Finding Stationary Subspaces in Multivariate Time Series: Week Ending 20 NOVEMBER 2009
Finding Stationary Subspaces in Multivariate Time Series: Week Ending 20 NOVEMBER 2009
Finding Stationary Subspaces in Multivariate Time Series: Week Ending 20 NOVEMBER 2009
week ending
20 NOVEMBER 2009
Identifying temporally invariant components in complex multivariate time series is key to understanding the underlying dynamical system and predict its future behavior. In this Letter, we propose a
novel technique, stationary subspace analysis (SSA), that decomposes a multivariate time series into its
stationary and nonstationary part. The method is based on two assumptions: (a) the observed signals are
linear superpositions of stationary and nonstationary sources; and (b) the nonstationarity is measurable in
the first two moments. We characterize theoretical and practical properties of SSA and study it in
simulations and cortical signals measured by electroencephalography. Here, SSA succeeds in finding
stationary components that lead to a significantly improved prediction accuracy and meaningful topographic maps which contribute to a better understanding of the underlying nonstationary brain processes.
DOI: 10.1103/PhysRevLett.103.214101
0031-9007=09=103(21)=214101(4)
non
stationary
stationary
stationary
nonstationary
214101-1
N
X
N
^
^ si :
^ s>
log det
i
i
Dd1
1
2
(4)
^ s k N 0; I
^ si ;
DKL N
i
i1
(2)
i1
Z> 0
@M M0
The components of this matrix gradient can be interpreted
as infinitesimal rotation angles; i.e., the entry in row i and
column j is the angle by which axis i will be rotated
towards axis j. Hence, the nonzero part Z 2 RdDd
corresponds to the rotations between coordinates of the
s- and n-space [8]. Note that the derivative with respect to
the rotations within the two spaces must vanish because
they do not change the solution. Thus, we can reduce the
number of variables to dD d. The optimization is then
carried out by conjugate gradient descend [9].
Error in degrees
LB0 R
week ending
20 NOVEMBER 2009
30
10
5
14 n-sources
w/ rand. mean
6 n-sources
w/ rand. mean
1
0.1
0.01
8 9 10 11 12 13 14 15 16 17 18 19 20
Number of epochs (N)
214101-2
60
30
10
1
2
5 6 7 8 910
Original EEG
50
Error in degrees
20
80
week ending
20 NOVEMBER 2009
15
20
40
60
80
Field Pattern
30
n-subspace
10
20
5
120
40
15
10
w/o with
SSA
0.5
0.67
0.83
1
1.17
1.33
Mean additional alpha power on test set
s-subspace
214101-3
week ending
20 NOVEMBER 2009
*buenau@cs.tu-berlin.de
meinecke@cs.tu-berlin.de
franz.kiraly@uni-ulm.de
x
klaus-robert.mueller@tu-berlin.de
[1] R. F. Engle and C. W. J. Granger, Econometrica 55, 251
(1987).
[2] It is straightforward to generalize SSA to nonsquare mixing matrices.
[3] See EPAPS Document No. E-PRLTAO-103-014948 for
supplementary material. For more information on EPAPS,
see http://www.aip.org/pubservs/epaps.html.
[4] This is also standard approach in ICA [5,6].
[5] A. Ziehe, P. Laskov, G. Nolte, and K.-R. Muller, J. Mach.
Learn. Res. 5, 777 (2004).
[6] A. Hyvarinen, J. Karhunen, and E. Oja, Independent
Component Analysis (Wiley, New York, 2001).
[7] E. T. Jaynes, Phys. Rev. 106, 620 (1957).
[8] A. Baker, Matrix Groups (Springer-Verlag, Berlin, 2002).
[9] M. D. Plumbley, Neurocomputing; Variable Star Bulletin
67, 161 (2005).
[10] The synthetic data is generated as follows: for each epoch,
the mean of the n-sources is drawn from N 0; 2I and the
covariance matrix of all sources 2 RDD is parametrized by Cholesky factors L> L with random entries
distributed uniformly on [0:5, 0.5]. The entries Lij with
i d correspond to the correlation of the s-sources and
are held constant between epochs.
[11] B. Blankertz, G. Dornhege, M. Krauledat, K.-R. Muller,
and G. Curio, NeuroImage 37, 539 (2007).
[12] In this experiment, the subject was instructed to imagine
movements of the right (left) hand for 3 s after seeing the
letter R (L) on a computer screen. The 272 trials of each
class were recorded with 49 EEG channels in a random
sequence. The data were split chronologically and evenly
into a calibration and test set.
[13] The applied state-of-the-art BCI method consists of a
dimensionality reduction to 6 dimensions by common
spatial patterns [14] followed by linear discriminant analysis [15].
[14] Z. J. Koles, Electroencephalogr Clin Neurophysiol 79, 440
(1991).
[15] G. Dornhege, J. R. Millan, T. Hinterberger, D. McFarland,
and K.-R. Muller, editors, Toward Brain-Computer
Interfacing (The MIT Press, Cambridge, MA, 2007).
[16] H. Berger, Archiv fur Psychiatrie und Nervenkrankheiten
87, 527 (1929).
[17] B. Blankertz, M. Kawanabe, R. Tomioka, F. Hohlefeld,
V. Nikulin, and K.-R. Muller, in Advances in Neural
Information Processing Systems 20, edited by J. Platt,
D. Koller, Y. Singer, and S. Roweis (MIT Press,
Cambridge, MA, 2008), pp. 113120.
[18] O. Yilmaz, Seismic Data Analysis (Society Of Exploration
Geophysicists, Tulsa, OK, 2001).
[19] R. K. Kaufmann and D. I. Stern, J. Geophys. Res. 107,
4012 (2002).
214101-4