Sie sind auf Seite 1von 11

998

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 4, APRIL 2011

Low Complexity Demapping Algorithms for


Multilevel Codes
Gkhan Gl, Student Member, IEEE, Aharon Vargas,
Wolfgang H. Gerstacker, Member, IEEE, and Marco Breiling

AbstractIn order to reduce the computational complexity


of maximumlikelihood symbol estimation (MLSE) demapping
of multilevel codes which is based on block partitioning and
which produces soft input for a multistage decoding (MSD)
process, three different demapping algorithms are proposed.
It is theoretically proven that the proposed algorithms can
reduce exponentially increasing computational complexity of the
MLSE demapping algorithm to a constant complexity (neglecting
comparisons). It is shown by extensive simulations for AWGN
and Rayleigh fading channels that the proposed low complexity
demapping algorithms can achieve near MLSE performance.
Index TermsMultilevel codes, multistage decoding, maximum
likelihood symbol estimation, log likelihood ratio.

I. I NTRODUCTION

ODED modulation is referred to the process of the


joint optimization of coding and modulation. The idea
of coded modulation has been introduced independently by
Ungerbck [1], [2] and Imai and Hirakawa [3] in 1976/77,
and generalized later by Pottie and Taylor [4] and Calderbank
[5]. Basically, the signal set (constellation) is divided into
several subconstellations using a criterion for set partitioning.
Maximization of minimum intrasubset Euclidean distance is
the criterion considered by Ungerbck and named afterwards
as Ungerbck partitioning or Ungerbck labeling (UL) [2].
Since the partitioning is accompanied by trellis codes, Ungerbcks approach to coded modulation is known as trellis
coded modulation (TCM).
The multilevel coding (MLC) approach of Imai is also a
form of coded modulation such as TCM. To each partitioning
level , an individual code rate is assigned to provide
different levels of protection. The set partitioning technique of
MLC can be UL or any other partitioning which is binary or
nonbinary. Additionally, without any constraint, block codes,
convolutional codes, or concatenated codes can be selected as
Paper approved by L. K. Rasmussen, the Editor for Iterative Detection Decoding and ARQ of the IEEE Communications Society. Manuscript received
August 20, 2009; revised April 9, 2010 and October 26, 2010.
G. Gl is with the Signal Processing Group, Institute of Telecommunications, Technische Universitt Darmstadt, 64283 Darmstadt, Germany (e-mail:
ggul@spg.tu-darmstadt.de). During this work, he was with the Fraunhofer
Institute for Integrated Circuits (IIS), Am Wolfsmantel 33, D-91058 Erlangen,
Germany.
A. Vargas and M. Breiling are with the Fraunhofer Institute for Integrated
Circuits (IIS), Am Wolfsmantel 33, D-91058 Erlangen, Germany (e-mail:
{aharon.vargas, marco.breiling}@iis.fraunhofer.de).
W. H. Gerstacker is with the Universitt Erlangen-Nrnberg, Chair of
Mobile Communications, Cauerstrae 7, D-91058 Erlangen, Germany (e-mail:
gersta@LNT.de).
Digital Object Identifier 10.1109/TCOMM.2011.020411.090497

the component codes for each level. Therefore, TCM can be


considered as a special case of MLC.
Huber and Wachsmann have proven in [7] that the capacity
of the modulation scheme can be achieved iff the individual
rates of the component codes are chosen according to the
individual level capacities, regardless of the mapping. Furthermore, in [4] and [5] it was clearly indicated that arbitrary
hierarchies of codes can be employed in a multilevel coding
approach. An efficient way of decoding multilevel codes by
individual decoding of each component code beginning from
the lowest level and using the decisions of lower levels for
the following levels is called multistage decoding (MSD).
However, until recently, MLC has achieved mainly theoretical
attention due to a severe degradation of the performance for
low levels. Although Cover already stated the idea of sub
constellations centered on centroids which are virtual points
at the center of mass of clusters of signal points, in [6],
it was Huber who showed that this degradation was due to
Ungerbcks set partitioning and alternatively proposed block
partitioning or block labeling (BL) to solve this problem [8].
He defined the minimum intrasubset variance as the criterion
for block partitioning. Indeed, the results in [8] and later in
[9] showed that BL is an optimum solution for a hierarchical
transmission system which uses MLC.
Since MLC has achieved also practical attention after
the proposal of block labeling, the tradeoff between the
computational complexity and the performance of demapping
has become a crucial factor. Although MSD is a fast solution
for decoding, the maximumlikelihood symbol estimation
(MLSE) demapping which is exploited in the MSD process introduces an exponentially increasing complexity with
increasing modulation order. This may cause a very slow
or inefficient decoding. Moreover, for a receiver with sufficient computational power, reduced complexity demapping
can provide a significant gain in terms of an efficient use
of power. Consequently, the reduction of the computational
complexity of the demapping stage, where the aposteriori
probabilities of the bits represented by the transmit signal
points are extracted after the observation of the received signal,
is of great importance and the main goal of this paper.
In summary, there are two main contributions of this
paper. First, the computational complexity of the MLSE algorithm for MLC in conjunction with MSD and BL is reduced
to a constant complexity (neglecting comparisons) while its
performance remains near to that of MLSE for additive
white Gaussian noise (AWGN) and Rayleigh fading channels.

c 2011 IEEE
0090-6778/11$25.00

GL et al.: LOW COMPLEXITY DEMAPPING ALGORITHMS FOR MULTILEVEL CODES

In this paper, we focus on MLSE for Mary amplitude


phaseshift keying (MAPSK) modulation, but the proposed
algorithms are also applicable to other modulation formats
such as quadrature amplitude modulation (QAM) or lattice
signal constellations. Second, it is shown that the proposed
approach is applicable to the case where the channel cannot
be modeled using typical channel models, such as AWGN or
Rayleigh. This property has a crucial importance for practice
since typical channel models are usually not valid in reality
and a tunable approach can provide a certain flexibility for
these kinds of scenarios.
The organization of this paper is as follows. In Section II,
an overview of the system model is given. Multilevel coding
in conjunction with multistage decoding is presented and
some set partitioning techniques are discussed. In Section III,
three different demapping algorithms are proposed to solve
the aforementioned complexity problem of MLSE. The limits
on the achievable lowest computational complexity of the
proposed two novel algorithms as well as of the maxlog algorithm, e.g., [10], are theoretically analyzed. In Section IV, the
performances of the demapping algorithms are investigated for
AWGN and Rayleigh fading channels using various evaluation
criteria. Finally, in Section V, the paper is concluded.
For the sake of clarity, the following notations are applied
throughout the paper. Uppercase calligraphic symbols denote
sets, e.g., . Upper case symbols in boldface denote vectors,
e.g., A. is used for the probability of a discrete event,
the probability density function (pdf) is referred to as ,
denotes the loglikelihood ratio (LLR), and refers to the a
priori information. stands for the order of computations,
() denotes the transpose of a matrix, refers to an
and Re and Im indicate the real
approximation, e.g.,
and imaginary parts of a complex number.

999

scheme with MSD is shown in Fig. 1. The information stream


is distributed to the coding levels according to the importance of the information bits. Subsequently, channel coding is
performed and each codeword , {0, 1, . . . , 1} is
interleaved by a bit interleaver . The mapper reads parallel
bits and maps them onto the signal points. Signal points are
then transmitted via a channel and received by the demapper.
A typical flat fading channel model comprises a factor as
well as an additive noise disturbance . The transmission
system given in Fig. 1 is reasonable; however, iff the MLC
scheme is designed such that the individual coding levels offer
different degrees of reliability. This holds e.g., if the capacity
or computational cutoff design rule is used but not for the
balanced distance rule where the various levels have equal
error rates, cf. [8]. In this paper, we consider only system
designs resulting in a different robustness of the individual
levels.
B. Multistage Decoding
For an MLC scheme followed by an MSD process, as given
in Fig. 1, the demapper receives the signal from the channel and accordingly outputs loglikelihood ratios (LLRs),
which can be interpreted as the soft information about a bit to
be one or zero, for decoding of a particular level . LLRs are
then deinterleaved and decoded by a forward error correction
(FEC) module. Decoded bits are fed back to the demapper by
reencoding and reinterleaving. After consecutive decoding
steps, the decoding process is completed. Obviously, the most
crucial point of the demapping process is the quality of the
soft decisions that are obtained by the demapper. An optimal
demapping algorithm is not invariant to the transmission
channel and derived for a particular channel. For an AWGN
channel with an input signal ,

II. S YSTEM OVERVIEW


A. Multilevel Coding
Multilevel coding is a general methodology to design
power and bandwidth efficient digital communication systems. In an MLC scheme, the signal set is divided into
some subsets adopting a set partitioning technique. Mathematically, = 2 , 1 signal elements ,
{0, 1, . . . , 1} are defined over a dimensional complex
constellation . There are information sequences
U = (u0 , u1 , . . . , ul1 ), where each component sequence
has a length for level , and the corresponding
codeword sequences Ck = (ck0 , ck1 , . . . , ckl1 ) with component
sequences , {0, 1, . . . , 1} each having a length
. A mapping process maps code bits to the constellation
points, Ck am . Using multilevel coded modulation, each
equivalent channel can be protected unequally and optimally.
Unequal error protection can be provided for different levels in
two different ways. The first way relies on the set partitioning
strategy to divide the signal set into several subsets, whereas
the second way is based on the assignment of proper code
rates to the component codes. In [8], it is proven that the
capacity of an ary digital modulation scheme is equal to
the sum of the capacities of the equivalent channels at the
individual coding levels of an MLC scheme. A typical MLC

= + ,

(1)

the output symbol contains additional complex white


Gaussian noise with twosided power spectral density 0
( = 1 in Fig. 1). By definition, the conditional LLR of a
code bit based on the observation of a received signal
is given as
( ) = log

P (ckq = 1yk )
.
P (ckq = 0yk )

(2)

Applying Bayes rule (2) can be written as


M
1
m=0
( ) = log M
1

m=0

where

q
p(yk am )bm
q
p(yk am )bm

(
) {
1 , 2 , 1
=

q1

n=0
q1

n=0

1
2

(
)
q1
n , bn ,
Sa bm
m ck
(
)
q1
n , bn ,
Sa bm
c
m k
(3)
if 1
=0

if 1
=1

(4)

is the feedback function which carries the a priori information


to the latter levels, 1
denotes the reencoded and inter
leaved hard decision of the channel decoder for demapping of

1000

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 4, APRIL 2011

Fig. 1: Multilevel coding and multistage decoding system.


1

at level 1, " denotes the logical not operation, and


stands for the bits of signal numbers which are defined as
=

21 .

0.8
0.6

(5)

=0

LLR calculation given by (3), which involves the calculation


of 2 distances per level via eliminating the constellation
symbols which contradict with the a-priori information, is optimal for the individual stages of MSD. The complexity of (3)
mainly depends on the computations of the conditional pdfs.
The factors of conditional pdfs in (3) (all other terms beside
the conditional pdfs either in numerator or in denominator) can
be simply calculated by 1 comparisons per each received
signal . Therefore, they contribute very little to the overall
complexity of demapping. Excluding the feedback functions,
the complexity decreases exponentially with increasing .
This means that the complexity of the first level dominates.
Therefore, any attempt to reduce the complexity of (3) should
consider the first level. For the first level of MSD, when the
conditional pdfs ( ) for the AWGN channel
( ) =

1 2 2
2

2 2

(
)
0 = log m=0
M
1
2
m=0

e
e

1 y 2
am
k
22

0 y 2
am
k

0.2
0
0.2
0.4
0.6
0.8
1
1

0.5

0
Real

0.5

Fig. 2: Representative points for level 0 (square points) and


level 1 (star points) of a block labeled 8-PSK constellation.

(6)

are inserted into (3), where 2 is the variance of the real


and imaginary part of the noise, respectively, optimal LLR
calculation is simplified to
M
2

Imaginary

0.4

(7)

22

where 0 0 and 1 1 are considered as the


signal points of the subsets after the first set partitioning step.
The two subsets, 0 and 1 , are extracted from and each
have /2 elements. A general description of set partitioning
strategies is presented in the following section.
C. Signal Set Partitioning
A signal set which has points in real dimensions
0
1
can be partitioned into two (subsets
) and each having

/2 signal points, where /2 different partitionings are


possible. Similarly 0 and 1 can be partitioned into 00 ,

(
)
01 , 10 and 11 with 4 /2
possible combinations. It/4
erating for 1 times, a dimensional constellation can
be partitioned into levels. Among various set partitioning
techniques, Ungerbck labeling (UL) [2] and block labeling
(BL) [8] are the most common set partitioning strategies.
In [11], it is shown that BL is more suitable for mobile
radio and broadcasting services. Accordingly, we adopt BL
for all derivations and simulations throughout the paper. As
an example to block partitioning, we consider an 8ary phase
shift keying (8PSK) constellation given in Fig. 2. Upper
half constellation points and lower half constellation points
which are linked with solid lines correspond to the elements
of 0 and 1 , respectively. Similarly, pairs of constellation
points which are linked with dotted lines correspond to 00 ,
01 , 10 , 11 starting from (-1,0) and proceeding in clockwise direction. For the third level, there are eight subsets:
000 ,. . . ,111 whose elements are neighboring constellation
points.

GL et al.: LOW COMPLEXITY DEMAPPING ALGORITHMS FOR MULTILEVEL CODES

III. L OW C OMPLEXITY D EMAPPING A LGORITHMS

corresponds to the log separation approximation (LSA).

High data rate digital communication systems can be


designed employing higher order modulation schemes. An
increase in the modulation order leads to an improvement
of the spectral efficiency as well as to an increase in the
computational complexity. Usually, the computational power
at the receiver side is limited for mobile radio systems. Low
complexity solutions are preferable because they consume less
power and hence provide longer battery life.
In [8], it is noted that there is a small difference in the
performance of optimum overall MLSE compared to MSD
with individual MLSE in each level. This small difference is
due to the fact that overall MLSE takes information from the
higher levels into account while MSD with individual MLSE
considers only the a priori information from the previous
levels. In spite of this small difference in performance, computational complexity is reduced considerably by MSD. However, the complexity of MSD with individual MLSE is still of
order (2 ). In the following, three different approaches for
complexity reduction are presented.

For (9), it is clear that we can get rid of the logarithm,


division, and the exponentials which have to be invoked in
true LLR calculation. However, the complexity of (9) is still
of order (2 ). The proof of Prop. III.1 shows that this
complexity can be further reduced to a constant complexity
as in (8).
Proposition III.3. For the first level of the two distance
calculations needed in (8) can be done with at most two
multiplications and one addition (neglecting the scaling factor
2
2 ).
Proof: Exploiting the symmetry 00 = 10 from (23) and
(24), 00 = + and 10 = can be written as the
representative points of 00 and 10 , respectively. Hence, (8)
can be rewritten as

1
2

0
2

+
2 2
2 2

(10)

which is equivalent to

A. Log Separation Algorithm


For any digital modulation scheme with constellation
points and dimensions, block labeled constellation points
have 2+1 , {0, 1, . . . , 1} distinct clusters at each
level and there exists a unique representative point
for each cluster which minimizes the error power =
2
1

between the cluster points and the


representative point where {0, 1, . . . , 2+1 1}.
Obviously, this point is the center of mass point of the related
(+1)
cluster and can be computed as = 2
. In
Fig. 2, representative points for the first and the second level
of an 8PSK constellation are shown. Square and star points
correspond to the representative points of level 0 and level 1,
respectively, and the circles correspond to constellation points.
For the last level, constellation points and representative points
are identical. In Fig. 2, solid lines and dotted lines show the
distances of the constellation points from the corresponding
representative points of level 0 and level 1, respectively.
Proposition III.1. For any received signal and a block
labeled symmetric constellation , {1, 2},
an exchange of the logarithm and summation terms in (7)
corresponds to two distance calculations to the representative
1
.
points with a scaling factor 2(1)
A proof of Prop. III.1 is given in Appendix A.
Lemma III.2. From the proof of Prop. III.1, we obtain an
approximate LLR by two distance calculations as

1 1

1 0

(1)

(1)

2
2

( )

=
+
2 2
2 2
2(1)
(8)
where
( (
)
(
))

am1 yk 2
am0 yk 2

0
) =
22
22
(
log e
log e

1001

(9)

1
(( Re{yk })2 ( + Im{yk })2 +
2 2
+ ( + Re{yk })2 + ( Im{yk })2 )
2
= 2 (Re{yk } Im{yk })

)
2 (
= 2 Re{r10 }Re{yk } + Im{r10 }Im{yk }

(11)

In order to calculate approximate LLRs using representative


points, it is possible to resort to two distance calculations for
each level independent of the modulation order. Although the
number of representative points increases with the level, effectively only two distances are needed since a priori information
eliminates the symbols which contradict with the decisions
from the previous levels. Therefore, the proof of Prop. III.3
for the first level can be transfered to all other levels. Hence,
without loss of generality, throughout the paper, only the first
level can be considered to determine the complexity of the
demapping algorithms.
The approximation in Lemma III.2 turns into equality iff
2
2
1/21 2
11
10
=
=
.
.
.
=
= 1 and
2
2
2
2
22
0/21 2
01 2
00 2
= 22
= ... =
= 2 where 1
22
22
and 2 are two arbitrary constants. Here, the only difference
between (8) and the LSA is a scaling factor. Furthermore, since
we assume that any scaling factor is negligible for a channel
decoder, representative points approach and log separation
algorithm are equivalent. Lemma III.2 reveals that the LSA
should deliver a good approximation of the true LLR when all
the distances are almost equal or, in other words, if there is no
dominant distance term. The distance terms are approximately
the same for low signaltonoise ratio (SNR) because then
typically >> 1 , >> 0 , . This indicates that LSA
has a noise power dependent performance.

1002

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 4, APRIL 2011

B. Jacobian Maximum Algorithm

(0 )

m=0

= log M
1
2

0.4
0.2
0
0.2

0.6

am1 yk 2
22

0 y 2
am
k

0.6

0.4

Applying (12) to (7) finally we derive


M
2

0.8

Imaginary

Another solution for the reduction of the complexity of the


computation of the logarithm of the sum of exponentials is
known from the literature and often referred to as maxlog
algorithm which is a special case of the Jacobian logarithm,
2
as the scaled squared
e.g., [10]. Defining =
22
Euclidean distance between the received symbol and the
constellation point , the maxlog algorithm is defined as
(
)

m
log
e
(12)
max (m ) .

(13)

22

(m=0
(

0
2
max
max
.

2 2
2 2
(14)
According to (14), we consider only the two nearest constellation points, which are elements of 0 and 1 , respectively to
the received signal. The advantage of (14) is that we need to
calculate neither exponentials nor the logarithm. However, the
direct way of calculating (14) requires still 2 squared distance
calculations, and this complexity increases exponentially with
the modulation order . In Prop. III.4, we introduce a low
complexity algorithm to calculate the LLR given in (14).
Proposition III.4. For any MPSK signal constellation with
2, the right hand side of the approximation in (14) can
be calculated with at most one division, two multiplications,
two additions, and comparisons.
The proof is not shown in details and uses the fact that


2
1
0
+
(15)

2 2
2 2
can be calculated with at most two multiplications, and two
additions for any arbitrary integers 0 and 1 . This follows
from rewriting (15) as


2
1
0

+
=
2 2
2 2
1
=
(2 + Re{yk }K3 + Im{yk }K4 ) ,
(16)
2 2
where 2 , 3 , and 4 are constant terms. Also, the maximum
operations in (14) can be calculated with one division and
comparisons determining the region of the constellation
point to which the received signal belongs using the decision
lines and the angle of iteratively (see Fig. 3). For M
APSK modulation systems, in addition to the calculations
for MPSK, the amplitude of needs to be calculated to
decide to which ring the received signal point belongs. After
the decision for the amplitude, the number of comparisons
necessary for the decision of the phase can be at most 1.
Notice that the maxlog approximation is precise when
< for some where is a small positive number.

0.8
1
1

0.5

0
Real

0.5

Fig. 3: A 16-PSK constellation with decision lines for the


first quadrant.

This condition is usually satisfied for good channel conditions


(high SNRs).
C. Mixed Algorithm
The basic idea behind the mixed algorithm can be expressed
as follows: if LSA is a good approximation for bad channel
conditions and maxlog algorithm is a good approximation
for good channel conditions, can we combine both so that
the resulting algorithm outperforms both single methods?. A
simple selector can choose the LLRs of either LSA or maxlog
algorithm depending on the channel conditions. It is obvious
that the outcome will outperform either of both methods.
However, this algorithm cannot perform better than the best
of both methods for each SNR. A combiner is desirable
with the aim of minimizing the error between a possible
combination of LLRs of LSA and maxlog algorithm and the
LLRs of optimum algorithm (3) so that the error variance of a
simple selector can be further decreased. For a particular noise
variance 2 , let LLSA be an 1 LLR vector calculated
by LSA using (9) and Lmaxlog be an 1 LLR vector
calculated by maxlog algorithm using (14) and Lopt be an
1 LLR vector of the optimum LLR calculation using (3).
Then, our objective is to find the optimum coefficient in
= aLmaxlog + (1 a)LLSA
L

(17)

such that

2

2

Lopt L
= (Lopt LLSA ) a (Lmaxlog LLSA )
(18)
is minimized. Denoting L1 = Lopt LLSA and L2 =
Lmaxlog LLSA , the optimum coefficients are known
from linear leastsquares solution [12] of (18) as
)1 ( T )
(
L1 L2 .
(19)
= LT
2 L2

GL et al.: LOW COMPLEXITY DEMAPPING ALGORITHMS FOR MULTILEVEL CODES

1003

TABLE I: Complexity and Memory Requirements of Low


Complexity Demapping Algorithms
Method

Complexity (2 ary modulation)

Memory Req.

MLSE

2 2 multiplications, 3 2 2 additions,
2 exponentials, 1 logarithm and 1 division
Two multiplications and one addition
One division, two multiplications,
two additions, and l comparisons
One division, six multiplications,
four additions, and l comparisons

Negligible

LSA
Max-Log
Mixed

Fig. 4: A possible selector-combiner scheme.

For any MPSK modulation scheme with 1, the LLRs


of the mixed algorithm can be calculated with at most one
division, six multiplications, four additions and comparisons.
( )
This holds because, assuming that the noise variance 2
dependent parameter (and 1 ) has already been calculated and is known, the complexity of the mixed algorithm
corresponds to the addition of the complexities of LSA and
maxlog algorithm with an extra combination cost. A linear
combination with one parameter has a complexity of two
multiplications and one addition as in (17). The complexity of
LSA from (9) is two multiplications and one addition and the
complexity of maxlog algorithm from (14) is one division,
two multiplications, two additions, and comparisons.
It is necessary to note that constant factors which
)
( are
neglected from the complexity
calculations
for
LSA
2/ 2
( ( 2 ))
and maxlog algorithm 1/ 2
are different from each
other. In (17), this effect can be canceled by including a
scaling factor of 1/4 in addition to 1 for LSA. The
accuracy of the LLRs of the mixed algorithm is at least as
good as that of the LLRs of either LSA or maxlog algorithm.
Here perfect channel knowledge has been assumed. For very
low SNR conditions or very high SNR conditions, either LSA
or maxlog algorithm can be preferred due to their reduced
computational complexity compared to the mixed algorithm.
Therefore, a selector and a combiner can be used together
for both best performance and least complexity. A possible
selectorcombiner scheme is given in Fig. 4.
D. Overview of the Complexity and Memory Requirements
For the sake of completeness, memory requirements of
LSA, maxlog, and mixed algorithms are evaluated. The LSA
should save 2 2 representative points before the demapping
process begins. If (16) is used to calculate the maxlog approximation, 32 real constants should be saved. Additionally,
2 1 slopes of decision lines have to be saved. If memory
is limited, (16) can be calculated without using memory. For
the mixed algorithm, obviously the memory requirement is
the addition of the memory requirements of LSA and max
log algorithms. Besides, for the SNR range of interest, the
mixed algorithm should save and 1 parameters
with a desired SNR spacing. Reasonably, the precision of the
SNR spacing should be in compliance with the precision of
the channel noise estimator. Even for higher order modulation

(2 )
(2 )
(2 )

formats, the aforementioned memory requirements are not


significant and in the same order as the requirements for
saving the constellation. It is also interesting that for LSA
and maxlog algorithm, exponentially increasing complexity
of the computation of true LLRs is reduced to a constant complexity (neglecting comparisons) with a cost of exponentially
increasing memory. Note that without using Prop. III.1 and
Prop. III.4, the complexities of LSA and maxlog algorithms
are of (2 ). A brief summary of the complexity and rough
memory requirements (neglecting the memory necessary to
store the signal points) of the proposed low complexity
demapping algorithms for MPSK modulation is given in
Table I. LSA offers the lowest complexity independently of
the modulation order, while the complexity of the maxlog
algorithm and mixed algorithm increase with the modulation
order. The MLSE complexity is much higher than that of the
suboptimal algorithms.
IV. E XPERIMENTAL R ESULTS
A. LLR Quality Assessment
For an uncoded transmission system, overall system performance depends on the modulation/demodulation units and
the channel. Similar to channel coding and channel decoding,
mapping from a set of bits to the signal points is unique
whereas the demapping process can vary according to the
system requirements or constraints. A low complexity demapper is usually acceptable if its performance loss compared
to the optimum demapping algorithm is limited. We consider
the error variance between the LLRs of the MLSE algorithm
and the proposed low complexity demapping algorithms for a
transmission with no channel coding. 106 bits per level are
transmitted using 32APSK modulation for an SNR range
from -5 dB to 20 dB with 1 dB spacing. Optimum parameters
for the mixed algorithm are determined for the same range
of SNR and for the same spacing before the transmission.
Figure 5 illustrates the error variances of the low complexity
demapping algorithms for a certain range of SNRs at level
= 0. It can be seen that the error variance of LSA steeply
increases with the increase of the SNR and that of the mixed
algorithm is the lowest for all SNRs. We note that similar
results have also been obtained for all other levels (not shown
here).
B. Transmission over AWGN Channel
In our first scenario, a 32APSK transmission with an
average /0 ( : average receive energy per bit) of 0

1004

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 4, APRIL 2011

0.45

10

LSA
Max Log
Mixed

10
BER

0.4
0.35
0.3

10

MLSE
LSA
Max Log
Mixed

Error Variance

10

0.25

2.5

2.45

2.4

2.35

2.3

2.25 2.2
SNR (dB)

2.15

2.1

2.05

4.05

4.1

4.15
4.2
SNR (dB)

4.25

4.3

4.35

4.4

12.6

12.8

13
13.2
SNR (dB)

13.4

13.6

13.8

14

10

0.2

0.15
BER

10

0.1

MLSE
LSA
Max Log
Mixed

10

0.05
6

10

0
5

10

15

20

SNR (dB)

3.95

10

MLSE
LSA
Max Log
Mixed

10

10

dB over an AWGN channel is considered. The same code


rate is assigned to each level. Information bits are encoded
block by block using 3GPP2 Turbo codes [13] with a block
size of 12282 bits. Three different code rates 1/5, 1/2, and
6/7 are employed. At the receiver side, the demapping has
been carried out with MLSE and low complexity demapping
algorithms, respectively. The iterative decoding process in the
Turbo decoder is repeated for 8 times. LLRs are quantized in
fixed point using 12 bits. The last level of 32APSK is not
considered in the experiments since the decoding methods of
the low complexity demapping algorithms become equivalent
to the optimum one. For a transmission of roughly 5 107 bits
per level, we calculated the BERs at the decoder output for all
levels.Figure 6 illustrates the BERs for the first level of 32
APSK with the aforementioned code rates. For code rate 1/5
(relevant SNR range: low SNRs), only the maxlog algorithm
introduces a noticeable SNR gap which is about 0.2 dB at
BER=104 . For code rate 1/2 (relevant SNR range: medium
SNRs) and for the other code rates, the mixed algorithm
is the best low complexity approximation which introduces
almost no SNR gap. For code rate 6/7 (relevant SNR range:
high SNRs), we observe that LSA introduces a noteworthy
performance loss compared to MLSE. This result is also in
agreement with the error variance analysis in Section IV-A
(for high SNRs, LSA has a high error variance). For the other
levels of 32APSK modulation, similar results are obtained.
Generally, the following interesting observations can be summarized from the simulations. First, LSA has a quite good
performance for low SNRs and maxlog algorithm is well
suited for high SNRs. Second, even if both LSA and max
log algorithms introduce some performance loss, the mixed
algorithm can still have a performance close to that of MLSE
(please cf. Fig. 6 (middle)). Third, the steeply increasing error
variance of LSA does not create a problem for the mixed

10
BER

Fig. 5: Error variance results for the optimum and the low
complexity demapping algorithms for level = 0 of MLC
scheme combined with 32-APSK modulation.

3.9

10

12

12.2

12.4

Fig. 6: BER for AWGN channel for level 0 with code rates
1/5 (top), 1/2 (middle), and 6/7 (bottom) of MLC scheme
combined with 32-APSK modulation.

algorithm. Fourth, we can claim that the mixed algorithm is


the best low complexity demapping algorithm and introduces
an SNR gap of not more than about 0.05 dB at BER=104
for all code rates and levels. During the MSD process, the
probability of error propagation is more pronounced. Error
propagation between the levels might appear if two levels
have about the same operation point; that means both levels
are almost errorfree for the same SNR region. This can be
overcome by separating the levels choosing appropriate code
rates, i.e., the previous level has to be assigned a lower code
rate than the next one. However, this is usually not necessary.
Typically, previous levels may have higher code rates than the
following ones since they exhibit higher capacity when BL
is used. The gap between the levels depends on the behavior
of the channel decoder, especially in the error floor region.
Channel decoders with a cliff BER behavior are desirable,
for example, the 3GPP2 turbo code performance falls from
BER=101 to BER=107 within 0.5 dB for code rates lower
than 1/2. The operation points of the levels can be designed
such that each lower level has a minimum required SNR which
is lower than the required SNRs of the higher levels with a
certain margin.
C. Transmission over Rayleigh Fading Channel
For the Rayleigh fading channel, the received signal is given
by
(20)
= +

GL et al.: LOW COMPLEXITY DEMAPPING ALGORITHMS FOR MULTILEVEL CODES

parameter of the mixed algorithm is varied between 0 and


1. The performance of the demapper is then evaluated with a
criterion such as BER. Accordingly, the parameter value which
gives the lowest BER is selected as the optimum choice.

10

BER

10

MLSE
LSA
Max Log
Mixed

10

10
1.5

1.45

1.4

V. C ONCLUSION
1.35

1.3

1.25 1.2
SNR (dB)

1.15

1.1

1.05

6.4

6.5

6.6
6.7
SNR (dB)

6.8

6.9

7.1

17.4

17.6

10

BER

10

MLSE
LSA
Max Log
Mixed

10

10

6.1

6.2

6.3

10

BER

10

10

10

10

16

1005

MLSE
LSA
Max Log
Mixed
16.2

16.4

16.6

16.8
SNR (dB)

17

17.2

Fig. 7: BER for Rayleigh fading channel for level 0 with


code rates 1/5 (top), 1/2 (middle), and 6/7 (bottom) of MLC
scheme combined with 32-APSK modulation.

where and are a zeromean complex Gaussian factor


and AWGN, respectively. Similar tests have been performed
for a fast Rayleigh fading channel with statistically independent coefficients for different time steps as for the AWGN
channel in Section IV-B. 5 107 bits per level are transmitted
with code rates 1/5, 1/2, and 6/7 using 32APSK modulation.
At the receiver side perfect channel knowledge is assumed.
At the output of the decoder, we determined BERs for all
levels. Figure 7 illustrates the BER curves for level 0 of
the MLC scheme. For low SNRs as in Fig. 7 (top), the
SNR gap between the MLSE algorithm and the proposed low
complexity demapping algorithms is more obvious compared
to the results for the AWGN channel and for the same code
rate. For a code rate of 1/2 (Fig. 7 (middle)), we observe
the highest degradation of the mixed algorithm. Besides, LSA
starts to introduce noticeable SNR gaps for code rates of 1/2
and higher. For a code rate of 6/7, maxlog algorithm and
mixed algorithm perform quite well. As a final remark, we
note that in none of the simulations the mixed algorithm has
introduced an SNR gap which exceeds 0.1 dB.
For the sake of completeness, it is worth noting that for
an unknown channel, the parameters of the mixed algorithm
can be determined by simulations so that an optimum performance can be achieved. The concept for determination of
the optimum parameters for an unknown channel is slightly
different from that of a known channel. A training sequence
should be transmitted periodically, and at the receiver side, the

In this paper, three different demapping algorithms for


decreasing the computational complexity of MLSE have been
considered for an MLC scheme combined with MSD and
BL. It has been shown that all proposed low complexity
demapping algorithms as well as the maxlog demapping
algorithm can be implemented with a constant complexity
(neglecting comparisons) per received symbol independent
of the modulation order. To analyze the tradeoff between the
complexity and the performance of the proposed algorithms,
in the first step, a transmission over an AWGN channel has
been considered without channel coding. For this case, the
variance of the error between LLRs of the MLSE algorithm
and those of the proposed demapping algorithms has been
evaluated. In the second step, the performance of the proposed
low complexity demapping algorithms have been tested for a
transmission over both the AWGN and the Rayleigh fading
channel using Turbo codes with different code rates as the
component codes of the considered multilevel coding scheme
employing a block partitioned 32APSK constellation. The
simulations have shown that the best algorithm among all
considered low complexity demapping algorithms exhibits
an SNR gap of at most 0.05 dB for the AWGN channel
and 0.1 dB for the Rayleigh fading channel, respectively at
BER=104 .
A natural extension of this work is the use of different
constellations, including other regular modulations such as
QAM, or even irregular mappings such as geometrical shapes.
For any kind of constellation, the optimum coefficients of the
mixed algorithm strongly depend on the channel estimation
process because they are a function of the received SNR.
In practical systems, where the channel estimation module
is not ideal, the mixed algorithm can select nonoptimal
coefficients. This leads to a performance degradation that
could be investigated in future work for different timevarying
channels.
R EFERENCES
[1] G. Ungerboeck and I. Csajka, On improving data-link performance by
increasing channel alphabet and introducing sequence coding," in Proc.
IEEE Int. Symp. Inf. Theory, June 1976.
[2] G. Ungerboeck, Channel coding with multilevel/phase signals," IEEE
Trans. Inf. Theory, vol. 28, pp. 55-67, Jan. 1982.
[3] H. Imai and S. Hirakawa, A new multilevel coding method using error
correcting codes," IEEE Trans. Inf. Theory, vol. 23, pp. 371-377, May
1977.
[4] G. J. Pottie and D. P. Taylor, Multilevel codes based on partitioning,"
IEEE Trans. Inf. Theory, vol. 35, pp. 87-98, Jan. 1989.
[5] A. R. Calderbank, Multilevel codes and multistage decoding," IEEE
Trans. Commun., vol. 37, pp. 222-229, Mar. 1989.
[6] T. Cover, Broadcast channels," IEEE Trans. Inf. Theory, vol. 18, pp.
2-14, Jan. 1972.
[7] J. Huber and U. Wachsmann, Capacities of equivalent channels in
multilevel coding schemes," IEEE Electron. Lett., vol. 30, pp. 557-558,
Mar. 1994.
[8] J. Huber, Multilevel-codes: distance profiles and channel capacity,"
Proc. ITG-Fachberichte 130, Oct. 1994, pp. 305-319.

1006

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 4, APRIL 2011

[9] A. Vargas, W. Gerstacker, and M. Breiling, Design and evaluation of


a multilevel decoder for satellite communications," in Proc. IEEE Int.
Conf. Commun., June 2009.
[10] G. Bauch and V. Franz, A comparison of soft-in/soft-out algorithms
for turbo detection," in Proc. Int. Conf. Telecommun., June 1998, pp.
259-263.
[11] U. Wachsmann, R. Fischer, and J. B. Huber, Multilevel codes: theoretical concepts and practical design rules," IEEE Trans. Inf. Theory, vol.
45, pp. 1361-1391, July 1999.
[12] C. L. Lawson, R. J. Hanson, Solving Least Squares Problems. PrenticeHall, 1974.
[13] C. Heegard and S. B. Wicker, Physical Layer Standard for CDMA2000
Spread Spectrum Systems, 3GPP2 Std. C.S0002-C, 2004.
[14] J. G. Proakis, Digital Communications. McGraw-Hill, 2001.
Gkhan Gl was born in Amasya, Turkey, in 1982.
He received the B.Sc. degree (with first class honor)
in electronic engineering from Uludag University,
Bursa, Turkey in 2005 and the M.Sc. degree in
Digital Communications from Christian Albrechts
University, Kiel, Germany in 2009. He wrote his
master thesis and worked as a research assistant
in Fraunhofer Institute (IIS), Erlangen. His research
interests include pattern recognition, steganography
and steganalysis.
Aharon Vargas was born in Tenerife, Spain, in
1979. He received the Dipl.-Ing. degree in telecommunication engineering from the University of Las
Palmas, Gran Canaria, Spain in 2003.
Since 2003, he has worked as system designer
in the Fraunhofer-Institut fr Integrierte Schaltungen
IIS in Erlangen, Germany. Since 2006, he is doing
his PhD at the University of Erlangen-Nuremberg
as an external student. His current research interests
include broadcasting systems, satellite communications, multilevel coding, and turbo codes.
Dr. Breiling works as a Senior Engineer at
the Fraunhofer Institute for Integrated Circuits.
He studied at the Technical University of
Karlsruhe/Germany, and within a pan-European
study programme at the Norwegian Institute
of Technology (NTH) Trondheim/Norway, the
Ecole Suprieure dIngnieurs en Electronique et
Electrotechnique (ESIEE) in Paris/France, and the
University of Southampton/England. He graduated
with a Diplom-Ingenieur degree (equiv. to Master)
in Electrical Engineering from the University of
Karlsruhe with a diploma thesis on Turbo Coding in 1997. Subsequently,
Dr. Breiling joined the research group in digital communications of Prof.
Huber at the University of Erlangen/Germany, where he conducted research
on channel coding techniques, diversity reception, crest factor reduction and
other topics in digital communications. He graduated with a Doktor-Ingenieur
degree (summa cum laude) in Digital Communications with a PhD thesis on
Interleaver Design for Turbo Codes in 2002. In December 2001, Dr. Breiling
started working for the Fraunhofer Institute for Integrated Circuits (IIS) in
Erlangen, where he is a member of the Communications Systems Design
group.

Wolfgang H. Gerstacker was born in Nuremberg,


Germany, in 1966. He received the Dipl.-Ing. degree
in electrical engineering, the Dr.-Ing. degree for
a thesis on equalization concepts for fast digital
transmission over twisted pair lines, and the Habilitation degree (Venia Legendi) for work on digital
transmission concepts for the EDGE mobile communications system from the University of Erlangen
Nuremberg, Erlangen, Germany, in 1991, 1998, and
2004, respectively.
From 1998 to 2002, he was a Consultant for
mobile communications and an External Lecturer in digital communications
at the University of ErlangenNuremberg. From 1999 to 2000, he was a
Postdoctoral Research Fellow for a six-month period at the University of
Canterbury, Christchurch, New Zealand, sponsored by a fellowship from the
German Academic Exchange Service (DAAD). Since 2002, he has been a
Senior Researcher and Lecturer at the Chair of Mobile Communications of
the University of ErlangenNuremberg. His current research interests include
wireless communications, detection, equalization and parameter estimation,
blind techniques, space-time processing, OFDM, MIMO systems, LTE and
WLANs.
In 2001, he was a co-recipient of the Research Award of the German Society
for Information Technology (ITG). For work on single antenna interference
cancellation for GSM, he was a co-recipient of the EEEfCOM Innovation
Award 2003 and of the Vodafone Innovation Award 2004. He is a Member
of the Editorial Boards of EURASIP Journal on Wireless Communications
and Networking and Elsevier PHYCOM; he has served as a TPC Member
for various conferences such as ICC, Globecom, EUSIPCO, and WCNC.

GL et al.: LOW COMPLEXITY DEMAPPING ALGORITHMS FOR MULTILEVEL CODES

1007

A PPENDIX A
T HE PROOF OF P ROPOSITION III.1
Proof: Exchange of the summation and the logarithm results in
M 1
2
am1 yk
2

( (
)
(
))
2
2
e


am1 yk 2
am0 yk 2
( 0 )
m=0

22
22
=log M
log e
log e

1
2
a 0 y 2
m
m22k
e
m=0
(

1
2

=
2 2
2 2

(
Re2 {a 1 yk } Im2 {a 1 yk } + Re2 {a 0 yk } + Im2 {a 0 yk } )
m
m
m
m
=
2 2

( Re2 {a 1 } Im2 {a 1 } + Re2 {a 0 } + Im2 {a 0 } )


m
m
m
m
=
2
2

( Re{a 1 }Re{yk } + Im{a 1 }Im{yk } Re{a 0 }Re{yk } Im{a 0 }Im{yk } )


m
m
m
m
+
2

(21)
Since the constellation is symmetric and block labeled, 0 0 there is a symmetric point 1 1 with respect to the real
and imaginary axis. For instance, for = 0, if 00 = 0 then there exists a point 10 = + 1 . According
to this,
( Re2 {a 1 } Im2 {a 1 } + Re2 {a 0 } + Im2 {a 0 } )
m
m
m
m
= 0,
(22)
2
2

1
Re{am
}+

1
Im{am
}+

0
Re{am
} = 0,

(23)

0
Im{am
}=0

(24)

0 ) (right hand side of (21)) is simplified to


is valid and (

0 ) =
(

Re{a 1 }Re{yk } + Im{a 1 }Im{yk } Re{a 0 }Re{yk } Im{a 0 }Im{yk }


m
m
m
m
.
2

From (23) and (24) it is obvious that


(
)2 (
)2 (
)2 (
)2

1
0
1
0

Re{am } +
Re{am }
Im{am } +
Im{am }
1

= 0,

2 2
2(1)

(25)

(26)

0 ) which results in
and this term can be added to (
0 ) =
(

( Re{a 1 }Re{yk } + Im{a 1 }Im{yk } Re{a 0 }Re{yk } Im{a 0 }Im{yk } )


m
m
m
m
2

(
)2 (
)2 (
)2 (
)2

1
0
1
0
Re{am } +
Re{am }
Im{am } +
Im{am }

+ (1)
.
2 2

2
(27)

1008

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 4, APRIL 2011

0 ) with 2(1) and adding another zero term yields


Scaling (

(
)
1
0
0
0 )
(
}Re{yk } + Im{a1m }Im{yk } Re{am
}Re{yk } Im{am
}Im{yk }
1 Re{am
= (1)
2
2(1)
2

(
)2 (
)2 (
)2 (
)2

1
0
1
0

Re{am } +
Re{am }
Im{am } +
Im{am }
1

+ 2(1)

2 2

2
(
+

Re2 {yk } + Re2 {yk } + Im2 {yk } Im2 {yk }


2 2

)
,
(28)

and (28) can be rewritten as

0 )
(
=
2(1)
+

2(1)

1
2(1)

corresponding to

1
Re{am
}

)2
Re{yk }

2(1)

1
Im{am
}

)2
Im{yk }

2 2
)2 (
)2

1
0
0
Re{am } Re{yk } + 2(1)
Im{am } Im{yk }

2 2

1 1

(1)

2 2

,
(29)

1 0

(1)

2 2

(30)

Das könnte Ihnen auch gefallen