Dominant Channel Estimation via MIPS for

Large-Scale Antenna Systems with One-Bit ADCs

In-soo Kim, Namyoon Lee, and Junil Choi
Department of Electrical Engineering
POSTECH
Email: {insookim, nylee, junil}@postech.ac.kr

Abstract—In large-scale antenna systems, using one-bit analog- past [8], although not widely, and begun to be in the spotlight
to-digital converters (ADCs) has recently become important since recently [9]–[11]. In [8], extensions of the conventional DOA
they offer significant reductions in both power and cost. However, estimators to one-bit ADCs, based on the arcsine law [12],
in contrast to high-resolution ADCs, the coarse quantization of
one-bit ADCs results in an irreversible loss of information. In [13], have been studied. The one-bit spatial smoothing MUSIC
the context of channel estimation, studies have been developed (SS MUSIC) proposed in [9], which also relies on the arcsine
extensively to combat the performance loss incurred by one- law, considers DOA estimation in sparse arrays. The drawback
bit ADCs. Furthermore, in the field of array signal processing, of the DOA estimators in [8], [9] is that many independent
direction-of-arrival (DOA) estimation combined with one-bit observations are needed to reconstruct the covariance matrix.
ADCs has gained growing interests recently to minimize the
estimation error. In this paper, a channel estimator is proposed The work in [10] proposed the gridless one-bit DOA estimator
for one-bit ADCs where the channels are characterized by their based on the support vector machine (SVM), which is cumber-
angular geometries, e.g., uniform linear arrays (ULAs). The goal some in practice since the coarse DOA estimate based on the
is to estimate the dominant channel among multiple paths. The SVM should be refined using the Taylor expansion to obtain
proposed channel estimator first finds the DOA estimate using the the gridless DOA estimate.
maximum inner product search (MIPS). Then, the channel fading
coefficient is estimated using the concavity of the log-likelihood In this paper, we propose a channel estimator for one-
function. The limit inherent in one-bit ADCs is also investigated, bit ADCs where the channels are characterized by their
which results from the loss of magnitude information. angular geometries, e.g., uniform linear arrays (ULAs). In
particular, the channel is composed of multiple paths where
I. I NTRODUCTION
each path is characterized by its channel fading coefficient
In large-scale antenna systems, known as massive multiple- and steering vector parameterized by the DOA. The goal
input multiple-output (MIMO), a significant performance gain is to estimate the dominant channel, known as the line-of-
is achieved by deploying a large number of antennas at the sight (LOS) channel. The proposed channel estimator finds
base station. However, the hardware cost and excessive power the DOA estimate via the maximum inner product search
consumption due to the large number of antennas make high- (MIPS). Then, the channel fading coefficient is estimated by
resolution analog-to-digital converters (ADCs) less attractive, maximizing the log-likelihood function at the MIPS DOA
and using low-resolution, or even the extreme case of one-bit, estimate using convex optimization. In contrast to [8], [9],
ADCs has gained popularity [1]–[4]. The drawback of one-bit the MIPS DOA estimator finds the DOA estimate using the
ADCs is that the information loss resulting from the coarse instantaneous received signal. The simulation results show
quantization is severe, acting as a performance bottleneck. that the proposed channel estimator performs close to the
In [5]–[7], channel estimators for one-bit ADCs have been pseudo ML (pML) channel estimator, which maximizes the
proposed. In [5], the near maximum likelihood (nML) channel likelihood function assuming that the sum of the non-line-
estimator relies on the concavity of the log-likelihood function of-sight (NLOS) channels and noise is white and Gaussian.
assuming that the channels have no predefined structure. The The MIPS-based channel estimator is also computationally
modified expectation-maximization (EM) algorithm proposed efficient compared to the pML channel estimator. In addition,
in [6] exploits the channel sparsity of millimeter wave sys- we investigate the performance limit of the pML channel
tems. However, the computational complexities of the channel estimator in one-bit ADCs.
estimators introduced in [5], [6] are expensive because both
are based on the ML channel estimator. The Bussgang linear II. S YSTEM M ODEL
minimum mean squared error (BLMMSE) channel estimator
introduced in [7] uses the second-order channel statistics to We consider a single-input multiple-output (SIMO) system
formulate the one-bit LMMSE channel estimator based on the where the receiver employs an array of antennas characterized
Bussgang decomposition, which is only applicable to Gaussian by its angular geometry. At each antenna, the received signal’s
distributed channels. real and imaginary parts are quantized by one-bit ADCs.
In the field of array signal processing, direction-of-arrival The single-antenna transmitter transmits a pilot sequence of
(DOA) estimation for one-bit ADCs has been studied in the length N to the M -antenna receiver. Therefore, the received

978-1-5386-4727-1/18/$31.00 ©2018 IEEE

signal Y ∈ CM ×N is Remark 1: In this paper, the steering vector is assumed to
√ be parameterized by the one-dimensional DOA to simplify the
Y= ρhxT + N (1) notations. However, the MIPS-based channel estimator can be
where the m-th row and n-th column correspond to the m-th applied to any angular geometry where the steering vector is
antenna and n-th time slot, respectively. The channel h ∈ CM parameterized by the multi-dimensional DOA as long as the
is formed by the combination of the LOS and NLOS channels elements of the steering vector obey the unit-power constraint,
h0 and h` where ` 6= 0, respectively, i.e., e.g., uniform planar arrays (UPAs) that are parameterized by
the horizontal and vertical DOAs.
L
X
h= c` h` , (2) III. P ML C HANNEL E STIMATION
`=0 In this section, we analyze the pML channel estimator. The
h` = g` a(θ` ), (3) observations established in this section builds the framework
of the MIPS-based channel estimator in Section IV. To develop
and (p the pML channel estimator, we write y as
K/(K + 1) if ` = 0
c` = p , (4) L
1/L(K + 1) 6 0
if ` = √ √ X
y= ρXc0 h0 + ρX c` h` + n
where L is the number of NLOS paths, K is the Rician K- √
`=1
factor, g` ∈ C is the zero-mean and unit-variance `-th channel = ρXc0 h0 + n̄. (10)
fading coefficient, θ` is the `-th DOA, and a(θ` ) ∈ CM is the In general, it is hard to estimate h0 because n̄ is neither white
steering vector parameterized by θ` . In particular, we assume nor Gaussian. Hence, the pML channel estimator circumvents
that a(θ` ) obeys the power constraint of |am (θ` )|2 = 1 where this problem by approximating n̄ by ñ, which is distributed
am (θ` ) is the m-th element of a(θ` ). For example, if the as CN (E{n̄}, diag(E{n̄n̄H })), i.e,
receiver is modeled as a ULA where the inter-element spacing
is d while the wavelength is λ, the steering vector is ñ ∼ CN (0M N , σ 2 IM N ) (11)
h
2πd 2πd
iT where
a(θ` ) = 1, e−j λ sin θ` , · · · , e−j λ (M −1) sin θ` . (5)
(
2 ρ/(K + 1) + 1 if L 6= 0
σ = , (12)
The support of g` and Θ ⊂ R, respectively. The 1 if L = 0
 θ` are C and
T
pilot sequence x = x1 , · · · , xN ∈ CN of length N follows and maximizes the likelihood function to estimate h0 . In fact,
|xn |2 = 1 to simplify the peak power constraint. The elements the pML channel estimator reduces to the ML channel esti-
of the additive white Gaussian noise (AWGN) N ∈ CM ×N mator when L = 0 because n̄ is both white and Gaussian. To
are independent and identically distributed (i.i.d.) as CN (0, 1). formulate the likelihood function assuming ñ, as a preliminary
Furthermore, g` , θ` , and N are assumed to be independent. step, we define ρ̃ = ρ/σ 2 . The k-th element of Xc0 a(θ0 ) is
The signal-to-noise ratio (SNR) is defined as ρ. denoted by Xk (θ0 ). In addition, define fR,k (θ0 ), fI,k (θ0 ), and
The vectorized received signal y = vec(Y) is g0 as
√  T
y = ρXh + n (6) fR,k (θ0 ) = Re(Xk (θ0 )), −Im(Xk (θ0 )) , (13)
 T
where X = x ⊗ IM and n = vec(N) ∼ CN (0M N , IM N ). At fI,k (θ0 ) = Im(Xk (θ0 )), Re(Xk (θ0 )) , (14)
each antenna, the real and imaginary parts of y are quantized  T
g0 = Re(g0 ), Im(g0 ) . (15)
by one-bit ADCs. The quantized received signal ŷ is
Then, assuming ñ, the k-th element of y is distributed as
ŷ = Q(y) (7) √
CN ( ρXk (θ0 )g0 , σ 2 ) (16)
where Q(·) is the element-wise one-bit quantization function.
conditioned on g0 and θ0 . Thus, because the elements of ñ are
In this paper, zero thresholds are used, i.e.,
independent whose real and imaginary parts are independent,
1 we obtain the log-likelihood function Lρ̃ (g00 , θ00 ) evaluated at
Q(y) = √ sgn(y) (8)
2 g00 ∈ R2 and θ00 ∈ Θ, which is
where sgn(·) is the sign function, which is applied to the real M
X N
Lρ̃ (g00 , θ00 ) = (θ00 )g00 )
p T
and imaginary parts element-wise. Therefore, the elements of (log Φ(2ŷR,k ρ̃fR,k
ŷ are constrained to the quadrature phase-shift keying (QPSK) k=1
(θ00 )g00 ))
p T
constellation points, i.e., + log Φ(2ŷI,k ρ̃fI,k (17)
1 where ŷR,k = Re(ŷk ), ŷI,k = Im(ŷk ), and Φ(·) represents the
ŷk ∈ √ {1 + j, 1 − j, −1 + j, −1 − j} (9)
2 cumulative distribution function (CDF) of N (0, 1). Then, the
pML channel estimate ĥ0,pML of h0 is defined as
where ŷk is the k-th element of ŷ. The goal is to estimate h0
from ŷ. ĥ0,pML = ĝ0,pML a(θ̂0,pML ) (18)
where Then, since
ĝρ̃2 (θ00 ) = argmaxLρ̃2 (g00 , θ00 )
(ĝ0,pML , θ̂0,pML ) = argmax Lρ̃ (g00 , θ00 ). (19) g00 ∈R2
g00 ∈R2 ,θ00 ∈Θ (a) 1
= √ argmaxLρ̃1 (g̃0 , θ00 )
k g̃0 ∈R2
ĝ0,pML contains the real and imaginary parts of ĝ0,pML . 1
In general, solving (19) is computationally cumbersome due = √ ĝρ̃1 (θ00 ) (22)
k
to the vast size of the search space and expensive objective √
function. To find ĥ0,pML more efficiently, we investigate the where (a) results from substituting g̃0 for kg00 , we observe
structure of Lρ̃ (g00 , θ00 ). First, note that Lρ̃ (g00 , θ00 ) is a concave that kĝρ̃2 (θ00 )k → 0 as k → ∞, which is equivalent to saying
function of g00 because the argument of each Φ(·) is an affine that kĝρ̃ (θ00 )k → 0 as ρ̃ → ∞, for any realization of ŷ
function of g00 , which preserves the log-concavity of Φ(·) [14]; evaluated at any θ00 . Hence, from ĝ0,pML = ĝρ̃ (θ̂0,pML ), we
a similar observation was established in [5]. Thus, ĥ0,pML can note that kĝ0,pML k → 0 as ρ̃ → ∞ for any realization of ŷ,
be found more efficiently by interpreting (19) as arriving at
   
Pr lim kĥ0,pML k = 0 = Pr lim kĝ0,pML k = 0
max Lρ̃ (g00 , θ00 ) = max
0
max Lρ̃ (g00 , θ00 ) (20) ρ̃→∞ ρ̃→∞
g00 ∈R2 ,θ00 ∈Θ 0 2
θ0 ∈Θg0 ∈R
= 1, (23)

where the inner optimization problem of (20) can be solved which completes the proof.
using convex optimization, e.g., the gradient descent method Lemma 1 shows that kĥ0,pML k becomes deterministic in the
(GDM) using the backtracking line search [14]. high SNR regime, implying that one-bit ADCs do not convey
Second, note that Lρ̃ (g00 , θ00 ) has no explicit structure with any channel quality information (CQI) in the absence of noise.
respect to θ00 . Thus, to solve the outer optimization problem of In contrast to full-resolution ADCs where noise is generally
(20), Θ should be searched exhaustively; that is, the number not welcomed, the presence of noise may enhance the channel
of convex optimization performed is proportional to the size of estimation performance in one-bit ADCs. The intuitive reason
Θ. For example, assume that the inner optimization problem is why the magnitude information of the received signal is lost
solved using the GDM while the outer optimization problem in one-bit ADCs can be explained by noting that
is solved using the B-bit grid search, thereby partitioning Θ √ √
into 2B uniform grid points. Then, the pML channel estimator Q( ρXc0 h0 + ñ) ≈ Q( ρXc0 h0 )
√
performs the GDM 2B times. In this paper, we assume that the = Q(k ρXc0 h0 ) (24)
pML channel estimator solves the inner and outer optimization in the high SNR regime where k > 0. According to (24),
problems of (20) using the GDM, which uses the backtracking the information embedded in k is indistinguishable in the
line search, and B-bit grid search, respectively. In Section V, high SNR regime because one-bit ADCs do not convey any
the computational complexity of the pML channel estimator magnitude information. Hence, we expect that the performance
is investigated by counting the number of real multiplications of the pML channel estimator falls as the SNR enters the high
performed. The explanation of how the number of real multi- SNR regime.
plications performed by the pML channel estimator is counted
is omitted because of the page limit. IV. MIPS-BASED C HANNEL E STIMATION
Before we proceed to the MIPS-based channel estimator, we In this section, we derive the MIPS-based channel estimator.
investigate the fundamental limit of one-bit ADCs in the high The problem of the pML channel estimator is that Lρ̃ (g00 , θ00 )
SNR regime through the lens of the pML channel estimator. has no explicit structure with respect to θ00 . Thus, the number
of convex optimization performed is proportional to the size
Lemma 1. kĥ0,pML k → 0 as ρ̃ → ∞ almost surely. of Θ. The MIPS-based channel estimator attempts to break
Proof. The goal of the proof is to show that kĝ this computational bottleneck by performing two-stage channel
√ 0,pML k → 0 estimation, which separates the search space into Θ and R2 . In
as ρ̃ → ∞ almost surely because kĥ0,pML k = M kĝ0,pML k.
Therefore, we proceed by showing that kĝρ̃ (θ00 )k → 0 as ρ̃ → this section, similar to the pML channel estimator, the MIPS-
∞ for any realization of ŷ evaluated at any θ00 where ĝρ̃ (θ00 ) based channel estimator considers ñ instead of n̄.
is defined as A. DOA Estimator
In this subsection, the first stage of the MIPS-based channel
ĝρ̃ (θ00 ) = argmaxLρ̃ (g00 , θ00 ). (21) estimator is described. In the first stage, θ0 is estimated in the
g00 ∈R2 search space Θ. Thus, we can interpret the problem as DOA
estimation in one-bit ADCs. In [8], DOA estimators for one-bit
First, consider ρ̃1 > 0 and ρ̃2 = k ρ̃1 where k > 0. From√(17), ADCs were proposed using the arcsine law, which relates the
observe that Lρ̃2 (g00 , θ00 ) = Lρ̃1 (g̃0 , θ00 ) where g̃0 = kg00 . covariance matrices of the quantized and unquantized received
signals using arcsine [12], [13]. In this paper, motivated by the Algorithm 1 MIPS-based channel estimator
approach of [8], we propose the MIPS DOA estimate of θ0 . 1: Find θ̂0,MIPS = argmax|(Xa(θ00 ))H ŷ|
To derive the MIPS DOA estimate of θ0 , we express θ0 in θ00 ∈Θ

terms of the conditional covariance matrix of ŷ given θ0 . The 2: Find ĝ0,MIPS = argmaxLρ̃ (g00 , θ̂0,MIPS )
g00 ∈R2
conditional covariance matrix of y given θ0 , namely Cy (θ0 ),
is 3: ĥ0,MIPS = ĝ0,MIPS a(θ̂0,MIPS )
√ √
Cy (θ0 ) = E{( ρXc0 h0 + ñ)( ρXc0 h0 + ñ)H |θ0 }
= ρc20 Xa(θ0 )a(θ0 )H XH + σ 2 IM N (25) to develop the DOA estimator. By replacing Cŷ (θ0 ) with Ĉŷ
in (30), we obtain the MIPS DOA estimate θ̂0,MIPS of θ0 ,
where n̄ was approximated by ñ. The diagonal matrix formed which is
π 
by Cy (θ0 ) is denoted by Σy (θ0 ), which is θ̂0,MIPS = argmax(Xa(θ00 ))H sin Ĉŷ Xa(θ00 )
θ00 ∈Θ 2
Σy (θ0 ) = diag(Cy (θ0 )) π 
= argmax(Xa(θ00 ))H sin ŷŷH Xa(θ00 )
= (ρc20 + σ 2 )IM N . (26) θ00 ∈Θ 2
(a)
Then, from the arcsine law, Cŷ (θ0 ), which is the conditional = argmax(Xa(θ00 ))H ŷŷH Xa(θ00 )
θ00 ∈Θ
covariance matrix of ŷ given θ0 , can be expressed by Cy (θ0 )
as = argmax|(Xa(θ00 ))H ŷ| (32)
θ00 ∈Θ
2 
−1 −1

Cŷ (θ0 ) = arcsin Σy 2 (θ0 )Cy (θ0 )Σy 2 (θ0 ) (27) where (a) is a direct consequence of the fact that the elements
π
of ŷŷH are constrained to {1, −1, j, −j}. The reason why the
and vice versa, i.e., proposed DOA estimator is named the MIPS DOA estimator
1
π  1 follows from (32); the receiver estimates θ0 by searching for
Cy (θ0 ) = Σy2 (θ0 ) sin Cŷ (θ0 ) Σy2 (θ0 ) the steering vector which maximizes the inner product. To
2
2 2 π  solve (32), an exhaustive search is required.
= (ρc0 + σ ) sin Cŷ (θ0 ) (28)
2 B. Channel Fading Coefficient Estimator
where arcsin(·) and sin(·) are the element-wise arcsine and In the second stage, g0 is estimated. The MIPS-based
sine functions applied to the real and imaginary parts, re- channel fading coefficient estimate ĝ0,MIPS of g0 is defined
spectively. In addition, based on the conventional beamformer as
approach, which attempts to maximize the output power [15], ĝ0,MIPS = argmaxLρ̃ (g00 , θ̂0,MIPS ) (33)
θ0 can be written as g00 ∈R2

where ĝ0,MIPS contains the real and imaginary parts of

argmax(Xa(θ00 ))H Cy (θ0 )Xa(θ00 ) ĝ0,MIPS . From (33), observe that ĝ0,MIPS is the maximizer
θ00 ∈Θ
(a)
of the inner optimization problem of (20) defined at θ̂0,MIPS ,
= argmax(ρc20 |(Xa(θ00 ))H Xa(θ0 )|2 + σ 2 kXa(θ00 )k2 ) which is a convex optimization problem. Therefore, we can
θ00 ∈Θ
find ĝ0,MIPS efficiently, which completes the estimation of h0 .
(b)
=argmax(ρc20 N 2 |a(θ00 )H a(θ0 )|2 + σ 2 M N ) The MIPS-based channel estimate ĥ0,MIPS of h0 is defined as
θ00 ∈Θ
ĥ0,MIPS = ĝ0,MIPS a(θ̂0,MIPS ). (34)
= argmax|a(θ00 )H a(θ0 )|2
θ00 ∈Θ In Algorithm 1, we give an outline of the MIPS-based channel
(c) estimator. The first stage requires an exhaustive search over
= θ0 (29)
Θ. The second stage can be solved using convex optimization
where (a), (b), and (c) are the consequences of (25), kxk2 = for a given θ̂0,MIPS . In this paper, we assume that the opti-
N , and the unit-power constraint imposed on the steering vec- mization problems in the first and second stages of the MIPS-
tor’s elements, respectively. At this point, θ0 can be expressed based channel estimator are solved by the B-bit grid search
in terms of Cŷ (θ0 ) by putting (28) into (29), i.e., and GDM where the backtracking line search is employed,
π  respectively.
θ0 = argmax(Xa(θ00 ))H sin Cŷ (θ0 ) Xa(θ00 ). (30) In Section V, it is shown that the computational complexity
θ00 ∈Θ 2 of the MIPS-based channel estimator is in the order of 1/2B
of the pML channel estimator’s since the GDM is performed
However, since the receiver has no prior knowledge of 2B times by the pML channel estimator. In addition, the sim-
Cŷ (θ0 ), we use the sample covariance matrix Ĉŷ , which is ulation results verify that the MIPS-based channel estimator
performs close to the pML channel estimator. Thus, the MIPS-
Ĉŷ = ŷŷH , (31) based channel estimator is computationally efficient while the
 101 TABLE I
Proposed MIPS: L=0 T HE AVERAGE NUMBER OF REAL MULTIPLICATIONS FOR M = 24,
ML: L=0 N = 15, AND L = 5 AT DIFFERENT SNR S .
LMMSE: L=0
Proposed MIPS: L=5
100 pML: L=5 SNR (dB) MIPS·10−6 pML·10−6
LMMSE: L=5
-10 1.8 514.7
-5 2.3 313.0
MSE

10-1 0 3.6 1873.5

5 5.1 684.3
10 13.8 1783.9
10-2

tors. The covariance matrices needed to compute the LMMSE

10-3 are found by 106 Monte-Carlo simulations. When L = 0, the
-10 -8 -6 -4 -2 0 2 4 6 8 10
pML channel estimator is equivalent to the ML channel esti-
SNR (dB)
mator because it is the maximizer of the likelihood function;
Fig. 1. MSE vs. SNR for M = 24, N = 15 at different L. that is, the pML channel estimator is optimal in the sense that
the likelihood function is maximized. The interesting point
is that the difference between the MSEs of the MIPS-based
performance loss is negligible compared to the pML channel and ML channel estimators is negligible, which shows that
estimator. the MIPS-based channel estimator achieves the performance
Remark 2: The performance of the MIPS-based channel of the ML channel estimator when L = 0.
estimator is expected to degrade in the high SNR regime be- When L = 5, the pML channel estimator becomes subopti-
cause ĝ0,MIPS shrinks to a meaningless estimate 02 according mal because it is the maximizer of the approximated likelihood
to Lemma 1. In fact, the poor performance of one-bit ADCs function, which is obtained by assuming that n̄ is both white
in the high SNR regime is a well-known phenomenon, which and Gaussian. The MSE of the MIPS-based channel estimator
is inevitable due to the loss of magnitude information. is close to the pML channel estimator’s, which shows that
the MIPS-based channel estimator performs as good as the
V. S IMULATION R ESULTS
pML channel estimator regardless of L. In addition, both the
In this section, Monte-Carlo simulations are performed to MIPS-based and pML channel estimators perform worse in
evaluate the performance of the MIPS-based channel estimator. the high SNR regime, which can be explained by Lemma 1.
It is assumed that the first and second stages of Algorithm 1 The LMMSE channel estimator is shown as a comparison.
are solved by the B-bit grid search and GDM, respectively. Table I shows the computational complexities of the MIPS-
In addition, we assume that the pML channel estimator finds based and pML channel estimators when M = 24, N = 15,
ĥ0,pML by solving the inner and outer optimization problems and L = 5 at different SNRs. The measure of the computa-
of (20) using the GDM and B-bit grid search, respectively. tional complexity is the average number of real multiplications
The parameter of the B-bit grid search is set to B = 8. The performed. The specific explanation of how they were counted
stopping critetion η and backtracking line search parameters is omitted because of the page limit. The ratio of the average
of the GDM in [14] are set to η = 0.01 and α = 0.1, β = 0.5, number of real multiplications performed by the MIPS-based
respectively. The zero-forcing (ZF) channel estimate in [5] is channel estimator to the pML channel estimator’s is at most
used as the starting point of the GDM. 0.7%, which is in the order of 1/2B . Therefore, the MIPS-
The Rician K-factor is set to 13.5 dB based on the measure- based channel estimator is computationally efficient while the
ments in [16], the channel fading coefficients and DOAs are performance gap is negligible compared to the pML channel
distributed as g` ∼ CN (0, 1) and θ` ∼ Unif(Θ), respectively, estimator.
where Θ = [−π/3, π/3], the receiver is modeled as a ULA in In Fig. 2 and Fig. 3, assuming that θ0 and g0 are determin-
(5) where d = λ/2, and the pilot sequence is selected as the istic, the Cramér-Rao bound (CRB) of θ0 and g0 is provided at
last column of the size N discrete Fourier Transform (DFT) different SNRs when L = 0, which was derived in [17]. The
matrix. k-th realizations of θ0 and g0 are denoted by θ0 (k) and g0 (k),
In Fig. 1, we compare the MSEs of the MIPS-based, pML, respectively. The MSEs of the MIPS-based and ML channel
and LMMSE channel estimators when M = 24, N = 15 at estimators conditioned on θ0 (k) and g0 (k), which are
different SNRs with L = 0 and L = 5. The MSE is defined
as MSE(θ0 (k)) = E{(θ̂0 − θ0 )2 |θ0 (k), g0 (k)}, (36)
1
MSE = E{kĥ0 − h0 k2 } (35) 2
MSE(g0 (k)) = E{|ĝ0 − g0 | |θ0 (k), g0 (k)}, (37)
M
where 1000 Monte-Carlo simulations are performed to com- are plotted as well. From Fig. 2, observe that the gap between
pute the MSEs of the MIPS-based and pML channel estima- the MSEs of the MIPS-based and ML channel estimators are
 101 101 using convex optimization. The simulation results showed that
Proposed MIPS Proposed MIPS
ML
CRB
ML
CRB
the MIPS-based channel estimator performs close to the pML
100 100
channel estimator.
MSE°

MSE
10-1 10-1 ACKNOWLEDGMENT
This work was supported by the Institute for Informa-
10-2 10-2
-5 0 5 10
SNR (dB)
15 20 -5 0 5 10
SNR (dB)
15 20 tion & communications Technology Promotion (IITP) un-
der grant funded by the MSIT of the Korea government
1 1
10 10
Proposed MIPS Proposed MIPS
(No.2018(2016-0-00123), Development of Integer-Forcing
ML ML
CRB CRB MIMO Transceivers for 5G & Beyond Mobile Communication
100 100
Systems).
MSE°

MSE
10-1 10-1
R EFERENCES
[1] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
10-2 10-2
-5 0 5 10 15 20 -5 0 5 10 15 20 MIMO for next generation wireless systems,” IEEE Communications
SNR (dB) SNR (dB)
Magazine, vol. 52, no. 2, pp. 186–195, 2014.
[2] E. Björnson, E. G. Larsson, and T. L. Marzetta, “Massive MIMO:
Fig. 2. MSE(θ0 (k)) and MSE(g0 (k)) vs. SNR for M = 8, N = 10, Ten myths and one critical question,” IEEE Communications Magazine,
and L = 0. Two realizations of θ0 and g0 are given with the corresponding vol. 54, no. 2, pp. 114–123, 2016.
CRBs. [3] C. Risi, D. Persson, and E. G. Larsson, “Massive MIMO with 1-bit
ADC,” arXiv preprint arXiv:1404.7736, 2014.
100 101 [4] F. Wang, J. Fang, H. Li, and S. Li, “Quantization Design and Channel
Proposed MIPS Proposed MIPS
ML ML
Estimation for Massive MIMO Systems with One-Bit ADCs,” arXiv
CRB CRB
preprint arXiv:1704.04709, 2017.
10-1 100
[5] J. Choi, J. Mo, and R. W. Heath, “Near maximum-likelihood detector
MSE°

MSE

and channel estimator for uplink multiuser massive MIMO systems with
10-2 10-1 one-bit ADCs,” IEEE Transactions on Communications, vol. 64, no. 5,
pp. 2005–2018, 2016.
10-3 10-2
[6] J. Mo, P. Schniter, N. G. Prelcic, and R. W. Heath, “Channel estima-
-5 0 5 10 15 20 -5 0 5 10 15 20
SNR (dB) SNR (dB)
tion in millimeter wave MIMO systems with one-bit quantization,” in
Signals, Systems and Computers, 2014 48th Asilomar Conference on.
100 101 IEEE, 2014, pp. 957–961.
Proposed MIPS
ML
Proposed MIPS
ML
[7] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, and
CRB CRB L. Liu, “Channel estimation and performance analysis of one-bit massive
10-1 100
MIMO systems,” IEEE Transactions on Signal Processing, 2017.
MSE°

MSE

[8] O. Bar-Shalom and A. J. Weiss, “DOA estimation using one-bit quan-

10-2 10-1 tized measurements,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 38, no. 3, pp. 868–884, 2002.
10-3 10-2
[9] C.-L. Liu and P. Vaidyanathan, “One-bit sparse array DOA estimation,”
-5 0 5 10 15 20 -5 0 5 10 15 20 in Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE
SNR (dB) SNR (dB)
International Conference on. IEEE, 2017, pp. 3126–3130.
[10] Y. Gao, D. Hu, Y. Chen, and Y. Ma, “Gridless 1-b DOA Estimation
Fig. 3. MSE(θ0 (k)) and MSE(g0 (k)) vs. SNR for M = 16, N = 12, Exploiting SVM Approach,” IEEE Communications Letters, vol. 21,
and L = 0. Two realizations of θ0 and g0 are given with the corresponding no. 10, pp. 2210–2213, 2017.
CRBs. [11] M. Stein, K. Barbé, and J. A. Nossek, “DOA Parameter Estimation with
1-bit Quantization Bounds, Methods and the Exponential Replacement,”
in Smart Antennas (WSA 2016); Proceedings of the 20th International
negligible as in Fig. 1. In addition, we point out that the gap ITG Workshop on. VDE, 2016, pp. 1–6.
between MSE(θ0 (k)) and the CRB is not increased, whereas [12] J. H. Van Vleck and D. Middleton, “The spectrum of clipped noise,”
Proceedings of the IEEE, vol. 54, no. 1, pp. 2–19, 1966.
the gap between MSE(g0 (k)) and the CRB is increased as the [13] G. Jacovitti and A. Neri, “Estimation of the autocorrelation function of
SNR exceeds the medium SNR regime. This follows because complex Gaussian stationary processes by amplitude clipped signals,”
one-bit ADCs convey only directional information, which is IEEE transactions on information theory, vol. 40, no. 1, pp. 239–245,
1994.
embedded in θ0 ; no magnitude information is provided, which [14] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge
is contained in |g0 |. university press, 2004.
[15] H. Krim and M. Viberg, “Two decades of array signal processing
VI. C ONCLUSION research: the parametric approach,” IEEE signal processing magazine,
vol. 13, no. 4, pp. 67–94, 1996.
We proposed the MIPS-based channel estimator for large- [16] Z. Muhi-Eldeen, L. Ivrissimtzis, and M. Al-Nuaimi, “Modelling and
scale antenna systems using one-bit ADCs where the array measurements of millimetre wavelength propagation in urban environ-
ments,” IET microwaves, antennas & propagation, vol. 4, no. 9, pp.
of antennas at the receiver is characterized by its angular 1300–1309, 2010.
geometry. The MIPS-based channel estimator finds the esti- [17] P. Wang, J. Li, M. Pajovic, P. Boufounos, and P. Orlik, “On Channel
mate of the dominant channel among multiple paths using Estimation for One-Bit Massive MIMO Systems with Fixed and Time-
Varying Thresholds,” in Signals, Systems and Computers, 2017 51st
two-stage channel estimation, which has low computational Asilomar Conference on. IEEE, 2017.
complexity. In the first stage, the DOA is estimated using the
MIPS. After finding the DOA estimate, the second stage is
performed, which finds the channel fading coefficient estimate