1
Near-Field Channel Modeling for
Electromagnetic Information Theory
Zhongzhichao Wan, Jieao Zhu, and Linglong Dai, Fellow, IEEE
Abstract—Electromagnetic information theory (EIT) is
one of the important topics for 6G communication due
to its potential to reveal the performance limit of wireless
communication systems. For EIT, the research foundation is
arXiv:2403.12268v1 [cs.IT] 18 Mar 2024
reasonable and accurate channel modeling. Existing channel
modeling works for EIT in non-line-of-sight (NLoS) scenario
focus on far-field modeling, which can not accurately capture
the characteristics of the channel in near-field. In this paper,
we propose the near-field channel model for EIT based on
electromagnetic scattering theory. We model the channel by
using non-stationary Gaussian random fields and derive the
analytical expression of the correlation function of the fields.
Furthermore, we analyze the characteristics of the proposed
channel model, e.g., the sparsity of the model in wavenumber
domain. Based on the sparsity of the model, we design a
channel estimation scheme for near-field scenario. Numerical
analysis verifies the correctness of the proposed scheme and
shows that it can outperform existing schemes like least
square (LS) and orthogonal matching pursuit (OMP).
Index Terms—Electromagnetic information theory (EIT),
near field, channel modeling, Gaussian random field, channel
estimation.
I. I NTRODUCTION
To improve the system performance, various promis-
ing technologies, including reconfigurable intelligent sur-
faces (RISs) [1], [2], continuous-aperture multiple-input
multiple-output (CAP-MIMO) [3], [4], and near-field
communications [5], [6], have been proposed for sixth-
generation (6G) communication. All these technologies are
trying to explore new sources of degrees of freedom (DoF)
or capacity gain. These performance gain actually come
from more accurate understanding and precise manipu-
lation of electromagnetic fields which convey information
[7]. Therefore, combining classical electromagnetic theory
and information theory to provide modeling and capacity
analysis tools is of great importance for exploring the
fundamental performance limit of wireless communica-
tion systems, which leads to the interdisciplinary subject
called electromagnetic information theory (EIT) [8]. By
integrating deterministic physical theory and stochastic
mathematically theory [9], EIT is expected to provide new
This work was supported in part by the National Key Research
and Development Program of China (Grant No. 2020YFB1807201), in
part by the National Natural Science Foundation of China (Grant No.
62031019).
All authors are with the Department of Electronic Engineering, Ts-
inghua University as well as Beijing National Research Center for In-
formation Science and Technology (BNRist), Beijing 100084, China (E-
mails: {wzzc20, zja21}@mails.tsinghua.edu.cn; daill@tsinghua.edu.cn).
insights into system models, degrees of freedom, capacity
limits, etc., from the electromagnetic perspective.
A. Prior works
The existing research directions of EIT includes chan-
nel modeling [10]–[12], DoF analysis [13]–[15], mutual
information and capacity analysis [16]–[18], etc. Among
these directions, channel modeling is the fundamental part.
Without precise channel model, DoF and capacity of EIT
can not be accurately analyzed.
For the channel modeling schemes of EIT, one approach
is line-of-sight (LoS) modeling scheme derived directly
from Maxwell’s equations, and the channel is expressed by
the Green’s function in free space [10], [11] or considering
reflection from a surface [12]. Another approach consid-
ers non-line-of-sight (NLoS) channel, which obeys the
electromagnetic scattering theory [19]. Compared to LoS
channel, NLoS channel is more general and can support
larger degrees of freedom in wireless communication.
Due to the complexity and uncertainty of the scattering
environment, the NLoS channel is often modeled by
random fields, whose statistical characteristics are derived
from the scattering environment [20]. For example, an
isotropic statistical channel model is derived in [21], where
correlation exists between different points of the received
field. This model can be viewed as an extension of the
traditional independent and identically distributed (i.i.d.)
Rayleigh fading channel model. Furthermore, a more gen-
eral scheme for constructing small-scale fading channel
is provided in [22], where the received electromagnetic
field is expanded by Fourier plane waves. This model
corresponds to far-field scattering scenario and provides
a useful way of wavenumber domain channel to spatial
correlation functions. An extended work [23] further de-
rives approximate analytical correlation function based on
[22], leading to a non-isotropic channel model.
The above works use far-field assumption to build chan-
nel models for EIT1 . However, for the use of middle-band,
millimeter-wave and terahertz technologies, or extremely
large aperture [24], such approximation is not accurate,
which may introduce non-negligible errors in DoF and
1 Far-field assumption here leads to the independence of the incoming
waves at different angles. Under this assumption, the received field is
spatially stationary, where the correlation function of the received field
only depends on the distance vector between two points. This is similar
to the stationary process in time-frequency domain.

capacity analysis, as well as system design such as chan-
nel estimation. Therefore, an accurate channel modeling
scheme for EIT in NLoS scenarios is needed.
B. Our contributions
Different from the existing works based on the far-field
assumption, in this paper, we propose a near-field channel
model and corresponding channel estimation scheme for
EIT2 . Specifically, the contributions of this paper are
summarized as follows:
• We propose a near-field channel model for EIT based
on the electromagnetic scattering theory. The channel
is modeled by zero-mean Gaussian random fields.
Then its correlation function can fully describe the
channel. An approximate analytical expression of the
correlation function of the channel is derived, and
its approximation accuracy is verified by numerical
simulation.
• We analyze the characteristics of the proposed near-
field model. We show how to generate one sample of
the random field channel model. Then, we analyze
how the parameters of the model affects the degree
of freedom (DoF) of the channel. Finally, we show
the sparsity of the channel model in the wavenumber
domain.
• We design a channel estimation scheme based on
the sparse property of the proposed model, which
provides electromagnetic prior information for chan-
nel estimation. Numerical simulations show that the
designed channel estimation scheme outperforms ex-
isting schemes.
C. Organization and notation
Organization: The rest of this paper is organized as fol-
lows. In Section. II, we provide the electromagnetic model
of NLoS channel based on the electromagnetic scattering
theory. Then, in Section. III, we derive an approximated
analytical expression of the correlation function of the
channel based on the electromagnetic model. In Section.
IV, we analyze the characteristics of the proposed model,
including DoF, sparsity, etc. Based on the sparsity of
the model in the wavenumber domain, in Section. V
we propose a near-field channel estimation scheme and
verify its correctness by numerical simulations. Finally, in
Section. VI we provide the conclusions and possible future
directions of our work.
Notation: bold uppercase characters denote matrices;
bold lowercase characters denote vectors; the dot · denotes
the scalar product of two vectors, or the matrix-vector
multiplication. E [x] denotes the mean of random variable
x; ϵ0 is the permittivity of a vacuum, µ0 is the permeability
of a vacuum, and c is the speed of light in a vacuum; ∗
2 Simulation codes will be provided to reproduce the results in this
paper: http://oa.ee.tsinghua.edu.cn/dailinglong/publications/publications.
html.
2
denotes the convolution operation, and F [f (x)] denotes
the Fourier transform of f (x); Jm (x) is the mth order
Bessel function of the first kind; Im (x) is the mth order
modified Bessel function. ⌊x⌋ represents rounding x down;
x%y represents modulo operation. det(·) denotes the
matrix determinant or the Fredholm determinant.
II. E LECTROMAGNETIC M ODEL FOR S CATTERING
F IELD
Maxwell’s equations are the fundamental physical laws
of the electromagnetic system. For the characteristics of
the scattering system, we can consider the scatterers as
spatial non-uniformity of the electromagnetic characteris-
tics like permittivity ϵ and permeability µ in the space.
From the Maxwell’s equations and the time-harmonic
assumption [25] we have
∇ × E(r) = jωµ(r)H(r), (1a)
∇ × H(r) = −jωϵ(r)E(r) + J(r), (1b)
∇ · (ϵ(r)E(r)) = ρ(r), (1c)
∇ · (µ(r)H(r)) = 0, (1d)
where ϵ(r), µ(r) and ρ(r) represents the permittivity,
permeability and charge density at the position r. In ho-
mogeneous media, ϵ(r) and µ(r) will be constant, which
is often used in light-of-sight channel modeling in the free
space. Now we are considering inhomogeneous media,
which can express the electromagnetic characteristics of
scattering fields [19].
By simplifying the Maxwell’s equations, we can obtain
the corresponding vector wave equation
(∇2 + k 2 (r))E(r) = q(r), (2)
p
where k(r) = ω µ(r)ϵ(r) represents the inhomogeneous
media over a finite domain V according to [19], and
q(r) = jωµ(r)J(r) represents the source field. Outside the
√
domain V , the wavenumber k(r) equals k0 = ω µ0 ϵ0 . By
jk∥r−r′ ∥
using the Green’s function g(r, r′ ) = 4π∥r−r
e
′ ∥ which is
the solution of (∇ + k )g(r, r ) = −δ(r − r′ ), we can
2 2 ′
derive
Z Z
E(r) = − g(r, r′ )q(r′ )dr′ + g(r, r′ )(k 2 (r′ ) − k02 )
Vs V
E(r′ )dr′ ,
(3)
where Vs is the source region which generates the signal,
and E(r′ ) is the induced electric field in the inhomo-
geneous regions in the space. The equation (3) is an
extension from the 2-dimensional case in [26]. Here we
can view the first item in (3) as the line-of-sight component
of the field which is fixed and well-studied. Then we will
focus on the second item in (3) which highly relies on
the characteristics of the inhomogeneity of the space. The
inhomogeneity of the space depends on the complicated
factors such as surface structure and material properties
of the medium which are hard to analytically model and
 3

y
Receiver
Scattering region
̂
Scattering region eˆx
| r  r1 |
rs Transmitter

r1
d 
O
x rs
z Td

Tr1 Tr2
Fig. 1. The three-dimensional near-field statistical channel modeling
where the scatterers are located in solid circles. O
Fig. 2. The rotated coordinate system with Tµ̂ = êx .
may change over time. Therefore, a statistical model will
be more suitable to capture the characteristics of the field
than deterministic modeling scheme. receivers at a certain angle. The radius of the circle is
rs . For the item (k 2 (r′ ) − k02 )2 , we view it as the gain
III. C HANNEL M ODEL BASED ON N ON -S TATIONARY of electromagnetic waves reflected from the surface of
R ANDOM F IELDS scatterer determined by the electromagnetic characteristics
In this section we will derive the channel model for EIT of the scattering surface. Specifically, we model it by
based on the electromagnetic scattering theory explained a+1 2
in the above section. The channel is modeled as Gaussian f (r′ ) = (k 2 (r′ ) − k02 )2 = (rs − ρ2 )a , (6)
πrs2a+2
random fields to tolerate the uncertainty of the inhomo-
geneity of the space [27]. Since in our model the received where ρ = r′ − d, ρ = ∥ρ∥, and a is a parameter
field can be viewed as weighted superposition of spherical characterizing the concentration of scatterer around the
waves other than plane waves in [22], it is suitable for both central point. When a = 0, the scattering region is a
near-field and far-field communications by considering uniform circular surface. When a → −1, the scatterer
distances between antenna array and scatterers besides the approximates a ring. When a → +∞, the scattering
azimuth and elevation angles. region shrinks to a single point. Then, we can express
the correlation function by
′ ′
ejk∥r1 −r ∥ e−jk∥r2 −r ∥
Z
A. Mathematical derivation of the analytical model
RE (r1 , r2 ) = β ′ ′
f (r′ )dr′ .
By omitting the first item in (3) which represents the V 4π∥r1 − r ∥ 4π∥r2 − r ∥
(7)
deterministic line-of-sight component, we have
To facilitate the derivation procedure, we choose a
hZ Z
RE (r1 , r2 ) =E g(r1 , r′1 )g ∗ (r2 , r′2 )(k 2 (r′1 ) − k02 ) coordinate rotation T which satisfies Tµ̂ = êx . Then
V V i we have a new rotated coordinate where µ is the x axis.
(k 2 (r′2 ) − k02 )E(r′1 )E ∗ (r′2 )dr′1 dr′2 . The center of the scatterer is located at Td, and the
receiving locations are Tr1 and Tr2 . One point in the
(4)
scattering region is located at Td + Tρ, where Tρ =
To derive a closed-form expression of the channel model
[ρ cos θ, ρ sin θ, 0]. Here we denote two directions µ̂1 and
we need to have some assumptions on the scattering
µ̂2 perpendicular to µ̂, which satisfies µ̂T 1 µ̂2 = 0. Then
field to do simplifications on (4). Considering the un-
we can denote Td by [d · µ̂, d · µ̂1 , d · µ̂2 ]. Similarly, we
certainty of the inhomogeneity of the space, we assume
have Tr = [r · µ̂, r · µ̂1 , r · µ̂2 ]. The point in the scattering
E [E(r′1 )E ∗ (r′2 )] = βδ(r′1 − r′2 ) for simplicity. Then, the
point is located at [d · µ̂, d · µ̂1 + ρ cos θ, d · µ̂2 + ρ sin θ].
channel correlation function reduces to
Z The rotated coordinate system is shown in Fig. 2. The
RE (r1 , r2 ) =β g(r1 , r′ )g ∗ (r2 , r′ )(k 2 (r′ ) − k02 )2 dr′ . distance between r and r′ is (8) where
V
(5)  r 2 r
We further assume that the scattering region V is dis- A(r) =1 + −2 (d̂ · µ̂)(r̂ · µ̂)
d d
tributed in a solid circle, centering at d and perpendicular ! (9)
to µ̂, which means that µ̂T (r′ − d) = 0. The practical
+ (d̂ · µ̂1 )(r̂ · µ̂1 ) + (d̂ · µ̂2 )(r̂ · µ̂2 ) ,
meaning of such assumption is that the scatterer faces
 4

q
2 2 2
∥r − r′ ∥ = (d · µ̂ − r · µ̂) + (d · µ̂1 + ρ cos θ − r · µ̂1 ) + (d · µ̂2 + ρ sin θ − r · µ̂2 )
r  ρ 2 (8)
ρ
= d A(r) + 2 B(r, ρ̂) + ,
d d

and
r
B(r, ρ̂) =d̂ · µ̂1 cos θ + d̂ · µ̂2 sin θ − r̂ · µ̂1 cos θ
d
r
− r̂ · µ̂2 sin θ.
d
(10)
Through mathematical derivations and simplifications,
we can derive the spatial correlation function of the
channel in the following lemma:
Lemma 1 (Correlation function of the channel). Assuming
that rs ≪ d, the correlation function of the channel can
be approximated by
√ √
β

j 2π R A(r1 )− A(r2 )
R̃(r1 , r2 ) = p e λ
8π 2 d2 A(r1 )A(r2 )
√ √
(a + 1)2a Γ(a + 1)( Crs )−(a+1) Ja+1 ( Crs ),
(11)
where Fig. 3. The correlation function plotted from the approximated analytical
expression.
 2
2π d̂ · µ̂1 d̂ · µ̂1 r1 r̂1 · µ̂1
C= p −p − p
λ A(r1 ) A(r2 ) d A(r1 )
!2 correlation between the received fields at two positions on
r2 r̂2 · µ̂1 the receiving array, we fix one position at the center of the
+ p array and another position at [0, ±ny dy , ±nz dz ] m, where
R A(r2 )
 2 ny , nz ∈ IN , IN = {1, · · · , 100}, dy = dz = 0.025 m.
2π d̂ · µ̂2 d̂ · µ̂2 r1 r̂1 · µ̂2 ∥R̃−R∥2
+ −p − p The wavelength λ is set to 0.05 m. We plot ∥R∥2 F ,
p F
λ A(r1 ) A(r2 ) d A(r1 ) which is the relative error between the approximated
!2 correlation matrix and the accurate correlation matrix,
r2 r̂2 · µ̂2 in Fig. 3. We can find that the approximation error is
+ p .
d A(r2 ) negligible compared to the value of the corresponding
(12) correlation function when d is large enough or rs is small
Proof: See Appendix A. enough. For example, when d is larger than 100 m and rs
For the channel with multiple scatterers, the correlation is smaller than 3.5 m, the relative approximation error is
function can be expressed by below 1%, which is tolerable in most cases.
M
X
R(r1 , r2 ) = R̃k (r1 , r2 ), (13) C. Fitness to the statistics of classical CDL model
k=1 In this part we will show the fitness of the proposed
where each R̃k (d1 , d2 ) is constructed according to model to the statistics of standard 3GPP TR 38.901 CDL
Lemma 1. model [28]. Since CDL model is now widely used in
5G new radio (5G NR) scenarios, the rationality of the
proposed analytical correlation function of the channel
B. Numerical verification of the accuracy of the analytical
model can be verified if it can well fit the statistics of the
model
CDL model. We simulate the field correlation of CDL-
In this subsection, we will show the accuracy of the A and CDL-D model, which represent strong scattering
analytical correlation function in Lemma 1. We set the di- and weak scattering scenarios separately. For the antenna
rection of the scattering region to the center of the array as array, we adopt 20 × 20 array with λ/8 antenna spacing.
d̂ = [ √13 , √13 , √13 ]. The scattering region is perpendicular We use the proposed analytical model with two scatterers
to the direction µ̂ = [− √13 , √13 , − √13 ]. The concentration to fit the field correlation of the CDL model, which is
parameter on the scatterer cluster is set to a = 0, which shown in Fig. 4. It is shown that the proposed model can
corresponds to uniform distribution on the circle. For the fit the statistical characteristics of CDL models with few
 5

(Ny Nz )×(Ny Nz )
hC , where Ri,j = R(rii, rj ), ri =
Ny −1 Nz −1
0, ⌊ i−1
Nz ⌋ − 2 , mod (i − 1, Nz ) − 2 , and rj =
h i
N −1 Nz −1
0, ⌊ j−1
Nz ⌋ −
y
2 , mod (j − 1, Nz ) − 2 . From the
correlation function of the received field, we can generate
1 1
the channel by h = R 2 N, where R 2 is the Cholesky
decomposition of the correlation matrix R, and N ∼
(a) CDL-D model (b) proposed model CN (0, I).

B. Impact of the parameters on the analytical model

(c) CDL-A model (d) proposed model
Fig. 4. Comparison between the field correlation of CDL-A, CDL-D
and the proposed model.
parameters, especially for CDL-D model, which verifies its
correctness and generalization capability. The benefits of
the proposed model is that it is analytical and can be used
to obtain the field correlation between any two positions
by direct and quick calculation.
IV. C HARACTERISTICS OF THE P ROPOSED C HANNEL
M ODEL
In this section, we will analyze and show the character-
istics of the derived channel model, which can reveal how
the scattering environment affects the system performance.
Moreover, the characteristics of the channel model may
inspire possible applications of it, e.g., channel estima-
tion schemes based on the sparsity of the model in the
wavenumber domain.
A. One realization of the random field
′
P∞ R(r, r ),
For the derived correlation function we have
∗ ′
the following expansion R = λ
i=1 i ϕ(r)ϕ (r ) from
Mercer’s theorem, where ϕ(r) is the solution of the
following integral equation
Z Then, we will discuss how the parameters influence the
λi ϕ(r) = R(r, r′ )ϕ(r)dr, (14)
V
according to [29]. Then the received field can be con-
structed by its Karhunen-Loève expansion
∞
X can support communication at a certain rate, leading
E(r) = ξi ϕ(r),
i=1
where λi = E[ξi ξi∗ ]. For a noisy received field Y (r) =
E(r) + N (r) where RN (r, r′ ) = σ 2 δ(r − r′ ), the in-
formation that
P can be obtained from the received field is
I(E; Y ) = i log(1 + σλ2i ) [18].
If we consider discrete samples of the contin-
uous fields, for a Ny × Nz array at the re-
ceiver, we can construct a correlation matrix R ∈
In this part, we will discuss the impact of parameters
on the accuracy and performance of the analytical model.
First we will show how scatterer size rs reflect the
accuracy of the model and when it can not be neglected.
It is well known that the Rayleigh distance, also called
8r 2
as Fraunhofer distance, is d = λm , where d is the
distance from the antenna array, and rm = max∥r∥
is the radius of the antenna array [30]. The Rayleigh
distance is defined by the distance where π8 phase error
is observed on the antenna array. If we further consider
the size of the scatterer, we have the channel response as
2π ′
h(r′ , r) = ej λ ∥r −r∥ , where r′ = d + ρ is the position
of one point on the scatterer. In Lemma 2 we extend the
Rayleigh distance considering scatterers
Lemma 2 (Extension of Rayleigh distance considering
scatterer size). The size of scatterer can be neglected
λ 8(rs2 +rm
2
) λ
when rs ⩽ 16 and d ⩽ λ−16r s
. If rs ⩾ 16 , the
scatterer size should be considered in2 the2 channel model.
8(rs +rm )
Under this scenario, when d ⩽ λ , the scatterer
and the antenna array are in the near-field region. When
8(r 2 +r 2 )
d ⩾ s λ m , the scatterer and the antenna array are in
the far-field region.
Proof: See Appendix B.
From Lemma 2 we know that unless the scatterer is
small enough (for frequency of 1 GHz the radius of the
scatterer should be smaller than 0.0187 m), neglecting the
size of the scatterer and simply view it as a point will be
inaccurate. Therefore, considering parameters of scatterer
is of necessity in channel modeling especially in near-field
communication scenarios.
performance of the system from the degree of freedom
(DoF) perspective. The DoF of the channel depends on
the eigenvalue distribution of the model. If the eigenvalue
decay rate is slow, there exist multiple subchannels that
(15) to higher DoF. On the contrary, if few eigenvalues are
obviously larger than other eigenvalues, the DoF will be
low [21]. Then, we plot the eigenvalues of the correlation
matrix when the radius and shape of the scattering region
vary in Fig. 5 and Fig. 6. We can observe that the DoF of
the channel will increase with the radius r of the scatterer,
and decrease when a increases. When a approximates −1,
which corresponds to the case that the scatterer tends to a
ring, the DoF of the channel reaches the maximum.
 6

0.7 function only relies on the distance vector between the

a = -0.99
a = -0.5 two points. If we perform Fourier transformation on the
0.6 a=0
a=3 correlation function, we will obtain its power spectrum in
the wavenumber domain. To be more specific, we have
0.5
Z +∞ Z +∞
0.4 R(∆r)ej(ky y+kz z) dydz
−∞ −∞
Z +∞ Z +∞
0.3 ′ ′ ′
= β0 e−j(r̂x x+r̂y y+r̂z z) ej(ky y+kz z) dydz
−∞ −∞
0.2
′
= β0 e−jr̂x x δ(ky − r̂′y )δ(kz − r̂′z ).
0.1 (17)
Therefore, the Fourier transform of the far-field correlation
0
1 2 3 4 5 6 7 8 9 10
function reveals the angular concentration of the scattering
regions in the wavenumber domain. A scattering region
with the azimuth angle θ and elevation angle ϕ will
Fig. 5. The eigenvalues of the correlation matrix in decreasing order lead to a single point [cos(θ) sin(ϕ), sin(θ) sin(ϕ), cos(ϕ)]
with d fixed to [−100, 100, −100] m, µ fixed to [− √1 , √1 , − √1 ],
3 3 3 in the wavenumber domain of the received field un-
and r = 5 m. The concentration parameter a varies.
der far-field assumption. If we sample the continuous
received electromagnetic fields to obtain a correlation
1
r=1 matrix R ∈ CNy ×Nz where Ri,j = R(r0 , r) and r =
0.9 r=2
r=4
[0, ±ny dy , ±nz dz ], we can use Fourier transform matrices
0.8 r=8 F1 and F2 instead continuous Fourier transform to find
0.7
the angular sparsity of the correlation matrix by F1 RFH 2.
Specifically, the Fourier transform matrix F1 and F2 can
0.6 2k Ny +1 Ny +1
j y (i− )(j− )d
be constructed by F1,i,j = e Ny −1 2 2
and
0.5 j Nz2k−1 (i− Nz2+1 )(j− Nz2+1 )dz
F2,i,j = e .
0.4
According to the law of large numbers, we know
0.3 that h has sparsity in wavenumber domain due to the
0.2 sparsity of the correlation function, which can be used in
channel estimation or channel reconstruction to improve
0.1
the accuracy. Specifically, if we reshape the vector h to a
0
1 2 3 4 5 6 7 8 9 10 matrix Hi,: = hT (i−1)∗Nz +1:i∗Nz ,1 , we have
 
2
2
X X
Fig. 6. The eigenvalues of the correlation matrix in decreasing order E[|(F1 HF2 )i,j | ] = E 

F1,i,j ′ Hi′ ,j ′ FH

with d fixed to [−100, 100, −100] m, µ fixed to [− √1 , √1 , − √1 ], 2,i′ ,j 
3 3 3 i′ j′
and a = 0. The radius r varies.
 
X
∗ ∗
= E F1,i,j1′ Hi′1 ,j1′ FH T
2,i′1 ,j F1,i,j2′ Hi′2 ,j2′ F2,i′2 ,j

C. Sparsity in the wavenumber domain
i′1 ,i′2 ,j1′ ,j2′
In this part we will discuss the sparsity of the proposed X j Ny2k−1 (i−
Ny +1
)(j1′ −j2′ )dy
channel model in the wavenumber domain, which may = e 2
R(r1 , r2 )
provide side information for channel estimation. First note i′1 ,i′2 ,j1′ ,j2′
that with far field approximation the correlation function 2k Nz +1
)(i′2 −i′1 )dz
ej Nz −1 (j− 2 ,
of the random channel can be simplified to (18)
N +1
′ ′ where r1 = [0, (i′1 − y2 )dy , (j1′ − Nz2+1 )dz ] and r2 =
ejk∥r1 −r ∥ e−jk∥r2 −r ∥
Z
R(r1 , r2 ) = β ′ ′
f (r′ )dr′ N +1
[0, (i′2 − y2 )dy , (j2′ − Nz2+1 )dz ] respectively. From (16)
V 4π∥r1 − r ∥ 4π∥r2 − r ∥
Z jk(r′ −r̂′ ·r1 ) −jk(r′ −r̂′ ·r2 ) it is easy to know that when rm and rs approximates 0,
e e 2
rm →0 E[|(F1 HF2 )i,j | ] will reaches a peak value compared to
≈ β 2 ∥r − r′ ∥2
f (r′ )dr′
V 16π 0 its neighbors, which is in the form of products of sinc
rs →0 βf (r′ ) ′ function as the discretized form of (17).
≈ e−jkr̂ ·(r1 −r2 ) , For the near-field scattering scenario, the scattering
16π ∥r0 − r′ ∥2
2
(16) region will correspond to an area rather than a point in
where r0 is the position of the center of the receiver array. the wavenumber domain. The shape of the area reflects
Therefore, the received field under such approximation the size, directions and concentration parameters of the
is a stationary field which implies that its correlation scattering region. We plot the correlation function of a
 7

Fig. 7. The correlation function of the received field.

generated channel in Fig. 7, and its Fourier transform in

Fig. 8. Three scattering regions are located in the space, Fig. 8. The Fourier transform of the correlation function of the received
with coordinates d = [25, 25, 25] m, d = [25, −25, 50] m, field. Three spots in the figure corresponds to three scattering regions in
and d = [25, −25, −50] m separately. While little in- the space.
formation can be directly observed from the figure of
the correlation function of the received field, the Fourier Algorithm 1 TDN OMP
transform of the correlation function reflects its sparsity Input:
in the wavenumber domain. Three shaded areas in Fig. 8 y % the received pilot
correspond to three scattering regions in the settings, with L % number of paths
their respective parameters labeled adjacent to the shaded W % the three dimensional codebook
areas. It is shown that when R increases and r decreases, Output:
the size of shaded areas will increase, which corresponds h̃ % the estimated channel
to larger angular expansion in the wavenumber domain. 1: Initialization: Y = y, γ = {∅}
When a tends to infinity, the shaded area tends to a single 2: for l ∈ {1, 2, · · · , L} do
point. When a tends to −1, the shaded area tends to a 3: Calculate the correlation matrix: Γ = WH Y
circle, which aligns with the definition of function f (r). 4: Detect new support: p∗ = argmaxp |Γp |
5: Update support set: γ = γ ∪ p∗
V. C HANNEL E STIMATION BASED ON THE P ROPOSED 6: Pseudo inverse: W† = (W:,γ T
W:,γ )−1 W:,γ
T
M ODEL 7: P
Orthogonal projection: h = W y †

After discussing the properties of the analytical channel
model, we will propose a near field channel estimation
scheme based on the model. By integrating the prior in-
formation of electromagnetic fields, the channel estimation
scheme based on the proposed model can achieve better
performance than the existing schemes. In the channel
estimation
√ procedure, the received field is denoted by
y = P h + n, where P is the signal-to-noise ratio,
h is generated from the channel correlation matrix, and
n ∼ CN (0, I) is the noise vector.
A. LS channel estimation
The simplest channel estimation scheme is the least
√ (LS) channel estimation, which leads to h̃ =
square
y/ P .
B. OMP based channel estimation
Another widely-used scheme is orthogonal matching
pursuit (OMP) [31], which performs well when the signal
8: Update residual: Y = Y − W:,γ hP
9: h̃ = W:,γ hP
10: return h̃
is sparse. For the transform matrix W in the three-
dimensional domain, we adopt the codebook in [32]. Since
the three-dimensional codebook provide approximate or-
thogonal basis for the near-field channel, it fully exploits
the angular and distance information of the channel. We
have y = Whm + n, where hm is sparse, and the
three-dimensional near-field (TDN) OMP algorithm for
the channel estimation problem is shown in Algorithm 1.
C. Subspace based channel estimation
From [23] it is known that we can estimate a channel by
using the subspace of an omni-directional channel model.
When isotropic scattering environment is considered, the
 8

correlation function at the receiver is assumed to be Algorithm 2 NFS correlation function generator

2∥r1 − r2 ∥
 Input:
R(r1 , r2 ) = sinc . (19) y % the received pilot
λ
Ny , Nz % number of antennas
Then, the correlation matrix R is sampled from the dy , dz % antenna spacing
correlation function. A compact eigenvalue decomposition F1 , F2 % Fourier transform matrix
is performed on R to obtain R = U1 Λ1 UH 1 , where λ % wavelength
Λ1 contains the non-zero eigenvalues of R. The channel η % threshold
estimator is expressed by d, r, a, µ % fixed parameters for simplicity
√ Output:
h̃ = U1 UH
1 y/ P . (20) R̂ % the constructed correlation matrix
T
In fact, if we further utilize the information contained 1: Vector to matrix: Yi,: = y(i−1)∗N
z +1:i∗Nz ,1
in the eigenvalues of R, the estimation precision can be 2: Fourier transform on both sides: Y ′ = F1 YFH 2
improved, which corresponds to the channel estimator 3: Average value: |Y ¯ ′ | = sum (|Y′ |) /Ny Nz
√ 4: Initialization: R = zeros(Ny Nz , Ny Nz ), Φ = ∅
h̃ = P U1 (P Λ1 + I)−1 UH 1 y. (21)
5: Peak value selection:
6: for i = 1 : Ny do
D. Proposed channel estimation scheme 7: for j = 1 : Nz do
8: ′
if |Yi,j | > max |Yi±1,j±1 ′ ¯ ′ |
then
|, η |Y
We propose a channel estimation scheme based on the N −1−2∗i
sparsity of the channel model. By reshaping channel vector 9: ky = yNy −1
to Hi,j = h(i−1)∗n+j , we obtain matrix H which has 10: kz = 2∗j−N Nq
z +1
z −1
sparsity in the wavenumber domain according to Section 2
11: d = d ∗ [ 2π 2 2
λ − ky − kz , ky , kz ]
T
IV. Therefore, we can detect the peaks in the wavenumber
12: Φ = Φ ∪ {d}
domain to capture the directions of the incident waves.
Then we can generate an approximate near-field corre- 13: for i = 1 : Ny Nz do
lation matrix of the electromagnetic field. The genera- 14: for j = 1h: Ny Nz do i
Ny −1 Nz −1
tion procedure of the approximated correlation matrix is 15: r1 = 0, ⌊ i−1Nz ⌋ − 2 , (i − 1)%N z − 2
shown in Algorithm 2. After obtaining the approximate
h i
Ny −1 Nz −1
16: r2 = 0, ⌊ j−1Nz ⌋ − 2 , (j − 1)%N z − 2
correlation
√ matrix, we then use the following estimator
h̃ = P R̂(P R̂ + I)−1 y. 17: Generate the correlation function R according
Here for simplicity we only design the correlation 18: to Lemma 1 based on Φ
matrix based on the estimated incident wave direction. 19: R̂i,j = R(r1 , r2 )
Obviously, the approximation of the correlation matrix 20: return R̂
can be improved by further estimating or optimizing the
parameters r, a and µ. In fact, the estimation error when
using an approximated correlation matrix R′ instead of In Fig. 9 we have shown the performance compari-
the true correlation matrix R can be expressed by the son between different channel estimation schemes shown
following lemma: above. Here we use the proposed near-field channel model
Lemma 3 (Estimated error when using the proposed to generate the channel realizations, where four scattering
scheme). The estimatedherror when using an regions exist in the space. We set the antenna array to
i approximated
 be 80 × 20 with λ/8 antenna spacing. The wavelength of
H
correlation matrix is E (h̃ − h) (h̃ − h) = tr P (P I +
the electromagnetic field is set to 0.2 m. Specifically, we
R̂−1 −1 −1 −1 −1 −1
 ) (P I + R̂ ) (P R + I) − 2P R(P I + R̂ ) + plot the normalized mean square error (NMSE) ∥h̃−h∥ ∥h∥
R . with the change of SNR P . It can be obeserved that
the proposed scheme outperforms traditional schemes like
Proof: See Appendix C.
OMP or subspace based channel estimation scheme. For
Corollary 1. When P → 0, the estimated error will example, when NMSE equals 3 × 10−3 , the proposed
approach tr(R). When P → ∞, the estimated error will scheme achieves 5dB performance gain compared to OMP
approach 0 whatever the approximate correlation matrix scheme whose support set is 20. The reason that the
R̂ is. Therefore, the performance limit with extremely high proposed scheme can outperform existing schemes can
or low SNR does not depend on the choice of R̂. However, be explained as follows. For the subspace based channel
for general P , the approximation error will reach the estimation scheme, it considers the incident waves from
minimum value when R̂ = R, which corresponds to the all directions, which covers the full wavenumber domain.
classical minimum mean square error (MMSE) channel On the contrary, the proposed scheme focuses on a smaller
estimator with full information of the distribution of the region in the wavenumber domain, thus providing a better
electromagnetic fields. approximation of the true correlation function. For the
 9

101 101
LS LS
OMP-4 Subspace
OMP-10 OMP-20
100 100 OMP-40
OMP-20
Subspace Proposed
Proposed
10-1 10-1

10-2 10-2

10-3 10-3

10-4 10-4

10-5 10-5
-10 -5 0 5 10 15 20 25 30 -10 -5 0 5 10 15 20 25 30

Fig. 9. Comparison of different channel estimation schemes for a 80 × Fig. 10. Comparison of different channel estimation schemes for a 40 ×
20 antenna array with λ/8 antenna spacing. Four scattering regions are 40 antenna array with λ/8 antenna spacing. The CDL-D channel model
distributed in the space. is used.

OMP scheme, note that it highly relies on the lattice points Further work can be done by integrating the proposed
of the electromagnetic field, it may not behave well in high model and traditional near-field model where some scat-
SNR region for h generated from the correlation function terers are invisible to part of the array.
in continuous space.
Furthermore, we have applied the proposed scheme on
CDL-D channel instead of the channel generated by our A PPENDIX A
correlation matrix to further verify its correctness. We P ROOF OF L EMMA 1
adopt a CDL-D channel model which has 40 × 40 size
Based on the assumption that rs = maxρ ≪ d, we
receiver array with λ/8 antenna spacing. The wavenumber
can use the Taylor expansion
to simplify ∥r1 − r′ ∥ and
of the electromagnetic field is set to be 0.4 m. The simu- 2
∥r2 − r′ ∥. First the item dρ can be ignored. Then  we
lation result is shown in Fig. 10. It can be observed that
B(r,ρ̂)
A(r) + dρ √
p
the proposed channel estimation algorithm can also work can approximate ∥r − r′ ∥ by d . The
A(r)
under classical channel model and outperform existing
Green’s function g(r, r′ ) has amplitude item 4π∥r−r
1
′ ∥ and
algorithms like LS and subspace based channel estimation ′
jk∥r−r ∥
algorithm. For OMP algorithm, its performance is better phase item e . For the distance item we further ap-
than the proposed algorithm when SNR is lower than proximate it by √1 . Then, the correlation function
4πd A(r)
10 dB. However, it will still face error platform in the high of the received field can be approximated by
SNR region, which can be solved by using the proposed Z
algorithm. For example, the proposed scheme can achieve 1
R(r1 , r2 ) ≈ R̃(r1 , r2 ) = β p
5dB performance gain compared to 40-points OMP when V 16π d
2 2 A(r1 )A(r2 )
NMSE is fixed to 2 × 10−3 . √ √  2π

B(r1 ,ρ̂) B(r ,ρ̂)

j 2π R A(r )− A(r ) j λ ρ √ −√ 2
e λ 1 2
e A(r1 ) A(r2 )
f (r′ )dr′ ,
(22)
VI. C ONCLUSION where f (r′ )dr′ = πra+1 ρ(r 2
− ρ 2 a
) dρdθ. Specifically,
2a+2
s
s

In this paper, we propose the near-field channel model if a = 0, we have uniform distribution on the scatterer,
for EIT based on electromagnetic scattering theory. Then, where f (r′ )dr′ = πrρ 2 dρdθ.
s

we derive the analytical expression of the correlation For the simple case with uniform distribution on the
function of the fields and analyze the characteristics it. circle, we have
Finally, we design a channel estimation scheme for near- √ √
1

j 2π
λ R A(r1 )− A(r2 )
field scenario. Numerical analysis verifies the correctness R̃(r1 , r2 ) =β p e
of the proposed scheme and shows that it can outper- 16π 2 d2 A(r1 )A(r2 )
 
form existing schemes like LS, OMP and subspace based rs Z 2π j 2π ρ √
Z B(r1 ,ρ̂) B(r ,ρ̂)
λ A(r1 )
−√ 2
A(d2 )
ρ
channel estimation schemes. Under CDL channel model, e 2
dθdρ.
0 0 πr s
the proposed scheme can achieve 5dB performance gain (23)
compared to 40-points OMP when NMSE is fixed to We first focus on the integral on the angle θ. Since B(r, ρ̂)
2 × 10−3 . contains the exponentionals of cos θ and sin θ, we adopt
 10

the [33, Eq. (3.937)] which shows that ing to [33, Eq. (6.567)] we have
√ √
β

j 2π
λ R A(r1 )− A(r2 )
R̃(r1 , r2 ) = p e
16π 2 d2 A(r1 )A(r2 )
Z 2π Z rs √
a+1
ep cos x+q sin x ej(a cos x+b sin x−mx) dx 2π 2a+2 J0 (ρ C)ρ(rs2 − ρ2 )a dρ
0 πrs 0 √
m m √
= 2π[(b − p)2 + (a + q)2 ]− 2 (A − jB) 2 Im ( C + jD),
p
β

j 2π R A(r1 )− A(r2 )
= e λ
(24)
p
16π 2 d2 A(r1 )A(r2 )
where A = p2 − q 2 + a2 − b2 , B = 2pq + 2ab, C = √
2(a + 1) 2a+2 1
Z
p2 + q 2 − a2 − b2 and D = −2ap − 2bq. Comparing with 2a+2 rs J0 (ρ′ rs C)ρ′ (1 − ρ′ )2 dρ′
rs 0 √
2π √d̂·µ̂1 − √
β

the integral, we have p = q = m = 0, a = λ ρ j 2π R A(r1 )− A(r2 )
A(r1 ) = p e λ
! 16π 2 d2 A(r1 )A(r2 )
√ √
√d̂·µ̂1 − rd1 √r̂1 ·µ̂1 + rR2 √r̂2 ·µ̂1 , b = − 2π
λ ρ
√d̂·µ̂2 − 2(a + 1)2a Γ(a + 1)( Crs )−(a+1) Ja+1 ( Crs ).
A(r2 ) A(r1 ) A(r2 ) A(r1 )
! (28)
√d̂·µ̂2 − r1
d
√r̂1 ·µ̂2 + r2
d
√r̂2 ·µ̂2 . Then, we can obtain
A(r2 ) A(r1 ) A(r2 )
A PPENDIX B
P ROOF OF L EMMA 2

√ √ We have
1

j 2π
λ R A(r1 )− A(r2 )
R̃(r1 , r2 ) = β p e
16π 2 d2 A(r1 )A(r2 ) ∥d + ρ − r∥
2
Z rs √ p
I0 (jρ C)ρdρ, = d2 + ρ2 + r2 + 2d · ρ − 2d · r − 2ρ · r
rs2 0 r
(25)
 ρ 2  r 2 d·ρ d·r ρ·r
=d 1+ + +2 2 −2 2 −2 2 .
where d d d d d
√ (29)
According to the Taylor expansion 1 + x ≈ 1+ 21 x− 18 x2
and the assumption that d ≫ max(r, ρ), we have that
 2
2π d̂ · µ̂1 d̂ · µ̂1 r1 r̂1 · µ̂1 ∥d + ρ − r∥
C= p −p − p
λ A(r1 ) A(r2 ) d A(r1 )
d·ρ d·r ρ2 r2 ρ·r
!2 ≈d 1+ − 2 + 2+ 2− 2
r2 r̂2 · µ̂1 d 2 d 2d 2d d (30)
+ p
R A(r2 ) !
(d · ρ)2 (d · r)2 (d · ρ)(d · r)
 2 − − + ,
d̂ · µ̂2
2π d̂ · µ̂2 r1 r̂1 · µ̂2 2d4 2d4 2d4
+ p −p − p
λ A(r1 ) A(r2 ) d A(r1 )
!2 where higher orders of dρ and dr are neglected. For far-
r2 r̂2 · µ̂2 field channel modeling without considering the size of
+ p .
d A(r2 ) scatterers, only the terms d 1 − d·r


d 2 is kept, which means
(26) that the rest terms should be small enough. If we adopt
Then, according to [33, Eq. (6.561)] we have the π8 phase error as the threshold, we have

2π d · ρ ρ2 r2 ρ · r (d · ρ)2 (d · r)2
+ + − − −
√ √ λ d 2d 2d d 2d 3 2d3
β

j 2π
λ R A(r1 )− A(r2 )
R̃(r1 , r2 ) = p e
16π 2 d2 A(r1 )A(r2 ) (d · ρ)(d · r)
+
2
Z rs √ 2d3
J0 (ρ C)ρdρ
rs2 0 2π

rs2 2
rm

π
√ √
β j 2π

⩽ rs + + ⩽ ,
λ R A(r1 )− A(r2 ) λ 2d 2d 8
= p e
8π 2 d2 A(r1 )A(r2 ) (31)
1 √ which leads to (λ−16rs )d ⩾ 8(rs2 +rm 2
). Therefore, when
√ J1 ( Crs ). λ
the radius rs of the scatterer is larger than 16 , the scattterer
rs C
(27) size has to be taken
 into consideration.
 Moreover, if we
If we adopt f (r′ )dr′ = a+1
ρ(rs2 −ρ2 )a dρdθ, accord- keep the terms d 1 + d·ρ d·r
d2 − d2 , which means that the
πrs2a+2
 11

antenna array and the scatterer are in each other’s far-field [7] M. Chafii, L. Bariah, S. Muhaidat, and M. Debbah, “Twelve
respectively, we have scientific challenges for 6G: Rethinking the foundations of com-
munications theory,” IEEE Comm. Surveys Tut., Feb. 2023.
2 2 2 2 [8] M. D. Migliore, “Horse (electromagnetics) is more important than
2π ρ + r ρ · r (d · ρ) (d · r) (d · ρ)(d · r) horseman (information) for wireless transmission,” IEEE Trans.
− + + −
λ 2d d 2d3 2d3 2d3 Antennas Propag., vol. 67, no. 4, pp. 2046–2055, Apr. 2018.
 2 [9] J. Zhu, Z. Wan, L. Dai, M. Debbah, and H. V. Poor, “Electromag-
2

2π rs + rm π netic information theory: Fundamentals, modeling, applications,
⩽ ⩽ , and open problems,” arXiv preprint arXiv:2212.02882, Dec. 2022.
λ 2d 8 [10] T. Gong, L. Wei, C. Huang, Z. Yang, J. He, M. Debbah, and
(32) C. Yuen, “Holographic MIMO communications with arbitrary
8(r 2 +r 2 )
which leads to d ⩾ s λ m . surface placements: Near-field LoS channel model and capacity
limit,” arXiv preprint arXiv:2304.05259, 2023.
[11] L. Wei, C. Huang, G. C. Alexandropoulos, Z. Yang, J. Yang, E. Wei,
A PPENDIX C Z. Zhang, M. Debbah, and C. Yuen, “Tri-polarized holographic
P ROOF OF L EMMA 3 MIMO surfaces for near-field communications: Channel modeling
and precoding design,” IEEE Trans. Wireless Commun., early
Note that the channel h is a random vector whose access, 2023.
distribution is controlled by its correlation matrix R. For [12] MIMO A. Pizzo, A. Lozano, S. Rangan, and T. L. Marzetta, “Wide-aperture
via reflection off a smooth surface,” IEEE Trans. Wireless
the difference between the estimated channel h̃ and the Commun., vol. 22, no. 8, pp. 5229–5239, Jan. 2023.
true channel h, we have [13] O. Bucci and G. Franceschetti, “On the spatial bandwidth of
h i scattered fields,” IEEE Trans. Antennas Propag., vol. 35, no. 12,
pp. 1445–1455, Dec. 1987.
E (h̃ − h)H (h̃ − h) [14] O. M. Bucci and G. Franceschetti, “On the degrees of freedom of
h√ √ i scattered fields,” IEEE Trans. Antennas Propag., vol. 37, no. 7, pp.
=E ( P R̂(P R̂ + I)−1 y − h)H ( P R̂(P R̂ + I)−1 y − h) 918–926, Jul. 1989.
h √ [15] M. Franceschetti, “On Landau’s eigenvalue theorem and informa-
=E tr ( P R̂(P R̂ + I)−1 y − h) tion cut-sets,” IEEE Trans. Inf. Theory, vol. 61, no. 9, pp. 5042–
√ i 5051, Jul. 2015.
( P R̂(P R̂ + I)−1 y − h)H [16] M. A. Jensen and J. W. Wallace, “Capacity of the continuous-space
electromagnetic channel,” IEEE Trans. Antennas Propag., vol. 56,
√ √ no. 2, pp. 524–531, Feb. 2008.
−1 −1 H
=tr P R̂(P R̂ + I) )(P R + I)( P R̂(P R̂ + I) ) [17] W. Jeon and S.-Y. Chung, “Capacity of continuous-space elec-
√ −1
√ tromagnetic channels with lossy transceivers,” IEEE Trans. Inf.
− P R̂(P R̂ + I) PR Theory, vol. 64, no. 3, pp. 1977–1991, Dec. 2017.
√ √  [18] Z. Wan, J. Zhu, Z. Zhang, L. Dai, and C.-B. Chae, “Mutual
− P R( P R̂(P R̂ + I)−1 )H + R information for electromagnetic information theory based on ran-
 dom fields,” IEEE Trans. Commun., early access, Feb. 2023,
=tr P (P R̂ + I)−1 R̂R̂(P R̂ + I)−1 (P R + I) 10.1109/TCOMM.2023.3247725.
 [19] W. C. Chew, Waves and fields in inhomogenous media. John Wiley
& Sons, 1999, vol. 16.
− 2P R(P I + R̂−1 )−1 + R [20] A. Pizzo, L. Sanguinetti, and T. L. Marzetta, “Fourier plane-wave
 series expansion for holographic MIMO communications,” IEEE
=tr P (P I + R̂−1 )−1 (P I + R̂−1 )−1 (P R + I) Trans. Wireless Commun., vol. 21, no. 9, pp. 6890–6905, Mar. 2022.
 [21] E. Björnson and L. Sanguinetti, “Rayleigh fading modeling and
− 2P R(P I + R̂−1 )−1 + R . channel hardening for reconfigurable intelligent surfaces,” IEEE
Wireless Commun. Lett., vol. 10, no. 4, pp. 830–834, Dec. 2020.
(33) [22] A. Pizzo, T. L. Marzetta, and L. Sanguinetti, “Spatially-stationary
model for holographic MIMO small-scale fading,” IEEE J. Sel.
R EFERENCES Areas Commun., vol. 38, no. 9, pp. 1964–1979, Sep. 2020.
[23] Ö. T. Demir, E. Björnson, and L. Sanguinetti, “Channel modeling
[1] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and channel estimation for holographic massive MIMO with planar
and R. Zhang, “Wireless communications through reconfigurable arrays,” IEEE Wireless Commun. Lett., vol. 11, no. 5, pp. 997–1001,
intelligent surfaces,” IEEE Access, vol. 7, pp. 116 753–116 773, May 2022.
Aug. 2019. [24] Y. Liu, Z. Wang, J. Xu, C. Ouyang, X. Mu, and R. Schober,
[2] Z. Wang, Z. Liu, Y. Shen, A. Conti, and M. Z. Win, “Location “Near-field communications: A tutorial review,” arXiv preprint
awareness in beyond 5G networks via reconfigurable intelligent arXiv:2305.17751, 2023.
surfaces,” IEEE J. Sel. Areas Commun., vol. 40, no. 7, pp. 2011 – [25] F. K. Gruber and E. A. Marengo, “New aspects of electromagnetic
2025, Mar. 2022. information theory for wireless and antenna systems,” IEEE Trans.
[3] C. Huang, S. Hu, G. C. Alexandropoulos, A. Zappone, C. Yuen, Antennas Propag., vol. 56, no. 11, pp. 3470–3484, Nov. 2008.
R. Zhang, M. Di Renzo, and M. Debbah, “Holographic MIMO [26] L. Li, L. G. Wang, F. L. Teixeira, C. Liu, A. Nehorai, and T. J.
surfaces for 6G wireless networks: Opportunities, challenges, and Cui, “DeepNIS: Deep neural network for nonlinear electromagnetic
trends,” IEEE Wireless Commun., vol. 27, no. 5, pp. 118–125, Jul. inverse scattering,” IEEE Trans. Antennas Propag., vol. 67, no. 3,
2020. pp. 1819–1825, Mar. 2018.
[4] Z. Zhang and L. Dai, “Pattern-division multiplexing for multi-user [27] G. Franceschetti and D. Riccio, Scattering, natural surfaces, and
continuous-aperture MIMO,” IEEE J. Sel. Areas Commun., vol. 41, fractals. Elsevier, 2006.
no. 8, pp. 2350–2366, Jun. 2023. [28] 3GPP TR, “Study on channel model for frequencies from 0.5 to
[5] M. Cui, Z. Wu, Y. Lu, X. Wei, and L. Dai, “Near-field MIMO 100 ghz,” 3GPP TR 38.901 version 14.0.0 Release, Dec. 2019.
communications for 6G: Fundamentals, challenges, potentials, and [29] J. Mercer, “Functions of positive and negative type, and their
future directions,” IEEE Comm. Mag., vol. 61, no. 1, pp. 40–46, connection with the theory of integral equations,” Philos. Trans.
Sep. 2022. Roy. Soc. London, vol. 209, no. 441-458, pp. 415–446, Jan. 1909.
[6] Z. Wu and L. Dai, “Multiple access for near-field communications: [30] K. T. Selvan and R. Janaswamy, “Fraunhofer and Fresnel distances:
SDMA or LDMA?” IEEE J. Sel. Areas Commun., vol. 41, no. 6, Unified derivation for aperture antennas.” IEEE Antennas and
pp. 1918–1935, May 2023. Propag. Mag., vol. 59, no. 4, pp. 12–15, Aug. 2017.
 12

[31] J. Wang, S. Kwon, and B. Shim, “Generalized orthogonal matching

pursuit,” IEEE Trans. Signal Process., vol. 60, no. 12, pp. 6202–
6216, Dec. 2012.
[32] Z. Wu and L. Dai, “Multiple access for near-field communications:
SDMA or LDMA?” IEEE J. Sel. Areas Commun., vol. 41, no. 6,
pp. 1918–1935, May 2023.
[33] D. Zwillinger, V. Moll, I. Gradshteyn, and I. Ryzhik, Eds., Table
of Integrals, Series, and Products (Eighth Edition). Boston:
Academic Press, 2014.