A Quick Tour through State Space Reconstruction

This tutorial presents the general idea of attractor reconstruction with synthetic data, and proposes two exercises with simulated data. For these activities, you will need the most recent version of R and the rEDM package installed in your work computer. To start, open R and load the package by typing the following code in the R console:

Before you proceed please read through the essential concepts by reading the introduction and the first section (“Empirical Dynamic Modelling”) of rEDM’s tutorial that you can find here or in your local R installation with the command

A simple case: harmonic oscillator

Let’s get the grips with R language and to figure out how to reconstruct a state space with only one time series and its lags. For the first example, we use a very well-known system and its representation on state space, the harmonic oscillator:

\(\frac{dx}{dt}=y\)

\(\frac{dy}{dt}=-x\)

which has the solution \(x(t)=A \sin(t), \ y(t)=A \cos(t)\), so we can draw the state space of this system by laying down in two Cartesian axis its two state variables, \(x(t)\) and \(y(t)\). The attractors of this system are circles, that is, all possible trajectories of a solution form a circle in the state space:

And here are the time series for the two state variables:

A reconstruction of the state space

Now suppose that all the information you have about this system is a time series of one of the state variables, say \(X(t)\). Because \(Y(t)\) is missing, we are not able to construct a state space as we see above, but Takens theorem tells us that we can do a reconstruction of the state space that recovers the information of the original state space. This state space reconstruction creates an embedding of a shadow manifold of the atractor. So we can define the shadow manifold as a valid topological reconstruction of the attractor.

To do a state space reconstruction we create new variables by lagging the time series by multiples of a time interval \(\tau\). In this case we will use the interval between observations (\(\tau=0.2\) time units). The R commands below create a data table object (which is called a data frame in R) with the original time series and its lagged version. The handy function make_block by the rEDM authors does this job. Use this command to download the function code and load it in your R workspace (the function will be included in the rEDM package in future releases):

And now you can create a data frame with the original and lagged series of data:

The most important arguments of this function are max_lag and tau; check the help page of this function (help(make_tutorial)) for more details. Anyway, a quick look at the first and last lines of the data frame can help to understand the lagged variable structure created by make_block:

##   time         x       x_1
## 1  0.0 0.0000000        NA
## 2  0.2 0.1986693 0.0000000
## 3  0.4 0.3894183 0.1986693
## 4  0.6 0.5646425 0.3894183
## 5  0.8 0.7173561 0.5646425
## 6  1.0 0.8414710 0.7173561
##       time           x         x_1
## 2496 499.0  0.49099533  0.65428133
## 2497 499.2  0.30813490  0.49099533
## 2498 499.4  0.11299011  0.30813490
## 2499 499.6 -0.08665925  0.11299011
## 2500 499.8 -0.28285377 -0.08665925
## 2501 500.0 -0.46777181 -0.28285377

By plotting the time series as a function of their lagged versions we get the shadow manifold, which in is this bi-dimensional case is easy to see:

As expected, the shadow is similar to the attractor. In fact this similarity is well defined in topological terms. A key property of a valid shadow manifold is that it is a one-to-one map of the attractor.

But a shadow manifold will only be valid in this sense if it is embedded in the correct number of dimensions. There is an optimal number of dimensions to build a valid state space reconstruction, which we call the embedding dimension of an manifold. In this simple case we obviously need two dimensions. Takens’ Theorem gives us an upper limit - it guarantees that the number of dimensions is up to 2D + 1, where D is the dimensionality of the attractor. Finding the optimal embedding dimension is part of a vast topic called Embedology (Sauer et al. 1991), and the package rEDM provide some methods to do this. We will learn more abou this in the following tutorials of this course.

Noisy data

How robust is the attractor reconstruction to random noise? There are two main types of noise or error in data, which we call process errors and observational errors. The first one is about random changes of the state variables, like environmental fluctuations that affect the dynamic. The second type of error is related to measurement error in the data, for instance when we miss individuals of some species when counting them, or have limited precision (inherent to all measurement). Observational errors do not affect the dynamics, only the portrait we make of it.

In this section, we check how robust the attractor is to measurement error by adding simulated noise to the original time series. Use the commands below to simulate that the time series has independent measurement errors with constant variance. To do this, we add to each observation a value drawn from a Gaussian distribution with zero mean and standard deviation which is 1/12 of the standard deviation of the time series itself.

We thus proceed by creating a new data frame with the lagged variables with noise and by plotting the new shadow manifold:

In this case the shadow manifold still looks like the original shape of the attractor, despite the noise caused by measurement errors.

Increasing the noise

Measurement errors make attractors more complex by spreading points (and thus trajectories) over the state space. Hence, the trajectories may cross in the shadow manifold, which breaks down the one-to-one mapping to the original attractor. Of course this effect gets larger as we increase the measurement error. We now will investigate a bit more how this affects the attractor reconstruction. The commands below double the measurement error of the time series and plots the shadow manifold:

The attractor reconstruction looks worse now, but we can overcome this problem by adding more dimensions to the state space reconstruction. By doing this we unfold the shadow manifold, by unfolding crossing trajectories. We do this adding more axes, that correspond to the time series lagged by the further time steps (\(t + \tau, \ t + 2\tau, \ t + 3\tau \dots\)).

You can check with the interactive plot below how an additional dimension unfolds a bit more the shadow manifold. 1 In the following days we will learn some methods to find the optimal dimension of the state space reconstruction. By now the take-home message is: the more complex the attractor, or the larger the error (observational or not), the more dimensions you will need to make a good reconstruction.

Exercises

What is the effect of sampling interval?

The commands below simulate a new time series of the bi-dimensional harmonic oscillator without measurement error but with a much smaller interval between observations (0.01 time units).

This change of the sampling frequency affects the shape of the reconstruction:

How do you explain that? Which practical issues does this fact bring and how to solve them?

Hints

1. The command below adds a line \(X(t) = X(t + \tau)\) to the state space reconstruction, which can help you to understand the problem with small values of \(\tau\):

2. Check this site: (http://www.physics.emory.edu/faculty/weeks//research/tseries4.html)

Attractor reconstruction

In this exercise you have to reconstruct the state space of a time series in this, checking if a three-dimensional plot improve the reconstruction. To load the data in R run the command:

To take a look in the time series plot run:

Hints

1. Recall from the previous problem that the sampling interval changes the shape of the shadow manifold. Using models, you control this sampling frequency; usually not so with data. Then you can only vary \(\tau\), which is by how many time steps each variable is lagged. For details see the help page of the make_block function.

2. To create a simple 3D plot in R you can use the scatterplot3d package. The example below does this for one of the noisy time series we examined above:

If the package is not available in your local R installation you can install it with:

Learn more

  • Sugihara, G., May, R., Ye, H., Hsieh, C.H., Deyle, E., Fogarty, M. and Munch, S., 2012. Detecting causality in complex ecosystems. Science, 338(6106), pp.496-500.
  • DeAngelis, D.L. and Yurek, S., 2015. Equation-free modeling unravels the behavior of complex ecological systems. Proceedings of the National Academy of Sciences, 112(13), pp.3856-3857.
  • Takens Theorem, a video by Sugihara’s Lab.
  • Takens, Floris. Detecting strange attractors in turbulence. Lecture notes in mathematics 898.1 (1981): 366-381.
  • Deyle, E.R. and Sugihara, G., 2011. Generalized theorems for nonlinear state space reconstruction. PLoS One, 6(3), p.e18295.
  • Sauer, T., Yorke, J.A. and Casdagli, M., 1991. Embedology. Journal of statistical Physics, 65(3), pp.579-616.

Glossary

  • Attractor: the set of points to which the trajectories of a dynamical system converges over time.

  • Lagged variables plot: a plot of a time series as a function of the same time series lagged by multiples of a length of time \(\tau\). For instance, the axes of a three-dimensional lagged variable plot are the original time series \(X(t)\), the lagged time series \(X(t + \tau)\), and \(X(t + 2\tau)\).

  • State space reconstruction: the process of making a representation of a state space of interest that preserves some important properties. In the context of EDM, the reconsctructed state space corresponds to the lagged variables plot of a time series.

  • Shadow Manifold: the projection of the original manifold (which is an attractor in our case) into an Euclidean space. This space has a dimension (the embedding dimension) that ensures a one-to-one map to the original manifold.

  • Embedding dimension: for our purposes, the dimension of the Euclidean space used to build the shadow manifold. In the case of a state space reconstruction, the embedding dimension is the dimension of this space, that is, the number of axes of the lagged variables plot.

  1. To run the R code that creates this plot in your computer you need the plotly package.↩

Brenno Cabella, Paulo Inácio Prado, Renato Coutinho, Marina Rillo, Rafael Lopes, Roberto Kraenkel

ICTP SAIFR, School on Physics Applications in Biology, January 2018