Krylov Subspace Method for Nonlinear Dynamical Systems with Random Noise

Authors: Yuka Hashimoto, Isao Ishikawa, Masahiro Ikeda, Yoichi Matsuo, Yoshinobu Kawahara

JMLR 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The empirical performance of our methods is investigated using synthetic and real-world healthcare data. Keywords: Nonlinear dynamical system, Transfer operator, Krylov subspace methods, Operator theory, Time-series data ... 7. Numerical Results We empirically evaluate the behavior of the proposed Krylov subspace methods in Subsection 7.1 then describe their application to anomaly detection using real-world time-series data in Subsection 7.2.
Researcher Affiliation Collaboration Yuka Hashimoto EMAIL NTT Network Technology Laboratories, NTT Corporation 3-9-11, Midori-cho, Musashinoshi, Tokyo, 180-8585, Japan / Graduate School of Science and Technology, Keio University 3-14-1, Hiyoshi, Kohoku, Yokohama, Kanagawa, 223-8522, Japan Isao Ishikawa EMAIL Faculty of Science, Ehime University 2-5, Bunkyo-cho, Matsuyama, Ehime, 790-8577, Japan / Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan Masahiro Ikeda EMAIL Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan / Faculty of Science and Technology, Keio University 3-14-1, Hiyoshi, Kohoku, Yokohama, Kanagawa, 223-8522, Japan Yoichi Matsuo EMAIL NTT Network Technology Laboratories, NTT Corporation 3-9-11, Midori-cho, Musashinoshi, Tokyo, 180-8585, Japan Yoshinobu Kawahara EMAIL Institute of Mathematics for Industry, Kyushu University 744, Motooka, Nishi-ku, Fukuoka, 819-0395, Japan / Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan
Pseudocode Yes Appendix C. Pseudo-codes of Arnoldi and shift-invert Arnoldi methods Let RS:T be the matrix composed of ri,t (S t T, 0 i S 1). The pseudo-codes for computing KS with the Arnoldi method and shift-invert Arnoldi method are shown in Algorithms 1 and 2, respectively. Algorithm 1 Arnoldi method for Perron-Frobenius operator K in an RKHS ... Algorithm 2 Shift-invert Arnoldi method for Perron-Frobenius operator K in an RKHS
Open Source Code No The paper does not provide explicit statements about open-source code availability or links to a code repository for the described methodology.
Open Datasets Yes We used electrocardiogram (ECG) data (Keogh et al., 2005).1 1. Available at http://www.cs.ucr.edu/~eamonn/discords/. We used chfdb chf01 275.txt , chfdb chf13 45590.txt and mitdbx mitdbx 108.txt in the experiment.
Dataset Splits Yes Using the synthetic data, K was first estimated, then the empirical abnormalities at,S,N were computed using all time-series data with t = 1601, N = 50, 75, 100 and S = 1, . . . , 12. We chose time points t = 1601 for evaluation because the estimation of K requires { x0, . . . , x N (S+1)} and 1601 > N (S + 1) for all N = 50, 75, 100 and S = 1, . . . , 12. ... We first computed KS,N with S = 10, N = 40 then computed the empirical abnormality ˆat,S,N for each t with the shift-invert Arnoldi method. ... KS,N was computed using the data { x0, . . . , x399}, and the empirical abnormalities ˆat,S,N at t = 430, 431, . . . were computed.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments. It only discusses the empirical evaluation of methods using synthetic and real-world data without mentioning computational resources.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., library names like PyTorch or scikit-learn with their versions) that were used to implement the methods or run experiments.
Experiment Setup Yes We used 100 synthetic time-series datasets { x0, . . . , x T } randomly generated by the following three dynamical systems: x0 = 2, xt+1 = 0.9995xt + 0.1ξt, (18) x0 = 0.5, xt+1 = 0.99xt cos(0.1xt) + ξt, (19) x0 = 0.1, x1 = x0 + 0.5x3 0 + ξ1, xt+1 = xt + 0.5(xt xt 1)3 + ξt, (20) where {ξt} is i.i.d with ξt N(0, 0.01). ... Using the synthetic data, K was first estimated, then the empirical abnormalities at,S,N were computed using all time-series data with t = 1601, N = 50, 75, 100 and S = 1, . . . , 12. ... The Gaussian kernel was used, and γ = 1+1i, where i denotes the imaginary unit, was set for the shift-invert Arnoldi method. ... We first computed KS,N with S = 10, N = 40 then computed the empirical abnormality ˆat,S,N for each t with the shift-invert Arnoldi method. The Laplacian kernel and γ = 1.25 were used. ... In this example, p was set as p = 15, 30, KS,N was computed using the data { x0, . . . , x399}, and the empirical abnormalities ˆat,S,N at t = 430, 431, . . . were computed. Also, the results obtained using long short-term memory (LSTM) (Malhotra et al., 2015) and autoregressive (AR) model (Takeuchi and Yamanishi, 2006) were evaluated for comparison. LSTM with 15, 30 time-series and 10 neurons, with the tanh activation function and the AR model xt+1 = Pp 1 i=0 cixt i + ξt with p = 15, 30 were used.