The nonlinearTseries package provides functionality for nonlinear time series analysis. This package permits the computation of the most-used nonlinear statistics/algorithms including generalized correlation dimension, information dimension, largest Lyapunov exponent, sample entropy and Recurrence Quantification Analysis (RQA), among others. Basic routines for surrogate data testing are also included. This vignette provides a brief overview of the most important routines contained in nonlinearTseries.
To explore the routines included in nonlinearTseries, we will study the famous Lorenz system. nonlinearTseries offers different routines for simulating the best well-known nonlinear systems:
suppressMessages(library('nonlinearTseries'))
library('plot3D')
# by default, the simulation creates a RGL plot of the system's phase space
lor = lorenz(do.plot = F)
# let's plot the phase space of the simulated lorenz system
scatter3D(lor$x, lor$y, lor$z,
main = "Lorenz's system phase space",
col = 1, type="o",cex = 0.3)
It must be noted that the lorenz
function returns the
simulated components of the system in a list
. Future
versions of the package will allow to obtain the same simulations as
ts
objects.
A complete list of the available functions for nonlinear systems
simulation can be found in the lorenz
help page
(?lorenz
command).
Usually, what we observe in a physical experiment is a single time series and not the complete phase space. For example, let’s assume that we have only measured the \(x\) component of the Lorenz system. Fortunately, we can still infer the properties of the phase space by constructing a set of vectors whose components are time delayed versions of the \(x\) signal \([x(t), x(t+\tau), ..., x(t + m\tau)]\) (This theoretical result is referred to as the Takens’ embedding theorem).
The nonlinearTseries package provides functions for estimating proper values of the embedding dimension \(m\) and the delay-parameter \(\tau\). First, the delay-parameter can be estimated by using the autocorrelation function or the average mutual information of the signal.
# suppose that we have only measured the x-component of the Lorenz system
lor.x = lor$x
old.par = par(mfrow = c(1, 2))
# tau-delay estimation based on the autocorrelation function
tau.acf = timeLag(lor.x, technique = "acf",
lag.max = 100, do.plot = T)
# tau-delay estimation based on the mutual information function
tau.ami = timeLag(lor.x, technique = "ami",
lag.max = 100, do.plot = T)
par(old.par)
Both techniques select a time-lag based on the behavior of the autocorrelation or the average mutual information function. Since the autocorrelation function is a linear statistic we usually obtain more appropriate values with the mutual information technique. Thus, for the remainder of this section, we will use the value obtained with this technique.
Once the time-lag parameter has been estimated, a proper embedding
dimension can be computed by using the well-known Cao’s algorithm (see
the documentation of the estimateEmbeddingDim
function for
references):
When applied to the Lorenz system, the Cao’s algorithm suggests the
use of an embedding dimension of 4. The final phase space reconstruction
can be obtained using the buildTakens
function:
tak = buildTakens(lor.x,embedding.dim = emb.dim, time.lag = tau.ami)
scatter3D(tak[,1], tak[,2], tak[,3],
main = "Lorenz's system reconstructed phase space",
col = 1, type="o",cex = 0.3)
Note that the reconstructed and the original phase space, although different, share similar topological features.
In practical applications, some of the best well-known nonlinear statistics (such as the Lyapunov exponent, the generalized correlation dimensions or the sample entropies) share a similar estimation process. This process could be summarized as follows:
plot
function can be used with all
the objects involved in the computation of these statistics.estimate
function.In the following sections we illustrate this procedure computing the correlation dimension, the sample entropy and the Lyapunov exponents of the Lorenz system.
The correlation dimension is a technique that measures the fractal
dimension of the phase space of a dynamical system. To verify that the
estimation of the correlation dimension does not depend on the embedding
dimension, we compute the correlation sums (corrDim
function) for several embedding dimensions. Once we have checked for the
existence of the linear regions in different embedding dimensions, we
obtain an estimation of the correlation dimension with the
estimate
function. This function allows to specify the
range in which the linear behavior appears
(regression.range
parameter) as well as the embedding
dimensions to be used for the estimation of the correlation dimension
(use.embeddings
parameter). The final estimation of the
correlation dimension is an average of the slopes obtained for each
embedding dimension.
cd = corrDim(lor.x,
min.embedding.dim = emb.dim,
max.embedding.dim = emb.dim + 5,
time.lag = tau.ami,
min.radius = 0.001, max.radius = 50,
n.points.radius = 40,
do.plot=FALSE)
plot(cd)
cd.est = estimate(cd, regression.range=c(0.75,3),
use.embeddings = 5:7)
cat("expected: 2.05 --- estimate: ",cd.est,"\n")
## expected: 2.05 --- estimate: 1.994026
The generalized correlation dimensions can also be computed with the
corrDim
function (by modifying the q
parameter). To estimate the information dimension,
nonlinearTseries provides the infDim
function (see ?infDim
for more information).
The sample entropy is a technique for measuring the unpredictability
of a time series. It is possible to use the correlation sums for
obtaining an estimation of the sample entropy of a time series. In this
case, the computations should yield a function with a clear plateau. The
value of this plateau is an estimation of the sample entropy. The next
chunk of code illustrates the procedure for estimating the sample
entropy from a previously computed corrDim
object.
se = sampleEntropy(cd, do.plot = F)
se.est = estimate(se, do.plot = F,
regression.range = c(8,15))
cat("Sample entropy estimate: ", mean(se.est), "\n")
## Sample entropy estimate: 0.1440437
One of the more important characteristics of a chaotic system is its
sensitivity to initial conditions. As a consequence of this sensitivity,
close trajectories diverge exponentially fast. The maximum Lyapunov
exponent measures the average rate of divergence of close trajectories
in the system. The maxLyapunov
function can be used for
computing this divergence through time. To define what is a close
trajectory we make use of the radius
parameter. After the
computation of the divergence rates we can get an estimate of the
maximum Lyapunov exponent by performing a linear regression
(estimate
function), just as we did with the correlation
dimension.
# get the sampling period of the lorenz simulation
# computing the differences of time (all differences should be equal)
sampling.period = diff(lor$time)[1]
ml = maxLyapunov(lor.x,
sampling.period=0.01,
min.embedding.dim = emb.dim,
max.embedding.dim = emb.dim + 3,
time.lag = tau.ami,
radius=1,
max.time.steps=1000,
do.plot=FALSE)
plot(ml,type="l", xlim = c(0,8))
## expected: 0.906 --- estimate: 1.005747
Although we have postponed its discussion until the end of this vignette, the first step before studying a system using nonlinear analysis techniques should be checking that the data shows indeed some degree of nonlinearity.
The preferred method for nonlinearity-test in literature is surrogate data testing. In surrogate data testing, a statistic \(\mu\) quantifying some nonlinear feature of the data is computed and compared with the resulting values for an ensemble of comparable linear processes.
nonlinearTseries includes basic functionality for surrogate data testing. The next example performs surrogate data testing by measuring the time asymmetry of the data and the surrogates (since linear stochastic processes are symmetric under time reversal, a deviation from the distribution of the surrogates would be a strong sign of nonlinearity). From the resulting figure, it is clear that our time series shows some degree of nonlinearity.
## Computing statistics
## Null Hypothesis: Data comes from a linear stochastic process
## Reject Null hypothesis:
## Original data's stat is significant larger than surrogates' stats
In this quickstart vignette we have only covered some of the main functions included in nonlinearTseries. Other interesting functions included in this package are enumerated below. The main reason for not including them in this quickstart guide is that these functions are quite simple to use.
rqa
: performs Recurrence Quantification Analysis.dfa
: performs Detrended Fluctuation Analysis.nonLinearNoiseReduction
: denoises a given time series
using phase space techniques.poincareMap
: computes a Poincare map of the
trajectories in phase space.spaceTimePlot
: shows the space time separation plot:
broadly-used method of detecting non-stationarity and temporal
correlations in time series.