Professional Documents
Culture Documents
Active RIS or Passive RIS
Active RIS or Passive RIS
ir
https://www.tarjomano.com https://www.tarjomano.com
Abstract— As a revolutionary paradigm for controlling wireless Index Terms— Reconfigurable intelligent surface (RIS), beam-
channels, reconfigurable intelligent surfaces (RISs) have emerged forming, active RIS, signal model.
as a candidate technology for future 6G networks. However,
due to the “multiplicative fading” effect, the existing passive
RISs only achieve limited capacity gains in many scenarios with
I. I NTRODUCTION
strong direct links. In this paper, the concept of active RISs is
proposed to overcome this fundamental limitation. Unlike passive
RISs that reflect signals without amplification, active RISs can
A S WIRELESS communications have advanced from the
first generation (1G) to 5G, the system capacity has
been significantly increased by improving the transceiver
amplify the reflected signals via amplifiers integrated into their designs, while the wireless channel has been considered
elements. To characterize the signal amplification and incorporate
the noise introduced by the active components, we develop and to be uncontrollable. Recently, due to the advances in
verify the signal model of active RISs through the experimental meta-materials, reconfigurable intelligent surfaces (RISs)
measurements based on a fabricated active RIS element. Based have been proposed [2], [3], [4] for the purpose of
on the verified signal model, we further analyze the asymptotic intelligently controlling wireless channels to achieve improved
performance of active RISs to reveal the substantial capacity gain communication performance. Specifically, an RIS is an array
they provide for wireless communications. Finally, we formulate
the sum-rate maximization problem for an active RIS aided composed of a very large number of passive elements
multi-user multiple-input single-output (MU-MISO) system and that reflects electromagnetic signals in a desired manner
a joint transmit beamforming and reflect precoding scheme is so as to reconfigure the propagation properties of wireless
proposed to solve this problem. Simulation results show that, in a environment [5]. Thanks to their high array gain, low cost,
typical wireless system, passive RISs can realize only a limited low power, and negligible noise [5], [6], [7], RISs promise to
sum-rate gain of 22%, while active RISs can achieve a significant
sum-rate gain of 130%, thus overcoming the “multiplicative improve channel capacity [8], extend coverage [9], and save
fading” effect. power [10] in future 6G networks. Additionally, RISs are also
projected to have other applications such as in WiFi [11],
Manuscript received 6 September 2022; revised 19 November 2022; precision measurement [12], and navigation [13].
accepted 19 December 2022. Date of publication 23 December 2022; date
of current version 17 March 2023. This work was supported in part by As an important advantage of RISs, the negligible noise
the National Key Research and Development Program of China (Grant introduced by passive RISs enables a high array gain.
No. 2020YFB1807201), in part by the National Natural Science Foundation Particularly, in an RIS aided single-user single-input single-
of China (Grant No. 62031019), in part by the European Commission through
the H2020-MSCA-ITN META WIRELESS Research Project under Grant output (SU-SISO) system, the achievable signal-to-noise ratio
956256, and in part by the U.S National Science Foundation under Grants (SNR) gain enabled by an N -element RIS is proportional to
CCF-1908308 and CNS-2128448. Robert Schober work was partly supported N 2 [14]. Benefiting from this advantage, RISs are expected
by the Federal Ministry of Education and Research of Germany under
the programme of “Souveran. Digital. Vernetzt.” joint project 6G-RIC to introduce significant capacity gains in wireless systems [8].
(project identification number: PIN 16KISK023) and by the Deutsche However, in practice, these capacity gains are typically only
Forschungsgemeinschaft (DFG, German Research Foundation) under grant observed in communication scenarios where the direct link
SCHO 831/15-1. An earlier version of this paper was presented in part at the
IEEE GLOBECOM’22, Rio de Janeiro, Brazil, December 2022. The associate between transmitter and receiver is completely blocked or very
editor coordinating the review of this article and approving it for publication weak [8], [9], [10], [15], [16], [17]. By contrast, in many
was A. García Armada. (Corresponding author: Linglong Dai.) scenarios where the direct link is not weak, conventional RISs
Zijian Zhang, Linglong Dai, Xibi Chen, Changhao Liu, and Fan Yang
are with the Department of Electronic Engineering, Tsinghua University, achieve limited capacity gains [18]. The reason behind this
Beijing 100084, China, and also with the Beijing National Research Center phenomenon is the “multiplicative fading” effect introduced
for Information Science and Technology (BNRist), Beijing 100084, China by RISs, i.e., the equivalent path loss of the transmitter-RIS-
(e-mail: zhangzj20@mails.tsinghua.edu.cn; daill@tsinghua.edu.cn; cxb17@
tsinghua.org.cn; liu-ch21@mails.tsinghua.edu.cu; fan_yang@tsinghua.edu.cn). receiver link is the product (instead of the sum) of the path
Robert Schober is with the Institute for Digital Communications, losses of the transmitter-RIS and RIS-receiver links, which
Friedrich-Alexander Universität Erlangen-Nürnberg (FAU), 91054 Erlangen, is usually thousands of times larger than that of the direct
Germany (e-mail: robert.schober@fau.de).
H. Vincent Poor is with the Department of Electrical and Computer link [18]. As a result, the “multiplicative fading” effect makes
Engineering, Princeton University, Princeton, NJ 08544 USA (e-mail: it almost impossible for passive RISs to achieve noticeable
poor@princeton.edu). capacity gains in many wireless environments. Many existing
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TCOMM.2022.3231893. works on RISs have bypassed this effect by only considering
Digital Object Identifier 10.1109/TCOMM.2022.3231893 scenarios with severely obstructed direct links [8], [9], [10],
0090-6778 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
[15], [16], [17]. Therefore, to advance the practicability of self-interference. Then, by utilizing alternating direction
RISs in future 6G wireless networks, a critical issue for RISs method of multipliers (ADMM) [22] and sequential
to be addressed is: How to break the fundamental performance unconstrained minimization techniques (SUMT) [23],
bottleneck caused by the “multiplicative fading” effect? an alternating optimization scheme is proposed to solve
To overcome the fundamental physical limitation, in this the formulated problem.
paper, a new RIS architecture called active RISs is proposed The rest of this paper is organized as follows. In Section II,
for wireless communication systems. Specifically, different the architectures as well as the signal models of passive RISs
from passive RISs that passively reflect signals without and active RISs are introduced, respectively. In Section III,
amplification, the key feature of active RISs is their ability the asymptotic performance of active RISs is analyzed and
to actively reflect signals with amplification, which can be compared to that of the passive RISs. In Section IV, a sum-
realized by integrating reflection-type amplifiers into their rate maximization problem is formulated for an active RIS
reflecting elements. At the expense of additional power aided MU-MISO system, and a joint beamforming and
consumption, active RIS can compensate for the large path precoding design is proposed to solve the formulated problem.
loss of reflected links, which is promising to overcome the In Section V, we extend the studied joint beamforming and
“multiplicative fading” effect. Some parallel works1 have precoding design to the practical case with self-interference.
revealed the potential benefits of active RISs for enhancing In Section VI, the convergence and complexity of the proposed
wireless systems. For example, in [20], an active RIS was schemes are analyzed. In Section VII, validation results
introduced into a single-user single-input multi-output (SU- are presented to validate the signal model of active RISs
SIMO) system to enhance the user’s SNR. In [21], the authors and evaluate the performance of active RISs in typical
considered an active RIS aided SU-SISO system with limited communication scenarios. Finally, conclusions are drawn and
RIS elements, and the impact of RIS location placement on future works are discussed in Section VIII.
the communication performance is analyzed. Notations: C, R, and R+ denote the sets of complex, real,
In this paper,2 we propose the concept of active RISs and positive real numbers, respectively; [·]−1 , [·]∗ , [·]T , and
and focuses on the signal model verification, asymptotic [·]H denote the inverse, conjugate, transpose, and conjugate-
performance analysis, and multi-user beamforming design of transpose operations, respectively; · denotes the Euclidean
active RISs aided communication systems. Specifically, our norm of the argument; ·F denotes the Frobenius norm of the
contributions are summarized as follows: argument; diag(·) denotes the diagonalization operation; R{·}
• We develop and verify the signal model of active RISs, denotes the real part of the argument; ⊗ denotes the Kronecker
which characterizes the amplification of the incident product; ∠[·] denotes the angle of the complex argument;
signal and accounts for the non-negligible thermal noise ln(·) denotes the natural logarithm of its argument; CN (μ, Σ)
introduced by the active elements. Particularly, the denotes the complex multivariate Gaussian distribution with
verification is made via the experimental measurements mean μ and variance Σ; IL is an L × L identity matrix, and
based on a designed and fabricated active RIS element. 0L is an L × 1 zero vector.
• Based on the verified signal model, we analyze the
asymptotic performance of an active RIS with extremely II. PASSIVE RIS S AND ACTIVE RIS S
large number of active elements, and we further compare
In this section, we introduce the architectures of RISs.
it to that of the existing passive RISs, which reveals the
First, in Subsection II-A, we review passive RISs and point
notable capacity gain enabled by the use of active RISs.
out their physical limitation imposed by the “multiplicative
• To evaluate the performance of active RISs in typi-
fading” effect. Then, in Subsection II-B, to overcome this
cal communication systems, we formulate a sum-rate
limitation, we propose the novel concept of active RISs
maximization problem for an active RIS aided multi-
along with their hardware structure and signal model. Finally,
user multiple-input single-output (MU-MISO) system.
in Subsection II-C, we present the transmission model for an
Then, by exploiting fractional programming (FP), a joint
active RIS aided MU-MISO system.
transmit beamforming and reflect precoding scheme is
proposed to solve this problem.
• To account for the non-ideal factors in practical A. Passive RISs
systems, we extend the studied beamforming design The RISs widely studied in most existing works are passive
in the scenario with the self-interference of active RISs [2], [3], [4], [5], [6], [7], [8], [9], [10]. Specifically,
RISs. We model the feedback-type self-interference of as shown in Fig. 1 (a), a passive RIS comprises a large number
active RISs, which allows us to formulate an mean- of passive elements each being able to reflect the incident
squared error minimization problem to suppress the signal with a controllable phase shift. In general, each passive
RIS element consists of a reflective patch terminated with an
1 In October 2019, we started to design an active RIS element integrating a
impedance-adjustable circuit for phase shifting [24]. Thanks to
reflection-type amplifier [19]. The fabrication of this active RIS element was
finished in August 2020. Subsequently, we set out to establish an experimental the passive mode of operation without active radio-frequency
environment for signal measurements with this element, and all measurements (RF) components, a passive RIS element practically consumes
were completed in February 2021. This paper first appeared on ArXiv in zero direct-current power [24], and the introduced thermal
March 2021 [DOI: 10.48550/arXiv.2103.15154].
2 Simulation codes are provided to reproduce the results presented in this noise is usually negligible [5], [6], [7], [8], [9], [10]. Thereby,
article: http://oa.ee.tsinghua.edu.cn/dailinglong/publications/publications.html. the signal model of an N -element passive RIS widely used in
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1709
Fig. 1. Illustration of the hardware architectures of (a) a passive RIS and (b) an active RIS.
the literature is given as follows [6]: the transceiver antennas is omnidirectional and RIS elements
y = Θx, (1) are tightly deployed with half-wavelength spacing [18]. Let
d = 200 m, dt = 110 m, and dr = 110 m denote the
where x ∈ CN denotes the incident signal, Θ := distances between transmitter and receiver, transmitter and
diag ejθ1 , · · · , ejθN ∈ CN ×N denotes the reflection RIS, RIS and receiver, respectively. Assume that all channels
coefficient matrix of the passive RIS with θn being the phase are line-of-sight (LoS) and the RIS phase shift is optimally
shift of the n-th passive element, and y ∈ CN denotes the configured to maximize the channel gain of the transmitter-
signal reflected by the RIS. Note that the impact of noise is RIS-receiver reflected link. Then, for carrier frequencies of
neglected in (1). As a consequence, by properly adjusting Θ 5/10/28 GHz, N = dλ 4
dt dr = 4034/8067/22587 RIS
to manipulate the N signals reflected by the N RIS elements elements are required to make the reflected link as strong as
to coherently add with the same phase at the receiver, a high the direct link [18]. The high signaling overhead introduced
array gain can be achieved. This is expected to significantly by the N pilots required for channel estimation [25] and the
increase the receiver SNR [5], [6], [7], which is one of the key high complexity of O(N 2 ) for real-time beamforming [26]
reasons for why RISs have attracted so much research interest make the application of such a large number of passive RIS
recently [8], [9], [10], [11], [12], [13], [15], [16], [17]. elements in practical wireless networks very challenging [18].
Unfortunately, in practice, this expected high capacity Consequently, many existing works have bypassed the
gain often cannot be realized, especially in communication “multiplicative fading” effect by only considering the scenario
scenarios where the direct link between the transmitter and where the direct link is completely blocked or very
the receiver is not weak. The reason for this negative result is weak [5], [6], [7], [8], [9], [10], [15], [16], [17].
the “multiplicative fading” effect introduced by passive RISs.
Specifically, the equivalent path loss of the transmitter-RIS-
receiver reflected link is the product (instead of the sum) of B. Active RISs
the path losses of the transmitter-RIS and RIS-receiver links, To overcome the fundamental performance bottleneck
and therefore, it is thousands of times larger than that of caused by the “multiplicative fading” effect of RISs, in this
the unobstructed direct link. Thereby, for an RIS to realize paper, we propose active RISs as a promising solution.
a noticeable capacity gain, thousands of RIS elements are As shown in Fig. 1 (b), similar to passive RISs, active RISs
required to compensate for this extremely large path loss. can also reflect the incident signals with reconfigurable phase
Remark 1: To illustrate the above fact, let us consider shifts. Different from passive RISs that just reflect the incident
an SU-SISO system aided by a passive RIS. Assume that signals without amplification, active RISs can further amplify
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
y= Ψx
+ Ψv
+ ns , (2)
Desired signal Dynamic noise Static noise
Fig. 2. Illustration of the downlink transmission in an active RIS aided
MU-MISO system.
where Ψ := diag p1 ejθ1 , · · · , pN ejθN ∈ CN ×N denotes
the reflection coefficient matrix of the active RIS, wherein C. Active RIS Aided MU-MISO System
pn ∈ R+ denotes the amplification factor of the n-th active
element and pn can be larger than one thanks to the integrated To characterize the performance gains enabled by our
reflection-type amplifier. Due to the use of active components, proposed active RISs in typical communication scenarios,
active RISs consume additional power for amplifying the we consider an active RIS aided downlink MU-MISO system
reflected signals, and the thermal noise introduced by active as shown in Fig. 2, where an M -antenna base station (BS)
RIS elements cannot be neglected as is done for passive RISs. serves K single-antenna users simultaneously with the aid of
Particularly, as shown in (2), the noise introduced at active an N -element active RIS.
T
RISs can be classified into dynamic noise Ψv and static noise Let s := [s1 , · · · , sK ] ∈ CK denotethe transmitted
ns , where Ψv is the noise introduced and amplified by the symbol vector for the K users with E ssH = IK .
reflection-type amplifier and ns is generated by the patch and We assume that multi-user linear precoding is employed at the
the phase-shift circuit [28]. More specifically, v is related to BS for downlink transmission. Then, according to (2), signal
the input noise and the inherent device noise of the active RIS rk ∈ C received at user k can be modeled as follows:
elements [28], while the static noise ns is unrelated to Ψ and is K
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1711
Furthermore, to obtain analytical results and find more B. Comparisons Between Passive RISs and Active RISs
insights, we assume that each active RIS element has the same
We can observe from Lemma 1 and Lemma 2 that, compared
amplification factor (i.e., pn := p for all n ∈ {1, · · · , N }),
to the asymptotic SNR for passive RISs γpassive in (4) which
while the power allocation among active elements will be
is proportional to N 2 , the asymptotic SNR for active RISs
considered in Section IV. For a fair comparison with the
γactive in (5) is proportional to N due to the noises additionally
asymptotic performance of passive RISs, similar to [14],
introduced by the use of active components. At first glance,
we assume Rayleigh-fading channels.
it seems that the SNR proportional to N 2 achieved by passive
For the above RIS aided SU-SISO system without direct
RISs γpassive always exceeds the SNR proportional to N
link, we first redefine the BS-RIS channel matrix and the RIS-
achieved by active RISs γactive . However, this is actually not
user channel vector as G := g ∈ CN ×1 and fk := f ∈ CN ×1 ,
the case in many scenarios.
respectively, to simplify the notations. Then, we recall the
The reason behind this counterintuitive behavior is that,
following lemma from [14] for the asymptotic SNR achieved
different from the denominator of (4) which depends on the
by passive RISs.
noise power σ 2 , the denominator of (5) is determined by the
Lemma 1 (Asymptotic SNR for passive RISs): Assuming
much smaller multiplicative terms composed of path losses
f ∼ CN 0N , 2f IN , g ∼ CN 0N , 2g IN and letting N → and noise power, i.e., PAmax σv2 2f , PBS
max 2 2
σ g , and σ 2 σv2 . In this
∞, the asymptotic SNR γpassive of a passive RIS aided SU- case, the denominator of (5) is usually much smaller than that
SISO system is given by of (4). Thus, even if the numerator of (5) is smaller than that
max 2 2 2 of (4) because of an N gain loss, the SNR gain of active
PBS π f g
γpassive → N 2 , (4) RISs can still be higher than that of passive RISs in many
16σ 2 scenarios.
max
where PBS denotes the maximum transmit power at the BS. Generally, due to the much smaller denominator of (5), only
Proof: The proof can be found in [14, Proposition 2]. when N is unaffordably large can passive RISs outperform
For comparison, under the same transmission conditions, active RISs. To illustrate this claim, let us consider two
we provide the asymptotic SNR of an active RIS aided SU- different SU-SISO systems, which are aided by an active RIS
SISO system in the following lemma. and a passive RIS, respectively. Then, the following lemma
Lemma 2 (Asymptotic SNR for active RISs): Assuming specifies the condition that has to be met for passive RISs to
outperform active RISs.
f ∼ CN 0N , 2f IN , g ∼ CN 0N , 2g IN and letting N →
Lemma 3 (Case when passive RISs outperform active RISs):
∞, the asymptotic SNR γactive of an active RIS aided SU-SISO
Assuming the number of RIS elements N is large, the required
system is given by
number of elements N for a passive RIS to outperform an
max max 2 2 2
PBS PA π f g active RIS has to satisfy
γactive → N , (5)
max
16 PAmax σv2 2f + PBS
max σ 2 2 + σ 2 σ 2 PBS-A PAmax σ 2
g v
N≥ max , (6)
PBS-P P max σ 2 2 + P max σ 2 2 + σ 2 σ 2
A v f BS-A g v
where PAmax denotes the maximum reflect power of the active
RIS. max
where PBS-A denotes the maximum BS transmit power for the
Proof: Please see Appendix A. max
active RIS aided system and PBS-P denotes that for the passive
Remark 3: From (5) we observe that the asymptotic SNR of
RIS aided system.
an active RIS aided SU-SISO system depends on both the BS
max Proof: Please see Appendix B.
transmit power PBS and the reflect power of the active RIS
max max Next, we consider a specific setup to compare the user’s
PA . When PBS → ∞, it can be proved that the asymptotic
achievable SNRs in the above two systems. For a fair
SNR of the active RIS aided system will be upper-bounded
P max π 2 2 comparison, we constrain the total power consumption P max
by γactive → N A16σ2 f , which only depends on the RIS- of the two systems to 2 W by setting PBS-P max
= 2 W for the
user channel gain 2f and the noise power at the user σ 2 . This passive RIS aided system and PBS-A = PAmax = 1 W for
max
indicates that, when the BS transmit power is high enough, the the active RIS aided system, respectively. Therefore, when
BS-RIS channel g and the noise power at the active RIS have σ 2 = σv2 = −100 dBm and 2f = 2g = −70 dB, the required
negligible impact on the user’s SNR. Similarly, if PAmax → ∞, number of elements N for the passive RIS to outperform the
the asymptotic SNR γactive in (5) will be upper-bounded by active RIS is 2.5×106 according to (6), which is impractical to
P max π 2 2
γactive → N BS16σ2 g . Compared with (5), this upper bound is realize with current technology. Besides, the high overhead for
v
independent of the RIS-user channel f and the noise power at channel estimation [25] and the high complexity for real-time
the user σ 2 . It indicates that, the negative impact of small f beamforming [26] also make the application of such a large
and large σ 2 can be reduced by increasing the reflect power number of RIS elements impractical [18]. Conversely, for a
of the active RIS PAmax , which may provide guidance for the more practical number of elements of N = 256, according
design of practical active RIS-aided systems. to (5) and (4), the SNR achieved by the passive RIS is
Next, we compare the asymptotic SNRs for passive RISs in γpassive ≈ 39.0 dB, while the SNR achieved by the active RIS
Lemma 1 and active RISs in Lemma 2 to reveal the superiority is γactive ≈ 79.0 dB, which is about 10, 000 times higher than
of active RISs in wireless communications. γpassive .
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Fig. 3. Asymptotic SNR as a function of the number of RIS elements N for different ranges of N .
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1713
IV. J OINT T RANSMIT B EAMFORMING AND R EFLECT Algorithm 1 Proposed Joint Transmit Beamforming and
P RECODING D ESIGN Reflect Precoding Scheme
To investigate the capacity gain enabled by active RISs in Input: Channels G, hk , and fk , ∀k ∈ {1, · · · , K}.
typical communication scenarios, in this section, we consider a Output: Optimized BS beamforming vector w, optimized
more general active RIS aided MU-MISO system. Specifically, RIS precoding matrix of active RIS Ψ, and optimized sum-
in Subsection IV-A, we formulate the problem of sum-rate rate Rsum .
maximization. Then, in Subsection IV-B, a joint transmit 1: Randomly initialize w and Ψ;
beamforming and reflect precoding scheme is proposed to 2: while no convergence of Rsum do
solve the problem. 3: Update ρ by (14);
4: Update by (15);
5: Update w by solving (17);
A. Sum-Rate Maximization Problem Formulation
6: Update Ψ by solving (20);
According to the MU-MISO transmission model in (3), the 7: Update Rsum by (11a);
signal-to-interference-plus-noise ratio (SINR) at user k can be 8: end while
obtained as 9: return Optimized w, Ψ, and Rsum .
H 2
h̄ wk
k
γk = K , (9)
h̄H wj 2 + f H Ψ2 σv2 + σ 2
j=1,j=k k k Due to the non-convexity and highly coupled variables in
h̄H hH H 1×M problem Po in (11), the joint design of w and Ψ is challenging.
wherein k = k +fk ΨG ∈C is the equivalent channel
Specifically, the introduction of the active RIS brings many
from the BS to user k, which includes both the direct link and
the reflected link. By solving the expectation of the squared difficulties to the beamforming design, such as the additional
Euclidean norm of the radiated signals, the BS transmit power, power constraint, the power allocation among active elements,
the cancellation of multi-user interference, and the amplified
PBS , and the reflect power of the active RIS, PA , can be
respectively derived as noise power. Therefore, to efficiently solve this problem,
⎧ we develop a joint beamforming and precoding scheme based
2 ⎫
⎨ K ⎬ K on alternating optimization and fractional programming (FP),
2
PBS = E wk s k = wk , (10a) as provided in the next subsection.
⎩ ⎭
k=1 k=1
Desired signal power B. Proposed Joint Beamforming and Precoding Scheme
⎧ 2 ⎫
⎨ K ⎬ To solve the problem efficiently, we reformulate the problem
PA = E ΨG wk sk +Ψv first. For simplicity, here we refer to w and Ψ as the
⎩ ⎭
k=1 BS beamforming vector and the RIS precoding matrix,
K respectively. In order to deal with the non-convex sum-of-
2 2
= ΨGwk + Ψ σ 2 . (10b) logarithms and fractions in (11), we exploit FP methods
F v
k=1 proposed in [34] to decouple the variables in problem Po
Amplified noise power
Desired signal power in (11), so that multiple variables can be optimized separately.
This leads to the following lemma.
Note that, different from the BS transmit power PBS which Lemma 5 (Equivalent problem for sum-rate maximization):
only includes the desired signal power, since the active RIS By introducing auxiliary variables ρ := [ρ1 , · · · , ρK ] ∈ RK
+
amplifies the noises as well, the additional power consumption and := [ 1 , · · · , K ] ∈ CK , the original problem Po
due to the noise amplification should be taken into account in in (11) can be equivalently reformulated as follows
the reflect power of active RIS PA .
K
Therefore, the original problem of sum-rate maximization,
subject to the power constraints at the BS and the active RIS, P1 : max Rsum (w, Ψ, ρ, ) = ln (1 + ρk )
w,Ψ,ρ,
k=1
can be formulated as follows: K K
K
− ρk + g(w, Ψ, ρk , k ),
Po : max Rsum (w, Ψ) = log2 (1 + γk ), (11a) k=1 k=1
w,Ψ
k=1 2 max
s.t. C1 : w ≤ PBS ,
K
wk 2 ≤ PBS
max K
s.t. C1 : , (11b)
C2 : ΨGwk 2 + Ψ2F σv2 ≤ PAmax , (12)
k=1
K k=1
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Then, for fixed RIS precoding matrix Ψ and auxiliary variables V. S ELF -I NTERFERENCE S UPPRESSION FOR ACTIVE RIS S
ρ and , problem P1 in (12) can be reformulated as follows
Since active RISs work in full-duplex (FD) mode, the
P2 : max R 2bH w − wH Aw, self-interference of active RISs occurs in practical systems.
w In this section, we extend the studied joint beamforming
s.t. C1 : w2 ≤ PBS
max
, and precoding design to the practical system with the self-
C2 : wH Ξw ≤ Pmmax . (17) interference of active RISs. Specifically, in Subsection V-A,
we first model the self-interference of active RISs, which
Since P2 in (17) is a standard quadratic constraint quadratic allows us to account for the self-interference suppression in
programming (QCQP) problem, by adopting the Lagrange the beamforming design. In Subsection V-B, we formulate a
multiplier method [22], the optimal solution wopt to P2 in (17) mean-squared error minimization problem to suppress the self-
can be obtained as follows interference of active RISs. In Subsection V-C, by utilizing
−1
ADMM [22] and SUMT [23], an alternating optimization
wopt = (A + λ1 IMK + λ2 Ξ) b, (18) scheme is proposed to solve the formulated problem.
where λ1 and λ2 are the Lagrange multipliers, which should
be chosen such that the complementary slackness conditions A. Self-Interference Modeling
of power constrains C1 and C2 are satisfied. The optimal The self-interference of FD relays and that of active RISs are
Lagrange multipliers λopt
1 and λopt
2 can be obtained via a two- quite different. Specifically, due to the long processing delay
dimensional grid search [22]. at relays, the self-interference of FD relay originates from the
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1715
−1
different symbols that transmitted in the adjacent timeslot [35], (IN − ΦH) ≈ IN + ΦH, thus (25) can be rewritten as
[36], [37]. In this case, the self-interference at relays is usually follows:
viewed as colored Gaussian noise, which can be canceled
by a zero-forcing suppression method [36]. Differently, since h̄H H H
k ≈ hk + fk (IN + ΦH) ΦG
(a) H
active RISs have nanosecond processing delay, the incident = hk + fkH + φH diag fkH H ΦG
and reflected signals carry the same symbol in a timeslot. Due (b) H
to the non-ideal inter-element isolation of practical arrays, part = hk + φH + φH Hk diag φH diag fkH G, (26)
of the reflected signals may be received again by the active Equivalent precoding vector for user k
RIS. In this case, the feedback-type self-interference occurs, H
which cannot be viewed as Gaussian noise anymore. wherein RIS precoding vector φ satisfies HΦ = diag(φ );
H H H
To distinguish the RIS precoding matrix in the ideal case (a) holds since φ diag(fk ) = fk diag φ ; (b) holds by
Ψ, we denote the RIS precoding matrix defining Hk = diag(fkH )H(diag(fkH ))−1 .
in the non-ideal case Comparing (26) and (19), the difference is that the RIS
with self-interference as Φ := diag p1 ejθ1 , · · · , pN ejθN .
Recalling (2) and ignoring the negligible static noise for precoding vector ψ in (19) is replaced by the equivalent
simplicity, the reflected signal of active RISs in the presence precoding vector φ + diag (φ) HH k φ for user k. Therefore,
of self-interference can be modeled as follows: an efficient way to eliminate the impact of self-interference
is to design a φ to make all φ + diag (φ) HH k φ approach
y= Φx + ΦHy + Φv , (23) the ideally optimized RIS precoding vector ψ opt as close
Desired signal Dynamic noise as possible. To achieve this, we temporarily omit the power
Self-interference
constraint of active RISs in (11c) and formulate the following
where H ∈ CN ×N denotes the self-interference matrix [35]. mean-squared error minimization problem:
In the general case without self-excitation (determinant of
1
K
opt 2
(IN −ΦH) is not zero), model (23) is a standard self-feedback P4 : min f (φ) = φ + diag (φ) HH
kφ −ψ ,
loop circuit, of which the output y naturally converges to the φ K
k=1
following steady state: (27)
−1
y= (IN − ΦH) Φ (x + v) . (24) where objective f (φ) is the cost function, defined as the mean
of the squared approximation errors.
Equivalent RIS precoding matrix
Comparing (24) and (2), one can observe that the difference C. Proposed Self-Interference Suppression Scheme
is that the RIS precoding matrix Ψ in (2) is replaced by
−1
(IN − ΦH) Φ. In particular, when all elements in H are To ensure the communication performance of active RIS
−1
zero, the equivalent RIS precoding matrix (IN − ΦH) Φ is aided systems, in this subsection, we propose a self-
equal to diagonal matrix Φ. interference suppression scheme to solve problem P4 in (27).
Obviously, in the ideal case without self-interference (i.e.,
self-interference matrix H is zero matrix), the optimal solution
B. Problem Formulation to problem P4 in (27) is φ = ψ opt and satisfies f (φ) = 0.
Here we focus on the non-ideal case with a non-zero H.
To account for the self-interference of active RISs in the In this case, problem P4 is challenging to solve due to the
beamforming design, according to the new signal model (24), three reasons. Firstly, the objective f (φ) is usually non-convex
an intuitive way is to replace the RIS precoding matrix Ψ in since Hk is asymmetric and indefinite. Secondly, f (φ) is in
problem P1 in (12) with the equivalent RIS precoding matrix quartic form with respect to φ thus P4 has generally no closed-
−1
(IN − ΦH) Φ and then solve P1 . Since this operation does form solution. Finally, the coupled term diag (φ) HH k φ is a
not influence the optimizations of w, ρ, and , here we focus non-standard quadratic thus is hard to be preprocessed and
on the optimization of Φ. optimized like (20).
−1
Consider replacing Ψ in (19) with (IN − ΦH) Φ, thus To tackle this issue, inspired by ADMM [22] and
H
the equivalent channel h̄k with self-interference can be written SUMT [23], we turn to find a feasible solution to problem
as: P4 by alternating optimization, as shown in Algorithm 2. The
−1 key idea of this algorithm includes two aspects: i) ADMM:
h̄H H H
k = hk + fk (IN − ΦH) ΦG. (25)
Fix some variables and then optimize the others, so that
However, due to the existence of self-interference matrix f (φ) becomes temporarily convex thus can be minimized
H, Φ to be optimized exists in an inversion, thus the by alternating optimizations. ii) SUMT: Introduce an initially
equivalent channel h̄H small but gradually increasing penalty term into the objective,
k cannot be processed like (19), which
makes Φ hard to be optimized. To address this challenge, so that the variables to be optimized can converge as an
we introduce the first-order Taylor expansion3 to approximate achievable solution to the original problem.
Following this idea, P4 in (27) can be reformulated as
2
P5 : min q (φ, φ ) = f (φ, φ ) + ζφ − φ ,
3 This approximation requires that the self-interference is not too strong
(28)
(i.e., the values in H are small), so that the high-order expansion items can φ,φ
be reasonably ignored. Penalty term
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
wherein f (φ, φ ) is defined as where (a) and (b) follow since the updates of Ψ and
w are the optimal solutions to subproblems P3 in (20)
1
K
opt 2 and P2 in (17), respectively; (c) and (d) follow because
f (φ, φ ) = φ + diag (φ ) HH
kφ −ψ (29)
K
k=1
the updates of and ρ maximize Rsum when the other
variables are fixed, respectively. Therefore, the objective
and ζ > 0 is the penalty coefficient that increases in each
Rsum is monotonically non-decreasing in each iteration.
iteration. For simplicity, here we assume ζ doubles in each Since the value of Rsum
is upper-bounded due to power
update. In particular, when ζ → ∞, problem P5 in (28) is constrains C1 and C2 , Algorithm 1 will converge to a local
equivalent to P4 in (27). optimum.
Observing (29), we note that q (φ, φ ) = f (φ, φ ) = f (φ) As an exterior point method, Algorithm 2 meets two
when φ = φ. Particularly, when φ (or φ ) is fixed, objective standard convergence conditions [23], which determines that
q (φ, φ ) becomes a convex quadratic as a function of φ it converges to a local optimal point where φ = φ and
(or φ). Therefore, for a given ζ, q (φ, φ ) in (28) can be q (φ, φ ) = f (φ, φ ) = f (φ). Firstly, for a given penalty
minimized by optimizing φ and φ alternatively. By solving coefficient ζ in each iteration, the value of q (φ, φ ) in (28)
∂q (φ, φ )/∂φ = 0 and ∂q (φ, φ )/∂φ = 0, we obtain the is lower-bounded by zero and experiences the following
updating formulas of φ and φ respectively, given by monotonically non-increasing update:
K
−1 K
1 1 (a) (b)
φ= BHk Bk + ζIN ζφ + BHkψ
opt
, q φt+1 , (φ )t+1 ≤ q φt+1 , (φ )t ≤ q φt , (φ )t , (34)
K K
k=1 k=1
(30) where (a) follows because the update of φ minimizes
−1 q (φ, φ ) in P5 in (28) when φ is fixed and (b) follows
1
K
1
K
since the update of φ minimizes q (φ, φ ) when φ is fixed.
φ = DH
k Dk +ζIN ζφ+ DH
k ψ
opt
−φ ,
K K Secondly, as penalty coefficient ζ increases to be sufficiently
k=1 k=1
(31) large (ζ → ∞), q (φ, φ ) in (28) is dominated by the penalty
2
term ζφ − φ . The updating formulas (30) becomes φ =
where Bk = IN + diag (φ ) HH k and Dk = diag Hk φ .
H
φ and (31) becomes φ = φ. It indicates that, φ and φ
2
Besides, due to the existence of penalty term ζφ − φ , do not update anymore and φ = φ always holds. As a
as ζ increases, the converged solution to P5 in (28) tends to 2
result, penalty term ζφ − φ is equal to zero and the
satisfy φ = φ. After several alternating updates, φ and φ converged objective q (φ, φ ) finally satisfies q (φ, φ ) =
will converge to the same value (φ = φ ), thus we obtain f (φ, φ ) = f (φ).
the desired RIS precoding matrix Φ = diag(φH ), which is
exactly the output of Algorithm 2.
Recall that we temporarily omitted the power constraint B. Computational Complexity Analysis
in (11c) while optimizing φ. Here we introduce a scaling factor The computational complexity of Algorithm 1 is mainly
τ > 0 for Φ to satisfy (11c), leading to the final solution Φopt , determined by the updates of the four variables ρ, , w,
i.e., and Ψ via (14), (15), (17), and (20), respectively. Specifically,
the computational complexity of updating ρ is O (KM ). The
Φopt = τ Φ. (32)
complexity of updating is O K 2 M + KN . Considering
According to (24) and (10), τ can be obtained by replacing the complexity of solving standard convex QCQP problem, for
−1
Ψ in PA in (10) with (IN − Φopt H) Φopt and then doing a a given accuracy tolerance
ε, the
√ computational complexity of
binary search to find a proper τ that satisfies PA = PAmax . This updating w is O log2 (1/ε) M K + 2 (1 + M K) M 3 K 3 .
completes the proposed self-interference suppression scheme. Similarly, the computational complexity of updating Ψ
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1717
Fig. 4. Experimental devices and environment for validating the signal model (2) of the active RIS.
√
is O log2 (1/ε) N + 1 (1 + 2N ) N 3 . Thus, the overall VII. VALIDATION R ESULTS
computational
complexity
of Algorithm
1 is given by In this section, we present validation results. To validate
O log2 (1/ε) Io M 4.5 K 4.5 + N 4.5 , wherein Io denotes the signal model (2), in Subsection VII-A, we present
the number of iterations required by Algorithm 1 for experimental results based on a fabricated active RIS element.
convergence. Then, in Subsection VII-B, simulation results are provided to
Similarly, the computational complexity of Algorithm 2 is evaluate the sum-rate of active RIS aided MU-MISO systems.
mainly determined by updating φ and φ via (30) and (31), Finally, in Subsection VII-C, the impact of active RIS self-
respectively. As closed-form updating
formulas, their compu-
interference on system performance is discussed.
tational complexity are both O (K + 1)N 3 + (K + 1)N 2 ,
which are mainly caused by matrix inversions. Thus,
the overall computational complexity of Algorithm 2 is A. Validation Results for Signal Model
O Is KN 3 , wherein Is is the number of iterations required To validate the signal model (2), we designed and fabricated
by Algorithm 2 for convergence. an active RIS element with an integrated reflection-type
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Fig. 5. Experimental measurement result for reflection gain G versus signal Fig. 6. Experimental measurement result for the density of noise power
frequency f . Gσv2 + σs2 versus reflection gain G.
amplifier for experimental measurements in [19]. Particularly, • Fig. 4 (d) shows a photo of the fabricated active RIS
since the phase-shifting ability of RISs has been widely element under test, which is connected by a waveguide
verified [24], we focus on studying the reflection gain and the for incident/reflected signal exchanges.
noise introduced by an active RIS element. Thus, the validation • Fig. 4 (e) shows a photo of the experimental environment
of signal model (2) is equivalent to validating with the required equipment for device driving and signal
Py = GPx + Gσv2 + σs2 , (35) measurement.
2) Reflection Gain Measurement: Using the measurement
Desired-signal power noise power system for the reflection gain depicted in Fig. 4 (b), we first
where Py is the power of the signals reflected by the active investigate the reflection gain G of the active RIS element.
RIS element; Px is the power of the incident signal; G := p2 is The reflection gain G can be reconfigured by the input power
the reflection gain of the active RIS element; Gσv2 and σs2 are of the pump source Pp . By setting the input power of the
the powers of the dynamic noise and static noise introduced vector network analyzer as Px = −50 dBm, the reflection
by the active RIS element, respectively. gain G as a function of the signal frequency can be directly
1) Hardware Platform: To validate the model in (35), measured via the vector network analyzer. Then, in Fig. 5,
we first establish the hardware platform used for our we show the measurement results for reflection gain G as a
experimental measurements, see Fig. 4. Specifically, we show function of signal frequency f for different input powers of
the following aspects: the pump source Pp . We observe that the active RIS element
• Fig. 4 (a) illustrates the structure of the fabricated active can achieve a reflection gain G of more than 25 dB, when
RIS element operating at a frequency of 2.36 GHz [19]. Pp = 18.24 dBm, which confirms the significant reflection
A pump input at a frequency of 4.72 GHz is used to gains enabled by active RISs. On the other hand, when Pp = 0,
supply the power required by the active RIS element. we observe that G falls to −6 dB, which is lower than the
The incident signal and the pump input are coupled in a expected 0 dB. This loss is mainly caused by the inherent
varactor-diode-based reflection-type amplifier to generate power losses of the circulator and transmission lines used for
the reflected signal with amplification. measurement.
• Fig. 4 (b) illustrates the system used for measuring the 3) Noise Power Measurement: We further study the noise
reflection gain G of the active RIS element. A direct- power introduced and amplified by the active RIS element,
current (DC) source is used to provide a bias voltage i.e., Gσv2 + σs2 in (35), where Gσv2 and σs2 are the powers
of 7.25 V for driving the active RIS element, and a of the dynamic noise and static noise introduced at the
controllable pump source is used to reconfigure the active RIS element, respectively. Using the noise measurement
reflection gain G. A circulator is used to separate the system in Fig. 4 (c), we show the measurement results for
incident signal and the reflected signal, and the reflection the spectral density of noise power Gσv2 + σs2 as a function
gain is directly measured by a vector network analyzer. of G for different operating frequencies in Fig. 6. We can
• Fig. 4 (c) illustrates the system for measuring the noises observe that the noise power increases nearly linearly with G,
introduced at the active RIS element, where a spectrum which verifies the noise model Gσv2 + σs2 in (35). Particularly,
analyzer is used to measure the noise power. The noise for f = 2.3601 GHz, the spectral density of σs2 is about
source is a 50 Ω impedance, which aims to simulate −174 dBm/Hz, while that of σv2 is about −160 dBm/Hz, which
a natural input noise of -174 dBm/Hz at each patch. is about 15 dB higher. The reason for this is that the input
The reflected signal is amplified by a low-noise amplifier noise is amplified by the noise factor [28], and additional
(LNA) so that the spectrum analyzer can detect it. noises are also introduced by the other active components the
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1719
Fig. 7. Two simulation scenarios with different channel conditions, where a BS aided by an active RIS serves four users.
Fig. 8. Simulation results for the sum-rate as a function of distance L in an RIS-aided MU-MISO system.
measurement equipment, such as the leakage noise from the where d is the distance between two devices. Path loss model
DC source. PLw is used to generate the weak BS-user link in scenario 1,
while PLs is used to generate the strong BS-user link in
B. Simulation Results for Joint Beamforming and Precoding scenario 2. For both scenarios in Fig. 7, PLs is used to
Design generate the BS-RIS and the RIS-user channels. To account
To evaluate the effectiveness of the proposed joint for small-scale fading, following [39], we adopt the Ricean
beamforming and precoding design, in this subsection, fading channel model for all channels involved. In this way,
we present simulation results for passive RIS and active RIS an arbitrary channel matrix H is generated by
aided MU-MISO systems, respectively.
1) Simulation Setup: For the simulation setup, we consider √ κ 1
H = PL HLoS + HNLoS , (37)
an active/passive RIS aided MU-MISO system operating at κ+1 κ+1
a frequency of 5 GHz as shown in Fig. 7. Particularly,
we consider two scenarios with different channel conditions. where PL is the corresponding path loss of H; κ is the Ricean
In Fig. 7 (a), the direct link is weak due to severe obstruction, factor; and HLoS and HNLoS represent the deterministic LoS
while the direct link is strong in Fig. 7 (b). To be specific, and Rayleigh fading components, respectively. In particular,
two different path loss models from the 3GPP standard [38, here we assume κ = 1.
B.1.2.1] are utilized to characterize the large-scale fading of As common settings, the BS and the active/passive RIS
the channels: are located at (0, -60 m) and (300 m, 10 m), respectively.
The locations of the four users will be specified later. Unless
PLs = 37.3 + 22.0 log d, specified otherwise, the numbers of BS antennas and RIS
PLw = 41.2 + 28.7 log d, (36) elements are set as M = 4 and N = 512, respectively.
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
Fig. 9. Simulation results for the sum-rate versus total power consumption P max in an RIS-aided MU-MISO system.
The noise power is set as σ 2 = σv2 = −100 dBm. Let PBS max
limited sum-rate gain, while the active RIS still realizes a
denote the maximum transmit power at the BS and PAmax noticeable sum-rate gain. For example, when L = 300 m,
denote the maximum reflect power of the active RIS, which the capacities without RIS, with passive RIS, and with
don’t include the hardware static power. For fair comparison, active RIS in scenario 1 are 2.98 bps/Hz, 13.80 bps/Hz, and
we constrain the total power consumption P max := PBS max
+ 33.39 bps/Hz respectively, while in scenario 2, these values are
max max max
PA to 10 dBm by setting PBS = 0.99 × P and 16.75 bps/Hz, 20.56 bps/Hz, and 38.45 bps/Hz, respectively.
PAmax = 0.01 × P max for the active RIS aided system, and For this position, the passive RIS provides a 363% gain in
max
PBS = 10 dBm for the other benchmark systems. To show scenario 1 and a 22% gain in scenario 2. By contrast, the active
the effectiveness of beamforming designs, here we consider RIS achieves noticeable sum-rate gains of 1020% in scenario
the following four schemes for simulations: 1 and 130% in scenario 2, which are much higher than those
• Active RIS (ideal case): In an ideal active RIS-aided achieved by the passive RIS in the corresponding scenarios.
MU-MISO system without self-interference, the proposed These results demonstrate that, compared with the passive RIS,
Algorithm 1 is employed to jointly optimize the BS the active RIS can overcome the “multiplicative fading” effect
beamforming and the precoding at the active RIS. and achieve noticeable sum-rate gains even when direct link
• Passive RIS [26]: In a passive RIS-aided MU-MISO is strong.
system, the algorithm proposed in [26] is adopted to 3) Sum-Rate Versus Total Power Consumption P max :
jointly optimize the BS beamforming and the precoding To evaluate the averaged performance in the coverage of
at the passive RIS. active/passive RIS, we assume that all users are randomly
• Random phase shift [40]: In a passive RIS-aided distributed in a large circle with a radius of 50 m from the
MU-MISO system, the phase shifts of all passive center (300 m, 0). We show the users’ sum-rate versus the total
RIS elements are randomly set. Then, relying on the power consumption P max in Fig. 9. From Fig. 9 we observe
equivalent channels from the BS to users, the weighted that the passive RIS achieves visible performance gains in
mean-squared error minimization (WMMSE) algorithm scenario 1 where the direct link is weak, while the passive RIS
from [40] is used to optimize the BS beamforming. only achieves limited sum-rate gains in scenario 2 where the
• Without RIS [40]: In an MU-MISO system without RIS, direct link is strong. By contrast, in both scenarios, the active
the WMMSE algorithm from [40] is adopted to optimize RIS realizes a high performance gain. Particularly, to achieve
the BS beamforming. the same performance as the passive RIS aided system, the
2) Coverage Performance of Active RISs: To observe the required power consumption for the active RIS aided system is
coverage performance of active RISs, we assume the four users much lower. For example, when the total power consumption
are randomly located in a circle with a radius of 5 m from of the passive RIS aided system is P max = 30 dBm, to achieve
the center (L, 0). In Fig. 8 (a) and (b), we plot the sum-rate the same sum-rate, the active RIS aided system only requires
versus distance L for the two considered scenarios, where the 7 dBm in scenario 1 and 12 dBm in scenario 2, which
direct link is weak and strong, respectively. Based on these correspond to power savings of 23 dB and 18 dB, respectively.
results, we have two observations. Firstly, in scenario 1 with The reason for this result is that, for the passive RIS, the total
a weak direct link, the passive RIS can indeed achieve power is only allocated to BS. Thus, all transmit power is
an obvious performance improvement, while the active RIS affected by the large path loss of the full BS-RIS-user link.
achieves a much higher sum-rate gain. Secondly, in scenario However, for the active RIS, part of the transmit power is
2 with a strong direct link, the passive RIS only achieves a allocated to the active RIS, and this part of the power is
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1721
Fig. 10. Simulation results for the sum-rate versus the number of RIS elements N in an RIS-aided MU-MISO system.
Fig. 11. Simulation results for the sum-rate versus the self-interference factor δ of the active RIS.
only affected by the path loss of the RIS-user link. Thus, the 14.78 bps/Hz), respectively. These results show that the sum-
active RIS is promising for reducing the power consumption rate increase of the active RIS aided system is much higher
of communication systems. than that of the passive RIS aided system. This indicates that,
4) Sum-Rate Versus Number of RIS Elements N : For the as long as the number of RIS elements N is not exceedingly
same setup as in Fig. 9, we plot the users’ sum-rate versus large (such as millions of elements), compared with the passive
the number of RIS elements N in Fig. 10. We observe that, RIS, increasing the number of elements of the active RIS
as the number of RIS elements N increases, both the passive is much more efficient for improving the communication
RIS and the active RIS achieve higher sum-rate gains, while performance, which is in agreement with the performance
the performance improvement for the active RIS aided system analysis in Section III.
is much larger than that for the passive RIS aided system. For
example, when N increases from 100 to 900, the sum-rate C. Simulation Results for Self-Interference Suppression
of the passive RIS aided system increases from 5.23 bps/Hz In this subsection, we present simulation results to verify
to 12.44 bps/Hz in scenario 1 (increase of 7.21 bps/Hz) and the effectiveness of the proposed self-interference suppression
from 17.57 bps/Hz to 20.85 bps/Hz in scenario 2 (increase scheme for active RISs.
of 3.28 bps/Hz), respectively. By contrast, the sum-rate of 1) Simulation Setup: To avoid the impact of other factors,
the active RIS aided system increases from 19.59 bps/Hz to we adopt the same setup in Subsection VII-B, which is used
34.91 bps/Hz in scenario 1 (increase of 15.32 bps/Hz) and in Fig. 9 and Fig. 10. Without loss of generality, we assume
from 23.81 bps/Hz to 38.59 bps/Hz in scenario 2 (increase of that each element in self-interference matrix H is distributed
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
as ∼ CN 0, δ 2 [35], [36], [37], where we name δ as the hardware design [27], prototype development [7], channel
self-interference factor, which is inversely proportional to the estimation [25], and energy efficiency analysis [10].
inter-element isolation of practical arrays [41], [42], [43].
To evaluate the impact of self-interference on sum-rate, we add A PPENDIX A
two new benchmarks for simulations: P ROOF OF L EMMA 2
• Active RIS (SI suppression): In a non-ideal active
RIS-aided MU-MISO system with self-interference, For notational simplicity, we rewrite some matrices and
T T
Algorithm 1 is employed to optimize the BS beamform- vectors in (3) as g = [g1 , · · · , gN ] , f = [f1 , · · · , fN ] , and
ing and the active RIS precoding, and Algorithm 2 is wk := w. Thus, the downlink transmission model in (3) can
employed to suppress the self-interference. Then, the be rewritten as
performance is evaluated under the condition of self-
r = pf H Θg ws + pf H Θv
interference.
• Active RIS (no suppression): In a non-ideal active Reflected link Noise introduced by active RIS
RIS-aided MU-MISO system with self-interference, only + z
, (38)
Algorithm 1 is employed to design the BS beamforming Noise introduced at user
and the active RIS precoding and the self-interference is
where r ∈ C is the signal received by the user. Based on
ignored. Then, the performance is evaluated under the
the transmission model in (38), the maximization of the user’s
condition of self-interference.
SNR γ, subject to the power constraints at the BS and the
2) Impact of Self-Interference on Sum-Rate: We plot the
active RIS, can be formulated as follows:
users’ sum-rate versus the self-interference factor δ in Fig. 11.
We observe that, when δ < −50 dB, the self-interference H
pf Θgw 2
has almost no impact on the sum-rate. However, as the max γ = 2 ,
self-interference strengthens, the active RIS aided system
w, p, Θ p2 f H Θ σv2 + σ 2
2
max
without self-interference suppression suffers an increasingly s.t. C1 : |w| ≤ PBS , (39)
high performance loss. Particularly, when δ = −35 dB, 2
C2 : p Θgw + p
2 2
N σv2 ≤ PAmax ,
the active RIS without self-interference suppression does not
max
even perform as well as the passive RIS in scenario 2. The where PBS and PAmax denote the maximum transmit power
reason is that, the existence of self-interference matrix H and the maximum reflect power at the BS and the active RIS,
makes the reflected signals unable to focus on the users, respectively. Then, the optimal solution of problem (39) can
or even worse, cancel the desired signals of the direct be obtained by the Lagrange multiplier method as follows:
link. Fortunately, thanks to our proposed Algorithm 2, the
active RIS aided system with self-interference suppression wopt = PBS max , (40a)
opt
can still hold a considerable performance. For example, when θn = ∠fn − ∠gn , ∀n ∈ {1, · · · , N }, (40b)
δ = −35 dB, compared with the active RIS aided system
PAmax
without self-interference suppression, the system with self- popt = 2
. (40c)
max N
+ N σv2
interference suppression can compensate for the sum-rate loss PBS n=1 |gn |
of 14.72 bps/Hz in scenario 1 and that of 18.52 bps/Hz in
By substituting (40) into (39), the user’s maximum achievable
scenario 2.
SNR for active RISs can be obtained as
2
VIII. C ONCLUSION AND F UTURE W ORKS max max
PBS PA N |f
n=1 n | |g |
n
In this paper, we have proposed the concept of active RISs γactive = .
2 2
PAmax σv2 n=1 |fn | +σ 2 PBSmax
N N 2
to overcome the fundamental limitation of the “multiplicative n=1 |gn | +N σv
fading” effect. Specifically, we have developed and verified a (41)
signal model for active RISs by a fabricated active RIS element
2
assume f ∼ CN (0N , f IN ) and g ∼
through experimental measurements. Based on the verified Note
that 2
we
signal model, we have analyzed the asymptotic performance CN 0N , g IN . Thus, by letting N → ∞ in (41), according
of active RISs and then formulated an optimization problem N
to the law of large numbers, we have n=1 |fn | |gn | →
to maximize the sum-rate in an active RIS aided MU- 2 2
N 4 , n=1 |gn | → N g , and n=1 |fn | → N 2f . After
2
πf g N N
MISO system. Subsequently, we have proposed a joint substituting these asymptotic equations into (41), we obtain
beamforming and precoding scheme to solve this problem. for the asymptotic SNR for active RISs the expression in (5),
Finally, experimental and simulation results have shown that, which completes the proof.
compared with the benchmark scheme without RIS, the
passive RIS can realize only a limited sum-rate gain of about
22% in a typical application scenario, while the proposed A PPENDIX B
active RIS can achieve a substantial sum-rate gain of about P ROOF OF L EMMA 3
130%, thus indeed overcoming the fundamental limitation of According to the related analysis in [14] and Appendix A,
the “multiplicative fading” effect. In the future, many research the user’s achievable SNR for an SU-SISO system aided by a
directions for active RISs are worth pursuing, including passive RIS and that aided by an active RIS can be respectively
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1723
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
[35] H. A. Suraweera, I. Krikidis, G. Zheng, C. Yuen, and P. J. Smith, Xibi Chen (Student Member, IEEE) received the
“Low-complexity end-to-end performance optimization in MIMO full- B.S. and M.S. degree from Tsinghua University,
duplex relay systems,” IEEE Trans. Wireless Commun., vol. 13, no. 2, Beijing, China, in 2017 and 2020, respectively. He is
pp. 913–927, Feb. 2014. currently pursuing the Ph.D. degree with the Depart-
[36] P. Lioliou, M. Viberg, M. Coldrey, and F. Athley, “Self-interference ment of Electrical Engineering and Computer Sci-
suppression in full-duplex MIMO relays,” in Proc. Asilomar Conf. ence (EECS), Massachusetts Institute of Technology
Signals, Syst. Comput., Nov. 2010, pp. 658–662. (MIT), Cambridge, MA, USA. From 2015 to 2017,
[37] P. Xing, J. Liu, C. Zhai, X. Wang, and L. Zheng, “Self-interference he was a Research Assistant at the Department
suppression for the full-duplex wireless communication with large- of Electronic Engineering, Microwave and Antenna
scale antenna,” Trans. Emerg. Telecommun. Technol., vol. 27, no. 6, Institute, Tsinghua University, where he became a
pp. 764–774, Feb. 2016. Graduate Student Researcher, from 2017 to 2019.
[38] Further Advancements for E-UTRA Physical Layer Aspects (Release 9), His current research interests include terahertz (THz) integrated electronic
document TS 36.814, 3GPP, Mar. 2010. systems, THz imaging/sensing, and CMOS electromagnetics/optics. He was
[39] H. Guo, Y.-C. Liang, J. Chen, and E. G. Larsson, “Weighted a recipient of Analog Devices Outstanding Student Designer Award.
sum-rate maximization for reconfigurable intelligent surface aided
wireless networks,” IEEE Trans. Wireless Commun., vol. 19, no. 5,
pp. 3064–3076, May 2020.
[40] Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An iteratively weighted
MMSE approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59, Changhao Liu (Graduate Student Member, IEEE)
no. 9, pp. 4331–4340, Sep. 2011. received the B.S. degree in electronic engineering
[41] M. Heino et al., “Recent advances in antenna design and interference from Tsinghua University, Beijing, China, in 2021,
cancellation algorithms for in-band full duplex relays,” IEEE Commun. where he is currently pursuing the Ph.D. degree in
Mag., vol. 53, no. 5, pp. 91–101, May 2015. electronic engineering.
[42] M. Heino, S. N. Venkatasubramanian, C. Icheln, and K. Haneda, From 2019 to 2021, he was a Research Assistant
“Design of wavetraps for isolation improvement in compact in-band full- at the Microwave and Antenna Institute, Department
duplex relay antennas,” IEEE Trans. Antennas Propag., vol. 64, no. 3, of Electronic Engineering, Tsinghua University. His
pp. 1061–1070, Mar. 2016. current research interests include reconfigurable
[43] D. Korpi, M. Heino, C. Icheln, K. Haneda, and M. Valkama, “Compact metasurfaces, surface electromagnetics, reflectarray
inband full-duplex relays with beyond 100 dB self-interference antennas, transmitarray antennas, terahertz metasur-
suppression: Enabling techniques and field measurements,” IEEE Trans. faces, and reconfigurable intelligent surfaces.
Antennas Propag., vol. 65, no. 2, pp. 960–965, Feb. 2017.
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
ZHANG et al.: ACTIVE RIS VS. PASSIVE RIS: WHICH WILL PREVAIL IN 6G? 1725
Robert Schober (Fellow, IEEE) received the Communication Society (ComSoc) in 2020 and 2021. Currently, he serves
Diplom (Univ.) and the Ph.D. degrees in electrical as Member of the Editorial Board of the P ROCEEDINGS OF THE IEEE,
engineering from Friedrich-Alexander University of as Member at Large of the ComSoc Board of Governors, and as ComSoc
Erlangen-Nuremberg (FAU), Germany, in 1997 and Treasurer. He is the ComSoc President-Elect for 2023.
2000, respectively. From 2002 to 2011, he was
a Professor and Canada Research Chair at the
University of British Columbia (UBC), Vancouver,
Canada. Since January 2012 he is an Alexander
von Humboldt Professor and the Chair for Digital
Communication at FAU. His research interests fall H. Vincent Poor (Life Fellow, IEEE) received the
into the broad areas of Communication Theory, Ph.D. degree in EECS from Princeton University in
Wireless and Molecular Communications, and Statistical Signal Processing. 1977. From 1977 to 1990, he was on the faculty
Robert received several awards for his work including the 2002 Heinz of the University of Illinois at Urbana-Champaign.
Maier Leibnitz Award of the German Science Foundation (DFG), the 2004 Since 1990, he has been on the faculty with
Innovations Award of the Vodafone Foundation for Research in Mobile Princeton, where he is currently the Michael Henry
Communications, a 2006 UBC Killam Research Prize, a 2007 Wilhelm Strater University Professor. From 2006 to 2016,
Friedrich Bessel Research Award of the Alexander von Humboldt Foundation, he served as the Dean of the Princeton’s School of
the 2008 Charles McDowell Award for Excellence in Research from UBC, a Engineering and Applied Science. He has also held
2011 Alexander von Humboldt Professorship, a 2012 NSERC E.W.R. Stacie visiting appointments at several other universities,
Fellowship, a 2017 Wireless Communications Recognition Award by the including most recently at Berkeley and Cambridge.
IEEE Wireless Communications Technical Committee, and the 2022 IEEE His research interests are in the areas of information theory, machine learning
Vehicular Technology Society Stuart F. Meyer Memorial Award. Furthermore, and network science, and their applications in wireless networks, energy
he received numerous Best Paper Awards for his work including the 2022 systems, and related fields. Among his publications in these areas is the
ComSoc Stephen O. Rice Prize. Since 2017, he has been listed as a Highly recent book Machine Learning and Wireless Communications (Cambridge
Cited Researcher by the Web of Science. Robert is a Fellow of the Canadian University Press, 2022). He is a member of the National Academy of
Academy of Engineering, a Fellow of the Engineering Institute of Canada, Engineering and the National Academy of Sciences and a Foreign Member of
and a Member of the German National Academy of Science and Engineering. the Chinese Academy of Sciences, the Royal Society, and other national and
He served as Editor-in-Chief of the IEEE T RANSACTIONS ON C OM - international academies. He received the IEEE Alexander Graham Bell Medal
MUNICATIONS from 2012 to 2015 and as VP Publications of the IEEE in 2017.
Authorized licensed use limited to: UNIVERSITAETSBIBL STUTTGART. Downloaded on July 18,2023 at 05:53:30 UTC from IEEE Xplore. Restrictions apply.