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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
1

Channel Estimation in One-Bit Massive MIMO

Systems: Angular vs. Unstructured Models
Shilpa Rao, Student Member, IEEE, Amine Mezghani, Member, IEEE,
and A. Lee Swindlehurst, Fellow, IEEE

Abstract—Millimeter wave (mmWave) massive MIMO cellular to the aforementioned limitations. Traditional channel models
systems will be characterized by increased bandwidths and the that assume Rayleigh fading are not suitable for mmWave sys-
ability to handle a large number of users. To cope with the tems because the scattering environment of mmWave channels
power consumption problem due to an increased number of
receive antennas, the idea of equipping one-bit ADCs at the base is not dense, but rather sparse with line-of-sight (LOS) and a
station has been proposed. The goal of this paper is to estab- few reflected propagation paths contributing to the effective
lish performance bounds on the channel estimation of one-bit channel [1], [2]. Hence, much of the work in mmWave has
mmWave massive MIMO receivers for different types of channel focused on direction-of-arrival (DOA)-based channel estima-
models. The Cramér-Rao bound (CRB) is a lower bound on the tion. Previous works in mmWave channel estimation exploit
performance of unbiased estimators and sets a benchmark for
the design of channel estimators. We consider both a structured the sparsity [3] of these channels in the delay and angle
channel model for a single user where the channel is composed of domains [4–11]. The authors of [1] focus on pilot-aided
a superposition of multipaths characterized by path delays and mmWave DOA-based channel estimation, channel subspace
directions-of-arrival (DOAs), and an unstructured channel model estimation is studied in [9] and mmWave channel estimation
where the channel is a generic FIR filter. The Fisher information with hybrid architectures is considered in [12] and [13].
matrix (FIM) for these channel models are derived in closed
form. The CRB is also extended to a dictionary-based channel The large bandwidths at mmWave limit the performance
model, where the path delays and DOAs are selected from small of analog-to-digital converters (ADCs) at the receiver. It has
perturbations on a discrete grid, and a sparsity constraint applies been shown that the power consumption of ADCs increases
to the vector of path loss components. We also derive the Bayesian
CRB when the array response is imperfectly known and is drastically with increasing quantization resolution and sam-
affected by perturbations in the sensor pattern or position. The pling rate requirements [14], [15]. One-bit ADCs are cost-
CRBs are evaluated numerically and the effects of various system effective, consume less power and do not require automatic
parameters on the CRB are studied. The dependencies between gain control [16], [17]. Moreover, it has been shown that one-
channel parameters and the effect of array perturbations are bit ADCs suffer from a low power penalty (approximately
also investigated.
π/2) at low SNRs [18]. Therefore, one-bit quantization is
Index Terms—Massive multiple input multiple-output, wide- pertinent to mmWave massive MIMO systems since they suffer
band channel estimation, one-bit analog-to-digital converters,
Cramér-Rao bounds, array calibration.
from low SNR per antenna. Prior work focuses on channel
estimation using the Bussgang decomposition with a Rayleigh
fading channel model [19], quantization design with non-
I. I NTRODUCTION zero threshold and channel estimation [20], throughput anal-
Millimeter wave (mmWave) massive multiple-input ysis [21], blind and semi-blind channel estimation with time
multiple-output (MIMO) is a promising area for next- division duplexing [21] in one-bit massive MIMO systems.
generation wireless communication systems. These systems The authors of [22] provide a blind sparse channel estimation
employ arrays with many antennas, of the order of a algorithm based on a maximum likelihood (ML) formulation.
hundred or more, at the base station (BS) and operate in the In [5], approximate message passing algorithms to exploit the
30-300GHz frequency range. Furthermore, the rising user joint sparsity of the broadband channel in the angle and delay
demands for capacity can be met through higher bandwidths domains with few-bit ADCs are proposed. However, in the
and spatial multiplexing. Massive MIMO is also capable above-mentioned DOA-based channel models, it is assumed
of concentrating energy in very selective directions, thus that the inter-element time delay between antennas in the
significantly increasing the energy efficiency. However, at array is small compared to the inverse signal bandwidth. For
mmWave, the benefits of massive MIMO are limited by mmWave massive MIMO systems, this assumption will typi-
the low SNR per antenna- a result of increased propagation cally not be true. This effect, sometimes referred to as (beam)
losses, diminished scattering, atmospheric absorption and “squint”, has been observed to cause a serious mismatch in
higher noise bandwidths. the array response, and if ignored can significantly degrade
Channel estimation, which is key for exploiting the potential performance [23–26].
gains offered by mmWave massive MIMO, is challenging due
In this paper, we focus on Cramér-Rao performance bounds
This work was supported in part by the National Science Foundation under for channel estimation in one-bit mmWave massive MIMO
Grant ECCS-1547155 and Grant CCF-1703635. systems. We will explore both deterministic and Bayesian
S. Rao and A. L. Swindlehurst are with the Center for Pervasive Commu-
nications and Computing, University of California, Irvine, CA 92697 USA Cramér-Rao bounds (CRBs) depending on whether or not
(e-mail: [email protected]; [email protected]). prior information about the parameters is available. In the
A. Mezghani was with the Center for Pervasive Communications and Bayesian setting, the a priori information is taken into account
Computing, University of California, Irvine, CA 92697 USA. He is now with
the Wireless Networking and Communications Group, The University of Texas in the computation of the joint probability density function
at Austin, Austin, TX 78712 USA (e-mail: [email protected]). (pdf). Misspecified Cramér-Rao bounds (MCRBs), when the

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
2

assumed data model is different from the true model, can multipaths in the structured case, or the maximum length of
also be derived [27]. In [28], a lower bound on the Fisher the FIR filter in the unstructured case) and the noise variance
information matrix (FIM) for the exponential family of dis- are known. For the structured model, these parameters are the
tributions is derived and the ML estimator based on the path loss components, DOAs, path delays and the array pertur-
“pessimistic” CRB for the DOA parameter is derived in [29]. bation parameters. The perturbations are modeled as complex
In [30] and [31], the CRB for the channel parameterized Gaussian with a known covariance matrix. The FIM is derived
by DOAs and path gains is derived when the channel and for the specific case of uniform linear arrays (ULAs), a
the array responses are frequency flat. In our work, we will root-raised-cosine shaping function, and sensor position and
take into account the channel estimation error for DOA-based pattern perturbations. For the unstructured channel model, the
models when the array response does not exactly match the parameters are simply the path loss components, whereas for
assumed array model, similar to the array perturbation studies the dictionary-based channel, the parameters are the path loss
of [32–34]. In particular, we are interested in the level of components, the grid mismatch error parameters and the array
array calibration accuracy needed for DOA-based methods perturbation parameters.
to maintain their advantage compared with less parsimonious A number of numerical experiments are performed to evalu-
unstructured models. ate the CRBs, and a comparison of the structured, unstructured
Our analysis focuses on a one-bit mmWave single-input- and dictionary-based channels as a function of the SNR is
multiple-output (SIMO) pilot-based single-carrier transmission performed. The effects of perturbation, bandwidth, the channel
system where a single transmitter is equipped with one antenna delay-tap length, and the number of receive antennas on the
and the base station employs an antenna array. We con- CRBs of the one-bit quantized system are also studied. We
sider channel models that are either “structured” (DOA-based, also include comparisons with the CRB obtained when there
arbitrary delays) or “unstructured” (FIR, uniformly-spaced is no quantization error. The numerical results provide insight
delays), and compare the resulting CRBs for the channel into the relative impact of the various factors that influence
estimates. Under the structured channel model, we assume of the channel estimate, including the precision of the array
that the channel is parameterized by the multipath fading calibration, the model parsimony, the one-bit quantization, size
coefficient, the DOA, and the delay associated with each of the array, SNR, bandwidth, etc.
of the paths. We also take into account perturbations to the Notation: Boldface lowercase a denotes a vector and bold-
array response, as mentioned above. The channel is modeled face uppercase A denotes a matrix. AT is the transpose of A.
as having a finite duration impulse response composed of a The ith element of a and the (i, j)th entry of A are represented
discrete number of arbitrary delays. To simplify the analysis, by ai and [A]ij respectively. The Hadamard (element-wise)
we do not assume any Doppler spread is present, so the angle product is represented by ⊙, the Kronecker product by ⊗ and
and delay parameters are assumed to be time-invariant over the convolution operation by ⊛. The operation vec(·) denotes
the channel estimation period. We also take into account the the vectorization operation, i.e. the stacking of the columns of
squint effect mentioned above, where for wideband signals a matrix one below the other. Real and imaginary parts are
the time delay from one end of the array to the other cannot given by Re(·) and Im(·) respectively. E[·] is the expectation
simply be represented as a phase shift. Incorporating these operator. The ith row and jth column of the matrix A are
temporal shifts is important because the end-to-end delay for given by A(i,) and A(,j) respectively. A  B and A ≻ B
the antenna array is of the same order as the symbol duration. mean that A − B is positive semidefinite and positive definite
For the unstructured model, the channel is represented as respectively. The cumulative distribution function (cdf) of the
an FIR filter. While this model has a much larger number standard normal distribution is given by Φ(x), and its derivate
of parameters than the structured case when the number of Φ′ (x) is the pdf of the standard normal distribution.
multipaths arrivals at each delay is not large, the model
depends linearly on the parameters, and thus it is easier to II. M M WAVE C HANNEL M ODEL
estimate. The interesting issue to be addressed is whether or
not the increased difficulty in estimating the structured channel We consider an uplink mmWave MIMO system with a
is worth the potential gain in estimation performance. single-antenna user terminal and M receive antennas at the
We also consider a dictionary-based channel model used base station. We assume that the wireless communication
in the compressive sensing literature [6], [10], [22]. The channel is linear and its properties change slowly with respect
dictionary is based on a discrete set of DOAs and path delays to the signal duration. The communication system block
obtained from a grid, where the grid size is greater than the diagram is illustrated in Fig. 1. The received baseband signal
number of antennas. This formulation is commonly used in at the mth antenna is given by
mmWave channel estimation since compressive sensing based p Z ∞
′
algorithms which exploit the underlying sparse multipath ym (t) = Pt h′m (t′ )s(t − t′ )dt′ + wm
′
(t), (1)
−∞
structure can be used. We consider a “dictionary mismatched”
channel model, where the multipath DOAs and delays are where Pt is the transmit signal power, h′m (t) is the impulse
matched to the nearest grid point and the difference between response of the channel from the transmitter to antenna m at
the dictionary and the true source parameters, or the grid ′
time t and wm (t) is the corresponding noise.
mismatch, are parameters to be estimated. The source signal is assumed to be band-limited to
′
We derive the FIM for the parameters of interest assuming [−B/2, B/2], and wm (t) is a complex circularly symmetric
that the “length” of the channel (i.e., the number of discrete and temporally white Gaussian process with power spectral

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
3

′
wm (t)

′ p


s(t) ym (t) ym n + P p
{s[k]} g ′ (t) h′m (t) + g ′′ (t) xm [n + P ]

Transmit Sampling 1-bit

Channel Receive
Filter Quantizer
Filter

p


hm k + P

Fig. 1: System block diagram.

density N0 . The source signal s(t) is encoded as a digital sig- gains of each path to the receiver. Let βr,m be the complex
nal with a common pulse period Ts , where the complex valued path gain of the rth path to the mth antenna so that the channel
symbols {s[k]} are modulated by a pulse shape function g ′ (t) between the source and antenna m is
as ∞
R
X
h′m (t) = βr,m δ(t − (τ0 + (r − 1)∆′ )),
X
s(t) = s[k]g ′ (t − kTs ).
r=1
k=−∞
where τ0 corresponds to the delay of the first multipath
The received signal in (1) is assumed to be fractionally
arrival, and the value of ∆′ is determined by the minimum
sampled by the filter g ′′ (t) by a factor of P , and can be
resolvable time difference between different paths. Under
equivalently represented in sampled time as
this FIR model, the discrete channel can be expressed as
∞ T
h pi p X h pi h pi h[k] = [h1 [k], h2 [k], . . . , hM [k]] , where k = l + Pp , and
ym n + = Pt hm l + s[n−l]+wm n + ,   
P P P β1,1 . . . βR,1 g(kTs − τ0 )
l=−∞
(2) ′
s − (τ0 + ∆ ))
   β1,2 . . . βR,2   g(kT 
  
where p = 0, 1, . . . , P − 1, hm l + Pp is the equivalent h[k] =  . .   . .
 .. ... .. ..
discrete-time channel and wm n + Pp is the discrete-time
 
noise, given by β1,M ... βR,M g(kTs − (τ0 + (R − 1)∆′ ))
| {z }
h Z ∞Z ∞
pi  p  β̃∈CM ×R
hm l + = g ′′ l + Ts − t′′ (3)
P −∞ −∞ P
The parameters of the unstructured model are comprised by
h′m (t′′ − t′ )g ′ (t′ )dt′ dt′′ β, the vector of all complex path gains, i.e. β = vec(β̃), and
Z ∞
we also define β Re = Re (β) ∈ RM R×1 and β Im = Im (β) ∈
h pi  p 
wm n + = g ′′ (t′′ )wm′
n+ Ts − t′′ dt′′ .
P −∞ P RM R×1 .
We will absorb the effects of the transmit and receive filters,
g ′ (t) and g ′′ (t), into g(t) = g ′′ (t) ⊛ g ′ (t). We assume that B. Structured Channel Model
the convolutive channel is frequency selective with maximum The structured channel model is a geometric channel pa-
delay length of L symbol periods, so that the FIR assumption rameterized by path loss components, path delays and DOAs.
means that at most L consecutive
  symbols play a role in the The response of the antenna array to a waveform arriving
received signal, i.e. hm l + Pp is zero outside the interval from direction θ is denoted by a(θ, ρ) ∈ CM ×1 . The vector
[0, L − 1]. Generally, oversampling w.r.t. the Nyquist rate ρ represents parameters on which the array response depends
results in noise correlation [35]. However, if the receive filter (e.g., antenna positions, gain and phase response, etc.). We will
g ′′ (t) is chosen to be a root-Nyquist pulse, the discrete-time assume these parameters are a priori known to be Gaussian
noise wm [n] is white [36]. Consequently, the root-raised- with some nominal mean value ρ0 and covariance Ω.
cosine filter is chosen for the transmit and receive filters in Let the DOA of the rth multipath, measured clockwise with
our analysis, so that respect to the y-axis, be θr , r = 1, 2, . . . , R, and assume that
the antenna elements are close enough together so that they
sin πt/Ts cos παt/Ts
g(t) = , share a common complex path gain γr for the rth path. The
πt/Ts 1 − 4α2 t2 /Ts2 path gain γr is an aggregate of the large-scale fading, namely
where α is the roll-off factor and the noise wm [n] ∼ the path loss and shadowing, as well as the small scale fading.
CN (0, σ 2 ), σ 2 = N0 B. In our analysis, we assume that the More specifically,
noise variance σ 2 is known, since the channel gains and σ 2 are
βr,m = γr qm (θr , ρ)exp(−jωc τr,m ), (4)
not separately identifiable when one-bit quantization is used,
which leads to a singular FIM [30], [31]. where ωc is the carrier frequency and qm (θr , ρ) is the sensor
pattern of the mth sensor in the direction θr . The channel
A. Unstructured Channel Model between the source and the mth antenna is
R
X
In the unstructured case, the channel is modeled as a h′m (t) = γr qm (θr , ρ)exp(−jωc τr,m )δ(t − τr,m ).
uniformly sampled FIR filter characterized by the complex r=1

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
4

Incorporating the transmit and receive pulse shaping, Note: Although the path delays, τr,m , are functions of ρ
PR discrete-time equivalent channel becomes hm [k] =
the (in the case of position perturbation, for example), we do not
r=1 γr qm (θr , ρ)g(kTs − τr,m )exp(−jωc τr,m ) and stacking
consider this effect in the derivation of the FIM since the
the discrete-time channels from all M antennas, we get derivative of τ with respect to ρ is negligible.
 
g(kTs − τr,1 )
X R  g (kTs − τr,2 )  C. Dictionary Based Channel Model
 
h[k] = γr  ..  ⊙ a(θr , ρ), (5) In mmWave transmission, the propagation channel is often
r=1
 . 
described using a sparse scattering model. The underlying
g (kTs − τr,M )
channel is still parameterized by DOAs, path delays and
where ⊙ is the Hadamard (element-wise) product and the complex path gains as in the structured model, but the DOAs
term exp(−jωc τr,m ) in (4) was absorbed into the expression and delays are assumed to lie on a fixed grid, and the channel
for the array response a(θr , ρ) by writing it in terms of estimation is formulated as a sparse recovery problem. This
the array sensor coordinates. Note that (5) uses time delays approach leverages tools available in compressive sensing to
rather than phase shifts in the pulse shaping functions because design efficient algorithms for determining the channel. A grid
the inverse bandwidth of the signals at mmWave frequencies mismatch occurs if a particular DOA is not present in the
may approach the inter-element delay between antennas in the possible DOA set. In our analysis, we model the true DOA as
array. a perturbation to the nearest DOA in the grid. Let the uniform
As an example of the array perturbation parameter ρ, grid of DOAs consist of Na points with Na ≥ M , so that
consider the case of sensor position and pattern perturbation. the DOA dictionary is the set θ′ ∈ {0, 2π/Na , . . . , 2π(Na −
The array response vector a(θr , ρ) can be written as 1)/Na }. Then, a Taylor interpolation of a(θ, ρ) around the
a(θr , ρ) = nearest DOA in the grid, θ′ , yields
 
q1 (θr , ρ)exp (−j2π(x1 (ρ) sin θr + y1 (ρ) cos θr )/λ) ∂a(θ, ρ)
a(θ′ + θ, ρ) = a(θ′ , ρ) + θ . (7)
 ..  ∂θ
 . , θ=θ ′

qM (θr , ρ)exp (−j2π(xM (ρ) sin θr + yM (ρ) cos θr )/λ) Here, we have used θ to denote the grid mismatch between
the nearest DOA grid point and the DOA of the corresponding
where (xm (ρ), ym (ρ)) are the coordinates of the mth sensor. multipath, rather than the multipath DOA as in the structured
In the presence of an antenna pattern perturbation only [34], channel.
we have qm (θr , ρ) = qm,0 (θr ) + ρm , where qm,0 (θr ) is the ′
∈
Similarly, the uniform delay
nominal pattern, ρm is the complex perturbation and ρ stacks n (L−1)Ts o grid is the set τ
(L−1)(Nd −1)Ts
0, Nd , . . . , and a Taylor interpolation
both the real and imaginary parts of the complex perturbation Nd
for all antennas, so that ρ ∈ R 2M ×1
. For sensor position similar to (7) can be performed to obtain
perturbations only, the sensor coordinates can be modeled in ∂g(kTs − τ )
a recursive manner as in [32]. The following piecewise linear g(kTs − τ ′ − τ ) = g(kTs − τ ′ ) + τ . (8)
∂τ τ =τ ′
model applies to flexible array structures: (xm (ρ), ym (ρ)) =
(xm−1 + δ sin φm (ρ), ym−1 + δ cos φm (ρ)), where φm (ρ) = Here, we have used τ to denote the grid mismatch between
φm−1 +ρm−1 , with initial conditions φ1 = π/2, x1 = y1 = 0, the nearest delay grid point and the delay of the corresponding
and δ is the spacing between antenna elements. Thus, ρm is multipath. Having taken into account the dictionary errors, the
the incremental angular perturbation of the mth sensor and discrete-time channel h[k] for the dictionary-based channel
ρ ∈ RM −1×1 . model is given by
A special case of the above modeling is a uniform linear h[k] = (GD (kTs , τ ) ⊗ AD (θ, ρ)) γ, (9)
array operating with a nominal omnidirectional sensor pattern,
qm (θr ) = 1 and ρ = 0. The path delay to the mth sensor where AD (θ, ρ) ∈ CM ×Na and GD (kTs , τ ) ∈ RP ×Nd are
is given by τr,m = τr + (m − 1) δc sinθr , where c is the the angular and delay domain perturbed dictionary matrices,
speed of light and τr = τr,1 is the time delay of the rth respectively, and θ and τ are the vectors of grid mismatch
propagation path to the first antenna element. Let gk (τr , θr ) = errors to be estimated. The columns of GD (kTs , τ ) are of
h  iT the form [g(kTs − τ ′ − τ ), . . . , g((k + PP−1 )Ts − τ ′ − τ )]T ,
g(kTs − τr ), . . . , g kTs − τr − (M −1)δsinθc
r
, a(θr ) = and the columns of AD (θ, ρ) are of the form a(θ′ + θ, ρ).
h i T
2πδ 2πδ
1, e−j λ sinθr , . . . , e−j λ (M −1)sinθr where λ is the wave- The complex unknown vector γ is a sparse Na Nd × 1 vector
T
that carries the path gains from the corresponding DOAs and
length, and γ = [γ1 , γ2 , . . . , γR ] . The lth delay-tap of the delays in the dictionary. The sparse formulation implies that γ
channel can then be written as only has R ≪ Na Nd non-zero elements. In our analysis of the
CRB for this model, and unlike the structured and unstructured
h[k] = [gk (τ1 , θ1 ), gk (τ2 , θ2 ), . . . , gk (τR , θR )] ⊙ models described above, we assume that the locations of the
| {z }
Gk ∈ RM ×R non-zero elements in γ, and therefore the nearest angle and
(6) delay grid points are known a priori. The benefit of this a priori
[a(θ1 ) a(θ2 ) . . . a(θR )] γ.
| {z } information will depend on the resolution of the grid, and the
A ∈ CM ×R ability of dictionary-based methods to correctly identify the
In the derivation of the CRB, we use the following notation: correct grid points. The assumption should be a reasonable
γ Re = Re (γ) and γ Im = Im (γ). one for dictionaries whose grids are not too finely spaced.

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
5

D. System Model A distinction between the unstructured, structured and dictio-

Gathering the received signals at the M antennas from (2) nary based models is that in the structured model and the
in y and the noise in w, dictionary based model, the channel parameters are estimated,
whereas in the unstructured model, the channel h is estimated.
h L−1
pi p X h pi h pi
y n+ = Pt h l+ s[n − l] + w n + ,
P P P III. C RAM ÉR - R AO B OUND
l=0
  In this section, we derive the CRB for parameters
where h l + Pp corresponds to (6) for the structured channel,
(3) for the unstructured channel and (9) for the dictionary of the spatially structured and unstructured channel
based channel model. We collect samples from N source models. For the spatially structured channel model, the
symbol periods, where the coherence time of the channel is parameter vector consists of both deterministic and stochastic
greater than N Ts , and sample the received signal at each components. The deterministic components are, namely,
antenna at P times the symbol rate. We collect the M P × N the DOAs, path delays and complex path gains. When
spatial and temporal samples of the received signal in the considering sensor position perturbations only, we have
M −1×1
matrix Y to get ρ ∈ R , and for pattern perturbations only, we have ρ = 
  Re(ρ1 ), Re(ρ2 ), . . . , Re(ρM ), Im(ρ1 ), Im(ρ2 ), . . . , Im(ρM )
y[0] y[1] ... y[N − 1] 2M ×1
 y[ 1 ] ∈ R Re .Im The  full list of parameters is Θ =
y[1 + P1 ] ... y[N − 1 + P1 ]  θ, τ , γ , γ , ρ under the structured
Y=

..
P
.. .. ..

 h and dictionary-
i
 . . . .  based channel models, and Θ = β Re , β Im for the
y[ PP−1 ] y[1 + PP−1 ] . . . y[N − 1 + PP−1 ] unstructured model. Assuming independent observations, the
 (10)
 log-likelihood for the spatially structured and unstructured
h[0] h[1] ... h[L − 1] models, l(x; Θ), can be derived in a manner similar to [19],
1
p   h[ P ] h[1 + P1 ] ... h[L − 1 + P1 ]  to obtain
= Pt  .. .. .. ..  2M NP  
 . . . .  X 2
l(x; Θ) = ln Φ xk uk ,
h[ PP−1 ] h[1 + P −1
h[L − 1 + P −1 σ
P ] ... P ] k=1
| {z }
H∈CM P ×L
T
  where xk is the kth element of x, uk = s(k,) h, s(k,) is the
s[0] s[1] ... s[N − 1] kth row of S, and Φ(x) is the cumulative distribution function

 s[−1] s[0] ... s[N − 2] 
 of the standard normal distribution.
 .. .. .. ..  +W.
 . . . .  Assuming that the regularity condition of the log-likelihood
s[−L + 1] s[−L + 2] ... s[N − L] holds, the FIM has the following form [37]:
| {z }
S̃ ∈ CL×N
J =JD + JP ,
Vectorizing (10) and taking the real and imaginary parts where JD and JP are the information matrices obtained from
separately, we have the data and the a priori information, respectively. The (i, j)th
  elements of JD and JP are given by
Re (vec(Y))
y= = Sh + w, (11) 
∂2

Im (vec(Y)) [JD ]i,j = − Ex,ρ l(x; Θ)
∂Θi ∂Θj
where,  
  √   √   ∂2
Re S̃T ⊗ Pt IM P −Im S̃T ⊗ Pt IM P [JP ]i,j =Eρ (ρ − ρ0 )T Ω−1 (ρ − ρ0 ) ,
∂Θi ∂Θj
S =  √   √  
Im S̃T ⊗ Pt IM P Re S̃T ⊗ Pt IM P where Ω is the covariance matrix of ρ. Since only ρ is random,
∈R 2M N P ×2LM P
, the matrix JP is given by
 
Re (vec(H)) JP = blkdiag{diag(0D ), Ω−1 },
h= ∈ R2LM P ×1 ,
Im (vec(H))
  where blkdiag{·} is a block-diagonal matrix where the argu-
Re (vec(W)) ments form the diagonal blocks, and D = 4R, D = 2M R
w= ∈ R2M N P ×1 ,
Im (vec(W)) and D = 4R, for the structured, unstructured and dictionary
2 based channel models respectively.
We note that w ∼ N (0, σ2 I). The per-antenna SNR at the
The expectation with respect to the joint distribution of x
receiver is defined as
and ρ in JD is difficult to compute. Instead, we follow the
L−1 P −1
Pt X X   approach in [32], [34]. If the perturbations are small, JD can
SNR = 2 E |hm [l + p/P ]|2 . be approximated to order O(1) around ρ0 , in which case
σ p=0 l=0  
∂2
As in [19], we define the quantization operation as Q(·) = [JD ]i,j ≈ −Ex l(x; Θ) . (12)
√1 (sign(·))), where the sign operation is performed separately ∂Θi ∂Θj ρ=ρ 0
2
for the real and imaginary parts, so the quantized output x is Then, the for the ith parameter of an unbiased estimator
h CRB
i
x = Q(y) = Q(Sh + w). with E Θ̂ = Θ is given by the (i, i) element of the inverse

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
6

of the FIM, where the FIM is computed at the “true” values IV. S IMULATION R ESULTS
of Θ and ρ. That is, We consider a ULA with half-wavelength spacing between
 
var(Θ̂i ) ≥ J−1 (Θ) i,i . antenna elements, i.e. δ = 0.5λ, λ = c/fc with the carrier
frequency fc = 60GHz. The source signal s[n], n = −L +
The approximation in (12) is sufficiently accurate for values of 1, . . . , N − 1, is taken to be a randomly generated quadrature
ρ commonly encountered in real calibrated systems (see [34] phase shift keying (QPSK) sequence. We choose r to be the
for more details). raised cosine filter with roll-off factor α = 0.8. For wideband
systems, the channel bandwidths can be as high as 1GHz [38].
A. Unstructured Channel For this reason, in our simulations the null-to-null bandwidth is
For the unstructured channel model, the FIM is given by set to 1GHz, the symbol duration Ts = 1ns, and the received
 
JβRe JβRe βIm signal is oversampled at the sampling frequency fs = P ×
J= ∈ R2M R×2M R , (13) null − to − null bandwidth. The oversampling factor is P =
JTβRe βIm JβIm
3 in all plots except Fig. 7.
with A comparison of the CRBs of the different channel models
 
JβRe =E ∇βRe l(x; Θ)(∇βRe l(x; Θ))T ∈ RM R×M R , is accomplished by translating the CRB of the parameters Θ
  into the CRB of the channel by the following transforma-
JβIm =E ∇βIm l(x; Θ)(∇βIm l(x; Θ))T ∈ RM R×M R ,
  tion [39]
JβRe βIm =E ∇βRe l(x; Θ)(∇βIm l(x; Θ))T ∈ RM R×M R . CRB(h)  ∇Θ h J−1 ∇TΘ h,
(14)
The expressions for the Jacobians as follows: where ∇Θ (·) is the Jacobian with respect to Θ. The parame-
" #
ters for the structured channel model are selected as follows.
∂h ∂h ∂h
DβRe = ,
Re ∂β Re
, . . . , Re , The angles of arrival of the multipaths are assumed to be
∂β1,1 1,2 ∂βR,M distributed independently and uniformly in [0, 2π). The R
∂h[k] complex gains of the multipaths γr are a. The path delays
Re
= g(kTs − (τ0 + (r − 1)∆′ )), τr , r = 1, 2, . . . , R, are chosen to be integer multiples of
∂βr,m
the sampling interval 1/fs . For the first few simulations, we
DβIm = jDβRe ,
ignore the effect of array perturbations. Similarly, for the
where em is the unit vector with a 1 at the mth index. unstructured channel, the elements of the spatial signatures
βm,r , m = 1, 2, . . . , M, r = 1, 2, . . . , R, are also assumed to
B. Structured and Dictionary Based Channels be generated independently and identically from a complex
The regularity condition for the pdf of x can be easily normal distribution with unit variance, and ∆′ = 1/fs .
verified. For the structured and the grid mismatched dictionary Finally, for the dictionary matrix in (9), we have selected
based channel models, JD is block-partitioned and symmetric Na = 2M and Nd = 2L − 1. The angular and delay domain
and is given by mismatch errors, θ and τ , are generated independently and

Jθ Jθτ Jθγ Re Jθγ Im Jθρ
 uniformly in [− 2N π
, π ] and [− (L−1)T
a 2Na 2Nd ,
s (L−1)Ts
2Nd ], respec-
 JθτT
Jτ Jτ γ Re Jτ γ Im Jτ ρ  tively. The square root of the trace of the CRB matrix for
 T
J J T
J J J
 each of the parameters is computed and averaged over 50
JD =  θγ
 Re τγ Re γ Re Re
γ γ Im γ Re ρ  ,
 JTθγ Im JTτ γ Im JTγ Re γ Im
 realizations of the channel.
Jγ Im Jγ Im ρ 
JTθρ JTτ ρ JTγ Re ρ JTγ Im ρ Jρ
(15) A. Performance vs. SNR
where the expression for each matrix block is provided in Fig. 2 (a) shows the square root of the trace of the channel

T
Appendix B. Let uk = s(k) h, (·)(i,) be the ith row of estimate CRB as a function of the SNR for a single line-of-
the argument and (·)(,i) be the ith column of the argument. sight path (R = 1) to the receiver and a single tap channel
Denoting the Jacobian of h with respect to θ by Dθ , we show (L = 1) with M = 32 receive antennas and pilot length N =
in Appendix A that Jθ can be written as 20. This is the frequency-flat fading case with the unstructured

∂l(x; Θ) ∂l(x; Θ)
  T model corresponding to the Rayleigh fading case.
(,i) (,j)
[Jθ ]i,j =E = Dθ ST Φ D S Dθ , The increase in the bound at high SNR in Fig. 2 (a)
∂θi ∂θj
(16) is commonly observed in one-bit sampled systems, and is
where ΦD √ is a diagonal matrix with [ΦD ]k,k = due to the loss of information in the channel gain as the
Φ′ σ2 uk
2 amplitude of the received signal grows, and the FIM becomes
2
2
σ Φ
√
2
 √  . In a similar manner, we can derive rank deficient. This illustrates the benefit behind dithering
σ uk 1−Φ σ2 uk
(stochastic resonance), where adding noise (lowering the SNR)
expressions for the other sub-matrices of (15) in terms of Dθ ,
can improve estimation performance with coarsely sampled
Dτ , Dγ Re and Dγ Im . For instance, the sub-matrix Jθτ is
data. We also observe that at low-moderate SNRs, which is
given by
T common in mmWave, the gap due to the quantization error
Jθτ = (Dθ ) ST ΦD S Dτ . between the unquantized and one-bit bounds is about 1.96dB
The expressions for the Jacobian matrices are derived in as expected [40]. At higher SNRs, however, the gap is much
Appendix B, and are evaluated at the nominal perturbation higher since the FIM becomes increasingly ill-conditioned due
value ρ0 . to reduced identifiability of the channel gains. The CRBs for

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
7

the unstructured one-bit and unquantized channels are higher

than the other CRBs since they reflect the estimation error
for a total of 64 real-valued parameters compared to only 3
structured and dictionary based channel parameters (2 real-
100
valued path gains and one DOA). Furthermore, the dictionary
based CRB is lower than the unstructured and structured
counterparts. This is because approximate knowledge of the
Tr(CRB( h))

1-bit unstr.
Ideal unstr. DOAs and delays is available and the CRB reflects the esti-
-1
10 1-bit str.
Ideal str.
mation error for the grid mismatch. Fig. 2 (a) provides a very
1-bit dict. interesting observation concerning the structured vs. unstruc-
1-bit dict.
tured models and one-bit quantization. Note that the channel
-2 estimation lower bound for the structured model under one-
10
bit quantization is significantly lower than the bound for the
unstructured model without quantization (perfect resolution),
-20 -10 0 10 20 30 40
provided that the SNR is below 10dB, which would be the
SNR (dB) typical case for mmWave systems. Thus, the gain in parsimony
(a)
provided by the structured model more than compensates for
the loss due to the coarse quantization, and this provides a
strong argument for the use of the structured model when the
10 0 propagation environment is relatively simple.
When the channel is strongly frequency selective (large de-
Tr(CRB( )) (deg)

-1 CRB( )
lay spread L), the received signal power is spread evenly over
10
time, the quantization noise becomes circularly symmetric, and
the resulting amplitude distortion caused by the quantization
10
-2
is reduced (see [41] for details). This is evident in the CRB
Tr(CRB( )),

results shown in Figs. 2 (b) and (c). Fig. 2 (b) illustrates

-3
the CRBs of γ and θ for different channel lengths L = 10,
10
L = 30 and L = 60. The number of pilots is set at N = 80
in all cases. It is seen that as the number of channel taps
-4
10 CRB( )
increases, the degradation in performance due to quantization
at higher SNRs also decreases. For the same number of pilots,
-20 -15 -10 -5 0 5 10 15 20 25
SNR (dB) no effect on the CRB is observed at low-to-moderate SNRs.
The performance of the ideal system is insensitive to L, and
(b) therefore, the dashed curves corresponding to the ideal system
overlap. A final observation regarding Fig. 2 (b) is that the
CRB result for θ is not inconsistent with the CRB results
1-bit L=10
in [30] and [31], but rather our definition of SNR is different
10
1
Ideal L=10 from theirs and, as a result, we are plotting the CRB over a
1-bit L=40
Ideal L=40 wider SNR range. A similar effect is seen Fig. 2 (c) where
1-bit L=60
Tr(CRB( h))

Ideal L=60 the CRB of the unstructured channel is plotted as a function

10
0 of the SNR for different values of L. Here, a single multipath
is considered with the number of pilots fixed to N = 80.
Since the size of the β grows linearly with R and L, the CRB
10
-1 is normalized by M R. Thus, the effect of quantization error
from one-bit ADCs in longer frequency-selective channels is
less severe at high SNRs.
-2
10
-15 -10 -5 0 5 10 15 20 25 30
SNR (dB) B. Performance vs Number of Antennas
(c) The effect of increasing the number of antennas on the CRB
is studied next. In Fig. 3 (a), we plot
p the CRB normalized by
Fig. 2: (a) CRB(h) with 32 antennas and a single LOS path the number of receive antennas ( Tr(CRB(h))/M ) for the
to the receiver. The dashed lines correspond to the CRBs of structured channel case, for varying values of M at different
unquantized (ideal) systems whereas solid lines are used to SNRs and delay tap lengths. The number of pilots is fixed
plot the CRBs of one-bit cases. (b) CRB(θ) and CRB(γ) to N = 60 and the number of multipaths is R = 4.
of the structured channel with 32 antennas and a single LOS As expected, increasing the number of antennas reduces the
path to the receiver. The CRB is plotted for different values average estimation error in each channel coefficient. In [41],
of the channel delay-tap length. (c) CRB(h) of unstructured it was found that, for a given SNR, the number of antennas
channels with 32 antennas for L = 10, L = 40 and L = 60. in one-bit systems with an unstructured channel model should
increase by approximately 2.5 times to meet the achievable

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
8

rate of an otherwise equivalent ideal unquantized system. At

-1
10
SNR = −5dB, fewer than twice the number of antennas
are required for the one-bit system to achieve the same
channel estimation performance as the unquantized system.
For example, the one-bit CRB at M = 80 is equal to the
Tr(CRB( h))/M

unquantized CRB at M = 50. At 5dB SNR, it is seen that the

-2
number of antennas should be increased by slightly more than
10
a factor of two; the one-bit CRB at M = 110 is equal to the
unquantized CRB at M = 50. At 5dB SNR, even the highly
frequency-selective channel (L = 45) has the same CRB as
a channel with L = 15 as seen by the overlapping blue and
black curves. At higher SNRs, it can be expected that one-
10
-3 bit systems are advantageous for highly frequency-selective
0 50 100 150
Number of antennas
channels.
We also plot the CRB as a function of M for unstructured
(a) channels in Fig. 3 (b). Since the number of parameters for
the unstructured channel scales with M (more specifically,
the number of parameters is 2M R), the CRB is normalized
by M R, and N = 80. As expected, increasing the number of
10 -1
antennas reduces the average estimation error in each channel
coefficient. At low SNRs, the number of antennas needed by
Tr(CRB( h))/M

one-bit systems has to again increase by almost two times to

cope with the loss due to quantization distortion. For instance,
at SNR = 0dB, the one-bit system with M = 100 achieves
10
-2 the same CRB as the unquantized system with M = 50.
As the SNR is increased, it is found that the number of
antennas should be increased even more to meet the CRB of
the unquantized system.
20 40 60 80 100 120 140
Fig. 3 (c) shows the CRB for θ as a function of the
Number of antennas number of antennas. The figure also shows the grid spacing
(b) of the dictionary based channel model scaled appropriately. At
SNR = −5dB, the number of antennas should be increased
by less than 1.5 times to achieve the unquantized CRB,
significantly less than for the channel itself in Fig. 3 (a). The
10
-1 one-bit CRB at M = 110 is equal to the unquantized CRB
at M = 90. At 5dB, the factor increases to 1.5, still fewer
than for the channel. Upon examining equation (V), it can
Tr(CRB( ))/M

be seen that the Jacobian scales linearly with M , and thus,

10 -2
the CRB reduces with a factor of M 2 . Comparing Figs. 3 (a)
and (c) with (b), it is seen that, to match the CRB of the
unquantized systems, the unstructured models require almost
10
-3 2.5 − 3 times the number of antennas, whereas the structured
models typically require much less than twice the number of
antennas. Thus, this example illustrates another advantage of
20 40 60 80 100 120 140 using the structured model together with one-bit quantization:
Number of antennas
fewer additional antennas are needed to achieve the same
(c) performance as an ideal unquantized system than in the case
Fig. 3: (a) CRB(h) of the structured channels as a function of an unstructured channel model. Note that in this example,
of the number of receive antennas. For the case with L = 45, the structured model provides DOA estimates that are well
the SNR is 5dB, and for the N = 100 case- L = 15 and beyond the resolution of the grid used in the dictionary-based
SNR = 5dB. For the other plots, L = 15. (b). CRB(h) of the model, especially for larger array sizes.
unstructured channels as a function of the number of receive
antennas for different values of L and SNRs. The constrained C. Effect of Array Calibration Errors
CRB is also illustrated for SNR = 5dB and L = 15. (c) We now consider the effect of array perturbations on the
CRB(θ) of the structured channels as a function of the number CRB. First, we assume a gain pattern perturbation only with
of receive antennas. As in (a), for the case of L = 45, the SNR Ω = σρ2 IM . Here, M = 64, the number of pilots is N = 50,
is fixed to 5dB. The final plot with N = 100 corresponds to the channel length L = 5 and the number of multipaths R = 5.
L = 15 and SNR = 5dB. The dotted line indicates the scaled Fig. 4 shows the CRB of the structured channel for a standard
angular resolution of the dictionary based channel model. deviation of ρ, σρ = 0.1. At σρ = 0.1, the structured model

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
9

1-bit unstr.
101 Ideal unstr.
1-bit str. SD = 0.1
Ideal str. SD = 0.1
1-bit str.
Ideal str.
Tr(CRB( h))

1-bit dict.
100

Tr(CRB( ))
Ideal dict.

10-1

10-1
1-bit SNR = 5dB
Ideal SNR = 5dB
1-bit SNR = 10dB
Ideal SNR = 10dB

10-2
-10 0 10 20 30 10-1 100
SNR (dB)
SD of

Fig. 4: CRB(h) as a function of the SNR for varying standard (a)

deviations (SDs) of the pattern perturbation.
1-bit SNR = 5dB
Ideal SNR = 5dB
1-bit SNR = 10dB
is an appropriate model to assess the channel performance 10-1 Ideal SNR = 10dB

at lower SNRs but the CRB degrades faster. On the other

Tr(CRB( )) (deg)
hand, the unstructured channel is a more suitable model when
the perturbations are large and unknown. Fig. 4 amplifies
the result of Fig. 2 (a), showing that the structured model
with one-bit quantization achieves better channel estimation
performance than the ideal unstructured model up to about
10dB SNR even when there are array perturbations at the level
of σρ = 0.1. Thus, for low-to-moderate SNRs where mmWave
systems operate, structured models that have even imprecise
10-1 100
calibration and use only one-bit quantization perform better SD of
than using unstructured models with perfect quantization.
(b)
We also study the effect of position perturbation on the
CRB. The CRBs of γ and θ as a function of the standard Fig. 5: (a) CRB(γ) as a function of the SD of ρ. (b) CRB(θ)
deviation, σρ , are plotted in Fig. 5 (a) and (b), respectively, antennas as a function of the SD of ρ.
for two different values of the SNR, 5dB and 10dB. It is seen
in Fig. 5 (a) that for small values of σρ , increasing the SNR
is advantageous since the CRB is lower at SNR = 10dB than system exhibits a lower estimation error for structured channel
at SNR = 5dB. However, more interestingly, increasing the parameters than a narrowband system. Similarly, increasing the
perturbation causes the CRB to degrade rapidly at SNR = bandwidth also provides better resolution for the estimation of
10dB due to the near-singularity of the FIM. The same effect the delay parameters τ as seen in Fig. 6(b).
was observed in Fig. 4. Similarly, the CRB of θ also degrades The effect of oversampling on the CRB is shown in Fig. 7.
for SNR = 10dB with an increasing perturbation in Fig. 5 Here, the gap between the one-bit CRB and the unquantized
(b). The gap between the one-bit and unquantized CRBs is CRB is plotted as a function of the oversampling factor. More
greater for SNR = 10dB and it becomes more pronounced specifically, we plot Tr[CRB1−bit (·)]/Tr[CRB∞ (·)] for τ , γ
upon increasing the perturbation. and θ. The number of antennas is kept fixed at 32 and the
oversampling factor above the Nyquist rate is varied from 1
to 5. It is seen that performance loss from quantization is
D. Effect of bandwidth and oversampling the least at SNR = −20dB and increases upon increasing
Fig. 6 (a) and (b) show the CRBs of γ and θ, and τ the SNR. However, the effect of oversampling is to reduce
respectively, as a function of the signal bandwidth for M = 32 the loss beyond the 2/π limit. At low SNRs, for example
and M = 64 in the structured channel model. The delay spread at −20dB, increasing the oversampling factor does not have
is fixed to 0.2µs and the SNR is 0dB. The number of pilots is any effect and the plots for all parameters coincide. However,
N = 60 and it is assumed that R = 4 multipaths are present. improvements would still be possible at low SNRs if the
The null-to-null bandwidth is varied from 1MHz to 1GHz and analog filter prior to sampling is optimized [42].
the oversampling factor is kept at P = 3. Since the delay
spread is kept constant, the channel length increases with the
bandwidth, making the channel more frequency-selective. For E. Effect of path separation and number of multipaths
both γ and θ, increasing the BW from 1MHz to 1GHz de- A disadvantage of the structured channel model is that the
creases the CRB by almost an order of magnitude. Therefore, FIM becomes ill-conditioned when two multipaths arrive with
for a given number of estimation parameters, a broadband similar DOAs and path delays. Fig. 8 illustrates the CRB as a

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
10

0
10 SNR = -5dB
SNR =-20dB
0.65

0.6
Tr(CRB( ), Tr(CRB( )) (deg)

10 -1 SNR = 0dB
0.55
, 1-bit M=32 , 1-bit M=32
, Ideal M=32 , Ideal M=32 0.5
, 1-bit M=64 , 1-bit M=64
, Ideal M=64 , Ideal M=64
-2 0.45 SNR = 5dB
10
0.4

0.35
10 -3 0.3

0.25

-4
0.2
10 1.5 2 2.5 3 3.5 4 4.5
10 6 10 7 10 8 10 9 Oversampling factor
BW (Hz)

(a) Fig. 7: χp = Tr[CRB 1−bit (p)]

Tr[CRB∞ (p)] for p ∈ {γ, θ, τ } of structured
channels as a function of the oversampling factor, P . The
10-4
plots with the ▽ and ∗ markers represent parameters τ and
-5
1-bit M=32
Ideal M=32
γ, respectively. The plots for θ are unmarked. For SNR =
10
1-bit M=64 −20dB, χ for all parameters are equal.
Ideal M=64
-6
Tr(CRB( )) (s)

10

-7
10
1-bit str.
Ideal str.
-8 1-bit unstr.
10
Ideal unstr.

10-9
Tr(CRB( h))

-10 100
10
106 107 108 109
BW (Hz)

(b)
Fig. 6: (a) CRB(γ) and CRB(θ) of the structured channels as
a function of the system bandwidth for a fixed delay spread.
(b) CRB(τ ) as a function of the system bandwidth. 10
-3
10
-2
10
-1
10
0

Fig. 8: The CRB of the channel as a function of the fractional

function of the fractional path delay difference. In this setup, separation between path delays. The SNR is 0dB, M = 64,
R = 2 and the two paths have different path gains but are N = 40 and L = 10.
configured to have the same DOA under the structured model.
It is seen that as the path delays become closer, the structured
model is not accurate and the unstructured model is better directions of arrival and path gains to the receiver, (2) the
suited for resolving the two paths. When the paths are well unstructured channel which is appropriate when the number
separated in time, the structured channel model yields better of paths to the receiver is large and estimation of the path
results. For the given example, it is seen that the structured parameters is too difficult, and (3) the dictionary-based formu-
model only begins to break down when the difference in path lation with grid mismatch which is useful for sparse mmWave
delays is on the order of 3 − 4% of the symbol period, but this multipath channels. Closed-form expressions for the FIMs
illustrates the necessity of choosing the proper model order for were derived when the receiver uses a uniform linear antenna
the structured case, which is the key drawback of this method. array and a root-raised cosine filter. The derivation for the FIM
pFig. 9 (a) and (b) show the normalized CRB for the structured and dictionary based channels also included
( Tr(CRB(h))/R) as a function of the number of the effect of array calibration errors, and the cases of gain
multipaths for M = 64, N = 80 and L = 25. It is seen that pattern and antenna position perturbations were considered as
the CRB increases almost linearly on the log scale as the specific examples. A number of numerical experiments were
number of parameters also increases linearly with R. performed to evaluate the CRBs. A comparison in the CRBs of
the structured, unstructured and dictionary-based channels as a
V. C ONCLUSION function of the SNR indicated that the structured CRB is lower
In this work, we considered performance bounds for channel than that of the unstructured channel since fewer parameters
estimation in one-bit mmWave massive MIMO systems. Three contribute to the expression of the CRB. The effects of
channel models were considered, namely, (1) the structured perturbation, bandwidth, the channel length, and the number of
channel where the scattering environment consists of a small receive antennas on the CRBs of the one-bit quantized system
number of scatterers characterized by distinct path delays, were also considered. It was found that pertubation caused the

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
11

A PPENDIX A
0
10
Tr(CRB( ))/R (deg)
For the structured channel,
 
∂l(x; Θ) ∂l(x; Θ)
10
-1
[Jθ ]i,j = E
∂θi ∂θj
"2M N 2M N
′ 2

X X 4 Φ ′ 2
-2 σ xk uk
Φ σ xl ul

10 =E xk xl
Tr(CRB( ))/R,

σ 2 Φ σ2 xk uk Φ σ2 xl ul
k=1 l=1
 T  T 
-3 (,i) (k,) (l,) (,j)
10 Dθ s s Dθ

!2 
2M
XN 2xk Φ′ 2 xk uk T  T
-4
10 σ (,i) (k,) (k,) (,j)
4 6 8 10 12 14 16 18 20 =E 
D θ s s Dθ
Number of multipaths σ Φ σ2 xk uk
k=1
(a) 2M


XN X 4 Φ′ 2 xk uk Φ′ 2 xl ul
σ σ
+

xk xl
σ 2 Φ σ2 xk uk Φ σ2 xl ul
k=1 l6=k
10
0  T  T 
(,i) (k,) (l,) (,j)
Dθ s s Dθ
 √ 2
Tr(CRB( h))/R

2M
XN 2 Φ′ σ2 uk
10 -1
1-bit unstr. SNR=-5dB = √  √ 
Ideal unstr. SNR=-5dB σ2 Φ 2
u 1−Φ 2
u
k=1 σ k σ k
1-bit unstr. SNR=5dB
Ideal unstr. SNR=5dB  T  T
(,i) (,j)
10
-2 1-bit unstr. M=128 Dθ s(k,) s(k,) Dθ
Ideal unstr. M=128
 T
(,i) (,j)
= Dθ ST Φ D S Dθ .
2 4 6 8 10 12 14 16 18 20
Number of multipaths
Other sub-matrices can be derived in a similar manner.
(b)
Fig. 9: (a) CRB(γ) and CRB(θ) as a function of the number A PPENDIX B
of multipaths. For the M = 128 plots, the SNR = 5dB. (b) The Jacobian matrices are computed below.
CRB(h) of unstructured channels as a function of the number
 
of multipaths. ∂h ∂h ∂h
Dθ = , ,...,
∂θ ∂θ ∂θ
 1  2  R
∂H
∂h  Re vec  ∂θr  
=
∂θr Im vec ∂θ ∂H
r

CRB to degrade and approach that of an unstructured model at  

∂h[0] ∂h[1] ∂h[L−1]
high SNRs. However, at low-to-moderate per-antenna SNRs, ∂θr ∂θr ... ∂θr
which is common in mmWave, the structured one-bit channel  ∂h[ P1 ] ∂h[1+ P1 ] ∂h[L−1+ P1 ] 
∂H  ∂θr ∂θr ... ∂θr

models have better channel estimation performance than the 
=

.
∂θr  .. .. .. ..
unquantized unstructured models. Furthermore, increasing the  . . . . 

bandwidth and the oversampling factor caused the estimation ∂h[ PP
−1
] ∂h[1+ PP−1
] ∂h[L−1+ PP
−1
]
∂θr ∂θr ... ∂θr
error variance to decrease due to improved temporal resolution.
It is also seen that to achieve the same error variance as an It is sufficient to compute one of the blocks since the other
unquantized system, the one-bit structured system required blocks can be computed in a recursive manner:
significantly less than twice the number of antennas.
 ∂g(kT s−τr,1 ) 
∂θr
One of the principal observations of our results is that a  ∂g(kTs −τr,2 ) 
significantly lower channel estimation error can be achieved ∂h[k]  ∂θr 
 ⊙ a(θr , ρ)+
=γr  ..
by using a structured rather than an unstructured channel ∂θr 
 .


model, even when the underlying array calibration is not ∂g(kTs −τr,M )
∂θr
precisely known. Thus, the extra computational cost required  
for the resulting non-linear optimization is often well worth the g(kT s − τr,1 )
effort. The study of dependencies between the various system 
 g (kTs − τr,2 )  ∂a(θ , ρ)
 r
parameters can be useful in the design of channel estimation
γr  .. ⊙ .
 .  ∂θr
for mmWave massive MIMO with one-bit quantizers at the g (kTs − τr,M )
receiver.

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
12

If we further assume a uniform linear array, the algebra is perturbations with the pattern perturbation independent of the
straightforward and, at the nominal array perturbation, we get DOA, ρ = [Re(ρ1 ), . . . , Re(ρM ), Im(ρ1 ), . . . , Im(ρM )]T and
 
  g(kT s − τr,1 )
0 R  g (kTs − τr,2 ) 
 ∂g ( kTs −τr − δ sin θr
)  ∂h[k] X  
 c
 = γr  .. ⊙
∂h[k]  ∂θr  ∂Re(ρm ) ρ=0  . 
=γr  ..  ⊙ a(θr , ρ0 )+ r=1
∂θr  .  g (kTs − τr,M )
 δ(M −1) sin θr

∂g (kTs −τr − c ) exp(−j2π(xm sin θr + ym cos θr )/λ)em ,
∂θr

g(kT s − τr )
 ∂h[k] ∂h[k]

=j .
 g kTs − τr − δ sinc θr  ∂Im(ρm ) ρ=0 ∂Re(ρm ) ρ=0
  ∂a(θr , ρ0 )
γr  .. ⊙ . Similarly, for nominal uniform linear arrays with position

  . 
  ∂θr
δ(M −1) sin θr perturbation only we have [32]
g kTs − τr − c  
g(kT s − τr,1 )
R
 g (kTs − τr,2 )

∂h[k] X  
The derivative of g kTs − τr − δ sinc θr with respect to θr is = γr  ..

 ⊙ (j2πδ cos θr /λ)
shown in (17) at the top of the next page. The derivatives of the ∂ρm ρ=0  r=1 . 
array steering vector can also be computed in a similar manner. g (kTs − τr,M )
For the above uniform linear array, under nominal perturbation M −m
!
X
values (qm,0 = 1, ρm = 0 for the pattern pertubation when the diag(a(θr , 0)) kek+m .
perturbation is not a function of the DOA, and ρm = 0 for k=1
the sensor position perturbation case), we have (18)
For arrays which are not uniform linear arrays, expressions for
∂a(θr , ρ0 ) −j2πδ cos θr the derivatives can be derived in a manner similar to (18).
=
∂θr λ
 T R EFERENCES
0, 1, . . . , M −1 ⊙ a(θr , ρ0 ).
[1] M. S. Ullah and A. Tewfik, “Pilot aided direction of arrival estimation
for mmWave cellular systems,” in Proc. IEEE Int’l Conf. on Acoustics,
The Jacobian of h with respect to τ , Dτ is Speech and Signal Processing (ICASSP), March 2016, pp. 3401–3405.
[2] A. M. Sayeed, “Deconstructing multiantenna fading channels,” IEEE
  Trans. Signal Process., vol. 50, no. 10, pp. 2563–2579, 2002.
∂h ∂h ∂h ∂h [3] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao,
Dτ = , ,..., , ,
∂τ1,1 ∂τ1,2 ∂τR,M −1 ∂τR,M and J. I. Tamir, “Broadband millimeter-wave propagation measurements
     and models using adaptive-beam antennas for outdoor urban cellular
∂h Re vec ∂τ∂H r,m
communications,” IEEE Trans. Antennas Propag., vol. 61, no. 4, pp.
=     , 1850–1859, 2013.
∂τr,m Im vec ∂τ∂H [4] L. You, X. Gao, A. L. Swindlehurst, and W. Zhong, “Channel acquisition
r,m for massive MIMO-OFDM with adjustable phase shift pilots,” IEEE
∂h[k] h iT Trans. Signal Process., vol. 64, no. 6, pp. 1461–1476, March 2016.
∂g(kTs −τr,m )
= γr 0 . . . ∂τr,m . . . 0 ⊙ a(θr , ρ). [5] J. Mo, P. Schniter, and R. W. Heath, “Channel estimation in broadband
∂τr,m millimeter wave MIMO systems with few-bit ADCs,” IEEE Trans.
Signal Process., 2017.
[6] K. Venugopal, A. Alkhateeb, N. G. Prelcic, and R. W. Heath, “Channel
The derivative of g (kTs − τr,m ) with respect to τr,m is estimation for hybrid architecture-based wideband millimeter wave
shown in (17) at the top of the next page. For a ULA, systems,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1996–2009,
τ = [τ1 , . . . , τR ]T , τr,m = τr + δ(m − 1) sin θr /c and the 2017.
[7] M. Kokshoorn, H. Chen, Y. Li, and B. Vucetic, “RACE: A rate adaptive
Jacobian can be derived in a straightforward manner. We can channel estimation approach for millimeter wave MIMO systems,” in
derive Dγ Re , Dγ Im and Dρ from the following expressions, Proc. Globecom, 2016.
[8] Y. Ding and B. D. Rao, “Dictionary learning-based sparse channel
  representation and estimation for FDD massive MIMO systems,” IEEE
g(kT s − τr,1 ) Trans. Wireless Commun., vol. 17, no. 8, pp. 5437–5451, 2018.
∂h[k]  g (kTs − τr,2 ) 
 ∂h[k] ∂h[k] [9] H. Ghauch, T. Kim, M. Bengtsson, and M. Skoglund, “Subspace esti-
= ..  ⊙ a(θr , ρ), = j Re , mation and decomposition for large millimeter-wave MIMO systems,”
∂γrRe  .  ∂γrIm ∂γr IEEE J. Sel. Topics Sig. Proc., vol. 10, no. 3, pp. 528–542, April 2016.
[10] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel
g (kTs − τr,M ) estimation and hybrid precoding for millimeter wave cellular systems,”
 
g(kT s − τr,1 ) IEEE J. Sel. Topics Sig. Proc., vol. 8, no. 5, pp. 831–846, 2014.
R [11] Z. Gao, C. Hu, L. Dai, and Z. Wang, “Channel estimation for millimeter-
 g (kTs − τr,2 )
∂h[k] X   ∂a(θ , ρ)
 r wave massive MIMO with hybrid precoding over frequency-selective
= γr  .. ⊙ . fading channels,” IEEE Commun. Lett., vol. 20, no. 6, pp. 1259–1262,
∂ρd r=1
 .  ∂ρd 2016.
g (kTs − τr,M ) [12] S. Haghighatshoar and G. Caire, “Enhancing the estimation of mmWave
large array channels by exploiting spatio-temporal correlation and sparse
scattering,” in Proc. Int. ITG Workshop on Smart Antennas, March 2016.
Let em ∈ RM ×1 be the unit vector with a 1 at the mth [13] S. Buzzi and C. D’Andrea, “Subspace tracking algorithms for millimeter
index. There are 2M R perturbation parameters in the case wave MIMO channel estimation with hybrid beamforming,” in Proc. Int.
ITG Workshop on Smart Antennas, March 2017.
of pattern perturbation and M − 1 parameters in the case [14] B. Murmann, “Energy limits in A/D converters,” in IEEE Faible Tension
of position perturbation. If we consider the pattern only Faible Consommation (FTFC). IEEE, 2013.

1932-4553 (c) 2019 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/JSTSP.2019.2933163, IEEE Journal
of Selected Topics in Signal Processing
13

 

π(kTs −τr − δ sin θr
) (kTs −τr − δ sin θr
)
∂g kTs − τr − δ sin θr
cos πα(kTs − τr − δ sinc θr )/Ts

δ cos θr
 cos Ts
c
− sinc Ts
c
c
= −
+
∂θr 1 − 4α2 (kTs − τr − δ sinc θr )2 /Ts2 c π(kTs − τr − δ sin θr
c )/Ts
" δ sin θr 2 δ sin θr
δ sin θr
(kTs − τr − c ) δ (1 − 4α2 (kTs − τr − c ) /Ts2 ) πα
Ts sin πα(kTs − τr − c )/Ts
sinc cos θr
Ts c (1 − 4α (kTs − τr − δ sinc θr )2 /Ts2 )2
2
#
8α 2
(kTs − τr − δ sinc θr )/Ts2 cos πα(kTs − τr − δ sinc θr )/Ts
−
(1 − 4α2 (kTs − τr − δ sinc θr )2 /Ts2 )2
(17)
 
s −τr )  cos π(kTs −τr,m ) − sinc (kTs −τr,m )
cos πα(kT

∂g(kTs − τr,m ) Ts −π T s T s
= 2 2 +
∂τr,m 1 − 4α (kT s −τr )
Ts2
T s π(kT s − τ r,m )/T s
h  i
4α2 (kTs −τr,m ) πα(kTs −τr,m ) 8α2 (kTs −τr,m ) πα(kTs −τr,m )
(kTs − τr,m ) 1 − Ts2 sin Ts − Ts2 cos Ts
sinc 4α2 (kTs −τr,m )2 2
Ts (1 − 2 ) Ts

[15] R. H. Walden, “Analog-to-digital converter survey and analysis,” IEEE with fixed and time-varying thresholds,” in Asilomar Conf. on Signals,
J. Sel. Areas Commun., vol. 17, no. 4, pp. 539–550, 1999. Systems, and Computers, Oct 2017, pp. 1056–1060.
[16] J. Mo and R. W. Heath, “Capacity analysis of one-bit quantized MIMO [32] M. Viberg and A. L. Swindlehurst, “A Bayesian approach to auto-
systems with transmitter channel state information,” IEEE Trans. Signal calibration for parametric array signal processing,” IEEE Trans. Signal
Process., vol. 63, no. 20, pp. 5498–5512, Oct 2015. Process., vol. 42, no. 12, pp. 3495–3507, 1994.
[17] J. Singh, S. Ponnuru, and U. Madhow, “Multi-gigabit communication: [33] B. Wahlberg, B. Ottersten, and M. Viberg, “Robust signal parameter
The ADC bottleneck,” in IEEE International Conference on Ultra- estimation in the presence of array perturbations,” in Proc. IEEE Int’l
Wideband, 2009, pp. 22–27. Conf. on Acoustics, Speech and Signal Processing (ICASSP). IEEE,
[18] J. Nossek and M. T. Ivrlač, “Capacity and coding for quantized MIMO 1991, pp. 3277–3280.
systems,” in Proc. Int. Wireless Commun. and Mobile Comput. ACM, [34] J.X. Zhu and H. Wang, “Effects of sensor position and pattern
2006, pp. 1387–1392. perturbations on CRLB for direction finding of multiple narrow-band
[19] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, sources,” in IEEE ASSP Workshop On Spect. Estim. and Modeling,
and L. Liu, “Channel estimation and performance analysis of one-bit 1988, pp. 98–102.
massive MIMO systems,” IEEE Trans. Signal Process., vol. 65, no. 15, [35] M. Schluter, M. Dorpinghaus, and G. P. Fettweis, “Bounds on channel
pp. 4075–4089, Aug 2017. parameter estimation with 1-bit quantization and oversampling,” in Proc.
[20] F. Wang, J. Fang, H. Li, Z. Chen, and S. Li, “One-bit quantization IEEE. Int. Workshop on Sig. Proc. Advances in Wireless Communica-
design and channel estimation for massive MIMO systems,” IEEE Trans. tions (SPAWC), June 2018.
Vehicular Technol., vol. 67, no. 11, pp. 10921–10934, Nov 2018. [36] C. Mollén, High-End Performance with Low-End Hardware, Linköping
[21] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer, Studies in Science and Technology. PhD Dissertation. 2018.
“Throughput analysis of massive MIMO uplink with low-resolution [37] H. L. Van Trees, Detection, estimation, and modulation theory, Part I,
ADCs,” IEEE Trans. Wireless Commun., vol. 16, no. 6, pp. 4038–4051, John Wiley & Sons, 2004.
June 2017. [38] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wide-
[22] A. Mezghani and A. L. Swindlehurst, “Blind estimation of sparse band millimeter-wave propagation measurements and channel models for
broadband massive MIMO channels with ideal and one-bit ADCs,” IEEE future wireless communication system design,” IEEE Trans. Commun.,
Trans. Signal Process., vol. 66, no. 11, pp. 2972–2983, June 2018. vol. 63, no. 9, pp. 3029–3056, 2015.
[23] M. Cai, K. Gao, D. Nie, B. Hochwald, J. N. Laneman, H. Huang, and [39] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
K. Liu, “Effect of wideband beam squint on codebook design in phased- Theory, Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 1993.
array wireless systems,” in Proc. Globecom, Dec 2016. [40] A. Mezghani, F. Antreich, and J. Nossek, “Multiple parameter estimation
[24] R. L. Haupt, “Antenna arrays in the time domain: An introduction to with quantized channel output,” in Proc. Int. ITG Workshop on Smart
timed arrays.,” IEEE Antennas Propag. Mag., vol. 59, no. 3, pp. 33–41, Antennas. IEEE, 2010.
June 2017. [41] C. Mollén, J. Choi, E. G. Larsson, and R. W. Heath, “Uplink
[25] B. Wang, F. Gao, S. Jin, H. Lin, and G. Y. Li, “Spatial- and frequency- performance of wideband massive MIMO with one-bit ADCs,” IEEE
wideband effects in millimeter-wave massive MIMO systems,” IEEE Trans. Wireless Commun., vol. 16, no. 1, pp. 87–100, 2017.
Trans. Signal Process., vol. 66, no. 13, pp. 3393–3406, July 2018. [42] M. Stein, F. Wendler, A. Mezghani, and J. A. Nossek, “Quantization-loss
[26] B. Wang, F. Gao, S. Jin, H. Lin, G. Y. Li, S. Sun, and T. S. Rappaport, reduction for signal parameter estimation,” in Proc. IEEE Int’l Conf.
“Spatial-wideband effect in massive MIMO with application in mmWave on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2013, pp.
systems,” IEEE Commun. Mag., vol. 56, no. 12, pp. 134–141, Dec 2018. 5800–5804.
[27] S. Fortunati, F. Gini, M. S. Greco, and C. D. Richmond, “Performance
bounds for parameter estimation under misspecified models: Fundamen-
tal findings and applications,” IEEE Signal Process. Mag., vol. 34, no.
6, pp. 142–157, Nov 2017.
[28] M. S. Stein, J. A. Nossek, and K. Barbé, “Fisher information lower
bounds with applications in hardware-aware nonlinear signal processing,
compression and inference,” CoRR, vol. abs/1512.03473, 2015.
[29] M. Stein, K. Barbe, and J. A. Nossek, “DOA parameter estimation with
1-bit quantization bounds, methods and the exponential replacement,”
in Proc. Int. ITG Workshop on Smart Antennas, March 2016.
[30] F. Liu, H. Zhu, J. Li, P. Wang, and P. V. Orlik, “Massive MIMO channel
estimation using signed measurements with antenna-varying thresholds,”
in IEEE Stat. Sig. Proc. Workshop (SSP), June 2018, pp. 188–192.
[31] P. Wang, J. Li, M. Pajovic, P. T. Boufounos, and P. V. Orlik, “On
angular-domain channel estimation for one-bit massive MIMO systems

1932-4553 (c) 2019 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.