Task and Bandwidth Allocation for UAV-Assisted Mobile Edge Computing with Trajectory Design

Extracted source text · arXiv 1907.03862

\documentclass[journal]{IEEEtran}
%\documentclass[12pt,draftclsnofoot,onecolumn]{IEEEtran}
%\documentclass[12pt,draftclsnofoot, onecolumn,journal]{IEEEtran}
%\renewcommand{\baselinestretch}{0.94}
%\setlength{\intextsep}{8pt plus 2pt minus 2pt}
\pagenumbering{gobble}
\usepackage{stfloats}
\usepackage{makecell}
\usepackage{graphicx}
\usepackage[cmex10]{amsmath}
\usepackage{cases}
\usepackage[tight,footnotesize]{subfigure}
\usepackage{amsthm} %proof
\usepackage{cite}
\usepackage{citesort}
\usepackage{amssymb}
\allowdisplaybreaks[4]
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{multirow}
\usepackage{amsmath}
\usepackage{xcolor}
\usepackage{CJK}
\usepackage{subeqnarray}
\usepackage{cases}
\usepackage{enumerate}

\usepackage{stfloats}

%\DeclareMathOperator*{\argmin}{argmin}         //argmin»òargmax¹«Ê½µÄÅÅ°æ

\newtheorem{definition}{\hskip\parindent\bf{Definition}}%[section]
\newtheorem{lemma}{\hskip\parindent\bf{Lemma}}%[section]
\newtheorem{proposition}{\hskip\parindent\bf{Proposition}}
\newtheorem{theorem}{\hskip\parindent\bf{Theorem}}
\newtheorem{remark}{\hskip\parindent\bf{Remark}}
\newtheorem{corollary}{\hskip\parindent\bf{Corollary}}

\ifCLASSINFOpdf
\else
\fi

\hyphenation{op-tical net-works semi-conduc-tor}
%\linespread{1.6}

\begin{document}
%\baselineskip 23.5pt
%\baselineskip 20pt

%\title{\LARGE{UAV-Assisted Relaying and Edge Computing: Scheduling and Trajectory Optimization}}
%\title{\LARGE{Task and Bandwidth Allocation for Mobile Edge Computing with the Assistance of UAV}}
\title{\LARGE{Task and Bandwidth Allocation for UAV-Assisted Mobile Edge Computing with Trajectory Design}}

\author{\normalsize{Xiaoyan Hu$^{\dag}$, Kai-Kit Wong$^{\dag}$, Kun Yang$^{\ddag*}$ and Zhongbin Zheng$^{\S}$} \\
$^{\dag}$Department of Electronic and Electrical Engineering, University College London, London, UK\\
$^{\ddag}$School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK \\
$^{*}$School of Communication and Information Engineering, UESTC, Chengdu, P. R. China\\
$^{\S}$East China Institute of Telecommunications, Shanghai, P. R. China\\
Email:$^{\dag}$\{xiaoyan.hu.16, kai-kit.wong\}@ucl.ac.uk, $^{\ddag*}$kunyang@\{essex.ac.uk, uestc.edu.cn\},$^{\S}$ben@ecit.org.cn
}
\maketitle
\vspace{-2cm}
\begin{abstract} %cellular-based
In this paper, we investigate a mobile edge computing (MEC) architecture with the assistance of an unmanned aerial vehicle (UAV). The UAV acts as a computing server to help the user equipment (UEs) compute their tasks as well as  a relay to further offload the UEs' tasks to the access point (AP) for computing. The total energy consumption of the UAV and UEs is minimized by jointly optimizing the task allocation, the bandwidth allocation and the UAV's trajectory, subject to
the task constraints, the information-causality constraints, the bandwidth allocation constraints, and the UAV's trajectory constraints. The formulated optimization problem is nonconvex, and we propose an alternating algorithm to optimize the parameters iteratively. The effectiveness of the algorithm is verified by the simulation results, where great performance gain is achieved in comparison with some practical baselines, especially in handling the computation-intensive and latency-critical tasks.
%The required optimization is nonconvex, and an alternating optimization algorithm is proposed to jointly optimize the computation resource scheduling, bandwidth allocation, and the UAV's trajectory in an iterative fashion. Numerical results demonstrate that significant performance gain is obtained over conventional methods. Also, the advantages of the proposed algorithm are more prominent when handling computation-intensive latency-critical tasks.
\end{abstract}
%\vspace{-0.2 cm}
% Note that keywords are not normally used for peerreview papers.
\begin{IEEEkeywords}
Mobile edge computing, UAV, task allocation, bandwidth allocation, trajectory design.
\end{IEEEkeywords}

\IEEEpeerreviewmaketitle

\vspace{-4mm}
\section{Introduction}\label{sec:Introduction}
With the prevail of smart devices, various mobile applications spring up which demand  more and more computing resources and will definitely increase the burden of the user equipment (UEs). % and the increasingly complex mobile applications, the computing demands at user equipment (UEs) are reaching an unprecedented level.
Mobile edge computing (MEC) aiming at liberating the resource-limited UEs from heavy computing workload through computation offloading, has been regarded as a key technology of future intelligent wireless networks \cite{S_P.Mach17Mobile,S_Y.Mao17ASurve}.
%The standardization organizations and industry associations such as  ETSI and 5GAA have identified several use cases for MEC, from the  intelligent video acceleration and application-aware performance optimization to vehicle-to-everything and massive machine-type communications, etc. \cite{W_Y.Hu15Mobile,5GAA-VISUAL}.
%The rationale behind MEC is that UEs' computing tasks can be offloaded and completed at the edge of wireless networks by deploying cloud servers at the access points (APs), so as to liberate the UEs from heavy computing workloads and prolong their battery lifetime~\cite{S_P.Mach17Mobile,S_Y.Mao17ASurve}.
Recently, MEC has been widely used in cellular networks, focusing on improving the energy efficiency or reducing the latency of various cellular-based MEC systems \cite{S.Sardellitti.15Joint,J_C.You17Energy,T.Q.Dinh17Offloading,J_X.Hu18Edge,J_H.Sun19Joint,J_X.Hu18Wireless}. %,J_F.Wang18Joint,J_L.Pu16D2DFogging


In order to further improve the system performance of edge computing, the technology of the unmanned aerial vehicle (UAV) has been leveraged in  MEC architectures, since the UAV communication has many attractive advantages, such as easy deployment, flexible movement, and line-of-sight (LoS) connections, etc.,  \cite{J_zeng16Wireless,J_M.M.Azari18Ultra,J_Xu18Uav}. %J_Zeng17Energy,
The performance improvement of the UAV-enabled MEC architecture has been shown to be substantial \cite{J_Jeong18Mobile,J_F.Zhou18Computation}. %C_X.Cao2018MobileEC
 A UAV-based MEC system is investigated in \cite{J_Jeong18Mobile}, where a moving UAV helps UEs compute their offloaded tasks and the total mobile energy consumption is minimized by optimizing the task-bit allocation and the UAV's trajectory. Later in \cite{J_F.Zhou18Computation}, a wireless powered UAV-enabled MEC system is studied, where the UAV provides energy as well as MEC services for the UEs. The computation rate maximization problems are addressed by alternating algorithms under both the partial and binary computation offloading modes. %In another study \cite{C_X.Cao2018MobileEC}, the UAV acted as a UE rather than an MEC server, which was served by multiple cellular ground base stations  to compute its offloaded tasks. The UAV's  mission completion time was minimized by optimizing the resource allocation as well as the UAV trajectory through an SCA algorithm.

%Due to the attractive advantages of unmanned aerial vehicle (UAV) for its easy deployment, flexible movement, and line-of-sight (LoS) connections, and so on, UAV-enabled wireless communication networks have been much researched in recent years \cite{J_zeng16Wireless,J_Zeng17Energy,J_M.M.Azari18Ultra,J_Xu18Uav}.


The aforementioned MEC works concentrate either on the cellular-based MEC networks or the UAV-enabled MEC architectures by exploiting the computing capability solely at the AP or the UAV. However, it is impossible to take full use of the computing resources at the AP  if the UEs' links to the AP are seriously degraded. Besides, it is risky to rely only on the UAVs for completing UEs' computation-intensive latency-critical tasks since the UAVs are also resource-limited \cite{X.Hu18UAV}. For these reasons, this paper studies a UAV-assisted MEC architecture, where the computing resources at the UAV and the AP are utilized at the same time. In addition, the energy-efficient LoS transmissions of the UAV have been fully exploited since the UAV is not only served as a mobile computing server to help the UEs compute their tasks  but also as a relay to further offload UEs' tasks to the AP for computing.
%To our best knowledge, this is the first work considering the UAV-assisted  MEC architecture by letting the UAV act as an MEC server and a relay simultaneously. %cellular-based

%{\em Notations}---In this paper, boldface lower case letters represent vectors, and $\|\mathbf{a}\|$ denotes the Euclidean norm of vector $\mathbf{a}$. $\left[x\right]^+\triangleq\max\left\{x,0\right\}$. $a\ll b$ means $a$ is much less than $b$. In addition, $\mathcal{O}(\cdot)$ is the big-O notation.

\vspace{-2mm}
\section{System Model and Problem Formulation}\label{sec:system}
\subsection{System Model}\label{sec:Channel Model}
A UAV-assisted  MEC system is considered in this paper, which consists of an AP, a cellular-connected UAV, and $K$ ground UEs, all being equipped with a single antenna. The UAV and UEs are all assumed to have limited computing resources  powered by their embedded battery, while the AP  is endowed with an ultra-high performance processing server.
%is capable of providing high-speed transmission rate  and
%which can be divided arbitrarily to facilitate parallel operation of computing.
A three-dimensional (3D) Euclidean coordinate system is adopted, whose coordinates are measured in meters. The locations of the AP and all the UEs are fixed on the ground with zero altitude, denoted as $\mathbf{\widetilde{v}}_0=(x_0,y_0,0)$ and $\mathbf{\widetilde{v}}_k=(x_k,y_k,0)$ for UE $k\in\mathcal{K}=\{1,\dots,K\}$.  %It is assumed that the locations of UEs are known to the UAV for designing its trajectory \cite{J_Zeng17Energy}. %\cite{J_Zeng17Energy,J_Jeong18Mobile,J_Xu18Uav}
We assume that the UAV flies at a fixed altitude $H>0$ during the task completion time $T$, which is the minimum altitude appropriate to the work terrain. % and can avoid buildings without  frequent descending and ascending.

For ease of exposition, the finite task completion time $T$ is discretized into $N$ equal time slots each with a duration of $\tau=T/N$,
where $\tau$ is sufficiently small such that the UAV's location can be assumed to be unchanged during each slot. % with \textcolor[rgb]{1.00,0.00,0.00}{$V_\mathrm{max}\tau\ll H$}, where $V_\mathrm{max}$ is the predetermined maximum speed of the UAV.
The initial and final horizontal locations of the UAV are preset as $\mathbf{u}_\mathrm{I}=(x_\mathrm{I}, y_\mathrm{I})$ and $\mathbf{u}_\mathrm{F}=(x_\mathrm{F}, y_\mathrm{F})$. Let $\mathcal{N}=\{1,\dots,N\}$ denote the set of the $N$ time slots. At the $n$-th time slot, the UAV's horizontal location is denoted as $\mathbf{u}[n]\equiv\mathbf{u}(n\tau)=(x[n],y[n])$ with $\mathbf{u}[0]=\mathbf{u}_\mathrm{I}$ and $\mathbf{u}[N]=\mathbf{u}_\mathrm{F}$. It is assumed that the UAV flies with a constant speed in each time slot, denoted as $\mathbf{v}[n]$, which should satisfy $\|\mathbf{v}[n]\|=\frac{\|\mathbf{u}[n]-\mathbf{u}[n-1]\|}{\tau}\leq V_\mathrm{max}$,
with a predetermined maximum speed $V_\mathrm{max}\geq\|\mathbf{u}_\mathrm{F}-\mathbf{u}_\mathrm{I}\|/T$.

Similar to \cite{J_Zeng17Energy}, %\cite{J_zeng16Wireless,J_Jeong18Mobile,J_Xu18Uav}
the wireless channels between the UAV and the AP as well as the UEs are assumed to be dominated by LoS links. Thus, the channel power gain between the UAV and the AP and between the UAV and  UE $k$ at the time slot $n$ can be, respectively, given by
\begin{align}
&h_{\mathrm{AP}}[n]=\frac{h_0}{\|\mathbf{u}[n]-\mathbf{v}_0\|^2+H^2}, \ n\in \mathcal{N}, \label{eq:channel_AP} \\
&h_k[n]=\frac{h_0}{\|\mathbf{u}[n]-\mathbf{v}_k\|^2+H^2}, \ k\in\mathcal{K}, n\in \mathcal{N}, \label{eq:channel_k}
\end{align}
where $h_0$ is the channel power gain at a reference distance of one meter; $\mathbf{v}_0=(x_0,y_0)$ and $\mathbf{v}_k=(x_k,y_k)$ denote the horizontal locations of the AP and UE $k\in\mathcal{K}$, respectively.
%$d_{\mathrm{AP}}$ and  $d_{k}$ are respectively the horizontal plane distances between the UAV and the AP  as well as the UE $k$ at the $n$-th time slot with $\mathbf{v}_0=(x_0,y_0)$ and $\mathbf{v}_k=(x_k,y_k)$ denoting the horizontal locations of the AP and UE $k$, $k\in\mathcal{K}$.
%It is assumed that the channel reciprocity establishes in our considered scenario, and thus the offloading and downloading channels between the UEs and the UAV are identical.
In this paper, the direct links between UEs and the AP are assumed to be negligible due to e.g., severe blockage, which means that the UEs cannot directly offload their task-input bits to the AP unless with the assistance of the UAV.

\vspace{-3mm}
\subsection{Computation Task Model and Energy Consumption}\label{sec:Task Model}
It is assumed that each UE has a  computation-intensive task, which is denoted as a positive tuple $[I_k, C_k, T_k]$. Here, $I_k$ denotes the size (in bits) of the computation task-input data which is bit-wise independent and can be arbitrarily divided, $C_k$ is the amount of required computing resource for computing 1-bit of input data,  and $T_k$ is the maximum tolerable latency with $T_k\leq T, k\in \mathcal{K}$.\footnote{ In our considered scenario, we assume that the output data sizes of the computation tasks are quite small that can be ignored especially compared with the input data sizes of the computation-intensive tasks.}
In this paper, we only consider the case that $T_k=T$ for all $k\in\mathcal{K}$.
The UAV acts as an assistant to  help the UEs complete their computation tasks by providing both  MEC and relaying services. For the MEC service, the UAV shares its computing resources with the UEs to help compute their tasks; while for the relaying service, the UAV forwards part of the UEs' offloaded tasks to the AP for computing.
Hence, the UEs can accomplish their computation tasks in a partial offloading fashion \cite{S_Y.Mao17ASurve} with the following three ways. %: local computing, task offloaded to the UAV or the AP for computing.
\subsubsection{Local Computing at UEs}
Each UE can perform local computing and computation offloading simultaneously. % since local computing at the UEs does not need radio resources such as bandwidth.
To efficiently use the energy for local computing, the UEs leverage a dynamic voltage and frequency scaling (DVFS) technique to adaptively  adjust their CPU frequency in each slot \cite{J_Zhang13Energy}. %\cite{J_Zhang13Energy}
The CPU frequency of  UE $k$ during time slot $n$ is denoted as $f_k[n]$ (cycles/second) for computing $l{_k^{\mathrm{loc}}}[n]=\tau f_k[n]/C_k$  task-input bits. Thus, the energy consumption of UE $k$ during time slot $n$ for local computing can be expressed as
%$\tau f_k[n]/C_k$ and $\tau\kappa_k f_k^3[n]$, where $\kappa_k$ is the effective capacitance coefficient of UE $k$ that depends on the of its processor's chip architecture. %, $k\in\mathcal{K}$, $n\in\mathcal{N}$.
\begin{align}
%&l{_k^{\mathrm{loc}}}[n]=\tau f_k[n]/C_k, \ k\in\mathcal{K}, n\in \mathcal{N}, \label{eq:Local_size_k}\\
&E{_k^{\mathrm{loc}}}[n]=\tau\kappa_k f_k^3[n]\equiv\frac{\kappa_kC_k^3}{\tau^2} (l{_k^{\mathrm{loc}}}[n])^3, \ k\in\mathcal{K}, n\in \mathcal{N}, \label{eq:Local_energy_k}
\end{align}
where $\kappa_k$ is the effective capacitance coefficient of UE $k$ that depends on  its processor's chip architecture. %, $k\in\mathcal{K}$, $n\in\mathcal{N}$.


\subsubsection{Task Offloaded to the UAV for Computing}
The UEs' remaining task-input data should be computed remotely, first by offloading to the UAV, and then one part of the data being computed at the UAV while the other part further offloaded to the AP for computing. In order to avoid interference among the UEs during the offloading process, we adopt the time-division multiple access (TDMA) protocol.
%we adopt a TDMA protocol as shown in \textcolor[rgb]{1.00,0.00,0.00}{Fig. 2}.
Each slot is further divided into $K$ equal durations $\delta=T/(NK)$, and UE $k$ offloads its task-input data in the $k$-th duration. Let $l_k^{\mathrm{off}}[n]$ denote the offloaded bits of UE $k$ in its allocated duration at time slot $n$, and thus  the corresponding energy consumption of UE $k$ at slot $n$ for computation offloading can be calculated  as
\begin{align}\label{eq:E_off_k}
E{_k^{\mathrm{off}}}[n]&=\delta p_k^{\mathrm{off}}[n]
\equiv\frac{\delta N_0}{h_k[n]}\bigg(2^{\frac{l_k^{\mathrm{off}}[n]}{\delta B{_k^\mathrm{off}}[n]}}-1\bigg), %k\in\mathcal{K}, n\in\mathcal{N}
\end{align}
where $p_k^{\mathrm{off}}[n]$ and $B{_k^\mathrm{off}}[n]$ are
the related transmit power and bandwidth,
%allocated to UE $k$  for offloading $l_k^{\mathrm{off}}[n]$ computation bits to the UAV at time slot $n$,
and $N_0$ denotes the noise power at the UAV.\footnote{Without loss of generality, we assume that the noise power at any node in the system is considered the same as $N_0$.} %(including the UAV, the AP and all the UEs)

Assume that the UAV also adopts the DVFS technique to improve its energy efficiency for computing, and its adjustable CPU frequency in the $k$-th duration of slot $n$  is denoted as $f_{\mathrm{U},k}[n]$ for computing UE $k$'s  $l_{\mathrm{U},k}[n]=\delta f_{\mathrm{U},k}[n]/C_k$ offloaded task-input bits. Hence, the energy consumption of the UAV for computing UE $k$'s task at slot $n$ can be  given as %helping UE $k$ during
\begin{align}
%l_{\mathrm{U},k}[n]&=\delta f_{\mathrm{U},k}[n]/C_k, \ k\in\mathcal{K}, n\in \mathcal{N}, \label{eq:UAV_size_k}\\
E_{\mathrm{U},k}[n]&=\delta\kappa_{\mathrm{U}} f_{\mathrm{U},k}^3[n]\equiv\frac{\kappa_{\mathrm{U}}C_k^3 }{\delta^2} l_{\mathrm{U},k}^3[n], \ k\in\mathcal{K}, n\in \mathcal{N}, \label{eq:UAV_energy_k}
\end{align}
where $\kappa_\mathrm{U}$ is the effective capacitance coefficient of the UAV.

\subsubsection{Task Offloaded to the AP for Computing}
Part of the UEs' offloaded task-input data at the UAV will be offloaded to the AP's processing server for computing. To better distinguish the offloading signals from different UEs, the TDMA protocol with $K$ equal time division ($\delta=T/(NK)$) is also adopted in this case. Let $l{_{\mathrm{U},k}^{\mathrm{off}}}[n]$ denote the number of UE $k$'s task-input bits being offloaded from the UAV to the AP at time slot $n$. Thus, the corresponding energy consumption of the  UAV for offloading UE $k$'s task at slot $n$ can be calculated as
\begin{align}\label{eq:E_UAV_off_k}
E{_{\mathrm{U},k}^{\mathrm{off}}}[n]&=\delta p{_{\mathrm{U},k}^{\mathrm{off}}}[n]
\equiv\frac{\delta N_0}{h_{\mathrm{AP}}[n]}\bigg(2^{\frac{l{_{\mathrm{U},k}^{\mathrm{off}}}[n]}{\delta B{_{\mathrm{U},k}^\mathrm{off}}[n]}}-1\bigg),
\end{align}
where $p{_{\mathrm{U},k}^{\mathrm{off}}}[n]$ and $B{_{\mathrm{U},k}^\mathrm{off}}[n]$ are the corresponding transmit power  and  the allocated bandwidth of the UAV.
%for offloading UE $k$'s tasks to the AP at time slot $n$.
As the AP is integrated with an ultra-high-performance processing server, the computing time is negligible. It is assumed that the UAV is equipped with a data buffer with sufficiently large size, and it is capable of storing each UE's offloaded data. % and the  corresponding computation results separately.

In fact, the energy consumption for UAV's propulsion is also considerable which is greatly affected by the UAV's trajectory, and hence should be taken into account. %Similar to the works in \cite{J_Zeng17Energy,J_Y.Zeng19Accessing}, the propulsion energy consumption model at the UAV due to flying
With the assumption that the time slot duration $\tau$ is  sufficiently small, the UAV's flying during each slot can be regarded as  straight-and-level flight with constant speed $\mathbf{v}[n]$.
Taking a fixed-wing UAV as an example \cite{J_Zeng17Energy,J_Y.Zeng19Accessing}, its propulsion energy consumption
at time slot $n$ can be expressed as%\textcolor[rgb]{1.00,0.00,0.00}{\footnote{\textcolor[rgb]{1.00,0.00,0.00}{The constant-altitude flying entails a constant hovering energy consumption for the UAV, and thus we just consider its varying energy consumption for flying in the formulated WSEC minimization problem. \eqref {eq:E_fly} shows a general flying energy consumption model for the UAV regardless the type of the UAV.}}}
\begin{align}\label{eq:E_fly}
E{_\mathrm{U}^{\mathrm{fly}}}[n]=\tau\bigg(\theta_1\|\mathbf{v}[n]\|^3+\frac{\theta_2}{\|\mathbf{v}[n]\|}\bigg), \ n\in \mathcal{N},
\end{align}
where $\theta_1$ and $\theta_2$ are two parameters related to the UAV's weight, wing area, wing span efficiency, and air density, etc.
%\textcolor[rgb]{0.00,0.00,1.00}{where $\theta=M\tau/2$  is the propulsion energy efficiency.} %$\theta=M\tau/2$ and $M$ is the mass of the UAV.
Combining with the above analysis, we obtain the total energy consumption of UE $k$ and the UAV in each time slot $n$ as
\begin{align}
\hspace{-2.5mm}E_{k}[n]&=E{_k^{\mathrm{loc}}}[n]+E{_k^{\mathrm{off}}}[n], \ k\in\mathcal{K}, n\in \mathcal{N},\label{eq:E_k_n}\\
\hspace{-2.5mm}E_{\mathrm{U}}[n]&=\sum\limits_{k{\rm{ = }}1}^K \Big(E_{\mathrm{U},k}[n]+E{_{\mathrm{U},k}^{\mathrm{off}}}[n]
\Big)+E{_\mathrm{U}^{\mathrm{fly}}}[n],\ n\in \mathcal{N}.\label{eq:E_UAV_k_N}
\end{align}

\vspace{-6mm}

\subsection{Problem Formulation}\label{sec:problem}
Considering that the battery-based UEs and  UAVs are usually power-limited, one major problem the UAV-assisted MEC system faces is energy. Hence, in this paper, we try to minimize the total energy consumption (TEC) of the UAV and all the UEs during the task completion time $T$. In the previous subsection, we have obtained the energy consumption of the UEs and the UAV for task offloading and computation.
In our considered scenario, the UEs' task allocation parameters in $\mathbf{l}\triangleq\{l_{k}^{\mathrm{loc}}[n],l{_{k}^{\mathrm{off}}}[n], l_{\mathrm{U},k}[n], l{_{\mathrm{U},k}^{\mathrm{off}}}[n]\}_{k\in\mathcal{K},n\in\mathcal{N}}$, the bandwidth allocation parameters in $\mathbf{B}\triangleq\{B{_k^\mathrm{off}}[n],$ $B{_{\mathrm{U},k}^{\mathrm{off}}}[n]\}_{k\in\mathcal{K},n\in\mathcal{N}}$ as well as the UAV's trajectory parameters in $\mathbf{u}\triangleq\{\mathbf{u}[n]\}_{n\in\mathcal{N}}$ will be optimized to minimize the TEC.
To this end, the TEC minimization problem is formulated as problem (P1) below%\footnote{Note that the constraints in
\begin{subeqnarray}\label{eq:WSECM1} %\ k\in\mathcal{K},
% ({\rm P1}):
&&\hspace{-7mm} \underset{\mathbf{l},\mathbf{B},\mathbf{u}}{\min}\
 \sum\limits_{n{\rm{ = }}1}^N\bigg(E_{\mathrm{U}}[n]+\sum\limits_{k{\rm{ = }}1}^KE_k[n]\bigg)\slabel{eq:WSECM1_0}\\
&&\hspace{-7mm}\mathrm{s.t.}\
 \sum\limits_{i{\rm{ = }}2}^n \left(l_{\mathrm{U},k}[i]+l{_{\mathrm{U},k}^{\mathrm{off}}}[i]\right)\leq \sum\limits_{i{\rm{ = }}1}^{n-1} l{_{k}^{\mathrm{off}}}[i],n\in\mathcal{N}_2, k\in\mathcal{K}, \quad~~ \slabel{eq:WSECM1_1}\\
&&\hspace{-7mm} ~~~~\sum\limits_{n{\rm{ = }}2}^{N} \left(l_{\mathrm{U},k}[n]+l{_{\mathrm{U},k}^{\mathrm{off}}}[n]\right)= \sum\limits_{n{\rm{ = }}1}^{N-1} l{_{k}^{\mathrm{off}}}[n], \ k\in\mathcal{K}, \slabel{eq:WSECM1_4}\\
&&\hspace{-7mm} ~~~~\sum\limits_{n{\rm{ = }}1}^{N}l{_{k}^{\mathrm{loc}}}[n]+\sum\limits_{n{\rm{ = }}1}^{N-1} l{_{k}^{\mathrm{off}}}[n]=I_k,\ k\in\mathcal{K}, \slabel{eq:WSECM1_3}\\
&&\hspace{-7mm}~~~~ B{_k^\mathrm{off}}[n]+B{_{\mathrm{U},k}^{\mathrm{off}}}[n]=B, \ n\in\mathcal{N}, \ k\in\mathcal{K}, \slabel{eq:WSECM1_12}\\
&&\hspace{-7mm}~~~~ l_k^{\mathrm{loc}}[n]\geq0, \ n\in\mathcal{N}, \ k\in\mathcal{K}, \slabel{eq:WSECM1_6}\\
&&\hspace{-7mm} ~~~~l_k^{\mathrm{off}}[N]=0, \ l_k^{\mathrm{off}}[n]\geq0, \ n\in\mathcal{N}_1,\  k\in\mathcal{K}, \slabel{eq:WSECM1_11}\\
&&\hspace{-7mm} ~~~~l_{\mathrm{U},k}[1]=0,\ l_{\mathrm{U},k}[n]\geq0, \ n\in\mathcal{N}_2,\  k\in\mathcal{K}, \slabel{eq:WSECM1_14}\\ %k\in\mathcal{K},
&&\hspace{-7mm}~~~~ l{_{\mathrm{U},k}^{\mathrm{off}}}[1]=0, \ l{_{\mathrm{U},k}^{\mathrm{off}}}[n]\geq0, \ n\in\mathcal{N}_2,\  k\in\mathcal{K}, \slabel{eq:WSECM1_7}\\ %k\in\mathcal{K},
&&\hspace{-7mm}~~~~ B{_k^\mathrm{off}}[N]=0, B{_k^\mathrm{off}}[n]\geq0,\ n\in\mathcal{N}_1, k\in\mathcal{K},\slabel{eq:WSECM1_13}\\
&&\hspace{-7mm} ~~~~B{_{\mathrm{U},k}^{\mathrm{off}}}[1]=0, B{_{\mathrm{U},k}^{\mathrm{off}}}[n]\geq0,
\ n\in\mathcal{N}_2, \  k\in\mathcal{K},\slabel{eq:WSECM1_15}\\ %k\in\mathcal{K},
&&\hspace{-7mm} ~~~~\mathbf{u}[0]=\mathbf{u}_\mathrm{I}, \ \mathbf{u}[N]=\mathbf{u}_\mathrm{F}, \slabel{eq:WSECM1_9}\\
&&\hspace{-7mm}~~~~ \|\mathbf{u}[n]-\mathbf{u}[n-1]\|\leq V_\mathrm{max}\tau, \ n\in\mathcal{N},  \quad \slabel{eq:WSECM1_10}
\end{subeqnarray}
where  $\mathcal{N}_1=\{1,\dots,N-1\}$.
%$\mathbf{l}\triangleq\{l_{k}^{\mathrm{loc}}[n],l{_{k}^{\mathrm{off}}}[n], l_{\mathrm{U},k}[n], l{_{\mathrm{U},k}^{\mathrm{off}}}[n]\}_{k\in\mathcal{K},n\in\mathcal{N}}$ and $\mathbf{B}\triangleq\{B{_k^\mathrm{off}}[n], B{_{\mathrm{U},k}^{\mathrm{off}}}[n]\}_{k\in\mathcal{K},n\in\mathcal{N}}$ denote the sets of the task allocation and the bandwidth allocation variables for UE $k$ in time slot $n$, $\mathbf{u}\triangleq\{\mathbf{u}[n]\}_{n\in\mathcal{N}}$ denotes the set of the UAV's horizontal locations for all the slots, i.e., the trajectory of the UAV, and $\mathcal{N}_1=\{1,\dots,N-1\}$.
%In (P1), \eqref{eq:WSECM1_0} is the objective function for minimizing the WSEC where $w_\mathrm{U}$ and $\{w_k\}_{k\in\mathcal{K}}$ represent the weights of the UAV and UEs, respectively. Also,
In (P1), \eqref{eq:WSECM1_1} is an information-causality constraints, indicating that the UAV can only compute or forward the task-input data that has already been received from the UEs with one slot processing delay. Hence, the UEs should not offload at the last slot, while the UAV should not compute or forward the received UEs' data  at the first slot.
\eqref{eq:WSECM1_4} and \eqref{eq:WSECM1_3} are the UEs' computation task constraints to make sure that all the UEs' computation task-input data have been computed. The bandwidth constraints are in \eqref{eq:WSECM1_12}.
%, while \eqref{eq:WSECM1_6}--\eqref{eq:WSECM1_15} ensure the non-negativeness of the optimization variables.
\eqref{eq:WSECM1_9} and \eqref{eq:WSECM1_10}  specify the UAV's initial and final horizontal locations, and its maximum speed constraints.

\vspace{-1mm}
\section{Algorithm Design}\label{algorithm_design1}
The problem (P1) is a complicated non-convex optimization problem because of the non-convex objective function where non-linear couplings exist among the variables $l{_{k}^{\mathrm{off}}}[n]$ and $B{_k^\mathrm{off}}[n]$, $l{_{\mathrm{U},k}^{\mathrm{off}}}[n]$ and $B{_{\mathrm{U},k}^{\mathrm{off}}}[n]$ which are further coupled with the trajectory of the UAV, i.e., $\mathbf{u}[n]$. To address these issues, we propose a three-step alternating optimization algorithm, where the task allocation $\mathbf{l}$, the bandwidth allocation $\mathbf{B}$, and the UAV's trajectory $\mathbf{u}$ are obtained alternatively.
%In the first step, the computation resource scheduling variables in $\mathbf{z}$ are optimized by solving the problem with  given UAV trajectory $\mathbf{u}$ and bandwidth allocation $\mathbf{B}$; and then in the second step, the bandwidth allocation variables in $\mathbf{B}$ will be optimized with the same given UAV trajectory $\mathbf{u}$ and the optimized  $\mathbf{z}$ obtained in the first step; and finally in the third step, we focus on designing the UAV trajectory $\mathbf{u}$ with the optimized variables $\mathbf{z}$ and $\mathbf{B}$.
The details for the three steps of the algorithm are presented as follows.

\vspace{-3mm}
\subsection{Task Allocation}\label{computation_resource_scheduling}
%\subsection{Computation Resource Scheduling with Fixed UAV Trajectory and Bandwidth Allocation}\label{computation_resource_scheduling}
A sub-problem of (P1) is the computation task allocation problem (P1.1), where the UAV's trajectory $\mathbf{u}$ and bandwidth allocation $\mathbf{B}$ are given as fixed. In this case, the time-dependent channels  $\{h_{\mathrm{AP}}[n]\}_{n\in\mathcal{N}}$ and $\{h_k[n]\}_{k\in\mathcal{K},n\in\mathcal{N}}$ defined in \eqref{eq:channel_AP} and \eqref{eq:channel_k} are  known with the fixed $\mathbf{u}$. Besides, the non-linear couplings among the offloading task-input \mbox{bits} ($l{_{k}^{\mathrm{off}}}[n], l{_{\mathrm{U},k}^{\mathrm{off}}}[n]$) with their corresponding allocated bandwidths ($B{_k^\mathrm{off}}[n], B{_{\mathrm{U},k}^{\mathrm{off}}}[n]$) no longer exist. The computation task allocation problem (P1.1) is convex with a convex objective function and constraints, which is expressed as
\begin{subeqnarray}\label{eq:RS1}
 ({\rm P1.1}):
 \underset{\mathbf{l}}{\min} &&\hspace{-4mm}
  \sum\limits_{n{\rm{ = }}1}^N\left(E{_{\mathrm{U}}^{(1)}}[n]+\sum\limits_{k{\rm{ = }}1}^KE_k[n]\right) \quad\quad\slabel{eq:RS1_0}\\
\mathrm{s.t.}
&&\hspace{-4mm} \eqref{eq:WSECM1_1}-\eqref{eq:WSECM1_3}, \ \eqref{eq:WSECM1_6}-\eqref{eq:WSECM1_7}, \slabel{eq:RS1_1} %\\
\end{subeqnarray}
where $E{_{\mathrm{U}}^{(1)}}[n]=\sum\limits_{k{\rm{ = }}1}^K \Big(E_{\mathrm{U},k}[n]+E{_{\mathrm{U},k}^{\mathrm{off}}}[n]\Big)$.
In order to gain more insights of the solution, we leverage the Lagrange method  \cite{B_Boyd04Convex} to solve problem (P1.1), and the optimal solution of problem (P1.1) is given in the following theorem.
%\begin{figure*}[t!]

\begin{theorem}\label{theorem_1}
The optimal solution of problem (P1.1) related to UE $k\in\mathcal{K}$ is given in \eqref{eq:f_k}--\eqref{eq:l_off_Uk} below

\vspace{-2mm}
{\small{
\begin{align}
&\hspace{-1mm}l{_{k}^{\mathrm{loc}*}}[n]=\frac{\tau}{C_k}\sqrt{\frac{[\beta{_k^*}]^+}{3C_k\kappa_k}}, \  n\in \mathcal{N}, \label{eq:f_k}\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
&\hspace{-1mm}l{_{k}^{\mathrm{off}*}}[n]=\left\{
\begin{aligned}
&\delta B{_k^{\mathrm{off}}}[n]\bigg[\log_2\rho_k^{\mathrm{off}}[n]\Big[\widehat{\lambda}_{k,n}^*+\beta{_k^*}-\eta{_k^*}\Big]^+\bigg]^+, n\in \mathcal{N}_1, \\
&0, \quad n=N, \\
\end{aligned}\right. \label{eq:l_k}\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
&\hspace{-1mm}l{_{\mathrm{U},k}^*}[n]=\left\{
\begin{aligned}
&\frac{\delta}{C_k}\sqrt{\frac{ \left[\eta{_k^*}-\widetilde{\lambda}_{k,n}^*\right]^+}{3C_k\kappa_\mathrm{U}}}, n\in \mathcal{N}_2, \\
&0, \quad n=1, \\
\end{aligned}\right. \label{eq:f_Uk}\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
&\hspace{-1mm}l{_{\mathrm{U},k}^{\mathrm{off}*}}[n]=\left\{
\begin{aligned}
&\delta B{_{\mathrm{U},k}^{\mathrm{off}}}[n]\left[\log_2\rho_{\mathrm{U},k}^{\mathrm{off}}\left[\eta{_k^*}-\widetilde{\lambda}_{k,n}^* \right]^+\right]^+, n\in \mathcal{N}_2,  \\
&0, \quad n=1, \\
\end{aligned}\right. \label{eq:l_off_Uk}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\end{align}
}}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\hspace{-1.3mm}where $\rho_k^{\mathrm{off}}[n]=\frac{B{_k^{\mathrm{off}}}[n]h_k[n]}{N_0\ln2}$, $\rho_{\mathrm{U},k}^{\mathrm{off}}=\frac{B{_{\mathrm{U},k}
^{\mathrm{off}}}[n]h_{\mathrm{AP}}[n]}{N_0\ln2}$,
$\widehat{\lambda}_{k,n}^*=\sum_{i{\rm{ = }}n+1}^{N-1}\lambda{_{k,i}^*}$ and $\widetilde{\lambda}_{k,n}^*=\sum_{i{\rm{ = }}n}^{N-1}\lambda{_{k,i}^*}$.
Here, $\lambda{_{k,n}^*}\geq0$  for $k\in\mathcal{K}, n\in\mathcal{N}$ are the optimal Lagrange multipliers (dual variables) associated with the inequality constraints \eqref{eq:WSECM1_1}  in problem (P1.1) (or P1), while  $\eta_k^*$ and $\beta^*_k$  are respectively the optimal Lagrange multipliers  associated with the equality constraints \eqref{eq:WSECM1_4} and \eqref{eq:WSECM1_3} for $k\in\mathcal{K}$.
\end{theorem}

%\begin{proof}
%See Appendix A.
%\end{proof}

It is necessary to obtain the optimal values of the Lagrange multipliers, i.e., $\boldsymbol{\lambda}^*=\{\lambda{_{k,n}^*}\}_{k\in\mathcal{K}, n\in\mathcal{N}}$, $\boldsymbol{\eta}^*=\{\eta{_{k}^*}\}_{k\in\mathcal{K}}$ and  $\boldsymbol{\beta}^*=\{\beta{_{k}^*}\}_{k\in\mathcal{K}}$  since they play important roles in determining the optimal task allocation $\mathbf{l}^*$ according to Theorem \ref{theorem_1}. In this paper, we adopt a subgradient-based algorithm to obtain the optimal dual variables in  $\boldsymbol{\lambda}^*$  related to the inequality constraints \eqref{eq:WSECM1_1} \cite{B_D.Bertsekas89Parallel}.
Besides,  the optimal dual variables in $\boldsymbol{\eta}^*$ and  $\boldsymbol{\beta}^*$  related to the equality constraints \eqref{eq:WSECM1_4} and \eqref{eq:WSECM1_3} can be obtained by bi-section search method.
Note that  the  convergence can be guaranteed according to \cite{B_Boyd04Convex}.

\vspace{-3mm}
\subsection{Bandwidth Allocation}\label{bandwidth_allocation}
%\subsection{Bandwidth Allocation with Fixed UAV Trajectory and Computation Resource Scheduling}\label{bandwidth_allocation}
Here, another sub-problem of (P1), denoted as the bandwidth allocation problem (P1.2) is considered to optimize $\mathbf{B}$ with the same given UAV's trajectory $\mathbf{u}$  and the optimized computation task allocation parameters in $\mathbf{l}$.
The bandwidth allocation problem (P1.2) is expressed as
\begin{subeqnarray}\label{eq:BA}
 ({\rm P1.2}):
 \underset{\mathbf{B}}{\min} &&\hspace{-4mm}
  \sum\limits_{n{\rm{ = }}1}^N\sum\limits_{k{\rm{ = }}1}^K
  \left(E{_k^{\mathrm{off}}}[n]+E{_{\mathrm{U},k}^{\mathrm{off}}}[n]\right) \quad\quad \slabel{eq:BA_0}\\
\mathrm{s.t.}
&&\hspace{-4mm} \eqref{eq:WSECM1_12},~ \eqref{eq:WSECM1_13}, ~\eqref{eq:WSECM1_15}. \slabel{eq:BA_1} %\\
\end{subeqnarray}
It can be easily proved that problem (P1.2) is convex with  convex objective function and constraints. To gain more insights on the structure of the optimal solution, we again leverage the Lagrange method \cite{B_Boyd04Convex} to solve this problem, and the optimal solution to problem (P1.2) is given in the following theorem.

\begin{theorem}\label{theorem_2}
The optimal solution of problem (P1.2)  related to UE $k\in\mathcal{K}$ is given by
{\small{
\begin{align}
\hspace{-2mm}B{_k^{\mathrm{off}*}}[n]&=\left\{
\begin{aligned}
%&B, \quad n=1, \\
&\frac{\frac{\ln2}{2}l_k[n]}{\delta W_0\Big[\frac{\ln2}{2}
\big( \frac{\nu{_{k,n}^*}h_k[n]l_k^{\mathrm{off}}[n]}{\delta^2N_0\ln2}\big)^{\frac{1}{2}}\Big]},~n\in\mathcal{N}_1, \\
&0, \quad n=N-1~\mathrm{or}~N, \\
\end{aligned}\right. \label{B_k_off}\\
\hspace{-2mm}B{_{\mathrm{U},k}^{\mathrm{off}*}}[n]&=\left\{
\begin{aligned}
&\frac{\frac{\ln2}{2}l{_{\mathrm{U},k}^{\mathrm{off}}}[n]}{\delta W_0
\Big[\frac{\ln2}{2}\big(\frac{\nu{_{k,n}^*}h_{\mathrm{AP}}[n]l{_{\mathrm{U},k}^{\mathrm{off}}}[n]}{\delta^2N_0\ln2}\big)
^{\frac{1}{2}}\Big]},~n\in\mathcal{N}_2, \\
&0, \quad n=1~\mathrm{or}~N, \\
\end{aligned}\right. \label{B_Uk_off}
\end{align}
}}
\hspace{-1mm}where $\{\nu{_{k,n}^*}\}_{k\in\mathcal{K}, n\in\mathcal{N}}$  are the optimal Lagrange multipliers associated with the equality constraints in \eqref{eq:WSECM1_12} of problem (P1.2), and  $W_0(x)$ is the principal branch of the Lambert $W$ function defined as the solution of $W_0(x)e^{W_0(x)}=x$ \cite{J_Corless1996LambertW}.
\end{theorem}

\begin{proof}
See Appendix A.
\end{proof}

The optimal Lagrange multipliers $\{\nu{_{k,n}^*}\}$  in Theorem \ref{theorem_2}  should make the  the optimal bandwidth allocation parameters satisfy the equality constraints in \eqref{eq:WSECM1_12}. In fact, $\{\nu{_{k,n}^*}\}$ can be obtained effectively with the bi-section search when the bandwidth is not exclusively occupied, i.e., both $l_k^{\mathrm{off}}[n]$ and  $l{_{\mathrm{U},k}^{\mathrm{off}}}[n]$ are positive,  since $\{B{_k^{\mathrm{off}*}}[n]\}_{n\in\mathcal{N}_1}$ and $\{B{_{\mathrm{U},k}^{\mathrm{off}*}}[n]\}_{n\in\mathcal{N}_2}$  are all monotonically decreasing functions with respect to (w.r.t.) $\{\nu{_{k,n}^*}\}$ according to the property of the $W_0$ function.
%Besides, we can obtain tight search ranges using the results in Lemma~\ref{lemma4}.

\vspace{-2mm}
\subsection{UAV Trajectory Design}\label{UAV_trajectory_design}
%\subsection{UAV Trajectory Design With Fixed Computation Resource Scheduling and Bandwidth Allocation}\label{UAV_trajectory_design}
Here, the sub-problem for designing the UAV's trajectory $\mathbf{u}$ is considered, which we refer to it as the UAV trajectory design problem (P1.3), by assuming that the computation resource scheduling  $\mathbf{z}$  and bandwidth allocation $\mathbf{B}$ are given as fixed with the previously optimized values. Hence, the  UAV trajectory design problem (P1.3) can be rewritten as
\begin{subeqnarray}\label{eq:UAV}
 ({\rm P1.3}):
 \underset{\mathbf{u}}{\min} &&\hspace{-4mm}
  \sum\limits_{n{\rm{ = }}1}^N\left(E{_{\mathrm{U}}^{(3)}}[n]+\sum\limits_{k{\rm{ = }}1}^KE{_k^{\mathrm{off}}}[n]\right)\quad\quad \slabel{eq:UAV_0}\\
\mathrm{s.t.}
&&\hspace{-4mm} \eqref{eq:WSECM1_9},~\eqref{eq:WSECM1_10}, \slabel{eq:UAV_1} %\\
\end{subeqnarray}
where $E{_{\mathrm{U}}^{(3)}}[n]=E{_\mathrm{U}^{\mathrm{fly}}}[n]+
\sum\limits_{k{\rm{ = }}1}^K E{_{\mathrm{U},k}^{\mathrm{off}}}[n]$.
Note that the $E{_\mathrm{U}^{\mathrm{fly}}}[n]$ defined in \eqref{eq:E_fly}  is not a convex function of $\mathbf{u}$. In order to address this issue, we define an upper bound of $E{_\mathrm{U}^{\mathrm{fly}}}[n]$ as follows
\begin{align}\label{eq:E_fly_up}
\widetilde{E}{_\mathrm{U}^{\mathrm{fly}}}[n]=\tau\bigg(\theta_1\|\mathbf{v}[n]\|^3+\frac{\theta_2}{\widetilde{v}[n]}\bigg), \ n\in \mathcal{N},
\end{align}
by introducing a variable $\widetilde{v}[n]$ and a constraint $\|\mathbf{v}[n]\|\geq \widetilde{v}[n]$, which is equivalent to $\|\mathbf{u}[n]-\mathbf{u}[n-1]\|^2\geq \widetilde{v}^2[n]\tau^2$.
This constraint  is still non-convex, and we leverage the successive convex approximation (SCA) technique to solve this issue. %successive convex approximation (SCA) technology
The left hand side of the constraint is convex versus $\mathbf{u}$ and can be  approximated as its linear lower bound by using the first-order Taylor expansion at a local point $\mathbf{u}_i$, where $i=1,2,\dots$ denotes the iteration index of the SCA method. Hence, the additional constraint can be approximated as a convex one as follows
\begin{align}\label{eq:speed_n_app}
&\widetilde{v}^2[n]\tau^2-2(\mathbf{u}_i[n]-\mathbf{u}_i[n-1])^T(\mathbf{u}[n]-\mathbf{u}[n-1]) \\ \nonumber
&\leq \|\mathbf{u}_i[n]-\mathbf{u}_i[n-1]\|^2, \ n\in \mathcal{N}.
\end{align}
The approximated problem of (P1.3) with $\{\widetilde{E}{_\mathrm{U}^{\mathrm{fly}}}[n]\}$, $\{\widetilde{v}[n]\}$ and the constraint \eqref{eq:speed_n_app} is convex w.r.t.~$\mathbf{u}$ and $\{\widetilde{v}[n]\}$, and we can resort to the software CVX \cite{M_Grant08CVX} to solve it.
%. However, the UAV's locations in different slots are coupled with each other as in \eqref{eq:WSECM1_10}, and thus it is difficult to obtain a closed-form solution of $\mathbf{u}$. In this case, and we resort to the software CVX \cite{M_Grant08CVX} to solve the approximated problem of (P1.3). %additional constraint \eqref{eq:speed_n_app}


\vspace{-2mm}
{\small{
\setlength{\tabcolsep}{0.3 pt}\begin{table}[thb]
\centering
\caption{Simulation Parameters}\label{table1}
\begin{tabular}{|l|l|l|}
\hline
~\textbf{Parameter }&~{\textbf{Symbol}} &~{\textbf{Value}} \\
\hline
~System bandwidth\quad~   &~$B$ \quad\quad &~20 MHz \quad\quad\quad\quad \\
\hline
~Task completion time  &~$T$ &~10 seconds\\
\hline
~Number of time slots &~$N$ &~50  \\
\hline
~Number of ground UEs &~$K$ &~4  \\
\hline
~The channel power gain& ~$h_0$ &~$-30\mathrm{dB}$ \\%\ $(-50\mathrm{dB})$
\hline
~The noise power  &~$N_0$ &~$-60$dBm \\ %($-174$ dBm/Hz,$10^{-9}W$)  spectral density
\hline
~Altitude of the UAV&~$H$ &~10 m \\
\hline
~UAV's maximum speed&~$V_{\mathrm{max}}$ &~10 m/s \\
\hline
%~UAV's flying energy efficiency&~$\xi$ &~0.2 \\
~UAV's propulsion parameters&~($\theta_1,\theta_2$) &~(0.00614,15.976)\\
\hline
~The switched capacitance &~$\kappa_\mathrm{U}$, $\kappa_k (k\in\mathcal{K})$ &~$10^{-28}$ \\
%\hline
%~The energy weight of the UAV &~$w_\mathrm{U}$ &~$0.2$ \\
%\hline
%~The energy weight of the UEs&~$w_k~(k\in\mathcal{K})$ &~$1$ \\
\hline
~UEs' task-input data size &~$I_k~(k\in\mathcal{K})$ &~400 Mbits \\
\hline
~Required CPU cycles per bit &~$C_k~(k\in\mathcal{K})$ &~1000 cycles/bit \\
\hline
~The tolerant thresholds &~$\epsilon_1$ and $\epsilon$ &~$10^{-4}$  \\
\hline
\end{tabular}
\end{table}
}}


\section{Simulation Results}\label{sec:simulation}

In this section, simulation results are presented to evaluate the effectiveness of the proposed algorithm in comparison with some practical baselines.
%The effects of the key parameters will be analyzed, including the relative location of the AP ($\mathbf{v}_0$),\footnote{In order to properly show the effects of the relative location of the AP to UEs on UAV's trajectory and the performance, we fix the locations of the UEs and vary the location  of AP even though AP is usually fixed in practice.} the computation task sizes of UEs ($I_k$ for $k\in\mathcal{K}$), the task completion time for UEs ($T$), the weight for energy consumption of the UAV ($w_\mathrm{U}$), \textcolor[rgb]{1.00,0.00,0.00}{and the iteration index of the alternating optimization algorithm ($\zeta$)}. %, especially on the total energy consumption and  UEs' cloud selections.
The basic simulation parameters are listed in Table~\ref{table1} unless specified otherwise. Besides, the initial and final positions of the UAV  are set as $\mathbf{u}_\mathrm{I}=(-5,-5)$, $\mathbf{u}_\mathrm{F}=(5,-5)$, and the horizontal positions of the UEs are set as $[\mathbf{v}_1$, $\mathbf{v}_2$, $\mathbf{v}_3$, $\mathbf{v}_4]=[(5,5), (-5,5), (-5,-5), (-5,5)]$.

\subsection{Trajectory of the UAV}\label{sec:Trajectory}
\begin{figure}[tbp]%[t!]
\centering
%\includegraphics[width=3.3 in]
\includegraphics[scale=0.4]{Trajectory_4Combinations.eps}%{Traj6_no.eps}%{Trajectory_5_5_d10.eps}%{Trajectory_Combine_4.eps}
\caption{The trajectories of the UAV in the situations with \mbox{different} location of the AP and task size allocation of the UEs:   $\mathbf{v}_0=(0,0)$ for (a) and (b), $\mathbf{v}_0=(10,5)$ for (c) and (d);  $[I_1,I_2,I_3,I_4]=[6,2,4,2]\times10^2$Mbits for (a) and (c), $[I_1,I_2,I_3,I_4]=[6,4,6,2]\times10^2$Mbits for (b) and (d).}
%\caption{The trajectories of the UAV in the situations with \mbox{different} location of the AP and task size allocation of the UEs:   $\mathbf{v}_0=(0,0)$ for (a), (b) and (c), $\mathbf{v}_0=(10,5)$ for (d), (e) and (f);  $[I_1,I_2,I_3,I_4]=[6,2,4,2]\times10^2$Mbits for (a) and (d), $[I_1,I_2,I_3,I_4]=[6,4,6,2]\times10^2$Mbits for (b) and (e), $[I_1,I_2,I_3,I_4]=[2,2,6,4]\times10^2$Mbits for (c) and (f).}
\label{Trajectory}
\end{figure}


In this subsection, numerical results for the trajectory of the UAV are given to shed light on the effects of  the task sizes of UEs
($[I_1,I_2,I_3,I_4]$) and the relative location of the AP ($\mathbf{v}_0$), as depicted in Fig. \ref{Trajectory}.
%In Fig. \ref{Trajectory}, the UAV's flying trajectories are depicted in different scenarios. It should be noted that the total task size of UEs is same for the cases in (a), (c), (d) and (f), i.e., 1400 Mbits, while the cases for (b) and (e) are with larger total task size, e.g., 1800 Mbits. From these results in Fig. \ref{Trajectory}, we can observe that the trajectory of the UAV is heavily reliant on the relative location of the AP and the distribution of UEs' task sizes.
For the scenario of $\mathbf{v}_0=(0,0)$, the AP is surrounded by the UEs and at the center of the UEs' distributed area. We can observe that the UAV tends to fly close to the UEs with large task sizes and tries to be not too far away from the AP when the total task sizes of UEs are moderate as the results in cases (a). When the total task size becomes larger and the distribution of UEs' task sizes becomes  more average, the UAV tends to fly close to the AP as the result in case (b).
These two cases  indicate that for the scenario where the AP is located at the center of UEs' distributed area, the distribution of the UEs' task sizes plays an important role on the UAV's trajectory, while the effect of the AP's location will become more dominant when the UEs' total task size becomes larger, which coincides with the intuition that more task-input data will be offloaded to the AP in this situation so as to reduce the TEC by making use of the super computing resources at the AP. For the scenario of $\mathbf{v}_0=(10,5)$, the AP is located outside the distributed area of the UEs and its average distance to the UEs is relatively larger than the above scenario. %In this situation, the effects of AP's location on the trajectories are more prominent for the cases with moderate total task size, where the comparison between (a) and (d), (c) and (f) can properly explain this.
In this situation, the effects of AP's location on the trajectories are more prominent, where the comparison between (a) and (c), (b) and (d), can properly explain this.
%On the other hand, for the case with larger total task size as case (e), the UAV's trajectory will be slightly moved closer to the UEs with large task sizes which can be seen from the comparison between (d) and (e).

The reason behind these results in Fig.~\ref{Trajectory} is that there exists a tradeoff between the distribution of UEs' task sizes and the relative location of the AP to the UEs. In other words, getting close to the UEs with large task sizes can reduce UEs' offloading energy consumption, while being closer to the AP will reduce the UAV's offloading energy consumption, and thus the UAV has to find a balance between these two factors meanwhile taking its own flying energy consumption into consideration, so as to minimize the TEC through optimizing its flying trajectory.
\vspace{-3mm}
\subsection{Performance Improvement}\label{sec:Performance}
Here, we focus on the performance improvement of the proposed algorithm in comparison with some baseline schemes in the scenario of $\mathbf{v}_0=(0,0)$. The baseline schemes include the the ``Local Computing" scheme where the UEs rely on their own computing resources without task offloading;
the ``Direct Trajectory" scheme where the UAV flies from its initial location to the final location directly with an average speed;
the ``Offloading Only" scheme where the UEs just rely on task offloading to the UAV and the AP for computing without any local computing; the ``Equal Bandwidth" scheme where the whole bandwidth are equally divided for UEs and the UAV's offloading without bandwidth optimization. % with equal bandwidth allocation, i.e.,
%\textcolor[rgb]{1.00,0.00,0.00}{and the ``Local Computing" scheme, where the UEs rely on their own computing resources to complete their computation tasks without offloading. Note that the former four schemes are all offloading schemes.}
%To better illustrate the effects of AP's relative location on the performance, we present all the results in two scenarios given in Fig.~\ref{Trajectory}, i.e., $\mathbf{v}_0=(0,0)$ and $\mathbf{v}_0=(10,5)$.


\begin{figure}[tbp]%[t!]
\centering
%\includegraphics[width=3.3 in]
\includegraphics[scale=0.4]{Energy_I00.eps}%{Energy_I.eps}%{E_I0.eps}%{UAV_Eg_I.eps}%[scale=0.52]{2Four_UAV_Eg__I400_20_500M.eps}%{Trajectory_Combine_4.eps}
\caption{The TEC of the UAV and UEs versus the uniform task size: $I=I_k$ for $k\in\mathcal{K}$.}
\label{Energy_I}
\end{figure}

Fig.~\ref{Energy_I} shows the TEC results versus the uniform task size $I=I_k$ for $k\in\mathcal{K}$. All the curves  increase with $I$ as expected since more energy will be consumed by completing tasks with more input data. It can be seen that great performance improvement can be achieved by leveraging the proposed solution in comparison with all the baseline  schemes.
%It is clear that the performance of the ``Local Computing" scheme is far worse that the other schemes with computation offloading, verifying the effectiveness of task offloading.
Specifically, the TEC of the ``Proposed Solution" are almost one thousandth of that for the ``Local Computing" scheme, presenting the tremendous benefits the UEs obtained by deploying the UAV as an assistant for computing and  relaying.
In addition, the TEC of the proposed solution are almost third less than those of the other offloading schemes for $I<460$ Mbits.
Moreover, the gaps between the proposed solution and the baseline schemes become larger as $I$ increases, and  the TEC of the proposed solution is  less than  half of the ``Offloading only" scheme when $I=500$ Mbits. All these results verify that the effectiveness of the proposed optimization on task allocation, bandwidth allocation and UAV's trajectory, which make the proposed algorithm superior in dealing with the computation-intensive tasks.


%\textcolor[rgb]{0.00,0.00,1.00}{In addition,  the WSECs of the proposed solution are half less than those of the ``Equal Bandwidth" scheme and they are almost quarter less than those of the ``Direct Trajectory" scheme. The ``Offloading Only" scheme performs well with relatively small task sizes, e.g., $I=400$ Mbits, but its gaps between the ``Proposed Solution" are even larger than those of the ``Direct Trajectory" scheme when task sizes are large, e.g., $I=500$ Mbits. All these results verify that the proposed optimization on bandwidth allocation and UAV's trajectory, as well as making full use of the computing resources at UEs  have great effect on  minimizing the WSEC of the UAV and UEs.}
%Note that the gaps between the proposed solution and the baselines become larger when $I$ increases, which further indicates that the proposed algorithm is more capable of handling the computation-intensive tasks.

\begin{figure}[tbp]%[t!]
\centering
%\includegraphics[width=3.3 in]
\includegraphics[scale=0.4]{Energy_T00.eps}%{Energy_T.eps}%{E_T0.eps}%{UAV_Eg_T.eps}%[scale=0.52]{2Four_UAV_Eg__T8_05_10s.eps}%{Trajectory_Combine_4.eps}
\caption{The TEC  of the UAV and UEs versus the total task completion time: $T$ (s).}
\label{Energy_T}
\end{figure}

In Fig.~\ref{Energy_T}, the TEC  w.r.t.~the total task completion time $T$ is depicted.
We can see that the TECs of all the schemes decrease with $T$, coinciding with the intuition that a tradeoff exists between the energy consumption and time consumption for completing the same tasks, and the energy consumption will decrease when the consumed time increases.  It is notable that the proposed solution performs better than the four baseline schemes, and the performance improvement is even more prominent with strict time restriction (small $T$), which further confirms that the proposed algorithm is good at dealing with the latency-critical computation tasks and can achieve a better energy-delay tradeoff. %Besides, some similar insights can also be obtained as from Fig.~\ref{Energy_I}.

%From the above results, we can observe that the WSEC for the scenario of $\mathbf{v}_0=(10,5)$ is higher than that for the scenario of $\mathbf{v}_0=(0,0)$ for all the schemes. It is easy to understand that more energy will be used for UAV's offloading transmission and flying because of the farther average distance between the AP and UEs. The performance of the proposed scheme is also more stable than that of the baseline schemes  considering the changing of the relative location of the AP to UEs since its relative WSEC increment is the smallest among the schemes.


\section{Conclusion}\label{sec:conclusion}
The UAV-assisted MEC architecture is studied in this paper, where the UAV acts as an MEC server and a relay to help the UEs compute their tasks or further offload their tasks to the AP for computing. The problem of minimizing the TEC of the UAV and the UEs  is solved  by an alternating algorithm, where  the task allocation, the bandwidth allocation, and the UAV's trajectory are iteratively optimized under some practical constraints. The simulation results have confirmed that
%the UAV's trajectory is greatly affected by the relative location of the AP and the distribution of UEs' task sizes. Besides,
significant performance improvement can be achieved by  the proposed algorithm over the baseline schemes especially in handling the computation-intensive and latency-critical tasks.

\section*{Acknowledgment}\label{sec:acknowledgment}
This work is supported  by the U.K. EPSRC under Grants EP/K015893/1.
%This work is supported in part by the U.K. EPSRC under Grants EP/K015893/1 and EP/L026031/1, and in part by the  NSFC  under Grants 61620106011 and U1705263.
%\newpage

\section*{Appendix A: Proof of {Theorem}~\ref{theorem_2}}
\label{App:theo_2}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}

The partial Lagrange function of (P1.2) is defined as
{\small{
\begin{align}\label{eq:L_P1_2}
  \mathcal{L}^{(2)}(\mathbf{B},\boldsymbol{\nu})= &
  \sum\limits_{k{\rm{ = }}1}^K \sum\limits_{n{\rm{ = }}1}^N \bigg(E{_k^{\mathrm{off}}}[n]+
  E{_{\mathrm{U},k}^{\mathrm{off}}}[n]\bigg)+ \nonumber \\
  &\sum\limits_{k{\rm{ = }}1}^K\sum\limits_{n{\rm{ = }}1}^N\nu_{k,n}\big(B-B{_k^{\mathrm{off}}}[n]-B{_{\mathrm{U},k}^{\mathrm{off}}}[n]\big),
\end{align}
}}
\hspace{-1.3mm}where $\boldsymbol{\nu}=\{\nu_{k,n}\}_{k\in\mathcal{K},n\in\mathcal{N}}$.
The Lagrangian dual function of problem (P1.2) can be presented as
{\small{
\begin{align}\label{Dual_fun_P1_2}
d^{(2)}(\boldsymbol{\nu})=~\underset{\mathbf{B}}{\min}~ \mathcal{L}^{(2)}(\mathbf{B},\boldsymbol{\nu}),~
~\mathrm{s.t.} ~~ \eqref{eq:WSECM1_13},~\eqref{eq:WSECM1_15}.
\end{align}
}}
\hspace{-1.2mm}Hence, the optimal solution of $\mathbf{B}$ with optimal dual variables $\boldsymbol{\nu}^*$ can be obtained by solving  \eqref{Dual_fun_P1_2}.  It is easy to note that the expressions of $E{_k^{\mathrm{off}}}[n]$ and $E{_{\mathrm{U},k}^{\mathrm{off}}}[n]$   have similar structures w.r.t.~$B{_k^{\mathrm{off}}}[n]$ and $B{_{\mathrm{U},k}^{\mathrm{off}}}[n]$, and thus the optimal solution of $B{_k^{\mathrm{off}}}[n]$ and $B{_{\mathrm{U},k}^{\mathrm{off}}}[n]$  should have similar structures according to problem \eqref{Dual_fun_P1_2}.
Next, we will take $B{_k^{\mathrm{off}}}[n]$  as an example to obtain its closed-form optimal solution versus $\nu_{k,n}^*$ for $k\in\mathcal{K}, n\in\mathcal{N}$.
Applying the KKT conditions  \cite{B_Boyd04Convex} leads to the following necessary and sufficient condition of  $B{_k^{\mathrm{off}*}}[n]$:
{\small{
\begin{align}\label{L_B_k_off}
\hspace{-2mm}\frac{\partial\mathcal{L}^{(2)}(\mathbf{B},\boldsymbol{\nu})}{\partial B{_k^{\mathrm{off}*}}[n]}=\nu_{k,n}^*-\frac{l_k^{\mathrm{off}}[n]N_0\ln2}{(B{_k^{\mathrm{off}*}}[n])^2h_k[n]}
2^{\frac{l_k^{\mathrm{off}}[n]}{B{_k^{\mathrm{off}*}}[n]\delta}}=0,
\end{align}
}}
\hspace{-1.2mm}where the optimal dual variable $\nu_{k,n}^*$ should make sure that the equality constraint  $B{_k^{\mathrm{off}*}}[n]+B{_{\mathrm{U},k}^{\mathrm{off}*}}[n]=B$ is satisfied.
It is not easy to obtain the closed-form solution of $B{_k^{\mathrm{off}*}}[n]$ through \eqref{L_B_k_off} directly.
By defining $\xi=\frac{l_k^{\mathrm{off}}[n]}{B{_k^{\mathrm{off}*}}[n]\delta}$, the equation in \eqref{L_B_k_off} can be re-expressed as $\xi^22^\xi=\frac{\nu_{k,n}^*h_k[n]l_k^{\mathrm{off}}[n]}{\delta^2N_0\ln2}\triangleq\Gamma$,
then by applying the natural logarithm at the both sides leads to $\ln\xi+\frac{\ln2}{2}\xi=\ln\Gamma^{\frac{1}{2}}$, which can be equivalently transformed into $\frac{\ln2}{2}\xi e^{\frac{\ln2}{2}\xi}=\frac{\ln2}{2}\Gamma^{\frac{1}{2}}$ by applying the exponential operation,
where $e$ is the base of the natural logarithm. According to the definition and property of Lambert function \cite{J_Corless1996LambertW}, we have $\frac{\ln2}{2}\xi=W_0(\frac{\ln2}{2}\Gamma^{\frac{1}{2}})$, and finally we can express $B{_k^{\mathrm{off}*}}[n]$ as
\begin{align}\label{express_B_k_off}
B{_k^{\mathrm{off}*}}[n]=\frac{\frac{\ln2}{2}l{_{k}^{\mathrm{off}}}[n]}{\delta W_0\big[\frac{\ln2}{2}(\frac{\nu_{k,n}^*h_k[n]l_k^{\mathrm{off}}[n]}{\delta^2N_0\ln2})^{\frac{1}{2}}\big]},~n\in\mathcal{N}_1.
\end{align}
Integrating with the cases $B{_k^{\mathrm{off}*}}[N]=0$, the complete solution of $B{_k^{\mathrm{off}^*}}[n]$ in \eqref{B_k_off} can be obtained. The solution of $B{_{\mathrm{U},k}^{\mathrm{off}*}}[n]$  in \eqref{B_Uk_off}  can be obtained in a similar way.


%\omega\theta

\bibliographystyle{IEEEtran}
\bibliography{UAV_MEC_Relay}


\end{document}



% Generated by IEEEtran.bst, version: 1.12 (2007/01/11)
\begin{thebibliography}{10}
\providecommand{\url}[1]{#1}
\csname url@samestyle\endcsname
\providecommand{\newblock}{\relax}
\providecommand{\bibinfo}[2]{#2}
\providecommand{\BIBentrySTDinterwordspacing}{\spaceskip=0pt\relax}
\providecommand{\BIBentryALTinterwordstretchfactor}{4}
\providecommand{\BIBentryALTinterwordspacing}{\spaceskip=\fontdimen2\font plus
\BIBentryALTinterwordstretchfactor\fontdimen3\font minus
  \fontdimen4\font\relax}
\providecommand{\BIBforeignlanguage}[2]{{%
\expandafter\ifx\csname l@#1\endcsname\relax
\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}%
\typeout{** loaded for the language `#1'. Using the pattern for}%
\typeout{** the default language instead.}%
\else
\language=\csname l@#1\endcsname
\fi
#2}}
\providecommand{\BIBdecl}{\relax}
\BIBdecl

\bibitem{S_P.Mach17Mobile}
P.~Mach and Z.~Becvar, ``Mobile edge computing: A survey on architecture and
  computation offloading,'' \emph{{IEEE} Commun. Surveys Tuts.}, vol.~19,
  no.~3, pp. 1628--1656, thirdquarter 2017.

\bibitem{S_Y.Mao17ASurve}
Y.~Mao, C.~You, J.~Zhang, K.~Huang, and K.~B. Letaief, ``A survey on mobile
  edge computing: The communication perspective,'' \emph{{IEEE} Commun. Surveys
  Tuts.}, vol.~19, no.~4, pp. 2322--2358, Fourthquarter 2017.

\bibitem{S.Sardellitti.15Joint}
S.~Sardellitti, G.~Scutari, and S.~Barbarossa, ``Joint optimization of radio
  and computational resources for multicell mobile-edge computing,''
  \emph{{IEEE} Signal Inf. Process. Netwo.}, vol.~1, no.~2, pp. 89--103, Jun.
  2015.

\bibitem{J_C.You17Energy}
C.~You, K.~Huang, H.~Chae, and B.~H. Kim, ``Energy-efficient resource
  allocation for mobile-edge computation offloading,'' \emph{{IEEE} Trans.
  Wireless Commun.}, vol.~16, no.~3, pp. 1397--1411, Mar. 2017.

\bibitem{T.Q.Dinh17Offloading}
T.~Q. Dinh, J.~Tang, Q.~D. La, and T.~Q.~S. Quek, ``Offloading in mobile edge
  computing: Task allocation and computational frequency scaling,''
  \emph{{IEEE} Trans. Commun.}, vol.~65, no.~8, pp. 3571--3584, Aug. 2017.

\bibitem{J_X.Hu18Edge}
X.~Hu, L.~Wang, K.-K. Wong, Y.~Zhang, Z.~Zheng, and M.~Tao, ``Edge and central
  cloud computing: A perfect pairing for high energy efficiency and
  low-latency,'' 2018.

\bibitem{J_H.Sun19Joint}
H.~{Sun}, F.~{Zhou}, and R.~Q. {Hu}, ``Joint offloading and computation energy
  efficiency maximization in a mobile edge computing system,'' \emph{{IEEE}
  Trans. Veh. Technol.}, vol.~68, no.~3, pp. 3052--3056, Mar. 2019.

\bibitem{J_X.Hu18Wireless}
X.~Hu, K.~Wong, and K.~Yang, ``Wireless powered cooperation-assisted mobile
  edge computing,'' \emph{{IEEE} Trans. Wireless Commun.}, vol.~17, no.~4, pp.
  2375--2388, Apr. 2018.

\bibitem{J_zeng16Wireless}
Y.~Zeng, R.~Zhang, and T.~J. Lim, ``Wireless communications with unmanned
  aerial vehicles: opportunities and challenges,'' \emph{{IEEE} Commun. Mag.},
  vol.~54, no.~5, pp. 36--42, May 2016.

\bibitem{J_M.M.Azari18Ultra}
M.~M. Azari, F.~Rosas, K.~Chen, and S.~Pollin, ``Ultra reliable uav
  communication using altitude and cooperation diversity,'' \emph{{IEEE} Trans.
  Commun.}, vol.~66, no.~1, pp. 330--344, Jan. 2018.

\bibitem{J_Xu18Uav}
J.~Xu, Y.~Zeng, and R.~Zhang, ``Uav-enabled wireless power transfer: Trajectory
  design and energy optimization,'' \emph{{IEEE} Trans. Wireless Commun.},
  vol.~17, no.~8, pp. 5092--5106, Aug. 2018.

\bibitem{J_Jeong18Mobile}
S.~Jeong, O.~Simeone, and J.~Kang, ``Mobile edge computing via a uav-mounted
  cloudlet: Optimization of bit allocation and path planning,'' \emph{{IEEE}
  Trans. Veh. Technol.}, vol.~67, no.~3, pp. 2049--2063, Mar. 2018.

\bibitem{J_F.Zhou18Computation}
F.~{Zhou}, Y.~{Wu}, R.~Q. {Hu}, and Y.~{Qian}, ``Computation rate maximization
  in uav-enabled wireless-powered mobile-edge computing systems,'' \emph{{IEEE}
  J. Sel. Areas Commun.}, vol.~36, no.~9, pp. 1927--1941, Sep. 2018.

\bibitem{X.Hu18UAV}
X.~Hu, K.-K. Wong, K.~Yang, and Z.~Zheng, ``Uav-assisted relaying and edge
  computing: Scheduling and trajectory optimization,'' \emph{arXiv preprint
  arXiv:1812.02658}, 2018.

\bibitem{J_Zeng17Energy}
Y.~Zeng and R.~Zhang, ``Energy-efficient uav communication with trajectory
  optimization,'' \emph{IEEE Trans. Wireless Commun}, vol.~16, no.~6, pp.
  3747--3760, Jun. 2017.

\bibitem{J_Zhang13Energy}
W.~Zhang, Y.~Wen, K.~Guan, D.~Kilper, H.~Luo, and D.~O. Wu, ``Energy-optimal
  mobile cloud computing under stochastic wireless channel,'' \emph{{IEEE}
  Trans. Wireless Commun.}, vol.~12, no.~9, pp. 4569--4581, Sep. 2013.

\bibitem{J_Y.Zeng19Accessing}
Y.~Zeng, Q.~Wu, and R.~Zhang, ``Accessing from the sky: A tutorial on uav
  communications for 5g and beyond,'' \emph{arXiv preprint arXiv:1903.05289},
  2019.

\bibitem{B_Boyd04Convex}
S.~Boyd and L.~Vandenberghe, \emph{Convex optimization}.\hskip 1em plus 0.5em
  minus 0.4em\relax Cambridge university press, 2004.

\bibitem{B_D.Bertsekas89Parallel}
D.~P. Bertsekas and J.~N. Tsitsiklis, \emph{Parallel and distributed
  computation: numerical methods}.\hskip 1em plus 0.5em minus 0.4em\relax
  Prentice hall Englewood Cliffs, NJ, 1989, vol.~23.

\bibitem{J_Corless1996LambertW}
R.~M. Corless, G.~H. Gonnet, D.~E. Hare, D.~J. Jeffrey, and D.~E. Knuth, ``On
  the lambertw function,'' \emph{Advances in Computational mathematics},
  vol.~5, no.~1, pp. 329--359, 1996.

\bibitem{M_Grant08CVX}
M.~Grant, S.~Boyd, and Y.~Ye, ``Cvx: Matlab software for disciplined convex
  programming,'' 2008.

\end{thebibliography}