Vous êtes sur la page 1sur 16

Purdue University

Purdue e-Pubs
Department of Electrical and Computer Department of Electrical and Computer
Engineering Technical Reports Engineering

2015

Multi-Resolution Codebook and Adaptive


Beamforming Sequence Design for Millimeter
Wave Beam Alignment
Song Noh
Purdue University, songnoh@purdue.edu

Michael D. Zoltowski
Purdue University, michael.d.zoltowski.1@purdue.edu

David J. Love
Purdue University, djlove@purdue.edu

Follow this and additional works at: http://docs.lib.purdue.edu/ecetr


Part of the Electrical and Computer Engineering Commons

Noh, Song; Zoltowski, Michael D.; and Love, David J., "Multi-Resolution Codebook and Adaptive Beamforming Sequence Design for
Millimeter Wave Beam Alignment" (2015). Department of Electrical and Computer Engineering Technical Reports. Paper 468.
http://docs.lib.purdue.edu/ecetr/468

This document has been made available through Purdue e-Pubs, a service of the Purdue University Libraries. Please contact epubs@purdue.edu for
additional information.
1

Multi-Resolution Codebook and Adaptive Beamforming


Sequence Design for Millimeter Wave Beam Alignment
Song Noh, Michael D. Zoltowski, and David J. Love

Abstract—Millimeter wave (mmWave) communication the works in [6]–[9] propose analog beam alignment
is expected to be widely deployed in fifth generation techniques in which phase shifters or discrete lens arrays
(5G) wireless networks due to the substantial bandwidth are employed. To provide increased multiplexing and
available at mmWave frequencies. To overcome the higher
path loss observed at mmWave bands, most prior work beamforming gains while using a limited number of
focused on the design of directional beamforming using RF chains, [10]–[14] present the design of hybrid beam
analog and/or hybrid beamforming techniques in large- alignment using a combination of analog and digital
scale multiple-input multiple-output (MIMO) systems. Ob- beamforming.
taining potential gains from highly directional beamform- Obtaining potential gains from highly directional
ing in practical systems hinges on sufficient levels of
channel estimation accuracy, where the problem of channel beamforming using a large number of antennas hinges
estimation becomes more challenging due to the substantial on sufficient levels of channel estimation accuracy. In
training overhead needed to sound all directions using a large-scale MIMO, much of the prior work exploits a
high-resolution narrow beam. In this work, we consider the priori knowledge of the underlying channel statistics
design of multi-resolution beamforming sequences to enable such as the channel covariance matrix and the spatial
the system to quickly search out the dominant channel
direction for single-path channels. The resulting design sparsity along with a specific antenna arrangement [15]–
generates a multilevel beamforming sequence that strikes [20]. Without the prior information, which requires ad-
a balance between minimizing the training overhead and ditional learning algorithms [16], [21]–[23], the problem
maximizing beamforming gain, where a subset of multilevel of channel estimation in the mmWave case becomes
beamforming vectors is chosen adaptively to provide an more challenging because narrow pencil beams capable
improved average data rate within a constrained time. We
propose an efficient method to design a hierarchical multi- of supporting high beamforming gain should be aligned
resolution codebook utilizing a Butler matrix, a generalized to dominant channel directions located randomly within
discrete Fourier transform (DFT) matrix implemented using a cell.
analog RF circuitry, which extends to three-dimensional To tackle the problem of channel estimation in a
beamforming. Numerical results show the effectiveness of mmWave system, there has been some work on channel
the proposed algorithm.
estimation using a hierarchical multi-resolution beam
Index Terms—Millimeter wave, MIMO, multi-resolution codebook design [8], [10], [11], [24]. When the angles
beam alignment, 3D beamforming, Butler matrix
of departure or arrival (AoDs/AoAs) of the base station
I. I NTRODUCTION and the user are divided into multiple partitions, the
beamforming vectors are sounded at increasing angular
The use of millimeter wave (mmWave) bands is
resolution to further divide either azimuth or elevation
considered a key technology for future wireless cellular
angular regions of interest. Ideally, the beamforming
communications due to the tens of GigaHertz of avail-
vectors should have beamwidths that grow narrow as the
able licensed and unlicensed spectrum bands in the range
search level increases.
of 20–100 GHz. The propagation properties of each of
Since the entire area spanned by a beam pattern
the mmWave bands are characterized by severe path loss
is bounded given a fixed transmit power, a wide
and sparse scattering conditions [2]. Compared to the
beamwidth beamformer (i.e., low angular resolution)
frequency bands used in most current wireless systems,
has relatively lower main-lobe directivity gain than a
directional beamforming design using large-scale MIMO
narrow beamwidth beamformer capable of supporting
systems is crucial to overcome the higher path loss at
high angular resolution. Therefore, the probability of
mmWave bands [3], [4].
successfully estimating the AoA/AoD ranges of the
Due to the radio frequency (RF) hardware limitations
previous level becomes lower and lower, and the initial
found in mmWave implementations, there is growing
estimation of angular partition can be the bottleneck to
interest in the design of channel estimation and pre-
minimize the beam misalignment probability.
coding algorithms [5]. To reduce hardware complexity,
Most prior work on multilevel beamforming [8], [10],
S. Noh, M. D. Zoltowski, and D. J. Love are with the [11], [24] is premised on similar estimation performance
School of Electrical and Computer Engineering, Purdue University, at each training stage, which is possible at high signal-
West Lafayette, IN 47907, USA (e-mail:songnoh@purdue.edu and
{mikedz,djlove}@ecn.purdue.edu). A preliminary version of this work to-noise ratio (SNR). When the total signal power used
will be presented in [1]. for the training period is large, power allocation can
2

improve the estimation performance during early train- either the pilot symbol during the training period (i.e.,
ing stages [11], which requires an increased dynamic 1 ≤ k ≤ Kp ) or the data symbol during the data
range for the power amplifiers. Since the performance transmission period (Kp < k ≤ K) with E{|sk |2 } = 1.
achievable using a multilevel beamforming framework In (1), the transmitter sounds Kp beamforming vectors
is closely related to both the training overhead and the and chooses a vector from them that archives the largest
beamforming gain, it is critical to understand how to received signal power, referred to as hard beam align-
design the multi-resolution beamforming sequences to ment [8]. The selected beamformer is then used for data
improve the data rate performance. transmission (i.e., channel uses satisfying Kp < k ≤ K).
In this work, we consider the design of a multilevel Transmission over mmWave frequencies using narrow
beamforming sequence that properly allocates a multi- transmit and receive beams can limit the multipath com-
resolution codebook to the training phase using a fixed ponents of the effective mmWave channel [25]. Thus, we
signal power. Given a set of multi-resolution codebooks, adopt a geometric-based channel model with L scatterers
a subset of the available codebooks is adaptively selected [26] where the channel matrix H for a ULA is given by
to provide an improved average data rate within a L
X
constrained time. We propose an efficient method to
p
H= Nt Nr gℓ ur (θℓr )uH t
t (θℓ ), (2)
design a hierarchical multi-resolution codebook utilizing ℓ=1
a Butler matrix implemented using analog RF circuitry.
where ut (·) and ur (·) denote the transmitter steering
We also show that this codebook design framework can
vector and the receiver response vector in the normalized
be extended to three-dimensional (3D) beamforming.
steering phases, i.e.,
This manuscript is organized as follows: The system
model and background are described in Section II. Sec- 1 h −j2π d θℓs d s T
i
us (θℓs ) = √ 1, e λ , · · · , e−j(Ns −1)2π λ θℓ ,
tion III describes the proposed multilevel beamforming Ns
sequence design method. A DFT-based codebook design (3)
is discussed in Section IV. Numerical results are pro- √
where j = −1, s ∈ {t, r}, λ is the carrier wave-
vided in Section V, followed by conclusions in Section
length, and d is the antenna spacing given by half-
VI.
wavelength spacing d = λ/2. For the possible physical
Notation Vectors and matrices are written in boldface angles φsℓ ∈ (−π/2, π/2) that cover the entire (one-
with matrices in capitals. All vectors are column vectors. sided) spatial horizon, we define the spatial frequency
For a matrix A, AT and AH indicate the transpose and θℓs ∈ (−1, 1) as θℓs = sin(φsℓ ) [27] where there is
Hermitian transpose of A, respectively. [A]i,: denotes a one-to-one mapping between θℓs and φsℓ . Denote the
the i-row of A. diag(d1 , . . . , dn ) is the diagonal matrix complex small-scale fading gain of the ℓ-th subpath by
composed of elements d1 , . . . , dn . For a vector a, a(i) gℓ modeled as independent and identically distributed
denotes the i-th element and kakp represents the p- (i.i.d.) complex Gaussian, i.e., gℓ ∼ CN (0, σg2 ) [25].
norm. For a real-valued scalar a, ⌈a⌉ and ⌊a⌋ denote We first consider the case of quantized AoAs/AoDs
the smallest integer greater than or equal to a and the assuming θℓs ∈ {−1 + (2n − 1)/Ns : 1 ≤ n ≤ Ns }
largest integer less than or equal to a, respectively. E{·} for s ∈ {t, r} as in [10], [11] and then extend to the
represents statistical expectation. N, R, and C denote the case of continuous angles in Sections IV-A and V. In
sets of natural, real, and complex numbers, respectively. this scenario, the MIMO channel matrix H in (2) can be
II. S YSTEM M ODEL rewritten as
A. System Setup H = Ur H̃UH t (4)
We consider a mmWave MIMO link composed of where Us = [us (θ1s ) · · · us (θN
s
s
)] ∈ CNs ×Ns for s ∈
uniform linear arrays (ULAs) of Nt transmit antennas {t, r}. Here,√H̃ has only L non-zero elements, i.e.,
and Nr receive antennas. We assume block transmission [H̃]m,n = Nt Nr gℓ where the indices (m, n) are
r
with K successive symbols for one block where the associated with the angles of θm and θnt of the ℓ-
block length K is smaller than the channel coherence th path. Thus, we may require at most Nr Nt channel
time, and Kp out of K channel uses are set aside for measurements to find the non-zero entries of H̃.
channel sounding. For a unit norm transmit beamforming To partition the set of AoAs into subintervals as in
vector fk ∈ CNt , the received signal yk ∈ CNr at a [24], [28], we assume that the receiver employs the filter
discrete symbol time k is given by bank UH r to the received signal in (1) given by
√ √
yk = P Hfk sk + zk , for k = 1, 2, . . . , K (1) ỹk = UH r yk = P H̃UH t fk sk + z̃k , (5)
where H ∈ CNr ×Nt denotes the MIMO channel matrix, where z̃k = UH r zk . For a ULA, the receive filter bank
an Nr × 1 Gaussian noise vector zk ∼ CN (0, σz2 INr ), can be implemented using a DFT-based filter [15]. We
P denotes the transmit power, and sk ∈ C represents convert a joint detection of the non-zero elements of H̃ in
3

Actual beam pattern AoD2


Approximate beam pattern C2
ψm Cm AoD1
AoD1,1 AoD1,2 AoD1,3 (2)
(2) f (2)
f1 2 f3
C1
Fig. 1: Sectored beam model.
(1)
(1) f2 (1)
the whole range of AoAs to separate detection problems f1 f3 ψ2
in small angle intervals.1 ψ1

B. Review of Prior Work on Multi-Resolution Beam Stage 1 Stage 2


Alignment
Fig. 2: Approximate beamforming patterns in a mul-
Much of the prior work considers a multi-resolution tilevel beamforming structure where V1 = V2 = 3,
beam alignment framework [8], [10], [11] where the hi- M = 2, and AoD1,2 = AoD2 .
erarchical multi-resolution codebook consists of M lev-
n For 1 ≤ m ≤ Mo,
els to achieve better angular resolution. distinct beamforming vectors in Fm required for search-
(m) (m)
the m-th level codebook Fm = f1 , . . . , fcard(Fm ) ing AoDm by Vm ∈ N given by n Vm = ⌈|AoDmo|/ψm ⌉.
(m) (m)
is designed to uniformly cover all the spatial fre- By abuse of notation, we write f1 , . . . , fVm ∈ Fm
quency range (−1, 1) (i.e., one-sided physical angles to denote the selected Vm codewords to cover the col-
(−π/2, π/2)) and to satisfy the relation card(F1 ) < lection of disjoint and mutually exclusive AoDm , i.e.,
· · · < card(FM ), where card(F ) denotes the cardinality AoDm = ∪Vv=1 m
AoDm,v .
of F . (An example of a multi-resolution codebook with At the m-th training stage, the system model of (5)
[card(F1 ), card(F3 ), card(F3 )] = [8, 16, 64] is shown in can be rewritten as
Fig. 5.) The beamwidth of each of the codewords in Fm h i √ h i
(m) (m) (m) (m)
is given by ỹ1 , . . . , ỹVm = P H̃UH t f1 , . . . , fVm
h i
(m) (m)
ψm = 2/card(Fm ), for 1 ≤ m ≤ M, (6) + z̃1 , . . . , z̃Vm , (8)

and satisfies ψ1 > · · · > ψM . Accordingly, the angular where the pilot symbols are assumed normalized to unit
domains are partitioned into card(Fm ) angular intervals power and omitted. From (8), the receiver selects a beam
for the m-th level codebook where we denote each index v ′ that achieves the largest received signal power
(m)
angular interval covered by fv by AoDm,v ⊂ (−1, 1) given by 2
card(Fm )
such that (−1, 1) = ∪v=1 AoDm,v and AoDm,v ∩ v ′ = argmax ỹv(m) (n) , (9)

AoDm,w = ∅ for v 6= w. Thus, the length of the interval 1≤v≤Vm

AoDm,v is given by |AoDm,v | = ψm for 1 ≤ v ≤ (m)


where ỹv (n) denotes the n-th element of the vector
card(Fm ) where |AoDi | denotes the length of an angular (m)
ỹv and the receiver selects theh index n that yields the
interval AoDi . The training period is partitioned into (m) (m)
i
M stages where the m-th training stage is allocated for largest average magnitude of ỹ1 (n), . . . , ỹVm (n) .
sounding a subset of beamforming vectors in Fm .2 For the sake of simplicity, we assume that the receiver
The codebook Fm is designed to provide an approx- knows the index n and will omit the dependence on (n).
(m)
imately constant main-lobe beamforming gain over the The beam index v ′ in (9) which corresponds to fv′ and
angle interval, which is mathematically described as AoDm,v′ is fed back to enable the transmitter to choose
 √ the next-level codewords in Fm+1 . (A hierarchical beam
Cm if θ ∈ AoDm,v

ut (θ)H fv(m) = , (7) alignment system requires feedback of the codebook

0 if θ ∈
/ AoDm,v index across the training stages.)
where Cm is a normalization constant that satisfies At training stage (m+1), the angle interval of interest
(m)
kfv k22 = 1. Therefore, the beam patterns of Fm is set to AoDm+1 = AoDm,v′ and the Vm+1 codewords
are characterized by the main-lobe directivity gain Cm required for probing AoDm+1 is selected from Fm+1 .
and the main-lobe beamwidth ψm with negligible side- Note from (A.1) that the value of Vm+1 is given by
lobe gain, in a manner similar to [29], [30], which is 
|AoDm+1 |
 
ψm

illustrated in Fig. 1. Vm+1 = = , (10)
ψm+1 ψm+1
Denote the angular interval to be searched at the m-th
training stage by AoDm ⊂ (−1, 1) and the number of where ψm = |AoDm,v′ | for 1 ≤ m < M given the
PM
constrained training period length Kp = m=1 Vm
1 The receiver can use a combiner for the received signals as the
within a block.
transmitter uses the beamformer in the same way.
2 We use the term “level” to refer the codebook and the term “stage” An example of a multilevel beamforming framework
to refer the training phase. with V1 = V2 = 3 and M = 2 is shown in Fig. 2.
4

(1) Feedback information


Given AoD1 , f2 is chosen at stage 1 and
 (2) then higher-
(2)
resolution beamformers f1 , . . . , f3 on AoD2 =
AoD1,2 are sounded at stage 2. K
KWm

III. P ROPOSED M ULTILEVEL B EAMFORMING Training period Data transmission


F RAMEWORK
In this section, We first explain the design of a (m)
(m) (m)
single-level beam alignment sequence that use a single f1 f2 fVm
codebook chosen from a set of multilevel codebooks Vm Wm codebook size
during the training period. We then extend the single-
Fig.
 (m)3: Training transmission where the codewords
level beam alignment framework to design a multilevel (m)
f1 , . . . , fVm ∈ Fm are retransmitted over Wm
beam alignment sequence where a subset of the set
channel uses.
of multilevel codebooks used for channel sounding is
selected adaptively aimed at improving the data rate
performance. training period requires substantial training overhead due
to its narrow beamwidth which can degrade the data
A. Adaptive Single Level Beam Alignment and Sequence rate performance. This tradeoff indicates that the index
Design m specifying the codebook level Fm used during the
To estimate an AoD partition during the training training period is a suitable design parameter.
period, the receiver chooses a codeword achieving the Given the codebook Fm , we introduce a new variable
largest received signal power as in (9). However, if the Wm ∈ N to specify the number of repetitions corre-
desired signal power is less than the uncertainty in the sponding to each of the codewords in Fm . Then, the
(m) (m)
noise power at moderate-to-low SNR, the receiver cannot codewords f1 , . . . , fVm ∈ Fm selected to cover the
robustly detect the desired signal. In contrast to the sector of interest denoted by AoD1 are sounded, where
(m)
previous methods explained in Section II-B, we do not the received signal ỹv ∈ C in (9) corresponding to
(m) (m) (m) T
use the expression in (9). Instead we employ a threshold fv is changed to ỹ1,v . . . , ỹWm ,v ∈ CWm given
technique that is adapted to the SNR. by
In this subsection, we consider a single-level beam h i √  
(m) (m)
alignment in which a single codebook selected from ỹ1,v , . . . , ỹWm ,v = P H̃n,: UH t 1 T
Wm ⊗ fv
(m)

{Fm : 1 ≤ m ≤ M } is exclusively used during the


h i
(m) (m)
+ z̃1,v , . . . , z̃Wm ,v , (12)
training period. For the mth -level codebook Fm , we
replace the alignment decision in (9) with a hypothesis where 1N ∈ RN denotes the all-ones vector, ⊗ denotes
test given by the Kronecker product, and recall that we omit the index
(n) for a concise notation. Here, the variable Vm is
(
(m)
H0m : ỹv ∼ CN (0, σz2 )
(m) , (11) redefined as
H1 : ỹv ∼ CN (Gm P |g|2 , σz2 )
m  
|AoD1 |
Vm = , (13)
where Gm := Nt Nr Cm and the path gain |g|2 may ψm
not be known a priori but follows a central chi-square
as to represent the size of the subset of Fm used to
distribution with 2ℓ degrees of freedom, i.e., |g|2 ∼
cover the angular interval AoD1 . The training overhead
(σg2 /2)χ22ℓ . Here, ℓ ≤ L represents the number of paths
becomes
within the corresponding beamwidth.3 For tractability of
Km = Vm Wm . (14)
the analysis, we consider the case of single dominant
path channels (i.e., L = 1 in (2)), then we make some In Fig. 3, we illustrate the process of training and data
comments regarding the multi-path case. transmission in a single level beam alignment frame-
(m)
Under a fixed signal power, the beamformer fv work. We will formulate an optimization problem over
(m+1)
has a lower main-lobe directivity gain than fv due the variables m and Wm where the objective is to
to its broader beamwidth (ψm > ψm+1 ). Thus, the provide an improved average data rate.
probability of successfully estimating the angular inter- For the present model in (11) and (12), we adopt a
val becomes lower as the index m decreases. However, (m)
Neyman-Pearson detector for fv given by [31]
using high-resolution beamforming vectors during the 2 m
1 X Wm H1
(m)
3 For the multi-path case, the number of paths covered by a beam- ỹn,v ≷ γm , (15)

Wm

forming vector depends on the propagation geometry, which makes Hm
n=1 0
the closed-form expression of the data rate performance required for
a sequence design intractable. where the threshold is set to γm = −σz2 ln(PF A )/Wm
5

given
 a target PF A , i.e., PF A =
false alarm probability
PWm (m) 2
does not admit a closed-form expression. Instead, we
P W1m n=1

ỹn,v > γm H0m . If there are multiple evaluate an average magnitude of the effective channel
codewords that satisfy (15), the receiver selects the single gain (i.e., |xm |2 ) to obtain a reasonable criterion for
codeword index among them that achieves the largest approximating the expression of (18). To gain a clear
received signal power (i.e., the left-hand side of (15)). insight to the main factor, we now provide a simple
The combined signal on the left-hand side of (15) can expression for the desired signal power in the high
be rewritten as SNR regime, while an exact expression without any
Wm Wm approximation is given in Appendix A.
1 X (m)
p 1 X
ỹn,v = Gm P g + z̃ (m) =: xm + zm , Lemma 1: The desired signal power in the high SNR
Wm n=1 Wm n=1 n,v
regime (i.e., Gm P σg2 ≫ σz2 /Wm ) is approximated as
(16)
√ − γm 2  PF A
E |xm |2 ≈ e Gm P σg Gm P σg2 + γm +
PWm (m) 
where xm = Gm P g and zm = (1/Wm ) n=1 z̃n,v
2
denote the desired signal and noise, respectively, i.e., − fσz2 ,PF A ,Wm , (19)
xm ∼ CN (0, Gm P σg2 ) and zm ∼ CN (0, σz2 /Wm ), and
2
σ ln(1/P ) √
the variables are assumed to be mutually independent.

where fσz2 ,PF A ,Wm = z Wm F A PF A − PF A .
That is, the magnitude of xm represents the instanta- Proof: See Appendix A.
neous effective channel gain achievable after a decision
H1m , while we operate as if |xm | = 0 for a decision Note that the average signal power given H1m is
H0m because there is no transmission. The beam mis- characterized by the first term on the right-hand side
alignment analysis, similar to [8], [11], can be performed of (19), which is the expected value of |xm |2 integrated
using the expressions of false alarm and miss detection over (γm , ∞), because the second term is negligible due
probabilities of (15). However, this approach cannot be to a small target false alarm rate and we numerically find
applied to the design of the optimal training length that the magnitude of fσz2 ,PF A ,Wm is less than 0.53σz2 for
because reducing the probability of beam misalignment all PF A ∈ (0, 1) and Wm ∈ {1, . . . , 102 }.
requires the maximum training length. Instead, we con- From Lemma 1, we consider the case that an instanta-
sider the average data rate as a performance metric and neous data rate is given by (17) when |xm |2 > γm (i.e.,
focus on designing a beam alignment sequence and the Gm P |g|2 > γm ), otherwise we have R(xm ) = 0. Then,
training length within a constrained time. Rm in (18) is approximated as
2 2
When a single propagation path is covered by a K − Km ∞  e−|xm | /σm
Z
(m)
beamformer fv ∈ Fm , an instantaneous data rate for Rm ≈ log 1 + |xm |2 2 /2
d|xm |
K γm σm
H1m is obtained given by
(20)
K − Km  
· log 1 + |xm |2 ,
 1
R(xm ) = (17) K − Km  1−β m e 2
σm
 
βm 
K = 2
e σm log βm + Γ 0, 2 ,
where the pre-log factor in (17) represents the cost of K ln 2 σm
channel sounding within K channel uses. Note that since (21)
a signal is detected when the squared absolute value of
2
the (combined) received signal is above the threshold, where σm := Gm P σg2 , βm = 1 + γm , and Γ(·, ·) de-
consideration of joint variations for the desired signal notes the incomplete gamma function [32, Eq. 8.350.2].
and noise is needed to provide an exact average data Maximization of Rm in (21) gives the design of the
rate, i.e., codebook level m and the number of retransmissions
of the corresponding codewords Wm .
Rm = E {R(xm )|A0 } p(A0 ) + E {R(xm )|A1 } p(A1 )
Problem 1: Given the codebooks {Fm : 1 ≤ m ≤
K − Km ∞ ∞
Z Z
= log(1 + |xm |2 )· M }, the variables m and Wm must be chosen to maxi-
K −∞ −∞ mize
p (xm , zm |A1 ) dxm dzm × p(A1 ), (18)
max Rm (22)
where A0 and A1 denotes mutually exclusive events {m,Wm }

described in terms of the random variables xm and zm subject to 1 ≤ m ≤ M and Km ≤ K,


(i.e., A0 = {|xm +zm |2 < γm } and A1 = {|xm +zm |2 ≥
where Km and Rm are defined in (14) and (21), respec-
γm }) and p (xm , zm |Am 1 ) is a conditional joint proba-
tively.
bility distribution function (pdf) of (xm , zm ) given A1 .
Here, E {R(xm )|H0m } = 0 because we set |xm |2 = 0 for Considering the integer constraint on the optimization
(xm , zm ) ∈ A0 from a consideration of no transmission. variables, we tackle Problem 1 by exhaustive search. Due
To the best of the authors’ knowledge, the integral in (18) to space limitations, we omit the details of our algorithm
6

for Problem 1. Instead, in the next subsection, we present Training period Feedback information
an algorithm for a multilevel beam alignment sequence
design that can be used to solve Problem 1. First Last
stage stage Data transmission
Remark 1: By applying Alzer’s inequality to the
mL
incomplete gamma function [33], Rm in (21) can be KmF

lower bounded in a closed form given by First (m )


(mF ) (mF )
stage f1 f2 fVmF
K − Km  1−β m F
Rm ≥ Rm =: e αm log βm
K VmF WmF codebook size
 C βm

−e αm log 1 − e−e ( αm ) , (23)
1

Last (mL ) (mL ) (m )


f2 fVmL
where C is Euler’s contant, i.e., C = 0.5772.... stage f1 L

VmL codebook size


To reduce the computational complexity required for
numerical calculation of Γ(·, ·), we can use the lower Fig. 4: Training and data transmissions where the code-
bound of Rm in (23) for a beam alignment sequence words sounded at the first training stage are retransmitted
design. Simulation results in Section V show that the over WmF channel uses, otherwise each codeword is
proposed algorithm using the lower bound in (23) per- transmitted once.
forms very closely to those results obtained using the
expression in (21).
Given this, the training overhead becomes
Note that the proposed beam alignment design can be
mL
directly used for the multi-path case where the selected mL
X
single codeword is expected to cover at least a single Km F
= VmF WmF + Vm , (24)
m=mF +1
dominant path. Alternatively, the codeword indices that
achieve the received signal power above the threshold where the variable Vm corresponding to the codebook
value in (15) can be fed back to the transmitter to Fm is redefined as
enable multi-stream transmission using the combined
VmF = ⌈|AoD1 |/ψmF ⌉ and Vm = ⌈ψm−1 /ψm ⌉ , (25)
codewords, which requires additional feedback. Joint
processing of the multiple training snapshots for the for mF < m ≤ mL . Here, AoD1 denotes the sector an-
multi-path case is beyond the scope of the current work. gle to be searched at the first training stage. The overall
process of training and data transmission is illustrated
in Fig. 4. As a result, we will optimize the codeword
B. Adaptive Multilevel Beam Alignment and Sequence
levels (mF , mL ) and WmF used during the consecutive
Design
training stages to provide an improved average data rate,
In this subsection, we focus on the design of a which will be shown shortly.
multilevel beam alignment sequence where a subset of We observe that the performance achievable by the
codebooks (i.e., {Fm : mF ≤ m ≤ mL }) can be used multilevel beam alignment framework is characterized
during the training period. Here, we define the indices by the detection performance at the first stage because
specifying the codeword levels used at the first and last the codewords in FmF have relatively low directivity
training stages as mF and mL for 1 ≤ mF ≤ mL ≤ M , gain than those in {Fm : mF < m ≤ mL }. For the
respectively. purpose of analysis, we assume that if a beamforming
Since the codewords in FmF sounded at the first vector in FmF is properly selected, then a beamforming
training stage have relatively lower directivity gain due to vector in Fm for mF < m ≤ mL that covers the
its broader beamwidth than those in {Fm : mF < m ≤ dominant channel path can be perfectly selected due to
mL }, the received signal power at the first training stage its relatively higher beamforming gain. We will refer
can be easily swayed by noise impact. Thus, we assume to this assumption as (A.1). (As the comparison with
retransmission of the codewords in FmF sounded during numerical results without (A.1), this approximation does
the first training stage where WmF represents the number not significantly affect our results.)4
of retransmissions. The system model and the detection
process at the first training stage can be described From (A.1), an (approximate) instantaneous data rate
by replacing the index of the codebook in (12) and
(15) with (mF ), respectively. During the intermediate
training stages allocated for sounding the codewords in 4 This assumption is not very restrictive in our view because for a

{Fm : mF < m ≤ mL }, we assume that each codeword single level codebook (i.e., mF = mL ), (A.1) is not involved in a
beam alignment sequence design, and for a multilevel codebook (i.e.,
is transmitted once, i.e., we set Wm = 1 in (12) and (15) mF < mL ) we also optimize the codeword levels (mF , mL ) to
for mF < m ≤ mL . improve the performance.
7

for |xmF |2 > γmF can be written, similar to (17), as Algorithm 1 Multilevel Beam Sequence Design
K − Km mL 
GmL
 1: Set R = 0
2
R(xmF , mL ) ≈ F
· log 1 + |xmF | , 2: for mL = M : −1 : 1 do
K GmF
(26) 3: for mF = mL : −1 : 1 do
where the operand in the log function denotes the 4: [VmhlF , . . . ,mVm lL ] m l mi
ψ mF ψmL −1
beamforming gain achievable at the last training stage = AoD ψ mF 
1
, ψmF +1 , . . . ,  ψmL
under (A.1). We then evaluate the (approximate) average P mL
K− m=m
F +1
Vm
throughput in the multilevel beamforming framework in 5: Wmax = Vm
F
semi-closed form given by 6: if mF < mL then
mL 7: Wmax = min{Wmax , ⌊GmF +1 /GmF ⌋}
Rm ≈ E {R(xmF , mL )} (27)
F
mL 
8: end if
K − K mF m
e(1−βmF )/α · log βm
mL
=
L
mL 9: for WmF = 1 : 1 : Wmax do
K F
10: Compute Rm mL
(Step †)
 mL
  F
+ e1/α /ln 2 · Γ 0, βm mL
F
/αmL
(28) 11: mL
if RmF > R then
mL
12: R = Rm F
where αmL = GmL P σg2 , βm mL
F
= 1 + (GmL /GmF )γmF , 13: [m̂F , m̂L ] ← [mF , mL ]
and the expectation in (27) is performed over |xmF |2 ∈ 14: Wm̂F ← WmF
(γmF , ∞), similar to (20). 15: end if
Therefore, using the objective function in (28), we 16: end for
design a multilevel beam alignment sequence optimized 17: end for
over the variables (mF , mL , WmF ) that strikes a balance 18: end for
between minimizing the training overhead and maximiz- (In Step †, the notations follow those of (28) or (31).
ing beamforming gain. Here, the hat notation denotes the output of the algo-
Problem 2: Given the codebooks {Fm : 1 ≤ m ≤ M } rithm. )
and (A.1), the level indices (mF , mL ) and WmF must be
chosen to maximize
mL
within the main-lobe beamwidth ψm . To design accurate
max Rm F
(29) approximations to the ideal beam patterns, much of
{mF ,mL ,WmF }
mL the prior work considers analog beamforming [6]–[8]
subject to 1 ≤ mF ≤ mL ≤ M, Km ≤ K,
F and hybrid beamforming techniques [10]–[12] where the
and 1 ≤ WmF ≤ ⌊GmF +1 /GmF ⌋ , codebook is assumed to be designed off-line constructed
(30) using a limited number of RF chains.
where Km mL mL
and Rm are defined in (24) and (28), However, when a user or scatterers are around the
F F
respectively. boundary of two adjacent (pre-designed) beamforming
patterns, it might be detected in both angular partitions.
The upper bound on WmF in (30) guarantees that the Then, shifting the main-lobe beam direction to the target
effective SNR at level mF is smaller than that of the next angle or redesigning the codebook is needed to provide
level, i.e., GmF P |g|2 /(σz2 /WmF ) ≤ GmF +1 P |g|2 /σz2 an improved beamforming gain. Considering many prior
because of WmF GmF ≤ GmF +1 . The corresponding works on DFT-based codebooks in [7], [34]–[36], we
algorithm is summarized in Algorithm 1 with reasonable present an efficient method to design a multi-resolution
computational complexity of O(M 2 K) as only three to codebook using the DFT vectors.
five levels are typically used in the multilevel beamform-
ing framework.
A. DFT-based Multilevel Codebook Design
Remark 2: Similar to Remark 1, the expression of (28) Considering the quantized spatial angles {θn = −1 +
can be lower bounded by Rm mL
F
≥ Rm L
mF where (2n − 1)/Nt : 1 ≤ n ≤ Nt }, the m-th level codebook
mL m Fm can be designed by combining the array steering
K − Km 1−βm L

Rm L
= F
e α mL
F
log β vectors corresponding to the subset of quantized angles
mF given by
K
m
βm L
  n o
1
−eC αmF (m) (m) (m)
−e αmL
log 1 − e L . (31) Fm = f1 , f2 , . . . , fN/Nm (36)
Nm 2N
(
m
1 X 1 X
= √ ut (θp )ejωm p, √ ut (θp )ejωm p,
Nm p=1 Nm p=Nm +1
IV. P ROPOSED C ODEBOOK D ESIGN T ECHNIQUE
Nt
)
Throughout the manuscript, we assumed that the beam 1 X jωm p
..., √ ut (θp )e , (37)
pattern is approximated by constant directivity gain Cm Nm p=Nt −Nm +1
8

−1(−90o) −1(−90o) −1(−90o)


where we introduce a new variable ωm to enable the 9 dB 12 dB 18 dB
(m)
beam patterns of fv to have a little to no variation 6.8 9 14

4.5 6 9
in main-lobe directivity gain as best as possible. Here, 2.3 3 4.5
we assume that Nm is a divisor of Nt for simplicity. o
0(0 )
o
0(0 )
o
0(0 )

The array steering vector on the quantized angles {θn }


in (37) becomes a generalized DFT vector with offset
−(1 + Nt )/2, i.e.,
1(90o) 1(90o) 1(90o)
1 h −j N2π (n− Nt2+1 ) −j N2π 2(n− Nt2+1 ) (a: Level 1) (a: Level 2) (a: Level 3)
ut (θn ) = √ 1, e t ,e t ,
Nt o
−1(−90 ) −1(−90 )
o
−1(−90 )
o

Nt +1
iT
. . . , e−j Nt (Nt −1)(n− 2 ) ,
2π 12 dB 14 dB 18 dB
(38) 9.1 10 14

6.1 7 9

which is the n-th column of the Butler matrix [37] and 3 3.5 4.5

0(0o) 0(0o) 0(0o)


can be easily implemented using millimeter microstrip
technology [38]–[40]. We denote the number of com-
bined vectors at level m by Nm and the number of active
RF chains by NRF , respectively. From (37), we need at 1(90o) 1(90o) 1(90o)

least NRF = Nm RF chains for combining the vectors. (b: Level 1) (b: Level 2) (b: Level 3)

The codewords in (37) form ideal beam patterns Fig. 5: Phase-shifted DFT beam patterns where Nt = 64,
over the quantized angles such that the codebook with M = 3, [card(F1 ), card(F3 ), card(F3 )] = [8, 16, 64],
arbitrary ωm ∈ R achieves the (constant) maximum and [N1 , N2 , N3 ] = [8, 4, 1]: (a) ideal case on the
beamforming gain Nt /Nm on the quantized angles {θn }. quantized spatial domain with an arbitrary ωm ∈ R and
For Nt = 64, M = 3, and {N1 , N2 , N3 } = {8, 4, 1}, (b) non-ideal case on the continuous spatial domain with
Fig. 5(a) shows the corresponding beam patterns over the ωm = 0 for 1 ≤ m ≤ M .
quantized angles, which is useful for theoretical analysis
and can be used as a benchmark showing what can be vectors as used in [24]) is shown in Fig. 5(b). Especially
achieved under idealized assumptions. For realizing this at lower level codewords, there are a large number of null
scenario in practice, large-scale antenna arrays can be points within the main beamwidth and picking points
applied in the transmitter and/or the receiver required near the boundary of beam patterns due to the overlapped
for finer spatial sampling of continuous angles. DFT beam patterns on the continuous angles.
(m)
Note that the codebook Fm in (37) is parameterized For the i-th codeword fi ∈ Fm in (36), we evaluate
by the phase shift ωm , and we refer to this scheme a closed-form expression for the array response over
as a phase-shifted DFT method. Optimizing the single the continuous angle θ in (33) where ut (θ)H ut (θp )
parameter ωm is important to achieve constant main-lobe represents the discrete Fourier transform using the p-th
directivity gain as possible. An example of the codebook vector in (38) evaluated at the spatial location parameter
for ωm = 0 (i.e., a simple addition of the array steering θ. Given this, the array response magnitude can be de-

 
iNm
(m) 1 X
ut (θ)H fi = ut (θ)H  √ ut (θp ) (32)
Nm p=(i−1)Nm +1
  
1
iNm
N −1

N (θ+1)+1
 sin π p − Nt (θ+1)+1
2
p− t 2
X jωm p −jπ N
= √ e e    (33)
Nt Nm p=(i−1)Nm +1 sin Nπt p − Nt (θ+1)+1
2

2
(m) (m)
fθ,i,ωm = ut (θ)H fi (34)

  
1  iNm
X sin2 π(p − θ̃)
= 2  
Nt Nm  sin2 Nπt (p − θ̃)
p=(i−1)Nm +1
    
iNm
X iNm
X  
N −1
 sin π(p − θ̃) sin π(p′ − θ̃) 
+ 2 cos (p′ − p) π − ωm ×     , (35)
N sin Nπt (p − θ̃) sin Nπt (p′ − θ̃) 
p=(i−1)Nm +1 p′ =p+1

Nt (θ+1)+1
where θ̃ := 2
.
9

(m) −1(−90o) −1(−90o) −1(−90o) −1(−90o)


scribed by a function fθ,i,ωm defined in (34). As a result, 9.5 dB 13 dB 18 dB 9.5 dB

we focus on an exhaustive search for the proper phase 7.1 9.4 14 7.1

shift ωm that minimizes the variation of the main-lobe 4.7 6.3 9 4.8

(m) 2.4 3.1 4.5 2.4


directivity gain evaluated with fi over the quantized o
0(0 )
o
0(0 )
o
0(0 )
o
0(0 )

version of main beam range given by


 
2((i − 1)Nm + 1) − 1 2(iNm ) − 1
−1 + , −1 + .
Nt Nt 1(90o) 1(90o) 1(90o) 1(90o)
(39) (a: Level 1) (a: Level 2) (a: Level 3) (a′ : Level 1)
In this case, it is important to reduce the computational
complexity of the exhaustive search, and thus we derive Fig. 6: Phase-shifted DFT beam patterns on the con-
(m) tinuous spatial domain where: (a) Nt = 64, M = 3,
a property of the objective function fθ,i,ωm .
{N1 , N2 , N3 } = {8, 4, 1}, and [ω1 , ω2 ] = [1.93, 2.24]
Theorem 1: Given the level m and 1 ≤ i ≤ N/Nm , the with any ω3 ∈ R, and (a′ ) Nt = 32, N1 = 4, and
(m)
function fθ,i,ωm in (35) satisfies the following properties: ω1 = 2.2.
(P.1) for ωm , 2π-periodic and an even function at π(Nt −
1)/Nt , where ut (·) denotes the steering vector defined in (3)
(P.2) for θ, an even function at −1 + (2i − 1)Nm /Nt , and (p, q) element of Γ in (41) is given by
(P.3) independent of the index i where 1 ≤ i ≤ N/Nm .  
π(p−q)(Nm −1)
Proof: See Appendix B. Nm (2i−1) sin Nt
[Γ]p,q = e−jπ(p−q)(−1+ Nt ) .
Remark 3: Theorem 1 yields a much simpler imple- π(p − q)
mentation of the exhaustive search without loss in per-
Since the power of the filtered signal is approximately
formance because: the search space for finding the phase
equal to the PSD value at θ, the auto-PSD estimation
shift ωm is reduced to the interval [−π/Nt , π(1−1/Nt )] PNm
ut (θ)H vp 2
over the interval is given by N1m p=1

of length π by (P.1); the continuous angular interval
for θ can be set to one-half of the main beam interval where vp ∈ CNt is the p-th dominant eigenvector of Γ,
of (39) by (P.2), e.g., [−1 + (2i − 1)Nm /Nt , −1 + referred to as the p-th dominant Slepian sequence. How-
(2iNm − 1)/Nt ]; the optimized phase shift ωm for the i- ever, simulations of a codebook design using the Slepian
(m) sequences as the basis vectors and optimizing over a
th codeword fi is also an optimal design for all other
single phase-shift design parameter, ωm , similar to the
codewords in Fm by (P.3).
codebook design for generalized DFT (Butler) beams,
Therefore, the problem of the codebook design at level yielded distorted beamforming patterns. Thus, multi-
m in (37) is formulated as dimensional optimization is needed to linearly combine
Slepian sequences to achieve a maximal constant level
 
(m)
min var fθb ,i,ωm , (40)
ωm ∈[− Nπ ,π (1− N1 )] of energy concentration over the sector of interest as
t t

b
best as possible. However, with respect to practical im-
for any fixed index i ∈ {1, . . . , N/Nm }. Here, θ b ∈ R2 plementation, our proposed codebook using DFT beams
denotes the uniformly quantized points of the half of according to (37) can be implemented via the well-
the main beam interval [−1 + (2i − 1)Nm /Nt , −1 + known Butler Matrix Beamformer [37] in analog RF
(2iNm − 1)/Nt ] using b bits where we set b = 9. Since circuitry which, as discussed previously, can be im-
the objective function of (35) can operate on the vector plemented using millimeter microstrip technology [38]–
input θ b , the design parameter ωm is swept over the [40]. In contrast, the values of the Slepian sequences are
possible range to minimize the sample variance in (40). not constant magnitude and also their phase values are
(m) arbitrary, i.e. are not confined to a finite alphabet.
Remark 4: Since each codeword fi ∈ Fm should
provide the maximal constant level of concentration of The proposed design obtained from (40) yields good
energy over the spatial interval [θ(i−1)Nm +1 , θiNm ], one performance over the continuous angles as shown in
can relate the codebook design problem in (40) to the de- Fig. 6 for {N1 , N2 , N3 } = {8, 4, 1}, M = 3, and
sign of bandpass filters for estimating the power spectral Nt = 64. The proposed codebook has constant main-
density (PSD) over (−π, π). The finite impulse response lobe directivity gain per beam similar to the ideal beam
(FIR) bandpass filter that concentrates its energy to the patterns in Fig. 5(a). For the codebook design in Fig.
interval [πθ(i−1)Nm +1 , πθiNm ] can be designed by using 6(a), we use at most NRF = 8 RF chains at the first
Nm dominant eigenvectors of Γ ∈ CNt ×Nt , referred to level codebook because of N1 = 8. To accommodate
as Slepian sequences [41], given by the case that we cover the same angular interval with the
Z πθiNm small number of RF chains, we can reduce the number
1
Γ= ut (ϑ)ut (ϑ)H dϑ, (41) of active antenna elements by turning off a subset of suc-
2π πθ(i−1)Nm +1 cessive antenna elements while having a coarse spatial
10

resolution [7]. In comparison to the beam patterns in Fig.


6(a: Level 1), Fig. 6(a′ ) shows the corresponding beam
patterns obtained with NRF = N1 = 4 and Nt = 32.
Although the estimation accuracy of the angular interval
is restricted by the reduced number of antenna elements,
the proposed beam patterns still yield reasonably good
performance, which will be further demonstrated in
Section V.

B. 3D Beamforming
When the array manifold of a two-dimensional uni- Fig. 7: 3D beam patterns of F in (42) on the continuous
form planar array (UPA) is modeled as the Kronecker angles for a UPA where NH,t = 128, NV,t = 32, NH =
product of vectors from the array manifolds of two (hori- 32, NV = 8, ωH = 2.04, ωV = 1.89, and card(F ) = 16.
zontal/vertical) ULAs as in [42]–[44], a 3D beamforming Each colored column block denotes each beam pattern.
vector can be constructed from the Kronecker product
of two (horizontal/vertical) codewords. From the phase-
In (47), if fH,v is a codeword designed for covering
shifted DFT method in (37), two (horizontal/vertical)
the azimuth angle φa ∈ (0◦ , 30◦ ) (i.e., sin(φa ) ∈
codebooks, denoted as FH and FV , are designed by opti-
(0, 0.5)) for a ULA, the first term of the right hand
mizing the horizontal and vertical phase shifts (ωH , ωV )
side of (47) yields the maximum beamforming gain over
separately. Mathematically, this means
sin(φa ) cos(φe ) ∈ (0, 0.5) for a UPA due to a similar
F = FH ⊗ FV , (42) structure of the array steering vectors in (3) and (45).
As a result, the range of the azimuth angle for a UPA
where for s ∈ {H, V} that achieves the maximum beamforming gain becomes
larger as much as the elevation angle φe changes from 0◦

Fs = fs,1 , fs,2 , . . . , fs,Ns,t /Ns
to ±90◦ (i.e., decreasing the value of cos(φe )). Similarly,
 
 1 iNs
X Ns,t 
ut,s (θp,s )ejωs p 1 ≤ i ≤

= √ . the area spanned by the codeword that covers off-center
 Ns Ns 
p=(i−1)N +1 s
in azimuth gets narrower by changing the elevation angle
(43) towards ±90◦ . This is a manifestation of the well-known
Here, ut,s (θp,s ) denotes the array steering vector on the reduced-aperture effect off boresite.
quantized angles {θp,s = −1 + (2p − 1)/Ns,t : 1 ≤ For a 32 × 128 UPA, Fig. 7 shows the beam patterns
p ≤ Ns,t }, and (Ns,t , Ns ) denote the number of antenna of F obtained from the proposed method. Note that
elements and combined vectors in each dimension, re- the beamwidth of the codeword designed for covering
spectively. center azimuth grows larger while maintaining (nearly)
For a size NV,t ×NH,t UPA with the azimuth φa (deg) the same bamforming gain as elevation angle φe goes to
and elevation φe (deg) AoDs, the channel response is ±90◦ . Each codeword yields reasonably constant beam-
given by [42]–[44] forming gain over the azimuth and elevation angular
partitions.
u(φa , φe ) = uH (φa , φe ) ⊗ uV (φe ), (44) Therefore, the proposed method provides an efficient
way to design the multi-resolution codebook on a need-
where
h based basis.5 Due to the separable 3D beamforming
2πd
uH (φa , φe ) = 1, e−j λ sin(φa ) cos(φe ) , . . . , structure in (42), Algorithm 1 can be used to generate
2πd
iT the multilevel beam alignment sequence in each of
e−j(NH,t −1) λ sin(φa ) cos(φe )
(45) the azimuth and elevation directions by averaging the
h 2πd beamforming gains within a sector angle to characterize
uV (φe ) = 1, e−j λ sin(φe ) , . . . , the main-lobe directivity gain of the codebooks.
2πd
iT
e−j(NV,t −1) λ sin(φe )
. (46) V. N UMERICAL R ESULTS
From (42) and (44), the array response for the codeword In this section, we provide some numerical re-
in F is evaluated as sults to evaluate the proposed beam sequence de-
u(φa , φe )H (fH,v ⊗ fV,w ) 2
sign and the corresponding algorithm. Throughout the
2 2 simulation, we fixed the noise power σz2 = 1 and
= uH (φa , φe )H fH,v × uV (φe )H fV,w ,

(47) K = 40. The sector angle at the transmitter is
where fH,v ∈ FH and fV,w ∈ FV defined in (43). 5 We observe that ωs = 2 can be used as a rule of thumb.
11

7
Fixed (m =m =1)
F L
6 Fixed (m =m =2)
6 F L
Fixed (m =m =3)
F L
5 Fixed (m =1, m =2)
5 F L
Fixed (m =2, m =3)
F L

Rate (bps/Hz)
Rate (bps/Hz)

4 Fixed (m =1, m =3)


4 F L
Proposed
3
3
Analysis in (28)
Algorithm 1 (L=1,(28)) 2
2 Algorithm 1 (L=2,(28))
Algorithm 1 (L=3,(28))
1 Algorithm 1 (L=1,(31)) 1
Algorithm 1 (L=2,(31))
Algorithm 1 (L=3,(31))
0
−14 −12 −10 −8 −6 −4 −2 0 2 −16 −14 −12 −10 −8 −6 −4 −2 0 2
SNR (dB) SNR (dB)
mL
Fig. 8: Data rate versus SNR for different L where Rm F
Fig. 9: Performance comparison with the methods that
in (28) and Rm L
mF in (31) are used in Algorithm 1 for the use the fixed levels (i.e., WmF = 1 and 1 ≤ mF ≤
same setup in Fig. 6(a). mL ≤ M ) where Nt = 64, M = 3, {N1 , N2 , N3 } =
{8, 4, 1}, and AoD1 = (−30◦ , 30◦ ).
given by AoD1 ∈ {(−0.5, 0.5), (−0.77, 0.77)} (i.e.,
{(−30◦ , 30◦ ), (−50◦ , 50◦ )}) and the angular interval at ods. The optimized beamforming sequence obtained by
the receiver is set to the entire range of quantized Algorithm 1 is summarized in Table I. At low SNR, the
angles. The performance was measured by averaging the proposed method retransmits the codewords sounded at
instantaneous data rate in (26) over 106 Monte Carlo the first training stage by two times aimed at improving
runs, where the false alarm probability for each AoD the effective SNR. As the operating SNR increases, the
direction is set to PF A = 0.01 and the SNR is defined proposed method uses all codebook levels with a single
as σg2 P to account for the propagation path gain and codeword transmission to decrease the training overhead.
the transmit power. Note that the SNR values in the We observed that the use of a multilevel codebook (i.e.,
abscissa of each plot does not include the beamforming mF < mL ) yields good performance compared to the
gain, which is substantial in a Massive MIMO setting. use of a single level codebook (i.e., mF = mL ) for most
Thus, in examining the results, it is important to look at of the SNR range.
SNRs significantly below 0 dB in consideration of path
loss attenuation. SNR(dB) M {mF , . . . , mL } {WmF , . . . , WmL }
We consider a MIMO system with Nt = 64, Nr = 4, -16 ∼ -12 2 {2, 3} {2, 1}
M = 3, and AoD1 = (−30◦ , 30◦ ) where we adopt -10 ∼ -2 2 {2, 3} {1, 1}
Laplacian distributed angles of channel paths with the 0∼2 3 {1, 2, 3} {1, 1, 1}
mean angle µθ = 0 and the standard deviation σθ = 7
TABLE I: Optimized beam alignment sequence
[45]. In Fig. 8, we examine the data rate performance
based on the proposed method along with the codebook Next, we compare the performance of the proposed
shown in Fig. 6(a) on the quantized angles. It is seen methods to those of several codebook design techniques
mL [8], [11]. For M = 3 and Nt = 64, we considered the
that the proposed algorithms using Rm F
in (28) and
mL
the lower bound RmF in (31) yield almost the same channel model with a number of path L = 1 and the
performance. Thus, the simpler algorithm with (31) can AoDs are assumed to take continuous values, i.e., not
be used without much performance loss. For L = 1, quantized and uniformly distributed in AoD1 given by
simulation result without (A.1) matches well the analytic AoD1 = (−50◦ , 50◦ ). Fig. 10 shows the beam patterns
result, i.e., the result in (28) gives expressions for the obtained by analog beamforming [8] (i.e., NRF = 1) and
average data rate for a single path channel L = 1. For hybrid beamforming [11] where we set NRF = 4 and
L > 1, simulation results are obtained by choosing a adopt enough phase quantization bits so that the phase
single codeword that achieves the largest received signal quantization error on the performance is negligible.
power. These should be compared to that obatined with our pro-
For purpose of comparison, we consider the case posed method in Fig. 6. Given {N1 , N2 , N3 } = {8, 4, 1},
that the transmitter fixes the codebook levels {Fm : the proposed codebook design (i.e., the phase-shifted
mF ≤ m ≤ mL } for 1 ≤ mF ≤ mL ≤ M with DFT method shown in Fig. 6(a)) needs at most NRF = 8
WmF = 1. With the same setup used in Fig 8, we RF chains because of N1 = 8. For a fair comparison in
compared the data rate performance of the proposed terms of NRF , we also considered a modified method
method to those of others using the analytic result of where we replace the codebook in Fig. 6(a: Level 1) by
(28) in Fig. 9. It appears that the proposed method using the codebook in Fig. 6(a′ ) in order to maintain NRF = 4.
adaptively selected codebooks outperforms other meth- To evaluate the performance of the phase-shifted DFT
12

−1(−90o) −1(−90o) −1(−90o)


8.8 dB 12 dB 18 dB By averaging the main-lobe directivity gain of the
6.6 9.3 14 multilevel codebook over the interval AoD1 , we in-
4.4 6.2 9

2.2 3.1 4.5


vestigate the performance of all the methods including
o
0(0 )
o
0(0 )
o
0(0 ) analog, hybrid, and the proposed methods in conjunction
with Algorithm 1. The results marked (b) in Fig. 11
show that the data rate performance is further enhanced
especially at moderate-to-low SNR due to beam se-
1(90o) 1(90o) 1(90o)
(a: Level 1) (a: Level 2) (a: Level 3)
quence optimization. The methods for analog and hybrid
beamforming yield similar performance up to the SNR
−1(−90o) −1(−90o) −1(−90o)
9.9 dB 12 dB 17 dB around -4dB since only the codebooks at levels (2, 3) are
7.5 9.1 13
used at low SNR and its beam patterns are comparable to
5 6.1 8.5

2.5 3 4.3
each other shown in Fig. 10. Similarly, the phase-shifted
0(0o) 0(0o) 0(0o) DFT methods that use NRF ∈ {4, 8} also show the same
performance up to mid-range SNR.

1(90 )
o
1(90 )
o
1(90 )
o VI. C ONCLUSIONS
(b: Level 1) (b: Level 2) (b: Level 3)
We considered multi-resolution beam alignment se-
Fig. 10: Beam patterns on the continuous spatial domain quence design to improve the data rate performance
where Nt = 64 and M = 3: (a) analog beamforming achievable in mmWave systems. The resulting multi-
(NRF = 1) [8] and (b) hybrid beamforming (NRF = 4) level beamformer alignment design generates a beam-
[11]. For each codebook level, there are eight, sixteen, former sequence that counterbalances minimizing train-
and sixty-four codewords, respectively. ing overhead and maximizing beamforming gain. We
5 then provided a phase-shifted DFT method that effi-
(a) Analog (N =1)

4.5
RF
(b) Analog (N =1)
RF
ciently designs the multi-resolution codebook and easily
(a) Hybrid (N =4)
4
RF
(b) Hybrid (NRF=4)
extends to three-dimensional beamforming codebook.
3.5
(a) Phase−shifted DFT (N =4)
RF The proposed codebook design and beam alignment
(b) Phase−shifted DFT (N =4)
RF
sequence design provide an efficient means to improve
Rate (bps/Hz)

(a) Phase−shifted DFT (N =8)


3 RF
(b) Phase−shifted DFT (N =8)
2.5
RF
the data rate performance in large-scale mmWave MIMO
2
systems.
1.5

1 A PPENDIX
0.5 A. Proof of Lemma 1
0
−14 −12 −10 −8 −6
SNR (dB)
−4 −2 0 2 Define random variables X and Y by X ∼ CN (0, σ 2 )
and Z ∼ CN (0, σ̃z2 ) where σ 2 := Gm P σg2 and σ̃z2 :=
Fig. 11: Performance comparison with the previous
σz2 /Wm . We will omit the codebook index (m) for
designs where Nt = 64, M = 3, L = 1, and
notational simplicity. The average power of X is given
AoD1 = (−50◦ , 50◦ ): (a) fixed beamforming sequence
by
(i.e., mF = 1, mL = M , and WmF = 1) and (b)  2  2
E |X| = E |X|2 |A0 p(A0 ) + E |X| |A1 p(A1 )

optimized beamforming sequence by Algorithm 1.
(48)
method without beam sequence optimization, we first Z ∞ Z ∞
fix the number of codewords levels as M = 3 where = p(A1 ) |x|2 · pXZ (x, z|A1 ) dxdz,
mF = 1, mL = 3, and WmF = 1, marked (a) in −∞ −∞
Fig. 11. Hybrid beamforming design achieves better (49)
performance than analog beamforming design due to the where A0 and A1 denotes mutually exclusive events
well-designed codebooks especially at level 1. For the described in terms of the random variables X and Z,
fixed levels, the phase-shifted DFT method is superior i.e., A0 = {|X + Z|2 < γ} and A1 = {|X + Z|2 ≥ γ}.
to the other methods over the whole SNR range relative pXZ (x, z|A1 ) denotes a conditional joint probability
to data rate performance, even when we keep the same distribution function (pdf) given by [46]
number of RF chains NRF = 4. Using more RF chains
pXZ (x, z)
NRF = 8 in the phase-shifted DFT method yields pXZ (x, z|A1 ) = , (50)
slightly better performance than using NRF = 4 due p(A1 )
to the different codebooks at level 1 shown in Fig. 6(a, RR∈ A1 , otherwise pXZ (x, z|A1 ) = 0, where
for (X, Z)
a′ ). p(A1 ) = (X,Z)∈A1 pXZ (x, z)dxdz. (When there is no
13

p
γ − z̃I2 and I2 :=

ambiguity, we omit thesubscript X from pX (x) and where  I1 := max p 0, −z̃ R −
write p(x).) Note that E |X|2 |A0 = 0 in (48) because max 0, −z̃R + γ − z̃I2 .

we set |X|2 = 0 for (X, Z) ∈ A0 from a consideration In (54), we consider two separated cases for z̃I :
of no transmission. if z̃I2 > γ, all values of z̃R and r satisfy A1 (i.e.,
For computational purposes, we separate the complex the 2
first term on the right-hand side of (54)), and if
random variables X and Z into real part and imaginary z̃ I < γ, there is a range of r for each z̃R as to
part as satisfy A 1 (i.e., the second term on the right-hand side
of (54)).
 The
pfirst term  of (54) can be evaluated as
X = XR + jXI and Z = ZR + jZI , 2σ 2 1 − Φ γ/(σ̃z2 /2) . If τ denotes the second
and
p change the random variable −1 from X to R = term of (54), we provide a complete form of τ in (55).
XR + XI ∈ (0, ∞) and Θ = tan (XI /XR ) where After some tedious integration, we obtain the analytical
2 2

pR (r) = σ2r/2 e−r /σ and pΘ (θ) = 1/(2π). Then, the formula for the third term of (55), denoted as τ , shown
2 2

event of A1 is rewritten as in (56).


Note that equation (54) depends on σ 2 , σ̃z2 , and γ
|X + Z|2 ≥ γ

as seen from equations (55) and (56). The result may
⇔ R2 + 2r(ZR cos Θ + ZI sin Θ) + (ZR 2
+ ZI2 ) ≥ γ look complicated, so we apply an approximation in high

SNR regime (i.e., σ 2 ≫ σ̃z2 ) to derive a simple formula
⇔ (R + Z̃R )2 ≥ γ − Z̃I2 =: Ã1 ,

(51)
of σ 2 , σ̃z2 , and γ, where we use σ 2 = Gm P σg2 and
where NR := ZR cos Θ + ZI sin Θ and NI := σ̃z2 = σz2 /Wm . In this case, (54) is given by
−ZR sin Θ + ZI cos Θ. When (XR , XI ) and (ZR , ZI )  2 γ  1 −γ
are initially aligned with the real and imaginary axes, Z̃R E |X| = e− σ2 σ 2 + γ + e σn2
2 
and Z̃I denote the noise components in the direction of  − γ
2 − σγ2
RejΘ and perpendicular to those of RejΘ , respectively. −γ e 2σn
−e n
γ
Note that Z̃R and Z̃I are two independent Gaussian −  PF A
= e Gm P σg Gm P σg2 + γ +
2
2 2
variables, i.e., Z̃R ∼ N (0, σ̃z /2) and Z̃I ∼ N (0, σ̃z /2) 2
due to circular symmetry assumed for noise modeling. σz2 ln(1/PF A ) p 
Therefore, (49) can be rewritten in terms of R, Z̃R , and − P F A − PF A , (57)
Wm
Z̃I given by ∞
where γ = σz2 ln(1/PF A )/Wm and γ r2 · p(r)dr =
R
Z ∞Z ∞Z ∞ γ
 2 −
E |X| = p(Ã1 ) r2 × 2
e Gm P σg Gm P σg2 + γ .


−∞ −∞ 0

p r, z̃R , z̃I |Ã1 drdz̃R dz̃I (52)
ZZZ B. Proof of Theorem 1
= r2 · p (r, z̃R , z̃I ) drdz̃R dz̃I , (P.1) Since only the cosine function in the second term
(m)
(r,z̃R ,z̃I )∈Ã1 in (35) depends on the phase shift ωn , fθ,i,ωm becomes
(53) periodic with the period 2π. Define c+ = π(N −1)/N +ǫ
 and c− = π(N − 1)/N − ǫ for ǫ > 0. Then, we show
where p r, z̃R , z̃I |Ã1 is defined similar to (50) and the
nuisance variable Θ is integrated out in the analysis. that the values of cosine function evaluated at c+ and c−
Since the variables R, Z̃R , and Z̃I are mutually indepen- are the same, i.e., cos ((p′ − p) (π(N − 1)/N − c+ )) =
dent, the joint pdf for R, Z̃R , and Z̃I can be factored cos ((p′ − p)ǫ) = cos ((p′ − p) (π(N − 1)/N − c− )).
(m)
as p (r, z̃R , z̃I ) = p(r)p(z̃R )p(z̃I ). Then, (53) can be (P.2) For the i-th codeword fi ∈ Fm , recall that the
rewritten as summation index p in (35) is given by p = (i−1)Nm +q
∞ ∞
for 1 ≤ q ≤ Nm . From the center of the main beam
Z Z
E |X|2 =

p (z̃I ) p (z̃R ) ×
range in (39) (i.e., −1 + (2i − 1)Nm /Nt ), we set the
Z −∞ −∞
continuous angle as
r 2 · p(r)drdz̃Rdz̃I
(r+z̃R )2 >γ−z̃I
2
Z Z ∞ θ+ = −1 + (2i − 1)Nm /Nt + ǫ (58)
= p (z̃I ) p (z̃R ) ×
2 >γ
z̃I −∞ for ǫ > 0. In (33), we consider the argument of the sin
Z ∞
2
and exponential functions that correspond to θ given by
r · p(r)drdz̃R dz̃I
Z0 Nt (θ+ + 1) + 1 1 + Nm + ǫNt
Z ∞ p− =q− , (59)
+ p (z̃I ) p (z̃R ) × 2 2
2 <γ
z̃I −∞
Z I1 Z ∞  where 1 ≤ q ≤ Nm . Then, we define p′ := Nm (2i−1)+
r 2 p(r)dr + r 2 p(r)dr dz̃R dz̃I , (54) 1 − p = −q + iNm + 1 ∈ {(i − 1)Nm + 1, . . . , iNm }
0 I2 that gives a one-to-one correspondence between p and
14

√ s
γ
Z !!
2 γ − z̃I2
τ = 2σ p(z̃I ) · 1−Φ dz̃I

− γ
σ̃z2 /2
Z √
γ
 Z −√γ−z̃I2  √ 2  q 2 !
2
− −z̃R − γ−z̃I 2 2
+ p(z̃I ) − p(z̃R ) · e σ + −z̃R − γ − z̃I dz̃R

− γ −∞
Z √γ−z̃I2  √ 2  q 2 ! 
2
− −z̃R + γ−z̃I
+ p(z̃R ) · e σ 2 + −z̃R + γ − z̃I2 dz̃R  dz̃I
−∞
2
e−γ/σ̃z
 r  
2 γ 1 − 2σ̃γ2
= 2σ −1 + Φ + e z + 2
σ̃z2 /2 2

σ̃2
2 1 + σz2
   

γ
v
γ−z̃ 2
Z
1  γ − z̃I2 1 γ − z̃I2 
u
− σ2 +σ̃I2 2
+ P (z̃I ) · e 2 + σ +    · erf t   dz̃I ,
z
 u
q 
σ̃2

1 1 √
2 σ12 + 1
σ̃ 2
σ̃z σ2 + σ̃z2 − γ 1 + σz2 σ̃z2 σ̃z2 1 + σz2
(55)
where erf(·) denotes the error function [32, Eq. 8.250.1].

  γ 

2 −2
   −1 ! r
2 2
σ̃z
!
1 σ̃
 γ 1+ z 1 1 1 − γ
σ̃z 
2
2σ̃z 1+
τ′ = + σ 2
+ + e 2 2
σ +σ̃z
1 + 1 − e σ2 
σ2 2 σ2 σ̃z2 σ2
q
1 1
σ̃z σ2 + σ̃z2
−2 3 3   r !#
σ̃ 2 σ̃z2 σ̃z2 2 − σ2 +γ σ̃2 σ̃z2 σ̃z2 2 − γ2 σ̃1z2 + σ2 +1 σ̃z2 σ̃z2 − σ̃γ2
  
− 1 + z2 1+ 2 e z − 1+ 2 e −γ 1+ 2e z .
σ 2 σ 2 σ σ
(56)

p′ . As in (59), we evaluate the same term for the angle codeword, we consider the indices p = (i − 1)Nm + q
θ− = −1 + (2i − 1)Nm /Nt − ǫ and the index p′ as and p′ = (i − 1)Nm + q ′ for q, q ′ ∈ {1, . . . , Nm }. Then,
(m)
Nt (θ− + 1) + 1

1 + Nm + ǫNt
 fθ,i,ωm can be rewritten as

p − =− q−
2 2

Nm
 +
 (m) 1 X sin2 (π q̃)
Nt (θ + 1) + 1 fθ,i,ωm = 2
=− p−
 
. Nt Nm q=1 sin2 π q̃
2 Nt
(60) Nm Nm   
X X N −1
Note from (60) that for each term in (59) derived at θ+ , + 2 cos (q ′ − q) π − ωm
q=1 q′ =q+1
N
there is the same term with a negative sign derived at 
θ− . Then, we have sin (π q̃) sin (π q̃ ′ ) 
 H ×     , (61)
(m)
ut (θ− )H fi = ejωm (Nm (2i−1)+1) ut (θ+ )H fi
(m)
, sin π q̃ sin π q̃ ′ 
Nt Nt

because for the summand in (33), two minus signs on where q̃ = q − (Nm + ǫNt − 1)/2 and q̃ ′ = q ′ − (Nm +
(m)
the sin function cancel each other and the phase of the ǫNt − 1)/2. Since the function fθ,i,ωm does not depend
exponential function changes from negative to positive. on the index i, we have the claim. 
(m)
Since the function fθ,i,ωm is obtained from the square
(m) (m)
of (33), fθ− ,i,ωm = fθ+ ,i,ωm holds.
(P.3) Given any i ∈ {1, . . . , N/Nm }, the function R EFERENCES
(m)
fθ,i,ωm is evaluated on the spatial angle in one-half
of the main beam interval in (39) by (P.2), e.g., set [1] S. Noh, M. D. Zoltowski, and D. J. Love, “Multi-resolution
codebook based beamforming sequence design in millimeter-
θ = −1 + (2i − 1)Nm /Nt + ǫ ∈ [−1 + (2i − wave systems,” in Proc. IEEE Global Commun. Conf., to be
1)Nm /Nt , −1 + (2iNm − 1)/Nt ] as in (58). For the i-th published, 2015.
15

[2] S. Sun, T. S. Rappaport et al., “MIMO for millimeter-wave [23] B. Gopalakrishnan and N. D. Sidiropoulos, “Cognitive transmit
wireless communications: Beamforming, spatial multiplexing, or beamforming from binary CSIT,” IEEE Trans. Wireless Com-
both?” IEEE Commun. Mag., vol. 52, no. 12, pp. 110 – 121, Dec. mun., vol. 14, no. 2, pp. 895 – 906, Feb. 2015.
2014. [24] J. Seo, Y. Sung et al., “Training beam sequence design for
[3] T. S. Rappaport, S. Sun et al., “Millimeter wave mobile commu- millimeter-wave MIMO systems: A POMDP framework,” sub-
nications for 5G cellular: It will work!” in Proc. IEEE Access, mitted for publication. [Online]. Available: http://arxiv.org/abs/
vol. 1, no. 1, 2014, pp. 335 – 349. 1410.3711.
[4] W. Roh, J.-Y. Seol et al., “Millimeter-wave beamforming as an [25] M. R. Akdeniz, Y. Liu et al., “Millimeter wave channel modeling
enabling technology for 5G cellular communications: Theoretical and cellular capacity evaluation,” IEEE J. Sel. Areas Commun.,
feasibility and prototype results,” IEEE Commun. Mag., vol. 52, vol. 32, no. 6, pp. 1164 – 1179, Jun. 2014.
no. 2, pp. 106 – 113, Feb. 2014. [26] A. M. Sayeed and V. Raghavan, “Maximizing MIMO capacity
[5] M. D. Renzo, H. Haas et al., “Spatial modulation for generalized in sparse multi path with reconfigurable antenna arrays,” IEEE
MIMO: Challenges, opportunities and implementation,” in Proc. J. Sel. Topics Signal Process., vol. 1, no. 1, pp. 156 – 166, Jun.
IEEE, vol. 102, no. 1, Jan. 2014, pp. 56 – 103. 2007.
[6] J. Wang, Z. Lan et al., “Beam codebook based beamforming [27] A. M. Sayeed, “Deconstructing multi antenna fading channels,”
protocol for multi-Gbps millimeter-wave WPAN systems,” IEEE IEEE Trans. Signal Process., vol. 50, no. 10, pp. 2563 – 2579,
J. Sel. Areas Commun., vol. 27, no. 8, pp. 1390 – 1399, Oct. Oct. 2002.
2009. [28] A. Tadaion, M. Derakhtian et al., “A fast multiple-source detec-
[7] J. Brady, N. Behdad, and A. Sayeed, “Beamspace MIMO for tion and localization array signal processing algorithm using the
millimeter-wave communications: System architecture, modeling, spatial filtering and ML approach,” IEEE Trans. Signal Process.,
analysis, and measurements,” IEEE Trans. Antennas Propag., vol. 55, no. 5, pp. 1815 – 1827, May 2007.
vol. 61, no. 7, pp. 2931 – 2941, Jul. 2013. [29] A. M. Hunter, J. G. Andrews, and S. Weber, “Transmission
[8] S. Hur, T. Kim et al., “Millimeter wave beamforming for wire- capacity of ad hoc networks with spatial diversity,” IEEE Trans.
less backhaul and access in small cell networks,” IEEE Trans. Wireless Commun., vol. 7, no. 12, pp. 5058 – 5071, Dec. 2008.
Commun., vol. 61, no. 10, pp. 4391 – 4403, Oct. 2013. [30] T. Bai, A. Alkhateeb, and R. W. H. Jr., “Coverage and capacity
of millimeter-wave cellular networks,” IEEE Commun. Mag.,
[9] J. Song, J. Choi et al., “Adaptive millimeter wave beam-
vol. 52, no. 9, pp. 70 – 77, Sep. 2014.
alignment for dual-polarized broadcast MIMO systems,” IEEE
[31] S. M. Kay, Fundamentals of Statistical Signal Processing: De-
Trans. Wireless Commun., vol. PP, no. 99, pp. 1 – 12, Jul. 2015.
tection Theory. Englewood Cliffs, New Jersey: Prentice-Hall,
[10] O. E. Ayach, S. Rajagopal et al., “Spatially sparse precoding in
1998.
millimeter wave MIMO systems,” IEEE Trans. Wireless Com- [32] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
mun., vol. 13, no. 3, pp. 1499 – 1513, Mar. 2014. Products. New York, NY: 7th ed., Academic Press, 2007.
[11] A. Alkhateeb, O. E. Ayach et al., “Channel estimation and hybrid [33] H. Alzer, “On some inequalities for the incomplete Gamma
precoding for millimeter wave cellular systems,” IEEE J. Sel. function,” Math. Comput., vol. 66, no. 218, pp. 771 – 778, Apr.
Topics Signal Process., vol. 8, no. 5, pp. 831 – 846, Oct. 2014. 1997.
[12] J. Song, J. Choi, and D. J. Love, “Codebook design for hybrid [34] D. J. Love and R. W. Heath Jr., “Equal gain transmission in
beamforming in millimeter wave systems,” in Proc. IEEE Int. multiple-input multiple-output wireless systems,” IEEE Trans.
Conf. on Commun., London, UK, Jun. 2015. Commun., vol. 51, no. 7, pp. 1102 – 1110, Jul. 2003.
[13] A. Liu and V. Lau, “Phase only RF precoding for massive MIMO [35] 3GPP TSG RAN WG1 #62, R1-105011, “Way forward on 8Tx
systems with limited RF chains,” IEEE Trans. Signal Process., codebook for Rel. 10 DL MIMO,” 2010.
vol. 62, no. 17, pp. 4505 – 4515, Sep. 2014. [36] C. Lim, T. Yoo et al., “Recent trend of multiuser MIMO in LTE-
[14] S. Noh, M. D. Zoltowski, and D. J. Love, “Training sequence Advanced,” IEEE Wireless Commun. Mag., vol. 51, no. 3, pp.
design for feedback assisted hybrid beamforming in massive 127 – 135, Mar. 2013.
MIMO systems,” submitted for publication. [Online]. Available: [37] J. Butler and R. Lowe, “Beamforming matrix simplifies design of
http://arxiv.org/abs/1407.1786. electrically scanned antennas,” Electron. Design, no. 9, pp. 170
[15] J. H. Kotecha and A. M. Sayeed, “Transmit signal design for – 173, Apr. 1961.
optimal estimation of correlated MIMO channels,” IEEE Trans. [38] K.-C. Huang and D. J. Edwards, Millimetre Wave Antennas for
Signal Process., vol. 52, no. 2, pp. 546 – 557, Feb. 2004. Gigabit Wireless Communications: A Practical Guide to Design
[16] S. Noh, M. D. Zoltowski et al., “Pilot beam pattern design for and Analysis in a System Context. New York, NY: Wiley, 2008.
channel estimation in massive MIMO systems,” IEEE J. Sel. [39] C. Chang, R. Lee, and T. Shih, “Design of a beam switch-
Topics Signal Process., vol. 8, no. 5, pp. 787 – 801, Oct. 2014. ing/steering Butler matrix for phased array system,” IEEE Trans.
[17] J. Choi, Z. Chance et al., “Noncoherent trellis coded quantiza- Antennas Propag., vol. 58, no. 2, pp. 367 – 374, Feb. 2010.
tion: A practical limited feedback technique for massive MIMO [40] K. Wu, Y. J. Cheng et al., “Substrate-integrated millimeter-wave
systems,” IEEE Trans. Commun., vol. 61, no. 12, pp. 5016 – and teraherz antenna technologies,” Proc. IEEE, vol. 100, no. 7,
5029, Dec. 2013. pp. 2219 – 2232, Jul. 2012.
[18] J. Choi, D. J. Love, and P. Bidigare, “Downlink training tech- [41] P. Stoica and R. L. Moses, Spectral Analysis of Signals. Engle-
niques for FDD massive MIMO systems: Open-loop and closed- wood Cliffs, NJ, USA: Prentice-Hall, 2005.
loop training with memory,” IEEE J. Sel. Topics Signal Process., [42] D. Ying, F. W. Vook et al., “Kronecker product correlation model
vol. 8, no. 5, pp. 802 – 814, Oct. 2014. and limited feedback codebook design in a 3D channel model,”
[19] X. Rao and V. Lau, “Distributed compressive CSIT estimation in Proc. IEEE Int. Conf. on Commun., Sydney, AUS, Jun. 2014.
and feedback for FDD multi-user massive MIMO systems,” IEEE [43] J. Choi, T. Kim et al., “Exploiting the preferred domain of FDD
Trans. Signal Process., vol. 62, no. 12, pp. 3261 – 3271, Jun. massive MIMO systems with uniform planar arrays,” in Proc.
2014. IEEE Int. Conf. on Commun., London, UK, Jun. 2015.
[20] J. So, D. Kim et al., “Pilot signal design for massive MIMO [44] A. Sayeed and N. Behdad, “Continuous aperture phased MIMO:
systems: A received signal-to-noise-ratio-based approach,” IEEE Basic theory and applications,” in Proc. of Allerton Conf. on
Signal Process. Lett., vol. 52, no. 5, pp. 549 – 553, May 2015. Commun., Control, and Computing, Allerton, IL, Oct. 2010.
[21] Y.-C. Liang and F. P. S. Chin, “Downlink channel covariance [45] A. Forenza, D. J. Love, and R. W. Heath Jr., “Simplified spatial
matrix (DCCM) estimation and its applications in wireless DS- correlation models for clustered MIMO channels with different
CDMA systems,” IEEE J. Sel. Areas Commun., vol. 19, no. 2, array configurations,” IEEE Trans. Veh. Technol., vol. 56, no. 4,
pp. 222 – 232, Feb. 2001. pp. 1924 – 1934, Jul. 2007.
[22] B. M. Hochwald and T. L. Marzetta, “Adapting a downlink array [46] L. C. Andrews and R. L. Phillips, Mathematical Techniques for
from uplink measurements,” IEEE Trans. Signal Process., vol. 49, Engineers and Scientists. Bellingham, Washington: SPIE Press,
no. 3, pp. 642 – 653, Mar. 2001. 2003.