Functional modulation of power-law distribution in visual perception

Masanori Shimono,
1
Takashi Owaki,
1
Kaoru Amano,
2
Keiichi Kitajo,
3
and Tsunehiro Takeda
1
1
Laboratory for Biological Complex Systems, Department of Complexity Science and Engineering, Graduate School of Frontier Science,
The University of Tokyo, 5-1-5, Kashiwanoha, Kashiwa-shi, Chiba 277-8561, Japan
2
Sensory and Motor Research Group, Human and Information Science Laboratory, NTT Communication Science Laboratories,
3-1, Morinosato Wakamiya Atsugi-shi, Kanagawa 243-0198, Japan
3
Laboratory for Dynamics of Emergent Intelligence, RIKEN, Brain Science Institute, 2-1, Hirosawa, Wako, Saitama 351-0198, Japan
Received 1 September 2006; revised manuscript received 6 January 2007; published 4 May 2007
Neuronal activities have recently been reported to exhibit power-law scaling behavior. However, it has not
been demonstrated that the power-law component can play an important role in human perceptual functions.
Here, we demonstrate that the power spectrum of magnetoencephalograph recordings of brain activity varies in
coordination with perception of subthreshold visual stimuli. We observed that perceptual performance could be
better explained by modulation of the power-law component than by modulation of the peak power in particu-
lar narrow frequency ranges. The results suggest that the brain operates in a state of self-organized criticality,
modulating the power spectral exponent of its activity to optimize its internal state for response to external
stimuli.
DOI: 10.1103/PhysRevE.75.051902 PACS numbers: 87.19.La, 05.45.Tp, 05.65.b, 87.19.Dd
I. INTRODUCTION
Synchronous neural oscillations are prominent features of
human brain activity and are thought to play a signicant
role in cognitive functions 1. When the functional roles of
such oscillations are being considered, their frequencies are
commonly categorized into bands , , , , Fig. 1a,
and most studies on the relationship between neural activities
and cognitive functions 2,3 have focused on band-specic
neural activity. However, neural activities in subjects at rest,
that is, when not responding to external stimuli, have been
frequently reported to exhibit a broad-band power-law distri-
bution with a scale-free property 46. Moreover, even in
subjects actively responding to external stimuli neuronal ac-
tivities can exhibit a power-law distribution Fig. 1a. The
relationship of such distributions to cognitive function, how-
ever, remains unknown. One possibility is that the observed
power-law distributions arise from a state of self-organized
criticality SOC in the brain, and that this state is maintained
by the brain in order to optimize cognitive function.
SOC 710, a concept in complexity physics that has
been studied in the context of a wide range of elds includ-
ing geophysics, economics and biology, also has been char-
acterized in the context of the brains neural oscillations.
SOC generally refers to a mechanism of slow energy accu-
mulation and fast energy redistribution that drives the system
toward a critical state wherein the spatiotemporal distribution
of elements in the system, interconnected nonlinearly, exhibit
power-law scaling behavior without ne-tuning. Recent stud-
ies have done well at accounting for the spectral exponent of
the power distribution, observed using noninvasive brain
measurement techniques such as Electroencephalography
EEG, magnetoencephalography MEG, or functional mag-
netic resonance imaging fMRI, using theoretical models of
the SOC state 8,11,12. However, in these studies, subjects
were at rest and were simply requested either to open or to
close their eyes.
Intriguingly, it is known that articial neural networks in
the SOC state are highly sensitive to weak external perturba-
tions 1315. This property could render SOC in the neural
networks of the brain an optimal state for detecting weak
stimuli. Therefore, we expected that the power-law compo-
nent PLC of the spectral distribution of brain activity could
be closely related to the perception of weak external stimuli.
Recently, it has been demonstrated that modulation in
band power in the MEG signal shows power-law scaling
behavior, and also that SOC plays an important role in de-
tecting weak somatosensory stimuli 1618. However, the
power-law behavior discussed in these previous studies was
limited to the modulation of the peak power only in a few
specic frequency bands. In contrast, we investigated the
possible functional roles of power-law scaling behavior
across a wide frequency range without being limited by a
priori assumptions regarding the functional roles of particu-
lar frequency bands in perception or cognition. Here we re-
port experimental results demonstrating that perceptual per-
formance can be explained better by variations in the PLC of
log (f)
10
(a)
(b)
log (f)
10

1 10 100
l
o
g
(
P
(
f
)
)
1
0

l
o
g
(
P
(
f
,
t
)
/
A
f
)
1
0
1 10 100
[Hz] [Hz]
FIG. 1. a A representative wavelet power spectrum of a mag-
netic eld on the surface of a human head, measured with MEG
black solid line. The subject was presented with weak ickering
visual stimuli. The power spectrum approximately obeys a power
law dashed line. Dotted lines demarcate the ve frequency bands
1.53.5 Hz, 3.56.5 Hz, 7.013 Hz, 1530 Hz,
30 Hz. b Residual component after subtraction of the power-
law component PLC from the raw wavelet power spectrum.
PHYSICAL REVIEW E 75, 051902 2007
1539-3755/2007/755/0519025 2007 The American Physical Society 051902-1
raw MEG data than by variations in peak power of particular
frequency components of those data.
We compared three models of modulations of the wavelet
power spectrum WPS. In model 1, perceptual performance
is assumed to correlate with peak amplitudes in particular
frequency ranges model 1. In model 2, perceptual perfor-
mance is assumed to correlate with two parameters, spectral
exponent and intensity A in a power-law distribution de-
scribed by Af

. These parameters are obtained by tting the

to each time point of Pf , t

af-
ter subtracting the PLC from the raw power spectrum on a
log scale model 3. We compared models 1 and 3 to conrm
the importance of the PLC model 2 in the explanation of
behavioral performance before directly comparing model 1
and model 2.
II. PREPARATION AND METHOD
Seven healthy males 2326 yr participated in the ex-
periment after giving informed consent 19. We investigated
their performance in detecting a ickering visual stimulus in
the dark. In this experiment, brain magnetic elds were re-
corded in a magnetically shielded room using a wholehead
MEG system PQ2440R; Yokogawa, Tokyo, Japan with 440
gradiometers. The sampling frequency of MEG was 500 Hz,
and the signals were bandpass ltered between 0.03 and
200 Hz. The size of the screen was 40 30. A xation
cross with a 3 diagonal was located at the center of the
screen and a target stimulus ickering at 10 Hz was pre-
sented in the lower left or lower right visual elds Figs. 2a
and 2c. Before the main experiment, we measured the de-
tection threshold step 1 and identied responsive sensors
for each stimulus position step 2.
In step 1, the target stimulus was presented at one of the
two possible target locations randomly and stimulus intensity
was changed to one of 6 levels in randomized order. Subjects
were asked to press one of the two buttons with their right
index nger when they detected the target in the left visual
eld, and with their right middle nger when they detected it
in the right visual eld. Each subject performed 200 trials.
The time course of the stimulus is shown in Fig. 2a. The
integral of the Gaussian probability density function was t-
ted to the averaged proportions of correct responses of the
six brightness levels by the least-squares method Fig. 2b.
The brightness of the target stimulus was adjusted to the
75% level of correct detections on the tted curve for each
subject in the main experiment. In step 2, subjects were pre-
sented with a ickering stimulus either in the lower right or
lower left visual elds, respectively, 350 times Fig. 2c.
From the resulting data we identied six planar gradiometers
that showed a strong evoked response at around 100 ms after
stimulus onset. The electrical current source in the primary
visual cortex was estimated by applying the steepest descent
method for the magnetic eld of all axial gradiometers Fig.
2d 20. This technique was used because the peak re-
sponse around 100 ms strongly correlates with contrast-
detection performance 21.
In the main experiment, we performed a similar experi-
ment as in step 1, setting the stimulus intensity at the
75%-correct detection threshold. We applied a wavelet trans-
form to extract information about the MEG power spectrum
from the six sensors chosen in step 2, using the Morlet wave-
let as the mother wavelet Eq. 1.
wt, f
0
=
1

exp

t
2
t
2

exp2if
0
, 1
where
t
is the variance along the time axis and f
0
represents
the center frequency.
t
is dened as ve times the size of
the period 1/ f
0
. To conrm that there was no articial cor-
relation over wide frequency ranges, we analyzed the WPS
obtained from harmonic articial data in 140 Hz with the
same wavelet and veried that the maximum value of vari-
ance was not over 2.5 Hz. We analysed trials remaining after
artifact rejection combined over all subjects. Artifacts, in-
cluding blinks, eye movements, and muscle movements were
rejected using an amplitude criterion 2.0 pT in the stimu-
lus period. Then we categorized all the trials into correct and
incorrect, and compared their WPSs. The total number of
trials analyzed was about 1200 17416 for each subject.
III. RESULTS OF WAVELET ANALYSES
Figure 3 shows the average WPS obtained across all sub-
jects and all trials from the left occipital sensors when the
stimulus was in the right visual eld. Figure 3a shows the
mean difference in the raw WPS between the correct and
incorrect trials. Similarly, Fig. 3b demonstrates the differ-
ence calculated for residual WPS derived by subtracting the
PLC from raw WPS for each trial. The PLC was obtained by
tting data points between 0.1 and 37 Hz by the method of
least squares. We chose this frequency range because the
power supply noise 60 Hz and its higher harmonics were
observed in the MEG signal when subjects were not in the
shielded room. We performed a two-tailed t test, with p val-
ues less than 0.05 considered signicant, to test for differ-
ences between correct and incorrect trials for measures of
WPS, such as the spectral exponent and raw WPS and re-
sidual WPS at each point in the time-frequency space. First,
we compared the results of statistical tests for the WPS based
on model 1 and those based on model 3 Figs. 3c and 3d.
There was an area where signicant differences were ob-
served in model 1 but not in model 3. This means that the
subtracted PLC includes a signicant difference between cor-
rect and incorrect trials.
Next, to examine how the differences in WPS between
correct and incorrect trials can be explained by model 2,
which assumes two parameters, spectral exponent and inten-
sity A, we focused on three time ranges A, B, and C, in which
the two groups of trials showed signicant differences Fig.
3c. In time window A, the WPS of the middle frequency
components and were signicantly different between
correct and incorrect trials. Since the spectral exponent
showed signicant differences between the correct and incor-
rect trials only in a short time period within window A Fig.
3e, the differences in WPS cannot be explained solely by
SHIMONO et al. PHYSICAL REVIEW E 75, 051902 2007
051902-2
modulation of the spectral exponent. To understand this phe-
nomenon, we analyzed the cooperative modulation of the
power in pairs of frequency bands e.g., and by calcu-
lating the product between averages of
Mf, t = log
10
P

cor
f, t/P

incor
f, t 2
across frequencies within the two frequency ranges at each
time point, where P

cor
f , t and P

incor
f , t is the average
power at frequency f on correct and incorrect trials, respec-
tively Fig. 4. Then, we calculated the products between
mean power in one frequency range and that in the other
frequency range at each time section, and tested whether the
data set of products in each time window were signicantly
different from zero two-tailed t test. In window A, the prod-
ucts had positive values in all pairs of frequency bands in
Subject3
Wavelet Transformation
(a)
(b)
(c)
C
o
r
r
e
c
t
R
a
t
e
1
0.75
0.5
0.25
stimulus intensity [cd/m ] ( ~10 )
0 2 4 6 8 10
-1
-0.5 0 4.0 5.5 [s]
-0.9 0 0.1
Prepare Stimulus
[s]
-5.0 7.0
~10
-14
3
2
1
0
-1
-2
-3
-4
4
L>R R>L
Answer Stimulus Prepare
[Tesla]
(d)
L > R
R > L
FIG. 2. Color a Time course of the stimulus presented to
identify the detection threshold. b Relationship between stimulus
amplitude and probability of correctly detecting the stimulus ob-
tained from one subject. Error bars correspond to SD. A probability
of 0.75 dotted line dened the detection threshold arrow. c
Time course of the stimulus for selection of sensors, which show
activity depending on the stimulus position. d The difference be-
tween magnetic elds activated by a stimulus presented in the lower
left and lower right visual elds. Positive negative areas are those
in which neural activity was stronger weaker when the stimulation
was in the left visual eld than when it was in the right visual eld.
FIG. 3. Color a Difference in mean WPS between correct
and incorrect trials. b The mean difference for data set across all
subjects in the residual WPS between the correct and incorrect tri-
als. The residual WPS are obtained by subtracting PLC from the
raw power spectrum for each trial. c Statistically signicant dif-
ference in mean WPS in a. Areas with signicant differences are
shown in white two-tailed t test with Bonferroni correction; p
0.005. d Statistically signicant difference in the residual WPS
in b p0.005. e Time courses of the averaged spectral expo-
nents on correct blue and incorrect trials red. Binary color map
below time courses shows the results of a two-tailed t test for spec-
tral exponents p0.005.
FIG. 4. Color Products of WPS modulations of frequency
bands a and , b and , c and , and d and in time
windows A, B, and C. Error bars correspond to SD. e Pairs of
frequency components shown in ad. Statistical test is per-
formed in each time-frequency window by comparing the averaged
WPS for correct trials and that for incorrect trials two-tailed t test,
*; p0.05, * * ; p0.01.
10 20 30 40
frequency [Hz]
40
30
20
10
f
r
e
q
u
e
n
c
y
[
H
z
]
-0.5
-1.0
-1.5
0
0.5
1.0
1.5
2.0
2.5
~10
-3
FIG. 5. Color The contour map for the average of the product
between Mf
1
, t and Mf
2
, t in time window C. Here, f
1
and f
2
are
two frequencies. Yellow and red areas indicate positive values, and
blue areas indicate negative values, of the average product.
FUNCTIONAL MODULATION OF POWER-LAW PHYSICAL REVIEW E 75, 051902 2007
051902-3
Figs. 4a4d. In other words, the WPSs in all frequency
bands in this window are different in the same direction
when comparing incorrect to correct trials. Moreover, in Fig.
3a, the direction is positive and the modulations do not
show remarkable peaks in specic frequency bands but
spread continuously over a wide frequency domain. Only the
and frequency bands showed a signicant difference
p0.005. When increasing the p value to 0.05, however,
the , and bands also showed signicant differences
data not shown.
In window B, all products of WPS in the various fre-
quency bands also had positive values. Figure 3a shows
that WPS in all frequency bands in window B are modulated
negatively from incorrect trials to correct ones. On the other
hand, in window C, the products between low frequency
bands such as and , and high frequency bands such as
and , showed positive values, while products between one
low frequency and one high frequency band showed negative
values Fig. 4. In addition, spectral exponents for correct
and incorrect trials also showed signicant differences in
window C Fig. 3e. To detect a xed frequency point in
the modulation of PLC in window C on the frequency axis
consecutively, we evaluated the modulation of PLC by the
product between Mf
1
, t and Mf
2
, t. Here f
1
and f
2
are two
specic frequencies. The averaged values of the products in
window C were then presented as a contour map Fig. 5.
In Fig. 5, in window C, products of only low frequency
bands and also those of only high frequency bands have
positive values, while those of low frequency with high fre-
quency bands have negative values, the sign reversing at
around 13 Hz. These results suggest that 13 Hz is a xed
point for the modulation of the spectral exponent. A part of
the modulation may correspond to the signicant differences
in the low frequency range in the above results based on
model 1 Fig. 3a. This interpretation is strongly supported
by two facts: 1 changing the signicance level of the p
value from 0.01 to 0.05 renders the differences in high fre-
quency ranges, such as and , signicant, and 2 the p
values for the entire data set of all points included in each
time-frequency window are as low as 0.005 data not
shown.
IV. DISCUSSION
In summary, two features stand out in three time windows
during which, over some frequency ranges, there were sig-
nicant differences in WPS between correct and incorrect
trials Fig. 3c. The rst feature is that the WPS in all
frequency ranges is modulated in the same direction, either
positive or negative in windows A and B. The second one is
that the WPS in low and high frequency ranges are modu-
lated in the opposite direction in window C. Furthermore, in
the rst case, WPS around the frequency boundaries, which
is about 13 Hz between the and bands, is modulated as
strongly as WPS within these frequency ranges. This clearly
shows that the modulation opposes the band-specic activity
model, model 1.
The most prominent feature of Fig. 3c is that specic
frequency components look activated. However, considering
that peak activities are eliminated by subtraction of PLC, and
the two features just described from the above analyses, the
difference between correct and incorrect behavioral perfor-
mance can be explained better by the modulation of widely
distributed PLC than by narrow peak activities in particular
frequency ranges. In similar previous experimental studies,
brain activities were measured only in resting states without
external stimulation 4,5. In this paper, however, we suggest
that the modulation of the power-law component affected
performance in detecting a weak visual stimulus. Our results
are consistent with the results of previous theoretical studies
showing that the neural networks in the SOC state show high
sensitivity to small environmental changes 1315. Our re-
sults suggest that the xed point for modulation of the spec-
tral exponent is around 13 Hz. However, further research
must be performed to understand the reason for this specic
value. In the present research, remarkable modulations of the
amplitude and of the spectral exponent in the power distri-
bution model were observed in different time windows.
However, in the general case, we must apply an appropriate
analytical method not only to differentiate the effects of these
components, but also to differentiate the effects of the PLC
from those of peak activity 22. Our results suggest that
PLC plays a major role in the brain activity reecting per-
ception of weak stimuli around detection threshold, making
the concept of SOC seem applicable to the maintenance of
optimal neural activity for perception and cognition.
ACKNOWLEDGMENTS
We are grateful to Professor Yoshiharu Yamamoto and Dr.
Kentaro Yamanaka for fruitful discussion, and to Professor
Lawrence M. Ward for detailed comments for this report.
1 Electroencephalography: Basic Principles, Clinical Applica-
tions, and Related Fields, edited by E. Niedermeyer and F. H.
Lopes da Silva Williams and Wilkins, Baltimore, 1998.
2 W. Klimesch, Brain Res. Rev. 29, 169 1999.
3 E. Basar, C. B-. Eroglu, S. Karakas, and M. Schurmann, Neu-
rosci. Lett. 259, 165 1999.
4 C. J. Stam and E. A. Bruin, Hum. Brain Mapp 22, 97 2004.
5 E. Novikov, A. Novikov, D. Shannahoff-Khalsa, B. Schwartz,
and J. Wright, Phys. Rev. E 56, R2387 1997.
6 V. G. Kiselev, K. R. Hahn, and D. P. Auer, Neuroimage 20,
1765 2003.
7 P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. Lett. 59, 381
1987.
8 L. deArcangelis, C. Perrone-Capano, and H. J. Herrmann,
Phys. Rev. Lett. 96, 028107 2006.
9 P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. A 38, 364
1988.
10 J. M. Beggs and D. Plenz, J. Neurosci. 23, 11167 2003.
11 P. A. Robinson, C. J. Rennie, J. J. Wright, H. Bahramali, E.
Gordon, and D. L. Rowe, Phys. Rev. E 63, 021903 2001.
SHIMONO et al. PHYSICAL REVIEW E 75, 051902 2007
051902-4
12 V. M. Eguiluz, D. R. Chialvo, G. A. Cecchi, M. Baliki, and A.
V. Apkarian, Phys. Rev. Lett. 94, 018102 2005.
13 P. Alstrom and D. Stassinopoulos, Phys. Rev. E 51, 5027
1995.
14 D. Stassinopoulos and P. Bak, Phys. Rev. E 51, 5033 1995.
15 D. R. Chialvo and P. Bak, Neuroscience 90, 1137 1999.
16 K. Linkenkaer-Hansen, V. V. Nikouline, J. M. Palva, and R. J.
Ilmoniemi, J. Neurosci. 21, 1370 2001.
17 K. Linkenkaer-Hansen, V. V. Nikouline, J. M. Palva, K. Kaila,
and R. J. Ilmoniemi, Eur. J. Neurosci. 19, 203 2004.
18 K. Linkenkaer-Hansen, V. V. Nikouline, J. M. Palva, R. J.
Ilmoniemi, and J. M. Palva, J. Neurosci. 24, 10186 2004.
19 I. Hashimono, R. Kakigi, T. Nagamine, N. Nakasato, H.
Shiraishi, and Y. Watanabe, J. Clin. Neurophysiol. 33, 4
2005.
20 The SQUID gradiometers of our MEG system consist of 300
axial gradiometers B
z
/ z and 703 planar gradiometers
B
x
/ z, B
y
/ z. The axes of B
z
/ z sensors circuits are in
vertical alignment with the scalp, and the axes of B
x
/ z and
B
y
/ z sensors circuits are horizontal to it. Therefore, the
magnetic eld map measured with B
z
/ z sensors shows a
negative peak and a positive peak for one dipole electrical
current source ECS running parallel to the scalp. However,
these peaks are not necessarily located near the ECS. On the
other hand, the root mean square rms value of B
x
/ z and
B
y
/ z sensors can detect the electrical current immediately
below them.
21 J. D. Haynes, G. Roth, M. Stadler, and H. J. Heinze,
J. Neurosci. 89, 2655 2003.
22 Y. Yamamoto and R. L. Hughson, Physica D 68, 250 1993.
FUNCTIONAL MODULATION OF POWER-LAW PHYSICAL REVIEW E 75, 051902 2007
051902-5