/4.

If, however, the wavelet transform should directly

yield the amplitude of the analyzed signal, a different
normalization factor needs to be used. The Morlet wavelet
transform is very similar to the Gabor transform
(windowed Fourier transform). The main difference is that
in the wavelet method the width of the data window is not
fixed but adapted to the analyzed frequency. In analogy
to the Gabor transform, the wavelet amplitude spectrum,
|Wx(t,f)|, yields the instantaneous amplitude of an
oscillation when the Gabor normalization factor Aψ is
used:

Aψ=σt-1(2/π)1/2

To represent phase-locked (evoked) activity in an ERP

experiment, the wavelet transform is computed on the
average over the single trials (i.e. on the ERP):

t −b 1 N
Evoked = AΨ ∫ Ψ *  ⋅
 a  N
∑ eeg (t ) dt
i =1
i

Note that absolute value (or absolute power) is

calculated. After calculating the evoked activity, the
frequency-specific baseline activity can be subtracted to
yield values that indicate oscillatory amplitude (or
power) relative to baseline.

When wavelet transforms are computed, the convolution

peaks at the same latency as the respective frequency
component in the raw data, although the peak width will
be smeared. Therefore, the baseline should be chosen to
precede the stimulation by half the width of the wavelet
(e.g. 75 milliseconds for six 25 millisecond cycles of a
 16

40 Hz wavelet) to avoid the temporal smearing of post-

stimulus activity into the baseline. To avoid distortions
by the rectangular window function that can result from
'cutting out' a single epoch from continuous raw data
(edge effects), the convolution should start and end one
half of the wavelet length before the baseline and after
the end of the assessed time interval, respectively.

The TF representation of the ERP contains only that

part of the activity that is phase-locked to the stimulus
onset. To compute the activity that is not phase-locked
to stimulus onset (and is therefore cancelled out in the
average), the total activity (sum of evoked and induced
activity) can be computed. To calculate the sum of all
activity at one frequency, the absolute values of the
wavelet transforms of the single trials are averaged,
which means that each single trial is at first
transformed and the absolute values (or alternatively the
power values) are averaged subsequently:

1 N
t −b
∑ A ∫ Ψ   ⋅ eegi (t ) dt .
*
Total = Ψ
N i =1 a 

The corresponding TF representation (sum) contains all

activity of one frequency that occurred after stimulus
onset, no matter whether it was phase-locked to the
stimulus or not (cf. Fig. 10). As above, the activity in
a pre-stimulus interval can be subtracted to obtain a
relative measure.

3. NECESSARY CONDITIONS FOR RECORDING OSCILLATIONS

The analysis of EEG frequencies requires some

precautions when data are recorded. These are discussed
next.

3.1 HARDWARE REQUIREMENTS

Two important parameters for the recording equipment

are critical to properly record oscillatory activity: (1)
 17

The sampling rate has to be set to a value that is at

least twice the highest frequency that should be analyzed
(four times is better and is required by some software).
For example, if gamma activity up to 80 Hz shall be
analyzed, a minimum sampling rate of 160 Hz is needed and
320 is recommended. (2) The low pass filter needs to be
set to a value higher than the highest frequency that
should be analyzed. The low pass filter is usually
integrated in the analog amplifier to prevent aliasing
errors when digitizing analog data. This step is
sometimes overlooked when trying to record high frequency
oscillations for the first time.

3.2 ARTIFACT REJECTION

All artifacts that contaminate traditional ERP averages

should be excluded from frequency analysis as well. In
addition, there are several specific artifact conditions
that are especially crucial when oscillatory activity is
analyzed.

When analyzing alpha activity subjects should keep

their eyes open even if they have no visual task to
perform. When they close their eyes strong alpha
oscillations will appear in the EEG which show no
correlation with the cognitive task and contaminate the
analysis.

A potential confound of human gamma activity is

electromyography (EMG). If subjects sit uncomfortably or
chew during an EEG session and innervate their muscles,
the EEG electrodes will record EMG activity. This high
frequency muscle-related activity (30-80 Hz) can be
mistaken for gamma EEG activity. Therefore, all epochs
that are subsequently averaged should be visually
evaluated for the presence of such EMG artifacts, which
should then be excluded from further analysis.
 18

Figure 11. Clean EEG data and its frequency spectrum (left) and an epoch
with EMG contamination leading to frequency peaks around 40 Hz.

Figure 11 shows ten seconds of clean EEG and the

corresponding frequency spectrum with a 0 Hz and a 12 Hz
alpha peak (left). EMG activity can easily be detected in
the time domain (right) but may be mistaken for gamma
activity in the spectrum.

3.3 STIMULUS SIZE

Exogenous parameters like physical stimulus properties

are known to influence the amplitude of sensory evoked
potentials. The same dependence upon exogenous parameters
can be observed for oscillatory EEG activity. Especially
for low-amplitude activity in the gamma range it is
crucial to present stimuli of sufficient size in order to
evoke reliable responses. Cognitive differences between
experimental conditions can only be observed when the
amplitude is sufficiently high.

Fig. 12 shows how the evoked gamma response depends

upon stimulus size. Large (9° vis. angle) and medium (5°
vis. angle) stimuli evoked gamma peaks of approx. 0.4 µV
over occipital cortex which clearly differ from baseline
activity. Small (1° vis. angle) stimuli, however, evoke
only weak gamma responses which are only twice the
amplitude of the baseline noise. The first peak of gamma
 19

activity is due to the onset of the visual stimuli and

the second one due to their offset.

Figure 12. Evoked gamma responses in response to visual stimuli of

different size. Large (solid) and medium (dotted) stimuli evoke
strong gamma peaks while small (dashed) stimuli evoke only weak
responses.

3.4 STIMULUS DURATION

Due to the fact that onset as well as offset of a

stimulus evoke significant gamma bursts the duration of a
stimulus plays an important role for the observed pattern
of oscillations.

Figure 13. Stimuli of three different durations evoke

approximately the same onset peak of gamma activity around 100 ms
but different offset peaks. Stimulus durations: 250 ms (solid), 150
ms (dotted), and 50 ms (dashed).

When stimuli are sufficiently long in duration their

offset responses can clearly be differentiated from their
onset peaks. This is illustrated in Fig. 13 for stimuli
of 250 ms duration (solid) and 150 ms duration (dotted).
If, however, the duration is very short onset peak and
 20

offest peak mix into each other and can not be told apart
(50 ms duration, dashed). This is also true for ERP
analysis but often disregarded. When analyzing late ERP
componets they should not be contaminated by offset
responses.

3.5 STIMULUS ECCENTRICITY

Gamma oscillations are mainly generated over sensory

cortices. In case of the visual cortex the central visual
field is represented by more neurons than the peripheral
visual field. This leads to an influence of the
eccentricity of visual stimuli on the evoked gamma
response.

Figure 14. Influence of eccentricity on the evoked gamma response.

Centrally presented stimuli (dashed) evoke much larger responses
than stimuli of identical size and shape which are presented more
eccentrically (dotted and solid).

Fig. 14 shows the responses to three identical stimuli

at different eccentricity. A centrally presented stimulus
(dashed) leads to the largest response. Already at 4
degrees eccentricity (dotted) the response is much lower.
At an eccentricity of 8 degrees it is even lower.
Therefore, it is advantageous if central presentation can
be applied.
 21

3.6 AGE OF SUBJECTS

It has been shown that age influences the amplitude of

ERPs (Polich, 1997). The same is true for evoked
oscillations, especially in the gamma frequency range.
Already at an age of around 45 years the amplitude of the
response begins to decrease (Böttger et al., 2002).
Therefore, subjects must be chosen such that they
represent a homogeneous age. Otherwise age might be a
confound for cognitive parameters.

3.7 ANALYSIS OF AN ERP

When all technical aspects have been taken care of,

interesting new findings can be observed in the
oscillatory EEG responses. As Makeig et al. (2002) have
nicely demonstrated, an ERP (Fig. 15) and the frequency
representation of the ERP (Fig. 16) are two alternative
ways of investigating the EEG in response to experimental
stimulation.

Figure 15. An ERP in response to a visual target stimulus exhibiting

a series of components: P1, N1, P2, N2, and P3.

Fig. 16 shows the alternative representation of the ERP

in Fig. 15 as a time-freqeuncy plot. The early ERP
components are visible as high-frequency blobs in the
gamma and beta range and the later components are visible
 22

as two overlapping big blobs in the theta and delta

range. The earliest frequency component around 36 Hz has
the shortest duration and terminates around 100 ms after
stimulation. The subsequent oscillation around 18 Hz is
already more widely spread across time and lasts
aproximately until 150 ms after stimulation. The theta
wave of about 7 Hz remains active until around 350 ms and
a delta wave (approx. 3 Hz) can be observed up to 700 ms
post-stimulus. Such a shift from early high-frequency
components to later low-frequency components is a typical
finding (Chen & Herrmann, 2001; Haenschel et al, 2000).
Sometimes they reveal a frequency relation of 4:2:1,
suggesting underlying neural resonance circuits that
trigger each other (Herrmann, 2001).

Figure 16. The time-frequency representation of the ERP in Fig.

15.

4. WAVELET-BASED DYNAMIC INTERDEPENDENCE MEASURES

In the neuroscience community there has been growing

interest not only in the modularization of brain
functions (i.e. the functional specialization of local
brain areas) but also in the cooperation between
specialized and widely distributed areas which is a
prerequisite of higher cognitive functions and large-
scale integration. This cooperation requires a certain
degree of dynamic synchronization between the involved
 23

neuronal assemblies which in turn should be reflected in

the EEG as a dynamic interrelation between the measured
brain signals (von Stein et al., 1999; Schack et al.,
1999).
Classical interrelation measures such as Fourier-based
coherence and correlation depend on the stationarity of
the measured signals, which is rarely fulfilled with
concurrent brain signals. Recently, alternative tools
based on wavelet analysis have been developed and
successfully applied to EEG/MEG signals (e.g. Lachaux et
al., 2002). They allow to track the time-course of
coherence in non-stationary neuronal signals with good
temporal and frequency resolution.

4.1 Wavelet Coherence

Analogous to classical coherence, wavelet coherence is

defined as the cross-wavelet spectrum of two signals x
and y normalized by their corresponding auto-spectra:

W xy (t , f )
cohw (t , f ) = ,
W xx (t , f ) ⋅ W yy (t , f )

where Wxy(t,f) is the cross-wavelet spectrum (see below)

at latency t and frequency f, while Wxx and Wyy denote the
auto-spectra of x and y, respectively. Wavelet coherence
ranges between 0 and 1. It is a measure of the degree of
linear relationship between x and y in a specific time-
frequency bin. The instantaneous cross-wavelet spectrum
can be estimated from the product of the corresponding
univariate wavelet coefficients, Wxy (t , f ) = Wx (t , f ) ⋅ W y* (t , f ) . As
is the case with Fourier cross-spectra, this estimate is
inconsistent and has to be smoothed in an appropriate
way in order to improve reliability.

In an event-related potential paradigm, the smoothing

can be done across trials:

N
1
W x y (t , f ) = ∑W (t , f ) ⋅ W (t , f ) ,
x
k
y
k*

N k =1
 24

where N is the total number of trials and Wk means the

wavelet coefficient calculated from a signal recorded
during the kth trial. This method yields a coherence
measure which is very similar to the event-related
coherence introduced by Rappelsberger et al. (1994). It
does not require stationarity across time but is based on
the (also questionable) assumption of stationarity across
trials.
The wavelet coherence method as introduced by Lachaux
et al.(2002) estimates Wxy by averaging over a time period
around the target latency,
t + δ /2
1
Wx y (t , f ) = ∫δ W (τ , f ) ⋅W (τ , f ) dτ .
*

δ
x y
t − /2

Using Morlet wavelets, this approach corresponds to the

WOSA (Welch overlapping segment averaging) estimate of
the cross-spectrum (Welch, 1967), with the exception that
in the wavelet-based method the length of the smoothing
window can be varied in dependence upon the target
frequency, f. The smoothing window can be chosen to
contain the same fixed number of cycles, ncy, at all
frequencies, i.e. δ=ncy/f. Due to the flexible integration
window the wavelet coherence measure yields a more
consistent time-frequency resolution than the WOSA
method. Moreover, the same statistical performance of
the coherence estimator can be achieved at all
frequencies. Bias and variance of the wavelet coherence
estimator have been shown (Lachaux et al. , 2002) to
depend only on the number of independent data epochs
entering into the calculation of coherence. The number of
independent (non-overlapping) data segments is given by
the ratio ncy:nco, where nco denotes the number of
significant wavelet cycles. In oder to gain statistical
power, this ratio should be chosen as high as possible.
However, a large ncy (i.e. a large integration window)
diminishes the temporal resolution for measuring
coherence and decreases the probability of detecting
short-lasting coherent epochs. Therefore, ncy should be
adapted to the length of the coherent epochs which are
searched for, using larger integration windows for longer
epochs of coherency. On the other hand, as discussed in
 25

Sec. 2.2, the parameter nco has influence on the frequency

resolution of the wavelet transform and thus, on the
frequency selectivity of the coherence measure, which
decreases for low values of nco. Hence, nco must be
chosen in accordance to the frequency range of interest.
Typical values proposed by Lachaux et al. (2002) are nco≥3
for wide frequency bands (more than 10Hz) and values up
to 8-10 for narrow bands.

4.2 Phase Synchronization

Coherence does not separate the effects of covariance

of the amplitude waveforms and of the phases of two
oscillatory signals. The recently developed concept of
phase synchronization of chaotic (and/or noisy) systems
(Rosenblum et al., 1996) is more general. It implies the
appearance of a certain relationship between the phases
of oscillatory (sub)systems but does not impose
restrictions on their irregular amplitudes, which may
remain non-correlated. This concept is based on the
well-known fact that weak coupling first affects the
phases of oscillators, not their amplitudes. Hence, the
detection of phase synchronization should be sufficient
in order to reveal an interaction between two weakly
coupled (sub)systems.
With respect to brain signals, phase synchronization
in certain frequency bands is supposed to be a central
mechanism in neuronal information processing (Varela et
al., 2001). There is evidence, mostly from animal
experiments, that synchronization of neuronal activity
within sensory cortex is involved in feature binding
(Eckhorn et al., 1988; Gray et al., 1989). Transient
synchronization between physically distant brain areas
has also been reported (Roelfsema et al., 1997). It was
revealed as a possible mechanism for large-scale
integration, establishing a dynamic link between neural
assemblies by temporarily adjusting their discharge
frequencies. Recent experimental results from
intracranial and scalp recordings support the assumption
that magnitude and phase of brain signals might indeed be
involved in a different manner during a cognitive process
(Rodriguez et al., 1999; Bruns et al., 2000).
 26

4.2.1 Instantaneous Phase Difference

The parameter for measuring phase synchronization is

the relative phase angle between two oscillatory systems.
Neuroelectrical recordings are broad-band signals and
their phase cannot thoroughly be defined. Formally, one
could apply the analytic signal approach (Gabor, 1946)
and assign an instantaneous phase and an instantaneous
amplitude via the Hilbert transform. However, the Hilbert
phase and Hilbert amplitude have direct physical meaning
only for band-limited signals.
The Morlet wavelet transform acts as a bandpass filter
and, at the same time, yields separate values for the
instantaneous amplitude a(t,f) and the phase θ(t,f) of a
time-series signal at a specific frequency f. Thus, the
wavelet phases of two neuronal signals x and y can be
utilized to determine their instantaneous phase
difference in a given frequency band

∆θ (t , f ) = θ x (t , f ) − θ y (t , f )

and to establish a synchronization measure which

quantifies the coupling of phases independent from
amplitude effects. (Note that according to the above
equation the phase difference has to be calculated from
the unfolded univariate phase angles.)

Transient phase entrainment (phase locking) is

observed if the phase difference remains approximately
constant over some time period (typically hundreds of
milliseconds in the context of neurocogniton). Due to the
noisy and/or chaotic nature of neuronal signals, their
relative phase is usually not bounded even when there
exists some phase coupling between them. For weak noise,
the phase difference fluctuates around some mean phase
shift with occasional rapid phase jumps of ±2π. For
strong and unbound noise (i.e. Gaussian noise), these
phase slips occur irregularly. That means that phase
locking can be detected in a statistical sense only (Tass
et al., 1998; Rosenblum et al., 2001). One has to analyze
 27

the distribution of the relative phase angles on the unit

circle (wrapped to the interval [0,2π]). For independent
signals, this distribution will be close to uniform,
while synchrony shows up as the appearance of a
dominating peak.

Fig. 17 shows the phase difference between the 8 Hz

oscillations recorded at electrodes O1 and F9 after
visual stimulation. While the phase difference varies
over time before stimulation, it remains stable at a
value of approx. 0.83 π (2.6 rad) after stimulation for
about 250 ms. This can be seen by the plateau of the
curves.
Filtered EEG(µV)

Filtered EEG(µV)

Fig 17. Two time courses of 8 Hz event-related oscillations after

wavelet decomposition for two electrodes (top row). The phase
differences (bottom row) reveal that after stimulation (at 0 ms)
there is a stable phase relation between the two electrodes in both
trials from 0 to approx. 250 ms.
 28

4.2.2 Phase-Locking Statistics

Different synchronization measures have been proposed,

based e.g. on the Shannon entropy, the mutual
information, a stroboscopic approach or directional
statistics; see e.g. Tass et al. (1998) and Rosenblum et
al. (2001) for a review. According to directional
statistics (Mardia & Jupp, 2000), the coherence of an
angular distribution θi can be quantified by estimating
the phase-locking index (PLI),

2 2
PLI = e jθ = cosθ + sin θ = 1 − CV ,

where brackets denote the expectation operator and CV is

the circular variance CV. It is easily confirmed that the
PLI ranges between zero for uniformly scattered phases
and one in the case of perfect phase locking.

In a repeated-stimulus design, the PLI can be estimated

by averaging over trials (Lachaux et al., 1999):

∑ exp{j∆θ (t , f )} ,
N
1
PLI (t , f ) = k

N k =1

where N is the total number of trials and ∆θk represents

the instantaneous phase difference of the two brain
signals recorded during the kth trial. The bivariate PLI
measures the intertrial variability of the frequency-
specific relative phase of two brain signals at a given
target latency, i.e. it quantifies the stability of a
linear phase relationship across trials.
 29

Figure 18. TF representation of the bivariate intertrial PLI

estimated from two simultaneous EEG recordings (from an occipital
(O1) and a frontal scalp electrode (F9)) during visual stimulation
of a human subject (stimulus onset at time t=0; N=59 trials; nco=6
significant wavelet cycles). After stimulus onset, a transient
period of phase coherence is selectively detected in the alpha band.

Fig. 18 shows the TF representation of the intertrial

PLI for two EEG scalp recordings in a visual ERP
experiment. A prominent epoch of transient phase
coherence can be detected shortly after stimulus onset.
The phase locking confines selectively to the alpha
band.

Fig. 19 shows how the phase of an oscillation is

influenced by an experimental stimulus. Before visual
stimulation the phase differences between the 8 Hz
oscillations in electrodes O1 and F9 were almost randomly
distributed (left panel). After a visual stimulus
occurred most phase differences showed a value of 150
degrees. This indicates that the stimulus affects the
phase of the oscillations.

Figure 19. The phase distribution of the phase differences of the

8 Hz oscillations between electrodes O1 and F9. The numbers on the
circle denote the phase difference in degrees and the extention of a
wedge indicates how many single trials showed this phase difference.
 30

Before visual stimulation the phases were randomly distributed

(left). In a time interval from 50 to 300 ms after stimulation the
phases were clustered around a value of about 150 degrees (0.83 π,
right).

The PLI measure offers a good temporal resolution,

which is limited only by the width of the wavelet
function applied for the univariate phase estimation. Due
to the trade-off between temporal and frequency
resolution of the wavelet transform it might be
advantageous to prefilter the signals in a narrow
frequency band around the target frequency prior to
estimating their wavelet phases (Lachaux et al., 1999).
This is recommended especially when dealing with high
target frequencies in the gamma band, where the frequency
resolution of the wavelet transform is rather poor.

When the PLI is estimated from a finite number of

samples, as is always the case in real situations, a non-
zero PLI value will be measured even if the samples are
drawn from a uniform distribution. For N samples, the
expected PLI value (i.e. the bias) is N-1/2. The Rayleigh
test (Mardia & Jupp, 2000) can be applied in order to
assess significance of the detected phase locking
against the null hypothesis of a uniform distribution.
Because the sampling distribution of the statistics is
usually unknown for brain signals, Lachaux et al. (1999)
have proposed a Monte Carlo approach based on the
shuffling of trials. Surrogate values are computed from
the same signals x and y used for original PLI
estimation, except that the order of trials for y is
permuted before calculating the relative phases. That
means that the instantaneous phase difference is computed
from signals which have been recorded during different
trials and can thus be considered to be uncorrelated.
For each permutation, the maximum PLI value is measured
and compared against the original PLI value. The
percentage of surrogate values which are greater than
the original PLI at a given latency is called phase-
locking statistics (PLS) (Lachaux et al., 1999). For PLS
values which are smaller than a chosen significance level
the measured synchrony is considered significant. The
 31

number n of permutations needed for PLS calculation

depends on the chosen significance level p; for a one-
sided test it is given by n=1/p-1 (Theiler et al., 1992;
Schreiber & Schmitz, 2000).
Although PLS is a powerfull method it has its caveats.
Note that PLS fails to reject the null hypothesis in the
important case, when both univariate signals have
constant phases across trials and thus, the bivariate
phases are perfectly locked (Lachaux et al., 1999).

However, phase-locking statistics cannot be applied to

single trials (or averaged signals like ERPs). Moreover,
it fails to detect periods of synchrony which occur with
varying phase delay across trials or at jittering
latencies. As an alternative, Lachaux et al. (2000)
proposed the single-trial phase-locking index

t +δ / 2
1
S − PLI (t , f ) = ∫ exp{ j∆θ (τ , f )}dτ ,
δ t −δ / 2

also referred to as smoothed phase-locking index (S-PLI),

where averaging of the phase vectors proceeds over
adjacent time points. Time smoothing diminishes the
temporal resolution of the S-PLI measure. As with
wavelet coherence, the width of the smoothing window
should be adapted to the target frequency, and to the
expected length of the coherent epochs. Surrogate data
for a statistical test can be obtained by data
scrambling, i.e. by the permutation of the temporal order
of the samples in each of the two signals. For a review
of surrogate data methods see Paluš (1997) and Schreiber
& Schmitz (2000).

5. CONCLUSION

We reviewed the nature of oscillations in human EEG and

how to analyze them via wavelet analysis. We hope that we
 32

were able to convince the reader that oscillations are a

valuable approach on how to look at electrophysiological
data in addition to computing event-related potentials.
Our attempt to name a few of the many experiments
investigating oscillations in the human EEG was by no
means complete. However, the list gives a short overview
of the different frequency bands and may give the
interested reader a link to further articles. In addition
to the frequency bands which were explicitely mentioned
here there are various others ranging from oscillations
close to 0 Hz up to 600 Hz (Curio, 1999).

We also hope that the reader has learned new ways to

investigate oscillatory activity in his EEG data. At the
same time we tried to show the limitations and caveats of
the introduced methods. Wavelets are not the only way to
analyze oscillations – but they have some advantages over
other methods. Especially the possibility to investigate
the time course of an oscillation and to compute time-
frequency representations with variable resolutions are
among the strengths of wavelet analysis. Also the
analysis of phases and their temporal characteristics is
easy to achieve via wavelet analysis. However, care needs
to be taken with some of the parameters, like the number
of cycles which determines the frequency resolution as
well as the temporal resolution of the analysis.

The interpretation of significant synchronies between

brain signals that have been detected is not
straightforward. Especially when dealing with EEG scalp
recordings, spurious synchronies may arise from volume
conduction and/or reference effects. Volume conduction
leads to an artificially high synchrony especially
between adjacent electrodes since their recorded
neuronal populations overlap in space (Srinivasan et al.,
1998). The effect of the choice of a specific reference
electrode can hardly be predicted without precise
knowledge of the source locations and of the volume
conductor (Nunez et al., 1997). It may lead to an
increase as well as to a decrease of measured synchrony
between EEG recordings due to adding or removing a common
signal, respectively. To circumvent these problems the
spatial resolution of EEG recordings can be enhanced by
 33

deblurring techniques (Le & Gevins, 1993), scalp current

density (SCD) calculation (Pernier et al., 1988;
Lagerlund et al., 1995) or cortical imaging (Nunez et
al., 1994) prior to wavelet analysis. However, it was
argued (Biggins et al., 1991) that SCD estimation could
also introduce spurious synchronies due to spatial
interpolation inherent in the mathematical algorithm. A
challenging approach could be to combine inverse methods
and TF methods in order to reconstruct the sources of
oscillatory neuroelectrical activity.

Of course, the approaches which we focused on are not

the only ones. There are a number of other fruitful
applications of wavelets in neurophysiology and some of
them will probably gain more importance in the future.

One approach which is very promising is using a

discrete wavelet analysis for denoising. An averaged ERP
may be decomposed into wavelet coefficients by a discrete
wavelet analysis. Then one can determine which
coefficients yield significant activity at the
corresponding frequency. In a second step only these
significant coefficients are considered and others are
set to zero. Now the ERP is reconstructed from the
remaining wavelet coefficients. This procedure results in
filtering out other frequencies which are considered
noise for the cognitive task (e.g. Wang et al., 1998;
Quiroga & Garcia, 2003).

REFERENCES
Adrian, E. (1942). Olfactory reactions in the brain of the hedgehog.
Journal of Physiology (London), 100, 459-473.
Başar, E., Başar-Eroglu, C., Karakas, S., & Schürmann, M., (1999).
Oscillatory brain theory: A new trend in neuroscience. IEEE
Engineering in Medicine and Biology, 18, 56-66.
Başar, E., Schürmann, M., Başar-Eroglu, C. & Karakas, S. (1997).
Alpha oscillations in brain functioning: an integrative theory.
International Journal of Psychophysiology, 26(1-3), 5-29.
Başar-Eroglu, C., Strüber, D., Schürmann, M., Stadler, M., & Başar,
E., (1996b). Gamma-band responses in the brain: A short review of
psychophysiological correlates and functional significance.
International Journal of Psychophysiology, 24, 101-112.
 34

Başar-Eroglu, C., Başar, E., Demiralp, T. & Schürmann, M. (1992).

P300-response: possible psychophysiological correlates in delta
and theta frequency channels. A review. International Journal of
Psychophysiology, 13(2), 161-179.
Biggins, C.A., Fein, G., Raz, J., & Amir, A. (1991). Artifactually
high coherences result from using spherical spline computation of
scalp current density. Electroencephalography and Clinical
Neurophysiology, 79, 413-419.
Böttger, D., Herrmann, C.S. & von Cramon, D.Y. (2002). Amplitude
differences of evoked alpha and gamma oscillations in two
different age groups. International Journal of Psychophysiology,
45(3), 245-251.
Brandt, M.E. (1997). Visual and auditory evoked phase resetting of
the alpha EEG. International Journal of Psychophysiology, 26, (1-
3), 319-340.
Berger, H. (1929). Über das Elektrenkephalogramm des Menschen.
Archiv für Psychiatrie und Nervenkrankheiten, 87, 527-570.
Bruns, A., Eckhorn, R., Jokeit, H., & Ebner, A. (2000). Amplitude
envelope correlation detects coupling among incoherent brain
signals. Cognitive Neuroscience, 11(7), 1509-1514
Chen, A.C.N. & Herrmann, C.S. (2001). Perception of pain coincides
with the spatial expansion of human EEG dynamics. Neuroscience
Letters, 297(3), 183-186.
Curio, G. (1999). High frequency (600 Hz) bursts of spike-like
activities generated in the human cerebral somatosensory system.
Electroencephalography and Clinical Neurophysiology, 49, 56-61.
Debener, S., Herrmann, C.S., Kranczioch, C., Gembris, D. & Engel,
A.K. (2003). Top-down attentional processing enhances auditory
evoked gamma band activity. Neuroreport, 14(5), 683-686.
Dumermuth, G. (1977). Fundamentals of spectral analysis in
electroencephalography. In: A. Remond (Ed.), EEG informatics. A
didactic review of methods and applications of EEG data processing
(pp. 83-105). Amsterdam: Elsevier.
Engel, A.K., Fries, P. & Singer, W. (2001) Dynamic predictions:
oscillations and synchrony in top-down processing. Nature Reviews
Neuroscience, 2(10), 704-716.
Eckhorn, R., Bauer, R., Jordan, W., Brosch, M., Kruse, W., Munk, M.,
& Reitboeck, H.J. (1988). Coherent oscillations: a mechanisms of
feature linking in the visual cortex? Multiple electrode and
correlation analyses in the cat. Biol. Cybern., 60, 121-130.
Fell, J., Klaver, P., Elfadil, H., Schaller, C, Elger, C.E. &
Fernandez, G. (2003). European Journal of Neuroscience, 17(5),
1082-1088.
Fell, J., Klaver, P., Lehnertz, K., Grunwald, T., Schaller, C.,
Elger, C.E., Fernandez, G. (2001). Human memory formation is
accompanied by rhinal-hippocampak coupling and decoupling. Nature
Neuroscience, 4, 1259-1264.
Gabor, D. (1946). Theory of communication, Journal of IEE Loondon,
93(3), 429-457.
 35

Galambos, R. (1992). A comparison of certain gamma band (40 Hz)

brain rhythms in cat and man. In E. Başar & T. Bullock (Eds.),
Induced rhythms in the brain (pp. 201-216). Boston: Birkhauser.
Gloor, P., Ball, G. & Schaul, N. (1977). Brain lesions that produce
delta waves in the EEG. Neurology, 27(4), 326-333.
Gray, C.M., König, P., Engel, A.K., & Singer, W. (1989). Oscillatory
responses in cat visual cortex exhibit intercolumnar
synchronization which reflects global stimulus properties. Nature,
338, 334-337.
Haenschel, C. (2000). Gamma and beta frequency oscillations in
response to novel auditory stimuli: A comparison of human
electroencephalogram (EEG) data with in vitro models. Proceedings
of the National Academy of Sciences USA, 97(13), 7645-7650.
Herrmann, C.S. (2001). Human EEG responses to 1-100 Hz flicker:
Resonance phenomena in visual cortex and their potential
correlation to cognitive phenomena. Experimental Brain Research,
137, 346-353.
Herrmann, C.S. & Knight, R.T. (2001). Mechanisms of human attention:
event-related potentials and oscillations. Neuroscience and
Biobehavioural Systems, 25(6), 465-476.
Herrmann, C.S., & Mecklinger, A. (2001). Gamma activity in human EEG
reflects attentional top-down processing. Visual Cognition, 8,
273-285.
Herrmann, C.S., Mecklinger, A., & Pfeiffer, E. (1999). Gamma
responses and ERPs in a visual classification task. Clinical
Neurophysiology, 110, 636 -642.
Jensen, O. & Tesche, C.D. (2002). Frontal theta activity in humans
increases with memory load in a working memory task, European
Journal of Neuroscience, 15(8), 1395-1399.
Keil, A., Müller, M.M., Ray, W., Gruber, T., & Elbert, T. (1999).
Human gamma band activity and perception of a gestalt. The Journal
of Neuroscience, 19, 7152-7162.
Klimesch, W., Doppelmayr, M., Russegger, H., Pachinger, T. &
Schwaiger, J. (1998a). Induced alpha band power changes in the
human EEG and attention. Neuroscience Letters, 13, 73-76.
Klimesch, W., Doppelmayr, M., Rohm, D. Pollhuber, D. & Stadler, W.
(1998b). Simultaneous desynchronization and synchronization of
different alpha responses in the human electroencephalogram: a
neglected paradox? Neurocience Letters, 244, 73-76.
Klimesch, W. (1997). EEG-alpha rhythms and memory processes.
International Journal of Psychophysiology, 26(1-3), 319-340.
Klimesch, W., Schimke, H. & Pfurtscheller, G. (1993). Alpha
frequency, cognitive load and memory performance, Brain
Topography, 5(3), 241-251.
Lachaux, J.-P., Rodriguez, E., Martinerie, J., & Varela, F.J.
(1999). Measuring phase synchrony in brain signals. Human Brain
Mapping, 8, 194-208
Lachaux, J.-P., Rodriguez, E., Le Van Quyen, M., Lutz, A.,
Martinerie, J., & Varela, F.J. (2000). Studying single-trials of
phase synchronous activity in the brain. International Journal of
Bifurcation and Chaos, 10(10), 2429-2439.
 36

Lachaux, J.-P., Lutz, A., Rudrauf, D., Cosmelli, D., Le Van Quyen,
M., Martinerie, J., & Varela, F. (2002). Estimating the time-
course of coherence between single-trial brain signals: an
introduction to wavelet coherence. Neurophysiologie Clinique, 32,
157-174.
Lagerlund, T.D., Sharbrough, F.W., Busacker, N.E., & Cicora, K.M.
(1995). Interelectrode coherences from nearest-neighbor and
spherical harmonic expansion computation of laplacian scalp
potential. Electroencephalography and Clinical Neurophysiology,
95, 178-188.
Le, J. & Gevins, A. (1993). Methods to reduce blur distortion from
EEGs using a realistic head model. IEEE Transactactions on
Biomedical Engineering, 40, 517-528.
Llinas, R. & Ribary, U. (1993). Coherent 40-Hz oscillation
charaecterizes dream state in humans. Proceeedings of the National
Academy of Sciences USA, 90(5), 2078-2081.
Lopes da Silva, F. (1991). Neural mechanisms underlying brain waves:
from neural membranes to networks. Electroencephalography and
Clinical Neurophysiology, 79(2), 81-93.
Makeig, S., Westerfield, M., Jung, T.P., Enghoff, S., Townsend, J.,
Courchesne, E. & Sejnowski, T.J. (2002) Dynamic brain sources of
visual evoked responses. Science, 295, 690-694.
Mardia, K.V. & Jupp, P.E. (2000). Directional Statistics.
Chichester: John Wiley & Sons Ltd.
Müller, M.M., Junghöfer, M., Elbert, T., & Rockstroh, B.
(1997).Visually induced gamma-band responses to coherent and
incoherent motion: A replication study. NeuroReport, 8, 2575-2579.
Müller, M.M., Teder-Sälejärvi, W., & Hillyard, S. (1998). The time
course of cortical facilitation during cued shifts of spatial
attention. Nature Neuroscience, 1, 631-634.
Müller, M.M., Gruber ,T., Keil, A., & Elbert, T. (2000). Modulation
of induced gamma band activity in the human EEG by attention and
visual processing. International Journal of Psychophysiology, 38,
283-299.
Neuper, C., Pfurtscheller G. (2001). Event-related dynamics of
cortical rhythms: frequency-specific features and functional
correlates. International Journal of Psychophysiology, 43, 41-58.
Niedermeyer, E. (1997). Alpha rhythms as physiological and abnormal
phenomena. International Journal of Psychophysiology, 26(1-3), 31-
49.
Nunez, P.L., Silberstein, R.B., Cadusch, P.J., Wijesinghe, R.S.,
Westdorp, A.F., & Srinivasan, R. (1994). A theoretical and
experimental study of high resolution EEG based on surface
Laplacians and corical imaging. Electroencephalography and
Clinical Neurophysiology, 90, 40-57.
Nunez, P.L., Srinivasan, R., Westdorp, A.F., Wijesinghe, R.S.,
Tucker, D.M., Silberstein, R.B., & Cadusch, P.J. (1997). EEG
coherency I: statistics, reference electrode, volume conduction,
Laplacians, cortical imaging, and interpolation at multiple
scales. Electroencphalography and Clinical Neurophysiology, 103,
499-515.
 37

Paluš, M. (1997). Detecting phase synchronization in noisy systems.

Physics Letters A, 235, 341-351.
Pernier, J., Perrin, F., Bertrand, O. (1988). Scalp current
densities: concept and properties. Electoencephalography and
Clinical Neurophysiology, 69, 385-389.
Polich, J. (1997). EEG and ERP assessment of normal aging.
Electroencephalography and Clinical Neurophysiology, 104, 224-256.
Quiroga, R.Q. & Garcia, H. (2003). Single-trial event-related
potentials with wavelet denoising, Clinical Neurophysiology, 114,
376-390.
Rappelsberger, P., Pfurtscheller, G., & Filz, O. (1994). Calculation
of event-related coherence – a new method to study short-lasting
coupling between brain areas. Brain Topography, 7, 121-127.
Rodriguez, E., George, N., Lachaux, J.-P., Martinerie, J., Renault,
B., & Varela, F.J. (1999). Perception’s shadow: long-distance
synchronization of human brain activity. Nature, 397, 430-433
Roelfsema, P.R., Engel, A.K., König, P., & Singer, W. (1997).
Visuomotor integration is associated with zero time-lag
synchronization among cortical areas. Nature 385, 157-161.
Rosenblum, M., Pikovsky, A., & Kurths, J. (1996). Phase
synchronization of chaotic oscillators. Physical Review Letters,
76, 1804-1807.
Rosenblum, M., Pikovsky, A., Kurths, J., Schaefer, C., & Tass, P.A.
(2001). Phase synchronization: from theory to data analysis. In F.
Moss & S. Gielen (Ed.), Handbook of Biological Physics, Volume 4,
Chapter 9(pp. 279-231). Elsevier Science.
Salmelin R. & Hari R. (1994). Spatiotemporal characteristics of
sensorimotor neuromagnetic rhythms related to thumb movement.
Neuroscience, 60, 537-550.
Samar, V.J., Bopardikar, A., Rao, R. & Swartz, K. (1999). Wavelet
analysis of neuroelectric waveforms. Brain and Language, 66, 7-60.
Sewards, T.V. & Sewards, M.A. (1999). Alpha-band oscillations in
visual cortex: part of the neural correlate of visual awareness?
International Journal of Psychophysiology, 33(2), 177-179.
Schack, B., Chen, A.C.N., Mescha, S. & Witte, H. (1999).
Instantaneous EEG coherence analysis during the Stroop task,
Clinical Neurophysiology, 110, 1410-1426.
Schiff, S.J., Aldroubi, A., Unser, M. & Sato, S. (1994). Fast
wavelet transformation of EEG, Electroencephalography and Clinical
Neurophysiology, 91(6), 442-455.
Schreiber, T. & Schmitz, A. (2000). Surrogate time series. Physica
D, 142, 346-382.
Srinivasan, R., Nunez, P.L., Silberstein, R.B. (1998). Spatial
filtering and neocortical dynamics: estimates of EEG coherence.
IEEE Transactions on Biomedical Engineering, 45, 814-826.
Steriade, M., Mc Cormick, D.A. & Sejnowski, T.J. (1993). Thalamo-
cortical oscillations in the sleeping and aroused brain, Science,
262, 679-685.
Steriade, M., Gloor, P., Llinás, R., Lopes da Silva, F. & Mesulam,
M., (1990). Basic mechanisms of cerebral rhythmic activities.
 38

Electroencephalography and Clinical Neurophysiology, 76(6), 481-

508.
Strüber, D. & Herrmann, C.S. (2002). MEG alpha activity decrease
reflects destabilization of multistable percepts. Cognitive Brain
Research, 14(3), 370-382.
Tallon, C., Bertrand, O., Bouchet, P., & Pernier, J. (1995). Gamma
range activity evoked by coherent visual stimuli in humans.
European Journal of Neuroscience, 7, 1285-1291.
Tallon-Baudry, C., Bertrand, O. & Fischer, C. (2001). Oscillatory
synchrony between human extrastriate areas during visual short-
term memory maintenance. Journal of Neuroscience, 21, 1-5.
Tallon-Baudry, C., & Bertrand, O. (1999). Oscillatory gamma activity
in humans and its role in object representation. Trends in
Cognitive Sciences, 3, 151-162.
Tallon-Baudry, C., Bertrand, O., Delpuech, C., & Pernier, J. (1996).
Stimulus specificity of phase-locked and non-phase-locked 40 Hz
visual responses in human. Journal of Neuroscience, 16, 4240-4249.
Tallon-Baudry, C., Bertrand, O., Delpuech, C., & Pernier, J. (1997).
Oscillatory gamma-band (30-70 Hz) activity induced by a visual
search task in humans. Journal of Neuroscience, 17, 722-734.
Tass, P., Rosenblum, M.G., Weule, J., Kueths, J., Pikovsky, A.,
Volkmann, J., Schnitzler, A., & Freund, H.J. (1998). Detection of
n:m phase locking from noisy data: application to
magnetoencephalography. Physical Review Letters, 81, 3291-3294.
Tesche, C.D. & Karhu, J. (2000). Theta oscillations index human
hippocampal activation during a working memory task. Proceedings
of the National Academy of Sciences USA, 97(2), 919-924.
Theiler, J., Eubank, S., Longtin, A., Galdrikian, B., & Farmer, J.D.
(1992). Testing for nonlinearity in time series: the method of
surrogate data. Physica D, 58, 77-94.
Tiitinen, H., Sinkkonen, J., Reinikainen, K., Alho, K., Lavikainen,
J., & Näätänen R. (1993). Selective attention enhances the
auditory 40-hz transient response in humans. Nature, 364, 59-60.
Varela, F., Lachaux, J.-P., Rodriguez, E., & Martinerie, J. (2001).
The brainweb : phase synchronization and large-scale integration,
Nature Reviews Neuroscience, 2, 229-239.
von Stein, A., Rappelsberger, P., Sarnthein, J & Petsche, H. (1999).
Synchronization between temporal and parietal cortex during
multimodal object processing in man, Cerebral Cortex, 9(2), 689-
692.
Wang, K., Begleiter, H. & Pojesz, B. (1998). Spatial enhancement of
event-related potentials using multiresolution analysis, Brain
Topography, 10(3), 191-200.
Welch, P.D. (1967). The use of fast Fourier transform for the
estimation of power spectra: a method based on time averaging over
short, modified periodogramms. IEEE Trans. Audio Electroacoust.,
AU-15, 70-73.
Yordanova, J., Kolev, V. & Polich, J. (2001). P300 and alpha event-
related desynchronization (ERD). Psychophysiology, 38, 143-152.
39