next up previous
Next: Phase space representation Up: Introduction Previous: Philosophy of the TISEAN

General computational issues

  The natural basis to formulate nonlinear time series algorithms from chaos theory is a multi-dimensional phase space, rather than the time or the frequency domain. It will be essential for the global dynamics in this phase space to be nonlinear in order to fulfill the constraints of non-triviality and boundedness. Only in particular cases this nonlinear structure will be easily representable by a global nonlinear function. Instead, all properties will be expressed in terms of local quantities, often by suitable global averages. All local information will be gained from neighborhood relations of various kinds from time series elements. Thus a recurrent computational issue will be that of defining local neighborhoods in phase space. Finding neighbors in multidimensional space is a common problem of computational geometry. Multidimensional tree structures are widely used and have attractive theoretical properties. Finding all neighbors in a set of N vectors takes O(logN) operations, thus the total operation count is O(NlogN). A fast alternative that is particularly efficient for relatively low dimensional structures embedded in multidimensional spaces is given by box-assisted neighbor search methods which can push the operation count down to O(N) under certain assumptions. Both approaches are reviewed in Ref. [20] with particular emphasis on time series applications. In the TISEAN project, fast neighbor search is done using a box-assisted approach, as described in Ref. [2].

No matter in what space dimension we are working, we will define candidates for nearest neighbors in two dimensions using a grid of evenly spaced boxes. With a grid of spacing tex2html_wrap_inline6495, all neighbors of a vector tex2html_wrap_inline6497 closer than epsilon must be located in the adjacent boxes. But not all points in the adjacent boxes are neighbors, they may be up to tex2html_wrap_inline6499 away in two dimensions and arbitrarily far in higher dimensions. Neighbors search is thus a two stage process. First the box-assisted data base has to be filled and then for each point a list of neighbors can be requested. There are a few instances where it is advisable to abandon the fast neighbor search strategy. One example is the program noise that does nonlinear noise filtering in a data stream. (This program is no longer maintained as part of TISEAN). It is supposed to start filtering soon after the first points have been recorded. Thus the neighbor data base cannot be constructed in the beginning. Another exception is if quite short (<500 points, say), high dimensional data are processed. Then the overhead for the neighbor search should be avoided and instead an optimized straight O(N²) method be used, like it is done in c2naive.

For portability, all programs expect time series data in column format represented by ASCII numbers. The column to be processed can be specified on the command line. Although somewhat wasteful for storing data, ASCII is the least common divisor between the different ways most software can store data. All parameters can be set by adding options to the command, which, in many programs, just replace the default values. Obviously, relying on default settings is particularly dangerous in such a subtle field. Since almost all routines can read from standard input and write to standard output, programs can be part of pipelines. For example they can be called as filters from inside graphics software or other software tools which are able to execute shell commands. Also, data conversion or compression can be done ``on the fly'' this way. The reader here realizes that we are speaking of UNIX or LINUX platforms which seems to be the most appropriate environment. It is however expected that most of the programs will be ported to other environments in the near future.

For those readers familiar with the programs published in Ref. [2] we should mention that these form the basis for a number of those TISEAN programs written in FORTRAN. The C programs, even if they do similar things, are fairly independent implementations. All C programs now use dynamic allocation of storage, for example.


next up previous
Next: Phase space representation Up: Introduction Previous: Philosophy of the TISEAN

Thomas Schreiber
Wed Jan 6 15:38:27 CET 1999