[IEEE 2014 European Conference on Networks and Communications (EuCNC) - Bologna, Italy...
Transcript of [IEEE 2014 European Conference on Networks and Communications (EuCNC) - Bologna, Italy...
Physical-layer Network Coding via Low Density Lattice Codes
Yi Wang and Alister Burr
Abstract—We present a new PLNC scheme based on therecently developed compute-and-forward (C&F) paradigm andlow density lattice codes (LDLC). LDLC possesses high codinggain and good algebraic structure which is inherently suitablefor C&F. We also show that the ring-based constellation can beused to improve the average rate per dimension.
I. INTRODUCTION
Physical-layer network coding (PLNC) has been shown to
be very effective in improving the throughput of a two-way
relay channel (TWRC) by some research groups in 2006 (e.g.
Zhang et al. [1] and Popovski et al. [2]). The core idea is
that the intermediate relay attempts to infer and forward linear
combinations of the simultaneously received signals, instead of
decoding the transmitted signals individually. The destination
node is able to recover the original signals successfully pro-
vided that the number of such linear combinations is sufficient.
The remarkable potential of PLNC was further exploited by
Nazer and Gastpar [3] who proposed a new approach to PLNC,
namely compute-and-forward (C&F) which extends TWRC
to a more general network topology. In this novel scheme,
the transmitted signals at each source node are lattice points
in a multi-dimensional lattice over integers, and the relay
decodes an integer combination of these lattice points based
on the noisy observations, which is again a lattice point. Their
approach relies on Loeliger’s type A construction of random
lattice ensembles which requires both a large field size pand infinite sequences of lattices, making C&F essentially an
information-theoretic approach, and hence less straightforward
for practical implementation. In a recent work, Feng et al. [4]
formulated a general algebraic framework to generate a large
class of efficient and finite-dimensional lattice partitions. They
showed that the lattice partitions for C&F closely related to
the algebraic theory of finite modules over a principal ideal
domain (PID), and hence the resulting lattice structure allows
the coset leaders to be employed as C&F codewords inherently
and efficiently.
One possible solution for the C&F design problem is to
use existing linear channel codes to construct lattices based
on finite dimensional R-lattice partitions, where R is a PID
of C. For example, by employing complex construction A
over the ring of Eisenstein integers Z[ω], Narayanan et al. [5]
implement the C&F protocol using low density parity check
(LDPC) codes over Fp and Sun et al. [6] implement it using
convolutional, BCH and Reed-Solomon codes over GF(4).
Another solution for C&F design is to employ practical
high coding gain lattice codes. Constructions of lattices on
the basis of the classical low-dimensional lattices or channel
codes attracted research attention in the past, but there was
less progress toward designing lattice codes directly in the
Euclidean space. Recent works from Sommer et al. filled this
gap by developing practically decodable lattice codes (e.g.
The work described in this paper was supported by the European Commis-sion Framework 7 Programme under grant agreement 318177 (DIWINE).
signal codes [7]). These lattice codes have a special algebraic
structure and this inherent merit makes them easily recovered
from a linear system over the finite field in a multiple access
channel (MAC), and promotes their use as a suitable and
practical coding approach for PLNC under the C&F paradigm.
In [8] [9], the authors have shown that signal codes applied
in PLNC achieve a good network throughput as expected. In
comparison to the lattices constructed from channel codes,
LDLC incorporates channel coding and modulation in a single
entity, and also its codeword is inherently a low-density lattice
point, which greatly simplifies the design in the PLNC system.
An interesting alternative to signal codes is the low density
lattice codes (LDLC) [10], in which the lattice is generated by
the inverse of a sparse matrix. LDLC was shown to be capable
of achieving error-free decoding within 0.6dB of the Shannon
bound. Inspired by its high coding gain and the algebraic
property, we use a modified Gaussian mixture model for low-
complexity LDLC detection, and present a demonstrator for
the LDLC-based C&F scheme with performance comparison
for both non-fading and fading cases, which constitute the
main contributions of this paper. To the best of our knowledge,
LDLC-based PLNC has not been proposed before, partly
because efficient shaping and decoding operations were not
well understood [8]. Our work here shows the remarkable
potential of LDLC-based PLNC, and may motivate further
research toward the remaining unsolved problems.
A. Notation Definitions
Throughout this paper, we use R and Z to denote the fields
of real numbers and integer, respectively and Fq , q > 1, q ∈ Z
to denote the algebraic ring with size q (which is a finite field
when q = p) where p is a prime number. We also use boldface
lowercase and boldface uppercase to denote column vector and
matrices, respectively (e.g. h = [h1, · · · , hn]T ). We denote
h\j = [h1, · · · , hj−1, hj+1, hn] and⊙
for element-wise
multiplication. Fnq denotes an n-dimensional ring where the
ring size of the ith dimension i ∈ {1, 2, · · · , n} is determined
by qj , and similarly for Fnp. Functions g(·) : Fn
p → Zn and
gr(·) : Fnq → Zn are used to map between the finite field
and the corresponding subset of integers, {0, 1, · · · , p− 1}n,
and between the ring and the corresponding subset of integers,
{0, 1, · · · , q − 1}n, respectively.
II. LOW DENSITY LATTICE CODES
A. Lattice Definitions
Definition 1 (Lattice): An n-dimensional lattice Λ is defined
as a discrete subgroup of Euclidean space Rn such that if λ1,
λ2 ∈ Λ, we have a1λ1 + a2λ2 ∈ Λ for any a1, a2 ∈ Z. A
lattice can be written in the form of:
Λ = {λ = Gb : b ∈ Zn} (1)
where G ∈ Rn×n including basis of n linearly independent
vectors in Rn.
Definition 2 (Quantizer): A lattice quantizer, QΛ: Rn → Λ,
maps a point s in Rn to the nearest point in Λ in Euclidean
distance:
QΛ(s) = argminλ∈Λ
‖s− λ‖ (2)
Definition 3 (Voronoi Region): The Voronoi region of a
lattice point λ, denoted by V(λ), is the set of points in Rn
closest to this point, i.e.,
V(λ) = {s : s ∈ Rn,QΛ(s) = λ} (3)
The fundamental voronoi region is defined as V(0), or simply
represented as V . Let Vol(V) denote the volume of V and
Vol(V) = Vol(Λ).Definition 4 (Goodness of Lattices): Let z be an n-
dimensional i.i.d. Gaussian vector, z � N (0, σ2In×n), a
sequence of lattices Λ is Poltyrev-good if
Pr(z /∈ V) ≤ e−nEp(γ) (4)
where Ep(·) is the Poltyrev exponent and γ is the volumn-to-
noise ratio (VNR),
γ =Vol(Λ)
2n
2πeσ2. (5)
B. LDLC Encoding and Shaping
Following Definition 1, we define an n-dimensional LDLC
in terms of its non-singular lattice generator matrix G ∈ Rn×n
such that its parity check matrix H = G−1 is sparse. Let
b denote a sequence of integer message vectors with each
element bi ∈ {0, 1, · · · ,m−1}. The integer linear combination
of the independent columns of G forms the unshaped LDLC
codeword over Rn which can be seen as a subset of an
infinite lattice over Rn. An (n, d) LDLC can be constructed
by designing a sparse parity check matrix H having constant
row and column weight d. Let
h = [1, τ, τ, · · · , τ︸ ︷︷ ︸(d−1) τ in total
, 0, · · · , 0] (6)
be a row/column vector with d−1 τs (τ > 0) and τ is designed
to ensure full convergence [10]. H should be designed such
that each row and column is a permutation of h followed by a
random sign change, and of course H must be cycle-free.The
encoding operation x = Gb has a computational complexity
of O(n2) based on the designed H, but the iterative Jacobi
method can be used to obtain the codewords by solving a
sparse linear system H · x = b.
Compared with finite-alphabet codes, the LDLC codeword
x = Gb may have very large energy and hence an efficient
shaping scheme for LDLC is necessary to limit the average
power of x. Tomlinson-Harashima precoding scheme allows us
to restrict the LDLC codewords into a hypercube. Recall that
hypercube shaping for each dimension of the signal codes is
based on the monic causal filter coefficients and their generator
matrix has a Toeplitz structure, which is similar to lower-
triangular structure [7]. Thus, signal codes are naturally suit-
able for shaping and the constellation size of each dimension
remains the same. However, to perform hypercube shaping
[11] for LDLC, we must reconstruct the H matrix based
−20 −15 −10 −5 0 5 10 15 200
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
x
GM
R
Fig. 1. A Gaussian mixture (black dash-dot) comprised of N = 10 Gaussiancomponents (blue dashed lines) is approximated by another Gaussian mixture(red thick line) with Nmax = 3.
on h to have a lower triangular structure. Obviously, this
means that the H matrix no longer has constant degree-dnow. Hence the row degree should be 1 at the top row and
gradually increases until it reaches d (similarly for column
degree). Since codewords corresponding to the low-degree row
in H are less protected, the messages whose check equations
have weak protection capability should belong to a small size
constellation mi, i ∈ {1, 2, · · · , n}. The shaping operation
maps the message integer bi of the ith dimension into ai by
performing
ai = bi −miki (7)
where ki is an integer which should be computed such that
the shaped codewords |xs,i| ≤ mi
2 . since Hxs = a, we have
xs,i = bi −miki −i−1∑l=1
Hi,lxs,l (8)
Let ki = ��i�, and hence |�i| ≤ 12 + ki, then we have
|(�i − ki)mi| ≤ mi
2, mi > 0 (9)
Following the shaping condition, (�i − ki)mi = xs,i, and
combining (8) and (9), ki is obtained from,
ki = ��i� =⌊bi −
∑i−1l=1 Hi,lxs,l
mi
⌉(10)
Another method which works for hypercube and nested
lattice shaping is based on the QR decomposition [11] of an
arbitrary structure H, where H = JQ can be decomposed
such that J is a lower triangular matrix and Q is orthonormal.
In this case, J has a different form from the lower triangular
H matrix described above, and ki should be calculated as
ki = �ϑi� =⌊(
bi −∑i−1
l=1 Ji,lxs,l
)/mi
⌉, and xs,i is
xs,i =bi −miki −
∑i−1l=1 Ji,lxs,l
Ji,i(11)
The derivation of the approach is similar to (7)-(10), except
that |(ϑi − ki)mi/Ji,i| ≤ mi
2·Ji,i, and hence, we need to define
φ1
a1L1
x1
S1
b1
E1
φ2
a2L2
x2S2
E2
+
z
R
αy
β
g(·)
b2g(·)
L−1v
φ−1
QΛ(y)
g−1(·)
u =
⊕2
�=1
(q′
�� b�
)
E−1
D1
D2
Fig. 2. Encoder and Decoder for LDLC-based PLNC at MAC.
a nominal constellation size m such that mi = �Ji,im� for
each dimension of b, to guarantee that the codewords are
constrained in a hypercube. In the rest of the paper, we will
consider QR decomposition-based hypercube shaping.
C. Decoding
We assume the noisy LDLC codeword y = x + z where
z is i.i.d. Gaussian noise z ∼ N (0, σ2In×1) . A sequential
maximum likelihood (ML) detector b = argminb ‖y−Gb‖2
obviously has large computational costs. Hence an iterative
detection method employing the belief propagation algorithm
over the bipartite graph of H was proposed [10]. Unlike
LDPC, the metrics exchanged between the variable nodes
(VN) and check nodes (CN) are real continuous functions over
(∞,−∞) rather than scalar values. At the check node: the ith
CN calculates function ηj(ω) for the jth VN connected to
it, based on the functions μ\j(ω) transmitted from all VNs
(except the jth one) that are connected to this CN. ηj(ω)should be computed in terms of the check equations, where
all functions in μ\j(ω) needs to be expanded, convolved and
stretched, followed by a periodic extension with period 1
hi,j. At
the variable node: the ith VN calculates the joint probability
density function (PDF) for the jth CN connected to it, with
pj(ω) ∝1
√2πσ2
e−(yj−xj)
2
2σ2
(⊙η\j(ω)
)(12)
The final decision occurs at the variable node where the ith
VN estimates the intrinsic messages pfinal(x) using all CN
functions connected to it, and xi should correspond to the
peak value of xi = argmaxxipfinal(x).
This decoder has been implemented in [11] where a quanti-
zation method, using Fourier transforms to combine messages,
was used to tackle the problem of the continuous function.
However, the discrete Fourier transform (e.g. 1024 point quan-
tization) taken at CN needs large computational complexity
and storage. To consider a practically feasible decoder, the
authors in [12] [13] use Gaussian mixture distributions as
the messages exchanged between CN and VN. Thus the
convolution and multiplication operations at CN and VN can
be performed in the form of Gaussian mixtures and hence
the output is another Gaussian mixture. The core problem is
to limit the exponential growth in the number of Gaussians as
the iteration proceeds. The Gaussian mixture reduction (GMR)
algorithm approximates a mixture of N Gaussians by another
mixture of Nmax Gaussians, where Nmax < N . In [13], the
GMR algorithm compares the distance metric of all possible
pairs of single Gaussians in the input list, and replaces the
pair having the minimum metric by a single Gaussian using
the second moment matching method.
In this paper we use a modified GMR algorithm based
on [13]. The LDLC decoder used for C&F has greater com-
putational cost than a traditional point-to-point system since
the periodic extension occurs over possible integers Z which
has a larger cardinality |Z|. Our modified GMR algorithm
further reduces the computational complexities of the metric
comparisons, and makes our LDLC decoder more suitable for
the demanding work at the intermediate relay using PLNC.
Fig.1 shows an example of the result of the GMR algorithm.
III. LDLC IN PLNC VIA C&F
We consider here a TWRC network as shown in Fig.2 where
two source nodes S1 and S2 transmit information messages to
the destination nodes D1 and D2 via a single relay node R. We
are mainly concerned with the MAC phase since it dominates
the overall system performance compared to the broadcast
phase. The n-dimensional message vector at the node S�,
� ∈ {1, 2} is denoted as b� = {b�,1, b�,2, · · · , b�,n} b� ∈ Fnp.
Without loss of generality, we consider that the two source
nodes have the same message length n. Each message vector
b� will be encoded to an n-dimensional LDLC codeword by
the transmitter encoder, E� : Fnp → Rn,
E�(b�) = L
(φ(g(b�)
))(13)
where E� is equipped with the mapping function g(·) : Fnp →
Zn, a shaping operator defined in (7) - (11) and a LDLC
encoder L� : Zn → Rn. The transmitted codeword is subject
to the average power constraint,
E[‖ E�(b�) ‖
2]� nP (14)
where P = Ω(Λ) is the normalized second moment over the
voronoi region V of the generated low density lattices Λ. Relay
R observes the noisy superimposed signals
y = x1 + βx2 + z (15)
where β is the relative channel fading coefficient β =|β| exp(jθ) between the two MACs. θ is the channel phase
difference, and z ∼ N (0, σ2In).
The relay attempts to decode y to a linear combination u
over Fnp,
u = (q′1 � b1)⊕ (q′2 � b2) (16)
where � and ⊕ are referred to as element-wise multiplication
and element-wise addition over Fnp, and q′� ∈ Fn
p. The
relay first attempts to decode a lattice equation v which is
an integer combination of lattice codewords based on some
integer coefficients c�, v =∑
2
�=1c�x�. In order to estimate
the lattice equation v, the LDLC decoder quantizes αy to the
closest lattice points in Λ, where α is the optimal scaling factor
minimizing the effective noise variance. Following Definition
2, the LDLC decoder is a lattice quantizer with respect to the
low density lattices Λ,
v = QΛ(αy) (17)
Then, the relay maps v to u which is the estimate of the
ideal linear combination u. We derive that:
u = E−1(v) = g−1(φ−1
(G−1v
))(18)
= g−1
(φ−1
(G−1
2∑�=1
c�x�
))(19)
(a)= g−1
(φ−1
(2∑
�=1
c� [g(b�)−m�k�]
))(20)
(b)= g−1
([2∑
�=1
c�g(b�)
]mod m�
)(21)
(c)= g−1
[g
(2⊕
�=1
(q′� � b�
))]mod m� (22)
=
2⊕�=1
(q′� � b�
)(23)
where (a) follows from the LDLC encoding and shaping
operations defined in section II-B, (b) follows from the inverse
shaping, and (c) follows from the algebraic property of the
bijective function g(·), and q′� = g−1([c�] mod p).⊕
denotes
element-wise summation over Fnp.
Non-Fading Case: Thus, β = 1, and y = x1 + x2 + z. In
this case, the original message b�,i at the ith dimension is not
necessarily a value of the prime-sized finite field Fp. It is easy
to verify that the algebra ring (thus b� ∈ Fnq) will suffice to
validate the derivations of (18) - (23). Now, the estimate u of
the linear combination is over ring Fnq, thus,
u = b1 � b1 (24)
where � denotes addition over Fnq.
This in fact relaxes the constraints of the information mes-
sages, and hence increases the flexibility of the constellation
selection. As mentioned in section II-B, the constellation size
at each dimension should be carefully selected in terms of
the shaping parity check matrix to guarantee the optimised
trade-off between the average transmission rate per dimension
and the reliability of the lattice decoding. In other words, a
message vector from an n-dimensional ring provides a more
flexible way in the design of the constellation size to reduce
the rate loss. Note that the hypercube-shaped signal codes are
not sensitive to the constellation selection since the shaping is
directly implemented in its generator matrix which naturally
has a toeplitz form. We also notice that the authors in [8] use
signal codes for C&F at TWRC based on the non-fading case.
Fading case: Due to the effects of the fading coefficients,
the linear combination of the transmitted lattice codewords is
distorted. A good solution [3] is to scale the received signal
by an MMSE factor α and hence to maximize the possibility
of correctly quantizing an integer linear combination of the
LDLC codewords. Following (15), the scaled y is
y = αy
=
2∑�=1
c�x� +
2∑�=1
(αβ� − c�)x� + αz (25)
The optimal scaling factor α is selected based on the MMSE
criterion such that the mean squared norm ε of the effective
noise zeff �∑2
�=1(αβ� − c�)x� + αz is minimized. By
simply setting the first derivative to zero with respect to α,
the optimal scaling factor is given by:
αopt =βT c�P
‖ β ‖2 P + σ2(26)
where β = [1, β] and c� = [1, c�,2] in terms of (15). The
integer coefficient vector c� is determined by maximizing the
computational rate
R(β, c�) = maxc�∈Z
2log
2
(P
‖ αoptβ − c� ‖2 P + |α|2σ2
)
(27)
which is equivalent to the shortest vector problem: c� =argminc�∈Z
2,c� �=0
{cT� Dc�
}, where D is a positive definite
matrix derived from (27).In the fading case, we find that the message vector from a
ring b� ∈ Fnq cannot lead to equation (22), and hence a linear
combination in the form of (23) cannot be obtained. The infor-
mation message must be the finite field value in order to get the
linear combination of (23). Due to the special characteristic of
LDLC as mentioned above, it would be of interest to examine
the possibility of relaxing the constraint from the finite field to
ring for the original messages. Apparently the relay needs to
recover another form of linear combination. Following LDLC
quantizer output, the estimate of the linear combination can
be described as:
ur = E−1
r(v)
= G−1v mod |c�,2|m�, |c�,2| > 1 (28)
=
[2∑
�=1
c�φ(gr(b�)
)]
mod |c�,2|m� (29)
where gr(·) : Fnq → Zn. Thus, the estimate of the fine lattice Λ
is mapped to a linear combination over a ring with increased
cardinality |c�,2|m�. This provides protection for those code-
words that have been servely faded. At the destination nodes
D1 and D2, they attempt to recover the original messages
(assuming error-free transmission over the broadcast phase).
Node D2 has the knowledge of the information message b2
and recovers b1 using:
b1 =[[ur − c�,2gr(b2)] mod |c�,2|m�
]mod m� (30)
17 18 19 20 21 22 23 24 2510−5
10−4
10−3
10−2
10−1
SNR [dB]
SE
R
Capacity of Hypercube Shapingβ=1, (1000,7) LDLC, 50 Iterationsβ=1.9, (1000,7) LDLC, 50 Iterationsβ=1, (100,5) LDLC, 10 Iterations
2.9 bits/dimension
Fig. 3. SER performance of the LDLC-based PLNC scheme.
Node D1 has the knowledge of b1 and attempts to recover b2
using:
b2 =
[ur −
([φ(gr(b1))] mod c�,2m�
)]mod |c�,2|m�
|c�,2|(31)
Note that node D1 needs to reshape gr(b1) for the successful
recovery of b2.
IV. SIMULATION RESULTS AND DISCUSSIONS
In this section we show and analyse the performance of
our LDLC-based PLNC scheme. Simulations were carried
out for the ring-type constellation b� ∈ Fnq where the con-
stellation size qi at the ith dimension is determined by the
shaping parity check matrix aforementioned in section II-B.
We set the nominal constellation size m = 8 and hence
qi = mi = �|Ji,im|� which gives an average transmission rate
of R = 2.90 bits/dimension. The symbol-to-noise ratio (SNR)
is defined as SNR = E[‖xi‖2]
σ2 . Fig.3 shows the symbol error
rate (SER) performance for non-fading and fading cases. In the
non-fading case, we observe that there is a loss around 1.7 dB
from the hypercube shaping capacity (HSC) at SER = 10−5
for (1000, 7) LDLC with 50 iterations, and around 4.1 dB
at SER = 10−5 for (100, 5) LDLC with 10 iterations. Note
that the number of iterations increases for (1000, 7) LDLC
to ensure full convergence at the BP decoder, and hence its
complexity is higher than (100, 5) LDLC. Since the SER
performance of LDLC improves as the length n increases, we
expect that the loss could be further reduced by 0.6 ∼ 0.7 dB
for (104, 7) LDLC. In the fading case, we set a fixed relative
fading coefficient β = 1.9 for the simulation. Fig.3 shows that
the gap to capacity at SER = 10−5 is 2.9 dB for (1000, 7)LDLC with 50 iterations. Thus a further 1.2 dB is required in
fading.
From Definition 4, we state that the necessary condition for
reliable lattice decoding at relay is:
γ =Vol(λ∑
2�=1 c�x�
)2n
2πeσ2zeff
> 1 (32)
where Vol(λ∑2�=1 c�x�
)2n is the volume of the Voronoi region
of the LDLC lattice point v =∑
2
�=1c�x�, and σ2
zeffis
the variance of the effective noise. This ensures the error
probability
Pr(u = u) = Pr(zeff /∈ V(λ∑2
�=1 c�x�
)(33)
goes to zero exponentially as n → ∞. When the LDLC
codewords are faded by non-integers, the probability that the
effective noise leaves the Voronoi region of V
(λ∑2
�=1 c�x�
)
increases. This requires a smaller receiver noise variance σ2,
and hence a larger SNR, to fulfil the condition of (33).
V. CONCLUSIONS
We have demonstrated that LDLC is a class of good lattice
codes suitable for PLNC via C&F. We proved that finite field
messages at each dimension works well, but also showed the
feasibility using ring-based constellation for the linear system
recovery. This effectively reduces the rate loss of hypercube-
shaped LDLC. Simulations were carried out for the ring-type
constellation, and showed good SER performance for both
non-fading and fading case, using a modified GMR model.
Our work in the future includes the extension of the proposed
PLNC scheme to a more general network topology.
VI. ACKNOWLEDGEMENTS
The authors would like to thank Chen Feng from University
of Toronto for useful discussions on lattice network coding.
REFERENCES
[1] S. Zhang, “Hot topic: physical-layer network coding,” in in Proc. ofACM Mobicom, 2006, pp. 358–365.
[2] P. Popovski and H. Yomo, “Bi-directional amplification of throughputin a wireless multi-hop network,” in Vehicular Technology Conference,2006. VTC 2006-Spring. IEEE 63rd, vol. 2, May 2006, pp. 588–593.
[3] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing interfer-ence through structured codes,” Information Theory, IEEE Transactionson, vol. 57, no. 10, pp. 6463–6486, Oct 2011.
[4] C. Feng, D. Silva, and F. Kschischang, “An algebraic approach tophysical-layer network coding,” Information Theory, IEEE Transactionson, vol. 59, no. 11, pp. 7576–7596, Nov 2013.
[5] N. Tunali, K. Narayanan, J. Boutros, and Y.-C. Huang, “Lattices overeisenstein integers for compute-and-forward,” in Communication, Con-trol, and Computing (Allerton), 2012 50th Annual Allerton Conferenceon, Oct 2012, pp. 33–40.
[6] Q. Sun, T. Huang, and J. Yuan, “On lattice-partition-based physical-layernetwork coding over gf(4),” Communications Letters, IEEE, vol. 17,no. 10, pp. 1988–1991, October 2013.
[7] O. Shalvi, N. Sommer, and M. Feder, “Signal codes: Convolutionallattice codes,” Information Theory, IEEE Transactions on, vol. 57, no. 8,pp. 5203–5226, Aug 2011.
[8] N. Tunali and K. Narayanan, “Concatenated signal codes with ap-plications to compute and forward,” in Global TelecommunicationsConference (GLOBECOM 2011), 2011 IEEE, Dec 2011, pp. 1–5.
[9] C. Feng, D. Silva, and F. Kschischang, “An algebraic approach tophysical-layer network coding,” in Information Theory Proceedings(ISIT), 2010 IEEE International Symposium on, June 2010, pp. 1017–1021.
[10] N. Sommer, M. Feder, and O. Shalvi, “Low-density lattice codes,”Information Theory, IEEE Transactions on, vol. 54, no. 4, pp. 1561–1585, April 2008.
[11] ——, “Shaping methods for low-density lattice codes,” in InformationTheory Workshop, 2009. ITW 2009. IEEE, Oct 2009, pp. 238–242.
[12] Y. Yona and M. Feder, “Efficient parametric decoder of low density lat-tice codes,” in Information Theory, 2009. ISIT 2009. IEEE InternationalSymposium on, June 2009, pp. 744–748.
[13] B. Kurkoski and J. Dauwels, “Message-passing decoding of latticesusing gaussian mixtures,” in Information Theory, 2008. ISIT 2008. IEEEInternational Symposium on, July 2008, pp. 2489–2493.