1

GMD-Based Hybrid Precoding For Millimeter-Wave

Massive MIMO Systems
Tian Xie, Linglong Dai, Xinyu Gao, Muhammad Zeeshan Shakir, and Zhaocheng Wang

Abstract—Hybrid precoding can significantly reduce the num- address this problem, a hybrid analog/digital precoding struc-
ber of required radio frequency (RF) chains and relieve the huge ture has been proposed [4], which decomposes the fully digital
arXiv:1605.04686v1 [cs.IT] 16 May 2016

energy consumption in mmWave massive MIMO systems, thus precoding matrix into one low-dimensional digital precoding
attracting much interests from academic and industry. However,
most existing hybrid precoding schemes are based on singular matrix and one high-dimensional analog precoding realized
value decomposition (SVD). Due to the very different sub- through analog circuits (e.g., phase shifter networks) [4]. In
channel signal-to-noise ratios (SNRs) after SVD, complicated bit this way, the hybrid precoding structure only needs a small
allocations is usually required to match the sub-channel SNRs. To number of RF chains to achieve the near-optimal performance
solve this problem, we propose a geometric mean decomposition [4].
(GMD)-based hybrid precoding scheme to avoid the complicated
bit allocation. Its basic idea is to seek a pair of analog and How to obtain the optimal precoding matrix is the key
digital precoding matrices that are sufficiently close to the optimal issue for hybrid precoding. In [5], the authors formulated the
unconstrained GMD precoding matrix. Specifically, we design hybrid precoding design problem as a sparse reconstruction
the analog (digital) precoding matrix while keeping the digital problem, and proposed to use the orthogonal matching pursuit
(analog) precoding matrix fixed. Further, the principle of basis (OMP) algorithm to obtain the analog precoding matrix. [6]
pursuit is utilized in the design of analog precoding matrix, while
we obtain the digital precoding matrix by projecting the GMD proposed a low-complexity version of the algorithm in [5]
operation on the digital precoding matrix. Simulation results based on the local search method, while [7] derived a lower
verify that the proposed GMD-based hybird precoding scheme bound on the numbers of RF chains to perfectly achieve the
outperforms conventional SVD-based hybrid precoding schemes performance of fully digital precoding by using the hybrid
and achieves much better bit error rate (BER) performance with precoding structures. The basic ideas of these works above are
low complexity.
seeking a pair of analog and digital precoding matrices that are
Index Terms—Massive MIMO, mmWave communications, hy- sufficiently close to the right singular matrix from the singular
brid precoding, geometric mean decomposition. value decomposition (SVD) of the channel matrix (which
are called SVD-based hybrid precoding in this paper). SVD-
I. I NTRODUCTION based hybrid precoding can achieve the capacity-approaching
Due to the vast unlicensed spectrum in millimeter-wave performance with the help of water-filling power allocation.
(mmWave) band (30-300 GHz), mmWave communications can Nevertheless, due to the very different SNRs for each sub-
offer huge spectrum resource [1]. On the other hand, massive channel in SVD-based precoding, complicated bit allocations
MIMO can provide enough array gain through a large number (i.e., allocating different modulation and coding schemes
of antennas (e.g., 256 antennas) to compensate for the severe (MCS) on different sub-channels) are needed to match the
attenuation of mmWave signals [2]. Thus, the combination SNRs of different sub-channels, which involve high or even
of mmWave communication and massive MIMO (mmWave unaffordable coding/decoding complexity in practice [8].
massive MIMO) is a promising technology for the future 5G In this paper, we propose a hybrid precoding scheme for
wireless communications, since it can significantly increase the mmWave massive MIMO systems based on the geometric
system throughput [3]. However, the conventional fully digital mean decomposition (GMD) to avoid the complicated bit
precoding structure, where one dedicated radio frequency (RF) allocations. Specifically, different from the conventional SVD-
chain is required for each antenna, is too energy-consumptive based hybrid precoding, we propose to treat the right semi-
due to the large number of RF chains (e.g., 256 RF chains unitary matrix in GMD as the optimal unconstrained precoding
for 256 antennas) in mmWave massive MIMO systems. To matrix, which can convert the mmWave massive MIMO chan-
nel into identical sub-channels and therefore naturally do not
T. Xie, L. Dai, X. Gao, and Zhaocheng Wang are with the Tsinghua require bit allocations any more. Furthermore, to efficiently
National Laboratory for Information Science and Technology (TNList), find a solution to the GMD-based hybrid precoding design
Department of Electronic Engineering, Beijing 100084, China (e-mail: ra-
[email protected], [email protected], [email protected], and problem, which is difficult to be solved, we use a decoupled
[email protected]). optimization method to design the analog and digital precoding
Muhammad Zeeshan Shakir is with the Department of Systems and matrix. Particularly, we solve the analog precoding matrix
Computer Engineering, Carleton University, Ottawa, Ontario, Canada. (e-mail:
[email protected]). design problem via the principle of basis pursuit inspired
This work was supported in part by the International Science & Technology by [5], while we obtain the digital precoding matrix by
Cooperation Program of China (Grant No. 2015DFG12760), the National Nat- projecting the GMD operation on the digital precoding matrix.
ural Science Foundation of China (Grant Nos. 61571270 and 61271266), the
Beijing Natural Science Foundation (Grant No. 4142027), and the Foundation Simulation results verify that the proposed GMD-based hybird
of Shenzhen government. precoding scheme outperforms the conventional SVD-based
2

hybrid precoding schemes and achieves much better bit error

RF chain
rate (BER) performance with low complexity.
Notation: Lower-case and upper-case boldface letters a and RF chain RF chain
T H
A denote a vector and a matrix. tr{A}, {A} , {A} , ||A||F , Ns Baseband N t Ns
Baseband
precoder
Nt
precoder
and {A}i,j denote the trace, transpose, conjugate transpose, RF chain
Frobenius norm, and the element in the ith row and the jth RF chain

column of A, respectively. a := b is the in-place update RF chain

operation, where a is replaced by b.
Digital RF Network Digital RF Network Analog
(a) (b)
II. S YSTEM M ODEL
Fig. 1. Comparison between two typical precoding structures: (a) The fully
A. Hybrid analog/digital precoding digital precoding; (b) The hybrid analog/digital precoding.
We consider a typical mmWave massive MIMO system where L denotes the number of paths, βi is the complex gain of
with hybrid precoding structure, where the base station (BS) the ith path, ar (ϕri ) and at (ϕti ) are the array response vectors at
with Nt transmit antennas sends Ns independent data streams the user and the BS, respectively, where ϕri denotes the angle
to the user with Nr receiving antennas. Furthermore, we of arrival (AoA) at the user, and ϕti is the angle of departure
assume that the BS and the user have NtRF and NrRF RF (AoD) at the BS. For the simple uniform linear line (ULA)
chains, respectively, which satisfy Ns ≤ NtRF ≤ Nt and antenna array of N elements, the array response vector is
Ns ≤ NrRF ≤ Nr [5]. In the hybrid precoding structure, as r h
1 iT
shown in Fig. 1, the hybrid precoding matrix P ∈ CNt ×Ns at aULA (ϕ) = 1, ejkd sin(ϕ) , · · · , ej(N −1)kd sin(ϕ) , (5)
the BS can be written as the product of two parts: the first part N
RF
is a low-dimension digital precoding matrix PD ∈ CNt ×Ns ; where k = 2π λ with λ the wavelength, and d denotes the
the second part is a high-dimension analog precoding matrix antenna spacing. Because of the limited spatial scattering in
RF
PA ∈ CNt ×Nt , i.e., P = PA PD , where C is the set of mmWave propagation, the mmWave massive MIMO channel
complex numbers. Note that the above remarks on precoding H is low-rank [10]. As a result, we can leverage a finite
matrix can also be applied to the combining matrix. Thus, the number of RF chains to achieve the near-optimal throughput
transmitted signal vector x is [10].
x = Ps = PA PD s, (1) III. P ROPOSED GMD-BASED H YBRID P RECODING
Ns ×1
where s ∈ C denotes the source signal vector. To meet the In this section, we first review the fully digital precoding
constraint of transmitting power, we bound the transmit power schemes. Next, we propose the GMD-based hybrid precoding.
at the BS as tr{PPH } ≤ Ns . After receiving the signal vector,
the user utilizes a hybrid combining matrix W to combine the A. Fully digital SVD-and GMD-based precoding
signal vector: In this subsection, we first briefly review the fully digital
√ precoding schemes in mmWave massive MIMO systems.
y = ρWH Hx + WH n
√ (2) Denote the SVD of the channel matrix H by
= ρWBH WAH HPA PD s + WDH WAH n,
H = UΣVH
where ρ is the average received power, H ∈ CNr ×Nt denotes   
  Σ1 0 V1H
the channel matrix between BS and the user, and n ∈ CNr ×1 is = U1 U2 (6)
0 Σ2 V2H
the additive white Gaussian noise (AWGN) vector at the user
following the circularly symmetric complex Gaussian distribu- = U1 Σ1 V1H + U2 Σ2 V2H ,
tion with zero mean vector and covariance matrix σ 2 INr , i.e., where U1 ∈ CNr ×Ns and V1 ∈ CNt ×Ns are semi-unitary
CN (0, σ 2 INr ), with σ 2 the noise covariance, and INr the Nr matrices containing the left Ns columns of unitary matrices
by Nr identity matrix. Since the analog precoder is realized U ∈ CNr ×Nr and V ∈ CNt ×Nt , respectively, and Σ1 =
through analog phase shifter network after up-converters, all diag(σ1 , · · · , σNs ) is a diagonal matrix with the largest Ns
elements of PA should have the same amplitude: singular values of H in decreasing order. With fully digital
1 SVD-based precoding, i.e., P = V1 and W = UH 1 , the
|{PA }i,j | = √ , (3) MIMO channels are converted into Ns parallel sub-channels,
Nt
of which the sub-channel gains are σ1 , · · · , σNs , i.e.,
where | · | denotes the modulus of a complex number. √
y = ρWH HPs + WH n
B. Channel Model √ √ (7)
= ρUH H
1 HV1 s + U1 n = ρΣ1 s + UH 1 n.
To capture the limited spatial selectivity or scattering char-
acteristics in mmWave massive MIMO channel, we adopt a Due to the limited spatial scattering in mmWave propagation,
widely used Saleh-Valenzuela (SV) model in this paper [5], the singular values of channel matrix H varies a lot [10],
[9], where the channel matrix H can be expressed as: which results in the very different SNRs among different sub-
r channels after water-filling power allocation1 as shown in Fig.
L
Nt Nr X
H= βi ar (ϕri )at (ϕti ), (4) 1 Note that the water-filling algorithm allocates less power on the sub-
L i channel with lower channel gain, which worsens the SNR for this sub-channel.
 3

Sub-channel gain Sub-channel gain This problem formulation can be intuitively understood as
Sub-channel gain
seeking a pair of QA QD sufficiently “close” to the optimal
Lowest sub- After averaged unconstrained precoding matrix Q1 .
channel gain The following Lemma 1 shows some insights about how
Noise Noise
power power
to efficiently solve (10).
Lemma 1. The GMD of H can be expressed as a rotation of
1 2 3 4 Index 1 2 3 4 Index
SVD as following
(a) (b)
Q1 = V1 SR ; G1 = U1 SL ; R1 = STL Σ1 SR (11)
Fig. 2. Intuitive illustration of sub-channel gains: (a) Conventional SVD-
based hybrid precoding; (b) Proposed GMD-based hybrid precoding. where SR ∈ CNs ×Ns and SL ∈ CNs ×Ns are unitary matrices
2 (a). Consequently, if we use the same MCS for all sub- dependent on Σ1 .
channels, the BER performance for the MIMO system will Proof: According to [12], the implementation algorithm of
be determined by the sub-channel with the lowest SNR. On GMD based on SVD can be briefly summarized as follow:
the other hand, if we want to guarantee the similar BER 1) Step1: Initialize R1 = Σ1 , Q1 = V1 , G1 = U1 , and
1
performance among all sub-channels, careful bit allocations compute the geometric mean value r̄ = (σ1 σ2 · · · σNs ) Ns .
are needed, which can bring the unaffordable coding/decoding 2) Step2: At stage i, where i varies from 1 to Ns − 1,
complexity in mmWave massive MIMO systems [8]. first check the ith diagonal element of R1 , i.e., ri,i : If ri,i ≥
To solve this problem, GMD has been proposed [11], where r̄, find certain element rp,p such that rp,p ≤ r̄, where p >
the channel matrix H is decomposed as: i; Otherwise, find certain element rp,p such that rp,p ≥ r̄,
  H  where p > i. Then, swap ri+1,i+1 and rp,p , as well as the
  R1 ∗ Q1
H = GRQH = G1 G2 , (8) corresponding (i+1)th and pth column in Q1 and G1 by
0 R2 QH2

where G1 ∈ CNr ×Ns and Q1 ∈ CNt ×Ns are semi-unitary R1 := {M(i) }T R1 M(i) , Q1 := Q1 M(i) , G1 := G1 M(i) ,
matrices containing the left Ns columns of unitary matrices (12)
G ∈ CNr ×Nr and Q ∈ CNt ×Nt , R1 is an upper triangular where M(i) is the corresponding permutation matrix.
(i) (i)
matrix with identical diagonal elements2 , i.e., the geometric 3) Step3: Construct two Givens matrices NL and NR
1 from INs by replacing the sub-matrix containing the four ele-
mean of singular values ri,i = (σ1 σ2 · · · σNs ) Ns = r̄, ∀i, and (i) (i) (i) (i)
∗ denotes an arbitrary matrix that we do not care. Note that the ments {NL }i,i , {NL }i+1,i , {NL }i,i+1 , and {NL }i+1,i+1
diagonal element in R1 is the geometric mean of the first Ns with the following 2 × 2 matrices:
   
singular values of H, since there are only Ns data streams. (i) 1 cri,i sri+1,i+1 (i) c −s
ΘL = , ΘR =
The fully digital GMD-based precoding employs Q1 as the r̄ −sri+1,i+1 cri,i s c
precoding matrix and GH 1 as the combining matrix, where we
(13)
can transform the effective channel matrix in (2) into an upper respectively, where
s
triangular matrix: 2
r̄2 − ri+1,i+1 p
x̂ = WH HPs + WH n c= 2 2 , s = 1 − c2 . (14)
ri,i − ri+1,i+1
(9)
= GH H H
1 HQ1 s + G1 n = R1 s + G1 n. Then, update R1 , Q1 , and G1 as
Then, by utilizing successive interference cancellation (SIC) at (i) (i) (i) (i)
R1 := NL R1 NR , Q1 := Q1 {NL }T , G1 := G1 NR .
the receiver [11], we can acquire Ns sub-channels with equal (15)
sub-channel gain rii as shown in Fig. 2 (b), which means that One can verify that
we can avoid the bit allocation caused by different SNRs for    
different sub-channels in precoding schemes based on SVD. (i) ri,i 0 (i) r̄ ∗
ΘL ΘR = r r , (16)
0 ri+1,i+1 0 i,i i+1,i+1
r̄
B. GMD-based hybrid precoding
(i) (i)
In this subsection, we turn to discuss the GMD-based hybrid which means that by NL R1 NR , we can make ri,i = r̄,
precoding. According to (1) and (4), we can formulate the while keeping other elements unchanged.
hybrid precoding problem as follow3 [5]: 4) Step4: Update i := i + 1 and go back to Step 2 until i
opt opt
equals Ns − 1.
QA , QD = arg min ||Q1 − QA QD ||F , (i)
Combining (15) and the fact that M(i) , NL , and NR
(i)
QA ,QD
1 are all unitary matrices, we can confirm Lemma 1 by let-
s.t. |{QA }i,j | = √ , (10) (1) (2) (N )
ting SL = M(1) NR M(2) NR · · · M(Ns ) NR s and SR =
Nt (1) (2) (N )
H H

M(1) {NL }T M(2) {NL }T · · · M(Ns ) {NL s }T .
tr QA QD QD QA ≤ Ns .
According to Lemma 1, the objective function in (10) can
2 Here, we use the simple form r
i,j to denote the element in the ith row
be written as
and jth row of R1 for simplicity. ||Q1 − QA QD ||F = ||V1 SR − QA QB ||F
3 To distinguish the GMD-based hybrid precoding from the SVD-based
(a)
precoding, we use Q1 , QA , and QD to denote the unconstrained GMD- = ||V1 − QA QD {SR }H ||F (17)
based precoding matrix, the GMD-based analog precoding matrix, and the
GMD-based digital precoding matrix, respectively. e D ||F ,
= ||V1 − QA Q
4

where (a) holds because the Frobenius norm is invariant under Require: H;
rotations, and Q e D = QD {SR }H . From (17), we can observe 1) Perform SVD of channel: [U1 Σ1 V1 ] = SVD(H);
that the optimal unconstrained precoding matrix in (10) will be 2) Initialize QA = Empty matrix, and Qres = V1
V1 , if we rotate QD as Qe D = QD {SR }H . So we reformulate For i ≤ NtRF
(10) as 3) Φ = AH t Qres ;

4) k = argmaxl=1,··· ,L (ΦΦH )l,l ;
Qopt , Qopt = arg min ||V1 − QA Qe D ||F ,
A D (k)
eD
QA ,Q 5) QA = [QA |At ];
†
1 6) QD = QA V1 ;
s.t. |{QA }i,j | = √ , (18) −QA QD
Nt 7) Qres = ||VV11−Q A QD ||F
;
  End for
tr QA Qe D QD Q
H eHA ≤ Ns .
8) Generate SR based on the algorithm in Lemma 1,
However, solving (18) is also highly complicated, since and QD = QD SR ;√
QA and Q e D are coupled, and the elemental-wise constraints 9) Normalize QD = Ns ||QAQQDD ||F ;
√
|{QA }i,j | = 1/ Nt is non-convex [5]. To this end, we 10) Return QA and QD ;
propose to solve to (18) by decoupling the design QA and Algorithm 1: The proposed GMD-based hybrid precoding
QD , i.e., treating QD as a fixed matrix while designing QA ,
and vice versa. Note that the main difficult lies in the design The overall algorithm for GMD-based hybrid precoding is
of the analog precoding matrix QA due to a non-convex summarized in Algorithm 1. Algorithm 1 can be divided
constraint on QA . To effectively design the analog precoding into two parts in general. The first part includes step 1 to
matrix, inspired by [5], we leverage the array response vectors step 7, which performs conventional spatially sparse hybrid
at (ϕti ), ∀i as the columns of QA , which can be explained precoding [5] based on the optimal unconstrained precoding
as follow. Firstly, at (ϕti ), ∀i has constant √ modulus elements, matrix V1 . The second part includes step 8 and step 9, which
which satisfy the constraint |{QA }i,j | = 1/ Nt . Secondly, the performs the GMD operation on the digital precoding matrix
columns of V1 form an orthogonal basis of the channel’s row QD and normalizes the effective precoding matrix to meet
space, while from (4), the array response vectors at (ϕti ), ∀i the transmitting power constraints. Note that in step 8, we
also form an orthogonal basis of the channel’s row space. do not need to explicitly compute SR , instead, we just apply
As a result, V1 can be expressed as a linear combination of the corresponding permutation and multiplication operations in
at (ϕti ), ∀i. In addition, the hybrid precoding matrix QA Q e D can each stage on QD . Note that the complexity of implementing
be also seen as a linear combination of QA . Since we want to GMD given SVD, i.e., Step 8, is only O((Ns + Nt )Ns ) [12],
approximate V1 via QA Q e D , it is reasonable to find the “best” since only a Ns by Ns rotation matrix SR is multiplied to
Nt array response vectors from at (ϕti ), ∀i as the columns of
RF QD , while the complexity of hybrid precoding algorithm, i.e.,
2
QA . So the corresponding analog precoding design problem Step 1-7, is O(NRF Nt Ns ), since we need to compute the
becomes pseudo-inverse matrix of QA , which indicates that the GMD-
T = arg min ||V1 − At TQ e D ||F , based hybrid precoding will only incur small extra complexity
T compared with conventional SVD-based hybrid precoding.
s.t. ||diag(TTH )||0 = NtRF , (19)
 
tr At TQ eH
e DQ H H
≤ Ns , IV. S IMULATION R ESULTS
D T At
In this section, we will evaluate the performance of the
where At = [at (φt1 ), at (φt2 ), · · · , at (φtL )] is an Nt ×L matrix, proposed GMD-based hybrid for mmWave massive MIMO
and T is a selecting matrix with NtRF non-zero rows. Problem communications. Consider typical mmWave massive MIMO
(19) is a sparse reconstruction problem, which can be effec- systems where Nt = 128-and Nt = 256-element ULAs with
tively solved via the principle of basis pursuit [5]. After we the antenna spacing d = λ/2 are utilized at the BS, while
determine the analog precoding matrix, the digital precoding Nr = 16-element ULA with also the antenna spacing d = λ/2
matrix design problem is a Frobenius norm minimization antennas is utilized at the user. The BS and the user both
problem: adopt NtRF = NrRF = 4 RF chains, and the carrier frequency
Qopt e is 28GHz. For the channel model, we utilize the SV model,
D = arg min ||V1 − QA QD ||F ,
eD
Q where the number of path is set to L = 4, the complex gain
  (20)
e D QH e H ≤ Ns . of each path follows the distribution CN (0, 1), and the AoAs
s.t. tr QA Q D QA
and AoDs are uniformly distributed in [−π/2, π/2]. Finally,
The optimal solution to (20) has a least square form [9] we use the water-filling power allocation scheme at the BS
e D = {QA }† V1 . [5], while the modulation scheme is 16QAM, which is utilized
Q (21)
in all sub-channels for both SVD and GMD-based precoding
Furthermore, we have schemes.
e D SR = {QA }† V1 SR . Fig. 3 shows the BER performance of the proposed GMD-
QD = Q (22)
based hybrid precoding in a 128 × 16 mmWave massive
which can be seen as projecting the GMD operation on the MIMO system, where we can observe that the proposed GMD-
e D to (20).
digital precoding matrix Q based precoding schemes (including fully digital GMD-based
 5

−1
10
array gains. However, the proposed GMD-based precoding
still outperforms the conventional SVD-based precoding. The
required SNR in GMD-based hybrid precoding to achieve
SVD−based precoding
BER = 10−3 is about 2.5dB and 3.5dB lower than that in the
−2
10

fully digital SVD-based precoding and the SVD-based hybrid

precoding [5]. In addition, the proposed GMD-based hybrid
BER

10
−3
precoding can still perform close to the the fully digital GMD-
based precoding with negligible gap.
GMD−based precoding

10
−4 V. C ONCLUSIONS
SVD−based hybrid precoding [5]
Fully digital SVD−based precoding In this paper, we propose a GMD-based hybrid precoding
Proposed GMD−based hybrid precoding
Fully digital GMD−based precoding
scheme for mmWave massive MIMO systems to avoid the
10
−5
complicated bit allocations in the conventional SVD-based
−30 −25 −20 −15 −10 −5 0
SNR (dB) hybrid precoding. With the help of SIC at the receiver, we can
convert the mmWave MIMO channel into multiple identical
Fig. 3. BER performance comparison of the proposed GMD-based hybrid sub-channels, which naturally do not need bit allocations.
precoding in a 128 × 16 mmWave massive MIMO system.
Specifically, we decouple the design of the analog and digital
−1 precoding matrix. The analog precoding matrix is designed
10
using the principle of basis pursuit, while we project the
GMD operation on the digital precoding matrix to obtain the
−2
digital precoding matrix. Simulation results verify that the
10
SVD−based precoding proposed GMD-based hybird precoding scheme outperforms
conventional SVD-based hybrid precoding schemes and is able
to achieve much better BER performance.
BER

−3
10

R EFERENCES
GMD−based precoding [1] T. Bai, A. Alkhateeb, and R. Heath, “Coverage and capacity of millimeter-
10
−4 wave cellular networks,” IEEE Commun. Mag., vol. 52, no. 9, pp. 70–77,
SVD−based hybrid precoding [5] Sep. 2014.
Fully digital SVD−based precoding [2] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors,
Proposed GMD−based hybrid precoding
and F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with
Fully digital GMD−based precoding
−5 very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60,
10
−30 −25 −20 −15 −10 −5 0 Jan. 2013.
SNR (dB) [3] L. Wei, R. Q. Hu, Y. Qian, and G. Wu, “Key elements to enable milli-
meter wave communications for 5G wireless systems,” IEEE Wireless
Fig. 4. BER performance comparison of the proposed GMD-based hybrid Commun., vol. 21, no. 6, pp. 136–143, Dec. 2014.
precoding in a 256 × 16 mmWave massive MIMO system. [4] S. Han, C.-L. I, Z. Xu, and C. Rowell, “Large-scale antenna systems
with hybrid precoding analog and digital beamforming for millimeter
wave 5G,” IEEE Commun. Mag., vol. 53, no. 1, pp. 186–194, Jan. 2015.
precoding and GMD-based hybrid precoding) can achieve a [5] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. Heath, “Spatially
better BER performance than the conventional SVD-based sparse precoding in millimeter wave MIMO systems,” IEEE Trans.
precoding (including fully digital SVD-based precoding and Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
[6] C. Rusu, R. Mendez-Rial, N. Gonzalez-Prelcicy, and R. W. Heath, “Low
SVD-based hybrid precoding [5]). This is because the GMD- complexity hybrid sparse precoding and combining in millimeter wave
based precoding converts the mmWave MIMO channel into MIMO systems,” in Proc. 2015 IEEE Int. Conf. Commun. (ICC), London,
multiple sub-channels with the identical SNR, which naturally UK, Jun. 2015, pp. 1340–1345.
[7] E. Zhang and C. Huang, “On achieving optimal rate of digital precoder
avoids the severe BER performance degradation over sub- by RF-baseband codesign for MIMO systems,” in Proc. 80th IEEE Veh.
channels with very lowest SNR for conventional SVD-based Technol. Conf. (VTC Fall), Sep. 2014, pp. 1–5.
precoding. Specifically, given the same BER performance, e.g., [8] C.-L. Chao, S.-H. Tsai, and T.-Y. Hsu, “Bit allocation schemes for mimo
equal gain precoding,” IEEE Trans. Wireless Commun., vol. 10, no. 5,
BER = 10−3 , the required SNR for GMD-based precoding pp. 1345–1350, Mar. 2011.
schemes is about −4dB, which is lower than that of conven- [9] X. Gao, L. Dai, S. Han, and R. W. Heath Jr, “Energy-efficient hybrid
tional SVD-based precoding. Furthermore, we can also see analog and digital precoding for mmwave MIMO systems with large
antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998–
that the performance gap between the fully digital GMD-based 1009, May. 2016.
precoding and the proposed GMD-based hybrid precoding is [10] Z. Gao, L. Dai, D. Mi, Z. Wang, M. A. Imran, and M. Z. Shakir,
negligible, which implies that the proposed GMD-based hybrid “Mmwave massive-MIMO-based wireless backhaul for the 5G ultra-
dense network,” IEEE Wireless Communications, vol. 22, no. 5, pp. 13–
precoding is able to well approximate the fully digital GMD- 21, Oct. 2015.
based precoding. [11] Y. Jiang, W. W. Hager, and J. Li, “The geometric mean decomposition,”
Fig. 4 is the BER performance of the proposed GMD- Linear Algebra and Its Applications, vol. 396, pp. 373–384, Feb. 2005.
[12] Y. Jiang, J. Li, and W. W. Hager, “Joint transceiver design for MIMO
based hybrid precoding in a 256 × 16 mmWave massive communications using geometric mean decomposition,” IEEE Tran. Sig-
MIMO system. In Fig. 4, the BER of the proposed GMD- nal Process., vol. 53, no. 10, pp. 3791–3803, Oct. 2005.
based precoding and the conventional SVD-based precoding
both improves, since more antennas at the BS provide more