5096 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO.

19, OCTOBER 1, 2018

Structured Random Compressed Channel Sensing for

Millimeter-Wave Large-Scale Antenna Systems
Cheng-Rung Tsai and An-Yeu Wu , Fellow, IEEE

Abstract—In millimeter-wave (mmWave) communications, abundant spectrum resource [1]. In mmWave communications,
large-scale antenna system (LSAS) is considered an essential tech- large-scale antenna system (LSAS) is fundamental at both base
nology to realize beamforming gain to compensate for huge prop- stations (BSs) and user equipment (UE) to provide beamforming
agation loss. However, channel estimation for LSASs poses a
formidable challenge, especially when hybrid analog–digital struc- gain to compensate for the huge propagation loss at mmWave
tures are adopted for ensuring reasonable complexity and cost. To bands [2], [3]. For ease of implementation, hybrid analog–digital
address this challenge, compressed channel sensing (CCS) is lever- (AD) LSAS [4]–[6] driven by a small number of radio frequency
aged to measure mmWave channels via a random sensing code- (RF) chains is a feasible solution to avoid the fabrication cost
book. By exploiting the sparse nature in mmWave channels, only and energy consumption of a large number of high-frequency
a small number of measurements is required for channel recovery.
However, signaling or storing the configuration of a full random mixed-signal components used in all antennas. Moreover, the
sensing codebook leads to a huge burden. Moreover, because the hybrid AD LSAS is capable of realizing both beamforming and
sensing beam of a full random sensing codebook always spreads its spatial multiplexing gains via the array signal processing in both
power over the channel, the CCS has a stringent requirement on the analog and digital domains.
signal-to-noise ratio (SNR) for robust channel recovery. To over- To exploit the full benefits of the LSAS, channel knowledge
come these issues, we propose a structured random sensing code-
book inspired by the random convolutional measurement process. is crucial for both BSs and UE. However, the large number of
Owing to its structured nature, signaling or storing overhead from antennas in the LSAS leads to a large size of the channel matrix,
the codebook configuration is significantly reduced. Additionally, which results in huge resource burdens for channel estimation.
the structured random sensing codebook can concentrate its power Moreover, traditional reference signal (RS)-based training is not
in a local angle coverage for a sectorized cell, thus improving the applicable to the hybrid AD LSAS because the baseband can-
robustness in low-SNR regimes. Simulation results demonstrate
that the recovery performance of the proposed structured random not directly access the individual signal corresponding to each
sensing codebook design is comparable to that of the full random antenna. By exploiting the fact that mmWave channels tend
design. Regarding the robustness in low-SNR regimes, the recovery to have few dominant propagation paths owing to the limited
performance is substantially improved by the structured random scattering effect [7], a closed-loop beam training referred to as
sensing codebook with power concentration for a local angle cov- hierarchical channel sensing (HCS) [8]–[10] is proposed to ad-
erage.
dress the challenge of channel estimation in the LSAS. In HCS,
Index Terms—Millimeter-wave communications, large-scale an- a divide-and-conquer search process is carried out across the
tenna systems, channel estimation, compressed channel sensing, hierarchy of multi-resolution beams, and the desirable beams
structured random compressed sensing.
are refined persistently according to the UE report until reach
the target resolution. HCS can greatly reduce the sweeping over-
I. INTRODUCTION head (i.e., number of employed resources for beam sweeping)
when the number of levels in the hierarchy is unlimited. How-
HE rapid growth of mobile traffic has imposed tremen-
T dous demands of high data rates on fifth-generation (5G)
cellular systems. Toward this end, millimeter-wave (mmWave)
ever, considering control complexity and feedback overhead,
the number of levels is limited in practice, thus the overhead
reduction of is minor. Moreover, the sweeping overhead of HCS
communications are the key enabling techniques that utilize the
linearly increases with the number of UEs since the BS needs
to sweep every possible beam according to the report from
Manuscript received January 13, 2018; revised May 13, 2018 and July 10,
2018; accepted July 21, 2018. Date of publication July 31, 2018; date of current each UE.
version September 4, 2018. The associate editor coordinating the review of Recently, an open-loop beam training called compressed
this manuscript and approving it for publication was Prof. Mathini Sellathurai. channel sensing (CCS) has been explored in [11]–[20] by lever-
This work was financially supported in part by the Ministry of Science and
Technology of Taiwan under Grants MOST 105-2622-8-002-002 and 106-2221- aging the tools of compressed sensing [21]–[23]. In CCS, the
E-002-20-MY3, and in part by the sponsored by MediaTek, Inc., Hsin-chu, random sensing codebook is swept to measure the channel. Each
Taiwan. (Corresponding author: Cheng-Rung Tsai.) sensing beam (i.e., random beamforming vector) of the sensing
The authors are with the Department of Electrical Engineering, Grad-
uate Institute of Electronics Engineering, and National Taiwan University, codebook is a set of unimodular entries with quantized random
Taipei 10617, Taiwan (e-mail:, [email protected]; andywu@ phases, and thus it can be realized in analog phase-shifting ar-
ntu.edu.tw). rays of the hybrid AD LSAS. By exploiting the sparse nature in
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. mmWave channels, CCS only requires a small number of mea-
Digital Object Identifier 10.1109/TSP.2018.2860545 surements for channel recovery. As CCS works in an open-loop

1053-587X © 2018 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.
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5097

manner, it does not need complicated control protocol and UE

report within beam sweeping, and enables simultaneous train-
ing for multiple UE without increasing the sweeping overhead.
Recently, CCS has been further extended to frequency-selective
wideband channel estimation [15], [16], and channel covari-
ance estimation [17], [18] for the LSAS. CCS for estimating
the mmWave channels with structured sparsity has also been
investigated in [19], [20]. However, two critical issues in CCS
need to be addressed in practice:
1) Signaling/storage overhead of codebook configuration:
In CCS, the knowledge of both transmit and receive sens-
ing codebooks is necessary for channel recovery. Thus,
the configuration of the sensing codebook adopted at the
BS should be known to the UE. However, the configura-
tion of the full random sensing codebook must be explic-
itly signaled to the UE or stored in the UE owing to its
completely unstructured nature. It leads to an unbearable
signaling or storage overhead. In many works, e.g., [24]–
[27], this issue has also been indicated, and some sensing
matrices with deterministic constructions have been pro-
posed to avoid it. However, these solutions usually require Fig. 1. Schematic of the compressed channel sensing (CCS) approach adopt-
ing (a) the full random sensing codebooks, and (b) the structured random sensing
additional overhead (e.g., hardware and latency) to gen- codebooks in the large-scale antenna systems (LSASs).
erate the sensing matrices on-the-fly. Moreover, most of
them cannot guarantee comparable recovery performance
to that of the full random sensing matrix [28]. and N denote the numbers of measurements and antennas,
2) Stringent requirement on SNR for robust channel recov- respectively.
ery: The sectorized cellular systems are typically adopted r We present a random convolutional measurement process
for outdoor deployments, especially when LSASs are con- extended from one class of compressed sensing tools [36]–
sidered [29]–[31]. In a sectorized cell, the BS adopts a [38] for the CCS adopting the proposed sensing codebooks.
LSAS to serve a local angle coverage instead of a full Then, a theoretical bound derived from this measurement
angle coverage to increase directivity gain, enhance ca- process is offered to show that the proposed sensing code-
pacity by spatial division [49], and avoid the limitation of book design compares favorably to the full random design.
beam scanning coverage [32]–[34]. However, as shown r We further design a sensing codebook with the capability
in Fig. 1(a), the sensing beam of the full random sens- of power concentration based on the proposed dual-stage
ing codebook always spreads its power over a full angle codebook structure. In the proposed sensing codebook de-
coverage. Without directivity gains, CCS has a stringent sign, we can configure the direction and the beamwidth of
requirement on signal-to-noise ratio (SNR) since channel each sensing beam by simply adjusting the direction of its
recovery has the minimum requirement on the ratio of sampling beam and the spreading factor of its spreading
signal power after measurement to noise power [35]. It sequence, respectively. As illustrated in Fig. 1(b), given a
severely limits the operating regime of CCS, especially local angle coverage for a sectorized cell, we can design
for mmWave with significant propagation loss. a structured random sensing codebook to concentrate its
In this paper, we aim to leverage the benefits of CCS for power in the intended angle coverage to increase directiv-
channel estimation in mmWave LSASs. To address the two ity gains. Thus, the robustness for channel recovery can be
aforementioned issues, we present a new class of CCS inspired improved in low-SNR regimes.
by the structured random compressed sensing [36]–[38]. The r We also show that the CCS adopting the proposed sens-
main contributions of this paper are summarized as follows: ing codebook with power concertation in a local angle
r We propose a novel dual-stage sensing codebook consist- coverage can also achieve stable recovery, and offer the
ing of a set of sampling beams randomly selected from the corresponding theoretical bound. Meanwhile, based on the
discrete Fourier transform (DFT) codebook for random theoretical bound, the design of spreading factor for a local
subsampling, and a modulation with a spreading sequence angle coverage is derived.
in the spatial domain for random convolution. Due to the Finally, we present simulation results on the recovery perfor-
structured nature, configuration of the proposed sensing mance of the proposed strategy and show that it can alleviate
codebook can be implicitly represented via (i) the sampling the stringent requirement on SNR for robust channel recovery.
indices of the selected DFT beams, and (ii) the spreading The rest of this paper is organized as follows. In Section II,
sequence, rather than explicit matrix. Thus, the overhead the system and channel models are introduced. The structured
from signaling or storing the proposed sensing codebook random sensing codebook design is presented in Section III.
can be reduced from O(M N ) to O(M log2 N ), where M In Section IV, the structured random sensing codebook with
5098 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

power concentration for a local angle coverage is introduced. nas, respectively. Both physical AoDs and AoAs are considered
Simulation results are presented in Section V. Finally, the continuous and uniformly distributed within the intended angle
conclusions are drawn in Section VI. coverages, respectively. The complex gains the complex gains
{αk }k =1,...,K are assumed to be i.i.d. (0, σα2 k ) for k = 1, . . . , K,
II. CHANNEL AND SYSTEM MODELS where σα2 k is the average power of the k-th cluster. To constrain
the total channel power, i.e., E [||H||2F ] = NT NR [39], [40], the
A. Notation 
sum of average cluster power should satisfy K k =1 σα k = K.
2
Throughout this paper, a normal-faced letter a denotes a To apply compressed sensing for channel estimation, we use
scalar, a bold-faced lowercase letter a denotes a vector, and the virtual angular domain (VAD) representation [26] to provide
a bold-faced uppercase letter A denotes a matrix. Other opera- a discrete approximation to the physical channel (1). Instead of
tions used in this paper are defined as follows: taking the AoDs and AoAs from arbitrary angles, the VAD repre-
– Ap,: , A:,q , and Ap,q is the p-th row vector, q-th column sentation quantizes them via fixed virtual angle grids with finite
vector, and (p, q)-th entry of A, respectively. resolutions. Using the VAD representation, we can approximate
– AΩ is the submatrix formed by collecting the columns with the physical channel (1) by
indices in the set Ω.
– AΩ,Ω is the submatrix formed by taking a block of the H ≈ H̄ = AR H̄ω AH
T , (2)
entries of A whose row and column are indexed by Ω.
– AT , A∗ , AH , A−1 , and A† , and vec(A) denote the trans- where H̄ω ∈ C G R ×G T is the VAD channel matrix, AT ∈
pose, conjugate, conjugate-transpose, inverse, pseudo- C N T ×G T and AR ∈ C N R ×G R denote the VAD transformation
inverse, and vectorization of A, respectively. dictionaries consisting of the steering vectors corresponding to
– A ⊗ B is the Kronecker product of A and B. the transmit and receive virtual angle grids with resolutions GT
– E[·], | · |, || · ||1 , || · ||2 , and || · ||F denote the expectation and GR , respectively. To effectively approximate the physical
operation, element-wise absolute value or cardinality for a channel (1), both transmit and receive virtual angle grids should
set, 1-norm, 2-norm, and Frobenius norm, respectively. be applied with sufficient high resolutions, such that GT ≥ NT
– FN denotes a unitary DFT matrix with size of N , where and GR ≥ NR . When GT = NT and GR = NR , AT and AR
Fp,q
N = (NT )
−1/2 −2π j (p−1)(q −1)/N
e . can be represented as the unitary DFT matrices FN T and FN R .
– C = Circ(c) denotes a circulant matrix specified by the The VAD representation (2) is defined for a channel with
sequence c, which appears as the first column of C. full transmit and receive angle coverages. However, in practical
cellular systems, because outdoor deployment typically adopts
B. Physical and Virtual Angular Domain Channel Models the sectorized cellular systems [29]–[31], the constraint on AoD
should be considered during channel estimation. Based on the
In an open-loop beam training, each UE can be considered an
prior knowledge of local transmit angle coverage corresponding
independent subsystem. Without loss of generality, we consider
to a sectorized cell, we can directly consider that the transmit
the LSAS with a BS with NT antennas and a reference UE with
virtual angle girds out of the local transmit angle coverage have
NR antennas. Both are equipped with half-wavelength-spaced
no energy. Thus, given a local transmit angle coverage, we have
uniform linear arrays (ULAs).
a subsequence ΘT ⊂ {1, . . . , GT } such that
In this study, we use the cluster-based physical channel model,
which is also widely used in the literatures [11]–[18], to cap- H̄:,q
ω = 0, if q ∈
/ ΘT . (3)
ture the limited scattering channel. In this model, the physical
channel is assumed to be the sum of K clusters. Since the In contrast, the AoA should be considered within a full angle
measurement campaigns in dense-urban environments [7] have coverage because the location and orientation of each UE are
revealed that mmWave channels typically have only a few scat- random. Finally, let H̄Θ T = H̄Θ
ω
T
∈ C G R ×|Θ T | , we can rewrite
tering clusters, and the angle spreading within each cluster is (2) as
also small. Thus, we can assume that K  NR ≤ NT , and each
cluster contributes a single dominant propagation path between H̄ = AR H̄Θ T (AΘ T H
T ) . (4)
the BS and the UE because most power of each cluster is con-
centrated around bore-sight direction. Under these assumptions, C. Channel Estimation by Compressed Channel Sensing
a narrowband block-fading channel can be expressed as
 In this work, we aim to study the sensing codebook design
K
NT NR  for CCS. As the sensing beams of the sensing codebook are
H= αk a(NR , θkA oA ) aH (NT , θkA oD ), (1)
K typically formed in the analog domain, they can be directly
k =1
adopted for both narrowband CCS [11]–[14] and wideband CCS
N R ×N T
where H ∈ C is the physical channel between the BS [15]–[18]. To focus on the sensing codebook design and make
and the UE, αk is the complex gain of the k-th propaga- the paper concise, we consider only the narrowband CCS, and
tion path. The steering vectors a(NR , θkA oA ) ∈ C N R ×1 and the extensions to wideband CCS are straightforward.
a(NT , θkA oD ) ∈ C N T ×1 correspond to the physical angle-of- In the m-th measurement of the CCS, the BS forms a unit-
arrival (AoA) θkA oA and the angle-of-departure (AoD) θkA oD of norm transmit sensing beam pm ∈ C N T ×1 , and the UE forms
the k-th propagation path for the ULAs with NT and NR anten- a unit-norm receive sensing beam qm ∈ C N R ×1 to measure the
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5099

channel. Considering the physical channel (1), we have TABLE I

FEATURE COMPARISON OF DIFFERENT SENSING CODEBOOK DESIGNS
√
ym = ρqH H
m Hpm sm + qm nm ∈ C
1×1
(5)
(a) √
= ρ(pTm ⊗ qH
m )vec(H)sm + qm nm ,
H
(6)
where (a) follows from the vectorization originally introduced
in [8], sm ∈ C 1×1 is the m-th RS transmitted through the m-th
transmit sensing beam, nm ∈ C N R ×1 is a Gaussian noise vector
with variance σn2 in the m-th measurement, and ρ denotes the
average received power before measurement. Thus, the SNR for
each measurement is represented as ρ/σn2 .
During the beam sweeping for CCS, the sensing code-
books P = [p1 , . . . , pM ] ∈ C N T ×M and Q = [q1 , . . . , qM ] ∈
C N R ×M with M transmit and receive sensing beams are swept
at the BS and the UE, respectively. Since CCS adopts sensing
beams instead of orthogonal RSs to measure the channel, the
transmitted RSs are typically identical within the beam sweep-
ing, i.e., sm = 1 for m = 1, . . . , M . After all the sensing beams Moreover, the RIP leads to stability with respect to the non-
are swept, we have a measurement vector y = [y1 , . . . , yM ]T ∈ ideal measurements introduced by the additive noise and the
C M ×1 by stacking the M measurements measurement matrix mismatch. Once ĥω is derived from chan-
⎡ T ⎤ ⎡ H ⎤ nel recovery, the estimated channel matrix can be given by
p1 ⊗ qH 1 q1 n1
√ ⎢ ⎥ ⎢ . ⎥ Ĥ = AR vec−1 (ĥω )AH
T .
y = ρ⎣ .. (11)
. ⎦ vec(H) + ⎣ .. ⎦ . (7)
pTM ⊗ qH
M qH
M nM In the CCS, it is worth mentioning that the VAD channel
To formulate a channel recovery problem from (7), we apply vector h̄ω typically exhibits compressibility rather than purely
the VAD representation (2) into (7) [13]. Then, we have sparsity because the physical AoAs {θkA oA }k =1,...,K and AoDs
⎡ T ⎤ ⎡ H ⎤ {θkA oD }k =1,...,K are actually continuous. Fortunately, as proved
p1 ⊗ qH 1 q1 n1 in [30], the RIP restriction guarantees stable recovery whether
√ ⎢ .. ⎥ ∗ ⎢ . ⎥
y ≈ ρ⎣ . ⎦ (AT ⊗ AR ) vec(H̄ω ) + ⎣ .. ⎦ the signal (or channel) is sparse or compressible.
pTM ⊗ qH
M qH
M nM
  D. Conventional Full Random Sensing Codebook Design
Φ e
√ It is known that if Φ is a full random measurement matrix
≈ ρΦh̄ω + e . (8)
(e.g., the Gaussian or Bernoulli random matrices), it can sat-
Herein, Φ ∈ C M ×N denotes the equivalent measurement ma- isfy the RIP in a high probability when M ≥ O(δ −2 K log N )
trix, e ∈ C M ×1 is the noise vector, and h̄ω = vec(H̄ω ) ∈ C N×1 for δ ∈ (1, 0) [44]. To achieve full randomness in Φ, both
is a VAD channel vector, where N = GT GR . Since we have {pm }m =1,...,M and {qm }m =1,...,M should be generated ran-
M < N in the CCS, (8) is underdetermined. However, when domly and independently. However, two critical issues arise
h̄ω exhibits sparsity/compressibility, it can be recovered by the from the full random design in practice. First, due to the full ran-
1 optimization domness, to construct Φ for channel recovery, the configuration
√ of P should be explicitly delivered to the UE or stored in the UE,
ĥω = arg min ||h̄ω ||1 , s.t. ||y − ρΦh̄ω ||22 ≤ ε . (9) leading to a huge signaling or storage overhead. For instance,
h̄ ω
when the Bernoulli random matrix is adopted as a transmit sens-
The variable ε bounds the noise effect in the measurements of ing codebook comprising M = 256 sensing beams for the BS
the CCS. Several greedy-based recovery algorithms [41]–[43] with NT = 128 antennas, 215 bits should be delivered or stored.
have been proposed to find approximated solutions of the 1 The signaling or storage overhead is obliviously unbearable,
optimization in low complexity. According to the compressed especially when multiple sensing codebooks are adopted at dif-
snesing theory [21]–[23], stable channel recovery is guaranteed ferent BSs in cellular systems. Second, since the sensing beam of
if the equivalent measurement matrix Φ satisfies the restricted the full random sensing codebook always spreads its power over
isometry property (RIP). a full angle coverage, sufficiently high SNR, which is only at-
Definition 1 ((K, δ)− RIP [21]): An M × N measurement tained in the short-range mmWave connectivity, is necessary for
matrix Φ satisfies the (K, δ)− RIP if the CCS to achieve robust channel recovery. In [45], the DFT-
MTC sensing codebook constructed by a random-permuting
(1 − δ)||h̄ω ||22 ≤ ||Φh̄ω ||22 ≤ (1 + δ)||h̄ω ||22 , (10)
DFT codebook and baseband processing for minimizing total
for all K-sparse h̄ω . coherence (MTC) is proposed to reduce the required number of
The RIP ensures that all submatrices of Φ with size M × K measurements while maintaining reasonable SNR for channel
are close to an isometry, and are therefore, distance-preserving. recovery. However, its sweeping overhead is scaled with the
5100 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

Fig. 2. Block diagram of the structured random sensing codebook design at the BS and UE.

size of the DFT codebook, and it has no theoretical guarantee

Algorithm 1: Structured Random Sensing Codebook
for stable recovery when only a few measurements are given.
Design.
To address the two issues, a new class of CCS inspired by the
structured random compressed sensing [36]–[38] is presented Input NT , NR , and M
in this paper. To clarify the differences between prior work and 1: Generate a subset ΩG with M different indices
the proposed approach, we briefly summarize the comparisons downselected from {1, . . . , NT NR } uniformly at random.
in Table I. return subsampling index set ΩG = [ΩG ,1 , . . . , ΩG ,M ]
Base Station:
Input NT and ΩG
III. STRUCTURED RANDOM SENSING CODEBOOK BASED ON 1: Compute ΩT ,m = (ΩG ,m − 1)/NR + 1, for
THE RANDOM CONVOLUTIONAL MEASUREMENT PROCESS
m = 1, . . . , M
A. Proposed Structured Random Sensing Codebook Design 2: Generate a spreading sequence σ T with the length of NT
3: Compute ΔT = diag(σ T )
In the proposed sensing codebook, the transmit sensing beam
4: Compute P = ΔT FN T RTΩ T
pm and receive sensing beam qm for the m-th measurement are
return transmit sensing codebook P = [p1 , . . . , pM ]
constructed by the sampling beams {wT ,m ∈ C N T ×1 , wR ,m ∈
User Equipment:
C N R ×1 } and the beam spreaders {ΔT ∈ C N T ×N T , ΔR ∈
Input NR and ΩG
C N R ×N R }, respectively. Then, the sensing beams can be ex-
1: Compute ΩR ,m = ((ΩG ,m − 1)modNR ) + 1, for
pressed as
m = 1, . . . , M
{pm , qm } = {ΔT wT ,m , ΔR wR ,m } . (12) 2: Generate a spreading sequence σ R with the length of NR
3: Compute ΔR = diag(σ R )
We now delve into the details of dual-stage codebook struc- 4: Compute Q = ΔR FN R RTΩ R
ture: return receive sensing codebook Q = [q1 , . . . , qM ]
r Sampling beamformer: In this stage, wT ,m and wR ,m
are selected from the transmit and receive DFT codebooks
through the sampling indices ΩT ,m and ΩR ,m , as illus- In the proposed design, the spreading sequence is should
trated by the corresponding beam patterns (blue) in Fig. 2, be a unimodular sequence with good autocorrelation prop-
respectively. The sampling beams can be expressed as erty, such as the maximum-length (ML) sequence or the
  Zadoff–Chu (ZC) sequence. After the beam spreader, the
:,Ω :,Ω
{wT ,m , wR ,m } = FN T T , m , FN RR , m sampling beams wT ,m and wR ,m are spread over full an-
  gle coverages, as illustrated by the corresponding beam
= FN T rTT ,m , FN R rTR ,m , (13) patterns (blue) in Fig. 2.
To generate a structured random sensing codebook consisting
where rT ,m ∈ R1×N T and rR ,m ∈ R1×N R denote the sam-
of M sensing beams, these sensing beams are constructed by M
pling vectors for selecting the columns of FN T and FN R
sampling beams and a same beam spreader based on (12). Thus,
indexed by ΩT ,m and ΩR ,m , respectively.
r Beam spreader: In this stage, ΔT and ΔR represent we can express the transmit and receive sensing codebooks as
 
the modulations in the spatial domain with the spreading {P, Q} = ΔT FN T RTΩ T , ΔR FN R RTΩ R , (15)
sequences σ T ∈ C N T ×1 and σ R ∈ C N R ×1 , respectively.
Then, they can be expressed as respectively. The matrices RΩ T = [rTT ,1 , . . . , rTT ,M ]T ∈
RM ×N T and RΩ R = [rTR ,1 , . . . , rTR ,M ]T ∈ RM ×N R denote
{ΔT , ΔR } = {diag(σ T ), diag(σ R )} . (14) the sampling operators for selecting M sampling beams
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5101

from FN T and FN R according to the sampling index sets

ΩT = [ΩT ,1 , . . . , ΩT ,M ] and ΩR = [ΩR ,1 , . . . , ΩR ,M ], respec-
tively. Similar to (7), we can stack the M measurements in a
vector form after the sweeping of P and Q in (15)
⎡ T H ⎤
(ΔT FN T rTT ,1 ) ⊗ (ΔR FN R rTR ,1 )
√ ⎢
⎢ .. ⎥
⎥ vec(H) + e
y= ρ⎣ . ⎦
T H
(ΔT FN T rTT ,M ) ⊗ (ΔR FN R rTR ,M )
⎡ ⎤
rT ,1 ⊗ rR ,1
√ ⎢ .. ⎥ 
= ⎢ ρ⎣ ⎥ (ΔT FN )T ⊗ (ΔR FN )H vec(H) + e
. ⎦ T R

rT ,M ⊗ rR ,M
√ T
 
= ρ [rTG ,1 , . . . , rTG ,M ] (ΔT FN T )T ⊗ (ΔR FN R )H vec(H)+e
 Fig. 3. Illustration of the practical implementation of the structured random
RΩ G sensing codebook design in the hybrid AD LSAS equipped with four RF chains.

√  
= ρRΩ G (ΔT FN T )T ⊗ (ΔR FN R )H vec(H) + e, (16)
for  = 1, . . . , NT . As a result, the sum of phase-shifting
where RΩ G ∈ RM ×N T N R is the subsampling operator
angles {ωm , + ϕ }=1,...,N T can be realized in the analog
containing the M subsampling vectors {rG ,m = rT ,m
phase-shifting array with log2 NT angle quantization bits.
⊗ rR ,m }m =1,...,M ∈ R1×N T N R . In the proposed design,
On the other hand, if σ T is a ZC sequence, we have
RΩ G is given by collecting M rows of IN indexed by
 
ΩG = [ΩG ,1 , . . . , ΩG ,M ], and ΩG is downselected from
πτ ( − 1)2 NT , for even NT ;
{1, . . . , NT NR } uniformly at random. Then, according to ΩG , ϕ = (20)
the corresponding sampling index sets for sampling operators πτ k( − 1)/ NT , for odd NT ,
RΩ T and RΩ R are respectively given by for  = 1, . . . , NT . The root index τ is a posi-
{ΩT ,m , ΩR ,m } tive integer relatively coprime to NT . In this case,
{ωm , + ϕ }=1,...,N T can be also realized in the analog
= { (ΩG ,m − 1)/ NR + 1 , ((ΩG ,m − 1)modNR ) + 1} , phase-shifting array with 1 + log2 NT angle quantization
(17) bits. As a result, the sensing beams of P in (15) naturally
for m = 1, . . . , M . Finally, the design of the structured ran- meet the practical constraints of the hybrid AD hardware.
dom sensing codebook for the BS and UE is summarized in In a similar manner, each sensing beam of Q in (15) can
Algorithm 1. be also realized in the analog phase-shifting arrays. In
the hybrid AD LSAS, each RF chain may be equipped
with one analog phase-shifting array. Thus, as illustrated
B. Practical Implementation in the LSAS
in Fig. 3, multiple sensing beams of the proposed sens-
The proposed sensing codebook deign has two advantages ing codebooks can be simultaneously swept to reduce the
for practical implementation in the LSAS: sweeping overhead, and the corresponding RSs are mul-
1) Realizable in the analog phase-shifting arrays: To per- tiplexed by frequency or code divisions.
form the beam sweeping in the analog domain of the hy- 2) Small signaling/storage overhead owing to the structured
brid AD LSAS, the sensing beam should meet the constant nature: In the proposed sensing codebook design, P can
modulus and quantized phase constraints for realizing in be implicitly represented and reconstructed by (15). In
the phase-shifting arrays. In the proposed sensing code- other words, instead of delivering or storing the explicit
book design, the m-th transmit sensing beam of P in (15) sensing codebook, we only have to signal or store the
can be rewritten as sampling index set and the spreading sequence, as shown
in Fig. 3. For instance, the configuration of P compris-
pm = (NT )−1/ 2 [e−j (ω m , 1 +ϕ 1 ) , . . . , e−j (ω m , N T +ϕ N T ) T
] ,
ing M = 256 transmit sensing beams for the BS with
(18)
NT = 128 antennas can be signaled or stored via ΩT
where {ωm , }=1,...,N T and {ϕ }=1,...,N T are the phase-
using 256 × 7 bits and σ T using 128 bits (if σ T is a bipo-
shifting angles resulting from wT ,m and ΔT , respectively.
lar sequence), thereby saving 94% signaling or storage
From (13), since wT ,m is selected from FN T according to
burden from that of the Bernoulli random matrix (215 bits
ΩT ,m , we have ωm , = 2π( − 1)(ΩT ,m − 1)/NT . From
should be delivered or stored). Moreover, the spreading
(14), if σ T is a bipolar sequence (e.g., ML sequence) in
sequence can be also represented via an implicit approach.
the beam spreader ΔT , we have
 For example, the ZC sequence is generated only according
0, σT , = +1; to the root index τ , which can be signaled or stored using
ϕ = (19)
π, σT , = −1, log2 NT bits. As a result, compared to the full random
5102 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

Fig. 4. Illustration of applying the random convolution C T to a VAD channel with a limited scattering effect.

design, the signaling or storage overhead is reduced from where UG ∈ C N ×N is a transform matrix with the unitary and
O(M NT ) to O(M log2 NT ). block-circulant properties, as shown in the following lemma.
Lemma 1 (UG is a unitary block-circulant matrix): Let CT
C. Equivalent Measurement Process of the Structured = Circ(cT ) and CR = Circ(cR ). Suppose that CT and CR
Random Compressed Channel Sensing are both unitary matrices; then UG = CTT ⊗ CH R is a unitary
To explain why stable channel recovery from the CCS ap- block-circulant matrix, which can be expressed as
proach adopting the structured random sensing codebooks (15) ⎡ ⎤
CG ,1 CG ,N T · · · CG ,2
can be achieved, we present an equivalent measurement process ⎢C CG ,1 · · · CG ,3 ⎥
⎢ G ,2 ⎥
for our design. Substituting the VAD representation (2) into UG = ⎢ ⎢ .. .. .. .. ⎥⎥, (24)
(16), we have ⎣ . . . . ⎦
√  
y ≈ ρRΩ G (ΔT FN T )T ⊗ (ΔR FN R )H CG ,N T CG ,N T −1 · · · CG ,1

where CG , = cT , C∗R for  = 1, . . . , NT . The proof can be

× vec(AR H̄ω AH T) + e found in Appendix A.
√  
≈ ρRΩ G (ΔT FN T )T ⊗ (ΔR FN R )H According to (23), we can interpret ΦU as a random convolu-
tional measurement process comprised of random convolution
× (A∗T ⊗ AR )vec(H̄ω ) + e followed by random subsampling:
  r Random convolution: This step corresponds to multiply-
√ T H
≈ ρRΩ G (AH H
T ΔT FN T ) ⊗ (AR ΔR FN R ) h̄ω + e . ing the VAD channel vector h̄ω with UG , and we have
(21) h̃ω = UG h̄ω , (25)
For simplicity, we consider that GT = NT and GR = NR in
this measurement process, and (21) can be rewritten as where h̃ω ∈ C N ×1 denotes the transformed VAD channel
⎛ ⎞ vector. To explain intuitively the action of UG , we rewrite
(25) as
√ ⎜ T H⎟
y≈ ρRΩ G ⎝(FH H
N T ΔT FN T ) ⊗ (FN R ΔR FN R ) ⎠ h̄ω + e
  h̃ω = vec(CH
R H̄ω CT ). (26)
CT CR
√ We can see that each row vector in H̄ω is first convolved
≈ ρRΩ G (CTT ⊗ CH
R )h̄ω + e . (22)
with the sequence cT via CT . Due to the randomness in cT ,
From the fact that a circulant matrix can be diagonalized by the highly localized information of propagation paths in the
a unitary DFT matrix, CT = Circ(cT ) ∈ C N T ×N T and CR = limited scattering channel is spread over the entire VAD
Circ(cR ) ∈ C N R ×N R are both circulant matrices, where cT = spectrum along the transmit virtual angle grid after the
[cT ,1 , . . . , cT ,N T ]T ∈ C N T ×1 and cR = [cR ,1 , . . . , cR ,N R ]T ∈ convolution, as illustrated in Fig. 4. Each column vector in
C N R ×1 are given by cT = (NT )−1/2 FH N T σ T and cR =
the convolved VAD channel, i.e., H̄ω CT , is then convolved
(NR )−1/2 FH σ R , respectively. Because σ T and σ R are with the sequence cR via CR . In a similar fashion, the
NR
spreading sequences, after the IDFT, the randomness will re- information of the propagation paths will be spread over the
main in cT and cR , respectively. In addition, CT and CR are entire VAD spectrum along the receive virtual angle grid.
both unitary matrices because σ T and σ R are unimodular se- As a result, by applying the circular convolutions via CT
quences, i.e., |σT , | = 1 and |σR , | = 1 for  = 1, . . . , NT . and CR , no matter where AoDs and AoAs corresponding
Let ΦU ∈ C M ×N denote the equivalent measurement matrix to the propagation paths in the physical channel are, the
for (22), where N = NT NR since GT = NT and GR = NR . information is uniformly spread out; in other words, we can
Then, ΦU can be written as ensure that each measurement on the transformed VAD
channel vector h̃ω must contain an approximately equal
ΦU = RΩ G (CTT ⊗ CH
R ) = RΩ G UG , (23) amount of information from each propagation path in the

UG channel.
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5103

r Random subsampling: This step corresponds to multiply- TABLE II

COHERENCE OF C WITH DIFFERENT SPREADING SEQUENCES σ[48]
ing the transformed VAD channel vector h̃ω with RΩ G ,
which denotes a subsampling at locations indexed by ΩG .
Thus, the action of RΩ G picks up M measurements out
of N entries of h̃ω uniformly at random. Since the in-
formation is uniformly preserved in each entry of h̃ω , we
can obtain a stable recovery of h̄ω from those subsampled
measurements on h̃ω .
The random convolutional measurement process described
above is one class of compressed sensing tools [36]–[38], and
it has been proved that this process can achieve comparable
From (30), the coherence of UG can be decomposed by the
recovery performance to that of the full random measurement
product of the coherences of CT and CR . The coherences of
process. In our design, this process is further extended to a
the circulant matrices constructed from different unimodular
Kronecker system.
spreading sequences have been investigated in [48]. Table II
lists the three widely used spreading sequences, along with the
D. Theoretical Bound for Stable Channel Recovery corresponding N and μ(C), where C = FH N diag(σ)FN .
To derive the theoretical bound for stable recovery from the From Table II, it can be seen that when the ZC sequences are
random convolutional measurement process of ΦU = RΩ G UG , adopted for both σ T and σ R at the BS and the UE, UG has the
we adopt the RIP bound of the subsampled unitary matrix. coherence μ(UG ) = (NT NR )−1/2 = (N )−1/2 . Then, we have
Theorem 1 (RIP of the subsampled unitary matrix [46]): the RIP bound of ΦU
Suppose that an M × N measurement matrix Φ is a sub-
M ≥ O(δ −2 K log2 K log N ). (31)
sampled unitary matrix, i.e., Φ = RΩ U, where U is an
N × N unitary matrix, and RΩ is a subsampling operator for Compared to the optimal bound offered by the full random de-
downselecting M samples out of N ones uniformly at random. sign, there is an extra log2 K factor in the structured random
Then, Φ satisfies the (K, δ)− RIP for δ ∈ (1, 0) in a high design. As we know, owing to the limited scattering mmWave
probability if channels, the channel sparsity level is far smaller than the chan-
M ≥ O(δ −2 μ2 (U)N K log2 K log N ). (27) nel dimension, such that K  N . Thus, the extra factor log2 K
is almost negligible. Note that although the RIP bound (31) is
The parameter μ(U) is defined as follows. only derived for GT = NT and GR = NR , it will be shown
Definition 2 (coherence of the unitary matrix [47]): The in the simulation that the structured random sensing codebook
coherence of an N-dimensional unitary matrix U describes the design has comparable performance to that of the full random
maximum magnitude of the entry in U, i.e., design for arbitrary resolutions. Moreover, when the virtual an-
gle grids with high resolutions are adopted, i.e., GT > NT and
μ(U) = max |Up,q | . (28) GR > NR , the quantization error caused by the discrete channel
1≤p,q ≤N
approximation can be effectively reduced [45].
The coherence has values in the rage N −1/2 ≤ μ(U) ≤ 1.
Using Theorem 1, we can obtain the RIP bound of ΦU IV. STRUCTURED RANDOM SENSING CODEBOOK DESIGN
−2 WITH POWER CONCENTRATION CAPABILITY
M ≥ O(δ μ (UG )N K log K log N ).
2 2
(29)
In the sectorized cell [29]–[31], a LSAS should only measure
In (29), we can see that the required number of measurements
and estimate the channel within a local transmit angle cover-
for stable channel recovery is dictated by μ(UG ). According to
age. In this scenario, if the CCS adopts the sensing beams with
the properties of UG stated in Lemma 1, we have
full angle coverages, most of the power will be wasted in the
μ(UG ) = max μ(CG , ) directions out of the intended angle coverage. To address this
1≤≤N T
issue, based on the previously proposed dual-stage codebook
= max μ(cT , C∗R ) structure, we propose a transmit sensing codebook with the ca-
1≤≤N T
pability of power concentration. Note that if the receive sensing
 1/2
= max (cT , c∗R ,k )(cT , c∗R ,k )∗ codebook also needs to concentrate its power in a local angle
1≤≤N T coverage, it can be designed in a similar fashion.
1≤k ≤N R
 1/2
= max (cT ,n c∗T ,n )(cR ,m c∗R ,m ) A. Structured Random Sensing Codebook Design for Local
1≤≤N T
1≤k ≤N R Angle Coverage
! !
To concentrate the power in a local angle coverage, both
= max |cT ,n | max |cR ,n | the sampling beamformer and the beam spreader of the trans-
1≤≤N T 1≤k ≤N R
mit sensing beam should be specifically designed. As shown
= μ(CT )μ(CR ). (30) in Fig. 5, the m-th transmit sensing beam pm = ΔT wT ,m
5104 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

Fig. 6. Beam patterns of the structured random sensing beams formed by the
Fig. 5. Block diagram of the transmit structured random sensing codebook
ULA with 64 antennas when (a) S T = 1, (b) S T = 1/2, and (c) S T = 1/4.
design with power concentration for a local angle coverage at the BS.

in the structured random sensing codebook is constructed as illustrated in Fig. 6(a), which has a relatively low direc-
follows: tivity gain under a fixed transmit power. In contrast, when
r Sampling beamformer: In this stage, the sampling beam
ST < 1 (i.e., NC < NT ), the beamwidth of the resultant
wT ,m should be selected from a candidate DFT codebook sensing beam is narrower as ST decreases, and its power
rather than a full DFT codebook. Consider that the candi- is concentrated in the direction of corresponding sampling
date DFT codebook consisting of NDL F T DFT beams that beam to gather a larger directivity gain, as illustrated in
cover the local angle coverage, the indices of these DFT Fig. 6(b) and Fig. 6(c). However, when ST is too small,
beams is defined in a subsequence ΓT ⊂ {1, . . . , NT }, the beam spreader has limited expansion of beamwidth, the
|ΓT | = NDL F T < NT , which is also regarded as the in- resultant sensing beam cannot achieve an effective mea-
tended sampling region. Then, wT ,m can be selected from surement for stable recovery. Therefore, in Section IV-C,
FΓNTT through the sampling index ΩT ,m , as illustrated by we will discuss the design of ST for a local angle coverage.
the corresponding beam pattern (blue) in Fig. 5. Thus, the Consequently, we can configure the direction and the beam-
sampling beam can be expressed as width of a structured random sensing beam by respectively ad-
wT ,m = FΓNTT rTT ,m , (32) justing the direction of its sampling beam and the spreading
factor of its spreading sequence to ensure that the power can be
where rT ,m ∈ R1×N T is the sampling vector for selecting concentrated in the intended angle coverage.
the columns of FΓNTT indexed by ΩT ,m . To generate a transmit structured random sensing codebook
r Beam spreader: In this stage, differing from the de- consisting of M sensing beams for a local angle coverage,
sign in Section III-A, the beam spreader ΔT should these sensing beams are constructed by M sampling beams
spread the sampling beam wT ,m over a limited angle selected from the candidate DFT codebook FΓNTT and a same
coverage instead of a full angle coverage. Consider a beam spreader adopting the spreading sequence with ST < 1.
unimodular sequence with good autocorrelation property Thus, the transmit sensing codebook can be expressed as
α = [α1 , . . . , αN C ]T ∈ C N C ×1 . The length of α should
be a factor of NT , i.e., NT = DT NC , where DT is a posi- P = ΔT FΓNTT RTΩ T , (35)
tive integer. Then, the spreading sequence σ T in the beam
spreader ΔT is constructed by where the sampling index set ΩT for RΩ T is also derived from
(17). However, since the size of the transmit DFT codebook is
σT , = α (−1)/ D T +1 , (33)
reduced, the subsampling index set ΩG for RΩ G in (17) should
for  = 1, . . . , NT . Throughout this paper, we define the be downselected from {1, . . . , N̄ }, where N̄ = NDL F T NR <
spreading factor ST ∈ [NT−1 , 1] of σ T as N , uniformly at random instead of {1, . . . , N }.
NC
ST = . (34)
NT B. Equivalent Measurement Process of the Structured Random
Compressed Channel Sensing for Local Angle Coverage
Similar to the direct-sequence spread spectrum (DSSS),
By adopting the transmit sensing codebook (35) at the BS,
where the expansion of signal bandwidth after the spreader
we can rewrite (16) as
is varied with different spreading factors, the expansion of
beamwidth after the beam spreader ΔT can be config- √  T

ured by adjusting ST . When ST = 1 (i.e., NC = NT ), the y= ρRΩ G (ΔT FΓNTT ) ⊗ (ΔR FN R )H vec(H) + e .
resultant sensing beam exhibits a full angle coverage, as (36)
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5105

Since the AoDs must locate in an intended angle coverage,

we can substitute the VAD representation (4) into (36). Thus,
√  T

y ≈ ρRΩ G (ΔT FΓNTT ) ⊗ (ΔR FN R )H

× vec(AR H̄Θ T (AΘ

T ) )+e
T H

!
√ H T H
≈ ρRΩ G ((AΘ
T ) ΔT
T
FΓNTT ) ⊗ (AH
R ΔR FN R )

× vec(H̄Θ T ) + e
!
√ H T H
≈ ρRΩ G ((AΘ ΓT H
T ) ΔT FN T ) ⊗ (AR ΔR FN R )
T

× h̄Θ T + e, (37)
where h̄Θ T = vec(H̄Θ T ) ∈ C |Θ T |G R ×1 is the VAD channel
vector with the constraint on the AoD. For simplicity, we con-
sider that GT = NT and GR = NR . In this case, we have
ΘT = ΓT , and (37) can be rewritten as
⎛ ⎞ Fig. 7. Illustration of (a) the VAD channel with the constraint on the AoD,
and (b) applying the random convolution C T constructed by σT with S T < 1
√ ⎜ T ⎟
⎜ H H⎟ on the VAD channel with the constraint on the AoD.
y ≈ ρRΩ G ⎜((FΓ T
) Δ T F ΓT
) ⊗ (F H
NR Δ R F N ) ⎟h̄Γ T + e
⎝ NT NT

R
 ⎠
Γ
CT T
,ΓT CR

⎛ ⎞ is limited, stable recovery still can be guaranteed for the CCS

in the local angle coverage. This is because, in (39), we only
√ ⎜ Γ ,Γ T ⎟
≈ ρRΩ G ⎝(CT T T ) ⊗ CH
R ⎠ h̄Γ T + e have to ensure that each measurement on the convolved VAD

TG
channel within the intended sampling region, i.e., h̃Γ T , should
√ preserve an approximately equal amount of the information.
≈ ρRΩ G TG h̄Γ T + e , (38) As illustrated in Fig. 7(b), when the information is only spread
within the intended sampling region ΓT rather than the entire
where RΩ G ∈ RM ×N̄ denotes the subsampling operator for
VAD spectrum, which can be achieved by choosing an appro-
downselecting M samples out of N̄ according to ΩT , and TG ∈
priately small ST , channel recovery from the subsampled h̃Γ T
C N̄ ×N̄ denotes the transform matrix.
still can perform stable embedding. On the other hand, although
As mentioned previously, before the subsampling by RΩ G ,
choosing a small ST for a sensing codebook can more concen-
the transform by TG on the VAD channel is the key action to
trate the power to gather larger directivity gains, it should be
achieve stable channel recovery from (38). Let h̃Γ T ∈ C N̄ ×1
sufficiently large to ensure the effect of information expansion
denote the transformed h̄Γ T ; then it can be expressed as
in the intended sampling region.
h̃Γ T = TG h̄Γ T
Γ T ,Γ T
R H̄Γ T CT
= vec(CH ) C. Theoretical Bound for Stable Channel Recovery
 
= vec CH
ΓT According to the definition of TG = (CΓT T ,Γ T )T ⊗ CHR , it
R (H̄ω CT ) . (39)
can be rewritten as a square submatrix of UG , such that
After the transform by TG , the information of each prop-
agation path in h̄Γ T is spread over the VAD spectrum along  Λ T ,Λ T
TG = CTT ⊗ CH
R
the transmit and receive virtual angle grids as broadly as pos-
sible. We can see that the difference between the transforms = UΛGT ,Λ T , (40)
of (39) and (26) is that the convolved VAD channel H̄ω CT is
restricted within a limited sampling region ΓT on the transmit
where ΛT is a subsequence of {1, . . . , N̄ } with the length of
virtual angle grid rather than the entire VAD spectrum. This
NDL F T NR for collecting the block of the entries in UG whose
is because we have prior knowledge that the information must
row and column are indexed by
locate in the intended sampling region ΓT , as illustrated in
Fig. 7(a), and the search space on the VAD spectrum can be "
reduced. When CT is constructed by σ T with ST = 1, because ΛT = (ΓT ,1 − 1)NR + 1, (ΓT ,1 − 1)NR + 2, . . . , ΓT ,1 NR ,
the information is spread over the entire VAD spectrum by CT ,
stable recovery from the subsampled h̃Γ T can be reasonably . . . , (ΓT ,N DL F T − 1)NR + 1, (ΓT ,N DL F T − 1)
#
achieved. On the other hand, when CT is constructed by σ T
NR + 2, . . . , ΓT ,N DL F T NR . (41)
with ST < 1, even the effect of information expansion by CT
5106 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

Let ΦT ∈ C M ×N̄ denote the equivalent measurement matrix V. SIMULATION RESULTS

for (38). It can be expressed as We evaluate the performance of the proposed sensing code-
ΦT = RΩ G TG book design for the CCS through several simulations. In
our simulations, the system and channel models described in
= RΩ G UΛGT ,Λ T , (42) Section II are set as follows:
– The BS and UE are equipped with NT = 128 and NR = 32
Since TG is embedded into UG , we can derive an analogous antennas, respectively.
bound of ΦT from the RIP bound of ΦU such that – The number of clusters in the physical channel is set
M ≥ O(δ −2 μ2 (TG )N̄ K log2 K log N̄ ), (43) K = 6.
– The physical AoDs are uniformly distributed in a lo-
where cal AoD coverage, such that {cos(θkA oD )}k =1,...,K ∈
 [−1/3, 1/3].
μ(UG ), μ(UG ) > N̄ −1/2 ;
μ(TG ) = −1/2 (44) – The physical AoAs are uniformly distributed in a full AoA
(N̄ ) , μ(UG ) ≤ N̄ −1/2 . coverage, such that {cos(θkA oA )}k =1,...,K ∈ [−1, 1].
In (30), we know that μ(UG ) = μ(CT )μ(CR ). When CT – Channel recovery is performed using the orthogonal
is constructed by σ T with ST = 1, since μ(UG ) = N −1/2 is matching pursuit (OMP) algorithm adopted in [45].
always smaller than N̄ −1/2 , we have the RIP bound of ΦT The normalized mean square error (NMSE) is adopted as the
recovery performance metric, which is defined as
M ≥ O(δ −2 K log2 K log N̄ ). (45)  "  #
NMSE = 10 log E ||Ĥ − H||2F ||H||2F . (47)
On the other hand, when CT is constructed by σ T with
ST < 1, the ideal coherence of CT can be given in the following We compare the recovery performance of channel recovery
lemma. from CCS that adopts the following different codebooks:
Lemma 2 (ideal coherence of CT when ST < 1): Let CT 1) Exhaustive DFT codebook (EXH-DFT): Both BS and UE
= Circ(cT ), where cT = (NT )−1/2 FH N T σ T , and σ T is con- adopt the DFT codebooks to exhaustively measure the
structed by a sequence α with the length of NC < NT using channel. At the BS, FΓNTT with NDL F T = 44 is adopted to
(33). If α is a unimodular sequence with the zero autocorrelation cover the local AoD coverage. At the UE, FN R is adopted
property, then CT has the ideal coherence μ(CT ) = (NC )−1/2 . to cover the full AoA coverage.
The proof can be found in Appendix B. 2) Full random Sensing codebook (FR-CS): At the BS and
From Table II and Lemma 2, it can be seen that when both the UE, full random sensing codebooks are adopted, where
α (for constructing σ T ) and σ R are the ZC sequences, we the transmit and receive sensing beams are drawn from the
have μ(UG ) = μ(CT )μ(CR ) = (NC NR )−1/2 . Thus, the RIP i.i.d. Bernoulli bipolar random variables, respectively.
bound of ΦT can be derived 3) Structured random sensing codebook (SR-CS) with
 $ ST = 1: At the BS and the UE, the sensing codebooks
O(δ −2 (NDL F T NC )K log2 K log N̄ ), NDL F T > NC ;
M> (15) are adopted, where the spreading sequences are ZC
O(δ −2 K log2 K log N̄ ), NDL F T ≤ NC < NT . sequences with the lengths of 128 and 32, respectively.
(46) 4) SR-CS with ST = 1/4 : At the BS, the transmit sens-
We can see that when NDL F T ≤ NC < NT , i.e., 1 > ing codebook with power concentration (35) is adopted,
ST ≥ NDL F T /NT , ΦT provides a bound the same with (45). where the spreading sequence is constructed by the ZC se-
However, when NDL F T > NC , i.e., NDL F T /NT > ST , more quence with the length of 32 using (33), and the sampling
measurements are required to guarantee stable channel re- beams are selected from the DFT codebook FΓNTT with
covery. This theoretical result agrees with our observation in NDL F T = 44. At the UE, the receive sensing codebook is
Section IV-B. the same as 3).
According to (46), the design of spreading factor for a local 5) SR-CS with ST = 1/8 : At the BS, the transmit sens-
angle coverage can be derived. By choosing an appropriately ing codebook with power concentration (35) is adopted,
small spreading factor satisfying ST ≥ NDL F T /NT for the trans- where the spreading sequence is constructed by the ZC se-
mit sensing codebook (35), it can achieve the same bound with quence with the length of 16 using (33), and the sampling
that of the fully spread sensing codebook (i.e., ST = 1) while beams are selected from the DFT codebook FΓNTT with
concentrating the power to gather larger directivity gains. On NDL F T = 44. At the UE, the receive sensing codebook is
the other hand, the design of spreading factor is also dictated the same as 3).
by the SNR. As we know, the RIP bound (46) only holds when Each simulation result is obtained from Monte Carlo trials
the minimum requirement on the ratio of signal power after over 5,000 independent channel realizations.
measurement to noise power is satisfied [35]. Therefore, when
the SNR is not sufficiently high, to improve the robustness of
A. NMSE Under Various SNRs
channel recovery, we should choose a spreading factor smaller
than NDL F T /NT to more concentrate the power of the sensing Fig. 8 depicts the recovery performance under various SNRs
codebook. with M = 320, which is approximately 23% of that in the
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5107

Fig. 8. NMSE against SNR when M = 320. Two resolution settings Fig. 10. Achievable rate against number of measurements when N R F =
{G T , G R } = {128, 32} and {G T , G R } = {512, 128} are considered in this 4, SNR = 0 dB, and {G T , G R } = {512, 128}.
simulation.
angle grids with {GT , GR } = {512, 128} to well approximate
the physical channel. As the EXH-DFT has full measurements
from the transmit and receive DFT codebooks, its recovery per-
formance remains constant for different values of M.
We can observe that the SR-CS with ST = 1 has comparable
recovery performance to that of the FR-CS with any number of
measurements. However, both of them are far worse than the
EXH-DFT, especially in low SNR. To reduce the performance
gaps, we can adopt the SR-CS with power concentration. As we
can see, when M ≥ 200, the SR-CS with ST = 1/8 exhibits
more than 4 dB and 2 dB NMSE improvements over the SR-
CS with ST = 1 for SNR = 0 dB and 10 dB, respectively. It is
worth noting that the improvement from the power concentration
is more significant when the SNR is lower.
Fig. 9. NMSE against number of measurements when SNR ∈{0 dB, 10 dB},
and {G T , G R } = {512, 128}.
C. Achievable Rate Under Various Numbers of Measurements
EXH-DFT. We also consider two different resolution settings, Fig. 10 depicts the achievable rates given by the estimated
{GT , GR } = {128, 32}, and {GT , GR } = {512, 128}. channels, which is the final aim of channel estimation, un-
We can first observe that all the codebooks exhibit significant der various numbers of measurements when SNR = 0 dB and
NMSE improvements when the virtual angle grids with higher {GT , GR } = {512, 128}. In this simulation, the achievable rate
resolutions are adopted. This is because when the impact of is derived from the optimal unconstrained precoder [40] for the
quantization error is reduced, the recovery performance of the hybrid AD LSAS with a limited number of RF chains
CCS can be effectively enhanced. Next, we can observe that the !
ρ −1 H H
SR-CS with ST = 1 has comparable recovery performance to R = log2 det IN S + R Wopt HFopt Fopt HWopt ,
that of the FR-CS. However, both of them exhibit huge perfor- NR F n
(48)
mance gaps compared to the EXH-DFT, especially in low SNR.
where the unconstrained precoder Fopt ∈ C N T ×N R F and com-
This is because the power of the fully spread sensing beams are
biner Wopt ∈ C N R ×N R F are designed by selecting the NR F -
not sufficient for robust channel recovery in low-SNR regimes.
dominant singular vectors of the estimated channel, Rn =
Fortunately, we can adopt the SR-CS with power concentration H
σn2 Wopt Wopt denotes the noise covariance matrix after com-
to enhance the robustness. As we can see, when SNR = 0 dB,
bining, and NR F denotes the number of RF chains.
the SR-CS with ST = 1/8 exhibits approximately 2 dB and 4 dB
We can observe that when NR F = 4, the SR-CS with ST =
performance gains over the FR-CS and SR-CS with ST = 1 in
1/8 achieves 95% optimal rate by employing only 140 measure-
the two different resolution settings, respectively. The perfor-
ments, which saves approximate 90% sweeping overhead from
mance gap between the SR-CS with ST = 1/8 and the EXH-
the EXH-DFT with negligible rate losses. In contrast, the FR-
DFT is less than 2 dB in any SNR.
CS and the SR-CS with ST = 1 exhibit significant rate losses
from the EXH-DFT. Even more measurements are employed
B. NMSE Under Various Numbers of Measurements for them, their performance saturates without any improvement.
Fig. 9 illustrates the recovery performance under various This is because channel recovery has the minimum requirement
numbers of measurements when SN R ∈ {0 dB, 10 dB}. In this on the ratio of signal power after measurement to noise power,
simulation, we only consider the transmit and receive virtual which is scaled as N regardless of M [35]. When both SNR and
5108 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

Fig. 11. NMSE of the SR-CS with various S T when M = 140, SNR ∈ Fig. 12. NMSE against scale factor r for the resolutions of the virtual angle
{−5 dB, 0 dB, 10 dB, 20 dB, 30 dB}, and {G T , G R } = {512, 128}. grids when M = 320 and SNR ∈ {0 dB, 10 dB}.

directivity gains are not sufficiently high, robust channel recov- the error caused by quantizing the physical AoDs/AoAs in fixed
ery cannot be achieved, and more measurements cannot reduce virtual angles is reduced by increasing the resolutions of virtual
the performance gap. angle grids. On the other hand, with higher resolutions, we can
see that the SR-CSs with power concentration exhibit a larger
D. NMSE Under Various Spreading Factors performance gain over the sensing codebooks with full angle
coverages.
Fig. 11 depicts the recovery performance of the SR-CS
with various ST when M = 140, {GT , GR } = {512, 128}, and
VI. CONCLUSION
SNR ∈ {−5 dB, 0 dB, 10 dB, 20 dB, 30 dB}.
We can observe that when SNR = 30 dB, the SR-CS with In this paper, a new class of CCS for the mmWave LSAS was
ST = 1/2 has the same performance with that of the SR- presented. First, we proposed a structured random sensing code-
CS with ST = 1. This is because the ST = 1/2 satisfies book design. Owing to its structured nature, instead of signaling
ST ≥ NDL F T /NT = 44/128, which is sufficiently large to en- or storing the explicit sensing codebook, the configuration of
sure the effect of information expansion in the intended sam- the proposed sensing codebook can be implicitly represented to
pling region. However, the SR-CS with spreading factor smaller reduce the signaling or storage overhead. For the CCS adopting
than NDL F T /NT = 44/128 (i.e., ST = 1/4, 1/8, 1/16, 1/32) the proposed sensing codebook design, we offered an RIP bound
suffers performance degradation since the measurements from derived from the equivalent measurement process, and showed
the structured random CCS becomes ineffective for channel re- that the proposed structured random design compares favor-
covery. When SNR decreases, the performance of SR-CS with ably to the full random design. Second, based on the proposed
ST = 1 becomes worse than that of the SR-CS with ST < 1. dual-stage codebook structure, we further developed a sensing
Meanwhile, the spreading factor of the SR-CS achieving the codebook with the capability of power concertation. For CCS
lowest NMSE becomes smaller as the SNR decreases, even it in a local angle coverage, we concentrated the sensing beams in
is smaller than NDL F T /NT = 44/128. This because when the the local angle coverage to increase directivity gains. Thus, the
SNR is not sufficiently high, the larger directivity gains of the stringent requirement on the SNR for robust channel recovery
sensing codebook can increase the ratio of signal power after could be alleviated. Simulation results showed that the proposed
measurement to noise power to improve the performance. Thus, sensing codebook design has comparable performance to that of
in the low-SNR regimes, we need the SR-CS with spreading the full random design. In low-SNR regimes, the performance
factor smaller than NDL F T /NT = 44/128 to more concentrate was substantially enhanced by the proposed sensing codebook
the power of each sensing beam. with power concentration.

E. NMSE Under Various Resolutions APPENDIX A

PROOF OF LEMMA 1
Fig. 12 shows the recovery performance under various res-
olutions of the virtual angle grids when M = 320 and SNR ∈ According to the definition of UG in Lemma 1, we have
{0 dB, 10 dB}. In this simulation, we set the resolutions to be UG = CTT ⊗ CH
R
scaled with the factor r, such that GT = rNT and GR = rNR .
(a)
We can observe that the SR-CS with ST = 1 has comparable = CT ⊗ C∗R
recovery performance to that of the FR-CS when GT > NT ⎡ ⎤
cT ,1 C∗R cT ,N T C∗R · · · cT ,2 C∗R
and GR > NR even when we only offer the RIP bound for ⎢ cT ,2 C∗R cT ,1 C∗R · · · cT ,3 C∗R ⎥
GT = NT and GR = NR . Furthermore, as we expect, when the =⎢
⎣ .. .. .. .. ⎥ ⎦
resolutions increase with r, the recovery performance of each . . . .
codebook is effectively enhanced. It comes from the fact that cT ,N T CR cT ,N T −1 C∗R
∗
· · · cT ,1 CR∗
TSAI AND WU: STRUCTURED RANDOM COMPRESSED CHANNEL SENSING FOR MILLIMETER-WAVE LARGE-SCALE ANTENNA SYSTEMS 5109

⎡ ⎤ '% N &'
CG ,1 CG ,N T · · · CG ,2 '  C '
1 ' j N2 π (k −1)(m −1) '
⎢ CG ,2 CG ,1 · · · CG ,3 ⎥ = max ' αm e C '
NT 1≤k ≤N T ' m =1 '
=⎢
⎢ .. .. ..
⎥
.. ⎥ , (A1) '% D
⎣ . ⎦ &'
. . . '  T '
CG ,N T CG ,N T −1 · · · CG ,1 ' j 2 π (k −1)(n −1) '
×' e NT '
' '
n =1
where (a) follows from CT and CR are both circulant matrices, ' %N &'
' 1  C '
and CG , = cT , CR ∈ C N R ×N R for  = 1, . . . , NT . As we can 1 ' j N2 π (k −1)(m −1) '
= max ' αm e C '
see, UG is a block-circulant matrix. On the other hand, DT 1≤k ≤N T ' NC m =1 '
'% D &'
UG UH T H T H H '  '
G = (CT ⊗ CR )(CT ⊗ CR )
T
' j 2 π (k −1)(n −1) '
×' e NT '
=(CTT ⊗ CH ∗ ' '
R )(CT ⊗ CR ) n =1
'% D &'
'  T '
(a)
= (CTT C∗T ) ⊗ (CH
1 ' j N2 π (k −1)(n −1) '
R CR ) = max |α̃ | ' e T ', (B3)
DT 1≤≤N C ' '
1≤k ≤N T n =1
= (CH ∗ H
T CT ) ⊗ (CR CR )
where α̃ = [α̃1 , . . . , α̃N C ]T = (NC )−1/2 FH
N C α. As we know,
(b) ∗
= (IN T ) ⊗ (IN R ) = IN , (A2) when α is a unimodular sequence with the zero autocorrelation
property (e.g., the ZC sequence), we have |α̃ | = (NC )−1/2 for
where (a) follows from (A ⊗ B)(C ⊗ D) = (AC) ⊗ (BD),  = 1, . . . , NC . Therefore, we can obtain
and (b) follows from CT and CR are both unitary matrices. In
a similar fashion, we can also derive UHG UG = IN . Thus, since '% D &'
'  T '
UG UH = U H
U G = I N , UG is a unitary matrix. Combining μ(CT ) =
1 ' j 2π
e NT
(k −1)(n −1) '
G G max ' '
the results of (A1) and (A2), we have proved Lemma 1. (NC ) DT 1≤k ≤N T ' n =1
1/2 '

APPENDIX B = (NC )−1/2 . (B4)

PROOF OF LEMMA 2
Since the CT is a circulant matrix specified by the sequence
REFERENCES
cT , the coherence of CT can be written as
[1] F. Boccardi, R. W. Heath, Jr., A. Lozano, T. L. Marzetta, and P. Popovski,
μ(CT ) = max |cT ,k | . (B1) “Five disruptive technology directions for 5G,” IEEE Commun. Mag.,
1≤k ≤N T −1 vol. 52, no. 2, pp. 74–80, Feb. 2014.
[2] L. Wei, R. Q. Hu, Y. Qian, and G. Wu, “Key elements to enable millimeter
According to the definition of cT in Lemma 2, the k-th entry wave communications for 5G wireless systems,” IEEE Wireless Commun.,
of cT can be rewritten as vol. 21, no. 6, pp. 136–143, Dec. 2014.
[3] W. Roh et al., “Millimeter-wave beamforming as an enabling technology
NT
1  j 2 π (k −1)(−1)
for 5G cellular communications: Theoretical feasibility and prototype
cT ,k = σT , e N T results,” IEEE Commun. Mag., vol. 52, no. 2, pp. 106–113, Feb. 2014.
NT [4] X. Zhang, A. F. Molisch, and S. Kung, “Variable-phase-shift-based RF-
=1
%N baseband codesign for MIMO antenna selection,” IEEE Trans. Signal
1  C
j 2 π (k −1)(m −1)
Process., vol. 53, no. 11, pp. 4091–4103, Nov. 2005.
= αm e N C [5] A. Alkhateeb, J. Mo, N. González-Prelcic, and R. W. Heath, Jr., “MIMO
NT m =1 precoding and combining solutions for millimeter-wave systems,” IEEE
Commun. Mag., vol. 52, no. 12, pp. 122–131, Dec. 2014.
NC
 [6] S. Han, I. Chih-Lin, Z. Xu, and C. Rowell, “Large-scale antenna systems
j N2 π (k −1)(m −1)+j N2 π k
+ αm e C T ... with hybrid analog and digital beamforming for millimeter wave 5G,”
m =1 IEEE Commun. Mag., vol. 53, no. 1, pp. 186–194, Jan. 2015.
NC
& [7] M. R. Akdeniz et al., “Millimeter wave channel modeling and cellular
 j N2 π (k −1)(m −1)+j N2 π (D T −1)k
capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1164–
+ αm e C T 1179, Jun. 2014.
m =1 [8] A. Alkhateeb, O. El Ayach, G. Leus, and R. Heath, Jr., “Channel estimation
%N &%D & and hybrid precoding for millimeter wave cellular systems,” IEEE J. Sel.
1  C
j N2 π (k −1)(m −1)
 T
j N2 π (k −1)(n −1) Topics Signal Process., vol. 8, no. 5, pp. 831–846, Oct. 2014.
= αm e C e T . [9] S. Noh, M. D. Zoltowski, and D. J. Love, “Multi-resolution codebook
NT m =1 n =1 based beamforming sequence design in millimeter-wave systems,” in Proc.
(B2) IEEE Global Commun. Conf., Dec. 2015, pp. 1–6.
[10] Z. Xiao, P. Xia, and X. G. Xia, “Codebook design for millimeter-wave
channel estimation with hybrid precoding structure,” IEEE Trans. Wireless
Substituting (B2) into (B1), we have Commun., vol. 16, no. 1, pp. 141–153, Jan. 2017.
'% N &
'  C
[11] D. E. Berraki, S. M. D. Armour, and A. R. Nix, “Application of com-
1 ' j N2 π (k −1)(m −1) pressive sensing in sparse spatial channel recovery for beamforming in
μ(CT ) = max ' α e C
NT 1≤k ≤N T ' m =1 mmWave outdoor systems,” in Proc. IEEE Wireless Commun. Netw. Conf.,
Apr. 2014, pp. 887–892.
%D &'
 T ' [12] A. Alkhateeb, G. Leus, and R. W. Heath, “Compressed sensing based
j N2 π (k −1)(n −1) ' multi-user millimeter wave systems: How many measurements are
× e T '
' needed?” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process.,
n =1 Apr. 2015, pp. 2909–2913.
5110 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 66, NO. 19, OCTOBER 1, 2018

[13] R. Mendez-Rial, C. Rusu, N. Gonzalez-Prelcic, A. Alkhateeb, and R. W. [38] H. Rauhut, J. Romberg, and J. Tropp, “Restricted isometries for partial
Heath, Jr., “Hybrid MIMO architectures for millimeter wave communi- random circulant matrices,” Appl. Comput. Harmon. Anal., vol. 32, no. 2,
cations: Phase shifters or switches?” IEEE Access, vol. 4, pp. 247–267, pp. 242–254, Mar. 2012.
2016. [39] V. Raghavan and A. M. Sayeed, “Sublinear capacity scaling laws for sparse
[14] Z. Marzi, D. Ramasamy, and U. Madhow, “Compressive channel esti- MIMO channels,” IEEE Trans. Inf. Theory, vol. 57, no. 1, pp. 345–364,
mation and tracking for large arrays in mmWave picocells,” IEEE J. Sel. Jan. 2011.
Topics Signal Process., vol. 10, no. 3, pp. 514–527, Apr. 2016. [40] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, Jr.,
[15] Z. Zhou, J. Fang, L. Yang, H. Li, Z. Chen and R. S. Blum, “Low-rank ten- “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
sor decomposition-aided channel estimation for millimeter wave MIMO- Trans. Wireless Commun., vol. 13, no.3, pp. 1499–1513, Mar. 2014.
OFDM systems,” IEEE J. Sel. Areas Commun., vol. 35, no. 7, pp. 1524– [41] J. A. Tropp and A. C. Gilbert, “Signal recovery from random measure-
1538, Jul. 2017. ments via orthogonal matching pursuit,” IEEE Trans. Inf. Theory, vol. 53,
[16] K. Venugopal, A. Alkhateeb, N. González-Prelcic, and R. W. Heath, Jr., no. 12, pp. 4655–4666, Dec. 2007.
“Channel estimation for hybrid architecture based wideband millimeter [42] D. Needell and J. A. Tropp, “CoSaMP: Iterative signal recovery from in-
wave systems,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1996– complete and inaccurate samples,” Appl. Comput. Harmon. Anal., vol. 26,
2009, Sep. 2017. no. 3, pp. 301–321, Mar. 2009.
[17] Y. Wang, Z. Tian, S. Feng, and P. Zhang, “Efficient channel statistics [43] W. Dai and O. Milenkovic, “Subspace pursuit for compressive sensing
estimation for millimeter-wave MIMO systems,” in Proc. IEEE Int. Conf. signal reconstruction,” IEEE Trans. Inf. Theory, vol. 55, no. 5, pp. 2230–
Acoust., Speech Signal Process., Mar. 2016, pp. 3411–3415. 2249, May 2009.
[18] S. Park and R. W. Heath, Jr., “Spatial channel covariance estimation for [44] E. Candès and T. Tao, “Near-optimal signal recovery from random pro-
mmWave hybrid MIMO architecture,” in Proc. Asilomar Conf. Signals, jections: Universal encoding strategies,” IEEE Trans. Inf. Theory, vol. 52,
Syst. Comput., Nov. 2016, pp. 1424–1428. no. 12, pp. 5406–5425, Dec. 2006.
[19] A. Liu, V. K. N. Lau, and W. Dai, “Exploiting burst-sparsity in massive [45] J. Lee, G. T. Gil, and Y. H. Lee, “Channel Estimation via Orthogo-
MIMO with partial channel support information,” IEEE Trans. Wireless nal Matching Pursuit for Hybrid MIMO Systems in Millimeter Wave
Comm., vol. 15, no. 11, pp. 7820–7830, Nov. 2016. Communications,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2370–2386,
[20] C. R. Tsai, Y. H. Liu, and A. Y. Wu, “Efficient compressive channel Jun. 2016.
estimation for millimeter-wave large-scale antenna systems,” IEEE Trans. [46] I. Haviv and O. Regev, “The restricted isometry property of subsampled
Signal Process., vol. 66, no. 9, pp. 2414–2428, May 2018. fourier matrices,” in Proc. 27th Annu. ACM-SIAM Symp. Discrete Algo-
[21] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Ex- rithms, Jan. 2016, pp. 288–297.
act signal reconstruction from highly incomplete frequency information,” [47] E. Candès and J. Romberg, “Sparsity and incoherence in compressive
IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006. sampling,” Inverse Probl., vol. 23, no. 3, pp. 969–985, 2007.
[22] D. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 2, [48] K. Li, L. Gan, and C. Ling, “Convolutional compressed sensing using
pp. 1289–1306, Apr. 2006. deterministic sequences,” IEEE Trans. Signal Process., vol. 61, no. 3,
[23] E. Candes and M. B. Wakin, “An introduction to compressive sampling,” pp. 740–752, Feb. 2013.
IEEE Signal Process. Mag., vol. 25, no. 2, pp. 21–30, Mar. 2008. [49] M. A. Khalighi, K. Raoof, and G. Jourdain, “Capacity of wireless com-
[24] L. Applebaum, S. D. Howard, S. Searle, and R. Calderbank, “Chirp sensing munication systems employing antenna arrays, a tutorial study,” Wireless
codes: Deterministic compressed sensing measurements for fast recovery,” Pers. Commun., vol. 23, no. 3, pp. 321–352, Dec. 2002.
Appl. Comput. Harmon. Anal., vol. 26, no. 2, pp. 283–290, Mar. 2009.
[25] R. Calderbank, S. Howard, and S. Jafarpour, “Construction of a large
class of deterministic sensing matrices that satisfy a statistical isometry
property,” IEEE Trans. Inf. Theory, vol. 4, no. 2, pp. 358–374, Apr. 2010. Cheng-Rung Tsai (S’15) received the B.S. degree
[26] A. Amini and F. Marvasti, “Deterministic construction of binary, bipo- in electrical engineering from Chang Gung Univer-
lar, and ternary compressed sensing matrices,” IEEE Trans. Inf. Theory, sity, Taoyuan City, Taiwan, in 2012, and the Ph.D.
vol. 57, no. 4, pp. 2360–2370, Apr. 2011. degree from the Graduate Institute of Electronics En-
[27] S. Li, F. Gao, G. Ge, and S. Zhang, “Deterministic construction of com- gineering, National Taiwan University, Taipei City,
pressed sensing matrices via algebraic curves,” IEEE Trans. Inf. Theory, Taiwan, in 2018. He is currently a Senior Engineer
vol. 58, no. 8, pp. 5035–5041, Aug. 2012. with MediaTek, Inc., Hsinchu, Taiwan. His research
[28] S. Li and G. Ge, “Deterministic sensing matrices arising from near orthog- interests include the area of signal processing for
onal systems,” IEEE Trans. Inf. Theory, vol. 60, no. 4, pp. 2291–2302, wireless communication, which includes MIMO sys-
Apr. 2014. tems, massive antenna technologies, and mmWave
[29] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband communications.
systems,” IEEE Commun. Mag., vol. 49, no. 6, pp. 101–107, 2011.
[30] P. Chen et al., “A multibeam antenna based on substrate integrated wave-
guide technology for MIMO wireless communications,” IEEE Trans. An- An-Yeu Wu (M’96–SM’12–F’15) received the B.S.
tenna Propag., vol. 57, no. 6, pp. 1813–1821, Jun. 2009. degree from National Taiwan University (NTU),
[31] G. Xu et al., “Full dimension MIMO (FD-MIMO): Demonstrating com- Taipei City, Taiwan, in 1987, and the M.S. and Ph.D.
mercial feasibility,” IEEE J. Sel. Areas Commun., vol. 35, no. 8, pp. 1876– degrees from the University of Maryland, College
1886, Aug. 2017. Park, College Park, MD,USA, in 1992 and 1995, re-
[32] M. Ettorre, R. Sauleau, and L. Le Coq, “Multi-beam multi-layer leaky- spectively, all in electrical engineering. In August
wave SIW pillbox antenna for millimeter-wave applications,” IEEE Trans. 2000, he was with the faculty of the Department of
Antennas Propag., vol. 59, no. 4, pp. 1093–1100, Apr. 2011. Electrical Engineering and the Graduate Institute of
[33] M. Jiang, Z. N. Chen, Y. Zhang, W. Hong, and X. Xuan, “Metamaterial Electronics Engineering, NTU, where he is currently
based thin planar lens antenna for spatial beamforming and multibeam a Distinguished Professor. His research interests in-
massive MIMO,” IEEE Trans. Antennas Propag., vol. 65, no. 2, pp. 464– clude VLSI architectures for signal processing and
472, Feb. 2017. communications and adaptive/multirate signal processing. He has published
[34] W. Hong et al., “Multibeam antenna technologies for 5G wireless commu- more than 190 refereed journal and conference papers in above research areas,
nications,” IEEE Trans. Antennas Propag., vol. 65, no. 12, pp. 6231–6249, together with five book chapters and 16 granted US patents. From August 2007
Dec. 2017. to December 2009, he was on leave from NTU and served as the Deputy General
[35] S. Aeron, V. Saligrama, and M. Zhao, “Information theoretic bounds for Director of SoC Technology Center, Industrial Technology Research Institute,
compressed sensing,” IEEE Trans. Inf. Theory, vol. 56, no. 10, pp. 5111– Hsinchu, Taiwan. In 2010, he received “Outstanding EE Professor Award” from
5130, Oct. 2010. The Chinese Institute of Electrical Engineering, Taiwan. From 2012 to 2014,
[36] J. Romberg, “Compressive sensing by random convolution,” SIAM J. Imag. he served as the Chair of VLSI Systems and Applications Technical Committee
Sci., vol. 2, no. 4, pp. 1098–1128, Dec. 2009. (TC), one of the largest TCs in IEEE Circuits and Systems Society. In 2015, he
[37] H. Rauhut, “Compressive sensing and structured random matrices,” Theor. was elevated to the IEEE Fellow for his contributions to DSP algorithms and
Found. Numer. Methods Sparse Recovery. Boston, MA, USA: De Gruyter, VLSI designs for communication IC/SoC. Since 2016, he serves as Director of
vol. 9, pp. 1–92, 2010. Graduate Institute of Electronics Engineering, National Taiwan University.