2018 14th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)

Smart Sorting in Massive MIMO Detection

Andrey Ivanov, Dmitry Yarotsky, Maria Stoliarenko, Alexey Frolov
Skolkovo Institute of Science and Technology
Moscow, Russia
[email protected], [email protected], [email protected], [email protected]

Abstract—In this paper, we proposed a discrete sorting op-

timization approach for uplink channel of Massive Multiple
Input, Multiple Output (MIMO) system. The algorithm includes
users (UEs) sorting before QR decomposition (QRD) and sorting-
reduced (SR) K-best detector for 48x64 MIMO uncoded systems.
Simulation results show that detector losses is about 1dB to a
Maximum Likelihood (ML) detection in low detector complexity.
The UEs sorting is required to sort diagonal elements of the R
matrix in ascending order to avoid error propagation in multi-
user (MU) scenario. Fast and low complexity method of online
discrete optimization is used to find the loss function minimum.
Sorting tracking is proposed, so that the pre-sorted interpolated
R matrix is used for further sorting optimization, resulting
in low sorting complexity. The proposed sorting demonstrates
huge performance gain compare to a common power-based one.
Simulation results in 5G QuaDRiGa channel are presented.
Fig. 1: MIMO detection scenario
SR-K-best detector is a variant of K-best detector. The SR-
K-best with (K,S,p) parameters results in significant losses in
scenarios with high correlated users, therefore we proposed a where y is the frequency domain received signal vector of size
new structure (K,S,p,v,q) of the SR-K-best algorithm and a new N for each subcarrier; N is the number of receive antennas;
discrete optimization method to increase performance. Discrete H is the propagation channel matrix of [N × M ] size for
stochastic optimization was done offline in QuaDRiGa channel to
find optimal (K,S,p,v,q) parameters for fixed detector structure.
each subcarrier, x is the transmitted signal vector of M size;
M is the number of single antenna UEs; n is an additive noise
I. I NTRODUCTION vector; xM L is the ML estimation of x. Complexity of ML
detection exponentially increases with the number of transmit
MU-MIMO detection, well described in [1] and [2], is
antennas.
a method of combining antennas digital signal to increase
Minimum mean square error (MMSE) estimator is a base-
user signal power in uplink channel of base station. It is a
line linear solution of the detection problem, described in [2]
key technology for future 5G systems [3], which provides
and [4]. However, huge performance losses in MU-MIMO
extremely high system capacity. Switching from a common
scenario limit the MMSE application, and a new algorithm is
MIMO to Massive MIMO requires efficient and fast signal
required. The MMSE estimation is calculated as:
processing algorithms since implementation complexity is the
main concern with a very large number of antennas (e.g., xM M SE = W y,
tens or hundreds). For MIMO detection there is always a
HH
trade-off between performance and complexity. Intuitively, W = 2
σn
,
computational resources should be utilized more efficiently HHH + 2 I
σx
by performing optimal detection. A typical MIMO detection
where xM M SE is the MMSE estimation of x; W is the
scenario is shown in Fig. 1.
weight matrix; σn2 is the receiving antenna noise power; σx2 is
Maximum Likelihood (ML) detector is the optimal detector a transmitted UE power; I is [M × M ] eye matrix. Equation
because it checks all possible combinations of TX symbols can be calculated in simplified form by using an extended
and chooses one with the minimum Euclidean distance. How- channel matrix:
ever, the complexity grows exponentially with high modula- HH
tion order (QAM16 or higher) or large number of TX antennas W = H ext ,
H ext H ext
(say > 16). The ML estimation is given by [2]:
h iT
y = Hx + n, (1) H ext = H T σσnx I , (2)
T
yT

xM L = arg min ky − Hxk , y ext = 0 ,

The research was carried out at Skoltech and supported by the Russian MMSE with ordered successive interference cancellation
Science Foundation (project no. 18-19-00673). (OSIC) MIMO detector is performed from the QRD of the

978-1-5386-6876-4/18/$31.00 ©2018
Authorized licensed use limited to: Ural FederalIEEE
University. Downloaded on June 21,2021 at 03:45:05 UTC from IEEE Xplore. Restrictions apply.
 2018 14th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)

permuted channel matrix H perm , which is defined as in [5], 1) variable extra complexity of cLLL implementation;
[6] and [7]: 2) slicing (rounding to the nearest point as shown in [1]
y ext = H perm xperm + n, and [5]) and candidates list generation in LR domain
since T transformation results in non-rectangle constel-
xperm = P x, lation and fixed candidates numeration is less efficient;
H perm = P H ext , (3) 3) log-likelihood ratio (LLR) calculation in LR domain
because most modern decoders require soft decision;
where matrix H ext is calculated by equation (2) and used 4) interpolation-based cLLL (i.e. how to interpolate T
instead of H for regularization reasons. QR factorization matrix from pilot symbol to the current one, such
plays a key role in the OSIC detection [6]: approach is described in [9] for QRD without LRA).
H perm = QR, Many researchers have proposed solutions to decrease
  LRA complexity or fix the processing iteration of the LR
QH y ext = QH Q Rxperm + QH n, algorithm in their hardware implementation, for example [1]
and [6], but LRA still requires a lot of computations and,
QH y ext = Rxperm + QH n, moreover, performance gap between ML and LRA solution
where R is a [(N + M ) × M ] upper triangular matrix; P is still big in highly correlated UEs scenarios. Nevertheless,
is the permutation matrix, required to reorder UEs before most researchers still ignore problems 1-4, only considering
the SIC detection. Iterative detection starts from xperm (M ) performance gain.
amplitude detection and stops with xperm (1) calculation Using MMSE or MMSE-OSIC MIMO detector after LR
due to upper triangle matrix R property. After that initial cannot achieve satisfying performance for 48×64 MIMO sys-
vector x can be calculated as x = P T xperm . MMSE-OSIC tem. Therefore, the processing K-best detector can be utilized
detection demonstrates improved performance compared to to achieve satisfying performance in adoptable complexity
the MMSE detection, but the improvement is limited due to [1]. Although the proposed distributed K-best (D-K-best, [8])
error propagation issues, caused by non-ideal permutation (or algorithm greatly reduces the sorting complexity compared
sorting). Most known sorting methods are too complex or with a common breadth-first K-best algorithm, the D-K-best
based on power reordering, resulting in performance losses in also results in significant performance losses in correlated
correlated channel. channel. In [1] a variant of D-K-best detector called SR-K-
Lattice reduction aided (LRA) type of methods used in best algorithm was proposed. Sorting the best K survivors
[1] and [8] are pre-processing procedures that operate on the from KM (M = 4) candidates is reduced to sorting the
channel matrix H only rather than pure detection methods best S candidates while other K − S survivors are called
that operate on the received data vector y as well. The LRA ”most expected” and selected without sorting. Thus, paper
MIMO detection algorithms build upon the idea of converting [1] proposes an SR-K-best with index (K, S, p) to reduce the
an ill-conditioned problem into an equivalent well-conditioned sorting operation from K to S times. The vector p defines
problem via a linear transform T that fulfills certain con- positions of the ”most expected” candidates. However, its
ditions. Lattice reduction reorganizes the extended channel architecture is not optimal in candidates selection for the next
matrix to more orthogonal one, lowering the likelihood of iteration step.
noise propagation: In addition, the MIMO detection problem can be refor-
mulated as how to achieve best performance in the fixed
y = H ext T T −1 x + n = H LR z + n, complexity? Alternatively, how to find optimal architecture
z = T −1 x, in fixed complexity? A general approach to this problem is
a machine learning application. Machine learning asymptoti-
H LR = H ext T , cally converges to ML solution and can increase performance

1
 in the receiver. There are many machine learning approaches
T
x = aP round (T z) + 1 (1 + i) , to MIMO detection problem: support vector regression (SVR)
2
is proposed in [10], convolutional neural networks (CNN) are
where H LR is the lattice-reduced channel matrix; T is the proposed in [11] and [12]. Authors achieve performance gain
uni-modular matrix; a is the scaling coefficient (a = √210 in high correlated scenarios and their solution requires less
for QAM16 and a = √242 for QAM64) . Thus, many LRA calculations. However, all simulations done with modulation
detectors use function round() to slice the invalid vector order up to QPSK, since phase modulation fits to neural
T z to the nearest point of x. Once the problem is well- networks architecture. With QAM16 or higher the machine
conditioned, common MMSE or MMSE-OSIC detectors can learning algorithm complexity grows significantly. Another
be used to achieve near-optimal performance. The complex problem appears when in uplink channel there are several
Lenstra–Lenstra–Lovász (cLLL) algorithm, described in [1], users with different modulation scheme, for example, N1 UEs
[6] and [8], is a well-known polynomial time LRA algorithm. with QPSK, N2 UEs with QAM16 and N3 UEs  with QAM64.
However, it is inefficient for direct implementation because It is clear that the number of combinations N1 N2 N3
of several problems: is rather big when Ni ∈ [1...64] and to learn all such scenarios

Authorized licensed use limited to: Ural Federal University. Downloaded on June 21,2021 at 03:45:05 UTC from IEEE Xplore. Restrictions apply.
 2018 14th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)

for each signal-to-noise ratio (SNR) and for each correlation III. D ETECTION APPROACH
between UEs the CNN should have huge size. Therefore, the
advantage of direct machine learning application in MIMO A functional scheme of the proposed approach is shown
detection is not evident; however, learning ideas could be use- in Fig. 3. It consists of pre-processing and processing parts.
ful to optimize existing algorithms structures. Our simulations Preprocessing is required to calculate sorted QRD in two
with CNN deep learning in QuaDRiGa channel show that steps.
learning error doesn’t go down when the number of layers
Step 1: QRD interpolation as described in [14].
exceeds the number of UEs. Therefore, CNN demonstrates
performance gain in MIMO detection due to clever iterative For QRD calculation in the MIMO system, we have to esti-
OSIC. mate the channel response of each subcarrier and then perform
In this paper, we show by simulations in QuaDRiGa the QRD for each subcarrier. In practice, the interpolation-
channel that MU-MIMO detection is very sensitive to sorting based QRD only computes the Q and R matrixes for the
operation and propose smart sorting optimization. In fact, pilot subcarriers to reduce a computational complexity. Then,
clever UE sorting before QRD and candidates sorting in SR- the Q and R of the data subcarriers are interpolated from
K-best stage demonstrate final performance similar to LRA those of the pilot subcarriers.
application, but in much less complexity, and the proposed Step 2: UE sorting (strings permutation in matrix H ext )
solution is more feasible in application. to find the P matrix for equation (3).
UE sorting problem is well-known and there are many ap-
II. S IMULATION TOOL proaches to overcome it, for example a post-sorting algorithm
and pre-sorting solution are analyzed in [6]. However, most
QuaDRiGa, short for QUAsi Deterministic Radio chan- methods require many computation resources or result in poor
nel GenerAtor [13], is a Matlab software used for gener- performance. We propose to use P matrix from QRD of the
ating realistic radio channel impulse responses for system- pilot symbols as the first step of P matrix calculation for the
level simulations of mobile radio networks. These simula- data symbol. Therefore, sorting tracking is considered. The
tions are used to determine the performance of new tech- P matrix slightly changes from one symbol to another and a
nologies in order to provide an objective indicator for the low sorting complexity is required to update it. Loss function
standardization process in bodies like the third generation L = L{diag (R)} of diagonal entries of interpolated R
partnership program 3GPP. Besides being a fully-fledged matrix is optimized to guarantee the least number of matrix P
three dimensional geometry-based stochastic channel model, updates (low complexity updates) from one symbol to another
QuaDRiGa contains a collection of channel models along with without performance losses. Processing part is represented by
novel modelling approaches, which provide features to enable
quasi-deterministic multi-link tracking of users movements in
changing environments. QuaDRiGa contains a couple of new
features and is furthermore calibrated against 3GPP channel
models like 3GPP-3D and the latest New Radio channel
model. The QuaDRiGa approach can be understood as a
statistical raytracing model. In our simulation we used non-
line of sight (NLOS) scenario with 48 single antenna UEs
moving with averaged speed of 3km per hour and 64 receiving
antennas of BS (MIMO 48×64). A short fragment (magnitude
spectrum) of generated channel model is shown in Fig. 2 for
600 subcarriers in 90 symbols.

Fig. 3: Functional MIMO detection scheme

optimized SR-K-best iterative detection. First of all, (K, S, p)

parameters were optimized in QuaDRiGa channel of 48 ×
64 MIMO system. Number of users = 48 is chosen as the
max realistic number of active UEs per subcarrier in 64
RX antennas system. Besides (K, S, p) optimization, a new
vectors (v, q) are proposed and optimized. The vector v
defines a set of sorting child nodes (i.e. a set for S sorted
candidates search); while the vector q defines location of S
sorted candidates in the final composed list of K candidates
Fig. 2: Magnitude spectrum of QuaDRiGa channel for the next detection iteration. Simulation demonstrates extra
gain without any additional computations.

Authorized licensed use limited to: Ural Federal University. Downloaded on June 21,2021 at 03:45:05 UTC from IEEE Xplore. Restrictions apply.
 2018 14th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)

IV. S MART UE SORTING BEFORE QRD 1) Given an interpolated pre-sorted QRD for the current
Let H be a N = 64 by M = 48 matrix of full rank M . symbol (sorting comes from the pilot symbol);
|R |2 +|Rn+1,n+1 |2
We consider the QRD of matrix H perm = P H ext according 2) We compute the ratios n,n+1 Rn,n 2 for all n
to equations (2) and (3) with permuted strings, i.e. equation and check if there are any less than 1 so that we can
H perm = QR. We assume without loss of generality that the decrease the loss;
matrix R in this decomposition is square. We are interested 3) If yes, we perform the respective elementary permuta-
in finding permutations such that an absolute value of the tion of neighboring strings, i.e. updating the matrices Q
diagonal entries |Rnn | tend to increase as n grows. In general, and R for the current symbol;
it is not possible to find a permutation achieving a complete |R |2 +|Rn+1,n+1 |2
4) After that, we update the ratios n,n+1 Rn,n 2
ordering, but we can fix a specific loss function for new matrix R;
L = L{diag (R)}, 5) Repeat until there are no ratios less than 1.
Let us estimate the complexity of one step of this algorithm.
penalizing the lack of order, and then consider the problem Updating the matrices Q and R requires O(N ) arithmetic
of optimizing the permutation with respect to this loss func- |R |2 +|Rn+1,n+1 |2
tion. One obvious choice for the loss function is the number operations. Updating the ratios n,n+1 Rn,n 2 requires
of pairs of diagonal entries with the wrong order: O(2) operations, since we only need to recompute it at
X n = k − 1 and n = k + 1, where k corresponds to the
L= (Rkk ≥ Rnn ). (4) elementary permutation of the previous step. Checking if there
1≤k<n≤M are ratios less than 1 requires O(M ) comparisons. Thus, the
However, equation (4) is a poor choice for several reasons. total number of one-step comparisons is O(N +M ), i.e. O(N )
First, this loss is excessively (discontinuously) sensitive to arithmetic operations. At each step of the algorithm, we can
the changes in R whenever two diagonal entries become ensure without changing the O(N ) complexity that we choose
equal, and is entirely insensitive otherwise. Second, even if n with the lowest currently available ratio, thus obtaining
we perform only some simple permutation on H ext (say the optimal gain. This can be done simply by minimizing
exchange a pair of neighboring strings), its effect on the loss the ratios over all n’s at each step or, more efficiently, by
is relatively hard to compute: we need to check for each maintaining (and keeping updated between steps) the sorted
diagonal entry, even if it is unaffected by the permutation, list of ratios smaller than 1.
whether any pairs of this entry with the affected diagonal
entries change their status (correct/wrong order). A much
more sensible choice for the loss function is defined as:
M
X
L=− (2n − M − 1)lnRnn . (5)
n=1

This loss (5) obviously has the locality property: if the

permutation only involves a range of strings, say I = [k, k +
1, .., m], then the change in L after applying the permutation
only depends on the old and new diagonal entries of R in
this range (the diagonal entries outside of I do not change).
On the other hand, this L can be equivalently written as:
X Rkk
L= ln , (6)
Rnn
1≤k<n≤M

which is similar to our first considered loss (4), but takes

into account the relative magnitude of the diagonal entries.
Consider the permutation of only two neighboring strings
of matrix H ext , say k and k + 1. Let R and R0 denote
the R factors in the QR decompositions of H ext before and
after the permutation, respectively.
q Then R0 nn =Rnn for all Fig. 4: BER comparison of smart and power-based orderings
0 2
n 6= k, k + 1, while R k,k = |Rk,k+1 | + |Rk+1,k+1 | and
2 in 48 × 64 uncoded MIMO system with QAM16
q
2 2
R0 k+1,k+1 = Rk,k Rk+1,k+1 / |Rk,k+1 | + |Rk+1,k+1 | . As a crude estimate of the number of steps, note that any
The increase in the loss that this permutation brings then permutation of M strings can be written as not more than
|R |2 +Rk+1,k+1 2 M2
equals ∆L = ln k,k+1 Rk,k 2 . 2 elementary permutations of pairs of neighboring strings.
This suggests the following algorithm of finding a locally However, in our procedure the optimization of the ordering
optimal permutation by multiplying a sequence of above starts with an already pre-sorted matrix H ext . In addition, our
elementary permutations. optimization (6) only finds a local optimum, which typically

Authorized licensed use limited to: Ural Federal University. Downloaded on June 21,2021 at 03:45:05 UTC from IEEE Xplore. Restrictions apply.
 2018 14th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)

does not need a long chain of elementary permutations. As

a result, the number of optimization steps is small, so the Initialization :
overall complexity of sorting optimization is low. Each father node produces 4 child nodes;
Comparison results of the proposed smart UE sorting with
power-based sorting are shown in Fig. 4 for QuaDRiGa Step 1 :
channel of 48 × 64 MIMO system. Simulation is done q = [13 14 15 16];
with (MMSE-OSIC + K-best (K = 8)) detection algorithm
P
p = [a1 a2 ... a16 ]; 0 ≤ ai ≤ 3; ai+1 ≤ ai ; ai = 12;
and QAM16 modulation. Power-based sorting considers only v = [4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4] − p;
power ratios between UEs and ignores correlation, i.e. in the Task: find p with the best gain in BER=10−2 ;
first iteration, the most powerful UE is detected and UE with
the lowest power is detected in the last iteration. Step 2 :
V. S MART SR-K- BEST q = [13 14 15 16]; optimized p;
The general idea of SR-K-best with parameters (K, S, p) v = [4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4] − p;
from [1] is to determine the (K − S) child nodes according Task: minimize those vector v elements, which cause the
to a fixed order and then apply distributed sorting S times least performance degradation in BER=10−2 and repeat the
to find the remaining nodes. The vector p is a vector that minimization until the least degradation < 0.1dB;
shows how to choose the first (K − S) child nodes without
sorting. For example, vector p = [2 1 1 1 1 0 0 0] means Step 3 :
that the detector selects two best child nodes of the first Optimized p; optimized v;
father node and one best child node from the second, third, q = [b1 b2 b3 b4 ]; bi ≤ 16; bi+1 > bi ;
fourth and fifth father nodes (6 candidates without sorting). Task: find q with the best gain in BER=10−2 ;
The remaining two (S = 2) nodes are selected by using a
distributed sorting method (D-K-best algorithm) from [8] over
the residual child nodes. In the D-K-best algorithm, to apply After 3 steps of optimization we achieved the result:
the sorting procedure we have to reserve the best residual child p= [2 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0];
node from each father node. In the first father node two best v= [2 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2];
child nodes are already pre-selected without sorting, therefore q = [2 4 6 8].
we reserve the third best child node. Then we reserve the
second best child node from the second father node because
one best child node is also pre-selected in the second father The optimized algorithm structure is shown in Fig. 5 and
node without sorting, and so on. On the last stage we reserve simulation results of optimized SR-K-best are shown in Fig. 6
the first best child node from the eight’s father node since this and Fig. 7 (smart UEs sorting before QRD is used). The
father node has no any pre-selected child nodes. Finally, in the algorithm outperforms K-best (K = 8) while requires much
D-K-best algorithm we have the vector v= [1 1 1 1 1 1 1 1] less computations, i.e. 8 times sorting in 8 ∗ 4 = 32 child
corresponding to the 8 reserved child nodes (1 child node per nodes in K-best instead of 4 times sorting in 22 child nodes
1 father node). The remaining two nodes (S=2) are selected by in the proposed one. It can be found that the initial SR-K-
sorting among the 8 reserved child nodes defined by the vector best structure (K, S, p) = (8, 2, [2 1 1 1 1 0 0 0]) requires
v. It is claimed that (K, S, p) = (8, 2, [2 1 1 1 1 0 0 0]) less computation, but demonstrates huge losses caused by
could achieve performance similar to K-best with K = 8, significant correlation between UEs. From Fig. 6 and Fig. 7
but in much less sorting resources. leads that OSIC has no losses to (LRA & OSIC) in high
To optimize the SR-K-best algorithm we applied discrete correlated scenario. With smart ordering before and after
stochastic optimization of p vector with fixed (K,S) param- QRD, the proposed low complexity Massive MIMO detection
eters in QuaDRiGa channel of 48 × 64 MIMO system. system has only 1 dB losses to ML solution in 48 × 64
In fact, a higher K and S parameters always result in scenario.
better performance in correlated channel, however, in practice
S = K/4 is a minimal number of sorted candidates to achieve VI. C ONCLUSION
performance similar to K-best. We fixed (K, S) = (16, 4) as We proposed a sorting solution consisting of a new UEs
a compromise between performance and sorting complexity, sorting in preprocessing unit and a new SR-K-best structure.
but also considered (K, S, p) = (8, 2, [2 1 1 1 1 0 0 0]) for The proposed method has a low complexity and results in only
further simulations. Direct D-K-best is inefficient since taking 1 dB losses compare to ML solution without using Lattice
only 1 child node per 1 father node always results in poor Reduction Aided (LRA) pre-processing. All simulation was
performance in high correlated channels. The proposed SR- done in QuaDRiGa NLOS channel with 48 UEs moving with
K-best algorithm consists of parameters (K, S, p, v, q) and averaged speed of 3 km per hour and 64 RX antennas (48 ×
simultaneous optimization of all of them takes too long time, 64 MIMO detection system), results are provided for uncoded
therefore, discrete stochastic optimization was implemented QAM16 and QAM64 modulations in comparison with other
in Matlab in 3 steps: detection methods.

Authorized licensed use limited to: Ural Federal University. Downloaded on June 21,2021 at 03:45:05 UTC from IEEE Xplore. Restrictions apply.
 2018 14th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)

Fig. 7: Simulation results in 48 × 64 MIMO with QAM64

times lower than the popular LRA-based & K-best solution in

the same performance according to our simulations. Moreover,
Fig. 5: Implementation structure of the smart SR-K-best our method is more feasible because of easy LLR calculation,
efficient in slicing and has a fixed complexity.
R EFERENCES
[1] C. F. Liao, J. Y. Wang, Y. H. Huang, A 3.1 Gb/s 8 x 8 sorting reduced
K-best detector with lattice reduction and QR decomposition, IEEE
Trans. Very Large Scale Integr. (VLSI) Syst., vol. 22, no. 12, 2014.
[2] S. Verdu, Multiuser detection, Cambridge university press, 1998.
[3] 5G PPP Architecture Working Group, View on 5G Architecture, Version
2.0, 2017.
[4] A. Ivanov, S. Kruglik, D. Lakontsev, Cloud MIMO for Smart Parking
System, IEEE 87th Vehicular Technology Conference (VTC-Spring),
2018.
[5] M. Li, B. Bougard, E. E. Lopez, A. Bourdoux, D. Novo,
L. Van Der Perre, F. Catthoor, Selective spanning with fast enumeration:
A near maximum-likelihood MIMO detector designed for parallel
programmable baseband architectures, Proc. IEEE ICC, 2008.
[6] A. M. Mussi, B. F. Costa, T. Abro, Wireless Pers Commun, Proc. IEEE
ICC, https://doi.org/10.1007/s11277-016-3807-6, 2017.
[7] Z. Guo, P. Nilsson, A VLSI implementation of MIMO detection for
future wireless communications, Proc. IEEE Pers. Indoor Mobile Radio
Commun., pp. 2852-2856, 2003.
[8] M. Shabany, P. G. Gulak, The application of lattice-reduction to the
K-best algorithm for near-optimal MIMO detection, IEEE International
Symposium on Circuits and Systems, 2008.
[9] P. L. Chiu, L. Z. Huang, Y. H. Huang, Scalable interpolation-
based QRD architecture for subcarrier-grouped-ordering MIMO-OFDM
system, Proc. 43th IEEE Asilomar Conf. on Signals Systems and
Fig. 6: Simulation results in 48 × 64 MIMO with QAM16 Computers, pp. 708-712, 2009.
[10] Ramanathan, R. & Jayakumar, A support vector regression ap-
proach to detection in large-MIMO systems, M. Telecommun Syst,
https://doi.org/10.1007/s11235-016-0202-2, 2017.
It is hard to estimate full implementation complexity since [11] N. Samuel, T. Diskin, A. Wiesel, Deep MIMO detection, Proc. IEEE
Int. Workshop Signal Process. Adv. Wireless Commun., 2017.
”honest” QRD is calculated for pilot symbols only, while [12] T. J. O’Shea, J. Hoydis, An introduction to deep learning for the
the SR-K-best is calculated for each data symbol. Similar physical layer, IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, 2017.
situation with UEs sorting before QRD, since its complexity [13] http://quadriga-channel-model.de/
[14] P.-L. Chiu, L.-Z. Huang, Li-Wei Chai, Yuan-Hao Huang, Interpolation-
depends on the difference between an ideal ordering in data Based QR Decomposition and Channel Estimation Processor for
symbol and power-based ordering in pilot symbol. We can MIMO-OFDM System, IEEE Trans. Circuits and Systems, vo1. 58,
only claim that the proposed algorithm complexity is at least 2 2011.

Authorized licensed use limited to: Ural Federal University. Downloaded on June 21,2021 at 03:45:05 UTC from IEEE Xplore. Restrictions apply.