Capacity Bounds and Power Allocation for Wireless Relay Channels

Anders Høst-Madsen∗ Junshan Zhang†

Department of Electrical Engineering Department of Electrical Engineering
University of Hawaii Arizona State University
Honolulu, HI 96813 Tempe, AZ 85287
e-mail: madsen@spectra.eng.hawaii.edu. e-mail: junshan.zhang@asu.edu

September 2, 2004

Abstract

We consider three-node wireless relay channels in a Rayleigh fading environment. Assuming

first transmitter channel state information (CSI), we study upper bounds and lower bounds on
the outage capacity and the ergodic capacity. Our studies take into account practical constraints
on the transmission/reception duplexing at the relay node and on the synchronization between
the source node and the relay node. We also explore power allocation. Compared to the direct
transmission and traditional multi-hop protocols, our results reveal that optimum relay channel
signaling can significantly outperform multi-hop protocols, and that power allocation has a
significant impact on the performance.

1 Introduction
Wireless ad-hoc networks consist of a number of terminals (nodes) communicating on a peer-to-peer
basis, without the assistance of wired networks or centralized infrastructure. In such systems, the
communications between nodes might take place through several intermediate nodes. In wireless
communications, channel impairments that limit capacity include multipath fading, shadowing,
and path loss.
Needless to say, high spectral efficiency is of vital importance in ad hoc wireless networks (see,
e.g., [11, 38, 5, 28, 19, 31]). One approach to increasing the capacity of wireless networks is to
∗
The work of A. Høst-Madsen was supported in part by NSF grant CCR03-29908
†
The work of J. Zhang is supported in part by NSF Grant ANI-0208135 and a grant from the Intel Research
Council.

1
use cooperative diversity. Earlier works on cooperative communications can be found in [46] and
[2]. Recently cooperative diversity has been studied in [34, 35, 36, 39] for cellular networks and
in [24, 23, 21, 25, 22] for ad-hoc networks. Roughly speaking, several terminals, each with one or
more antennas, form a kind of “coalition” to cooperatively act as a large transmit or receive array.
The channel can therefore exhibit some characteristics of the MIMO channel.
In this paper, we take steps to obtain a fundamental understanding of cooperative diversity. In
particular, we investigate the 3-node relay channel shown in Figure 1. The desired transmission is
from the source node (node 1) to the destination node (node 3), while the relay node (node 2) aids
the communication by using its “capture” of the transmission between node 1 and node 3. The
Gaussian relay channel was introduced by van der Meulen [44, 45] and used in connection with the
Aloha system [32]. It was thoroughly analyzed by Cover and El Gamal in [3], where the capacity of a
degraded type of relay channel was found, and upper bounds and lower bounds for the general relay
channel were presented. The work [49] used a simpler coding argument based on [46] to obtain the
same achievable rate as in [3]. Very recently, partly spurred by the work on cooperative diversity,
the relay channel has garnered much attention [13, 15, 48, 12, 14, 16, 42, 27, 30, 33, 43, 7, 9, 8].
This paper focuses on wireless relay channels in a Rayleigh fading environment. We take into
account some practical constraints for wireless transceivers, such as synchronization and transmis-
sion duplexing. For convenience, we say that the relay node operates in the full duplex mode if
the relay node can receive and transmit simultaneously on the same frequency channel. While
there might exist some RF techniques making this possible [29], it is in general regarded unrealistic
in practical systems, due to the dynamic range of incoming and outgoing signals and the bulk of
ferroelectric components like circulators. We will therefore consider the cases where the relay node
operates in a frequency-division (FD) manner or time-division (TD) manner.
We study upper bounds and lower bounds on the channel capacity. Specifically, we start with
examining the capacity for the fixed channel gain case, building on which we explore the outage
capacity. We then concentrate on the ergodic capacity.
The rest of the paper is organized as follows: we start with the channel model in Section II. In
Section III we present capacity bounds for the fixed channel gain case and outage capacity for fading
channels, and Section IV contains the main thrust on the ergodic capacity. Section V contains our
conclusions. The proofs are relegated to Appendices A-E.

2 System Model
Consider the relay channel in Figure 1. The transmitter at the source node sends a message w ∈
{1, . . . , M } to the destination node, where the message is encoded into n symbols x1 (w)[1], . . . , x(w)[n]

2
 2
c21 c32

1 c31 3

Figure 1: The relay channel.

n
and transmitted over the channel, under the power constraint 1
n i=1 x1 [i]
2 ≤ P1 . (The rate is
log M/n.) Let y2 denote the received signal at the relay node. Building on prior received signals,
the relay node then transmits a message x2 that is intended to aid the transmission between source
and destination, where x2 [i] is encoded based on {y2 [i − 1], . . . , y2 [1])}, subject to the power con-


straint n1 ni=1 x2 [i]2 ≤ P2 . Thus, the received signals at the relay node and the destination node
are given by

y2 [i] = c21 x1 [i] + z2 [i]

y3 [i] = c31 x1 [i] + c32 x2 [i] + z3 [i] (1)

where z2 and z3 are independent additive white Gaussian noise with unit variance (after normaliza-
tion). As is standard, the channel gains {cij } are modeled as independent (flat) fading processes,
and are in general complex.
A few more words on relaying techniques. As noted before, we say that the relay node operates
in the full duplex mode if the relay can receive and transmit simultaneously on the same frequency
channel. In contrast, if the relay node operates in a TD (time-division) manner, then for a given
time window D, the relay node is in the receive mode for a fraction of the time αD (we call
this period the relay-receive period ), and in the transmit mode for the rest (1 − α)D (we call this
period the relay-transmit period ). Similarly, if the relay node operates in a FD (frequency-division)
manner, the bandwidth W is divided into a bandwidth of αW over which the relay node listens,
and a bandwidth of (1 − α)W over which the relay node transmits. The destination node listens
over the whole bandwidth W . Clearly, from an information-theoretic point of view, the TD mode
and the FD mode are equivalent for the fixed channel gain case. In fading channels, however, the
TD mode has an advantage over the FD mode because α can be adjusted to the instantaneous
channel conditions, whereas α is usually fixed in the FD mode. Thus motivated, we focus on the
TD relay channels.
In the synchronized channel model, it is assumed that all nodes have complete channel state
information (CSI), i.e., each node knows the instantaneous values (magnitudes and phases) of all
cij as well as their statistics. It is furthermore assumed that all nodes are perfectly synchronized.
It is relatively straightforward to obtain symbol (timing) synchronization between different nodes;

3
however, carrier synchronization requires phase-locking separated microwave oscillators, which is
very challenging in practical systems [37]. In light of this observation, we also consider the asyn-
chronous channel model in which there is a random phase offset θ[i] between the source and the
relay, and the corresponding received signal at the destination is

y3 [i] = c31 x1 [i] + ejθ[i] c32 x2 [i] + z3 [i], (2)

where θ[i] is random and ergodic, and is uniformly distributed in in [0, 2π) (the case where θ[i]
is constant during the transmission is considered in [20]). We assume that only the destination
knows (i.e., can estimate) θ[i]. Note that the phase factor ejθ[i] can be incorporated into c32 , that
is, c32 = ejθ[i] c32 would have the same distribution as c32 . Modelled in this way, the phase of c32 is
unknown at the source and the relay. In summary, in the asynchronous channel model, the source
and the relay know the amplitudes of cij , but not necessarily the phases.
We now outline the cases to be considered. We consider all 4 combinations of full duplex/TD
and synchronized/asynchronous transceiver models. We first consider the case where the channel
gains are fixed, and each node has a certain average (per frame) power constraint. We then
consider outage capacity, also with an average power constraint per node. We finally consider
ergodic capacity with optimum power allocation over time and among the nodes; and the issue
we investigate here is: what is the minimum network power (energy) needed to transmit a given
amount (say one bit) of information at a given rate under a bandwidth constraint, and how much
can be saved (by the network in total) by using a relay?
As a baseline for comparison, we will consider two strategies commonly applied in networks:
direct transmission between source and destination and multihop transmission (routing). By the
latter we mean that a packet is first transmitted to the relay; the relay decodes the packet, re-
encodes it, and transmits it to the destination in the next time slot; that is, the destination only
“uses” the transmission from the relay.
A few words about notation. The function log denotes the base 2 logarithm. For convenience
of notation we will use c2ij to denote |cij |2 . We define the average channel gain on each link by

sij = E[|cij |2 ] (3)

For notational convenience, we let C + denote an upper bound on the channel capacity, and C − for
a lower bound. We will use R to denote the rate of a specific signaling scheme, i.e., an achievable
rate, which also constitutes a lower bound. A quantity (function, parameter) x used in an upper
bound is denoted x+ and x− for a lower bound. Expressions common to upper and lower bounds
are written as C ± , which involves quantities associated either with an upper bound or a lower
bound.

4
3 Preliminary: The Fixed Channel Gain Case
We start with the bounds on the capacity of the relay channel, assuming fixed channel gain coeffi-
cients cij . The results will also be used to find the outage capacity in a fading channel.

3.1 Full Duplex Relaying

The capacity of the full duplex relay channel was studied in [3] (see also [4]). In particular, using
the “max-flow-min-cut” Theorem [4, Theorem 14.10.1] or [3, Theorem 4] yields the following upper
bound:

C+ = max min {I(X1 ; Y2 , Y3 |X2 ), I(X1 , X2 ; Y3 )} (4)

f (X1 ,X2 )

1   
= max min log 1 + (1 − β)P1 c221 + c231 ,
0≤β≤1 2
  
1 2 2 2 2
log 1 + c31 P1 + c32 P2 + 2 βc31 c32 P1 P2 (5)
2

If the relay node decodes and re-encodes the message, the following rate is achievable [3, Theorem
1]:

R1 = max min {I(X1 ; Y2 |X2 ), I(X1 , X2 ; Y3 )} (6)

f (X1 ,X2 )

1
= max min log(1 + (1 − β)c221 P1 ),
0≤β≤1 2
  
1
log 1 + c231 P1 + c232 P2 + 2 βc231 c232 P1 P2 (7)
2

Note that if c21 < c31 , the above achievable rate boils down to the rate of the direct transmission
between source and destination. On the other hand, applying [3, Theorem 6]1 to the Gaussian
relay channel gives a rate
 
1 c221 P1
R2 = log 1 + c231 P1 + c31 P1 +c221 P1 +1
2
 (8)
2 1+ c232 P2

Then, the following rate serves as a lower bound:

C − = max{R1 , R2 } (9)
1
[3, Theorem 6] only applies to the discrete alphabet relay channel. However, using equations (76a) and (76b) in
[3] for Gaussian distributions yields (8). Alternatively, the method of proof we use for Proposition 3 in the Gaussian
case can also be applied to the full duplex relay channel and will result in (8) in the Gaussian case.

5
 Upper and lower bounds in the asynchronous case can be found by setting β = 0 (the proof
of this is similar to the proofs of Propositions 1 and 2 below). Interestingly, for the asynchronous
case, the upper bound (5) and lower bound (7) meet if c32 is sufficiently large, indicating that the
exact capacity of the relay channel can be found.

3.2 Time-Division Relaying

Recall that in the TD mode, given a time window D, the relay is in the receive mode for a fraction of
the time αD (the relay-receive period ), and in the transmit mode for (1 − α)D (the relay-transmit
period ). The source node can transmit in both the relay-receive period and the relay-transmit
(1)
period. Suppose that the source node transmits with power P1 during the relay-receive period,
(2)
and with power P1 during the relay-transmit period; the relay node transmits with power P2 .
(1) (2)
Denote the corresponding capacity as CR (α, P1 , P1 , P2 ). We then have the following upper
bound on the capacity of the TD relay channel. This is a relatively straightforward application of
[4, Theorem 14.10.1] or [3, Theorem 4] (the proof can be found in Appendix A).

Proposition 1 (Upper bound) The capacity of the TD relay channel is upper bounded by

C+ = max min C1+ (β), C2+ (β) (10)
0≤β≤1

with
1   1  
(1) (2)
C1+ (β) = α log 1 + (c231 + c221 )P1 + (1 − α) log 1 + (1 − β)c231 P1 , (11)
2  2  
1   1
(1) (2) (2)
C2+ (β) = α log 1 + c231 P1 + (1 − α) log 1 + c231 P1 + c232 P2 + 2 βc231 P1 c232 P2 (12)
2 2
Furthermore, an upper bound for the asynchronous relay channel is obtained by putting β = 0.

Consider the case where the relay can decode what it has received during the relay-receive period,
re-encode it, and transmit during the relay-transmit period. We have the following proposition on
the achievable rate (we call this rate the decode-forward rate).

Proposition 2 (Achievable rate: decode-forward) The capacity of the TD relay channel is

lower bounded by

C − = max min C1− (β), C2− (β)
0≤β≤1
with
1   1  
C1− (β) = α log 1 + c221 P1
(1) (2)
+ (1 − α) log 1 + (1 − β)c231 P1 (13)
2 2   
1   1
C2− (β) = α log 1 + c231 P1
(1) (2) (2)
+ (1 − α) log 1 + c231 P1 + c232 P2 + 2 βc231 P1 c232 P2 (14)
2 2

6
The optimum value of β can be found in closed form (see Appendix A.) Furthermore, an achievable
rate for the asynchronous relay channel is obtained by putting β = 0 in (2).

The proof of proposition 2 in Appendix A shows that, as opposed to the full duplex channel of [3],
block-Markov coding is not needed for the TD relay channel.
Instead of using “decode-forward,” an alternative approach is to let the relay use Wyner-Ziv
lossy source coding [47] on the received signal. The compressed signal is then transmitted to the
destination using an (error-free) channel encoding. The following proposition gives the rate for
both the synchronized and the asynchronous cases (the proof is in Appendix A). The proposition
can be viewed as a generalization of [3, Theorem 6].

Proposition 3 (Achievable rate: compress-forward) The rate R is achievable if

 
1 c 2 P (1) 1  
(1) (2)
R = α log 1 + c231 P1 + 21 1 2 + (1 − α) log 1 + c231 P1 (15)
2 1 + σw 2

2 is the “compression noise” given by

where σw
(1) (1)
c221 P1 + c231 P1 + 1
2
σw =  (1−α)/α  (16)
c232 P2 (1)
1+ 2 (2) − 1 (c231 P1 + 1)
1+c31 P1

If c21 < c31 , the decode-forward approach would not work, and only the compress-forward
approach can be used. The compress-forward method can be used for all channels and always gives
a rate gain over the direct transmission, but as c21 becomes larger compared to c31 , the decode-
forward rate would be eventually larger than the compress-forward rate (see Figure 2). In general,
a higher rate is obtained by taking the maximum of the rates of Propositions 2 and 3.
(1) (2)
In the above, we have assumed that the parameters α, P1 , P1 , P2 are fixed. We now consider
that the source node and the relay node are subject to average power constraints P1 and P2 . Since
the relay only transmits during the relay-transmit period of length (1 − α)D, it can use power P2
1−α
κP1
during the transmission period. Similarly, the source node transmits with power α during the
(1−κ)P1
relay-receive period and with power 1−α during the relay-transmit period, where 0 ≤ κ ≤ 1 so
that the average power constraint is satisfied. The capacity is given by
 
κP1 (1 − κ)P1 P2
CT D (P1 , P2 ) = max CR α, , , (17)
0≤α≤1,0≤κ≤1 α 1−α 1−α

The optimum value of β in Propositions 2 and 1 can be found in closed form (see [17]), but
the optimization of α and κ needs to be done numerically. The upper and lower bounds of the

7
propositions are then be obtained. Unfortunately, as opposed to the full duplex case, the upper
and lower bounds never meet in non-trivial cases, even in the asynchronous case.
To illustrate the above results, Figure 2 shows upper and lower bounds for different fixed values
of c32 .
2 2
c32=10.0 dB c32=0.0 dB
2.6 1.9
TD Sync. Upper bound TD Sync. Upper bound
TD async. upper bound TD async. upper bound
Sync. decode−forward Sync. decode−forward
2.4 1.8
Async. decode−forward Async. decode−forward
Compress−forward Compress−forward

1.7
2.2
Full Duplex
Full Duplex 1.6 Upper/Lower
Upper/Lower
2

1.5
Rate

Rate
1.8 Direct 2xP

1.4

1.6
1.3
Direct 2xP
1.4
1.2

1.2 1.1

Direct Direct
1 1
−10 −5 0 5 10 −10 −5 0 5 10
c2 (dB) c2 (dB)
21 21

Figure 2: Capacity bounds for TD relay channels with fixed channel coefficients, P1 = P2 = 5dB
and c31 = 1. The curve ‘Direct’ is the rate for direct transmissions, and ‘Direct 2 × P ’ is the rate
for direct transmissions when the source has twice the power.

3.3 An Application to Outage Capacity

The bounds on capacity for fixed channel gains can be used to find bounds on outage capacity. The
outage capacity of a fading channel is defined by [40]

Co = arg max P (R(cji ) ≥ C) ≥ p (18)

C

where R(cji ) is the rate for a specific realization of the fading process. Thus, to find the outage
capacity, we first find the rate for specific values of cji , and the outage capacity can then be
calculated as the (1 − p) percentile of the random variable R(cji ). By using the upper and lower
bounds from the previous section we can then find upper and lower bounds on the outage capacity.
Figure 3 plots outage capacity versus the SNR in the direct link when all links experience
independent Rayleigh fading. The curves are calculated based on an ensemble of R(cji ) by Monte
Carlo simulations for each value of the SNR. We also compare the results above with multihop

8
transmissions for which the achievable rate is
  
1  2
 1  2
 1  2

R = max log 1 + c31 P1 , min log 1 + 2c21 P1 , log 1 + 2c32 P2 (19)
2 4 4

While the figures shows only two examples, they are typical for the many other examples we
have considered. It should be noticed that upper and lower bounds are close, that the gain from
synchronization is limited, and that relaying clearly outperforms the traditional multihop strategy,
in particular at high SNR.
Outage Capacity Outage Capacity
4.5 5
Relay (sync) Relay (sync)
Relay (async) Relay (async)
TD Relay (sync) 4.5 TD Relay (sync)
4 TD Relay (async) TD Relay (async)
Multihop Multihop
2 Tx Antenna 4 2 Tx Antenna
3.5 Direct 2×P Direct 2×P
Direct Direct
3.5
3

3
2.5
Rate

Rate

2.5
2
2

1.5
1.5

1
1

0.5 0.5

0 0
−10 −5 0 5 10 15 20 −10 −5 0 5 10 15 20
SNR (dB) SNR (dB)

Figure 3: Outage capacity for the relay channel with independent Rayleigh fading. Source and relay
have equal power, and the x axis corresponds to the SNR for the direct link. The solid curve is for
the upper bound, and the dashed curve is for the lower bound. Figure (a) is for s21 = s32 = 12dB
and Figure (b) for s21 = s32 = 20dB (sij is defined in (3)). ’Direct 2 × P ’ is the rate for direct
transmission when the source has twice the power. ‘2 Tx Antenna’ is the rate for direct transmission
when the source has two antennas.

4 Bounds on Ergodic Capacity

We now turn to study the ergodic capacity for wireless relay channels. As stated in the introduction,
the problem here is what is the minimum network power (energy) needed to transmit a given amount
(say one bit) of information at a certain rate under a bandwidth constraint. This is equivalent to
finding the rate under a total power constraint, P1 + P2 ≤ P .

9
4.1 Ergodic Capacity: The General SNR Case
(1) (2) (1)
First consider the full duplex case. Set P1 = P1 + P1 , where P1 is used for transmission to
(2) (1)
the relay, and P1 for transmission to the destination (i.e, in earlier notation P1 = (1 − β)P1 ,
(2) (1) (2)
P1 = βP1 ). Power allocation dictates that P1 , P1 and P2 are functions of c = (c31 , c21 , c32 ).
Using the upper bound (4), together with the fact that the optimum distribution for fixed c is
Gaussian [10, 40], we have (a formal argument is given in the proof of Proposition 4)
 
(1) (1) (2)
C+ = max min R1+ (P1 ), R2+ (P1 , P1 , P2 ) (20)
(1) (2)
P1 ,P1 ,P2

with
  
(1) 1 (1)  
R1+ (P1 ) = E log 1 + P1 (c) c221 + c231 , (21)
2
   
+ (1) (2) 1 (1) (2)
R2 (P1 , P1 , P2 ) = E log 1 + c231 P1 (c) + P1 (c)
2
 
2 2 2 (2)
+c32 P2 (c) + 2 c31 c32 P1 (c)P2 (c) (22)

(1) (2)
It remains to find the optimum functions P1 (c), P1 (c), and P2 (c), and this is a power allocation
(2)
problem. To this end, first fix c. For fixed Pj (c) = P1 (c) + P2 (c), the rate is maximized when
(2) c231 c232
we set P1 (c) = P (c)
c231 +c232 j
and P2 (c) = P (c)
c231 +c232 j
(a beamforming solution). We then get a
simplified problem
 
(1) (1)
C+ = max min R1+ (P1 ), R2+ (P1 , Pj ) , (23)
(1)
P1 ,Pj

with
  
(1) 1 (1)  2 
R1+ (P1 )= E log 1 + P1 (c) c21 + c312
(24)
2
  
+ (1) 1 2 (1) 2 2
R2 (P1 , Pj ) = E log 1 + c31 P1 (c) + (c31 + c32 )Pj (c) (25)
2
For the achievable rate, we similarly get a decode-forward solution (see Appendix B for details
about the coding used),
 
C− = max min R1− (P1 ), R2− (P1 , Pj )
(1) (1)
(26)
(1)
P1 ,Pj

with
  
1 2 2 
R1− (P1 )
(1) (1)
= E log 1 + P1 (c) max c21 , c31 (27)
2
  
− (1) 1 2 (1) 2 2
R2 (P1 , Pj ) = E log 1 + c31 P1 (c) + (c31 + c32 )Pj (c) (28)
2

10
 For asynchronous signaling, we get a similar solution by replacing c231 + c232 by c232 in (25) and
(1) (2)
(28). Since there is no beamforming gain, P1 = P1 , P1 = 0, and Pj = P2 is the optimum. To
unify the treatment of the different cases, we define

c+ (c221 , c231 ) = c221 + c231 (29)

c− (c221 , c231 ) = max{c221 , c231 } (30)

c231 + c232 synchronized
c̃(c231 , c232 ) = (31)
c232 asynchronous
The optimization problems for both upper and lower bounds can be solved by using “water-
filling” techniques (see Appendix B), and the solutions are stated in the following proposition.

Proposition 4 (Ergodic capacity: full duplex relaying) Define

  
± 1 (1)  ± 2 2 
R1 (t, µ) = E log 1 + P1 (t, µ, c) c (c21 , c31 ) (32)
2
  
± 1 2 (1) 2 2
R2 (t, µ) = E log 1 + c31 P1 (t, µ, c) + c̃(c31 , c32 )Pj (t, µ, c) (33)
2

where the functions P1 (t, µ, c) and Pj (t, µ, c) are given in Table 1. Let (t± , µ± ) be the solution to
(1)

the equations
 
(1)
E P1 (t, µ, c) + Pj (t, µ, c) = P (34)
R1± (t, µ) = R2± (t, µ). (35)

The capacity of the relay link Cr is then bounded by

R1− (t− , µ− ) ≤ C ≤ R1+ (t+ , µ+ ) (36)

In the TD relay channel, the relay cannot receive while it transmits. Accordingly, an upper
bound on the capacity and a lower bound can therefore be obtained by the following modification
of the full duplex problem:
 
C± = max min R1± (P1 ), R2± (P1 , Pj ) ,
(1) (1)
(37)
(1)
P1 ,Pj
  
1  2 2 
R1± (P1 )
(1)
= E
(1) ± 
log 1 + P1 (c)c c21 , c31  Pj (c) = 0
2

1   
(1) 2 
+E log 1 + P1 (c)c31  Pj (c) > 0 (38)
2
  
± (1) 1 2 (1) 2 2
R2 (P1 , Pj ) = E log 1 + c31 P1 (c) + (c31 + c32 )Pj (c) (39)
2
The details of the above optimization can be found in Appendix D. The solutions are stated as
follows.

11
 (1)
Table 1: Power allocation algorithm for the full duplex case. The procedure calculates P1 and Pj
for specific values of t and µ
1. Set
(1)
P1 = 0
1
Pj = tµ −
c̃(c231 , c232 )
if
c̃(c231 , c232 )(1 − t)µ 1
0 > 2 − ± 2 2
c32 c (c21 , c31 )
0 < Pj
exit;
2. Set
(1) c̃(c231 , c232 )(1 − t)µ 1
P1 = 2 − ± 2 2
c32 c (c21 , c31 )
2 (1)
1 + c31 P1
Pj = tµ −
c̃(c231 , c232 )
(1)
if P1 > 0, Pj > 0 exit;
(1)
3. Set Pj = 0 and P1 equal to the largest solution for x in
 
−µ−1 c231 c± (c221 , c231 )x2 + c± (c221 , c231 )(c231 −µ−1 ) − µ−1 c231 x
−µ−1 + c231 + (1 − t)c± (c221 , c231 ) + tc231 = 0
(1)
if P1 > 0 and
1
0 < tµ − OR
c̃(c231 ,c232 ) 
(1 − t) log 1 + c± (c221 , c231 )P1
(1)
0 <
 
− (µ ln 2)−1 P1
(1) (1)
+t log 1 + c231 P1
exit;
(1)
4. Set P1 = Pj = 0

12
Proposition 5 (Ergodic capacity: time-division relaying) Define
   
1  ± 2 2  
±
R1 (t, µ) = E
(1) 
log 1 + P1 (t, µ, c) c (c21 , c31 )  Pj (t, µ, c) = 0
2

1   
(1) 2 
+E log 1 + P1 (t, µ, c)c31  Pj (t, µ, c) > 0 (40)
2
  
± 1 2 (1) 2 2
R2 (t, µ) = E log 1 + c31 P1 (t, µ, c) + c̃(c31 , c32 )Pj (t, µ, c) (41)
2

where P1 (t, µ, c), and Pj (t, µ, c) are given in Table 2. Let (t± , µ± ) be the solution to the equations
(1)

 
(1)
E P1 (t, µ, c) + Pj (t, µ, c) = P (42)
R1± (t, µ) = R2± (t, µ). (43)

The capacity of the relay link Cr is then bounded by

R1− (t− , µ− ) ≤ C ≤ R1+ (t+ , µ+ ) (44)

We also consider the asynchronous channel without power allocation; by this we mean that the
network does not allocate power based on instantaneous channel state, while it still allocates power
between source and relay based on channel statistics. It is then easily seen that the capacity bounds
for full duplex are given by
  
± 1  ±
 2 2 
C = max min E log 1 + κP c c21 , c31 ,
κ∈[0,1] 2
 
1  
E log 1 + c31 κP + c32 (1 − κ)P
2 2
(45)
2
For the TD case, we also assume that the network allocates the relay-receive and relay-transmit
time intervals based on average channel conditions, and we then get the bounds
  
± 1  ±
 2 2  1  
C = max min E α log 1 + κP c c21 , c31 + (1 − α) log 1 + κP c31 ,
2
α∈[0,1],κ∈[0,1] 2 2
  
1   1 2 1−κ
E α log 1 + κP c31 + (1 − α) log 1 + c31 κP + c32
2 2
P (46)
2 2 1−α
We now illustrate our findings on ergodic capacity via numerical results . We compare our
results with simple multi-hop forwarding. For fair comparison, we allow the network to optimize
power allocation and relay receive/transmit durations base on channel statistics. Accordingly, we
have that
   
1 2 P
R = max min E α log 1 + c21 κ ,
κ∈[0,1],α∈[0,1] 2 α
  
1 P
E (1 − α) log 1 + c32 (1 − κ)
2
(47)
2 1−α

13
 (1)
Table 2: Power allocation algorithm for the TD case. The procedure calculates P1 and Pj for
specific values of t and µ.
1. Set
(1)
P1,1 = 0
1
Pj,1 = tµ −
c̃(c231 , c232 )
if
c̃(c231 , c232 )(1 − t)µ 1
0 > 2 − 2
c32 c31
0 < Pj
go to 3.
2. Set
(1) c̃(c231 , c232 )(1 − t)µ 1
P1,1 = 2 − 2
c32 c31
2 (1)
1 + c31 P1
Pj,1 = tµ −
c̃(c231 , c232 )
(1)
3. Set Pj,2 = 0 and P1,2 equal to the largest solution for x in
 
−µ−1 c231 c± (c221 , c231 )x2 + c± (c221 , c231 )(c231 − µ−1 ) − µ−1 c231 x
−µ−1 + c231 + (1 − t)c± (c221 , c231 ) + tc231 = 0
(1)
if P1 > 0 and
1
0 < tµ − OR
c̃(c231 ,c232 ) 
(1 − t) log 1 + c± (c221 , c231 )P1
(1)
0 <
 
− (µ ln 2)−1 P1
(1) (1)
+t log 1 + c231 P1
go to 5
(1)
4. Put P1,2 = Pj,2 = 0
5. Calculate
   
(1) (1)
Rt,1 = (1 − t) log 1 + P1,1 c231 + t log 1 + c231 P1,1 + c̃(c231 , c232 )Pj,1
 
− (µ ln 2)−1 P1,1 + Pj,2
(1)
    
(1 − t) log 1 + P1,2 c± c221 , c231 + t log 1 + c231 P1,2
(1) (1)
Rt,2 =
− (µ ln 2)−1 P1,2
(1)

(1) (1) (1)

if P1,1 ≤ 0 or Rt,2 > Rt,1 put P1 = P1,2 , Pj = Pj,2 ,
(1) (1)
otherwise P1 = P1,1 , Pj = Pj,1

14
Figure 4 shows the different rates versus SNR for s21 = s32 = 12dB relative to s31 (see (3); all
channel gains are relative to s31 ) and all paths are independently Rayleigh fading. The curves are
generated using Monte-Carlo simulation to generate an ensemble of values c which is then used to
calculate upper and lower bounds.
We note that the upper and lower bounds are close in many cases of interest. Moreover, it
should be noted that traditional multihop signaling is clearly suboptimum compared to the relay
signaling, in the sense that the slope of the rate increase is smaller than that for the relay signaling,
especially at high SNR. For SNR larger than 5dB, all curves (except multihop) are approximately
parallel lines. The offset between the lines (either in terms of rate or power) therefore can be used
to characterize the gain of relaying at (reasonably) high SNR. It turns out that this offset can be
calculated explicitly, and this is studied in the following section.
Ergodic Capacity
6
Full duplex relay upper
Full duplex relay lower
TD relay upper
TD relay lower
5 TD no power allocation upper
TD no power allocation lower
2 Tx antennas
Direct transmission
Multihop
4
Rate

0
−10 −5 0 5 10 15 20
SNR (dB)

Figure 4: Ergodic capacity of the relay channel in a Rayleigh fading channel with s21 = s32 = 12dB
relative to s31 , which is also the SNR. All curves are for synchronized transmission, except the
curve for no power allocation

4.2 Ergodic Capacity in the High SNR Regime

Clearly, the solutions for the bounds and the power allocation for general cases are complicated. It
turns out, however, that in the high SNR regime there exist simple closed-form solutions (cf. the high
SNR results for the MIMO channel in [6]). As mentioned in the previous section, in the high SNR
regime, the capacity of the relay channel, Crelay (SNR) satisfies Crelay (SNR) ≈ Cdirect (SNR) + Rg ,
where Cdirect (SNR) is the capacity for the direct link and Rg is a constant. To formalize this, we

15
define the asymptotic relay gain by

Rg = lim (Crelay (SNR) − Cdirect (SNR)) (48)

SN R→∞

By SN R → ∞, we mean that the noise power σ 2 → 0 while the total power constraint is kept
constant. Our goal is to find bounds for Rg . It turns out that at high SNR, the complicated power
allocation rules in Tables 1 and 2 become much simpler.
Since only bounds on the capacity have been found, it is not clear that the limit (48) exists for
the capacity. More formally, we can define bounds Rg± for the the relay gain by

lim sup (Crelay (SNR) − Cdirect (SNR)) ≤ Rg+ (49)

SN R→∞
lim inf (Crelay (SNR) − Cdirect (SNR)) ≥ Rg− (50)
SN R→∞

Another way to look at the gain by relaying is in terms of the additional power needed for direct
transmission to get the same rate as with a relay; we call this high SNR limit the asymptotic power
gain. It is easily seen that the relationship between the asymptotic relay gain and the asymptotic
power gain (in dB) is
20
Pg (in dB) = Rg ≈ 6Rg (51)
log 10
We will now state the main results in the high SNR regime. The asymptotic relay gain for full
duplex is found in Appendix C and stated in the following proposition

Proposition 6 (High SNR relay gain: full duplex synchronized case) In the high SNR regime
(1)
(i.e., σ 2 → 0) the optimum values of P1 and Pj for both upper and lower bounds in the synchro-
nized case converge towards
     2
 2 2 2
 c31 + c32 (1 − t)P, t − c31 (1 − t) P c31
(1) 2 2 <κ
(P1 (c), Pj (c)) = c32 c32 c232 (52)

 (P, 0) otherwise
t
κ = (53)
1−t
In a Rayleigh fading channel κ is given by (for s31 = s32 )
s21
κ− = achievable rate (54)
s32
2l(s21 ,s31 ) − s31
κ+ = upper bound (55)
s32
with

 a log a − b log b a = b
l(a, b) = a−b (56)
 log a + (ln 2)−1 a=b

16
In addition, the asymptotic relay gain is bounded by
 
1 s 1 + κ±
Rg± =
32
log s31 = s32 (57)
2 s31 − s32 1 + κ± ss32
31

Note that power allocation rule (52) does not depend on the characteristic of the fading channel
(i.e., it is not specific to Rayleigh fading), and it is the same for upper and lower bounds. (The
above proposition is for the case s31 = s32 ; and results for s31 = s32 can be found by taking the
limit s31 → s32 .)
For asynchronous signalling, we have that

Proposition 7 (High SNR relay gain: full duplex asynchronous case) In the high SNR regime
(i.e., σ 2 → 0), the optimum values of P1 and P2 for both upper and lower bounds in the asynchronous
case converge towards
     2

 c232 c231 c31
(1 − t)P, t − 2 (1 − t) P <t
(P1 (c), P2 (c)) = c32 − c31
2 2 c32 − c31
2 c232 (58)

 (P, 0) otherwise

In Rayleigh fading the value of t is given by

±
± 2x − s31
t = (59)
2x± + s32
(s31 + s32 ) log(s21 + s31 ) − s31 log(s31 )
x− = achievable rate (60)
s32
(s31 + s32 )l(s21 , s31 ) − s31 log(s31 )
x+ = upper bound (61)
s32
and the asymptotic rate gain is bounded by
 
1 ± s32
Rg± = log t +1 (62)
2 s31
The above results give upper and lower bounds on the capacity (not the exact capacity). How-
ever, it is possible to bound the gap between the upper and lower bounds for both the synchronized
and asynchronous cases (Propositions 6-7).

Corollary 1 The rate difference and the power difference between the upper and lower bounds
satisfy that
ln 2 − 1
rate : Rg+ − Rg− ≤ ≈ 0.221 (63)
2 ln 2
10(ln 2 − 1)
power : Pg+ − Pg− ≤ ≈ 1.33dB (64)
log 10 ln 2
and the maximum of the differences occurs when (s21 , s32 ) = (s31 , ∞).

17
For most values of sij , Pg+ − Pg− is smaller than 1.33 dB. A closer elimination of the difference
reveals that it is only in a small neighborhood of the ridge s21 = s31 that the difference can be
significant.
The achievable rates in Propositions 6-7 are based on using “decode-forward”. As seen for the
fixed channel case in Section 3, a compress-forward scheme, based on Wyner-Ziv rate-distortion
theory, can yield a higher rate than the decode-forward scheme in certain regions. We now consider
a compress-forward scheme for ergodic capacity. As shown by Propositions 6-7, finding ergodic
capacity is much simplified in the high SNR regime, and we consider the compress-forward scheme
in this regime. The power control rule (58) combined with Gaussian Wyner-Ziv compression is
applied for each given c. The details of the approach can be found in Appendix E. We have the
following proposition.

Proposition 8 (High SNR relay gain: full duplex compress-forward) Define Σ(t) as the
solution to the equation
   s32 
Σ(t)s21 Σ(t)s21 s32 1+ s31 t
log + log(1 + Σ(t)) = log (65)
Σ(t)s21 − (1 + Σ(t))s31 (1 + Σ(t))s31 s31 + s32 1−t

For Rayleigh fading, the following rate gain is achievable using “compress-forward” in the high SNR
regime (i.e., σ 2 → 0):
  
   s log(1 − t) + s log 1 + t s32 
1 Σ(t)s21 Σ(t)s21 32 31 s31
Rg = max log + (66)
2 t∈[0,1]  Σ(t)s21 − (1 + Σ(t))s31 (1 + Σ(t))s31 s31 + s32 

Numerical evaluation of this compress-forward rate shows that it can give a gain of around 0.5dB
over decode-forward in the region s32 > 5dB, s21 < 0dB, but otherwise is inferior to decode-forward.
For the asynchronous channel without power allocation, bounds for the asymptotic relay gain
in Rayleigh fading can be easily found to be
1 1
Rg+ = max min {l(κs21 , κs31 ), l(κs31 , (1 − κ)s32 )} − log s31 (67)
2 κ∈[0,1] 2
1 1
Rg− = max min {log((s21 + s31 )κ), l(κs31 , (1 − κ)s32 )} − log s31 (68)
2 κ∈[0,1] 2

where the maximization over κ must be done numerically.

We now turn to the TD case. As in the full duplex case, we get a simple power allocation rule

18
for the synchronized case for high SNR,


  2    s(t, c31 , c21 , c32 ) > 0
 2 2
 c31 + c32 (1 − t)P, t − c31 (1 − t) P
(1) c2 t
(P1 (c), Pj (c)) = c232 c232 and 31 2 < 1−t
(69)

 c


32
(P, 0) otherwise
 2 2   2 
c31 (c31 + c232 ) c31 + c232
s(t, c31 , c21 , c32 ) = (1 − t) log 2 ± 2 2 (1 − t) + t log t (70)
c32 c (c21 , c31 ) c231
Unfortunately, because of the complicated condition for joint transmission (70) we cannot evaluate
the Rayleigh integrals analytically, and we have to find t as well as the asymptotic relay gain Rg
numerically, see appendix D.
For the asynchronous TD case we get the asymptotic power allocation rule,


     s(t, c31 , c21 , c32 ) > 0

 c232 c231
(1 − t)P, t − (1 − t) P c2
(P1 (c), P2 (c)) = c232 − c231 c232 − c231 and 31 <t (71)

 c232


(P, 0) otherwise
 2 2
  2 
c31 c32 c
s(t, c31 , c21 , c32 ) = (1 − t) log ±
(1 − t) + t log 32 t (72)
(c32 − c31 )c (c21 , c31 )
2 2 2 2 c231
Since we do not have closed form expressions for the bounds in the TD case, we cannot exactly
bound the difference between upper and lower bounds as in Corollary 1. However, extensive
numerical evaluation of the bounds show that the difference also seems to be bounded, and that

Pg+ − Pg− ≤ 1dB (73)

For the TD case without power allocation, we can find the following bounds for the asymptotic
relay gain in Rayleigh fading using the results from Propositions 1-2 and taking limits
1 
Rg+ = max min αl(κs21 , κs31 ) + (1 − α) log (κs31 ) ,
2 κ∈[0,1],α∈[0,1]
 
1−κ 1
α log (κs31 ) + (1 − α)l κs31 , s32 − log s31 (74)
1−α 2
1 
Rg− = max min α log(κs21 + κs31 ) + (1 − α) log (κs31 ) ,
2 κ∈[0,1],α∈[0,1]
 
1−κ 1
α log (κs31 ) + (1 − α)l κs31 , s32 − log s31 (75)
1−α 2

4.3 Discussions on the High SNR Case

The asymptotic relay gain characterizes the gain from relaying at reasonably high SNR. Further-
more, since the difference between upper and lower bounds is bounded (see Corollary 1), we can use

19
the achievable rates to characterize the gain of relaying. Figures 5-6 compare the asymptotic relay
gain for a number of different signaling schemes over Rayleigh fading. The comparison is done for
values of s21 and s32 between −10dB and 20dB. Figure 5(a) shows the relay gain for full duplex
synchronized relaying. Figures 5(b)-6 shows the loss by different constraints on the signaling.

18 8

16 7

14 6
12

Power gain (dB)

5
Power gain (dB)

10
4
8
3
6
2
4

2 1

0 0
20 20
15 15
20 20
10 15 10 15
5 10 5 10
0 5 0 5
0 0
−5 −5 −5
−5
−10 −10 −10 −10
s32 (dB) s32 (dB)
s21 (dB) s21 (dB)

Figure 5: (a) High SNR relay gain for full duplex synchronized (b) Difference between full duplex
and TD for synchronized signaling

5 Conclusion
In this paper, we have studied upper bounds and lower bounds on the ergodic capacity (and outage
capacity) for a variety of wireless relay channel models. In many cases of interest, the gap between
the upper bounds and lower bounds is small compared to the relaying gain (as shown by both
analytical and numerical examples), which sheds insights on the channel capacity. A number of
conclusions can be drawn based on our results:

• Relaying yields performance gain in both ergodic capacity and outage capacity, relative to
direct transmission without a relay.

• Optimal Relaying outperforms traditional multihop protocols.

• Power allocation can yield a significant gain in wireless relay channels, in particular when the
relay operates in the time-division mode.

• Since transmitter CSI makes power allocation possible, transmitter CSI leads to higher rates,
even at high SNR. This is in contrast to the case of the point-to-point single-antenna channel
[10, 1], but is in line with the point-to-point MIMO channel [18, 26].

20
 2.5

1
2
0.9

Power gain (dB)

0.8 1.5
0.7
Power gain (dB)

0.6 1

0.5
0.5
0.4

0.3 20
0
0.2 20
10 15
0.1 20
10 15
0 0 5 10
20 0 5
15 10 5 0
0 −10 −5 −5
−5 −10 s21 (dB)
−10 −10
s32 (dB)
s32 (dB) s21 (dB)

Figure 6: High SNR relay gain, achievable rate for TD: (a) Difference between synchronized and
asynchronous signaling (b) Difference between asynchronous signaling with power allocation and
asynchronous signaling without power allocation.

Appendix A: Proofs for The Fixed channel Gain Case

Proof of Proposition 1

Let X1,1 , X1,2 , . . . , X1,n be the transmitted sequence from node 1. The relay, node 2, receives
the first αn transmitted signals as Y2,1 , Y2,2 , . . . , Y2,αn , and after that transmits a sequence
X2,αn+1 , X2,αn+2 , . . . , X2,n . The result in [3, Theorem 4] or [4, Theorem 14.10.1] gives the
following non-single letter bound (i.e., equation (53) in [3]) on the rate R
 n #
" "
n
nR ≤ min I(X1,k ; Y2,k , Y3,k |X2,k ), I(X1,k , X2,k ; Y3,k ) + nn (76)
k=1 k=1

In the relay-receive period the relay does not transmit and therefore for k ≤ αn I(X1,k ; Y2,k , Y3,k |X2,k ) =
I(X1,k ; Y2,k , Y3,k ) and I(X1,k , X2,k ; Y3,k ) = I(X1,k ; Y3,k ). On the other hand, in the relay-transmit
period the relay does not receive, and therefore for k > αn I(X1,k ; Y2,k , Y3,k |X2,k ) = I(X1,k ; Y3,k |X2,k ).
We then get

αn
" "
n
nR ≤ min I(X1,k ; Y2,k , Y3,k ) + I(X1,k ; Y3,k |X2,k ),

k=1 k=αn+1

αn
" "
n 
I(X1,k ; Y3,k ) + I(X1,k , X2,k ; Y3,k ) + nn (77)

k=1 k=αn+1

21
By letting n → ∞ and using standard arguments as in [4] we get the following single letter bound
on the capacity

(1) (1) (1) (2) (2)
C ≤ max min αI(X1 ; Y2 , Y3 ) + (1 − α)I(X1 ; Y3 |X2 ),
(1) (2)
X1 ,X1 ,X2

(1) (1) (2) (2)
αI(X1 ; Y3 ) + (1 − α)I(X1 , X2 ; Y3 ) (78)

In the Synchronized Gaussian case the same argument as in [3] shows that this bound is maximized
(1) (1) (2) (2)
by letting X1 be Gaussian with
 power P1 , X2 and X1 Gaussian with powers P2 and P1
(2) ∗ (2)
respectively and E[X1 X2 ] = βP1 P2 . Theorem 1 follows.
In the asynchronous case the phases of the source and relay transmissions cannot be synchro-
(2)
nized, i.e., the argument of the complex random variables X1 and X2 are independent. To derive
an upper bound in this case, first notice that
(2) (2) (2) (2) (2)
I(X1 ; Y3 |X2 ) = H(Y3 |X2 ) − H(Y3 |X1 , X2 ) (79)
(2)
≤ H(Y3 ) − H(Z3 ) (80)
(2) (2)
since conditioning reduces entropy. Next, lemma 1 below shows that I(X1 , X2 ; Y3 ) is maximized
(2)
for X1 , X2 independent Gaussian. This shows that (78) provides an upper bound in the asyn-
(2)
chronous case by letting X1 , X2 be independent Gaussian, and this equivalent to putting β = 0
in the expressions (11) and (12).

Lemma 1 Let X and Y be complex random variables with an arbitrary joint distribution satisfying
E[|X|2 ] ≤ Px , E[|Y |2 ] ≤ Py , let Z be Gaussian independent of X and Y , and let θ be uniform over
[0, 2π], and independent of X and Y . Put R = X + Y ejθ + Z. Then the differential entropy h(R)
is maximized for X and Y independent Gaussian.

Proof: Notice that

E[|R|2 ] = E[|X|2 ] + E[|Y |2 ] + 2

{E[XY ∗ ]E[ejθ ]} + E[|Z|2 ] (81)
≤ Px + Py + E[|Z|2 ] (82)

But among all random variables with power bounded by Px + Py + E[|Z|2 ] the Gaussian random
variable with power Px + Py + E[|Z|2 ] maximizes entropy [4], and that is achieved when X and Y
are independent Gaussian. 2

Proof of Proposition 2

There are two coding methods for the relay channel that give the same rate (and the capacity for
the degraded relay channel [3]): The original list decoding method in [3] and backward decoding

22
in [49] based on [46]. A simpler argument using parallel (Gaussian) channel arguments, due to
Matthew Valenti [41], gives the same rate as the two methods above. It can be shown that also in
the TD case, all three coding methods give the same rate; we will provide an argument based on
parallel (Gaussian) channels. It can be noticed that it is not necessary to use block-Markov coding
[3], as opposed to the full duplex relay channel.
We split the message into two independent parts: A part wd which is transmitted directly to
the destination without the help of the relay at a rate Rd , and a part wr which is transmitted
through the relay to the destination at a rate Rr . The total rate is then R = Rr + Rd .
(1)
We define three codebooks: X1 (wr ) has αn elements distributed i.i.d. according to a Gaussian
(1) (2) (2)
distribution with power P1 , X2 (wr ) has (1 − α)n elements with power βP1 and X1 (wd ) has
(2)
(1 − α)n elements with power (1 − β)P1
The transmission and decoding scheme is as follows. During the relay-receive period, the source
(1)
transmits X1 (wr ). The relay can decode wr if
1  
(1)
Rr < α log 1 + c221 P1 , (83)
2
$
P2
During the relay-transmit period, the relay transmits (2) X2 (wr ). The source transmits X2 (wr )+
βP1
(2)
X1 (wd ).
(1) (2) (2)
The destination starts by decoding wr from (y3 , y3 , treating X1 (wd ) as noise. Using the
theory for parallel Gaussian channel [4], it can do so if
  2 
(2) √
   βc231 P1 + c32 P2 
1 (1) 1  
Rr ≤ α log 1 + c31 P1
2
+ (1 − α) log 1 +  (84)
2 2  (2)
1 + (1 − β)c31 P1
2 

 
(2) (2) (2) √
It then subtracts X2 (wr ) from the received signal, ỹ3 = y3 − 2
βc31 P1 + c32 P2 X2 (wr ),
(2) (2)
and decodes X1 (wd ) from ỹ3 ; it can do so if
1  
(2)
Rd ≤ (1 − α) log 1 + c231 (1 − β)P1 (85)
2
Adding (85) and (84) gives (13), (83) and (84) gives (14).

Proof of Proposition 3

As for the proof of Proposition 2 we split the message into two parts: wd transmitted directly and
wr transmitted through the relay, with a total rate R = Rd + Rr .

23
 The encoding and transmission schemes are as follows. The message wr is encoded using the
(1)
codebook X1 (wr ) with αn elements distributed i.i.d. according to a Gaussian distribution with
(1) (1)
power P1 . The signal X1 (wr ) is transmitted during the relay-receive period. The resulting
received signal at the relay, Y2 is compressed to the index s ∈ {1, . . . , 2nR0 } (how this is done
is outlined below). During the relay-transmit period the relay transmits X2 (s), with a Gaussian
(2)
codebook with (1 − α)n elements with power P2 . Simultaneously the source transmits X1 (wd )
(2)
with (1 − α)n with power P1 .
The destination starts by decoding s and wd . During the relay-transmit period the channel is
a multiple access channel, and it can therefore do so if2
 
1 c232 P2
R0 ≤ (1 − α) log 1 + (2)
(86)
2 1 + c231 P1
1  
(2)
Rd ≤ (1 − α) log 1 + c231 P1 (87)
2
(1)
It now uses s and Y3 to decode wr . To see how this is done, we will outline the compression
method in detail.
The compression at the relay is done using Wyner and Ziv’s rate distortion theory with side
information at the decoder, [47] [4, Theorem 14.9.1]. We will outline how this is done following
the steps in the proof of [4, Theorem 14.9.1]. The first steps are identical to the steps in that
proof (except for the continuous alphabet): we generate an i.i.d. codebook W(t) with 2αnR1 ,
R1 = I(Y2 ; W ) + , codewords. The distribution of W, Y2 is chosen as a joint Gaussian distribution
given by

W = Y2 + ZW (88)
2 , independent of Y ,Y and X . The 2αnR1 codewords are
where ZW is Gaussian with power σw 2 3 1

divided into 2αnR2 , R2 = I(Y2 ; W ) − I(Y3 ; W ) + 5, bins. The relay finds W(t) so that Y2 and
W(t) are jointly typical, and then transmits the bin index s(t) to the destination. The destination
(1)
looks for a W(t) so that t is in bin number s and so that W(t) and Y3 are jointly typical. The
proof of [4, Theorem 14.9.1] shows that this can be done with small probability of error if n is large.
(1)
Since they are jointly typical, the distribution of (Y2 , Y3 , W(t)) is close to the original distri-
(1)
bution (step 5 in the proof of [4, Theorem 14.9.1]). The destination looks for a codeword X1 (ŵr ) so
(1) (1)
that (X1 (ŵr ), Y3 , W(t)) are jointly typical. Since ZW and Z3 are independent, this is equivalent
(1)
to maximal ratio combining (MRC) on W and Y3 , and we get the following rate
 
1 c221 P1
Rr ≤ α log 1 + c31 P1 +
2
2
(89)
2 1 + σw
2
R0 and Rd can be chosen anywhere on the ’sum rate side’ in the MAC pentagon (between C and B in [4, Figure
14.14]) with the same resulting rate for the relay channel.

24
 2 of Z . This is determined by the constraint that the
The last step is to find the power σw W
number of bins 2αnR2 should be less than the number of codewords 2nR0 the relay has available, so

R0 /α ≥ I(Y2 ; W ) − I(Y3 ; W ) + 5 (90)

= H(W |Y3 ) − H(W |Y2 ) + 5 (91)
 
1 (c31 c21 P1 )2 1
= log c21 P1 + 1 + σw − 2
2 2
− log(σw
2
) + 5 (92)
2 c31 P1 + 1 2
c231 P1 + c221 P1 + 1
2
σw ≥  2(R /α−5)  (93)
2 0 − 1 (c231 P1 + 1)

Inserting (86) in (93) gives (16) and (87) and (89) with R = Rr + Rd give (15).

Appendix B: Bounds on Ergodic Capacity: The Full Duplex Re-

laying Case
For later use in calculating limits for SNR → ∞, we will re-introduce the explicit noise power σ 2 .
Let {c[i]}∞
i=1 be a specific realization of the fading process. The max-flow-min-cut bound gives

"
n
nR ≤ I(X1 [i]; Y2 [i], Y3 [i]|X2 [i], c[i]) + nn (94)
i=1
 
"n
1
(1)
(c221 [i] + c231 [i])P1 [i]
≤ log 1 + +nn (95)
2 σ2
i=1
' () *
R1n

and
"
n
nR ≤ I(X1 [i], X2 [i]; Y3 [i]|c[i]) + nn (96)
i=1
 
"n
1
(1)
c231 [i]P1 [i] + (c231 [i] + c232 [i])Pj [i]
≤ log 1 + +nn (97)
2 σ2
i=1
' () *
R2n

n (1) (1)
The power constraint is i=1 (P1 [i] + Pj [i]) ≤ nP , together with P1 [i] ≥ 0, Pj [i] ≥ 0. Notice
(1)
that we can always find an optimum solution with R1n = R2n . If we have a solution (P1 [i], Pj [i])
(1) (1) (1)
with R1n > R2n , put (P̃1 [i], P̃j [i]) = ((1 − ε)P1 [i], Pj [i] + εP1 [i]), and let the corresponding rates
be R̃1n , R̃2n . Then R̃1n is a decreasing function of ε, R̃1n = 0 for ε = 1, and R̃2n an increasing function.
By continuity we then have R̃1n = R̃2n ≥ R2n for some ε. Similarly for R1n < R2n .

25
 Thus, we can state the optimization problem as the Lagrange problem of maximizing3

λ " (1)
n
J = R1n + t(R2n − R1n ) + (P1 [i] + Pj [i]) (98)
2 ln 2
i=1
λ "
n
(1)
= (1 − t)R1n + tR2n + (P1 [i] + Pj [i]) (99)
2 ln 2
i=1

A few comments on this Lagrange optimization problem. Fix λ < 0, and suppose that for some t
we have found an unconstrained maximum to J satisfying R1n = R2n . Then this is also a solution
to the original constrained problem for some power constraint. In addition, the problem splits up
in individual coordinate problems:
"
n
J = J[i] (100)
i=1
 
(1)
1−t (c221 [i] + c231 [i])P1 [i]
J[i] = log 1 +
2 σ2
 
(1)
t c231 [i]P1 [i] + (c231 [i] + c232 [i])Pj [i]
+ log 1 +
2 σ2
λ (1)
+ (P [i] + Pj [i]) (101)
2 ln 2 1
Thus the problem reduces to the problem maximizing J[i] for given t, λ.
(1)
Differentiating with respect to P1 [i] and Pj [i] and equating to zero, we get the equations
 
tc231 [i] (1 − t) c221 [i] + c231 [i]
(1)
+ (1)   +λ = 0 (102)
σ 2 + c231 [i]P1 [i] + (c231 [i] + c232 [i])Pj [i] σ 2 + P1 [i] c221 [i] + c231 [i]
 
t c231 [i] + c232 [i]
(1)
+λ = 0 (103)
σ 2 + c231 [i]P1 [i] + (c231 [i] + c232 [i])Pj [i]
From the second equation we get
(1)
t σ 2 + c231 [i]P1 [i]
Pj [i] = − − (104)
λ c231 [i] + c232 [i]
(1)
There are four possibilities, depending on whether P1 [i] and Pj [i] are zero, or strictly positive.
(1)
First consider the case Pj [i] > 0, P1 [i] > 0. Then

(1) (1 − t)(c231 [i] + c232 [i]) σ2

P1 [i] = − 2 − 2 >0 (105)
c32 [i]λ c21 [i] + c231 [i]
(1)
t σ 2 + c231 [i]P1 [i]
Pj [i] = − − >0 (106)
λ c231 [i] + c232 [i]
3 (1)
The explicit Lagrange multipliers corresponding to the constraints P1 [i] ≥ 0, Pj [i] ≥ 0 have been omitted

26
 (1)
Second consider Pj [i] > 0, P1 [i] = 0. Then

t σ2
Pj = − − 2 >0 (107)
λ c31 [i] + c232 [i]

with the condition

(1 − t)(c231 [i] + c232 [i]) σ2
− − ≤0 (108)
c232 [i]λ c221 [i] + c231 [i]
(1)
Consider now the cases with Pj [i] = 0. P1 [i] is then the solution to
 
tc231 [i] (1 − t) c221 [i] + c231 [i]
(1)
+ (1)   +λ = 0 (109)
σ 2 + c231 [i]P1 [i] σ 2 + P1 [i] c221 [i] + c231 [i]

Or (with σ 2 = 1)

(1)
λc231 [i](c221 [i] + c231 [i])P1 [i]2 + (c431 [i] + c221 [i]c231 [i] + λc221 [i]
(1)
+2λc231 [i])P1 [i] + λ + c231 [i] + (1 − t)c221 [i] = 0 (110)

By looking at the coefficients of this second order equation, it is seen that if

1 1
≤ − (111)
c231 [i] + (1 − t)c221 [i] λ

it has one negative and one positive root; otherwise, it has two negative roots. If (111) is satisfied,
(1) (1)
and (107) is not a solution, both P1 [i] = 0 and P1 [i] equal to the positive solution to (110)
(both with Pj [i] = 0) are solutions to the Lagrange problem. The correct solution can be found by
inserting each of the solutions in the expression (101) for J[i] and choosing the one that maximizes
J[i].
Then, by letting n → ∞ and using the ergodicity of the fading process, we arrive at the solution
in table 1. For simplicity of notation we have used µ = − λ1 for the Lagrange parameter; µ can be
interpreted as a water filling level similar to ν in [4, Section 10.4].
For the achievable rate we consider decode-forward. If c31 > c21 the source bypasses the relay.
If c31 < c21 the source transmits to the relay, which decodes the transmission, puts it into a queue
and transmits the output of the queue to the destination. Call the direct rate Rd , the rate through

27
the relay Rr with total rate R = Rd + Rr . We then get the following rate
 
1 2
Rd = P (c31 > c21 ) Ec31 >c21 log(1 + c31 P1 (c)) (112)
2
  
1 2
Rr = min P (c31 < c21 ) Ec31 <c21 log(1 + c21 P1 (c)) ,
2
 
1
P (c31 < c21 ) Ec31 <c21 log(1 + c231 P1 (c))
2
  
1 (c231 + c232 )Pj (c)
+Ec31 ,c32 log 1 + (113)
2 1 + c231 P1 (c)
  
1
R = min Ec31 ,c21 log(1 + max{c21 , c31 }P1 (c)) ,
2 2
2
 
1 2 2 2
Ec31 ,c32 log(1 + c31 P1 (c) + (c31 + c32 )Pj (c)) (114)
2
(1)
For achievable rate, any power allocation rule (Pj (c), P1 (c)) will give a valid rate as seen by
the arguments in [10]. The results in Table 1 is obtained by replacing c221 + c231 with max{c221 , c231 }
everywhere in the Lagrange optimization problem.
Finally, the asynchronous solution follows along the same lines by replacing c231 + c232 with c232
everywhere.

Appendix C: Bounds on Asymptotic Rate Gains: The Full Duplex

Relaying Case

Proof of Proposition 6

First we will prove the power allocation rule (52). Notice that for σ 2 suitably small, the condition
(1)
(108) is not satisfied, and Pj [i] > 0, P1 [i] = 0 is not a solution. On the other hand, (105) is
satisfied, and inserting (105) in (106) we get
 
t (1 − t)c231 [i] 1 σ2
Pj [i] = − + − 1 − (115)
λ c232 [i]λ c221 [i] + c231 [i] c231 [i] + c232 [i]

We see that Pj [i] > 0 for arbitrary small σ 2 if and only if

c231 [i] t
≤ = κ. (116)
2
c32 [i] 1−t

= −λ−1 , i.e.,
(1)
For σ 2 → 0 the largest solution of the equation (109) converges towards P1
P1 (σ 2 ) = −λ−1 + (σ 2 ). The condition that this is a solution to the Lagrange problem is that
(1)

28
Pj [i] given by (104) is negative,

t σ 2 + (σ 2 ) c231 [i] 1

− −− 2 2 + 2 2 <0 (117)
λ c31 [i] + c32 [i] c31 [i] + c32 [i] λ

For σ 2 → 0 this is the opposite condition of (116).

Thus, we have proven that for a specific value of c and fixed λ, the Lagrange solution converges
towards (again with µ = −λ−1 )

(P1∞ (c), Pj∞ (c)) =
(1)
lim (P1 (c), Pj (c))
σ 2 →0
    
 c231 +c232 (1 − t)µ, t − c231
(1 − t) µ
c231
<κ
c232 c232 c232
= (118)
 (µ, 0) otherwise

Furthermore, the solution (118) gives a constant total transmit power µ, and it follows (i.e., Lebe-
 
(1)
gues dominated convergence) that limσ2 →0 E P1 + Pj = µ, and the solution (52) follows. The
proof for the achievable rate is similar.
Let f (cij ) denote the probability measure for c2ij . If c2ij has a probability density function f  (c2ij )
df (cij ) = f  (c2ij )dc2ij in the following.
Now, from continuity and the Lebesgue dominated convergence,
+  
1 c(c221 , c231 ) (1) 1
R1∞ (t) = lim log 1 + P1 (c, σ 2
) df (c31 )df (c21 )df (c32 ) + log(σ 2 ) (119)
σ→0 2 σ2 2
+
1    
= log c c221 , c231 P1∞ (c) df (c31 )df (c21 )df (c32 ) (120)
2
+  
(1)
∞ 1 c2 P (c, σ 2 ) + (c231 + c232 )Pj (c, σ 2 )
R2 (t) = lim log 1 + 31 1 df (c31 )df (c21 )df (c32 )
σ→0 2 σ2
1
+ log(σ 2 ) (121)
2
+
1  
= log c231 P1∞ (c) + (c231 + c232 )Pj∞ (c) df (c31 )df (c21 )df (c32 ) (122)
2
  2 
c
We have to find t so that R1∞ (t) = R2∞ (t). Define A = (c31 , c32 )  c31 2 < 1−t
t
. Then
32

+ +  
c2 + c2
R1∞ (t) = log c(c21 , c31 ) 31 2 32 (1 − t)P df (c31 )df (c32 )df (c21
A R c32
+ +
+ log (c(c21 , c31 )P ) df (c31 )df (c32 )df (c21 ) (123)
R2 −A R
+ +
∞
 2 2
   
R2 (t) = log c31 + c32 tP df (c31 )df (c32 ) + log c231 P df (c31 )df (c32 ) (124)
A R2 −A

29
The equation R1∞ (t) = R2∞ (t) now reduces to
+ +
 
log (t) df (c31 )df (c32 ) + log c231 df (c31 )df (c32 ) (125)
A R2 −A
+ +
 
= log (c(c21 , c31 )) df (c31 )df (c21 ) − log c232 df (c31 )df (c32 )
R2
+ A

+ log (1 − t) df (c31 )df (c32 ) (126)

A

Or
+ + +
   
log (κ) df (c31 )df (c32 ) + log c231 df (c31 )df (c32 ) + log c232 df (c31 )df (c32 )
A R2 −A A
+
= log (c(c21 , c31 )) df (c31 )df (c21 ) (127)
R2

For Rayleigh fading, where c2ij is exponentially distributed, all the integrals involved can be evalu-
ated in closed form. For example
+ + ∞ + κc232 2   2 
c31 c
df (c31 )df (c32 ) = −s−1 2 −1
dc31 s32 exp − 32 dc232
31 exp (128)
A 0 0 s31 s32
+ ∞  2   2 
c κ c32
= 1 − exp − 32 s−1
32 exp − dc232 (129)
0 s31 s32
s32 κ
= (130)
s32 κ + s31
The other integrals needed are evaluated below. The evaluations and following simplifications have
been done by a symbolic analysis program (Maple) aided by calculations by hand, and are rather
straightforward but tedious
+
  s31 (ln(s31 ) − γ)
ln 2 log c231 df (c31 )df (c32 ) =
R2 −A s32 κ + s31
s32 κ(ln(s32 κ + s31 ) − ln(κ) − ln(s32 ))
+ (131)
s32 κ + s31
+
  κs32 (ln(s32 ) − γ)
ln 2 log c232 df (c31 )df (c32 ) =
A s32 κ + s31
s31 (ln(s32 κ + s31 ) − ln(s31 ))
+ (132)
s32 κ + s31
+
 
ln 2 log max(c221 , c231 ) df (c31 )df (c21 ) = ln (s21 + s31 ) − γ (133)
R2

+
 2   s21 ln s21 − s31 ln s31 − γ s = s
21 31
ln 2 2
log (c21 + c31 ) df (c31 )df (c21 ) = s21 − s31 (134)
R2 
ln s21 + 1 s21 = s31

where γ is Euler’s constant. Inserting this in (127) we then get (55).

30
 We finally calculate the asymptotic relay gain. First notice that in the high SNR regime, power
control for direct transmission is not needed (i.e., it does not give any gain). This follows directly
from the results in [10]. We can therefore define the following rate for direct transmission
+  
∞ 1 c231 P 1
Rd = lim log 1 + 2 df (c31 ) + log(σ 2 ) (135)
σ→0 2 R σ 2
+
1  
= log c231 P df (c31 ) (136)
2 R
The rate gain relative to direct transmission for σ → 0 can then be found as R2∞ (t) − Rd∞ (the term
1
2 log(σ 2 ) appears in both definitions, and therefore cancels out). We get

Rg = R2∞ (t) − Rd∞ (137)

+ +
1    1  
= log c231 + c232 tP df (c31 )df (c32 ) − log c231 P df (c31 )df (c32 ) (138)
A 2 A 2
+  2
 +
1 c32 1
= log 1+ 2 df (c31 )df (c32 ) + log (t) df (c31 )df (c32 ) (139)
A 2 c31 2 A
 
1 s32 1+k
= log (140)
2 s32 − s31 1 + k ss32
31

Proof of Proposition 7

For the asynchronous case, we can go through the same set of calculations, replacing Pj with P2
and c231 + c232 with c232 . Notice that for c232 < c231 it does not pay off to let the relay transmit: the
destination knows at least as much about the message as the relay, and it has a better connection.
By going through the Lagrange solution, taking limits, and excluding solutions with c232 < c231 and
P2 > 0, we get the asymptotic power allocation law

(P1∞ (c), P2∞ (c)) =
(1)
lim (P1 (c), P2 (c))
σ 2 →0
  2   
 c32 c231 c231
c232 −c231
(1 − t)µ, t− c232 −c231
(1 − t) µ c232
<t
= (141)
 (µ, 0) otherwise

As for the synchronized case, the transmit power is constant, and we therefore
 arrive at the solution
 
∞ ∞  c231
(58). We have to find t to solve R1 (t) = R2 (t). Denote by A = (c31 , c32 )  c2 < t . Then
32

+ +  
∞ c232
R1 (t) = log c(c21 , c31 ) 2 (1 − t)P df (c31 )df (c32 )df (c21
A R c32 − c231
+ +
+ log (c(c21 , c31 )P ) df (c31 )df (c32 )df (c21 ) (142)
R2 −A R
+ +
∞
 2   
R2 (t) = log c32 tP df (c31 )df (c32 ) + log c231 P df (c31 )df (c32 ) (143)
A R2 −A

31
The equation R1∞ (t) = R2∞ (t) reduces to
+ +
 
log (t) df (c31 )df (c32 ) + log c231 df (c31 )df (c32 ) (144)
A R2 −A
+ +
 
= log (c(c21 , c31 )) df (c31 )df (c21 ) − log c232 − c231 df (c31 )df (c32 )
R2 + A

+ log (1 − t) df (c31 )df (c32 ) (145)

A

Or
 + +
t  
log df (c31 )df (c32 ) + log c231 df (c31 )df (c32 )
1−t A R2 −A
+
 2 
+ log c32 − c231 df (c31 )df (c32 )
+A
= log (c(c21 , c31 )) df (c31 )df (c21 ) (146)
R2

To solve this we need the following integral in addition to (130-134)

+ + ∞ + t y
 2  1−t
ln 2 log c32 − c31 df (c31 )df (c32 ) =
2
ln(y) fc32 (c231 + y)fc31 (c231 )dc231 dy (147)
A 0 0

The integral can be evaluated in closed form, but gives a complicated expression that we will not
write down here. However, inserting this in (146) and reducing, we get the following left hand side
of (146)
 
ts32 +s31
s32 ln 1−t + s31 ln(s31 )
−γ (148)
s31 + s32
while the right hand side is unchanged from the synchronized case. The rate gain relative to direct
transmission for σ → 0 can now be evaluated to
+ +
1  2  1  
Rg = log c32 tP df (c31 )df (c32 ) − log c231 P df (c31 )df (c32 ) (149)
A 2 A 2
 
1 s32
= log t +1 (150)
2 s31

Proof of Corollary 1

We will briefly outline the proof of the Corollary. For convenience we normalize all sij by s31 ,
s̃ij = sij /s31 . It is first proven that for each fixed value of s̃21 , Rg+ − Rg− is an increasing function
of s̃32 for both the synchronous and asynchronous case; therefore Rg+ − Rg− ≤ lims32 →∞ Rg+ − Rg−
(provided the limit exists). In both cases
 
1 2l(s̃21 ,1)
lim Rg+ − Rg− = log (151)
s̃32 →∞ 2 s̃21 + 1

32
It is now easy to prove that the expression inside the log has its maximum at s̃21 = 1. The
expression (151) actually gives an upper bound on the difference Rg+ − Rg− for any fixed value of
s̃21 = s21 /s31 , that can be used to show how close the bounds are away from the worst case.

Appendix D: Bounds on Ergodic Capacity: The Time-Division Re-

laying Case
Let {c[i]}∞
i=1 be a specific realization of the fading process. The max-flow-min-cut bound gives

"
n
nR ≤ I(X1 [i]; Y2 [i], Y3 [i]|X2 [i], c[i]) + nn (152)
i=1
"n  
(1)
≤ f1 [i] P1 [i], Pj [i] +nn (153)
i=1
' () *
R1n
"
n
nR ≤ I(X1 [i], X2 [i]; Y3 [i]|c[i]) + nn (154)
i=1
 
"n
1
(1)
c231 [i]P1 [i] + (c231 [i] + c232 [i])Pj [i]
≤ log 1 + +nn (155)
2 σ2
i=1
' () *
R2n
  
 1 (c221 [i]+c231 [i])P1
log 1 + 2 Pj = 0
f1 [i](P1 , Pj ) =
2  σ
(156)
 1 c231 [i]P1
2 log 1 + σ2
Pj > 0

n (1) (1)
The power constraint is i=1 (P1 [i] + Pj [i]) ≤ nP , together with P1 [i] ≥ 0, Pj [i] ≥ 0. We
will argue that we can restrict attention to solutions with R1n = R2n . First, if we have a solution
with R1n < R2n we can find a solution which is at least as good with R1n = R2n by decreasing Pj [i].
Therefore consider a sequence of solutions {R1n , R2n }∞
n=1 with R1 ≥ R2 and limn→∞ n R2 = R.
n n 1 n

R1n is a sum of terms where the relay receives (Pj [i] = 0) and terms where the relay transmits
(Pj [i] > 0). By ”switching off” the relay reception for some of the terms where the relay receives,
we can decrease R1n to R̃1n ≤ R2n . Furthermore, for any ε > 0 and for n large enough we can do
this so that 1 n
n (R2− R̃1n ) < ε. By decreasing R2n as above we then have a solution R̃1n = R̃2n and
 ∞
1
(R n − R̃n ) < ε. We can therefore find a sequence of solutions R̃n , R̃n with R̃1n = R̃2n and
n 2 1 1 2
n=1
limn→∞ n1 R̃2n = R.

33
 As for full duplex, the problem reduces to maximization of J[i], where
"
n
J = J[i] (157)
i=1
 
  t (1)
c231 [i]P1 [i] + (c231 [i] + c232 [i])Pj [i]
(1)
J[i] = (1 − t)f1 [i] P1 [i], Pj [i] + log 1 +
2 σ2
λ (1)
(P [i] + Pj [i])
+ (158)
2 ln 2 1
The difference is that J[i] is no longer continuous in the variable Pj , see (156). We therefore have
(1)
to find both the maximum on the boundary Pj [i] = 0 (this corresponds to the solution P1,2 , Pj,2 in
(1)
Table 2) and in the interior Pj [i] > 0 (this corresponds to the solution P1,1 , Pj,1 in Table 2). These
two solution should then be compared to see which one maximizes J[i], which is done in Table 2
by comparing Rt,1 and Rt,2 . The detailed calculations to find the solutions is along the same lines
as for full duplex.
For the asymptotic solution when σ → 0, we can argue as for full duplex that only two of the
solutions are relevant for σ small enough, and that the solution with Pj = 0 (joint transmission)
c231 t
is selected if c232
< 1−t , with the additional condition that the joint transmission solution is only
used if it maximizes J[i], which in the asymptotic limit reduces to the condition
 2 2 
c31 (c31 + c232 )  2 
(1 − t) log (1 − t)P + t log (c 31 + c 2
32 )tP
c2
 2 32 2   
> (1 − t) log (c21 + c31 )P + t log c231 P (159)

or
   
c231 (c231 + c232 ) c231 1
(1 − t) log (1 − t) > t log 2 (160)
c232 (c221 + c231 ) c31 + c232 t
We then get (69-70).
Define A as the region of (c31 , c21 , c32 ) where joint transmission is used. We can then write the
rates as
+  
c2 + c2
R1∞ (t) = log c231 31 2 32 (1 − t)P df (c31 )df (c32 )df (c21
A c32
+
+ log (c(c21 , c31 )P ) df (c31 )df (c32 )df (c21 ) (161)
R2 −A
+ +
 2    
R2∞ (t) = 2
log c31 + c32 tP df (c31 )df (c32 ) + log c231 P df (c31 )df (c32 ) (162)
A R2 −A
The equation R1∞ (t) = R2∞ (t) reduces to
 + +
t  
log df (c31 )df (c32 ) + log c231 df (c31 )df (c32 ) (163)
1−t A R2 −A
+ +  2 
c
= log (c(c21 , c31 )) df (c31 )df (c21 ) + log 31 df (c31 )df (c32 ) (164)
R2 −A A c232

34
It appears this equation (and the integrals) must be solved numerically to find t because of the
complicated shape of A. The rate gain can be found (numerically) from (139) with the new
definition of A.
For the asynchronous case we use joint transmission if
 2 2 
c31 c32  2   2 
(1 − t) log 2 (1 − t)P + t log c 32 tP > (1 − t) log (c(c21 , c31 )P ) + t log c31 P (165)
c32 − c231

Or
   2 
c232 c32
(1 − t) log c231 (1 − t) + t log t >0 (166)
(c232 − c231 )c(c21 , c31 ) c231

The rates are now

+ +  
c232
R1∞ (t)= 2
log c31 2 (1 − t)P df (c31 )df (c32 )df (c21
A R c32 − c231
+ +
+ log (c(c21 , c31 )P ) df (c31 )df (c32 )df (c21 ) (167)
2
+ R −A R +
∞
 2   
R2 (t) = log c32 tP df (c31 )df (c32 ) + log c231 P df (c31 )df (c32 ) (168)
A R2 −A

The equation R1∞ (t) = R2∞ (t) reduces to

 + +
t  
log df (c31 )df (c32 ) + log c231 df (c31 )df (c32 ) (169)
1−t A R2 −A
+ +  
c231
= log (c(c21 , c31 )) df (c31 )df (c21 ) + log 2 df (c31 )df (c32 ) (170)
R2 −A A c32 − c231

And the rate gain can be found from (149).

Appendix E: Proof of Proposition 8

We use the asymptotic power allocation rule from the asynchronous case, i.e., P1 (c) and P2 (c) are
functions of c given by (58). For now we let the value of t be open (i.e., some fixed unknown value);
the value of t is found be optimizing over t the solution found in the following. Consider a fixed
c. The received signal at nodes 2 and 3 is then Gaussian, and we can use the same coding method
as in Proposition 3, modified to full duplex. Thus, the received signal at node 2 is Wyner-Ziv
compressed to a rate Rs (c), and the power of the resulting ’compression noise’ given by (93), which
for full duplex and with σ 2 explicitly reintroduced is

c231 P1 (c) + c221 P1 (c) + σ 2

σw (c) = σ 2   (171)
2Rs (c) − 1 (c231 P1 (c) + σ 2 )

35
The problem is how to choose the function Rs (c). We suggest to choose Rs (c) so that σw becomes
a constant independent of c. Solving for Rs (c) results in
 
1 σ 2 c231 P1 (c) + c221 P1 (c) + σ 2
Rs (c) = log 1 + 2 (172)
2 σw c231 P1 (c) + σ 2

The compressed signal is transmitted on the channel between the relay and the destination. The
capacity of this channel, with the power allocation rule (58) is given by
+  
1 c232 P2 (c)
Rf = log 1 + 2 df (c32 )df (c31 ) (173)
2 A σ + c231 P1 (c)
  2 
c
with A = (c31 , c32 )  c31
2 < t . Now, as in the proof of proposition 3, the rate of the compressed
32
signal must match the capacity of the channel between relay and the destination, i.e., we must
have Ec [Rs (c)] = Rf . Since the power allocation rule is determined (for specific t), this amounts to
finding the value of the constant σw . We will restrict ourselves to solving this problem in the high
SNR regime σ 2 → 0 for Rayleigh fading. First, the capacity of the forwarding channel between
relay and destination is
+  
1 c232 P2
Rf∞ = lim log 1 + 2 df (c32 )df (c31 ) (174)
σ 2 →0 2 A σ + c231 P1
  
+ c231
1 c 2
32 t − c232 −c231
(1 − t)
= log 1 + 2
 df (c32 )df (c31 ) (175)
2 A c
c31 c2 −c2 (1 − t)
2 32

+   2 32  31

1 t c32
= log −1 df (c32 )df (c31 ) (176)
2 A 1 − t c231
 
1 s32 1 + ss32
31
t
= log (177)
2 s31 + s32 1−t

Clearly, σw is also a function of σ 2 . As will be seen just below, we get a valid solution for
σ 2 → 0 if we put σw = Σ−1 σ 2 , where Σ is a constant independent of σ 2 . Rs (c) then becomes
  
∞ 1 c221
Rs (c) = log 1 + Σ 1 + 2 (178)
2 c31
 
1 Σ c221 1
= log 1 + + log(1 + Σ) (179)
2 1 + Σ c231 2

and further
+   
1 c221
Ec [Rs∞ (c)] = log 1 + Σ 1 + 2 df (c21 )df (c31 ) (180)
2 R2 c31
 
1 Σs21 Σs21 1
= log + log(1 + Σ) (181)
2 Σs21 − (1 + Σ)s31 (1 + Σ)s31 2

36
For given t we can then find Σ by equalizing Rf∞ given by (177) and Ec [Rs∞ (c)] given by (181).
We will finally calculate the asymptotic relay gain achieved. First, by using maximum ratio
combining at the destination as in the proof of Proposition 3, we get the following rate
+  
1 c231 P1 (c) c221 P1 (c)
R = log 1 + + 2 df (c21 )df (c31 )df (c32 ) (182)
2 R3 σ2 σ + σw 2

and
1
R∞ = lim (R + log(σ 2 )) (183)
2
σ →0 2
+  
1 2 Σ 2
= log c31 P1 (c) + c P1 (c) df (c21 )df (c31 )df (c32 ) (184)
2 R3 1 + Σ 21

The asymptotic relay gain can now be found as R∞ − Rd∞ with Rd∞ given by (136)
+   
2 Σ 2
2Rg = log c31 + c P df (c21 )df (c31 ) (185)
R2 −A 1 + Σ 21
+   
Σ 2 c232
+ log 2
c31 + c (1 − t)P df (c21 )df (c31 ) (186)
A 1 + Σ 21 c232 − c231
+
 
− log c231 P df (c31 ) (187)
R
+  
Σ c221
= log 1 + df (c21 )df (c31 ) (188)
R2 1 + Σ c231
+  
c232
+ log (1 − t) df (c31 )df (c32 ) (189)
A c232 − c231
 
  s log(1 − t) + s log 1 + t s32
Σs21 Σs21 32 31 s31
= log + (190)
Σs21 − (1 + Σ)s31 (1 + Σ)s31 s31 + s32

This should then be maximized with respect to t, bearing in mind that Σ is also a function of t.

References
[1] M.S. Alouini and A.J. Goldsmith. Capacity of rayleigh fading channels under different adaptive
transmission and diversity-combining techniques. IEEE Transactions on Vehicular Technology,
48(4):1165 –1181, July 1999.

[2] A.B. Carleial. Multiple-access channels with different generalized feedback signals. IEEE
Transactions on Information Theory, 28(6):841–850, November 1982.

[3] T.M. Cover and A.A. El Gamal. Capacity theorems for the relay channel. IEEE Transactions
on Information Theory, 25(5):572–584, September 1979.

37
 [4] T.M. Cover and J.A. Thomas. Information Theory. John Wiley & Sons, 1991.

[5] A. Ephremides. Energy concerns in wireless networks. IEEE Wireless Communications, pages
48–59, August 2002.

[6] G.J. Foschini. Layered space-time architecture for wireless communication in a fading environ-
ment when using multi-element antennas. Bell Labs Technical Journal, pages 41–59, Autumn
1996.

[7] A. El Gamal and S. Zahedi. Minimum energy communication over a relay channel. In Proceed-
ings of IEEE International Symposium on Information Theory: ISIT’03 (Yokohama, Japan),
2003.

[8] M. Gastpar, G. Kramer, and P. Gupta. The multiple-relay channel: Coding and antenna-
clustering capacity. In Proceedings of the IEEE International Symposium on Information
Theory: ISIT’02 (Lausanne, Switzerland), June 2002.

[9] M. Gastpar and M. Vetterli. On the asymptotic capacity of Gaussian relay networks. In
Proceedings of the IEEE International Symposium on Information Theory: ISIT’02 (Lausanne,
Switzerland), June 2002.

[10] A.J. Goldsmith and P.P. Varaiya. Capacity of fading channels with channel side information.
IEEE Transactions on Information Theory, 43(6):1986–1992, November 1997.

[11] A.J. Goldsmith and S.B. Wicker. Design challenges for energy-constrained ad-hoc wireless
networks. IEEE Wireless Communications, pages 8–27, August 2002.

[12] M.O. Hasna and M. Alouini. Application of the harmonic mean statistics to the end-to-
end performance of transmission systems with relays. In Proceedings of the IEEE Global
Telecommunications Conference: GLOBECOM’02 (Taipei, Taiwan, November 2002.

[13] M.O. Hasna and M. Alouini. Optimal power allocation for relayed transmissions over rayleigh
fading channels. In Proceedings of the 57th IEEE Semiannual Vehicular Technology Conference:
VTC’03-Spring, April 2003.

[14] M.O. Hasna and M. Alouini. Outage probability of multihop transmission over nakagami
fading channels. IEEE Communications Letters, May 2003.

[15] M.O. Hasna and M. Alouini. A performance study of dual-hop transmissions with fixed gain
relays. In Proceedings of IEEE International Conference on Acoustics, Speech, and Signal
Processing. ICASSP ’03, April 2003.

38
[16] M.O. Hasna and M. Alouini. Outage probability of multihop transmission over Nakagami
fading channels. In Proceedings of the IEEE International Workshop on Advances in Wireless
Communications: ISWC’02 (Taiwan), Fall, 2002.

[17] A. Høst-Madsen. On the capacity of wireless relaying. In Proceedings of the IEEE Vehicu-
lar Technology Conference: VTC’02Fall (Vancouver, BC), 2002. Available at http://www-
ee.eng.hawaii.edu/∼madsen/publist2.htm.

[18] S. K. Jayaweera and H. V. Poor. Capacity of multiple-antenna systems with both receiver and
transmitter channel state information. IEEE Transactions on Information Theory, 49(10):2697
– 2709, October 2003.

[19] O.H. Koymen, V. Rodoplu, and T.H. Meng. Throughput characteristics of a minimum energy
wireless network. In Proceeding of the IEEE International Conference on Communications:
ICC’01 (Helsinki, Finland), volume 8, pages 2568–2572, June 2001.

[20] G. Kramer, P. Gupta, and M. Gastpar. Information-theoretic multihopping for relay networks.
In 2004 International Zurich Seminar on Communications, 2004.

[21] J. N. Laneman, G. W. Wornell, and D. N. C. Tse. Cooperative diversity in wireless net-

works, efficient protocols and outtage behavior. IEEE Transactions on Information Theory,
Submitted for publication.

[22] J.N. Laneman, E. Martinian, G.W. Wornell, J.G. Apostolopoulos, and J.W. Wee. Comparing
application- and physical-layer approaches to diversity on wireless channels. In Proceedings of
IEEE International Conference on Communications: ICC’03, 2003.

[23] J.N. Laneman and G.W. Wornell. Energy-efficient antenna sharing and relaying for wireless
networks. In In Proceedings of the IEEE Wireless Communications and Networking Confer-
ence: WCNC’00 (Chicago, IL, September 2000.

[24] J.N. Laneman and G.W. Wornell. Exploiting distributed spatial diversity in wireless networks.
In In Proceedings of the 38th Allerton Conference on Communication, Control, and Computing,
October 2000.

[25] J.N. Laneman and G.W. Wornell. Distributed space-time coded protocols for exploiting coop-
erative diversity in wireless networks. In Proceedings of Global Telecommunications Conference:
GLOBECOM’02, November 2002.

39
[26] J. Liu, J. Chen, M. P.C. Fossorier, and A. Høst-Madsen. Capacity-approaching multiple
coding for MIMO Rayleigh fading systems with transmit channel state information. IEEE
Transactions on Communications, Submitted.

[27] I. Maric and R. Yates. Efficient multihop broadcast for wideband systems. In Proceddings of
the 40th Annual Allerton Conference on Communication, Control, and Computing, October
2002.

[28] R. Min, M. Bhardwaj, S. Cho, N. Ickes, E. Shih, A. Sinha, A. Wang, and A. Chandrakasan.
Energy-centric enabling technologies for wireless sensor networks. IEEE Wireless Communi-
cations, pages 28–39, August 2002.

[29] D. M. Pozar. Microwave Engineering. John Wiley & Sons, 1998.

[30] A. Reznik, S.R. Kulkarni, and S. Verdu. Capacity and optimal resource allocation in the
degraded Gaussian relay channel with multiple relays. In Proceedings of the 40th Annual
Allerton Conference on Communication, Control, and Computing, October 2002.

[31] V. Rodoplu and T.H. Meng. Minimum energy mobile wireless networks. IEEE Journal on
Selected Areas in Communications, 17(8):1333–1344, August 1999.

[32] H. Sato. Information transmission through a channel with relay. Technical Report B76-7,
University of Hawaii, March 1976.

[33] B. Schein and R. Gallager. The Gaussian parallel relay network. In Proceedings of the IEEE
International Symposium on Information Theory: ISIT’00 (Sorrento, Italy), June 2000.

[34] A. Sendonaris, E. Erkip, and B. Aazhang. Increasing uplink capacity via user cooperation di-
versity. In Proceedings of the IEEE International Symposium on Information Theory: ISIT’98
(Boston, MA), page 156, August 1998.

[35] A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity-part I: System descrip-
tion. IEEE Transactions on Communications, 51(11):1927 –1938, November 2003.

[36] A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity-part II: Implementation
aspects and performance analysis. IEEE Transactions on Communications, 51(11):1939 –1948,
November 2003.

[37] Wayne Shiroma and Michael P. De Lisio. Quasi-optical circuits. In Wiley Encyclopedia of
Electrical and Electronics Engineering, pages 523–533. John Wiley & Sons, Inc., New York,
1999.

40
[38] W. Stark, H. Wang, A. Worthen, S. Lafortune, and D. Teneketzis. Low-energy wireless com-
munication network design. IEEE Wireless Communications, pages 60–72, August 2002.

[39] A. Stefanov and E. Erkip. Cooperative information transmission in wireless networks. In

Proceedings of the Asian-European ITW 2002 (Breisach, Germany), June 2002.

[40] I.E. Telatar. Capacity of multi-antenna Gaussian channels. European Transaction on Telecom-
munications, 10(6):585–595, November-December 1999.

[41] M. C. Valenti and B. Zhao. Capacity approaching distributed turbo codes for the relay chan-
nel. In Proceedings of the 57th IEEE Semiannual Vehicular Technology Conference: VTC’03-
Spring, April 2003.

[42] M.C. Valenti and N. Correal. Exploiting macrodiversity in dense multihop networks and
relay channels. In Proceedings of the IEEE Vehicular Technology Conference: VTC’02 Fall
(Vancouver, BC), October 2002.

[43] P. Vamroose and E.C. van der Meulen. Uniquely decodable codes for deterministic relay
channels. IEEE Transactions on Information Theory, 38(4):1203–1212, July 1992.

[44] E.C. van der Meulen. Three-terminal communication channels. Advances in Applied Probabil-
ity, 3:120–154, 1971.

[45] E.C. van der Meulen. A survey of multi-way channels in information theory: 1961-1976. IEEE
Transactions on Information Theory, IT-23(1):1–37, January 1977.

[46] F.J. Willems and E.C. van der Meulen. The discrete memoryless multiple-access channel with
cribbing encoders. IEEE Transactions on Information Theory, 31(3):313–327, May 1985.

[47] A.D. Wyner and J. Ziv. The rate-distortion function for source coding with side information
at the decoder. IEEE Transactions on Information Theory, IT-22(1):1–10, January 1976.

[48] L. Yang, M.O. Hasna, and M. Alouini. Average outage duration of multihop communication
systems with regenerative relays. In Proceedings of the 57th IEEE Semiannual Vehicular
Technology Conference: VTC’03-Spring, April 2003.

[49] C. Zeng, F. Kuhlmann, and A. Buzo. Achievability proof of some multiuser channel coding
theorems using backward decoding. IEEE Transactions on Information Theory, 35(6):1160–
1165, November 1989.

41