Channel Estimation and Transmission for Intelligent
Reflecting Surface Assisted THz Communications
Boyu Ning∗ , Zhi Chen∗ , Wenrong Chen∗ , and Yiming Du∗
∗ National
Key Laboratory of Science and Technology on Communications
University of Electronic Science and Technology of China (UESTC), Chengdu 611731, China
Emails: boydning@outlook.com; chenzhi@uestc.edu.cn; wenrongchen@std.uestc.edu.cn;
Abstract—Intelligent reflecting surface (IRS) is envisioned as
a promising technology to broaden signal coverage and enhance
transmission in terahertz (THz) communications. Due to the
passivity of IRS, the channel measurement can not be achieved by
traditional pilot manner and the subsequent cooperative trans-
mission design remains an open problem. This paper investigates
the channel estimation and transmission solutions for massive
multiple input multiple output (MIMO) IRS-assisted THz system.
The channel estimation is realized by beam training and the
quantization error is analyzed for evaluating performance. In
addition, a novel hierarchical search codebook design is proposed
as a low-complexity basis of beam training. Based on above foun-
dations, we propose a cooperative channel estimation procedure
to tactfully acquire the channel knowledge. Finally, by leveraging
obtained channel information, the designs of IRS and transceivers
are directly provided in closed form without reconstructing the
full channel matrix or additional optimization. Simulation and
numerical results are presented to illustrate the minimum signal
to noise ratio (SNR) required for beam training and the efficacy
of the proposed transmission solutions.
Index Terms—Terahertz communications, intelligent reflecting
surface, sparse channel estimation, hybrid beamforming, hierar-
chial codebook, multiple input multiple output (MIMO).
I. I NTRODUCTION
As a promising technology to support explosive growth
of data traffic, terahertz (THz) communication emerges as a
key candidate in future generations [1]. Due to the severe
signal attenuation and narrow wave spread in THz frequencies
(0.1T-10T), the number of effective propagation paths is quite
limited and the THz channel is sparse with line-of-sight (LoS)
dominant [2], [3]. Therefore, the transmission is extremely
sensitive to LoS-cut obstacles and the enhancement of signal
coverage for THz communication remains a critical problem.
Recently, an emerging hardware technology for broaden-
ing signal coverage with reduced energy consumption and
low-cost implementation is the so-called intelligent reflecting
surface (IRS) [4]. What makes the IRS attractive compared
to fixed reflector is the possibility to induce a certain phase
shift independently on the incident electromagnetic by con-
trollable meta-material [5], [6]. Compared to conventional
relay schemes, IRS-assisted schemes enhance the transmission
without receiving or transmitting signals, however, by altering
the propagation environments. Hence, with the emergence
of IRS, the cooperative transmission design for IRS-assisted
schemes has attracted much attention and concerns from
researchers. Prior IRS-related works [7]–[12] mainly aim
to maximize the spectral efficiency by optimizing the IRS
and the transmitter beamformer for multi-input single-output
(MISO) systems under perfect channel state information (CSI).
However, these solutions can not be applied to IRS-assisted
THz communication due to the following concerns.
• To compensate severe signal attenuation in THz commu-
nication, massive multiple input multiple output (MIMO)
This work was supported by the National Key R&D Program of China
under Grant 2018YFB1801500.
978-1-7281-5089-5/20/$31.00 ©2020 IEEE
Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.
implementation needs to be considered to provide enough
array gain. The algorithms in [7]–[12] for cooperative
optimization are oriented for single antenna receiver, and
can not be extended to MIMO systems.
• Traditional fully digital beamformer (FDB) requires one
dedicated radio-frequency (RF) chain for each antenna
in the MIMO systems, which results in huge energy
consumption and unaffordable hardware cost. Thus, an
appealing low-complexity hybrid beamforming architec-
ture should be retorted as an alternative [13]–[15].
• Owing to the passivity of IRS (unable to send and receive
signals), the CSI of IRS-assisted system is inaccessible by
traditional channel estimation approaches [16]–[18]. As a
result, the acquisition of the overall channel information
in IRS-assisted systems becomes the mandatory and
primary issue. Meanwhile, the cooperative transmission
solution should be derived based on the form of obtained
channel knowledge.
For above-mentioned reasons, three state-of-the-art technolo-
gies, i.e., THz communication, IRS-assisted transmission, and
massive MIMO techniques are possible and necessary to be
integrated in wireless system for expected orders of magnitude
gain. Thus, the development of such a triad-specific transmis-
sion scheme is imperative and significant.
In this paper, we investigate the channel estimation and
transmission solutions for THz massive MIMO IRS-assisted
system with hybrid beamforming architectures. By exploiting
the sparsity of THz channel and the characteristics of massive
antenna array, a beam training manner is proposed as the basis
for realizing channel measurement, and the beam patterns (or
beam directions) distribution as well as the quantization error
caused by beam training are analyzed and evaluated. To reduce
the complexity of the channel estimation, a novel hierarchical
search codebook design is proposed, in which beams in the
same stage are homogeneous in different directions. Based on
the above foundations, a cooperative estimation procedure is
proposed to measure the channel of IRS-assisted scenarios.
Specifically, the deaf-mute IRS carries out a set of prescribed
operations step by step in estimation phase, which are known
to active transceivers. By leveraging the prescribed prior in-
formation, corresponding channel knowledges can be obtained
by active terminals and sent to IRS controller. Then, the
IRS design is provided to optimally bridge the transmission
links, in which the effective channel is further estimated by
beam training. Without CSI matrix reconstruction, the hybrid
precoder/combiner and IRS designs are straightforwardly pro-
vided in closed form via obtained channel knowledge. Simu-
lations results are presented to illustrate the minimum signal
to noise ratio (SNR) required for beam training and numerical
results indicate that the spectral efficiencies achieved by the
proposed designs significantly outperform non-IRS-assisted
benchmark and approach the rate achieved by optimal fully
digital precoder/combiner with perfect CSI.
 ,QWHOOLJHQWUHIOHFWLQJVXUIDFH,56JURXS THz communication [1], [3]. As shown in Fig. 1, this paper
,56 ,56 ,56. I considers such an indoor scenario that LoS path is blocked
t
,QFLGHQFHSDWK 4 4  4. I and Ni IRSs (≤ NRF ) are installed on surrounding wall to
assist communication by providing tunable strong reflecting
5HIOHFWLQJSDWK
-  ..
component. Since the scattering components (reflected by
- I other surfaces) are negligible, the overall channel H can be
$QDORJ . $QDORJ
SUHFRGHU further expressed as

FRPELQHU

5) 5) 
Ni
FKDLQ -. . FKDLQ
H(f, d) = HRef (f, dl ), (4)



I

l=1
.U
'LJLWDO 'LJLWDO
.T







SUHFRGHU FRPELQHU
/R6
&"
T
. 2& U
. 2& 7" in which {HRef (f, dl )}N l=1 are the strong reflecting channels
i

5) 5)
provided by IRSs. f is the carrier frequency and d =
FKDLQ FKDLQ [d1 , d2 , ..., dNi ] is the vector of transmit distance in corre-



$OLFH %RE
sponding propagation paths. Each IRS can dynamically ad-
&2& 3K\VLFDOREVWDFOH 72& just Nr intelligent elements of the meta-surface to achieve
phase shifts on the reflecting signals by a passive way.
Fig. 1. A massive MIMO IRS-assisted system with the hybrid beamforming Therefore, the received signal vector at the lth IRS can be
architecture at Alice and Bob terminal.
linearly transformed by a diagonal phase-shift matrix Θl =
l l l
II. S YSTEM M ODEL diag(βejθ1 , βejθ2 , · · √ · , βejθNr ), l = 1, ..., Ni , and reflected
A. Massive MIMO with Hybrid Architectures to Bob, where j = −1, {θil }N i=1 ∈ [0, 2π), and β ∈ [0, 1]
r

are imaginary unit, PSs, and amplitude reflection coefficient

As depicted in Fig. 1, we consider a narrowband downlink on the combined incident signal, respectively. Thus, the Ni
massive MIMO system where Alice with Nt antennas intends strong reflecting channels in (4) can be expressed as
to send Ns data streams to Bob with Nu antennas. Since the
implementation of fully digital precoder/combiner suffers from 
Ni 
Ni
high cost and power consumption, we consider hybrid digital HRef (f, dl ) = ηl Gt Gr Nl (f, dN M
l )Θl Ml (f, dl ), (5)
t u
and analog architecture. NRF RF chains and NRF RF chains l=1 l=1
are implemented at Alice and Bob such that Ns ≤ NRF t
 Nt
and Ns ≤ NRF  Nu .
u where ηl is the path-loss compensation factor (see9). Gt and
At Alice with hybrid precoder, a symbol vector s ∈ C Ns is Gr are the transmit and receive antenna gains. Ml is the Alice-
t IRSl channel, and Nl is the IRSl-Bob channel. dM N
l (resp. dl )
first processed by a digital precoder FB ∈ CNRF ×Ns , and then
Nt ×NRFt is the distance between Alice and lth IRS (resp. lth IRS and
up-converted to analog procoder FRF ∈ C . The ana- Bob) and dl = dM N
l + dl . The area of each IRS is limited and
log beamformer is implemented by variable phase shifters and only provide one strong reflecting propagation path. Thus, the
subjected to magnitude constraint, i.e., |FRF (i, j)| = 1, ∀i, j. separate components in (5) can be explicitly written as [2]
Here, we assume that E[ssH ] = INs . Given the normalized
 l  A  l H
power constraint FRF FB 22 = 1 and total transmit power P , Ml (f, dM M R
l ) = a(f, dl )aNr ϕR,M aNt ϕA,M ,
the received signal can be expressed as  l  R  l H (6)
√ N N B
Nl (f, dl ) = a(f, dl )aNu ϕB,N aNr ϕR,N ,
y = P HFRF FB s + n, (1)
where l = 1, ..., Ni , and a(f, d) is the path loss which is
where H ∈ CNu ×Nt represents the overall channel matrix drastically affected by the molecular absorption. Thus, a(f, d)
between Alice and Bob. n ∈ CNu are zero-mean additive consists of the free-spread loss and the molecular absorption
white Gaussian noise, i.e., n ∼ CN (0, σn2 INu ). The received loss as [3]
signal y ∈ CNu uis first processed
 k by an  analog combiner c − 1 τ (f )d
a(f, d) = e 2 , (7)
WRFk
∈ CNu ×NRF such that WRF (i, j) = 1, ∀i, j, and 4πf d
u
then down-converted to digital combiner WB k
∈ CNRF ×Nq ,
which results the final signals in where c stands for the speed of light and τ (f ) is the medium
√ absorption factor. According to the path loss of the far-field
y = P WB H
WRFH
HFRF FB s + WB H
WRFH
n. (2) RIS-assisted beamforming case [6], the cascade path loss of
BS-RISl-userk link is supposed to satisfy
The overall spectral efficiency is given as [15]
 Gt Gr GNr c − 1 τ (f )(dM N
Gt Gr ηa(f, dMl )a(f, dk,l ) = √ 3
N l +dk,l ) .
e 2
R = log2 det INs + P C−1 WB H
WRFH
Hk FRF FB M
8 π f dl dk,lN
 (3) (8)
×FH H H
B FRF Hk WRF WB , where G is the RIS element gain. Thus, the path-loss com-
pensation factor is given as
where C = σn2 WB H
WRFH
WRF WB is the noise covariance. √
We mention that the uplink expression of considered system 2 πf GNr
η= . (9)
is identical to (2) with the roles of the precoder and combiner c
switched under transposed reciprocal channel, i.e., H → HT . In (6), the variable ϕ is the path’s azimuth angles of departure
and arrival (AoD/AoA)1 . In addition, aE Na (φ) is the normalized
B. IRS Assisted THz Channel Model antenna array response vectors at terminal E with Na antenna
The large-scale antenna arrays and high pathloss at THz 1 For explicit ϕ, the subscript A, R, and B represent Alice, IRS, and
frequencies significantly limit scattering. Thus, the signal Bob; and the subscript H, M, and N represent the channel H, M, and N
coverage is LoS dominant and is easily blocked by obstacles in respectively. The superscript l represents the lth IRS.

Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.
 M 
elements. For simplicity of exposition, (6) considers a uniform
linear array (ULA) configuration such as MI

M
1
aNa (φ) = √ [1, ejkda sin(φ) , ..., ejkda (N −1) sin(φ) ]T , (10)
M
M
Na M
M
where k = 2π/λ and λ is the wavelength. da is the antenna  S MI M++ 
S
M
spacing. 
 
$UUD\
3ODQH
III. C HANNEL E STIMATION P ROCEDURE M

In prior works for IRS-assisted systems [7]–[12], perfect

CSI of all channels involved is assumed known at every
terminal. Unfortunately, owing to the passivity of IRSs, the %HDP'LUHFWLRQ

traditional channel estimation approaches [16]–[18] are not &RYHUDJH(GJH'LUHFWLRQ M S

applicable any more to IRS-assisted systems. Thanks to the
natural sparsity of THz channels (see (6)), techniques of Fig. 2. Narrow-beam patterns in different directions with common ρ.
low-complexity beamspace channel estimation, such as beam 7 [ ML

training [19], [20], can be invoked to facilitate the estimation %HDPGLUHFWLRQ
of the multipath components (MPCs) in (6). U
MLL  MLL  &RYHUDJHHGJHGLUHFWLRQ

A. Quantization Error by Beam Training

Given a propagation path with arbitrary continuous angle ϕ∗  
 F F [ VLQ\  VLQ ML
in practice, the discrete AoA/AoD of the codewords’ beams 1D

1D

in channel estimation leads to quantization error. In order

Fig. 3. Normalized ith beam power distribution in different directions.
to measure the quantization error, the characteristics of the
narrow beams are analyzed as follows.
For terminal with Na antennas, the codeword of normalized and
sin[(Na π)/2K]
narrow beam in direction ϕ is simply given as aNa (ϕ) (see ρ= , (15)
(10)). The normalized beam power of aNa (ϕ) in the direction Na sin[π/2K]
ψ is given as where ρ is monotonically increasing with the increase in K
 
  when K ≥ Na . The beam direction of {ϕi }Ki=1 is given as
A (aNa (ϕ), ψ) = aNa (ϕ)H aNa (ψ)
  ⎧ 
  ⎪ 2i − 1
 1  jkda (n−1)[sin(ψ)−sin(ϕ)] 
Na (11) ⎪
⎨ arcsin − 1 , front range (FR)
= e  ≤ 1, K
 Na  ϕi =  . (16)
n=1 ⎪
⎪ 2i − 1
⎩π − arcsin − 1 , back range (BR)
and we denote the beam coverage of aNa (ϕ) as K
CV (aNa (ϕ)) = {ψ | A (aNa (ϕ), ψ) ≥ ρ } (12) Proof: For narrow beam in direction ϕ, the normalized
where ρ is the coverage-edge energy, and the length of beam power in (11) can be further expressed as
 
coverage is named beam width hereafter.  1  Na 
Lemma 1: Narrow beam in direction ϕ is equivalent to that  jkda (n−1)[sin(ψ)−sin(ϕ)] 
A (aNa (ϕ), ψ) =  e 
in direction π − ϕ in term of terminal, i.e.,  Na 
n=1
  
aNa (ϕ) = aNa (π − ϕ). (13)  j Na kda m
j Na kda m
− −j Na kda m

 1 e 2 e 2 e 2

=    
Besides, the beam coverage of aNa (ϕ) is front-back mirror N kda m kda m kda m 
symmetrical about the array plane.  a j
e 2 e 2 −e
j −j 2 
Proof: It is easy to verify that aNa (ϕ) = aNa (π−ϕ). Ac-  
 1 j (Na −1)kda m sin[(Na kda m)/2] 
cording to (11), we get A (aNa (ϕ), ψ) = A (aNa (ϕ), π − ψ), =  e 2  (17)
i.e., CV (aNa (ϕ)) = π − CV (aNa (ϕ)). In general, the angle Na sin[(kda m)/2] 
perpendicular to the plane is defined as 0 and the front range where m = sin(ψ)−sin(ϕ). Thus, for half-wavelength antenna
is [−π/2, π/2]. Thus, the beam coverage is front-back mirror spacing, i.e., da = λ/2, the normalized beam power of
symmetrical about array plane. aNa (ϕi ) is given as
According to Lemma 1, beams with front-back mirror  
symmetry about array plane is isomorphic. For simplicity  sin[ Na π (sin(ψ) − sin(ϕ ))] 
 2 i 
of exposition hereafterin, we use the same notation ϕi for A (aNa (ϕi ), ψ) =   (18)
a pair of front-back symmetrical beams. Considering half-  Na sin[ π2 (sin(ψ) − sin(ϕi ))] 
wavelength antenna spacing, we have following Propositions
and Corollaries for narrow beam design. Define a function T(x) = | sin[ N2a π x]/{Na sin[ π2 x]}| as
Proposition 1: For K (K ≥ Na ) narrow beams covering all shown in Fig. 3, it can be observed that the beam power mono-
directions with common ρ, we use {ϕi }K i=1 to represent the
tonically decreases with the increase in |x| when |x| ≤ 2/Na .
ith beam direction and {ϕi±1 }K
to represent the coverage- Thus, the common ρ yields common |x| for all beams, and
i i=1
edge direction of ith beam. As shown in Fig. 2, despite the for ∀ ith beam, we have
common ρ, each beam yields different beam width. For ∀ ith | sin ϕi+1 − sin ϕi | = | sin ϕi − sin ϕi−1 | = χ(ρ), (19)
beam, we have i i

1 where χ is a constant and subjected to the inverse function

| sin ϕi+1
i − sin ϕi | = | sin ϕi − sin ϕi−1
i |= , (14) of ρ = T(χ). When all the K beams cover all directions as
K

Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.
 SHV S %HVWFDVHV :RUVWFDVHV 5RRW 1RGH /HDI


U . 
§ S ·
 1 D VLQ ¨ ¸
© 1D ¹
Ȧ  Ȧ  Ȧ 

Ȧ  Ȧ  Ȧ  Ȧ  Ȧ  Ȧ  Ȧ  Ȧ  Ȧ 

         
ȦȦȦ ȦȦȦ Ȧ ȦȦ Ȧ ȦȦ ȦȦ Ȧ Ȧ Ȧ Ȧ Ȧ Ȧ Ȧ  Ȧ      


. 1D  




     VLQ M
S S S
1D 1D 1D 1D 1D 1D 6LQJOH%HDP:LGWK 7RWDO%HDP:LGWK Sˈ$R$˖ >  @
  
Fig. 4. Accuracy of measurement versus direction of AoA angle ϕ∗ .
Fig. 5. The structure of 3-tree hierarchical beam sweeping when K = 22.

shown in Fig. 2, we have 2K number of χ uniformly divide

the interval [-1,1] with χ = 1/K, and the condition K ≥ more. As such, the number of total stages is given as {SM ∈
Na ⇒ χ < 1/Na . Thus, ρ monotonically increases with the N+ | logK K
M ≤ SM < logM +1} and the beam in sth stage
increase in K. In Fig. 2, sin ϕ1 , ..., sin ϕK uniformly divide SM −s
covers M narrow beam(s).
the interval [-1,1] and it is easy to obtain the beam direction Let ωns denotes the nth beam candidate in the sth stage.
by combining the feasible domain of the arcsine function. Considering the arbitrary K which may not exactly be
Proposition 2: Given a narrow beam with arbitrary AoA M n , n ∈ N+ , zero vectors need to be added in the bottom
angle ϕ∗ to terminal with Na antennas. On condition that K stage for constructing full tree. A example 3-tree structure of
(K ≥ Na ) narrow beams covering all directions with common such a procedure when K = 22 is demonstrated in Fig. 5.
ρ are used to estimate the power, the best cases are that ϕ∗ lies In each stage, we find and follow the best beam (node) for
on the any one of the beam directions and the worst cases are next-stage search, until the best narrow beam (leaf) is found.
that ϕ∗ lies on the any one of the coverage-edge directions. Considering half-wavelength antenna spacing, the narrow
Both cases are presented in Fig. 4, in which p∗es /p∗ is the beam candidates (leaves) in the bottom stage is exactly given
normalized received energy (see (34) in appendix). as the receiver’s array response vector (10), such as
Corollary 1: Given a narrow beam with arbitrary AoA angle 1
ϕ∗ to terminal with Na antennas. On condition that K (K ≥ ωnS3 = aNa (ϕn ) = √ [1, ejπ sin(ϕn ) , ..., ejπ(N −1) sin(ϕn ) ]T ,
Na ) narrow beams covering all directions with common ρ are N
⎧  a 
used to estimate the power, the normalized quantization error ⎪ 2n − 1
of worst case is given as ⎪
⎨ arcsin − 1 , FR
K
ϕn =   , n = 1, ..., K.
sin[(Na π)/2K] ⎪
⎪ 2n − 1
eworst = 1 − , (20) ⎩π − arcsin − 1 , BR
Na sin[π/2K] K
(22)
and the average normalized quantization error is given as
For wide beam candidates (nodes) design, we first propose the
2−K
K  Na π   following Criterion for arbitrary K.
sin 2 y − 1−K
Criterion 2: To design the wide beam candidates, i.e.,
eaver =1 −    K
 dy M −1
1 − y 2 Na π sin π2 y − 1−K
K
{ωns }Ss=1 ∈ C Na ×1 in a M -tree structure, we expect that
−1+ 
2n−K
 Na π 
 1, ωiSM is a descendant of ωns

K−1  K (ωiSM )H ωns = (23)
sin y − 2n−1−K 0, otherwise
−    2
K
 dy
n=2 2n−2−K 1 − y 2 Na π sin π2 y − 2n−1−K
K holds true for all i when n = 1, ..., SM − 1.
K Based on Criterion 2 for M-tree search, we can write the
1−  Na π   design of wide beam candidates matrix of sth stage, i.e.,
sin y − K−1
−    K2
 dy. (21) Ξs = [ω1s ω2s ... ωM
s
s] for s = 1, ..., SM − 1
1 − y 2 Na π sin π2 y − K−1
K
K−2
K in a more compact form as
Proof: The proof of Proposition 2 and Corollary 2 are [ω1SM ω2SM ... ωK
SM H
] Ξs = LH Ξ = D s , (24)
referred to Appendix A. Based on numerical results, we note
that a proper K should be ≥ 2Na where eaver < 4%. where Ds is an K × M matrix. Define a function as
s

a, a ≤ K
UK [a] = .
B. Hierarchical Sweeping and Reflecting Modes Realization 0, a > K
To find the direction with strongest beam power, hierarchical The ith column of Ds contains element 1 in locations
beam sweeping is an efficient way to reduce search complexity
without loss optimality. M -tree search schemes [16], [17], {UK [(i − 1)M SM −s + n] | n = 1, ..., M SM −s },
[20], [21] realize the beam sweeping stage by stage with and element 0 in other locations. As a result, the wide beam
decreasing beam range. Above studies assumed the number candidate can be projectively given as
K of total narrow beams is exactly K=M S and the number
of stages is given as S ∈ N+ . Here, we propose a search ωns = (LLH )−1 LDS (:, n). (25)
Criterion for arbitrary K. Let Ns = 1 and s = 1, the wide beam in hybrid beamforming
Criterion 1: For K narrow beams covering all direction, architecture is realized as
the defined narrow beam is implemented in the bottom stage.
Each stage up, the beam implemented covers M -fold the range ωns = FRF fBs,n s = FRF fBs,n . (26)

Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.
 %HDP3DWWHUQVE\7UDGLWLRQDO$SSURDFK  
  
      %RE 3KDVH $OLFH
6LOHQW 6LOHQW
     
,56L ,56L
 
$OLFH %RE
 L L
     
M 2-
M 2.

     

2PQL7UDQV 5HFHLYH 6ZHHSLQJ5HWXUQ 6ZHHSLQJ5HWXUQ 2PQL7UDQV 5HFHLYH
     
   ,56L ,56L
6WDJH 6WDJH 6WDJH
M2L - M2L . 3KDVH M2L - M2L .
%HDP3DWWHUQVE\3URSRVHG&ULWHULRQ
  
     
M"L . M!L -
         
$OLFH %RE $OLFH %RE

      

2PQL7UDQVPLW 6ZHHSLQJ5HFHLYH 6ZHHSLQJ5HFHLYH 2PQL7UDQVPLW

     

     

   Fig. 7. Phases of the cooperative channel estimation procedure.
6WDJH 6WDJH 6WDJH
strongest receiving signal. Repeating operation by switching
Fig. 6. Comparison of beam patterns by traditional approach [16], [17] and the manner of Alice and Bob, we obtain ϕ̂lR,M and ϕ̂lR,N .
proposed Criterion 2 with the setting that M = 2, Na = 32, and K = 64.
Phase 2: With the obtained ϕ̂lR,M and ϕ̂lR,N , we fix IRSl to
We state that (26) can be always achieved by using two RF optimally bridge the Alice-IRSl-Bob link by direction mode,
chains, and the solution is given in Proposition 2 of [15]. i.e., Θdir (ϕ̂lR,M , ϕ̂lR,N ). Fix Alice to be in an omni-beam
Remark 1: We mention that the proposed Criterion 2 is transmitting mode and Bob hierarchically sweeps the beam
inspired by prior works [16], [17] which happen to have the to find the best direction. Repeating operation by switching
similar concept but contain following shortages. the manner of Alice and Bob, we obtain ϕ̂lB,N and ϕ̂lA,M .
n∈N+ Above procedure is illustrated in Fig. 7. It is easy to follow
• The narrowest beam number must satisfy K = M .
• The derived narrowest beams are with different coverage- above description for implementation and full algorithms are
edge energy, which may causes direction misalignment. omitted due to the limited pages.
To overcome above shortcomings, we first define the beam
with common coverage-edge energy at the bottom stage (see IV. D ESIGN OF IRS AND H YBRID P RECODER /C OMBINER
(22)). For arbitrary K, Criterion 2 is proposed based on leaves With the obtained estimated angles, the optimal design of
in perspective of graph theory (unlike the principle based on IRS for communication is given as
aNa (ϕn ) in above works). Fig. 6 illustrates the comparison of
beam patterns by different approaches. Θl = Θdir (ϕ̂lR,M , ϕ̂lR,N ), l = 1, ..., Ni .
(29)
For purpose of matching the reflecting state to specific IRS  
design, we provide the following proposition. By training beams
 on above angles, i.e., x = aR l
Nr ϕ̂B,M ,
Proposition 3: Given an incoming narrow beam aNr (ϕin ) w = aU l
Nu ϕ̂U,N , we can directly obtain the effective com-
with direction ϕin to IRS, to achieve a reflecting narrow beam posite loss
in return path, the phase-shift matrix of return mode is given
as al ≈ ηl Gt Gr a(f, dM N
l )a(f, dl ), l = 1, ..., Ni ,

Θret (ϕin ) = diag(βejθ1 , βejθ2 , · · · , βejθNr ), on Alice-IRSl-Bob link by calculating the energy of received
(27) beam. Next, a key lemma is introduced to provide the basis
θn = −2kda (n − 1) sin ϕin , n = 1, ..., Nr .
of the derivations that follows.
To achieve a reflecting narrow beam in direction ϕout , the Lemma 2 ([14]): For a ULA system with azimuth angles of
phase-shift matrix of direction mode is given as arrival or departure drawn independently from a continuous
distribution, the transmit and receive array response vectors
Θdir (ϕin , ϕout ) = diag(βejθ1 , βejθ2 , · · · , βejθNr ), (28) are orthogonal, i.e., we have a(φk )⊥span({a(φl )|∀l = k}) as
θn = kda (n − 1)(sin ϕout − sin ϕin ), n = 1, ..., Nr . the number of antenna elements, N , tends to infinity and the
number of paths in the channel is L = o(N ).
Proof: The proof are referred to Appendix B, and above Here, we straightforwardly provide the design of hybrid
defined modes are used in the description below. precoder and combiner as
 1   
Ni
C. Cooperative Channel Estimation Procedure FRF = [aA A
Nt ϕ̂A,M , ..., aNt ϕA,M , 0, ..., 0],
The main parameters of the MPCs in (6) include four  1   
Ni
angles ϕlA,M , ϕlR,M , ϕlR,N , ϕlB,N . Owing to the grid effects WRF = [aB B
Nu ϕB,N , ..., aNu ϕB,N , 0, ..., 0], (30)
of channel estimation, we use ϕ̂ to represent the estimated  
FB = diag( S1 , ..., SNi , 0, ..., 0), WB = INRF u ,
direction. To obtain the four angles, two phases are developed
to achieve measurements as following. where {Si }N
i=1 is the power allocation factors. According to
i
Phase 1: Keep Bob silent and fix Alice to be concurrently Lemma 2, the transceivers design (30) creates a set of parallel
in an omni-beam transmitting and receiving mode. Then, IRSl independent reflecting channels and the spectral efficiency is
successively sweeps the narrow beam in return mode, i.e., degraded as
changing Θret (ϕ), of which the sequence is predefined on 
time slots and known to all terminals. The best return direction P
is informed to Alice by determining the time slot with the R = log2 det INi + 2 FH diag(a 2
, ..., a 2
)F B . (31)
σn B 1 Ni

Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.
 0.8 50
0.7
Misalignment Probability

45
0.6

Spectral Efficiency [bps/Hz]

Na=32,K=2Na
Na=32,K=3Na
0.5
Na=64,K=2Na 40
Na=64,K=3Na
0.4
Na=128,K=2Na
0.3 Na=128,K=3Na 35 Upperbound with IRS-assisted
Perfect CSI: Proposed Design
0.2 Estimated CSI: Proposed Design
30 Upperbound w/o IRS-assisted
0.1

0 25
-10 -8 -6 -4 -2 0 2 4
Signal to Noise Ratio [dB]
20
-5 -4 -3 -2 -1 0 1 2 3 4 5
Fig. 8. Misalignment probability on bottom stage versus signal to noise ratio. Power [dBm]

,56 ,56 ,56

Fig. 10. Spectral efficiency versus transmit power, where the terminal
elements number Nt = Ni = Nu = 32 and K = 2Na .
-
. 60

- .
P 55
.

Spectral Efficiency [bps/Hz]

-
50

Upperbound with IRS-assisted

  45 Perfect CSI: Proposed Design
$OLFH %RE Estimated CSI: Proposed Design
 40
Fig. 9. The simulated IRS-assisted MIMO communication scenario.
The remaining objective of Alice is to determine the power
factor on these subchannels for maximizing total efficiency,
i.e,

Ni
max log
Sl
l=1
Ni
s.t.
l=1
Problem (32) admits a convex form of classic water-filling
problem and the optimal solution can be obtained by resorting
the Karush-Kuhn-Tucker method, which is given by
 +
1 σn2
Sl = − , l = 1, ..., Ni , (33)
ln 2 · μ P a2l
+
where · = max{·, 0}. The Lagrange parameter Ni μ > 0 is
chosen by 1-D search to satisfy the constraint l=1 Sl = 1.
V. S IMULATION AND N UMERICAL R ESULTS
In this section, we first provide simulation results to illus-
trate the performance of the proposed training codebook for
channel estimation. The results were averaged over 10,000
random channel realizations. Considering a ULA with half-
wavelength antenna spacing, Fig. 8 plots the misalignment
probability (MP) on bottom stage versus SNR with different
implementation parameters. As we can see, the minimum
SNR that satisfies zero MP is related to the implementation.
It decreases with the increase of antennas number Na and
increases with the increase of narrow beam number K.
Then, numerical results are presented to validate the ef-
fectiveness of proposed cooperative designs for IRS-assisted
massive MIMO system. We consider such a indoor scenario
as illustrated in Fig. 9, in which Alice and Bob are randomly
located at (0, dA ∈ [0, 5]) and (0, dB ∈ [5, 10]). Three
IRS are located at (5, 4), (5, 5), and (5, 6). The results were
averaged over 10,000 random location settings of transceivers.
In each result that follows, the operating frequency is set
as 0.3THz and the background noise power at the receiver
is σn2 = −80dBm. The antenna spacing for all terminals
is da = λ/2 and the IRS reflection coefficient is β = 1.
Absorption coefficient τ (0.3T ) = 0.0033/m and antenna
Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.
Upperbound w/o IRS-assisted
P
35
30
25
-5 -4 -3 -2 -1 0 1 2 3 4 5
Power [dBm]
Fig. 11. Spectral efficiency versus transmit power, where the terminal
P elements number Nt = Ni = Nu = 64 and K = 3Na .
1 + 2 a2l Sl
σn (32) gains Gt = Gr = 18dBi for Na = 32 (Gt = Gr = 21dBi
for Na = 64). In the following, we respectively plot the
Sl = 1, {Sl }N
l=1 ≥ 0.
i
spectral efficiency of the receivers in different settings of
antenna number and narrow beam number. For comparison,
we also plot the upper-bound rate of IRS-assisted scheme
and the upper-bound rate of non-IRS-assisted scheme, as two
benchmarks. The former is realized by the optimal FDB with
the optimal IRSs designs under perfect CSI. The latter one is
realized by the optimal FDB with random IRSs setting.
Fig. 10 plots the spectral efficiency versus transmit power,
in which Nt = Ni = Nu = 32 and K = 2Na . It is
seen that the performance gain of the IRS-assisted schemes
is significant compared to the non-IRS-assisted benchmark,
since the LoS link is blocked and the reflecting links provide
additional strong propagation paths. Moreover, the spectral
efficiency achieved by our proposed design with perfect CSI
is close to the upperbound, which validates the effectiveness
of our proposed solutions. The gap between the performance
with perfect CSI and that with estimated CSI is due to the
AoAs/AoDs quantization error.
Fig. 11 plots the spectral efficiency versus transmit power, in
which Nt = Ni = Nu = 64 and K = 3Na . Compared to Fig.
10, the performance gain of the IRS-assisted schemes in Fig.
11 increases due to the increase of terminal elements, and the
proposed design with perfect CSI is closer to upperbound. In
addition, with the increased estimation ratio K/Na (increased
quantization resolution), the performance gap resulted from
AoAs/AoDs quantization error is decreasing. Thus, with large-
scale antenna array implementation and adequate quantiza-
tion resolution, our proposed channel estimation procedure
and cooperative transmission solutions exhibit a comparable
performance in the IRS-assisted MIMO system.
VI. C ONCLUSION
We considered the channel estimation and transmission de-
sign in THz massive MIMO IRS-assisted system. Specifically,
we invoked the beam training manner to facilitate the channel
 measurement. Analysis of beam distribution as well as the expect that the reflecting signal achieves narrow beam in
quantization error by beam training are provided. To reduce direction ϕout , i.e.,
the complexity of estimation, a hierarchical codebook design
is provided as the basis of beam training. Finally, a cooperative aNr (ϕout )
channel estimation procedure is proposed to achieve the mea- 1
surement of IRS-assisted system. Subsequently, the IRS and = √ [1, ejkda sin ϕout , ..., ejkda (Nr −1) sin ϕout ]
Nr
hybrid precoder/combiner designs are given in closed form. = ΘaNr (ϕin ) (38)
Numerical results showed significant performance improve-
ment of proposed scheme as compared to non-IRS-assisted 1  jθ1 j[kda sin ϕin +θ1 ] 
=√ e ,e , ..., ej[kda (Nr −1) sin ϕin +θNr ] .
one and the achieved spectral efficiency can approach the Nr
upperbound benchmark.
Thus, the solution can be derived from the equation as
kda (n − 1) sin ϕin + θn = jkda (n − 1) sin ϕout
A PPENDIX A (39)
P ROOF OF P ROPOSITION 2 ⇒ θn = kda (n − 1)(sin ϕout − sin ϕin ), ∀n.
Let ϕout = π + ϕin , we reach the solution of return mode as
The quantization error is existing due to the fact that the
intended power p∗ is not fully estimated by alignment beam, θn = −2kda (n − 1) sin ϕin , n = 1, ..., Nr . (40)
but estimated as

p∗es = max p∗ A (aNa (ϕ∗ ), ϕn ) . (34)

n=1,...,K R EFERENCES
[1] I. F. Akyildiz, J. M. Jornet, and C. Han, “Terahertz band: Next frontier
By Propositon 1, the worst quantization error is given as (20), for wireless communications,” Physical Communication (Elsevier), vol.
where we left out the modulus operator because the right- 12, pp. 16-32, Sep. 2014.
hand side is always positive when K ≥ Na . Considering the [2] C. Lin and G. Y. Li, “Indoor terahertz communications: How many
arbitrary AoA angle ϕ∗ follows a uniform distribution ϕ∗ ∼ antenna arrays are needed?” IEEE Trans. Wireless Commun., vol. 14,
no. 6, pp. 3097-3107, Jun. 2015.
U [−π/2, 3π/2], the probability density function (PDF) of it [3] J. M. Jornet and I. F. Akyildiz, “Channel modeling and capacity analysis
is given as for electromagnetic wireless nanonetworks in the Terahertz band,” IEEE
Trans. Wireless Commun., vol. 10, no. 10, pp. 3211-3221, Oct. 2011.
[4] L. Zhang et al., “Space-time-coding digital metasurfaces,” Nat. Commun.,
1
fΦ (ϕ∗ ) = , ϕ∗ ∈ [−π/2, 3π]. (35) vol. 9, pp. 1-11, Oct. 2018.
[5] T. J. Cui, M. Q. Qi, X. Wan, J. Zhao, and Q. Cheng, “Coding metama-
2π terials, digital metamaterials and programmable metamaterials,” Light:
Science & Applications, vol. 3, no. 10, pp. e218, Oct. 2014.
The sine function y = sin ϕ is monotonic at the interval [6] W. Tang, M. Z. Chen, X. Chen, J. Y. Dai, Y. Han, M. Di Renzo, Y.
[−π/2, π/2] and [−π/2, 3π/2]. The inverse function can be Zeng, S. Jin, Q. Cheng, and T. J. Cui, “Wireless communications with
written as ϕ1 (y) = arcsin(y) and ϕ2 (y) = π − arcsin(y). reconfigurable intelligent surface: Path loss modeling and experimental
measurement,” arXiv:1911.05326, 2019.
Thus, the PDF of y = sin ϕ is given as [7] Q. Wu and R. Zhang,“Intelligent Reflecting Surface-Enhanced Wireless
Network via Joint Active and Passive Beamforming,” IEEE Trans. Wire-
fsin Φ (y) = fΦ [ϕ1 (y)]|ϕ1 (y)| + fΦ [ϕ2 (y)]|ϕ2 (y)|. (36) less Commun., DOI 10.1109/TWC.2019.2936025, 2019.
[8] C. Huang et al., “Achievable rate maximization by passive intelligent
mirrors,” in Proc. IEEE ICASSP, pp. 3714-3718, 2018.
The center (resp. coverage edge) angle of the beams 1 to K [9] C. Huang et al., “Reconfigurable intelligent surfaces for energy efficiency
mapped to sin ϕ is given as (2n − 1 − K)/K (resp. (2n − 2 − in wireless communication,” IEEE Trans. Wireless Commun., vol. 18, no.
8, pp. 4157-4170, Aug 2019.
K)/K, (2n − K)/K) for n = 1, ..., K. Considering the PDF [10] Q.-U.-A. Nadeem et al.,“Large intelligent surface assisted MIMO com-
interval y ∈ (−1, 1) in (36), the average accuracy is given on munications,” arXiv:1903.08127, 2019.
three pieces of integral by expectation formula as [11] H. Guo et al., “Weighted sum-rate optimization for intelligent reflecting
surface enhanced wireless networks,” preprint arXiv:1905.07920, 2019.
[12] B. Ning et al., “Intelligent Reflecting Surface Design for MIMO System
2−K

K  Na π   by Maximizing Sum-Path-Gains,” preprint arXiv:1909.07282, 2019.
p∗es sin 2 y − 1−K [13] A. Alkhateeb, G. Leus, and R. W. Heath, “Limited feedback hybrid
= fsin Φ (y)  K  dy precoding for multi-user millimeter wave systems,” IEEE Trans. Wireless
p∗ Na sin π
2 y − 1−K
K
Commun., vol. 14, no. 11, pp. 6481-6494, Nov. 2015.
−1+ [14] O. El Ayach, R. W. Heath, Jr., S. Abu-Surra, S. Rajagopal, and Z.
Pi, “The capacity optimality of beam steering in large millimeter wave
2n−K
  Na π   MIMO systems,” in Proc. IEEE SPAWC, pp. 100-104, 2002.

K−1 K
sin 2 y − 2n−1−K [15] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design
+ fsin Φ (y)  K  dy for large-scale antenna srrays” IEEE J. Sel. Topics Signal Process., vol.
n=2 2n−2−K
Na sin π
2 y − 2n−1−K
K
10, no. 3, pp. 501-513, Apr. 2016.
[16] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., “Channel
K
estimation and hybrid precoding for millimeter wave cellular systems,”
1−  Na π   IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831-846, Oct. 2014.
sin 2 y − K−1 [17] M. E. Eltayeb, A. Alkhateeb, R. W. Heath, Jr., and T. Y. Al-Naffouri,
+ fsin Φ (y)  K  dy (37) “Opportunistic beam training with hybrid analog/digital codebooks for
Na sin π
2 y − K−1
K
mmWave systems,” in Proc. IEEE Global Conf. Signal Inf. Process.
K−2 (GlobalSIP), Orlando, FL, USA, pp. 315-319, Dec. 2015.
K
[18] Y. Peng, Y. Li, and P.Wang, “An enhanced channel estimation method for
millimeter wave systems with massive antenna arrays,” IEEE Commun.
Substituting (36) into (37), the average quantization error can Lett., vol. 19, no. 9, pp. 1592-1595, Sep. 2015.
be written as (21), which completes the overall proof. [19] Z. Xiao, L. Bai, and J. Choi, “Iterative joint beamforming training with
constant-amplitude phased arrays in millimeter-wave communications,”
IEEE Commun. Lett., vol. 18, no. 5, pp. 829-832, May 2014.
[20] Z. Xiao, T. He, P. Xia, and X.-G. Xia, “Hierarchical codebook design for
A PPENDIX B beamforming training in millimeter-wave communication,” IEEE Trans.
Wireless Commun., vol. 15, no. 5, pp. 3380-3392, May 2016.
P ROOF OF P ROPOSITION 4 [21] S. Noh, M. D. Zoltowski, and D. J. Love, “Multi-resolution codebook
and adaptive beamforming sequence design for millimeter wave beam
We first derive the IRS solution for direction mode. Given alignment,” IEEE Trans. Wireless Commun., vol. 16, no. 9, pp. 5689-
an incoming narrow beam aNr (ϕin ) with direction ϕin , we 5701, Sep. 2017.

Authorized licensed use limited to: POLO BIBLIOTECARIO DI INGEGNERIA. Downloaded on February 15,2023 at 14:53:34 UTC from IEEE Xplore. Restrictions apply.