P( A B ) = P ( B A) P ( A) P( B )
Each term in Bayes's theorem has a conventional name: P(A) is the prior probability or marginal probability of A. It is "prior" in the sense that it does not take into account any information about B. P(A|B) is the conditional probability of A, given B. It is also called the posterior probability because it is derived from or depends upon the specified value of B. P(B|A) is the conditional probability of B given A. P(B) is the prior or marginal probability of B, and acts as a normalizing constant.
posterior = likelihood prior normalizin gcons tan t
In words: the posterior probability is proportional to the prior probability times the likelihood. In addition, the ratio P(B|A)/P(B) is sometimes called the standardised likelihood, so the theorem may also be paraphrased as
Two ways to lose information 1. 2. Cost C0 information lost when a transmitted digital 1 is received digital 0 (error) Cost C1 information lost when a transmitted digital 0 is received digital 1 (error)
There is no cost when no information is lost, i.e., when correct decisions are made.
2.0
Expected conditional cost, C ( 0 v ) , incurred when a detected voltage is interpreted as digital 0 is given by C ( 0 v ) = C 0 P (1 v ) (1) C 0 = cost if decision is in error ( v ) is interpreted as digital 0; P(1 v ) = the a posteriori probability that decision is in error.
Similarly, by symmetry,
C (1 v ) = C1 P( 0 v ) ,
(2)
is interpreted as digital 1.
3.0
Rational rule: interpret each detected voltage, the expected conditional cost, i.e.,
v , as either a 0 a 1 so as to minimize
(3)
< C (1 v ) C ( 0 v ) >
0
Interpretation of (3): If upper inequality holds, decides binary 1; if lower inequality holds, decides binary 0. Substituting (2) into (3) gives the following equation.
(4)
If costs of both types of error are same, then (4) represents a maximum a posteriori (MAP) probability decision criteria (one form of Bayess decision criterion). Bayess theorem:
P ( v,0 ) = p ( v ) P ( 0 / v ) = P ( 0) p ( v 0 ) and P ( v,1) = p ( v ) P (1 v ) = P (1) p ( v 1)
(5)
(6)
Figure below illustrates the conditional probability density function for the case of zero mean Gaussian noise.
(7)
(8)
or
(9)
and
threshold Lth . If C 0 = C1 and P ( 0) = P (1) (or C 0 P(1) = C1 P ( 0) ), then Lth =1 and (9) is called maximum likelihood decision criterion. Relationship between maximum likelihood, MAP and Bayess decision criteria: Receiver A priori probabilities known Yes Decision costs known Yes Assumptions Decision criterion
Bayes
None
MAP
Yes
No
C 0 = C1
Max. Likelihood
No
No
C 0 P(1) = C1 P ( 0)
p( v 0) < 1 p ( v 1) >
0
4.0
Bayess decision criterion sets optimum reference (threshold voltage), v th in receiver decision circuit. v th minimizes the expected conditional cost of each decision and satisfies: p ( v 0) C 0 P (1) = = Lth (10) p( v 1) C1 P( 0 ) If Lth =1 (for statistically independent, equiprobable symbols with equal error costs), then voltage threshold would occur at the intersection of the two conditional pdfs. Assume voltage levels for binary 0 and 1 are represented by 0 volt and V volts, respectively. If binary 1 and 0 are equi-probable the decision threshold is located exactly hslfway between the voltage levels representing 1 and 0. If 0 is transmitted more often than 1 the threshold moves towards the transmitted voltage representing 1. Once the decision threshold has been established the total probability of error can then be calculated. If binary 0 (0 volt) is sent, the probability that it will be received as a 1 is the probability that noise will exceed +
P0 = P ), i.e., 1
A volts (assuming 2
Pe 0 =
exp ( v / 2 2 )
A/ 2
( 2 )
2
dv
(11)
If a 1 is sent (A volts) the probability that it will be received as a 0 is the probability that the noise voltage will be between A / 2 and , i.e.,
Pe1 =
A / 2
exp v / 2 2
( 2 )
2
)dv
(11)
These types of error are mutually exclusive, since sending a 0 precludes sending a 1. Symmetry of Gaussian function results in Pe 0 = Pe1 , and the total probability of error is P0 Pe 0 +P Pe1 . This can be reduced to 1
Pe = Pe1 ( P0 + P1 ) = Pe1
i.e.,
Pe =
A / 2
exp v / 2 2
( 2 )
2
)dv
(12)
exp v / 2 2
( 2 )
2
)dv
exp v 2 / 2 2
A / 2
( 2 )
2
) dv
(13)
Using the fact that Gaussian distribution is symmetrical about its mean value, equation (13) reduces to 2 2 A / 2 exp ( v / 2 ) 1 Pe = + dv 2 0 ( 2 2 ) Substituting y = v / 2 2 , the equation reduces to
Pe = 1 1 + 2
A / 2 2
exp y 2 dy
i.e.,
Pe =
1 1 erf A / 2 2 2
[ (
)]}
(14)
The error probability therefore depends solely on the ratio of the peak pulse voltage A to the rms noise voltage .