HR: 10:55h
AN: NG22A-03 [Abstracts]
TI: Computing Mutual Information Based Nonlinear Dependence Among Noisy and Finite Geophysical Time
Series
AU: * Khan, S
EM: skhan4@eng.usf.edu
AF: University of South Florida, Civil and Environmental Engineering
311 Kopp Engineering Building, Tampa, FL 33620
United States
AU: Bandyopadhyay, S
EM: sharba@jhu.edu
AF: Johns Hopkins University, Department of Biomedical Engineering
, Baltimore, MD 21205
United States
AU: Ganguly, A R
EM: gangulyar@ornl.gov
AF: Oak Ridge National Laboratory, Computational Sciences and Engineering
MS 6085; 1 Bethel Valley Road, Oak Ridge, TN 37831
United States
AB:
Linear correlation measures are widely used but may not be adequate for many geophysical problems, especially those that are
dominated by nonlinear dynamics and nonlinear interactions. Mutual information (MI), which originated in communications and
information theory, can be utilized to obtain measures of complete dependence, or "nonlinear correlation" (NLC). However,
while the computation of MI is conceptually straightforward when the full probability density function (PDF) is available,
there is no one best approach to compute MI or NLC from finite data sets. The state-of-the-art and emerging approaches used
to compute the MI or NLC range from methods that are based on ranking of variables (RANKS), kernel density estimation (KDE),
k-nearest neighbors (KNN) and what is called the "Edgeworth approximation" (Edgeworth). However, the emerging literature does
not point to a clear winner that outperforms the other methods for real data sets and all the methods can be extremely
sensitive to the presence of significant amount of noise. Thus, developing a better estimate often reduces to a better
judgmental choice of the model parameters like the number of kernels or neighbors, even though preliminary guidelines may be
available. The estimation problem, especially for uncertainty bounds, becomes even more difficult for time series data, where
approaches like bootstrapping need to be applied with care. This study implements the four approaches (RANKS, KDE, KNN,
Edgeworth) and investigates their relative performance, specifically for limited amount of noisy time series data, as a
function of the signal-to-noise ratios and the size of the data. The datasets range from simulations (e.g., time series
generated from the Lorenz system of equations contaminated with various noise levels) to real geophysical problems. The
relative performance of the methodologies, as well as the insights gained over and above linear correlation approaches, is
presented. The impacts of these insights on predictive modeling and scientific understanding are discussed.
Acknowledgment: Shiraj Khan would like to thank Professor Sunil Saigal at the University of South Florida. Auroop R Ganguly
gratefully acknowledges the Laboratory Directed Research and Development Program (SEED money funds) of the Oak Ridge National
Laboratory (ORNL), managed by UT-Battelle, LLC for the U.S. DOE under Contract No. DE-AC05-00OR22725.
DE: 3200 MATHEMATICAL GEOPHYSICS (0500, 4400, 7833)
DE: 4400 NONLINEAR GEOPHYSICS (3200, 6944, 7839)
SC: Nonlinear Geophysics [NG]
MN: Fall Meeting 2005