1

Near-Field Localization and Sensing with Large-Aperture

Arrays: From Signal Modeling to Processing
Zhaolin Wang, Parisa Ramezani, Yuanwei Liu, and Emil Björnson

The signal processing community is currently witnessing a growing interest in near-field signal pro-
cessing, driven by the trend towards the use of large aperture arrays with high spatial resolution in the
arXiv:2406.10941v1 [eess.SP] 16 Jun 2024

fields of communication, localization, sensing, imaging, etc. From the perspective of localization and
sensing, this trend breaks the basic far-field assumptions that have dominated the array signal processing
research in the past, presenting new challenges and promising opportunities.

I. I NTRODUCTION

Localization and sensing (L&S) are critical functionalities in modern wireless systems, particularly
with the advent of advanced multiantenna technologies. Localization refers to the process of acquiring
the states, such as location and velocity, of connected devices within wireless networks. Tracing back to the
inception of first-generation (1G) standards, this capability has been historically viewed as an optional
function of wireless cellular networks [1]. Over the decades, various localization methods have been
developed which include terrestrial techniques that utilize time of arrival (TOA) or angle of arrival (AOA)
measurements, and satellite-based approaches represented by the Global Positioning System (GPS). These
approaches rely on sending signals from a source to a receiver, and processing the received signal to
extract relevant state information. Sensing, on the other hand, refers to the estimation of the states of
passive, non-cooperative reflecting objects, similar to traditional radar systems. Sensing is expected to
become a native function in next-generation wireless networks because it enables new services and is
made possible since base stations will have similar multiantenna and spectrum features as digital radars.
L&S techniques have advanced with multiantenna technologies, evolving from phased arrays to digital
antenna arrays, and continuously moving towards larger multiantenna systems. The concept of using
multiple antennas, or arrays, for enhancing target localization and detection dates back to the early
20th century, with significant advancements during World War II for radar applications. These early
developments laid the groundwork for understanding how antenna arrays could be manipulated to improve
directionality and gain, which are crucial for L&S. The development of phased-array antennas for radar

Zhaolin Wang and Yuanwei Liu are with the School of Electronic Engineering and Computer Science, Queen Mary University
of London, London, U.K. (email: {zhaolin.wang, yuanwei.liu}@qmul.ac.uk). Parisa Ramezani and Emil Björnson are with KTH
Royal Institute of Technology, Stockholm, Sweden (email:{parram,emilbjo}@kth.se).
 2

in the 1950s was a significant milestone [2]. In this era, a new initiative is to steer the direction of
the beam electronically without moving the physical antennas, thereby greatly enhancing the speed
and flexibility of L&S. With the advancement of antenna array technologies, array signal processing
gradually became an active research area since the 1980s [3]. This period, extending into the 1990s,
witnessed a surge of classical L&S methods based on array signal processing, including the famous
MUSIC (MUltiple SIgnal Classification) method, parametric MLE (Maximum-Likelihood Estimation)
method, and ESPRIT (Estimation of Signal Parameters via Rotational Invariance Technique) method.
These methods laid an important foundation for array signal processing techniques for L&S. The turn
of the century brought about a digital revolution, further transforming the capabilities of antenna arrays.
The integration of digital signal processing with antenna array technology led to the development of
adaptive antenna systems capable of dynamically adjusting their patterns to optimize L&S accuracy.
This era also witnessed the advent of Multiple-Input Multiple-Output (MIMO) systems [4], which use
multiple antennas at both the transmitter and receiver to improve communication performance through
beamforming, spatial multiplexing, and diversity. Building on these advancements, the concept of MIMO
radar emerged in the early 2000s [5], allowing each antenna to transmit different waveforms and thus
providing more adaptive array configuration and higher L&S resolution.
The electromagnetic (EM) radiation field emitted by antennas or antenna arrays can be divided into
two regions: the far-field (Fraunhofer) region and the radiative near-field (Fresnel) region [6], [7]. The
demarcation between these regions is traditionally defined by the phase error-based Fraunhofer distance
(also known as the Rayleigh distance), which is given by
2D2
dFA = , (1)
λ
where D denotes the aperture of the antenna (array) and λ is the signal wavelength. The exact boundary
between these regions can be defined differently depending on the application. Regardless, when the
emitted signal propagates into the far-field region, its wavefront is approximated as planar, which implies
that the receiving antenna array can only resolve the direction of the signal. However, the accurate
spherical-wavefront model should be used to characterize the signal propagation within the near-field
region, where the curvature is distance-dependent. This property allows the antenna array to determine
both the direction and the propagation distance of the signal. Research on L&S of sources or targets in
the near-field region dates back to as early as the 1980s [8]. However, this research topic did not attract
widespread attention during that era. This is because the array aperture and the signal frequency were
generally limited, rendering the sources or targets normally located in the far-field region.
The landscape shifted with the introduction of the massive MIMO concept by Marzetta in 2010 [9].
 3

Target Target

Wavefront

Wavefront
Origin Origin
Array Array

(a) Near-field model (b) Far-field model

Figure 1: The schematic of the near-field and far-field models under spherical wavefront and planner
wavefront assumptions, respectively.

The core idea of massive MIMO is to equip base stations with numerous antennas connected to fully
digital transceiver branches, allowing them to generate many simultaneous narrow beams with high
controllability. Massive MIMO was developed for communications but builds on enhancing the spatial
resolution compared to smaller arrays, which also significantly improves L&S accuracy. These favorable
features have made an elementary form of massive MIMO with 32-64 antennas the fundamental building
block of 5G networks, typically operating in the 3.5 GHz band. Since the emergence of 5G massive
MIMO, researchers have continued to envision using even larger arrays with various form factors and
characteristics, inevitably increasing the physical size of the antenna arrays. Coupled with the trend
towards exploiting high-frequency bands (centimeter-, millimeter-waves, and beyond), which have very
short wavelengths, the near-field region around base stations can be significantly expanded. For instance,
an antenna array with a 1-meter aperture operating at 30 GHz can have a near-field region extending up
to 200 meters, making it highly likely for targets to be located within this near-field region. This article
aims to provide a tutorial review on near-field L&S from signal modeling to signal processing. Our goal is
to highlight the most significant changes from traditional far-field L&S and to elucidate the key intuitive
ideas for leveraging the advantages of near-field L&S and addressing the associated challenges.

II. E ND - TO - END SIGNAL MODELING

For L&S, the location and velocity of targets are of the most interest to acquire. The method of obtaining
these parameters depends highly on the channel structure; more structure leads to higher accuracy. In this
section, we step through the end-to-end signal modeling of near-field L&S from fixed to moving targets,
and reveal how near-field effects influence the way antenna arrays observe target parameters.
 4

A. Fixed targets

We begin with signal modeling for the L&S system with a fixed target, using a two-dimensional
coordinate system for simplicity. As shown in Figure 1, assume the L&S node, equipped with an N -
antenna large-aperture array, is centered at the origin with its n-th antenna located at sn ∈ R2 . The
aperture of the array is defined as D = maxm,n ∥sn − sm ∥, where ∥ · ∥ is the Euclidean norm. The
target’s location is denoted by p = ru(θ), where r is the distance to the target from the origin, θ is the
angular direction from the origin, and u(θ) = [cos θ, sin θ]T is the direction vector. We focus on a simple
L&S system with a single-antenna transmitter, allowing the L&S to be described using a unified model.
Let xpb (t) represent the transmitted passband pilot signal at time t ∈ R, modulated to a carrier frequency
fc . In the considered free-space propagation scenario, the passband signal received by the n-th antenna
from the target can be modeled as follows:
β
ypb,n (t) = xpb (t − τn + τo ) , (2)
rn
where β denotes the channel gain at a reference distance of 1 meter, τn is the signal delay from the
target to the n-th antenna, and τo is the sampling clock offset at the receiver (which can be used for
synchronization). The signal delay τn = rn /c depends on the distance rn from the target to the n-th
antenna and the speed of light c. As shown in Figure 1(a), the distance rn in near-field systems must be
calculated as the exact Euclidean distance [7]:
q
rn = ∥sn − p∥ = r2 − 2rsTn u(θ) + ∥sn ∥2 . (3)

In wireless systems, most processing is conducted at the baseband, making it beneficial to abstract the
carrier frequency and develop a baseband equivalent of the signal model. The baseband equivalents of
√
the transmit and receive signals, yn (t) and x(t), can be defined via ypb,n (t) = ℜ 2yn (t)ej2πfc t and
√ √
xpb (t) = ℜ 2x(t)ej2πfc t , where ℜ{·} represents the real part of a complex number and j = −1.
Including the thermal noise wn (t), the baseband equivalent model is
β
yn (t) = x(t − τn + τo )e−j2πfc (τn −τo ) + wn (t). (4)
rn
In narrowband systems, it then holds approximately that x(t − τn + τo ) ≈ x(t − r/c + τo ) due to the
slow variation of the baseband signal across the array aperture. Then, by combining the signals from all
antennas into a vector y(t) = [y1 (t), . . . , yN (t)]T and defining s(t) = β/r · x(t − r/c + τo )e−j2πfc (r/c−τo )
as the common equivalent source signal received at all antennas, we obtain the narrowband model for
near-field L&S:
y(t) = a(θ, r)s(t) + w(t), (5)

where w(t) = [w1 (t), . . . , wN (t)]T and a(θ, r) is the near-field array response vector
 5

 T
r −j 2π (r1 −r) r −j 2π (rN −r)
a(θ, r) = e λ ,..., e λ (6)
r1 rN
with λ = c/fc denoting the carrier wavelength. Note that we write the near-field array response vector as
a function of the target’s direction θ and distance r due to the dependence of rn on them. The amplitude
variations in the near-field array response vector are often omitted since r/rn ≈ 1 holds whenever r is
substantially larger than the array aperture length [6], [7]. The model in (5) can be easily extended to
multi-target cases. If there are K targets, the receive signal becomes

y(t) = A(θ, r)s(t) + w(t), (7)

where A(θ, r) = [a(θ1 , r1 ), . . . , a(θK , rK )] and s(t) = [s1 (t), . . . , sK (t)]T denote the array response
matrix and equivalent source signals of K targets. The noise w(t) is generally assumed to be a white
circular-symmetric complex Gaussian random process with a covariance matrix σ 2 IN .

N EAR - FIELD VS . FAR - FIELD FOR FIXED TARGETS

In far-field L&S, the target’s distance is much larger than the array aperture. Therefore, the distance
in (3) can be approximated as the following planar wave-based distance (see Figure 1(b)):

rn ≈ r − sTn u(θ). (8)

√
This approximation comes from the Taylor series 1 + x = 1+x/2+· · · for small x = (−2rsTn u(θ)+
∥sn ∥2 )/r2 , ignoring higher order terms. Similar to (4), we obtain the baseband far-field model as
β far
ynfar (t) = x(t − τnfar + τo )e−j2πfc (τn −τo ) + wn (t), (9)
r
where τnfar = (r − sTn u(θ))/c. Here, the channel gain β/r is approximately the same at all antennas
due to the large r. Inspecting (4) and (9), we notice that both the far-field and near-field signals
depend on the target’s location (θ and r). This raises the question: What unique impacts does the
near-field have on L&S? To explore this, we need to identify which parameters are estimatable. A
parameter is considered estimatable if it meets the consistency condition [10], meaning that there
exists an estimator that can approximate the true value to any accuracy as the number of samples
goes to infinity. In wireless systems, samples can be collected across three dimensions: time, space,
and frequency. Time-domain samples, acquired through sampling signals at different times t, provide
an overall view of the noiseless received signal but fail to resolve θ and r, as these parameters are
time-invariant. Therefore, the space- and frequency-domain samples are required. Applying the Fourier
transform to the signals in (4) and (9) yields:
 6

√
β (fc +f ) r 2 −2rsT
n u(θ)+∥sn ∥
2

Yn (f ) = X(f )e−j2π c + Wn (f ), (10)

rn
β (fc +f )r (fc +f )sTn u(θ)
Ynfar (f ) = X(f )e−j2π c ej2π c + Wn (f ), (11)
r
where the sampling offset τo is excluded for brevity. In the far-field model Ynfar (f ), the observation of
r depends solely on the frequency f . Accurately estimating r requires a wide bandwidth in X(f ). In
a narrowband system, distance estimation is infeasible. However, the direction θ can be determined
from space-domain samples taken at different antennas sn . Thus, the narrowband far-field signal is
typically modeled as [11]:
y far (t) = b(θ)s(t) + w(t), (12)
2π 2π
sT sT
where b(θ) = [ej λ 1 u(θ) , . . . , ej λ N u(θ) ]T is the far-field array response vector (also known as
steering vector). The situation is different for the near-field signal Yn (f ), for which the distance r is
related to both frequency f and the antenna coordinate sn . Therefore, near-field provides a new way
of estimating the target distance r in the space domain, reducing the need for broad bandwidth.

B. Moving targets

We now extend the model to a moving target within a millisecond-scale coherent processing interval
(CPI), where the target approximately follows uniform motion along a straight line. Let v = vu(ϕ)
denote the velocity vector of the target, where v and ϕ represent the speed and direction, respectively.
Within a CPI, the time-variant location of the moving target is po (t) = p + vt. The distance from the
target to the n-th antenna is thus given by
q
ro,n (t) = ∥sn − po (t)∥ = rn2 − 2vt(sn − p)T u(ϕ) + v 2 t2 , (13)
√
where rn is given in (3). This expression is usually simplified using the Taylor series 1+x = 1+
x/2 + · · · for small x. Within a CPI, where vt is almost always much less than the initial distance rn ,
the distance ro,n (t) can be approximated as [12]
(sn − p)T u(ϕ)vt
ro,n (t) ≈ rn − . (14)
rn
In this formula, the term (sn − p)T u(ϕ)/rn represents the projection from the velocity direction onto
the line-of-sight direction of the n-th antenna. To highlight the difference from the far-field model, we
redefine the velocity vector in terms of its components along and orthogonal to the line-of-sight path
with respect to the origin: v = vr u(θ) + vθ u⊥ (θ), where u⊥ (θ) = [− sin θ, cos θ]T is a unit vector
orthogonal to u(θ), and vr = vuT (θ)u(ϕ) and vθ = vuT⊥ (θ)u(ϕ) denote the radial and transverse
velocity, respectively. Substituting the new velocity vector into (14) yields
 7

(r − sTn u(θ))vr −sTn u⊥ (θ)vθ

ro,n (t) = rn + vn t, vn = + . (15)
rn rn
| {z } | {z }
Projection of vr Projection of vθ

Following (3) through (5), we obtain the narrowband near-field model for a moving target as [13]

yo (t) = a(θ, r) ⊙ d(vθ , vr , t)s(t) + w(t), (16)

2π 2π
where d(vθ , vr , t) = [e−j λ
v1 t
, . . . , e−j λ
vN t T
] is the near-field Doppler vector at time t and ⊙ signifies
the Hadamard product. For K moving targets, the signal model becomes

yo (t) = (A(θ, r) ⊙ D(vθ , vr , t)) s(t) + w(t), (17)

where D(vθ , vr , t) = [d(vθ,1 , vr,1 , t), . . . , d(vθ,K , vr,K , t)] denotes the Doppler matrix.

N EAR - FIELD VS . FAR - FIELD FOR MOVING TARGETS

In far-field L&S, where r ≫ D, we have the following approximations are tight:
r − sTn u(θ) −sTn u⊥ (θ)
≈ 1, ≈ 0. (18)
rn rn
Combining this property with (8), the far-field time-variant distance can be approximated as

ro,n (t) ≈ r − sTn u(θ) + vr t. (19)

Thus, the narrowband far-field model for a moving target is

2π
yofar (t) = b(θ)e−j λ
vr t
s(t) + w(t). (20)

Comparing (16) and (20), we see that near-field effects cause a non-uniform Doppler frequency across
the antenna array, providing additional information about the transverse velocity vθ . This allows us to
determine the full motion state of the target in near-field L&S. In contrast, far-field L&S only provides
information about the radial velocity vr , making it challenging to ascertain the full motion state.

III. N EAR - FIELD L&S FOR FIXED TARGETS

The primary task for near-field L&S with fixed targets is to estimate the location vectors θ and r from
the noisy signal in (7), which is different from far-field L&S where we can only estimate θ . However,
the core nature of the estimation problem remains similar, allowing us to use conventional estimation
techniques. Specifically, we aim to find the location vectors θ and r that minimize or maximize an
objective function based on a data matrix X , derived from the observation matrix Y = [y(1), . . . , y(L)] =
A(θ, r)S + W where S = [s(1), . . . , s(L)] and W = [w(1), . . . , w(L)]. In the following, we describe
the basic estimation methods for near-field L&S, which estimate θ and r either jointly or sequentially.
 8

A. Subspace fitting methods

Many classical estimation methods, such as deterministic maximum likelihood (DML) and MUSIC,
are specific instances of the subspace fitting method. These methods, widely used in far-field L&S, can
be extended to near-field cases. The general form of the subspace fitting problem for near-field L&S is

{θ̂, r̂} = argmin min ∥X − A(θ, r)Φ∥2F , (21)

θ,r Φ

where Φ is any unknown nuisance parameter and ∥ · ∥F is the Frobenius norm. This problem aims to fit
the subspace spanned by A(θ, r) to the data matrix X in a least-squares manner. For fixed A(θ, r), the
optimal nuisance parameter is Φ̂ = A† (θ, r)X . Substituting this pseudoinverse solution back into (21),
we get the equivalent problem

{θ̂, r̂} = argmax tr PA (θ, r)XX H ,

(22)
θ,r

where PA (θ, r) = A(AH A)−1 AH is the projection matrix onto the column space of A(θ, r) and tr(·)
represents the trace of a matrix. The complex structure of A(θ, r) makes this problem highly non-convex,
often resulting in many local maxima near the global maximum. Therefore, ensuring estimation accuracy
typically requires a 2K -dimensional high-resolution grid search of θ and r , which can be computationally
prohibitive in many application scenarios.
The single-target approximation is a popular method to reduce the search space dimension [11]. It
estimates based on the single-target model while keeping the multi-target data matrix X unchanged,
resulting in a simplified objective function: ∥X − a(θ, r)ϕT ∥2F . Here, the multi-target array response
matrix A(θ, r) is replaced by the single-target array response vector a(θ, r). The locations of the K
targets are then found as the K deepest local minima of this function. Substituting the optimal ϕ with
fixed a(θ, r), the following equivalent problem can be obtained:
K
argmax tr Pa (θ, r)XX H ,


{θ̂, r̂} = (23)
θ,r

where K argmax f (·) identifies the K arguments at which the function f (·) reaches its K highest local
maxima. While the single-target approximation reduces the search space dimension from 2K to 2, it
may struggle with coherent signals from different targets. Such cases might require spatial smoothing to
decorrelate the signals [14]. The DML and MUSIC are the two most popular subspace fitting techniques
that follow the general framework described above, as elaborated below.
1) DML: In the DML method, the data matrix is directly chosen as the observation matrix, i.e., X = Y .
The objective function in (21) thus becomes ∥Y −A(θ, r)S∥2F . Minimizing this objective function essen-
tially maximizes the likelihood function of the observation signal Y , assuming the source signal is deter-
ministic and the noise is zero-mean complex circular Gaussian such that y(t) ∼ CN (A(θ, r)s(t), σ 2 IN ).
 9

The equivalent DML problem and its single-target approximation are given by

{θ̂, r̂} = argmax tr PA (θ, r)Y Y H ,

(24)
θ,r
K
argmax tr Pa (θ, r)Y Y H .


{θ̂, r̂} = (25)
θ,r

2) MUSIC: In the MUSIC method, the data matrix X is derived from the covariance matrix of
the observation signal y(t), theoretically given by R = E[y(t)y H (t)] = ARs AH + σ 2 IN , where
Rs = E[s(t)sH (t)] and E[·] denotes the statistical expectation. The rank of the signal subspace matrix
ARs AH is at most K when there are K targets. In practice, we can estimate the covariance matrix
as the sample average R̂ = Y Y H /L. The signal and noise subspaces are then estimated from the
eigendecomposition R̂ = Us Σs UsH + Un Σn UnH , where Σs and Σn contain the K largest and (N − K)
smallest eigenvalues of R̂, and Us and Un contain the eigenvectors corresponding to the signal and noise
subspaces, respectively. If K is unknown, it can be estimated by identifying how many eigenvalues are
substantially larger than the noise floor.
Following the subspace fitting concept, the data matrix can be selected as X = Us and the objective
function can be designed as ∥Us − A(θ, r)Γ∥2F , where Γ is a nuisance matrix related to Rs . The
corresponding equivalent problem and its single-target approximation are thus given by

{θ̂, r̂} = argmax tr PA (θ, r)Us UsH ,

(26)
θ,r
K
argmax tr Pa (θ, r)Us UsH .


{θ̂, r̂} = (27)
θ,r

The methods described in (26) and (27) are variants of the classical MUSIC method based on signal
subspace [15]. It is easy to transform (27) into the classical MUSIC based on the noise subspace [16].
Specifically, let p(θ, r) = tr Pa (θ, r)Us UsH , which satisfies 0 < p(θ, r) < 1. Therefore, maximizing

p(θ, r) is equivalent to maximizing 1/(1 − p(θ, r)), resulting in the following classical MUSIC method:

K ∥a(θ, r)∥2
{θ̂, r̂} = argmax . (28)
θ,r aH (θ, r) (IN − Us UsH ) a(θ, r)
PN
The value of the numerator ∥a(θ, r)∥2 = n=1 r/rn is approximately constant if the target is much
further from the antenna array than the aperture length, so that r/rn ≈ 1. Using the orthogonality
between the signal and noise subspaces, i.e., IN − Us UsH = Un UnH , we can simplify (28) as

K 1
{θ̂, r̂} = argmax . (29)
θ,r aH (θ, r)Un UnH a(θ, r)
While the DML method generally has superior asymptotic properties compared to MUSIC [15],
the latter performs better under single-target approximation, as illustrated in Figure 2. This improved
performance of the MUSIC method is due to its effective utilization of the orthogonality between the
 10

(a) DML (b) MUSIC

Figure 2: An example of the two-dimensional DML and MUSIC spectra obtained using the single-target
approximation for five targets, with a uniform linear array (ULA) composed of N = 128 antennas and
SNR = 0 dB. The ULA antenna spacing is set to λ/2. The MUSIC spectrum exhibits clearer peaks
around the true target locations compared to the DML spectrum, particularly for closely located targets.

signal and noise subspaces, which helps mitigate the interference between different targets. Due to the
favorable properties, many MUSIC variants have been proposed in the literature for far-field L&S to
either reduce complexity or enhance estimation performance. Some variants rely on far-field properties,
like the linear phase of the far-field array response used in the root-MUSIC method [17], making them
unsuitable for near-field cases. However, other variants, such as weighted MUSIC [15], which do not
depend on specific array response properties, remain effective for enhancing near-field L&S performance.

B. Array-structure-based methods

The structural information of the antenna array is important prior information that can be exploited to
simplify the estimation problem and reduce computational complexity. For instance, the ESPRIT method,
a well-known far-field L&S technique designed for shift-invariant arrays, can eliminate the need for a
grid search. In near-field L&S, several modifications to the classical MUSIC method have been proposed
to decouple the estimation of θ and r by leveraging the symmetric structure of the array.
Assume the L&S node has N = 2M + 1 antennas arranged in a symmetric ULA, with the center
antenna at the origin. The coordinate of the n-th antenna is sn = [δn d, 0]T , where d is the inter-antenna
spacing and δn = n−M −1. The application of the symmetric ULA structure in the MUSIC method relies
on two approximations: the amplitude variation of the near-field array response is negligible (r/rn ≈ 1),
and the distance rn is approximated using the so-called Fresnel approximation as [7]
s 
2sTn u(θ) ∥sn ∥2 δn2 d2 sin2 θ

2
rn = r 1 − + ≈ r − δ d cos θ + , (30)
n
r r2 2r
√
which is obtained using the second-order Taylor approximation 1 + x = 1 + x/2 − x2 /8 + · · · and
ignoring some higher-order terms. Setting γ = −2πd cos θ/λ and ψ = πd2 sin2 θ/(λr), the near-field
 11

array response vector can be reformulated from (6) and (30) as

h 2 2
iT
a(θ, r) = ej(−M γ+M ψ) , . . . , 1, . . . , ej(M γ+M ψ) . (31)

Writing the array response vector in this form allows for decomposing the location estimation problem
into separate direction and distance estimation problems. Specifically, with this symmetric structure of
the near-field array response vector, the terms related to ψ that involve the distance parameters cancel out
in the anti-diagonal elements of the covariance matrix of the observation signal, i.e., R = E{y(t)y H (t)}.
Collecting all the anti-diagonal entries of R in a vector and defining ϵk = E[sk (t)s∗k (t)], we have
K
X  −jM γk T
ȳ = ϵk e , . . . , ϵk ejM γk , (32)
k=1
where we temporarily ignore the noise term for brevity. The elements in ȳ are independent of the target
distances and only contain their direction information. This enables us to decouple the direction and
distance estimation problems. To effectively estimate the K directions via the MUSIC method, we need
a covariance matrix of rank greater than K . This is to construct the signal and noise subspaces of
sufficient dimensions. To this end, the vector ȳ in (32) is divided into J overlapping subvectors, each
having 2M + 2 − J entries. The i-th subvector is
K h
X iT
ỹ(i) = ϵk e−j2(M −i+1)γk , . . . , ϵk e−j2(J−i−M )γk = Ã(γ)s̃(i), (33)
k=1
T
where s̃(i) = ϵ1 ej2iγ1 , . . . , ϵK ej2iγK and Ã(γ) = [ã(γ1 ), . . . , ã(γK )] with


h iT
ã(γk ) = e−j2(M +1)γk , . . . , e−j2(J−M )γk . (34)

The model in (33) exhibits a similar form to (7). Therefore, the MUSIC method can be applied for
estimating the K directions (encapsulated in γk ) by exploiting the signal and noise subspaces of the
sample covariance matrix R̃ = Ji=1 ỹ(i)ỹ H (i)/J . Then, for each estimated direction, the corresponding
P

distance can be found by employing the MUSIC method described in (29) but only need to search the
distance dimension. This way, the two-dimensional grid search over θ and r is replaced by K + 1
one-dimensional grid searches, thereby remarkably reducing the computational complexity.
Note that there are three necessary conditions for the use of the modified MUSIC method: 1) a
symmetric array structure to cancel out distance parameters in the anti-diagonal elements of the covariance
p
matrix, 2) the Fresnel approximation in (30) is tight, typically requiring r ≥ 0.5 D3 /λ [7], and 3) a
small inter-antenna spacing d ≤ λ/4 to avoid direction ambiguity, c.f. (34). An example of the direction
and distance spectra obtained using the modified MUSIC when the above conditions are met is depicted
in Figure 3. While the modified MUSIC algorithm reduces the computational complexity involved in
estimating directions and distances, it may still require a significant number of searches, particularly
 12

(a) Direction spectrum (b) Distance spectrum

Figure 3: An example of the direction and distance spectra obtained using the modified MUSIC for five
targets, with N = 257, d = λ/4, J = 50, and SNR = 0 dB. (a) The one-dimensional direction spectrum
obtained by classical MUSIC method based on the model in (33). (b) The five one-dimensional distance
spectra obtained using the MUSIC method described in (29) at each estimated direction.

when dealing with many targets. There are other array-structure-based methods that further simplify the
estimation process by establishing a relationship between the estimation parameters, such as [18] and
[19]. These methods decompose the array response expression in (31) into a matrix-vector product with
the matrix and vector only related to γ and ψ , respectively, enabling closed-form estimators.
In this section, we have introduced the parametric subspace-based methods for near-field L&S with
fixed targets, like DML and MUSIC. These methods build on some common presumptions, such as
the knowledge of the number of targets, low coherence between source signals, Gaussianity of signals
and noise, and a large number of data samples, which limit their applicability to specific scenarios.
The sparsity-based method, an important advancement inspired by compressive sensing in array signal
processing, relaxes these assumptions by using a sparse signal model representation and is applicable to
both near-field and far-field L&S. However, it can be sensitive to hyperparameters and may bias estimation
solutions. For more details on the sparsity-based method, see [20].

IV. N EAR - FIELD L&S FOR MOVING TARGETS

In a near-field L&S system involving moving targets, it is crucial to account for both the targets’
locations θ and r and their velocities vθ and vr encapsulated in the signal model in (17). This is
a dynamical signal model due to the time-variant Doppler matrix D(vθ , vr , t), which significantly
complicates our estimation problem and renders many conventional methods inapplicable. This section
will describe several potential estimation methods for near-field L&S with moving targets.

A. Subspace fitting method

The subspace fitting method remains useful for L&S systems with moving targets and is widely used
for estimating the direction θ and radial velocity vr in the far-field model (20), as shown in [21]. In the far-
 13

field model, both DML and MUSIC can be effectively applied because all antennas experience the same
2π
Doppler frequency caused by vr . For example, to estimate θ from (20), we can define s̃(t) = e−j λ
vr t
s(t)
as an equivalent source signal, ensuring the signal subspace is always spanned by b(θ) at different times.
Thus, both DML and MUSIC can be used to estimate θ. Moreover, vr can be estimated by DML, MUSIC,
or Fourier analysis.
For the near-field model (17), the challenge is that different antennas observe different velocity com-
ponents of targets, resulting in non-uniform Doppler frequencies and, thus, a time-varying signal sub-
space. This makes the covariance matrix and corresponding subspaces nearly immeasurable, render-
ing covariance matrix-based methods like MUSIC unusable. However, we can still apply the DML
method. Specifically, the observation matrix over L, with a sampling interval of Ts , can be derived
from (17) as Yo = Ao (θ, r, vθ , vr )S̃ + W , where Ao (θ, r, vθ , vr ) = (1TL ⊗ A(θ, r)) ⊙ D̃(vθ , vr ) with
D̃(vθ , vr ) = [D(vθ , vr , Ts ), . . . , D(vθ , vr , LTs )] and S̃ = blkdiag{s(1), . . . , s(L)}. Here, ⊗ signifies
the Kronecker product, blkdiag{·} indicate block diagonal matrices, and 1L is an all-one vector of length
L. Then, the DML method can be described as

{θ̂, r̂, v̂θ , v̂r } = argmin min ∥Yo − Ao (θ, r, vθ , vr )S̃∥2F . (35)
θ,r,vθ ,vr S̃

The computational cost of the above multi-dimensional DML is impractical for many applications due
to the necessity of a 4K -dimensional grid search. Additionally, calculating the projection matrix for the
large-dimensional matrix Ao (θ, r, vθ , vr ) presents further challenges. While reducing the DML search
space dimension using single-target approximation is possible, it often leads to significant performance
loss due to interference between targets. Partial relaxation offers a promising alternative by considering
signals from all targets while maintaining a low-dimensional search space [22]. However, extending this
method to models with time-variant signal subspaces remains an open problem.

B. Nonlinear filtering method: from estimation to tracking

As mentioned in the previous section, estimating target locations and velocities within a single CPI
is challenging due to the lack of statistical information and the non-uniform Doppler frequencies caused
by transverse velocity effects on the array response. However, if we consider a series of CPIs instead of
focusing on a single CPI, this feature can actually facilitate target estimation. Let us take a single-target
case as an example, where ri , θi , vri , and vθi represent the target state in the i-th CPI and Tc denotes
the CPI length. The target state between successive CPIs is highly correlated and we assume it can be
described by the following state evolution model:
v i Tc
ri+1 = ri + vri Tc + zri , θi+1 = θi + θ i + zθi , vri+1 = vri + zvi r , vθi+1 = vθi + zvi θ , (36)
r
 14

where zri , zθi , zvi r , and zvi θ are error parameters accounting for the perturbations caused by accelerations.
The imposition of the above correlation constraint allows us to estimate the target state using both current
CPI observations and previous CPI estimates, turning the estimation into a tracking problem over multiple
CPIs. Note that the state evolution model (36) applies to both near-field and far-field scenarios. However,
in the far-field, the absence of transverse velocity vθ information in the observation signal makes it
challenging to track the target accurately based on real-world measurements.
Nonlinear filtering is a powerful tool to estimate and track the state of a nonlinear system from noisy
observations. The extended Kalman filter (EKF) is the standard technique typically applied to a nonlinear
system under additive Gaussian noise. We refer to [23, Chapter 13.7] for a detailed EKF derivation. The
EKF requires a state evolution model and an observation model. The state evolution model is given in
(36), which can be written into a nonlinear function as

qi+1 = h(qi ) + zi , zi ∼ CN (0, Rz ), (37)

where qi = [ri , θi , vri , vθi ]T and zi = [zri , zθi , zvi r , zvi θ ]T are the state and noise vectors in the i-th CPI,
respectively, h(·) is the nonlinear state evolution function, and CN (0, C) denotes the circularly symmetric
complex Gaussian distribution with covariance C . The observation model is given in (16). In each CPI,
numerous observation samples can be obtained by sampling the observation signal at different times.
Exploiting more samples typically results in better tracking performance, especially for multi-target cases
[24]. Despite this, it is still possible to use only one sample in each CPI. Without loss of generality,
we consider the observation sample obtained at time t = Tc /2. Under this simplified condition, the
observation model in the i-th CPI is given by

yo,i = g(qi )si + wi , wi ∼ CN (0, Rw ), (38)

where g(q) = a(θ, r) ⊙ d(vθ , vr , Tc /2) is the nonlinear array response function and si is the equivalent
source signal at t = Tc /2 in the i-th CPI. We are now prepared to formulate the EKF method, which
can be summarized as the following iterative process:

1) A priori state estimate: q̂i|i−1 = h(q̂i−1 ).

2) Linearization: Hi = ∇q h(q)|q=q̂i−1 and Gi = ∇q (g(q)si )|q=q̂i|i−1 .
3) A priori covariance estimate: Qi|i−1 = Hi Qi−1 H H + Rz .

−1
4) Kalman gain: Ki = Qi|i−1 GH i Gi Qi|i−1 GH
i + Rw .


5) A posteriori estimates: q̂i = q̂i|i−1 + Ki yo,i − g(q̂i|i−1 )si and Qi = (I − Ki Gi ) Qi|i−1 .

In Steps 2 and 5 of the EKF, knowledge of the equivalent source signal si is required, which may
be unavailable in practice. However, it can be estimated based on the a priori state estimate q̂i|i−1 and
the observation yo,i [24], which is essentially a signal detection problem under channel estimation error.
 15

(a) Location tracking (b) Velocity tracking

Figure 4: An example of applying the EKF to tracking a near-field moving target, with a 128-antenna
ULA, 28 GHz carrier frequency, Tc = 1 ms, and SNR = 0 dB. The ULA antenna spacing is set to λ/2
and the mean-squared estimation error of si is set to −3 dB. The covariance matrices of the state and
observation noises are set to Rz = diag(0, 0, 0.01, 0.01) and Rw = σ 2 IN , respectively.

An example of the EKF performance under imperfect knowledge of si is shown in Figure 4. Besides
the robust tracking performance, another key advantage of the EKF method is its adaptability to multi-
target scenarios. By following (36) and (17), we can obtain the multi-target state evolution model and
the observation model, respectively. Then, the multi-target EKF can be carried out following the standard
process. Apart from the EKF, other effective nonlinear filters based on Bayesian estimation are available
in the literature [25], including other variants of the Kalman filter, such as the unscented Kalman filter
and the Monte Carlo Kalman filter, as well as the non-Gaussian filters, like particle filters, which can be
helpful to improve the near-field L&S performance in different scenarios.

V. ACCURACY, RESOLUTION , AND EFFECTIVE NEAR - FIELD REGION

Accuracy and resolution are two critical performance metrics for L&S systems. Accuracy refers to
the degree to which the estimated value of a parameter matches the true value. The Cramér-Rao Bound
(CRB) is an important indicator of accuracy, defined as the inverse of the Fisher information. It provides a
theoretical lower bound on the mean-squared error of estimators for deterministic parameters. Resolution,
on the other hand, refers to the system’s ability to distinguish between two or more closely spaced targets.
This is typically measured by the ambiguity function (AF), which is the matched filter response. Although
these two metrics serve different purposes, they are highly correlated: optimal accuracy is typically a
scaling of the resolution by a factor related to the SNR. The accuracy and resolution for far-field L&S
have been extensively studied in the literature and can typically be represented by separate, simple, and
closed-form CRBs and AFs [26]. This simplicity stems from the fact that the key target parameters, such
as distance, direction, and velocity, are not correlated in the far-field model.
 16

The situation is much more complicated in near-field systems because the target parameters become
highly coupled and spatially variant, c.f. (5) and (16). However, it is still possible to obtain relatively
simple CRBs using the standard formula to evaluate the accuracy of near-field L&S. In [27], the closed-
form CRBs for estimating θ and r of a fixed target using a ULA with antenna spacing d are derived by
applying the Fresnel approximation in (30) and ignoring amplitude variation. These CRBs are given by
3N1
CRBθ = , (39)
2γLN2 π 2 sin2 θ
6r2 (15r2 + 30rd(N − 1) cos θ + d2 N1 cos2 θ)
CRBr = , (40)
γLN2 π 2 d2 sin4 θ
where γ = SNR · d2 /λ2 , N1 = (8N − 11)(2N − 1), and N2 = N (N 2 − 1)(N 2 − 4). Although the above
CRBs are derived for a specific case, they provide qualitative insights that hold in general situations:

• The distance between the target and the L&S node generally has a negligible impact on the accuracy
of direction estimation (i.e., CRBθ is not related to r).
• The accuracy of distance estimation decreases unboundedly as the distance between the target and
the L&S node increases (i.e., CRBr increases unboundedly with r).

These insights align with intuition because as the distance approaches infinity, the near-field model
converges to the far-field model, which only includes direction information. A recent study suggests a
similar trend for the near-field velocity estimation with a moving target [28]: a larger distance between
the target and the L&S node has negligible impact on the accuracy of radial velocity estimation but
reduces the accuracy of transverse velocity estimation unboundedly.
Regarding resolution, the AF can be characterized based on the matched filter response. Let’s first
consider the case of a fixed target located at (θ0 , r0 ) so that the array response is time-invariant. Hence,
we can represent the matched filter response of a single time-domain sample as the AF function, which
is given by
AF(θ, r) = aH (θ, r)a(θ0 , r0 ). (41)

The resolution can then be measured by the half-power mainlobe wideth (HPMW), i.e., the distance
between the half-power intercepts of the AF mainlobe. It can be challenging to distinguish two targets if
they fall within the HPMW of each other. A comprehensive analysis of the function AF(θ, r) for ULAs is
provided in [29] from the codebook sampling perspective. It suggests that the direction resolution can be
evaluated on a curve where sin2 θ/r = sin2 θ0 /r0 , leading to the AF function for the direction resolution

r0 sin2 θ (c) sin (πN d(ϑ − ϑ0 )/λ)

 
Λ1 (ϑ) ≜ AF θ, ≈ , (42)
sin2 θ0 sin (πd(ϑ − ϑ0 )/λ)

where ϑ = cos θ and ϑ0 = cos θ0 . The step (c) uses the Fresnel approximation and omits the amplitude
variation. |Λ1 (ϑ)| reaches its maximum value of N at ϑ = ϑ0 . Then, the HPMW in the direction
 17

Figure 5: Illustration of the distance-domain AF function Λ2 (r) when θ0 = π/2.

dimension is obtained as
N2
 
2 1.4λ 1.4λ
|Λ1 (ϑ0 + ∆ϑ)| ≥ → ∆ϑ ∈ − , . (43)
2 πN d πN d
This is the same as the far-field resolution on the direction and is measured in radians, thus, the physical
width is smaller at short distances. Moreover, the distance resolution can be evaluated by the following
AF function:
(d) N C(η) + jN S(η)
Λ2 (r) ≜ AF(θ0 , r) ≈ , (44)
η
Rη Rη
where C(η) = 0 cos(πx2 /2)dx and S(η) = 0 sin(πx2 /2)dx are Fresnel functions and η is defined as
s
N 2 d2 sin2 θ0 1 1
η= − . (45)
2λ r r0
The proof of (d) can be found in [29, Appendix A]. Similarly, |Λ2 (r)| reaches its maximum value of N
at r = r0 . Then, the HPMW in the distance dimension is obtained as [7, Appendix B]
N2 r02 r02
 
2
|Λ2 (r0 + ∆r)| ≥ → ∆r ∈ − , , (46)
2 dT + r0 max{dT − r0 , 0}
where dT is given by
N 2 d2 sin2 θ0
dT ≈ . (47)
5λ
An important observation from (46) is that the HPMW in the distance dimension is not always finite, as
illustrated in Figure 5. Specifically, if r0 ≥ dT , the upper limit in (46) becomes ∞. This implies that any
two targets in the same direction are almost indistinguishable once their distances from the L&S node
are greater than dT . The resolution of moving targets can be more complicated since the AF function
must consider all time-domain samples in a CPI, which remains an open question.
The above observations prompt us to reconsider the near-field region for L&S. The classical Fraunhofer
distance is defined based on the phase error from the far-field approximation relative to the accurate
near-field model, which may not effectively reflect L&S characteristics. Instead, we should reevaluate
the effective near-field region by considering the accuracy and resolution of L&S. Since the study of
 18

accuracy and resolution for moving targets is still in its early stages, we will focus on the fixed target
case. For fixed targets, the key benefit of near-field L&S is enabling distance estimation in narrowband
systems. An important conclusion is that the effective near-field region is infinite for accuracy but finite
for resolution. Specifically, for distance accuracy evaluated by CRBr in (40), although CRBr increases
indefinitely with r, this increase can always be offset by higher SNR or more data samples L. As
SNR → ∞, CRBr → 0, indicating infinite accuracy regardless of the distance from the L&S node. This
provides a useful practical guideline: if there are multiple targets with non-overlapping AF mainlobes,
we can always effectively estimate their distances as long as the SNR and the number of data samples
are sufficiently large. However, for distance resolution, the effective near-field region is limited by the
distance dT in (47), beyond which targets become indistinguishable in the distance domain. Importantly,
dT is much smaller than the Fraunhofer distance dFA . More specifically, in the case of θ0 = π/2 where
dT attains its maximum, we have
N 2 d2 1
dT ≈ ≈ dFA , (48)
5λ 10
based on the expression of ULA’s aperture D = (N − 1)d. This result suggests that, from a distance
resolution perspective, the effective near-field region is at most one-tenth of the Fraunhofer distance.

VI. C ONCLUSIONS AND FUTURE DIRECTIONS

In this article, we have provided a general review of the near-field L&S techniques, from the funda-
mental signal models to basic signal processing methods for both fixed and moving targets. We have
also discussed the accuracy and resolution from a theoretical perspective. We could not delve into many
advanced methods for near-field L&S due to space limitations, but we hope this article will inspire further
literature reading, such as [11], [20], [25]. Despite the long history of near-field L&S research, many
challenges remain to be addressed. We highlight several promising directions for future research.

• Wideband systems: Although the near-field effect allows for estimating target distances in the spatial
domain using narrowband signals, the achievable accuracy and resolution diminish as the target
distance increases. As previously mentioned, using wideband signals, such as orthogonal frequency-
division multiplexing (OFDM) and linear frequency modulation (LFM) signals, is an effective
solution to counteract this diminishing accuracy. This approach introduces challenges such as large-
dimensional space-frequency signals, spatial-wideband effects, and imperfect synchronization.
• Hybrid and low-resolution arrays: Considering the large number of antennas and the high frequency
in near-field systems, implementing digital antenna arrays presents significant challenges. Hybrid
analog-digital arrays and low-resolution arrays are two popular alternatives to reduce hardware
complexity and energy consumption [4]. However, the use of these arrays can result in significantly
 19

distorted measurements or, in some cases, measurements with only one-bit resolution. Addressing
such distortion for near-field L&S remains an open problem.
• Sparse array design: Sparse arrays, such as coprime and nested arrays, have been widely studied
for far-field L&S to improve direction estimation and increase the number of resolvable targets
using virtual coarrays. In near-field L&S, sparse arrays offer the additional benefit of generating a
large aperture with fewer antennas, which improves the estimation performance of target distance
and transverse velocity due to the more pronounced near-field effect. Tailored antenna placement
optimization may be needed to fully reap the advantage of sparse arrays in near-field L&S systems.
• Tensor-based signal processing: Tensors are high-dimensional extensions of matrices. In near-field
L&S, the target states (distance, direction, radial velocity, transverse velocity) and data measurements
(space, time, frequency) are multidimensional. Unlike far-field L&S, where different target states
can be estimated separately across different data dimensions, near-field L&S involves highly coupled
target states that require joint processing of the high-dimensional data. In this context, higher-order
tensors can be a powerful tool [30].
• Data-driven and model-based machine learning: The near-field L&S methods discussed in this
article are model-based and thus limited to specific scenarios. In particular, we have considered the
free-space line-of-sight model, which is seldom exact in terrestrial use cases. In contrast, model-free
data-driven approaches can generalize to various training data statistics but are typically data-hungry.
In recent years, model-based machine learning has become a hot topic, which combines data and
model knowledge to reduce data requirements and enhance robustness to model uncertainty and
scattering and is thus promising to address the complicated near-field L&S problems.

R EFERENCES

[1] G. D. Ott, “Vehicle location in cellular mobile radio systems,” IEEE Trans. Veh. Technol., vol. 26, no. 1, pp. 43–46, Feb.
1977.
[2] A. J. Fenn et al., “The development of phased-array radar technology,” Lincoln Lab. J., vol. 12, no. 2, pp. 321–340, 2000.
[3] H. Krim and M. Viberg, “Two decades of array signal processing research: the parametric approach,” IEEE Signal Process.
Mag., vol. 13, no. 4, pp. 67–94, Jul. 1996.
[4] E. Björnson et al., “Twenty-five years of signal processing advances for multiantenna communications: From theory to
mainstream technology,” IEEE Signal Process. Mag., vol. 40, no. 4, pp. 107–117, Jun. 2023.
[5] E. Fishler et al., “MIMO radar: An idea whose time has come,” in Proc. IEEE Radar Conf., Apr. 2004, pp. 71–78.
[6] P. Ramezani and E. Björnson, “Near-field beamforming and multiplexing using extremely large aperture arrays,” in
Fundamentals of 6G Communications and Networking. Springer, 2023, pp. 317–349.
[7] Y. Liu et al., “Near-field communications: A tutorial review,” IEEE Open J. Commun. Soc., vol. 4, pp. 1999–2049, 2023.
[8] Y. Rockah and P. Schultheiss, “Array shape calibration using sources in unknown locations–part II: Near-field sources and
estimator implementation,” IEEE Trans. Acoust., Speech, Signal. Process., vol. 35, no. 6, pp. 724–735, Jun. 1987.
 20

[9] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Trans. Wireless
Commun., vol. 9, no. 11, pp. 3590–3600, Nov. 2010.
[10] E. L. Lehmann and G. Casella, Theory of point estimation. Cham, Switzerland: Springer, 2006.
[11] M. Pesavento, M. Trinh-Hoang, and M. Viberg, “Three more decades in array signal processing research: An optimization
and structure exploitation perspective,” IEEE Signal Process. Mag., vol. 40, no. 4, pp. 92–106, Jun. 2023.
[12] M. A. Richards et al., Fundamentals of radar signal processing. New York, NY, USA: McGraw-Hill Educ., 2005, vol. 1.
[13] Z. Wang et al., “Near-field velocity sensing and predictive beamforming,” arXiv preprint arXiv:2311.09888, 2023.
[14] T.-J. Shan, M. Wax, and T. Kailath, “On spatial smoothing for direction-of-arrival estimation of coherent signals,” IEEE
Trans. Acoust. Speech Signal Process., vol. 33, no. 4, pp. 806–811, Aug. 1985.
[15] M. Viberg and B. Ottersten, “Sensor array processing based on subspace fitting,” IEEE Trans. Signal Process., vol. 39,
no. 5, pp. 1110–1121, May 1991.
[16] R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Trans. Antennas Propag., vol. 34, no. 3,
pp. 276–280, 1986.
[17] B. D. Rao and K. S. Hari, “Performance analysis of root-MUSIC,” IEEE Trans. Acoust. Speech Signal Process., vol. 37,
no. 12, pp. 1939–1949, Dec. 1989.
[18] X. Zhang, W. Chen, W. Zheng, Z. Xia, and Y. Wang, “Localization of near-field sources: A reduced-dimension MUSIC
algorithm,” IEEE Commun. Lett., vol. 22, no. 7, pp. 1422–1425, Jul. 2018.
[19] C. Huang, J. Xu, W. Xu, X. You, C. Yuen, and Y. Chen, “Low-complexity channel estimation for extremely large-scale
MIMO in near field,” IEEE Wireless Commun. Lett., vol. 13, no. 3, pp. 671–675, Mar. 2024.
[20] D. Malioutov, M. Cetin, and A. S. Willsky, “A sparse signal reconstruction perspective for source localization with sensor
arrays,” IEEE Trans. Signal Process., vol. 53, no. 8, pp. 3010–3022, Aug. 2005.
[21] J. Lee, J. Park, and J. Chun, “Weighted two-dimensional root MUSIC for joint angle-Doppler estimation with MIMO
radar,” IEEE Trans. Aerosp. Electron. Syst., vol. 55, no. 3, pp. 1474–1482, 2019.
[22] M. Trinh-Hoang, M. Viberg, and M. Pesavento, “Partial relaxation approach: An eigenvalue-based DOA estimator
framework,” IEEE Trans. Signal Process., vol. 66, no. 23, pp. 6190–6203, Dec. 2018.
[23] S. M. Kay, Fundamentals of statistical signal processing: estimation theory. Englewood Cliffs, NJ: Prentice-Hall, 1993.
[24] Y. I. Kamil and A. Manikas, “Multisource spatiotemporal tracking using sparse large aperture arrays,” IEEE Trans. Aerosp.
Electron. Syst., vol. 53, no. 2, pp. 837–853, Apr. 2017.
[25] A. J. Haug, “A tutorial on Bayesian estimation and tracking techniques applicable to nonlinear and non-Gaussian processes,”
MITRE Corporation, McLean, VA, USA, 2005.
[26] A. Liu et al., “A survey on fundamental limits of integrated sensing and communication,” IEEE Commun. Surv. Tut.,
vol. 24, no. 2, pp. 994–1034, 2022.
[27] M. N. El Korso, R. Boyer, A. Renaux, and S. Marcos, “Conditional and unconditional Cramér–Rao bounds for near-field
source localization,” IEEE Trans. Signal Process., vol. 58, no. 5, pp. 2901–2907, May 2010.
[28] C. Giovannetti, N. Decarli, and D. Dardari, “Performance bounds for velocity estimation with large antenna arrays,” arXiv
preprint arXiv:2405.05234, 2024.
[29] M. Cui and L. Dai, “Channel estimation for extremely large-scale MIMO: Far-field or near-field?” IEEE Trans. Commun.,
vol. 70, no. 4, pp. 2663–2677, Apr. 2022.
[30] A. Cichocki et al., “Tensor decompositions for signal processing applications: From two-way to multiway component
analysis,” IEEE Signal Process. Mag., vol. 32, no. 2, pp. 145–163, Mar. 2015.