IEEE COMMUNICATIONS LETTERS, VOL. 20, NO.
6, JUNE 2016
Channel Estimation for Millimeter-Wave Massive MIMO With Hybrid
Precoding Over Frequency-Selective Fading Channels
Zhen Gao, Chen Hu, Linglong Dai, and Zhaocheng Wang
Abstract— This letter proposes a multi-user uplink channel
estimation scheme for mmWave massive MIMO over frequency
selective fading (FSF) channels. Specifically, by exploiting the
angle-domain structured sparsity of mmWave FSF channels, a
distributed compressive sensing-based channel estimation scheme
is proposed. Moreover, by using the grid matching pursuit strat-
egy with adaptive measurement matrix, the proposed algorithm
can solve the power leakage problem caused by the continuous
angles of arrival or departure. Simulation results verify the good
performance of the proposed solution.
Index Terms— Millimeter-wave massive MIMO, frequency-
selective fading, channel estimation, compressive sensing.
I. I NTRODUCTION
M ILLIMETER-WAVE (mmWave) massive MIMO has
been considered as a promising key technique for
5G [1]. For mmWave massive MIMO with the phase shifter
network based hybrid precoding [1]–[3] or electromagnetic
lens based beamspace MIMO [4], [5], the number of radio
frequency (RF) chains is usually much smaller than that of
antennas for reduced hardware cost and power consumption.
However, such architectures will lead to the challenging chan-
nel estimation [6].
To date, several channel estimation schemes [6]–[8] have
been proposed for mmWave massive MIMO with hybrid pre-
coding. Specifically, the reference signal was designed in [6]
for the estimation of angles of arrival or departure (AoA/AoD),
but it assumes the discrete AoA/AoD. [8] can directly estimate
the singular subspace with Krylov subspace method, but mul-
tiple amplify-and-forward operations between the transmitter
and receiver will introduce much noise. More importantly,
existing schemes [6]–[8] only consider the narrow-band flat
fading channel model, while the practical mmWave channels
exhibit the broad-band frequency-selective fading (FSF) due to
the very large bandwidth and different delays of multipath [9].
To this end, this letter proposes a multi-user uplink channel
estimation scheme for mmWave massive MIMO systems,
where the broad-band FSF channel is converted to multiple
parallel narrow-band flat fading channels when OFDM is con-
sidered. Specifically, the mmWave channels exhibit the obvi-
ously angle-domain sparsity due to the much higher path loss
Manuscript received March 20, 2016; accepted April 11, 2016. Date
of publication April 20, 2016; date of current version June 8, 2016.
This work was supported by the Beijing Natural Science Foundation
(Grant No. 4142027), the National Natural Science Foundation of China
(Grant Nos. 61271266 and 61571270), the International Science & Technology
Cooperation Program of China (Grant No. 2015DFG12760), and the Foun-
dation of Shenzhen government. The associate editor coordinating the review
of this letter and approving it for publication was K. J. Kim. (Corresponding
author: Linglong Dai.)
The authors are with the Tsinghua National Laboratory for Information
Science and Technology, Department of Electronic Engineering, Tsinghua
University, Beijing 100084, China (e-mail: gaozhen010375@foxmail.com;
daill@tsinghua.edu.cn; zcwang@tsinghua.edu.cn).
Digital Object Identifier 10.1109/LCOMM.2016.2555299
1558-2558 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1259
Fig. 1. Illustration of a broad-band mmWave massive MIMO system.
for non-line-of-sight (NLOS) paths than that for line-of-sight
(LOS) paths [7]. Moreover, this sparsity is almost unchanged
within the system bandwidth according to our derivation. By
exploiting such angle-domain structured sparsity of mmWave
FSF channels, we propose a distributed compressive sensing
(DCS)-based channel estimation scheme, where both the trans-
mit pilot signal and receive channel estimation algorithm are
elaborated under the DCS theory for improved performance.
Moreover, by using the grid matching pursuit strategy with
adaptive measurement matrix, the proposed algorithm can
solve the power leakage problem caused by the continuous
AoA/AoD. To the best of our knowledge, this is the first paper
to investigate the FSF channel estimation for mmWave massive
MIMO systems.
Notation: The boldface lower and upper-case symbols
denote column vectors and matrices, respectively. The
Moore-Penrose inversion, transpose, conjugate transpose, inte-
ger ceiling, and expectation operators are given by (·)† , (·)T ,
(·)∗ , ·, and E{·}, respectively. || is the cardinality of the
set . The support set of the vector a is denoted by supp{a}.
⊗ is the Kronecker product, and vect (·) is the vectorization
operation according to the columns of the matrix. [a]i denotes
the i th entry of the vector a, and [A]i, j denotes the i th-row
and j th-column element of the matrix A.
II. S YSTEM M ODEL
We consider a typical mmWave massive MIMO-OFDM sys-
tem over FSF channels as shown in Fig. 1 [9], where the base
station (BS) employs NaBS antennas but only NRF BS RF chains
with Na  NRF = K to support K user equipments (UEs),
BS BS
and each UE has NaUE antennas but only NRF UE RF chain with
Na  NRF = 1. The hybrid analog-digital precoding at the
UE UE
BS can be used to realize the spatial multiplexing with low
hardware cost and energy consumption [9]. Particularly, the
uplink FSF channel associated with the kth user in the delay
domain can be modeled as [9]
 L k −1  
Hkd (τ ) = d
Hl,k δ τ − τl,k , (1)
l=0
where L k is the number of multipath, τl,k is the delay of the
∈ C Na ×Na is given by
d BS UE
lth path, Hl,k
d
  ∗  
Hl,k = αl,k aBS d sin(θl,k )/λ aUE d sin(ϕl,k )/λ , (2)
1260 IEEE COMMUNICATIONS LETTERS, VOL. 20, NO. 6, JUNE 2016

αl,k is the complex gain of the lth path, and θl,k ∈ [0, 2π] indicates small L k and large K factor in mmWave systems, e.g.,
and ϕl,k ∈ [0, 2π] are azimuth AoA/AoD if we consider L k = 4 and K factor = 20 dB [2]. Hence, we can transform the
f
the typical uniform linear array (ULA). For path gains, we frequency-domain channel matrix H p,k in (5) into the sparse
consider Rician fading channels consisting of one LOS path a
angle-domain channel matrix H p,k as [7]
(the 0th path) and L k − 1 NLOS paths (the lth path for
Hap,k = A∗BS H p,k AUE ,
f
1 ≤ l ≤ L k − 1), where path gains follow the mutually (6)
independent complex Gaussian distribution with zero means,
where ABS ∈ C Na ×Na and AUE ∈ C Na ×Na are the
BS BS UE UE
and K factor denotes the ratio between the power of LOS path
and the power of NLOS paths. In addition, DFT matrices by quantizing the virtual angular domain with
   T the resolutions of 1/NaBS at the BS and 1/NaUE at the user,
d sin(θl,k ) f
aBS = e j 2πnBS d sin(θl,k )/λ respectively. By vectorizing H p,k , we can further obtain
 λ   n BS ∈[0,1,··· ,NaBS −1]
d sin(ϕl,k ) T f f
h p,k = vect H p,k = Ahap,k , (7)
aUE = e j 2πnUE d sin(ϕl,k )/λ
λ n UE ∈[0,1,··· ,NaUE −1]  
(3) where A = (A∗UE )T ⊗ ABS and hap,k = vect Hap,k . Due to the
sparsity of Hap,k , only a minority of elements of hap,k dominate
are steering vectors at the BS and the kth user, respectively,
the majority of the channel energy, and thus we have
where λ is wavelength and d is the antenna spacing.
p,k = supp hap,k = Sk NaBS NaUE , (8)
III. DCS-BASED C HANNEL E STIMATION S CHEME
A. Uplink Pilot Training where p,k is the support set, and Sk is the sparsity level in
the angular domain.
We consider that the training sequences used for
According to (6)-(8), (4) can be further expressed as
channel estimation adopt OFDM to combat the FSF K (t )
channels, where lengths of cyclic prefix (CP) and r(tp ) = (Z(tp ) )∗ ABS Hap,k A∗UE f p,k + v(tp )
discrete Fourier transform (DFT) are L CP > (max k=1
L k −1,K
{τl,k }l=0,k=1 L k −1,K
− min{τl,k }l=0,k=1 ) f s and P > L CP , respectively, = (Z(tp ) )∗ ABS H̄ap Ā∗UE f̄ (t ) (t )
p + vp
 
T
where f s is the sampling rate. At the BS, after the CP ∗ (t ) (t ) ∗
= ĀUE f̄ p ⊗ (Z p ) ABS vect H̄ap + v(tp )
removal and DFT operation, the received signal at the
pth (1 ≤ p ≤ P) subcarrier of the tth OFDM symbol in the = (t ) a
p h̄ p + v(tp ), (9)
frequency domain can be expressed as
K where
 
r(tp ) = (Z(t ) (t ) ∗
H p,k F(t ) (t ) (t ) (t )
f
RF ZBB, p ) RF,k FBB, p,k s p,k + v p , (4) H̄ap = Hap,1 , Hap,2 , · · · , Hap,K ∈ C Na ×K Na ,
BS UE
k=1
(t )  
where r p ∈ C NRF ×1 is the received signal dedicated to the pth
BS
Ā∗UE = diag A∗UE , A∗UE , · · · , A∗UE ∈ C K Na ×K Na ,
UE UE

pilot subcarrier in the tth OFDM symbol, Z(t ) BS ×N BS

BB, p ∈ C
NRF
f̄ (tp ) = [(f (t ) T (t ) T (t ) T T
RF UE ×1
p,1 ) , (f p,2 ) , · · · , (f p,K ) ] ∈ C ,
K Na
(t ) NaBS ×NRF
BS
is the digital combining matrix, ZRF ∈ C is the K NaBS NaUE ×1
h̄ap = vect H̄ap ∈C ,
RF combining matrix, Z(tp ) = Z(t ) (t ) Na ×NRF is the
BS BS
RF ZBB, p ∈ C T
(t )
= Ā∗UE f̄ (t )
⊗ (Z(tp ) )∗ ABS ∈ C NRF ×K Na
BS BS N UE
composite combining matrix at the BS, p p
a . (10)
 L k −1
f d − j 2π f s τl,k p/ P
H p,k = Hl,k e Furthermore, by jointly using the received pilot signals in
 L k −1
l=0     G successive OFDM symbols, we can obtain
j 2π f s τl,k p d sin(θl,k ) ∗ d sin(ϕl,k )
= αl,k e− P aBS aUE , ˜ p h̄ap + ṽ p ,
r̃ p =  (11)
l=0 λ λ
(5) (1) (2) (G)
where r̃ p = [(r p )T , (r p )T , · · · , (r p )T ]T ∈ CG NRF ×1 is the
BS

denotes the frequency-domain channel matrix associated with aggregate received signal,  ˜ p = [( (1) (2) T
p ) ,( p ) ,··· ,
T
(t )
the pth pilot subcarrier for the kth UE, FRF,k ∈ C Na ×NRF ,
UE UE
(G)
( p )T ]T ∈ CG NRF ×K Na Na is the aggregate measurement
BS BS UE
(t ) (t )
FBB, p,k ∈ C NRF ×NRF , s p,k ∈ C NRF ×1 are the RF pre-
UE UE UE
matrix, and ṽ p = [(v(1) (2) T (G) T T
p ) , (v p ) , · · · , (v p ) ] is aggregate
T
coding matrix, digital precoding matrix, and transmitted AWGN.
(t )
training sequence for the kth UE, respectively, f p,k =
(t ) (t ) (t ) UE ×1
FRF,k FBB, p,k s p,k ∈ C Na is considered as the pilot sig- B. DCS-Based Channel Estimation
nal transmitted by the kth user, and v(tp )
is the additive To accurately estimate channels from (11), G in conven-
white Gaussian noise (AWGN) at the BS. Note that RF tional algorithms, is heavily dependent on the dimension
BS ≥ K N UE N BS is
of h̄ap , i.e., K NaUE NaBS . Usually, G NRF
precoding/combining is the same for all subcarriers, since the a a
RF phase shifter network can provide constant phase shift required, which leads G to be much larger than the channel
response over a wide frequency range [9]. coherence time [7]. Fortunately, the sparsity of mmWave
Since the path loss for NLOS paths is much larger than that massive MIMO channels motivates us to leverage the CS
for LOS paths in mmWave systems, the mmWave channels theory to estimate channels with much reduced pilot overhead.
f
appear the obvious sparsity in the angular domain, which Moreover, according to (5), it can be observed that {H p,k } Pp=1
GAO et al.: CHANNEL ESTIMATION FOR MMWAVE MASSIVE MIMO WITH HYBRID PRECODING OVER FSF CHANNELS 1261

Algorithm 1 Proposed DGMP Algorithm outer loop and inner loop. In each iteration of outer loop
Input: Received signals r̃ p and sensing matrices ˜ p in (11) ∀ p, (steps 2.1-2.3 and 2.19-2.21), according to correlation oper-
AoA/AoD resolution factor J , and error threshold ε. ation (step 2.1), the UE index k̃ (step 2.2) and adaptive
Output: The steering vector estimation of kth UE’s LOS path measurement matrix ϒ̄ p (step 2.3) associated with the most
k,LOS
âBS k,LOS
and âUE , and the estimation set of path gains α̂ ∈ possible path are acquired and input to inner loop; according
  to the output of inner loop, the k̃th UE’s transmit/receive
C1×K , where α̂ k denotes the gain estimate of kth UE’s LOS
steering vectors are acquired (steps 2.19-2.20), and |K| UEs’
path.
• Step 1 (Initialization) The residue b p = r̃ p , the iteration index LOS path gains and residue b p are updated (step 2.21).
    
   The iteration of outer loop stops when AoA/AoD and path
k = 1, ˜p
 ˜p
=  / 
˜p  for 1 ≤ j ≤ K NaUE NaBS , ∀ p,
:, j :, j  :, j 2 gains of all K UEs’ LOS paths are estimated. For inner loop
and the matrix p and set K are set to be empty. (steps 2.4-2.18), the AoA/AoD estimation associated with
• Step 2 (Estimate steering vectors and gains of K UEs’ LOS paths) the k̃th UE’s LOS path is improved with the grid matching
for k ≤ K do  
 P 
 
∗  2   strategy. Specifically, according to the inputs k̃ and ϒ̄ p from
1. ρ = arg max  
˜ p bp  ,  ρ /(N UE N BS ) ∈/ K ;
p=1 
ρ
  a a 
ρ 2 outer loop, AoA/AoD indices n BS and n UE of the most
  possible path are acquired (step 2.6), and the corresponding
2. k̃ = ρ/(NaUE NaBS ) , K = K ∪ k̃ ;
  correlation value is recorded as β (step 2.5); we construct the
˜p
3. ϒ̄ p = 
:,(k̃−1)NaBS NaUE +1:k̃ NaBS NaUE
; local over-complete measurement matrix ϒ̃ p (steps 2.7-2.11),
repeat   2  where the local resolution of AoA associated with the
P  ∗  
4. ρ = arg max p=1 
 ϒ̄ p b
p 
ρ
 ; index n BS and AoD associated with the index n UE is
ρ

increased by (2 J −1) times; according to correlation operation
2
 
P  ∗  2
5. βlast = β , β =  
p=1  ϒ̄ p b p ρ  ; (step 2.12), finer AoA/AoD indices m BS and m UE can be
2
  acquired (step 2.13); finally, ϒ̄ p is updated, where the grid of
6. n UE = ρ/NaBS , n BS = ρ − (n UE − 1)NaBS ;

j
 AoA/AoD candidates is adjusted according to m BS and m UE
7. ÃUE = aUE (n UE + 2UE J )/Na
UE ;
  jUE ∈[− J,− J +1,··· ,J ] (step 2.14-2.18). The inner loop stops if |βlast − β| < ε.
8. ÃBS = aBS (n BS + 2BS
j
J )/N a
BS ; With the joint process of  ˜ p and b p for 1 ≤ p ≤ P,
jBS ∈[− J,− J +1,··· ,J ]
 T
9. (t) ∗
ϒ̃ p = ÃUE f
(t) (t) ∗
⊗ (Z p ) ÃBS ;
the DGMP algorithm exploits the structured sparsity for
p,k̃ improved performance, which can be found in steps 2.1,
(1) (2) (G)
10. ϒ̃ p = [(ϒ̃ p )T , (ϒ̃ p  )T , · · · ,(ϒ̃ p )T ]T ; 2.4, and 2.12. Moreover, the adaptive measurement matrix
      
11. ϒ̃ p = ϒ̃ p /  2
 ϒ̃ p :, j  , 1 ≤ j ≤ (2J − 1) , ∀ p; ϒ̄ p with grid matching pursuit strategy can achieve high
:, j  :, j  2  
P  
∗
 2 resolution estimation of AoA/AoD. Additionally, the near-
η = arg max   ;
12.

η p=1  (ϒ̃ p ) b p  η LOS mmWave channel property is exploited, where only
2

13. m UE = η/(2J − 1), m BS = η − (m UE − 1)(2J − 1);

K UEs’ LOS paths are estimated. Compared to the adaptive
   CS algorithm [7] estimating single sparse narrow-band channel
UE −1
14. ÃUE = aUE (n UE + − J +m )/NaUE  ;
 
2J UE
 n ∈ 0,1,··· ,NUE −1
from single received signal, the proposed DGMP algorithm
BS
15. ÃBS = aBS (n BS + − J +m 2J
−1 )/N BS
a  ;
jointly estimates multiple sparse subchannels from multiple
n BS ∈ 0,1,··· ,NBS −1
 T received signals. Moreover, the grid matching pursuit strategy
(t) (t) (t)
16. ϒ p = Ã∗UE f ⊗ (Z p )∗ ÃBS ; (steps 2.4-2.18) with adaptive ϒ̄ p can solve the problem of
p,k̃
(1) (2) (G) power leakage caused by the continuous AoA/AoD, which is
17. ϒ p = [(ϒ p )T , (ϒ p  )T , · · · ,(ϒ p )T ]T ;
      
18. ϒ̄ p = ϒp / ϒp
 , 1 ≤ j ≤ N UE N BS , ∀p ; different from the classical DCS algorithms [10].
:, j :, j :, j 
2
a a

until |βlast − β| < ε C. Pilot Design According to DCS Theory

19. k̃,LOS
âBS = aBS ((n BS + −J +m
BS −1
)/NaBS ); The measurement matrices  ˜ p , ∀ p in (11) are very
2J
important for guaranteeing the reliable channel estimation.
((n UE + −J +m
k̃,LOS UE −1
20. âUE = aUE )/NaUE ); ˜p =
   K NaUE NaBS . Since 
2J Usually, we have G NRF BS
†
21. p = p , ϒ p :,η , α̂K = ( p ) r̃ p , b p = r̃ p − αK p; (1) T (2) T (G) T T (t ) (t )
[( p ) , ( p ) , · · · , ( p ) ] , p = (Ā∗UE f̄ p )T ⊗
end for  
(Z(tp ) )∗ ABS , Ā∗UE = diag A∗UE , A∗UE , · · · , A∗UE , and AUE ,
share the same AoA/AoD, and thus {hap,k } Pp=1 obtained after ABS are determined by the geometrical structure of the antenna
(t ) ,G
(6) and (7) have the structured sparsity, i.e., arrays, both {f p,k } P,K
p=1,k=1,t =1 transmitted by the K users and
 a   a    {Z(tp )} P,G
supp h1,k = supp h2,k = · · · = supp haP,k = k . (12) p=1,t =1 at the BS should be elaborated to guarantee the
desired robust channel estimation.
Specifically, given (11) and the sparse constraints of According to [10], a measurement matrix whose elements
(8) and (12), the channels can be estimated with standard DCS follow an independent identically distributed (i.i.d.) Gaussian
tool. However, due to the continuous AoA/AoD and the limited distribution can achieve the good performance for sparse sig-
angle-domain resolution of ABS and AUE , the sparsity of h̄ap nal recovery. Furthermore, diversifying measurement matrices
may be impaired due to the power leakage problem [7], which ˜ p , ∀ p can further improve the recovery performance of
will result in the poor channel estimation performance. sparse signals according to DCS theory [10]. This enlight-
To this end, we propose a distributed grid matching pur- ens us to appropriately design pilot signals for mmWave
suit (DGMP) algorithm as listed in Algorithm 1 including massive MIMO systems. Specifically, as discussed above,
1262
Fig. 2. Comparison of spectral efficiency performance of different channel
estimation schemes against the training overhead G and SNR.
(t ) (t ) (t ) (t ) (t )
Z(tp ) = Z(t ) (t )
RF ZBB, p , f p,k = FRF,k FBB, p,k s p,k = FRF,k s̃ p,k if
(t ) (t ) (t )
we define s̃ p,k = FBB, p,k s p,k (1 ≤ k ≤ K , 1 ≤ t ≤ G,
1 ≤ p ≤ P). Hence, we propose that each element of
 ) j φi1 , j ,t
pilot signals is given by Z(t RF i1 , j1 = e
1 1 , 1 ≤ i1 ≤

BS , F(t )
 j φ 2
NaBS , 1 ≤ j1 ≤ NRF RF,k i2 , j2 = e
i2 , j2 ,t,k , 1 ≤ i
 (t )  j φi4 , j , p,t
Na , 1 ≤ j2 ≤ NRF , ZBB, p i , j = e 4 4 , 1 ≤ i 4 ≤
UE UE
 (t )  4 4 j φ 3
NRF , 1 ≤ j4 ≤ NRF , s̃ p,k i = e i3 , p,t,k , 1 ≤ i 3 ≤ NRF
BS BS
3
where φi11 , j1 ,t , φi22 , j2 ,t,k , φi33 , p,t,k , and φi44 , j4 , p,t follow the
i.i.d. uniform distribution U [0, 2π). Note that elements of
RF precoding/combining matrices should meet the constant
modulus property, and different subcarriers share the same
RF precoding/combining. It is readily seen that the designed
pilot signals guarantee that the elements of  ˜ p obey the i.i.d.
complex Gaussian distribution with zero mean. Moreover, 
with different p are diversified. Hence, the proposed pilot
signal design is optimal in terms of the joint recovery of multi-
user’s sparse angle-domain channels in the uplink.
IV. S IMULATION R ESULTS
In simulations, carrier frequency f c = 30 GHz, f s =
0.25 GHz, the maximum delay spread τmax = 100 ns, L CP =
τmax f s = 25, P = 32, NaUE = 32, NRF UE = 1, N BS = 128,
a Oct. 2015.
NRF = 4, d = λ/2, K factor = 20 dB, J = 10, ε = 10−3 ,
BS
K = 4, L k = 4 for 1 ≤ k ≤ K . The case with the
ideal AoD/AoD known at the BS is used as the performance
benchmark for comparison. The adaptive CS-based channel
estimation scheme [7] is also adopted for comparison.
Fig. 2 investigates the downlink spectral efficiency by using
the hybrid analog-digital precoding scheme in [3], where the
channels were estimated by the adaptive CS scheme [7] and
the proposed DGMP algorithm. The case with ideal AoA/AoD
was adopted as the performance bound. From Fig. 2, it can
be observed that the adaptive CS scheme performs poorly,
since it does not exploit the structured sparsity of mmWave
massive MIMO channels. In contrast, the proposed DGMP
algorithm can approach the performance bound with ideal
AoA/AoD when G ≥ 20. This is because the proposed
DCS-based channel estimation scheme can leverage the angle-
domain structured sparsity of mmWave FSF channels within
the system bandwidth.
Fig. 3 compares the downlink bit error rate (BER) per-
formance, where 16-QAM is used, and G for adaptive CS
IEEE COMMUNICATIONS LETTERS, VOL. 20, NO. 6, JUNE 2016
Fig. 3. BER performance comparison of different schemes.
algorithm and DGMP algorithm are 40, and 30, respectively.
(t )
It can be observed that the proposed channel estimation
scheme outperforms its counterpart with reduced training
overhead, and its BER performance is very close to the
performance bound with ideal AoA/AoD.
V. C ONCLUSIONS
In this letter, we have proposed a DCS-based uplink channel
2 ≤ estimation scheme for the multi-user mmWave massive MIMO
with hybrid precoding, and it can effectively combat mmWave
FSF channels. Specifically, we have designed an efficient
UE ,
pilot scheme and proposed a reliable DGMP algorithm under
the framework of DCS theory, whereby the angle-domain
structured sparsity of mmWave FSF channels is exploited for
the reduced training overhead. Moreover, by using the grid
matching pursuit strategy with adaptive measurement matrix,
the proposed algorithm can effectively solve the power leakage
problem. Our scheme can accurately estimate the FSF channels
˜p in mmWave massive MIMO with much lower pilot overhead
than the existing scheme.
R EFERENCES
[1] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath, Jr., “Energy-efficient
hybrid analog and digital precoding for mmWave MIMO systems with
large antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4,
Apr. 2016.
[2] Z. Gao, L. Dai, D. Mi, Z. Wang, M. A. Imran, and M. Z. Shakir,
“mmWave massive-MIMO-based wireless backhaul for the 5G ultra-
dense network,” IEEE Wireless Commun., vol. 22, no. 5, pp. 13–21,
[3] A. Alkhateeb, G. Leus, and R. W. Heath, Jr., “Limited feedback hybrid
precoding for multi-user millimeter wave systems,” IEEE Trans. Wireless
Commun., vol. 14, no. 11, pp. 6481–6494, Nov. 2015.
[4] J. Brady, N. Behdad, and A. M. Sayeed, “Beamspace MIMO for
millimeter-wave communications: System architecture, modeling, analy-
sis, and measurements,” IEEE Trans. Antennas Propag., vol. 61, no. 7,
pp. 3814–3827, Jul. 2013.
[5] Y. Zeng and R. Zhang, “Millimeter wave MIMO with lens antenna array:
A new path division multiplexing paradigm,” IEEE Trans. Commun.,
vol. 64, no. 4, pp. 1557–1571, Apr. 2016.
[6] S. Han, C.-L. I, Z. Xu, and S. Wang, “Reference signals design for
hybrid analog and digital beamforming,” IEEE Commun. Lett., vol. 18,
no. 7, pp. 1191–1193, Jul. 2014.
[7] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., “Channel
estimation and hybrid precoding for millimeter wave cellular systems,”
IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846,
Oct. 2014.
[8] H. Ghauch, M. Bengtsson, T. Kim, and M. Skoglund, “Subspace
estimation and decomposition for hybrid analog-digital millimetre-
wave MIMO systems,” in Proc. IEEE 16th SPWAC, Jun./Jul. 2015,
pp. 395–399.
[9] A. Alkhateeb and R. W. Heath, Jr., “Frequency selective hybrid pre-
coding for limited feedback millimeter wave systems,” IEEE Trans.
Commun., to be published.
[10] Y. C. Eldar and G. Kutyniok, Eds., Compressed Sensing: Theory and
Applications. Cambridge, U.K.: Cambridge Univ. Press, May 2012.