this is for holding javascript data
Kim H. Parker edited subsection_information_theory_Information_theory__.tex
about 8 years ago
Commit id: 03f83e15c86f5bba2bd63486bde6fbc0e3a7dc69
deletions | additions
diff --git a/subsection_information_theory_Information_theory__.tex b/subsection_information_theory_Information_theory__.tex
index 5acbdb2..2ce5d9f 100644
--- a/subsection_information_theory_Information_theory__.tex
+++ b/subsection_information_theory_Information_theory__.tex
...
\[
H(A_x) = -\sum_{x \in A} \phi(A_x) \log \phi(A_x)
\]
where $\phi(A_x)$ is the probability density function of $A$. It is a measure of the uncertainty of $A$ and its units depend on the base of the logarithm. We will use log base 2
in which
means that case the unit of entropy is bits.
Given two probability density functions $A(x)$ and $B(x)$ which are defined over the same variable $x$, the distance between them can be measured in several different ways. One of the first measures of the difference is the Kullback-Leibler divergence
\[
...
\]
This measure of distance has several disadvantages; it is not symmetric and it is not a metric. The Jensen-Shannon divergence is defined using the Kullback-Leibler divergence in a way that makes it symmetric
\[
JS(A;B) = \frac{1}{2}
\big(D(A||M) D(A||M) +
D(B||M)\big) \frac{1}{2} D(B||M)
\]
where $\phi(M_x) = \frac{1}{2}\big(\phi(A_x) + \phi(B_x)\big)$
From its definition it can easily be shown that
\[
JS(A,B) = H(M) -
\frac{1}{2}\big(H(A) + H(B)\big) \frac{1}{2}(H(A) - H(B))
\]
That is, the Jensen-Shannon divergence is equal to the entropy of the average distribution of the two distributions minus the average of the entropies of the individual distributions.