Professional Documents
Culture Documents
I: Fundamentals
Tim Sauer
July 2006
Goal
Find invariant set and dynamics
Attractor reconstruction
Example
Periodic orbit in R 3
E (x, y , z) = (x, y ) to R 2 may or may not unfold periodic orbit
To guarantee embedding of a periodic orbit, 3 independent
observations are need, generically.
Whitney
There is a C 1 - open and dense set of maps into R 2d+1 which
embed A.
Consider m = 2d + 1 independent measurements as a map
Embedding means individual states are distinguished by the
observations
Fractal attractor
-2
-2
-2
-2
log2(N(eps))
14
12
10
8
6
0 2 4 6 8
log2(1/eps)
1/3
2/3
dbox = lim
log 2i
log 2
=
0.63
i
log 3
log 3
1/5
4/5
Theorem
Let A be a compact subset of R k of box-counting dimension d.
Let n be an integer, n > 2d. Then for a dense set of smooth maps
F : R k R n ,
1. F is one-to-one on A
2. F is an immersion on each compact subset C of a smooth
manifold contained in A
Dense can be replaced by prevalent.
Intersection theory in R 1
1/3
2/3
v +1
2
[0, 1]
v +1
2
v + 1 = .02222022 . . . + .02002002 . . .
= c1 + c2
Therefore v + 1 c1 = c2 .
1/5
4/5
Hausdorff dimension
Example
There exists a set of Hausdorff dimension 0 in R m such that all
linear projections to R k , k < m, fail to be one-to-one.
(Ittai Kan)
Idea
Replace independent observations with time delays
Define H : R k R m by
x [h(x), h(f (x), h(f2 (x), . . . , h(f(m1) (x)]
Mathematical translation
Is
H(x) = [h(x), h(f (x)), . . . , h(f(m1) (x))]
one-to-one on A for generic h : R k R?
Example
Periodic orbit, period . For each point on the orbit,
xt = xt = xt2 = . . .
and the orbit is projected to a line segment. H cannot be 1-1.
Example
Periodic orbit, period 2 . The function
h(x) h(f (x))
has at least one zero crossing x0 on A. Then
h(x0 ) = h(f (x0 )) = h(f2 (x0 )) = h(f3 (x0 )) = . . .
so x0 and f (x0 ) are mapped together. F cannot be 1-1.
20
x(t)
10
0
-10
-20
10
20
30
x(t-2T)
x(t-T)
x(t)
Henon map
-2
-2
0
x
h(i)
0
-2
100
i
200
0
h(i-1)
-4
-2
-4
-4
-2
h(i)
Belousov-Zhabotinskii reaction
Br+ concen.
2
1.8
1.6
1.4
6
time
12
1.8
1.8
B(t-T)
B(t-T)
1.6
1.4
1.4
1.6
1.6
B(t)
1.8
1.4
1.4
1.8
1.8
1.6
B(t)
1.8
1.8
B(t-T)
B(t-T)
1.6
1.4
1.4
1.6
1.6
B(t)
1.8
1.4
1.4
1.6
B(t)
B(t-2T)
B(t-T)
B(t)
Applications - Analysis
Calculating system invariants from data stream
I
dimension
I
I
Lyapunov exponents
I
determinism tests
I
Applications
I
Noise reduction
Control of chaos
10000
0.08
2000
4000
6000
8000
10000
Noise reduction
Sample technique:
Embedding threshold estimator used in Fourier frame
10
10000
0.12
0.12
0.1
0.1
0.1
0.08
0.08
0.06
0.06
0.04
0.08
0.06
0.04
0.02
0.04
0.02
0.02
0
0.02
0
0.02
0.04
0.02
0.04
0.06
0.04
0.06
0.08
0.06
0.08
d scaled
0.1
Fig. 16.
2000
4000
6000
8000
10000
0.08
2000
4000
6000
8000
2000
4000
6000
8000
10000
10000
10000
0.08
0.04
0.02
0
0.02
0.04
An+1
normal
perturbed
As
An
Attractor Reconstruction
II: Extensions
Tim Sauer
July 2006
Extensions
10
-10
-20
-30
0
10
20
30
40
50
time
function of time. The lower trace shows the times at which spikes are generated
according to equation (2.1), with S (t) = (x(t) + 2)2 and = 60.
2. Nonautonomous Takens
g
D
fd
-
h
-
Goal
Reconstruct D X , recording only from X .
3. Stochastic Takens
f
D
fd
-
= D
= (. . . , d1 , d0 , d1 , . . .)
is shift map and xi+1 = f (xi , )
Goal
Reconstruct fibers over .
h
-
4. Driver reconstruction
:
X1
fd1
f1
XX
z
X
f2
X2
h1
h2
fd2
di+1 = g (di )
1
= f 1 (xi1 , di )
xi+1
2
= f 2 (xi2 , di )
xi+1
Goal
Reconstruct D, recording from X1 and X2 .
10
-10
-20
-30
0
10
20
30
40
50
time
function of time. The lower trace shows the times at which spikes are generated
according to equation (2.1), with S (t) = (x(t) + 2)2 and = 60.
Figure
2 Reconstruction
of Lorenz
dynamics from
a spike train. ThreeReconstruction
coordinates
are interspike
intervals
dimensional vectors of successive interspike time intervals are plotted, connected by line segments.
[Ti+1 The
Tiseries
, Ti+2oftime
Ti+1intervals
, Ti+3 used
T to]. make the recontruction was created using
integrate-and-re
model
(2.1),i+2
with an input signal
from the Lorenz attractor.
Ti+1
S(t)dt = .
Ti
= (y x)
= x y xz
= x + xy
Ti+1
S(t)dt = .
Ti
20
15
10
-5
-10
-15
-20
-5
(a)
10
15
(b)
the control
nuous signal
through an
in Fig. 1~a!,
is the meagenerates a
an integral
rom
Experimental control
Experimental control of a chaotic point process using interspike
intervals
G. M. Hall, S. Bahar, D. Gauthier at Duke University
G. MARTIN HALL, SONYA BAHAR, AND DANIEL J. GAUTHIER
a! Projection in phase space of the chaotic attractor of the electronic circuit shown in Fig. 1~b!. ~b! Reconstruction of the
g the interspike intervals.
16 MARCH 1998
Nonautonomous Takens
g
D
fd
-
h
-
Goal
Reconstruct D X , recording only from X .
Nonautonomous Takens
Nonautonomous Takens
Stochastic Takens
f
D
fd
-
= D
= (. . . , d1 , d0 , d1 , . . .)
is shift map and xi+1 = f (xi , )
Goal
Reconstruct fibers over .
h
-
Stochastic Takens
Stochastic Takens
Driver reconstruction
:
X1
fd1
f1
XX
z
X
f2
X2
h1
h2
fd2
di+1 = g (di )
1
= f 1 (xi1 , di )
xi+1
2
= f 2 (xi2 , di )
xi+1
Goal
Reconstruct D, recording from X1 and X2 .
Driver reconstruction
Algorithm based on Nonautonomous Takens Theorem uses
observed time series to identify states of driver as equivalence
classes.
Equivalence classes give semi-conjugacy with driver dynamics g .
VOLUME 93, N UMBER 19
FIG. 4.
for syste
struction
similar in
determine
with the
driver A.
Sauer
(2004) X1 and X2 were reconstructed from a
The attractors
Future directions