RESOURCE ALLOCATION FOR OFDM-BASED DOWNSTREAM MULTI-USER SIMULTANEOUS WIRELESS INFORMATION AND POWER TRANSFER
Richard Ketcham, Jeroen Verdyck, and Marc Moonen KU Leuven, Department of Electrical Engineering (ESAT),
STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics
ABSTRACT
This paper considers a resource allocation method to simultaneously optimize the beamforming vectors and power allocation for down- stream multi-user simultaneous wireless information and power transfer (SWIPT) systems that employ orthogonal frequency di- vision multiplexing (OFDM). The method utilizes a novel MAC- BC duality to simplify an optimization problem composed of a weighted-sum of data rates subject to power harvesting and trans- mit power constraints. A simple method to dynamically choose the minimum harvesting power instead of using a constant value is also introduced. In simulations, it is observed that the resulting rate-power region appears to be convex.
Index Terms— SWIPT, OFDM, Resource Allocation, Opti- mization, Precoding
1. INTRODUCTION
This paper presents a method to simultaneously optimize the beam- forming vectors and power allocation for downstream multi-user SWIPT systems that employ OFDM. The concurrent nature of SWIPT makes optimal resource allocation challenging and an area of ongoing research, some of which examines using OFDM ( [1], [2], [3]). Most similar to our work, [2] considers a network using beamforming vectors, but does not optimize them and instead assumes the beamforming vectors are given. This assumption may lead to a suboptimal solution since the design of the beamforming vectors changes significantly depending on if the system seeks to promote information transfer, power transfer, or a combination of the two. There has been work in which the beamforming vector op- timization and power allocation is jointly considered for the single carrier case ( [4], [5], [6]), but to the authors’ knowledge beam- forming optimization has not been considered for multi-user SWIPT networks using OFDM.
OFDM enables the use of a Lagrange dual-based algorithm since multicarrier systems have zero duality gap, which then enables the use of a MAC-BC duality. First an optimization problem is formu- lated in which a weighted-sum of the data rates is maximized subject This research work was carried out at the ESAT Laboratory of KU Leu- ven, in the frame of:
Fonds de la Recherche Scientifique - FNRS and the Fonds Wetenschap- pelijk Onderzoek - Vlaanderen under EOS Project no 30452698 ’(MUSE- WINET) MUlti-SErvice WIreless NETwork’
Research Project FWO nr. G.0B1818N ’Real-time adaptive cross-layer dynamic spectrum management for fifth generation broadband copper access networks’
The scientific responsibility is assumed by its authors.
to power harvesting and transmit power constraints. A novel simpli- fication via a MAC-BC duality is then utilized to solve the Lagrange dual of the resource allocation problem. Lastly, a simple method to dynamically choose the minimum harvesting power is utilized in the simulations.
2. SYSTEM MODEL
A SWIPT system using OFDM is considered, where the OFDM symbols are comprised of K orthogonal sub-carriers. The envi- sioned system is modeled by a network containing one access point using N T antennas to deliver data and power to N remote devices each with one antenna, where N ≤ N T . Additionally, it is assumed that there is no inter-carrier interference present, which allows each sub-carrier to be modeled independently and that perfect channel state information is available.
2.1. Downstream Channel
The access point applies a precoder to the symbol vector before transmitting to individual remote devices. This is commonly referred to as the broadcast channel (BC) and may be modeled on each sub- carrier (denoted with subscript k) as
y k = H k H T k x k + z k . (1) Where H H k is the the N × N T channel matrix, with [H k ] H m,n being the transfer function between receiver n and transmitter m, evaluated at sub-carrier k. Moreover, z k is the N × 1 additive white Gaussian noise, x k is the N ×1 vector of transmit symbols, T k is the N T ×N precoder matrix, and y k is the N × 1 vector of symbols as received by the remote devices.
2.2. Performance Metrics
For information transfer, the downstream data rate (DR) for user n on sub-carrier k may be modeled using Shannon’s channel capacity equation divided by the sub-carrier bandwidth:
α n log 2
1 + γ DS k,n (T k , s k )
. (2)
The downstream SINR for user n on sub-carrier k, γ DS k,n (T k , s k ), is
a function of the precoder matrix, T k , and symbol power (SP) vec-
tor, s k . The SP for user n (i.e. the n-th element of s k ) is denoted by
s k,n , where s k,n , E |x k,n | 2 with E [·] being the expected value
operator. The term α n is ∈ [0, 1] and is included in the formulation
to reflect reduced efficiencies due to the receiver architecture type
(e.g. time-splitting), but will be omitted (i.e. effectively assumed
α n = α ∀ n) in Section (3) due to space. The downstream SINR is given by:
γ k,n DS (T k , s k ) = |t H k,n h k,n | 2 s k,n
P N m=1
m6=n |t H k,m h k,n | 2 s k,m + σ k,n
, (3)
where t k,n is the precoder vector for user n (i.e. the n-th column of T k ). The transfer function for user n (i.e. the n-th column of H k ) is h k,n . The additive white Gaussian noise power spectral density experienced by user n and sub-carrier k is given by σ k,n .
The power harvested by user m on sub-carrier k is
P k,m (T k , s k ) = (1 − α m )
N
X
n=1
|t H k,n h k,m | 2 s k,n , (4)
where α m is the same as α n . The total harvested power (HP) for user m is P k,m (T k , s k ) summed ∀ k. For the allocation over the entire OFDM symbol, the total transmit power constraint must also be considered.
3. OPTIMUM RESOURCE ALLOCATION 3.1. Primal Maximization
The resource allocation problem is formulated as the maximization of a weighted sum of DRs over all users and sub-carriers, subject to the total transmit power constraint P T otal and the minimum per-user harvested power C m , and is given as
maximize
{T k ,s k } ∀k K
X
k=1 N
X
n=1
ω n log 2
1 + γ k,n DS (T k , s k )
Subject to
K
X
k=1
P k,m (T k , s k ) ≥ C m ∀ m ∈ [1, ..., N ]
K
X
k=1 N
X
n=1
s k,n ||t k,n || 2 2 ≤ P Total .
(5)
The optimal point is chosen by selecting a precoder matrix (T k ∀ k) and symbol power vector (s k ∀ k) which maximizes the weighted sum. The weights ω n are used to weight the importance of individual user DRs as determined by the use-case.
3.2. Lagrange Dual
Dual decomposition is used to split the primal problem in (5) into K smaller problems to be solved in parallel, and enables the use of MAC-BC duality theory. The dual problem is formulated by dualiz- ing the constraints and combining them with the objective function.
The Lagrange dual function is defined as
q(λ) , maximize T ,s L(s, T, λ) (6) where the Lagrangian, L(s, T, λ), is
L(s, T, λ) = λ 0 P Total −
N
X
m=1
λ m C m +
K
X
k=1 N
X
n=1
ω n log 2 (1 + γ k,n DS (T k , s k )) −
K
X
k=1 N
X
n=1
t H k,n (λ 0 I −
N
X
m=1
λ m (h k,m h H k,m ))t k,n s k,n , (7)
and λ = [λ 0 , ..., λ N ] T . Additionally we will define Υ k (λ) as:
Υ k (λ) ,
λ 0 I −
N
X
m=1
λ m h k,m h H k,m
. (8)
If Υ k (λ) is not positive definite for all k then (6) will become un- bounded and q(λ) will be +∞, in which case the corresponding λ cannot be dual optimal. The Lagrange dual problem is then given by
minimize
λ q(λ)
Subject to λ ≥ 0
Υ k (λ) 0 ∀ k.
(9)
Note that the primal problem in (5) is not convex and so the duality gap may be non-zero. A non-zero duality gap would mean that a solution to the dual problem is not guaranteed to lead to an optimal solution for the primal problem. However, [7] indicates that the time-sharing property of a practical multi-carrier system provides for zero duality gap.
3.3. Dual Function Evaluation
MAC-BC duality allows the construction of a multiple access chan- nel (MAC) problem that has a known solution [8] linked to the so- lution of the BC maximization problem (∀ k) in (6). The duality asserts that provided Υ k (λ) 0, then for any positive s k and T k
there exists a dual SP vector (δ k ) that achieves the same SINR in the MAC (upstream channel H k with encoder T H k ) as in the BC (down- stream channel H H k with precoder T k ). In our case, we will convert our BC problem to a MAC problem by equating the SINRs, optimize over the dual SPs, solve for the optimal equalizer, and then convert back to attain the optimal precoder. Equating the SINRS results in
γ DS k,n (T k , s k ) = γ k,n US (T k , δ k ), ∀ k and n, (10) where γ k,n DS (T k , s k ) is defined in (3) and the dual upstream (US) SINR is given by:
γ US k,n (T k , δ k ) = |t H k,n h k,n | 2 δ k,n
P N m=1
m6=n |t H k,n h k,m | 2 δ k,m + t H k,n Υ k (λ)t k,n
. (11) In (11), δ k is the SP vector in the US system and Υ k (λ) is seen to act as a noise covariance matrix. Since Υ k (λ) 0, (10) may be expressed as:
Z k s k = σ k ◦ δ k , (12) where ◦ is the Hadamard product and Z k is given by:
[Z k ] ij =
N
X
n=1 n6=j
|t H k,j h k,n | 2 δ k,n + t H k,j Υ k (λ)t k,j , for i = j
[Z k ] ij = −δ k,i |t H k,j h k,i | 2 , for i 6= j.
(13)
Z k is an M-matrix [9]. Therefore, it is guaranteed that for any δ k ≥ 0, solving (12) yields a s k ≥ 0. Summing the rows of (12) yields:
N
X
n=1
t H k,n Υ k (λ)t k,n s k,n = σ H k δ k . (14) The maximization problem (∀ k) in (6) may then be simplified fur- ther by redefining the problem in terms of the dual upstream DR and using (14) to arrive at:
maximize
T k ,δ k N
X
n=1
ω n log 2 (1 + γ US k,n (T k , δ k )) − σ H k δ k . (15)
3.4. Equalizer Optimization
Equation (15) is separable when maximizing over T k (given δ k ) since the SINR of user n (given by γ k,n US ) only depends on the receive filter (t k,n ) of user n and σ k H δ k does not. This, along with the monotonicity of log(1 + ·), allows us to only consider γ k,n US (T k , δ k ) in the maximization:
maximize
T k γ US k,n (T k , δ k ). (16) The optimal linear equalizer for a MAC is the linear MMSE equal- izer and, given the US SINR in (11), is readily shown to be:
t ∗ k,n = (
N
X
m=1 m6=n
δ k,m h k,m h H k,m + Υ k (λ)) −1 h k,n . (17)
The expression in (17) enables the dual upstream SINR to be ex- pressed as only a function of δ k , i.e.
γ k,n US (δ k ) = δ k,n h H k,n
N X
m=1 m6=n
δ k,m h k,m h H k,m + Υ k (λ)
−1
h k,n .
(18) 3.5. Dual Symbol Power Optimization
Successive convex approximation [10] is employed to optimize the dual SP vector (δ k ). Successive approximations of (15) are made by fixing constant the dual upstream SP values δ k,m of all users m 6= n and by approximating the DRs with their first order Taylor expan- sion for all users m 6= n. This leads to the maximization being approximated with:
maximize
δ k,n ω n log 2 (1 + δ k,n q ¯ k,n ) − ( ¯ φ k,n + σ k,n )δ k,n
Subject to δ k,n ≥ 0,
(19)
where ¯ q k,n is the interference-plus-noise covariance,
¯
q k,n = h H k,n
N
X
m=1 m6=n
¯ δ k,m h k,m h H k,m + Υ k (λ)
−1
h k,n , (20)
φ ¯ k,n is from the first order Taylor series expansion,
− ¯ φ k,n = ∂
∂δ k
[
N
X
m=1 m6=n
ω m log 2 (1 + γ k,m US (δ k ))]
δ k =¯ δ k
, (21)
and σ k,n is the noise. Terms that are constant over the maximiza- tion were removed from the approximation. KKT conditions enable calculating the optimal value for δ k,n via the following equation:
δ ∗ k,n =
"
0,
ω n log(2)
φ ¯ k,n + σ n
− 1
¯ q k,n
# +
, (22)
where [x, y] + = max{x, y}.
3.6. Solving the Dual Problem
Since (9) is convex, the deep-cut ellipsoid algorithm can be used to find values for λ as described in [11]. Once λ is selected, the resource allocation solves for the optimal dual SP, which is then used to calculate the optimal precoder and SP for the BC. This is done iteratively until convergence.
3.6.1. Ellipsoid Initialization
An ellipsoid E can be described as E = {z | (z−λ) T P −1 (z−λ) ≤ 1}, where λ is the center of the ellipsoid and P gives the size and shape of E. The initial ellipsoid may be defined with P = λ 2 U B I, where λ U B is an N +1×1 vector of upper-bounds for the optimal λ.
The center for each successive ellipsoid defines λ for each iteration.
3.6.2. Upper-bound for Lagrange Multipliers
Since q(λ) is minimized when λ = λ ∗ , any other λ would result in a larger value for q(λ), yielding
q(λ) ≥ q(λ ∗ ). (23)
Similarly, if s = s ∗ and T = T ∗ maximizes L(s, T, λ) then any other s and T (provided that s ≥ 0) will yield a smaller L, giving
L(s, T, λ) ≤ L(s ∗ , T ∗ , λ). (24) Using an approach detailed in [12] (p. 43), (23) and (24) are refor- mulated to solve for the λ ∗ upper-bound. The equation is further modified by plugging in values for a feasible s = s 0 and T = T 0 as well as the solution for q(λ) when λ = [1, ..., 1] T . This yields an expression for the λ ∗ m upper-bound:
λ ∗ m ≤ q(1) − P K k=1
P N
n=1 ω n log 2 (1 + γ k,n DS (T k 0 , s 0 k )) P m (s 0 , T 0 ) − C m
, (25) as well as one for the λ ∗ 0 upper-bound:
λ ∗ 0 ≤ q(1) − P K k=1
P N
n=1 ω n log 2 (1 + γ k,n DS (T k 0 , s 0 k )) P T otal − P K
k=1
P N
n=1 s 0 k,n ||t 0 k,n || 2 2 . (26) Feasible s 0 and T 0 are found by minimizing the transmit power subject to meeting the constraints to the primal problem (5), which can be posed as a semidefinite program:
minimize
{Q H k } ∀k K
X
k
< Q H k , I >
Subject to
K
X
k=1
< Q H k , (h k,m h H k,m ) > ≥ C m ∀ m Q k < 0 ∀ k,
(27)
where Q k , P N
n t k,n s k,n t H k,n = T k s k I NxN T H k and < A, B > , trace(A H B) is the Frobenius inner product. By setting s 0 k = 1 ∀ k, T 0 k may be found by factoring Q k . The upper-bound for λ can then be calculated and used to initialize the ellipsoid algorithm.
3.6.3. Ellipsoid Cuts
The deep-cut ellipsoid method iteratively reshapes the ellipsoid us- ing either a feasibility cut or an objective cut. A feasibility cut is made when a dual constraint (specified in (9)) is violated and is based on the normalized subgradient of the violated constraint. En- tries associated with the non-violated constraints are set to zero. The Lagrange multiplier constraint function is of the form:
m n (λ) = − 0, ..., λ n , ..., 0 T
. (28)
The subgradient of the Lagrange multiplier constraint function is of the form:
∇m n (λ) = 0, ..., −1, ..., 0 T . (29)
The constraint associated with Υ k (λ) is equivalent to
f k (λ) = λ 0 − max{ eigenvalues {
N
X
m
λ m h k,m h H k,m }} ≥ 0 ∀ k.
(30) If υ σ is the normalized eigenvector related to the maximum eigen- value of P N
m λ m h k,m h H k,m , then the subgradient [13] of f k (λ) is:
∇f k (λ) =
∂f (λ)
∂λ 0
∂f (λ)
∂λ 1
.. .
∂f (λ)
∂λ N
=
1
−||h H k,1 υ σ || 2 2
.. .
−||h H k,N υ σ || 2 2
. (31)
An objective cut is made when the dual constraints are met but the primal constraints are not. The objective cut is based on the subgradient of the objective function (6) and is:
∇L(s, T, λ) =
P tot − P K k=1
P N
n=1 t H k,n t k,n s k,n
−C 1 + P K k=1
P N
n=1 t H k,n (h k,1 h H k,1 )t k,n s k,n
.. .
−C N + P K k=1
P N
n=1 t H k,n (h k,N h H k,N )t k,n s k,n
. (32)
The optimal s and T are obtained via a method outlined in (3.3).
4. SIMULATIONS
A network of remote devices receiving data and power from one access point is modeled to show the resource allocation algorithm meeting its data rate and energy consumption needs. A simple method from queue stability theory is used to determine the DR weights, ω n , and the HP constraints, C n . In particular, instead of predetermining a constant C n for each device, the perturbed queue size is used to dynamically choose these values.
4.1. Network Model
The network operates in slotted time and each device has a data queue and energy queue. The data queues are located at the access point and store data before transmitting to a specific device. The size of the data queue at slot t + 1 for device n may be modeled with
Q n (t + 1) = [0, Q n (t) − R n (t)] + + A n (t), (33) where Q n (t) is the value of the queue in bits, R n (t) is the amount of data transmitted to the device, and A n (t) is the amount of data added to the queue. An energy queue is located at each remote device n for energy storage, the size of which at slot t + 1 may be modeled as
B n (t + 1) = [0, B n (t) − D n (t)] + + E n (t), (34) where B n (t) is the value of the queue in joules, D n (t) is the amount of energy consumed by the device, and E n (t), given by
E n (t) = τ
K
X
k=1
P k,n (T k (t), s k (t)), (35) is the amount of energy added to the queue. τ is the slot duration in seconds. T k (t) and s k (t) are determined at the beginning of slot t. Neither queue type is allowed to be negative. The data arrival rates, A n (t), follow a Poisson distribution with a mean value that is within the rate region. The energy consumption rates, D n (t), are kept constant.
4.2. Setting Resource Allocation Parameters
Max-Weight scheduling has been shown to stabilize network queues for systems using a linear sum of weighted attributes for resource al- location [14], which is the type of formulation this paper uses. Based on Max-Weight scheduling, the DR weights are assigned the queue size, i.e.
ω n (t + 1) = [1, Q n (t + 1)] + ∀ n, (36) and the minimum HP is assigned the perturbed queue size, i.e.
C n (t + 1) = θ n − B n (t + 1) ∀ n, (37) where θ n is a target value [15]. Other parameters are in the footnote 1 . 4.3. Results
60 80 100 120 140 160 180 200
Avg. Received Power, W
00.5 1 1.5 2 2.5 3
Avg. Data Rate (bps/Hz)
Average Rate - Average Power Region
(a) Projected Paretto Edge
0 50 100 150 200 250 300 350 400
Received Power, W
00.5 1 1.5 2 2.5 3 3.5
Data Rate (bps/Hz)
Average Rate - Average Power Region
Avg. User Resource Allocation User Arrival/Consumption Points Avg. Arrival/Consumption Point
(b) Simulation Operating Points
Fig. 1. Projected Rate-Power Region
0 10 20 30 40 50 60 70 80 90 100
Time Slots 0.9996
0.9997 0.9998 0.9999 1 1.0001 1.0002 1.0003 1.0004 1.0005 1.0006
Energy Queue Size (Joules)
10-3 Energy Queue Size
User=1 User=2 User=3 User=4 User=5 User=6 Average Energy Queue
(a) Energy Queue
0 10 20 30 40 50 60 70 80 90 100
Time Slots 0
1 2 3 4 5 6
Data Queue Size (Bits)
104 Data Queue Size Over Time
User=1 User=2 User=3 User=4 User=5 User=6 Average Data Queue