Professional Documents
Culture Documents
net/publication/292672531
CITATIONS READS
4 28
1 author:
SEE PROFILE
All content following this page was uploaded by Shama Naz Islam on 18 October 2017.
Research Article
ISSN 1751-8628
Achievable rate and error performance of an Received on 3rd October 2014
Revised on 17th August 2015
amplify and forward multi-way relay network Accepted on 8th October 2015
doi: 10.1049/iet-com.2015.0484
in the presence of imperfect channel www.ietdl.org
estimation
Shama Naz Islam ✉
Research School of Engineering, College of Engineering and Computer Science, The Australian National University, Canberra 2601, ACT,
Australia
✉ E-mail: shama.islam@anu.edu.au
Abstract: In this paper, we investigate the impact of channel estimation error on the achievable common rate and error
performance of amplify and forward (AF) multi-way relay networks (MWRNs). Assuming lattice codes with large
dimensions, we provide the analytical expressions for the end-to-end SNR at the users and obtain upper bounds on
the achievable common rate for an AF MWRN. Moreover, considering binary phase shift keying (BPSK) modulation as
the simplest case of lattice codes, we obtain the average bit error rate (BER) for a user in an AF MWRN. The analysis
shows that the average BER is a linearly increasing function and the achievable common rate is a linearly decreasing
function of the channel estimation error. On the other hand, the average BER decreases and the achievable common
rate increases with increasing correlation between the true and the estimated channel. Also, we observe that the AF
protocol is robust against increasing number of users in terms of error performance. We show that when the decoding
user has better channel conditions compared to other users, AF relaying gives a better error performance and common
rate. Finally, simulation results are provided to verify the validity of our analysis.
1 Introduction signals, and then amplifies and forwards the received signal to the
users. It has been shown in [12, 13, 19] that for complete
Recently, multi-way relay networks (MWRNs) have attracted cancellation of self-interference from the received signal so that
significant research interest due to their improved capacity and other users’ messages can be extracted, perfect channel state
spectral efficiency benefits [1–3]. MWRNs are characterised by information (CSI) at both the users and the relay are required.
multiple users and a single relay, where the users can exchange However, practical system models for AF MWRNs need to
information with the help of the relay terminal. MWRNs have consider imperfect CSI.
interesting potential applications in teleconferencing, sensor Recently, the impact of imperfect channel estimation on AF
networks or satellite communication networks [2]. MWRNs allow TWRNs has been quantified in terms of error performance with
the benefits of network coding to be realised for multiple users relay selection [20, 21], optimum power allocation [19] and design
through generalising the concept of two-way relay networks of suitable channel estimation algorithms [22, 23]. For MWRNs,
(TWRNs) [4–8] for multi-user networks. Another candidate the impact of imperfect CSI on the error performance of FDF
solution for realising the benefits of TWRNs in a multi-user relaying protocol has been investigated in [24]. However, the
scenario is a multi-user TWRN, where multiple user pairs impact of imperfect channel estimation on AF MWRNs has not
exchange messages within the pairs [9–11]. In this paper, we limit been addressed in the prior works on MWRNs.
our study to the performance analysis of an MWRN, as this is a To improve the achievable data rate of an MWRN, structured
more generalised version of a multi-user TWRN. codes such as lattice codes have been proposed recently for FDF
Prior research on MWRNs has focused on the performance relaying [14, 25]. Lattice codes can achieve the capacity of
analysis of different relaying protocols such as amplify and additive white Gaussian noise (AWGN) channel [16, 26] and
forward (AF) [12, 13], functional decode and forward (FDF) [3, allow higher data rates compared with uncoded transmission.
14] and compute and forward [15, 16]. However, in this paper, we Though the achievable rate of pairwise FDF MWRNs with lattice
consider only AF MWRNs for their simpler implementation and codes has been investigated in [14] for perfect CSI, the rate
lower signal processing complexity at the relay. Recent studies on analysis for lattice-coded AF MWRNs has not been performed yet
AF MWRNs investigate error performance [13, 17], as well as in the literature.
outage probability and sum rate [13, 18]. For complete message In an AF MWRN with pairwise transmission, if a user incorrectly
exchange among all users, the users in an AF MWRN can decodes another user’s message, then the mean of the received signal
transmit either in a pairwise or a non-pairwise manner. It was of the following user is shifted, leading to incorrect detection in the
shown in [13] that a non-pairwise transmission strategy can offer subsequent decoding operations. This problem, termed as error
larger spectral efficiency at the cost of additional signal processing propagation, has been studied for FDF and AF MWRNs in [17]
complexity at the relay. However, pairwise transmission offers for AWGN and fading channels with perfect CSI and it has been
lower complexity and simpler implementation which motivate us shown that the error performance degrades significantly due to
to investigate pairwise transmission strategy in this paper. error propagation. Moreover, when there is imperfect CSI, the
To improve the energy efficiency of a pairwise transmission-based additional estimation error worsens the error performance by a
MWRN, network coding schemes such as physical layer network larger degree. To the best of our knowledge, the impact of channel
coding (PNC) have been proposed in the literature. In AF estimation error on the error performance of AF MWRNs with
MWRNs with PNC, at first, the relay receives the sum of the lattice codes has not been addressed yet in the literature.
Fig. 1 System model for an L-user MWRN, where the users exchange We assume that Xp = 1 and nr is a zero-mean complex valued AWGN
information with each other via the relay R at the relay with variance s2nr = (N0 /2) per dimension, where N0 is
(10)
2.2 Broadcast phase
where
2.2.1 Pilot transmission: Prior to data transmission in the
broadcast phase, the relay broadcasts its own pilot and the
estimated channel coefficients from the multiple access phase. n′ = Prd ĥr, m n + (bm − 1) Prd ĥr, m Xr
Then the m th (m ∈ [1, L]) user performs MMSE estimation similar
to (3), to obtain the estimate ĥr, m . The channel estimation error in
+ bm nm + bm Prd h̃r, m Xr
this case is h̃r, m = hr, m − ĥr, m , with variance s2h̃ = s2h̃ . Similar
r, m m, r
to (5), the correlation between hr, m and ĥr, m is represented as
ρr,m = ρm,r. and βm is chosen to minimise the noise variance [25].
The users then utilise a lattice quantiser [25] and detect the
received signal to obtain the estimate X̂i, i+1 = (c(Wi ) +
2.2.2 Data transmission: At the end of pilot transmission, the c(Wi+1 )) mod L, where it is assumed that the lattice dimension is
factor α
relay amplifies the received signal with an amplification large enough such that Pr (n′ Ó V) 0 and V denotes the voronoi
[16] and removes the dithers di, di+1 scaled by Psd ĥi, r and region of the lattice Λ. After performing all the decoding
c(Ŵ L ) = (X̂ L−1, L − c(Ŵ L−1 )) mod L. Prd Psd r2r, m s2hr, m r2i, r s2hi, r
+
Psd (s2h̃ + s2h̃ ) + N0
i, r i+1, r
Then the upward message extraction is performed to recover the
messages of the (i − 1)th user to the first user. It can be clearly Prd Psd r2r, m s2hr, m r2i, r s2hi, r
+ (15)
identified from (24) that if Ŵi+1 = Wi+1 , then Wi+2 will be Prd s2h̃ + N0
r, m
incorrectly decoded. Thus, an incorrect decision about any user
can lead to incorrect decision about the preceding or the following
users which may cause error propagation in the sequential
3.1 Special case: perfect channel estimation
decoding process.
When the channel estimation is perfect, the SNR can be obtained by
Remark 1: It is clear from (8) and (10) that the error performance of replacing the channel estimates with the actual channel coefficients,
an AF MWRN with imperfect CSI depends on the channel i.e. setting the estimation error variances to zero and correlation
estimation error. The expressions of the channel estimates [see (3)] coefficients to one. Thus, if we set ρa, b = 1 and s2h̃ = 0 in (15),
and estimation error [see (4)] show that these are functions of the where a, b ∈ {r, i, i ± 1}, the modified SNR gpe
a, b
i, m is given by
noise variance and the channel variance. The channel variance and
the noise variance depend on the distance and the SNR,
respectively. Thus, we can expect the distance of users and the Prd s2hr, m s2hi, r Prd Psd s2hr, m s2hi, r Prd Psd s2hr, m s2hi, r
gpe
i, m = + + . (16)
SNR to play a key role in determining the error performance of s2hi, r + s2hi+1, r N0 N0
AF MWRNs. In a practical wireless system, tuning the channel
gain of the users through physical parameters can be quite useful
for the best relay selection in the presence of multiple relays [20,
38] or optimal user ordering [39] for a pairwise transmission scheme. 4 Common rate analysis
where the numerator represents the power of the signal terms in (10) where the factor (L − 1) comes from the fact that there are (L − 1)
and the denominator time slots in each of the multiple access phase (MAC) and the
broadcast phase (BC) phases and Ri,m denotes the achievable
information rate with which the i th user’s message is received at
N ′ = Prd r2r, m s2hr, m |a − 1|2 Psd (r2i, r s2hi, r + r2i+1, r s2hi+1, r )
the m th user, as upper bounded in the following theorem.
+ |a|2 N0 + |a|2 Psd (s2h̃ + s2h̃ ) Theorem 1: The maximum possible information rate from the i th user
i, r i+1, r
to the m th user can be upper bounded by
+ |bm |2 N0 + Prd |bm − 1|2 r2r, m s2hr, m + |bm |2 Prd s2h̃
r, m
⎛
d 2 2 2 2
′ 1 ⎝ Pr rr, m shr, m ri, r shi, r Prd Psd r2r, m s2hr, m r2i, r s2hi, r
denotes the variance of the noise terms n in (10). R = log 2 2 +
The optimum value of α can be obtained by differentiating N′ with 2 ri, r shi, r + r2i+1, r s2hi+1, r Psd (s2h̃ + s2h̃ ) + N0
respect to α and then setting it to zero (i.e. (dN′ /dα) = 0) and can be
i, r i+1, r
⎞
d d 2 2 2 2
given as Pr Ps rr, m shr, m ri, r shi, r
+ ⎠ (18)
Prd s2h̃ + N0
Prd r2i, r s2hi, r + Prd r2i+1, r s2hi+1, r
r, m
a= (13)
Prd s2hi, r + Prd s2hi+1, r + N0
Proof: see Appendix 1. □
5.1 Data transmission with BPSK modulation Finally, the users perform maximum-likelihood detection on the
th
signals in (24) and (25) to extract other users’ messages.
In the BPSK modulated AF MWRN, at the ith time slot, the i and To obtain the error performance of AF MWRN with channel
the (i + 1)th users transmit their messages Wi and Wi+1 which are estimation error, we need to follow the general steps in [17] which
BPSK modulated to Xi and Xi+1, respectively, where Xi, Xi+1 ∈ are outlined below:
[−1, 1]. The relay receives the signal as
† Obtain the probability of incorrectly decoding a network-coded
ri, i+1 = hi, r Psd Xi + hi+1, r Psd Xi+1 + nr , (19) message.
† Investigate different error cases for the k th error event, where the
k th error event can be denoted as the occurrence when exactly k
and amplifies and broadcasts the network-coded signal, which is number of users’ messages are incorrectly decoded.
given as † Obtain the probability of the above error cases.
† Add the probabilities of different error cases to obtain the
probability of the k th error event Pi (k) and express it in terms of
Zi, i+1 = a hi, r Psd Xi + hi+1, r Psd Xi+1 + nr , (20)
the probability of incorrectly decoding a network-coded message
(from step 1).
The i th user receives † Since, there are L − 1 possible error events in an L-user MWRN,
find the expected probability of all these error events to obtain the
average BER.
Yi, i+1 = Prd hr, i Zi, i+1 + nm , (21)
Now, we obtain the probability of incorrectly decoding a user’s
and subtracts its own signal Xi multiplied by aĥr, i Psd Prd ĥi, r to message given that the previous user’s message is correct and the
th
detect the (i + 1) user’s signal as probability of incorrectly decoding a user’s message given that the
previous user’s message is also incorrect. These error probabilities
can be derived from the SNR expressions obtained in Section 3.
X̂i+1 = Yi, i+1 − aĥr, i Psd Prd ĥi, r Xi
Lemma 1: The probability that the ith user incorrectly decodes the
= aĥr, i ĥi+1, r Psd Prd Xi+1 + ni+1 (22) p th( p ≠ i) user’s message, given that the ( p ± 1)th user’s message is
correctly decoded, can be obtained as
where ni+1 denotes the noise terms, given as
PAF (i, p) = Q 2gi, p , (26)
ni+1 = 2a Psd Prd ĥr, i h̃i, r Xi + Psd Prd ah̃r, i h̃i, r Xi
where gi, p is the SNR of the p th user’s signal received at the i th user
+ a Psd Prd ĥr, i h̃i+1, r Xi+1 and can be given as
+ a Psd Prd h̃r, i ĥi+1, r Xi+1 + a Psd Prd h̃r, i h̃i+1, r Xi+1 Psd Prd r2r, i s2hr, i r2p, r s2h p, r
gi, p = . (27)
+ a Prd ĥr, i n1 + a Prd h̃r, i n1 + n2 . (23) Np
Then the user detects the signal of the (i + 2)th user to the L th user in where Np denotes the variance of the noise terms in the p th user’s
the downward [In the downward message extraction process, the signal, received at the i th user and can be given by
user extracts other users’ messages who transmitted after it.]
message extraction. The process can be shown as
Proof: see Appendix 2. □
X̂ i+2 = Yi+1, i+2 − aĥr, i Psd Prd ĥi+1, r X̂ i+1 , . . . , Lemma 2: The probability that the i th user incorrectly decodes the
p th( p ≠ i, i ± 1) user’s message, given that the ( p ± 1)th user’s
X̂ L = YL−1, L − aĥr, i Psd Prd ĥL−1, r X̂ L−1 . (24) message is incorrectly decoded, can be obtained as
Similarly, the upward [In the upward message extraction process, the PAF (i, p) = Q 2gei, p , (28)
user extracts other users’ messages who transmitted before it.]
⎧
⎪
⎪ 2Psd Prd r2r, i s2hr, i s2h̃ + Psd Prd r2r, i s2hr, i s2h̃ + Psd Prd r2i+1, r s2hi+1, r s2h̃ + (Psd + Prd )
⎪
⎪ p=i+1
⎪
i, r i+1, r r, i
⎪
⎨ ×r2r, i s2h N0 + Psd r2i+1, r s2h N
i+1, r 0
Np =
r, i
⎪
⎪ Psd Prd r2r, i s2hr, i s2h̃ + Psd Prd r2p+1, r s2h p+1, r s2h̃ + Psd Prd r2r, i s2hr, i s2h̃
⎪
⎪
⎪
⎪
p+1, r r, i p, r
p = i, i + 1
⎩ + Psd Prd r2p, r s2h s2 + Prd r2r, i s2h N0 + Psd r2p+1, r s2h N + P d 2 2
s r p, r sh p, r N0
p, r h̃
r, i r, i p+1, r 0
Proof: see Appendix 4. □ Fig. 2 Achievable common rate for different channel conditions in L = 6
and 10 user AF MWRNs
In this section, we provide insights from the achievable rate and error Fig. 2 shows the achievable common rate with lattice codes for L = 6
performance analysis. We also verify the error performance results and 10 user AF MWRNs for the cases (i) di < dm, (ii) di = dm and (iii)
with Monte Carlo simulations. We consider two cases L = 6 and di > dm in the presence of imperfect channel estimation. The above
10 users where each user transmits a message packet of T = 1000 cases represent that the i th user has (i) better, (ii) same and (iii)
bits. For each of the message packets, one pilot bit is transmitted. poorer channel conditions, respectively, compared with the m th
We assume equal and unity power for the data signal and the pilot users. As defined in (4), the estimation error varies linearly with
signal at the users and the relay. Note that when the data and the SNR. From Fig. 2, it is clear that for larger number of users the
pilot powers are optimised, one pilot symbol transmission is common rate will be lower, which can be identified from (17). It
⎧
⎪
⎪
L−1 k−1
L−k+1
⎪
⎪ k P′ AF (i, p + t)PAF (i, p)
⎪
⎪
⎪
⎪ k=1 p=2 t=1
i = 1, 2, 3
⎪
⎪
⎪
⎪
k−2
⎪
⎪ + ′ ′
P AF (i, L)P AF (i, L − t)PAF (i, L − k + 1)
⎪
⎪
⎪
⎪
t=1
⎪
⎪
⎪
⎪
L−1 k−1
L−1 ′
⎪
⎪ P AF (i, p − t)PAF (i, p)
⎪
⎪
k
⎪
⎪ k=1 p=k t=1
⎪
⎪ i = L, L − 1, L − 2
⎪
⎪
⎪
⎪
k−2
′ ′
⎨ + P AF (i, 1)P AF (i, 1 + t)PAF (i, k)
1
Pi, avg, AF =
t=1 (30)
L − 1⎪
⎪ L−k+1 k−1 ′
⎪
⎪
L−1
⎪
⎪ k P AF (i, p + t)PAF (i, p)
⎪
⎪
⎪
⎪
k=1 p=i+1 t=1
⎪
⎪
⎪
⎪
k−2
⎪
⎪
⎪ + P′ AF (i, 1)P′ AF (i, 1 + t)PAF (i, k)
⎪
⎪ i Ó {1, 2, 3, L − 2, L − 1, L}.
⎪
⎪
t=1
⎪
⎪ k−1
i−1 ′
⎪
⎪ + P AF (i, p − t)PAF (i, p)
⎪
⎪
⎪
⎪ p=k t=1
⎪
⎪
⎪
⎪
k−2
⎪
⎩ + P′ AF (i, L)P′ AF (i, L − t)PAF (i, L − k + 1)
t=1
can be noted that when the decoding user has better channel Fig. 4 shows the achievable common rate for L = 10 user AF
conditions, the achievable common rate is higher than that when MWRN for the cases when (i) 10% of the users have distances
the decoding user has poorer channel conditions. This is because, below 0.1d0 (corresponds to the case when most of the users have
the SNR received at the decoding user increases when the poor channel conditions) and (ii) 90% of the users have distances
decoding user has a good channel gain [see (15)], and as a result below 0.1d0 (corresponds to the case when most of the users have
the achievable rate improves [see (18)]. good channel conditions). It can be observed that when most of
Figs. 3a and 3b show the impact of different levels of channel the users experience good channel conditions, the achievable
estimation errors and the correlation coefficient, respectively, on common rate improves. Moreover, it can be noted that the
the achievable common rate for L = 6 and 10 user MWRNs. To degradation in the overall channel conditions leads to performance
plot the results in Fig. 3a, the estimation errors are set using the loss in AF MWRN by nearly the same degree for the cases when
following technique, which is illustrated for two estimation errors perfect CSI is available or not.
of 0.1 and 0.01% of the combined variance of the fading channel
and the complex AWGN noise. These values of channel
estimation errors have been introduced by setting s2nr equal to 6.2 Average BER
0.001 and 0.0001, respectively, in (4) and noting that
s2hi, r ≃ s2hi, r + s2nr at high SNR and n = 3. To plot the results in Fig. 5 shows the average BER for AF MWRN with L = 6 and 10
Fig. 3b, we consider the worst channel conditions for every user users for the cases (i) di < dm, (ii) di = dm and (iii) di > dm in the
(i.e. di = d0) since the impact of the correlation is more visible presence of imperfect channel estimation. Here, the analytical
when users have smaller channel gains [see (5)]. Using (5), we set results are plotted using (30) and compared with the simulation
the noise variance s2nr = (s2hi, r /r2i, r ) − s2hi, r and substitute it in results. It can be seen from this figure that the analytical results
(18) to calculate the achievable common rate for different match with the simulations at medium to high SNR. It can be seen
values of correlation coefficients. Fig. 3a shows that the from this figure that the average BER remains almost the same
achievable rate decreases with increasing estimation error. For with the increasing number of users. This is because larger
different channel condition cases and a certain number of users, number of error events in AF MWRN are less probable, as
the achievable common rate decreases with the same proportion. explained in Appendix 4. For this reason, the average BER does
On the other hand, Fig. 3b shows that the achievable common rate not increase with increasing number of users which is expected
increases with the correlation coefficient at the same rate for from (30). Moreover, it can be seen that when the decoding user
different number of users. has better channel conditions compared with other users, the error
Fig. 4 Achievable common rate when 10 and 90% users’ distances below Fig. 5 Average BER for different channel conditions in L = 6 and 10 user
0.1d0 in L = 10 user AF MWRN AF MWRNs
From (34), the SNR of the (i ± 1)th user’s signal, when received at the
i th user, is given by
9 Appendices
9.1 Appendix 1: Proof of Theorem 1 a2 Psd Prd r2r, i s2hr, i r2i+1, r s2hi+1, r
gi, i+1 = . (36)
Ni+1
The proof can be obtained following the steps in [16]. Since, the
lattice dimension D is large enough to ensure Pr (n′ Ó V) 0, the
volume-to-noise ratio of the lattice, μ > 2πe is satisfied. To ensure Here, Ni±1 represents the variance of the noise terms present in ni±1
Ni+1 = 2a2 Psd Prd r2r, i s2hr, i s2h̃ + a2 Psd Prd s2h̃ s2h̃ Psd Prd r2r, i s2hr, i r2p, r s2h p, r
i, r r, i i, r
gi, p = , (44)
+ a2 Psd Prd r2r, i s2hr, i s2h̃ Np
i+1, r
+ a2 Psd Prd r2i+1, r s2hi+1, r s2h̃ + a2 Psd Prd s2h̃ s2h̃ where
r, i r, i i+1, r
where ρa, b represents the correlation between the channel estimate + Psd Prd r2r, i s2hr, i s2h̃
ĥa, b and the true channels ha, b and s2h̃ represents the variance of p, r
a, b
the estimation error h̃a, b , a, b ∈ {r, i, i ± 1}. + Psd Prd r2p, r s2h p, r s2h̃ + Prd r2r, i s2hr, i N0
After substituting the value of α from (13), the expression of the
r, i
SNR in (15) can be given as + Psd r2p+1, r s2h p+1, r N0 + Psd r2p, r s2h p, r N0 .
Psd Prd r2r, i s2hr, i r2i+1, r s2hi+1, r The exact probability density function (pdf) of the noise ni±1 in (35)
gi, i+1 = ′ , (38) and np in (43) is not Gaussian due to the presence of product terms of
Ni+1
two Gaussian variables [i.e. the first six terms in (35) and the first
seven terms in (43)]. However, the pdf of the noise can be
where numerically shown to match closely to that of a Gaussian
distribution at high transmit SNR [The results are not provided for
′
Ni+1 = 2Psd Prd r2r, i s2hr, i s2h̃ + Psd Prd r2r, i s2hr, i s2h̃ brevity.]. Thus, the probability of incorrectly decoding the p th
i, r i+1, r
user’s message, given that the ( p ± 1)th user’s message is correctly
+ Psd Prd r2i+1, r s2hi+1, r s2h̃ + (Psd + Prd )r2r, i s2hr, i N0 decoded, can be obtained from (26) under Gaussian noise
r, i
approximation.
+ Psd r2i+1, r s2hi+1, r N0 + N02 + Psd s2h̃ N0 + Psd s2h̃ N0
i, r i+1, r
9.3 Appendix 3: Proof of Lemma 2
+ Psd Prd s2h̃ s2h̃ + Psd Prd s2h̃ s2h̃ .
r, i i, r r, i i+1, r
(39) When X̂ p+1 = X p+1 , the noise terms in the p th user’s signal can be
written as
Since, the higher-order noise terms have a negligible contribution to
the SNR, we can ignore the second, third and fourth-order noise nep = 2a Psd Prd ĥr, i ĥ p+1, r + a Psd Prd ĥr, i h̃ p+1, r X p+1
terms [i.e. the last five terms in (39)], and thus the expression of
the SNR can be approximated as + a Psd Prd h̃r, i ĥ p+1, r X p+1
Psd Prd r2r, i s2hr, i r2i+1, r s2hi+1, r + a Psd Prd h̃r, i h̃ p+1, r X p+1 + a Psd Prd ĥr, i h̃ p, r Xp
gi, i+1 ≃ , (40)
N ′′ i+1
+ a Psd Prd h̃r, i ĥ p, r Xp
where
+ a Psd Prd h̃r, i h̃ p, r Xp + a Prd ĥr, i n1 + a Prd h̃r, i n1 + n2 .
′′
Ni+1 = 2Psd Prd r2r, i s2hr, i s2h̃ + Psd Prd r2r, i s2hr, i s2h̃ (45)
i, r i+1, r
For p ≠ i, i ± 1, from (24) and (25), the signal of the p th user can be 9.4 Appendix 4: Proof of Theorem 2
written as
First, we need to investigate different error cases for the k th error
event in an AF MWRN. For k = 1, the possible error cases are
X̂ p = Y p+1, p − aPsd Prd ĥr, i ĥ p+1, r X̂ p+1
illustrated in [17] as:
= aPsd Prd ĥr, i ĥ p, r Xp + np , (42)
† when a middle user’s message is wrongly estimated with correct
where p ± 1 denotes the user whose signal is detected before (or decision about the following user.
after) the p th user in the downward (or upward) extraction process † when an error occurs in the estimated signal of one of the end
and np denotes the noise terms present at the extracted signal of users.
the p th user’s signal and is given as
For larger values of k, there will be many more error cases and
considering all the possible error cases would make the analysis
np = aPsd Prd ĥr, i ĥ p+1, r (X p+1 − X̂ p+1 )
complicated. For a tractable analysis, we consider only the
+ aPsd Prd ĥr, i h̃ p+1, r X p+1 + Psd Prd ah̃r, i ĥ p+1, r X p+1 dominating error cases that influence the k th error event at high
′2
SNR. That is, we consider the higher-order error terms (e.g. PAF
+ aPsd Prd h̃r, i h̃ p+1, r X p+1 + aPsd Prd ĥr, i h̃ p, r Xp 2
and PAF ) and the corresponding error cases negligible.
+ aPsd Prd h̃r, i ĥ p, r Xp + aPsd Prd h̃r, i h̃ p, r Xp At high SNR, the dominating cases for the k th error event are
either k consecutive errors in the middle users or k consecutive
+ aPrd ĥr, i n1 + aPrd h̃r, i n1 + n2 . (43) errors involving one end user and k − 1 following (or preceding)
users. For example, if k = 2, i = 5 and L = 10, the error cases would
If X̂ p+1 = X p+1 (i.e. no error propagation), the SNR of the p th user’s be either two consecutive errors in the middle users [i.e. the fifth
(46a)
k −2
′ ′
k −1 PD = PAF (i, 1)PAF (i, 1 + t)PAF (i, k), (47c)
′
PC ′ (p) = PAF (i, p − t)PAF (i, p) t=1
t=1
L−1
′
(1 − PAF (i, p − k))(1 − PAF (i, m)),
k −2
′ ′
m=1, m=i, p, p−t, p−k
PD′ = PAF (i, L)PAF (i, L − t)PAF (i, L − k + 1). (47d)
t=1
(46b)
k−2
′ ′
Now, there are L − k − 1 number of user combinations, where
PD = PAF (i, 1)PAF (i, 1 + t)PAF (i, k) exactly k number of middle users’ messages are incorrectly
t=1 decoded, and one combination where k errors occur involving the
L−1 end user. Then, adding the expressions (47a) and (47d) or (47b)
(1 − PAF (i, m)), (46c) and (47c) for the possible user combinations would give the
m=1, m=i, t+1, k probability of exactly k error events. Thus, the probability of
exactly k error events can be asymptotically approximated as
k −2
′ ′
PD ′ = PAF (i, L)PAF (i, L − t)PAF (i, L − k + 1)
t=1
⎧
⎪
⎪
L−k+1
L−1 ⎪
⎪ PC (p) + PD′ , i = 1, 2, 3
⎪
⎪
(1 − PAF (i, m)). (46d) ⎪
⎪ p=2
⎪
⎪
m=1, m=i, L−t, L−k+1 ⎪
⎪
L−1
⎪
⎪ PC′ (p) + PD , i = L, L − 1, L − 2
⎪
⎪
where the term C (C′ ) represents the case of k errors involving the ⎨ p=k
Pi (k) =
L−k+1
middle users for the downward (upward) extraction process and D ⎪
⎪ PC (p) + PD
(D′ ) represents the case of k errors involving the end user for i ≠ ⎪
⎪
⎪
⎪
1, 2 (i ≠ L, L − 1). At high SNR, the terms ⎪
⎪
p=i+1
⎪
⎪ i Ó {1, 2, 3, L − 2, L − 1, L}.
⎪
⎪
i−1
⎪
⎪ +
L−1 ⎪
⎪ PC′ (p) + PD′ ,
′ ⎩
(1 − PAF (i, p + k))(1 − PAF (i, m)), p=k
m=1, m=i, p, p+t, p+k (48)
L −1
′
(1 − PAF (i, p − k))(1 − PAF (i, m)),
Now, the average BER of an L-user AF MWRN can be obtained by
m=1, m=i, p, p−t, p−k
taking the expectation over all the L − 1 error events, as follows
L−1
(1 − PAF (i, m))
m=1, m=i, t+1, k 1 L−1
Pi, avg, AF = kP (k) (49)
L − 1 k=1 i
and
L −1 Substituting (48) and (47) in (49), the average BER of an AF
(1 − PAF (i, m)) MWRN can be asymptotically expressed as in (30), which
m=1, m=i, L−t, L−k+1 completes the proof.