Analysis of spatiotemporal signals: A method based on perturbation theory A. Hutt, 1, * C. Uhl, 1 and R. Friedrich 2 1 Max-Planck-Institute of Cognitive Neuroscience, Stephanstrasse 1a, 04103 Leipzig, Germany 2 Institute for Theoretical Physics, University of Stuttgart, Pfaffenwaldring 57, 70550 Stuttgart, Germany Received 18 December 1998 We present a method of analyzing spatiotemporal signals with respect to its underlying dynamics. The algorithm aims at the determination of spatial modes and a criterion for the number of interacting modes. Simultaneously, a way of filtering of nonorthogonal noise is shown. The method is discussed by examples of simulated stable fixpoints and the Lorenz attractor. S1063-651X9901908-X PACS numbers: 05.45.-a, 02.50.Sk I. INTRODUCTION In various scientific fields the analysis of spatiotemporal patterns emerging from complex systems plays an important role. An investigation of measured multidimensional data al- lows us to learn more about the internal dynamics of the system. It represents the basis for microscopic modeling of interactions in investigated systems e.g., 1. Some typical fields of application are chemical reactions 2, meteorology e.g., 3 and hydrodynamics 4or biological systems as analyzing electroencephalography EEGor magnetoen- cephalography MEGdata 5–7. Depending on the intended use, different kinds of data processing techniques can be applied. An often used method for linear data analysis is known as principal component analysis PCA8or Karhunen-Loe ` ve expansion. Spatial modes are calculated based on maximizing signal projections on these modes. It leads to orthogonal spatial and temporal modes and gives a measure for the contribution of each mode to the signal. Modes with a signal contribution above a certain threshold are considered as relevant, those below the threshold as irrelevant. However, this method fails to sepa- rate signal from noise, if signal and noise are not orthogonal on each other, and if noise parts contribute more than parts of the relevant signal to the data. Furthermore an estimation of the number of interacting modes depends on the choice of the threshold. Underlying dynamic structures are neglected by this linear data technique. A nonlinear approach aiming at extracting interacting modes and the underlying dynamics has been presented, e.g., in 9,10. However, the numerical effort of these nonlinear approaches is considerably high, especially with an increas- ing dimensionality of the underlying dynamical system. An estimation of the number of interacting modes is also still an open question. In this paper we will present a nonlinear technique based on linearperturbation theory, which focuses on internal deterministic dynamic patterns and extracts signal dynamics from noisy data sets. It improves PCA suspending the con- dition of orthogonality and allows an objective estimation of interacting spatial modes. Due to the linear equations to be solved, the method leads to a fast and robust algorithm. The perturbational approach is based on a ground state of the PCA modes, which represents the exact solution of mini- mizing a cost function leading to a complete orthogonal ba- sis. We introduce a perturbation by an additional term in the cost function for a determination of signal dynamics. Using a mathematical methodology similar to Hartree and Fock 11,12, we obtain dynamically coupled spatial modes. A criterion for the estimation of the number of interacting modes can be derived. We obtain the relevant signal sub- space independent of an orthogonality relation between sig- nal and noise, due to our special choice of a biorthogonal basis. II. METHOD A. Principal component analysis PCA A N-dimensional spatiotemporal signal can be described by a vector q( t ) of dimension N. In order to determine sig- nificant parts of the signal, one can decompose the signal into spatial and temporal modes v i and x i ( t ) by PCA. The properties are determined by a cost function V = i =1 N qt -qv i v i 2 q 2 + i , j =1 N ij v i v j - ij , V v k =0, 2.1 where ¯denotes time average and ij are Lagrange mul- tipliers to fulfill the orthogonality constraint. Standard cost functions of PCA lead to degenerated solu- tion spaces. To obtain the known equations of PCA directly, here one sums up the single errors to the signal and fixes the amplitudes as projections on the modes. This breaks the in- variance with respect to linear transformation. It leads to an eigenvalue problem Cv k = k v k , 2.2 with C=q( t ) q( t ) / q 2 , orthogonal spatial modes v k and amplitudes x k ( t ) =q( t ) v k , where denotes the dyadic product. They obey Eqs. 2.3, v k v l = kl , x k t x l t = k kl . 2.3*Electronic address: hutt@cns.mpg.de PHYSICAL REVIEW E AUGUST 1999 VOLUME 60, NUMBER 2 PRE 60 1063-651X/99/602/13509/$15.00 1350 © 1999 The American Physical Society