node24.html

时间序列工具

we can compute the centre of mass of these points in the neighborhood of <IMG WIDTH=15 HEIGHT=14 ALIGN=MIDDLE ALT="tex2html_wrap_inline7173" SRC="img79.gif">. Let us call it <IMG WIDTH=48 HEIGHT=24 ALIGN=MIDDLE ALT="tex2html_wrap_inline7221" SRC="img96.gif">. Under
fairly mild assumptions this point has twice the distance from the manifold
then <IMG WIDTH=36 HEIGHT=24 ALIGN=MIDDLE ALT="tex2html_wrap_inline7215" SRC="img95.gif">. To correct for the bias the origin of
the local coordinate system is set to the point: <IMG WIDTH=114 HEIGHT=24 ALIGN=MIDDLE ALT="tex2html_wrap_inline7225" SRC="img97.gif">.
<P>
The implementation and use of locally projective noise reduction as realized
in <a href="../docs_f/project.html">project</a> and <a href="../docs_c/ghkss.html">ghkss</a> is described in detail in Refs.&nbsp;[<A HREF="citation.html#on">61</A>, <A HREF="citation.html#buzug">62</A>].
Let us recall here the most important parameters that have to be set
individually for each time series. The embedding parameters are usually chosen
quite differently from other applications since considerable over-embedding may
lead to better noise averaging. Thus, the delay time is preferably set to unity
and the embedding dimension is chosen to provide embedding windows of
reasonable lengths. Only for highly oversampled data (like the
magneto-cardiogram, Fig.&nbsp;<A HREF="node25.html#figmcgnoise"><IMG  ALIGN=BOTTOM ALT="gif" SRC="icons/cross_ref_motif.gif"></A>, at about 1000 samples per cycle),
larger delays are necessary so that a substantial fraction of a cycle can be
covered without the need to work in prohibitively high dimensional spaces.
Next, one has to decide how many dimensions <I>q</I> to leave for the manifold
supposedly containing the attractor. The answer partly depends on the purpose
of the experiment. Rather brisk projections can be optimal in the sense of
lowest residual deviation from the true signal. Low rms error can however 
coexist with systematic distortions of the attractor structure. Thus for a
subsequent dimension calculation, a more conservative choice would be in order.
Remember however that points are only moved <EM>towards</EM> the local linear
subspace and too low a value of <I>q</I> does not do as much harm as may be though.
<P>
<P><blockquote><A NAME="5290">&#160;</A><IMG WIDTH=222 HEIGHT=493 ALIGN=BOTTOM ALT="figure1159" SRC="img98.gif"><BR>
<STRONG>Figure:</STRONG> <A NAME="fignoise_opt_raser">&#160;</A>
   Two-dimensional representation of the NMR Laser data  (top) and the 
   result of the <a href="../docs_c/ghkss.html">ghkss</a> algorithm (bottom) after three iterations.<BR>
</blockquote><P>
<P>
The noise amplitude to be removed can be selected to some degree by the choice
of the neighborhood size. In fact, nonlinear projective filtering can be seen
independently of the dynamical systems background as filtering by amplitude
rather than by frequency or shape. To allow for a clear separation of noise and
signal directions locally, neighborhoods should be at least as large as the
supposed noise level, rather larger. This of course competes with curvature
effects. For small initial noise levels, it is recommended to also specify a
minimal number of neighbors in order to permit stable linearizations.
Finally, we should remark that in successful cases most of the filtering is
done within the first one to three iterations. Going further is potentially
dangerous since further corrections may lead mainly to distortion.
One should watch the rms correction in each iteration and stop as soon as it
doesn't decrease substantially any more.
<P>
As an example for nonlinear noise reduction we treat the data obtained from an
NMR laser experiment&nbsp;[<A HREF="citation.html#raser">63</A>]. Enlargements of two-dimensional delay
representations of the data are shown in Fig.&nbsp;<A HREF="node24.html#fignoise_opt_raser"><IMG  ALIGN=BOTTOM ALT="gif" SRC="icons/cross_ref_motif.gif"></A>. The
upper panel shows the raw experimental data which contains about 1.1% of
noise. The lower panel was produced by applying three iterations of the noise
reduction scheme. The embedding dimension was <I>m</I>=7, the vectors were projected
down to two dimensions. The size of the local neighborhoods were chosen such
that at least 50 neighbors were found.  One clearly sees that the fractal
structure of the attractor is resolved fairly well.
<P>
<P><blockquote><A NAME="5376">&#160;</A><IMG WIDTH=236 HEIGHT=488 ALIGN=BOTTOM ALT="figure1231" SRC="img99.gif"><BR>
<STRONG>Figure:</STRONG> <A NAME="fignoise_opt_breath">&#160;</A>
   Two-dimensional representation of a pure Gaussian process (top) and the
   outcome of the <a href="../docs_c/ghkss.html">ghkss</a> algorithm (bottom) after 10 iterations. Projections
   from <I>m</I>=7 down to two dimensions were performed.<BR>
</blockquote><P>
<P>
The main assumption for this algorithm to work is that the data is well
approximated by a low-dimensional manifold. If this is not the case it is
unpredictable what results are created by the algorithm. In the absence of a
real manifold, the algorithm must pick statistical fluctuations and spuriously
interprets them as structure.  Figure&nbsp;<A HREF="node24.html#fignoise_opt_breath"><IMG  ALIGN=BOTTOM ALT="gif" SRC="icons/cross_ref_motif.gif"></A> shows a result
of the <a href="../docs_c/ghkss.html">ghkss</a> program for pure Gaussian noise. The upper panel shows a delay
representation of the original data, the lower shows the outcome of applying
the algorithm for 10 iterations. The structure created is purely artifical and
has nothing to do with structures in the original data. This means that if one
wants to apply one of the algorithms, one has to carefully study the results.
If the assumptions underlying the algorithms are not fulfilled in principle
anything can happen. One should note however, that the performance of the
program itself indicates such spurious behavior. For data which is indeed well
approximated by a lower dimensional manifold, the average corrections applied
should rapidly decrease with each successful iteration. This was the case with
the NMR laser data and in fact, the correction was so small after three
iteration that we stopped the procedure. For the white noise data, the
correction only decreased at a rate that corresponds to a general shrinking of
the point set, indicating a lack of convergence towards a genuine low
dimensional manifold. Below, we will give an example where an approximating
manifold is present without pure determinism. In that case, projecting onto the
manifold does reduce noise in a reasonable way. See Ref.&nbsp;[<A HREF="citation.html#danger">64</A>] for
material on the dangers of geometric filtering.
<P>
<HR><A NAME="tex2html316" HREF="node25.html"><IMG WIDTH=37 HEIGHT=24 ALIGN=BOTTOM ALT="next" SRC="icons/next_motif.gif"></A> <A NAME="tex2html314" HREF="node22.html"><IMG WIDTH=26 HEIGHT=24 ALIGN=BOTTOM ALT="up" SRC="icons/up_motif.gif"></A> <A NAME="tex2html308" HREF="node23.html"><IMG WIDTH=63 HEIGHT=24 ALIGN=BOTTOM ALT="previous" SRC="icons/previous_motif.gif"></A>   <BR>
<B> Next:</B> <A NAME="tex2html317" HREF="node25.html">Nonlinear noise reduction in </A>
<B>Up:</B> <A NAME="tex2html315" HREF="node22.html">Nonlinear noise reduction</A>
<B> Previous:</B> <A NAME="tex2html309" HREF="node23.html">Simple nonlinear noise reduction</A>
<P><ADDRESS>
<I>Thomas Schreiber <BR>
Wed Jan  6 15:38:27 CET 1999</I>
</ADDRESS>
</BODY>
</HTML>