1
Enhanced Tracking and Beamforming Codebook
Design for Wideband Terahertz Massive MIMO
System
Xu Shi, Student Member, IEEE, Jintao Wang, Senior Member, IEEE, Jian Song, Fellow, IEEE
arXiv:2211.13933v1 [cs.IT] 25 Nov 2022
Abstract—True-time-delay (TTD) lines are recently applied
inside Terahertz (THz) hybrid-precoding transceiver to acquire
high beamforming gain against beam squint effect. However,
beam tracking turns into a challenging puzzle where enormous
potential beam directions bring about unacceptable overhead
consumption. Frequency-scanning-based beam tracking is ini-
tially explored but still imperfect in previous studies. In this
paper, based on TTD-aided hybrid precoding structure, we
give an enhanced frequency-scanning-based tracking scheme.
Multiple beams are generated and utilized simultaneously via
several subcarriers for tracking at one timeslot. The squint
beams’ angular coverage at all subcarriers can be flexibly
controlled by two different subcarrier-angular mapping policies,
named forward-pairing and backward-pairing. Then multiple
physical directions can be simultaneously searched in one timeslot
for lower overhead consumption. Besides, closed-form searching
radius bound, parameter configuration and interferences are
theoretically analyzed. Furthermore, we provide the coupled
codebook design for TTDs and phase shifters (PSs), with joint
consideration of both beamforming and tracking. Analytical
and numerical results demonstrate the superiority of the new
frequency-scanning-based tracking scheme and beamforming
codebook.
Index Terms—Wideband THz, tracking and beamforming,
codebook design, beam squint, true-time-delay
I. I NTRODUCTION
Exploitation of Terahertz (THz) spectrum is expected to
accommodate enormous emerging wireless applications with
high data throughput requirements [1]–[5]. Benefited from the
abundant band resources and high frequency, the supported
devices and transmission rate are prominently enlarged in THz
wireless communication [6]. To overcome the severe signal
attenuation in THz band, massive multiple-input multiple-
output (MIMO) is indispensable to generate narrow ’pencil’
beam with enhanced array gain [7], [8]. Besides, given the
limited hardware and power consumption in radio frequency
(RF) chain, hybrid beamforming structure has been widely
adopted in millimeter wave (mmWave) and THz systems to
further upgrade energy efficiency [9], [10].
Due to the massive number of antennas and huge spectrum
bandwidth, beam squint effect is deadly harmful in wideband
This work was supported in part by Tsinghua University-China Mobile
Research Institute Joint Innovation Center. (Corresponding author: Jintao
Wang.)
Xu Shi, Jintao Wang and Jian Song are with the Department of Elec-
tronic Engineering, Tsinghua University, Beijing 100084, China and Beijing
National Research Center for Information Science and Technology (BN-
Rist). (e-mail: shi-x19@mails.tsinghua.edu.cn; wangjintao@tsinghua.edu.cn;
jsong@tsinghua.edu.cn).
THz massive MIMO system [11]–[14]. In beam squint effect,
one propagation path points in distinct beam directions at
different subcarriers since the shifted phase in realistic channel
varies with frequency. Unfortunately, the conventional phase
shifter (PS)-aided analog precoder is frequency-independent
and beam mismatch then appears in THz wideband beam-
forming, which may severely weaken communication quality.
Up to now, the existing schemes against beam squint can
be divided into two categories. One is based on the conven-
tional PS-aided hybrid precoding structure, including window
compensation [15] and beam broadening [16]. The former
is to scale signals in spatial-frequency domain via energy-
focusing window, and the latter is to establish several virtual
sub-arrays for evenly distributed array gain across the whole
frequency band. Other related hybrid precoding designs can
also be found in [17], [18]. The other category is true time
delay (TTD)-based [19], [20]. Inspired by large-bandwidth
phased radar system, we can perfectly mitigate beam squint
by replacing PSs with TTDs, which can adaptively control
the time delay of received signals from each antenna [21],
[22]. Given that TTDs occupy quite high hardware cost and
power consumption, [19] established a hybrid network with a
few TTDs and abundant PSs, which can approach near-optimal
performance on sum-rate and array gain across the whole THz
frequency band. All in all, the PS-aided schemes only partially
relieve incompatibility inside beam squint, but fortunately the
TTD-based method can be regarded as an exhaustively perfect
enabler for beam squint compensation [16].
Besides, the channel state information (CSI) acquirement
appears urgently challenging due to the lengthy pilot overhead
as antennas grow in THz operating band [4]. As antennas
increase and beam width gets narrower, much more beams
should be searched in a fixed angular interval for the opti-
mal beamforming direction, which requires unaffordable pilot
overhead due to the monogamy between timeslot and beam.
Therefore, traditional mmWave beam searching schemes are
no longer appropriate such as hierarchical sweeping [23], user
mobility-assisted tracking [24] like Kalman filtering [25], and
so on [26]. Fortunately, by adopting beam squint effect and
TTD-aided beamforming structure, [27] innovatively designed
a specific pairing between frequency and beam angle, which is
quite similar to the frequency-scanned radar. Multiple beams
are generated and utilized simultaneously via several subcar-
riers with only one RF chain for tracking, which also means
that several beams can be searched at one timeslot and pilot
overhead is then reduced by 90%.

However, there still exist some puzzles for TTD-aided THz
wireless communications. From point of view of tracking,
limited by the beam pattern and pairing rule, the zooming
pairing scheme [27] only obtains good performance on semi-
angular domain [−1, 0], but suffers from severe performance
degradation when spatial direction is located inside [0, 1].
And power leakage problem herein seems more challenging
since the pairing for beam direction is non-uniform and
previous fine-tune methods like spatial rotation [25] cannot be
employed directly. On the other hand, joint consideration of
tracking and beamforming is necessary for TDD-aided system
due to the same hardware support. Since TTDs and PSs are
conjugate and both restricted to finite value sets in practice,
quantized codebook design and resolution analysis are sig-
nificant in TTD-PS-aided transceiver, especially to support
both the beamforming and tracking modules simultaneously.
Unfortunately, to our best knowledge, there exist few studies
that thoroughly addressed the above puzzles.
Correspondingly, the contributions of this article are sum-
marized as follows:
• Based on the TTD-aided frequency-dependent beams,
we provide a new pairing policy between frequency
and angular directions (called forward-backward-pairing),
and further propose an enhanced wideband frequency-
scanning-based tracking scheme. More efficient beam
patterns can be generated to simultaneously search several
beam angles via subcarrier indices at one timeslot, and the
whole angular domain [−1, 1] can be robustly supported
with lower pilot overhead.
• We develop a novel compressive phase retrieval (CPR)-
based method for power leakage compensation after
wideband pairing-based tracking. In this gridless com-
pensation scheme, we exploit the commonalities among
multiple pilot timeslots and several subcarriers to further
improve tracking accuracy. Simulation results confirm
the superiority of the compensation scheme in wideband
frequency-scanning-based tracking.
• Besides, we further theoretically expand the tracking cov-
erage under one timeslot for lower overhead. We derive
closed-form upper bound of angular searching radius. In
certain utmost directions, the frequency-dependent beam
coverage, i.e., the angular tracking range, can be enlarged
by 50%.
• Moreover, we design a quantized decoupled codebook for
both TTDs and PSs in TDD-aided beamforming archi-
tecture, which is realistic and modified from the basic
discrete Fourier transform (DFT) codebook. The new
quantized codebook can simultaneously support TTD-
aided beamforming and tracking with quite low degrada-
tion. The resolutions among codewords are then analyzed
theoretically.
The rest of this paper is organized as follows. In Section II,
we introduce the TTD-aided beamforming structure and ana-
lyze the propagation channel in time and frequency domain. In
Section III, we propose the coupled codebook for TTDs and
PSs in equivalent baseband model and further provide analysis
for array gain and beam characteristics. The enhanced forward-
2
backward wideband-mapping tracking algorithm is proposed
in Section IV, while we design a novel CPR-based energy-
focusing scheme for power leakage problem. In Section V
we give thorough analysis and evaluation for the proposed
tracking scheme and obtain closed-form searching ranges. In
Section VI we provide the codebook resolution requirement
with joint consideration of TTD-based beamforming and track-
ing. Finally, we show the simulation results in Section VII and
conclude this paper in Section VIII.
Notations: Lower-case and upper-case boldface letters a
and A denote a vector and a matrix respectively. The op-
erators (·)∗ , (·)T and (·)H denote conjugate, transposition and
conjugate transposition of matrix. We use (·)−1 and (·)† to
denote inverse and pseudo-inverse. Tr(·) denotes matrix trace.
⊗ and ◦ denote the Kronecker product and Hadamard product,
respectively. IM is identity matrix with size M × M . The
operator diag(a) is a square diagonal matrix with entries of
a on its diagonal. ||∆||F and ||∆||2 denote Frobenius norm
and l2 norm for matrix ∆. CN (µ, Σ) stands for a circularly
symmetric complex Gaussian distribution with mean µ and
variance Σ.
II. S YSTEM M ODEL
Let us consider a wideband TTD-aided massive MIMO
system as shown in Fig. 1(a), where the base station (BS)
is configured with uniform linear array (ULA) of NBS anten-
nas and NRF RF chains. The carrier frequency and system
bandwidth are set as fc and B, respectively. Denote d = 2fcc
as the fixed antenna spacing and c is the propagation veloc-
ity of electromagnetic waves. For convenience, we assume
the 2M + 1 baseband subcarriers are symmetric to zero
B
with subcarrier spacing fd = 2M , which can be written
b
as fm = mfd , m = −M, . . . , M . Furthermore, the up-
conversion subcarrier frequencies can be written as
fm = fc + mfd , m = −M, −M + 1, . . . , M. (1)
In the TTD-aided hybrid beamforming transceiver as shown
in Fig. 1(a), NTTD TTD lines are equipped between each RF
chain and corresponding NBS PSs. Specifically, overall NBS
phase shifters assigned to each RF chain are divided into NTTD
groups, i.e., the number of phase shifters assigned to each TTD
is marked as P and satisfies that NTTD ·P = NBS . According to
[16], the k-th TTD line’s time-domain response can be written
as δ(t − t̂k ). Correspondingly, the frequency responses for PS
and TTD can be formulated as follows.
fm
VPS (φ) = e−jπφ , VTTD (t̂, fm ) = e−j2πfm t̂ = e−jπ fc t , (2)
where we define auxiliary time delay as
t , 2fc t̂ (3)
and t̂ is the real delay time caused by TTD. By utilizing
the TTD line network, different phase biases under several
frequencies, caused by beam squint phenomenon, can be
effectively compensated 1 . And we can express the received
1 The tradeoff for TTD number has been studied in [19] and near-optimal
beamforming performance was theoretically confirmed with limited TTD
lines.
 3

Antennas
Data Stream and beamforming codebook design of wideband THz system.
1
1
. Without loss of generality, we only consider one UE (one
. . .
RF Chain .
.
.
.
.
RF chain) scenario in this paper 2 . Then the corresponding
1
Baseband !澔
.
. Digital . simplified signal in l-th timeslot can be rewritten as:
. .
.
Precoder
!澔 . .
.
.
ym,l = hH
m [fPS (φl ) ◦ (fTTD,m (tl ) ⊗ 1P ×1 )] sl + nm . (9)
RF Chain . . .
. .
!" 澔 where φ ∈ CNBS ×1 is phase shift vector and t ∈ CNTTD ×1 is
the time delay vector.
: "#$%& ' 澔 : "#$( 澔 TTD Network PS Network

Fig. 1. Block diagram of TTD-aided beamforming model. III. C ODEBOOK D ESIGN FOR TTD- AIDED M ODEL
Data Stream
Antennas In this section we give the codebook design for TTD-aided
system, which is the basis for all practical beamforming and
signal at the subcarrierRFfChain
m as: tracking modules next. Moreover, it is necessary to analyze
Baseband
y
Digital m
= HH
m Fm Dm sm + nm ,
澔 (4) the codebook-based array gain and corresponding beam char-
Precoder acteristics in this section, since the analysis and inference
where澔 Hm = [h1,m , . . . , hK,m ] ∈ CNBS ×K represents are extremely instructive for the following beam tracking,
THz channel matrixRFfor
Chain
K users in the frequency fm . codebook design and performance analysis.
澔
NBS ×NRF
Fm ∈ C and Dm ∈ CNRF ×NRF denote analog Inspired by the derivation from [20], [27], we establish an
(PS+TTD) and baseband澔 digitalBaseband
precoders,Equivalent
respectively. And equivalent beamforming model as shown in Fig. 1(b). We
2
n ∼ CN (0, σN IK ) representsTTD
additive
Network white Gaussian noise
PS Network fm
divide the TTD phase e−jπ fc tn into frequency-independent
2
(AWGN) at all K users with noise power σN . The analog b
fm

beamforming matrix Fm can be further decomposed as PS part e−jπtn and baseband part e−jπ fc tn , and merge the
and TTD two parts as follows former part into PS network. Define the equivalent PS values
as ψ = φ + (t ⊗ 1P ×1 ), and the equivalent signal model 3 can
Fm = FPS (Φ) ◦ F̂TTD,m (T ) be further rewritten as
(5)
y m = hH b
 
= FPS ◦ (FTTD,m ⊗ 1P ×1 ). m fPS (ψ) ◦ (fTTD,m (t) ⊗ 1P ×1 ) s + nm , (10)
b b
where baseband time delay fm (t) is defined as [fm (t)]n =
where FPS ∈ CNBS ×NRF and FTTD,m ∈ CNTTD ×NRF denote the b
fm

frequency response matrix of phase shifters and TTD lines
fm
at fm , respectively. [FTTD,m ]n,k = e−jπ fc tn,k denotes the
response for the n-th TTD of k-th RF chain at subcarrier fm ,
with auxiliary delay time set as tn,k . Similarly, [FPS ]n,k =
e−jπφn,k is respective shifted phase through PS network.
The widely utilized ray-based time-domain channel impulse
response [28] between the k-th UE and the n-th wideband THz
BS antenna hk,n (t) is expressed as
L 
ψk,l
X
−j2π(n−1)ψk,l 0
hk,n (t) = gk,l e δ t − τk,l − (n − 1) ·
fc
l=1
(6)
where δ(·) denotes the Dirac delta function, ψk,l is the l-
th beam angle of arrival (AoA) and (n − 1) fψc is the delay
difference among BS antennas which causes the spatial-
wideband effect [13]. By transforming the impulse response
into frequency domain and reshaping it into a vector along all
antennas, we can obtain the corresponding frequency response:
L Remark 1: The equivalent signal model and corresponding
X b 0
hk,m = gk,l e−j2πfm τk,l aNBS (fm , ψk,l ).
l=1
and the frequency-dependent steering vector
fm
aN (fm , ψ) , [1, e−jπ fc ψ , . . . , e−jπ fc
e−jπ fc tn . Intuitively speaking, the equivalent PS array fPS
provides the rough beam direction while the baseband TTD
b
array fTTD,m undertakes a fine-tune for compensation of beam
squint. Therefore, the codebook for ψ and t can be designed
as discrete Fourier transform (DFT)-style:
t = P [0, t, . . . , (NTTD − 1)t]T ,
(11)
ψ = [0, ψ, . . . , (NBS − 1)ψ]T , ψ ∈ UQ
Rψ ([−1, 1])

where UQ
, R ([a, b]) denotes uniform quantization for interval
[a, b] with sampling spacing (resolution) R. After the code-
word selection of ψ and t, we can directly calculate realistic
PS value φ and delay time t̂ to further control the initial
TTD-aided transceiver. The quantization resolution design of
t and ψ in (11) will be shown in Section VI, with joint
consideration of beamforming and tracking. So for brevity,
we don’t illustrate too much about codebook performance in
this section.
(7) codebook generation for TTD-aided beamforming simultane-
ously support wideband beamforming and tracking. To the
best of our knowledge, it is first clearly proposed and jointly
fm
considered from the two aspects in this paper, which is
(N −1)ψ T
] . (8)
2 Multi-user scenario can be similarly extended with additional consideration

Due to the severe path loss under extreme high frequency of Inter-User Interference (IUI) in beam selection module, which has been
and large antenna number, line-of-sight (LoS) beam is domi- deeply studied in conventional mmWave hybrid beamforming.
3 Notice that the tuple {f (ψ), f b (t)} is a bijection to {f (φ), f (t)}
PS m PS m
nant among all channel paths in THz communication. Besides, and is reversible from the latter, which also means the two models are
it is worth noting that we mainly focus on beam tracking equivalent.
 4

single beam pattern at subcarrier !" 澔

period, semi-width and the global-maximum point location are
2fc 2fc fc fb
(2)
Tm = , Wm (2)
= (2)
, θc,m = ψ + mt
P fm NBS fm fm fm
(15)
(1)
It is worth noting that the former period Tm is much
(2)
larger than Tm and near to the width of argument interval
[−1, 1]. Therefore, we can regard the former Dirichlet function
ΞP ( ffmc θ − ψ) as a window for the second term ΞNTTD (·)
and almost all components outside the main lobe of ΞP are
filtered to zero, as depicted in Fig. 2. Besides, notice that the
(1) (2)
former period Tm is equal to the latter semi-width Wm ,
which means that there must exist only two peaks inside main
(1) (2)
lobe of ΞP (when θc,m 6= θc,m ). The maximum array gain at
(2) (2)
subcarrier fm almost always appears at θ = θm,c + kTm and
(1) (1)
Fig. 2. The beam patterns of ΞP (·), ΞNTTD (·) and final beam G under single must have |θ − θc,m | < Wm .
subcarrier f−M . For wideband beamforming, we set PS and TTD parameters
equal to the physical user direction θ0 ideally, i.e., ψ = t = θ0 .
In this way, the global-maximum point for array gain G locates
necessary and useful, although some brilliant studies have been at θ = θ0 for all subcarriers fm , which means beam squint is
done from partial aspects such as [20], [27]. perfectly eliminated via TTD line.
As for TTD-aided wideband tracking, we configure the
Next, we analyze the array gain based on the codebook
equivalent PS/TTD values ψ, t bias to real beam angle. No-
above, which can be calculated as follows:
tice the maximum array gain at subcarrier fm must lie in
(2) (2)
1 θm,c (+kTm ) and thus we can strictly design the values of
G(fm , θ) = aN (fm , θ)H fm ψ and t to form particular beam patterns for tracking, i.e.,
NBS
P N TTD
establish a monotonical bijection between subcarriers {fm }
1 X X fm
and quantized angle set {θm } via maximum array gain location
= ejπ[(q−1)P +(p−1)] fc θ ×
NBS p=1 q=1
(15), i.e.,
b fc fb
e −jπ[(q−1)P +(p−1)]ψ
×e −jπ(q−1)
fm
fc Pt θm = ψ + m t, m = −M, . . . , M. (16)
fm fm
P NTTD
1 X jπ(p−1)( ffm θ−ψ) X fm fmb where θm denotes the angular location with maximum beam-
= e c ejπP (q−1)( fc θ−ψ− fc t) forming gain at frequency fm .
NBS p=1 q=1 For clearer illustration of the frequency-scanning tracking,
fb
1 sin P2π ( ffmc θ − ψ) sin NBS π fm
2 ( fc θ − ψ − fc t)
m we provide the holistic gain distribution of overall subcarriers
= · in Fig. 4(b), where 2M + 1 beamforming curves are plotted,
P NTTD sin π2 ( ffm θ − ψ) fb
sin P2π ( ffmc θ − ψ − fmc t)
c corresponding to the whole 2M + 1 subcarriers. Each curve

fm
  
fm fb
 here displays a similar beam pattern to the gray region of
= ΞP θ − ψ · ΞNTTD P θ − ψ − mt Fig. 2. We can obviously observe the monotonical bijective
fc fc fc
(12) relationship between subcarriers {fm } and quantized angular
where values {θm } in Fig. 4(b). Following this mechanism, we can
1 sin N2π a easily screen the maximum beamforming gain, corresponding
ΞN (a) , (13) subcarrier and angular direction to implement tracking proce-
N sin π2 a
dure in one timeslot.
is the normalized Dirichlet function, with period TDir = 2 and
main lobe’s semi-width WDirsemi
= N2 [29].
IV. E NHANCED T RACKING
Based on the derivation in (12), we exploit some interesting
properties of the two Dirichlet functions ΞP (·) and ΞNTTD (·), A. Drawback of previous frequency-scanned tracking
which may be helpful for the following beam tracking and Before we propose the enhanced tracking scheme, it is
performance analysis. We herein regard real channel angle θ necessary to illustrate the conventional zooming-based method
as the function argument. As for the former ΞP ( ffmc θ − ψ), the [27] and its drawback in some specific scenarios. Here we only
period, semi-width and the global-maximum point location are consider one timeslot for tracking, i.e., we only send pilot sl
calculated as: once, with L = 1. Assume the initial beam searching range is
(1) 2fc (1) 2fc (1) fc set as interval [θ0 − α, θ0 + α], where θ0 denotes the central
Tm = , Wm = , θc,m = ψ (14) angle and 2α is the whole searching length in angular domain.
fm P fm fm
b
Based on the TTD array gain derived in (12), [27] projects all
fm
Similarly, as for the latter ΞNTTD (P ( ffmc θ − ψ − fc t)), the subcarriers to beam angles forwardly as shown in Fig. 3(a).
 5

Fix the peripheral mapping f−M → θ0 − α and fM → θ0 + α Frequency

color bar # $0 +
" '
%澳 # $0 ! %澳 # $0 !
" '
%澳
!" # $0 ! %澳 &
&
" &
and we have !" 澳
&

 ­=0
fc M fd .
.
­=0
 θ0 − α= ψ fw − tfw

 .
fc − M fd fc − M fd , (17) 澳
fc M fd #

 θ0 + α= ψ fw + tfw

 " # $0 + %澳 !" # $0 + %澳
fc + M fd fc + M fd .
.
.
and the final configuration of equivalent PS/TTD can be
澳
directly calculated via (17): "

 (a) Forward Pairing (b) Backward Pairing

M fd
 ψ fw = θ0 + α


fc . (18) Fig. 3. The tracking procedures of forward and backward pairings between
fw fc subcarriers and angular values.
 t = θ0 + α


M fd
Then frequency-dependent beams can be further generated to between realistic physical direction and tracked angle.
simultaneously track several physical directions, i.e., the angle-
Therefore, it is necessary to design enhanced frequency-
domain coverage [θ0 − α, θ0 + α], as depicted in Fig. 4. scanning-based tracking scheme with enlarged and fixed
However, the coverage radius α here cannot be arbitrarily searching range α, also with consideration of power
configured but should satisfy certain strict constraints. To avoid leakage compensation.
passive influence of sidelobes inside ΞNTTD (·) and periodic
(2)
beams inside ΞP (·), we should ensure that θc,m is located
around the strong mainlobe of windowing function ΞP (·) with B. Enhanced Tracking Design
(1)
maximum angular radius Wm /2, i.e., To overcome the disadvantages analyzed above, we propose
"
(1) (1)
# a novel wideband THz tracking scheme named Forward-
(2) (1) Wm (1) Wm Backward-pairing based algorithm.
θc,m ∈ θc,m − , θc,m + , ∀m. (19)
2 2 As illustrated in Section III, we should establish a bijection
Substituting (14), (15) and (18) into (19), then the constraint between subcarriers {fm } and angular searching set {θm } ⊂
for α is simplified as follows [θ0 − α, θ0 + α], where the sampled angular points satisfy
1 M fd θ0 − α = θ−M < · · · < θm < · · · < θM = θ0 + α. (21)
αfw < − θ0 . (20)
P fc The conventional zooming tracking only considered forward
When we further enlarge the range α > P1 − Mf fc θ0 , there ap-
d projection fm → θm . However, there exists another projection
pears a diffusion in the multi-frequency beam pattern as shown approach named as backward mapping, i.e., fm → θ−m
in Fig. 4(a), which causes a mismatch between subcarriers and as shown in Fig. 3(b). By jointly adopting the forward and
quantized angles and thoroughly wrecks the zooming-based backward mappings in different scenarios, we can give a better
tracking. Furthermore, from (20) we can easily observe the frequency-scanning-based tracking scheme.
drawbacks of conventional zooming-based method (we only First, let’s provide some analysis about the backward map-
consider interval θ0 ∈ [0, 1] here): ping similarly to (17). The peripheral backward mappings
f−M → θM and fM → θ−M can be established as
• Tracking fails when θ0 → 1: Actually, the upper bound 
1 fc M fd
P can only be approached when central angle θ0 = 0.  θ0 − α= ψ bw + tbw


When UE moves to certain specific directions that θ0 ≈ 1, fc + M fd fc + M fd , (22)
the maximum searching range α declines to P1 − Mf d fc M fd
fc ,  θ0 + α= ψ bw − tbw


which may be quite small and consume unaffordable time fc − M fd fc − M fd
resource for tracking. For example, when we set P = 16, and based on (22), the configuration {ψ, t} is solved as
fc = 100 GHz, B = 12.5 GHz and AoD 72◦ (θ0 = 0.95), 
the corresponding searching radius α|θ0 =0.95 ≈ 3×10−3. M fd
 ψ bw = θ0 − α


Almost 34 pilot timeslots are needed to thoroughly scan fc . (23)
the angular interval [0.9, 1]. fc
 tbw = θ0 − α


• Angle-dependent range α: On the other hand, we cannot M fd
find a constant scanning range for all directions, because Substituting (23) into (19), the maximum angular searching
α is angle-dependent here and the minimum searching range is enlarged to
radius P1 − Mf d
fc is too small for pilot consumption. The 1 M fd
adaptive adjustment for α is redundant and hardware- αbw < + θ0 . (24)
P fc
intricate during tracking process.
• Power leakage problem: The zooming-based tracking Compared with the forward-pairing range αfw , we can easily
[27] still only searched part quantized angular values, observe that as central angle θ0 increase, forward range αfw
which causes beam mismatch since there may exist biases drop off while backward range αbw rise up. And they intersect
 6

Theta=0.8, alpha=0.08

The same searching range:

[-0.72,0.88]

beam gain

beam gain
Strong interference:
Cause tracking failure

angular value angular value

(a) Forward Pairing (b) Forward-Backward Pairing

Fig. 4. The beam pattern comparison of conventional forward mapping and forward-backward pairing schemes with overall 128 subcarriers, where the central
angle θ0 = 0.8 and searching radius α = 0.08.

at θ0 = 0, which inspires us to retain the maximum value process is provided in Algorithm 1.

inside αbw and αfw as
 Algorithm 1 Proposed Wideband THz Tracking: Forward-
 αfw , θ0 < 0

Backward-Pairing Tracking
b-f bw fw
α = max{α , α } =
 αbw , θ0 ≥ 0 Input: Central direction for searching θ0 ; Searching radius α;
(25)

Pilot overhead length L; Transceiver configuration NBS ,
1 M fd NTTD , P ; Frequency info fc , fd and M .
< + |θ0 |.
P fc Output: Physical beam angle θ̂
1: for time fraction index l = 1 to L do
When tracking UEs with angle range [θ0 − α, θ0 + α], to 2: fraction info θ0,l ← θ0 − α + (2l−1)α ; αl ← α/L
L
pursue lower pilot overhead for UE tracking and approach the 3: if θ0,l ≥ 0 then
bound ᾱb,f = P1 + Mffc |θ0 |, we divide it into two categories
d
4: Backward pairing: calculate ψl , tl via (23)
θ0 < 0 and θ0 ≥ 0. If the central angle θ0 is a negative 5: else
number, we adopt the forward mapping fm → θm and then 6: Forward pairing: calculate ψl , tl via (18)
the corresponding tracking process is the same with [27], as 7: end if
formulated in (20). On the contrary, if θ0 ≥ 0, we utilize the 8: calculate precoding matrix fl,m (ψl , tl ) via (11)
backward mapping fm → θ−m instead of forward one. 9: obtain received signals yl = [yl,−M , . . . , yl,M ]T via
Generally speaking, we now assume there exist L timeslots (26)
for tracking, and only uniformly-quantized angular fraction 10: end for
[θ0 −α+ 2α 2α
L (l−1), θ0 −α+ L l] is tracked in the l-th timeslot.
11: pick the strongest beam and corresponding subcarrier/time
After we set up the subcarrier-angle mapping configuration fraction index {l̂, m̂} from all received signals Y via (27)
(25), i.e., the parameters ψl and tl , the signal processing for
enhanced TTD-aided tracking is then shown as follows. The 12: calculate the final physical beam angle θ̂ via (28)
received signal at l-th frame is formulated as
yl,m = hH b-f b-f
m fl,m (ψl , tl )sl,m + nl,m . (26)
Without loss of generality, the pilot sequences sl,m of all C. Power Leakage Elimination
L timeslots are fixed as 1. Then the strongest beam and Although the forward-backward-pairing tracking can signif-
corresponding timeslot index triple {m, l} can be selected icantly reduce pilot overhead consumption, there still exists
from overall received signals Y ∈ CL×(2M+1) : power leakage problem since subcarriers for direction mapping
are finite. When the real physical direction θreal don’t match the
{m̂, l̂} = arg max k[Y]l,m k2 . (27) quantized searching grid θ̂ ∈ {θl,m } exactly, power leakage
l=1,...,L, m=−M,...,M
effect then appears. Under certain extreme situations such
The final beam angle can be calculated as as when massive antennas are configured, the beamforming
fc fb mainlobe amplitude is sharply reduced and side-lobe inter-
θ̂ = ψl̂ + m̂ tl̂ , (28) ference increases, which is quite harmful to communication
fm̂ fm̂
capacity and beamforming gain. To tackle this problem, based
where ψl̂ , tl̂ are controlled by l̂-th time fraction’s central angle on our proposed TTD-aided tracking above, we further provide
θ0l̂ , θ0 − α + (2l̂−1)α
L and tracking radius L α
. The detailed a CPR-based scheme for power leakage compensation.
 7

Based on the channel model (8) and only considering the After several iterations we can obtain the accurate beam
dominant LoS path, the initial objective can be established as direction to further overcome power leakage problem. The
follows, detailed process is provided in Algorithm 2 as follows.
L M
X X
H 2 Algorithm 2 Proposed Power Leakage Compensation for
θ̂P = arg min yl,m − fl,m hm (θ)
θ
l=1 m=−M
Wideband THz frequency-scanning-based Tracking
(29)
L
X M
X 2
Input: Received signals yl,m , Precoding matrix fl,m and
H
= arg min yl,m − gm fl,m am (θ) initial tracking result θ̂(0) obtained from Alg 1; Maximum
θ
l=1 m=−M iteration T ; Update threshold ξ.
where gm denotes the complex gains for LoS path at frequency Output: Accurate physical direction θ̂
fm . Notice that the path gain’s amplitude |gm | is invariable 1: Calculate the reshaped signal vector ŷm and precoders
under all subcarriers due to the only LoS path, which means Bm in (30)
we can divide it into two separate parts gm = gejτm . Besides, 2: for iteration num t = 0 to T do
the path gain’s phase τm is irrelevant to timeslot fraction index 3: Update the common path gain amplitude g (t+1) among
l, and thus we can merge all timeslots and further reformulate subcarriers and timeslots via (32)
(t+1)
the objective (29) into 4: Update frequency-dependant beam gain phase τm
M
via (33)
X 2 5: Compensate power leakage and update the real beam
min ǫ= ŷm − gejτm BH
m am (θ) , (30)
θ,g,τm
m=−M
direction θ̂(t+1) via (35)
2
T 6: if [g (t+1) , θ̂(t+1) ] − [g (t) , θ̂(t) ] < ξ then
where ŷm = [y1,m , y2,m , . . . , yL,m ] and Bm =
[f1,m , f2,m , . . . , fL,m ] denote received signals and joint beam- 7: Break;
forming matrix along all timeslots L for the m-th subcarrier, 8: end if
respectively. 9: end for
10: Final beam direction θ̂ ← θ̂(T )
It is difficult to directly calculate the global optimum for
problem (30) due to its non-convexity. Therefore, we adopt al-
Furthermore, in some low-SNR scenarios, the signal-based
ternating minimization method searching for the near-optimal
tracking isn’t always completely reliable while prior infor-
solution. Take the result from Algorithm 1 as an initial value,
mation on UE trajectory is more helpful for beam tracking.
marked as 0-th iteration’s result θ̂(0) . During each iteration,
Certain Gauss-Markov UE-mobility models can be further
we optimize beam gain amplitude g, phase ejτm and beam
adopted here, and extended Kalman filter (EKF) or unscented
direction θ successively.
Kalman filter (UKF) can be similarly utilized based on TTD-
First, we neglect the phase information and only consider aided tracking. The corresponding basic work on EKF/UKF
the modulus to optimize the common gain amplitude g. The tracking can be found in [25], [26]. The combination of TTD-
problem is derived as follows: aided tracking and EKF/UKF is herein omitted since the joint
M
X 2
tracking is simple and quite intuitive.
min |ŷm | − g|BH (t)
m am (θ̂ )| , (31)
g(t+1)
m=−M V. S EARCHING R ADIUS E NHANCEMENT
which can be easily calculated via least square (LS) as In this section, we focus on the two tracking algorithms’
M
X h i performance comparison and take the tracking range α as
H
2|ŷm | · |BH (t)
m am (θ̂ )|
one important criterion, which directly controls the overhead
(t+1) m=−M consumption. Furthermore, one critical question is whether
gopt = (32)
M
X h i the range in (25) is maximum or not. Fortunately the answer
2BH (t) (t)
m am (θ̂ )am (θ̂ )Bm is negative. An exquisite enlarged range with closed form is
m=−M provided in this section 4 .
Next, we optimize the phase parameters separately for all
subcarriers m = −M, . . . , M from (30): A. Forward-mapping tracking [27] (as comparison)
! Based on the above analysis, we have given a upper-
(t+1) am (θ̂(t) )H Bm ŷm
τm,opt = angle (33) bounded searching range (20) and obtained the beam-
|am (θ̂(t) )H Bm ŷm | independent maximum range αfw = P1 − Mf d
fc . Neverthe-
less, another criterion should also be considered in forward-
Finally, based on the updated beam gain information gejτm ,
mapping tracking [27] as comparison for more fairness.
we optimize the beam direction θ̂(t+1) via gradient descent
(t) Consider one-timeslot tracking here for convenience. When
method [30]. With g (t) and τm fixed in (30), the gradient of the period of initial beam ΞNTTD (·) in (15) is larger than
beam direction is calculated as (34), and the update process
is formulated as 4 Without loss of generality, we only consider the spatial beam direction
θ ∈ [0, 1] and neglect the negative interval [−1, 0) in Section V and Section
θ̂(t+1) = θ̂(t) − η∇ǫ . (35) VI, due to its symmetry properties.
θ̂ (t)
 8

M H  
H ∂am (θ)
h (t+1)
(t+1) jτ (t+1) H (t+1) jτm
X
∇ǫ |θ = ŷm − gopt e m Bm am (θ) · ŷm − gopt e Bm +
∂θ
m=−M (34)
H
   
(t+1) (t+1) ∂am (θ) (t+1) (t+1)
H
ŷm − gopt e−jτm Bm · ŷm − gopt ejτm BH a
m m (θ)
∂θ

total searching range 2α, the side lobes of ΞNTTD (·) at over- correctly via the mainlobe group, which means the searching
all subcarriers are all located outside the searching interval radius α can be further enlarged from αf-b (25). An intuitive
[θ0 − α, θ0 + α], which also means there exists no interference example is shown in Fig. 5.b, where the central angle is set
inside searching interval. Under this criteria, the constraint can as 0.8 and the coverage of mainlobe envelope is full and
(2)
be formulated as Tm > 2α, ∀m = −M, . . . , M , and then acceptable.
we have
fc 1 fc The next question is what the accurate upper bound for the
α′f-ind < min = . (36) tracking range is. Consider the marginal situation, i.e., when
fm P fM P
the sidelobe group touches mainlobe envelope, as depicted in
When the former bound (20) is smaller than (36), i.e., Side
fc Fig. 5(c). It is worth noting that the sidelobe θM at fM is
θ0 > fM P , the tolerable searching radius increases via (36) Side
smaller than θc , and similarly we have θc < θ−M < θ−M .
but is still weaker than P1 . And it is worth noting that this
bound only holds in single-timeslot scenario. When overall Theorem 1: With a large central direction θ0 → 1, the
L > 1 pilot timeslots are adopted with each angular-mapping maximum searching radius αmax in forward-backward-pairing
fraction radius as α/L, there exists inter-fraction interference tracking scheme is bounded by
(IFI) caused by the sidelobes of windowing beam ΞP (·), which 
means corresponding tracking range should be further reduced. 1.620 M fd
αbound = min + θ0 ,
Otherwise, the side lobe is extremely enlarged and harmfully P fc
(37)
M 2 fd2

influences other time fractions’ beam pattern, which causes 2 M fd
+ + θ0 .
the wideband tracking failure. P P (fc2 − M 2 fd2 ) 2fc

B. Forward-Backward-Pairing Tracking (Algorithm 1) Proof: The proof is provided in Appendix B.

Similarly, based on the above basic criteria, we have ob- From Theorem 1, we can observe that the searching range
tained standard searching range as (25), which is always larger can be further extended when with large central angle situa-
than P1 . Next we analyze and provide further enhancement on tion. For some small searching angle θ0 → 0, we suggest the
tracking radius α. tracking range α could be set as (25) and when θ0 → 1 we
Lemma 1: The periodic sidelobes 5 of ΞP (·) in (14) set the range as (37). Furthermore, according to the overall
at subcarriers f−M and fM are located in different sides analysis above, when we consider the fixed searching radius
with respect to vertical line θ = θ0 − Mf fc α. And the all-
d
α (irrelevant to angle θ0 ) for lower hardware/power consump-
frequency joint beam pattern (as shown in Fig. 4(b)) gets tion, we suggest the global-angular-independent range to be
global maximum gain at θ = θ0 − Mf fc α.
d
set as
Proof: The proof is provided in Appendix A. 1
αfixed = . (38)
Side
From Lemma 1 we consider a specific situation, i.e., θ−M = P
Side
θM = θc as shown in Fig. 5(b). The searching range can be Otherwise, quasi-angular-independent searching range can
2fc2 2 b-f be set via certain piecewise functions for instance:
calculated as α = P (f 2 −M 2f 2) > P > α . Corresponding
c d
Mfd Mfd
sidelobe gain GSide Side
−M = GM = Ξ( fc (θ0 − fc α)) is almost

1
|θ0 | < θ1th ,

decreasing as central angle θ0 raises. When θ0 is small, for ,


P


example, setting θ0 = Mf fc α ≈ 0, the sidelobe gain can
d



αquasi-fixed = 1.620 (39)
approach GSide
M ≈ Ξ(0) = 1, which means the mainlobe , θ1th < |θ0 | < θ2th
 P
group cannot effectively cover sidelobe group and cause a


2 M 2 fd2


tracking failure. Therefore, when the direction is small enough, ) , |θ0 | > θ2th


 (+
the maximum range is bounded by (25) exactly. However, as P P f1 fM
central angle θ0 increases, the sidelobe gain fades as shown where θ1th and θ2th denote threshold angles for piecewise func-
in Fig. 2. In this way, the mainlobe’s envelope can absolutely tion, which is related with system equipment such as antennas
cover the sidelobes and then strongest beam can be tracked N , TTD number NTTD and frequency configurations M , fc , B
5 Notice that the definitions of sidelobe and mainlobe here don’t depend on
and so on. Conservatively speaking, we can uniformly divide
(2) the AoD domain at AoD = 30◦ and AoD = 60◦ , i.e., set
the beam amplitudes but the locations in (15). The exact location of θc,m in
(15) is defined as mainlobe while the peak location with periodic shifting is θ1th = 0.5 and θ2th = 0.866 for the quasi-fixed searching radius
defined as sidelobe. in (39).
 9

* !
(a ) .= = ). )-!,澔 (b) .= = ). *!,)澔
" "

Sidelobe group:
Must be covered inside mainlobe

! #! $!% #$%
(c) critical condition = + + ( = ). *+,-澔
" "($!& ' #! $!% ) !$& )

(d) diffusion exists δtracking failureε

Fig. 5. The beam patterns of forward-backward pairing scheme under different searching ranges α.

VI. F INITE -C ODEBOOK - BASED T RACKING easily yielded to

In this section, we give a joint consideration of TTD- ψ BF ∈ UQ

2/NBS ([−1, 1]) , t
BF
∈ UQ
2/NBS ([−1, 1]) , (40)
aided tracking and beamforming. The above analysis is based which denotes uniform quantization for the whole angular
on the infinite resolution of ψ and t in (11). However the domain [−1, 1] with codebook size as NBS (resolution as
beamforming codebook size is always countable in practice, ∆tBF = 2/NBS ). Then at each subcarrier, NBS almost-
which inspires us to evaluate the performance with finite orthogonal beams are generated to provide full communication
beamforming codeword configuration. direction coverage, just like the conventional narrowband DFT-
As for beamforming module, given the DFT-based code- based analog beamforming.
word pattern derived in (11), the quantization for ψ and t are
 10

As for tracking module, in the forward-pairing scheme, sub-

stituting the radius constraint (20) to tfw (18), we can get that -2

tfw = Mf d 1 1 NTTD Mfd

fc P ≥ 1 since P = NBS ≥ fc according to [19, -3
equ.25]. We can observe that the codewords in beamforming
(|tBF | ≤ 1) and tracking (tfw > 1) are incompatible if we -4

adopt the tracking radius upper bound (20), and thus the fixed
-5
radius αfixed,fw should be further reduced from P1 − Mf d
fc . From

NMSE
another aspect, if we utilize the beamforming codebook (40) -6
for forward-pairing track, we should ensure that tfw ≤ 1 and
then get αfw ≤ Mf fc (1 − θ0 ), which means the fixed searching
d -7

radius doesn’t exist at all. -8 Forward-pairing tracking scheme

On the other hand, next we give an analysis of the proposed Proposed Forward-backward-pairing scheme
forward-backward-pairing scheme. We also only consider the -9 Forward-backward + power leakage compensation
Conventional beam searching
angular interval [0, 1] (i.e. the backward-pairing method) and
-10
neglect [−1, 0] part due to its symmetry property. If we still -10 -5 0 5 10 15 20
utilize the above codebook (40) for tracking, we have tbw = SNR
fc
θ0 − Mf d
α ∈ [−1, 1], then the corresponding fixed tracking
fixed Fig. 6. NMSE performance comparison against SNR with pilot overhead
radius α̂ should be modified as number L = 4 and whole searching range α = 0.2.
M fd M fd
α≤ (1 + θ0 ) ⇒ α̂fixed = ≤ αfixed . (41)
fc fc
From (41) we can get that the fixed tracking radius still P = 16 and NRF = K, respectively. The carrier frequency
exists but decreases when initial beamforming codebook (40) is fc = 100 GHz. Baseband semi-bandwidth and subcarrier
is adopted. To enlarge the radius to αfixed or even αquasi-fixed number are set as B = 10 GHz and M = 64. As for THz wire-
with quantized ψ and t, the quantization interval of t should be less channel, due to the quasi-optical signal transmission from
further entended. For example, if we select the fixed maximum massive antennas and high frequency, we only consider LoS
radius αfixed = P1 , the minimum value of t in tracking is path for each UE, i.e., Nkp = 1. The path gains are randomly
calculated as generated with distribution gp ∈ C(1, σp2 ). The initial physical
(0)
fc fixed fc directions of different users θk are randomly generated by
tmin = min θ0 − α =− (42) the distribution following U(−1, 1). In each frame t, the real
θ0 M fd M fd · P (t) (t−1) (t)
physical direction changes as θk = θk + ∆ζk , where we
Similarly the maximum value is tmax = Mffdc ·P assume the whole system has already known the last direction
As far as the quantization interval is designed, we should θk
(t−1) (t)
at t-th frame, and ∆ζk denotes the statistical mobil-
further optimize the resolution for t correspondingly. Notice ity parameter, which is always bounded by maximum bias
that the quantized step of θ0 is 2αfixed in tracking, and ζmax = 0.2. In another word, at t-th frame, we only need to
fc /M fd αfixed ≫ αfixed . Therefore, the latter term Mf
fc α
d fixed
track the angular interval with overall central angle θk
(t−1)
and
is dominant for resolution design of ψ while the former term (t−1) (t−1)
total angular length 2ζmax , i.e., [θk −ζmax , θk +ζmax ]. In
θ0 is dominant for t in (23). Then we can obtain that for t,
each frame, we configure L timeslots for TTD-based tracking
the resolution ∆t should satisfy that
and in each timeslot we search angular sub-interval with length
2 2ζmax /L. The normalized mean square error (NMSE) and
∆tTrack ≤ 2αfixed = (43)
P beamforming gain here are defined as follows:
and therefore, with joint consideration of TTD-aided beam- K
" #
forming and tracking, the quantization of t can be reformulated 1 X |θ̂ − θr |2
NMSE = E (45)
as follows via (40) and (43) K |θr |2
k=1
  K M
Q fc 1 X X
t ∈ U2/P − , −1 ∪ GBF = |hH
k,m fm (θ̂)|
2
(46)
M fd · P KM
  (44) k=1 m=−M
Q Q fc
U2/NBS ([−1, 1]) ∪ U2/P 1, where θ̂ and θr are the tracking result and real channel beam
M fd · P
direction, respectively.
VII. S IMULATION R ESULTS Fig. 6 shows the average tracking accuracy (NMSE) by
In this section, we provide some simulations to confirm different kinds of THz tracking approaches. The method in
the brilliant performance of our proposed forward-backward [27] is displayed as a benchmark, which only forwardly
tracking scheme and power leakage compensation method. In mapped subcarriers and beam direction. It can be seen that
the simulation, we assume BS is equipped with ULA antennas the NMSE of the proposed forward-backward-pairing scheme
NBS = 256 and only consider K = 4 single-antenna UEs with α = 0.2 becomes lower with the increase of SNR. At
for tracking and beamforming. TTD line number and other high SNR, the leakage error dominates and hampers further
corresponding system parameters are set as NTTD = 16, decline of NMSE to −6 dB. Fortunately, through the additional

-1
-2
-3
-4
NMSE
-5
-6 Forward-pairing tracking scheme
Proposed forward-backward-pairing scheme
-7 Forward-backward + power leakage compensation
Conventional beam searching
-8
-9
0 2 4 6 8 10 12 14
Pilot length
Fig. 7. NMSE performance comparison against pilot overhead length with
SNR 10 dB and whole searching range α = 0.2.
application of leakage compensation method (Algorithm 2),
the tracking error can be significantly solved as shown in Fig.
6. Besides, the conventional beam searching method is also
provided here for comparison, which only sweeps one beam
direction using all subcarriers in one timeslot. In Fig. 6 with
pilot overhead L = 4, without the TTD-aided simultaneous
angular pairing of several subcarriers, the tracking accuracy
and efficiency is deadly reduced to −2.4 dB.
Fig. 7 presents the NMSE performance under different pilot
overhead lengths L. The whole angular searching interval
is fixed as [0.4, 0.8], i.e., θ0 = 0.6 and α = 0.2. This
interval is firstly divided into L fractions and then only
the corresponding fraction is swept in each pilot timeslot
l = 1, . . . , L. The benchmarks here are the same with Fig. 6.
With abundant pilot overhead consumption (for instant L = 16
in Fig. 7), the method in [27] performs quite similarly to
our proposed scheme. However, with low pilot length, our
proposed forward-backward-pairing scheme extremely outper-
forms method in [27] with NMSE gap almost 1.7 dB at pilot
length L = 6. Besides, it is worth noting that only 2 or 3 pilot
timeslots are needed to sweep a large angular interval and
achieve quite low tracking loss through our proposed scheme.
The beamforming gains against pilot length are shown in
Fig. 8(a) and 8(b). All curves here adopt the same precoding
method that exactly align to the tracked beam direction,
while the tracking accuracy varies among different approaches.
Frequency-dependent fully-digital precoder with perfect CSI
is displayed as an upper bound for comparison. SNR is
set as 10 dB and the whole searching range is fixed as
α = 0.2. As shown in Fig. 8, there exists a huge gap
between frequency-scanning-based tracking and conventional
searching-based method [23], which has been acknowledged
and analyzed in previous study [27]. To approach beamform-
ing gain 7.5, only 2 pilot timeslots are needed through our
proposed forward-backward-pairing tracking, while 4 pilots
should be equipped for forward-pairing method. Almost 50%
pilot overhead can be further saved for wideband THz beam
tracking.
11
Next in Fig. 9, we elaborate the beamforming gain under
different UE locations, i.e., at different physical angular di-
rections θr . Interestingly, it can be observed that our proposed
scheme’s performance is symmetric with respect to vertical
line θ0 = 0, while the tracking accuracy of [27] rapidly drops
off when the UE direction θr > −0.2. From the two figures
we can get that the two tracking schemes obtain the same
beamforming gain in the negative angular zone, and when
θr > 0, our proposed scheme still works well but [27] may
carry tracking failure and consequently consumes larger pilot
overhead. As for our proposed scheme, it can be seen that the
worst performance appears when the physical UE direction is
θr = 0. This can be explained by the maximum searching
range in (25) and (37). Under our simulation parameter
16 settings, when UE direction angle is small (|θr | → 0), the
maximum searching range in one timeslot is bounded by
1/P = 0.0625. To fully sweep the whole interval with radius
α
α = 0.2, more than 1/P = 3.2 pilot timeslots are needed.
As shown in Fig. 9(a) and Fig. 9(b), when pilot length is
enlarged from L = 2 to 3, the beamforming gain at direction
θr = 0 is drastically ameliorated. On the other hand, for the
directions aside (|θr | → 1), since the searching range bound
in one timeslot is calculated from (37) as
2 M 2 fd2 M fd
+ + ≈ 0.1502, (47)
P P (fc − M 2 fd2 )
2 2fc
it is sufficient to configure 2 pilot overhead timeslots for
tracking as shown in Fig. 9(a).
VIII. C ONCLUSION
In this paper, based on TTD-aided hybrid precoding struc-
ture, we give an enhanced frequency-scanning-based track-
ing scheme, where we flexibly control the angular coverage
from all squint beams over the whole bandwidth with two
different pairing policies, to simultaneously search multiple
physical directions in one timeslot. Closed-form searching
radius, parameter configuration and interference are theoret-
ically analyzed. Pilot overhead is further reduced and tracking
robustness is also largely improved via our proposed scheme.
Besides, we provide coupled codebook design and codeword
resolution design for TTDs and PSs, with joint consideration
of both beamforming and tracking. Analytical and numerical
results demonstrate the superiority of the new coupled TTD-
aided codebook from both tracking error and beamforming
performance. Overall, the tracking accuracy and beamforming
gain in wideband THz system are largely enhanced via our
proposed codebooks and corresponding algorithms.
A PPENDIX A
P ROOF OF L EMMA 1
Proof: Substituting (15) and (23) into the beam pattern,
the periodic sidelobes’ angular locations of ΞP ( ffmc θ − ψ) at
 12

8 8

7 7

Optimal digital beamforming

6 6
Forward-pairing tracking scheme
Proposed forward-backward-pairing scheme
Beamforming gain

Beamforming gain
5 Conventional beam searching 5 Optimal digital beamforming
Forward-pairing tracking scheme
Proposed forward-backward-pairing scheme
4 4 Conventional beam searching

3 3

2 2

1 1

0 0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Pilot length Pilot length

(a) Path direction θ0 = 0.4 (AoD = 23.6◦ ) (b) Path direction θ0 = 0.8 (AoD = 53.1◦ )

Fig. 8. Beamforming gain comparison (based on tracking results) against path direction, with SNR 10 dB and searching range α = 0.2.

3.5 7

3 6

2.5 5
Beamforming gain

Beamforming gain

2 4

1.5 3

Optimal digital beamforming Optimal digital beamforming

1 2
Forward-pairing tracking scheme Forward-pairing tracking scheme
Proposed forward-backward-pairing scheme Proposed forward-backward-pairing scheme
Conventional beam searching Conventional beam searching
0.5 1

0 0
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
angular value 0 angular value 0

(a) Pilot length L = 2 (b) Pilot length L = 3

Fig. 9. Beamforming gain comparison (based on tracking results) against beam direction, with SNR 10 dB and the searching range α = 0.2.

f−M , fM and the mainlobe’s central location at fc = f0 are ΞNTTD (0) = 1, i.e., the upper bound of normalized Dirichlet
respectively written as function. Then we finish the proof for Lemma 1.




Side
θ−M = θ0 + α − P2ffc1

 A PPENDIX B

Side
θM = θ0 − α + P2ffM
c . (48) P ROOF OF T HEOREM 1


θc = θ0 − Mf

fc α
 d
Proof: The tracking range is limited by two aspects: intra-
Reweight the top two angles with normalized factors fraction interference and inter-fraction interference. We will
f−M
and f−MfM+fM , then we have provide analyses of the two aspects, respectively.
f−M +fM
• Inter-fraction interference:
f−M Side fM Side M fd
θ−M + θM = θ0 − α = θc . Among several timeslots, the interference gain is mainly
f−M + fM f−M + fM fc
(49) controlled by sidelobes of window function ΞP (·), while the
Side Side
and we can further draw the conclusion that θ−M and θM interference angular location is limited by initial beam ΞTTD (·)
are alwaye located in different sides of θc . in (12). To confirm that the (l − 1)-th fraction’s window
Besides, substitute (23) and θ = θc back into (12), and sidelobes couldn’t hold deadly interference on l-th angular
we can calculate the beam gain at fc as G(fc , θc ) = ΞP (0) · fraction’s tracking, the array gains of all subcarriers should
 13

be set larger than the maximum sidelobe gain GSide

win,P in is equivalent to
windowing beam ΞP (·), i.e.,  ′
f Side
 
fM Side

ΞP θM − ψ bw > ΞP θM − ψ bw . (53)
  fc fc
fm
ΞP θm − ψ bw > GSide
win,P , ∀m Side
fc According to Lemma 1, θM and θM are located in different
(1)
(a)

M fd
 sides with respect of θc,M (14), which means θM Side
> ffMc ψ bw .
⇒ ΞP θ0 − α > GSide
win,P Similarly we have θM Side f
< fc′ ψ bw . According to the mono-
fc (50) tonicity and duality of Dirichlet function Ξ(·), inequality (53)
(b) M fd can be rewritten as
⇒ α< Ξ−1 Side
P (Gwin,P ) + θ0
fc f ′ Side fM Side
ψ bw − θM > θ − ψ bw . (54)
(c) 1.620 M fd fc fc M
≈ + θ0 ,
P fc Substituting (23), (48) to (54) yields that
where GSide
win,P is nearly a constant value only related to P since 1 1 4fc fc fc
fm /fc can be approximated to 1 in (15), and (a) is because 2α − M fd ( ′
+ )α − < (2 − ′ − )θ0 . (55)
f fM P fM f fM
the array gains are minimized at peripheral subcarriers fM and
And here (55) is a quadratic inequality for searching radius α,
f−M . (b) is derived due to the monotone increasing property
which means the solution is not that intuitive. Due to the fact
of Dirichlet function ΞP (x) inside [− P2 , 0]. And x = Ξ−1 (a)
f ′ < fc , we approximate the right side of (55) and then yield
denotes the angular location x > 0 in the mainlobe of Ξ(·)
that
with relationship a = Ξ(x). Besides, (c) is approximated due
to the similarity between Dirichlet function ΞP (x) and sinc 1 1 4fc fc
2α − M fd( ′ + )α − < (1 − )θ0 (56)
function sinc( P2 x). f fM P fM fM
Then substituting f ′ (52) to (56), we have
• Intra-fraction interference:
2fc M 2 fd2 2M fd 4fc M fd
For each angular fraction, we only consider intra-fraction (2 − 2 α − f )α < + θ0
P f1 fM M P f M fM
interference inside windowing mainlobe. Notice that in the
window ΞP (·)’s mainlobe there exist two beams correspond- 2fc 4fc 2fc M 2 fd2 M fd (57)
⇒ α< + 2 + θ0
ing to each subcarrier as analyzed in Section III. One is utilized fM P fM P f1 fM fM
here positively for tracking but the other unfortunately de- 2 M 2 fd2 M fd
grades other subcarriers’ tracking performance, just as shown ⇒ α< + + θ0 .
P P f1 fM 2fc
in Fig. 5. We name the harmful beam sets along all subcarriers
(circled in Fig. 5) as sidelobe groups. Similarly, we consider f−M side and can derive the same
closed-form upper bound in (57). With joint consideration of
The marginal location of sidelobe groups (corresponding intra-fraction (57) and inter-fraction interference (50) we can
to maximum sidelobe gains), at subcarriers fM and f−M , finish this proof.
are formulated in (48). First, we only consider one side
fM . Substituting (48) to (16), we rewrite the relationship R EFERENCES
Side
between angular location θM and the corresponding mainlobe [1] H. Sarieddeen, N. Saeed, T. Y. Al-Naffouri, and M.-S. Alouini, “Next
′
frequency f as follows: generation terahertz communications: A rendezvous of sensing, imaging,
and localization,” IEEE Communications Magazine, vol. 58, no. 5,
Side fc bw f ′ − fc bw pp. 69–75, 2020.
θM = ψ + t . (51) [2] Z. Chen, X. Ma, B. Zhang, Y. Zhang, Z. Niu, N. Kuang, W. Chen,
f′ f′ L. Li, and S. Li, “A survey on terahertz communications,” China
Communications, vol. 16, no. 2, pp. 1–35, 2019.
Substituting ψ bw and tbw (23) to (51) and we can solve the [3] C. Lin, G. Y. Li, and L. Wang, “Subarray-based coordinated beamform-
mainlobe frequency f ′ as ing training for mmwave and sub-thz communications,” IEEE Journal
2
on Selected Areas in Communications, vol. 35, no. 9, pp. 2115–2126,
P f−M fM 2017.
f′ = 2Mfd fc
. (52) [4] H. Elayan, O. Amin, B. Shihada, R. M. Shubair, and M.-S. Alouini,
P f−M fM + α “Terahertz band: The last piece of rf spectrum puzzle for communication
systems,” IEEE Open Journal of the Communications Society, vol. 1,
f ′ here denotes corresponding subcarrier that form strongest pp. 1–32, 2020.
Side
beam at location θM , via configuration of ψ bw and tbw in [5] K. Tekbıyık, A. R. Ekti, G. K. Kurt, and A. Görçin, “Terahertz
Side
(23). Due to the fact θM < θc , we can project it back to band communication systems: Challenges, novelties and standardization
efforts,” Physical Communication, vol. 35, p. 100700, 2019.
subcarriers and easily obtain f ′ > fc . [6] K. David and H. Berndt, “6g vision and requirements: Is there any need
It is not difficult to know, when the mainlobe envelope fully for beyond 5g?,” IEEE vehicular technology magazine, vol. 13, no. 3,
pp. 72–80, 2018.
covers sidelobe groups, frequency-scanning-based tracking [7] J. Hoydis, S. ten Brink, and M. Debbah, “Massive mimo in the ul/dl of
can always succeed (regardless of noise). Next, consider the cellular networks: How many antennas do we need?,” IEEE Journal on
critical condition as shown in Fig. 5c, the corresponding gains Selected Areas in Communications, vol. 31, no. 2, pp. 160–171, 2013.
[8] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
should be set as G(f ′ , θM Side Side
) > G(fM , θM ). And according bers of base station antennas,” IEEE Transactions on Wireless Commu-
to the gain domination of windowing beam ΞP (·) in (12), it nications, vol. 9, no. 11, pp. 3590–3600, 2010.
 14

[9] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,

“Spatially sparse precoding in millimeter wave mimo systems,” IEEE
Transactions on Wireless Communications, vol. 13, no. 3, pp. 1499–
1513, 2014.
[10] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath, “Energy-efficient hy-
brid analog and digital precoding for mmwave mimo systems with large
antenna arrays,” IEEE Journal on Selected Areas in Communications,
vol. 34, no. 4, pp. 998–1009, 2016.
[11] M. Wang, F. Gao, N. Shlezinger, M. F. Flanagan, and Y. C. Eldar, “A
block sparsity based estimator for mmwave massive mimo channels with
beam squint,” IEEE Transactions on Signal Processing, vol. 68, pp. 49–
64, 2020.
[12] B. Wang, M. Jian, F. Gao, G. Y. Li, and H. Lin, “Beam squint
and channel estimation for wideband mmwave massive mimo-ofdm
systems,” IEEE Transactions on Signal Processing, vol. 67, no. 23,
pp. 5893–5908, 2019.
[13] B. Wang, F. Gao, S. Jin, H. Lin, and G. Y. Li, “Spatial- and frequency-
wideband effects in millimeter-wave massive mimo systems,” IEEE
Transactions on Signal Processing, vol. 66, no. 13, pp. 3393–3406, 2018.
[14] S. Targonski and D. Pozar, “Minimization of beam squint in microstrip
reflectarrays using an offset feed,” in IEEE Antennas and Propagation
Society International Symposium. 1996 Digest, vol. 2, pp. 1326–1329
vol.2, 1996.
[15] B. Wang, X. Li, F. Gao, and G. Y. Li, “Power leakage elimination for
wideband mmwave massive mimo-ofdm systems: An energy-focusing
window approach,” IEEE Transactions on Signal Processing, vol. 67,
no. 21, pp. 5479–5494, 2019.
[16] F. Gao, B. Wang, C. Xing, J. An, and G. Y. Li, “Wideband beamforming
for hybrid massive mimo terahertz communications,” IEEE Journal on
Selected Areas in Communications, vol. 39, no. 6, pp. 1725–1740, 2021.
[17] Q. Wan, J. Fang, Z. Chen, and H. Li, “Hybrid precoding and combining
for millimeter wave/sub-thz mimo-ofdm systems with beam squint
effects,” IEEE Transactions on Vehicular Technology, vol. 70, no. 8,
pp. 8314–8319, 2021.
[18] Y. Chen, Y. Xiong, D. Chen, T. Jiang, S. X. Ng, and L. Hanzo, “Hybrid
precoding for wideband millimeter wave mimo systems in the face of
beam squint,” IEEE Transactions on Wireless Communications, vol. 20,
no. 3, pp. 1847–1860, 2021.
[19] J. Tan and L. Dai, “Delay-phase precoding for thz massive mimo
with beam split,” in 2019 IEEE Global Communications Conference
(GLOBECOM), pp. 1–6, 2019.
[20] K. Dovelos, M. Matthaiou, H. Q. Ngo, and B. Bellalta, “Channel
estimation and hybrid combining for wideband terahertz massive mimo
systems,” IEEE Journal on Selected Areas in Communications, vol. 39,
no. 6, pp. 1604–1620, 2021.
[21] E. Carrasco, J. A. Encinar, and M. Barba, “Bandwidth improvement
in large reflectarrays by using true-time delay,” IEEE Transactions on
Antennas and Propagation, vol. 56, no. 8, pp. 2496–2503, 2008.
[22] E. Ghaderi, A. Sivadhasan Ramani, A. A. Rahimi, D. Heo, S. Shekhar,
and S. Gupta, “An integrated discrete-time delay-compensating tech-
nique for large-array beamformers,” IEEE Transactions on Circuits and
Systems I: Regular Papers, vol. 66, no. 9, pp. 3296–3306, 2019.
[23] C. Qi, K. Chen, O. A. Dobre, and G. Y. Li, “Hierarchical codebook-
based multiuser beam training for millimeter wave massive mimo,” IEEE
Transactions on Wireless Communications, vol. 19, no. 12, pp. 8142–
8152, 2020.
[24] X. Gao, L. Dai, Y. Zhang, T. Xie, X. Dai, and Z. Wang, “Fast
channel tracking for terahertz beamspace massive mimo systems,” IEEE
Transactions on Vehicular Technology, vol. 66, no. 7, pp. 5689–5696,
2017.
[25] J. Zhao, F. Gao, W. Jia, S. Zhang, S. Jin, and H. Lin, “Angle domain hy-
brid precoding and channel tracking for millimeter wave massive mimo
systems,” IEEE Transactions on Wireless Communications, vol. 16,
no. 10, pp. 6868–6880, 2017.
[26] G. Arredondo, W. Chriss, and E. Walker, “A multipath fading simulator
for mobile radio,” IEEE Transactions on Communications, vol. 21,
no. 11, pp. 1325–1328, 1973.
[27] J. Tan and L. Dai, “Wideband beam tracking in thz massive mimo
systems,” IEEE Journal on Selected Areas in Communications, vol. 39,
no. 6, pp. 1693–1710, 2021.
[28] C. Han and Y. Chen, “Propagation modeling for wireless communica-
tions in the terahertz band,” IEEE Communications Magazine, vol. 56,
no. 6, pp. 96–101, 2018.
[29] R. J. Mailloux, Phased array antenna handbook. Artech house, 2017.
[30] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex optimization.
Cambridge university press, 2004.