Professional Documents
Culture Documents
2 PDF
2 PDF
AA BB
Email: erik.g.larsson@liu.se MM
SINR P
Abstract—Intelligent reflecting surface (IRS) is a promising
solution to build a programmable wireless environment for 1 2 N
future communication systems, in which the reflector elements IRS
IRS
steer the incident signal in fully customizable ways by passive G
1 2 M
beamforming. This work focuses on the downlink of an IRS-
aided multiuser multiple-input single-output (MISO) system. A
practical IRS assumption is considered, in which the incident hd,k hr,k
User 1
signal can only be shifted with discrete phase levels. Then, the
weighted sum-rate of all users is maximized by joint optimizing
the active beamforming at the base-station (BS) and the passive BS User k
with closed-form solutions using the multi-ratio quadratic where the superscript H denotes the conjugate transpose, and
transform [18]. The RC adjustment subproblem is reformu- uk ∼ CN (0, σ02 ) denotes the additive white Gaussian noise
lated as the non-convex quadratically constrained quadratic (AWGN) at the k-th user receiver. The k-th user treats all
program (QCQP), and is further solved via the alternating the signals from other users as interference. Accordingly, the
direction method of multipliers (ADMM) instead of the high- decoding SINR is
complexity semidefinite relaxation (SDR) technique suggested 2
H
by [16]. The main contributions of this work are summarized (hd,k + hHr,k Θ H
G)wk
as follows: γk = ∑ 2 . (2)
K H
i=1,i̸=k (hd,k + hr,k Θ G)wi + σ0
H H 2
• Firstly, this paper investigates the WSR maximization
problem for the multiuser downlink MISO system aided
by IRS. An iterative algorithm with closed-form solutions III. WSR M AXIMIZATION FOR D OWNLINK
is proposed to jointly optimize the active and passive T RANSMISSION
beamforming. In this paper, we aim to maximize the WSR of all the K
• Secondly, we design a low-complexity algorithm for the users by jointly designing the beamforming vectors at the BS
passive beamforming, which is applicable to both the and the RC∑matrix at the IRS, subject to the transmit power
K
k=1 ∥wk ∥ ≤ PT of BS. Mathematically, the
discrete and the continuous phase-shift IRS. constraint 2
• Finally, simulation results have verified the effectiveness WSR maximization problem is thus formulated as
of the proposed algorithm. In particular, the 2-bit phase
shifter may work well with only a small performance ∑
K
(P1) max f1 (W, Θ) = ωk log(1 + γk )
degradation compared with the performance achieved by W,Θ
k=1
the ideal phase shifter assumption.
s.t. θn ∈ FD , ∀n = 1, · · · , N, (3a)
The rest of the paper is organized as follows. Section
II outlines the system model. The joint active and passive ∑
K
∥wk ∥2 ≤ PT , (3b)
beamforming problem is formulated and solved in Section III.
k=1
Simulation results are provided in Section IV, and Section V
concludes the paper. where W = [w1 , w2 , · · · , wK ] ∈ CM ×K , and the constant
weight ωk is used to represent the priority of user k.
II. S YSTEM M ODEL
Despite the conciseness of (P1), it is generally very hard
As shown in Fig. 1, we investigate an IRS-aided multiuser to obtain the optimal solution due to the non-convex objective
MISO communication system, in which one multi-antenna function f1 (W, Θ) and the non-convex constraint sets FD . In
BS serves K single-antenna users. The numbers of transmit this paper, we propose a low-complexity algorithm to solve
antennas at the BS and reflector elements at the IRS are (P1) sub-optimally based on the fractional programming tech-
denoted by M and N . In addition, the baseband equivalent nique [18]. Applying the Lagrangian dual transform proposed
channels from BS to user k, from BS to IRS, and from IRS in [18], (P1) can be equivalently written as
to user k are denoted by hd,k ∈ CM ×1 , G ∈ CN ×M , and
hr,k ∈ CN ×1 , respectively. We assume a quasi-static flat- (P1′ ) max f1a (W, Θ, α)
W,Θ,α
fading model, and the channel state information (CSI) of all
channels involved is perfectly known by the BS and the IRS. s.t. (3a), (3b),
Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on October 28,2020 at 11:42:01 UTC from IEEE Xplore. Restrictions apply.
User
Power allocation
User
SINR
where α refers to [α1 , · · · , αk , · · · , αK ]T , and αk is an where β is the vector of the auxiliary variables [β1 , · · · , βK ]T ,
auxiliary variable
User for the decoding SINR γk ; and the new and the superscript ∗ denotes the conjugate of scalar. Then,
IRS phase adjustment
objective function is defined by β and W can be updated with closed-form solutions.
∑ K ∑ K Lemma 1: The beamforming vector wk is updated by
f1a (W, Θ, α) = ωk log(1 + αk ) − ωk αk ( )−1
k=1 k=1 √ ∑K
◦
wk = α̃k βk λ0 IM + |βi | hi hi
2 H
hk , (8)
∑K
ωk (1 + αk )γk i=1
+ . (4)
1 + γk
k=1 where λ0 is the dual variable introduced for the transmit power
′ Update SINR ◦
alize 0 0In (P1 ), when W and Θ hold fixed, the optimal αk is αk = constraint, which is determined by
γk . Then, for aafixed
k g kα, optimizing W and Θ is reduced to { }
∑K
∑K
α̃ γ
◦
λ0 = min λ0 ≥ 0 : ∥wk ∥ ≤ PT ,
2
(9)
(P1′′ ) max , ( i +k1) , k ( i +1)
(i )
, (i ) , (i ) (i )
W,Θ 1 + γk k=1
k=1
s.t. (3a), (3b), and βk is the fractional-programming auxiliary variable,
which is given by
BSwhere α̃k = ω( ik) ,(1 (+
Transmit i 1)
α, k ().
i)
IRS RC √
To sum up, the original problem (P1) is firstly transformed α̃k hH
k wk
Beamforming Optimization
to (P1′ ), in which the active beamforming at BS and the βk◦ = ∑K . (10)
hH wi 2 + σ 2
i=1 k 0
passive beamforming at the IRS can be decomposed. Then,
we can solve (P1′ ) in an iterative manner by alternatively Proof: βk◦ in (10) and wk◦ in (8) can be obtained by
optimizing α, W and Θ as illustrated in Fig. 2. setting ∂f2a /∂βk and ∂f2a /∂wk to zero, respectively. The
detailed derivations can be found in [18, Section IV].
∑
K
α̃k |(hH
d,k + hr,k Θ G)wk |
H H 2
A. Transmit Beamforming f3u (Θ) = ∑K . (11)
For a given Θ, the composite channel for user k is hk = k=1 i=1 |(hH
d,k + hr,k Θ G)wi | + σ0
H H 2 2
Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on October 28,2020 at 11:42:01 UTC from IEEE Xplore. Restrictions apply.
where the new objective function is Then, (P4) is equivalently represented as
µ
∑
K √ { } (P4a) max f4 (q) − ∥q − θ∥22
f3a (θ, ε) = 2 α̃k Re ε∗k θ H ak,k + εH
k bk,k
θ,q 2
k=1 s.t. q = θ, (20a)
( )
∑K
2
∑
K
θn ∈ FD , ∀n = 1, · · · , N, (20b)
− |εk | |bi,k + θ H ai,k |2 + σ02 , (14)
k=1 i=1 where µ > 0 is the penalty parameter. Then, we have the
Lagrangian of (P4c):
and ε refers to the auxiliary variable vector [ε1 , · · · , εK ]T .
Similarly, we optimize θ and ε alternatively. The optimal ∑
N
µ
εk for a given θ can be obtained by setting ∂f3a /∂εk to zero,
G(q, θ, λR , λI ) = −q H U q − 1F (θn ) − ∥q − θ∥22
2
i.e., { n=1
}
√ ( ) H
α̃k bk,k + θ H ak,k + Re 2q ν + (λR + jλI ) (q − θ) ,
H
ε◦k = ∑K . (15) (21)
i=1 |bi,k + θ ai,k | + σ0
H 2 2
where λR = [λR,1 , · · · , λR,N ]T and λI = [λI,1 , · · · , λI,N ]T
Then the remaining problem is optimizing θ for a given ε. It are the Lagrange variables for Re{q − θ} = 0 and Im{q −
is known that, |bi,k + θ H ai,k |2 in (14) can be further written θ} = 0, respectively, and 1F (·) is the indicator function of
as set F (i.e., 1F (θn ) = 0 if θn ∈ F ; otherwise, equals infinity).
( )( ) The dual problem for (P4a) is formulated as
|bi,k + θ H ai,k |2 = bi,k + θ H ai,k b∗i,k + aH
i,k θ
{ ∗ H } (P5) min L(λR , λI ) = max {G(q, θ, λR , λI )} .
i,k θ + 2Re bi,k θ ai,k + |bi,k | . (16)
= θ H ai,k aH 2
λR ,λI θ,q
Substituting (15) and (16) into (14), and dropping irrelevant The ADMM proposed in [19] aims to find the optimal
constant terms, the optimization problem for θ is represented solution for the dual problem (P5), which has the following
as follows iterative forms:
Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on October 28,2020 at 11:42:01 UTC from IEEE Xplore. Restrictions apply.
Proof: The detailed proof can be referred to [19]. 18
It is worth noting that, the ADMM algorithm needs not
16
to converge to the global or even local optimum. Thus, we
14
need to check whether the solution is better than the initial
point. Also note that, if we relax the non-convex constraint 12
R (bps/Hz)
10
convex, and the ADMM converges to the global optimum.
8
This optimal objective value may serve an upper bound of
6 Upper bound
(P4a) for performance evaluation. Proposed, d=+
4 Proposed, d=4
Proposed, d=2
D. Algorithm Development 2 Baseline 2
Baseline 1
Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on October 28,2020 at 11:42:01 UTC from IEEE Xplore. Restrictions apply.
25
ACKNOWLEDGEMENT
Upper bound
Proposed, d=+
This work was supported in part by the National Natu-
Proposed, d=4 ral Science Foundation of China under Grants U1801261,
Proposed, d=2
20
Baseline 2 61631005, and 61571100.
Baseline 1
R EFERENCES
R (bps/Hz)
Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on October 28,2020 at 11:42:01 UTC from IEEE Xplore. Restrictions apply.