This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
1

Near-Optimal Hybrid Processing for Massive

MIMO Systems via Matrix Decomposition
Weiheng Ni, Student Member, IEEE, Xiaodai Dong, Senior Member, IEEE, and Wu-Sheng Lu, Fellow, IEEE

Abstract—For practical implementation of massive multiple- For practical implementation of massive MIMO systems,
input multiple-output (MIMO) systems, the hybrid processing the number of antennas required for large antenna array gains,
(precoding/combining) structure is promising to reduce the high typically in the order of a hundred or more, is determined
implementation cost and power consumption rendered by large
number of radio frequency (RF) chains of the traditional pro- by examining the convergence properties over the antenna
cessing structure. The hybrid processing is realized through low- number [7]. However, to exploit such a large antenna array
dimensional digital baseband processing combined with analog in massive MIMO systems, the amplitudes and phases of
RF processing enabled by phase shifters. We propose to design the complex transmit symbols are traditionally modified at
hybrid RF and baseband precoders/combiners for multi-stream the baseband, and then upconverted to the passband around
transmission in massive MIMO systems, by directly decomposing
the pre-designed unconstrained digital precoder/combiner of a the carrier frequency after passing through radio frequency
large dimension. This approach is fundamental and general in (RF) chains (performing the analog radiowave/digital baseband
the sense that any conventional full RF chain precoding solution conversion, signal mixing, power amplifying). In this setting,
of a MIMO system configuration can be converted to a hybrid all outputs of the RF chains are connected to the antenna
processing structure by matrix decomposition. The constant elements, which means that the number of RF chains must
amplitude constraint of analog RF processing results in the
matrix decomposition problem non-convex. Based on an alternate exactly equal the number of antenna elements. Under this
optimization technique, the non-convex matrix decomposition circumstance the fabrication cost and energy consumption of
problem can be decoupled into a series of convex sub-problems such a massive MIMO system become unbearable due to the
and effectively solved by restricting the phase increment of tremendous number of RF chains [9].
each entry in the RF precoder/combiner within a small vicinity To deal with the aforementioned problem, smaller number
of its preceding iterate. A singular value decomposition based
technique is proposed to secure an initial point sufficiently close to of RF chains are used in the large scale MIMO systems,
the global solution of the original non-convex problem. Through with cost-effective variable phase shifters employed to handle
simulation, the convergence of the alternate optimization for the mismatch between the number of RF chains and of
such a matrix decomposition based hybrid processing (MD-HP) antennas [9], where high-dimensional analog RF (phase only)
scheme is examined, and the performance of the MD-HP scheme processing is enabled by using phase shifters while digital
is demonstrated to be near-optimal.
baseband processing is performed in a very low dimension.
Index Terms—Massive MIMO, hybrid processing, limited RF This type of structure is generally named hybrid processing
chains, matrix decomposition, alternate optimization.
in the literature. The limited number of RF chains in a hybrid
analog RF and digital baseband processing structure leads to
I. I NTRODUCTION less power consumption and higher energy efficiency com-
Massive multiple-input multiple-output (MIMO) is poten- pared with the conventional full RF chain configuration. There
tially one of the key technologies to achieve high capacity is extensive existing work on the precoding and combining
performance in the fifth generation of mobile cellular sys- design in a hybrid transceiver structure. The pioneering paper
tems [1]-[4]. In the limit of an infinite number of anten- [8] is one of the first to consider using variable phase shifter
nas, the massive MIMO propagation channel becomes quasi- network to compensate for the performance degradations in
static, where the effects of uncorrelated noise and fast fading a hard antenna selection system which enables limited RF
vanish, and such favorable characteristics enables arbitrarily chains. As the millimeter wave (mmWave) frequencies appear
small energy per transmitted bit [2]. Prominently, in massive more promising as candidates for future generation wireless
multiuser MIMO systems simple linear processing schemes, systems, massive MIMO becomes more relevant as smaller
such as zero-forcing (ZF) and linear minimum mean-square wavelengths lead to compact antenna footprint even for large
error (MMSE), are shown to approach the optimal capacity scale antenna arrays. The hybrid baseband and RF processing
performance achieved by the dirty paper coding in the down- (transmit precoding/receive combining) is particularly suitable
link communication [5]. The spectral efficiency performance for mmWave MIMO communications as it effectively reduces
of massive MIMO systems with several linear processing the excessive cost of RF chains [9]-[12]. Many hybrid de-
schemes, including ZF, MMSE and maximum-ratio combining sign in mmWave channels aim to employ the sparse nature
(MRC), with perfect or imperfect channel state information of mmWave channels. In the point-to-point (P2P) scenario,
(CSI) has been analyzed in [6]. the idea is to capture the “dominant” paths by choosing
RF control phases from array response vectors [9], [10],
W. Ni, X. Dong and W.-S. Lu are with the Department of Electrical
and Computer Engineering, University of Victoria, Victoria, BC V8W 3P6, while in multiuser systems, RF precoders are designed to
Canada. (e-mail: [email protected], [email protected], [email protected]). form multiple beams pointing to multiple users [11], [12].

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
2

For more general propagation environments including both on heuristic approaches and some using optimizations. One
rich scattering Rayleigh channels and mmWave channels, commonality of the majority hybrid processing schemes in
[13] proposed an equal gain transmission [14]-[15] based the literature [8]-[19] is that they approach the problem by
RF precoder design to realize large antenna gains, while directly designing the RF precoding matrix and the baseband
baseband processing performs in low-dimensional equivalent matrix within the given hybrid structure, regardless of their
channels, for single antenna multi-user systems. This work is detailed design methodology. In this paper, we tackle the
extended to multi-antenna multi-user systems in [16] where problem from a different perspective. The starting point is a
hybrid block diagonalization is proposed. Moreover, some known optimal or near optimal precoding/combining matrix
researchers design the hybrid RF and baseband precoders solution in the full RF chain scenario. This is reasonable as
employing certain optimization formulations, such as in [17] precoding/combining design in traditional MIMO structures
which maximizes the data rates of multi-users based on a bi- are well studied in the literature. Given this desired large di-
convex approximation approach and [18] for P2P systems. mension unconstrained digital precoder/combiner, we propose
More limited RF chain architectures that use subsets of to directly decompose the matrix into the product of a high
antennas with phase shifters or switches are studied in [19], dimension phase only RF matrix and a low dimension base-
which shows that antenna subset selection lowers the system band matrix as the hybrid solution. Similar to [9], we consider
complexity at the sacrifice of the performance. Another di- multi-stream transmission in P2P massive MIMO systems as
rection of reducing the massive MIMO implementation cost one example to develop our proposed matrix decomposition
is to use low resolution analog-to-digital converter (ADC) approach. The general goal of this paper and [9] is similar,
for each RF chain connected to an antenna [20]. Mixed i.e., designing the mixed analog/digital precoder/combiner.
high and low resolution ADCs can also be used to offer The methods used, however, are completely different. Our
trade off between performance and cost [21]. Both [20] and approach can be considered as a general solution that can
[21] demonstrate that using low resolution ADC degrades the convert any existing known precoder of a conventional full
performance but the performance loss can be compensated by RF chain MIMO structure into a hybrid structure. It is suited
using more antennas, however at an exponential scale increase for hybrid processing design over any MIMO channels and
of antennas, especially in the moderate to high SNR range. any MIMO system configurations. The orthogonal matching
Moreover, the large dimension digital processing after ADC pursuit scheme in [9] is efficient but valid only for mmWave
consume resources, which may be relatively moderate if very channels as it exploits the sparsity in the limited scattering
low resolution matrices are involved. mmWave channels. Moreover, it requires to know not only
CSI is the prerequisite to perform any processing at trans- the channel matrix, but also each multipath component in-
mitter and receiver, whether it is a type of unconstrained high- cluding the corresponding path gain and the array responses
dimensional baseband processing for the traditional design of AoA/AoD. Moreover, [9] limits the columns of its RF
with one antenna element coupled with one dedicated RF chain precoder/combiner to take the form of array response vectors
or it is a type of hybrid processing. In [22], training sequences of AoD/AoA, thus narrowing the feasible region into a smaller
and closed-loop sounding vectors are designed to estimate a one and degrading the performance compared to our approach.
massive multiple-input single-output (MISO) channel through Recently, [27] also treats the hybrid design problem from a
the alignment of transmit beamformer with true channel di- matrix decomposition perspective. A subspace decomposition
rection. In [23], a compressive sensing (CS) based low-rank based approach is presented, which has a different RF pre-
approximation problem for estimating massive MIMO channel coder/combiner design in each iterative step from this work.
matrix is formulated, and is solved by semidefinite program- While the idea is original as in this work, [27] acknowledges
ming. Considering the massive MIMO channels with limited that it is extremely difficult to make claims on the convergence
scattering feature (especially when mmWave channels are of its algorithm, while our proposed scheme can securely
involved), the parameters of paths, such as the angles of depar- converge to a local optimum.
ture (AoDs), angles of arrival (AoAs) and the corresponding Starting with an optimal unconstrained precoder built on
path loss are estimated by designing beamforming codebook a set of right singular vectors (associated with the largest
so as to obtain the pathloss of all paths whose AoDs/AoAs singular values) of the channel matrix, our hybrid precoders
are spatially quantized in the entire angular domain [24], [25], are designed by minimizing the Frobenius norm of the matrix
while [25] performs the beamforming in a hybrid processing of difference between the unconstrained precoding matrix and
setting. Furthermore, some adaptive CSI acquisition schemes products of the hybrid RF and baseband precoding matrices,
with reduced time slot overhead are proposed exploiting the subject to the aforementioned constraints on the RF precoder.
spatial sparsity of the massive MIMO channels in [26]. In this Technically, solving this matrix decomposition problem is
paper however, as in much other literature, perfect CSI, i.e., rather challenging because it is a highly nonconvex constrained
the channel matrix is assumed known during the design of the problem involving a fairly large number of design param-
hybrid precoder/combiner, and channel estimation is out of the eters. Here we present an alternate optimization technique
scope of this paper. to approach the solution in that the hybrid precoders are
Designing hybrid precoders/combiners is a difficult problem alternately optimized in a relaxed setting so as to ensure all
which needs to be tackled for every different system config- sub-problems involved are convex. We stress that the convex
uration involving number of users, antennas, RF chains, data relaxation technique utilized here includes not only properly
streams, etc. Many possible solutions exist, with some based grouping design parameters for alternate optimization, but

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
3

also restricting the phase increment of each entry in the RF The transmitted symbols are processed by a baseband
precoder to within a small vicinity of its preceding iterate. precoder FB of dimension Mt × Ns , then up-converted to the
Under these circumstances, it is critical to start the proposed RF domain through the Mt RF chains before being precoded
decomposition algorithm with a suitable initial point that is by an RF precoder FR of dimension Nt × Mt . Note that
sufficiently close to the global solution of the original non- the baseband precoder FB enables both amplitude and phase
convex matrix decomposition problem. To this end a singular- modifications, while only phase changes can be realized by
value-decomposition (SVD) based technique is proposed to FR as it is implemented by using analog phase shifters. We
(i,j)
secure a satisfactory initial point that with high probability normalize each entry of FR to satisfy |FR | = √1N , where
t
allows our decomposition algorithm to yield near-optimal |(·)(i,j) | denotes the amplitude of the (i, j)-th element of (·).
hybrid precoders. Concerning the hybrid combiners design, Furthermore, to meet the constraint on total transmit power,
a linear MMSE combiner is selected as the unconstrained FB is normalized to satisfy ||FR FB ||2F = Ns , where || · ||F
reference matrix for matrix decomposition, and the hybrid RF denotes the Frobenius norm [28].
and baseband combiners can be obtained in the same way as We assume a narrowband flat fading channel model and the
the hybrid precoder design. received signal before analog RF processing is given by
We remark that the proposed matrix decomposition method,
termed MD-HP, provides a general solution to convert any y = HFR FB s + n, (1)
existing known precoder/combiner of a conventional full RF
where y ∈ CNr ×1 is the received signal vector, s ∈ CNs ×1
chain MIMO structure into a hybrid structure. It is suited for
is the signal vector such that E[ssH ] = NPs INs where (·)H
hybrid processing design over any MIMO channels and any
denotes conjugate transpose, E[·] denotes expectation, INs is
MIMO system configurations. Simulations are presented to
the Ns × Ns identity matrix and P is the average transmit
examine the convergence of the alternate optimization for the
power. H ∈ CNr ×Nt is the channel matrix, normalized as
MD-HP scheme and to demonstrate its near-optimal perfor-
E[||H||2F ] = Nt Nr , and n is the vector of i.i.d. CN (0, σ 2 )
mance by comparing to the optimal unconstrained baseband
addictive complex Gaussian noise. To perform the precoding
processing based on the SVD technique.
and combining, we assume the channel is known at both the
transmitter and the receiver, thus the processed received signal
II. S YSTEM M ODEL after combining is given by
In this section, we introduce the hybrid processing structure H
ỹ = WB H
WR H
HFR FB s + WB H
WR n, (2)
for P2P massive MIMO systems and the channel models
considered in this paper. where WR is the Nr × Mr RF combining matrix and WB
is the Mr × Ns baseband combining matrix. Since WR is
also implemented by the analog phase shifters, all elements
A. System Model of WR are constrained to have constant amplitude such that
(i,j)
We consider a communication scenario from a transmitter |WB | = N1r . If Gaussian inputs are employed at the
with Nt antennas and Mt RF chains to a receiver equipped transmitter, the instantaneous spectrum efficiency achieved
with Nr antennas and Mr RF chains, where Ns data streams shall be
are supported. The system model of the transceiver is shown
 
 P −1 H

in Fig. 1, employing the full-array structure proposed in R(FR , FB , WR , WB ) = log2 INs +
 Rn H̃H̃  ,
Ns
[9], where each RF chain output is phase shifted multiple (3)
times and combined to generate Nt different copies before where Rn = σ 2 WB H
WR H
WR WB is the covariance matrix of
being fed to Nt transmit antennas, and similarly the signals H H
the noise and H̃ = WB WR HFR FB .
gathered from all the receive antennas are phase shifted prior to
combining for baseband processing at the receiver1 . To ensure
B. Channel Model
effectiveness of the communication driven by the limited
number of RF chains, the number of the communication In this paper, we seek to find optimal hybrid precoders (FR ,
streams is constrained to be bounded by Ns ≤ Mt ≤ Nt FB ) as well as hybrid combiners (WR , WB ) based on a
for the transmitter and by Ns ≤ Mr ≤ Nr for the receiver. general channel matrix H. To measure the performance of our
MD-HP scheme, we examine two types of channel models in
simulation studies to be presented in Section IV, namely
RF Chain 1 1) Large Rayleigh fading channel Hrl with i.i.d. CN (0, 1)
...

Analog
...

Digital ...
Data Baseband RF entries;
Streams Processing RF Chain P Processing
Nt / N r
2) Limited scattering mmWave channel Hmmw .
...

...
...

FB / WB FR / WR We remark that several hybrid processing schemes for

RF Chain0
mmWave communications have been studied, where a large
antenna array is implemented to combat high free-space
Fig. 1. System model of the transceiver with the hybrid processing structure
pathloss and reflection loss [10]-[12]. Thus the mmWave chan-
1 In Fig. 1, the number of RF chains M represents M and M respectively
t r
nel model Hmmw is an appropriate instance for comparing the
for transmitter and receiver performance of the proposed scheme with related recent work

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
4

in the literature. Because of the limited (sparse) scattering optimization problem to maximize the spectral efficiency,
characteristic of a mmWave channel, we decide to introduce which is given by
a clustered mmWave channel model to characterize its key
max R(FR , FB , WR , WB )
features [29]. The mmWave channel Hmmw is assumed to
be the sum of all propagation paths that are scattered in s.t. ||FR FB ||2F = Ns , (6)
Nc clusters with each cluster contributing Np paths. Under FR ∈ FR , WR ∈ WR ,
these circumstances, the normalized channel matrix can be
where FR (WR ) is the set of matrices with all constant
expressed as
amplitude entries, which is √1N ( √1N ). However, this type
t r
s Np
Nc X of joint optimization problems is often intractable [30], due
Nt Nr X
Hmmw = αil ar (θil )at (φil )H , (4) to the presence of non-convex constraints FR ∈ FR and
Nc Np i=1 WR ∈ WR that obstruct the regular progress of securing a
l=1
globally optimal solution. Before gaining an insight into the
where αil is the complex gain of the i-th path in the l-th solution of this joint optimization problem (6), we introduce
cluster, which follows CN (0, 1)2 . For the (i, l)-th path, θil and the optimal unconstrained precoder F? and combiner W? for
φil are the azimuth angles of arrival/departure (AoA/AoD), achieving maximum capacity of a general MIMO channel,
while ar (θil ) and at (φil ) are the receive and transmit array based on which a procedure for the design of near-optimal
response vectors at the azimuth angles of θil and φil respec- hybrid precoders/combiners is developed. Assume that the
tively, and the elevation dimension is ignored3 . Within the i-th channel matrix H is well-conditioned to transmit Ns data
cluster, θil and φil have the uniformly-distributed mean values streams, namely, rank(H) ≥ Ns . To obtain the optimal
of θi and φi respectively, while the lower and upper bounds F? and W? , we perform the SVD of the channel matrix
of the uniform distribution for θi and φi can be defined as H = UΣVH , where U and V are Nr × Nr and Nt × Nt
[θmin , θmax ] and [φmin , φmax ]. The angular spreads (standard unitary matrices, respectively, and Σ is an Nr × Nt diagonal
deviations) of θil and φil among all clusters are assumed to matrix with singular values on its diagonal in descendant order.
be constant, denoted as σθ and σφ . According to [9], we use Without incorporating the waterfilling power allocation, the
truncated Laplacian distribution to generate all the θil ’s and optimal unconstrained precoder and combiner are given by
φil ’s based on the above parameters.
As for the array response vectors ar (θil ) and at (φil ), we F? = V1 , W? = U1 , (7)
choose uniform linear arrays (ULAs) in our simulations, while
where V1 and U1 are constructed with the first Ns columns
the precoding scheme to be developed in Section III can
of V and U, respectively, and the corresponding spectral
directly be applied to arbitrary antenna arrays. For an N -
efficiency by using such unconstrained F? and W? is given
element ULA, the array response vector can be given by
by  
1 h j 2π d sin(θ) iT  γ 2 
aULA (θ) = √ 1, e λ
2π
, · · · , ej(N −1) λ d sin(θ) , R̃ = log2 INs +
 Σ , (8)
Ns 1 
N
(5) where Σ1 represents the first partition of dimension Ns × Ns
where λ is the wavelength of the carrier, and d is the dis- of Σ by defining that
tance between any two adjacent antenna elements. The array  
response vectors at both the transmitter and the receiver can Σ1 0
Σ= , (9)
be written in the form of (5). In addition, due to the generality 0 Σ2
of the proposed decomposition scheme, its effectiveness is where γ = σP2 is the signal-to-noise ratio (SNR).
not just limited to the above channel models of Rayleigh
Actually, R̃ sets an upper bound for the spectral efficiency
fading channels and limited-scattering mmWave channels. No
R(FR , FB , WR , WB ) in problem (6) where the ranges of
matter what kind of antenna array the channel is associated
the matrix products FR FB and WR WB are respectively
with, such as ULA, uniform planar array (UPA), uniform
the subsets of feasible regions of the unconstrained precoder
circular array (UCA) and so on, we can decompose the desired
and combiner, namely, CNt ×Ns and CNr ×Ns . Considering
precoder/combiner by directly applying the proposed scheme
the non-convex nature of the problem (6), it is impractical
as long as the channel matrix is known.
to insist upon securing its global solution. One apparently
viable approach is to construct hybrid precoders (FR , FB ) and
III. H YBRID P RECODING /C OMBINING D ESIGN FOR A combiners (WR , WB ) such that the optimal unconstrained
G ENERAL M ASSIVE MIMO C HANNEL precoder F? and combiner W? can be sufficiently closely
approached by FR FB and WR WB respectively. In what
The design of hybrid precoders (FR , FB ) and combiners follows, the design of such hybrid precoders is substantiated
(WR , WB ) based on a general massive MIMO channel H via matrix decomposition.
may be achieved by formulating a joint transmitter-receiver
A. Hybrid Precoders Design via Matrix Decomposition
2 Thepower gain of the channel matrix is normalized such that
E[||Hmmw ||2F ] = Nt Nr Given hybrid precoding structure and constraint on RF
3 Only 2D beamforming is considered in this mmWave channel model. precoder FR , there is no guarantee that a pair (FR , FB ) can be

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
5

found such F? = FR FB holds exactly. However, by relaxing Alternate minimization is an iterative procedure with each
the strict equality in (6), the matrix decomposition can be iteration be carried out in two steps. In each of these steps one
accomplished through reformulating the original problem as set of design parameters are fixed while the objective function
min ||F? − FR FB ||F is minimized with respect to the other set of parameters
FR ,FB and the role of design parameters alternates as the design
s.t. ||FR FB ||2F = Ns , (10) step switches. For the design problem at hand, naturally the
FR ∈ FR . components in FR and those in FB are the two parameter sets,
the alternate minimization is performed as follows: 1) solve
To look closely at the physical implication of this problem problem (10) with respect to FB with FR given; and 2) solve
re-formulation, recall that our design objective is essentially to problem (10) with respect to FR with FB given.
approximate F? by the product of hybrid precoding matrices, We begin by examining a simplified version of problem
namely FR FB . Thus a natural question arising at this point (10) by temporarily removing the normalization constraint
is how sensitive the spectral efficiency R(FR , FB , WR , WB ) ||FR FB ||2F = Ns , which leads (10) to
to any deviation of FR FB from F? , because small residue
||F? −FR FB ||F at a solution of problem (10) is inevitable and min ||F? − FR FB ||F
FR ,FB
this residue would divert the optimal unconstrained combiner (13)
W? away from the SVD-based solution U1 . s.t. FR ∈ FR .
Bearing the analysis above on mind, we begin the design
Denote the hybrid precoders at the k-th iteration by
of hybrid precoders by assuming that the Nr –dimensional (k) (k) (0)
(FR , FB ), and assume the initial FR is given. We up-
minimum distance decoding can be performed at the receiver, (k)
which implies that the achieved spectral efficiency is equiva- date FB by solving the unconstrained convex problem
(k) (k)
lent to the mutual information over the MIMO channel when minFB ||F? −FR FB ||F whose closed-form solution is given
Gaussian inputs are used, which is given by by
 
γ (k) (k) H (k) (k) H
H H H FR )−1 FR F? , k = 0, 1, 2, · · · .
 
I(FR , FB ) = log2 INs +
 HFR FB FB FR H  . FB = (FR (14)
Ns
(11) In turn, we update the RF precoder to FR by solving the
(k+1)
Next, we obtain the hybrid precoders by maximizing the (k)
non-convex problem below while FB is given as a constant
mutual information between the transmit data streams and the
matrix:
receive antennas in (11). The problem of mutual information (k+1) (k)
maximization problem has been investigated in [9], where the min ||F? − FR FB ||F
(k+1)
FR (15)
mutual information is approximated as (k+1)
s.t. FR ∈ FR .
I(FR , FB )
(k+1)
 
γ 2  (12) To deal with the nonconvex constraint FR ∈ FR in (15),
Σ1  − Ns + ||V1H FR FB ||2F ,

≈ log2 INs + (k) (k)
Ns we update FR with a local search in a small vicinity of FR .
(k) (k) (k)
and max I(FR , FB ) ≈ max ||V1H FR FB ||2F is approximately Denote the phase of the (m, n)-th entry of FR as φm,n , FR
(k)
equivalent to minimizing ||F? − FR FB ||F . Consequently, can be represented as √1N {ejφm,n }, m = 1, · · · , Nt , n =
t
designing (FR , FB ) so as to maximize the mutual information 1, · · · , Mt . To characterize the relation between FR
(k+1)
and
over the massive MIMO channel can be accomplished by (k) (k+1)
FR , we write FR as
solving the matrix decomposition problem (10). Once the
hybrid precoders (FR , FB ) are optimized, we can proceed (k+1) 1 (k+1) 1 (k) (k)
FR = √ {ejφm,n } = √ {ej(φm,n +δm,n ) }, (16)
to design the hybrid combiners (WR , WB ) to maximally Nt Nt
increase the system’s spectral efficiency.
(k)
The second constraint in (10) requiring that the entries of where δm,n is the phase increment of the (m, n)-th entry of
(k) (k) (k)
FR have constant amplitude √1N is evidently non-convex, FR . Note that the approximation ejδm,n ≈ 1 + jδm,n holds
t
which prevents the use of efficient convex optimization algo- (k) (k)
as long as |δm,n | is sufficiently small, e.g. |δm,n | ≤ 0.1. Based
rithms and makes it extremely challenging to secure a globally on Taylor’s expansion, therefore, we have
optimal solution. Under the circumstance, our design searches
for a near-optimal solution so that the spectral efficiency (k+1) 1 (k) (k)
FR ≈ √ {(1 + jδm,n )ejφm,n }
achieved by the obtained hybrid precoders (as well as the Nt
hybrid combiners) is comparable with the upper bound R̃. The (k) j (k) (k)
= FR + √ {δm,n · ejφm,n } (17)
design method described below has three main ingredients: it Nt
employs an alternate optimization strategy that separates the (k) (k) j (k)
= FR + {δm,n } ◦ √ {ejφm,n },
two sets of design parameters in a natural manner; a local Nt
convexification technique ensures that each sub-problem be
(k) (k)
solved in a convex setting; and the use of a carefully chosen where {δm,n } is the matrix whose (m, n)-th entry is δm,n
initial point that facilitates the alternate iterates to converge to and “◦” denotes the Hadamard product (entrywise product).
(k+1)
a satisfactory design. It follows that the problem in (15) for seeking FR can

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
6

be reformulated as an optimization problem with respect to scenario b) reduces δ̄ (k) to a smaller δ̄ (k+1) when εk increases
(k)
{δm,n } as due to that the previous large phase increment has made the
(k) (k)
    approximation ejδm,n ≈ 1+jδm,n invalid, or when εk is close
 ? (k) (k) j jφ(k) (k) 
min F − FR + {δm,n } ◦ √ {e
  m,n } FB  to the required ε̄ suggesting that higher precision is required.
(k)
{δm,n } Nt F In Section IV we shall come back to this matter again in terms
2
of specific adjustments on δ̄ (k) .
  
j (k) (k) 
⇔ min Q(k) − {δm,n (k)
} ◦ √ {ejφm,n } FB  ,

(k)
{δm,n } Nt 3) Re-formulation of Problem (19): Another issue concern-
F
(18) ing problem (19) is that its formulation in terms of Hadamard
(k) (k) product is not suited for many convex-optimization solvers
where Q(k) = F? − FR FB . We remark that problem (18)
has a convex quadratic objective function, and the constant that require standard and explicit formulations. Denote the p-
(k)
amplitude constraint FR
(k+1)
∈ FR has been taken into account th row of Q(k) by qp , we can write the objective function
because FR
(k+1)
here assumes the form of √1N {ej(φm,n ) }.
(k+1)
in (19) as
t
However, the above formulation is based on the approximation 
Nt  
" (k) (k)
# 2
(k) (k) (k) jδp,1 jφ(k) jδp,Nt jφ(k) 
ejδm,n ≈ 1 + jδm,n , hence it is valid only if |δm,n | is suffi- (k) 
X  (k)
q p − √ e p,1 ,··· , √ e p,M t FB 
ciently small. Therefore, linear constraints on the smallness of 
p=1
Nt Nt 
2
(k)
|δm,n | need to be imposed, thus problem (18) is modified to Nt  
X 2
 (k)
   2 = qp − ∆(k) (k) 
p Gp  ,
j (k) (k)  2
min Q(k) − {δm,n (k)
} ◦ √ {ejφm,n } FB 

p=1
(k)
{δm,n } Nt F (19) (20)
(k) (k) (k) (k) (k)
(k) (k) where ∆p = [δp,1 , δp,2 , · · · , δp,Mt ] and Gp =
s.t. |δm,n | ≤ δ̄ , ∀m, n,  (k) (k)

j jφp,1 , · · · , e
jφp,M (k)
(k)
√
N
diag e t FB , hence
t
where δ̄ (k) > 0 is sufficiently small such that ejδm,n ≈ 1 +
(k) Nt  
jδm,n holds. Problem (19) is a convex quadratic programming X  (k)
2
min − ∆(k) G(k)

(QP) problem whose unique global solution can be calculated (k)
 q p p p 
2
(k) {δm,n } p=1
efficiently [32]. Once the solution {δm,n } is obtained, the (21)
(k+1) Nt
FR can be updated by (16). X  2
= min q(k) − ∆(k) (k) 
Gp  .

There are several issues that remain to be addressed. These (k)
p p
2
p=1 ∆p
include defining an error measure to be used in the algorithm’s
stopping criterion and elsewhere; selection of a good initial It follows that problem (19) can be solved by separately
point to start the algorithm; adaptive thresholding for phase solving Nt sub-problems
(k)
increments δm,n and derivation of an explicit formulation for  2
problem (19); and a treatment of the constraint ||FR FB ||2F = min q(k)

p − ∆ (k) (k) 
p Gp 
(k) 2
Ns in problem (10). ∆p (22)
1) An Error Measure: The relative distance between F? (k) (k)
s.t. |δp,n | ≤ δ̄ , n = 1, 2, · · · , Mt ,
(k) (k)
(k) (k) ||F? −FR FB ||F
and FR FB , namely εk = ||F? ||F , will be used as
for p = 1, 2, · · · , Nt . Note that each problem in (22) is an
an error measure. In the proposed algorithm, alternate itera-
explicitly formulated convex QP problem to which efficient
tions continue until εk falls below a prescribed convergence
interior-point algorithms apply [32]. To implement the interior-
tolerance ε̄ (or the iteration index k exceeds a preset upper
(k) (k) point method for a convex optimization problem with inequal-
bound Ku ), and when this occurs, the last iterate (FR , FB )
ity constraints, we can use the logarithmic barrier function
is taken to be a solution of problem (13).
(k) to incorporate those inequality constraints implicitly in the
2) Adaptive Thresholding for Phase Increments δm,n : The objective function, and then obtain the solution through the
constraints on the magnitude of phase increment in (19) limit iterations for searching the global optimum along one central
(k+1) (k)
FR to within a small neighborhood of FR that usually path inside the feasible region based on Algorithm 11.1 in
affects the algorithm’s convergence rate. This is however less [32].
problematic for (19) because the effective range for each phase 4) Choosing an Initial Point: Choosing an appropriate
parameter in the RF precoder is limited to [0, 2π). In addition, initial point to start the proposed algorithm is of critical
the issue can be addressed by making the upper bound importance because the original problem (10) is a non-convex
(threshold) in (19) adaptive to the current error measure so as problem which typically possesses multiple local minimizers.
to improve the algorithm’s convergence rate. The adaptation As far as gradient-based optimization algorithms are con-
of threshold δ̄ (k) is performed as follows: cerned, the likelihood of capturing global minimizer or a good
a) set δ̄ (k+1) slightly larger than δ̄ (k) if εk is far greater than local minimizer is known to be highly dependent on how close
ε̄, and εk < εk−1 holds; the initial point to the desired solution.
b) set a smaller δ̄ (k+1) than δ̄ (k) if εk is close to ε̄, or εk ≥ Note that the objective function in (10), namely ||F? −
εk−1 holds. FR FB ||F , measure the difference between the optimal un-
Scenario a) allows a larger phase increment while the algo- constrained RF precoder F? and an actual decomposition
rithm converges in the right direction (εk is decreasing), while FR FB in the feasible region. If we temporarily neglect the

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
7

constant amplitude constraint on FR , the perfect decompo- B. Hybrid Combiners Design

sition of F? can be performed through SVD decomposition
F? = UF ΣF VFH . This motivates us to construct an initial The hybrid precoders are designed under the assumption
point based on the SVD of F? . As F? comes from the first that the Nr –dimensional minimum distance decoding can be
Mt right singular vectors of the channel matrix H, F? has performed at the receiver. However, such a decoding scheme is
the full column rank, which means all Ns entries along the difficult to implement in practice due to its high complexity.
diagonal of ΣF are non-zero. Note that UF ΣF is an Nt × Ns In this paper, we employ linear combining at the receiver.
matrix with full column rank, VFH is an Ns × Ns matrix and As we know, if the hybrid precoders would be equivalent to
FR consists of NtRF columns. To construct an initial point that the unconstrained optimal precoder F? = V1 , the optimal
conforms to the dimensions of FR and FB , we generate an unconstrained combiner W? would be U1 . However the error
Nt ×(Mt −NS ) matrix F̂R where the amplitude of each entry ||F? − FR FB ||F can never be absolutely zero, hence U1
is equal to √1N and the phase of each entry obeys a uniform deviates from the optimal unconstrained combiner W? . The
t
distribution over [0, 2π). In this way, a decomposition of F? linear MMSE combiner WM M SE will achieve the maximum
is found to be spectral efficiency for linear and separate detection of each
 H  data stream, which would be a better solution compared with
VF that in [33]. The unconstrained linear MMSE combiner is
F? = [UF ΣF F̂R ] , (23)
0 given in [34] as

and (FR = [UF ΣF F̂R ], FB = [VF 0]H ) is exactly a W? = WM M SE = arg min E [||s − Wy||2 ]
W
global solution for min ||F? − FR FB || when no constraints √  −1
are imposed. We stress that FR = [UF ΣF F̂R ] is infeasible P P
= HFR FB FH F
B R
H H
H + σ 2
I Nr HFR FB .
when the constant amplitude constraint of FR is imposed. Ns Ns
(0) (24)
Nevertheless, we can select a feasible initial point FR that
Once W? is obtained, the designed hybrid combiners WR
is close to the above [UF ΣF F̂R ] by modifying the first
and WB can also be obtained through the decomposition of
partition UF ΣF as follows:
W? , which leads to the problem
1) retaining the phases of all entries in UF ΣF ;
2) enforcing the amplitudes of all entries in UF ΣF into min ||W? − WR WB ||F
(0) WR ,WB
√1 to make FR feasible. (25)
N t
s.t. WR ∈ WR .
Since the modified UF ΣF still incorporates the information
of the phases in decomposition (23), it is intuitively clear that
(0) Replacing the symbols, F∗ , FR and FB , in Algorithm 1
the FR generated above is reasonably near the global solution by W∗ , WR and WB , the alternate optimization method
of problem (13) and for this reason we shall chose it as the presented in Section III-A can be directly applied to solve
initial point for the proposed algorithm. the above decomposition problem (25). We will evaluate
Finally, the constraint ||FR FB ||2F = Ns in the original the performance of the proposed hybrid processing scheme
matrix decomposition problem (10) is treated by performing
√ through simulations in Section IV.
a normalization step where FB is multiplied by ||FR FNBs ||F .
The normalization assures that the transmission power remains
consistent after precoding. A step-by-step summary of the
hybrid precoder design is given below as Algorithm 1. C. Approach To Waterfilling Spectral Efficiency

To reach a capacity-achieving processing scheme, the wa-

Algorithm 1 The Hybrid Precoders Design via Matrix De- terfilling power allocation should be applied to the precoder.
composition based on Alternate Optimization In this case, the optimal unconstrained precoder and combiner
Require: F? , FR , Ku and ε̄
(0) in Section III are updated to F? = V1 Γ and W? = U1
(0) (0) H (0) (0) H respectively, where Γ is a diagonal matrix that performs the
1: FB = (FR FR )−1 FR F?
||F ? (0) (0)
−FR FB ||F
waterfilling power allocation. The precoder so produced can
2: ε0 = , ε−1 = ∞
||F? ||F directly be decomposed through Algorithm 1. However, there
3: k=0 may be cases where no power is allocated to some data
4: while εk > ε̄ and k ≤ Ku do streams corresponding to the lowest singular values of H,
5: k =k+1 especially when the SNR is small. In other words, we may
(k)
6: obtain FR by solving (15) end up with F∗ = [F0 , 0], where F0 is the non-zero columns
(k) (k) H (k) (k) H
7: FB = (FR FR )−1 FR F? of F? = V1 Γ after waterfilling power allocation. In this
||F? (k) (k)
−FR FB ||F case, we can apply the MD-HP scheme to the F0 part first,
8: εk = ||F? ||F F0 = FR F0B . And then the whole decomposition for F∗ is
9: end while
√
given by F∗ = [F0 , 0] = [FR F0B , 0] = FR [F0B , 0] = FR FB .
10: FB = ||FRNFs F B
B ||F In this way, the zero-power allocation part is realized through
11: return FR , FB
the baseband precoding rather than the phase shift in the RF
domain.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
8

D. An Analysis of Complexity and Convergence examined in the case of multiple single-antenna users in [13],
Below we present an analysis concerning complexity and while the azimuth and elevation angles are quantized to limit
convergence matters for the hybrid precoders design described feedback overhead in [9].
in Algorithm 1.
1) Complexity: The complexity of the alternate optimiza- IV. S IMULATION R ESULTS
tion is dominated by the calculations of the baseband and In this section, we report the results of the simulations
RF precoders/combiners. In each iteration, baseband pre- conducted, where the convergence of the proposed matrix
coder is computed using (14), which can be expressed as decomposition method based on alternate optimization are ex-
(k) (k)
FB = (FR )† F? where (·)† denotes the Moore-Penrose amined and the performance of the proposed MD-HP scheme
pseudo-inverse. Thus the complexity of (14) is equal to the are evaluated.
complexity of computing the pseudo-inverse of an Nt × Mt
matrix, which is O(Nt Mt2 ) for a Cholesky factorization based A. Implementation Issues for Algorithm 1
method [35], plus that of computing the product of a matrix Before we apply Algorithm 1 to design the hybrid precoders
of size Mt × Nt with a matrix of size Nt × Ns , which is and combiners, it is necessary to examine whether it will
O(Nt Mt Ns ). Since Ns ≤ Mt , the complexity of (14) is in the converge to a level where the error εk is acceptably small,
order of O(Nt Mt2 ). Concerning the complexity of updating this is because the original optimization problem (10) to be
(k+1)
the RF precoder to FR , that involves solving a total of solved is non-convex and there is no guarantee that Algorithm
Nt convex QP problems, reliable polynomial-time interior- 1 will certainly result in a satisfactory matrix decomposition.
point algorithms exist [36], with complexity in the order of We took a 256 × 64 MIMO system as example, and set
O(Mt3 ), for each convex QP problem as seen in (22). Hence Ns = 4, Mt = 6. An i.i.d Rayleigh fading channel matrix Hrl
the complexity for this part of the iteration is of O(Nt Mt3 ). with each entry obeying CN (0, 1) was randomly generated.
Consequently, the complexity of a single iteration in Algorithm From Section III-A, the optimal unconstrained precoder F?
1 is O(Nt Mt3 ), and the complexity of Algorithm 1, assuming was obtained by selecting the first Ns right singular vectors
its convergence in K iterations, is in the order of O(KNt Mt3 ). based on the SVD decomposition on Hrl . An initial RF pre-
2) Convergence: The alternate minimization employed by (0)
coder FR was chosen by following the technique described
Algorithm 1 ensures that it is a descent algorithm in the in Section III-A4. The tolerance was set to ε̄ = 0.1 with
sense that the objective function is monotonically decreasing Ku = 100 and the first phase increment threshold was set
as the iterations proceed. Moreover, the objective function to δ̄ (1) = 0.1, in order to observe the convergence property
has a nonnegative lower bound as it is a matrix norm. It of Algorithm 1 with an enough number of iterations. In the
follows that the objective function’s values at its iterates simulations, two options for δ̄ (k) were examined:
constitute a convergent sequence approaching to a certain
1) δ̄ (k) =  0.1, ∀k;
non-negative minimum. Given a tolerance ε̄, Algorithm 1 is
(k) 1.25 · δ̄ (k−1) , when |εk−1 − ε̄| > 0.15
claimed to converge as the relative error εk (see Step 7 of 2) δ̄ = .
0.8 · δ̄ (k−1) , when |εk−1 − ε̄| ≤ 0.15
the algorithm) falls below ε̄. The number of iterations K
For option 2) with adaptive phase increment threshold, the
required for the algorithm to converge is closely related to
adjustment of δ̄ (k) depends on how close the previous error
the selection of phase increment threshold δ̄ (k) in (22), and a
(k) indicator εk−1 and its tolerance ε̄ are. When the difference
strategy to adaptively thresholding the phase increment δm,n
|εk−1 − ε̄| is smaller than 0.15, which means Algorithm 1
in each iteration so as to reduce K has been proposed in
is going to converge, δ̄ (k) should be reduced to enhance the
Section III-A2. We shall revisit this issue in Section IV when
precision of the solution by guaranteeing the effectiveness of
presenting our simulation results. (k) (k)
the approximation ejδm,n ≈ 1 + jδm,n . Otherwise, δ̄ (k) can
be augmented to accelerate the algorithm by enlarging the
E. Quantized RF Phase Control feasible region of (19). Moreover, we need to decrease δ̄ (k)
It is difficult to assign arbitrary value to the phase of each whenever εk−1 > εk−2 which means the previous δ̄ (k−1) is
(k−1) (k−1)
entry in the RF precoder FR or combiner WR is difficult to too large to guarantee ejδm,n ≈ 1 + jδm,n . We restricted
be set to be an arbitrary value due to the limited precision in (k) (k)
δ̄ ∈ [0.1, 0.5] by clamping δ̄ to 0.1(0.5) when it was
the practical implementation. To address the problem here, we smaller (larger) than 0.1(0.5) in case that the feasible region
also introduce the quantized phase implementation of FR and for (19) was too small or too large4 .
(k+1)
WR . Assume the phase of each entry in FR and WR can To examine the effectiveness of approximation ejφm,n =
(k) (k) (k)
be quantized up to L bits of precision by choosing the closet (k)
ej(φm,n +δm,n ) ≈ (1 + jδm,n )ejφm,n , we compare the traces
neighbor based on the shortest Euclidean distance, which is (k) (k) (k) (k)
of ej(φm,n +δm,n ) and (1 + jδm,n )ejφm,n within Ku = 100
given by iterations in Fig. 2, where the red dash line indicates the
2πn̄
φ= L , (26) unit circle on the complex plane. It is observed that the
2
  points of the two traces (m = 1, n = 5) update simul-
where n̄ = arg minn∈{0,··· ,2L −1} φ − 2πn 2L
. In the literature,
taneously and two corresponding points remain very close,
using the quantized phased control to lower the system com- (k+1)
which suggests that the iteration ejφm,n = ej(φm,n +δm,n ) may
(k) (k)

plexity has been considered in [13][9]. The spectral efficiency

performance with the quantized RF phase control has been 4 All parameters given in this section can be revised for other specific cases.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
9

(k)
be regarded as a linear operation over δm,n . By performing all simulations. In addition, the water-filling power allocation
(k)
adaptive δ̄ (k) , ejφm,n updates with relatively larger step size in Section III-C is adopted in the simulations.
at the beginning when the iterate is far from the solution 1) Large i.i.d Rayleigh Fading Channels: The MD-HP
e−j1.0026 ≈ 0.5381 − j0.8429, and then gradually gets close scheme is compared in Fig. 4 against the optimal uncon-
to it. In Fig. 3, we show how the error measure εk converges strained SVD based processing scheme when Ns = 8 data
to about 0.2 as the number of iterations increases when the streams are transmitted in a 256 × 64 massive MIMO system.
adaptive and constant δ̄ (k) are applied respectively. It can be For the MD-HP scheme, the situations of using 8 and 12
observed that the adaptive threshold δ̄ (k) helps the algorithm RF chains (along with their quantized versions) are examined.
converge more quickly because it allows the algorithm to When 12 RF chains are implemented at both the transmitter
conduct a search over a larger part of the feasible region when and receiver, the performance of the MD-HP scheme is near-
the error εk is relative small. The above parameters will also optimal compared with the optimal unconstrained SVD based
be used in the next simulations. scheme. Even though we reduce the number of the RF chains
to the number of the data streams, namely, 8 RF chains
0.2
are employed, the spectral efficiency achieved by the MD-
Trace of ej(φk +δk )

Trace of (1 + jδk )ejφk

HP scheme slightly decreases by around 3 bps/Hz. As for
0 the heavily quantized versions (L = 2 bits with the phase
candidates {0, 21 π, π, 23 π}) corresponding to the 8 and 12 RF
-0.2 chains settings, the spectral efficiency suffers less than 2.5
dB loss (from the view of SNR). Fig. 5 further demonstrates
imaginary part

-0.4
the spectral efficiency performance by also setting the number
-0.6 0.5381-j0.8429
of transmit data streams to 4 while 8 RF chains are used.
Compared with the case of 4 transmit data streams, the
-0.8 performance of the 8 data stream case is evidently improved
thanks to the multiplexing gain. Notably, there is a small
-1
gap between the MD-HP scheme and the SVD based scheme
which can be eliminated by properly increasing the number of
-1.2
-0.2 0 0.2 0.4
real part
0.6 0.8 1 1.2 RF chains, e.g., double the number of the data streams in the
case of Ns = 4. In addition, the quantized versions (L = 2)
(k) (k) (k) (k)
Fig. 2. The traces of ej(φm,n +δm,n ) and (1+jδm,n )ejφm,n on the complex also results in 2.5 dB loss in performance. Under a critical
plane. condition that the numbers of RF chains of the transmitter and
receiver are set to Mt = Mr = Ns , Fig. 6 shows the spectral
efficiency of the above schemes with Ns = 2, 4 and 8. It is
0.5
Adaptive threshold δ̄ (k) ∈ [0.1, 0.5] observed that the MD-HP scheme (including the quantized
Constant threshold δ̄ (k) = 0.1
0.45
version) consistently remains close to the optimal spectral
efficiency as Ns increases, which implies that the MD-HP
0.4 scheme can probably achieve the near-optimal performance
even when a large number of data streams are conducted.
Error measure εk

0.35
50
Optimal Unconstrained SVD
0.3 45 MD-HP scheme, Mt = Mr = 12
Quantized MD-HP, Mt = Mr = 12, L=2
0.25
40 MD-HP scheme, Mt = Mr = 8
Quantized MD-HP, Mt = Mr = 8, L=2
Spectral Efficiency (bps/Hz)

35
0.2
30

25
0 20 40 60 80 100
Number of iterations
20

Fig. 3. The convergence property of Algorithm 1 when applying the adaptive 15

and constant δ̄ (k) respectively
10

B. Spectral Efficiency Evaluation 0

−40 −35 −30 −25 −20 −15 −10 −5 0
In this part of simulation section, we illustrate the spectral SNR (dB)

efficiency performance of the proposed MD-HP scheme by Fig. 4. Spectral efficiency achieved by different processing schemes of a
comparing it with several other options under large i.i.d. 256 × 64 massive MIMO system in i.i.d. Rayleigh fading channels where
Rayleigh channel and mmWave channel settings respectively. Ns = 8 data streams are transmitted through 8 and 12 RF chains respectively.
The SNR γ = σP2 range was set to be from -40 dB to 0 dB in

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
10

50 In Fig. 7, the spectral efficiency performance is demon-

Optimal Unconstrained SVD
45 MD-HP scheme strated in a 256 × 64 mmWave MIMO system, where Ns = 8
Quantized MD-HP, L=2 data streams are transmitted through 8 or 12 RF chains.
40
Our proposed MD-HP scheme apparently outperforms the
Ns = 8
Spectral Efficiency (bps/Hz)

35 spatially sparse processing scheme when the same number of

30 RF chains are implemented. Moreover, the MD-HP scheme
can even achieve higher spectral efficiency with only 8 RF
25
chains than the spatially sparse processing scheme with 12 RF
20 chains. Particularly, the SVD based processing is sufficiently
15
approached by the MD-HP scheme given 12 RF chains. It is
shown that our proposed MD-HP scheme can better capture
10 the characteristics of the mmWave channel than the spatially
Ns = 4

5 sparse processing scheme. In addition, we examine the impact

of phase shifter resulotion on these two hybrid precoding
0
−40 −35 −30 −25 −20 −15 −10 −5 0 schemes in Fig. 8, while the number of data streams is set
SNR (dB)
to Ns = 8 and 12 RF chains are deployed on the transmitter
Fig. 5. Spectral efficiency achieved by different processing schemes of a and the receiver. The performance degradation of the MD-HP
256 × 64 massive MIMO system in i.i.d. Rayleigh fading channels where scheme is basically close to that of the spatially sparse coding
Ns = 4 and 8 data streams are transmitted through 8 RF chains respectively. in mmWave channels with 2-bit phase quantization, and even
the quantized version of the MD-HP scheme (red dash line)
50
Optimal Unconstrained SVD
can approach the spectral efficiency achieved by the spatially
45 MD-HP scheme sparse coding scheme without phase quantization. This result
Quantized MD-HP, L=2
40
again supports the performance of the MD-HP scheme since
it does not have higher RF component requirements in terms
Spectral Efficiency (bps/Hz)

35 Ns = 8 of the phase shifter resulotion.

30
60
25 Optimal Unconstrained SVD
Ns = 4 MD-HP scheme, Mt = Mr = 12
20
50 Sparse precoding&combining, Mt = Mr = 12
MD-HP scheme, Mt = Mr = 8
15 Ns = 2 Sparse precoding&combining, Mt = Mr = 8
Spectral Efficiency (bps/Hz)

10 40

5
30
0
−40 −35 −30 −25 −20 −15 −10 −5 0
SNR (dB)
20
Fig. 6. Spectral efficiency achieved by different processing schemes of a
256 × 64 massive MIMO system in i.i.d. Rayleigh fading channels where
Ns = 2, 4 and 8 data streams are transmitted respectively and the numbers 10
of RF chains are set to Mt = Mr = Ns .

0
−40 −35 −30 −25 −20 −15 −10 −5 0
2) Large mmWave Channels: Our proposed MD-HP SNR (dB)

scheme can also be applied to the large mmWave channels

Fig. 7. Spectral efficiency achieved by different processing schemes of a
where a certain number of hybrid processing schemes have 256 × 64 massive MIMO system in mmWave channels where Ns = 8 data
been studied in the literature. In simulations, the clustered streams are transmitted through 8 and 12 RF chains respectively.
mmWave channel model (4) was adopted to characterize
its limited scattering feature. Apart from the unconstrained 3) Effect of Channel Errors: Although we assume the
SVD based processing and our MD-HP schemes, we employ perfect channel state information is known during design,
the spatially sparse processing [9] which designs the hybrid some channel errors would certainly occur in the procedures
precoders/combiners by capturing the characteristics of the of channel estimation, channel feedback, RF calibration, an-
dominant paths. The propagation model mainly follows the tenna calibration, etc. It is therefore meaningful to discuss
settings in [9]: 1) the mmWave channel incorporates Nc = 8 how the channel errors affect the performance of our MD-
clusters, each of which has Np = 10 paths; 2) the transmitter HP scheme via comparing the effects of channel errors on
angle sector is assumed to be 60◦ -wide in the azimuth while both full-complexity and hybrid processing schemes. Since
the receiver with a smaller omni-directional antenna array; 3) we do not enforce any specific channel estimation, channel
the angle spreads of the transmitter and receiver σθ and σφ feedback, and component calibration models, the input of the
are all set to be 7.5◦ ; 4) the antenna spacing d is equal to MD-HP scheme is just a general channel matrix H and the
half-wavelength. channel errors are directly added to it. In the simulation, the

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
11

60 30
Optimal Unconstrained SVD
MD-HP scheme
Quantized MD-HP, L=2
50 25
Optimal Unconstrained SVD (5% Channel Error)
MD-HP scheme (5% Channel Error)
Spectral Efficiency (bps/Hz)

Spectral Efficiency (bps/Hz)

Quantized MD-HP, L=2 (5% Channel Error)
40 20

30 15

20 10

10 5

0 0
-40 -35 -30 -25 -20 -15 -10 -5 0 -40 -35 -30 -25 -20 -15 -10 -5 0
SNR (dB) SNR (dB)

Fig. 8. Spectral efficiency achieved by different processing schemes (with Fig. 9. Spectral efficiency achieved by different processing schemes of a
phase quantization) of a 256×64 massive MIMO system in mmWave channels 256×64 massive MIMO system with/without 5% channel errors where Ns =
with Ns = 8 and Mt = Mr = 12. 4 data streams are transmitted respectively and the numbers of RF chains are
set to Mt = Mr = Ns .

range of channel errors is set to ±5%. More specifically,

denoting the (m, n)-th element in H as hm,n , the channel as i.i.d. Rayleigh fading channels and mmWave channels.
error m,n of hm,n is set to m,n = crm,n · real(hm,n ) + cim,n · By providing enough number of RF chains (e.g., double
imag(hm,n ), where crm,n and cim,n are uniformly distributed the number of the transmit data streams), the pre-designed
among [−5%, 5%] while real(hm,n ) and imag(hm,n ) indicate unconstrained digital precoder/combiner of a large dimension
the real and imaginary parts of hm,n . In Fig. 9, we revisit can be sufficiently approached and thus the near-optimal
the spectral efficiency of 1) optimal unconstrained SVD, performance is achieved. The proposed matrix decomposition
2) MD-HP and 3) MD-HP (2-bit quantized) schemes in a method is a general algorithm. Any existing full complexity
256 × 64 massive MIMO system with Ns = 4 data streams precoder/combiner design can be decomposed into a hybrid
and Mt = Mr = Ns RF chains under i.i.d. Rayleigh fading RF phase shifter and digital baseband structure using this
channels. The solid lines show the performance of these approach.
schemes with accurate channel state information while the
dash lines present the corresponding performance with 5% R EFERENCES
channel errors. Even with the optimal unconstrained (full- [1] J. G. Andrews, S. Buzzi, C. Wan, S. V. Hanly, A. Lozano, A. C.
complexity) SVD solution, the spectral efficiency performance K. Soong, and J. C. Zhang, “What will 5G be,” IEEE J. Sel. Areas
Commun., vol. 32, pp. 1065–1082, June 2014.
would be degraded to some extent by the channel errors. As for [2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
the MD-HP schemes with fewest RF chains Mt = Mr = Ns , bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
the degree of performance loss is somehow approximate to pp. 3590–3600, Nov. 2010.
[3] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
the optimal unconstrained SVD solution, which means that O. Edfors, and F. Tufvesson, “Scaling up MIMO: opportunities and
the effects of channel errors on both full-complexity and challenges with very large arrays,” IEEE Sig. Process. Mag., vol. 30,
hybrid processing structures are virtually similar. In this sense, pp. 40–60, Jan. 2013.
[4] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
the techniques that are used to mitigate channel errors in MIMO for next generation wireless systems,” IEEE Commun. Mag., pp.
traditional MIMO systems can possibly be migrated to the 186–195, vol. 52, no. 2, Feb. 2014.
hybrid structure. [5] U. Erez, S. Shamai, and R. Zamir, “Capacity and lattice strategies for
canceling known interference,” IEEE Trans. Info. Theory, vol. 51, pp.
3820–3833, Nov. 2005.
V. C ONCLUSION [6] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral
efficiency of very large mmultiuser MIMO systems,” IEEE Trans.
In this paper, we have designed the hybrid RF and baseband Commun., vol. 61, pp. 1436–1449, Apr. 2013.
precoders/combiners for multi-stream transmission in P2P [7] P. J. Smith, C. Neil, M. Shafi, and P. A. Dmochowski, “On the
massive MIMO systems via solving a non-convex matrix convergence of massive MIMO systems,” in Proc. IEEE Intl. Conf.
Commun. (ICC), pp. 5191–5196, June 2014.
decomposition problem. Based on an alternate optimization [8] X. Zhang, A. F. Molisch, and S. Y. Kung, “Variable-phase-shift-based
technique, we have transformed the non-convex matrix de- RF-baseband codesign for MIMO antenna selection,” IEEE Trans. Sig.
composition problem into a series of convex sub-problems. Process., vol. 53, pp. 4091–4103, Nov. 2005.
[9] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
Careful handling of the phase increment of each entry in RF “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
precoders and combiners in each iteration and smart choice Trans. Wireless Commun., vol. 13, pp. 1499–1513, Mar. 2014.
of an initial point have allowed our algorithm to yield near- [10] O. E. Ayach, R. W. Heath, S. Abu-Surra, S. Rajagopal, and Z. Pi, “The
capacity optimality of beam steering in large millimeter wave MIMO
optimal solution with high probability. The MD-HP scheme systems,” in Proc. IEEE 13th Intl. Workshop on Sig. Process. Advances
can be applied to any general massive MIMO channels such in Wireless Commun. (SPAWC), pp. 100–104, June 2012.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2017.2699643, IEEE
Transactions on Signal Processing
12

[11] C. Kim, T. Kim, and J.-Y. Seol, “Multi-beam transmission diversity [36] A. Nemirovski, “Interior point polynomial time methods in convex pro-
with hybrid beamforming for MIMO-OFDM systems,” IEEE Globecom gramming,” 2012 [Online], Available: http://www2.isye.gatech.edu/ ne-
Workshops, pp. 61-65, Dec. 2013. mirovs/Lect IPM.pdf.
[12] A. Sayeed and J. Brady, “Beamspace MIMO for high-dimensional
multiuser communication at millimeter-wave frequencies,” in Proc. IEEE
Global Commun. Conf. (GLOBECOM), pp. 3679–3684, Dec. 2013.
[13] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding in
massive multiuser MIMO systems,” IEEE Wireless Commun. Lett., vol. Weiheng Ni (S’14) received the B.E. degree in
3, pp. 653–656, Dec. 2014. Communication Engineering from Beijing Univer-
[14] D. J. Love and R. W. Heath, “Equal gain transmission in multiple-input sity of Posts and Telecommunications, China, in
multiple-output wireless systems,” IEEE Trans. Commun., vol. 51, pp. 2012, and the M.A.Sc degree in Electrical and
1102–1110, July 2003. Computer Engineering from University of Victoria,
[15] X. Zheng, Y. Xie, J. Li, and P. Stoica, “MIMO transmit beamforming Canada, in 2015. He is currently pursuing the Ph.D.
under uniform elemental power constraint,” IEEE Trans. Sig. Process., degree with the University of California, San Diego,
vol. 55, pp. 5395–5406, Nov. 2007. USA. From 2015 to 2016, he was with Fortinet
[16] W. Ni and X. Dong, “Hybrid block diagonalization for massive multiuser Technologies, Burnaby, BC, Canada and worked on
MIMO systems,” in IEEE Trans. Commun., vol. 64, pp. 201-211, Jan. the development of the FortiCloud system.
2016. His research interests include cognitive radio,
[17] A. Liu and V. Lau, “Phase only RF precoding for massive MIMO energy harvesting, massive multiple antenna systems, millimeter wave tech-
systems with limited RF chains,” IEEE Trans. Sig. Process., vol. 62, niques and some machine learning based designs.
pp. 4505–4515, Sept. 2014.
[18] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design
for large scale MIMO systems,” in Proc. IEEE Intl. Conf. Acoustics,
Speech and Signal Processing (ICASSP), pp. 2929–2933, Apr. 2015.
[19] R. Mendez-Rial, C. Rusu, N. Gonzalez-Prelcic, A. Alkhateeb, and R. Xiaodai Dong (S’97–M’00–SM’09) received her
W. Heath, “Hybrid MIMO architectures for millimeter wave communi- B.Sc. degree in Information and Control Engineering
cations: phase shifters or switches?”, IEEE Access, pp. 249–267, Mar. from Xi’an Jiaotong University, China in 1992, her
2016. M.Sc. degree in Electrical Engineering from Na-
tional University of Singapore in 1995 and her Ph.D.
[20] O. Orhan, E. Erkip and S. Rangan, “Low power analog-to-digital con-
degree in Electrical and Computer Engineering from
version in millimeter wave systems: impact of resolution and bandwidth
Queen’s University, Kingston, ON, Canada in 2000.
on performance,” in Proc. Info. Theory and Appl. Workshop (ITA), pp.
Since January 2005 she has been with the University
191-198, Oct. 2015.
of Victoria, Victoria, Canada, where she is now
[21] J. Liu, J. Xu, W. Xu, S. Jin and X. Dong, “Multiuser massive MIMO a Professor at the Department of Electrical and
relaying with mixed-ADC receiver,” IEEE Signal Processing Lett., vol. Computer Engineering. She was a Canada Research
24, pp. 76-80, Jan. 2017. Chair (Tier II) in 2005-2015. Between 2002 and 2004, she was an Assistant
[22] A. J. Duly, T. Kim, D. J. Love, and J. V. Krogmeier, “Closed-loop Professor at the Department of Electrical and Computer Engineering, Univer-
beam alignment for massive MIMO channel estimation,” IEEE Commun. sity of Alberta, Edmonton, AB, Canada. From 1999 to 2002, she was with
Letters, vol. 18, pp. 1439–1442, Aug. 2014. Nortel Networks, Ottawa, ON, Canada and worked on the base transceiver
[23] S. L. H. Nguyen and A. Ghrayeb, “Compressive sensing-based chan- design of the third-generation (3G) mobile communication systems.
nel estimation for massive multiuser MIMO systems,” in Proc. IEEE Dr. Dong’s research interests include wireless communications, radio prop-
Wireless Commun. and Network. Conf. (WCNC), pp. 2890-2895, Apr. agation, ultra-wideband radio, machine to machine communications, cyber
2013. physical systems, wireless security, ehealth, smart grid, nano-communications
[24] D. Ramasamy, S. Venkateswaran, and U. Madhow, “Compressive adap- and signal processing for communication applications. She served as an Editor
tation of large steerable arrays,” in Proc. Info. Theory and App. Workshop for IEEE Transactions on Wireless Communications in 2009-2014, IEEE
(ITA), pp. 234–239, Feb. 2012. Transactions on Communications in 2001-2007, Journal of Communications
[25] A. Alkhateeb, O. E. Ayach, G. Leus, and R. W. Heath, “Channel and Networks in 2006-2015, and is currently an Editor for IEEE Transactions
estimation and hybrid precoding for millimeter wave cellular systems,” on Vehicular Technology.
IEEE J. Sel. Topics Sig. Process., vol. 8, pp. 831–846, Oct. 2014.
[26] Z. Gao, L. Dai, Z. Wang, and S. Chen, “Spatially common sparsity based
adaptive channel estimation and feedback for FDD massive MIMO,”
IEEE Trans. Sig. Process., vol. 63, no. 23, pp. 6169–6183, Dec. 2015.
[27] G. Hadi, M. Bengtsson, T. Kim, and M. Skoglund, “Subspace estimation Wu-Sheng Lu (F’99) received the B.Sc. degree
and decomposition for hybrid analog-digital millimetre-wave mimo sys- in Mathematics from Fudan University, Shanghai,
tems,” in IEEE Int. Workshop on Signal Process. Advanced in Wireless China, in 1964, and the M.S. degree in Electrical
Commun. (SPAWC), June 2015. Engineering and the Ph.D. degree in Control Sci-
[28] G. H. Golub and C. F. Van Loan, Matrix Computations, 3rd ed., Johns ence from the University of Minnesota, Minneapolis,
Hopkins University Press, 1996. USA, in 1983 and 1984, respectively.
[29] Q. H. Spencer, B. D. Jeffs, M. A. Jensen, and A. L. Swindlehurst, He was a post-doctoral fellow at the University
“Modeling the statistical time and angle of arrival characteristics of an of Victoria, Victoria, B.C., Canada, in 1985 and a
indoor multipath channel,” IEEE J. Sel. Areas Commun., vol. 18, pp. visiting assistant professor with the University of
347–360, Mar. 2000. Minnesota in 1986. Since 1987, he has been with the
[30] R. Escalante and M. Raydan, “Alternating projection methods”, Society University of Victoria where he is a professor. His
for Industrial and Applied Mathematics, vol. 8, 2011. current teaching and research interests are in the general areas of digital signal
[31] H. Kim and H. Park, “Nonnegative matrix factorization based on alter- processing and application of optimization methods. He is the co-author with
nating nonnegativity constrained least squares and active set method”, A. Antoniou of Two-Dimensional Digital Filters (Marcel Dekker, 1992) and
Society for Industrial and Applied Mathematics, vol. 30, July 2008. Practical Optimization – Algorithms and Engineering Applications (Springer,
2007). Dr. Lu is a Fellow of Engineering Institute of Canada, and a Fellow
[32] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge Uni-
of the IEEE.
versity Press, 2004.
[33] Z. Gao, L. Dai, D. Mi, Z. Wang, M. Imran, and M. Shakir, “MmWave
massive MIMO based wireless backhaul for 5G ultra-dense network,”
IEEE Wireless Commun., vol. 22, no. 5, pp. 13-21, Oct. 2015.
[34] T. Kailath, A. H Sayed, and B. Hassibi, Linear Estimation, Prentice
Hall, vol. 1, 2000.
[35] P. Courrieu, “Fast computation of Moore-Penrose inverse matrices,”
Neural Info. Process. - Letters and Reviews, vol. 8, Aug. 2005.

1053-587X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.