SOFT DECODE AND FORWARD

IMPROVES COOPERATIVE COMMUNICATIONS

Harold H. Sneessens and Luc Vandendorpe
Communications and Remote Sensing Laboratory, Universit catholique de Louvain,
Louvain-la-Neuve, Belgium. {sneessens, vandendorpe}@tele.ucl.ac.be
ABSTRACT

to keep soft information. This new technique can be seen as

a soft flavor of DF, and will be referred to as Soft-DF in the
sequel. Simulations show that this technique outperforms
AF and DF.
Section 2 introduces the system model and section 3
presents the previously known schemes, AF and DF. The
new cooperative technique, Soft-DF, is developed in section 4. Finally, section 5 compares the performance of the
new scheme with the two previous ones.

Mobile receivers can benefit from an increased spatial diversity through cooperation. Two relaying techniques are
typically used : either Amplify and Forward (AF) or Decode
and Forward (DF). This contribution presents a new technique which combines the main advantages of both AF and
DF. This technique amounts to a DF scheme where all operations are performed in a soft-input soft-output fashion. In
particular, an encoding technique which uses soft values to
incorporate reliability information is proposed. The resulting Soft-DF relay can be seen as an analog signal-to-noise
ratio enhancer. Simulations show that this Soft-DF scheme
outperforms both AF and DF.

2. SYSTEM MODEL

1. INTRODUCTION
Cooperation between mobiles provides spatial diversity and
thus improves communications on fading wireless channels.
This cooperative diversity is particularly suited to cases where
the mobiles cannot support multiple antennas.
User cooperation was first investigated in [1, 2], where
the authors show the interest of such techniques through an
information-theoretic analysis and propose an implementation. Two relaying techniques are presented and compared
in [3, 4]. In the first technique, called Amplify and Forward
(AF), the relay simply amplifies its partners signal and forwards it to the destination. The second technique is called
Decode and Forward (DF) and consists in decoding, reencoding and forwarding the partners signal to the destination. These two techniques were the object of many analyses and improvements in the literature, for instance in [57].
While these works demonstrate the advantages of user
cooperation, DF lacks the main advantages of AF and vice
versa : DF regenerates the signal while AF does not lose
soft information. This contribution presents a new relaying
technique which enables both to regenerate the signal and
The work reported in this paper has partly been funded by the Belgian interuniversity projects PAI MOTION and by the european Network
of Excellence NEWCOM .
Harold H. Sneessens would like to thank the Belgian FNRS for its
financial support.

0-7803-9323-6/05/$20.00 2005 IEEE.

157

We consider the particular case of two users transmitting to

the same destination. The inter-users channels as well as the
uplink channels are mutually independent Rayleigh block
fading channels. They are frequency flat and orthogonal.
After the transmission by each source user in a first time
slot, each partner relays the signal in a second time slot.
The ratio of the second time slot duration to the total duration is defined as the cooperation ratio. The channels are
assumed to be constant during both time slots, and are perfectly known by the corresponding receivers so that coherent detection is used. The destination uses both the direct
and relayed signals to make a joint decision. Although both
users cooperate in a symmetric fashion, we will focus on
one source signal in order to simplify notations. We consider thus one user as the source S and the other as its relay
R, transmitting to a destination D.
For BPSK modulation, the baseband-equivalent discretetime signal of user i {S, R} received by user j {R, D}
(i = j) at time n is


(1)
rij (n) = hij Es,i bi (n) + nj (n),
which describes the transmission of BPSK symbols bi (n)
with an energy per symbol Es,i , experiencing fading of magnitude hij and additive noise at the reception nj (n). The
model assumes that the fading magnitudes hij are independent Rayleigh-distributed samples and that the noise terms
nj (n) are samples of a zero-mean white gaussian noise process with two-sided power spectral density Nj /2.

3. KNOWN COOPERATIVE SCHEMES :

AF AND DF
Relays using the AF technique simply amplify and re-transmit
their partner coded data. The relay chooses a constant amplification factor such that the average energy per symbol
over the consecutive frames equals its average symbol energy Es,R . The destination receives the relayed signal and
combines it optimally with the direct path before decoding.
With the DF technique, the relay decodes its partners
signal, makes hard decisions and checks if errors occured
thanks to a parity check code. Erroneous frames are not
relayed since it would break down the performance at destination, but successfully decoded frames are re-encoded,
possibly with another code, and forwarded to the destination. An adequate choice of the codes makes possible to
construct a cooperative turbo coded scheme, for which an
appropriate iterative decoding algorithm is used at the destination. This happens if both the source user and the relay
use convolutional codes, provided that the relay interleaves
the data before re-encoding. This leads to the interpretation
of DF as the partition of a users codeword, whose the first
part is transmitted by the source user and the second by the
relay.
The main advantage of AF over DF is that no hard decisions are made, but on the contrary AF does not regenerate
the signal.

soft-DF relay keeps the soft information onthe information


bits, i.e. the a posteriori probabilities P u(k)|rSR and
produces
 a sequence of a posteriori probabili from these
ties P x(n)|rSR on the coded bits x(n). For the sake of
notational simplicity in this section, we assume that the encoder produces only one parity bit per information bit (i.e.
N = K). Generalization to other rates is straightforward.
For a convolutional code, the representation of this soft
encoding operation within the framework of factor graphs
and the application of the sum-product algorithm lead to an
encoding algorithm quite similar to that of the well-known
MAP decoding algorithm. More precisely, the soft encoder
aims to compute



P (x|rSR )
P x(k)|rSR =

k = 1, . . . , K, (2)

x(k)


where x(k) denotes the summation on all variables but
x(k). A convenient factorization of P (x|rSR ) must be found
to enable an efficient computation of the marginals P (x(k)|rSR )
thanks to the sum-product algorithm. Since x does not depend on rSR once u is assumed, rSR u x forms a
Markov chain and we have :



P (x|u) P (u|rSR ),
(3)
P x(k)|rSR =
x(k)

where P (x|u) equals 1 if the codeword x corresponds to

the information bits sequence u and 0 otherwise. For a convolutional code, this factor expands in the same way as for
a SISO decoder. In order to obtain the second factor of (3)
and to factorize it further we need to make the assumption
that the u(k) are independent given rSR :

4. NEW COOPERATIVE TECHNIQUE : SOFT-DF

The new Soft-DF technique presented here keeps the advantages of both previous techniques, i.e. it regenerates the
signal while keeping soft information.
A relay using this technique first decodes its partners
signal, keeping soft information, interleaves and re-encodes
it with a soft-input soft-output (SISO) encoder. Finally the
relay deducts the signal to transmit thanks to an equivalent
channel assumption on the outputs of the encoder. This
equivalent channel assumption results in interpreting the relay as an analog channel signal-to-noise ratio (SNR) enhancer. In this sense, the soft nature of the output signal
enables to convey information about its reliability.
We successively describe below the new SISO encoder,
the equivalent channel modeling its output and finally the
receiver at destination.

P (u|rSR ) =

K




P u(k)|rSR ,

(4)

k=1

where the P (u(k)|rSR ) are computed by the SISO decoder.

Simulations suggest that this approximation does not have
any noticeable impact.
For a convolutional code, the application of the sumproduct algorithm to the factor graph described by (3)-(4)
gives slightly modified versions of the standard forward recursion () and reverse recursion () of the MAP decoding algorithm. For an encoder of state s(k) at time k, we
can write the following recursions where each term in the
summations has to correspond to a combination of (s(k
1), u(k), x(k), s(k)) satisfying the code constraints:

4.1. Soft-Input Soft-Output Encoder

The proposed SISO encoding algorithm has some similarities with a SISO decoding algorithm.
Given a sequence of noisy coded symbols rSR , a DF
relay would compute estimates of the K information bits
u = [u(1), . . . , u(K)]T and from these produce a sequence
of N coded bits x = [x(1), . . . , x(N )]T . In contrast, the



s(k) =



s(k 1) =





s(k 1) P u(k)|rSR ,

s(k)

s(k1)

158





s(k) P u(k)|rSR .

We finally compute
 
 
 



s(k 1) s(k) P u(k)|rSR .
P x(k)|rSR =
x(k)

4.2. Equivalent Channel Model

We need
the signal lying behind the probabil
 to re-transmit
ities P x(k)|rSR . Since we lack any analytical model for
the output of the SISO encoding/decoding process, we deduce this signal from the log likelihood ratios (LLRs) under
the assumption that they have the same form as at the output
of an equivalent AWGN channel with BPSK modulation :




P (x(n) = 0|rSR )
= Leq bR (n)+neq (n) .
LLR x(n)
log
P (x(n) = 1|rSR )
(5)
The amplitude factor Leq reflects the signal reliability, bR (n) =
2x(n) 1 is the symbol corresponding to the bit x(n) and
neq (n) is a zero-mean white gaussian noise sample of vari2
2
. Both amplitude Leq and variance eq
of the noise
ance eq
depend on the signal-to-noise ratio inside a block on the
inter-user channel and therefore they change at each frame.
Since no analytical characterization exists their relation with
the noise variance on the inter-user channel has to be computed empirically.
From this interpretation of the LLRs, the relay constructs
and forwards to the destination the signal R (bR (n)+neq (n))
which contains the payload plus an additive gaussian noise.
The factor R is constant over consecutive frames and is
computed in order to normalize the transmitted energy averaged over the realizations of the inter-user channel :

Es,R
,
(6)
R =
2 ]
E[b2 + eq

Normalization
at the relay

hRD

rRD (n) = hRD (n)R (bR (n) + neq (n)) + nD (n),

(7)

which is simply under the previous assumptions the output of a cascade composition of two channels corrupted by
gaussian noise. Of course the receiver needs to estimate the
parameters of the compound channel only and not the parameters of both channels separately, thus no channel state
information for the individual channels is required.
The decoding process of this signal jointly with that
from the direct path is achieved by a standard turbo decoding operation.

Eb/Neq (output SNR) [dB]

Equivalent
inter-user channel

nD
R

The destination receives the following signal from the relay :

Relay to destination
channel

15
10
5
0
5
5

Fig. 1. Equivalent userrelaydestination channel for soft

decode and forward.
This model suggests to interpret the transmission of coded
data on the inter-user channel followed by the soft decoding/reencoding process as the transmission of coded data on a
global super-channel with better SNR. This interpretation
is depicted in figure 1.
The SNR enhancement is quantified in figure 2. This
figure presents in its upper part the empirical input-output

159

Probability density

4.3. Reception and decoding

20

where b2 is the variance of the symbols b(n).

neq

relation of the SISO decoder/re-encoder in terms of SNR,

for the simulation parameters of section 5 and under the
previous assumption of a gaussian i.i.d. noise corrupting
the output. For a coded signal received with a SNR of 5 dB
for example, the relay produces an output signal with a SNR
of about 12 dB.
The limitations of the i.i.d. gaussian noise assumption
appear distinctly below 0 dB and above 15 dB where the
output SNR is lower than the input SNR. However for common channel parameters, input SNRs lie mainly in the range
of interest, as shown in the lower part of figure 2 by the
probability density function of the input SNR for a Rayleigh
fading channel with mean SNR of 8 dB.

10

15

20

15

20

0.2
0.1
0
5

5
10
E /N (input SNR) [dB]
b

Fig. 2. Input-output relation of the decoding/re-encoding

process, in terms of instantaneous SNRs (i.e. for a particular
realization of the fading).

perfect information at the relay, and the performance of no

cooperation (turbo-code of rate 1/3 with the same constituent
codes as for DF, and an equal total energy expense).
These results reflect the advantages of the principles underlying soft-DF : in comparison to DF, soft-DF gains from
an increased cooperation level and compared to AF, it wastes
less power to transmit noise.

10

BER

10

10

6. CONCLUSION
3

10

10

No cooperation
AF
DF
SoftDF
Perfect cooperation
4

8
12
Uplink E /N [dB]
b

16

The Soft-DF cooperation technique presented in this paper

achieves performance reflecting its advantages over AF and
DF. The use of exclusively soft information at the relay as
well for decoding as for re-encoding is a key point for this
result. This soft process is enabled thanks to the design of a
SISO encoder.
Dedicated strategies and protocols can be designed to
further improve its performances, in the same way as for
other cooperative techniques. These strategies can take advantage of the soft nature of this new cooperation technique.

20

Fig. 3. BER curves for AF, DF and Soft-DF.

5. PERFORMANCE COMPARISON

7. REFERENCES

Simulation results confirm the advantages of the Soft-DF

relay technique through the following analysis of bit error
rate (BER) curves. Figure 3 compares the performance of
AF, DF and Soft-DF with the following parameters. Both
users encode their own data blocks of length K = 128 bits
with a recursive systematic convolutional (RSC) encoder
of rate 1/2, constraint length 4 and generator polynomials
(138 , 178 ). The partner relays these data with a cooperation rate 1/3. In order to achieve such a cooperation rate,
the AF relay forwards the parity bits only, the DF relay reencodes the decoded bits with the same RSC code and forwards the computed parity bits, and the Soft-DF relay forwards the K parity bits computed using the non-recursive
convolutional soft-encoder of constraint length 2 and generator polynomial 38 . To decide whether to cooperate or
not, the DF relay checks if a frame is error-free thanks to
a 16-bit CRC code with generator polynomial given by the
coefficients 0x1021 (hexadecimal notation). The bandwidth
overhead required to transmit the 16 CRC bits is neglected.
The figure compares the different schemes for equal overall
energy consumption per information bit. The BER is computed for equal SNRs on the uplinks (abscissa in figure 3)
and a SNR fixed to 10 dB on the inter-user channel. Denoting Eb,S the energy spent by the source per information
bit, the relay spends an additional average energy Eb,S /2 to
cooperate. Each point is obtained after the observation of
1000 frame errors. The turbo-decoders of DF and Soft-DF
perform three iterations, which is enough to achieve convergence.
To achieve a BER of 103 Soft-DF requires about 2.5 dB
less power than AF and 2 dB less than DF. For the sake of
comparison, the figure shows the performance of DF with

[1] A. Sendonaris, E. Erkip, and B. Aazhang, User Cooperation DiversityPart I: System Description, IEEE
Trans. Commun., vol. 51, no. 11, Nov. 2003.
[2] , User Cooperation DiversityPart II: Implementation Aspects and Performance Analysis, IEEE Trans.
Commun., vol. 51, no. 11, Nov. 2003.
[3] J. N. Laneman and G. W. Wornell, Energy-efficient antenna sharing and relaying for wireless networks, in
Proc. IEEE WCNC, Chicago, IL, 2000, pp. 7 12.
[4] J. N. Laneman, G. W. Wornell, and D. N. C. Tse, An
efficient protocol for realizing cooperative diversity in
wireless networks, in Proc. IEEE ISIT, Washington,
DC, 2001, p. 294.
[5] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, Cooperative diversity in wireless networks: Efficient protocols and outage behavior, IEEE Trans. Inform. Theory, vol. 50, no. 12, pp. 30623080, Dec. 2004.
[6] T. Hunter and A. Nosratinia, Diversity through coded
cooperation, IEEE Trans. Wireless Commun., submitted for publication.
[7] M. Janani, A. Hedayat, T. Hunter, and A. Nosratinia, Coded cooperation in wireless communications:
Space-time transmission and iterative decoding, IEEE
Trans. Signal Processing, vol. 52, no. 2, pp. 362 371,
Feb. 2004.

160