메뉴 건너뛰기




Volumn 36, Issue 9, 1987, Pages 4456-4462

Efficient algorithm for estimating the correlation dimension from a set of discrete points

Author keywords

[No Author keywords available]

Indexed keywords


EID: 0000818930     PISSN: 10502947     EISSN: None     Source Type: Journal    
DOI: 10.1103/PhysRevA.36.4456     Document Type: Article
Times cited : (227)

References (32)
  • 4
    • 35949021570 scopus 로고
    • ``Chaotic'' is defined by the property of nearby trajectories to diverge.
    • (1981) Rev. Mod. Phys. , vol.53 , pp. 655
    • Ott, E.1
  • 5
    • 0000954445 scopus 로고
    • A method of distinguishing stochastic and (infinitely differentiable) deterministic time series directly from the power spectrum is discussed by
    • (1987) Phys. Rev. A , vol.35 , pp. 2276
    • Sigeti, D.1    Horsthemke, W.2
  • 6
    • 84926837044 scopus 로고    scopus 로고
    • However, involves the asymptotic high-frequency behavior and this may not be available in a discrete time series. Also, does not measure the number of degrees of freedom in a deterministic system.
  • 13
    • 84926858778 scopus 로고    scopus 로고
    • The correlation dimension was discussed independently by F. Takens, Invariants Related to Dimension and Entropy, in Atas do 13degr. Colóqkio Brasiliero de Matemática, Rio de Janeiro, 1983 (unpublished).
  • 16
    • 84926793323 scopus 로고    scopus 로고
    • J. Theiler, Ph.D. thesis, California Institute of Technology, 1987.
  • 17
    • 84926858777 scopus 로고    scopus 로고
    • F. Takens, in Detecting Strange Attractors in Turbulence, Vol. 898 of Lecture Notes in Mathematics, edited by D. A. Rand and L.-S. Young (Springer-Verlag, Berlin, 1981).
  • 19
    • 0018999905 scopus 로고
    • A general discussion of related problems from a computer science viewpoint can be found in
    • (1980) Commun. ACM , vol.23 , pp. 214
    • Bentley, J.L.1
  • 20
    • 84926837042 scopus 로고    scopus 로고
    • We are indebted to J. D. Farmer for this reference.
  • 21
    • 84926858776 scopus 로고    scopus 로고
    • This assumes the Linf or ``maximum'' metric. For the L1 or ``taxicab'' metric, the appropriate inequality is r <= 2mr0; for the L2 or Euclidean metric, is r <= 2 sqrt m r0.
  • 22
    • 84926815309 scopus 로고    scopus 로고
    • In the ``lexicographic'' ordering, we say that a vec < b vec, where in coordinate notation a vec =(a1, a2,..., am) and b vec =(b1, b2,..., bm), if and only if ak< bk for some k <= m, and ai=biinssA i < k.
  • 23
    • 84926858775 scopus 로고    scopus 로고
    • On our IBM PC, we find τdist=(0.5 +0.2m) msec. In the Takens method, we have also to take a logarithm of all distances less than r0; these cost 4.3 msec each (with an 8087 floating point co-processor). Also τsearchapprox (1.5+0.6m) msec and τsort=(0.44+0.09m) msec, with some leveling off at large m. Finally, τread=6.7 msec. There is a small memory compiler option which cuts these times in half, but can only be used for N<5000.
  • 24
    • 84926815308 scopus 로고    scopus 로고
    • Takens (in Ref. 17) says the error bar on nu will scale as 1/ Dstar, where Dstar is the number of distances less than r0.
  • 25
    • 84926858774 scopus 로고    scopus 로고
    • Properly we should write B approx (2R/r0)d, where d is the ``capacity'' of the set. Though capacity and correlation dimension are not precisely the same thing, our approximations do not distinguish between them.
  • 26
    • 84926837041 scopus 로고    scopus 로고
    • What we really assume is that points are distributed uniformly among those boxes which cover the attractor; this presumes not only that the distribution of points is uniform over the attractor, but that the intersection of the attractor with boxes is uniform—in fact, there are often a lot of ``clipped edges.'' For details of this effect in another context, see W. E. Caswell and J. A. Yorke, in Dimensions and Entropies in Chaotic Systems, Vol. 32 of Springer Series in Synergetics, edited by G. Mayer-Kess (Springer-Verlag, Berlin, 1986), p. 123.
  • 27
    • 84926858773 scopus 로고    scopus 로고
    • The source code, further documentation, and executable files which run on an IBM PC (or compatible) are available from the author.
  • 28
    • 84926815307 scopus 로고    scopus 로고
    • In our Grassberger-Procaccia routine, we use 8N +4xr0 bytes, where x is the ``expansion factor'' which is multiplied by the floating point input before discretizing into integers. In our Takens maximum likelihood routine, we use 14N bytes since the floating point input is stored in double precision.
  • 30
    • 84926815306 scopus 로고    scopus 로고
    • The mapping is xi+1=yi+1-axi2;
  • 31
    • 84926837040 scopus 로고    scopus 로고
    • yi+1=bxi. Following Hénon, we use a=1.4, b=0.3.
  • 32
    • 84926837039 scopus 로고    scopus 로고
    • For low-dimensional attractor data, we find that the search time increases so slowly with m that O(N) distances can always be computed faster with the box-assisted algorithm than with the standard algorithm. But even in these cases, we find that the ``prism-assisted'' variant provides still further improvement.


* 이 정보는 Elsevier사의 SCOPUS DB에서 KISTI가 분석하여 추출한 것입니다.