3620

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009

Bounds and Lattice-Based Transmission Strategies

for the Phase-Faded Dirty-Paper Channel
Amir Bennatan, Member, IEEE, Vaneet Aggarwal, Student Member, IEEE, Yiyue Wu, Student Member, IEEE,
A. Robert Calderbank, Fellow, IEEE, Jakob Hoydis, Student Member, IEEE,
and Aik Chindapol, Senior Member, IEEE

AbstractWe consider a fading version of the dirty-paper

problem, as proposed by Grover and Sahai. In this formulation, the various signals involved are complex-valued, and the
interference (known only to the transmitter) is multiplied by a
random complex-valued coefficient, whose phase is known only
to the receiver. We focus on a compound channel formulation,
and seek to maximize the worst-case performance. We present
an achievable strategy modeled on the lattice-based approach
of Erez, Shamai and Zamir and propose heuristic methods to
optimize its parameters. We also derive an upper bound on the
maximum achievable transmission rates. Our bounds are shown
to be tight in some settings, yielding a complete characterization
of capacity. We also provide simulation results, indicating the
practical effectiveness of our approaches.
Index TermsDirty-paper, cognitive radio, fading channels,
nested lattices.

I. I NTRODUCTION
IRTY-PAPER problems have attracted substantial attention recently, due to a variety of applications including
MIMO broadcast [5],[16], digital watermarking [6] and cognitive radio [15]. In the classic formulation of the problem [7],
the channel equation is given by,

Y =X +S+Z

(1)

where X is the transmitted signal, subject to an average

power constraint PX , Z is additive noise and Y is the
channel output. S is random interference, whose realization
is assumed to be known to the encoder (non-causally). We
assume that Y, X, S, and Z are complex-valued (Costa [7]
assumed that the random variables are real-valued, but his
discussion is straightforwardly applicable to this setting as
well). We model S and Z to be complex, zero-mean circularlysymmetric Gaussian random variables, with variances PS and
Manuscript received April 24, 2008; revised January 7, 2009; accepted
February 28, 2009. The associate editor coordinating the review of this paper
and approving it for publication was A. Grant.
A. Bennatan is with the Program in Applied and Computational
Mathematics, Princeton University, Princeton, NJ 08544, USA (e-mail:
abn@math.princeton.edu).
V. Aggarwal, Y. Wu and A. R. Calderbank are with the Department of
Electrical Engineering, Princeton University, Princeton, NJ 08544, USA (email: {vaggarwa, yiyuewu}@princeton.edu; calderbk@math.princeton.edu).
J. Hoydis is with SUPELEC, Plateau de Moulon 3 rue Joliot-Curie, 91192
GIF SUR YVETTE CEDEX, France (e-mail: jakob.hoydis@rwth-aachen.de).
A. Chindapol is with Nokia Siemens Networks, 575 Herndon Pkwy, Ste
200, Herndon, VA 20170 (email: aik@ieee.org).
The work of A. R. Calderbank is supported in part by NSF under grant
0701226, by ONR under grant N00173-06-1-G006, and by AFOSR under
grant FA9550-05-1-0443.
Digital Object Identifier 10.1109/TWC.2009.080569

PZ , respectively. Since S is known to the transmitter, we

would expect that it would be able to pre-subtract it from
its transmitted signal, thus eliminating its interference. Doing
so, however, would involve violating its power constraint. A
remarkable result by Costa [7] shows that using an elaborate
transmission method, it is possible to effectively pre-subtract
S without violating the power constraint.
Mitran et al. [14] generalized Costas formulation, by
multiplying X and S by random fading coefficients 1 and
2 (respectively) whose realizations are known only to the
receiver. This models the fading typically experienced over
wireless channels. Specifically, in cognitive radio, X and S
are produced at different transmitters and thus experience
different fades (unlike wireless broadcast, see e.g. [2], [20]).
The receiver is often able to estimate the values of 1 and 2
from its received signal and is thus aware of their realizations.
In some formulations of the problem, a feedback link to the
transmitter is assumed to exist whereby the transmitter is made
aware of the realizations of 1 and 2 . In this paper we
focus on the setting where such a link does not exist, or the
values of 1 and 2 fluctuate too quickly for the link to be
useful. Mitran et al. [14] modelled 1 and 2 as quasi-static,
meaning that they are random, but remain constant for the
duration of the transmission. They assumed that 1 and 2
were real-valued Rician random variables. This modeled lack
of knowledge of the magnitude of the fade at the transmitter.
Khina and Erez [12] considered a similar formulation where
1 was constant and 2 was real-valued and random. In their
work, they showed that proactively injecting randomness into
the transmission strategy can produce a substantial benefit.
In this paper, we focus on the following channel model,
as introduced by Grover and Sahai [11], which focuses on
lack of phase knowledge at the transmitter. We thus refer to
it as the phase-faded dirty-paper channel model. The channel
is defined by the following equation.
Y = X + ej S + Z

(2)

where is quasi-static and random. Different values of

correspond to different channel realizations, each capable
of supporting different transmission rates. In this paper, we
follow [11] (as well as Weingarten et al. [17]) and focus
on a compound channel formulation1 (rather that outage1 The terminology of [11] is somewhat confusing. Although they mention an
outage probability formulation, a close examination of their work reveals that
they focus on C(), which corresponds to a compound channel formulation.

c 2009 IEEE
1536-1276/09$25.00 

BENNATAN et al.: BOUNDS AND LATTICE-BASED TRANSMISSION STRATEGIES FOR THE PHASE-FADED DIRTY-PAPER CHANNEL

probability). With this formulation, a rate is achievable if a

transmission scheme exists that supports it for every possible
realization of (i.e., we are interested in worst-case performance). Capacity is defined as the maximal such achievable
rate. Specifically, we seek to design transmission schemes
that maximize the minimal achievable rate, evaluated over all
possible . Note that in this context, the precise distribution
of is of no consequence, and only its support matters (the
range of values that may take). We let A denote this support,
and C(A) denote the corresponding capacity.
Grover and Sahai [11] obtained lower bounds on capacity
for the case of A being an interval of the form [, ].
In this paper, we derive bounds which are based on an
adaptation of the nested-lattices approaches of Erez, Shamai
and Zamir [8] (for the classic no-fading channel) and Khina
and Erez [12] (for the case of magnitude rather than phase
fading). Our bounds are valid for arbitrary A, and outperform
those of [11] in the case of A = [, ]. The bounds
are based on practical transmission strategies, and we present
simulation results that confirm their effectiveness.
Grover and Sahai [11][Theorem 1] also obtained an upper
bound on capacity for the case2 of A = {0, }. In this paper,
we extend their bounds to arbitrary A, and improve them at
high values of interference power PS .
Our lower bound is provided in Sec. II and our upper bound
in Sec. III. In Sec. IV we examine the bounds in various
settings, and provide simulation results. Sec. V concludes the
paper.
In this paper, we use boldface to denote vectors and normal
face to denote scalar values. (x) denotes the real part of a
complex value x. All logarithms are assumed to have base 2.
II. ACHIEVABLE R ATES FOR P HASE -FADED D IRTY PAPER
We begin in Sec. II-A by describing an algorithm for
transmission of the phase-faded dirty paper channel. Our
analysis of this algorithm in Sec. II-B will produce a lower
bound on capacity.
A. Algorithm
The proposed algorithm is modeled on the approach of Erez,
Shamai and Zamir [8]. Letting n denote the transmission block
length, we begin by breaking up the transmission into n/k
blocks of time, each of length k (we assume that n is an
integer multiple of k). We use a k-dimensional lattice to
construct the transmitted signal at each time block. Unlike [8],
our lattice is defined over the complex number field rather than
the real number field. This choice results from the channel
model (Sec. I), which is different from the channel of [8]. We
assume the use of a code C which will be defined later. The
encoding and decoding processes now proceed as follows.
Transmitter. The transmitter begins by selecting a codeword c C. At time block i, i = 1, ..., n/k, the
transmitter sends the signal xi , constructed as follows.
xi = [ci i si di ] mod

(3)

2 The notation of [11] is slightly misleading, but a close examination of the

proof of Theorem 1 reveals that it is valid for this choice of A alone.

3621

where ci and si are the portions of c and s (respectively)

corresponding to time block i, and di is a dither signal,
which is randomly selected with uniform distribution
within V (independently between time blocks). di is
assumed to have been selected in advance and is thus
known to the decoder. i is a parameter which will be
determined later. For an arbitrary vector u, we define [u]

mod = u Q (u), where Q (u) is the lattice vector

that is nearest to u in terms of Euclidean distance.
Receiver. The receiver obtains y = (y1 , ..., yn/k ). At
time block i, it computes yi = [i yi + di ] mod . i
is a parameter that will be determined later, whose value
may depend on (which is assumed to be known to the
receiver) and on i .
Evaluating yi (using (2) and (3)), we obtain the sequence
of equalities ending with (4), on the following page. Note
that in this derivation, we have applied the modulo-lattice
distributive property [3],
[(x

mod ) + y]

mod = [x + y]

mod

The receiver now attempts to decode the codeword c C


]. Methods for doing so will be
using y = [y1 , ..., yn/k
discussed in the subsequent sections.
The second moment per-dimension of a lattice , is defined
by,

1 1
x2 dx
(5)
2 =
kV V

where V = Vol(V). With this definition, k 2 is the second

moment of a random vector which is uniformly distributed
within V. By virtue of the above-mentioned distribution of
di , we obtain that xi is also distributed uniformly in V.
Following [8], we restrict to satisfy 2 = PX , thus ensuring
that x complies with the average power constraint PX .
In [8], i was constant, independent of i and set to
PX /(PX + PZ ), and i was chosen to equal i . Following
the example of [12] (in the case of perfect phase knowledge and partial magnitude knowledge), we select parameters
1 , ...n/k randomly, as explained in the following section.
This choice is also similar to the time-sharing concept applied
by Grover and Sahai [11][Sec. V.B]. Parameters 1 , ..., n/k
will be non-random functions of 1 , ...n/k and of the channel
realization (which is known to the receiver).
B. Performance Analysis and Parameter Selection
The selection of parameters = [1 , ..., n/k ] and =
[1 , ..., n/k ] of the algorithm of Sec. II-A determine decoders
performance. Recall from Sec. I that we are interested in
the worst-case performance (the minimal achievable rate,
evaluated over all values of ) for given and . Our
objective is to select and so as to maximize it.
To guide our selection, we derive a lower bound on the
achievable rate for given and . Our bound is based on
lattice analysis techniques, as in [8]. We begin by rewriting (4)
as,
yi = [ci + zi ] mod

(6)

3622

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009

yi

[i yi + di ] mod

[i xi + i ej si + i zi + di ] mod

=
=

[i si + (i si + di ) + (i ej i ) si + i zi ] mod
[i si + (ci si ) + (i ej i ) si + i zi ] mod

[ci (1 i ) si + (i ej i ) si + i zi ] mod

where zi is the effective noise, given by zi = (1 i )

xi + (i ej i ) si + i zi . Note that the role of xi in
this context is equivalent to that of noise (it is referred to as
self-noise in [8]). zi can easily be shown to be independent of
ci . To see this, observe that its components si , zi and xi are
independent of one another and of ci (xi is independent of ci
by virtue of it being a function of the random dither di ). For
given , zi has variance,


PZ = k |1 i |2 PX + |i ej i |2 PS + |i |2 PZ
(7)
Following [8], we define the normalized lattice second moment G() by,

G() =

2
V 1/k

where 2 and V are defined as in Sec. II-A. Note that

this definition, as well as the discussion below, is slightly
different from the one of [8], and is tailored to the use of
lattices over the complex number field, rather than the real
number field. However, the translation between the two is
immediate [10][Sec. II.D]. Armed with these definitions, we
are now ready for the following lemma.
Lemma 1: Assume a fixed value of the the channel coefficient . Assume the components of are selected randomly
and independently according to some distribution, and that the
components of are given by i = f (i , ) where f () is
a fixed, non-random function. Then the following value is a
lower bound on capacity,
 


PX
log
RLB () = E
log(eG())
PZ (, , )
+
(8)

(4)

also assumes the use of maximum-likelihood decoding to

recover c from y .
G() can be shown [19] to be lower bounded by 1/(e),
and thus the second term within the brackets in (9) is always negative. However, Zamir and Feder [18] proved the
existence of optimal sequences of lattices {(k) } satisfying
limk G((k) ) = 1/(e) (The actual figure in [18] was
1/(2e), but this figure was normalized to account for the
switch from the real number field to the complex number
field). With such sequences, this term can be made arbitrarily
small.
Although we have not proven that the lower bound (9) is
tight, we still rely on it as a useful figure of merit when
selecting the values of and . We begin by defining f ()
which determines . By (9), maximizing RLB () is equivalent
to minimizing PZ (, , ).
For given values of and as well as the fixed problem
parameters PX , PZ and PS , simple algebraic manipulations
show that the value of PZ is minimized by,
= f (, ) =

PX + ej PS
PX + PS + PZ

(10)

With this choice,

PZ (, , ) = (PX + ||2 PS )

|PX + ej PS |2
PX + PS + PZ

(11)

Determination of a good choice for , however, is more

involved. Unlike , which is computed at the decoder, is
used at the encoder, which does not have knowledge of the
realization of . Its value must therefore be determined based
on the distribution of rather than its realization. In this paper
we have found the following choice for to be useful,

=
ej

(12)

where is a random variable whose distribution is that of the

components i of , and = f (, ). PZ (, , ) is given
by the right hand side of (7), with i and i being replaced

by and . For any real-valued x, [x]+ = max(x, 0).

where is distributed identically as (and is independent

of it) and
is a real-valued non-random scalar in the range
[0, 1]. In Sec. IV-A we selected
so as to maximize RLB (A)
as given by (9).

In the sequel, we will be interested in the worst-case lower

bound, RLB (A), defined as,

III. U PPER B OUND ON THE ACHIEVABLE R ATES

RLB (A) = min {RLB ()}

A

(9)

The proof of Lemma 1 generalizes the proof of [8][Theorem 3]

and is omitted. It considers as a form of side information
that is available to the receiver. The code C in the proof
is constructed randomly. That is, each codeword is broken
up into its component ci , i = 1, ..., n/k and ci is selected
independently, with uniform probability within V. The proof

We now turn to provide an upper bound on the achievable

rates over the phase-faded dirty paper channel. Our bound is
valid for any set A of values for .
Theorem 1: The capacity of the compound phase-faded
dirty paper channel, where is confined to a set A, is upper
bounded by expression (13) on the following page, where
J(; PX , PS , PZ ) is defined by (14) on the same page, and
where the operation PS  defines the largest non-ascending

BENNATAN et al.: BOUNDS AND LATTICE-BASED TRANSMISSION STRATEGIES FOR THE PHASE-FADED DIRTY-PAPER CHANNEL

function of PS which is upper bounded by the contents of the

brackets. Specifically,

4.5

tx

Rate (bits/channel use)

3.5

2.5
2
1.5

0.5
0

10
P (dB)

15

20

25

30

Fig. 1.

Bounds for A = {0, }

4.5

Upper bound
Lower bound
Lower bound, constant
Time sharing

4
3.5
3
2.5
2
1.5
1

IV. N UMERICAL R ESULTS

0.5

A. Comparisons of the Various Bounds

In Figs. 1, 2 and 3, we have plotted the bounds considered
in this paper. In all plots, PX = 12 dB and PZ = 0 dB, and
performance is plotted as a function of PS .
Fig. 1 corresponds to a choice of A = {0, }. The lower
bound is based on the analysis of Sec. II and the upper bound
is based on Sec. III. For the lower bound, parameters
and were computed as in Sec. II-B, with
determined
by exhaustive search over a fine grid. The constant- curve
corresponds the lower bound of Lemma 1 with constant nonrandom, equal to
ej0 where 0 is fixed at the mean value
of A. is computed using (10). Also plotted is the trivial
time-sharing curve. This curve corresponds to the achievable
rates when the available time is split equally between the two
possible realizations of . Specifically, with time-sharing, the
transmitter sends its message twice, once for each possible
value of and the receiver listens only in the interval corresponding to the true realized . Lastly, we have plotted the
results of simulation results, which are discussed in Sec. IV-B.
At high values of the interference PS , the upper and lower
bound curves coincide with the time sharing curve. Using (13),
this can easily be shown to hold for values of PS greater
than PX + PZ . At such values of PS , the channel capacity
is completely characterized and is achieved by time sharing.
At low values of PS , the achievable rates of Sec. II as well
as the simulation results outperform the time-sharing curve.
Even in this regime, the gap between the lower and upper
bound is small, indicating the effectiveness of our bounds in

Rate (bits/channel use)

Examining (13), it is easy to observe that J(2 1 ) is

minimized by the pair 2 and 1 that have maximal angular
separation, defined as one minus the cosine of their difference.
The proof of this theorem is provided in Appendix A. It is a
generalization of the bound of Grover and Sahai [11] for the
case of A = {0, }, which in turn is a generalization of the
Khisti et al. [13][Theorem 1].
The operation PS  in (13) converts the contents of the
brackets, which is generally not a non-ascending function
of PS , to a non-ascending function. Grover and Sahai [11]
(for the special case of {0, }) did not apply this
operation, but nonetheless conjectured that the capacity is a
non-ascending function of the interference PS . This conjecture
is easily justified using the following argument, which assumes
common-randomness at the transmitter and receiver. A transmitter and receiver operating on a channel characterized by
interference power PS can emulate a channel characterized

by PS for any PS > PS by randomly generating a vector S


with power PS PS . The receiver, knowing , would add
to its received Y, and the transmitter would simply
ej S

to be a virtual interference. Thus, any rate

redefine S = S + S
achieved over a channel characterized by PS can be achieved
at PS < PS , and therefore the channel capacity is necessarily
non-ascending as a function of PS .

Upper bound
Lower bound
Lower bound, constant
Time sharing
Simulations

[f (x)]x = min f (t)

3623

Fig. 2.

10
PS (dB)

15

20

25

30

Bounds for A = {0, 2/3 }

this setting. The gap between the lower bound curve and the
constant- lower bound curve is an indication of the merit of
selecting randomly.
Fig. 2 corresponds to a choice of A = {0, 2/3}. The three
curves are similar to the ones in Fig. 1. Our bounds in this case
are less tight, and a gap exists between the upper and lower
bounds. The lower bound approaches the time-sharing curve
at high values of PS , while the upper bound is substantially
higher. We conjecture that the true capacity is achieved by
time-sharing as PS . Interestingly, the constant- lower
bound curve outperforms the non-constant- lower bound at
low values of PS , indicating that the distribution by which
is generated may still be optimized.
Finally, Fig. 3 corresponds to a choice of A = [/2, /2].
Beside the bounds of Sections II and III, we have also plotted
the lower bound of Grover and Sahai [11][Theorem 2]. Our
lower bound coincides with that of [11] at low values of PS ,
but outperforms it substantially at the limit of high PS . The
bound of [11] coincides almost precisely with our constant-
lower bound, indicating that their bound could possibly be
improved by similarly generating its parameters randomly.
Interestingly, at an interval surrounding PS = 2 dB, the

3624

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009



C + (A) = min J(2 1 ; PX , PS , PZ ),

J(; PX , PS , PZ ) =

4.5

Rate (bits/channel use)

3.5

1
log

2PZ PS (1 cos )

2.5
2
1.5
1
0.5

Fig. 3.

10
PS (dB)

15

20

25

(13)

(14)

(5, 6, 7, 8, 20) = (0.09973, 0.02331, 0.5885, 0.1833,

0.1051). These edge distributions determine the symbol-rate
of the code, which is 0.6 symbols per channel use. The
corresponding bit rate depends on the alphabet used. We
experimented with two constellations: 16-QAM and 32QAM, corresponding to rates 2.4 and 3 bits per channel
use, respectively. The constellations points are provided in
Appendix B. The block lengths were 100,000 and 50,000
symbols, for the rate 2.4 and rate 3 simulations, respectively.
The decoder was modeled on the decoder of [1]. In the
computation of the initial messages [1][Equation (7)], the
probabilities were computed as follows,


 PS 
PX
log 1 +
PZ


  
2

PX + PS + PZ + 2 cos 2  PX PS

Upper bound
Lower bound
Lower bound, constant
Lower bound of Grover and Sahai

2 , 1 A,

30

Bounds for A = [/2, /2]

various bounds are remarkably tight, indicating that capacity

is almost completely characterized in this regime.
B. Simulations
We now describe simulation results using the algorithm of
Sec. II-A. In these simulations, we confined our attention
to a support set A = {0, }. We experimented at PX =
12 dB, and PZ = 0 dB. We focus on two rates: 2.4 and 3
bits per channel use. The interference powers were set to
PS = -2.89 dB and PS = -7.17 dB for the rate 2.4 and
rate 3 simulations, respectively. Our choice to focus on the
low interference regime follows the fact that at high levels of
interference, capacity is easily achieved by time sharing (see
Sec. IV-A). The results are plotted in Fig. 1. The gap between
the performance we obtained and the theoretical lower bound
curve is primarily attributed to the sub-optimal choice of lattice
and code that were used, as explained below. These choices
were necessary to enable a manageable decoding complexity.
We let be a scalar lattice (i.e. k = 1), = A(Z + j
Z) where A is a constant selected so that
the lattice second
moment 2 (see (5)) equals PX , i.e. A = 6PX . A similar
choice for the non-fading dirty-paper channel was made by [9]
and [4]. Parameters and , explained in Sec. II-A, were
selected as in Sec. II-B, with
= PX /(PX + PZ ).
To obtain rates greater than 1 bit ber channel use (as
required to approach capacity), we used nonbinary codes.
We chose the code C to be a non-binary LDPC code over
a Galois field (GF), see [1]. The edge distributions that
were used (borrowed from [1]) were, (2, 5, 6, 16, 30) =
(0.5768, 0.1498, 0.07144, 0.1045, 0.09752)
and

P [y | (i)] = f (y (i) mod A; PZ )

where PZ is given by (11), f (x; PZ ) is the pdf of a zero-mean,
circularly-symmetric complex Gaussian random variable with
variance PZ . () denotes the mapping from the code alphabet
GF(2m ) to the channel alphabet (2m -QAM) (see Appendix B).
50 simulations were run, 25 per channel realization. The
average symbol errors experienced were 104 with the rate
2.4 simulations, and zero (no simulation symbol was lost) with
the rate 3 simulations.
V. C ONCLUSION
We have examined the compound phase-fading dirty-paper
channel, where the interference, which is known to the transmitter, is multiplied by a complex coefficient whose phase is
random and unknown to the transmitter. In Sec. II and Sec. III
we obtained lower and upper bounds on capacity, which were
shown to be tight in some settings. For example, for the case
when is distributed in {0, }, we have obtained a complete
characterization of capacity whenever PS PX + PZ . For
the case of {0, 2/3}, our bounds are less tight. As
mentioned in Sec. IV-A, we conjecture that at the limit of
high PS , capacity in this setting is achieved by time-sharing.
Further work will focus on proving this by tightening our
upper bound.
A PPENDIX A
P ROOF OF T HEOREM 1
The PS  operation in (13) is easily justified using the
discussion of Sec. III. The second term in the minimization
is simply the achievable rate when the receiver has full
knowledge of the interference S and is able to eliminate it.
We thus focus on the first term. Our bound is a generalization
of the bounds of [11] and [13][Theorem 1].
Let W be a random variable corresponding the message to
be conveyed to the destination in a scheme with rate R and

BENNATAN et al.: BOUNDS AND LATTICE-BASED TRANSMISSION STRATEGIES FOR THE PHASE-FADED DIRTY-PAPER CHANNEL

block length n. W is uniformly distributed in {0, ..., 2nR }.

The transmitted X is a function of W and the interference S.
We define random variables Y , A satisfying,
Y = X + ej S + Z

(15)

For each , Y corresponds to an hypothesis for . Transmission can be viewed as multicast to set of different users,
indexed by A, each receiving a different Y . We assume
that the users experience the same realizations of S and Z.
By Fanos inequality, for transmission to user to be possible
with vanishing probability of error (as a function n), we must
have,
R

1
( I(W ; Y ) + o(n) )
n

where o(n) is a term that approaches zero with n. In the

sequel, to simplify our notation, we neglect this term. We are
interested in worst-case performance (as explained in Sec. I),
and thus we are interested in the minimum of I(W ; Y ).
nR

min I(W ; Y )
A

min min[ I(W ; Y1 ), I(W ; Y2 ) ]

that |ej2 ej1 |2 = 2 (1 cos(2 1 )).

Combining (17), (18) and (19), we obtain:
1
[I(W ; Y1 ) + I(W ; Y2 )]
2
n J(2 1 ; PX , PS , PZ )
where J() is given by (14). Combining this with (16) concludes the proof.
A PPENDIX B
QAM C ONSTELLATIONS U SED IN S IMULATIONS
In this appendix, we provide the QAM constellations used
in the simulations of Sec. IV-B. A list of the points is provided
on the following page. They correspond to a choice of A = 2
(the parameter A was introduced Sec. IV-B), and must be
scaled to suit other choices of A (by multiplying their value
by A/2). The points are ordered using a representation of
GF(2m ) (where m is 4 or 5) as binary vectors of length m
(see [1]). For example, letting () denote the mapping from
GF(16) to the 16-QAM constellation, then (0000) = 0.75
0.75 i, (0001) = 0.75 0.25 i, etc.
ACKNOWLEDGEMENT

1 ,2 A

1
[I(W ; Y1 ) + I(W ; Y2 )]

min
1 ,2 A 2

3625

(16)

We now focus on bounding 1/2 [I(W ; Y1 ) + I(W ; Y2 )].

To further simplify our notation, we let Y1 and Y2 denote
Y1 and Y2 , respectively.
1
[I(W ; Y1 ) + I(W ; Y2 )]
2
1
[h(Y1 ) + h(Y2 ) h(Y1 | W ) h(Y2 | W )]
2
1
1
[h(Y1 ) + h(Y2 )] h(Y1 , Y2 | W )
(17)
2
2
We now bound both terms on the right hand side above, in
the string of equations which ends with (18) on the following
page. In (a), Y,t denotes the component of Y corresponding
to time t. In (b), we have bounded h(Y,t ) by the entropy
of a complex circularly-symmetric Gaussian random variable
with the same variance. PX,t denotes the second moment of
Xt and t denotes the (complex-valued) correlation coefficient
we have applied Jensens inequality.
between Xt and St . In (c) 
n
In (d) we have bounded t=1 PX,t by the power constraint
t as |t |ejt . In (e) we have
nPX and have
2rewritten
j(t  )
evaluated 1/2 =1 e
= 1/2(cos(t 2 )+cos(t
1 )), which can be shown to be maximized (over t ) by
| cos ((2 1 )/2) |. In (f) we have upper bounded |t | by
1. In (g) we have applied 
Jensens inequality, and in (h) we
n
have again upper bounded t=1 PX,t by the power constraint
nPX .
We now turn to the second term in (17) and bound
h(Y1 , Y2 | W ), see the string of equations ending with (19)
on the following page. We begin in (a) by applying a unitary
transformation on the components of Yk and Yl . In (b) we
applied (15). In (c) we have relied on the fact that S is
independent of W , and that Z is independent of X, W and
S. In (d) we have simply evaluated the entropies, observing

Discussions with Oren Somekh are very gratefully acknowledged.

R EFERENCES
[1] A. Bennatan and D. Burshtein, Design and analysis of nonbinary LDPC
codes for arbitrary discrete-memoryless channels," IEEE Trans. Inform.
Theory, vol. 52, pp. 549-583, Feb. 2006.
[2] A. Bennatan and D. Burshtein, On the fading paper achievable region
of the fading MIMO broadcast channel," IEEE Trans. Inform. Theory,
vol. 54, pp. 100-115, Jan. 2008.
[3] R. E. Blahut, Theory and Practice of Error Control Codes. Reading, MA:
Addison-Wesley, 1987.
[4] G. Caire, D. Burshtein and S. Shamai, LDPC coding for interference
mitigation at the transmitter," in 40th Annual Allerton Conf. Commun.,
Cont. Comp., Monticello, IL, Oct. 2002.
[5] G. Caire and S. Shamai, On the achievable throughput of a multiantenna
Gaussian broadcast channel," IEEE Trans. Inform. Theory, vol. 49, July
2003.
[6] A. S. Cohen and A. Lapidoth, The Gaussian watermarking game," IEEE
Trans. Inform. Theory vol. 48, no. 6, June 2002.
[7] M. Costa, Writing on dirty paper," IEEE Trans. Inform. Theory, vol.
29, no. 3, pp. 439-441, May 1983.
[8] U. Erez, S. Shamai, and R. Zamir, Capacity and lattice-strategies for
cancelling known interference," IEEE Trans. Inform. Theory, Nov. 2005.
[9] U. Erez and S. ten Brink, A close-to-capacity dirty paper coding
scheme," IEEE Trans. Inform. Theory, Oct. 2005.
[10] G. D. Forney, Jr., Coset codespart I: introduction and geometrical
classification," IEEE Trans. Inform. Theory, vol. 34, no. 5, pp. 1123-1151,
Sept. 1988.
[11] P. Grover and A. Sahai, What is needed to exploit knowledge of
primary transmissions?" available at arXiv:cs/0702071.
[12] A. Khina and U. Erez, On robust dirty paper coding," in Proc. IEEE
Inform. Theory Workshop 2008, pp. 204-208, May 2008.
[13] A. Khisti, U. Erez, A. Lapidoth, and G. Wornell, Carbon copying onto
dirty paper," IEEE Trans. Inform. Theory., vol. 53, no. 5, pp. 1814-1827,
May 2007.
[14] P. Mitran, N. Devroye, and V. Tarokh, On compound channels with
side information at the transmitter," IEEE Trans. Inform. Theory, vol. 52,
no. 4, pp. 1745-1755, Apr. 2006.
[15] N. Devroye, P. Mitran, and V. Tarokh, Achievable rates in cognitive
radio channels," IEEE Trans. Inform. Theory, vol. 52, no. 5, pp. 18131827, May 2006.
[16] H. Weingarten, Y. Steinberg, and S. Shamai The capacity region of the
Gaussian multiple-input multiple-output broadcast channel," IEEE Trans.
Inform. Theory, vol. 52, no. 9, pp. 3936-3964, Sept. 2006

3626

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 7, JULY 2009

h(Y1 ) + h(Y2 )

(a)

n
2 


(b)

=1 t=1
n
2 


=1 t=1

(c)

(d)

(e)

(f )

(g)

(h)

h(Y1 , Y2 | W )

h(Y,t )
log{ (e)[PX,t + PS + PZ + 2(t ej


PX,t PS )] }




2
2
n
n
1 
1   j 
2n log (e)
PX,t + PS + PZ + 2
t e
PX,t PS
2n
2n
=1 t=1
=1 t=1


 n
 2



1
1  j(t  ) 
2n log (e) PX + PS + PZ + 2
|t |
e
PX,t PS
n t=1
2
=1




 n
 


2 1  
1

2n log (e) PX + PS + PZ + 2
|t | cos
 PX,t PS
n t=1
2




 
  
 1 n 


2
1

2n log (e) PX + PS + PZ + 2 cos
PX,t PS
 n
2
t=1
"


 
 #
n


#


2 1  $ 1
2n log (e) PX + PS + PZ + 2 cos
PX,t PS


2
n t=1
 



 

2 1 
2n log (e) PX + PS + PZ + 2 cos
 PX PS
2

(a)

(b)

=
(c)

(d)

(18)

1
1
h( (Y1 Y2 ), (Y1 + Y2 ) | W )
2
 2


1 j1
1 
j2
j1
j2
h (e e )S, 2X + (e + e )S + 2Z | W
2
 2



1 j1
1 
j2
h (e e )S | W + h 2X + (ej1 + ej2 )S + 2Z | W, S
2
 2



1 j1
h (e ej2 )S + h
2Z
2
n log[(e) (1 cos(1 2 ))PS ] + n log[(e) 2PZ ]

16 QAM
[0.75 0.75 i, 0.75 0.25 i, 0.75 + 0.25 i, 0.75 + 0.75 i, 0.25 0.75 i, 0.25 0.25 i,
0.25 + 0.25 i, 0.25 + 0.75 i, 0.25 0.75 i, 0.25 0.25 i, 0.25 + 0.25 i, 0.25 + 0.75 i,
0.75 0.75 i, 0.75 0.25 i, 0.75 + 0.25 i, 0.75 + 0.75 i]
32 QAM
[0.875 0.875 i, 0.875 0.375 i, 0.875 + 0.125 i, 0.875 + 0.625 i, 0.375 0.875 i,
0.375 0.375 i, 0.375 + 0.125 i, 0.375 + 0.625 i, 0.125 0.875 i, 0.125 0.375 i,
0.125 + 0.125 i, 0.125 + 0.625 i, 0.625 0.875 i, 0.625 0.375 i, 0.625 + 0.125 i,
0.625 + 0.625 i, 0.625 0.625 i, 0.625 0.125 i, 0.625 + 0.375 i, 0.625 + 0.875 i,
0.125 0.625 i, 0.125 0.125 i, 0.125 + 0.375 i, 0.125 + 0.875 i, 0.375 0.625 i,
0.375 0.125 i, 0.375 + 0.375 i, 0.375 + 0.875 i, 0.875 0.625 i, 0.875 0.125 i, 0.875 + 0.375 i,
0.875 + 0.875 i]

(19)

BENNATAN et al.: BOUNDS AND LATTICE-BASED TRANSMISSION STRATEGIES FOR THE PHASE-FADED DIRTY-PAPER CHANNEL

[17] H. Weingarten, S. Shamai, and G. Kramer, On the compound MIMO

broadcast channel," in Proc. 2007 Workshop Inform. Theory Appl., UCSD
Campus, La Jolla, CA, Jan. 2007.
[18] R. Zamir and M. Feder, On lattice quantization noise," IEEE Trans.
Inform. Theory, vol. 42, no. 4, pp. 1152-1159, July 1996.
[19] R. Zamir, S. Shamai, and U. Erez, Nested linear/lattice codes for
structured multiterminal binning," IEEE Trans. Inform. Theory, vol. 48,
no. 6, pp. 1250-1276, June 2002.
[20] W. Zhang, S. Kotagiri, and J. N. Laneman, Writing on dirty paper with
resizing and its application to quasi-static fading broadcast channels," in
Proc. 2007 IEEE International Symp. Inform. Theory (ISIT07), Nice,
France, June 2007.
Amir Bennatan (S02 - M07) received the B.Sc.
degree (summa cum laude) in mathematics and
computer science, from Tel-Aviv University, TelAviv, Israel, in 1994. He received the M.Sc. (magna
cum laude) and the Ph.D. degrees, both in electrical
engineering, from Tel-Aviv University, in 2002 and
2006, respectively.
During 1995-2000, he worked at the Israel Air
Force Information Systems unit, including three
years as a software team leader and systems analyst.
He is currently a Postdoctoral Researcher at the
Program in Applied and Computation Mathematics (PACM) at Princeton
University, Princeton, NJ. His fields of interest include wireless networks
and software engineering methodologies.
Dr. Bennatan was a recipient of a scholarship from the Wolf Foundation in
2002, of the Weinstein award in 2002, 2003, and 2004, and of an Intel study
award in 2004.
Vaneet Aggarwal (S08) received the B.Tech. degree in 2005 from Indian Institute of Technology,
Kanpur, India and the M.A. degree in 2007 from
Princeton University, Princeton, NJ, USA, both in
Electrical Engineering. He is currently pursuing the
Ph.D. degree in Electrical Engineering at Princeton University, Princeton, NJ, USA. He received
the Porter Ogden Jacobus Honoric Fellowship, the
highest honoric fellowship awarded by the Graduate School of Princeton University, in 2009. His
research interests are in applications of information
and coding theory to wireless systems and quantum error correction.

3627

Yiyue Wu (S08) received the B. S. degree in

2006 from Shanghai Jiao Tong University, Shanghai,
China and the M.A degree in 2008 from Princeton
University, Princeton, NJ, USA, both in Electrical
Engineering. He is current a Ph.D. student advised
by Prof. Robert Calderbank in Electrical Engineering at Princeton University, Princeton, NJ, USA.
His research interests include space-time coding,
information theory and cross-layer issues in wireless
communication systems.
A. Robert Calderbank (M89 - SM97 - F98)
received the BSc degree in 1975 from Warwick
University, England, the MSc degree in 1976 from
Oxford University, England, and the PhD degree in
1980 from the California Institute of Technology, all
in mathematics.
Dr. Calderbank is Professor of Electrical Engineering and Mathematics at Princeton University
where he directs the Program in Applied and Computational Mathematics. He joined Bell Telephone
Laboratories as a Member of Technical Staff in
1980, and retired from AT&T in 2003 as Vice President of Research. Dr.
Calderbank has made significant contributions to a wide range of research
areas, from algebraic coding theory and quantum computing to wireless
communication and active sensing.
Dr. Calderbank served as Editor in Chief of the IEEE T RANSACTIONS
ON I NFORMATION T HEORY from 1995 to 1998, and as Associate Editor for
Coding Techniques from 1986 to 1989. He was a member of the Board of
Governors of the IEEE Information Theory Society from 1991 to 1996 and
began a second term in 2006. Dr. Calderbank was honored by the IEEE
Information Theory Prize Paper Award in 1995 for his work on the Z4
linearity of Kerdock and Preparata Codes (joint with A. R. Hammons Jr., P.
V. Kumar, N. J. A. Sloane, and P. Sole), and again in 1999 for the invention
of space-time codes (joint with V. Tarokh and N. Seshadri). He received the
2006 IEEE Donald G. Fink Prize Paper Award and the IEEE Millennium
Medal, and was elected to the US National Academy of Engineering in 2005.

Jakob Hoydis (S08) received the Diploma degree in electrical engineering

and information technology from RWTH Aachen University, Germany, in
2008. He is currently pursuing a Ph.D. degree in electrical engineering at
SUPELEC, Gif-sur-Yvette, France. His research interests are in network
MIMO and cooperative communications.

Aik Chindapol (M00 - SM07) received the Ph.D. in Electrical Engineering

from University of Washington in 2000. During 2000-2007, he was with
Siemens Mobile Networks R&D and Siemens Corporate Research, where
he was involved in the PHY/MAC design of GPRS/EDGE, WiFi (IEEE
802.11), Mobile WiMAX (IEEE 802.16e) and Mobile Multihop Relay (IEEE
802.16j). He is with Nokia Siemens Networks, where he leads the research and
standardization activities for the revision of Mobile WiMAX standards. His
current research interests include network coding, error control coding, radio
resource management and performance optimization of mobile broadband
systems.