A Dirty-paper Coding Scheme for the Cognitive

Radio Channel
Momin Uppal

, Guosen Yue

, Yan Xin

, Xiaodong Wang

, and Zixiang Xiong

Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX 77843

NEC Laboratories America Inc, Princeton, NJ 08540

Department of Electrical Engineering, Columbia University, New York, NY 10027

Emails: momin@tamu.edu, {yueg,yanxin}@nec-labs.com, wangx@ee.columbia.edu, zx@ece.tamu.edu
AbstractWe implement a dirty-paper coded framework for
the cognitive radio channel. We assume that the cognitive user has
non-causal knowledge about the primary users transmissions.
Thus the secondary receiver can employ dirty-paper coding to
counter the effect of any interference from the primary user.
In addition, we consider a situation where the introduction of
the cognitive user should not affect the performance of the
primary system nor should the primary system have to change
its encoding/decoding process. For the primary user we use
a low-density parity-check code and a 4-ary pulse amplitude
modulation format. For the cognitive user, we propose a dirty-
paper coding scheme which employs trellis-coded quantization
as the source code and an irregular repeat-accumulate code as
the channel code. At a transmission rate of 1.0 bits/sample, the
designed dirty-paper coding scheme operates within 1.23 dB of
the theoretical limit.
I. INTRODUCTION
Recent years have seen a growing trend in the use of
wireless devices, with an ever increasing appetite for band-
width intensive applications. In order to efciently use the
current available bandwidth, the idea of deploying secondary
wireless devices has been proposed with the secondary
devices using the same frequency band as the existing primary
systems. In order to have a minimal effect on the operation and
performance of the primary systems, these secondary devices
have to be opportunistic in nature. In other words, they should
be able to cognitively adapt to their environment, while at
the same time communicating effectively with their respective
base-stations. Because of this required cognition capability, the
term cognitive radio has been widely used in the literature.
The simplest form of the cognitive radio channel (CRC)
consisting of one primary and one secondary/cognitive user is
shown in Fig. 1. Each user wishes to communicate some infor-
mation to its respective base-station. However, since the two
users share the same frequency band, the signal transmitted by
a user interferes with the transmission of the other. A number
of works have focused on deriving the information theoretic
achievable rates for the CRC when the primary user message
is known at the cognitive user non-causally [1], [2], as well as
causally [3]. For the non-causal case, it was shown in [1] that
when the channel from the cognitive user to its respective base-
station is stronger than the cognitive user to primary receiver
channel (referred to as the low-interference regime), a scheme
which achieves capacity involves dirty-paper coding (DPC)
[4]. For the general CRC as well, the inner bounds on the
capacity region rely heavily on dirty-paper coding [1], [2],
[3]. In this work, we design and simulate a DPC based coding
scheme for the CRC. As a rst step towards developing a
practical coding strategy, we assume that the cognitive user has
non-causal knowledge of the primary user message as well as
its codebook. In addition, we consider the low interference
regime where the channel gain from the cognitive user to
its base-station is stronger than it is to the primary base-
station. This scenario is of practical interest since the cognitive
user will typically be closer to its base-station than that of
the primary user. As mentioned earlier, the cognitive radio
should be minimally intrusive in the operation of the primary
system. In this work, we consider the extreme situation where
the introduction of the cognitive radio should have no effect
whatsoever on the primary users operation and performance,
i.e. the primary system should be oblivious to the presence of
the cognitive user. For the primary user, we use a low-density
parity-check (LDPC) code with the codeword bits mapped
to a 4-ary pulse amplitude modulation (PAM) constellation,
whereas for the cognitive user, we propose a DPC scheme
which uses trellis coded quantization (TCQ) for source coding
and an irregular repeat-accumulate (IRA) code for channel
coding. Simulations indicate that at a transmission rate of 1.0
bits/sample (b/s) and a block length of 50,000, the proposed
DPC scheme for the cognitive user performs within 1.23 dB
of the theoretical limit [1].
Primary Base-Station
Secondary Base-Station
Secondary/Cognitive User
Primary User
+
+
Xp
Xc
Zp
Zc
Yp
Yc
1
1
a
b
Fig. 1. The cognitive radio channel.
978-1-4244-6404-3/10/$26.00 2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings
II. CODING SCHEME
In this section, we briey describe the coding scheme for
the CRC, and provide the information theoretical performance
limits. We consider the CRC in its standard form [1] as
depicted in Fig. 1. The primary user encodes its message using
a rate-R
l
p
LDPC code. The coded bits are then mapped to the
symbols X
p
[n], n = 1, . . . , L, which belong to a constellation
of size M = 2
m
. Thus the overall transmission rate for the
primary user is given as R
p
= mR
l
p
b/s and the average
transmission power as P
p
=
1
L

L
n=1
X
2
p
[n]. The primary
base station decodes its message using belief-propagation (BP)
algorithm on the LDPC decoding graph. We assume that
the primary user always transmits with the minimum power
required for the BP decoder to satisfy a given bit-error rate
(BER) requirement, with the decoder assuming that the noise
(plus any interference) has Gaussian statistics.
On the other hand, the cognitive user allocates a fraction
of its power for transmitting the coded symbols of the primary
user. With the remaining power, it encodes its own message
using a DPC scheme (the details of which will be discussed
in Section III) with S[n] =
_
b +
_
Pc
Pp
_
X
p
[n] treated as
the side-information. The dirty paper coded output

X
c
[n],
n = 1, . . . , L, satises
1
L

L
n=1

X
2
c
[n] = (1 )P
c
and the
cognitive user transmits X
c
[n] =

X
c
[n] +
_
Pc
Pp
X
p
[n]. A key
property of the DPC encoder is that the coded output

X
c
is
uncorrelated with the side-information X
p
[4] a property
veried by the simulations of our DPC scheme of Section
III. Thus, the total transmission power at the cognitive user is
1
L

L
n=1
X
2
c
[n] = P
c
. The signals received at the primary and
secondary base-stations are given as
Y
p
[n] =
_
1 +a

P
c
P
p
_
X
p
[n] +a

X
c
[n] +Z
p
[n],
Y
c
[n] =

X
c
[n] +
_
b +

P
c
P

p
_
X
p
[n] +Z
c
[n],
respectively, where Z
p
[n], Z
c
[n] are the unit variance additive
white Gaussian noise, and a, b are the real channel gains
from a user to the other users base-station. Note that in the
standard from of the CRC, the low interference regime implies
a < 1. Since the primary base-station is oblivious to the
presence of the cognitive user, the term a

X
c
[n] is treated as
interference. In order for the signal to noise (plus interference)
ratio (SNR) at the primary base-station to remain the same as
in the absence of the cognitive user, the choice of the power
allocation parameter should be [1]

(P
p
, P
c
) =
_
_
_
P
p
_
_
1 +a
2
P
c
(1 +P
p
) 1
_
a

P
c
(1 +P
p
)
_
_
2
. (1)
Thus if the BER of the primary decoder were a function of
only the SNR, choosing =

(P
p
, P
c
) would ensure that
the BER performance of the primary decoder is not affected
by the introduction of the cognitive user. However, note that
the primary decoder will calculate its channel log-likelihood
ratios (LLR) by assuming that the noise plus interference term
is Gaussian which might not be the case in a practical
setup. Thus even though the SNR remains the same, the BER
performance of the primary decoder might not be the same as
that without the cognitive user. Fortunately, our simulations
indicate that the dirty-paper coded output

X
c
is close to
Gaussian and the performance of the primary decoder is not
adversely affected by assuming that the interference is Gaus-
sian. Because of this reason, we always choose =

(P
p
, P
c
)
in our coding setup.
A. Performance Limits
With the choice of =

(P
p
, P
c
), the information
theoretic achievable rate for the primary user is given by
1
2
log (1 +P
p
). Thus for a given rate requirement R
p
, the
theoretical minimum for the primary user power is
P
p,min
= 2
2Rp
1. (2)
On the other hand, the secondary receiver does not see any
interference from the primary user transmission because of
DPC. Thus the achievable rate for the cognitive user is given
as
1
2
log
_
1 +
_
1

(P
p
, P
c
)
_
P
c
_
. Using simple algebra, it
can be shown that the received SNR at the secondary base-
station is an increasing function of the cognitive user power
P
c
. Hence, for a given rate requirement R
c
, the theoretical
minimum for the cognitive user power is
P
c,min
= {P
c
|
_
1

(P
p
, P
c
)
_
P
c
= 2
2Rc
1}. (3)
III. A DPC SCHEME FOR THE COGNITIVE USER
Since DPC is a source-channel coding problem [5], [6], [7],
the DPC encoder contains a channel as well as source coding
component. Focusing on the high rate regime, [6] proposed a
DPC scheme based on nested turbo codes. Because of the
nested nature of the scheme, the presence of the random
interleaver in the turbo-channel code negatively impacts the
performance of the TCQ source code. Whereas this does not
affect the overall performance too much at higher transmission
rates, the impact on lower rates is much more pronounced.
Indeed, using a 256-state TCQ as the source code, the scheme
in [6] performs only 1.42 dB from capacity at a transmission
rate of 1.0 b/s, whereas the gap to capacity increases to 2.65 dB
at a lower transmission rate of 0.5 b/s. On the other hand, the
schemes in [5], [7] employ IRA codes as the channel code and
TCQ as the source code. In contrast with [6], these schemes do
not suffer loss in source coding performance and were shown
to perform near capacity at low transmission rates of 0.25
b/s. However, in their original format, these schemes cannot
achieve a rate higher than 1.0 b/s. In this work, we use the
coding framework of the scheme in [5] and adapt it to higher
transmission rates. Our extension to higher rates is analogous
to the extension of an LDPC code with BPSK modulation
to that over a higher order constellation using bit-interleaved
coded modulation (BICM) [8]. The motivating factors behind
using an IRA code based DPC scheme over a turbo-coded
scheme are:
1) Since the source code does not suffer any degradation in
performance (as the scheme in [6] does), we are able to reduce
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings
the gap to capacity. At a transmission rate of 1.0 b/s, a block
length L = 50, 000 and a 64-state TCQ as the source code,
we are able to achieve a gap of 1.23 dB, as opposed to 1.42
dB in [6] with the same block length and a 256-state TCQ
1
.
2) Since the basic coding framework for our high-rate scheme
is the same as that for the low-rate scheme of [5], an obvious
advantage is the ability to switch between the low and high rate
regime without changing the coding setup. This is particularly
attractive in situations where the transmission rate needs to be
adapted to the instantaneous channel quality.
In the following, we give the details of our proposed DPC
scheme by discussing the encoding and decoding separately.
A. Encoding
The encoding scheme for our DPC scheme is shown in Fig.
2, where we use an IRA code and TCQ as the channel and
source coding component, respectively. In the following, we
briey discuss the two coding components.
1
+
+
+
+
3
2
M
U
X
M
U
X
Conv.
Code
TCQ
Symbol
Mapper
ACC
ACC
TCQ(Viterbi)
+
TCQbit
(ToChannel)
Type-2Paritybits
Degree distribution: 2(x)
Type-1Paritybits
Degree distribution: 1(x)
P
r
e
c
o
d
e
d

I
n
f
o
r
m
a
t
i
o
n

b
i
t
s
Degree distribution: (x)
TCQ
Codebook
Shift
+
,
C
1 2 b b
3 4
2 b b K ( ) + +
d
2
d
1
b
2
b
1
b
3
b
4
S+D
U X
Fig. 2. The proposed DPC encoder. The LDPC precode is not shown here
for the sake of clarity.
Channel Coding: The message sequence is rst encoded using
a rate-R
pre
LDPC precode, the purpose of which is to avoid
error oors, especially for nite block lengths. The precoded
information is then encoded with a rate-R
I
IRA code, which
can be thought of as a concatenation of a low density generator
matrix (LDGM) code and an accumulator. The degree distribu-
tion (from the edge perspective) of the (precoded) information
nodes is denoted by (x) =

Di
d=2

d
x
d1
, where D
i
is the
maximum information node degree, and
d
is the fraction of
edges connected to degree d information nodes. There are a
total of 2L parity bits, with the rst L bits categorized as type-
1, and the remaining L as type-2. The two types of parity bits
follow different degree distributions
i
(x) =

Di,p
d=1

d(i)
x
d1
,
i = 1, 2, where D
i,p
is the maximum degree for parity nodes
of type-i and
d(i)
is the fraction of type-i edges connected
to parity nodes of degree d. Each parity bit stream is rst
1
We expect the gap of our scheme to be approximately 1.10 dB with a
256-state TCQ. By using a multi-stage decoding approach instead of BICM,
we expect the gap to reduce even further.
randomly interleaved and then accumulated. At any given time
instance, elements from these two accumulated bit streams are
used to select one out of four TCQ codebooks, as will be
explained in the source coding part. The overall rate of the
dirty-paper code can be calculated as R = 2 R
pre
R
I
b/s.
Note that with the proposed scheme, the maximum achievable
rate is 2 b/s as opposed to 1 b/s for the schemes of [5], [7].
The same methodology can be used to devise schemes for
even higher transmission rates by increasing the number of
types of parity bits.
Source Coding: For practical source coding, we employ TCQ,
the strongest quantization code known in the literature. The
input to be quantized is V = S + D, where is Costas
MSE scaling factor [4], S is the side-information sequence,
and D is the random dither shared by the encoder and the
decoder. The basic component of the TCQ code is a 1/2
convolutional code, which outputs the bit streams b
3
and b
4
.
At a time instance n, n = 1, . . . , L, we let the channel coded
bits b
1
[n] and b
2
[n] select one out of four TCQ codebooks
to which the convolutional code outputs are mapped. Since
the interference can have any arbitrary variance, each TCQ
codebook is replicated innitely in both directions. Amongst
the replicated copies of the codebook, the one closest to
V [n] is chosen. Since the performance of the underlying TCQ
quantization codebook is independent of any shifts, we let
the four TCQ codebooks be shifted versions of each other, as
indicated by Fig. 2. Mathematically, the shifted TCQ symbol
at time n is given by (we omit the time indices of the bits
b
i
[n], i = 1, 2, 3, 4 for notational convenience)
U[n] =
_
(b
3
+ 2b
4
) +C
b1,b2
+ 4K[n]

, (4)
n = 1, . . . , L, where is the step size of the TCQ codebook,
C
b1,b2
R is the relative shift of the TCQ codebook as a
function of the channel coded bits b
1
and b
2
, and K[n] is in-
dicative of the fact that the codebooks are replicated innitely
with the copy closest to V [n] selected. Mathematically,
K[n] = arg min
kZ

(b
3
+2b
4
)+C
b1,b2
+4kV [n]

2
. (5)
Given the channel coded bit streams b
1
and b
2
, the quanti-
zation procedure uses the Viterbi algorithm to choose the bit
streams b
3
and b
4
such that the overall MSE is minimized.
Finally, the error sequence X = UV is transmitted over the
channel. Note that the transmitter power is in fact the overall
MSE. Thus for a given constraint on the transmit power, one
needs to search for the TCQ step size for which the MSE
is equal to the required power.
B. Decoding
The DPC decoding scheme is shown in Fig. 3. The received
sequence is rst scaled by followed by dither removal to
obtain Y

= Y + D. The decoding is done by iterative

message passing between the BCJR algorithm (on the com-
bined trellis of the convolutional code and the accumulators)
and the BP algorithm (on the combined decoding graph of the
LDGM code and the LDPC precode). The BCJR algorithm
evaluates the extrinsic LLRs on the data sequences d
1
and d
2
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings
with the respective a priori LLRs coming from the decoding
graph (these a-priori LLRs are initialized to zero for the rst
iteration) and with Y

as the channel input. The extrinsic LLRs

from the BCJR algorithm are then fed into the BP algorithm as
the a-priori LLRs. The extrinsic LLRs from the BP algorithm
are then used as a-priori LLRs to the BCJR completing one
iteration of message passing. The iterations continue until
some stopping criteria is met
2
. The BP algorithm also outputs
the LLRs on the original (precoded) information bits which
are then used to form the hard decisions.
BCJR
TCQ + ACC
x
Lap(2)

Y
r
+
D
r
' Y
r
BP
LDGM + Precode
Lap(1)
Lext(2)
Lext(1)
L
ap(2)
L
ap(1)
I
ext(1)
I
ext(2)

3
-1

2
-1

3
m
Fig. 3. DPC decoder.
C. DPC Design
As mentioned earlier, the DPC scheme consists of a source
coding and a channel coding component. The design for
the source code involves choosing appropriate shifts C
b1,b2
,
b
1
, b
2
{0, 1} of the TCQ codebook, whereas the channel
code design involves choosing appropriate degree distributions
for the LDGM code. In the following, we discuss these two
aspects separately.
Source Code Design The relative shifts C
b1,b2
, b
1
, b
2
{0, 1}
of our scheme are analogous to the TCQ codebook shift men-
tioned in [5]. The shifts characterize the TCQ induced channel
and control the amount of information the induced channel
conveys about the original message. In order to evaluate this
information as a function of these relative shifts, we utilize the
EXIT chart strategy and the area theorem [9]. If I
EB
(I
AB
)
is the extrinsic information conveyed by the BCJR decoder
as a function of the a-priori information I
AB
, the channel
capacity can be approximated by the area under the curve.
For some given shifts C
b1,b2
, b
1
, b
2
{0, 1}, and the a-priori
information I
AB
on the two bit-streams d
1
and d
2
, we evaluate
the extrinsic information I
EB
by Monte-Carlo simulations.
The a-priori LLRs on the two bit streams are assumed to have
an identical Gaussian distribution, with mean = J
1
(I
AB
)
and variance 2, where I = J() is the information conveyed
by a Gaussian LLR of mean and variance 2. We point out
that since I
EB
is the sum of information conveyed about two
bits, the maximum value it takes is two.
In order to illustrate the signicant difference that an appro-
priate choice of the shifts can make, we show the EXIT curves
I
EB
(I
AB
) for two sets of shifts in Fig. 4. It can be observed
that the second set of shifts gives a gain of approximately
0.37 bits over the uniform shifts of the rst set. The choice
in the second set was made with the help of a very limited
search, and we believe that additional gains can be attained by
performing a more exhaustive search over the relative shifts.
2
In our simulations, we let the iterations continue until a maximum number
of iterations is reached, or when a valid codeword has been decoded.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
I
AB
I
E
B
C
0,0
= 0.5; C
1,0
= 0.75; C
0,1
= 1.00; C
1,1
= 1.25 : Approx. Cap = 0.627
C
0,0
= 0.5; C
1,0
= 1.30; C
0,1
= 0.7; C
1,1
= 0.9 : Approx. Cap = 0.998
Fig. 4. EXIT curves for the BCJR decoder for two sets of shifts of the TCQ
codebook. The SNR is xed at 5.2 dB and the TCQ uses a 64-state feedback
convolutional code dened by the polynomials h
0
= 127 and h
1
= 42 in
octal form.
Channel Code Design We once again use the Gaussian
assumption and the EXIT chart strategy to design the degree
distributions for the LDGM code. We x the check node
proles
1
(x) and
2
(x) and optimize the variable node
prole (x). Note that the TCQ induced channel might not
be symmetric, and hence, in general the usual assumption of
the all-zero codeword being transmitted is not valid. However,
we use the concept of i.i.d. channel adapters [10] based on
which one can use the all-zero codeword assumption. Given
an a-priori information I
AC
to the check nodes, we nd the
probability mass function (PMF) P
C(i)BCJR
of the extrinsic
LLRs from the type-i check nodes (i = 1, 2) to the BCJR
algorithm using discretized density evolution (DDE) [11]. The
a-priori LLRs to the check nodes are assumed to follow a
Gaussian distribution P
AC
corresponding to the information
I
AC
. We then use Monte-Carlo simulations to nd the PMFs
P
BCJRC(1)
and P
BCJRC(2)
of the LLRs from the BCJR
decoder to the two types of check nodes. Using these PMFs,
as well as P
AC
, we once again utilize DDE to evaluate the
extrinsic information I
EC
from the check nodes.
At the variable node decoder, we also employ the EXIT
chart strategy coupled with the Gaussian assumption. If I
AV
is the a-priori information to the variable nodes, the extrin-
sic information as a function of the a-priori information is
evaluated as I
EV
(I
AV
) =

Di
d=1

d
J
_
(d 1)J
1
(I
AV
)
_
.
For convergence of BP decoding (in the Gaussian assumption
and EXIT function sense), the following constraint should be
satised for all I
AC
[0, I
max
]
I
EV
_
I
EC
(I
AC
)
_
> I
AC
, (6)
where it is assumed that the precode is able to correctly decode
the message sequence when I
AC
> I
max
. In addition to the
convergence constraint, we require

Di
d=1

d
= 1, and
d
> 0,
d = 1, . . . , D
i
. For these constraints one should maximize the
rate of the LDGM code, which is equivalent to maximizing

Di
d=1

d
d
. By discretizing the interval I
AC
[0, I
max
] and
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings
requiring the constraint (6) to be satised for all values of
the discretized values, the optimization problem can be easily
solved using linear programming.
IV. SIMULATION RESULTS
We provide simulation results when the channel parameters
are xed at a = b = 0.5. For the primary user, we x the
transmission rate at R
p
= 1 b/s. We use the BICM scheme
with a 4-PAM modulation and a rate-1/2 LDPC code with the
optimized degree proles of [8]. Simulation results in Fig. 5
indicate that in the absence of the cognitive user, and a block
length of L = 50, 000, the transmission power required to
achieve a target BER of at least 1 10
5
is approximately
5.95 dB. Thus, in our setup we always x P
p
= 5.95 dB. For
the DPC scheme, we use a 64-state feedback convolutional
code for TCQ with polynomials h
0
= 127 and h
1
= 42 in
octal form. The relative shifts of the TCQ codebook are xed
to the second set in Fig. 4, i.e. C
0,0
= 0.5, C
1,0
= 1.3,
C
0,1
= 0.7, and C
1,1
= 0.9. For the LDGM code
in the DPC scheme, we x the check node degree proles
to
1
(x) = 0.1667 + 0.6667x + 0.1667x
2
, and
2
(x) =
0.1364 + 0.1818x + 0.6818x
2
. The optimized variable node
degree prole for I
max
= 0.95 is given as (x) = 0.2864x +
0.3188x
2
+0.0254x
3
+0.1314x
23
+0.0061x
26
+0.2320x
136
,
such that the LDGM code rate is R
LDGM
0.526. For the
precode, we use a rate-0.95 LDPC code with degree proles

pre
(x) = x
79
and
pre
(x) = 0.6x
2
+0.4x
7
. Thus the overall
transmission rate for the cognitive user is R
c
= 1.0 b/s.
The theoretical limit for the cognitive user required for this
transmission rate is evaluated from (3) as P
c,min
= 6.67 dB,
which ensures that the received SNR at the secondary receiver
is (1

)P
c,min
= 4.77 dB. The BER for the cognitive user
as a function of the received SNR (1

)P
c
is shown in Fig.
5. We also plot the BER for the primary user as a function of
(1

)P
c
when the primary user power is xed at P
p
= 5.95
dB. Since the primary user BER remains below 1 10
5
, we
satisfy the requirement that the primary user performance is
not affected by the presence of the cognitive user. For the
cognitive user we simulated a total of 500 blocks of length
L = 50, 000 with a cognitive user power of P
c
= 8.25
dB and found no errors at the corresponding decoder. The
aforementioned P
c
results in received SNR at the secondary
destination being (1

)P
c
= 6.0 dB, which is 1.23 dB from
capacity.
V. CONCLUSIONS AND FUTURE WORK
We have implemented the DPC based scheme of [1] for
the CRC. We have considered the case where the cognitive
radio is already aware of the primary user message and that
the primary user and destination are unaware of the existence
of the cognitive radio. For DPC, we have proposed a scheme
which employs an IRA code for channel coding and TCQ for
source coding. Using a 64-state TCQ, the scheme operates
within 1.23 dB of the capacity at a transmission rate of 1.0
b/s and a block length of 50,000. Possible extensions to this
work is to consider a scenario where the cognitive user does
not initially know the primary user message, but knows its
5.8 5.85 5.9 5.95 6 6.05 6.1
10
6
10
5
10
4
10
3
10
2
10
1
P
p
/ (1*)P
c
(dB)
B
E
R
Primary User BER No Cognitive user
Cognitive User BER
Primary user BER vs. (1*)Pc; Pp = 5.95 dB
Fig. 5. Bit-error rates for the primary and cognitive users at a block length
of L = 50, 000. The channel parameters are xed at a = b = 0.5. The
dashed part for the cognitive user curve is indicative of the fact that we got
no bit errors at (1

)Pc = 6 dB after simulating a total of 500 blocks.

The theoretical limit for both users is 4.77 dB.
codebook. Hence the rst phase for the cognitive user would
be to listen in on the primary user transmission and attempt
to decode its message before it begins its own transmissions.
Such a scheme would obviously require the primary user code
to be rate-compatible. Another direction of future research
is to improve the performance of the dirty-paper code by
considering multi-stage decoding involving two IRA codes
instead of one in the current scheme. Yet another extension
is to consider the general scheme of [2] which involves joint
decoding of multiple dirty-paper codes.
REFERENCES
[1] A. Jovicic and P. Viswanath, Cognitive Radio: An Information-
Theoretic Perspective,, IEEE Trans. Info Theory, vol. 55, no. 9, pp.
39453958, Sept. 2009.
[2] J. Jiang and Y. Xin, On the achievable rate regions for interference
channels with degraded message sets, IEEE Trans. Info Theory, vol.
54, no. 10, pp. 47074712, Oct. 2008.
[3] S. Seyedmehdi, J. Jiang, Y. Xin, and X. Wang, An improved achievable
rate region for causal cognitive radio, Proc. ISIT 2009, Seoul, Korea,
July 2009.
[4] M. Costa, Writing on dirty paper, IEEE Trans. Inform. Theory, vol.
29, pp. 439-441, May 1983.
[5] Y. Sun, Y. Yang, A.D. Liveris, V. Stankovi c, and Z. Xiong, Near-
capacity dirty-paper code design: A source-channel coding approach,
IEEE Trans. Info. Theory, vol. 55, no. 7, pp. 30133031, July 2009.
[6] Y. Sun, M. Uppal, A. Liveris, S. Cheng, V. Stankovi c , and Z. Xiong,
Nested turbo codes for the Costa problem, IEEE Trans. Comm., vol.
56, no. 3, pp. 388-399, Mar. 2008.
[7] U. Erez and S. ten Brink, A close-to-capacity dirty paper coding
scheme, IEEE Trans. Info. Theory, vol. 51, pp. 34173432, Oct. 2005.
[8] H. Sankar, N. Sindhushayana, and K. Narayanan, Design of low-density
parity check codes for high order constellations, Proc. Globecom 2004,
Dallas, TX, Nov. 2004.
[9] A. Ashikhmin, G. Kramer, and S. ten Brink,Extrinsic information
transfer functions: Model and erasure channel properties, IEEE Trans.
Info. Theory, vol. 50, no. 11, pp. 26572673, Nov. 2004.
[10] J. Hou, P.H. Siegel, L.B. Milstein, and D. Pster, Capacity approaching
bandwidth-efcient coded modulation schemes based on low density
parity-check codes, IEEE Trans. Info. Theory, vol. 49, no. 9, pp. 2141
2155, Sept. 2003.
[11] S. Chung, G. Forney Jr., T. Richardson, and R. Urbanke, On the design
of low-density parity-check codes within 0.0045 dB of the Shannon
limit, IEEE Commun. Letters, vol. 5, no. 2, pp. 5860, Feb. 2001.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings