IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO.
5, MAY 2021
Reconfigurable ULAs for Line-of-Sight
MIMO Transmission
Heedong Do, Graduate Student Member, IEEE, Namyoon Lee , Senior Member, IEEE,
and Angel Lozano , Fellow, IEEE
Abstract— This paper establishes an upper bound on the
capacity of line-of-sight multiantenna channels over all pos-
sible antenna arrangements and shows that uniform linear
arrays (ULAs) with an SNR-dependent rotation of transmitter
and/or receiver can closely approach such capacity—and in fact
achieve it at low and high SNR, and asymptotically in the
numbers of antennas. Then, as an alternative to mechanically
rotating ULAs, we propose to electronically select among multiple
ULAs having a radial disposition at either transmitter or receiver,
and we bound the shortfall from capacity as a function of
the number of such ULAs. With only three ULAs, properly
angled, 96% of the capacity can be achieved. Finally, we further
introduce reduced-complexity precoders and linear receivers that
capitalize on the structure of the channels spawned by these
configurable ULA architectures.
Index Terms— Line-of-sight transmission, MIMO, multi-
antenna channels, reconfigurable arrays, mmWave frequencies,
terahertz frequencies.
I. I NTRODUCTION
A N UNRELENTING trend in the evolution of wireless
communication is the move to ever higher frequencies,
so as to gain access to ever wider bandwidths. The current
frontier is at about 90 GHz, but researchers already have
their eyes set on sub-terahertz bands where new applica-
tions await, including kiosk information transfers [2], wire-
less backhaul [3]–[5], and wireless interconnections within
datacenters [6]. Another consolidated feature of wireless sys-
tems are multiple-input multiple-output (MIMO) techniques,
Manuscript received April 24, 2020; revised August 12, 2020 and
October 19, 2020; accepted December 7, 2020. Date of publication Decem-
ber 24, 2020; date of current version May 10, 2021. The work of Hee-
dong Do and Namyoon Lee was supported in part by the Samsung
Research Funding and Incubation Center of Samsung Electronics under
Project SRFC-IT1702-04, in part by the National Research Foundation of
Korea (NRF) Grant funded by the Korea Government (MSIT) under Grant
2020R1C1C1013381, and in part by the BK21Plus Program. The work
of Angel Lozano was supported in part by the European Research Coun-
cil through the H2020 Framework Programme/ERC under Grant 694974,
in part by the MINECO’s under Project RTI2018-102112 and Project
RTI2018-101040, and in part by the ICREA. This article was presented in part
at the 2020 IEEE International Symposium on Information Theory (ISIT’20).
The associate editor coordinating the review of this article and approving it
for publication was C. Shen. (Corresponding author: Namyoon Lee.)
Heedong Do and Namyoon Lee are with the Department of Electri-
cal Engineering, POSTECH, Pohang 37673, South Korea (e-mail: dohee-
dong@postech.ac.kr; nylee@postech.ac.kr).
Angel Lozano is with the Department of Information and Communication
Technologies, Pompeu Fabra University, 08018 Barcelona, Spain (e-mail:
angel.lozano@upf.edu).
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TWC.2020.3045588.
Digital Object Identifier 10.1109/TWC.2020.3045588
1536-1276 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2933
which bring about major improvements in spectral and energy
efficiency [7].
At microwave frequencies, MIMO enables spatial multi-
plexing by virtue of multipath propagation, whereby the envi-
ronment acts as a lens that delivers a high-rank channel [8].
As we move up in frequency, into the mmWave realm and then
into sub-terahertz territory [9]–[11], the transmission range
necessarily shrinks and the propagation becomes mostly line-
of-sight (LOS). The multipath lensing effect dwindles. At the
same time, because the wavelength also shrinks dramatically,
it becomes progressively possible to span a high-rank chan-
nel based only on the array apertures themselves [12]–[14].
In particular, parallel uniform linear arrays (ULAs) can give
rise to a channel with all-equal singular values, ideal for
spatial multiplexing [15], provided the antenna spacing is
d = λD/N where λ is the wavelength, D the trans-
mission range, and N the number of antennas at each end.
With this so-called Rayleigh spacing within the ULAs, direc-
tional signals can be launched and then resolved at the
receiver without cross-talk. And, at sub-terahertz frequencies,
Rayleigh spacing become feasible within reasonably com-
pact arrays: at 300 GHz, for instance, a 16-antenna ULA
designed for D = 5 m would occupy 26.52 cm, and far
less if arranged as a two-dimensional uniform rectangular
array (URA).
With spatial multiplexing as an objective, the antenna
spacings that yield a channel with all-equal singular values
have been determined, not only for ULAs as detailed above,
but for a variety of regular array geometries [8], [16]–[23].
Also, a genetic algorithm has been employed to identify
suitable nonregular geometries [24]. Moreover, the efficacy
of LOS spatial multiplexing has been demonstrated experi-
mentally, for now with up to N = 4 antennas at mmWave
frequencies [25]–[32]. Spatial multiplexing, however, is the
desirable transmission strategy only at high SNR. At low SNR,
maximizing the received power is of essence [7, ch. 5], and
that demands beamforming over a channel whose maximum
singular value is as large as possible, rather than having
all-equal singular values [33], [34]. This contrast suggests that
the ULA antenna spacings, and more generally the antenna
arrangements, should depend on the SNR so as to strike a bal-
ance between spatial multiplexing and beamforming [35]. As a
step in this direction, [33] proposes switching, as a function of
the SNR, among three ULAs with distinct antenna spacings.
Also recognizing that both spatial multiplexing and beam-
forming are relevant ingredients, other works such as [15],
2934 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

[36]–[38] propound the use of arrays-of-subarrays, which mix placements that respect the condition of apertures much
small and large antenna spacings and, advantageously, require smaller than D.
a reduced number of RF chains [39]. Let us define, as the unprecoded signal-to-noise ratio per
The present paper shows how a single ULA can be recon- receive antenna,
figured, through simple rotation, to closely approach the LOS λ2 Gt Gr Pt
capacity at any desired SNR. Precisely: SNR = (4)
(4πD)2 BN0
• An information-theoretic footing is established in the
form of an upper bound on the LOS capacity over all where Pt is the total radiated power and BN0 the noise
possible antenna arrangements. power. For a range D and specific parameters (wavelength,
• The ULA antenna spacings that are optimum as a function antenna gains, powers), the SNR becomes determined. The
of the SNR are determined, and it is shown that ULAs information-theoretic capacity of a channel H is then [40]
with such spacings approach the LOS capacity within a min −1
N

gap that vanishes as the number of antennas increases, C(H, SNR) = max log2 1 + pn σn2 (H)
Nmin −1
and also at low and high SNR. n=0 pn =SNR n=0
pn ≥0
• An architecture is proposed in which ULAs can be  + 
min −1
N
configured without changing their antenna spacing, rather 1 1
= log2 1+ − 2 σn2 (H)
through a mere rotation of transmitter and/or receiver,
n=0
γ σn (H)
or else by electronically selecting among various ULAs
(5)
in a radial disposition. Nmin −1
• For such configurable architectures, low-complexity pre- with γ such that n=0 pn = SNR and [z]+ = min(0, z).
coders and linear receivers are also put forth. Achieving C(H, SNR) requires a precoder aligned with the
The paper is organized as follows. Section II introduces right singular vectors of H and whose powers along those
the LOS channel model and its capacity. Then, Section III directions, p0 , . . . , pNmin −1 , are dictated by waterfilling on
specializes the channel model to parallel and non-parallel 1/σ02 , . . . , 1/σN
2
min −1
; the linear receiver, in turn, must be
ULAs. The upper bound on the LOS capacity is presented in aligned with the left singular vectors of H. Altogether,
Section IV, proved in an appendix. Section V sets the stage for the precoder and receiver depend on H, but the capacity
the configurable architectures presented in Section VI. Finally, itself rests only on its singular values. There are up to Nmin
the low-complexity precoders and receivers are described in spatial degrees of freedom (DOF) over which to communicate,
Section VII and the paper concludes in Section VIII, with with the number of active DOFs determined by the number
further proofs relegated to subsequent appendices. of nonzero powers within p0 , . . . , pNmin−1 . Beamforming in
particular corresponds to a single nonzero power.
II. C HANNEL M ODEL AND C APACITY From C(H, SNR), the problem of establishing the LOS
Consider Nt transmit and Nr receive antennas connected capacity broadens to that of identifying the antenna placements
by an LOS channel. The far-field complex baseband channel that yield the channel H ∈ H whose capacity is largest, i.e.,
coefficient from the mth transmit to the nth receive antenna,
separated by a distance Dn,m , is C(SNR) = max C(H, SNR). (6)
H∈H
√
Gt Gr λ −j 2π Dn,m
hn,m = e λ
4π Dn,m
III. A RRAY S TRUCTURES
n = 0, . . . , Nr − 1 m = 0, . . . , Nt − 1 (1)
This section presents compact expressions for the LOS
where Gt and Gr are the respective antenna gains. Provided channels spawned by different ULA configurations.
the antenna apertures are small relative to Dn,m , the magni-
tude |hn,m | is approximately constant across m and n while A. Parallel ULAs
Dn,m ≈ D such that only the phase variations need to be
modeled. These are captured by the normalized matrix For parallel transmit and receive ULAs with respective
⎡ −j 2π D ⎤ antenna spacings dt and dr , basic trigonometry leads to [22]
e λ 0,0 ···
2π
e−j λ D0,Nt −1 2π
√ 2 2
⎢ .. .. .. ⎥ hn,m = e−j λ D +(ndr −mdt ) , (7)
H=⎣ . . . ⎦, (2)
2π 2π
e−j λ DNr −1,0 ··· e−j λ DNr −1,Nt −1 which, under our proviso that the antenna apertures are small
relative to D, satisfies [18], [20], [41]
which, letting σn (·) denote the nth singular value of a matrix, D n2 2 nm m2 2
satisfies hn,m ≈ e−j2π λ e−jπλD dr ej2π λD dr dt e−jπλD dt . (8)
min −1
N RX phase shifts TX phase shifts
σn2 (H) = Nr Nt . (3) The constant phase shift in the leading term does not affect
n=0 the singular values, and neither do the phase shifts across the
For the sake of compactness, we define Nmin = min(Nt , Nr ) transmit and receive arrays, which amount to diagonal unitary
and Nmax = max(Nt , Nr ). We further define H as the set matrices pre-and post-multiplying the central term. (Besides
of normalized matrices H generated by all possible antenna not affecting the singular values, all these shifts can be easily

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION 2935

are required, for instance (see Fig. 1) the elevation angles of

the transmit and receive arrays, θt and θr , and the relative
azimuth angle between them, φr . Parameterized by these
angles, hn,m is given by (11), as shown at the bottom of the
page, and, invoking again the premise that the apertures are
small relative to D [20],
D −jπ
 2n n
sin θr cos φr + λD
2
d2r (1−sin2 θr cos2 φr )

hn,m ≈ e−j2π λ e λ dr

nm −jπ
 2m
sin θt + m
2 2

· ej2π λD dr dt cos θr cos θt e λ dt λD dt
. (12)
The fixed phase and the phase shifts affecting only either the
transmitter or the receiver are again immaterial to the singular
values, hence the relevant term is the phase
nm
2π dr dt cos θr cos θt . (13)
λD
The ensuing normalized channel matrix is identical to H ULA
in (9), only with
(dr cos θr )(dt cos θt )Nmax
η= . (14)
λD
Under this more general definition of η, then, any
ULA-induced channel can be represented by H ULA in (9)
as far as the singular values are concerned. Interestingly,
the relative azimuth orientations represented by φr play no
role in the singular values, and therefore in the capacity. This
is capitalized on later in the paper.
Fig. 1. Non-parallel ULAs. We hasten to emphasize that, while the analysis throughout
the paper relies on the approximation in (12) and the ensuing
definition of η, all the results are generated with channel
compensated for.) As far as the singular values are concerned, matrices whose entries correspond to (11).
the channel can be reduced to the central term in (8), which
gives rise to the Vandermonde matrix IV. C APACITY U PPER B OUND
⎡ 0×0 (Nt −1)×0
⎤
We commence by establishing an upper bound on the LOS
ej2πη Nmax ··· ej2πη Nmax
⎢ ⎥ capacity over all possible channels belonging to the set H,
H ULA = ⎢ ⎣
..
.
..
.
..
.
⎥ (9)
⎦ a technical result that, besides serving as a benchmark in the
0×(Nr −1) (Nt −1)×(Nr −1)
e j2πη Nmax
··· e j2πη N max sequel, has intrinsic relevance. As detailed in Appendix A,
such upper bound corresponds to a channel having ρ identical
where we have introduced nonzero singular values and Nmin −ρ zero singular values with
dr dt Nmax ρ depending on the SNR via
η= (10)
λD ⎧
⎪
⎪ 1 SNR < ζ1
as a parameter that compactly describes the parallel ULA ⎪
⎪
⎪
⎪ 2 ζ1 ≤ SNR < ζ2
configuration. Rayleigh antenna spacings correspond to η = 1, ⎨
whereby the matrix becomes column-orthogonal if Nt ≤ Nr ρ(SNR) = 3 ζ2 ≤ SNR < ζ3 (15)
⎪
⎪ .
and row-orthogonal if Nt ≥ Nr . The singular values are then ⎪ ..
⎪
⎪
⎪
all identical and the singular vectors are Fourier basis vectors, ⎩
making for easy-to-compute precoders and receivers [42]. Nmin ζNmin −1 ≤ SNR
where ζn is a threshold equal to the unique positive solution
B. Non-Parallel ULAs of
   
In order to determine in complete generality the relative Nr Nt Nr Nt
f ζn = f ζn (16)
position of two non-parallel ULAs, three geometric parameters n2 (n + 1)2

⎛  ⎞
 2  2  2

⎜ 2π ⎟
hn,m = exp ⎝−j  D + ndr sin θr cos φr − mdt sin θt + ndr cos θr −mdt cos θt + ndr sin θr sin φr ⎠
λ         
z-axis x-axis y-axis
(11)

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2936 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

given the function

1
f (x) = √ log2 (1 + x). (17)
x
We then have that
 
Nr Nt
C(SNR) ≤ ρ(SNR) log2 1 + SNR (18)
ρ(SNR)2
whose right-hand side is the capacity of such channel with ρ
identical nonzero singular values. Referring back to (5), this
corresponds to ρ active DOF and, from (3), σn2 = Nt Nr /ρ for
n = 0, . . . , ρ − 1; then, pn = SNR/ρ for n = 0, . . . , ρ − 1.
By relaxing the integer ρ into a real-valued counterpart ρ̃,
a (very slightly) looser upper bound can be obtained in explicit
form as detailed again in Appendix A. Precisely,
 
Nr Nt
C(SNR) ≤ ρ̃(SNR) log2 1 + SNR (19)
ρ̃(SNR)2
where
⎧ c
⎪
⎪ 1 SNR <
⎪
⎪
⎨$
⎪ Nmin Nmax
SNR c Nmin c  √
ρ̃(SNR) = Nmin Nmax ≤ SNR < Fig. 2. Spectral efficiencies of ULAs with η = 0, 1/ N , 1 for
⎪
⎪ c N min N max Nmax Nt = Nr = N = 256, D = 10 m, and λ = 1 mm. Also shown are
⎪
⎪
⎪
⎩ Nmin N min c
≤ SNR
the corresponding capacity upper bound and its relaxed counterpart.
Nmax
(20)
with
2
c = −1 − ≈ 3.92, (21)
W0 (−2/e2 )
given W0 (z) as the principal branch of a Lambert W function,
i.e., the solution to W0 eW0 = z. Within the precision of this
slightly relaxed bound, then, the transition away from pure
beamforming (ρ̃ = 1) takes place when Nmin Nmax SNR, which
is the effective SNR including the beamforming gains, equals
precisely c. Combining (19) and (20), we can express the
relaxed upper bound in the more compact form
⎧ c
⎪
⎪ log2 (1 + Nr Nt SNR) SNR <
⎪$
⎪ Nt Nr
⎪
⎨ SNR c Nmin c
C≤ Nt Nr log2 (1 + c) ≤ SNR <
⎪
⎪ c
  N t N r Nmax
⎪
⎪
⎪
⎩ Nmin log2 1 +
Nmax
SNR
Nmin c
≤ SNR.
Nmin Nmax
(22)
A comparison between the upper bounds in (18) and (22)
is provided in Fig. 2 for Nt = Nr = 256, evidencing their Fig. 3. Spectral efficiencies of ULAs with η ∈ [0, 1] for Nr = Nt = 256,
tightness. Only for very small numbers of antennas do the D = 10 m, and λ = 1 mm. Also shown is the corresponding capacity upper
two bounds become distinguishable. bound.

V. O PTIMUM C ONFIGURATION FOR PARALLEL ULA S
Parallel ULAs adopting three SNR-based configurations√are
proposed in [33], namely η = 0 for low SNRs, η = 1/ N
for medium SNRs, and η = 1 for high SNRs. Shown in Fig. 2
are the spectral efficiencies achieved by these configurations
for Nt = Nr = N = 256, computed via (5), alongside the
capacity upper bounds. The approach is seen to be effec-
tive at very low and at high SNR, less so at intermediate
SNRs. At SNR = −5 dB, for instance, the spectral efficiency
reaches only about 55% of the capacity.
By releasing η and allowing it to take any value within
[0, 1], the upper bound can be hugged much more closely
(see Fig. 3). This involves fine-tuning the antenna spacings
depending on the SNR, computing the channel’s singular-value
decomposition (SVD) to obtain the precoding directions and
the receiver, and solving (5) for the transmit powers.

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION 2937

To interpret the effectiveness of ULAs with antenna spac- nonzero singular value. Such low-SNR capacity, achieved by
ings adapted to the SNR, let us examine the (n, m)th entry of beamforming, expands as
H ∗ULA H ULA , namely
C(SNR) = log2 (1 + Nmin Nmax SNR) (29)
% &
N r −1
(n−m)
= Nmin Nmax SNR log2 e + O(SNR2 ). (30)
H ∗ULA H ULA = e−j2πη Nmax (23)
n,m

=0
 In contrast, Rayleigh-spaced ULAs would feature ρ(SNR) =
sin πη (n−m)N
Nmax
r
(n−m)(Nr −1)
Nmin and
=
 e−jπη Nmax ,  
sin πη Nmax
n−m Nmax
C(H ULA , SNR) = Nmin log2 1 + SNR (31)
Nmin
(24)
= Nmax SNR log2 e + O(SNR2 ). (32)
and let us denote by λ0 , . . . , λNmin −1 the Nmin largest eigen-
values of Nηmax H ∗ULA H ULA . For any  > 0, as both Nt and Nr With tight ULAs, an Nmin -fold improvement is obtained in
grow large with some ratio Nt /Nr > 0, these eigenvalues can low-SNR capacity relative to Rayleigh-spaced ULAs designed
be shown to satisfy [43], [44] for high-SNR operation.
'( )'
'  | λ <  '
lim =1−η (25) B. High-SNR Regime
Nmin →∞ Nmin
For SNR ≥ ζNmin −1 , the upper bound reduces to the capacity
and
'( )' of Rayleigh-spaced ULAs, which in this regime become
'  | 1 −  < λ < 1 +  '
lim =η (26)
Nmin →∞ Nmin
where |{·}| indicates the cardinality of a set. As observed
in [15], [41], then, these eigenvalues polarize (asymptotically
in the numbers of antennas) into the states 1 and  0, and
therefore the singular values of H ULA polarize into Nmax /η
and 0. It follows from (3)that (asymptotically) we have ηNmin
singular values equal to Nmax /η and (1 − η)Nmin singular
values equal to 0. By adjusting the antenna spacings such that
ρ(SNR)
η= , (27)
Nmin
we (asymptotically) obtain a channel matrix featuring ρ(SNR)
identical nonzero singular values and (Nmin − ρ) zero singular
values. This is the precise disposition that yields the capacity
upper bound, with ρ(SNR) as in (15).
Making matters precise, it is shown in Appendix B that
the capacity of H ULA with η properly adjusted converges
pointwise to the LOS capacity, i.e.,
maxη∈[0,1] C(H ULA , SNR)
lim =1 (28)
Nmin →∞ C(SNR)
at every SNR.
For finite numbers of antennas, the polarization of the sin-
gular values is not complete and thus the LOS capacity cannot
be strictly attained by ULAs in general, but, as illustrated
in Fig. 3, it is approached very closely for values of interest
(256 in the figure). And, as documented in Table I, it is
approached closely with as few as 32 antennas. Moreover,
there are specific regimes where ULAs can achieve the LOS
capacity regardless of the numbers of antennas, as seen next.
A. Low-SNR Regime
As the SNR drops, ρ(SNR) shrinks and, for SNR ≤
ζ1 = 8/Nr Nt , it floors at ρ(SNR) = 1. Then, ULAs with
tightly spaced antennas become capacity-achieving regardless
of Nt and Nr as the channel they create exhibits a single
Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
optimum irrespective of Nt and Nr . Such high-SNR capacity
expands as
 
Nmax
C(SNR) = Nmin log2 1 + SNR (33)
Nmin
= Nmin log2 SNR + O(1). (34)
VI. R ECONFIGURABLE ULA S
Adapting η to the SNR by means of adjusting the antenna
spacings in parallel ULAs is implementationally very chal-
lenging because separate moving parts would be required
for each individual antenna. This justifies the proposition of
having several ULAs with distinct but fixed spacings, and
the possibility of switching among them, with a performance
shortfall that depends on the number thereof [33].
This section presents an alternative method to tune η as a
function of the SNR with a single fixed-spacing ULA at each
end of the link.
A. Adapting η via ULA Rotation
To operate with fixed antenna spacings, we seek to adapt η
by modifying the relative orientation of the ULAs depending
on the SNR. This can be realized by rotating one of the two
ULAs, either transmitter or receiver, while keeping the other
one fixed. Since, as seen in Section III-B, the relative azimuth
angle is immaterial to the channel singular values, we set it
to φr = π/2 (preferable from the standpoint of the directivity
of the individual antennas). For starters, we further set θt = 0
and consider rotating only the receiving ULA by θr . Then,
(14) reduces to
(dr cos θr ) dt Nmax
η= (35)
λD
and, by fixing the antenna spacings at the Rayleigh values,
further to
η = cos θr , (36)
which indeed can be controlled by merely varying θr .
2938 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

Selecting the adequately angled ULA as a function of the

SNR (see Fig. 5), the performance of a rotatable ULA can
be matched. The number of RF chains continues to be Nr ,
but Nmin (Nr − 1) + 1 antennas are now required—the central
antenna is common to all the radially arranged ULAs—
as opposed to Nr in the case of a single rotatable ULA.
To keep the number of additional antennas to a minimum,
the radial architecture should be deployed at the end of the link
(transmitter or receiver) with the smallest number of antennas.
Now, a natural next step, especially with a view to hav-
ing large arrays, is to reduce the number of radial ULAs
below Nmin . Suppose that the values of η are taken from
the geometric series 1, r, r2 , . . . where the ratio r < 1 is
a design parameter that determines the trade-off between
number of ULAs and performance. To span the entire range
η ∈ [1/Nmin , 1], the number of ULAs required for a given r is
* +
log Nmin
k =1+ (40)
log 1/r

where · rounds down to the closest integer. We know that, for
Fig. 4. ULA spectral efficiency as a function of η ∈ [0, 1] for Nt = Nr =  exhibits (asymptotically) ηNmin singular
a given η, the channel
N = 32 and 256 at SNR = 0 dB with D = 10 m and λ = 1 mm. Also values equal to Nmax /η for a spectral efficiency of
shown, in dashed, are the respective capacity upper bounds.
 
Nmax
ηNmin log2 1 + 2 SNR (41)
Recalling (27) and (28), we can infer that capacity can η Nmin
be achieved (asymptotically in the numbers of antennas) by
setting φr = π/2, θt = 0, dt dr = λD/Nmax , and meaning that, with k radial ULAs, we can hope for a spectral
efficiency of
ρ(SNR)
θr = arccos . (37)  
Nmin Nmax
R(SNR) ≈ max ηNmin log2 1 + 2 SNR
Since ρ(SNR) ∈ [1, Nmin ], every SNR maps to a feasible θr . η∈{1,r,...,r k−1 } η Nmin
And θr need only be modified if the SNR varies, which for (42)
the envisioned applications occurs on a slow time scale or
possibly not at all. where the approximation sharpens with the numbers of anten-
Presented in Fig. 4 are two examples of spectral efficiency nas and the maximization determines the best possible ULA
as a function of η ∈ [0, 1], corresponding to θr ∈ [90◦ , 0◦ ], at each SNR. If we make η dependent on the SNR via
at SNR = 0 dB. The optimum configuration is η = 0.504, cor- ⎧
responding to θr = 59.7◦ , at which point capacity is essentially ⎪
⎪ rk−1 SNR ≤
Nmin 2k−3
cr
⎪
⎪
achieved with both 32 and 256 antennas. The performance is ⎪
⎪
Nmax
⎪
⎪ .. ..
seen to be robust to small angular misadjustments. ⎪
⎪ . .
⎪
⎪
If the transmit elevation angle is not set to θt = 0, but rather ⎨ 2 Nmin 5 Nmin 3
to some arbitrary value, then (37) generalizes to η= r c r < SNR ≤ cr (43)
⎪ Nmax Nmax
⎪
⎪
ρ(SNR) ⎪
⎪ Nmin 3 Nmin
θr = arccos . (38) ⎪
⎪ r c r < SNR ≤ cr
Nmin cos θt ⎪
⎪ Nmax Nmax
⎪
⎪
⎪
⎩1 Nmin
c r < SNR
Nmax
B. Radial ULAs With Electronic Selection
where, recall, c ≈ 3.92, then (see Appendix C) the achievable
Since ρ(SNR) takes on integer values, namely 1, . . . , Nmin , spectral efficiency satisfies
the angles required as per (37) to reconfigure a rotatable
ULA for every possible SNR are discrete. This suggests that, R(SNR) log(1 + c r)
as an alternative to the SNR-based mechanical rotation of ≥ √ (44)
C(SNR) r log(1 + c)
the receiver, complete reconfigurability is also possible by
deploying at the receiver Nmin radial ULAs angled at at every SNR. As r approaches unity, the capacity is
n approached ever more closely, albeit at the expense of a larger
θr,n = arccos n = 1, . . . , Nmin . (39)
Nmin number of radial ULAs as stipulated by (40).

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION 2939

Fig. 5. Reduced-complexity configurable architecture.

If it is known that SNR ≥ SNRmin with SNRmin > 0, we situation at high enough SNR, when η = 1 is optimum. More
can truncate (43) into generally, though, η < 1; then, obtaining the precoder and
⎧ receiver requires subjecting
⎪ Nmin 2k−1 Nmin 2k−3 . 2 the channel
/ to an SVD, with a
⎪
⎪ rk−1 cr < SNR ≤ cr computational cost of O Nmin Nmax . This is not a limitation
⎪
⎪ N max N max
⎪
⎪ .. .. in applications where the precoder and receiver recomputation
⎪
⎪
⎪
⎪ . . is sporadic, but in more dynamic settings, especially with
⎪
⎨ 2 Nmin 5 Nmin 3
c r < SNR ≤ large arrays, less complex alternatives might be welcome. In
η= r N N
cr (45)
⎪
⎪ max max this section, we present one such alternative that very closely
⎪
⎪ Nmin 3 Nmin
⎪
⎪ r c r < SNR ≤ cr approaches the LOS capacity with a computational cost of
⎪
⎪ N N O(Nmax log Nmax ).
⎪
⎪ max max
⎪
⎪ Nmin
⎩1 c r < SNR While, for the purpose of capacity calculations, only the
Nmax singular values of the channel were relevant hitherto, to devise
where k now satisfies the precoder and the receiver the singular vectors become
Nmin 2k−1 Nmin 2k−3 equally relevant. Then, (9) no longer suffices to represent
cr < SNRmin ≤ cr , (46) nonparallel transmit and receive ULAs, but rather we need
Nmax Nmax
to generalize H ULA to have its (n, m)th entry be given
ensuring that the range [SNRmin , ∞) is covered. Equivalently, by (12). For this more general form of H ULA , it is shown
, - in Appendix D that
log Nmax
Nmin c
SNRmin 3
k= + . (47)
2 log 1/r 2 Nmax . /
F ∗ D∗tx H ∗ULA H ULA Dtx F ≈ diag 1, . . . , 1, 0, . . . , 0
This truncation markedly reduces the number of necessary η   
ηNmin
ULAs, eliminating those that would correspond to SNRs below
(48)
the operating range of interest. For instance, with k = 3 radial
ULAs, Nt = Nr = N , and SNRmin = −10 dB, the largest
possible ratio r is, applying (47), r = 0.48. Plugging this value where Dtx is a diagonal matrix with entries
into (44), we infer that at least 95.9% of the LOS capacity
jπ
 2m 2
sin θt + m 2

can be achieved (for large N ) with only three ULAs angled λ dt λD dt
[Dtx ]m,m = e (49)
at θr,0 = 0, θr,1 = 61◦ , and θr,2 = 77◦ .
while F is a Fourier matrix; the approximation in (48) tightens
VII. R EDUCED -C OMPLEXITY A RCHITECTURE as the dimensionality grows large.
For η = 1, meaning with Rayleigh antenna spacings at the Asymptotically then, D ∗tx H ∗ULA H ULA D tx is diagonalized
ULAs, the precoder and the receiver are straightforward to by a Fourier matrix. This is illustrated in Fig. 6, which depicts
compute as the channel adopts a Fourier structure. This is the the power concentrated on the ensuing diagonal entries relative

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2940 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

where Drx is a diagonal matrix with entries

 2n n2
sin θr cos φr + λD d2r (1−sin2 θr cos2 φr )

j2π D jπ λ dr
[D rx ]n,n = e λ e .
(54)

The application of F ∗ and Drx to a vector require com-

plexities of O(Nt log Nt ) and O(Nr ), respectively. In turn,
(D rx H ULA D tx )∗ has entries
nm
[(Drx H ULA Dtx )∗ ]n,m = e−j2πη Nmax (55)

and, from the relationship

nm n2 (n−m)2 m2
e−j2πη Nmax = e−jπη Nmax ejπη Nmax e−jπη Nmax , (56)

it follows that (D rx H ULA Dtx )∗ is a Toeplitz matrix

Fig. 6. Diagonal power of F ∗ D∗tx H∗ULA HULA Dtx F normalized by its
total power for various number of antennas (Nt = Nr = N ).
to the total power, i.e.,
Nt −1 ∗ ∗ ∗ 2 Table I compares the performance of the low-complexity
n=0 [F D tx H ULA H ULA D tx F ]n,n
∗ ∗ ∗ .
F D tx H ULA H ULA Dtx F
2F
Notice how, except for very small η, the ratio in (50) is indeed
very close to unity, increasingly so as N does from 32 to 64.
(This asymptotic trend is then disrupted for larger N , as the
small-aperture approximation underpinning the formulation
becomes loose.)
Let us next see how to take advantage of (48). Inserting
between the precoder and the antennas a bank of phase
shifts corresponding to the diagonal entries of D tx , a Fourier
precoder F yields at the receiver
y = H ULA Dtx F x + v.
With straight maximum ratio combining (MRC) at the receiver,
(H ULA Dtx F )∗ y = F ∗ D ∗tx H ∗ULA H ULA D tx F x + ṽ
where the noise at the output of the MRC is
H ∗ULA H ULA
2F = λ2n (57)
. /
ṽ = (H ULA Dtx F )∗ v ∼ NC 0, F ∗ D∗tx H ∗ULA H ULA D tx F .
From (48), this transmit-receive architecture (asymptotically)
decomposes the channel into ηNmin parallel subchannels,
each with power gain (Nmax /η)2 and noise variance Nmax /η.
By transmitting ηNmin equal-power signal streams over these
subchannels and separately decoding them, the spectral effi-
ciency (asymptotically) equals (41).
Applying the MRC matrix (H ULA D tx F )∗ to the vector y
entails a complexity of O(Nr Nt ) = O(Nmin Nmax ), already
2
markedly lower than the O(Nmin Nmax ) that an SVD necessi-
tates. A further reduction can be attained by rearranging the
receiver into [42]
(H ULA Dtx F )∗ = F ∗ (D rx H ULA Dtx )∗ D rx
pre- and post-multiplied by diagonal matrices. Therefore,
the complexity of applying (D rx H ULA D tx )∗ to a vector
is O(Nmax log Nmax ) [45]. Altogether, with a precoding
complexity of O(Nt log Nt ) and a receiver complexity of
O(Nmax log Nmax ), the performance equals (asymptotically)
that of an SVD-based architecture.
Fig. 7 illustrates how the reduced-complexity architecture
consisting of a Fourier precoder and a bank of phase shifts at
the transmitter, and an MRC at the receiver, can tightly track
the LOS capacity upper bound. Complementing the figure,
(50) architecture against those of its SVD-based counterpart and
of parallel Rayleigh-spaced ULAs, for various SNRs and
numbers of antennas. For SNR = −10 dB and Nr =
Nt = 256, for instance, parallel ULAs attain only 37.5%
of the capacity upper bound whereas, by applying a rotation
of θr = 80.8◦ , that share increases to 99.5% (with the
SVD-based architecture) or 96.4% (with the lower-complexity
architecture).
A. Comparison With Other Architectures
(51) The presence of the Fourier precoder is very consequential
in the proposed reduced-complexity architecture. Without it,
an MRC receiver would perform poorly because, from (26),
(52)
min −1
N
n=0
 2
Nmax
≈ ηNmin (58)
η
2
=
Nmin Nmax
(59)
η
while the power on the diagonal entries adds up to
min −1
N
[H ∗ULA H ULA ]2n,n = Nmin Nmax
2
(60)
n=0
regardless of η. Since η ≤ 1, there is more power on the
off-diagonal entries (which represent interference) than on the
(53) diagonal entries (which represent intended signals).

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION
Fig. 7. Spectral efficiencies for η ∈ [0, 1] achievable by the proposed
reduced-complexity architecture with Nt = Nr = 128, D = 10 m, and
λ = 1 mm. Also shown, in dashed, is the capacity upper bound.
TABLE I
C ONFIGURATION PARAMETERS AND S PECTRAL E FFICIENCY ( AS A S HARE
OF THE LOS C APACITY U PPER BOUND ) W ITH R AYLEIGH -S PACED
ULAs AND Nr = Nt = N
2941
The small-aperture approximation underpinning the deriva-
tions turns out to be highly robust; some of the examples
in the paper push its boundaries, yet the performance is
wholly consistent with the theory developed on the basis of
that approximation. This is welcome news for short-range
applications, with the reduced-complexity architecture as the
only result that appears sensitive to the limits imposed by
this approximation. The robustness extends to the individual
antenna patterns being directive, say as per 3GPP recommen-
dations; in Fig. 4, for instance, 99.8% of the capacity upper
bound can be attained with omnidirectional antennas, while
with the antennas specified in [46, Table 7.3-1] that would
diminish only slightly to 98.7%.
Potential follow-up research directions include the extension
to a variety of antenna configurations such as URAs [47]
or uniform circular arrays (UCAs); both offer superior form
factors, and UCAs provide a connection with orbital angu-
lar momentum techniques [48], [49]. Another worthwhile
direction is the generalization to channels exhibiting some
degree of multipath propagation [50]. Reducing the sensi-
tivity to angular misadjustments is yet another avenue for
further work, which could perhaps benefit from designs of
nonuniform arrays intended to increase the resilience to range
variations [51], [52].
A PPENDIX A
A proof of (18) can be found in [1]. Here, we provide an
alternative proof that relies on the following intuitive result.
Lemma 1: Consider N real numbers x1 , . . . , xN . If we
successively replace two of them with their average, it is
possible to make all of them to be equal in the limit. Precisely,
letting t be an iteration counter, there exists a sequence
satisfying
 N 

 N
(t) (t) t→∞ i=1 xi i=1 xi
x1 , . . . , xN → ,..., . (61)
N N
Proof: Averaging the largest and smallest values does the
trick. Since the sum of N numbers is invariant to averaging,
VIII. C ONCLUSION we can let x1/+ · · · + xN.= 0/without loss of generality. Now,
. (t) 2 (t) 2
This paper has shown that, through an SNR-dependent rota- observe x1 + · · · + xN at iteration t. After one more
tion, ULAs with Rayleigh antenna spacings can be configured iteration, this sum decreases by
to closely approach the LOS channel capacity. Such capacity  

2
2 (t) (t) 2
is actually attained asymptotically in the numbers of antennas, (t) (t) max xi + min xi
max xi + min xi −2
and at low/high SNR. The same performance can be achieved, 2
avoiding the need for mechanical rotations, by selecting among
2
(t) (t)
Nmin radially disposed ULAs. As the number of radial ULAs max xi − min xi
shrinks, the performance declines very gradually such that, = . (62)
2
with only a few properly oriented ULAs, the vast majority of
(t) (t)
the capacity is still within reach. From the zero sum constraint, max xi ≥ 0 ≥ min xi , and
In comparison with structures consisting of multiple ULAs thus
featuring distinct antenna spacings, the proposed architecture
2
' (t) '2
max 'xi '
(t) (t)
is more compact and better performing. A structure with three max xi − min xi
distinct-spacing ULAs, for instance, performs very well at ≥
2 2
some SNRs, but drops down to only 55% of capacity at others.
2
2
(t) (t)
In contrast, the proposed architecture ensures at least 96% of x1 + · · · + xN
≥ . (63)
the LOS capacity at every SNR. 2N

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2942 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

 √
where we capitalized on the fact that the maximum is no Introducing xn = Nr Nt /SNR p̄n and X = Nr Nt SNR,
smaller than the average. Altogether, the above becomes

2
2 min −1
N
. /
(t+1) (t+1)
x1 + · · · + xN max log2 1 + x2n
 
2
2 
Nmin −1
xn =X
1 (t) (t)
n=0
xn ≥0
n=0
≤ 1− x1 + · · · + xN (64)  
2N X2
 t+1
2
2  = max ρ log2 1 + 2 . (72)
1 (0) (0) ρ∈{1,2,...,Nmin } ρ
≤ 1− x1 + · · · + xN (65)
2N We start from the base case Nmin = 2, which boils down to a
t→∞
→ 0. (66) single variable optimization given that x2 = X − x1 . From
. / . /
log2 1 + x21 + log2 1 + (X − x1 )2

. /. /
Combining (5) and (6) we obtain, as starting point, = log2 1 + x21 1 + (X − x1 )2 , (73)
the capacity
Nmintaken over all possible channels normalized to
−1 2
satisfy n=0 σn = Nr Nt , that is, it suffices.to show/ .that, when the domain
/ is [0, X], the quartic
function 1 + x21 1 + (X − x1 )2 can attain its maximum
C only at 0, X/2, or X. This quartic function is symmetric
min −1
N
. / with respect to x1 = X/2. If the quartic function has its
= max max log2 1 + σn2 pn .
Nmin −1 2
σn =Nr Nt
Nmin −1
pn =SNR
local maximum at x̄ = 0, X/2, X, then it has one more local
n=0 n=0 n=0
2
σn ≥0 pn ≥0 maximum at X−x̄ from symmetry. This contradicts the quartic
(67) function having at most one local maximum, completing the
proof for the base case.
Then, defining Assuming now that (72) holds for Nmin = N , let us prove
it for Nmin = N + 1. Let x0 , x1 , . . . , xN −1 , whose sum is X,
σn2 /Nr Nt + pn /SNR
σ̄n2 = Nr Nt be given. Now, we replace the maximum
. / and minimum
. /ones
2 with x̄1 and x̄2 maximizing log2 1 + x̄21 + log2 1 + x̄22 and
σ 2 /Nr Nt + pn /SNR
p̄n = n SNR, satisfying
2 x̄1 + x̄2 = max xi + min xi . (74)
i i
we have that the argument of (67) satisfies, by virtue of the
inequality between the arithmetic and geometric means, From the base case, (x̄1 , x̄2 ) turns out to equal either



min −1
N
. / max xi + min xi , 0 , 0, max xi + min xi (75)
log2 1 + σn2 pn i i i i

n=0 or
  2   
min −1
N max xi + min xi max xi + min xi
Nr Nt SNR σn2 pn
≤ log2 1+ + i i
, i i
. (76)
n=0
2 Nr Nt SNR 2 2
min −1
N
. / Over the successive replacements, if at some iteration t it
= log2 1 + σ̄n2 p̄n (68) (t) (t)
happens that (x̄1 , x̄2 ) equal (75), the induction argument
n=0
becomes complete. Otherwise, with the averaging operation
under constraints that are preserved, namely describe in the lemma being performed at each iteration,
min −1
N min −1
N we know that
σ̄n2 = Nr Nt p̄n = SNR. (69) N
 N

n=0 n=0 log2 1 + (x(0)
n )2
≤ log 2 1 + (x(t) 2
n )
n=0 n=0
From the relationship σ̄n2 = Nr Nt  
SNR p̄n , then,
t→∞ X2
→ (N + 1) log2 1+ .
min −1
N   (N + 1)2
Nr Nt 2
C ≤ max log2 1 + p̄ . (70) (77)
Nmin −1
n=0 p̄n =SNR n=0
SNR n
p̄n ≥0 Tracing out footsteps back to (70), and having established (71),
Now, armed with Lemma 1, we set out to prove by induction we altogether have that
 
that Nr Nt SNR
C ≤ max ρ log2 1 + · . (78)
min −1
N   ρ∈{1,2,...,Nmin } ρ ρ
Nr Nt 2
max log2 1 + p̄ Let us now see how this relationship is in fact a strict equality,
Nmin −1
n=0 p̄n =SNR n=0 SNR n
p̄n ≥0 for which purpose it suffices to prove that
   
Nr Nt  Nr Nt SNR
= max ρ log2 1 + SNR . (71) C ≥ ρ log2 1 + · (79)
ρ∈{1,2,...,Nmin } ρ2 ρ ρ

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION 2943

for ρ ∈ {1, 2, . . . , Nmin }. This follows from the right-hand (Otherwise, from Rolle’s theorem there would be another
side being an evaluation of critical value, leading to a contradiction). Therefore, ζn is a
min −1
N solution of
. /    
log2 1 + σn2 pn (80) Nr Nt Nr Nt
n=0 f x −f x =0 (86)
(n + 1)2 n2
at
on the interval
(σ02 , . . . , σN
2
min −1
, p0 , . . . , pNmin −1 ) 
  c c
Nr Nt Nr Nt SNR SNR n2 , (n + 1)2 . (87)
= ,..., , 0, . . . , 0, ,..., , 0, . . . , 0 , Nr Nt Nr Nt
ρ ρ    ρ ρ   
   Nmin −ρ    Nmin −ρ
ρ ρ The left-hand side of (86) is an increasing function of x.
 Since it is negative at NrcNt n2 and positive at NrcNt (n + 1)2 ,
which is a set contained within the set over which C is
ζn uniquely exists. Furthermore, for subsequent reference,
optimized in (67). We therefore have that
  we have that
 Nr Nt SNR    
C = max ρ log2 1 + · (81) Nr Nt Nr Nt
ρ∈{1,2,...,Nmin } ρ ρ f SNR < f SNR (88)
n2 (n + 1)2
where, recall, C  is the capacity over all channels satisfying
 Nmin −1 2
n=0 σn = Nt Nr . Since H is a subset of such channels, for SNR > ζn and
the corresponding capacity is upper-bounded as per (18).    
Let us next establish the value of ρ that yields the upper Nr Nt Nr Nt
f SNR > f SNR (89)
bound at every SNR. With f (·) as defined in (17), the upper n2 (n + 1)2
bound can be rewritten as
  for SNR < ζn .
 Nr Nt
Nr Nt SNR max f SNR . (82) We are now ready to prove (18). Let ζn−1 ≤ SNR < ζn .
ρ∈{1,2,...,Nmin } ρ2 On this SNR interval, we have to show that
A routine manipulation concludes that f is increasing on (0, c]    
Nr Nt Nr Nt
and decreasing on [c, ∞): max f SNR = f SNR , (90)
  ρ∈{1,2,...,Nmin } ρ2 n2
1 2 log(1 + x)
f  (x) = √ − >0 which, by virtue of the unimodality, is equivalent to
(2 log 2) x 1 + x x
2x    
⇔ > log(1 + x) Nr Nt Nr Nt
1+x f SNR ≤ f SNR (91)
2 2x (n − 1)2 n2
⇔− = − 2 > log(1 + x) − 2
1+x 1+x and
2 1+x    
⇔ e− 1+x > Nr Nt Nr Nt
e2 f SNR ≤ f SNR . (92)
2 − 1+x 2 2 (n + 1)2 n2
⇔− e >− 2
1+x  e   
2 2 −2 2 The condition ζn−1 ≥ SNR, plus (88) with the substitution
⇔− > W0 − 2 or < W−1 − 2 n ← (n − 1), give (91). Similarly, the condition SNR < ζn
1+x e 1+x e
  plus (89) give (92).
2 2 With (18) proved, we can now proceed to (19). Let us define
⇔− > W0 − 2
1+x e the sets
2 + W0 (−2/e2) 0   1
⇔x<− = c, (83) Nr Nt
W0 (−2/e2 ) A = ρ log2 1 + SNR : ρ ∈ {1, 2, . . . , N min }
ρ2
where W−1 (·) is the nonprincipal branch. /of a Lambert W
function and (83) holds because W−1 − e22 = −2, whereby and
the second condition is infeasible. 0   1
Nr Nt
Let ζn for n = 1, . . . , Nmin be a threshold equal to the B = ρ̃ log2 1 + SNR : ρ̃ ∈ [1, Nmin] .
ρ̃2
unique positive solution of (16). First of all, let us verify the
existence and the uniqueness of ζn . From the unimodality of Since A ⊂ B, it follows that sup A ≤ sup B. Hence, relaxing
f (·), we have that ρ into ρ̃ can only increase the upper bound. The relaxed upper
Nr Nt Nr Nt bound also can be represented via f (·) as
ζn ≤ c ≤ ζn (84)
(n + 1)2 n2   
Nr Nt
and, subsequently, Nr Nt SNR max f SNR (93)
ρ̃∈[1,Nmin ] ρ̃2
c c
n2 ≤ ζn ≤ (n + 1)2 (85)
Nr Nt Nr Nt and (22) follows from the unimodality of f (·).

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2944 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

A PPENDIX B A PPENDIX C
Suppose that
For SNR ≥ N min c
Nmax , the LOS capacity upper bound is Nmin 2+1 Nmin 2−1
achieved by setting η = 1, hence we concentrate on prov- cr < SNR < cr , (101)
ing (28) for SNR < N min c Nmax Nmax
Nmax . The numerator of (28) is lower
bounded by any achievable spectral efficiency, and in particular such that η in (43) equals r ( = 0, 1, . . . , k − 1). In this
the one obtained by setting interval, from (22) and (42),


$ r 
N min log 1 + Nmax
SNR
Nmax SNR R(SNR) 2 Nmin r 2

η = η̄ = (94) ≥√ 4 . (102)
Nmin c C(SNR) N N SNR
log (1 + c)
min max c 2

and driving only the subchannels with gain larger than With the substitution x = Nmax
SNR,
the right-hand side
Nmin r 2
Nmax (1 − )/η̄. This gives becomes f (x)/f (c) with the above SNR range mapping to
cr ≤ x ≤ c/r and with f (·) as defined in (17). Since f (·) is
max C(H ULA , SNR) unimodal with maximum attained at c,
η∈[0,1]
'2 3' . /
' ' f (x) min f (cr), f (c/r) f (cr)
≥ '  | 1 −  < λ < 1 +  ' min = = (103)
  cr≤x≤c/r f (c) f (c) f (c)
Nmax (1 − ) SNR
· log2 1 + '( )' where the last step follows from f (cr) ≤ f (c/r), as can be
η̄ '  | 1 −  < λ < 1 +  '
'( verified.
)'
'  | 1 −  < λ < 1 +  '
= Nmin A PPENDIX D
 Nmin 
This appendix builds on [41] to establish (48). Letting T =
Nmin (1 − )
Nmax
SNR
· log2 1 + |{ | 1− <λ <1+ }|
, (95) (D rx H ULA D tx )∗ (D rx H ULA D tx ) and denoting by tn−m its
η̄
Nmin (n, m)th entry—the indexing depends only on the difference
between n and m because T is a Toeplitz matrix—we have
which, from the asymptotic polarization of the eigenvalues,
r −1
leads to
N
5 6 5 6
tn−m = (Drx H ULA Dtx )∗ n,
Drx H ULA D tx ,m
maxη∈[0,1] C(H ULA , SNR) =0
lim
N r −1
Nmin →∞ Nmin n m
  = e−j2πη Nmax · ej2πη Nmax (104)
Nmax SNR
≥ η̄ log2 1 + (1 − ) (96) =0
Nmin η 2 (n−m)Nr

 1 − e−j2πη Nmax

= η̄ log2 1 + (1 − ) c (97) = (n−m)

(105)
1 − e−j2πη Nmax


where the last step follows from (94). sin πη (n−m)N r
−jπη (n−m)(N r −1) Nmax
Turning now to the denominator of (28), it is upper =e Nmax
 . (106)
bounded—recall Section IV—by sin πη n−m
Nmax

$ We aim to show that

Nmax SNR 7 . /7
Nmin log2 (1+c) = Nmin · η̄ log2 (1+c). (98) 7 ∗ Nmax 72
Nmin c 7F T F − diag 1, . . . , 1, 0, . . . , 0 7
η    F

ηNmin 
Altogether then, (28) satisfies 7 . / 7
7 Nmax 72
. / = 7T − F diag 1, . . . , 1, 0, . . . , 0 F ∗ 7 (107)
maxη∈[0,1] C(H ULA , SNR) log2 1 + (1 − ) c η    F
lim ≥
ηNmin 
Nmin →∞ C(SNR) log2 (1 + c)
(99) is O(Nt2 log Nt ) whereas
T
2F = Θ(Nt3 ), meaning that
the relatively difference between the terms being subtracted
where  can be arbitrarily small, making the ratio arbitrarily in (107) diminishes as Nr and Nt grow large with a fixed
close to 1. We note that this pointwise convergence does not ratio. Defining
imply uniform convergence, and in fact Nmax . /
C= F diag 1, . . . , 1, 0, . . . , 0 F ∗ , (108)
η   
maxη∈[0,1] C(H ULA , SNR)
ηNmin 
lim min <1 (100)
Nmin →∞ SNR C(SNR) the (n, m)th entry of C, also Toeplitz, is given by


(n−m)
ηNmin 
because of the behavior around the SNR thresholds identified (n−m)(ηNmin −1) sin π N
cn−m = e−jπ
t
in Section IV. Notwithstanding that, for the purposes of this Nt
 . (109)
paper ULAs asymptotically achieve the LOS capacity.
ηNt
Nmax sin π n−m
Nt

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION 2945

With T and C being Toeplitz matrices, their squared differ- concavity of sin(·) on [0, π||/Nt ], whereby the mentioned
ence can be written as argument satisfies


T − C
2F = (Nt − ||) |t − c |2 . (110)    
||<Nt ηNt π|| ηNt
sin · + 1− ·0
Nmax Nt Nmax
By means of the quantities defined as  
ηNt π|| ηNt
πη ηNt π ≥ sin + 1− sin 0. (128)
x = sin x + Δx = sin (111) Nmax Nt Nmax
Nmax Nmax Nt
πηNr πηNmin 
y = sin y + Δy = sin (112) Subsequently, (127) descends from Taylor’s theorem.
Nmax Nt
Moving on to (117), it holds by virtue of | sin y| ≤ 1
(Nr − 1) (ηNmin  − 1)
θ = −πη θ + Δθ = −π , (113) while (118) follows from
Nmax Nt
we rewrite the entries of T and C as ' '
' πηNr πηNmin  ''
y + Δy '
|Δy| = 'sin − sin (129)
y '
t = ejθ c = ej(θ +Δθ ) . (114) '
Nmax Nt
'
x x + Δx ' πηNr '
πηNmin  '
Before examining this squared difference in (110), we intro- ≤ '' − ' (130)
Nmax Nt
duce the proceeding lemma. π||
Lemma 2: The quantities defined above satisfy ≤ (131)
Nt
2η|| πη||
≤ |x |, |x + Δx | ≤ (115)
Nmax Nmax where, in (130), the mean value theorem has been applied.
 3
ηNt 1 π|| Then,
|Δx | ≤ · (116)
Nmax 6 Nt
'   '
|y |, |y + Δy | ≤ 1 (117) ' ηNr ηNmin  1 1 '
|Δθ | = π|| '' − + − ' (132)
π|| Nmax Nt Nt Nmax '
|Δy | ≤ (118)  
Nt 2 1
2π|| ≤ π|| − (133)
|Δθ | ≤ (119) Nt Nmax
Nt 2π||
≤ (134)
for || ≤ Nt /2, and Nt
 2  
1 1 Nmax 1 1
, ≤ + where (133) derives from the triangle inequality. Finally, (120)
x2 (x + Δx )2 2ηNt (||/Nt )2 (1− ||/Nt )2
can be obtained from
(120)
 
for || < Nt . We note that (115), (117), and (120), compactly 1 π2 1 1
≤ + − π ≤ x ≤ π. (135)
indicate inequalities that apply to two distinct quantities. sin2 x 4 |x|2 (π − |x|)2
Proof: Eq. (115) can be directly derived from
2
|x| ≤ | sin x| ≤ |x| − π/2 ≤ x ≤ π/2, (121)
π Utilizing the Toeplitz structure of T and C, we compute
while (116) follows from the squared difference in (110) for three cases: 1)  = 0, 2)
' ' || ≤ Nt /2, and 3) || > Nt /2. In case of  = 0 where both
' πη ηNt π ''
'
|Δx | = 'sin − sin ' (122) t and c are indeterminate form, |t − c |2 is bounded by a
Nmax Nmax Nt
' ' constant:
' πη|| ηN π|| '
'
= 'sin −
t
sin ' (123)
Nmax Nmax Nt '  2  2
ηNmin  Nmax
= sin
πη||
−
ηNt
sin
π||
(124) Nr − ≤ . (136)
Nmax Nmax Nt ηNt /Nmax ηNt
πη|| ηNt π||
≤ − sin (125)
Nmax  Nmax Nt  For 0 < || ≤ Nt /2,
ηNt π|| π||
= − sin (126) ' '2
Nmax Nt Nt ' y y + Δy ''
 3 |t −c |2 = '' − ejΔθ
ηNt 1 π|| x x + Δx '
≤ · (127)  2
Nmax 6 Nt y y + Δy Δθ y y + Δy
= − + 4 sin2 · .
where (124) holds because the argument of | · | in (123) x x + Δx 2 x x + Δx
is always positive. This positivity, in turn, follows from the (137)

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2946 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 20, NO. 5, MAY 2021

Using Lemma 2 and the triangle inequality, the first term can ACKNOWLEDGMENT
be further bounded by a constant: The handling by the Associate Editor, and the feed-
' ' ' '
' y ' ' Δy ''
' − y + Δy ' ≤ ' y Δx back provided by the anonymous reviewers, are gratefully
' x − (138)
x + Δx ' ' x (x + Δx ) x + Δx ' acknowledged.
' ' ' '
' y Δx ' ' Δy '
≤ '' '+' ' (139) R EFERENCES
x (x + Δx ) ' ' x + Δx ' [1] H. Do, N. Lee, and A. Lozano, “Capacity of line-of-sight MIMO

3 channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Jun. 2020,
1 π||
ηNt
Nmax 6· Nt π||/Nt pp. 2044–2048.
≤ + (140) [2] H.-J. Song, H. Hamada, and M. Yaita, “Prototype of KIOSK data
(2η||/Nmax )2 2η||/Nmax downloading system at 300 GHz: Design, technical feasibility, and
π 3 ||Nmax πNmax results,” IEEE Commun. Mag., vol. 56, no. 6, pp. 130–136, Jun. 2018.
= + (141) [3] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and
24ηNt 2 2ηNt A. Ghosh, “Multilevel millimeter wave beamforming for wireless back-
π 3 Nmax πNmax haul,” in Proc. IEEE GLOBECOM Workshops (GC Wkshps), Dec. 2011,
≤ + . (142) pp. 253–257.
48ηNt 2ηNt [4] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and
A. Ghosh, “Millimeter wave beamforming for wireless backhaul and
The second term also is bounded by a constant: access in small cell networks,” IEEE Trans. Commun., vol. 61, no. 10,
' ' ' '
' ' ' ' pp. 4391–4403, Oct. 2013.
'4 sin2 Δθ · y y + Δy ' ≤ '(Δθ )2 y y + Δy ' [5] D. Cvetkovski, T. Halsig, B. Lankl, and E. Grass, “Next generation mm-
' 2 x x + Δx ' ' x x + Δx ' wave wireless backhaul based on LOS MIMO links,” in Proc. German
 2 Microw. Conf. (GeMiC), Mar. 2016, pp. 69–72.
πNmax
≤ . (143) [6] D. Halperin, S. Kandula, J. Padhye, P. Bahl, and D. Wetherall, “Aug-
ηNt menting data center networks with multi-gigabit wireless links,” in Proc.
ACM SIGCOMM Conf. SIGCOMM - SIGCOMM, 2011, pp. 38–49.
Altogether, for || ≤ Nt /2 the aggregate squared difference [7] R. W. Heath Jr and A. Lozano, Foundations of MIMO Communication.
satisfies Cambridge, U.K.: Cambridge Univ. Press, 2018.
[8] P. F. Driessen and G. J. Foschini, “On the capacity formula for multiple
(Nt − ||) |t − c |2 = O(Nt2 ). (144) input-multiple output wireless channels: A geometric interpretation,”
IEEE Trans. Commun., vol. 47, no. 2, pp. 173–176, Feb. 1999.
||≤Nt /2 [9] S. Priebe and T. Kurner, “Stochastic modeling of THz indoor radio chan-
nels,” IEEE Trans. Wireless Commun., vol. 12, no. 9, pp. 4445–4455,
Proceeding to || > Nt /2, we begin by noting that |t −c |2 ≤ Sep. 2013.
2 (|t |2 + |c |2 ). Then, using again Lemma 2, we have that [10] C. Gentile et al., “Millimeter-wave channel measurement and modeling:
A NIST perspective,” IEEE Commun. Mag., vol. 56, no. 12, pp. 30–37,
(Nt − ||) |t |2 Dec. 2018.
[11] M. Shafi et al., “Microwave vs. millimeter-wave propagation channels:
||>Nt /2 Key differences and impact on 5G cellular systems,” IEEE Commun.
⎛ ⎞
 2
Mag., vol. 56, no. 12, pp. 14–20, Dec. 2018.
Nmax ⎝ Nt − || Nt − || ⎠ [12] J.-S. Jiang and M. A. Ingram, “Spherical-wave model for short-
≤ + range MIMO,” IEEE Trans. Commun., vol. 53, no. 9, pp. 1534–1541,
2ηNt (||/Nt )2 (1 − ||/Nt )2
||>Nt /2 ||>Nt /2 Sep. 2005.
[13] F. Bohagen, P. Orten, and G. Oien, “On spherical vs. Plane wave
(145) modeling of line-of-sight MIMO channels,” IEEE Trans. Commun.,
vol. 57, no. 3, pp. 841–849, Mar. 2009.
where the first right-hand-side term satisfies [14] H. Do, S. Cho, J. Park, H.-J. Song, N. Lee, and A. Lozano, “Terahertz
 2 line-of-sight MIMO communication: Theory and practical challenges,”
Nmax Nt − ||
= O(Nt2 ) (146) IEEE Commun. Mag., to be published.
2ηNt (||/Nt )2 [15] E. Torkildson, U. Madhow, and M. Rodwell, “Indoor millimeter wave
||>Nt /2 MIMO: Feasibility and performance,” IEEE Trans. Wireless Commun.,
vol. 10, no. 12, pp. 4150–4160, Dec. 2011.
while the second one satisfies [16] D. Gesbert, H. Bolcskei, D. A. Gore, and A. J. Paulraj, “Outdoor MIMO
 2 wireless channels: Models and performance prediction,” IEEE Trans.
Nmax Nt − || Commun., vol. 50, no. 12, pp. 1926–1934, Dec. 2002.
2ηNt (1 − ||/Nt )2 [17] T. Haustein and U. Kruger, “Smart geometrical antenna design exploiting
||>Nt /2
the LOS component to enhance a MIMO system based on Rayleigh-
 2
Nmax 1 fading in indoor scenarios,” in Proc. 14th IEEE Proc. Pers., Indoor
= Mobile Radio Commun. PIMRC, Sep. 2003, pp. 1144–1148.
2η Nt − || [18] F. Bohagen, P. Orten, and G. E. Oien, “Construction and capacity
||>Nt /2
analysis of high-rank line-of-sight MIMO channels,” in Proc. IEEE
= O(Nt2 log Nt ), (147) Wireless Commun. Netw. Conf., Mar. 2005, pp. 432–437.
[19] I. Sarris and A. R. Nix, “Maximum MIMO capacity in Line-of-Sight,”
with the last equality holding by virtue of the logarithmic in Proc. 5th Int. Conf. Inf. Commun. Signal Process., Dec. 2005,
pp. 1236–1240.
growth of the harmonic series. The same result can be obtained [20] F. Bohagen, P. Orten, and G. E. Oien, “Design of optimal high-rank
for c , hence line-of-sight MIMO channels,” IEEE Trans. Wireless Commun., vol. 6,
 no. 4, pp. 1420–1425, Apr. 2007.

T − C
2F = ||<Nt (Nt − ||) |t − c |2 = O(Nt2 log Nt ). [21] I. Sarris and A. R. Nix, “Design and performance assessment of
high-capacity MIMO architectures in the presence of a line-of-sight
This scaling, in conjunction with
T
2F = Θ(Nt3 ), indicates component,” IEEE Trans. Veh. Technol., vol. 56, no. 4, pp. 2194–2202,
that the relatively difference between the terms subtracted Jul. 2007.
[22] P. Larsson, “Lattice array receiver and sender for spatially OrthoNor-
in (107) diminishes as Nr and Nt grow large with a fixed mal MIMO communication,” in Proc. IEEE 61st Veh. Technol. Conf.,
ratio. Therefore, (48) holds with asymptotic equality. May 2005, pp. 192–196.

Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
DO et al.: RECONFIGURABLE ULAs FOR LOS MIMO TRANSMISSION
[23] X. Song and G. Fettweis, “On spatial multiplexing of strong line-of-
sight MIMO with 3D antenna arrangements,” IEEE Wireless Commun.
Lett., vol. 4, no. 4, pp. 393–396, Aug. 2015.
[24] N. Matsumura, K. Nishimori, R. Taniguchi, T. Hiraguri, T. Tomura,
and J. Hirokawa, “Novel unmanned aerial vehicle-based line-of-sight
MIMO configuration independent of transmitted distance using millime-
ter wave,” IEEE Access, vol. 8, pp. 11679–11691, 2020.
[25] C. Sheldon, E. Torkildson, M. Seo, C. P. Yue, U. Madhow, and
M. Rodwell, “A 60GHz line-of-sight 2×2 MIMO link operating at
1.2Gbps,” in Proc. IEEE Antennas Propag. Soc. Int. Symp., Jul. 2008,
pp. 1–4.
[26] C. Sheldon, E. Torkildson, M. Seo, C. P. Yue, U. Madhow, and
M. Rodwell, “Spatial multiplexing over a line-of-sight millimeter-
wave MIMO link: A two-channel hardware demonstration at 1.2Gbps
over 41m range,” in Proc. Eur. Conf. Wireless Technol., Oct. 2008,
pp. 198–201.
[27] C. Sheldon, M. Seo, E. Torkildson, M. Rodwell, and U. Madhow, “Four-
channel spatial multiplexing over a millimeter-wave line-of-sight link,”
in IEEE MTT-S Int. Microw. Symp. Dig., Jun. 2009, pp. 389–392.
[28] L. Bao and B.-E. Olsson, “Methods and measurements of channel phase
difference in 2×2 microwave LOS-MIMO systems,” in Proc. IEEE Int.
Conf. Commun. (ICC), Jun. 2015, pp. 1358–1363.
[29] C. Hofmann, K.-U. Storek, R. T. Schwarz, and A. Knopp, “Spatial
MIMO over satellite: A proof of concept,” in Proc. IEEE Int. Conf.
Commun., May 2016, pp. 1–6.
[30] T. Halsig, D. Cvetkovski, E. Grass, and B. Lankl, “Measurement results
for millimeter wave pure LOS MIMO channels,” in Proc. IEEE Wireless
Commun. Netw. Conf. (WCNC), Mar. 2017, pp. 1–6.
[31] Y. Yan et al., “11-Gbps broadband modem-agnostic line-of-sight MIMO
over the range of 13 km,” in IEEE Global Commun. Conf., Dec. 2018,
pp. 1–7.
[32] G. Sellin et al., Ericsson Microwave Outlook. Göteborg, Sweden:
Ericsson, 2019.
[33] M. Matthaiou, A. M. Sayeed, and J. A. Nossek, “Maximizing LoS
MIMO capacity using reconfigurable antenna arrays,” in Proc. Int. ITG
Workshop Smart Antennas (WSA), Feb. 2010, pp. 14–19.
[34] N. Chiurtu and B. Rimoldi, “Varying the antenna locations to optimize
the capacity of multi-antenna Gaussian channels,” in Proc. IEEE Int.
Conf. Acoust., Speech, Signal Process., Jun. 2000, pp. 3121–3123.
[35] S. Sun, T. Rappaport, R. Heath, A. Nix, and S. Rangan, “MIMO for
millimeter-wave wireless communications: Beamforming, spatial multi-
plexing, or both?” IEEE Commun. Mag., vol. 52, no. 12, pp. 110–121,
Dec. 2014.
[36] E. Torkildson, B. Ananthasubramaniam, U. Madhow, and M. Rodwell,
“Millimeter-wave MIMO: Wireless links at optical speeds,” in Proc.
Allerton Conf. Commun., Control Comput., 2006, pp. 1–9.
[37] X. Song, C. Jans, L. Landau, D. Cvetkovski, and G. Fettweis, “A 60
GHz LOS MIMO backhaul design combining spatial multiplexing and
beamforming for a 100 Gbps throughput,” in IEEE Global Commun.
Conf., Dec. 2015, pp. 1–6.
[38] H. Sarieddeen, M.-S. Alouini, and T. Y. Al-Naffouri, “Terahertz-band
ultra-massive spatial modulation MIMO,” IEEE J. Sel. Areas Commun.,
vol. 37, no. 9, pp. 2040–2052, Jul. 2019.
[39] C. Lin and G. Y. Li, “Terahertz communications: An array-of-subarrays
solution,” IEEE Commun. Mag., vol. 54, no. 12, pp. 124–131, Dec. 2016.
[40] A. Tulino, A. Lozano, and S. Verdú, “MIMO capacity with channel state
information at the transmitter,” in Proc. IEEE Int. Symp. Spread Spectr.
Techn. Applic. (ISSSTA), Aug. 2004, pp. 22–26.
[41] P. Wang, Y. Li, X. Yuan, L. Song, and B. Vucetic, “Tens of gigabits
wireless communications over E-band LOS MIMO channels with uni-
form linear antenna arrays,” IEEE Trans. Wireless Commun., vol. 13,
no. 7, pp. 3791–3805, Jul. 2014.
[42] X. Song, D. Cvetkovski, W. Rave, E. Grass, and G. Fettweis, “Sequential
channel equalization in strong line-of-sight MIMO communication,”
IEEE Trans. Wireless Commun., vol. 18, no. 1, pp. 340–356, Jan. 2019.
[43] Z. Zhu, S. Karnik, M. A. Davenport, J. Romberg, and M. B. Wakin,
“The eigenvalue distribution of discrete periodic time-frequency limiting
operators,” IEEE Signal Process. Lett., vol. 25, no. 1, pp. 95–99, 2017.
[44] A. Edelman, P. McCorquodale, and S. Toledo, “The future fast Fourier
transform,” SIAM J. Sci. Comput., vol. 20, no. 3, pp. 1094–1114, 1998.
[45] I. Gohberg and V. Olshevsky, “Fast algorithms with preprocessing for
matrix-vector multiplication problems,” J. Complex., vol. 10, no. 4,
pp. 411–427, Dec. 1994.
[46] Study on Channel Model for Frequencies From 0.5 to 100 GHz (Release
16), Standard ETSI TR 138 901 V14.0.0 (2017-05), 3GPP Technical
Specification Group Radio Access Networks, Tech. Rep., Dec. 2019.
Authorized licensed use limited to: EURECOM. Downloaded on May 24,2023 at 12:38:15 UTC from IEEE Xplore. Restrictions apply.
2947
[47] F. Bøhagen, P. Orten, and G. Øien, “Optimal design of uniform rectan-
gular antenna arrays for strong line-of-sight MIMO channels,” EURASIP
J. Wireless Commun. Netw., vol. 2007, no. 1, pp. 1–10, Dec. 2007.
[48] Q. Bai, A. Tennant, and B. Allen, “Experimental circular phased array
for generating OAM radio beams,” Electron. Lett., vol. 50, no. 20,
pp. 1414–1415, Sep. 2014.
[49] K. A. Opare, Y. Kuang, J. J. Kponyo, K. S. Nwizege, and Z. Enzhan,
“The degrees of freedom in wireless line-of-sight OAM multiplexing
systems using a circular array of receiving antennas,” in Proc. 5th Int.
Conf. Adv. Comput. Commun. Technol., Feb. 2015, pp. 608–613.
[50] X. Song, W. Rave, N. Babu, S. Majhi, and G. Fettweis, “Two-level spa-
tial multiplexing using hybrid beamforming for millimeter-wave back-
haul,” IEEE Trans. Wireless Commun., vol. 17, no. 7, pp. 4830–4844,
Jul. 2018.
[51] E. Torkildson, C. Sheldon, U. Madhow, and M. Rodwell, “Nonuniform
array design for robust millimeter-wave MIMO links,” in Proc. IEEE
Global Telecommun. Conf. (Globecom), Nov. 2009, pp. 1–7.
[52] P. Wang, Y. Li, Y. Peng, S. C. Liew, and B. Vucetic, “Non-uniform linear
antenna array design and optimization for millimeter-wave communica-
tions,” IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7343–7356,
Nov. 2016.
Heedong Do (Graduate Student Member, IEEE)
received the B.S. degree in mathematics and the
M.S. degree in electrical engineering from the
Pohang University of Science and Technology
(POSTECH), Pohang, South Korea, in 2018 and
2020, respectively. He is currently pursuing the
Ph.D. degree.
Namyoon Lee (Senior Member, IEEE) received
the Ph.D. degree from The University of Texas at
Austin, in 2014. He was with the Communications
and Network Research Group, Samsung Advanced
Institute of Technology (SAIT), South Korea, from
2008 to 2011, and also with the Wireless Com-
munications Research (WCR), Intel Labs, Santa
Clara, CA, USA, from 2015to 2016. He is currently
an Associate Professor with POSTECH, Pohang,
Gyeongbuk, South Korea. He was a recipient of
the 2016 IEEE ComSoc Asia-Pacific Outstanding
Young Researcher Award and the 2020 IEEE Best YP Award (Outstanding
Nominee). He is also an Editor of the IEEE T RANSACTIONS ON W IRELESS
C OMMUNICATIONS, the IEEE T RANSACTIONS ON C OMMUNICATIONS,
the IEEE C OMMUNICATIONS L ETTERS , and the IEEE T RANSACTIONS ON
V EHICULAR T ECHNOLOGY.
Angel Lozano (Fellow, IEEE) received the Ph.D.
degree from Stanford University in 1998. From
1999 to 2008, he was with joined Bell Labs
(Lucent Technologies, now Nokia). He is cur-
rently a Professor with Pompeu Fabra University
(UPF), Barcelona, and the coauthor of the textbook
Foundations of MIMO Communication’ (Cambridge
University Press, 2019). He serves as an Area
Editor for the IEEE T RANSACTIONS ON W IRE -
LESS C OMMUNICATIONS. He received the 2009
Stephen O. Rice Prize, the 2016 Fred W. Ellersick
prize, and the 2016 Communications Society and Information Theory Society
Joint Paper Award. He holds an Advanced Grant from the European Research
Council and was a 2017 Clarivate Analytics Highly Cited Researcher.