Вы находитесь на странице: 1из 6

A hybrid precoding design in mmWave massive

MIMO systems
Homeyra Rahbari Kahjough† , Reza Abdolee‡ , Behzad Mozaffari Tazehkand†
Emails:h_rahbari@alum.sharif.edu, rabdolee@csub.edu, mozaffary@tabrizu.ac.ir
†University of Tabriz, Iran ‡California State University, Bakersfield

Abstract—Millimeter wave (mmWave) communication using and simplified beam steering systems. They have proposed
massive multiple input multiple output (MIMO) techniques is the gradient pursuit (GP) algorithm as a novel solution that
one of the key enabling technologies for high capacity 5G is cost-effective and fast, also perform well compared to the
cellular networks. However, the hardware complexity and the
high power consumption in massive-MIMO array hinder its OMP algorithm.
integration. Hybrid precoding technique, which combine large- In [6], a hybrid precoding algorithm with low computa-
dimensional analog preprocessing with low-dimensional digital tional complexity has been developed by exploiting the semi-
processing, is a promising approach for reducing the hardware unitary optimum precoder. Effective alternating minimization
cost and power consumption in massive MIMO systems. In this
paper, we propose a novel algorithm for hybrid minimum mean- (AltMin) algorithms was proposed in [7] in which the hybrid
square error (MMSE) precoding, inspired by two algorithms precoders are designed using a matrix factorization scheme.
namely the effective alternating minimization (AltMin) and The authors of [8] have proposed a hardware efficient im-
orthogonal matching pursuit (OMP) algorithm. These algorithms plementation for hybrid precoding, named the fixed phase
optimize the digital and analog precoders. In addition, the shifter (FPS) implementation. In this technique, a dynamic
flexible partially-connected architecture that used in the proposed
structure will be more energy efficient, since it requires fewer switch network which is adaptive to the channel states, provide
phase shifters. Numerical simulation results demonstrate that connections from a small number of phase shifters to the
the proposed hybrid precoder with low-complexity scheme can antenna array. The hybrid precoder and switch mode are
significantly perform near-optimal fully-connected structure. It determined based on an effective alternating minimization
can provide an appropriate trade-off between performance and (AltMin) algorithm.
complexity compared to the conventional structures.
Index Terms—Hybrid precoding, mmWave MIMO systems,
RF chain, phase shifter, alternating minimization approach, low-
complexity. A. Contributions

I. INTRODUCTION The main purpose of this paper is to find a low-complex


Millimeter wave (mmWave) multiple input multiple out- hybrid precoder with near-optimal performance. We provide
put (MIMO) systems which use hybrid analog and digital an algorithmic hybrid precoding solution inspired from the
precoding are one of the promising techniques for future concept of orthogonal matching pursuit (OMP) algorithm and
communication networks. In mmWave MIMO systems due to alternating minimization (AltMin) algorithm in combination.
the use of small wavelengths, a large number of antennas can The proposed algorithm takes an optimal unconstrained pre-
be packed in a small area. The number of radio frequency (RF) coder as input and approximates it as a linear combination of
chains increases when the large antenna arrays are used, that beam steering vectors that can be applied to RF chain. Further,
results in higher cost and power consumption of the system. we extend this sparse precoding approach by designing hybrid
To reduce the number of RF chain in mmWave massive minimum mean-square error (MMSE) precoders. To have low
MIMO system, the hybrid precoding scheme was proposed RF-complexity technique for mmWave MIMO systems, we
[1–3] is promising solution to these problems. In this scheme use the flexible partially-connected architecture in computing
analog precoders in the RF domain are combined with digital the analog RF precoding matrix. Our results indicate that
precoders in baseband to reduce the number of RF chains. the performance gains of the proposed scheme with fewer
The decomposition of optimal precoder into hybrid pre- number of phase shifters is near to the fully-connected and
coders is realized using an algorithmic precoding solution conventional structures.
based on the concepts of compact sensitivity such as or- We organize the remainder of this paper as follows. In
thogonal matching pursuit (OMP) [4]. It takes an optimal section II, we introduce the system model and formulate
unconstrained precoder as input and approximates it as a the corresponding optimization problem. Section III discusses
linear combination of beam steering vectors applied at analog a new hybrid precoding design algorithm. In this section,
RF. They have developed sparse precoding in receiver-side we represent the proposed hybrid precoding algorithm based
by designing hybrid minimum mean-square error (MMSE) on flexible partially-connected architecture and analyze its
combiners via basis pursuit. Reference [5] compares the computational complexity. Simulation results are presented in
performance of hybrid precoder design with optimal digital section IV, followed by concluding remarks in section V.
B. Notations: C. Architecture of Analog Precoding
The lowercase and uppercase bold letters are used for The analog part of the hybrid structure is equipped by basic
vectors and matrices, respectively. [.]∗ signifies the conjugate analog components such as analog adders and analog phase
transpose. IN is the N ×N identity matrix and the expectation shifters which can only change the phase of signals. This
operator and the real part of a complex variable are represented causes some constraints on designing the analog precoding
by E[·] and Re[.]. T r(.) and vec(.) indicate the trace and matrix depending on the structure of the analog precoder.
vectorization. The motivation behind the RF domain precoder design is
coordination of the channel elements to obtain great power
efficiency. Hence, the (ı, )th element of Va is given by
II. FORMULATION OF THE HYBRID PRECODER 1
DESIGN PROBLEMS va (ı, ) = aı, eφı, = √ eφı, (4)
M
A. System Model where aı, denotes the magnitude of the (ı, )th element of Va
and φı, is the phase of the (ı, )th element of H. The ideal
In this section, we describe a system model for the downlink phase shifters harvest continuous phases without quantization,
hybrid precoding system. We assume that the BS is equipped however such phase shifters are difficult to implement.
with M antennas to transmit data to K users and there are 1) Fully-Connected Architecture: In the fully-connected
NtRF RF chains at the transmitter. architecture as shown in Fig. 1 (a), there is a connection
The channel vector of k th user is denoted as hk ∈ from each RF chain to all of the antenna elements via a
M ×1
C , k = 1, . . . , K whose entries are i.i.d. CN (0, δk ) where large number of phase shifters. Therefore, each RF chain
δk represents large scale fading. To model favorable propaga- is connected to all the antennas at BS. This connection
tion, these channel vectors should be pair-wisely orthogonal. requires NtRF analog phase shifters and one RF adder for
The downlink channel from BS to all users is defined as each antenna. However, this structure can provide preferable
H = [h1 , h2 , . . . , hK ]. spectral efficiency compare with the partially-connected struc-
In the hybrid precoding structure, Va = ture. The implementation of conventional hybrid precoding is
RF
[va,1 , va,2 , . . . , va,NtRF ] ∈ CM ×Nt is the analog RF less feasible in practical implementation [10].
RF
precoder and Vd = [vd,1 , vd,2 , . . . , vd,K ] ∈ CNt ×K is 2) Partially-Connected Architecture: In the partially-
the digital precoder. connected structure [11], each RF chain is connected to a
Let s ∈ CK×1 , where E{ss∗ } = IK is the signal subset with M/NtRF antennas, where each antenna element
vector transmitted by the BS antenna array and n is Gaussian is equipped with one analog phase shifter. Therefore, in
noise signal with zero mean and unit variance. The sampled the partially-connected architecture the analog precoding ma-
baseband signal received by all users is given by: trix needs to have M analog phase shifters. The partially-
connected structure that is more appropriate for practical
y = H∗ Va Vd s + n (1) application, is illustrated in Fig. 1 (b).
The received signal by the k th user can be written as 3) Flexible Partially-Connected Architecture: In the flex-
ible partially-connected structure as presented in Fig. 1 (c),
yk = h∗k Va Vd s + nk (2) each RF chain is connected to M/NtRF analog phase shifters.
th Unlike partially-connected method, the antennas associated
where nk is the additive noise at the k user.
with each RF chains are not fixed. Each phase shifter is
connected to all of the antennas through M independent on/off
switches.
B. Millimeter Wave Channel Model
We consider a clustered channel model to characterize the D. Architecture of Digital Precoding
sparse scattering features. Thus, the narrow-band mmWave
Inter-user interference at the baseband, is neutralized by
channel matrix can be written as
applying zero-forcing precoding given by
Ncl Nray
1 X X
H= p αkl a∗t (φta,kl , φte,kl )ar (φra,kl , φre,kl ) (3) Vd = H∗eq (Heq H∗eq )−1 (5)
Ncl Nray k=1 l=1
where Heq = H∗ Va is a low dimension K × NtRF matrix.
The channel consists of Ncl clusters and each cluster contains
The final digital processed signal is:
Nray propagation paths. In this equation, αkl , the complex
gain of the lth ray in the k th cluster, is assumed to be a y = H∗ Va Vd s + n
Gaussian random variable with zero mean and unit variance. = Heq H∗eq (Heq H∗eq )−1 s + n = s + n (6)
The azimuth (elevation) angle of departure φta,kl (φte,kl ) and
azimuth (elevation) angle of arrival φra,kl (φre,kl ) are of a where H  CM ×K is the channel matrix. Assume that the
truncated Laplacian distribution. The vectors a∗t (φta,kl , φte,kl ) transmitted signals of different users are independent from
and ar (φra,kl , φre,kl ) are, respectively, the transmit and receive each other and from the noise signals. The received signal
antenna array response vectors [9]. at the k th user can be written as:
K
X
yk = h∗k Va Vd,k sk + h∗ı Va Vd,ı sı + nk (7)
ı=1,ı6=k
2
where nk ∼ CN (0, σn,k ) denotes the additive white Gaus-
th
sian noise at the k user. Hence, the downlink signal-to-
interference-plus-noise-ratio (SINR) at the k th user can be
represented as:
|h∗k Va Vd,k |2
SIN Rk = PK (8)
2
|h∗ı Va Vd,ı |2 + σn,k
ı=1,ı6=k

Gaussian symbols are transmitted over the mmWave channel,


the spectral efficiency achieved is given by:
K
X
Rsum−rate = log2 (1 + SIN Rk ) (9)
k=1

(a) Fully-connected
E. Problem Formulation
In this paper, we will mainly focus on the precoder design.
Instead of maximizing the spectral efficiency, we design Va
and Vd to maximize the mutual information achieved by Gaus-
sian signaling over the mmWave channel [4]. The precoder
optimization problem can be stated as:

{Va,opt , Vd,opt } = arg min kVopt − Va Vd k (10)


Va ,Vd

s.t. kVa Vd k2 = K
where Vopt stands for the optimal fully digital precoder,
and Va , Vd respectively, are the analog and digital precoders
to be optimized.

III. HYBRID PRECODING DESIGN


In this section, we introduce a novel hybrid precoder
(b) Partially-connected
structure to optimize the digital and analog precoders. The
proposed hybrid precoding algorithm as described in Algo-
rithm 1, uses the flexible partially-connected structure (low-
complexity architecture) that reduces total number of phase
shifters compared to the fully-connected structure. In addition,
the proposed algorithm is combined with the sparse MMSE
solution to find linear MMSE estimators with complex struc-
tural constraints for mmWave systems.
The hybrid precoder optimization problem can further be
stated as follows:

{Va,opt , Vd,opt } = arg min kVopt − Va Vd k (11)


Va ,Vd

s.t. Vaı , ∀ı ∈ {at (φta,kl ), 1 ≤ k ≤ Ncl , 1 ≤ l ≤ Nray }


kVa Vd k2 = K
The constraint of Vaı can be inserted directly into the opti-
mization objective to obtain Vd,opt as follows
Vd,opt = arg min kVopt − At Vd k (12)
Vd
(c) Flexible partially-connected
s.t. kdiag(Vd Vd∗ )k = NtRF
Fig. 1: Architectures of analog precoding kAt Vd k2 = K
where At = [at (φta,11 , φte,11 ), ..., at (φta,Ncl Nray , φte,Ncl Nray )]
is an t × Ncl Nray matrix consisting of array response vectors.  Pı
=1 x̃

 Pn ı
, x̄ı < 0 and x̄ı ∈ [2x̃ı , 2x̃ı+1 ]
x̃
A. Proposed Architecture x̄ı = =ı+1
, x̄ı > 0 and x̄ı ∈ [2x̃ı , 2x̃ı+1 ] (22)
 n−ı
+∞, otherwise

In the proposed hybrid precoding implementation based on
the flexible partially-connected structure, M phase shifters and
in total M 2 adaptive switches with binary states are required. Algorithm 1 Proposed Algorithm: Spatially Sparse MMSE
Accordingly, the analog precoding matrix Va can be expressed precoding under flexible partially-connected structure
as:
Require: Vopt , SN R
Ṽa = SVa (13)
1: Va = ∅
M ×M ı=0
where S ∈ {0, 1} is the switch matrix, and the boolean 2: Vres = Vopt
constraints are induced by the switches with binary states. By 3: FOR ı ≤ NtRF do
replacing the digital precoder matrix Vd with Vd = αVD 4: ψı−1 = A∗t Vres ı−1
, {At = [at (φta,kl , φte,kl )], 1 ≤ k ≤
where VD is a semi-unitary matrix, the precoder optimization Ncl , 1 ≤ l ≤ Nray }

problem can be written as: 5: `= arg min (ψı−1 ψı−1 )k,l
∗ ∗
k ∈ {1, . . . , N cl }
kVopt − Ṽa Vd k2 = T r(Vopt Vopt ) − T r(Vopt Ṽa Vd ) l ∈ {1, . . . , Nray }
− T r(Vd∗ Ṽa∗ Vopt ) + T r(Vd∗ Ṽa∗ Ṽa Vd ) 6: Va = [Va |A`t ]
∗ ∗
2
= kVopt k − 2αRe(T r(VD )Vopt Ṽa )∗ 7: VD = √ 1
SN R
(VD ∗
Va∗ Vopt Vopt Va VD +
RF
2 2 Nt ∗ −1 ∗ ∗ ∗
+ α kṼa VD k (14) SN R Va Va ) VD Va Vopt

IN RF −Vopt Va VD
ı
We can derive an upper bound for the last term by utilizing 8: Vres = t
∗ V V k
kIN RF −Vopt a D
t
the semi-unitary property of Ṽa VD 9: END FOR
∗ 10: REPEAT
kṼa VD k 2
= T r(VD Ṽa∗ Ṽa VD )

11: Fix VD and optimize α and S according to (20) and (21),
= T r(VD Va∗ S∗ SVa VD ) respectively.
2
< kSk (15) 12: Fix α and S and update VD with (23);
13: until a stopping criterion triggers.
Thus, we will have:
14: For the digital
√ precoder at the transmit end, normalize
∗ K
kVopt k2 − 2αRe(T r(VD Vopt SVa )) + α2 kSk2 (16) Vd = α kVa V Dk
VD
15: RETURN Ṽa = SVa , Vd
So, the precoder optimization problem can be reformulated as:

min {α2 kSk2 − 2αRe(T r(VD Vopt SVa )} (17) For the digital precoder at the transmit end, Vd is normal-
α,S,VD
ized to reduce the complexity of the algorithm. After updating
∗ the analog precoder Va , the digital precoder Vd = αVD will
VD V D = IK
be optimized with the semi-orthogonal constraint. The solution
When the switch matrix S and α are fixed, the optimization which is a typical semi-Orthogonal Procrustes Problem (OPP),
problem can be written as: can be expressed as

max{αRe(T r(VD Vopt SVa ))} (18) VD = VU∗ (23)
VD

It is shown recently in [8] that the subproblem of updating S where αVopt SVa = UΣV∗ follows the singular value
and α can be recast as: decomposition (SVD) and Σ is a diagonal matrix.

min kRe(T r(Vopt VD Va ∗ )) − αSk (19)
α,S B. Computational Complexity Comparison
The optimal solution to (19) is given by: In this subsection, we compare the computational complex-
?
α = arg min f(x̃ı )f(x̄ı ) (20) ities for the three involved architectures, namely the fully-
{x̃ı x̄ı }n
ı
connected, partially-connected and flexible partially-connected
(the proposed architecture). Their computational complexities
where, f (α) is a quadratic function within each interval
of these architectures are listed, in TABLE I. It is worth not-
[2x̃ı , 2x̃ı+1 ].
ing that our proposed architecture reduces the computational
complexity compare to the fully-connected structure.

Va ∗ ) > α

1Re(T r(Vopt VD 2 1M ×M , α>0
S? = (21)

1Re(T r(Vopt VD Va ∗ ) < α
2 1M ×M , α<0 IV. SIMULATION RESULTS
where x = ∗
vec{Re(Vopt VD Va ∗ )} and 1 is the indicator In this section, we evaluate the performance of the proposed
function and 1m×n denotes an m × n matrix with all entries hybrid structure. In the clustered channel model, we use Ncl =
equal to one. In addition, x̃ı is the ıth smallest entry in x, 3 and Nray = 10, and the simulation results are averaged
TABLE I: Computational complexity

Architecture Phase shifters Adders Switches

Fully-connected M NtRF M -

Partially-connected M - -

Proposed architecture 2M M M2

over 10000 channel realizations. This paper demonstrates the


average rate and bit error rate (BER) performance of the
proposed algorithm for downlink transmission scenario when
M = 64 and K = 16. The results of other algorithms are also
provided for comparison. We assume that the number of RF
chains equals the number of users K.
Fig. 2 compares the BER performance of the proposed
Fig. 3: Average transmission rate versus SNR
structure and other structures (OMP [4], AltMin [7], Fully-
digital structure). Fig. 3 illustrates the average transmission
rate versus SNR achieved by the proposed structure (Proposed
ALG), the hybrid precoder design via OMP algorithm (OMP) results have shown that the proposed algorithm outperforms
[4], alternating minimization algorithm (AltMin) [7] and our the other algorithms and the performance gain of this struc-
proposed architecture under two configurations in analog pre- ture under the flexible partially-connected configuration with
coding (Fully, Flexible Partially). We observe that the rate of reducing the number of phase shifters for a given total number
the proposed structure (Proposed ALG-Fully) is superior than of RF chains, is near to the performance of fully-connected
the other algorithms and it is near to the rate of fully-digital architecture.
structure.
R EFERENCES
[1] Y. Y. Lee, C. H. Wang, and Y. H. Huang, “A hy-
brid RF/baseband precoding processor based on parallel-
index-selection matrix-inversion-bypass simultaneous or-
thogonal matching pursuit for millimeter wave MIMO
systems,” IEEE Transactions on Signal Processing,
vol. 63, no. 2, pp. 305–317, 2015.
[2] M. M. Molu, P. Xiao, M. Khalily, K. Cumanan, L. Zhang,
and R. Tafazolli, “Low-complexity and robust hybrid
beamforming design for multi-antenna communication
systems,” IEEE Transactions on Wireless Communica-
tions, vol. 17, no. 3, pp. 1445–1459, 2018.
[3] X. Zhang, A. F. Molisch, and S.-Y. Kung, “Variable-
phase-shift-based RF-baseband codesign for MIMO an-
tenna selection,” IEEE Transactions on Signal Process-
ing, vol. 53, no. 11, pp. 4091–4103, 2005.
[4] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and
R. W. Heath, “Spatially sparse precoding in millimeter
Fig. 2: BER performance wave MIMO systems,” IEEE transactions on wireless
communications, vol. 13, no. 3, pp. 1499–1513, 2014.
[5] A. Kaushik, J. Thompson, and M. Yaghoobi, “Sparse
V. CONCLUSIONS hybrid precoding and combining in millimeter wave
In this paper, we have studied hybrid precoding schemes for MIMO systems,” IET Radio Prop. Tech. 5G, pp. 1–7,
narrow-band downlink massive MIMO systems and demon- Durham, UK, Oct. 2016.
strated the advantages of the recently proposed hybrid pre- [6] C. Rusu, R. Méndez-Rial, N. González-Prelcicy, and
coder. The analog precoder of the proposed algorithm uses the R. W. Heath, “Low complexity hybrid sparse precoding
flexible partially-connected configuration. We showed that the and combining in millimeter wave MIMO systems,”
same framework can be applied to the problem of designing in IEEE International Conference on Communications
practical MMSE precoders for mmWave systems. Simulation (ICC), pp. 1340–1345, IEEE, 2015.
[7] X. Yu, J. C. Shen, J. Zhang, and K. B. Letaief, “Al- modeling and cellular capacity evaluation,” IEEE journal
ternating minimization algorithms for hybrid precoding on selected areas in communications, vol. 32, no. 6,
in millimeter wave MIMO systems,” IEEE Journal of pp. 1164–1179, 2014.
Selected Topics in Signal Processing, vol. 10, no. 3, [10] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid
pp. 485–500, 2016. precoding in massive multiuser MIMO systems,” IEEE
[8] X. Yu, J. Zhang, and K. B. Letaief, “A hardware- Wireless Communications Letters, vol. 3, no. 6, pp. 653–
efficient analog network structure for hybrid precoding 656, 2014.
in millimeter wave systems,” IEEE Journal of Selected [11] F. Sohrabi and W. Yu, “Hybrid analog and digital beam-
Topics in Signal Processing, vol. 12, no. 2, 2018. forming for mmwave OFDM large-scale antenna arrays,”
[9] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, IEEE Journal on Selected Areas in Communications,
T. S. Rappaport, and E. Erkip, “Millimeter wave channel vol. 35, no. 7, pp. 1432–1443, 2017.

Вам также может понравиться