Iterative Soft Decision Based Complex K-best MIMO Decoder

Mehnaz Rahman & Gwan S. Choi
Signal Processing: An International Journal (SPIJ), Volume (9) : Issue (5) : 2015 54
Iterative Soft Decision Based Complex K-best MIMO Decoder
Mehnaz Rahman mehnaz@tamu.edu
Department of ECE
Texas A&M University
College Station, Tx- 77840, USA
Gwan S. Choi gchoi@ece.tamu.edu
Department of ECE
Texas A&M University
College Station, Tx- 77840, USA
Abstract
This paper presents an iterative soft decision based complex multiple input multiple output
(MIMO) decoding algorithm, which reduces the complexity of Maximum Likelihood (ML) detector.
We develop a novel iterative complex K-best decoder exploiting the techniques of lattice
reduction for 8 × 8 MIMO. Besides list size, a new adjustable variable has been introduced in
order to control the on-demand child expansion. Following this method, we obtain 6.9 to 8.0 dB
improvement over real domain K-best decoder and 1.4 to 2.5 dB better performance compared to
iterative conventional complex decoder for 4th iteration and 64-QAM modulation scheme. We
also demonstrate the significance of new parameter on bit error rate. The proposed decoder not
only increases the performance, but also reduces the computational complexity to a certain level.
Keywords: Complex K-best Algorithm, MIMO, Lattice Reduction, Iterative Soft Decoding, SE
Enumeration.
1. INTRODUCTION
With the advancement of wireless system, MIMO has been acclaimed by different wireless
standards such as IEEE 802.11n, IEEE 802.16e to achieve high data rates and performance with
ML or near-ML algorithms. Most of these standards have a specified minimum error rate to
guarantee quality of service (QoS), which is either in bit error rate (BER) or packet error rate
(PER) (e.g., 10ି଺
is specified as maximum tolerable BER according to IEEE 802.11n standard
[1]).
The main challenge behind MIMO system is maintaining the performance of the receiver with low
complexity. Several algorithms have been proposed to address the issue, offering different
tradeoffs between complexity and performance. The ML detector minimizes BER performance
through exhaustive search. However, with increased number of transmitting and receiving
antennas, and bits in modulation, the complexity grows exponentially [2, 3]. In contrast, sub-
optimal detectors with polynomial complexity such as zero forcing (ZF), minimum mean square
error (MMSE) detectors etc. have been developed with significant performance loss.
Recently, lattice reduction (LR) has been proposed in order to achieve high performance, yielding
much less complexity than the conventional K-best decoder [4, 5, 6]. LR-aided detector can
achieve the same diversity as of ML at the cost of some performance loss [7, 8]. Later, it is
implemented in complex domain [9]. All of these suboptimal detectors mentioned above were
based on hard decision. Therefore, soft input-soft output (SISO) detectors, suitable for
subsequent iterative decoding are introduced in [10].

Researchers further improved these SISO detectors with low density parity check (LDPC)
decoder [10, 11]. The output of LDPC decoder is fed back to the detector for updating the soft
value in order to achieve better performance. This is called iterative decoding. It can achieve near
Shannon performance with less computational complexity compared to other near Shannon
decoders [12].
This paper presents an iterative soft decision based complex K-best decoder, which enables the
utility of lattice reduction and complex SE enumeration in MIMO decoder. For soft decoding, the
log likelihood ratio (LLR) values for LDPC decoder are first computed from the K best candidates
and then, they are fed back to LLR update unit as inputs to the next iteration. This process of
iterations is continued until the gain of subsequent iteration becomes saturated. Then, the last
updated LLR values are forwarded to the LDPC decoder for final detection. Besides list size K, a
new tunable parameter Rlimit is introduced in order to enable adaption of the computation of on-
demand child expansion for choosing the list candidates.
We compare the results of our proposed decoder with those of iterative conventional complex
decoder in [11] and LR-aided real decoder in [13]. For 8 × 8 MIMO, it achieves 6.9 to 8.0 dB
improvement over real domain K-best decoder and 1.4 to 2.5 dB better performance comparing
to conventional complex K-best decoder for 4th iteration and 64 QAM modulation scheme. If we
consider only 1st iteration, the gain increases to more than 9.0 dB and 2.9 dB comparing with
iterative real and complex decoder respectively. This provides significant gain in terms of practical
execution. The effect of Rlimit is also analyzed to achieve the maximum performance. The
introduction of Rlimit also leads to complexity reduction significantly.
The rest of the paper is organized as follow. In Section II we introduce soft decision based
complex MIMO decoding algorithm. Then, Section III presents the results of our studied cases
and Section IV concludes this paper with a brief overview.
2. SYSTEM MODEL
Let us consider a MIMO system operating in M-QAM modulation scheme and having N୘ transmit
antenna and Nୖ receiving antenna as:
‫ݕ‬ = ‫ݏܪ‬ + ݊, (1)
where s = [sଵ , sଶ, …. s୒౐
]୘
is the transmitted complex vector, ‫ܪ‬ is complex channel matrix and
y = [yଵ , yଶ, …. y୒౎
]୘
is the symbol of Nୖ dimensional received complex vector. Noise is a
Nୖ dimensional complex additive white Gaussian noise (AWGN) with variance and power σଶ
and
ܰ଴ respectively. Noise can be represented by ݊ = [nଵ , nଶ, …. n୒౎
]୘
.
The detector solves for the transmitted signal by solving non-deterministic hard problem:
sො = argୱ෤∈ୗొ౐ min‖y − Hs෤‖ଶ
. (2)
Here, s෤ is the candidate complex vector, and sො is the estimated transmitted vector [8]. In the
expression, ∥. ∥ denotes 2-norm. This MIMO detection problem can be represented as the closest
point problem in [14], which is an exhaustive tree search through all the set of all possible lattice
points in ‫̃ݏ‬ ∈ S୒౐ for the global best in terms of Euclidean distance between ‫ݕ‬ and ‫̃ݏܪ‬. Each
transmit antenna offers two level of search for real-domain MIMO detection: one for real and the
other for imaginary part. However, in complex domain detection method, only one level of search
is required for each antenna.

ML detector performs a tree search through the set of all possible branches from root to node,
thereby achieves the best performance. However, its complexity increases exponentially with the
number of antennas and constellation bits. Therefore, suboptimal detectors such as LR-aided
detector comes into consideration.
2.1 LR-aided Decoder
Lattice reduction provides more orthogonal basis with short basis vector from a given integer
lattice. Hence, it effectively reduces the effects of noise and mitigates error propagation in MIMO
detection. Since lattice reduction is most effective for unconstrained boundary, the following
change is made to (2) to obtain a relaxed search.
sො = argୱ෤∈࣯ొ౐ min‖y − Hs෤‖ଶ
, (3)
where ࣯ is unconstrained complex constellation set ሼ… , −3 + j, −1 − j, −1 + j, 1 − j, … ሽ. Hence,
sො may not be a valid constellation point. This is resolved by quantizing sො = ࣫ ሺsොሻ , where ࣫ሺ. ሻ is
the symbol wise quantizer to the constellation set S.
However, this type of naive lattice reduction (NLD) does not acquire good diversity multiplexing
tradeoff (DMT) optimality. Hence, MMSE regularization is employed as proposed in [15, 16],
where the channel matrix and received vector are extended as ‫ܪ‬ഥ and ‫ݕ‬ത:
Hഥ = ൦
H
ඨ
N଴
2σଶ
ଶ I୒౐
൪, yത = ൤
y
0୒౐×ଵ
൨, (4)
where 0୒౐×ଵ is a N୘ × 1 zero matrix and I୒౐
is a N୘ × N୘ complex identity matrix [17, 18]. Then,
Eq. (3) can be represented as:
sො = argୱ෤∈࣯ొ౐ min‖yത − Hഥs෤‖ଶ
.
(5)
Hence, lattice reduction is applied to ‫ܪ‬ഥ to obtain ‫ܪ‬෩ = ‫ܪ‬ഥܶ , where ܶ is a unimodular matrix. Eq. (5)
then become:
sො = T arg min୸෤∈࣯ొ౐ ቀฮy෤ − H෩z෤ฮ
ଶ
+ ሺ1 + jሻ୒౐×ଵቁ, (6)
where y෤ = ሺyത − Hഥሺ1 + jሻ୒౐×భ
ሻ/2 is the complex received signal vector and 1ଶ୒౐×ଵ is a 2N୘ × 1
one matrix. After shifting and scaling, (6) became the following one.
sො = Tz෤ + ሺ1 + jሻ୒౐×ଵ . (7)
Lattice reduction is an NP complete problem. However, polynomial time algorithms such as
Lenstra-Lenstra-Lovasz (LLL) algorithm in [19] can find near orthogonal short basis vectors.
2.2 Complex K-Best LR-Aided MIMO Detection
Complex K-best LR-aided detection offers a breadth first tree search algorithm, which is
performed sequentially starting at ܰ௧௛-level. First, it requires QR decomposition on H෩ = QR,

Signal Processing: An International
where Q is a ሺNୖ + N୘ሻ × ሺN
triangular matrix. Then (6) is reformulated as
sො = T arg min୸෤∈࣯ొ౐
where y෬ = Q୘
y෤ . The error at each step is measured by the partial
which is an accumulated error at a given level of
selected and passed to the next lev
tree are evaluated to find the one
parent in LR-aided K-best algorithm is
children issue is addressed by calculating K best candidates using complex on
expansion.
2.3 Complex On-demand Expansion
Complex on-demand expansion ex
20]. The strategy employs expanding of a node (child) if and only if all of it
already been expanded and chosen as the partial candidates
non-decreasing error, K candidates are selected. In conventional complex SE enumeration,
expansion of a child can be of two types: Type I, where the expanded child has same imaginary
part as its parent, i.e. enumerating
example of conventional complex on
First received symbol is rounded to the nearest integer as shown in Fig. 1(a), which includes
quantizing of both real and imaginary components of the signal to the nearest integer. Type
candidate will be expanded two times along real and imaginary axis u
the two expanded nodes are considered candidates, as demonstrated in Fig. 1(b). Then, the one
with the minimum PED is chosen, and expanded for further calculation depending on the type. As
in Fig. 1(c), the chosen node is of type I
is of Type II, as shown in Fig. 1(d), it will be expanded only along imaginary axis.
The number of nodes needs expanded
complexity analysis. The worst case scenario will be if all the nodes chosen are of type I. Then, at
an arbitrary level of tree, the number of expanded nodes is bounded by
the entire tree, the complexity for the search becomes
: An International Journal (SPIJ), Volume (9) : Issue (5) : 2015
ሺNୖ + N୘ሻ orthonormal matrix and R is a ሺNୖ + N୘
) is reformulated as
൫‖y෬ − Rz෤‖ଶ
+ ሺ1 + jሻଶ୒౐×ଵ൯,
The error at each step is measured by the partial Euclidean distance (PED),
r at a given level of the tree. For each level, the K best nodes are
and passed to the next level for consideration. At the end, all the K paths through the
to find the one with minimum PED. The number of valid children for each
best algorithm is infinite. Hence, in our proposed algorithm
children issue is addressed by calculating K best candidates using complex on
Expansion
demand expansion exploits the principle of Schnorr-Euchner (SE) enumeration [
expanding of a node (child) if and only if all of its better siblings have
and chosen as the partial candidates [21, 22]. Hence, in an order of strict
decreasing error, K candidates are selected. In conventional complex SE enumeration,
part as its parent, i.e. enumerating along the real axis; and Type II for all other cases. The
example of conventional complex on-demand SE enumeration is shown in Fig. 1.
FIGURE 1: Complex SE Enumeration.
quantizing of both real and imaginary components of the signal to the nearest integer. Type
candidate will be expanded two times along real and imaginary axis using SE enumeration, and
in Fig. 1(c), the chosen node is of type I, it will be expanded to 2 more nodes. If the chosen node
is of Type II, as shown in Fig. 1(d), it will be expanded only along imaginary axis.
expanded at any level of the tree is considered as measurement of
The worst case scenario will be if all the nodes chosen are of type I. Then, at
an arbitrary level of tree, the number of expanded nodes is bounded by K + 2ሺK − 1
the entire tree, the complexity for the search becomes 3N୘K − 2N୘. Comparing with the
57
୘ሻ × N୘ upper
(8)
Euclidean distance (PED),
, the K best nodes are
paths through the
The number of valid children for each
lgorithm, the infinite
-demand child
enumeration [9,
better siblings have
. Hence, in an order of strict
decreasing error, K candidates are selected. In conventional complex SE enumeration,
ll other cases. The
quantizing of both real and imaginary components of the signal to the nearest integer. Type-I
sing SE enumeration, and
, it will be expanded to 2 more nodes. If the chosen node
considered as measurement of
The worst case scenario will be if all the nodes chosen are of type I. Then, at
1ሻ. Taken over
Comparing with the real

domain detection algorithm in [13], the number of the expanded nodes is 4N୘K − 2N୘. For
instance, with K as 4 and N୘ equal to 8, the number of expanded node is 80 and 112 considering
complex and real decoder respectively. Hence, complex SE enumeration requires less
calculation, thereby reduces hardware complexity.
In this paper, we introduce another parameter, Rlimit while performing the complex on demand
child expansion. In contrast with the conventional one, the type of a child is not considered for
further expansion. The example of improved complex SE enumeration with Rlimit as 3 is given in
Fig. 2.
FIGURE 2: Improved Complex SE Enumeration with Rlimit as 3.
As shown in Fig. 2, after rounding the received symbol to the nearest integer, first real SE
enumeration is performed to calculate Rlimit candidates. Hence, it means that, all the calculated
nodes up to Rlimit will have same imaginary value, as demonstrated in Fig. 2(b). Then, the one
with minimum PED is selected and expanded only along the imaginary axis using imaginary
domain SE enumeration. This process is continued till K nodes are selected at that level of tree
as presented in Fig. 2(c)-(d).
The complexity analysis of the improved child expansion proceeds as follows. At any level of tree
search, first KRlimit nodes need to be expanded. After that, only imaginary domain SE
enumeration will be performed. Hence, considering the worst case, the total number of nodes
calculated at each level is KRlimit + ሺK − 1ሻ. For N୘ levels, the complexity becomes
N୘K ሺRlimit + 1ሻ − N୘. Therefore, introduction of Rlimit may increase the complexity as
evidenced in result section, although offers better BER performance comparing to the
conventional one. However, comparing with the real domain detection, the total complexity is still
less. We have used improved complex on demand expansion to perform the list calculation and
then the chosen K paths are passed to the iterative soft input soft output (SISO) decoder.
2.4 Iterative Soft Decoding
LDPC decoder in [12] calculates approximate LLR from the list of possible candidates using (9).
L୉ሺx୩|Yሻ ≈
ଵ
ଶ
max
x ∈ X୩, ାଵ
ቄ−
ଵ
஢మ ∥ y − Hs ∥ଶ
+ x[୩]
୘
. L୅,[୩]ቅ −
ଵ
ଶ
max
x ∈ X୩, ିଵ
ቄ
ଵ
஢మ ∥ y − Hs ∥ଶ
+
x[୩]
୘
. L୅,[୩]ቅ,
(9)
here ‫ܠ‬[‫]ܓ‬
‫܂‬
and ‫ۺ‬‫]ܓ[,ۯ‬ are the candidates values {-1 or 1} and LLR values except k-th candidate
respectively. In order to perform the soft decoding, the LLR values are first computed at the last

layer of K-best search. Then, the soft values are fed into the iterative decoder for the subsequent
iteration. This process continues until the difference in error levels between the last two iterations
becomes negligible. Lastly, the updated LLR values are used for hard decision.
From the perspective of hardware design as proposed in [16, 23], the LLR calculation unit takes
one of the candidates at a given time and computes the LLR value. Then, the new LLR is
compared to the maximum of previous LLRs. Hence, this unit has to keep track of 2 values for
each LLR; one for those whose k-th bit of the candidate list is 1 (Lambda-ML), and the other for 0
(Lamda-ML-bar). After that, the LLR values are calculated as the subtraction of Lambda-ML and
Lambda-ML-bar divided by 2.
3. SETUP AND RESULTS
This section demonstrates the performance of the proposed iterative soft decision based complex
K-best decoder. The test and simulation environment has been implemented using IEEE 802.11n
standard. All the simulations are for 8 × 8 MIMO with different modulation schemes. The ratio of
the signal and noise power is considered as signal to noise ratio (SNR).
We first analyze the performance of four iterations of our proposed decoder for different
modulation scheme. Then, the effect of Rlimit on BER performance is shown for 64QAM
modulation scheme. Finally, we demonstrate the comparison of performance of our proposed
work with that of iterative conventional complex decoder and real decoder for 64QAM modulation
scheme.
The total number of the nodes expanded for 8 × 8 MIMO is considered as measurement of the
complexity analysis. For iterative real decoder, as shown in [13] the improvement gained from 3rd
to 4th iteration is limited and negligible for iterations beyond that. Hence, we consider BER versus
SNR curve up to four iterations in order to perform comparison among maximum performance.
3.1 Simulation and Analysis
The performance of four iterations of our proposed soft decision based complex decoder for
QPSK modulation scheme is presented in Fig. 3.
FIGURE 3: BER vs SNR curve of the first 4 iterations of iterative complex decoder for 8 x 8 MIMO system
with K as 4 and QPSK modulation scheme.

As shown in Fig. 3, for QPSK modulation with list size, K of 4 and Rlimit of 4, we observe 0.4 dB
improvement in BER due to the 2nd iteration at the BER of 10ି଺
. When we compare the
performance of 1st iteration with 3rd and 4th one, the improvement increases to 0.7 dB and 1.0
dB respectively.
(a) 16 QAM (b) 64 QAM
FIGURE 4: BER vs SNR curve of the first 4 iterations of iterative complex decoder for 8 x 8 MIMO system
with K as 4.
Next, Fig. 4 represents the performance curve of 4th iteration for 16 QAM and 64 QAM
modulation scheme. As demonstrated in Fig. 4(a), the performance of 2nd iteration is
approximately 0.4 dB better than the 1st one with K as 4 and Rlimit set to 4 for 16 QAM
modulation scheme. When increasing the iteration, the performance improves by 0.8 dB for the
3rd and 1.1 dB for the 4th iteration compared to the 1st one.
For 64QAM having same K as 16QAM, the improvement due to the 2nd iteration is 0.4 dB, as
shown in Fig. 4(b). If we then compare the 3rd and 4th iteration with respect to the 1st one, the
improvements are 0.8 dB and 1.0 dB respectively. By extensive simulation, we observe that the
performance does not improve beyond 4th iteration. Therefore, with iteration number, the
performance between i-th and (i+1)-th iteration gets saturated.
3.2 Effect of Rlimit on BER
The effect of Rlimit, as discussed in section 3.2 for proposed complex on demand child expansion
is shown in Fig. 5. It represents BER performance for the 4th iteration over different SNR
considering 8 × 8 MIMO and 64 QAM modulation scheme with list size, K as 4.

Signal Processing: An International
FIGURE 5: BER vs SNR curve of the
It is evident that if the value of Rlimit
saturates with Rlimit. On the other hand, decreasing Rlimit
Fig. 4, when Rlimit increases from 4 to 6
the Rlimit to 2 and then 1, degrades the performance 0.3 dB and 1.1
Similar curves can be obtained considering 1st, 2nd and 3rd iteration of proposed i
decoder for different Rlimit. By extensive simulation, we a
modulation schemes, Rlimit set to 4
Rlimit is increased, the performance does not improve.
3.3 Comparison of Performance
The comparison of the performance of different iteration
iterative conventional complex decoder and real decoder for 64QAM modulation scheme with K
as 4 is presented in Fig. 6.
For proposed iterative compl
performance evaluation. Simulation with Rlimit higher than 4
minimum value required to achieve the maximum performance. We consider BER versus SNR
curve up to four iterations in order to perform comparison among maximum performa
shown in [14] .
: An International Journal (SPIJ), Volume (9) : Issue (5) : 2015
BER vs SNR curve of the 4th iteration of iterative complex decoder for 8 x 8 MIMO with 64QAM
modulation scheme having K as 4.
that if the value of Rlimit is increased, the performance improves and then, it
On the other hand, decreasing Rlimit will degrade BER. Hence, as shown in
Rlimit increases from 4 to 6, the performance get saturated. However, decreasing
the Rlimit to 2 and then 1, degrades the performance 0.3 dB and 1.1 dB respectively.
Similar curves can be obtained considering 1st, 2nd and 3rd iteration of proposed i
. By extensive simulation, we also observe that, for QPSK and 16 QAM
modulation schemes, Rlimit set to 4 can obtain the maximum performance. Even if the value of
is increased, the performance does not improve.
parison of Performance
The comparison of the performance of different iterations of our proposed work with those
x decoder and real decoder for 64QAM modulation scheme with K
tive complex decoder, we have considered Rlimit as 1, 2 and 4
ce evaluation. Simulation with Rlimit higher than 4 is not considered, since it is the
to four iterations in order to perform comparison among maximum performa
61
MIMO with 64QAM
es and then, it
ce, as shown in
rated. However, decreasing
dB respectively.
Similar curves can be obtained considering 1st, 2nd and 3rd iteration of proposed iterative
lso observe that, for QPSK and 16 QAM
formance. Even if the value of
ed work with those of
x decoder and real decoder for 64QAM modulation scheme with K
ex decoder, we have considered Rlimit as 1, 2 and 4 for
is not considered, since it is the
to four iterations in order to perform comparison among maximum performance, as

(a) (b) (c)
FIGURE 6: BER vs SNR curves of proposed iterative complex decoder, conventional complex and
real decoder for 8 x 8 MIMO with 64QAM modulation scheme having K as 4. For proposed
decoder, Rlimit is set to 1, 2, and 4.
As demonstrated in Fig. 6(a), a 3.4 dB improvement in performance can be achieved comparing
the 1st iteration of proposed decoder with that of conventional iterative complex decoder with
Rlimit as 4 at the BER of 10ି଺
. When Rlimit is changed to 2 and 1, the improvements become 3
dB and 2.9 dB respectively. We also compare the performance of proposed decoder with that of
the iterative real decoder for 1st iteration [14]. As presented in Fig.6(a), 9.0 dB to 9.5 dB
improvement can be achieved using Rlimit as 1 to 4.
Next, as shown in 6(b), a 1.5 dB improvement can be obtained if we consider the performance of
1st iteration of proposed decoder with the 4th iteration of conventional complex one using Rlimit
as 4. Decreasing Rlimit to 2 and 1 results in 1 dB and 0.8 dB improvement. Comparing to the 4th
iteration of iterative real decoder, 6.1 dB to 6.8 dB SNR gain can be achieved using Rlimit set to 1
to 4 respectively.
Fig. 6(c) presents the comparison curves considering the 4th iteration of iterative decoders. As
demonstrated in the figure, a 2.4 dB improvement can be obtained using Rlimit as 4 at the BER of
10ି଺
comparing the conventional iterative complex decoder. In addition, when simulating for
Rlimit as 2 and 1, the gain becomes 2.2 dB and 1.4 dB respectively. Similar analysis can be
performed comparing to the 4th iteration of iterative real decoder. A gain of 6.9 dB to 8.0 dB can
be achieved for Rlimit set to 1 to 4.
Then, we have performed the computational complexity analysis for the presented work. The total
number of the nodes expanded for 8 x 8 MIMO is considered as measurement of the analysis.
Complexity analysis of proposed and conventional complex decoder is shown in Table 1.

K
Proposed Conv. Complex Proposed vs Conv. (in dB)
Rlimit Node Node 1st vs 1st 4th vs 4th 1st vs 4th
4 1 56 80 2.9 1.4 0.8
4 2 88 80 3.0 2.2 1.0
4 4 152 80 3.4 2.5 1.5
TABLE 1: Complexity Analysis of Conventional and Proposed Complex Decoder.
As tabulated in Table 1, for iterative conventional complex decoder, we need to perform 80
calculations for K equal to 4. Although our proposed decoder calculates 56, 88 and 152 nodes
using same list size and Rlimit set to 1, 2, and 4 respectively. Hence, with less computational
complexity, the proposed decoder can achieve 1.4 dB better performance than that of
conventional one for the 4th iteration. However, 2.2 to 2.5 dB gain can be achieved by tolerating
higher computational complexity using proposed complex decoder. Considering 1st iteration with
same level of complexity, 2.9 dB to 3.4 dB gain can be achieved using proposed decoder. Next,
complexity analysis of proposed and iterative real decoder is presented in Table 2.
K
Proposed Real Proposed vs Real (in dB)
Rlimit Node Node 1st vs 1st 4th vs 4th 1st vs 4th
4 1 56 112 9.0 6.9 6.1
4 2 88 112 9.1 7.7 6.5
4 4 152 112 9.5 8.0 6.8
TABLE 2: Complexity Analysis of Iterative Real and Proposed Complex Decoder.
As shown in Table 2, the number of the nodes need to be expanded for LR-aided real decoder
[14] for list size 4 is equal to 112. Considering the same list size, proposed complex decoder
requires 56, 88 and 152 node expansion for Rlimit set to 1, 2 and 4 respectively. Hence,
proposed decoder can achieve 6.9 dB to 7.7 dB better performance even with less computational
complexity comparing with the iterative real one. Allowing more complexity, can increase the
performance to 8.0 dB. If we consider the performance of only 1st iteration, with same level of
complexity the proposed decoder can attain 9.0 to 9.5 dB improvement comparing with the real
one.
Therefore, our iterative soft complex decoder with Rlimit offers a trade-off between performance
and complexity for different iteration. It not only increases the performance, but also can reduce
complexity to a certain level.

4. CONCLUSION
In this paper, an iterative soft decision based complex domain K-best decoder is proposed
exploiting the improved complex on-demand child expansion. It includes the use of LR algorithm
in order to achieve orthogonality among the constellation points reducing the effect of noise. An
additional parameter, Rlimit is introduced to tune the complexity of computation with improvement
in BER performance. Reduction of computational complexity directly results to less power
consumption of the decoder as well.
We also compare the result of 4th iteration of our proposed decoder with iterative conventional
complex decoder and obtain 1.4 to 2.5 dB improvement at the BER of 10ି଺
for 8 × 8 MIMO and
64 QAM modulation scheme with comparable complexity. Comparing with iterative LR-aided real
domain decoder, the improvement increases more than 7.0 dB with less computational
complexity. Although more than 2.9 dB and 9.0 dB gain can be achieved with same level of
complexity comparing 1st iteration of proposed decoder with that of conventional iterative
complex and real decoder respectively.
Future work of this proposed decoder includes evaluating the detector performance using
additional channel and simulation environment and also implementing the algorithm on FPGA
and so on.
5. REFERENCES
[1] "IEEE Standard for Information Technology- Local and Metropolitan Area Networks-
Specific Requirements- Part 11: Wireless LAN Medium Access Control (MAC) and Physical
Layer (PHY) Specifications Amendment 5: Enhancements for Higher Throughput." IEEE
Standard 802.11n-2009 (Amendment to IEEE Standard 802.11-2007 as amended by IEEE
Standard 802.11k-2008, IEEE Standard 802.11r-2008, IEEE Standard 802.11y-2008, and
IEEE Standard 802.11w-2009), pp. 1-565, Oct. 2009.
[2] J. Jalden and B. Otterston. “On the Complexity of Sphere Decoding in Digital
Communications.” IEEE Transaction on Signal Processing, vol. 53, no. 4, pp. 1474-1484,
Apr. 2005.
[3] I. Lai, G. Ascheid, H. Meyr and T.-D. Chiueh. "Low-Complexity Channel-Adaptive MIMO
Detection with Just-Acceptable Error Rate." IEEE 69th Vehicular Technology Conference:
VTC-2009 Spring, Apr. 2009, pp. 1-5.
[4] M. Shabany and P. Glenn Gulak. “The Application of Lattice-Reduction to the K-Best
Algorithm for Near-Optimal MIMO Detection.” IEEE International Symposium on Circuits
and Systems (ISCAS), May 2008, pp. 316-319.
[5] C. Windpassinger and R. Fischer. “Low-Complexity Near-Maximum Likelihood Detection
and Precoding for MIMO Systems Using Lattice Reduction.” Proceeding IEEE Information
Theory Workshop, Mar. 2003, pp. 345-348.
[6] Q. Zhou and X. Ma. “An Improved LR-aided K-Best Algorithm for MIMO Detection.”
Proceeding IEEE International Conference on Wireless Communication and Signal
Processing (WCSP), Oct. 2012, pp. 1-5.
[7] X. Qi and K. Holt. “A Lattice-Reduction-Aided Soft Demapper for High-Rate Coded MIMO-
OFDM Systems.” IEEE Signal Processing Letters, vol. 14, no. 5, pp. 305-308, May 2007.
[8] B. M. Hochwald and S. Ten Brink. "Achieving Near-Capacity on a Multiple-Antenna
Channel." IEEE Transactions on Communications, vol. 51, no. 3, pp. 389-399, Mar. 2003.

[9] F. Sheikh, E. Wexler, M. Rahman, W. Wang, B. Alexandrov, D. Yoon, A. Chun and A.
Hossein. "Channel-Adaptive Complex K-Best MIMO Detection Using Lattice Reduction."
IEEE Workshop on Signal Processing Systems (SiPS), pp. 1-6, Oct. 2014.
[10] M. Rahman, E. Rohani and G. Choi. "An Iterative Soft Decision Based Adaptive K-Best
Decoder Without SNR Estimation." Asilomer Conference on Signals, Systems and
Computers, Nov. 2014, pp. 1016-1020.
[11] Q. Wen, Q. Zhou, C. Zhao and X. Ma. “Fixed-Point Realization of Lattice-Reduction Aided
MIMO Receivers With Complex K-Best Algorithm.” IEEE International Conference on
Acoustics, Speech and Signal Processing, May 2013, pp. 5031-5035.
[12] K. Gunnam, G. Choi, W. Weihuang and M. Yeary. "Multi-Rate Layered Decoder
Architecture for Block LDPC Codes of the IEEE 802.11n Wireless Standard." IEEE
International Symposium on Circuits and Systems (ISCAS), May 2007, pp. 1645-1648.
[13] M. Rahman, E. Rohani and G. Choi. "An Iterative LR-Aided MMSE Extended Soft MIMO
Decoding Algorithm." International Conference on Computing, Networking and
Communications, California, Feb. 2015.
[14] E. Agrell, T. Eirksson, A. Vardy and K. Zeger. “Closest Point Search in Lattices.” IEEE
Transaction on Information Theory, vol. 48, no. 8, pp. 2201-2214, Aug. 2002.
[15] J. Jalden and P. Elia. "DMT Optimality of LR-Aided Linear Decoders for a General Class of
Channels, Lattice Designs, and System Models." IEEE Transaction on Information Theory,
vol. 56, no. 10, pp. 4765-4780, Oct 2010.
[16] M. Rahman, E. Rohani, J. Xu and G. Choi. "An Improved Soft Decision Based MIMO
Detection Using Lattice Reduction." International Journal of Computer and Communication
Engineering, vol. 3, no. 4, pp. 264-268, Apr. 2014.
[17] J. Jalden and P. Elia. “DMT Optimality of LR-Aided Linear Decoders for a General Class of
Channels, Lattice Designs, and System Models.” IEEE Transaction on Information Theory,
vol. 56, no. 10, pp. 4765-4780, Oct. 2010.
[18] M. Taherzadeh and A. Khandani. "On the Limitations of the Naive Lattice Decoding." IEEE
Transaction on Information Theory, vol. 56, no. 10, pp. 4820-4826, Oct. 2010.
[19] A. K. Lenstra, H. W. Lenstra and L. Lovasz. "Factoring Polynomials with Rational
Coefficients." Mathematische Annalen, vol. 261, no. 4, pp. 515-534, Dec. 1982.
[20] M. Mahdavi and M. Shabany. "Novel MIMO Detection Algorithm for High-Order
Constellations in the Complex Domain." IEEE Transaction on VLSI Systems, vol. 21, no. 5,
pp. 834-847, May 2013.
[21] M. Shabany and P. Glenn Gulak. “The Application of Lattice-Reduction to the K-Best
Algorithm for Near-Optimal MIMO Detection.” IEEE International Symposium on Circuits
and Systems (ISCAS), May 2008, pp. 316-319.
[22] C. P. Schnorr and M. Euchner. “Lattice basis reduction: Improved practical algorithms and
solving subset sum problems.” Mathematical Programming, vol. 66, pp. 181–191, Aug.
1994.
[23] M. Rahman and Gwan S. Choi. "Fixed Point Realization of Iterative LR-Aided Soft MIMO
Decoding Algorithm." An International Journal on Signal Processing, vol. 9, issue 2, pp. 14-
24, May 2015.

Iterative Soft Decision Based Complex K-best MIMO Decoder

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Iterative Soft Decision Based Complex K-best MIMO Decoder (20)

Recently uploaded (20)

Iterative Soft Decision Based Complex K-best MIMO Decoder