LDPC Coding For The Three-Terminal Erasure Relay Channel

Abstract

A three terminal erasure relay channel is considered. It has been shown that appropriately designed maximum distance separable codes achieve the cut-set upper bound on capacity of the three terminal erasure relay channel. This paper presents low-density parity-check (LDPC) coding alternatives for this channel. Design rules for constructing LDPC codes that perform close to the cut-set upper bound on capacity are provided for the general erasure relay channel and also the degraded erasure relay channel, wherein all the information available at the receiver are also available at the relay

Full text

Page 1

LDPC Coding For The Three-Terminal Erasure
Relay Channel
Deepak Sridhara
Institut f¨ ur Mathematik
Universit¨ at Z¨ urich
CH8057 Z¨ urich, Switzerland.
Email: sridhara@math.unizh.ch
Christine A. Kelley
Department of Mathematics
University of Notre Dame
Notre Dame, IN 46556, USA.
Email: ckelley1@nd.edu
Abstract—A three terminal erasure relay channel is considered.
It has been shown that appropriately designed maximum distance
separable codes achieve the cut-set upper bound on capacity of
the three terminal erasure relay channel. This paper presents low-
density parity-check (LDPC) coding alternatives for this channel.
Design rules for constructing LDPC codes that perform close
to the cut-set upper bound on capacity are provided for the
general erasure relay channel and also the degraded erasure relay
channel, wherein all the information available at the receiver are
also available at the relay.
I. INTRODUCTION
A simple relay network comprising of one sender, one
receiver, and one intermediate node, called the relay, that
participates in the communication by relaying packets from
the sender to the receiver, is considered. Relay channels that
have been considered are predominantly noisy channels with
interferences between the sender and relay transmissions at
the receiver. In this paper, we consider a simplistic relay
channel, where there is no interference and the only channel
impairment are losses or erasures. Consequently, this network
with three nodes will be referred to as the three-terminal
erasure relay channel. From the standpoint of higher layers
in a communication network, the erasure channel model is
appealing, since they view communication as over a packet
network where packets arrive either error-free or are erased by
the lower link-layer error-detection protocols.
Erasure channels are particularly simple to analyze. In [2], it
is shown that simple closed form expressions can be derived for
erasure relay channels as opposed to relay channels with errors.
[3] presents a cut-set upper bound for the achievable rate for
the erasure relay channel without interference and shows that
this bound is achievable for the degraded relay channel using
maximum-distance-separable (MDS) codes. However, MDS
codes of long block lengths are impractical for use in practical
system design since their decoding complexity increases as
O(n3) for a block length n. Furthermore, a typically large
field size is necessary to construct an MDS code of a large
block length n.
Hence, motivated by the remarkable success of low-density
parity-check (LDPC) codes in performing close to capacity
over several point-to-point communication channels, and in
particular over the binary erasure channel [4], we propose,
as a first step, to use LDPC codes instead to come close to
the capacity of the degraded erasure relay channel with no
interference. (The no-interference assumption at the receiver
can be realized practically if the transmissions from the sender
to the receiver and those from the relay to the receiver
occur at different frequencies or at different times following
an FDMA/TDMA protocol.) Following the density evolution
1
Y
1
X
0
Y
D
2
Y
0
X
2
Y
0
Y
Y=( , )
p1
p2
p
C1
C2
S
R
Fig. 1.Erasure Relay Channel
techniques of [4], [5], we derive the necessary conditions on
the degree distributions of LDPC codes to be used in the
relay channel to come close to the cut-set upper bound of [3].
Further, LDPC-based coding extensions to the non-degraded
case is also proposed.
II. CHANNEL MODEL
The communication channel model that is considered in this
paper is shown in Figure 1. It comprises of a sender node S, a
receiver (or, destination) node D, and an intermediate or relay
node R. The relay channel is described by five random vari-
ables X0,X1,Y0,Y1, and Y2and a conditional probability
density function p(y0,y1,y2|x0,x1). This probability density
function specifies the probability that when x0is sent by S and
x1is sent by R, y0and y2are received at D and y1is received
at R. Furthermore, we can view the above relay channel as two
separate channels: an erasure-broadcast channel (X0;Y0,Y1)
[6] and a point-to-point erasure channel (X1;Y2). Further, we
assume that the erasure loss probabilities on the links from
S to R, from R to D, and from S to D are p1, p2, and p,
respectively.
III. CUT-SET UPPER BOUND ON ACHIEVABLE RATE
The cut-set upper bound on the achievable rate for the three-
terminal relay channel, obtained in [1], is given here.
Theorem 3.1 (Capacity region bound): [1] The capacity re-
gion of the relay channel in Figure 1 is bounded by
R ≤
sup
p(x0,x1)
min{I(X0; Y0, Y1) − I(X1; Y0, Y1), I(X0; Y0) + I(X1; Y2|X0)}
If the relay channel is a degraded channel, i.e., X0→ Y1→
(Y0,Y2), is a Markov chain, then the bound becomes
R ≤
sup
p(x0,x1)
min{I(X0; Y1) − I(X1; Y1), I(X0; Y0) + I(X1; Y2|X0)}
For the erasure relay channel, the capacity bound simplifies:
Theorem 3.2: [3] The capacity region over an erasure relay
channel is bounded as:
R ≤ max
α
min{(1 − p · p1),(1 − p) + α(1 − p2)}

Page 2

where p1,p2,p are the loss probabilities between sender and
relay, relay and destination, and sender and destination, re-
spectively, and 0 ≤ α ≤ 1 is a coupling parameter that
describes the proportion of information sent by the sender that
is available at the relay. (α = 1 if R < (1 − p1) and α = p
otherwise.)
Under the degraded channel situation, the bound becomes
R ≤ max
α
min{(1 − p1),(1 − p) + α(1 − p2)}
When p < p1, the relay can be bypassed and the maximum
rate achievable is R ≤ (1 − p). ([3] provides more details.)
A. Achievability of capacity bound [3]: Degraded case
Consider the relay channel in Figure 1. Suppose S has k
symbols (or, packets) to transmit. Let us suppose that there
exists an [k+m+?,k] MDS code defined over an appropriate
finite field with generator matrix G = [Ik×k| Ak×m| Bk×?].
The source S uses the generator matrix G1= [Ik×k | Ak×m]
to encode. Thus, the rate R =
Asymptotically, for large n, R receives around n(1 − p1)
symbols. Under the condition that R ≤ 1 − p1, an MDS
code will, asymptotically with large k and n, ensure perfect
communication between S and R. R then re-encodes the k
information packets of S by multiplying with generator matrix
G2 = [Bk×?]. At the destination D, asymptotically, at large
n = k + m and ?, n(1 − p) symbols arrive from S, and
?(1 − p2) symbols arrive from R. D decodes the received
symbols using the MDS code corresponding to G. As long
as at least k symbols are available at D, decoding at D will
result in complete recovery of the k information symbols.
Thus, the above coding scheme achieves a rate R ≤ min{1 −
p1,n(1−p)+?(1−p2)
n
} = min{1 − p1,(1 − p) + α(1 − p2)}.
IV. DESIGN RULES FOR LDPC CODES
In this section, we use LDPC codes of large block lengths
in the relay channel to come close to the cut-set upper bound.
k
k+m. (Let n = k + m.)
A. LDPC coding for the binary erasure channel
Let H be a parity-check matrix of a binary LDPC code
C and GH the corresponding Tanner graph. Further, let λ,ρ
be the variable node and check node degree distributions
of GH [5]. Adopting the notation in [5], we have λ(x) =
?dv
is the fraction of edges in GH that are incident with degree i
variable (resp., check) nodes and dvand dcare the maximum
degrees of variable (resp., check) nodes in GH. An LDPC code
constructed randomly and having a degree distribution (λ,ρ)
has a rate R satisfying
?1
?1
A random LDPC code typically achieves equality in the above.
Suppose an LDPC code of large block length n described
by a parity-check matrix H with degree distribution (λ,ρ) is
used for transmission across a binary erasure channel (BEC).
Then, under the assumption that there are no cycles of length
2? or less in the LDPC constraint graph GH, we can find the
probability of sending an erasure message along an edge from
a variable node to a check node in GH in the ?thiteration as
a function of the probability of sending an erasure message
along an edge from a variable node to a check node in the
(? − 1)thiteration and the erasure probability of the BEC.
i=2λixi−1and ρ(x) =?dc
i=2ρixi−1, where λi (resp., ρi)
R ≥ 1 −
0ρ(x)dx
0λ(x)dx
.
Let p(0)denote the erasure probability of the BEC and let p(?)
denote the probability of sending an erasure along an edge from
a variable node to a check node in the ?thiteration. Assuming
that GH has no cycles of length 2? or less, it is easy to show
that the following relation holds [4]
p(?)= p(0)λ(1 − ρ(1 − p(?−1)))
Defining a function f(x) = p(0)λ(1−ρ(1−x)), we can obtain
the condition for an LDPC code of degree distribution (λ,ρ)
to recover all erasures occurring at a BEC erasure probability
of p(0)in the asymptotic case where the LDPC code has an
arbitrarily large block length and there are no cycles of length
2? or less in the LDPC constraint graph for as large an ? as
required. Under these assumptions, the condition for recovering
all erasures is
f(x) < x, ∀x ∈ (0,p(0)]
This condition implies that as long as f(x) is strictly less than
x for all x in the range (0,p(0)], p(?)→ 0 as ? → ∞.
Thus, a rule in designing LDPC codes is to find a degree dis-
tribution pair (λ,ρ) that satisfies the following two constraints:
capacity of BEC = 1 − p(0)= rate of LDPC ≥ 1 −
?1
?1
0ρ(x)dx
0λ(x)dx
.
p(0)λ(1 − ρ(1 − x)) < x, ∀x ∈ (0,p(0)]
B. LDPC coding for the relay channel
In the following, we adapt the same encoding techniques
as with MDS codes in the previous section to obtain design
rules for the degree distribution of binary LDPC codes for the
relay channel. Let us suppose we have a generator matrix G =
[Bk×?| Ak×m| Ik×k]. The sender S encodes the information
to be sent using an LDPC code C1that is described by a parity
check matrix H1of size m × (k + m). Let H1have a degree
distribution (λ(a),ρ(a)). Suppose reducing H1 to systematic
form yields a matrix H1,sys = [Im×m| AT
[Ak×m| Ik×k] is a generator matrix for the code C1. Since
H1is designed randomly, we can assume that it has full rank
and therefore there exists a sparse m × m invertible matrix
D such that H1= DH1,sys. Thus, H1can also be written as
H1= [D | DAT]. Let D have a degree distribution (λ(1),ρ(1))
and DAThave a degree distribution (λ(2),ρ(2)).
We will now form a new parity check matrix H of size
(m + ?) × (k + m + ?) having the following form:
?
0m×?
where X and Y are sparse matrices and X is invertible.
Suppose reducing H to systematic form yields a matrix Hsys
of the form
m×k], then G1 =
H =
X?×?
0?×m
Dm×m
Y?×k
(DAT)m×k
?
,
Hsys=
?
I?×?
0
0
BT
AT
?×k
m×k
Im×m
?
Then, H
?
0
check matrices for the generator matrix G. The sub-matrix A
in G is obtained by converting H1to systematic form, and the
sub-matrix B in G is obtained from H as B = X−1Y .
Let us suppose that while constructing H, we choose X to
be a sparse matrix with degree distribution (λ(3),ρ(3)) and the
is obtained by leftmultiplying Hsys
with
X
0
D
?
. Therefore, both H and Hsys are valid parity-

Page 3

sub-matrix Y to be a sparse matrix with degree distribution
(λ(4),ρ(4)).
The encoding scheme is the same as before: S uses the
generator matrix G1= [Ak×m| Ik×k] to encode. The sender’s
packets arrive at R via an erasure channel with loss probability
p1and at D via an erasure channel with loss probability p.
Suppose the following two conditions hold:
rateR =
k
k + m
= 1 −
?1
?1
0ρ(a)(x)dx
0λ(a)(x)dx
≤ 1 − p1= capacity(1)
f(x) = p1λ(a)(1 − ρ(a)(1 − x)) < x, ∀x ∈ (0,p1]
Then, asymptotically, using an LDPC code of large block
length n = k+m, R will be able to reconstruct the sender’s in-
formation perfectly. R then encodes the information sequence
that it has reconstructed by multiplying with generator matrix
G2= [Bk×?]. D receives two sequences of data: one from R
over an erasure channel with loss probability p2and the other
from S over an erasure channel with loss probability p. D
uses the sparse parity-check matrix H to iteratively decode the
received stream. We will now derive conditions on the degree
distributions so that D can reconstruct the sender’s information
perfectly and at the same time operate close to capacity.
The analysis of the iterative decoder assuming no-cycles
in the LDPC constraint graph for H is as follows: Let L1
refer to the first ? columns of H, L2 refer to the next m
columns, and L3 refer to the last k columns. Similarly, let
R1 refer to the first ? rows of H and R2 to the remaining
m rows. Note that D receives information from R for the
code symbols corresponding to L1 and from S for the code
symbols corresponding to L2and L3. Hence, the code symbols
correspondingto L1arrive at D with erasure probability p2and
the code symbols corresponding to L2and L3arrive at D with
erasure probability p.
Let p(t,i)
an erasure message along an edge from a variable node of
degree i, belonging to Ls, to a check node in the tthiteration.
Similarly, let q(t,i)
an erasure message along an edge from a check node of degree
i, belonging to Rs, to a variable node in the tthiteration.
Further, let p(t)
of sending an erasure message along on edge from a variable
node, belonging to Ls, to a check node in the tthiteration.
Define q(t)
free assumption, we obtain the following conditions on how the
probability of erasures, defined above, evolve with the iteration
t.
p(t+1,i)
L1
Averaging over all i , we obtain
p(t+1)
L1
Similarly,
p(t+1)
L2
(2)
Ls, for s = 1,2,3, be the probability of sending
Rs, for s = 1,2, be the probability of sending
Ls, for s = 1,2,3, be the average probability
Rs, for s = 1,2, analogously. Then, under the cycle-
= p2· (q(t)
R1)i−1
= p2λ(3)(q(t)
R1)
= pλ(1)(q(t)
R2)
For calculating p(t+1,r)
degree r may be receiving messages along i edges that are
connected to check nodes in R1 and along j edges that are
connected to check nodes in R2, where i + j = r − 1. The
probability of having such a connection at a degree r variable
node in L2is δi,j(λ(4),λ(2)) = 0.5(λ(4)
the degree r variable node may be passing a message to a
check node in R1(or to a check node in R2) after receiving
i messages from nodes in R1and j messages from nodes in
L3
, observe that a variable node in L3of
i+1λ(2)
j+λ(4)
iλ(2)
j+1), since
R2. Thus, averaging over r, we obtain the final recursion for
p(t)
?dv4+dv2
r=2
i+j=r−1
where dv4and dv2are the maximum variable node degrees in
the sub-matrices Y and DAT, respectively.
Define γi,j(ρ(1),ρ(2)) := 0.5(ρ(1)
deriving the recursions for q(t)
we obtain
dr3+dr4
?
r=2
i+j=r−1
dr1+dr2
?
r=2
i+j=r−1
where dr1,dr2,dr3,dr4are the maximum row degrees in
D,DAT,X,Y , respectively.
Thus, combining the above recursions, and setting x =
p(t)
p(t+1)
L2
L3
obtain a final set of recursions for p(t)
conditions must hold for the iterative decoder at D to recover
all the erasures:
? dr3+dr4
r=2
i+j=r−1
? dr1+dr2
r=2
i+j=r−1
? dv4+dv2
r=2
i+j=r−1
where q(t)
L2as
p(t)
L3= p ·
??
δi,j(λ(4),λ(2))(q(t−1)
R1
)i(q(t−1)
R2
)j?
,
i+1ρ(2)
R2in a similar manner,
j
+ ρ(1)
iρ(2)
j+1). Then,
R1and q(t)
q(t)
R1=
?
γi,j(ρ(3),ρ(4))
?
1 − (1 − p(t)
L1)i(1 − p(t)
L3)j?
q(t)
R2=
?
γi,j(ρ(1),ρ(2))
?
1 − (1 − p(t)
L2)i(1 − p(t)
L3)j?
L1,y = p(t)
, and h(x,y,z) = p(t+1)
L2,z = p(t)
L3and f(x,y,z) = p(t+1)
, we can eliminate the q(t)
L1,p(t)
L1
, g(x,y,z) =
Rs’s and
L3. The following
L2,p(t)
f(x, y, z)=p2· λ(3)
??
γi,j(ρ(3), ρ(4))
?
1 − (1 − x)i(1 − z)j??
g(x, y, z)=p · λ(1)
??
γi,j(ρ(1), ρ(2))
?
1 − (1 − y)i(1 − z)j??
h(x, y, z)=p ·
??
δi,j(λ(4), λ(2))(q(t)
R1
)i(q(t)
R2
)j
?
,
R1and q(t)
R2are functions of x, y, and z.
f(x,y,z) < x, ∀x ∈ (0,p2],∀y,z ∈ (0,p]
g(x,y,z) < y, ∀x ∈ (0,p2],∀y,z ∈ (0,p]
h(x,y,z) < z, ∀x ∈ (0,p2],∀y,z ∈ (0,p]
The remaining conditions to be satisfied are on the rate R
of the LDPC code from S and the sizes of the sub-matrices
D,DAT, X, Y . For convenience, we choose ? = n. Then,
?1
?1
1-
?1
1-
?1
Condition (1) holds as long as (6)-(8) hold. And further, using
the analysis for decoding at D, condition (2), for recovering
all erasures at R, can be reduced to the following:
fR(x,y)=p1· λ(1)?dr1+dr2
r=2
gR(x,y)=p1· λ(2)?dr1+dr2
r=2
(3)
(4)
(5)
1 −
0ρ(1)(x)dx
0λ(1)(x)dx
= 0, 1 −
?1
?1
0ρ(3)(x)dx
0λ(3)(x)dx
= 0
(6)
?1
?1
0ρ(2)(x)dx
0λ(2)(x)dx
0ρ(4)(x)dx
=k − m
k
=2 −
1
R
≥ 2 − max{
1
1 − p1
,
1
1 − p + 1 − p2
}
(7)
0λ(4)(x)dx
=k − ?
k
=1 −
1
R
≥ 1 − max{
1
1 − p1
,
1
1 − p + 1 − p2
}
(8)
??
i+j=r-1
γi,j(ρ(1),ρ(2))
?
1 − (1-x)i(1-y)j??
??
i+j=r-1
γi,j(ρ(1),ρ(2))
?
1 − (1-x)i(1-y)j??
fR(x,y) < x, ∀x,y ∈ (0,p1]
gR(x,y) < y,
(9)
(10)
∀x,y ∈ (0,p1]

Page 4

The conditions (3)-(10) yield the design rules for constructing
LDPC codes to achieve or come close to the cut-set upper-
bound of capacity for the three-terminal degraded erasure relay
channel.
C. Simplification: Decode and Forward
A simpler strategy to adopt for coding in the degraded relay
case is the following: S uses an LDPC code H of rate
to transmit to R and D. As long as the rate
an LDPC code with an appropriate degree distribution (λ,ρ)
can be designed to recover all erasures at R. R decodes
the S’s transmission and simply forwards all the n symbols
to D. Asymptotically, at large n, D receives approximately
n(1 − p) symbols from S and n(1 − p2) symbols from R.
However, approximately only np(1 − p2) symbols from the
R’s transmission are new information to D. The total number
of unerased symbols at D is approximately n(1−p)+np(1−
p2) = n(1 − pp2). Thus, D views the transmission of the
sender’s codeword as over an erasure channel with effective
erasure probability pp2. D also uses the LDPC matrix H
to decode. This coding scheme achieves a maximum rate
R ≤ min{1 − p1,1 − pp2} which is lower than the cut-set
bound in Theorem 3.2 for the degraded case. The design rules
for the degree-distribution of H are
p1λ(1 − ρ(1 − x)) < x, ∀x ∈ (0,p1] to decode at R
pp2λ(1 − ρ(1 − x)) < x, ∀x ∈ (0,pp2] to decode at D
?1
?1
Although, the cut-set bound is not achievable using this coding
scheme, the simplicity in the LDPC design makes up for
the loss in capacity achievable by this method. Furthermore,
the design rules say that any (capacity-achieving) LDPC code
designed on the binary erasure channel with the desired rate
will work well in this scenario.
V. THE NON-DEGRADED CASE
The difference between the degraded case and the non-
degraded case is that in the degraded case, the relay has access
to all the information from the sender that are available to the
destination. The maximum-achievable rate for the erasure relay
channel in the non-degraded case is given by
R ≤ min{1 − pp1,(1 − p) + β(1 − p2)}
for 0 ≤ β ≤ 1. Here β = 1 if R ≤ 1−p1and β = p otherwise.
The case when 1 − p1 ≥ 1 − p + 1 − p2 reduces to
the degraded case and has already been considered. We will
consider some other interesting cases and review the coding
method that achieves this upper-bound on the achievable rate
using MDS codes as discussed in [7]. Further, we will propose
LDPC coding alternatives for these cases.
The general coding scheme is as follows: S uses an [n,k]
MDS code with generator matrix [Ik×k| Ak×(n−k)]. In the
general erasure channel, R asymptotically, at large n, receives
n(1 − p1) symbols sent from S and D receives n(1 − p)
symbols. Of these, asymptotically, n(1−p1)(1−p) are received
by both R and D, np(1−p1) symbols are received by R alone
and not by D, and np1(1−p) symbols are received by D alone
and not by R.
A. Case I: 1 − p1≤ 1 − p2
For R > 1 − p1, these conditions imply that the maximal
achievable rate must satisfy R ≤ min{1 − pp1,1 − p + p(1 −
p2)} = 1 − pp1.
k
n
k
n≤ 1 − p1,
(11)
(12)
1 −
0ρ(x)dx
0λ(x)dx
≤ min{1 − p1,1 − pp2}
(13)
1) MDS coding: Suppose the [n,k] MDS code used by S
has rate R =
k∗= n(1 − p1) symbols from S. R then encodes these k∗
symbols using an [n∗,k∗] MDS code of rate R2 =
1 − p2. This ensures that D can reconstruct the k∗symbols
since the capacity of the R to D channel is 1 − p2. Out of
these, approximately pk∗symbols have not been received by
D via the direct link from S. Thus, as long as n(1−p)+pk∗=
n(1 − pp1) ≥ k, the destination D will be able to reconstruct
the k information symbols of S.
2) LDPC coding: S uses an [n,k] LDPC code C described
by a sparse parity-check matrix H with degree-distribution
(λ,ρ). For a large block length n, R, asymptotically receives
k∗= n(1 − p1) symbols. The relay uses an [n∗,k∗] LDPC
code C2of rate R2=k∗
check matrix H2with degree-distribution (λ(2),ρ(2)). D first
decodes the LDPC code C2using iterative decoding. From the
previous sections, the design-rules for (λ(2),ρ(2)) to ensure
perfect recovery of the k∗encoded symbols are
f(x) = p2λ(2)(1 − ρ(2)(1 − x)) < x, ∀x ∈ (0, p2]
?1
?1
Once D has decoded the k∗= n(1 − p1) symbols of the
encoded codeword from R, it obtains knowledge of approx-
imately pk∗symbols that were not received via the direct
link from S. Thus, D now has information of n(1 − p) +
np(1−p1) = n(1−pp1) symbols. Hence, the effective erasure
probability between S and D, for decoding the LDPC code C,
is pp1. D now uses the parity-check matrix H to recover all the
erased symbols. Hence, the design rules for (λ,ρ), to ensure
perfect recover of all the erased symbols of the codeword sent
from S, are
k
n> 1 − p1, then R receives approximately
k∗
n∗ ≤
n∗ ≤ 1 − p2and described by a parity-
(14)
R2= 1 −
0ρ(2)(x)dx
0λ(2)(x)dx
≤ 1 − p2
(15)
f(x) = pp1λ(1 − ρ(1 − x)) < x,∀x ∈ (0,pp1]
(16)
R = 1 −
?1
?1
0ρ(x)dx
0λ(x)dx
≤ 1 − pp1
(17)
From the above rules, it is clear that the design of the two
LDPC codes used at S and R are independent. Hence, LDPC
codes with desired rates designed for the BEC will work well
in this scenario. Figure 2 shows the performance of two rate
1/2 LDPC codes of a specific irregular degree profile used at S
and at R for the non-degraded relay channel. The performance
for different blocklengths n is shown as a function of the S−D
erasure probability p. (p1 and p2 are fixed at 0.625 and 0.5,
respectively. Case I situation arises for p ≥ 0.8. The irregular
LDPC codes that were constructed have poor distances and
reveal error floors.)
B. Case II: 1 − p2≤ 1 − p1
For R > 1 − p1, these conditions imply that the maximal
achievable rate must satisfy R ≤ min{1 − pp1,1 − p + p(1 −
p2)} = 1 − pp2.
1) MDS coding: Suppose the [n,k] MDS code used by S
has rate R =k
p1) symbols from S. R then randomly chooses k∗= n(1−p2)
symbols out of the n(1−p1) symbols and encodes them using
an [n∗,k∗] MDS code of rate R2=k∗
that D can reconstruct the k∗symbols since the capacity of
the R to D channel is 1−p2. Out of these approximately pk∗
symbols have not been received by D via the direct link from
S to D. Thus, as long as n(1 − p) + pk∗= n(1 − pp2) ≥ k,
n> 1−p1, then R receives approximately n(1−
n∗ ≤ 1−p2. This ensures

Page 5

the destination D will be able to reconstruct the k information
symbols of S.
2) LDPC coding: Again, S uses an [n,k] linear code C
described by a sparse parity-check matrix H with degree-
distribution (λ,ρ). For a large block length n, R asymptotically
receives k?= n(1 − p1) symbols. R randomly chooses k∗=
n(1 − p2) of these k?symbols and encodes them using an
[n∗,k∗] LDPC code C2 of rate R2 =
described by a parity-check matrix H2with degree-distribution
(λ(2),ρ(2)). D first decodes the LDPC code C2using iterative
decoding. The design-rules for (λ(2),ρ(2)) to ensure perfect
recovery of the k∗encoded symbols remain the same:
f(x) = p2λ(2)(1 − ρ(2)(1 − x)) < x, ∀x ∈ (0,p2]
?1
?1
Once D has decoded the k∗= n(1 − p2) symbols of the
encoded codeword from S, it obtains knowledge of approx-
imately pk∗symbols that were not received via the direct
link from S. Thus, D now has information of n(1 − p) +
np(1−p2) = n(1−pp2) symbols. Hence, the effective erasure
probability between S and D, for decoding the LDPC code C,
is now pp2. The decoding rule remains the same as before and
the design rules for (λ,ρ) become
k∗
n∗ ≤ 1 − p2 and
(18)
R2= 1 −
0ρ(2)(x)dx
0λ(2)(x)dx
≤ 1 − p2
(19)
f(x) = pp2λ(1 − ρ(1 − x)) < x,∀x ∈ (0,pp2]
(20)
R = 1 −
?1
?1
0ρ(x)dx
0λ(x)dx
≤ 1 − pp2,
(21)
Here also, the design of the two LDPC codes is independent
and therefore, good LDPC codes with desired rates designed
on the BEC will work well in this scenario. Figure 3 shows
the performance of a rate 1/2 LDPC code of a specific
irregular degree profile at S and a rate 0.375 LDPC code at
R for the non-degraded relay channel. The performance for
different blocklengths n is shown as a function of the S − D
erasure probability p. (p1 and p2 are fixed at 0.5 and 0.625,
respectively. Case II situation arises for p ≤ 0.8.)
C. Case III: 1−p+1−p2≥ 1−p1≥ 1−p+p(1−p2) = 1−pp2
This case has not received adequate attention in [7] and we
believe the cut-set bound is not achievable for this case. This is
because if a code of rate R = min{1−pp1,(1−p)+β(1−p2)}
is used at S and R ≤ 1 − p1, then we must have β = 1 by
Theorem 3.2. However, to achieve the cut-set bound, we must
have R = min{1 − pp1,1 − p + 1 − p2} (≥ 1 − p1 by the
hypothesis), which is a contradiction. On the other hand, if a
code of rate R > 1−p1is used at S, then the coding scheme
presented in section III-A will fail since the relay will fail to
reconstruct the sender’s transmission perfectly.
Hence, the best rate that may be achievable in this case
is R ≤ 1 − pp2 and the coding scheme to achieve this rate
is the same as for the “Decode and Forward” strategy in
the degraded situation. Consequently, the design rules for the
LDPC code at S remain the same as in that case. Figure 4
shows the performance of a rate 1/2 LDPC code of a specific
irregular degree profile used at S for both the non-degradedand
the degraded relay channel. The relay uses the “Decode and
Forward” strategy. The performance for different blocklengths
n is shown as a function of the S − D erasure probability
p. (p1 and p2 are fixed at 0.1 and 0.625, respectively. Case
III situation in the non-degraded case arises when p ≥ 0.16.
It is interesting to note that at intermediate values of p, the
00.10.2 0.30.4
p (S−D erasure rate)
0.50.60.70.80.91
10−8
10−7
10−6
10−5
10−4
10−3
10−2
10−1
100
Bit erasure rate
Non−degraded case: p1=0.625, p2=0.5, 1−p1 < 1−p2
N=1000
N=10000
N=100000
Cut−set bound
Fig. 2.
as a function of p: p1= 0.625,p2= 0.5, Case I for p ≥ 0.8
Performance of irregular rate 1/2 length N LDPC codes at S and R
00.10.2 0.30.4
p (S−D erasure rate)
0.50.6 0.70.8 0.91
10−7
10−6
10−5
10−4
10−3
10−2
10−1
100
Bit erasure rate
Non−degraded case: p1=0.5, p2=0.625, 1−p1 > 1−p2
N=1000
N=10000
N=100000
Cut−set bound
Fig. 3.
R as a function of p. p1= 0.5,p2= 0.625, Case II for p ≤ 0.8
Performance of rate 1/2 length N LDPC at S, rate 0.375 LDPC at
performance of the same code in the degraded channel differs
from that in the non-degraded channel.)
REFERENCES
[1] T. M. Cover and A. A. El Gamal. Capacity Theorems for Relay Channels.
IEEE Trans. Inform. Theory, 25:572–584, 1979.
[2] S. Boucheron and M. R. Salamation. About Priority Encoding Transmis-
sion. IEEE Trans. Inform. Theory, March 2000.
[3] R. Khalili and K. Salamation. On the achievability of cut-set bound for
a class of erasure relay channels. in Proceedings of IWWAN 2004, Oulu,
Finland, 2004.
[4] P. Oswald and A. Shokrollahi.
erasure channel. IEEE Trans. Inform. Theory, 48(12):3017–3028, 2002.
[5] T. J. Richardson and R. L. Urbanke. Capacity of low-density parity-check
codes under message-passing decoding.
47(2):599–618, 2001.
[6] T. M. Cover and J. A. Thomas. Elements of Information Theory. John
Wiley and Sons, New York, Second Edition, 1991.
[7] R. Khalili and K. Salamation. On the achievability of cut-set bound for
a class of erasure relay channels: The non-degraded case. in Proceedings
of the Intl. Symposium on Information Theory and its Applications,
(CDROM), Parma, Italy, Oct. 10–13, 2004.
[8] R. M. Tanner. A recursive approach to low complexity codes. IEEE Trans.
Inform. Theory, 27(5):533–547, 1981.
Capacity-achieving sequences for the
IEEE Trans. Inform. Theory,
0 0.10.20.3 0.4
p (S−D erasure rate)
0.5 0.60.70.8 0.91
10−7
10−6
10−5
10−4
10−3
10−2
10−1
100
Bit erasure rate
"Decode & Forward": p1=0.1, p2=0.625, 1−p1 > 1−pp2
N=10000 (degraded)
N=10000 (non−degraded)
N=100000 (degraded)
N=100000 (non−degraded)
Cut−set bound
Fig. 4.
p. p1= 0.1,p2= 0.625, Case III for p ≥ 0.16
Performance of rate 1/2 length N LDPC code at S as a function of