Testing for general dynamical stationarity with a symbolic data compression technique

Physical Review E - Statistical Physics, Plasmas, Fluids, and Related Interdisciplinary Topics

Volumn 61, Issue 3, 2000, Pages 2563-2568

  Kennel, Matthew B ^a   Mees, Alistair I ^b  

^a UNIVERSITY OF CALIFORNIA   (United States)

^b ISAAC NEWTON INSTITUTE FOR MATHEMATICAL SCIENCES   (United Kingdom)

Author keywords

[No Author keywords available]

Indexed keywords

DYNAMICAL SYSTEMS;

DATA STREAM; EVOLUTION LAW; HYPOTHESIS TESTS; SERIAL CORRELATION; STATIONARITY; SYMBOLIC DATA;

DATA COMPRESSION;

EID: 0000281130     PISSN: 1063651X     EISSN: None     Source Type: Journal    
DOI: 10.1103/PhysRevE.61.2563     Document Type: Article

References (25)

1. A. N. Kolmogorov, Dokl. Akad. Nauk (SSSR) 119, 861 (1958)
2. Dokl. Akad. Nauk (SSSR)Y. Sinai, 124, 768 (1959).
3. A. M. Fraser, IEEE Trans. Inf. Theory 35, 245 (1989)
4. A. M. Fraser and H. L. Swinney, Phys. Rev. A 33, 1134 (1986).
5. The most common application has been for entropy estimation, usually via string matching methods inspired by Lempel-Ziv compression. A context method similar to the present one was presented in T. Schurmann and P. Grassberger, Chaos 6, 414 (1996).
6. M. B. Kennel, Phys. Rev. E 56, 316 (1997)
7. T. Schreiber, Phys. Rev. Lett. 78, 843 (1997)
8. A. Witt, J. Kurths, and A. Pikovsky, Phys. Rev. E 58, 1800 (1998).
9. T. Cover and J. Thomas, Elements of Information Theory (Wiley Interscience, New York, 1991).
10. C. E. Shannon, Bell Syst. Tech. J. 27, 379 (1948).
11. L. Mason, A. I. Mees, and K. Judd (unpublished);A. I. Mees and M. B. Kennel (unpublished).
12. J. Rissanen, IEEE Trans. Inf. Theory 29, 656 (1983)
13. S. Bunton, in A Percolating State Selector for Suffix-Tree Context Models, Proceedings of the Data Compression Conference, 1997, edited by J. A. Storer and M. Cohn (IEEE Computer Society Press, Los Alamitos, CA, 1997), pp. 32-41.
14. F. M. Willems, Y. M. Shtarkov, and T. J. Tjalkens, IEEE Trans. Inf. Theory 41, 653 (1995).
15. Y. M. Shtarkov, T. J. Tjalkens, and F. M. J. Willems, Probl. Inf. Transm. 33, 17 (1997). Although the Krischevsky-Trofimov (KT) estimator has nice analytically computable asymptotic properties, other estimators may be superior for finite length observed data. We explored a number of ansatzes used by the text compression community but none were consistently superior to the KT estimator with a variable “ballast” (Formula presented) optimization parameter: (Formula presented). Often (Formula presented) gave results close to the minimum code length.
16. K. Judd and A. I. Mees, Physica D (to be published).
17. The (Formula presented) analytics degrade for small bin counts. For those bins, we merge any symbols whose expected count—for either set one or two—is less than five, reducing the degrees of freedom appropriately. If, after merging, there remain fewer than two symbols passing this criterion, then this node is wholly excluded.
18. In this discrete case, the sum will encounter tables exactly as likely as the observed one (such as the observed table itself); the summed (Formula presented) for these tables is weighted by a uniform random deviate (Formula presented).
19. E. N. Lorenz, Tellus, Ser. A 36, 98 (1984). The model is (Formula presented) (Formula presented) (Formula presented) Each set was 5000 points long sampled every (Formula presented).
20. C. S. Daw, C. E. A. Finney, M. Vasudevan, N. A. van Goor, K. Nguyen, D. D. Bruns, E. J. Kostelich, C. Grebogi, E. Ott, and J. A. Yorke, Phys. Rev. Lett. 75, 2308 (1995).
21. K. E. Trenberth and T. J. Hoar, Geophys. Res. Lett. 23, 57 (1996).
22. D. E. Harrison and N. K. Larkin, Geophys. Res. Lett. 24, 1779 (1997).
23. X. Z. Tang, E. R. Tracy, A. D. Boozer, A. deBrauw, and R. Brown, Phys. Rev. E 51, 3871 (1995).
24. C. S. Daw, M. B. Kennel, C. E. A. Finney, and F. T. Connolly, Phys. Rev. E 57, 2811 (1998).
25. Unfortunately the state-selecting context methods that are amenable to rigorous proof of universality are slightly different from those that are most practical to implement and have the best performance, though most of the time the differences are small.