I Introduction
Simultaneous wireless information and power transfer (SWIPT) is a new technology where information and energy flows coexist, coengineered to simultaneously provide communication connectivity and energy sustainability [1, 2]. It has been considered as a new promising solution to transmit information and energy to low power devices and to extend the battery lifetime of wireless networks, especially in wireless sensor networks and Internet of Things (IoT) applications. Compared to the traditional energy harvesting (EH) and green communication techniques, which collect energy from natural and manmade sources such as solar, wind or mechanical vibration, SWIPT can be fully controlled and optimized by harvesting energy from the radiofrequency (RF) signals. From the seminal work of Varshney [36], who introduced the concept of SWIPT and the fundamental tradeoff between information and energy transfer (i.e., informationenergy capacity region), substantial works appear in the literature that study SWIPT from different perspectives.
Initial works in SWIPT assume a linear channel to transfer both information and power and investigate sophisticated transmission techniques and/or receiver architectures for SWIPT. The work in [4] characterizes the fundamental tradeoff between information and energy for a basic multipleaccess channel with information and EH receivers; it has been shown that a feedback channel is a critical mechanism to increase the informationenergy capacity region. To enable practical techniques to convey information and power, the authors in [5] propose two practical receiver approaches, namely, a) “time switching (TS)”, where the receiver switches between decoding information and harvesting energy, and b) “power splitting (PS)”, where the receiver splits the received signal into two parts for decoding information and harvesting energy, respectively. Using the PS approach and multiple transmit antennas at the transmitter, the optimal multipleinput singleoutput (MISO) beamforming with both qualityofservice (QoS) and EH constraints was studied in [6] and [7] for an interference channel and a downlink MISO channel, respectively. To reduce the complexity of the SWIPT receivers, some studies overcome the use of RFtoDC circuits and achieve information transfer by embedding information into the shape of the transmitted signal rather than modulating radio waveforms. The work in [8] introduces the integrated receiver, where information is embedded in the amplitude of energy signals, while decoding is performed by taking samples at the output of the rectification circuit. The authors in [9] also exploit the concept of precoded spatial modulation and convey information both at the energy pattern of the transmitted signal and the index of the received antenna. More recent works take into account the nonlinearity of the rectification process and focus on the waveform design for SWIPT. By introducing a simple and tractable model of the diode nonlinearity (i.e., physicsbased diode model), the work in [10] deals with the design of waveforms that maximizes the DC power at the output of the rectifier. This work is extended in [11] to convey both information and energy by transmitting a superposition of unmodulated and modulated multitone signals.
On the other hand, wireless technologies are characterized by terminals that are subject to strict regulations on the level of RF radiation that users of the terminals are exposed to. Since RF radiation has been proven harmful for humans and the environment, these regulations minimize the potential biological effects (e.g., tissue heating, metabolic changes in the brain, carcinogenic effects) caused by RF radiation [12, 13]. Two widely adopted regulations/measures on RF exposure are mainly considered i.e., the first one is the maximum permissible exposure (MPE) in [W/], which is defined as the highest level of electromagnetic radiation (EMR) to which a user may be exposed without incurring an established adverse health effect. This is a particularly important issue in the applications of wireless power transfer (WPT) and various relevant studies have been carried out concerning EMR. For instance, the problem of scheduling the power chargers is investigated in [14] so that the charging utility for all rechargeable devices is maximized with a constraint on EMR. The works in [15] and [16] deal with the problem of maximizing the harvested energy and wireless charging tasks scheduling, respectively, when the transmitted signals guarantee a welldefined EMR constraint. The work [17] discusses the security issues in RFbased WPT systems and investigates techniques that ensure confidentiality, security and safety for different types of attacks. The second one is the specific absorption rate (SAR) that measures the absorbed power in a unit mass of human tissue by using units of Watt per kilogram [W/kg]. The Federal Communications Commission (FCC) enforces an SAR limitation of W/kg averaged over one gram of tissue on partial body exposure [18]; other regulatory agencies (e.g., Comité Européen de Normalisation Électrotechnique in European Union) also adopt a similar limitation of W/kg averaged over 10 grams of tissue on SAR measurements [19]. For short distances (e.g., the distance between the transmitter and the nearest user is less than cm), the SAR measure dominates the RF exposure and therefore becomes a more critical factor than the MPE for the design of efficient uplink communications. SAR is obviously different from MPE in that it is a different measure of RF exposure and applies to short distances. In addition, it needs to account for various exposure constraints on the whole body, partial body, hands, wrists, feet, ankles, etc., with various measurement limitations according to FCC regulations [18], therefore multiple SAR constraints are needed even for a single transmit device. For instance, the iPhone 6 Model A1549 has a whole body SAR of 1.14 W/kg and head SAR of 1.08 W/kg [20].
The integration of RF exposure constraints, and specifically the SAR constraint, in the design of wireless communication systems is a timely research area with significant impact; however, few works in the literature take into account SAR regulations. In conventional singleantenna wireless communication systems, the SAR exposure limitation simply introduces an additional transmit power constraint, and can be easily guaranteed by reducing the transmit power below a specific threshold. However, it becomes more involved when there are multiple transmit antennas and signals can be beamformed towards certain directions. In this case, the exposure limitation is coupled with the multiantenna channel vector/matrix, so it needs to be incorporated into the transmitter optimization. The measurements and simulations carried out in
[21] demonstrate SAR as a function of the phase different between two transmit antennas. The SAR reduction and modeling in multiantenna systems has been studied in [22, 23]. The SAR constraints are incorporated into the transmit signal design for a multipleinput multipleoutput (MIMO) uplink channel in [24], where the quadratic model for the SAR measurements is first proposed. In [25], a SAR code is proposed to handle a SARconstrained channel, which is shown to have significant improvement over the conventional Alamouti spacetime code under SAR limitations. It is also revealed that the SAR measurement is a function of the quadratic form of the transmitted signal with the SAR matrix. The SARaware beamforming and transmit signal covariance optimization methods are first presented in [26] and [27]. Capacity analysis with multiple SAR constraints on singleuser MIMO systems is intensively examined in [28]. Sumrate analysis for a multiuser MIMO system with SAR constraints is performed in [29] with both perfect and statistical channel state information (CSI). Form an information theoretical perspective, SARconstrained multiantenna transmit covariance optimization can be considered as the classical MIMO channel capacity optimization problem subject to generalized linear transmit covariance constraints studied in [30].The integration of SAR exposure constraints in the design of wireless power transfer or SWIPT systems is an unexplored research area [31]. Although SWIPT corresponds to a controlled transmission of RF radiation to communicate and energize, and may significantly contribute to the electromagnetic pollution (electrosmog), no work in the literature discusses the integration of SAR in SWIPT. For example, one of the main application areas of SWIPT is for medical devices in wireless body area networks, where an access point will support the communication connectivity and the power sustainability of a shortrange sensor network in, on, or around the human body [32, 33, 34]. Since the accesspoint may be placed within cm of the body, SAR measurements are mandatory and should be taken into account in the SWIPT design. To the best of our knowledge, the impact of the SAR on SWIPT has not been studied before.
In this paper, we focus on the PS approach and study the beamforming design in a multiuser MISO downlink channel, where the receivers are characterized by both communications QoS and EH constraints with additional SAR limitations; the nonlinearity of the rectification circuit is also taken into account. The contributions of this paper are summarized as follows:

We introduce and formulate the SARaware beamforming optimization problem with simultaneous QoS and EH requirements.

We derive the optimal beamforming solution in the general case by leveraging semidefinite programming (SDP) together with rank relaxation. More importantly, we prove that the proposed approach always gives rank solutions, such that the exactly optimal beamforming solutions can be derived. In the special case of a single receiver with a single SAR constraint, we propose a fast algorithm to achieve the optimal solution.

We develop lowcomplexity suboptimal solutions using both fixed beamforming schemes based on the criterion of zeroforcing (ZF) or regularized ZF (RZF), and a hybrid beamforming scheme that combines the ZF and maximum ratio transmission (MRT) to achieve a better performancecomplexity tradeoff.

Furthermore, when the CSI and/or the SAR matrices are imperfectly known at the transmitter, we devise a robust beamforming scheme suitable for imperfect CSI and SAR matrices with bounded errors, to guarantee the QoS and EH requirements, as well as the SAR constraints in the worse case.
In addition, we have made the following assumptions:
Notation: All boldface letters indicate vectors (lower case) or matrices (upper case). The superscripts , , , ,
denote the transpose, the conjugate transpose, the matrix inverse, the MoorePenrose pseudoinverse and the diagonal elements respectively. A circularly symmetric complex Gaussian random variable
with meanand variance
is represented as. The identity matrix of size
, and the zero matrix of size
, are denoted by and , respectively. and denote the and Fibonacci norms of a complex vector , denotes the magnitude of a complex variable , denotes the trace of a matrix , while indicates that matrix is positive semidefinite. denotes the imaginary unit.This paper is organized as follows: Section II sets up the system model and introduces the optimization problem considered. In Section III, we investigate the joint optimization of the optimal beamforming design and power splitting ratios. Section IV discusses solutions for the fixed and hybrid beamforming schemes, while Section V develops the robust solution under imperfect estimation of the channels and the SAR matrices. Simulation results are presented in Section VI, followed by conclusions in Section VII.
Ii System Model and Problem Formulation
We assume a MISO downlink channel consisting of an antenna transmitter (e.g., a base station (BS)) and singleantenna receivers that employ singleuser detection as illustrated in Fig. 1. This ensures simple receiver processing due to size limitation or limited processing power. We assume that the BS transmits with a total power and let be its transmitted data symbol to receiver
, which is Gaussian distributed with zero mean and unit variance, i.e.,
. The transmitted data symbol is mapped onto the antenna array elements by the beamforming/precoding vector . Other properties of the signal (e.g., modulation, waveform and input distribution) can also be optimized to improve the efficiency of RFDC conversion [10, 35], so the CSCG distribution of the input signal is in general suboptimal. However, the joint design of signal and beamforming is a challenging problem and out of the scope of this paper.All wireless links exhibit independent fading and additive white gaussian noise (AWGN) with zero mean and certain variances. The fading is assumed to be frequency nonselective block fading. This means that the fading coefficients in the vector channel remain constant during one slot, but may change independently from one slot to another. The mean and variance of the channel coefficients capture largescale degradation effects such as pathloss and shadowing. The received baseband signal at receiver can be expressed as
(1) 
where denotes the AWGN component with zero mean and variance . Therefore, the received power at receiver is equal to
(2) 
The receivers have RFEH capabilities and therefore can harvest energy from the received RF signal based on the power splitting technique [5], which is a mature SWIPT technique and does not require strict time synchronization between information and power transfer^{1}^{1}1 In case of nonlinear energy harvesting, neither the power splitting technique nor the time splitting technique always performs better. In the lower power regime, the time switching technique performs better than the power splitting technique [38].. With this approach, each receiver splits its received signal into two parts: a) one part is converted to a baseband signal for further signal processing and data detection, and b) the other part is driven to the required circuits for conversion to DC voltage and energy storage. Let denote the power splitting parameter for the th receiver; this means that of the received power is used for data detection, while the remaining amount is the input to the RFEH circuitry. More specifically, after reception of the RF signal at the receiver, a power splitter divides the power into two parts according to , so that is directed towards the decoding unit and towards the EH unit. During the baseband conversion, additional circuit noise, , is present due to phase offsets and nonlinearities which is modeled as AWGN with zero mean and variance .
The signaltointerferenceplusnoise ratio (SINR) metric characterizing the data detection process at the th receiver is given by
(3) 
On the other hand, the total power that can be harvested is equal to , where is a nonlinear parametric EH function. Note that in general should be a function rather than just the power of the energy signal. In this paper we adopt a simplified model to highlight the dependency on the power of the energy signal only. More details about will be discussed later.
As discussed in the introduction, wireless communication devices are subject to SAR limitations. Previous reported results such as [25] have shown that the pointwise SAR value with multiple transmit antennas can be modeled as a quadratic form of the transmitted signal, and the SAR matrix fully describes the SAR measurement’s dependence on the transmitted signals; entries of the SAR matrix have units of W/kg. Because the SAR measurements are always real positive numbers, the SAR matrices are positivedefinite conjugatesymmetric matrices.
Since SAR is a quantity averaged over the transmit signals, we model the th SAR constraint with a timeaveraged quadratic constraint given by
(4) 
where is the th SAR matrix and is the th SAR limit.
Iia Problem Formulation
We study the following problem of maximizing the ratios of the received SINR and EH over the target requirements, i.e., subject to the SAR and total power constraints, where are the SINR and the EH requirements, respectively. The choice of the objective function will balance the received SINR and the EH between users. To make the problem more tractable, we introduce an auxiliary variable , and formulate the optimization problem as follows
P1:  
where is the maximum total transmit power, and is the number of SAR constraints.
In this paper, we adopt a nonlinear parametric EH model, so the output DC power at the th receiver can be represented by the nonlinear function , where
is the input RF power. The nonlinear function can take many forms to capture the relationship between the input and output power at the energy receiver, such as the sigmoid function
[40], [41], the linear fraction [42], the piecewise linear function [43], the second order polynomial model [44], and a heuristic expression by curve fitting in
[45]. We use to denote a general nonlinear energy harvesting function; our analysis is general and independent of the specific function . In general, the nonlinear EH function is monotonically increasing, therefore we can find the inverse mapping , and the EH constraint (IIA) can be rewritten as(9) 
In general, it is difficult to solve the above problem P1, because both SINR and EH constraints (IIA) and (IIA) are nonconvex, and we also have additional multiple SAR constraints. Instead, we solve the following power minimization problem P2 below
(10)  
s.t.  
Clearly, the problem P2 is closely related to the problem P1. For instance, if P1 is solved and the optimal is achieved, then the same beamforming and power splitting solutions are also optimal for P2 to achieve the same SINR and EH, and the optimal minimum power will be . Because P1 is a quasiconvex problem in , once P2 is solved, P1 can be solved via a bisection search algorithm over as follows (note that in P1 is embedded in of P2):
Algorithm 1 to Solve P1:

Set the upper and lower bounds of as and . Repeat the following steps until convergence.

Calculate , and solve P2 with and .

If P2 is feasible and , ; otherwise .
Therefore, in the rest of the paper we will focus on solving the problem P2.
Iii The Optimal Scheme
In this section, we investigate the optimal beamforming and power splitting solutions to P2, and we begin with developing a fast solution for the special case of P2 referring to a singlereceiver system with a single SAR constraint, i.e., .
Iiia Fast Solution to the Special Case of
When we have only a single user and a single SAR constraint, the problem P1 reduces to the problem below, where the receiver and SAR indices are dropped for the sake of simplicity,
(11)  
s.t.  (14)  
We find the following Lemma is useful to derive the optimal solution to P1.
Lemma 1
Proof: It can be proved by contradiction. If (14) is not tight, can be reduced without affecting (14); if (14) is not tight, can be increased without affecting (14). This completes the proof.
Base on Lemma 1, we know that the optimal solution to P1 falls into the following two cases:

Case I: Only the SINR and the EH constraints are satisfied with equality.
In this case, we first solve the optimal . From the equality constraint (14), we can get
(15) Substituting (15) into (14), we can have the following equation, i.e.,
(16) which leads to a quadratic equation of given by
(17) Its solution is written as
(18) where . It is easy to verify that is between 0 and 1, because the left hand side of the quadratic equation (17) is strictly negative when and strictly positive when . Note that is the optimal solution regardless of as long as P3 is feasible.
Once is found, the optimal can be solved by satisfying the equality (15) and minimizing the objective simultaneously, which is given by
(19) After solving , we can substitute it into (14) and check whether the SAR constraint is satisfied: if so, is the optimal beamforming solution; otherwise, we need to consider the next case.

Case II: All three SINR, EH and SAR constrains are satisfied with equality.
In this case, we can simplify P3 to the following problem where is given by (18)
(20) s.t. (21) We keep the inequality constraints for the convenience of the derivation. In general, it is difficult to find the optimal solution to P4 directly. We first study the analytical beamforming structure that facilitates the development of an efficient solution.
The Lagrangian of P4 is given by
(22) where and are dual variables. The dual problem is then formulated as:
(23) s.t. (24) and the optimal admits the following form
(25) Let the eigenvalue decomposition of
be , whereis a unitary matrix and
is a diagonal matrix with elements being the eigenvalues of . Then we have , and(26) and
(27) where is defined in P4. From (27), we can solve
(28) Now we proceed to consider the constraint (21) and define the function below:
(29) where is due to (26), is used to obtain , and we use (28) to get .
Clearly, the parameter that satisfies uniquely determines the optimal solution. We find the following theorem useful to develop an efficient numerical solution to solve .
Theorem 1
in (29) is a decreasing function in .
Proof: Suppose and . Then after some algebraic manipulation, becomes
(30) From (30), it can be seen that is a decreasing function in , and this completes the proof.
Theorem 1 can be verified by the simulation result in Fig. 2.
Based on Theorem 1, we propose the following bisection search algorithm to solve .
Algorithm 2 to Solve .

Initialize and . Repeat the following steps until convergence.

Set .

If , ; otherwise .

IiiB The Optimal Solution using SDP in the General Case
The general case of the problem P2 is nonconvex and difficult to solve. In this section, we develop an efficient SDP based algorithm that jointly optimizes the beamforming vectors and power splitting parameters. To tackle P2, we first introduce new matrix variables , and then P2 can be reformulated as
(31)  
s.t.  (32)  
The problem P5 is convex because it is linear in all , and both terms and are convex in . It can be efficiently solved using numerical software package such as CVX [52]. Once P5 is optimally solved, if the resulting solutions are all rank1, then they are the exactly optimal solutions; otherwise, the solutions only provide a lower bound for the minimum required transmit power.
It has been proved in [7] that without the SAR constraints (32), the above SDP with rank relaxation will produce rank1 solutions which means they are also optimal to the problem P2. However, whether the SDP with rank relaxation can generate the optimal solution highly depends on the problem structure. With the additional SAR constraints, it is unknown whether this property remains true. In the following theorem, we prove that this is still the case.
Theorem 2
The optimal solution to P5 satisfies , i.e., the SDP relaxation is tight, and the optimal solution to the problem P2 can be recovered from via eigenvalue decomposition.
The proof is given in Appendix A.
Iv Suboptimal beamforming schemes
Although the algorithm based on SDP in Section III ensures the optimal beamforming and power splitting solutions, its complexity is high. In this section, we focus on lowcomplexity suboptimal solutions including both the fixed heuristic beamforming schemes and the hybrid beamforming scheme that considers a linear combination of fixed beamforming strategies.
Iva Solutions using Fixed Beamforming
Suppose the fixed beamforming vector is given by , where is the power for the th receiver. Let denote the link gain between the BS and the th receiver, and denote the radiation channel gain due to the transmission intended for the th receiver.
Using the fixed beamforming and rearranging the terms in the problem P2, we have
P6: min{pk,ρk} && ∑k=1Kpk
s.t. (1γk+1) Gk,k pk & ≥& 1ρkNC + ∑j=1KGk,jpj + N0 , ∀k,
∑j=1KGk,jpj +N0 & ≥& λk1ρk, ∀k,
pk ≥0, && 0 ≤ρk ≤1, ∀k,
∑k=1K pk Fk,l &≤& Pl, ∀l.
The optimization problem P6 is convex because it is comprised of a linear objective function and convex constraints. Constraint (IVA) is convex because the term is convex for and the other terms are linear, (IVA) is a restricted hyperbolic constraint and the term is also convex. Note that by solving the problem P6, we obtain optimal values for the splitting parameters, as well as an optimal power allocation for any given fixed beamforming vectors.
Commonly used fixed beamforming vectors include the MRT, ZF and RZF criteria [51], defined respectively below:
(33)  
(34)  
(35) 
where we have defined , and . The MRT beamforming and the ZF beamforming aim to enhance the received signal strength and remove interference, respectively, while the RZF beamforming provides a tradeoff between them. Note that in the RZF beamforming, the SAR matrices are included to incorporate the radiation constraints.
As a special case, for the ZF beamforming, the optimization problem can be further simplified because , for . Hence the optimization problem P6 simplifies into the following formulation
(36)  
The problem P7 is also convex but is much easier to solve than P6 because of the simplified constraints. Unlike the previous works [6, 7] where closedform solutions to P7 were derived, the SAR constraint does not permit a closedform solution, therefore we use CVX [52] to solve both P6 and P7.
IvB Hybrid Beamforming
In this subsection, we introduce the hybrid beamforming scheme to provide a tradeoff between enhancing the received SINR and energy harvested, dealing with interference, and accounting for the SAR constraints, which admits the following expression
(37) 
where and are combining coefficients.
Then
(38) 
Define , , and , then
The transmit power intended for the th receiver is
(39)  
where , and we have defined and .
The th SAR power can be reformulated as
(40)  
where we have defined and .
Then the power minimization problem P2 becomes
P8:  (41)  
s.t.  
The problem P8 is not convex because of the nonconvex term introduced by the EH and the SAR constraints, which makes the problem more challenging to solve. To deal with this difficulty, we introduce and relax it to or , then we have the following secondorder cone programming (SOCP) problem formulation
P9:  (42)  
s.t.  (43)  
The problem P9 is convex and can be optimally solved. We observe that in most cases, holds true which means it is also the optimal solution to P8; otherwise, we will employ (37) to form the hybrid beamforming, and then use the fixed beamforming scheme in Section IV.A to find the optimal power vector and power splitting by solving P6.
Complexity Analysis: The complexity of the optimally solution to the general problem P5 is dominated by the SDP constraints, and according to [53, 6.6.3], the complexity of the interiorpoint algorithm for solving P5 is . While the for solving the fixed beamforming optimization problem P6 and the hybrid beamforming problem P9, their complexities are dominated by the linear and SOCP constraints, and can be expressed as [53, 6.6.1] and [53, 6.6.2], respectively. It can be clearly seen that the suboptimal beamforming schemes reduces the complexity significantly when compared to the optimal beamforming.
V Robust beamforming schemes
In this section, we study the robust beamforming design, when the channel information and the SAR matrices are imperfectly known due to estimation and measurement errors. Without a robust solution, the SINR constraints and the EH constraints may not be satisfied; in addition, the SAR constraints may be violated.
We model the th receiver’s actual channel as
(44) 
where denotes the CSI estimate known to the BS. represents the CSI uncertainty that belongs to the set below
(45) 
where denotes the error bound.
Similarly, the th SAR matrix is modelled as
(46) 
where is know to the BS. is the SAR uncertainty within the set below
(47) 
where denotes the error bound.
We assume that the BS has no knowledge about the error channel vectors or the error SAR matrices, except for their error bounds and . Thus we take a worstcase approach for the beamforming design to guarantee the resulting solution is robust to all possible channel and SAR uncertainties within the given error sets. The specific robust design problem is to minimize the overall transmit power for ensuring the receivers’ worstcase individual SINR, EH and SAR constraints, i.e.,
P10:  (48)  
s.t.  (49)  
We first ignore the rank constraint (49) and define . Then P10 becomes
P11:  (50)  
s.t. 
Comments
There are no comments yet.