Scribd
Upload
38 views

Final Sol

This document contains solutions to problems on channel capacity with state information, rate-distortion theory under log loss, and analyzing a maximum mutual information (MMI) decoding scheme. For the channel capacity problem, the solutions find capacity for different cases depending on whether the transmitter and receiver know the state information. For rate-distortion, the rate-distortion function is derived for log loss and a time-sharing scheme is described. For MMI decoding, it is shown the decoder does not need channel statistics and analyzing error probability leads to the capacity being the mutual information.

Uploaded by

Islam El Bakoury
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
38 views

Final Sol

This document contains solutions to problems on channel capacity with state information, rate-distortion theory under log loss, and analyzing a maximum mutual information (MMI) decoding scheme. For the channel capacity problem, the solutions find capacity for different cases depending on whether the transmitter and receiver know the state information. For rate-distortion, the rate-distortion function is derived for log loss and a time-sharing scheme is described. For MMI decoding, it is shown the decoder does not need channel statistics and analyzing error probability leads to the capacity being the mutual information.

Uploaded by

Islam El Bakoury
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

EE376A: Final Solutions

1. Channel Capacity in Presence of State Information (32 points)


Consider a binary memoryless channel with state whose input-output relation is described
as follows. The received channel output Y is given by
Y = X S Z,
where X is the channel input, S is the channel state, and Z is the channel noise. X, S, Z
and Y all take their values in {0, 1}, and denotes modulo-2 addition. S Bernoulli(q)
and Z Bernoulli(p) are independent, and jointly independent of the channel input X.
Thus, when we employ n-block encoding and decoding, we have for each 1 i n
Yi = Xi Si Zi ,
where S n are i.i.d. Bernoulli(q) and Z n are i.i.d. Bernoulli(p), where S n and Z n are
independent, and jointly independent of the channel input sequence X n .
Find the capacity of the channel in the following cases:
(a) (8 points) Neither the transmitter nor the receiver knows the state sequence.
(b) (8 points) The state sequence is known only to the receiver, i.e., the n-block decoder
gets to base its decision on Y n and S n .
(c) (8 points) The state sequence is known to both the transmitter and the receiver, i.e.,
the n-block encoder and decoder know S n prior to communication.
(d) (8 points) The state sequence is known to both the transmitter and the receiver, as
in the
Pprevious part but, in addition, the encoding must adhere to the cost constraint
E[ n1 ni=1 Xi ] 0.25.
Solution:
(a) The state can be considered as a part of noise, i.e., S Z can be treated as a
noise. Therefore, the channel is a binary symmetric channel with crossover probability
p q = p(1 q) + (1 p)q, ant the capacity is Ca = 1 h2 (p q) where h2 (x) is a
binary entropy function.
(b) The state can be considered as a part of output, i.e., (S, Z) can be treated as an
output. Therefore, the capacity is
Cb = max I(X; Y, S)
PX

= max H(Y, S) H(Y, S|X)


PX

= max H(X Z, S) H(Z, S)


PX

= max H(X Z) + H(S) H(Z) H(S)


PX

= 1 h2 (p).
Final

Page 1 of 7

where the capacity achieving distribution is Bernoulli( 21 ).


(c) Since the state is given to both the encoder and the decoder, the capacity is
Cc = max I(X; Y |S)
PX|S

= max H(Y |S) H(Y |X, S)


PX|S

= max H(X Z|S) H(Z)


PX|S

= 1 h2 (p).
(d) The capacity with cost constraint is
Cd =
=

I(X; Y |S)

max
PX|S :E[X]0.25

H(Y |S) H(Y |X, S)

max
PX|S :E[X]0.25

H(X Z|S) H(Z)

max
PX|S :E[X]0.25

H(X Z) H(Z)

max
PX|S :E[X]0.25

max

H(X Z) H(Z)

PX :E[X]0.25

1
= h2 ( p) h2 (p).
4
It is clear that the equality can be achieved by X Bern( 14 ) that are independent
to S. Therefore,
1
Cd = h2 ( p) h2 (p).
4

2. Rate Distortion under Log Loss (40 points)


Let X be distributed i.i.d. according to distribution P P(X ), where P(X ) denotes
the set of all probability mass functions on the finite alphabet X . Distortion functions
usually ask for hard reconstructions. For example, the reconstruction alphabet is the
same as the source alphabet, Y = X , and the distortion is Hamming d(x, y) = 1(x 6= y).
Suppose we are instead interested in a soft reconstruction: instead of a finite reconstruction alphabet we let Y = P(X ), i.e., the reconstruction is a PMF on X . Thus, a
reconstruction q P(X ) may be construed as a distribution of belief q(x) by the decoder
over the values x X that the source symbol might take. A natural distortion function
for such soft reconstructions is the log-loss:
d` (x, q) = log q(x).

Final

Page 2 of 7

(a) (8 points) Show that


E[d` (X, q)] H(X),
for all q P(X ). When is equality achieved?
(b) (8 points) Suppose X and U are jointly distributed and denote the conditional PMF
of X given U by PX|U . Show that


E d` X, PX|U (|U ) = H(X|U ),
where PX|U (|u) denotes the conditional PMF of X given U = u and, therefore,
PX|U (|U ) is a P(X )-valued random variable.
(c) (8 points) Consider the random PMF Q that is, it is a random variable on
P(X ). Let Q be jointly distributed with X according to PX,Q . Furthermore, let
= PX|Q (|Q) where Q
is again a random variable on P(X ). Show that
Q
h
i

E [d` (X, Q)] E d` (X, Q) = H(X|Q).


Also, argue why
I(X; Q).
I(X; Q)
(d) (8 points) Find and plot the rate distortion function under log-loss. I.e., find
R(D) =

min

I(X; Q).

E[d` (X,Q)]D

Hint: use the previous part to show that the constraint set for the minimization can
be taken to be H(X|Q) D instead of E[d` (X, Q)] D.
(e) (8 points) For a given distortion level D, describe a concrete implementable scheme
(not based on a random coding argument) for achieving R(D).

Solution: Rate Distortion under Log Loss.


(a) Since X is distributed according to P ,
E[d` (X, q)] H(X) = E[ log q(X)] E[ log P (X)]
P (X)
]
= E[log
q(X)
= D(P ||q)
0.
The equality holds if and only if q = P .

Final

Page 3 of 7

(b) By definition of d` (, ) and the tower property,


E[d` (X, PX|U (|U ))] = E[E[d` (X, PX|U (|U ))|U ]]
= E[E[ log PX|U (X|U )|U ]]
= H(X|U ).
(c) By the tower property,
E[d` (X, Q)] = E[E[d` (X, Q)|Q]]
E[E[d` (X, PX|Q (|Q))|Q]]
= E[d` (X, PX|Q (|Q))]

= E[d` (X, Q)]


= H(X|Q).
Note that E[d` (X, Q)] H(X) is not true for random pmf Q.
is a function of Q,
Since Q

H(X|Q) H(X|Q).
Therefore,
= H(X) H(X|Q)
H(X) H(X|Q) I(X; Q).
I(X; Q)
(d) In part (c), we have seen that E[d` (X, Q)] D implies H(X|Q) D. Therefore,
R(D) =

min

I(X; Q)

E[d` (X,Q)]D

min

I(X; Q)

H(X|Q)D

H(X) H(X|Q)

min
H(X|Q)D

= H(X) D.
= H(X|Q), we have
On the other hand, since E[d` (X, Q)]
H(X) D =

min

I(X; Q)

H(X|Q)D

min

E[d` (X,Q)]D

I(X; Q)

= R(D).
Therefore, R(D) = H(X) D which is a straight line between (H(X), 0) and
(0, H(X)).
(e) Since the rate-distortion curve is a straight line, we can use the time sharing argument.
First, we have a concrete scheme that achieves that losslessly compress the source
using the rate H(X) (Enumerating all typical sequences). Also, we have a concrete
scheme that achieves the distortion H(X) using zero-rate (the reconstruction is simply
PMF of X).
D
By using first scheme for H(X)D
fraction of time and using second scheme for H(X)
H(X)
fraction of time, we can achieve the distortion D with a rate R = H(X) D.
Final

Page 4 of 7

3. The MMI decoder (40 points)


Valery has discovered an amazing new channel decoder. He claims it needs to know
nothing about the channel! Imre and Janos are suspicious and need your help checking
the validity of his claim.
The (standard random coding) assumptions:
2nR codewords are drawn i.i.d. according to distribution PX .
Message J is chosen uniformly from the set {1, 2, . . . , 2nR }, and the corresponding
codeword X n (J) is sent through the channel.
The channel is discrete memoryless, characterized by the conditional PMF PY |X ,
where both X and Y take values in the respective finite alphabets X and Y.
Now suppose sequence Y n Y n is received by the decoder. Let PX n (j),Y n denote the joint
empirical distribution of the jth codeword X n (j) and Y n . Valerys decoder produces as
its estimate the codeword with maximum empirical mutual information with Y n :
J = arg max I(PX n (j),Y n )
j{1,2,...,2nR }

where, for QX,Y P(X , Y), I(QX,Y ) denotes the mutual information between X and Y
when distributed according to QX,Y (and ties in the maximization are broken arbitrarily).
(a) (8 points) Does Valerys decoder have to know the channel statistics PY |X in order
to implement this decoding scheme?
(b) (8 points) Prove that, for any QX,Y P(X , Y),
D (QX,Y kPX PY ) I(QX,Y ).
Hint: Use the fact that I(QX,Y ) = D (QX,Y ||QX QY ).
(c) (8 points) Using the previous part, prove that f () , where

f () =

min

QX,Y P(X ,Y):I(QX,Y )

D (QX,Y kPX PY ) .

(d) (8 points) Using the method of types show that for any 2 j 2nR


P I(PX n (j),Y n ) J = 1 = 2nf ()
(e) (8 points) What is the supremum of rates R for which



P J 6= J 0 as n
under Valerys scheme? How does it compare to the supremum of rates for which
joint typicality decoding would have achieved reliable communication?

Final

Page 5 of 7

Hint: First justify why for any





P J 6= J


= P J 6= J|J = 1




P I(PX n (1),Y n ) < J = 1 + P I(PX n (j),Y n ) for some 2 j 2nR J = 1




P I(PX n (1),Y n ) < J = 1 + 2nR P I(PX n (2),Y n ) J = 1 .
Then, with the help of previous parts, establish for which values of R and the
expression in the last line vanishes.

Solution:
(a) It only needs to know the X n (j) for all 1 j 2nR and the output Y n . Unlike to
the joint typicality decoding, the decoder does not have to know about the channel
statistics PY |X .
(b)
D(QX,Y ||PX PY ) I(QX,Y )
X
X
QX,Y (x, y)
QX,Y (x, y)
QX,Y (x, y) log

QX,Y (x, y) log


=
PX (x)PY (y)
QX (x)QY (y)
x,y
x,y
=

QX,Y (x, y) log

x,y

QX (x) log

QX (x)QY (y)
PX (x)PY (y)

QY (y)
QX (x) X
+
QY (y) log
PX (x)
PY (y)
y

= D(QX ||PX ) + D(QY ||PY )


0.
(c)

f () =

min

D (QX,Y kPX PY )

min

I (QX,Y )

QX,Y P(X ,Y):I(QX,Y )


QX,Y P(X ,Y):I(QX,Y )

.
(d) Let En = {QX,Y P(X , Y) : I(QX,Y ) }. For j 2, the joint law of X n (j) and

Final

Page 6 of 7

Y n is PX PY . Therefore,



P I(PX n (j),Y n ) J = 1 = P PX n (j),Y n En |J = 1
X
=
P(T (QX,Y ))
QX,Y En

(PX PY )n (T (QX,Y ))

QX,Y En

.
= 2n minQX,Y En D(QX,Y ||PX PY )
.
= 2nf ()
(e) Recall the hint:


P J 6= J


(i)
= P J 6= J|J = 1
(ii)

(iii)





P I(PX n (1),Y n ) < J = 1 + P I(PX n (j),Y n ) for some 2 j 2nR J = 1




P I(PX n (1),Y n ) < J = 1 + 2nR P I(PX n (2),Y n ) J = 1 .

This is because,
(i) is because of symmetry.
(ii) is because: J 6= 1 if I(PX n (1),Y n ) < or I(PX n (J),Y n ) I(PX n (1),Y n ) for
some other J 6= 1.
(iii) is because of symmetry again.
Suppose that the inequalities R < < I(X; Y ) = I(PX,Y ) hold. Then, by the law of
large number,
lim I(PX n (1),Y n ) = I(PX,Y ) > .



This implies that limn P I(PX n (1),Y n ) < J = 1 = 0.
On the other hand, by part (c) and (d),

 .
2nR P I(PX n (2),Y n ) J = 1 = 2n(Rf ())
2n(R)
which vanishes as n grows.


Thus, we can argue that P J 6= J converges to zero as n grows. Finally, using this
scheme, we can achieve any rate below supPX I(X; Y ) which is the channel capacity
that we achieved using joint typicality.

Final

Page 7 of 7

You might also like