9.8: Averaging information ( Entropy )

As one of applications (of Bayes theorem), we now study the "entropy of the measurement. This section is due to the following refs.

$(\sharp):$ S. Ishikawa, A Quantum Mechanical Approach to Fuzzy Theory, Fuzzy Sets and Systems, Vol. 90, No. 3, 277-306, 1997
$(\sharp):$ S. Ishikawa, Mathematical Foundations of Measurement Theory, Keio University Press Inc. 2006.



Let us begin with the following definition.

Definition 9.19 [Entropy ] Assume:

\begin{align} \mbox{ Classical basic structure $[ C_0(\Omega ) \subseteq L^\infty (\Omega, \nu ) \subseteq B(L^2 (\Omega, \nu ))] $ } \end{align}

Consider a mixed measurement ${\mathsf M}_{L^\infty(\Omega, \nu )} $ $({\mathsf O} = (X , 2^X , F ), $ $ {\overline S}_{[\ast ]}({w_0}) ) $ with a countable measured value space $X=\{x_1,x_2,\ldots\}$. The probability $P(\{ x_n \})$ that a measured value $x_n$ is obtained by the mixed measurement ${\mathsf M}_{L^\infty(\Omega)} ({\mathsf O} , {\overline S}_{[\ast ]}({w_0})) $ is given by

\begin{align} P(\{ x_n \}) = \int_{\Omega} [{}F(\{ x_n \})] (\omega) {w_0}(\omega) \nu (d \omega) \tag{9.25} \end{align}

Furthermore, when a measured value $x_n$ is obtained, the information $I(\{ x_n \})$ is, from Bayes' theorem 9.11, is calculated as follows.

\begin{align} I(\{ x_n \}) & = \int_\Omega \frac{ [{}F(\{ x_n \})] (\omega ) } {{ \int_{\Omega} [{}F(\{ x_n \})] (\omega ) {w_0}(\omega) \nu (d \omega ) } } \log \frac{ [{}F(\{ x_n \})] (\omega ) } {{ \int_{\Omega} [{}F(\{ x_n \})] (\omega ) {w_0} (\omega) \nu(d \omega ) } } {w_0}(\omega) \nu (d \omega) \end{align}

Therefore, the averaging information $ H \big( {\mathsf M}_{L^\infty(\Omega)} ({\mathsf O} , {\overline S}_{[\ast ]}({w_0})) \big) $ of the mixed measurement ${\mathsf M}_{L^\infty(\Omega)} $ $({\mathsf O}, $ $ {\overline S}_{[\ast ]}({w_0})) $ is naturally defined by

\begin{align} & \; \; H \big( {\mathsf M}_{L^\infty(\Omega)} ({\mathsf O} , {\overline S}_{[\ast ]}({w_0})) \big) = \sum\limits_{n=1}^\infty P(\{ x_n \}) \cdot I(\{ x_n \}) \tag{9.26} \end{align}

Also, the following is clear:

\begin{align} H \big({\mathsf M}_{L^\infty(\Omega)} ({\mathsf O} , {\overline S}_{[\ast ]}({w_0})) \big) = & \sum\limits_{n=1}^\infty \int_\Omega [{}F(\{ x_n \})] (\omega) \log [{}F(\{ x_n \})] (\omega) {w_0}(\omega) \nu (d \omega) \nonumber \\ & \qquad \qquad - \sum\limits_{n=1}^\infty P(\{ x_n \}) \log P(\{ x_n \}) \tag{9.27} \end{align} }




Example 9.20 [The offender is man or female?$\;$fast or slow?] Assume that

$(a):$ There are 100 suspected persons such as $\{ s_1 , s_2 ,\ldots, s_{100} \}$, in which there is one criminal.

Define the state space $\Omega$ $ = $ $\{ \omega_1 , \omega_2 ,\ldots, \omega_{100} \}$ such that

\begin{align} {\mbox{state}} \omega_n \cdots \mbox{the state such that suspect } {s_n} \mbox{ is a criminal} \qquad (n=1,2,...,100) \end{align}

Assume the counting measure $\nu$ such that $\nu(\{\omega_k \})=1 (\forall k=1,2, \cdots, 100)$ Define a male-observable ${\mathsf O}_{\rm m}$ $=$ $(X = \{ y_{\rm m} , n_{\rm m} \} , 2^{X} , M )$ in $L^\infty (\Omega)$ by

\begin{align} & [M({\{ y_{\rm m} \} } )](\omega_n) = m_{ y_{\rm m} }(\omega_n) = \left\{\begin{array}{ll} 0 \quad & (n \; \mbox{is odd}) \\ 1 & (n \; \mbox{is even}) \end{array}\right. \quad \\ & [M({\{ n_{\rm m} \} } )](\omega_n) = m_{ n_{\rm m} }(\omega_n) = 1- [M({\{ y_{\rm m} \} } )](\omega_n) \end{align}

For example,

$\quad$Taking a measurement ${\mathsf M}_{L^\infty (\Omega )} ({\mathsf O}_{\rm m} , S_{[\omega_{17} ]}) $ --- the sex of the criminal $s_{17}$ ---, we get the measured value $n_{\rm m}$(=female).

Also, define the fast-observable ${\mathsf O}_{\rm f}$ $ = $ $(Y= \{ y_{\rm f} , n_{\rm f} \} , 2^{Y} , F )$ in $L^\infty(\Omega)$ by

\begin{align} & [F({\{ y_{\rm f} \} } )](\omega_n) = f_{ y_{\rm f} }(\omega_n) = {\displaystyle \frac{n-1}{99} }, \qquad \\ & [F({\{ n_{\rm f} \} } )](\omega_n) = f_{n_{\rm f} }(\omega_n) = 1- [F({\{ y_{\rm f} \} } )](\omega_n) \end{align}

According to the principle of equal weight (=Theorem 9.18 ), there is a reason to consider that a mixed state ${w_0}$ $( \in L^1_{+1} (\Omega))$ is equal to the state $w_e$ such that ${w_0} ( \omega_n ) =w_e ( \omega_n ) = 1/100$ $(\forall n)$. Thus, consider two mixed measurement ${\mathsf M}_{L^\infty (\Omega )} ({\mathsf O}_{\rm m} , {\overline S}_{[\ast ]}(w_e)) $ and ${\mathsf M}_{L^\infty (\Omega )} ({\mathsf O}_{\rm f} , {\overline S}_{[\ast ]}(w_e)) $. Then, we see:

\begin{eqnarray*} H \big({\mathsf M}_{L^\infty (\Omega )} ({\mathsf O}_{\rm m} , {\overline S}_{[\ast ]}(w_e))\big) &=& \int_\Omega m_{ y_{\rm m} } (\omega) w_e (\omega ) \nu (d \omega) \cdot \log \int_{\Omega} m_{ y_{\rm m} } (\omega) w_e (\omega ) \nu(d \omega) \\ & & - \int_\Omega m_{\{ n_{\rm m} \} } (\omega) w_e (\omega ) \nu(d \omega) \cdot \log \int_\Omega m_{ n_{\rm m} } (\omega) w_e (\omega ) \nu(d \omega)\\ &=& - \frac{1}{2} \log \frac{1}{2} - \frac{1}{2} \log \frac{1}{2} =\log_2 2 = 1 \; \; \mbox{(bit)} \end{eqnarray*}

Also,

\begin{eqnarray*} H \big({\mathsf M}_{L^\infty (\Omega )} ({\mathsf O}_{\rm f} , {\overline S}_{[\ast ]}(w_e)) \big) &=& \int_\Omega f_{ y_{\rm f} } (\omega) \log f_{ y_{\rm f} } (\omega) w_e (\omega ) \nu(d \omega) \\ && \hspace{-4cm}+ \int_\Omega f_{ n_{\rm f} } (\omega) \log f_{ n_{\rm f} } (\omega) w_e (\omega ) \nu(d \omega) - \int_\Omega f_{ y_{\rm f} } (\omega) w_e(\omega ) \nu (d \omega) \cdot \log \int_{\Omega} f_{ y_{\rm f} } (\omega) w_e (\omega ) \nu(d \omega) \\ && \hspace{-4cm} - \int_\Omega f_{ n_{\rm f} } (\omega) w_e (d \omega) \cdot \log \int_\Omega f_{ n_{\rm f} } (\omega) w_e(\omega ) \nu (d \omega) \\ && \hspace{-4cm}{\doteqdot} 2 \int_0^1 \lambda \log_2 \lambda d \lambda +1 = - \frac{1}{ 2 \log_e 2} +1 = 0.278 \cdots \mbox{(bit)} \end{eqnarray*} Therefore, as eyewitness information, "male or female" has more valuable than "fast or slow".