You are on page 1of 14

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

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
1

Throughput Analysis of Massive MIMO Uplink


with Low-Resolution ADCs
Sven Jacobsson, Student Member, IEEE, Giuseppe Durisi, Senior Member, IEEE, Mikael Coldrey, Member, IEEE,
Ulf Gustavsson, and Christoph Studer, Senior Member, IEEE

AbstractWe investigate the uplink throughput achievable BS antenna elements leads to high spatial resolution; this makes
by a multiple-user (MU) massive multiple-input multiple-output it possible to simultaneously serve several user equipments
(MIMO) system in which the base station is equipped with a large (UEs) in the same time-frequency resource, which brings large
number of low-resolution analog-to-digital converters (ADCs).
Our focus is on the case where neither the transmitter nor capacity gains. The improvements in terms of radiated energy
the receiver have any a priori channel state information. This efficiency are enabled by the array gain that is provided by
implies that the fading realizations have to be learned through the large number of antennas.
pilot transmission followed by channel estimation at the receiver, Equipping the BS with a large number of antenna elements,
based on coarsely quantized observations. We propose a novel however, increases considerably the hardware cost and the
channel estimator, based on Bussgangs decomposition, and a
novel approximation to the rate achievable with finite-resolution power consumption of the radio-frequency (RF) circuits [4].
ADCs, both for the case of finite-cardinality constellations and This calls for the use of low-cost and power-efficient hardware
of Gaussian inputs, that is accurate for a broad range of system components, which, however, reduce the signal quality due to
parameters. Through numerical results, we illustrate that, for the an increased level of impairments. The aggregate impact of
1-bit quantized case, pilot-based channel estimation together with hardware impairments on massive MIMO systems has been
maximal-ratio combing or zero-forcing detection enables reliable
multi-user communication with high-order constellations in spite investigated in, e.g., [5][8], where it is found that massive
of the severe nonlinearity introduced by the ADCs. Furthermore, MIMO providesto a certain extentrobustness against signal
we show that the rate achievable in the infinite-resolution (no distortions caused by low-cost RF components. However, most
quantization) case can be approached using ADCs with only of these analyses rely on the assumption that the distortion
a few bits of resolution. We finally investigate the robustness caused by the hardware imperfections can be modeled as an
of low-ADC-resolution MU-MIMO uplink against receive power
imbalances between the different users, caused for example by additive Gaussian random variable that is independent of the
imperfect power control. transmit signal. It is prima facie unclear how accurate such
modeling assumption is, especially for the distortion caused by
Index TermsAnalog-to-digital converter (ADC), channel ca-
pacity, linear minimum mean square error (LMMSE) chan- low-resolution analog-to-digital converters (ADCs). This has
nel estimation, low-resolution quantization, multi-user massive been noted in [8, Sec. IV.A] where it is pointed out that such
multiple-input multiple-output (MIMO). modeling assumption targets ADCs with high resolution.

I. I NTRODUCTION
A. Quantized Massive MIMO
Massive multiple-input multiple-output (MIMO) is a promis-
In this paper, we consider an uplink massive MU-MIMO
ing multi-user (MU) MIMO technology for next generation
system and focus on the signal distortion caused by the use
cellular communication systems (5G) [2]. With massive MIMO,
of low-resolution ADCs at the BS. An ADC with sampling
the number of antennas at the base station (BS) is scaled up
rate fs Hz and a resolution of b bits maps each sample of
by several orders of magnitude compared to traditional multi-
the continuous-time, continuous-amplitude baseband received
antenna systems with the goals of enabling significant gains in
signal to one out of 2b quantization labels, by operating fs 2b
capacity and energy efficiency [2], [3]. Increasing the number of
conversion steps per second. In modern high-speed ADCs (e.g.,
S. Jacobsson is with Ericsson Research and Chalmers University of with sampling rates larger than 1 GS/s), the dissipated power
Technology, Gothenburg, Sweden (e-mail: sven.jacobsson@ericsson.com) scales exponentially in the number of bits and linearly in the
G. Durisi is with Chalmers University of Technology, Gothenburg, Sweden
(e-mail: durisi@chalmers.se) sampling rate [9], [10]. This implies that for wideband massive
M. Coldrey and U. Gustavsson are with Ericsson Research, Gothenburg, MIMO systems where hundreds of high-speed converters are
Sweden (e-mail: {mikael.coldrey,ulf.gustavsson}@ericsson.com) required, the resolution of the ADCs may have to be kept low
C. Studer is with Cornell University, Ithaca, NY (e-mail: studer@cornell.edu)
The work of S. Jacobsson and G. Durisi was supported in part by the in order to maintain the power consumed at the BS within
Swedish Foundation for Strategic Research under grant ID14-0022, and by acceptable levels.
the Swedish Government Agency for Innovation Systems (VINNOVA) within An additional motivation for reducing the ADC resolution
the competence center ChaseOn.
The work of C. Studer was supported in part by Xilinx Inc., and by the is to limit the amount of data that has to be transferred over
US National Science Foundation (NSF) under grants ECCS-1408006, CCF- the link that connects the RF components and the baseband-
1535897, and CAREER CCF-1652065. processing unit. For example, consider a BS that is equipped
The material in this paper was presented in part at the IEEE International
Conference on Communications (ICC) Workshop on 5G and Beyond: Enabling with an antenna array of 500 elements. At each antenna element,
Technologies and Applications, London, U.K., June 2015 [1]. the in-phase and quadrature samples are quantized separately

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
2

using a pair of 10-bit ADCs operating at 1 GS/s. Such a values that are below this threshold, on-off QPSK is capacity
system would produce 10 Tbit/s of data. This exceeds by achieving [19].
far the rate supported by the common public radio interface For the 1-bit quantized MIMO case, the capacity-achieving
(CPRI) used over todays fiber-optical fronthaul links [11]. distribution is unknown. In [20], it is shown that QPSK is
Alleviating this capacity bottleneck is of particular importance optimal at low SNR, again under the assumption of perfect CSI
in a cloud radio access network (C-RAN) architecture [12], at the receiver. Mo and Heath Jr. [21] derived high-SNR bounds
where the baseband processing is migrated from the BSs to a on capacity, and showed that high-order modulations are
centralized unit, which may be placed at a significant distance supported. However, their analysis relies on the assumption that
from the BS antenna array. the transmitter has access to perfect CSI, which is unrealistic in
An implication of lowering the ADC resolution is that low-resolution architectures. Their contribution leaves open the
the requirement on accurate radio-frequency circuitry can be question on whether high-order modulations are supported
relaxed. The reason is that the quantization noise may dominate in training based schemes where the receiver has partial
the noise introduced by other components such as mixers, knowledge of the channel and the transmitter (in our case,
oscillators, filters, and low-noise amplifiers. Hence, further the UE) has no channel knowledge.
power-consumption reductions may be achieved by relaxing Channel estimation on the basis of quantized observations is
the quality requirements on the RF circuitry. considered in, e.g., [22], [23] (see also [24] for a compressive-
The 1-bit resolution case, where the in-phase and quadrature sensing version of this problem). A closed-form solution for the
components of the continuous-valued received samples are maximum likelihood (ML) estimate in the 1-bit case is derived
quantized separately using a pair of 1-bit ADCs, is particularly in [23], under the assumption of time-multiplexed pilots.
attractive because of the resulting low hardware complex- The use of 1-bit ADCs in massive MIMO was considered
ity [13], [14]. Indeed, a 1-bit ADC can be realized using in [25]. There, the authors examined the achievable uplink
only a simple comparator. Furthermore, in a 1-bit architecture, throughput for the scenario where the UEs transmit QPSK
there is no need for automatic gain control circuitry, which is symbols, and the BS employs a least squares (LS) channel
otherwise needed to match the dynamic range of the ADCs. estimator, followed by a maximal ratio combining (MRC)
or zero-forcing (ZF) detector. Their results show that large
sum-rate throughputs can be achieved despite the coarse
B. Previous Work
quantization. The results in [25] were extended to high-order
Receivers employing low-resolution ADCs need to cope with modulations (e.g., 16-QAM) by the authors of this paper in [1].
the severe nonlinearity introduced by the coarse quantization, There, we showed that one can detect not only the phase, but
which may render signaling schemes and receiver algorithms also the amplitude of the transmitted signal, provided that the
developed for the case of high-resolution ADCs suboptimal. number of BS antennas is sufficiently large, hence, answering
The impact of the 1-bit ADC nonlinearity on the performance positively the question left open in [21]. Choi et al. [26]
of communication systems has been previously studied in the recently developed a detector and a channel estimator capable
literature under various channel-model assumptions. In [15], it of supporting high-order constellations such as 16-QAM. Again
is proven that BPSK is capacity achieving over a real-valued for the case of 1-bit ADCs, Li et al. [27], [28] proposed a
nonfading single-input single-output (SISO) Gaussian channel. linear minimum mean square error (LMMSE) channel estimator
For the complex-valued Gaussian channel, QPSK is optimal. based on Bussgangs decomposition that was shown to be
These results hold under the assumption that the 1-bit superior to the one proposed in [26]. Furthermore, they derive
quantizer is a zero-threshold comparator. It turns out that an approximation on the rates achievable with Gaussian inputs.
in the low-SNR regime, a zero-threshold comparator is not The accuracy of this approximation is not fully validated in [28],
optimal [16]. The optimal strategy involves the use of flash- since no comparison with actual achievable rates is provided.
signaling [17, Def. 2] and requires an optimization over the Wen et al. [29] proposed a joint channel- and data-estimation
threshold value. Unfortunately, the power gain obtainable using algorithm that offers significant improvement compared to the
this optimal strategy manifests itself only at extremely low case when channel estimation and data detection are treated
values of spectral efficiency. separately. However, as noted in [29], the complexity of the
For the Rayleigh-fading case, under the assumption that proposed algorithm is too high for practical implementations.
the receiver has access to perfect channel state information A mixed-ADC architecture where many 1-bit ADCs are
(CSI), it is shown in [18] that QPSK is capacity achieving complemented with few high-resolution ADCs is proposed in
(again for the SISO case). The assumption that perfect CSI is [30]. It is found that the addition of a relatively small number
available may, however, be unrealistic in the 1-bit quantized of high-resolution ADCs increases the system performance
case, since the nonlinear distortion caused by the 1-bit ADCs significantly. Specifically, the authors of [30] present an
makes channel estimation challenging. In particular, if the achievability bound under Gaussian signaling and minimum
fading process evolves rapidly, the cost of transmitting training distance decoding that holds for the setup where channel
symbols cannot be neglected. For the more practically relevant estimates are acquired through the high-resolution ADCs. This
case when the channel is not known a priori to the receiver, relies on the assumption that each high-resolution ADC can be
but must be learned (for example, via pilot symbols), QPSK is linked to several RF chains through a switch. The disadvantage
optimal when the SNR exceeds a certain threshold that depends of such architecture is that ADC switches increase hardware
on the coherence time of the fading process [19]. For SNR complexity. Furthermore, the time needed to acquire channel

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
3

estimates increases dramatically. Through a numerical study, we determine the minimum


In all of the contributions reviewed so far, low-resolution ADC resolution needed to make the performance gap to
quantized massive MIMO systems have been investigated the infinite-resolution case negligible. Our simulations
solely for communication over frequency-flat, narrowband, suggest that only few bits (e.g., 3 bits) are required to
channels. A spatial-modulation-based massive MIMO system achieve a performance close to the infinite-resolution case
over a frequency-selective channel was studied in [31]. The for a large range of system parameters. This holds also
proposed receiver employs LS estimation followed by a when the users are received at vastly different power levels
message-passing-based detector. The performance of a low- (imperfect power control).
resolution quantized massive MIMO system using orthogonal This paper complements the analysis previously reported
frequency division multiplexing (OFDM) and operating over in [1] by generalizing it to ZF receivers, to multi-bit quantiza-
a wideband channel was investigated in [32]. There, it is tion, and to the case of imperfect power control. Furthermore,
found that using ADCs with only 4 to 6 bits resolution is the proposed channel estimator and the rate approximations
sufficient to achieve performance close to the infinite-resolution are novel.
(i.e., no quantization) case, at no additional cost in terms of
digital signal processing complexity. A capacity lower bound
D. Notation
for wideband channels and 1-bit ADCs has been recently
reported in [33]. The analysis in [33] relies on the same Lowercase and uppercase boldface letters denote column
signal decomposition used in [27], [28] for the frequency- vectors and matrices, respectively. The identity matrix of size
flat case. However, differently from [27], [28], the temporal N N is denoted by IN . We use tr() and diag() to denote
correlation of the quantization noise in the channel-estimation the trace and the main diagonal of a matrix, and kk to denote
phase is ignored. the `2 -norm of a vector. The multivariate normal distribution
All the results reviewed so far hold under the assumption of with mean and covariance is denoted by N (, ). Further-
Nyquist-rate sampling at the receiver. It is worth pointing out more, the multivariate complex-valued circularly-symmetric
that Nyquist-rate sampling is not optimal in the presence of Gaussian probability density function with zero mean and
quantization at the receiver [34][36]. For example, for the 1-bit covariance is denoted by CN (0, ). The operator Ex []
quantized complex AWGN channel, high-order constellations stands for the expectation over the random variable x. The
such as 16-QAM can be supported even in the SISO case, if mutual information between two random variables x and y is
one allows for oversampling at the receiver [37]. indicated by I(x; y). The real and imaginary parts of a complex
scalar s are <{s} and ={s}. The superscripts T , , and H
C. Contributions denote transpose, complex conjugate, and Hermitian transpose,
Focusing on Nyquist-rate sampling, and on the scenario respectively. The function (x) is the cumulative distribution
where neither the transmitter nor the receiver have a priori function (CDF) of a standard normal random variable.
CSI, we investigate the rates achievable over a frequency-flat
Rayleigh block-fading MU-MIMO channel, when the receiver E. Paper Outline
is equipped with low-resolution ADCs. Our contributions are The rest of the paper is organized as follows. In Section II,
summarized as follows: we introduce the massive MIMO system model and the channel-
We propose a novel channel estimator for the case of multi-
estimation and data-detection problems. In Section III, we
bit ADCs and nonuniform quantization regions using derive an approximation on the rate achievable with finite-
Bussgangs decomposition. This estimator recovers the cardinality constellations and Gaussian inputs. In Section IV,
LMMSE estimator proposed in [27], [28], [33] for the we validate the accuracy of our approximations for different
case of 1-bit ADCs. scenarios and determine the ADC resolution required to
We present a easy-to-evaluate approximation on the rates
approach the rate achievable in the infinite-resolution case.
achievable with finite-cardinality constellations under the We conclude in Section V.
assumption of training-based channel estimation. The
approximation is explicit in the number of pilots used to
II. C HANNEL E STIMATION AND DATA D ETECTION WITH
estimate the channel and in the resolution of the ADCs; by
L OW-R ESOLUTION ADC S
comparing it with a numerically computed lower bound
on the achievable rates, we show that this approximation A. System Model and Sum-Rate Capacity
is accurate for a large range of SNR values. We consider the single-cell uplink system depicted in Fig. 1.
We also obtain a closed-form approximation on the rates Here, K single-antenna users are served by a BS that is
achievable with Gaussian inputs that is derived using equipped with an array of N > K antennas. We model
Bussgangs decomposition. This approximation recovers the subchannels between each transmit-receive antenna pair
for the 1-bit case the approximation recently presented as a Rayleigh block-fading channel (see, e.g., [38]), i.e., a
in [27], [28]. A comparison with a numerically computed channel that stays constant for a block of T channel uses, and
lower bound on the achievable rates reveals that, in the changes independently from block to block. We shall refer to
1-bit case, this Gaussian approximation is accurate at low T as the channel coherence interval. We further assume that
SNR, but overestimates the achievable rate at high SNR the subchannels are mutually independent. The discrete-time
in the multiuser scenario. complex baseband received signal over all antennas within

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
4

Re 2N b-bit ADCs at the BS by the function Qb () : CN RN b


RF Low-resolution ADC

UE 1
Im
Low-resolution ADC that maps the received signal yt with entries {yn,t } to the
Re quantized output rt with entries {rn,t } in the following way:
RF Low-resolution ADC
Im
Low-resolution ADC if <{yn,t } (k , k+1 ] and ={yn,t } (l , l+1 ], then rn,t =
UE 2
... ... `k + j`l . Using this convention, the quantized received signal
... can be written as
Re
RF
Im
Low-resolution ADC rt = Qb (yt ) = Qb (Hxt + wt ) , t = 1, 2, . . . , T. (4)
Low-resolution ADC
UE K
RF
Re
Low-resolution ADC
Finding the optimal quantization labels, i.e., the ones
Im
Low-resolution ADC that minimize the mean square error (MSE) between the
nonquantized received vector yt and the quantized vector rt ,
Fig. 1. Quantized massive MIMO uplink system model. requires one to determine the probability density function (PDF)
of yt . Note that such PDF depends on the choice of the input
constellation xt . Since adapting the quantization labels to the
an arbitrary coherence interval and before quantization is
choice of the input constellation appears to be impractical from
modeled as
an implementation point of view, in this paper we shall consider
yt = Hxt + wt , t = 1, 2, . . . , T. (1) the following suboptimal choice of the set of quantization
labels Lb and thresholds Tb : we first approximate each entry
Here, xt CK denotes the channel input from all users at of the nonquantized channel output vector yt by a complex
time t, and H CN K is the channel matrix connecting the K Gaussian random variable with variance K + 1 and then use
users to the N BS antennas. The entries of H are independent the Lloyd-Max algorithm [43], [44] to find a set of labels
and CN (0, 1)-distributed. Furthermore, the vector wt CN , Lb = {`0 , `1 , . . . , `2b 1 } that minimize the mean square error
whose entries are independent and CN (0, 1)-distributed, stands between the nonquantized and the quantized signal. Then, we
for the AWGN. rescale the labels such that the variance of each entry of rt is
Throughout the paper, we consider the case where CSI is K + 1. Specifically, we multiply each label in the set Lb by
not available a priori to the transmitter or to the receiver, the factor
i.e., they are both not aware of the realization of H. This v
scenario captures the cost of learning the fading channel [39]
u K + 1
=u u 2b 1  q 2  q  (5)
[41], an operation that has to be performed using quantized t P 2 2i+1 2i2
observations and may yield significant performance loss in 2 `i K+1 K+1
i=0
the case of low-resolution ADCs. We further assume that
coding can be performed over many coherence intervals. Let to produce the set of labels Lb = Lb .
X = [x1 , x2 , . . . , xT ] be the K T matrix of transmitted sig- Some comments on our choice are in order. The Gaussian
nals within a coherence interval, and let R = [r1 , r2 , . . . , rT ] approximation is accurate at low SNR or when the number of
be the corresponding N T matrix of received quantized UE is sufficiently large. When such conditions are not fulfilled,
samples. For a given quantization function, the ergodic sum- it may result in a suboptimal choice of the quantization labels.
rate capacity is [38] The rescaling of the labels by in (5) turns out to simplify
1 the performance analysis (see Sections II-D and III-C).
C() =
sup I(X; R). (2) In the 1-bit case, we can write the quantized received signal
T
at the nth antenna, at discrete time t, as follows:
Here, the supremum is over all probability distributions on X r
for which X has independent rows and the following average K + 1
power constraint is satisfied: Q1 (yn,t ) = (sgn(<{yn,t }) + jsgn(={yn,t })) . (6)
2
E tr(XXH ) KT .
 
(3) Here, sgn() denotes the signum function defined as
(
Since the noise variance is normalized to one, we can think of 1, if x < 0
sgn(x) = (7)
as the SNR. The sum-rate capacity in (2) is, in general, not +1, if x 0.
known in closed form, even in the infinite-resolution case, for
which tight capacity bounds have been reported recently in [42].
C. Signal Decomposition using Bussgangs Theorem
B. Quantization of a Complex-Valued Vector
The in-phase and quadrature components of the received The quantization of a vector using finite-resolution ADCs
signal at each antenna are quantized separately by an ADC causes a distortion that is correlated with the input to the
of b-bit resolution. We characterize the ADC by a set of 2b + quantizer. When the input to the quantizer is Gaussian, we
1 quantization thresholds Tb = {0 , 1 , . . . , 2b }, such that can use Bussgangs theorem [45] to decompose the quantized
= 0 < 1 < < 2b = , and a set of 2b quantization signal in the convenient form detailed in the following theorem.
labels Lb = {`0 , `1 , . . . , `2b 1 } where `i (i , i+1 ]. Let Theorem 1: Let r = Qb (y) denote the output from a set
Rb = Lb Lb . We shall describe the joint operation of the of ADCs described by the set of labels Lb and the set of

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
5

thresholds Tb . Assume that y CN (0N , K) where K the nth BS antenna. Furthermore, let Xp = [x1 , . . . , xP ]T
CN N . Then, the quantized vector r can be decomposed as denote the matrix containing the P pilot symbols transmitted
(p) (p)
by the K UEs. Finally, let yn = Xp hn + wn and
r = Gb y + d (8) (p) (p)
rn = Qb (yn ) denote the nonquantized and quantized pilot
where the quantization distortion d and y are uncorrelated. sequences received at the nth antenna during the training phase.
(p)
Furthermore, Gb RN N is the following diagonal matrix:1 Here, wn = [wn,1 , . . . , wn,P ]T is the additive noise. For
2Xb
1  the 1-bit case, the LMMSE estimator of hn was obtained
1/2 `  
in [28]. Proceeding similarly to [28], we generalize the
Gb = diag(K) i exp i2 diag(K)1
i=0
LMMSE estimator [28] to the multi-bit case. Specifically, let
(p) (p)

1
 Cy(p) and Cr(p) be the covariance matrices of yn and rn ,
2 n n
exp i+1 diag(K) . (9) respectively. Using Bussgangs decomposition (8) (recall that
both additive noise and fading are Gaussian) and the fact
Here, `i corresponds to the ith element of the set of labels Lb that diag(Cy(p) ) = (K + 1)IP , which follows from (12)
and i to the ith element of the set of thresholds Tb . n
and implies that Gb = Gb IP (see (9)), we conclude that the
Proof: See Appendix A. LMMSE estimator for the multi-bit case is
Bussgangs theorem has been used previously in the literature
1 (p)
to decompose the quantized signal in the 1-bit-ADC case (see, hn = Gb XH
p C (p) rn . (13)
rn
e.g., [27], [28]). A generalization of this result to the case of
The computation of (13) requires knowledge of the covari-
multi-bit uniform ADCs has been recently proposed in [46] in
ance matrix Cr(p) . For the case of 1-bit ADCs, one can
the context of downlink precoding. The more general result n
compute Cr(p) in closed form, as shown in [28]. For the
in Theorem 1 allows one to handle the case of nonuniform n
multi-bit case, however, Cr(p) is not known in closed form.
quantizers as well. For the special case when diag(K) = n
To overcome this issue, we shall next present an alternative
(K + 1)IN , which will turn out relevant for our analysis in
channel estimator, which is an approximation of (13), but
Section II-D, the matrix Gb in (9) reduces to
admits a simple closed-form expression. Using Bussgangs
Gb = Gb IN (10) decomposition (8), we write Cr(p) as
n

with Cr(p) = G2b Cy(p) + Cd(p) = G2b Xp XH


p + G2b IP + Cd(p) . (14)
n n n n
b
2X 1 
i2
2 
`i K+1
i+1
Here, Cd(p) denotes the covariance matrix of the quantization
Gb = e e K+1
n . (11)
distortion. To simplify (14), we shall next assume that the off-
p
i=0
(K + 1)
diagonal elements of Cd(p) are zero, i.e., we shall ignore the
Note that in the infinite-resolution case (b = ), it follows temporal correlation of the
n
quantizaton distortion. Specifically,
from (8) that G = IN and, hence, G = 1 (see p (10)). we assume that
For the 1-bit-ADCs case, we have that G1 = 2/, a
well-known result used recently in [28], [33] to analyze the Cd(p) = (1 G2b )(K + 1)IP . (15)
n
throughput achievable with 1-bit ADCs. We shall use the The assumption in (15) is accurate in the low-SNR regime or
Bussgang decomposition to develop a channel estimator in when the number of UEs K is large, and it is actually exact if
the next section as well as an approximation on the rates the number of pilots P coincides with the number of UEs K.
achievable with Gaussian inputs in Section III-C. The constant on the right-hand side of (15) follows from the
power normalization (5). Substituting (15) into (14) and (14)
D. Channel Estimation into (13), we obtain
A common approach to transmitting information over fading hn = Gb XH 2 H
p Gb Xp Xp
channels whose realizations are not known a priori to the  1 (p)
receiver is to reserve a certain number of time slots in each + G2b + (1 G2b )(K + 1) IP rn (16)
coherence interval for the transmission of pilot symbols. These = Gb G2b XH
p Xp
pilots are then used at the receiver to estimate the fading  1 H (p)
+ G2b + (1 G2b )(K + 1) IK Xp rn (17)
channel. Assume that P pilot symbols are used in each
coherence interval (K P T ). We shall assume that the = Gb G2b P
pilot sequences used by different UEs are pairwise orthogonal, 1
+ G2b + (1 G2b )(K + 1) XH (p)
p rn . (18)
i.e., that
P
Rewriting (18) in matrix form, we obtain the following
simplified estimator, which we shall use in the remainder of
X
xt xH
t = P IK . (12)
t=1
the paper:
PP
Let hn denote the channel vector whose entries contain Gb t=1 rt xH t
the channel gain between the kth UE, k = 1, . . . , K, and H = 2 . (19)
Gb P + G2b + (1 G2b )(K + 1)
1 We use the convention that the function exp() applied to a diagonal matrix
acts element-wise on its diagonal entries. Some comments on (19) are in order. For the case of 1-

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
6

bit ADCs, the estimator (19) coincides with the one derived 1 1
in [33]. Under the assumption that the number of pilots P is
equal to the number of UEs K, the covariance matrix Cd(p) is

Quadrature
Quadrature
n
indeed diagonal, and the estimator (19) is actually the LMMSE 0 0
estimator (13). This fact has been observed in [28] for the
1-bit case. For the infinite-resolution case (G = 1), (19)
coincides with the classic minimum mean square error (MMSE) 1 1
estimator (see, e.g., [47]). Let H = H + H where H denotes 1 0 1 1 0 1
the estimation error. Under the assumption that (15) holds, the In-phase In-phase
variance of the channel estimate and of the estimation error (a) N = 20 antennas, = 0 dB. (b) N = 200 antennas, = 0 dB.
take the following forms:
1 h  i 1
2 = E tr HHH (20)
NK

Quadrature
G2b P
= 2 (21) 0
Gb P + Gb + (1 G2b )(K + 1)
2

and
1 h  i 1
2 = E tr HHH (22)
NK 1 0 1
G2 + (1 G2b )(K + 1) In-phase
= 2 b . (23)
Gb P + G2b + (1 G2b )(K + 1) (c) N = 200 antennas, = 20 dB.

For the case P = K, (21) and (23) are exact. Note that in the Fig. 2. Single-user MRC outputs for 16-QAM inputs as a function of the
number of receive antennas N and the SNR . The channel estimates are
infinite-resolution case (G = 1), (21) and (23) recover well- based on P = 20 pilot symbols.
known results for MMSE estimation (see, e.g., [47, Eq. (19)]).

estimate is acquired using P = 20 pilots. As the size of


E. Data Detection the BS antenna array increases, the 16-QAM constellation
We shall focus on the practically relevant case when the BS becomes distinguishable (see Fig. 2b), provided that is not
employs a linear receiver. Linear receiver processingalthough too high. Indeed, additive noise is one of the factors that
inferior to nonlinear processing techniques such as successive enables the detection of the 16-QAM constellation. The other
interference cancellationis less computationally demanding is the different phase of the fading coefficients associated with
and has been shown to yield good performance if the number of each receive antenna. The explanation is as follows: in the 1-
antennas exceeds significantly the number of active users [48]. bit ADCs case, the quantized received output at each antenna
We shall consider two types of linear receivers, namely MRC belongs to the set R1 of cardinality 4. These 4 possible outputs
and ZF. Using either of the two methods, a soft estimate xk,t are then averaged by the MRC filter to produce an output (a
of the transmitted symbol xk,t from the kth user at time t = scalar) that belongs to an alphabet with much higher cardinality.
P + 1, P + 2, . . . , T is obtained as follows: The cardinality depends on the number of pilots and on the
H number of receive antennas. The key observation is that the
xk,t = ak rt . (24)
inner points of the 16-QAM constellation, which are more
Here, ak CN denotes the receive filter for the kth user, susceptible to noise, are more likely to be erroneously detected
which is given by at each antenna. This results in a smaller averaged value after
( MRC than for the outer constellation points.
hk /khk k2 , for MRC To highlight the importance of the additive noise, we
ak = (25)
(H )k , for ZF consider in Fig. 2c the case when = 20 dB. Since the
additive noise is negligible, the output of the MRC filter lies
where (H )k is the kth column of the pseudo-inverse of the approximately on a circle, which suggests that the amplitude
channel estimate matrix H = H(HH H)1 . of the transmitted signal cannot be used to convey information.
However, the phase of the 16-QAM symbols can still be
F. High-Order Modulation Formats with 1-bit ADCs: Why detected. Indeed, consider the following argument. At high
Does it Work? SNR and in the single-user case, the signal received at the nth
antenna can be well-approximated by2
Although for 1-bit ADCs, QPSK is optimal in the SISO
case [19], the use of multiple antennas at the receiver opens rn = Q1 (hn x + wn ) Q1 (hn x) = Q1 (ej(n +) ). (26)
up the possibility of using higher-order modulation schemes
to support higher rates. This observation is demonstrated in Here, n and denote the phase of hn and of x, respectively.
Fig. 2 where we plot the MRC receiver output (for 300 Furthermore, again at high SNR, the nth entry an of the MRC
different channel fading realizations) for the case when a 2 In the remainder of this section, we shall drop the time index t and the
single user transmits 16-QAM data symbols. The channel user index k because they are superfluous.

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
7

filter a in (25) is well-approximated by over a rectangular grid in the complex plane. With this
1 1 technique, one obtains a lower bound on R(k) () [52, p. 3503]
Q1 ejn .

an Q1 (hn ) = (27) that becomes increasingly tight as the grid spacing is made
2N 2N
smaller.3 Note that (31) holds for every choice of input
Using (26) and (27), we can approximate the MRC output (24)
distribution and for ADCs with arbitrary resolution.
at high SNR by
N
1 X B. Sum-Rate Approximation for Finite-Cardinality Inputs
Q1 ejn Q1 ej(n +) .
 
x (28)
2N n=1 The evaluation of (31) using the method just described
To analyze (28), let us assume without loss of generality is extremely time consuming. We next provide an accurate
that 0 < < /2. Since n is uniformly distributed on approximation of (31) for finite-cardinality constellations that
[0, 2] (recall that we assumed hn to be Rayleigh dis- is easier to evaluate, although still not in closed form (note that
tributed), one can show that the phase of the random variable even for the infinite-resolution case, no closed-form expression
Q1 ejn Q1 ej(n +) is equal to 0 with probability 1
 for the rate achievable with finite-cardinality constellations
2/ and is equal /2 with probability 2/. Hence, its mean is available). The approximation relies on the following
R
is . Since the fading coefficients {hn }, and, hence, also their assumption: the real part x k = <{x k } and the imaginary
I
phases {n }, are independent, the phase of x in (28) converges part xk = ={xk } of the soft estimate xk of the transmitted
to as N grows large, due to the central limit theorem. symbol xk are conditionally jointly Gaussian given xk and H,
As shown in Fig. 2c, N = 200 antennas are sufficient to with conditional mean (xk , H) and conditional covariance
distinguish the phase of 16-QAM constellation points at 20 dB (x ,
k H). We use this assumption to approximate (29) as

of SNR. Note that independence between the {hn } is crucial follows (see Appendix B):
T P
 
for the central limit theorem to hold and for the phases to be 
R(k) () h xRk , x I
k | H
distinguishable. T
1 h  i
2
III. ACHIEVABLE R ATE A NALYSIS Exk ,H log2 (2e) det (xk , H) . (32)
2
In this section, we shall characterize the rate achievable in Here, h( | ) denotes the conditional differential entropy [53].
a low-resolution quantized massive MIMO uplink system. In Note that under the above Gaussian approximation, the condi-
contrast to [27], [28], [49], [50] we shall mainly focus on finite- tional probability of [xR , xI ]T given H is a Gaussian mixture,
k k
cardinality constellations. Using Bussgangs decomposition, we for which the differential entropy is not known in closed
also provide a closed-form approximation of the achievable rate form but can be computed efficiently. The accuracy of the
with Gaussian inputs, which turns out accurate at low SNR. approximation in (32) depends crucially on the choice of
(xk , H) and (xk , H). In Appendix B, we provide suitable
A. Sum-Rate Lower-Bound for Finite-Cardinality Inputs choices for (xk , H) and (xk , H) for the MRC case (see
(k) (56)(62)). For the ZF case, (xk , H) is provided in (56)
It follows from, e.g., [51], that the achievable rate R ()
whereas, to improve the accuracy of the approximation, we
for user k = 1, 2, . . . , K with pilot-based channel estimation
resort to the numerical method described in Appendix B to
and MRC or ZF detection is
compute (xk , H). As we shall illustrate in Section IV, the
T P
R(k) () = I(xk ; xk | H) (29) resulting approximation (32) turns out to be accurate for all
T system parameters considered in this paper.
where xk and xk are distributed as xk,t and xk,t respectively.
It follows that the sum-rate capacity can be lower-bounded as
C. Sum-Rate Approximation for Gaussian Inputs
follows:
K
Next, we present an approximation on the achievable
C()
X
R(k) (). (30) rate (29) assuming Gaussian inputs. In contrast to [28], [33],
k=1
where a similar approximation is derived for the 1-bit case, we
shall consider the case of multi-bit ADCs.
In order to compute the achievable rate, we expand the mutual The approximation relies on Bussgangs decomposition and
information in (29) as follows: on the assumption that the quantizer input y can be modeled
as a Gaussian random vector4 and that its covariance matrix
" #
Pxk |xk ,H (xk |xk , H)
I(xk ; xk | H) = Exk ,xk ,H log2 . (31) satisfies
Pxk |H (xk |H)
Cy = (K + 1)IN . (33)
To compute (31), one needs the conditional probability
Both the Gaussian assumption and (33) are accurate at low SNR
mass functions Pxk |xk ,H (xk |xk , H) and Pxk |H (xk |H) =
h i or when the number of UEs is large. Under these assumptions,
Exk Pxk |xk ,H (xk |xk , H) . Since no closed-form expressions
3 The numerical routines used to evaluate (29) can be downloaded at https:
are available for these quantities, we estimate them by Monte-
//github.com/infotheorychalmers/1-bit massive MIMO.
Carlo sampling. Specifically, we simulate many noise and 4 To keep notation compact, we drop in this section the time index t since
interference realizations, and map the resulting xk to points it is superfluous.

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
8

we can use Bussgangs theorem to decompose the received 0 0


signal as
r = Qb (y) = Gb y + d (34) 5 5

where d is the quantization distortion. Here, we have used that

MSE [dB]

MSE [dB]
Gb = Gb IN , which follows from (33). Furthermore, due to the b = 1, 2, 3,
10 10
power normalization (5) and due to (33), the covariance matrix
Cr of r satisfies Cr = (K + 1)IN . Hence, the covariance b = 1, 2, 3,
matrix Cd of the quantization distortion d must be equal to 15 15

Cd = Cr G2b Cy = 1 G2b (K + 1) IN .

(35)
20 20
Substituting (34) into (24), we obtain 30 20 10 0 10 30 20 10 0 10
SNR, [dB] SNR, [dB]
xk = aH H H
k (Gb y + d) = Gb ak Hx + ak n (36)
(a) P = K pilots. (b) P = 3K pilots.
where H is the channel estimate (19), and H is the cor- Fig. 3. MSE of the channel estimator (19) as a function of the SNR ;
responding estimation error. Here, we have defined n = QPSK pilots, N = 200, K = 10. The solid lines correspond to the MSE
approximation (23) and the marks correspond to the exact MSE, which was
Gb Hx + Gb w + d. Note that the noise n and the input computed numerically.
vector x are uncorrelated provided that both (33) and (15) hold.
Assuming that this is indeed the case, we can approximate
the mutual information (29) by using the auxiliary channel IV. N UMERICAL R ESULTS
lower bound [52, p. 3503] and treating the additive noise aHk n We now assess the rates achievable with the above detailed
in (36) as a Gaussian random variable. Specifically, let channel estimation and data-detection schemes detailed in the
G2b 2 previous section on a massive MU-MIMO uplink system where
= (37) the receiver is equipped with low-resolution ADCs. We assume
G2b K 2 + G2b + (1 G2b )(K + 1)
that the users are able to coordinate the transmission of their
where 2 and 2 are given by (21) and (23), respectively. pilots: when one of the UEs transmits pilots, the other UEs
In (37), the three terms in the denominator correspond to remain idle. In other words, pilots are transmitted in a round
the estimation error, the additive noise, and the quantization robinPfashion.5 The use of time-interleaved pilots ensures
distortion, respectively. Since the channel input x is Gaussian, P H
that t=1 xt xt = P IK . Also, because of the idle time,
using [52, p. 3503] we obtain the following approximation: each user can transmit its pilots at a power level that is K
times higher than the power level for the data symbols, while

!
H 2
|ak hk | still satisfying the average-power constraint (3).
I(xk ; xk |H) E log2 1 + P . (38)

H 2
|ak hj | +kak k 2
j6=k A. Channel Estimation
Under the additional assumption that H is Gaussian, we can We start by validating the accuracy of the approximation
use [54, Eq. (16) and (20)] to further lower-bound (38) and for the MSE of the simplified channel estimator (19) given
obtain the following closed-form Gaussian approximations for in (23). Specifically, we compare in Fig. 3 the exact MSE of
the rates achievable with MRC and ZF, respectively: the estimator (19), which is evaluated numerically, with the
T P

(N 1)
 approximation (23), for different values of SNR , number of
RMRC () log2 1 + (39) pilots P , and ADC resolution b.
T (K 1) + 1
We note that if P = K (Fig. 3a), then the MSE approxima-
and tion (23) is indeed exact (as we claimed in Section II-D). For
T P the case of P = 3K (Fig. 3b), the approximation (23) turns
RZF () log2 (1 + (N K)) . (40) out to be accurate at low SNR. Furthermore, the accuracy
T
Here, we have multiplied the log terms by (T P )/T to of (23) increases with the resolution of the ADCs. Indeed, (23)
take into account the pilot overhead. Note that for the infinite- relies on the assumption that the off-diagonal elements of the
resolution case (G = 1), we recover from (39) and (40) the covariance matrix in (15) are zero and these entries vanish as
achievable rate with imperfect CSI reported in [54, Eq. (39)] the ADC resolution increases (see, e.g., [55, p. 541] for more
and [54, Eq. (42)] for the MRC and ZF receiver, respectively. details).
p
For the case of 1-bit ADCs (G1 = 2/), we recover
from (40) the achievable rate approximation with ZF recently B. Achievable Rate
reported in [28]. 1) Single-User Case, 1-bit ADCs: In Fig. 4 we compare
for the single-user 1-bit ADC case, the rates achievable with
As we shall demonstrate in Section IV, despite the several QPSK, 16-QAM, and 64-QAM as a function of for the
assumptions invoked to obtain (39) and (40), these approxima- 5 This pilot-transmission method is chosen for convenience; it may be
tions turn out to be accurate in the low-SNR regime. suboptimal.

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
9

6 6 6
64-QAM

Rate per user [bits/channel use]

Rate per user [bits/channel use]


16-QAM inf. res.
QPSK inf. res.
Rate [bits/channel use]

4 4 4

inf. res.
2 2 2
1-bit ADCs 1-bit ADCs
1-bit ADCs 64-QAM 64-QAM
16-QAM 16-QAM
0 0 0
30 20 10 0 10 20 30 30 20 10 0 10 30 20 10 0 10
SNR, [dB] SNR, [dB] SNR, [dB]
(a) MRC receiver. (b) ZF receiver.
Fig. 4. Single-user achievable rate with MRC as a function of the SNR ; N =
200, K = 1, T = 1142; the number of pilots P is optimized for each value Fig. 5. Per-user achievable rate as a function of the SNR ; N = 200,
of . The solid lines correspond to the finite-cardinality approximation (32), K = 10, T = 1142; the number of pilots P is optimized for each value of
the dashed lines corresponds to the Gaussian approximation (39), and the . The solid lines correspond to the finite-cardinality approximation (32), the
marks correspond to the rates computed via (29) and (31). dashed lines corresponds to the Gaussian approximations (39), (40), and the
marks correspond to the rates computed via (29) and (31).

MRC receiver.6 We depict both the rates achievable with 6


1-bit ADCs and the ones for the infinite-resolution case.
The rates with 1-bit ADCs, which are computed using (29) Rate per user [bits/channel use]

and (31), are compared with the approximation for finite-


4
cardinality constellations provided in (32) and the Gaussian
approximation (39) to verify their accuracy. The infinite- inf. res.
resolution rates are computed using (29) and (31) and are also
compared with the Gaussian approximation (39) The number 2
of receive antennas is N = 200 and the coherence interval is
T = 1142.7 The number of transmitted pilots P is numerically 1-bit ADCs
optimized for every value of . We see that, despite using 1-bit
0 0
ADCs, higher-order modulations outperform QPSK already at 10 101 102 103 104 105
SNR values as low as = 15 dB. Coherence interval, T
Note that the achievable rate does not increase monotonically
with in the 16-QAM and 64-QAM case. Indeed, as gets Fig. 6. Per-user achievable rate with 64-QAM and ZF as a function of T ;
= 10 dB, N = 200, K = 10; the number of pilots P is optimized for each
large the constellation gets projected onto the unit circle and value of T . The solid lines correspond to the finite-cardinality approximation
the number of distinguishable constellation points becomes (32), the dashed lines corresponds to the Gaussian approximation (40), and
smaller (see Fig. 2c). Note also that the approximation (32) for the marks correspond to the rates computed via (29) and (31).
finite-cardinality constellations closely tracks the simulation
results for all SNR values. This approximation enables us
to accurately predict the SNR value beyond which the rates 2) Multi-User Case, 1-bit ADCs: In Fig. 5, we plot the
achievable with a given constellation saturates. This, in turn, rates achievable with MRC and ZF for both the 1-bit-ADC
allows us to identify the most appropriate constellation for a and the infinite-resolution case when K = 10 users are active.
given SNR value. Motivated by the results in Fig. 4, we only compare the rates
The Gaussian approximation (39) tracks the rates achievable achievable with 16-QAM and 64-QAM. Note again that the
with finite-cardinality constellations accurately in the low- approximation (32) turns out to be accurate for a all SNR
SNR regime. values, whereas the Gaussian approximation is accurate only
We note that, when QPSK is used, the difference in the at low SNR. Note also that rates with 16-QAM and 64-QAM
achievable rates between the 1-bit quantized case and the saturate at the same level at high SNR for both MRC and
infinite-resolution case is marginalan observation that was ZF. This implies that the system is effectively distortion and
already reported in [25]. In contrast, the rate loss is more interference limited, and that the Gaussian approximations (39),
pronounced for higher-order constellations. (40) overestimate the rate for high SNR values.
3) Dependence on the Coherence Interval: In Fig. 6, we
6 To evaluate the mutual information (31), we have simulated 300 random plot the per-user achievable rates with ZF, as a function of the
fading channel realizations. For each channel realization we have considered coherence interval T for = 10 dB, N = 200, K = 10, and
3000 random noise realizations for each symbol in the constellation. 64-QAM constellation. We observe that the reduction in the
7 For an LTE-like system operating at 2 GHz, with symbol time equal to
66.7 s, and with UEs moving at a speed of 3 km/h, the duration of the coher- achievable rate when T is made smaller is similar for both the
ence interval according to Jakes model is approximately T = 1142 symbols. 1-bit and infinite-resolution case. Hence, operating in a high-

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
10

6 TABLE I
S UMMARY OF SIMULATION PARAMETERS
Rate per user [bits/channel use]

Description Assumption
4
Cell layout Circular cell
b = 1, 2, 3,
Cell radius 335 meters
Minimum distance between UE and BS 35 meters
2
Path loss 35 + 35 log10 (d) dB
Number of BS antennas (N ) 200 antennas
Number of single-antenna users (K) 10 users
Coherence interval (T ) 1142 channel uses
0 Number of pilots per user (P/K) 10 pilots per user
30 25 20 15 10 5 0 5 10
Carrier frequency 2 GHz
SNR, [dB]
System bandwidth 20 MHz
UE transmit power 8.5 dBm
Fig. 7. Per-user achievable rate with 64-QAM and ZF as a function of
the SNR ; N = 200, K = 10, T = 1142; the number of pilots
Noise spectral efficiency 174.2 dBm/Hz
P is optimized for each value of . The solid lines correspond to the Noise figure 5 dB
finite-cardinality approximation (32), the dashed lines corresponds to the
Gaussian approximation (40), and the marks correspond to the rates computed
via (29) and (31).
4 4

10% worst throughput [bits/channel use]

10% worst throughput [bits/channel use]


mobility scenario leads to similar performance losses in both
cases. Note also that the achievable rate is zero when T 10. 3 3
In fact, when orthogonal pilot sequences are transmitted, at
least 10 pilot symbols are required when K = 10.
4) Dependence on ADC Resolution: Focusing on 64-QAM 2 2
b = 1, 2, 3,
and ZF, we compare in Fig. 7 the achievable rate as a function b = 1, 2, 3,
of the ADC resolution and the SNR. We observe that with 2-bit
1 1
ADCs, the achievable rate increases significantly compared to
the 1-bit-ADC case. For example, at = 10 dB, we achieve
90% of the infinite-resolution rate, compared to 71% with 1-bit 0 0
ADCs. Increasing the ADC resolution beyond 3 bits seems 0 25 50 75 100 125 150 0 25 50 75 100 125 150
d [meters] d [meters]
unnecessary for the system parameters considered in Fig. 7.
This conclusion is supported by both the approximation for (a) MRC receiver. (b) ZF receiver.
finite-cardinality constellations and the one for Gaussian inputs. Fig. 8. The 10% worst throughput with 16-QAM for a user located d1 = 185
We note that the Gaussian approximation (40) is again accurate meters away from the BS as a function of d for the parameters specified
in Table I.
at low SNR. Furthermore, as expected its accuracy increases
with the ADC resolution.
185 meters from the BS, results in a SNR of approximately
C. Impact of Large-Scale Fading and Imperfect Power Control 1 = 10 dB. The remaining K 1 users in the cell are
So far, we have considered only the case when all users randomly dropped according to a uniform distribution on the
operate at the same average SNR. This corresponds to the circular ring of inner radius d1 d meters and outer radius
scenario where perfect power control can be performed in d1 + d meters, for a distance spread 0 < d < 150 meters.
the uplink, which is clearly favorable for low-resolution ADC The case d = 0 corresponds to the scenario when power
architectures. If, however, the received signal powers are vastly control is executed perfectly. The case d = 150 meters
different, low-power signals may not be distinguishable from corresponds to the worst-case scenario of uncoordinated uplink
high-power interferers for cases in which the ADCs resolution transmission, where no power control is performed by the UEs.
is too low. In the latter case, the SNR for each interfering user lies in the
In practical systems, large spreads in the received power range [19.0 dB, 15.3 dB].
is typically avoided through power control. However, perfect In Fig. 8, we plot the 10% worst throughput (i.e., the
power control may be impossible to achieve in practice due throughput corresponding to the 10% point of the CDF of
to limitations on the UE transmit power, for example. We throughputs), for the intended user located d1 = 185 meters
next investigate how relaxing the accuracy of the UE transmit away from the BS, as a function of d. We focus on 16-
power control will impact the system performance. We consider QAM and assume that the received signal power level for
a single-cell scenario and adapt the urban-macro path loss each user is known to the BS. To attain the curves, we have
model in [56]. The simulation parameters for this study are considered 103 random interfering user drops for each d
summarized in Table I. The transmit power for all UEs is value. As expected, the gap to the infinite-resolution rate grows
set to 8.5 dBm, which for the first user that is located d1 = as d increases. In the uncoordinated case, with 1-bit ADCs

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
11

and ZF, we attain 57% of the rate achievable with perfect We obtain (9) by substituting (44) in (42) and by using
power control. The corresponding number for the 3-bit-ADC that n2 = [K]n,n .
case is 79%. This shows that high rates are achievable with
low-resolution ADCs even in absence of power control.
V. C ONCLUSIONS A PPENDIX B
D ERIVATION OF (32)
We have analyzed the performance of a low-resolution
quantized uplink massive MU-MIMO system operating over
a frequency flat Rayleigh block-fading channel whose realiza- To keep the notation compact, we set xR I
k = <{xk } and xk =
R I R I
tions are not known a priori to transmitter and receiver. In ={xk }. By Rletting Ian,k = an,k + jan,k , hn,k = hn,k + jhn,k ,
particular, we have shown that for the 1-bit massive MIMO and rn = rn + jrn , where an,k denotes the nth entry of the
case, high-order constellations, such as 16-QAM, can be used receive filter ak and rn the nth entry of the received vector r,
to convey information at higher rates than with QPSK; this we can express the real components of the received signal as
holds in spite of the nonlinearity introduced by the 1-bit XN XN

xR aR R I I

ADCs. Furthermore, reliable communication can be achieved k = <{a n,k r n } = n,k rn + an,k rn . (45)
by using simple signal processing techniques at the receiver, n=1 n=1
i.e., pilot-based channel estimation based on the Bussgang Similarly, for the imaginary part we can write
decomposition (19) and MRC detection. By increasing the N N
resolution of the ADCs by only a few bits, e.g., to 3 bits, we xIk =
X
={an,k rn } =
X
aR I I R

n,k rn an,k rn . (46)
can achieve near infinite-resolution performance for a broad n=1 n=1
range of system parameters; furthermore, the system becomes
robust against differences in the received signal power from Now, R
we collect the real and imaginary components in a vector
I T
the different users, due for example, to large-scale fading or [x k , x k ] and approximate their conditional distribution given

imperfect power control. the channel input and the channel estimate as a bivariate Gaus-
An extension of our analysis to a OFDM based setup for sian random vector with mean (xk , H) and 2 2 covariance
transmission over frequency-selective channels is currently matrix (x k , H). Under this assumption, we have that
 
under investigation. Such an extension could be used to I(x ; x | H) = h xR , xI | H
k k k k
benchmark the results recently reported in [32] in which the
authors reported that, with a specific modulation and coding 1 h 
2
i
Exk ,H log2 (2e) det (xk , H) . (47)
scheme taken from IEEE 802.11n, 4 to 6 bits are required to 2
2
achieve a packet error rate below 10 at an SNR close to It is worth emphasizing that, differently from (31), the dif-
the one needed in the infinite-resolution case. We conclude ferential entropy h(xR I
k , xk | H) in (47) is evaluated under the
R I T
that for a fair comparison between the performance attainable assumption that [xk , xk ] is conditionally Gaussian given xk
using low-resolution versus high-resolution ADCs, one should and H. The conditional probability of [xR I T
k , xk ] given H is a
take into account the overall power consumption, including Gaussian mixture. The achievable rate in (32) follows from (47)
the power consumed by RF and baseband processing circuitry. by taking into account the rate loss due to the transmission of
P pilot symbols to estimate the channel. We shall next discuss
A PPENDIX A how to choose (xk , H) and (xk , H).
P ROOF OF T HEOREM 1
We start by finding a suitable approximation for the proba-
It follows from Bussgangs theorem [45] that bility mass functions of the random variables rnR and rnI . For
R
E ryH = Gb E yyH (41) rn , it holds that
   

pR
 R
where Gb is a N N diagonal matrix with n,i = Pr rn = `i (48)
 R  R
1 = Pr rn < i+1 Pr rn i (49)
[Gb ]n.n = 2 E[Qb (yn )yn ] . (42) R R
n ( i+1 ) ( i ) (50)
Here, yn denotes
h the i nth entry of the vector y, n = 1, . . . , N , where Pin the last step we have approximated the interference
2
and n2 = E |yn | = [K]n,n . It follows from (41) that we term j6=k <{hn,j xj } P
by a zero-mean Gaussian random
2
can write the quantized signal as r = Gb y + d, where d and variable with variance j6=k |hn,j | and defined
y are uncorrelated. Note now that the quantizer output Qb (yn ) v  
is equal to `i + j`i if and only if <{yn } [i , i+1 ) and
u
u 2 hR xR + hI xI
i n,k k n,k k
={yn } [j , j+1 ). Thus, iR = t
u
P 2 . (51)
1 + j6=k |hn,j |
Z  2
2Qb (z)z z
Eyn [Qb (yn )yn ] = exp 2 dz (43) For the single-user case, the approximation (50) is exact since
n n
there is no interference. Proceeding in an analogous way, we
can show that
L1 2
!
X `i n i22
i+1

= 2
e n e n . (44)
i=0
pIn,i (i+1
I
) (iI ) (52)

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
12

where Specifically, we find (xk , H) by simulating several random


v 
u  noise and interference realizations for each point in the
u 2 hR xI hI xR symbol constellation. Obtaining a sufficiently accurate estimate
i n,k k n,k k
iI = t
u
P 2 . (53) of (xk , H) requires orders of magnitude fewer noise and
1 + j6=k |hn,j |
interference realizations compared to estimating the probability
mass functions in (31).
Next, we use these approximations to derive (xk , H) and
(xk , H). The conditional mean of xR
k given both xk and H
can be written as ACKNOWLEDGEMENTS
h i XN N The authors would like to thank Dr. Fredrik Athley at
 X
E xR
 R R I I
k | x k , H = E a r
n,k n + a r
n,k n = mRn,k (54)
Ericsson Research for fruitful discussions.
n=1 n=1
P2b 1  R R  R EFERENCES
where mRn,k = i=0 `i a p
n,k n,i + aI
pI
n,k n,i . Similarly, for
[1] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer, One-
the imaginary part it holds that bit massive MIMO: Channel estimation and high-order modulations, in
N Proc. IEEE Int. Conf. Commun. Workshop (ICCW), London, U.K., June
2015, pp. 13041309.
h i X
E xIk | xk , H = mIn,k (55) [2] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, Massive
n=1 MIMO for next generation wireless systems, IEEE Commun. Mag.,
P2b 1  R I  vol. 52, no. 2, pp. 186195, Feb. 2014.
where mIn,k = I R
i=0 `i an,k pn,i an,k pn,i . The sought-
[3] T. L. Marzetta, Noncooperative cellular wireless with unlimited numbers
of base station antennas, IEEE Trans. Wireless Commun., vol. 9, no. 11,
after mean vector can thus be written as pp. 35903600, Nov. 2010.
N 2Xb
1  R R [4] H. Yang and T. L. Marzetta, Total energy efficiency of cellular large
an,k pn,i + aIn,k pIn,i

scale antenna system multiple access mobile networks, in Proc. IEEE
X
(xk , H) = `i R I . (56)
an,k pn,i aIn,k pRn,i
Online Conf. Green Commun. (OnlineGreenComm), Piscataway, NJ, Oct.
n=1 i=0 2013, pp. 2732.
[5] U. Gustavsson, C. Sanchez-Perez, T. Eriksson, F. Athley, G. Durisi,
We next move to (xk , H). Assuming that the received P. Landin, K. Hausmair, C. Fager, and L. Svensson, On the impact
signal is conditionally uncorrelated over the antenna array, we of hardware impairments on massive MIMO, in Proc. IEEE Global
Telecommun. Conf. (GLOBECOM), Austin, TX, Dec. 2014, pp. 294300.
obtain that [6] X. Zhang, M. Matthaiou, E. Bjornson, M. Coldrey, and M. Debbah, On
R 2
h i h i
the MIMO capacity with residual transceiver hardware impairments, in
= E xR

(xk , H) k k | xk , H (57) Proc. IEEE Int. Conf. Commun. (ICC), Sydney, Australia, Jun. 2014, pp.
1,1
b b
52995305.
N 2X
X 1 2X
1 [7] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, Massive MIMO
= pR I
n,i pn,j
systems with non-ideal hardware: Energy efficiency, estimation, and
n=1 i=0 j=0
capacity limits, IEEE Trans. Inf. Theory, vol. 11, no. 60, pp. 7112
2 7139, Nov. 2014.
aR I R
n,k `i + an,k `j mn,k . (58) [8] E. Bjornson, M. Matthaiou, and M. Debbah, Massive MIMO with non-
ideal arbitrary arrays: Hardware scaling laws and circuit-aware design,
Analogously, it holds that IEEE Trans. Wireless Commun., vol. 14, no. 8, pp. 43534368, Apr.
2015.
h 2 i
[9] R. H. Walden, Analog-to-digital converter survey and analysis, IEEE
[(xk , H)]2,2 = E xIk Ik | xk , H (59) J. Sel. Areas Commun., vol. 17, no. 4, pp. 539550, Apr. 1999.
b b [10] B. Murmann, ADC performance survey 1997-2016. [Online]. Available:
N 2X
1 2X
1
X http://web.stanford.edu/murmann/adcsurvey.html
= pR I
n,i pn,j [11] Ericsson AB, Huawei Technologies, NEC Corporation, Alcatel Lucent,
n=1 i=0 j=0 and Nokia Siemens Networks, Common public radio interface (CPRI);
2 Interface Specification, CPRI specification v6.0, Aug. 2013.
aR I I
n,k `j an,k `i mn,k . (60) [12] S.-H. Park, O. Simeone, O. Sahin, and S. Shamai (Shitz), Fronthaul
compression for cloud radio access networks, IEEE Signal Process.
Furthermore, Mag., vol. 31, no. 6, pp. 6979, Nov. 2014.
h i [13] I. D. ODonnell and R. W. Brodersen, An ultra-wideband transceiver
[(xk , H)]1,2 = E xR R
 I
xk Ik | xk , H

architecture for low power, low rate, wireless systems, IEEE Trans. Veh.
k k
Technol., vol. 54, no. 5, pp. 16231631, Sep. 2005.
N 2X
1 2Xb
1 b [14] S. Hoyos, B. M. Sadler, and G. R. Arce, Monobit digital receivers for
ultrawideband communications, IEEE Trans. Wireless Commun., vol. 4,
X
= pR I
n,i pn,j no. 4, pp. 13371344, Jul. 2005.
n=1 i=0 j=0 [15] J. Singh, O. Dabeer, and U. Madhow, On the limits of communication
with low-precision analog-to-digital conversion at the receiver, IEEE
aR I
mR

n,k `i + an,k `j n,k Trans. Commun., vol. 57, no. 12, pp. 36293639, Dec. 2009.
[16] T. Koch and A. Lapidoth, At low SNR, asymmetric quantizers are
aR I I

n,k `j an,k `i mn,k . (61) better, IEEE Trans. Inf. Theory, vol. 59, no. 9, pp. 54215445, Sep.
2013.
Finally, because of symmetry, [17] S. Verdu, Spectral efficiency in the wideband regime, IEEE Trans. Inf.
Theory, vol. 48, no. 6, pp. 13191343, Jun. 2002.
[(xk , H)]2,1 = [(xk , H)]1,2 . (62) [18] S. Krone and G. Fettweis, Fading channels with 1-bit output quantiza-
tion: Optimal modulation, ergodic capacity and outage probability, in
For the ZF receiver, computing the covariance using (58) IEEE Inf. Theory Workshop (ITW), Dublin, Ireland, Aug. 2010.
[19] A. Mezghani and J. A. Nossek, Analysis of Rayleigh-fading channels
(62) does not yield a satisfactory approximation. Therefore, with 1-bit quantized output, in Proc. IEEE Int. Symp. Inf. Theory (ISIT),
we resort to Monte-Carlo simulations to obtain the covariance. Toronto, ON, Jul. 2008, pp. 260264.

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
13

[20] , On ultra-wideband MIMO systems with 1-bit quantized outputs: [45] J. J. Bussgang, Crosscorrelation functions of amplitude-distorted Gaus-
Performance analysis and input optimization, in Proc. IEEE Int. Symp. sian signals, Res. Lab. Elec., Cambridge, MA, Tech. Rep. 216, Mar.
Inf. Theory (ISIT), Nice, France, Jun. 2007, pp. 12861289. 1952.
[21] J. Mo and R. W. Heath Jr., Capacity analysis of one-bit quantized [46] S. Jacobsson, G. Durisi, M. Coldrey, T. Goldstein, and C. Studer,
MIMO systems with transmitter channel state information, IEEE Trans. Quantized precoding for massive MU-MIMO, Oct. 2016. [Online].
Signal Process., vol. 63, no. 20, pp. 54985512, Oct 2015. Available: https://arxiv.org/abs/1610.07564
[22] T. M. Lok and V. K.-W. Wei, Channel estimation with quantized [47] B. Hassibi and B. M. Hochwald, How much training is needed in
observations, in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Cambridge, multiple-antenna wireless links? IEEE Trans. Inf. Theory, vol. 49, no. 4,
MA, Aug. 1998, p. 333. pp. 951963, Apr. 2003.
[23] M. T. Ivrlac and J. A. Nossek, On MIMO channel estimation with [48] E. Bjornson, E. G. Larsson, and T. L. Marzetta, Massive MIMO: 10
single-bit quantization, in Int. ITG Workshop on Smart Antennas (WSA), myths and one critical question, IEEE Commun. Mag., vol. 54, no. 2,
Vienna, Austria, Feb. 2007. pp. 114123, Aug. 2016.
[24] A. Zymnis, S. Boyd, and E. Candes, Compressed sensing with quantized [49] J. Zhang, L. Dai, S. Sun, and Z. Wang, On the spectral efficiency of
measurements, IEEE Signal Process. Lett., vol. 17, no. 2, pp. 149152, massive MIMO systems with low-resolution ADCs, IEEE Commun.
Feb. 2010. Lett., vol. 20, no. 5, pp. 842 845, Dec. 2015.
[25] C. Risi, D. Persson, and E. G. Larsson, Massive MIMO with 1-bit [50] L. Fan, S. Jin, C.-K. Wen, and H. Zhang, Uplink achievable rate for
ADC, Apr. 2014. [Online]. Available: http://arxiv.org/abs/1404.7736 massive MIMO systems with low-resolution ADC, IEEE Commun. Lett.,
vol. 19, no. 12, pp. 2186 2189, Dec. 2015.
[26] J. Choi, J. Mo, and R. W. Heath Jr., Near maximum-likelihood detector
[51] L. Tong, B. M. Sadler, and M. Dong, Pilot-assisted wireless transmis-
and channel estimator for uplink multiuser massive MIMO systems with
sions: general model, design criteria, and signal processing, IEEE Signal
one-bit ADCs, IEEE Trans. Commun., vol. 64, no. 5, pp. 20052018,
Process. Mag., vol. 21, no. 6, pp. 1225, Nov. 2004.
May 2016.
[52] D. M. Arnold, H.-A. Loeliger, P. O. Vontobel, A. Kavcic, and W. Zeng,
[27] Y. Li, C. Tao, L. Liu, G. Seco-Granados, and A. L. Swindlehurst, Simulation-based computation of information rates for channels with
Channel estimation and uplink achievable rates in one-bit massive memory, IEEE Trans. Inf. Theory, vol. 52, no. 8, pp. 34983508, Aug.
MIMO systems, in IEEE Sensor Array and Multichannel Signal Process. 2006.
Workshop (SAM), Rio de Janeiro, Brazil, Jul. 2016. [53] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed.
[28] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, Wiley, 2006.
and L. Liu, Channel estimation and performance analysis of [54] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, Energy and spectral
one-bit massive MIMO systems, Mar. 2017. [Online]. Available: efficiency of very large multiuser MIMO systems, IEEE Trans. Commun.,
https://arxiv.org/abs/1609.07427 vol. 61, no. 4, pp. 1436 1449, April 2013.
[29] C.-K. Wen, C.-J. Wang, S. Jin, K.-K. Wong, and P. Ting, Bayes-optimal [55] B. Widrow and I. Kollar, Quantization Noise: Roundoff Error in
joint channel-and-data estimation for massive MIMO with low-precision Digital Computation, Signal Processing, Control, and Communications.
ADCs, IEEE Trans. Signal Process., vol. 64, no. 10, pp. 25412556, Cambridge, UK: Cambridge Univ. Press, 2008.
Jul. 2015. [56] 3GPP, Spatial channel model for multiple input multiple output (MIMO)
[30] N. Liang and W. Zhang, Mixed-ADC massive MIMO, IEEE J. Sel. simulations, Tech. Rep. 25.996 ver. 12.0.0 rel. 12, Sep. 2014.
Areas Commun., vol. 34, no. 4, pp. 983997, May 2016.
[31] S. Wang, Y. Li, and J. Wang, Multiuser detection in massive spatial
modulation MIMO with low-resolution ADCs, IEEE Trans. Wireless
Commun., pp. 21562168, Dec. 2014.
[32] C. Studer and G. Durisi, Quantized massive MU-MIMO-OFDM uplink,
IEEE Trans. Commun., vol. 64, no. 6, pp. 23872399, Jun. 2016. Sven Jacobsson received the M.Sc. degree in com-
[33] C. Mollen, J. Choi, E. G. Larsson, and R. W. Heath Jr., Uplink munication engineering from Chalmers University of
performance of wideband massive MIMO with one-bit ADCs, IEEE Technology, Gothenburg, Sweden, in 2015. He joined
Trans. Wireless Commun., vol. 16, no. 1, pp. 87100, Oct. 2016. Ericsson Research in 2015 and is currently pursuing
[34] S. Shamai (Shitz), Information rates by oversampling the sign of a the Ph.D. degree at the Department of Electrical
bandlimited process, IEEE Trans. Inf. Theory, vol. 40, no. 4, pp. 1230 Engineering, Chalmers University of Technology,
1236, Jul. 1994. Gothenburg Sweden, as an industrial Ph.D. candi-
[35] T. Koch and A. Lapidoth, Increased capacity per unit-cost by oversam- date. His research interests are within the areas of
pling, in Proc. IEEE 26th Conv. Electrical and Electronics Engineers advanced antenna systems and hardware-constrained
in Israel (IEEEI), Eliat, Israel, Nov. 2010, pp. 684688. communications, with a special emphasis on the
impact of coarse quantization.
[36] W. Zhang, A general framework for transmission with transceiver
distortion and some applications, IEEE Trans. Commun., vol. 60, no. 2,
pp. 384399, Feb. 2012.
[37] S. Krone and G. Fettweis, Capacity of communications channels with 1-
bit quantization and oversampling at the receiver, in Proc. IEEE Sarnoff
Symp. (SARNOFF), Newark, NJ, May 2012. Giuseppe Durisi (S02M06SM12) received the
[38] T. L. Marzetta and B. M. Hochwald, Capacity of a mobile multiple- Laurea degree summa cum laude and the Doctor
antenna communication link in Rayleigh flat fading, IEEE Trans. Inf. degree both from Politecnico di Torino, Italy, in
Theory, vol. 45, no. 1, pp. 139157, Jan. 1999. 2001 and 2006, respectively. From 2006 to 2010 he
[39] A. Lapidoth, On the asymptotic capacity of stationary Gaussian fading was a postdoctoral researcher at ETH Zurich, Zurich,
channels, IEEE Trans. Inf. Theory, vol. 51, no. 2, pp. 437446, Feb. Switzerland. In 2010, he joined Chalmers University
2005. of Technology, Gothenburg, Sweden, where he is
[40] G. Durisi, T. Koch, J. Ostman, Y. Polyanskiy, and W. Yang, Short-packet now associate professor and co-director of Chalmers
communications over multiple-antenna Rayleigh-fading channels, IEEE information and communication technology Area of
Trans. Commun., vol. 64, no. 2, pp. 618629, Feb. 2016. Advance. He is also guest researcher at Ericsson,
[41] W. Yang, G. Durisi, and E. Riegler, On the capacity of large-MIMO Sweden. Dr. Durisi is a senior member of the IEEE.
block-fading channels, IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. He is the recipient of the 2013 IEEE ComSoc Best Young Researcher Award
117132, Feb. 2013. for the Europe, Middle East, and Africa Region, and is co-author of a paper
[42] R. Devassy, G. Durisi, J. Ostman, W. Yang, T. Eftimov, and Z. Utkovski, that won a student paper award at the 2012 International Symposium on
Finite-SNR bounds on the sum-rate capacity of Rayleigh block-fading Information Theory, and of a paper that won the 2013 IEEE Sweden VT-
multiple-access channels with no a priori CSI, IEEE Trans. Commun., COM-IT joint chapter best student conference paper award. In 2015, he joined
vol. 63, no. 10, pp. 36213632, Oct. 2015. the editorial board of the IEEE Transactions on Communications as associate
editor. From 2011 to 2014, he served as publications editor for the IEEE
[43] J. Max, Quantizing for minimum distortion, IRE Trans. Inf. Theory,
Transactions on Information Theory. His research interests are in the areas of
vol. 6, no. 1, pp. 712, Mar. 1960.
communication and information theory.
[44] S. P. Lloyd, Least squares quantization in PCM, IEEE Trans. Inf.
Theory, vol. 28, no. 2, pp. 129137, Mar. 1982.

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2691318, IEEE
Transactions on Wireless Communications
14

Mikael Coldrey received the M.Sc. degree in applied Christoph Studer (S06M10SM14) received
physics and electrical engineering from Linkoping his Ph.D. degree in Information Technology and
University, Linkoping, Sweden, in 2000 and the Electrical Engineering from ETH Zurich in 2009. In
Ph.D. degree in electrical engineering from Chalmers 2005, he was a Visiting Researcher with the Smart
University of Technology, Gothenburg, Sweden, in Antennas Research Group at Stanford University.
2006. He joined Ericsson Research in 2006, where From 2006 to 2009, he was a Research Assistant
he is currently a Master Researcher. He has been in both the Integrated Systems Laboratory and the
working with 4G research and for several years Communication Technology Laboratory (CTL) at
with 5G research. His main research interests in- ETH Zurich. From 2009 to 2012, Dr. Studer was a
clude advanced antenna systems, channels, models, Postdoctoral Researcher at CTL, ETH Zurich, and the
algorithms, and millimeter-wave communications for Digital Signal Processing Group at Rice University.
both radio access and wireless backhaul systems. Since 2012, he has been an In 2013, he has held the position of Research Scientist at Rice University. Since
Adjunct Associate Professor with Chalmers University of Technology. 2014, Dr. Studer is an Assistant Professor at Cornell University and an Adjunct
Assistant Professor at Rice University. Dr. Studers research interests include
the design of very large-scale integration (VLSI) circuits, as well as wireless
communications, signal and image processing, and convex optimization.
Dr. Studer received ETH Medals for his M.S. and Ph.D. theses in 2006 and
Ulf Gustavsson received the M.Sc. degree in electri- 2009, respectively. He received a Swiss National Science Foundation fellowship
cal engineering from Orebro University, Sweden, in for Advanced Researchers in 2011 and a US National Science Foundation
2006 and a Ph.D. degree from Chalmers University CAREER Award in 2017. Dr. Studer won a Michael Tien 72 Excellence in
of Technology, Sweden, in 2011. Dr. Gustavsson Teaching Award from the College of Engineering, Cornell University, in 2016.
is currently a senior researcher with Ericsson Re- He shared the Swisscom/ICTnet Innovations Award in both 2010 and 2013.
search, Gothenburg, where his main interests lie Dr. Studer was the winner of the Student Paper Contest of the 2007 Asilomar
in radio signal processing and behavioral modeling Conf. on Signals, Systems, and Computers, received a Best Student Paper
of radio hardware for advanced antenna systems. Award of the 2008 IEEE Int. Symp. on Circuits and Systems (ISCAS), and
Dr. Gustavsson is currently the lead scientist for shared the best Live Demonstration Award at the IEEE ISCAS in 2013.
Ericsson Research within the Marie Skodowska-
Curie Innovative Training Network, SILIKA.

1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.