Professional Documents
Culture Documents
Throughput Analysis of Massive MIMO Uplink
Throughput Analysis of Massive MIMO Uplink
6, JUNE 2017
An additional motivation for reducing the ADC resolution case, since the nonlinear distortion caused by the 1-bit ADCs
is to limit the amount of data that has to be transferred makes channel estimation challenging. In particular, if the
over the link that connects the RF components and the fading process evolves rapidly, the cost of transmitting training
baseband-processing unit. For example, consider a BS that symbols cannot be neglected. For the more practically relevant
is equipped with an antenna array of 500 elements. At each case when the channel is not known a priori to the receiver,
antenna element, the in-phase and quadrature samples are but must be learned (for example, via pilot symbols), QPSK
quantized separately using a pair of 10-bit ADCs operating is optimal when the SNR exceeds a certain threshold that
at 1 GS/s. Such a system would produce 10 Tbit/s of data. depends on the coherence time of the fading process [19].
This exceeds by far the rate supported by the common public For SNR values that are below this threshold, on-off QPSK is
radio interface (CPRI) used over today’s fiber-optical fronthaul capacity achieving [19].
links [11]. Alleviating this capacity bottleneck is of particular For the 1-bit quantized MIMO case, the capacity-achieving
importance in a cloud radio access network (C-RAN) architec- distribution is unknown. In [20], it is shown that QPSK is
ture [12], where the baseband processing is migrated from the optimal at low SNR, again under the assumption of perfect
BSs to a centralized unit, which may be placed at a significant CSI at the receiver. Mo and Heath Jr. [21] derived high-SNR
distance from the BS antenna array. bounds on capacity, and showed that high-order modulations
An implication of lowering the ADC resolution is that are supported. However, their analysis relies on the assump-
the requirement on accurate radio-frequency circuitry can tion that the transmitter has access to perfect CSI, which is
be relaxed. The reason is that the quantization noise may unrealistic in low-resolution architectures. Their contribution
dominate the noise introduced by other components such as leaves open the question on whether high-order modulations
mixers, oscillators, filters, and low-noise amplifiers. Hence, are supported in training based schemes where the receiver
further power-consumption reductions may be achieved by has partial knowledge of the channel and the transmitter (in
relaxing the quality requirements on the RF circuitry. our case, the UE) has no channel knowledge.
The 1-bit resolution case, where the in-phase and quadrature Channel estimation on the basis of quantized observations is
components of the continuous-valued received samples are considered in, e.g., [22], [23] (see also [24] for a compressive-
quantized separately using a pair of 1-bit ADCs, is particularly sensing version of this problem). A closed-form solution for
attractive because of the resulting low hardware complex- the maximum likelihood (ML) estimate in the 1-bit case is
ity [13], [14]. Indeed, a 1-bit ADC can be realized using derived in [23], under the assumption of time-multiplexed
only a simple comparator. Furthermore, in a 1-bit architecture, pilots.
there is no need for automatic gain control circuitry, which is The use of 1-bit ADCs in massive MIMO was consid-
otherwise needed to match the dynamic range of the ADCs. ered in [25]. There, the authors examined the achievable
uplink throughput for the scenario where the UEs transmit
QPSK symbols, and the BS employs a least squares (LS)
B. Previous Work channel estimator, followed by a maximal ratio combin-
Receivers employing low-resolution ADCs need to cope ing (MRC) or zero-forcing (ZF) detector. Their results show
with the severe nonlinearity introduced by the coarse quantiza- that large sum-rate throughputs can be achieved despite the
tion, which may render signaling schemes and receiver algo- coarse quantization. The results in [25] were extended to
rithms developed for the case of high-resolution ADCs sub- high-order modulations (e.g., 16-QAM) by the authors of this
optimal. paper in [1]. There, we showed that one can detect not only
The impact of the 1-bit ADC nonlinearity on the perfor- the phase, but also the amplitude of the transmitted signal,
mance of communication systems has been previously studied provided that the number of BS antennas is sufficiently large,
in the literature under various channel-model assumptions. hence, answering positively the question left open in [21].
In [15], it is proven that BPSK is capacity achieving over Choi et al. [26] recently developed a detector and a chan-
a real-valued nonfading single-input single-output (SISO) nel estimator capable of supporting high-order constella-
Gaussian channel. For the complex-valued Gaussian channel, tions such as 16-QAM. Again for the case of 1-bit ADCs,
QPSK is optimal. Li et al. [27], [28] proposed a linear minimum mean
These results hold under the assumption that the 1-bit square error (LMMSE) channel estimator based on Bussgang’s
quantizer is a zero-threshold comparator. It turns out that decomposition that was shown to be superior to the one pro-
in the low-SNR regime, a zero-threshold comparator is not posed in [26]. Furthermore, they derive an approximation on
optimal [16]. The optimal strategy involves the use of flash- the rates achievable with Gaussian inputs. The accuracy of this
signaling [17, Definition 2] and requires an optimization over approximation is not fully validated in [28], since no compar-
the threshold value. Unfortunately, the power gain obtainable ison with actual achievable rates is provided. Wen et al. [29]
using this optimal strategy manifests itself only at extremely proposed a joint channel- and data-estimation algorithm that
low values of spectral efficiency. offers significant improvement compared to the case when
For the Rayleigh-fading case, under the assumption that the channel estimation and data detection are treated separately.
receiver has access to perfect channel state information (CSI), However, as noted in [29], the complexity of the proposed
it is shown in [18] that QPSK is capacity achieving (again algorithm is too high for practical implementations.
for the SISO case). The assumption that perfect CSI is A mixed-ADC architecture where many 1-bit ADCs are
available may, however, be unrealistic in the 1-bit quantized complemented with few high-resolution ADCs is proposed
4040 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 6, JUNE 2017
in [30]. It is found that the addition of a relatively small num- lower bound on the achievable rates, we show that this
ber of high-resolution ADCs increases the system performance approximation is accurate for a large range of SNR
significantly. Specifically, Liang and Zhang [30] present an values.
achievability bound under Gaussian signaling and minimum • We also obtain a closed-form approximation on the rates
distance decoding that holds for the setup where channel achievable with Gaussian inputs that is derived using
estimates are acquired through the high-resolution ADCs. This Bussgang’s decomposition. This approximation recovers
relies on the assumption that each high-resolution ADC can be for the 1-bit case the approximation recently presented
linked to several RF chains through a switch. The disadvantage in [27] and [28]. A comparison with a numerically
of such architecture is that ADC switches increase hardware computed lower bound on the achievable rates reveals
complexity. Furthermore, the time needed to acquire channel that, in the 1-bit case, this Gaussian approximation is
estimates increases dramatically. accurate at low SNR, but overestimates the achievable
In all of the contributions reviewed so far, low-resolution rate at high SNR in the multiuser scenario.
quantized massive MIMO systems have been investigated • Through a numerical study, we determine the minimum
solely for communication over frequency-flat, narrowband, ADC resolution needed to make the performance gap
channels. A spatial-modulation-based massive MIMO sys- to the infinite-resolution case negligible. Our simulations
tem over a frequency-selective channel was studied in [31]. suggest that only few bits (e.g., 3 bits) are required to
The proposed receiver employs LS estimation followed by a achieve a performance close to the infinite-resolution case
message-passing-based detector. The performance of a low- for a large range of system parameters. This holds also
resolution quantized massive MIMO system using orthogonal when the users are received at vastly different power
frequency division multiplexing (OFDM) and operating over a levels (imperfect power control).
wideband channel was investigated in [32]. There, it is found This paper complements the analysis previously reported
that using ADCs with only 4 to 6 bits resolution is sufficient in [1] by generalizing it to ZF receivers, to multi-bit quantiza-
to achieve performance close to the infinite-resolution (i.e., tion, and to the case of imperfect power control. Furthermore,
no quantization) case, at no additional cost in terms of the proposed channel estimator and the rate approximations
digital signal processing complexity. A capacity lower bound are novel.
for wideband channels and 1-bit ADCs has been recently
reported in [33]. The analysis in [33] relies on the same
signal decomposition used in [27] and [28] for the frequency- D. Notation
flat case. However, differently from [27], [28], the temporal
correlation of the quantization noise in the channel-estimation Lowercase and uppercase boldface letters denote column
phase is ignored. vectors and matrices, respectively. The identity matrix of
All the results reviewed so far hold under the assumption size N × N is denoted by I N . We use tr(·) and diag(·) to
of Nyquist-rate sampling at the receiver. It is worth pointing denote the trace and the main diagonal of a matrix, and
out that Nyquist-rate sampling is not optimal in the presence · to denote the 2 -norm of a vector. The multivariate
of quantization at the receiver [34]–[36]. For example, for the normal distribution with mean μ and covariance is denoted
1-bit quantized complex AWGN channel, high-order constel- by N (μ, ). Furthermore, the multivariate complex-valued
lations such as 16-QAM can be supported even in the SISO circularly-symmetric Gaussian probability density function
case, if one allows for oversampling at the receiver [37]. with zero mean and covariance is denoted by CN (0, ). The
operator Ex [·] stands for the expectation over the random vari-
able x. The mutual information between two random variables
C. Contributions x and y is indicated by I (x; y). The real and imaginary parts
Focusing on Nyquist-rate sampling, and on the scenario of a complex scalar s are {s} and {s}. The superscripts T ,
∗ , and H denote transpose, complex conjugate, and Hermitian
where neither the transmitter nor the receiver have a priori
CSI, we investigate the rates achievable over a frequency-flat transpose, respectively. The function (x) is the cumulative
Rayleigh block-fading MU-MIMO channel, when the receiver distribution function (CDF) of a standard normal random
is equipped with low-resolution ADCs. Our contributions are variable.
summarized as follows:
• We propose a novel channel estimator for the case of
E. Paper Outline
multi-bit ADCs and nonuniform quantization regions
using Bussgang’s decomposition. This estimator recovers The rest of the paper is organized as follows. In Section II,
the LMMSE estimator proposed in [27], [28], and [33] we introduce the massive MIMO system model and the
for the case of 1-bit ADCs. channel-estimation and data-detection problems. In Section III,
• We present a easy-to-evaluate approximation on the rates we derive an approximation on the rate achievable with finite-
achievable with finite-cardinality constellations under the cardinality constellations and Gaussian inputs. In Section IV,
assumption of training-based channel estimation. The we validate the accuracy of our approximations for differ-
approximation is explicit in the number of pilots used ent scenarios and determine the ADC resolution required to
to estimate the channel and in the resolution of the approach the rate achievable in the infinite-resolution case.
ADCs; by comparing it with a numerically computed We conclude in Section V.
JACOBSSON et al.: THROUGHPUT ANALYSIS OF MASSIVE MIMO UPLINK WITH LOW-RESOLUTION ADCs 4041
b −1
2
2 2
Cr( p) = G 2b Cy( p) + Cd( p) = G 2b X p X H + G 2b I P + Cd( p) . (14)
p
i τ
− i
τ
− i+1 n n n n
Gb = √ e Kρ+1 − e Kρ+1 . (11)
i=0
π(Kρ + 1) Here, Cd( p) denotes the covariance matrix of the quantization
n
distortion. To simplify (14), we shall next assume that the off-
Note that in the infinite-resolution case (b = ∞), it follows diagonal elements of Cd( p) are zero, i.e., we shall ignore the
from (8) that G∞ = I N and, hence, G ∞ =√1 (see (10)). n
temporal correlation of the quantizaton distortion. Specifically,
For the 1-bit-ADCs case, we have that G 1 = 2/π, a well- we assume that
known result used recently in [28] and [33] to analyze the
throughput achievable with 1-bit ADCs. We shall use the Cd( p) = (1 − G 2b )(Kρ + 1)I P . (15)
n
Bussgang decomposition to develop a channel estimator in
the next section as well as an approximation on the rates The assumption in (15) is accurate in the low-SNR
achievable with Gaussian inputs in Section III-C. regime or when the number of UEs K is large, and it is actually
exact if the number of pilots P coincides with the number of
1 We use the convention that the function exp(·) applied to a diagonal matrix UEs K . The constant on the right-hand side of (15) follows
acts element-wise on its diagonal entries. from the power normalization (5). Substituting (15) into (14)
JACOBSSON et al.: THROUGHPUT ANALYSIS OF MASSIVE MIMO UPLINK WITH LOW-RESOLUTION ADCs 4043
ĥn = G b X H 2 H
p GbX p X p
−1
( p)
+ G 2b + (1 − G 2b )(Kρ + 1) I P rn (16)
= G b G 2b X H
p Xp
−1
( p)
+ G 2b + (1 − G 2b )(Kρ + 1) I K XH p rn (17)
= G b G 2b Pρ
−1
( p)
+ G 2b + (1 − G 2b )(Kρ + 1) p rn .
XH (18)
outputs are then averaged by the MRC filter to produce an A. Sum-Rate Lower-Bound for Finite-Cardinality Inputs
output (a scalar) that belongs to an alphabet with much higher It follows from, e.g., [51], that the achievable rate R (k) (ρ)
cardinality. The cardinality depends on the number of pilots for user k = 1, 2, . . . , K with pilot-based channel estimation
and on the number of receive antennas. The key observation is and MRC or ZF detection is
that the inner points of the 16-QAM constellation, which are
T−P
more susceptible to noise, are more likely to be erroneously R (k) (ρ) = I (x k ; x̂ k | Ĥ) (29)
detected at each antenna. This results in a smaller averaged T
value after MRC than for the outer constellation points. where x k and x̂ k are distributed as x k,t and x̂ k,t respectively.
To highlight the importance of the additive noise, we con- It follows that the sum-rate capacity can be lower-bounded as
sider in Fig. 2c the case when ρ = 20 dB. Since the follows:
additive noise is negligible, the output of the MRC filter lies
K
approximately on a circle, which suggests that the amplitude C(ρ) ≥ R (k) (ρ). (30)
of the transmitted signal cannot be used to convey information. k=1
However, the phase of the 16-QAM symbols can still be In order to compute the achievable rate, we expand the mutual
detected. Indeed, consider the following argument. At high information in (29) as follows:
SNR and in the single-user case, the signal received at the nth
antenna can be well-approximated by2 Px̂k |xk ,Ĥ (x̂ k |x k , Ĥ)
I (x k ; x̂ k | Ĥ) = Exk ,x̂k ,Ĥ log2 . (31)
rn = Q 1 (h n x + wn ) ≈ Q 1 (h n x) = Q 1 (e j (φn +θ) ). (26) Px̂k |Ĥ (x̂ k |Ĥ)
To compute (31), one needs the conditional probabil-
Here, φn and θ denote the phase of h n and of x, respectively.
ity mass functions Px̂k |xk ,Ĥ (x̂ k |x k , Ĥ) and Px̂k |Ĥ (x̂ k |Ĥ) =
Furthermore, again at high SNR, the nth entry an of the MRC
filter a in (25) is well-approximated by Exk Px̂k |xk ,Ĥ (x̂ k |x k , Ĥ) . Since no closed-form expressions
are available for these quantities, we estimate them by
1 1
an ≈ Q 1 (h n ) = Q 1 e j φn . (27) Monte-Carlo sampling. Specifically, we simulate many noise
2N 2N and interference realizations, and map the resulting x̂ k to
Using (26) and (27), we can approximate the MRC output (24) points over a rectangular grid in the complex plane. With
at high SNR by this technique, one obtains a lower bound on R (k) (ρ) [52,
p. 3503] that becomes increasingly tight as the grid spacing is
1 − j φn j (φn +θ)
N
made smaller.3 Note that (31) holds for every choice of input
x̂ ≈ Q1 e Q1 e . (28) distribution and for ADCs with arbitrary resolution.
2N
n=1
To analyze (28), let us assume without loss of generality B. Sum-Rate Approximation for Finite-Cardinality Inputs
that 0 < θ < π/2. Since φn is uniformly distributed on
[0, 2π] (recall that we assumed h n to be Rayleigh distrib- The evaluation of (31) using the method just described
uted), one is extremely time consuming. We next provide an accurate
can show that the phase of the random vari- approximation of (31) for finite-cardinality constellations that
able Q 1 e− j φn Q 1 e j (φn +θ) is equal to 0 with probability
1 − 2θ/π and is equal π/2 with probability 2θ/π. Hence, its is easier to evaluate, although still not in closed form (note that
mean is θ . Since the fading coefficients {h n }, and, hence, also even for the infinite-resolution case, no closed-form expression
their phases {φn }, are independent, the phase of x̂ in (28) con- for the rate achievable with finite-cardinality constellations
verges to θ as N grows large, due to the central limit theorem. is available). The approximation relies on the following
As shown in Fig. 2c, N = 200 antennas are sufficient to assumption: the real part x̂ kR = {x̂ k } and the imaginary
distinguish the phase of 16-QAM constellation points at 20 dB part x̂ kI = {x̂ k } of the soft estimate x̂ k of the transmitted
of SNR. Note that independence between the {h n } is crucial symbol x k are conditionally jointly Gaussian given x k and Ĥ,
for the central limit theorem to hold and for the phases to be with conditional mean μ(x k , Ĥ) and conditional covariance
distinguishable. (x k , Ĥ). We use this assumption to approximate (29) as
follows (see Appendix B):
T−P
R I
III. ACHIEVABLE R ATE A NALYSIS R (k) (ρ) ≈ h x̂ k , x̂ k | Ĥ
T
In this section, we shall characterize the rate achievable
1
in a low-resolution quantized massive MIMO uplink system. − Exk ,Ĥ log2 (2πe)2 det (x k , Ĥ) . (32)
In contrast to [27], [28], [49], [50] we shall mainly focus on 2
finite-cardinality constellations. Using Bussgang’s decompo- Here, h(· | ·) denotes the conditional differential entropy [53].
sition, we also provide a closed-form approximation of the Note that under the above Gaussian approximation, the condi-
achievable rate with Gaussian inputs, which turns out accurate tional probability of [x̂ kR , x̂ kI ]T given Ĥ is a Gaussian mixture,
at low SNR. for which the differential entropy is not known in closed form
2 In the remainder of this section, we shall drop the time index t and the 3 The numerical routines used to evaluate (29) can be downloaded at
user index k because they are superfluous. https://github.com/infotheorychalmers/one-bit_massive_MIMO.
JACOBSSON et al.: THROUGHPUT ANALYSIS OF MASSIVE MIMO UPLINK WITH LOW-RESOLUTION ADCs 4045
but can be computed efficiently. The accuracy of the approx- distortion, respectively. Since the channel input x is Gaussian,
imation in (32) depends crucially on the choice of μ(x k , Ĥ) using [52, p. 3503] we obtain the following approximation:
and (x k , Ĥ). In Appendix B, we provide suitable choices for ⎡ ⎤
μ(x k , Ĥ) and (x k , Ĥ) for the MRC case (see (56)–(62)). For ⎢ ρ̄|ak ĥk |
H 2
⎥
the ZF case, μ(x k , Ĥ) is provided in (56) whereas, to improve I (x k ; x̂ k |Ĥ) ≈ E ⎣log2 1 + ⎦.
ρ̄ |akH ĥ j |2 + ak 2
the accuracy of the approximation, we resort to the numerical j =k
method described in Appendix B to compute (x k , Ĥ). As we (38)
shall illustrate in Section IV, the resulting approximation (32)
turns out to be accurate for all system parameters considered Under the additional assumption that Ĥ is Gaussian, we can
in this paper. use [54, eqs. (16) and (20)] to further lower-bound (38) and
obtain the following closed-form Gaussian approximations for
the rates achievable with MRC and ZF, respectively:
C. Sum-Rate Approximation for Gaussian Inputs
T−P (N − 1)ρ̄
Next, we present an approximation on the achievable RMRC (ρ̄) ≈ log2 1 + (39)
T (K − 1)ρ̄ + 1
rate (29) assuming Gaussian inputs. In contrast to [28], [33],
where a similar approximation is derived for the 1-bit case, and
we shall consider the case of multi-bit ADCs. T−P
RZF (ρ̄) ≈ log2 (1 + (N − K )ρ̄) . (40)
The approximation relies on Bussgang’s decomposition and T
on the assumption that the quantizer input y can be modeled Here, we have multiplied the log terms by (T − P)/T to
as a Gaussian random vector4 and that its covariance matrix take into account the pilot overhead. Note that for the infinite-
satisfies resolution case (G ∞ = 1), we recover from (39) and (40) the
Cy = (Kρ + 1)I N . (33) achievable rate with imperfect CSI reported in [54, eq. (39)]
and [54, eq. (42)] for the MRC and ZF receiver,
√ respectively.
Both the Gaussian assumption and (33) are accurate at low For the case of 1-bit ADCs (G 1 = 2/π), we recover
SNR or when the number of UEs is large. Under these
from (40) the achievable rate approximation with ZF recently
assumptions, we can use Bussgang’s theorem to decompose reported in [28].
the received signal as As we shall demonstrate in Section IV, despite the several
r = Q b (y) = G b y + d (34) assumptions invoked to obtain (39) and (40), these approxi-
mations turn out to be accurate in the low-SNR regime.
where d is the quantization distortion. Here, we have used that
Gb = G b I N , which follows from (33). Furthermore, due to the IV. N UMERICAL R ESULTS
power normalization (5) and due to (33), the covariance matrix
We now assess the rates achievable with the above detailed
Cr of r satisfies Cr = (Kρ + 1)I N . Hence, the covariance
channel estimation and data-detection schemes detailed in
matrix Cd of the quantization distortion d must be equal to
the previous section on a massive MU-MIMO uplink system
where the receiver is equipped with low-resolution ADCs.
Cd = Cr − G 2b Cy = 1 − G 2b (Kρ + 1) I N . (35)
We assume that the users are able to coordinate the trans-
Substituting (34) into (24), we obtain mission of their pilots: when one of the UEs transmits pilots,
the other UEs remain idle. In other words, pilots are trans-
x̂ k = akH (G b y + d) = G b akH Ĥx + akH n (36) mitted in a round robin
P fashion.
5 The use of time-interleaved
TABLE I
S UMMARY OF SIMULATION PARAMETERS
Fig. 7. Per-user achievable rate with 64-QAM and ZF as a function C. Impact of Large-Scale Fading and Imperfect Power
of the SNR ρ; N = 200, K = 10, T = 1142; the number of pilots Control
P is optimized for each value of ρ. The solid lines correspond to the
finite-cardinality approximation (32), the dashed lines corresponds to the So far, we have considered only the case when all users
Gaussian approximation (40), and the marks correspond to the rates computed operate at the same average SNR. This corresponds to the
via (29) and (31).
scenario where perfect power control can be performed in
the uplink, which is clearly favorable for low-resolution ADC
architectures. If, however, the received signal powers are vastly
values, whereas the Gaussian approximation is accurate only at different, low-power signals may not be distinguishable from
low SNR. Note also that rates with 16-QAM and 64-QAM sat- high-power interferers for cases in which the ADCs resolution
urate at the same level at high SNR for both MRC and ZF. This is too low.
implies that the system is effectively distortion and interfer- In practical systems, large spreads in the received power
ence limited, and that the Gaussian approximations (39), (40) is typically avoided through power control. However, perfect
overestimate the rate for high SNR values. power control may be impossible to achieve in practice
3) Dependence on the Coherence Interval: In Fig. 6, due to limitations on the UE transmit power, for example.
we plot the per-user achievable rates with ZF, as a func- We next investigate how relaxing the accuracy of the UE
tion of the coherence interval T for ρ = −10 dB, transmit power control will impact the system performance.
N = 200, K = 10, and 64-QAM constellation. We We consider a single-cell scenario and adapt the urban-macro
observe that the reduction in the achievable rate when T path loss model in [56]. The simulation parameters for this
is made smaller is similar for both the 1-bit and infinite- study are summarized in Table I. The transmit power for
resolution case. Hence, operating in a high-mobility sce- all UEs is set to 8.5 dBm, which for the first user that is
nario leads to similar performance losses in both cases. located d1 = 185 meters from the BS, results in a SNR of
Note also that the achievable rate is zero when T ≤ 10. approximately ρ1 = −10 dB. The remaining K − 1 users
In fact, when orthogonal pilot sequences are transmitted, in the cell are randomly dropped according to a uniform
at least 10 pilot symbols are required when K = 10. distribution on the circular ring of inner radius d1 − d
4048 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 6, JUNE 2017
Fig. 8. The 10% worst throughput with 16-QAM for a user located d1 = 185 A PPENDIX A
meters away from the BS as a function of d for the parameters specified
in Table I.
P ROOF OF T HEOREM 1
It follows from Bussgang’s theorem [45] that
meters and outer radius d1 + d meters, for a distance spread E ry H = Gb E yy H (41)
0 < d < 150 meters. The case d = 0 corresponds to the
scenario when power control is executed perfectly. The case where Gb is a N × N diagonal matrix with
d = 150 meters corresponds to the worst-case scenario of 1
[Gb ]n.n = 2 E Q b (yn )yn∗ . (42)
uncoordinated uplink transmission, where no power control is σn
performed by the UEs. In the latter case, the SNR for each
interfering user lies in the range [−19.0 dB, 15.3 dB].
Here, yn denotes
the nth entry of the vector y, n = 1, . . . , N,
and σn2 = E |yn |2 = [K]n,n . It follows from (41) that we can
In Fig. 8, we plot the 10% worst throughput (i.e.,
write the quantized signal as r = Gb y + d, where d and y
the throughput corresponding to the 10% point of the CDF
are uncorrelated. Note now that the quantizer output Q b (yn )
of throughputs), for the intended user located d1 = 185
is equal to i + j i if and only if {yn } ∈ [τi , τi+1 ) and
meters away from the BS, as a function of d. We focus on
{yn } ∈ [τ j , τ j +1 ). Thus,
16-QAM and assume that the received signal power level for
∞ 2
each user is known to the BS. To attain the curves, we have ∗
2 Q b (z)z z
considered 103 random interfering user drops for each d E yn Q b (yn )yn = √ exp − 2 dz (43)
−∞ σ n π σ
value. As expected, the gap to the infinite-resolution rate grows ⎛ n⎞
τi2 τ2
as d increases. In the uncoordinated case, with 1-bit ADCs i σn − 2
L−1
− 2i+1
= √ ⎝e σn − e σn ⎠ . (44)
and ZF, we attain 57% of the rate achievable with perfect π
i=0
power control. The corresponding number for the 3-bit-ADC
case is 79%. This shows that high rates are achievable with We obtain (9) by substituting (44) in (42) and by using
low-resolution ADCs even in absence of power control. that σn2 = [K]n,n .
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 To keep the notation compact, we set x̂ kR = {x̂ k } and
a frequency flat Rayleigh block-fading channel whose real- = {x̂ k }. By letting an,k = an,k
x̂ kI R + jaI , h
n,k n,k = h n,k + j h n,k ,
R I
izations are not known a priori to transmitter and receiver. and rn = rn + jrn , where an,k denotes the nth entry of the
R I
In particular, we have shown that for the 1-bit massive MIMO receive filter ak and rn the nth entry of the received vector r,
case, high-order constellations, such as 16-QAM, can be used we can express the real components of the received signal as
to convey information at higher rates than with QPSK; this
N N
∗
holds in spite of the nonlinearity introduced by the 1-bit x̂ kR = {an,k rn } = R R
an,k rn + an,k
I
rnI . (45)
ADCs. Furthermore, reliable communication can be achieved n=1 n=1
by using simple signal processing techniques at the receiver, Similarly, for the imaginary part we can write
i.e., pilot-based channel estimation based on the Bussgang
N N
decomposition (19) and MRC detection. By increasing the ∗
resolution of the ADCs by only a few bits, e.g., to 3 bits, x̂ kI = {an,k rn } = R I
an,k I
rn − an,k rnR . (46)
we can achieve near infinite-resolution performance for a n=1 n=1
broad range of system parameters; furthermore, the system Now, we collect the real and imaginary components in a vector
becomes robust against differences in the received signal [x̂ kR , x̂ kI ]T and approximate their conditional distribution given
JACOBSSON et al.: THROUGHPUT ANALYSIS OF MASSIVE MIMO UPLINK WITH LOW-RESOLUTION ADCs 4049
2b −1
R I
the channel input and the channel estimate as a bivariate I
where m n,k = i=0 i a p
n,k n,i − a I p R . The sought-
n,k n,i
Gaussian random vector with mean μ(x k , Ĥ) and 2×2 covari- after mean vector can thus be written as
ance matrix (x k , Ĥ). Under this assumption, we have that
−1
N 2 b % R R I pI &
an,k pn,i + an,k
I (x k ; x̂ k | Ĥ) = h x̂ kR , x̂ kI | Ĥ μ(x k , Ĥ) = i R pI − a I pR
n,i . (56)
an,k
1
n=1 i=0 n,i n,k n,i
− E xk ,Ĥ log2 (2πe)2 det (x k , Ĥ) .
2 We next move to (x k , H). Assuming that the received
(47) signal is conditionally uncorrelated over the antenna array,
we obtain that
It is worth emphasizing that, differently from (31), the dif-
ferential entropy h(x̂ kR , x̂ kI | Ĥ) in (47) is evaluated under the %
2 &
assumption that [x̂ kR , x̂ kI ]T is conditionally Gaussian given x k (x k , Ĥ) = E x̂ kR − μkR | x k , Ĥ
1,1
and Ĥ. The conditional probability of [x̂ kR , x̂ kI ]T given Ĥ −1 2
−1 b b
is a Gaussian mixture. The achievable rate in (32) follows
N 2
= R I
pn,i pn, j (57)
from (47) by taking into account the rate loss due to the
n=1 i=0 j =0
transmission of P pilot symbols to estimate the channel.
2
We shall next discuss how to choose μ(x k , Ĥ) and (x k , Ĥ). R
× an,k I
i + an,k R
j − m n,k . (58)
We start by finding a suitable approximation for the proba-
bility mass functions of the random variables rnR and rnI . For Analogously, it holds that
rnR , it holds that %
&
2
R
pn,i = Pr rnR = i (48) [(x k , H)]2,2 = E x̂ kI − μkI | x k , Ĥ
−1 2
N 2 −1 b b
= Pr rnR < τi+1 − Pr rnR ≤ τi (49) R I
= pn,i pn, j (59)
R
≈ (ζi+1 ) − (ζiR ) (50) n=1 i=0 j =0
2
wherein the last step we have approximated the interference × an,k
R
j − an,k
I
i − m n,k
I
. (60)
term j =k {h n, j x j } by a zero-mean Gaussian random vari-
! !2
able with variance ρ j =k !h n, j ! and defined Furthermore,
"
# [(x k , H)]1,2 = E x̂ kR − μkR x̂ kI − μkI | x k , Ĥ
# 2 τi − h R x R + h I x I
# n,k k n,k k
ζiR = $ ! !2 . (51) −1 2
N 2 −1b b
1 + ρ j =k h n, j !
!
= R I
pn,i pn, j
n=1 i=0 j =0
For the single-user case, the approximation (50) is exact since
there is no interference. Proceeding in an analogous way, × an,k
R
i + an,k
I
j − m n,k
R
[49] J. Zhang, L. Dai, S. Sun, and Z. Wang, “On the spectral efficiency of Mikael Coldrey received the M.Sc. degree in
massive MIMO systems with low-resolution ADCs,” IEEE Commun. applied physics and electrical engineering from
Lett., vol. 20, no. 5, pp. 842–845, May 2016. Linköping University, Linköping, Sweden, in 2000,
[50] L. Fan, S. Jin, C.-K. Wen, and H. Zhang, “Uplink achievable rate for and the Ph.D. degree in electrical engineering from
massive MIMO systems with low-resolution ADC,” IEEE Commun. Chalmers University of Technology, Gothenburg,
Lett., vol. 19, no. 12, pp. 2186–2189, Dec. 2015. Sweden, in 2006. He joined Ericsson AB in 2006,
[51] L. Tong, B. M. Sadler, and M. Dong, “Pilot-assisted wireless trans- where he is currently a Master Researcher. He
missions: General model, design criteria, and signal processing,” has been involved in 4G research and for sev-
IEEE Signal Process. Mag., vol. 21, no. 6, pp. 12–25, Nov. 2004. eral years with 5G research. Since 2012, he has
[52] D. M. Arnold, H.-A. Loeliger, P. O. Vontobel, A. Kavcic, and W. Zeng, been an Adjunct Associate Professor with Chalmers
“Simulation-based computation of information rates for channels with University of Technology. His main research inter-
memory,” IEEE Trans. Inf. Theory, vol. 52, no. 8, pp. 3498–3508, ests include advanced antenna systems, channels, models, algorithms, and
Aug. 2006. millimeter-wave communications for both radio access and wireless backhaul
[53] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. systems.
New York, NY, USA: Wiley, 2006.
[54] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi-
ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
[55] B. Widrow and I. Kollár, Quantization Noise: Roundoff Error in Ulf Gustavsson received the M.Sc. degree in electri-
Digital Computation, Signal Processing, Control, and Communications. cal engineering from Örebro University, Sweden, in
Cambridge, U.K.: Cambridge Univ. Press, 2008. 2006, and the Ph.D. degree from Chalmers Univer-
[56] “Spatial channel model for multiple input multiple output (MIMO) sim- sity of Technology, Sweden, in 2011. He is currently
ulations,” 3GPP, Sophia Antipolis, France, Tech. Rep. 25.996 ver. 12.0.0 a Senior Researcher with Ericsson AB , Gothenburg.
rel. 12, Sep. 2014. He is also the Lead Scientist for Ericsson AB within
the Marie Skłodowska-Curie Innovative Training
Network, SILIKA. His main interests lie in radio
signal processing and behavioral modeling of radio
hardware for advanced antenna systems.