Exercise 4.4: Maximum–a–posteriori and Maximum–Likelihood
From LNTwww
Channel transition probabilities
To illustrate "maximum–a–posteriori" $\rm (MAP)$ and "maximum likelihood" $\rm (ML)$ decision, we now construct a very simple example with only two possible messages $m_0 = 0$ and $m_1 = 1$, represented by the signal values $s_0$ resp. $s_1$:
- $$s \hspace{-0.15cm} \ = \ \hspace{-0.15cm}s_0 = +1 \hspace{0.2cm} \Longleftrightarrow \hspace{0.2cm}m = m_0 = 0\hspace{0.05cm},$$
- $$s \hspace{-0.15cm} \ = \ \hspace{-0.15cm}s_1 = -1 \hspace{0.2cm} \Longleftrightarrow \hspace{0.2cm}m = m_1 = 1\hspace{0.05cm}.$$
- Let the probabilities of occurrence be:
- $${\rm Pr}(s = s_0) = 0.75,\hspace{0.2cm}{\rm Pr}(s = s_1) = 0.25 \hspace{0.05cm}.$$
- The received signal can – for whatever reason – take three different values, i.e.
- $$r = +1,\hspace{0.2cm}r = 0,\hspace{0.2cm}r = -1 \hspace{0.05cm}.$$
- The conditional channel probabilities can be taken from the graph.
After transmission, the message is to be estimated by an optimal receiver. Available are:
- the maximum likelihood receiver $\rm (ML$ receiver$)$, which does not know the occurrence probabilities ${\rm Pr}(s = s_i)$, with the decision rule:
- $$\hat{m}_{\rm ML} = {\rm arg} \max_i \hspace{0.1cm} \big[ p_{r |s } \hspace{0.05cm} (\rho |s_i ) \big]\hspace{0.05cm},$$
- the maximum-a-posteriori receiver $\rm (MAP$ receiver$)$; this receiver also considers the symbol probabilities of the source in its decision process:
- $$\hat{m}_{\rm MAP} = {\rm arg} \max_i \hspace{0.1cm} \big[ {\rm Pr}( s = s_i) \cdot p_{r |s } \hspace{0.05cm} (\rho |s_i ) \big ]\hspace{0.05cm}.$$
Notes:
- The exercise belongs to the chapter "Optimal Receiver Strategies".
- Reference is also made to the chapter "Structure of the Optimal Receiver.
- The necessary statistical principles can be found in the chapter "Statistical Dependence and Independence" of the book "Theory of Stochastic Signals".
Questions
Solution
(1) The receiver side occurrence probabilities we are looking for are
- $${\rm Pr} ( r = +1) \hspace{-0.1cm} \ = \ \hspace{-0.1cm} {\rm Pr} ( s_0) \cdot {\rm Pr} ( r = +1 \hspace{0.05cm}| \hspace{0.05cm}s = +1) = 0.75 \cdot 0.8 \hspace{0.05cm}\hspace{0.15cm}\underline { = 0.6}\hspace{0.05cm},$$
- $${\rm Pr} ( r = -1) \hspace{-0.1cm} \ = \ \hspace{-0.1cm} {\rm Pr} ( s_1) \cdot {\rm Pr} ( r = -1 \hspace{0.05cm}| \hspace{0.05cm}s = -1) = 0.25 \cdot 0.6 \hspace{0.05cm}\hspace{0.15cm}\underline {= 0.15}\hspace{0.05cm},$$
- $${\rm Pr} ( r = 0) \hspace{-0.1cm} \ = \ \hspace{-0.1cm} 1 - {\rm Pr} ( r = +1) - {\rm Pr} ( r = -1) = 1 - 0.6 - 0.15 \hspace{0.05cm}\hspace{0.15cm}\underline {= 0.25}\hspace{0.05cm}.$$
- For the last probability also holds:
- $${\rm Pr} ( r = 0) = 0.75 \cdot 0.2 + 0.25 \cdot 0.4 = 0.25\hspace{0.05cm}.$$
(2) For the first inference probability we are looking for holds:
- $${\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = +1) = \frac{{\rm Pr} ( r = +1 \hspace{0.05cm}|\hspace{0.05cm}s_0 ) \cdot {\rm Pr} ( s_0)}{{\rm Pr} ( r = +1)} = \frac{0.8 \cdot 0.75}{0.6} \hspace{0.05cm}\hspace{0.15cm}\underline {= 1}\hspace{0.05cm}.$$
- Correspondingly, we obtain for the other probabilities:
- $${\rm Pr} (s_1 \hspace{0.05cm}| \hspace{0.05cm}r = +1) \hspace{-0.1cm} \ = \ 1 - {\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = +1) \hspace{0.05cm}\hspace{0.15cm}\underline {= 0}\hspace{0.05cm},$$
- $${\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = -1) \hspace{0.05cm}\hspace{0.15cm}\underline {= 0}\hspace{0.05cm},$$
- $${\rm Pr} (s_1 \hspace{0.05cm}| \hspace{0.05cm}r = -1) \hspace{0.05cm}\hspace{0.15cm}\underline {= 1}\hspace{0.05cm},$$
- $${\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = 0) \hspace{-0.1cm} \ = \ \hspace{-0.1cm}\frac{{\rm Pr} ( r = 0 \hspace{0.05cm}|\hspace{0.05cm}s_0 ) \cdot {\rm Pr} ( s_0)}{{\rm Pr} ( r = 0 )}= \frac{0.2 \cdot 0.75}{0.25} \hspace{0.05cm}\hspace{0.15cm}\underline {= 0.6}\hspace{0.05cm},$$
- $${\rm Pr} (s_1 \hspace{0.05cm}| \hspace{0.05cm}r = 0) \hspace{-0.1cm} \ = \ \hspace{-0.1cm} 1- {\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = 0) \hspace{0.05cm}\hspace{0.15cm}\underline {= 0.4} \hspace{0.05cm}.$$
(3) Let $r = +1$. Then decides
- the MAP receiver for $s_0$, because ${\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = +1) = 1 > {\rm Pr} (s_1 \hspace{0.05cm}| \hspace{0.05cm}r = +1)= 0\hspace{0.05cm},$
- the ML receiver likewise for $s_0$, since ${\rm Pr} ( r = +1 \hspace{0.05cm}| \hspace{0.05cm}s_0) = 0.8 > {\rm Pr} ( r = +1 \hspace{0.05cm}| \hspace{0.05cm}s_1) = 0 \hspace{0.05cm}.$
So the correct answer is NO.
(4) NO is also true under the condition "$r = \, –1$", since there is no connection between $s_0$ and "$r = \, –1$".
(5) Solutions 1 and 4 are correct:
- The MAP receiver will choose event $s_0$, since ${\rm Pr} (s_0 \hspace{0.05cm}| \hspace{0.05cm}r = 0) = 0.6 > {\rm Pr} (s_1 \hspace{0.05cm}| \hspace{0.05cm}r = 0) = 0.4 \hspace{0.05cm}.$
- In contrast, the ML receiver will choose $s_1$, since ${\rm Pr} ( r = 0 \hspace{0.05cm}| \hspace{0.05cm}s_1) = 0.4 > {\rm Pr} ( r = 0 \hspace{0.05cm}| \hspace{0.05cm}s_0) = 0.2 \hspace{0.05cm}.$
(6) The maximum likelihood receiver
- decides for $s_0$ only if $r = +1$,
- thus makes no error if $s_1$ was sent,
- only makes an error when "$s_0$" and "$r = 0$" are combined:
- $${\rm Pr} ({\rm symbol\hspace{0.15cm}error} ) = {\rm Pr} ({\cal E } ) = 0.75 \cdot 0.2 \hspace{0.05cm}\hspace{0.15cm}\underline {= 0.15} \hspace{0.05cm}.$$
(7) The MAP receiver, on the other hand, decides for $s_0$ when "$r = 0$". So there is a symbol error only in the combination "$s_1$" and "$r = 0$". From this follows:
- $${\rm Pr} ({\rm symbol\hspace{0.15cm}error} ) = {\rm Pr} ({\cal E } ) = 0.25 \cdot 0.4 \hspace{0.05cm}\hspace{0.15cm}\underline {= 0.1} \hspace{0.05cm}.$$
- The error probability here is lower than for the ML receiver,
- because now also the different a-priori probabilities ${\rm Pr}(s_0)$ and ${\rm Pr}(s_1)$ are considered.
