Chapter 2- Taken’s Theorm and Time Delayed Embedding

Since we have discussed phase space and its role in analyzing nonlinear systems, let’s discuss about Taken’s theorm.

In 1981, Floris Takens published the paper “Detecting Strange Attractors in Turbulence”, which introduced this concept. Takens demonstrated that when a system involves multiple interconnected variables driving its dynamics (i.e., multidimensional dynamics), and if we can only observe a single variable \(x\) from the system (i.e., measuring one dimension), then we can reconstruct the multidimensional dynamics of the system. This is achieved by plotting the observable \(x\) against itself at specific time intervals and delays (refer to Figure 1). The process begins with the measured values of the variable \(x\):

\[\mathbf{x} = (x_1, x_2, x_3, \ldots, x_n)\]

Let \(\mathbf{x}\) be a vector containing values \(x_1\) to \(x_n\), representing the time-series of the variable \(x\) sampled at regular intervals \(t_1, t_1 + \delta t, t_1 + 2\delta t, \ldots, t_1 + (n - 1)\delta t\). If we are aware of (or can estimate) the true dimension \(D\) of the dynamical system from which we have sampled \(x\), then we can form \(D\)-dimensional vectors in the following manner:

\[\boldsymbol{V_{1}} = (x_{1}, x_{1+\tau}, x_{1+2\tau}, \ldots, x_{1+(D-1)\tau})\]

The elements of \(\mathbf{V_{1}}\) are derived from the vector \(\mathbf{x}\), starting with \(x_{1}\) sampled at time \(t_{1}\), and then including values at later times, such as \(x_{1 + \tau}\) sampled at \(t_{1} + \tau\delta t\). Here, \(\tau\) represents the time-lag as the later times are delayed relative to \(t_{1}\) by an integer multiple of \(\tau\delta t\).

A similar vector \(\mathbf{V_{2}}\) can be constructed by starting with \(x_{2}\) sampled at \(t_{2} = t_{1} + \delta t\). In fact, we can construct \(n - (D - 1)\tau\) such vectors, which can then be arranged in a matrix.

\[\begin{split}\boldsymbol{V} = \begin{bmatrix} \boldsymbol{V_{1}} \\ \boldsymbol{V_{2}} \\ \vdots \\ \boldsymbol{V_{n-(D-1)\tau}} \end{bmatrix} = \begin{bmatrix} x_{1}, x_{1+\tau}, x_{1+2\tau}, \ldots, x_{1+(D-1)\tau} \\ x_{2}, x_{2+\tau}, x_{2+2\tau}, \ldots, x_{2+(D-1)\tau} \\ x_{3}, x_{3+\tau}, x_{3+2\tau}, \ldots, x_{3+(D-1)\tau} \\ \vdots \\ x_{n-(D-1)\tau}, x_{n-(D-2)\tau}, x_{n-(D-3)\tau}, \ldots, x_{n} \end{bmatrix}\end{split}\]

FInding \(\tau\) value

For practical purpose it is important to compute the appropriate value of the the delay(\(\tau\)) in the first place. For this we had a multidimensional time series in which we com- puted a multidimensional mutual information and used it’s first minima(and global minima,in case the first minima doesn’t exist) in a plot between time delay and mutual information.

Let the time series be \(x_{n}\) having length N The time delayed versions are given by:

\[x_{n}^{(0)} = x_{n}[1 : N-\tau]\]

\[x_{n}^{(\tau)} = x_{n}[\tau : N]\]

And the mutual information is given by:

\[I(x_{n}^{(0)}, x_{n}^{(\tau)}) = H(x_{n}^{(\tau)}) - [H(x_{n}^{(0)}, x_{n}^{(\tau)}) - H(x_{n}^{(0)})] = H(x_{n}^{(0)}) + H(x_{n}^{(\tau)}) - H(x_{n}^{(0)}, x_{n}^{(\tau)})\]

Where \(H\) denoted entropy.

We Implemented the Mutual information calculation as follows:

mutualinfo()

mutualinfo(X, Y, n, d)

Function to calculate mutual information between to time series

Parameters:

• x (ndarray) – double array of shape (n,d). Think of it as n points in a d dimensional space

• y (ndarray) – double array of shape (n,d). second time series

• n (int) – number of samples or observations in the time series

• d (int) – number of measurements or dimensions of the data

Returns:

MI (double) – mutual information between time series

References

• Shannon, Claude Elwood. “A mathematical theory of communication.” The Bell system technical journal 27.3 (1948): 379-423.

Note that, here in the source code, np.histogramdd is a function to compute d-dimensional histogram in numpy module. This function might lead to numpy core memory error if the input is of higher dimension

And then we compute the mutual information for a specific time delay by using the following function:

timedelayMI()

timedelayMI(u, n, d, tau)

Function to calculate mutual information between a time series and a delayed version of itself

Parameters:

• u (ndarray) – double array of shape (n,d). Think of it as n points in a d dimensional space

• n (int) – number of samples or observations in the time series

• d (int) – number of measurements or dimensions of the data

• tau (int) – amount of delay

Returns:

MI (double) – mutual information between u and u delayed by tau

References

• Shannon, Claude Elwood. “A mathematical theory of communication.” The Bell system technical journal 27.3 (1948): 379-423.

For understanding this function, let’s consider the case of a simpe sine wave, as depicted below:

from SMdRQA.RQA_functions import timedelayMI
import numpy as np
import matplotlib.pyplot as plt

angles = np.linspace(0,10*np.pi, 1000)

angles = np.reshape(angles,(len(angles),1))

sign = np.sin(angles)

plt.figure(figsize = (12,9))
plt.plot(range(len(sign)),sign,'r')
plt.xlabel('time')
plt.ylabel('signal')
plt.title('signal')
plt.show()

Now, let’s compute the mutual inormation for different value of \(\tau\) from 1 to n-1

MI = []
TAU = []
for tau in range(len(angles)-1):
  TAU.append(tau)
  MI.append(timedelayMI(sign,len(sign),1, tau))

plt.figure(figsize = (12,9))
plt.plot(TAU, MI)
plt.xlabel('$\\tau$')
plt.ylabel('mutual information')
plt.show()

We can see 10 peaks in this plot, but, it is not easy to understand what that codes for. For that we need to get the angular values corresponding to each of these \(\tau\) values.

angles_from_tau = np.array(TAU)*((10*np.pi)/1000)
angles_2pi = angles_from_tau % (2*np.pi)
plt.figure(figsize = (12,9))
plt.figure()
plt.plot(angles_2pi, MI,'b.')
plt.axvline(x = np.pi)
plt.xlabel('angle')
plt.ylabel('mutual information')
plt.show

We can see that the odd multiples of \(\frac{\pi}{2}\) gives the minumum value of mutual information, and this is easy to understand in the case of sine wave as a sin wave shufted by \(\frac{\pi}{2}\) or its multiples results in a cos wave. findtau function, in this context, finds the first minima of the mutual information vs \(\tau\) curve.

timedelayMI()

findtau(u, n, d, grp)

Function to calculate correct delay for estimating embedding dimension based on the first minima of the tau vs mutual information curve

Parameters:

• u (ndarray) – double array of shape (n,d). Think of it as n points in a d dimensional space

• n (int) – number of samples or observations in the time series

• d (int) – number of measurements or dimensions of the data

Returns:

tau (double) – optimal amount of delay for which mutual information reaches its first minima(and global minima, in case first minima doesn’t exist)