Shannon.jl

A collection of quantifications related to Shannon's information theory and methods to discretise data.

using Shannon
xy = hcat([sin(x) + randn() * .1 for x=0:0.01:2pi], [cos(x) + randn() * .1 for x=0:0.01:2pi])
bxy = bin_matrix(xy, -1.0, 1.0, 10)
c=combine_binned_matrix(bxy)
c=relabel(c)
H = entropy(c)
I = MI(bxy)

A faster way is to call

unary_of_matrix(xy, -1.0, 1.0, 10)

which is a short cut for the lines below

bxy = bin_matrix(xy, -1.0, 1.0, 10)
c=combine_binned_matrix(bxy)
c=relabel(c)

The estimators are implemented from the following list of publications:

[1] A. Chao and T.-J. Shen. Nonparametric estimation of shannon’s index of diversity when there are unseen species in sample. Environmental and Ecological Statistics, 10(4):429–443, 2003.

and the function call is

entropy(data, base=2, mode="ML")

where

###Maximum Likelihood Estimator This is the default estimator.

entropy(data)
entropy(data, mode="ML")
entropy(data, mode="Maximum Likelihood")

###Maximum Likelihood Estimator with Bias Correction (implemented from [1])

entropy(data, mode="MLBC")
entropy(data, mode="Maximum Likelihood with Bias Compensation")

###Horovitz-Thompson Estimator (implemented from [1])

entropy(data, mode="HT")
entropy(data, mode="Horovitz-Thompson")

###Chao-Shen Estimator (implemented from [1])

entropy(data, mode="CS")
entropy(data, mode="Chao-Shen")
entropy(data, mode="ChaoShen")

entropy(data, base=2) [ this is the default ]
entropy(data, mode="HT", base=10)

Currently, only the maximum likelihood estimator is implemented. It can be used with different bases:

MI(xy, base=2) [ this is the default ]
MI(xy, base=10)

xy is a two-dimensional matrix with n rows and two columns.

This in an implementation of the one-step predictive information, which is given by the mutual information of consecutive data points. If x is the data vector, then:

PI(x) = MI(hcat(x[1:end-1], x[2:end]))
PI(x,[base],[mode]) = MI(x[1:end-1], x[2:end], base, mode)

This function calculates the KL-Divergence on two probability distributions, and is essentially given by:

KL(p,q)=  sum([(p[i] != 0 && q[i] != 0)? p[i] * log(base, p[i]/q[i]) : 0 for i=1:length(p)])

p,q must be valid probability distributions, i.e.

x >= 0 for x in p
y >= 0 for y in q
sum(p) == sum(q) == 1.0

Implementation of measures from

Quantifying Morphological Computation, Zahedi & Ay, Entropy, 2013: [pdf]

and

Quantifying Morphological Computation based on an Information Decomposition of the Sensorimotor Loop, Ghazi-Zahedi & Rauh, ECAL 2015: [pdf]