2274 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO.
4, APRIL 2017
Hybrid Architectures With Few-Bit ADC Receivers:
Achievable Rates and Energy-Rate Tradeoffs
Jianhua Mo, Member, IEEE, Ahmed Alkhateeb, Member, IEEE, Shadi Abu-Surra, Member, IEEE,
and Robert W. Heath, Jr., Fellow, IEEE
Abstract Hybrid analog/digital architectures and receivers
with low-resolution analog-to-digital converters (ADCs) are two
low power solutions for wireless systems with large antenna
arrays, such as millimeter wave and massive multiple-input
multiple-output systems. Most prior work represents two extreme
cases in which either a small number of radio frequency (RF)
chains with full-resolution ADCs, or low-resolution ADC with
a number of RF chains equal to the number of antennas is
assumed. In this paper, a generalized hybrid architecture with
a small number of RF chains and a finite number of ADC bits
is proposed. For this architecture, achievable rates with channel
inversion and singular value decomposition-based transmission
methods are derived. Results show that the achievable rate is
comparable to that obtained by full-precision ADC receivers at
low and medium SNRs. A trade-off between the achievable rate
and power consumption for the different numbers of bits and
RF chains is devised. This enables us to draw some conclusions
on the number of ADC bits needed to maximize the system
energy efficiency. Numerical simulations show that coarse ADC
quantization is optimal under various system configurations.
This means that hybrid combining with coarse quantization
achieves better energy-rate trade-off compared with both hybrid
combining with full-resolutions ADCs and 1-bit ADC combining.
Index Terms Massive MIMO, millimeter wave commu-
nication, analog-to-digital converters, hybrid analog/digital
architecture.
I. I NTRODUCTION
M ASSIVE multiple-input multiple-output (MIMO) is
a key feature of next-generation wireless systems.
At low-frequencies, massive MIMO supports many users
simultaneously and achieves large sum-rates with relatively
simple multi-user processing [2][4]. At mmWave frequen-
cies, the large antenna arrays, deployed at both the base
station and mobile users, guarantee sufficient received signal
power [5][8]. This allows signal transmission with ultra high
data rates thanks to large bandwidths available at the mmWave
Manuscript received May 1, 2016; revised November 2, 2016; accepted
January 15, 2017. Date of publication March 13, 2017; date of current version
April 7, 2017. This work was supported in part by the National Science
Foundation under Grant NSF-CCF-1319556 and Grant NSF-CCF-1527079.
The work of J. Mo was partly supported by Samsung Research America-
Dallas. This paper was presented at the 20th International ITG Workshop
on Smart Antennas, Munich, Germany, March 2016 [1]. The associate
editor coordinating the review of this paper and approving it for publication
was M. Uysal.
J. Mo, A. Alkhateeb, and R. W. Heath, Jr. are with The University
of Texas at Austin, Austin, TX 78712 USA (e-mail: jhmo@utexas.edu;
aalkhateeb@utexas.edu; rheath@utexas.edu).
S. Abu-Surra is with PHAZR Inc., Allen, TX, 75002 USA (e-mail:
sc.shadi@gmail.com).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TWC.2017.2661749
1536-1276 2017 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.
frequency band [3], [9]. Unfortunately, the high hardware cost
and power consumption of mixed-signal components makes
a fully-digital transmission solution, that allocates an RF
chain per antenna, difficult to realize in practice [10], [11].
To overcome this challenge, new architectures that relax the
requirement of associating an RF chain per antenna are being
developed [7]. Hybrid analog/digital architectures [12], [13],
and 1-bit ADC receivers [14] are two potential solutions.
Those two solutions, though, represent two extreme cases in
terms of the number of bits and RF chains. In this paper,
we explore a generalization of these two architectures, where
finite resolution ADCs are used with hybrid combining.
A. Related Work
Hybrid analog/digital architectures divide the precod-
ing/combining processing between analog and digital domains.
They have been proposed for both mmWave and low-
frequency massive MIMO systems [12], [15][23]. Hybrid
architectures employ many fewer radio frequency (RF)
chains than the number of antennas, relying on RF beam-
forming that is normally implemented using networks of
phase shifters [12], [15], [16]. Hybrid precoding for diver-
sity and multiplexing gain was investigated in [15], and
for interference management in [16], considering general
MIMO systems. These solutions, however, did not make use
of the special large MIMO characteristics in the design.
For mmWave massive MIMO systems, the sparse nature
of the channels was exploited to design low-complexity
hybrid precoding algorithms [12], assuming perfect channel
knowledge at the transmitter. Extensions to the case when
only partial channel knowledge is required was considered
in [17] and [18]. Hybrid precoding algorithms that do not
rely on channel sparsity were proposed in [19] and [20],
with the aim of maximizing the system spectral efficiency.
Hybrid precoding was also shown to achieve performance near
that of the fully-digital solutions in low-frequency massive
MIMO systems when the number of RF chains is large enough
compared to the number of users [21][23]. A common
limitation of the hybrid architectures adopted in [12], [15],
and [17][23] is the assumption that the receive RF chains
include high-resolution analog-to-digital converters (ADCs),
which consume high power, especially at mmWave [10].
Another limitation is the extra power consumption of the ana-
log phase shifters, which can have high impact on the energy
efficiency of the hybrid combiner. Compared to the conven-
tional fully-digital receiver, the power saved by reducing the
number of RF chains in hybrid receiver may be offset by the
additional power consumed by the phase shifters.
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS
An alternative to high resolution ADCs is to live with ultra
low resolution ADCs (1-4 bits), which reduces power since
ADC power grows exponentially with resolution [24], [25].
In [14] and [26][41], receiver architectures where the
received signal at each antenna is directly quantized
by low resolution ADCs without any analog combining is
considered. At present, the exact capacity of quantized MIMO
channel is unknown, except for special cases like
the multiple-input single-output (MISO) and single-input
multiple-output (SIMO) channels in the low or high SNR
regime [14], [26], [36]. Transmitting independent QAM
signals [26] or Gaussian signals [28], [33], [34] from each
antenna nearly achieves the capacity at low SNR, but is not
optimal at high SNR. The case with CSIT was studied in
our previous work [14] where two methods were proposed to
design the input constellation and precoder to maximize the
channel capacity. It was shown that the proposed methods
achieve much larger rate than QAM signaling, especially at
high SNR. There is also interest in using 1-bit ADCs for the
massive MIMO receiver where a large number of ADCs are
needed [35], [37][39]. The achievable rate of the multiuser
uplink massive MIMO channel with 1-bit ADCs was analyzed
in [38] and [39]. Symbol detection algorithms in a similar
setup were proposed in [35] and [37]. The architecture
in [14] and [26][41], though, assume that the number of RF
chains is equal to the number of antennas, which means that
the hardware cost may be high, and no gain is made from
possible beamforming processing in the RF domain.
B. Contribution
In this paper, we propose a generalized hybrid architecture
with few-bit ADC receivers and draw important conclusions
about its energy-rate trade-off. The hybrid architecture and
1-bit ADC receiver architecture studied in the past represent
two extreme points in terms of the number of ADC bits and
RF chains. In prior work, the hybrid architecture employs a
small number of RF chains but with high resolution ADCs,
while the 1-bit ADC receivers assume that the number of RF
chains equals the number of antennas. The contributions of
this paper are summarized as follows.
For the transceiver architecture with hybrid precod-
ing/combining and low resolution ADCs, we propose
two transmission methods and derive their achievable
rates in closed forms. For the channel inversion based
method, the inter-stream interference is canceled before
quantization and hence the channel can be separated
into several parallel SISO channels. For the SVD based
method, the additive quantization noise model is used to
derive a lower bound of the achievable rate by assuming
Gaussian input distribution. We also derive an upper
bound of the channel capacity for one-bit quantization.
The bound is achieved by the proposed two transmission
methods under certain conditions. In simulations, we
show that the proposed architecture with few-bit ADCs
can achieve a performance comparable to that obtained
with fully-digital or hybrid architecture with infinite-bit
ADC receiver in the low-to-medium SNR range, which
is of a special importance for mmWave communications.
2275
Fig. 1. MIMO system with hybrid precoding and few-bit ADCs. The
t (N r ) RF chains. the
transmitter (receiver) has Nt (Nr ) antennas and NRF
RF
transmitter has full-precision DACs while the receiver has only few-bit low
resolution ADCs.
We characterize the trade-off between the achievable rate
and power consumption in the proposed hybrid architec-
ture with few-bit ADC receivers. This allows us to make
important conclusions about the energy efficiency of the
considered hybrid architecture for different numbers of
bits. This also enables us to explore the performance
of the considered architecture compared with the fully-
digital transceiver and the conventional hybrid architec-
ture with full-resolution ADCs. Using numerical results
and adopting a power consumption model from recent
research [10], [42], we draw insights into the optimal
number of quantization bits from an energy efficiency
perspective. A key finding is that coarse quantization
(4-5 bits) normally achieves the maximum energy effi-
ciency. The reason is that very low quantization (1-2 bits)
suffers from a severe rate loss, while high quantization
(7-8 bits) has high power consumption. Hence, both of
the two regimes result in very low energy efficiency.
In conclusion, this paper draws a complete picture about the
generalized hybrid architectures with few-bit ADC receivers
by analyzing both their achievable spectral efficiency and
their energy-rate trade-off. After the submission of our orig-
inal work, there were some recent contributions from other
researchers [43], [44] which also consider different combining
methods with low resolution ADCs. They also found that
medium precision ADC achieves the highest energy efficiency
and their results are similar to ours.
Notation: a is a scalar, a is a vector and A is a matrix.
tr(A), A and A F represents the trace, conjugate transpose
and Frobenius norm of a matrix A, respectively. I stands for
an identity matrix. I (a; b) represents the mutual information
between a and b.  (a) is the phase of the complex number a.
II. S YSTEM M ODEL
We propose a MIMO architecture that combines hybrid
analog/digital precoding and combining with few-bit ADCs,
as shown in Fig. 1. The transmitter and receiver are equipped
with Nt and Nr antennas, respectively. The transmitter is
assumed to have NRF t RF chains with full-precision digital-
to-analog converters (DACs), while the receiver employs
r
NRF RF chains with few-bit (1-4 bits) ADCs. Further, the
numbers
 t of antennas and RF chains are assumed to satisfy
NRF Nt , NRFr N . The transmitter and receiver commu-
r  t 
nicate via Ns data streams, with Ns min NRF , NRF
r .
Compared to the fully-digital architecture where the receiver
has Nr pairs of high resolution ADCs, the proposed receiver
2276 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 4, APRIL 2017

r pairs of few-bit ADCs, which

architecture contains only NRF III. P ROBLEM F ORMULATION
greatly reduces both the hardware cost and power consump- Since analog precoding (combining) is implemented by
tion. Note that the transmitter has high-resolution digital-to- analog phase shifters, the entries of FRF (WRF ) are limited
analog converters (DACs) in our model. Analyzing a hybrid to have same norm. The optimization problem is to maximize
transceiver architecture with both low resolution ADCs and the mutual information between s and r as follows.
DACs is left for future work.
In this paper, we denote FRF C Nt NRF as the frequency
t
P1 : max I (s; r|H) (5)
FBB , FRF , WRF ,
band analog precoder and FBB C NRF Ns as the baseband
t
p(s), Q( )
 
digital precoder. Assuming a narrowband channel and perfect   1
s.t. [FRF ]mn  = , m, n, (6)
synchronization, the complex baseband signal prior to com- Nt
bining can be written as  
  1
[WRF ]mn  = , m, n, (7)
Nr
y = HFRF FBB s + n, (1)
FRF FBB  F = Ns ,
2
(8)
where s is the digital baseband signal with the covari- where (8) is due

to the transmission power constraint, i.e.,
ance E[ss ] = NPts I where Pt is the transmission power, E FRF FBB s2 = Pt .
n CN (0, N2 I) is the white Gaussian noise with variance N2 . It is very non-trivial to solve the problem P1. First, it is
After the analog combining, quantization and digital com- hard to optimize the mutual information over so many para-
bining, the received signal is meters simultaneously. Second, the quantization function Q( )
  is nonlinear and also related to the input distribution p(s) and

v = WBB Q WRF HFRF FBB s + WRF n , (2) renders it difficult to analyze the mutual information. Third,
the equality constraints in (6) and (7) is non-convex and hard
where WBB C NRF NRF is the baseband combiner, WRF
r r
to deal with.
C Nr NRF is the analog combiner, and Q( ) is a scalar quanti-
r
Throughout the paper, the analog precoder FRF is assumed
zation function which applies component-wise and separately to satisfy FRF FRF = I. Under this assumption, the coupled
to the real and imaginary parts. power constraint (8) involving the digital and analog precoding
Since this paper focuses on capacity analysis and the become a simple constraint on the digital precoder FBB . The
choice of baseband combiner WBB does not affect the channel similar assumption also appeared in [19] where the digital and
capacity as long as WBB is invertible, we ignore the baseband analog precoders are designed separately. In addition, we also
W
combiner in this paper. Further, we assume perfect channel assume that WRF RF = I and therefore the effective noise
knowledge at the transmitter and receiver. Developing efficient is still white Gaussian noise. This assumption simplifies the
channel estimation techniques for the proposed transceiver computation of mutual information and a similar idea appeared
architecture is an interesting problem for future work. These in [47]. Further, FRF FRF and WRF W
RF are close to identity
techniques may leverage the previously designed channel matrices when Nt and Nr is large.
estimation algorithms for hybrid architectures [18], [45] and To sum up, we assume both the analog precoder and
MIMO systems with low-resolution ADCs [31]. We denote combiner are semi-unitary matrices. Consequently, the opti-
the signal without digital combining as mization problem P1 is reformulated as

 P2 : max I (s; r|H) (9)
r = Q WRF HFRF FBB s + WRF n , (3) FBB , FRF , WRF ,
p(s), Q( )
 
where the effective noise  n has covariance W W .
n  WRF   1
RF RF s.t. [FRF ]mn  = , m, n, (10)
With known CSI at the transmitter, the capacity of this Nt
 
channel is   1
[WRF ]mn  = , m, n, (11)
Nr
C = I (s; r|H)
max FRF FRF = I, WRF WRF = I, (12)
FBB , FRF , WRF ,
p(s), Q( )
  FBB  F Ns .
2
(13)
Pr(r|s; H)
= max p(s)Pr(r|s; H) log2 ds In this paper, we develop two transmission strategies,
FBB , FRF , WRF ,
p(s), Q( )
s r Pr(r)
including the precoding techniques, the distribution of
(4) signal s, and the quantization design. We will investigate their
achievable rates and show that their performances are close to
where p(s) represents the probability distribution of s, optimum in certain cases.
Pr(r|s; H) is the transition
 probability between s and r
given H, and Pr(r) = s p(s)Pr(r|s; H) ds. Note that the
IV. U PPER B OUND OF THE ACHIEVABLE R ATE
maximization is also over the quantization function Q( ), for
example, thresholds of the ADCs [36], [46]. If the simple In this section, we provide upper bounds of the achievable
uniform quantization is assumed, then the stepsize  is the rate for one-bit quantization. The upper bounds are used as
only parameter in quantization function Q( ). benchmarks for evaluating our proposed transmission methods.
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS 2277

For multi-bit quantization, the upper bounds are unknown and A. Channel Inversion Based Precoding Algorithm
left for future work. For digital precoding design, we propose to use channel
Proposition 1: An upper bound on the achievable rate with inversion precoding assuming that NRFt NRF
r = Ns . The
hybrid precoding and one-bit quantization is digital precoder is


r 2
Ns  1
R 1bit,ub = 2NRF 1 Hb Q r
1
, (14) FBB = G GG (18)
NRF
where Hb (x) = x log2 x (1 x) log2 (1 x) is the where
binary entropy function, Q() is tail probability of the standard   2 
normal distribution,  P2t is the SNR and 1 is the maximum = tr G GG GFRF FRF (19)
N
singular value of the effective channel matrix G  WRF HF . such that the power constraint (8) is satisfied. As it is assumed
RF
Proof: Please see the appendix. that FRF FRF = I, is simplified to be
The upper bound is achieved when the effective channel G is  1 
r identical singular values, or equivalently,
full rank and has NRF = tr GG . (20)

GG = 1 I. 2
Since there is no interference among streams because of
At low SNR, this upper bound is approximated as
channel inversion precoding, each stream of data can be
2 12 detected separately. The received signal is
R 1bit,ub = + o(), (15)
ln 2 

r = Q WRF HFRF FBB s + WRF n (21)
following the facts Q(t) = 12 1 t + o(t 2 ) and Hb ( 12 + t) =  
2 Ns
1 ln22 t 2 + o(t 2 ). Therefore the bound increases linearly with
=Q   s + WRF n . (22)
the power at low SNR. But at high SNR, the upper bound tr (GG )1
converges to 2NRF r bps/Hz, which is due to the finite number
The channel is converted to 2Ns parallel sub channels, each
of quantization output bits. of which is a quantized real-valued single-input single-output
Note that the upper bound given in (14) is related to the (SISO) channel. The SNR of each sub-channel is given by
choice of analog precoding WRF and FRF . Next, we give
another bound, which is looser but independent of the analog SNRCI =  , (23)
precoding. tr (GG)1
Corollary 1: An upper bound of the achievable rate with where  P2t .
hybrid precoding and one-bit quantization is N
Maximizing the SNR is equivalent to maximizing the
12 following term
 1bit,ub
= 2NRF 1 Hb Q
r ,  
R r
NRF
(16) 1 1
(G)  tr GG (24)
where 1 is the maximum singular value of H.  1
W 1 1 1
Proof: Under the constraint WRF RF = I and = + 2 + + 2 , (25)

FRF FRF = I, it is proved that 1 12 in [48, Th. 2.2].
2 1
2 2 Ns
Therefore, we have where 1 , 2 , , Ns are the singular values of the effective

2 channel G in descending order. Therefore, WRF and FRF
r
R 1bit,ub 2NRF 1 Hb Q r
1 1bit,ub .
R should be chosen to maximize the harmonic mean of the
NRF
squared singular values of G, or equivalently the harmonic
(17) mean of the eigenvalues of GG .
This completes the proof of Corollary 1. To maximize (G), the optimal choice of WRF and FRF are
Remark 1: For a channel with infinite-bit the singular vectors associated with the largest Ns singular
 ADCs, a simple values of H [49]. Although such choice satisfies the semi-
12
2 1+
r log
upper bound of the capacity is NRF , which
r
NRF unitary constraints (12), the norm constraints (10)(11) are
is achieved when the effective channel G has same singular violated.
values. Compared to (14) and (16), the bound for infinite-bit In this paper, we use alternating projection algorithm [50] to
ADCs increases to infinity as the power increases to infinity. find an approximate solution satisfying both the constant-norm
and semi-unitary constraints. The algorithm is summarized
V. ACHIEVABLE R ATE W ITH C HANNEL I NVERSION in Algorithm 1. In steps 2a)-2b), the semi-unitary matrices
BASED T RANSMISSION  RF and 
W FRF are projected to the sets of matrices satisfy-
In this section, we propose channel inversion based trans- ing the norm constraints (10)-(11), resulting in the solutions
mission. In this method, there is no interference among data  RF and 
W FRF respectively. Each element of W  RF  FRF has
streams at the receiver and each stream is quantized separately. the same phase of the corresponding element in W  RF  FRF
Therefore, the exact achievable rate of this method can be but normalized amplitude. In steps 2c)-2d), W  RF and  FRF
found in closed-form. are projected back to the sets of semi-unitary matrices. The
2278 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 4, APRIL 2017

Algorithm 1 Alternating Projection Algorithm for Analog semi-unitary constraints of analog precoder are both satisfied.
Precoding Design However, searching the best combination of columns has
 (0) = U and  higher complexity than the alternating projection method when
1) Initialize W RF F(0)
RF = V where H = U V is

the number of antennas is large.
the singular value decomposition of H. Set k = 1 and

= 105 .
2) Alternating B. Rate Analysis With One-Bit Quantization
 projection method    
 (k)  (k1) In this subsection, we focus on the special case of one-bit
RF mn = Nr exp j , m, n,
a) W 1 W RF
     mn quantization and derive the gap between the achievable rate
(k) (k1)
b)  FRF = 1N exp j  FRF , m, n, and the upper bound given in Proposition 1.
mn t mn
   12 If one-bit ADCs are used at the receiver, the capacity of each
c) W  (k) = W  (k) W  (k) W  (k) , real-valued SISO sub-channel is achieved by binary antipodal
RF RF RF RF
   12 signaling and is given by [36]
d) F(k)
RF = FRF
(k)  F(k)
RF F(k)

RF ,   
 
  (k)  (k) 
W W 
1 Hb Q SNRCI . (26)
e) If the normalized distance
RF
r
RF F
<
and
  NRF The total sum rate therefore is
(k) (k) 
FRF FRF     
F  (k) and 
<
, return W F(k) 1bit
= 2Ns 1 Hb Q .
t
NRF RF RF ; RCI SNRCI (27)
else, k = k + 1 and go back to step (a). Next, noting that
 1
1 1 1 N2 s
+ + + , (28)
12 22 N2 s Ns
a lower bound of the SNR of the proposed precoding design is
N2 s
SNRCI  SNRlb . (29)
Ns
Based on the SNR lower bound in (29), a lower bound of
the achievable rate is

2
= 2Ns 1 Hb Q
1bit,lb Ns
RCI (30)
Ns
!
! 2 N2
= 2Ns 1 Hb Q " 1 2s . (31)
Fig. 2. Typical convergence behavior of the alternating projection method. Ns 1
The normalized error decreases exponentially with the iteration number k.
In the figure, FRF C648 and WRF C84 . Note that the floor of the Comparing (31) and (14), we find that the power gap
normalized distance around 1015 is due to the precision limitation of our 1bit,lb 2
computer. By default, MATLAB uses 16 digits of precision. between R 1bit,ub and RCI is 10 log10 21 dB. Therefore, we
Ns
conclude that compared to optimal digital precoding (which is
 
(k) (k) 
   
FRF FRF   (k) 
 (k) W
W unknown), the power loss of the channel inversion precoding
RF 

RF
projection process continues until F F 2
t
NRF r
NRF is at most 10 log10 21 dB. It also implies that for a well-
Ns
is smaller than a specified threshold
. The convergence of conditioned effective channel, the power loss is small.
the alternating projection algorithm is discussed in details Furthermore, if there is only one RF chain at the receiver,
r = N = 1, the achievable rate in (27) and the upper
i.e., NRF
in [50]. In Fig. 2, we show a typical convergence behaviour s
when the transmitter is assumed to have 64 antennas and bound in (14) are exactly equal and it implies that the channel
8 RF chains, while the receiver employs 8 antennas and 4 RF capacity is achieved by the proposed transmission method.
F2 RF Ct  RF2 C  . Note
chains. 648 and W 84 At last, if there are more RF chains at the receiver than that
 2 Therefore,
  that
FRF  F =  FRF  F = NRF and W  RF  = W  RF 2 = r > N t , then only N t (or less
at the transmitter, i.e., NRF
F F r t r
RF RF
r . So the distance is normalized by t than NRF ) out of NRF receive RF chains are used, otherwise
NRF NRF and NRF ,
GG in (18) does not have the inverse. The power consump-
respectively. It is seen the algorithm converges very fast,
tion also decreases by turning off some receive RF chains.
within less than 100 iterations, to a normalized distance of
less than 105 .
Another choice for designing analog precoder is to assume C. Rate Analysis With Few-Bit Quantization
that FRF and WRF consist of columns from the DFT For a real-valued SISO channel with b-bit quantizer,
matrices [1]. This is inspired by the virtual channel rep- a benchmark design was proposed in [36] where input signal
resentation [51]. Note that the DFT matrix has constant- are equiprobable, equispaced 2b -PAM (pulse amplitude modu-
norm entries and orthogonal columns, therefore the norm and lated), and quantizer thresholds are chosen to be the mid-points
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS 2279

of the input mass point locations. Although this combination of Transmission Method 2 Channel Inversion Based Transmis-
input and quantization is suboptimal, it is shown in [36] to be sion Method
close to the optimum which is obtained by high-complexity 1) Analog precoding design: Find the approximate solu-
iterative algorithm. In addition, this combination is actually tion WRF and FRF by the alternating projection shown
optimal for one-bit quantization. Therefore, in the proposed in Algorithm 1.
method, we assume the simple 22b -QAM signaling at the 2) Digital precoding design:
transmitter and uniform quantization at the receiver.
a) Compute the effective channel G  WRF HFRF .
The channel inversion based transmission, including the b) Set the digital precoder FBB as
analog and digital precoding design, signaling and quantiza-
!  
tion, is summarized in Transmission Method 2. ! Ns 1
FBB = ! %  & G G G .
We next show an example of two-bit quantization. The " 1
tr G G
derivation with multi-bit quantization is similar.
 The set of
  3
input signals is S = 3 2 , ,
2 2 , 2 where  is the
stepsize. The transition probability matrix is given in (32), 3) Signaling: s is chosen to be 22b -QAM symbols.
shown at the bottom of this page, where 2 denotes the noise 4) Quantization: Uniform quantization.
variance and () is the cumulative distribution function of the
standard normal distribution. The transition probability matrix
of higher resolution ADCs could be obtained similarly. A simple lower bound of (35) can be found by Fanos
The SNR of each sub-channel must be equal to the value inequality [52, Sec. 2.10]. The conditional entropy is upper
given in (23). Therefore, bounded by
   2 
1
SNRCI = b+1 2 +(3)2 + + 2b 1  / 2 Hb (s|r ) Hb (Pe ) + Pe log(|S| 1), (36)
2
1 1 2
SNRCI = b+1 2b1 (22b 1) 2 (33) where Pe is the error probability. Hence, the mutual
2 3
# information between s and r is
12SNRCI 
= (34)
22b 1 I (s; r ) = Hb (s) Hb (s|r ) (37)
 
where (33) is from the fact that 12 + 32 + + (2n 1)2 = b Hb (Pe ) Pe log 2b 1 . (38)
3 n(4n 1).
1 2

The achievable rate can therefore be computed as A lower bound of the sum rate of 2Ns sub-channels therefore is
 Pr(r |s)   
b bit
RCI = 2Ns Pr(s) Pr(r |s) log b bit,lb
Pr(r ) RCI = 2Ns b Hb (Pe ) Pe log 2b 1 , (39)
s r
 Pr(r |s)
= 2Ns Pr(s) Pr(r |s) log $

where the error probability Pe for 2b -PAM signaling is [53]
s r s
Pr(s ) Pr(r |s )
 1    
(a) Pr(r |s) 1 
2Ns Pr(r |s) log 1 $ Pe = 2 1 b Q (40)

s
Pr(r |s )
2 b
2 2
s r 2b     # 
1  Pr(r |s) 1 3 SNRCI
= 2Ns b + b Pr(r |s) log $ , = 2 1 b Q . (41)
2 s r

s
Pr(r |s )
2 22b 1
(35)
From (39), we find that the as SNRCI increases, Pe
where (a) follows from that the equal transmission b bit,lb
decreases to zero and RCI converges to 2Ns
b bps/Hz.
probability Pr(s) = 21b . The achievable rate in (35) is In addition, note that for the one-bit case, Pe = Q SNRCI
complicated and does not provide us with enough intuition. and therefore (39) degrades to (27).

 3  5 3 5 3
( 2 ) (
2
) ( )
2
(
2
) (
2
) 1 ( )
2 s=
2
   3  3 
( ) ( ) ( ) ( ) ( ) 1 ( ) s=
2 2 2 2 2 2
Pr(r |s) = 3  3    
2 (32)
( ) ( ) ( ) ( ) ( ) 1 ( ) s=

2 2 2 2 2 2 2
5 3 5  3  3
( ) ( ) ( ) ( ) ( ) 1 ( ) s=
2 2 2 2 2 2 2
3   3
r = r = r= r=
2 2 2 2
2280 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 4, APRIL 2017

VI. ACHIEVABLE R ATE W ITH S INGULAR VALUE Transmission Method 3 SVD Based Transmission Method
D ECOMPOSITION BASED T RANSMISSION 1) Analog precoding design: Find the approximate solu-
The channel inversion precoding generally works well at tion WRF and FRF by the alternating projection method
high SNR but has poor performance at low SNR. In the shown in Algorithm 1.
second transmission method, we use the singular value 2) Digital precoding design:

decomposition (SVD) digital precoding. Since the interference a) Compute the effective channel G  WRF HFRF .
between each streams can not be completely eliminated before b) Set the digital precoder FBB by SVD of G and
quantization as in the channel inversion case, the exact achiev- waterfilling method.
able rate is unknown. We therefore choose to apply the additive 3) Signaling: s is chosen to follow Gaussian signaling.
quantization noise model (AQNM) [28], [33], [54], which is 4) Quantization: The thresholds of ADC are determined
accurate enough at low SNR, to find a lower bound of the by Max-Lloyd algorithm [55], [56] which minimizes the
achievable rate. MSE of Gaussian distributed input.
Applying the additive quantization noise model, the
equivalent channel is

 For higher SNR, the optimal FBB maximizing the rate
r = (1 b ) GFBB s + WRF n + nQ , (42)
RAQNM in (43) is unknown. We therefore use the conventional
where b is the distortion

factor for b-bit ADC defined SVD precoding and waterfilling power allocation as done
E (Q(y)y)2
as b = E[|y|2 ]
and nQ is the quantization noise with in [27] and [34]. The baseband digital precoder is
   
the variance b (1 b )diag NPts GFBB FBB G + N2 I . Note FBB = V diag p , (48)
that we ignore the cross-correlation among the elements of nQ where V is obtained from the singular value decomposition
here for simplicity as done in [27], [33], and [34]. As shown of the matrix G, i.e., G = U V and p denotes the power
in the simulation results of [44], this diagonal approximation allocation factor obtained from the waterfilling method.
is accurate enough for the MIMO channel, especially at For the analog precoding and combining, the optimal
low SNR. FRF and WRF maximizing the rate in (43) are unknown. We
Assuming s is Gaussian distributed and the quantization adopt the same alternating projection method in Algorithm 1
noise is the worst case of Gaussian distributed, a lower bound to find a suboptimal solution. As a result, the analog precoding
of the achievable rate is given in (43), shown at the bottom of is designed by alternating projection method with initial values
the page. obtained by SVD of channel and the digital precoding is got
At low SNR, the achievable rate is approximated to be by SVD of the baseband channel. The proposed SVD based
 
  design is summarized in Transmission Method 3.
RAQNM = log2 I + (1 b ) FBB G GFBB  + o() Last, we show why the AQNM model is not accurate enough
Ns
(1 b )   at high SNR to model the quantization channel. At high SNR,
= tr FBB G GFBB + o(). (44) RAQNM converges as follows.
Ns ln 2
 
To maximize the term N1s tr GFBB FBB G under the RAQNM
 
 1 b   
1

constraint FBB 2F Ns , the optimal choice of FBB is the 
log2 I + FBB G diag GFBB FBB G
GFBB 
eigenmode beamforming, i.e., b
 % & % & 
 1 b 1 1 
Ns = 1, and FBB = v1 , (45) 
= log2 I + A diag diag A
b ||ai || ||ai ||
 
where v1 is the right singular vector corresponding to the  1 
= log2 I + 
A 

A
b
largest singular value of G. The resulting rate is b
(1 b )12 Ns  
1 b  
RAQNM = + o(). (46) = log2 1 + i A A (49)
ln 2 b
i=1
For one-bit quantization, 1 = 1 2
and the rate is
where A  GFBB , ai is the i -th row of A, and 
A is obtained
2 12 by normalizing each each row of A. Since each row of A has
RAQNM = + o(). (47)
ln 2 unit norm, then
Notice that the upper bound of one-bit quantized chan- 
Ns
   
nel at low SNR given in (15) is achieved by eigenmode i 
A 
A = tr 
A 
A = Ns . (50)
beamforming. i=1

  % &1 
 
 
RAQNM = log2 I + (1 b ) FBB G I + b diag GFBB FBB G GFBB  . (43)
 Ns Ns 
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS 2281

Therefore, we have
 
(a) 1 b
RAQNM Ns log2 1 + (51)
b
1
= Ns log2 , (52)
b
where (a) follows from Jensens inequality and that log2 (1+x)
is concave in x. When the ADC resolution b is large (b 3),
the distortion factor b can be approximated as [57]

3 2b
b 2 . (53)
2
As a result, the rate obtained by AQNM model is upper
bounded by

3
RAQNM 2Ns b Ns log2 (54)
2
2Ns b 1.44Ns . (55)
However, we know the achievable rate of quantized MIMO
channel is upper bounded by 2Ns b bps/Hz and the channel
inversion method can achieve the bound at high enough SNR
as shown in (39). Therefore, the AQNM is not an accurate
model at high SNR. The reason is threefold. First, the input
signal s is assumed to follow suboptimal Gaussian distribution.
Second, the quantization noise is assumed to be the worst-case
Gaussian noise. Third, the Max-Loyd quantizer minimizing the
MSE is not necessarily optimum for maximizing the channel
capacity.

VII. S IMULATION R ESULTS

We evaluate the performance of the proposed methods in
a mmWave MIMO channel with large antenna arrays and
limited numbers of transmit and receive RF chains. According
to measurement results [58], [59], the number of clusters
tends to be lower in the mmWave band compared with lower
frequencies. The mmWave channel will mostly consist of
the line-of-sight (LOS) and a few NLOS clusters. In the
simulations, the wireless channel is assumed to have 4 clusters,
each of which consists of 5 rays. The angle spread is 7.5
degrees. These numbers are chosen according to the urban
macro (UMa) NLOS channel measurement results at 28 GHz
given in the white paper [60]. The results are obtained by
averaging over 100 channel realizations.
A. Achievable Rates
In this subsection, we evaluate the achievable rates of the
proposed architecture by numerical simulations. The transmit-
ter is assumed to have Nt = 64 antennas and NRF t = 8 RF
chains, while the receiver employs Nr = 8 antennas. The
number of data streams is assumed to same as the number
of receive RF chains, i.e., Ns = NRF
r .
Fig. 3 shows the achievable rates when the receiver has
1 and 4 RF chains, respectively. In Fig. 3(a), it is seen that
when there is only one RF chains at the transmitter, the channel
inversion and SVD method has close performance at low SNR.
At high SNR, however, the channel inversion method achieves
the rate 2b bps/Hz while the rate of SVD method saturates to
Fig. 3. Rates versus SNR of different transmission methods. The trans-
mitter is assumed to have 64 antennas and 8 RF chains. The receiver
employs 8 antennas. In Fig. (a), the receiver has 1 RF chains while in Fig. (b),
the receiver has 4 RF chains.
log2 1b which is about 1.46, 3.09, 4.86 bps/Hz for b = 1,
2, 3, respectively. In Fig. 3(b), another case when NRF r = 4
is shown. It is found that although at high SNR the channel
inversion method achieves larger rate than SVD method, its
performance at low SNR is much worse. The reason is that
since the channel has only 4 clusters, the fourth largest singular
2
value 4 is small and the power loss log2 12 is large. Last, the
4
lower bound provided in (39) is also plotted. It is seen that
the lower bound is tight for 1-bit ADC case. For other cases,
the lower bound is tight at high SNR.
Fig. 4 compares the achievable rate of fully-digital and
hybrid architecture. The rate of hybrid architecture is the
maximum of the CI and SVD method. The gap between
the curves of Digital-b bit and Hybrid-b bit represents
the loss due to limited number of RF chains while the gap
between the curves of bit and bbit is the loss due to low
resolution ADCs. It is seen that the digital architecture is much
better than the hybrid architecture with only one RF chain
because the hybrid architecture can only support single stream
transmission while the digital architecture can support at most
2282
Fig. 4. Rates versus SNR of different transmission methods. The transmitter
is assumed to have 64 antennas and 8 RF chains. The receiver employs
8 antennas. In Fig. (a), the receiver has 1 RF chains while in Fig. (b), the
receiver has 4 RF chains.
8 streams simultaneously. However, when there are 4 receive
RF chains in the hybrid architecture, the gap between these
two architectures is small since the channel has only 4 clusters.
Last, we can see the loss due to low resolution ADCs is small
at low and medium SNRs. For example, the gap between the
curve Hybrid-bit and Hybrid-3bit is less than 3 dB when
the SNR is less than 0 dB in Fig. 4(b).
Fig. 5 shows the achievable rate with respect to the ADC
resolution. First, as expected, the rates of the finite-bit ADC
receiver increase with resolution. Second, with multi-bit ADCs
(5-bit when SNR = 10 dB and 7-bit when SNR = 10 dB),
the SVD method achieves the performance similar to that of
hybrid architecture with -bit ADCs. This implies that high
resolution ADCs do not provide much gain compared to the
few-bit ADCs when the SNR is low. Third, when the ADC
resolution is low, channel inversion method is better than the
SVD method while with high resolution quantization, the SVD
method is better. This is reasonable since with high resolution
ADC, the channel is close to the one without quantization and
in a unquantized channel, SVD method is optimum.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 4, APRIL 2017
Fig. 5. Rates versus ADC resolution for different transmission methods. The
transmitter is assumed to have 64 antennas and 8 RF chains. The receiver
employs 8 antennas and 4 RF chains. In Fig. (a), the SNR is 10 dB while
in Fig. (b), the SNR is 10 dB.
Fig. 6 presents the achievable rates versus the number
of RF chains at the receiver. First, we find that the rate
of SVD method always increases with NRF r . Second, at
low SNR (10 dB), the channel inversion method achieve
the largest rate when NRF t = 2. This means at low
SNR, it is better to turn off some RF chains and transmit
fewer number of streams. Note that the power consump-
tion also decreases by turning off some RF chains. Third,
we find that compared to fully-digital architecture where
NRFt = Nt = 64 and NRF r = Nr = 8, the hybrid
architecture with limited number of RF chains (NRF t = 8
and NRF = 2) and low resolution ADCs (4-bit) incurs about
r
20%-30% spectral efficiency loss. As shown in the next
subsection, however, the energy efficiency of the proposed
receiver is much higher than the fully-digital architecture.
B. Energy Efficiency
In this subsection, we evaluate the performance of the
different receiver architectures by investigating the trade-off
between their achievable rates and power consumptions. First,
we formulate a generic power consumption model for the
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS
Fig. 6. Rates versus the number of receive RF chains for different transceiver
architecture and transmission methods. The transmitter is assumed to have
64 antennas and 8 RF chains, while the receiver employs 8 antennas.
In Fig. (a), the SNR is 10 dB while in Fig. (b), the SNR is 10 dB.
hybrid architecture with low-resolution ADCs. Then, we use
this model in the performance evaluation. Consider the system
model in Fig. 1 with a b-bit ADC receiver having Nr antennas
and NRFr RF chains. Let PLNA , PPS , PRFchain , PADC , PBB
denote the power consumption in the LNA, phase shifter, RF
chain, ADC, and baseband processor, respectively. Then, the
consumed power by the hybrid combining receiver in Fig. 1
can be approximated as [61]
Ptot = Nr PLNA + NRF
r
(Nr PPS + PRFchain + 2PADC ) + PBB ,
(56)
where the power consumed by ADCs can be further expressed
in terms of the number of bits as
PADC = F O MW f s 2b , (57)
where f s is the Nyquist sampling rate, b is the number of bits,
and F O MW is Waldens figure-of-merit for evaluating ADCs
power efficiency with resolution and speed [10], [24],
Next, we present some simulation results in Figs. 7-8 that
illustrate the power consumption-rate trade-off of the hybrid
2283
Fig. 7. Trade-off between the achievable rate and power consumption for
hybrid combiners with different number of RF chains and ADC bits. The
results also include the fully-digital architecture with finite-bit ADCs. The
transmitter is assumed to have 64 antennas and 8 RF chains, while the receiver
employs 8 antennas and 1, 2 or 4 RF chains. The SNR is 10 dB and the
bandwidth is 1 GHz.
combining receiver for different number of RF chains and
ADC quantization bits at the receiver. In these simulations,
we adopt the same clustered channel model, and consider
the system model in Section II with a transmitter having
64 antennas and 8 RF chains, and a receiver deploying
8 antennas.
First, we plot in Fig. 7 the power consumption versus the
achievable rate for the fully-digital and proposed hybrid archi-
tecture with few-bit ADC receivers. The power consumption
in the different receivers are calculated based on the power
model in (56)-(57). The consumed power in the RF compo-
nents are assumed to be PLNA = 20 mW, PPS = 10 mW,
PRFchain = 40 mW, PBB = 200 mW [61], [62]. For the
ADCs, we adopt the power consumption model in (57)
with F O MW = 500 fJ/conversion-step, which is a typical
achievable value at 1 GHz [10], [63]. 1
The achievable rate for the hybrid architecture with few-
bit ADCs in Fig. 7 is calculated as the maximum of the two
achievable rate expressions in (35) and (43) for CI and SVD
based precoding. Several important insights can be obtained
from Fig. 7. First, for all the four cases in the figure, the
achievable rate significantly increases when the number of bits
increases from 1 to 4 5 bits, with a negligible increase in
the receiver power consumption. This implies that with these
system configurations, having 4 5 bits ADCs can lead to
a better trade-off between the rate and power consumption
compared with having 1-bit ADC receivers. Second, going
from 4 5 bits to 7 8 bits for the ADCs slightly improves
the achievable spectral efficiency, but with significantly more
power consumption. This means that 4 5 ADC bits can
be better than 7 8 ADC bits when investigating the rate-
power consumption trade-off. Third, it is found that among
the four cases considered, the hybrid one with 1 RF chain
gives the best power-rate trade-off when the rate is less than
1 The minimum achievable value of F O M
W can be as low as
5 fJ/conversion-step as shown in the Figure Walden FOM vs. Speed in [10].
2284 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 4, APRIL 2017
Fig. 8. Energy efficiency of the digital and hybrid architecture with few-bit
ADC receivers for different numbers of bits. The transmitter is assumed to
have 64 antennas and 8 RF chains, while the receiver employs 8 antennas and
2 RF chains. In Fig. (a), the SNR is 10 dB while in Fig. (b), and the SNR
is 10 dB.
11 Gbps, the hybrid one with 2 RF chains gives the best trade-
off when the rate is between 11 Gbps and 18 Gbps, and the
fully-digital architecture is best when the rate is larger than
18 Gbps. Last, when there are 4 RF chains, the performance
of hybrid architecture is dominated by that of the fully-digital
architecture. To achieve the same rate, the power consumption
of hybrid receiver with 4 RF chains is always larger than of
fully-digital one because of the high power consumption of
the large number of analog phase shifters and RF chains.
Based on the observations made from Fig. 7, we investigate
the energy efficiency of the different receiver architectures in
Fig. 8. The energy spectral efficiency EE is defined as
R W
EE = bits/Joule, (58)
Ptot
where R is the achievable spectral efficiency and W is
the transmission bandwidth. The channel and system models
adopted in Fig. 8(a)-Fig. 8(b) are similar to that in Fig. 7
with 2 RF chains used at the receiver and with W = 1 GHz
bandwidth. In Fig. 8(a), the energy efficiency of the few-bit
ADC hybrid architecture is plotted for two different precoding
techniques versus the number of ADC bits at SNR = 10 dB.
The result is compared with the energy efficiencies of the
fully-digital transceiver. Fig. 8(a) shows that the number of
bits maximizing the energy efficiency is finite and equals to
4 bits. This maximum energy efficiency is achieved with the
SVD precoding in (43). It is worth noting from Fig. 8(a)
that the choice of the precoding technique depends on the
number of ADC quantization bits. While the channel inversion
achieves a better energy efficiency with 1 2 ADC bits,
SVD precoding can lead to a better energy efficiency for
higher numbers of ADC quantization bits. Fig. 8(b) has the
same comparison in Fig. 8(a) with the only difference of
operating at SNR = 10 dB. The energy efficiency trends
in Fig. 8(b) are similar to that in Fig. 8(a) with the main
difference that the maximum energy efficiency is achieved
with the channel inversion precoding in (35). This implies
that determining the precoding techniques should take the
SNR into consideration. We note though the 4 bits for the
ADC quantization still achieves the optimal energy efficiency.
Further, Fig. 8(a)-Fig. 8(b) also show that energy efficiency of
the proposed few-bit ADC hybrid architecture can be much
better that that of the fully-digital architectures that employ
high resolution ADCs. The reason is that the ADC will
dominate the power consumption when the resolution is high
and hybrid architecture saves power by reducing the number
of high-resolution ADCs.
VIII. C ONCLUSION
In this paper, we derived the achievable spectral efficiency
and energy-rate trade-off of a generalized hybrid architecture
with few-bit ADC receivers. First, we considered channel
inversion and SVD precoding based transmission methods
and derived their achievable rates. The transmission methods
include three elements: the design of analog and digital
precoding, the choice of the transmit signal distribution and the
setup of quantizer. Simulation results showed that at the low
and medium SNRs, the proposed architecture and precoding
can achieve a comparable rate to the fully-digital solution.
Second, we explored the trade-off between the achievable rate
and the power consumption, which is particularly important
in massive MIMO and mmWave systems. Adopting energy
efficiency as a performance metric, the proposed hybrid archi-
tecture with few-bit ADC receiver is compared with the fully-
digital transceiver and the conventional hybrid architecture
with full-resolution ADCs. Numerical results showed that the
generalized hybrid architecture with coarse ADC quantization
achieves a potential energy efficiency gain compared to the
other architectures. The results illustrated that coarse ADC
quantization with 4 5 bits normally achieves the maximum
energy efficiency in various system configurations.
A PPENDIX
P ROOF OF P ROPOSITION 1
The proof is similar to the proof of [14, Proposition 4].
The channel without any constraint on digital precoding and
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS 2285

Pt
combining is, where  N2
. Therefore, the upper bound in (14) is obtained.


r = sgn WRF HFRF x + WRF n . (59)
We want to find an upper bound of the mutual information R EFERENCES
I (r; x) subject to the power constraint E[FRF x2 ] Pt . [1] J. Mo, A. Alkhateeb, S. Abu-Surra, and R. W. Heath, Jr., Achievable
First, we have, rates of hybrid architectures with few-bit ADC receivers, in Proc. 20th
Int. ITG Workshop Smart Antennas, Mar. 2016, pp. 18.
I (r; x) = H(r) H(r|x) (60) [2] T. L. Marzetta, Noncooperative cellular wireless with unlimited num-
bers of base station antennas, IEEE Trans. Wireless Commun., vol. 9,
(a)
r
2NRF H(r|x), (61) no. 11, pp. 35903600, Nov. 2010.
[3] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
where inequality (a) follows from that there are at most 2NRF
r Five disruptive technology directions for 5G, IEEE Commun. Mag.,
vol. 52, no. 2, pp. 7480, Feb. 2014.
quantization outputs. [4] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, Massive
Next, we derive a lower bound of H(r|x). For a given MIMO for next generation wireless systems, IEEE Commun. Mag.,
transmitted signal x = x
, denote z
= WRF HF x
and vol. 52, no. 2, pp. 186195, Feb. 2014.
RF



[5] T. S. Rappaport et al., Millimeter wave mobile communications for 5G

z = [Re(z ) , Im(z ) ] . The conditional entropy of r given
T T T
cellular: It will work! IEEE Access, vol. 1, pp. 335349, May 2013.
x = x
is, [6] W. Roh et al., Millimeter-wave beamforming as an enabling tech-
r
NRF
nology for 5G cellular communications: Theoretical feasibility and
(a)  prototype results, IEEE Commun. Mag., vol. 52, no. 2, pp. 106113,

H(r|x = x ) = H(r j |x
) (62) Feb. 2014.
j =1 [7] A. Alkhateeb, J. Mo, N. Gonzalez-Prelcic, and R. W. Heath, Jr., MIMO
r
2NRF  
 precoding and combining solutions for millimeter-wave systems, IEEE
(b)  2
Commun. Mag., vol. 52, no. 12, pp. 122131, Dec. 2014.
= Hb Q 
z (63) [8] R. W. Heath, Jr., N. Gonzlez-Prelcic, S. Rangan, W. Roh, and
j =1
N2 j A. M. Sayeed, An overview of signal processing techniques for
millimeter wave MIMO systems, IEEE J. Sel. Top. Signal Process.,

!   vol. 10, no. 3, pp. 436453, Mar. 2016.
! 2
z
j
r
! 2 
2NRF
(c) 
[9] T. Bai and R. W. Heath, Jr., Coverage and rate analysis for millimeter-
= Hb " wave cellular networks, IEEE Trans. Wireless Commun., vol. 14, no. 2,
Q 2
(64) pp. 11001114, Feb. 2015.
j =1 N [10] B. Murmann. (2016). ADC Performance Survey 19972016, accessed
on Jul. 12, 2016. [Online]. Available: http://www.stanford.edu/ mur-
where (a) follows from that the effective noise WRF n has mann/adcsurvey.html
[11] W. Hong, K.-H. Baek, Y. Lee, Y. Kim, and S.-T. Ko, Study and
independent entries, (b) follows from that the in-phase and prototyping of practically large-scale mmWave antenna systems for 5G
quadrature parts of the effective noise are independent and cellular devices, IEEE Commun. Mag., vol. 52, no. 9, pp. 6369,
2 Sep. 2014.
Gaussian distributed with variance 2N , (c) follows from that [12] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
Hb (Q(x)) is an even function of x. Next, noting that Jr., Spatially sparse precoding in millimeter wave MIMO systems,
r IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 14991513,
  2
2NRF
Mar. 2014.
z
j = z
2 x
2 12 ,
 (65) [13] S. Han, I. Chih-Lin, Z. Xu, and C. Rowell, Large-scale antenna
j =1 systems with hybrid analog and digital beamforming for millime-
ter wave 5G, IEEE Commun. Mag., vol. 53, no. 1, pp. 186194,
where 1 is the maximum singular value of the effective Jan. 2015.
HF , we have
channel WRF [14] J. Mo and R. W. Heath, Jr., Capacity analysis of one-bit quantized
RF
 
 MIMO systems with transmitter channel state information, IEEE Trans.
Signal Process., vol. 63, no. 20, pp. 54985512, Oct. 2015.

r x
2 12
H(r|x = x ) 2NRF Hb Q r 2
(66) [15] X. Zhang, A. F. Molisch, and S.-Y. Kung, Variable-phase-shift-based
NRF RF-baseband codesign for MIMO antenna selection, IEEE Trans. Signal
N
   Process., vol. 53, no. 11, pp. 40914103, Nov. 2005.
by Jensens inequality since Hb Q x is decreasing and [16] V. Venkateswaran and A. van der Veen, Analog beamforming in
MIMO communications with phase shift networks and online channel
convex in x (see [64] for the proof of convexity). estimation, IEEE Trans. Signal Process., vol. 58, no. 8, pp. 41314143,
The conditional entropy of r is, Aug. 2010.
/  
0 [17] A. Alkhateeb, O. El Ayach, G. Leus, and R. Heath, Jr., Hybrid
r x2 12 precoding for millimeter wave cellular systems with partial channel
H(r|x) = Ex 2NRF Hb Q r 2
. (67) knowledge, in Proc. Inf. Theory Appl. Workshop (ITA), Feb. 2013,
NRF N pp. 15.
[18] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., Channel
Since Hb (Q( z)) is convex, then estimation and hybrid precoding for millimeter wave cellular systems,
 
 IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831846,
r E[x2 ]12 Oct. 2014.
H(r|x) 2NRF Hb Q r 2
. (68) [19] C. Rusu, R. Mndez-Rial, N. Gonzlez-Prelcic, and R. W. Heath, Jr.,
NRF N Low complexity hybrid sparse precoding and combining in millimeter
wave MIMO systems, in Proc. IEEE Int. Conf. Commun. (ICC),

minimize H(r|x) subject to the power
We want to
constraint Jun. 2015, pp. 13401345.
E FRF x2 Pt , which is equivalent to E x2 Pt by
[20] C. E. Chen, An iterative hybrid transceiver design algorithm for
the assumption FRF FRF = I. As Hb (Q( z)) is decreasing in millimeter wave MIMO systems, IEEE Wireless Commun. Lett., vol. 4,
z, we have no. 3, pp. 285288, Jun. 2015.
 #  [21] T. E. Bogale and L. B. Le, Beamforming for multiuser mas-
sive MIMO systems: Digital versus hybrid analog-digital, in
12
r H
H(r|x) 2NRF b Q r
NRF , (69) Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2014,
pp. 40664071.
2286 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 4, APRIL 2017
[22] F. Sohrabi and W. Yu, Hybrid digital and analog beamforming design
for large-scale MIMO systems, in Proc. IEEE Int. Conf. Acoust., Speech
Signal Process. (ICASSP), Brisbane, Qld., Queens., Australia, Apr. 2015,
pp. 29292933.
[23] L. Liang, W. Xu, and X. Dong, Low-complexity hybrid precoding
in massive multiuser MIMO systems, IEEE Wireless Commun. Lett.,
vol. 3, no. 6, pp. 653656, Dec. 2014.
[24] R. H. Walden, Analog-to-digital converter survey and analysis, IEEE
J. Sel. Areas Commun., vol. 17, no. 4, pp. 539550, Apr. 1999.
[25] L. Bin, T. W. Rondeau, J. H. Reed, and C. W. Bostian, Analog-to-
digital converters, IEEE Signal Process. Mag., vol. 22, no. 6, pp. 6977,
Nov. 2005.
[26] A. Mezghani and J. A. Nossek, On ultra-wideband MIMO systems with
1-bit quantized outputs: Performance analysis and input optimization,
in Proc. IEEE Int. Symp. Inf. Theory, Jun. 2007, pp. 12861289.
[27] A. Mezghani, M.-S. Khoufi, and J. A. Nossek, A modified MMSE
receiver for quantized MIMO systems, in Proc. ITG/IEEE WSA, Vienna,
Austria, 2007, pp. 15.
[28] A. Mezghani and J. Nossek, Capacity lower bound of MIMO channels
with output quantization and correlated noise, in Proc. IEEE Int. Symp.
Inf. Theory, 2012, pp. 15.
[29] W. Zhang, A general framework for transmission with transceiver
distortion and some applications, IEEE Trans. Commun., vol. 60, no. 2,
pp. 384399, Feb. 2012.
[30] J. Mo, P. Schniter, N. G. Prelcic, and R. W. Heath, Jr., Channel
estimation in millimeter wave MIMO systems with one-bit quantiza-
tion, in Proc. 48th Asilomar Conf. Signals, Syst. Comput., Nov. 2014,
pp. 957961.
[31] J. Mo, P. Schniter, and R. W. Heath, Jr. (Oct. 2016). Channel estimation
in broadband millimeter wave MIMO systems with few-bit ADCs.
[Online]. Available: https://arxiv.org/abs/1610.02735
[32] J. Mo and R. W. Heath, Jr., Limited feedback in multiple-antenna
systems with one-bit quantization, in Proc. 49th Asilomar Conf. Signals,
Syst. Comput., Nov. 2015, pp. 14321436.
[33] Q. Bai and J. Nossek, Energy efficiency maximization for 5G multi-
antenna receivers, Trans. Emerg. Telecommun. Technol., vol. 26, no. 1,
pp. 314, 2015.
[34] O. Orhan, E. Erkip, and S. Rangan, Low power analog-to-digital con-
version in millimeter wave systems: Impact of resolution and bandwidth
on performance, in Proc. Inf. Theory Appl. (ITA) Workshop, 2015,
pp. 18.
[35] S. Wang, Y. Li, and J. Wang, Multiuser detection in massive spatial
modulation MIMO with low-resolution ADCs, IEEE Trans. Wireless
Commun., vol. 14, no. 4, pp. 21562168, Apr. 2015.
[36] J. Singh, O. Dabeer, and U. Madhow, On the limits of communication
with low-precision analog-to-digital conversion at the receiver, IEEE
Trans. Commun., vol. 57, no. 12, pp. 36293639, Dec. 2009.
[37] J. Choi, J. Mo, and R. W. Heath, Jr., Near maximum-likelihood detector
and channel estimator for uplink multiuser massive MIMO systems with
one-bit ADCs, IEEE Trans. Commun., vol. 64, no. 5, pp. 20052018,
May 2016.
[38] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer.
(May 2015). One-bit massive MIMO: Channel estimation and
high-order modulations. [Online]. Available: https://arxiv.org/abs/
1504.04540
[39] C. Mollen, J. Choi, E. G. Larsson, and R. W. Heath, Jr., Uplink
performance of wideband massive MIMO with one-bit ADCs, IEEE
Trans. Wireless Commun., vol. 16, no. 1, pp. 87100, Jan. 2017.
[40] C. Studer and G. Durisi, Quantized massive MU-MIMO-OFDM
uplink, IEEE Trans. Commun., vol. 64, no. 6, pp. 23872399, Jun. 2016.
[41] J. Zhang, L. Dai, S. Sun, and Z. Wang, On the spectral efficiency of
massive MIMO systems with low-resolution ADCs, IEEE Commun.
Lett., vol. 20, no. 5, pp. 842845, May 2016.
[42] B. Murmann, Energy limits in A/D converters, in Proc. IEEE Faible
Tension Faible Consommation (FTFC), Jun. 2013, pp. 14.
[43] W. B. Abbas, F. Gomez-Cuba, and M. Zorzi. (Jul. 2016). Millime-
ter wave receiver efficiency: A comprehensive comparison of beam-
forming schemes with low resolution ADCs. [Online]. Available:
https://arxiv.org/abs/1607.03725
[44] K. Roth and J. A. Nossek. (Oct. 2016.).Achievable rate and energy effi-
ciency of hybrid and digital beamforming receivers with low resolution
ADC. [Online]. Available: https://arxiv.org/abs/1610.02909
[45] H. Ghauch, T. Kim, M. Bengtsson, and M. Skoglund, Subspace esti-
mation and decomposition for large millimeter-wave MIMO systems,
IEEE J. Sel. Top. Signal Process., vol. 10, no. 3, pp. 528542, Apr. 2016.
[46] U. S. Kamilov, V. K. Goyal, and S. Rangan, Message-passing de-
quantization with applications to compressed sensing, IEEE Trans.
Signal Process., vol. 60, no. 12, pp. 62706281, Dec. 2012.
[47] A. F. Molisch and X. Zhang, FFT-based hybrid antenna selection
schemes for spatially correlated MIMO channels, IEEE Commun. Lett.,
vol. 8, no. 1, pp. 3638, Jan. 2004.
[48] C. R. Rao, Separation theorems for singular values of matrices and
their applications in multivariate analysis, J. Multivariate Anal., vol. 9,
no. 3, pp. 362377, 1979.
[49] D. P. Palomar, J. M. Cioffi, and M. A. Lagunas, Joint Tx-Rx beam-
forming design for multicarrier MIMO channels: A unified framework
for convex optimization, IEEE Trans. Signal Process., vol. 51, no. 9,
pp. 23812401, Sep. 2003.
[50] J. A. Tropp, I. S. Dhillon, R. W. Heath, and T. Strohmer, Designing
structured tight frames via an alternating projection method, IEEE
Trans. Inf. Theory, vol. 51, no. 1, pp. 188209, Jan. 2005.
[51] A. M. Sayeed, Deconstructing multiantenna fading channels, IEEE
Trans. Signal Process., vol. 50, no. 10, pp. 25632579, Oct. 2002.
[52] T. M. Cover and J. A. Thomas, Elements of Information Theory.
Hoboken, NJ, USA: Wiley, 2012.
[53] J. G. Proakis, Digital Communications. New York, NY, USA:
McGraw-Hill, 2008.
[54] A. Fletcher, S. Rangan, V. Goyal, and K. Ramchandran, Robust predic-
tive quantization: Analysis and design via convex optimization, IEEE
J. Sel. Topics Signal Process., vol. 1, no. 4, pp. 618632, Dec. 2007.
[55] J. Max, Quantizing for minimum distortion, IRE Trans. Inf. Theory,
vol. 6, no. 1, pp. 712, Mar. 1960.
[56] S. Lloyd, Least squares quantization in PCM, IEEE Trans. Inf. Theory,
vol. IT-28, no. 2, pp. 129137, Mar. 1982.
[57] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression,
vol. 159. Springer, 2012.
[58] M. R. Akdeniz et al., Millimeter wave channel modeling and cellular
capacity evaluation, IEEE J. Sel. Areas Commun., vol. 32, no. 6,
pp. 11641179, Jun. 2014.
[59] T. S. Rappaport, G. R. Maccartney, M. K. Samimi, and S. Sun, Wide-
band millimeter-wave propagation measurements and channel models for
future wireless communication system design, IEEE Trans. Commun.,
vol. 63, no. 9, pp. 30293056, Sep. 2015.
[60] 5G channel model for bands up to100 GHz in Proc. 3rd Workshop
Mobile Commun. Higer Freq. Bands (MCHFB), Dec. 2015. [Online].
Available: http://www.5gworkshops.com/5GCM.html
[61] R. Mndez-Rial, C. Rusu, N. Gonzlez-Prelcic, A. Alkhateeb, and
R. W. Heath, Jr., Hybrid MIMO architectures for millimeter wave
communications: Phase shifters or switches? IEEE Access, vol. 4,
pp. 247267, 2016.
[62] Y.-H. Lin and H. Wang, A low phase and gain error passive phase
shifter in 90 nm CMOS for 60 GHz phase array system application, in
Proc. IEEE MTT-S Int. Microw. Symp. (IMS), May 2016, pp. 14.
[63] A. M. A. Ali et al., A 14 Bit 1 GS/s RF sampling pipelined ADC with
background calibration, IEEE J. Solid-State Circuits, vol. 49, no. 12,
pp. 28572867, Dec. 2014.
[64] O. Dabeer, J. Singh, and U. Madhow, On the limits of communica-
tion performance with one-bit analog-to-digital conversion, in Proc.
IEEE 7th Workshop Signal Process. Adv. Wireless Commun., Jul. 2006,
pp. 15.
Jianhua Mo (S12) received the B.S. and M.S.
degrees in electronic engineering from Shanghai
Jiao Tong University, Shanghai, China, in 2010 and
2013, respectively, and the M.S. degree in elec-
trical and computer engineering from the Georgia
Institute of Technology, Atlanta, GA, USA. He is
currently pursuing the Ph.D. degree with the Wire-
less Networking and Communications Group, The
University of Texas at Austin, Austin, TX, USA. His
research interests include physical layer security and
mmWave communications.
He was a co-recipient of the Heinrich Hertz Award for Best Communi-
cations Letter in 2013. He was recognized as an Exemplary Reviewer of
the IEEE C OMMUNICATIONS L ETTERS in 2012, and the IEEE W IRELESS
C OMMUNICATIONS L ETTERS in 2015.
MO et al.: HYBRID ARCHITECTURES WITH FEW-BIT ADC RECEIVERS
Ahmed Alkhateeb (S08) received the B.S. (Hons.)
and M.S. degrees from Cairo University, Cairo,
Egypt, in 2008 and 2012, respectively. He is cur-
rently pursuing the Ph.D. degree with the Wireless
Networking and Communication Group, Department
of Electrical and Computer Engineering, The Uni-
versity of Texas at Austin, Austin, TX, USA. His
research interests include the broad area of network
information theory, communication theory, signal
processing, and in the context of wireless com-
munication, mmWave communication, and massive
MIMO systems.
Shadi Abu-Surra received the Ph.D. degree from
the University of Arizona in 2009.
In 2016, he joined PHAZR Inc. as a Principle
Engineer to bring 5G-Gigabit-communication sys-
tems to market. Prior to that, he was a Research
Engineer with Samsung Research America, since
2009, where he contributed to several standards and
projects, of them the 60 GHz connectivity standards
(Wireless Gigabit Alliance, IEEE 802.11ad, and
IEEE 802.11ay), and 5G mm-Wave mobile broad-
band system design and prototyping. His research
interests include digital communications, in particular next generation
mm-Wave wireless communication systems, and channel coding with empha-
sis on the design of LDPC codes and their decoders.
2287
Robert W. Heath, Jr. (S96M01SM06F11)
received the B.S. and M.S. degrees from the Univer-
sity of Virginia, Charlottesville, VA, USA, in 1996
and 1997 respectively, and the Ph.D. degree from
Stanford University, Stanford, CA, USA, in 2002,
all in electrical engineering. From 1998 to 2001, he
was a senior member of the technical staff and then
a Senior Consultant with Iospan Wireless Inc, San
Jose, CA, USA, where he was involved in the design
and implementation of the physical and link layers of
the first commercial MIMO-OFDM communication
system. Since 2002, he has been with the Department of Electrical and
Computer Engineering, The University of Texas at Austin, where he is a
Cullen Trust for Higher Education Endowed Professor, and a member of
the Wireless Networking and Communications Group. He is also President
and CEO of MIMO Wireless Inc. He is co-author of the book Millimeter
Wave Wireless Communications (Prentice Hall, 2014) and author of Digital
Wireless Communication: Physical Layer Exploration Lab Using the NI USRP
(National Technology and Science Press, 2012).
Dr. Heath has been a co-author of 15 award winning conference and
journal papers, including recently the 2010 and 2013 EURASIP Journal
on Wireless Communications and Networking Best Paper Awards, the 2012
Signal Processing Magazine Best Paper Award, the 2013 Signal Processing
Society Best Paper Award, the 2014 EURASIP Journal on Advances in Signal
Processing Best Paper Award, the 2014 Journal of Communications and
Networks Best Paper Award, the 2016 IEEE Communications Society Fred W.
Ellersick Prize, and the 2016 IEEE Communications and Information Theory
Societies Joint Paper Award. He was a Distinguished Lecturer with the IEEE
Signal Processing Society and is an ISI highly cited researcher. He is also an
elected member of the Board of Governors for the IEEE Signal Processing
Society, a licensed amateur radio operator, a private pilot, and a registered
professional engineer in Texas.