Low Density Parity Check codes in OFDM system

Dileep M.K, Aravind Iyengar, Andrew Thangaraj, Srikrishna Bhashyam

Department of Electrical Engineering,
Indian Institute of Technology, Madras, Chennai, India 600036
andrew,skrishna@iitm.ac.in
AbstractIn this work, we study the performance of Low Den-
sity Parity Check (LDPC) codes over an Orthogonal Frequency
Division Multiplexing (OFDM) channel. We state a concentra-
tion theorem which shows that no Gaussian approximation is
required in the analysis of LDPC codes over OFDM. Then we
propose a rigorous density evolution method (without Gaussian
approximations) to prove the existence of thresholds for LDPC
codes over OFDM and evaluate the thresholds for various regular
and irregular LDPC codes. We calculate the capacity of OFDM
channel and compare LDPC threshold with this theoretical limit
and show that for irregular codes, LDPC thresholds are very
close to capacity at higher rates. We also compare the LDPC
threshold in OFDM with LDPC threshold in an ISI channel
with BCJR equalization. Using the feedback to the transmitter,
we apply Mercury/Waterlling power allocation to improve the
OFDM capacity and LDPC thresholds. We show that with
Mercury/Waterlling power allocation LDPC thresholds are very
close to capacity even at moderate rates .
I. INTRODUCTION
Low Density Parity Check (LDPC) codes exhibit a threshold
phenomenon when iteratively decoded using a sum-product
message passing decoder over many channels. Arbitrarily
low bit-error rates (BERs) can be obtained whenever the
channel noise level is below a particular threshold value
by increasing the blocklength and number of iterations. An
algorithm for nding the threshold, Density Evolution (DE),
has been proposed by Richardson et al [1]. For channels
with additive Gaussian noise, threshold DE was simplied
by Chung et al [2] using a Gaussian approximation. DE has
been extended to binary-input inter-symbol interference (ISI)
channels by Kavcic et al [3], who show that LDPC codes
provide near capacity performance over discrete-time Inter
Symbol Interference (ISI) channels.
The general problem of channel coding for OFDM systems
has been addressed in [4]. Design optimization of LDPC codes
for Multiple Input Multiple Output-OFDM (MIMO-OFDM)
system for a xed target data rate has been addressed in [5].
Mannoni et al proposed a linear criterion for the optimization
of irregular LDPC codes for an OFDM system [6]. Baynast
et al [7] have proposed a two-step optimization of irregular
LDPC codes for OFDM channels. All of these previous works
employ a Gaussian approximation for threshold estimation
and do not completely prove the existence of thresholds. In
[11] we proposed a density evolution algorithm without Gaus-
sian approximation and presented initial results on thresholds
calculated for LDPC codes over OFDM. In this paper, we
address the following: (1) Proof of the concentration theorem,
(2) Mercury/Waterlling power allocation across the OFDM
subcarriers, (3) More accurate computation of thresholds for ir-
regular and regular LDPC codes, (4) Comparison with OFDM,
ISI capacities.
Specically, we state a concentration theorem which shows
that no Gaussian assumption is necessary in the analysis of
LDPC codes over OFDM. Using this result we then propose a
rigorous density evolution algorithm to compute threshold for
LDPC codes over an ISI channel under OFDM. We assume
that one code block is transmitted using a single OFDM
symbol. In the algorithm, we allow the block length to tend to
innity. Consequently, the subcarrier spacing reduces and the
number of subcarriers tend to innity for the same bandwidth.
Since the number of subcarriers tend to innity, the nite
cyclic prex results in no additional overhead. We calculate
the OFDM channel capacity and compare OFDM thresholds
obtained by our density evolution with this theoretical limit.
We show that for higher rates (rates higher than 0.6) the
thresholds are very close to the theoretical limit. An optimum
power allocation scheme, Mercury/Waterlling, for parallel
Gaussian channel with arbitrary input constellation has been
proposed by Lozano et al [10]. We use this power allocation
scheme to improve the OFDM capacity. We apply LDPC
codes with this power allocation and demonstrate that LDPC
thresholds also show considerable improvement. We show that,
with this optimum power allocation, LDPC thresholds are very
close to capacity even at moderate rates ( rates higher than
0.2). We also make a comparison between the time-domain
BCJR algorithm and the frequency-domain OFDM method
for equalizing an ISI channel. To this end, we compare the
threshold for LDPC codes under OFDM with that of the BCJR
thresholds obtained by using the algorithm given in [3].
II. SYSTEM DESCRIPTION
In this work, we focus our attention on an Orthogonal
Frequency Division Multiplexing (OFDM) system. We assume
transmission over an ISI channel
1
with L xed taps. The
channel is modeled as
z = H.c + N, (1)
where c is the input vector, z the output vector, H the
Discrete Fourier Transform (DFT) of the Channel Impulse
Response and N the normalized DFT of the random noise
vector, and . denotes the dot product of two vectors. All
these vectors mentioned here are of length N
c
, the number of
1
This can be easily extended to OFDM over a block fading channel.
NCC 2009, January 16-18, IIT Guwahati 103
subcarriers. We have input alphabet X = F
2
and consider
BPSK modulated input 0 +1, 1 1. Therefore, the
power in every input symbol is the same and equal to unity.
The analysis of the decoder requires letting the blocklength
of the code to tend to innity. The motive behind the as-
sumption is that the cyclic prex involved in the OFDM
transmission remains an overhead of xed length (given by
the number of taps in the channel), while the number of
information symbols increases as the OFDM symbol length
tends to innity. This increases the throughput, and in the limit
the cyclic prex gives no overhead. The above assumption
would also imply that the number of subcarriers N
c
would
tend to innity. In the OFDM system, the length of all
vectors would also tend to innity and the Inverse Discrete
Fourier Transform (IDFT) and DFT computed would also
be innite-point versions. In the mathematical model for the
channel, the factor multiplying the signal is now a sample of
the Discrete Time Fourier Transform (DTFT) of the channel
impulse response instead of a DFT sample in the nite version.
III. ANALYSIS OF LDPC CODES OVER AWGN
For irregular LDPC codes of blocklength n, the parameters,
the variable and check node degree distributions are specied
as polynomials, denoted (x) and (x), respectively. The
triplet, (, , n), thus species an ensemble of LDPC codes.
A. Density Evolution
Density Evolution is an algorithm that analyzes an ensemble
of LDPC codes by tracking the probability density function
(pdf) of the message passed on a random edge of a graph in the
ensemble averaged over the entire ensemble. It can therefore
specify the average error probability as a function of iteration
number, thereby serving as a performance metric for LDPC
codes.
Consider an AWGN channel, with binary input and BPSK
modulation. The actual message passed on the edges in case
of AWGN channel in the zeroth iteration of message passing
is the Log-Likelihood Ratio (LLR) of the received value from
the channel. The pdf of the LLR is termed the L-density and
the L-density at iteration l is denoted f
l
. It is assumed that
there are no cycles up to depth l. Density evolution, in this
case, states
l(y) f
0
N
_
2

2
,
4

2
_
f
l
= f
0
( (f
l1
)) , (2)
where for L-density f
(f) :=

i
f
(i1)
, (f) :=

i
f
(i1)
.
Here, and denote the convolutions carried out in the
L-domain and G-domain, respectively. These domains are
dened in [9].
The probability of error obtained is a monotone function
with respect to the channel parameter (noise variance
2
) and
with respect to iteration number. Also, there exists a well-
dened supremum of for which probability of error 0 as
the number of iterations l , and this supremum is called
the threshold of the decoder, denoted

[9].
IV. ANALYSIS OF LDPC CODES OVER OFDM
A. Log-Likelihood Ratio
The OFDM channel described is clearly a binary memory-
less channel. The LLR dened as
u
i
= L(z
i
) := ln
_
p
Zi|Ci
(z
i
|c
i
= 1)
p
Zi|Ci
(z
i
|c
i
= 1)
_
forms a sufcient statistic with respect to decoding for all
binary memoryless channels.
B. Channel Symmetry
We see that
p
Zi|Ci
(z
i
|c
i
= 1) = p
Zi|Ci
(z
i
|c
i
= 1)
for z
i
given by (1). Thus, the channel is symmetric and can
be modelled [9] as
c
i
fU
i
Z
i
, Z
i
= c
i
U
i
, U
i
f
Ui
(3)
where f
Ui
is the distribution of u
i
conditioned on c
i
= 1 and
channel gain H[i]:
f
Ui
(u
i
) =

4|H[i]|

exp
_

(
2
u
i
4|H[i]|
2
)
2
16|H[i]|
2

2
_
(4)
i.e. U
i
N
_
4|H[i]|
2

2
,
8|H[i]|
2

2
_
. We see that
f
Ui
(u
i
) = exp(u
i
).f
Ui
(u
i
) (5)
and thus, the LLR distribution is symmetric. Under these
symmetry conditions, we can assume that the transmitted code
word is all-one codeword
2
.
C. Concentration Theorem
In an OFDM channel, message passing decoding with an
irregular LDPC code raises an interesting question. OFDM
system can be considered as a set N parallel AWGN channel,
each with a different SNR. For irregular codes, the degree of
each bit node can be different. The incoming bit from the ith
channel can be assigned to jth bit node for decoding. Since
we are assuming that the length of OFDM symbol is same as
the code length, there are N! such assignments possible (all
of them need not be different since there are many number
of bit nodes with same degree). For example, the incoming
bit from the highest SNR channel can be assigned to bit node
with lowest degree and the incoming bit from the lowest SNR
channel can be assigned to bit node with highest degree. All
the other bits follows this order. We can think of another
arrangement in the opposite order too. So, the problem is to
nd out an assignment which gives the optimum performance
2
BPSK modulated
NCC 2009, January 16-18, IIT Guwahati 104
in terms of BER and LDPC threshold. This is equivalent to
the problem of designing an optimum interleaver.
However, in our study, we observed that BER performance
(and LDPC threshold) of an irregular LDPC code is almost
the same for different random interleaving (incoming bits
are assigned to the bit nodes in a random manner). More
specically, out of the N! interleavers available, if we select
interleaver u and v uniformly at random, their performance
turned out to be remarkably close. We present this result as a
concentration theorem. The proof is given in [12]
Consider a degree distribution pair (, ) and transmission
over an OFDM channel with N subcarriers. We assume that
the block length of the LDPC code n is same as the number
of OFDM subcarrier ie, n = N. We denote each of these
interleaver (or,assignment) as vectors G
u
, 1 u N!.
1) Theorem: Let Z
l
Gu
be the random variable that denotes
the number of erroneous variable-to-check node messages after
l rounds of the message-passing decoding algorithm when the
code graph is chosen uniformly at random from the ensemble
of graphs with degree distribution pair (, ) and when the
the interleaver chosen uniformly at random is G
u
. Let p
l
Gu
be the expected number of incorrect messages along an edge
with a tree-like neighborhood of depth atleast 2l at the lth
iteration when the interleaver chosen uniformly at random.
Let n
e
be the number of edges in the graph. For an arbitrary
small constant > 0, there exists a positive constant =
(, , l), such that if n >
2

, then
P
_

Z
l
Gu
n
e
p

_
4e

2
n
(6)
where the error concentration probability p is dened as
p =
1
N!
N!

i=1
p
l
Gu
.
The theorem shows that Z
l
Gu
is highly concentrated around
p. This result ensures that we need not consider any particular
interleaver for the analysis of LDPC over OFDM since the
performance given by an interleaver selected uniformly at
random from the set of all arrangement is close to the average
performance and hence it is enough to study this average be-
haviour. This eliminates the need for Gaussian approximation
in the density evolution and enables us to propose a rigorous
density evolution which analyze this average behavior.
D. Density Evolution
Theorem: Consider an OFDM channel with N
c
subcarriers
with code of blocklength n = N
c
, with associated L-densities

f
i
, i {1, 2, . . . , N
c
} (4). Then, the initial message density
f
0
=
1
N
c
Nc

i=1

f
i
, (7)
and for l 1,
f
l
= f
0
( (f
l1
)) , (8)
where for L-density f
(f) :=

i
f
(i1)
, (f) :=

i
f
(i1)
.
Proof: The initial density of the LLR f
0
is the only step of
the algorithm that differs from the AWGN channel case. It can
be easily seen that the f
0
given by (7) is still symmetric, i.e. it
still satises (5). We now prove that the initial density of the
LLR is given by (7). The proof of the rest of the algorithm is
exactly the same as in the AWGN channel case and is given
in [9].
Let e
r
be a random edge and v
i
the ith variable node in the
Tanner graph G of the code specied by the degree distribution
pair (, ). Let N
e
be the total number of edges in G. The LLR
distribution of the message received at v
i
from the channel is
given by

f
i
. The probability density function of the message
carried by this edge in the variable-to-check message passing
step of the zeroth iteration, averaged over the ensemble of
graphs characterized by (, ) is given by
f
0
:=E
[G(,)]
(f(e
r
))
=
Nc

i=1
P(v
i
e
r
)

f
i
=
Nc

i=1
_

m
P(d
G
(v
i
) = m) .P(v
i
e
r
|d
G
(v
i
) = m)
_

f
i
=
Nc

i=1
_

m
_
N
e

m
mN
c
__
m
N
e
_
_

f
i
=
1
N
c
Nc

i=1

f
i
which is the same as (7).
E. Monotonicity and Threshold
As the update equation involved in the density evolution
algorithm is the same as that in the AWGN case, the mono-
tonicity arguments made there apply here also. We therefore
have a supremum

for the noise variance beyond which the

error probability does not converge to zero even after innite
rounds of message passing.
We estimate these thresholds for various regular LDPC
codes over OFDM channels. We describe the method em-
ployed for threshold estimation in the following section.
V. THRESHOLD ESTIMATION
Since we are assuming a single OFDM symbol over the
blocklength, we let N
c
tend to innity. Equation (7) is no
more a summation but an integral. As described earlier, the
LLR distribution is now a continuous function of the angular
frequency through its dependence on H(e
j
) the DTFT of the
NCC 2009, January 16-18, IIT Guwahati 105
channel impulse response. Equation (4) therefore becomes
f(u, ) =

4|H(e
j
)|

exp
_

(
2
u 4|H(e
j
)|
2
)
2
16|H(e
j
)|
2

2
_
H(e
j
) =

i=
h[i]e
ji
and (7) becomes
f
0
(u) =
1
2
_
2
0
f(u, ).d (9)
Unfortunately, the function f(u, ) is not always well be-
haved - it tends to the continuous Dirac-Delta function when
|H(e
j
)| = 0 and therefore f
0
(u) is not directly obtainable
from (9) when the channel has spectral nulls. This difculty
can be overcome by calculating f
0
(u) through the character-
istic function of f(u, ) [11].
The density given by this method is now used in the density
evolution to estimate the threshold.
VI. OPTIMUM POWER ALLOCATION USING
MERCURY/WATERFILLING
OFDM can be considered as a set on N
c
parallel AWGN
channel. On the ith channel, the input-output relation is
Y
i
= H
i
X
i
+W
i
, (10)
where the complex scalar H
i
is the deterministic gain while
the noise W
i
is a zero mean unit variance complex Gaussian
random variable independent of the noise of the other channel.
The aggregate power constraint is
1
Nc

Nc
i=1
E
_
|X
i
|
2
_
P.
Once we know the expected value of the signal power in
each channel, calculating the capacity of that channel is fairly
straight forward. Since we can calculate the capcity for each
parallel AWGN channel, the capacity of OFDM system can
also be computed easily.
The SNR corresponds to each channel can be different due
to different H
i
which scales the signal. Therefore, the trivial
power allocation, equal power on all subcarrier, will not be the
optimum. If the input to the parallel channels are mutually
independent and Gaussian, the optimum power allocation is
simple and given by the waterlling policy. However the inputs
are usually drawn from a discrete constellation, and the wa-
terlling policy is no longer optimum. The main difculty in
the formulation is the lack of explicit expression for the corre-
sponding mutual information. Recently, Mercury/Waterlling
scheme which is the optimum power allocation for parallel
Gaussian channels with arbitrary input constellation has been
proposed by Lozano et al [10].
We used the Mercury/Waterlling power allocation scheme
and calculated the OFDM system capacity. OFDM system
capacity shows 2-4 dB improvement compared to equal power
allocation case. Calculation of LDPC thresholds with this
power allocation in OFDM subcarriers may look like a
new challenge. We assumed that the input to each OFDM
subcarriers is binary with BPSK modulation and hence the
signal power in each subcarrier is constant. We formulated
the density evolution algorithm (which gives us the LDPC
thresholds) with these assumptions. So, how we calculate the
LDPC thresholds when the signal power in each subcarriers is
different is not very clear. But, this difculty can be overcome
by a simple manipulation. We can write X
i
=

p
i
PS
i
, where
where S
i
is unit power input. The the normalized powers p
i
are constrained by
1
Nc

Nc
i=1
p
i
1 so that overall power
constraint is satised. Now, we can subsume this

p
i
P factor
to the H
i
and rewrite (10) as
Y
i
= H

i
S
i
+W
i
, (11)
Now LDPC thresholds can be calculated with the same algo-
rithm.
VII. RESULTS
We give a few results of the thresholds estimated using the
proposed algorithm.
A. Threshold Evaluation
We consider a 2-tap channel,
{h[i]} = {
1

2
,
1

2
}.
We use this channel to compare results with those published
in [3].
(L, R) Rate ISI threshold [3] OFDM threshold

SNR

SNR

(3, 4) 0.250 1.196 1.555 1.175 1.413

(3, 5) 0.400 0.945 0.491 0.897 0.9376
(3, 6) 0.500 0.822 1.703 0.751 2.478
(3, 10) 0.700 0.631 3.999 0.479 6.3754
(3, 15) 0.800 0.547 5.240 0.3217 9.840
TABLE I
REGULAR LDPC CODE THRESHOLD FOR ISI & OFDM
Rate Equal Power Mercury/Waterlling

SNR

SNR

0.10 1.713 4.677 2.485 7.906

0.30 1.176 1.408 1.493 3.484
0.60 0.666 3.511 0.737 2.653
0.80 0.365 8.757 0.472 6.515
TABLE II
IRREGULAR LDPC CODE THRESHOLD FOR OFDM .
Table I gives the thresholds obtained for different rate
regular LDPC codes. It gives the degree distribution of the
LDPC code, its design rate, the threshold values obtained
3
for
OFDM and also lists the thresholds obtained for ISI channel
using the BCJR algorithm in [3].
Table II gives the thresholds for irregular LDPC codes.
These codes are optimized for AWGN channel.
3
The threshold values are scaled by

2 to compare with the corresponding

values of ISI thresholds, wherein the values correspond to the variance of
the real part of the noise.
NCC 2009, January 16-18, IIT Guwahati 106
15 10 5 0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR (dB)
R
a
t
e

(
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
)
Channel with spectral Null


OFDMRegular LDPC Threshold
OFDMIrregular LDPC Threshold
OFDM Capacity BPSK
ISIRegular LDPC Threshold
ISI Capacity BPSK
Fig. 1. OFDM and ISI Thresholds
10 5 0 5 10 15
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR (dB)
R
a
t
e

(
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
)


CapacityQPSK,Equal Power
CapacityQPSKM/WF
LDPC ThresholdsEqual Power
LDPC ThresholdsM/WF
Capacity Gaussian WF
Fig. 2. LDPC Thresholds for OFDM with Mercury/Waterlling power
allocation.
Figure 1 summarizes the results. It gives the LDPC thresh-
olds (for regular and irregular codes) over OFDM and LDPC
thresholds over an ISI channel. For comparison, we also plot
OFDM capacity and ISI capacity. We see that irregular codes
gives an improvement of 1dB over the regular codes. At higher
rates LDPC threshold (for irregular codes) are very close to
the theoretical limit, the OFDM capacity. LDPC thresholds
for ISI channel is also very close to ISI channel capacity at
higher rates. In both case, the gap between the capacity and
LDPC thresholds increases as the rate decreases. We can also
compare OFDM-LDPC thresholds and ISI-LDPC thresholds.
We see that for lower rate OFDM-LDPC thresholds are better
than that of ISI-LDPC thresholds. But as the rate increases,
ISI-LDPC thresholds get better and at higher rates ISI-LDPC
threshold are very much superior to OFDM-LDPC thresholds.
One can observe a fundamental reason behind this. In this
region, ISI capacity is much better than the OFDM capacity,
which can be seen clearly from the gure. OFDM-LDPC
threshold can get maximum upto the OFDM channel capacity
and at higher rate they do so. However the ISI capacity is much
better than these values and ISI-LDPC codes perform close to
those. So, ISI-LDPC thresholds are fundamentally better than
OFDM-thresholds in this region.
We apply Mercury/Waterlling power allocation policy to
OFDM and calculate the LDPC threshold in this case. Table
II gives the OFDM-LDPC thresholds for irregular codes.
OFDM capacity with Mercury/Waterlling power allocation
and LDPC thresholds in this case are plotted in the gure 2
along with equal power case (We have considered QPSK mod-
ulation here). LDPC thresholds show a 2-4 dB improvement.
More interestingly, LDPC thresholds in this case are very close
to the capacity even at moderate rates.
VIII. CONCLUSION
In this work, we have proposed a rigorous density evolution
(without Gaussian approximation) to analyze the performance
of LDPC codes over an OFDM system. We proved the exis-
tence of LDPC threshold in an OFDM system and calculated
the LDPC threshold for various regular and irregular codes.
We have shown that, for higher rates the irregular LDPC
thresholds are very close to OFDM capacity. Then, we used
Mercury/Waterlling power allocation for OFDM subcarriers
and calculated the LDPC thresholds in this case. In this case,
we demonstrated that, irregular LDPC threshold are close to
capacity even at moderate rates.
REFERENCES
[1] T.Richardson, R.Urbanke, The capacity of low-density parity check
codes under message passing algorithm, IEEE Transactions on Infor-
mation Theory, vol. 47, pp.599618,Feb 2001.
[2] S.Y.Chung, T.Richardson, R.Urbanke, Analysis of sum-product decod-
ing of low density parity check codes using a Gaussian approximation,
IEEE Transactions on Information Theory, vol. 47, pp.657670,Feb
2001.
[3] A.Kav ci c, Xiao Ma, M.Mitzenmacher, Binary Intersymbol Interference
Channels: Gallager Codes, Density Evolution and Code Performance
Bounds, IEEE Transactions on Information Theory, pp.100118, Feb
2002.
[4] W.Y. Zou, Y. Wu, COFDM : An Overview, IEEE Transaction on
broadcasting, vol.41, no. 1, pp.1-8,March 1995
[5] B.Lu, G.Yue, X.Wang, Performance Analysis and Design Optimization
of LDPC coded MIMO OFDM system, IEEE Transaction On Signal
Processing, vol.52, issue 2, pp.348-361, Feb.2004
[6] V.Mannoni, G.Gelle, D.Declercq, A Linear Criterion to Optimize
Irregular LDPC Codes for OFDM Communications, IEEE Vehicular
Technology Conference, vol.1, pp.100118, May 2005.
[7] A.Baynast, A.Sabharwal, B.Aazhang, LDPC Code Design for OFDM
channel: Grapgh Connectivity and Information Bits Positioning, ISSCS
2005. International Symposium on Signals, Circuits and Systems, vol. 2,
pp.649652,July 2005.
[8] D.Tse, P.Viswanath Fundamentals of Wireless Communication, Cam-
bridge University Press, First South Asian Edition 2006
[9] T.Richardson, R.Urbanke, Modern Coding Theory
http://lthcwww.epfl.ch/mct/index.php
[10] A.Lozano, A.M.Tulino, S. Verdu, Optimum Power Allocation for
Parallel Gaussian Channels With Arbitrary Input Distributions, IEEE
Transactions on Information Theory, Vol. 52, July 2006.
[11] A. Iyengar, M. K. Dileep, S. Bhashyam, A. Thangaraj, TThresholds for
LDPC codes over OFDM, Proceedings of IEEE COMSWARE 2008,
Workshop on WIreless Systems: Advanced Research and Development
(WISARD 2008), January 2008.
[12] M. K. Dileep,, LDPC codes over an OFDM system, Mtech Thesis,
IIT Madras, June 2008.
NCC 2009, January 16-18, IIT Guwahati 107