Lecture
Let there be two systems and generally dependent. Suppose the system took the state . Denote conditional probability that the system will take the state provided that the system is able to :
. (18.4.1)
We now define the conditional entropy of the system provided that the system is able to . Denote it . By general definition, we have:
(18.4.2)
or
. (18.4.2 ')
The formula (18.4.2) can also be written in the form of a mathematical expectation:
, (18.4.3)
where is familiar denotes the conditional expectation of the value in brackets, provided .
Conditional entropy depends on which state adopted the system ; for some states, it will be more, for others - less. Determine the average, or total, entropy of the system taking into account the fact that the system can take different states. To do this, each conditional entropy (18.4.2) must be multiplied by the probability of the corresponding state and all such works add up. Denote the full conditional entropy :
(18.4.4)
or, using the formula (18.4.2),
.
Bringing in under the sign of the second sum, we get:
(18.4.5)
or
. (18.4.5 ')
But by the probability multiplication theorem , Consequently,
. (18.4.6)
The expression (18.4.6) can also be given the form of a mathematical expectation:
. (18.4.7)
Magnitude characterizes the degree of system uncertainty remaining after system state fully defined. We will call it the complete conditional entropy of the system. regarding .
Example 1. There are two systems. and merged into one ; probabilities of system states set by table
0.1 | 0.2 | 0 | 0.3 | |
0 | 0.3 | 0 | 0.3 | |
0 | 0.2 | 0.2 | 0.4 | |
0.1 | 0.7 | 0.2 |
Determine the total conditional entropy and .
Decision. Adding probabilities by columns, we get probabilities :
; ; .
Write them in the bottom, extra row of the table. Similarly, folding in rows, we find:
; ;
and write to the right an additional column. Sharing on , we will receive the table of conditional probabilities :
By the formula (18.4.5 ') we find . Since the conditional entropy at and equal to zero then
.
Using table 7 of the application, we find
(two units).
Similarly, we define . From the formula (18.4.5 '), changing places and , we get:
.
Make a table of conditional probabilities . Sharing on we will receive:
From here
(two words).
Using the concept of conditional entropy, one can determine the entropy of the combined system through the entropy of its constituent parts.
We prove the following theorem:
If two systems and combined into one, the entropy of the combined system is equal to the entropy of one of its components plus the conditional entropy of the second part relative to the first:
. (18.4.8)
For proof, we write in the form of mathematical expectation (18.3.3):
.
By the probability multiplication theorem
,
Consequently,
,
from where
or, according to the formulas (18.2.11), (18.3.3)
,
Q.E.D.
In the particular case of systems and independent and we get the one already proven in the previous one entropy addition theorem:
.
In general
. (18.4.9)
The ratio (18.4.9) follows from the fact that the total conditional entropy can not surpass the unconditional:
. (18.4.10)
Inequality (18.4.10) will be proved in 18.6. Intuitively, it seems pretty obvious: it is clear that the degree of uncertainty of the system cannot increase because the state of some other system has become known.
It follows from relation (18.4.9) that the entropy of a complex system reaches a maximum in the extreme case when its components are independent.
Consider another extreme case where the state of one of the systems (for example ) completely determines the state of another ( ). In this case and the formula (18.4.7) gives
.
If the state of each system uniquely identifies the state of another (or, as they say, systems and are equivalent)
.
The entropy theorem of a complex system can be easily extended to any number of systems to be merged:
, (18.4.11)
where the entropy of each subsequent system is calculated under the condition that the state of all the previous ones is known.
Comments
To leave a comment
Information and Coding Theory
Terms: Information and Coding Theory