288 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.

1, JANUARY 2009

Downlink Scheduling and Resource Allocation for

OFDM Systems
Jianwei Huang, Member, IEEE, Vijay G. Subramanian, Member, IEEE,
Rajeev Agrawal, and Randall A. Berry, Member, IEEE

Abstract—We consider scheduling and resource allocation for allowed to be traded-off. The “Max Weight” policy (e.g. [7]–
the downlink of a cellular OFDM system, with various practical [9]) can also be viewed as a gradient-based policy, where the
considerations including integer tone allocations, different sub- utility is also a function of the user’s queue-size or delay.
channelization schemes, maximum SNR constraint per tone, and
“self-noise” due to channel estimation errors and phase noise. In TDM systems, one only needs to schedule one user in
During each time-slot a subset of users must be scheduled, and a time-slot and choose the modulation and coding scheme
the available tones and transmission power must be allocated for that user. However, in many current systems, multiple
among them. Employing a gradient-based scheduling scheme pre- users may be multiplexed within a time-slot using Orthog-
sented in earlier papers reduces this to an optimization problem
to be solved in each time-slot. Using a dual formulation, we onal Frequency Division Multiplexing (OFDM) (e.g. IEEE
give an optimal algorithm for this problem when multiple users 802.16/WiMAX [11] and 3GPP LTE [12]). This paper ad-
can time-share each tone. We then give several low complexity dresses gradient-based scheduling and resource allocation for
heuristics that enforce integer tone allocations. Simulations are the downlink of such a system where in addition to determin-
used to compare the performance of different algorithms. ing which users are scheduled, the allocation of physical layer
Index Terms—Orthogonal frequency division multiplexing resources (e.g. transmission power and subcarriers) must also
(OFDM), WiMax, cellular downlink, scheduling, resource allo- be specified.
cation, nonlinear optimization, wireless communications.
Our approach is motivated by [10], where a gradient-based
scheduling algorithm is used for a system which multiplexes
I. I NTRODUCTION users in a time-slot via code division multiple access (CDMA).

M OST recent high-speed wireless data systems dynam-

ically schedule users and allocate physical layer re-
sources among them based on the users’ channel conditions
Compared to CDMA, OFDM offers more degrees of free-
dom to allocate resources across (i.e., tone allocation in the
frequency domain). This enables exploiting both multi-user
and quality of service (QoS) requirements. Many of the diversity and frequency diversity at a finer granularity, but also
scheduling algorithms considered can be viewed as “gradient- significantly increases the complexity of the optimization.
based” algorithms, which select the transmission rate vec- At the beginning of each scheduling interval, the gradient-
tor that maximizes the projection onto the (time-varying) based scheduling algorithm maximizes the weighted through-
gradient of the system’s total utility [1]–[4]. Several such put sum over the current set of feasible rates. In Section II, we
algorithms have been studied for time-division multiplexed give a model for this rate region, taking into account the fol-
(TDM) systems, including the “proportionally fair rule” [4], lowing important practical considerations for OFDM systems:
[6] which is based on a logarithmic utility function of each 1) different subchannelization techniques in which resource
user’s throughput. A larger class of throughput-based utilities allocation is performed at a larger granularity (i.e, groups of
is considered in [2], [5], where efficiency and fairness are tones or symbols) to reduce the channel measurement and
Manuscript received November 13, 2007; revised April 12, 2008 and July feedback overhead; 2) constraints that each subchannel/tone
31, 2008; accepted September 17, 2008. The associate editor coordinating the can be allocated to at most one user; 3) constraints on the
review of this paper and approving it for publication was D. I. Kim. maximum rate per tone to model a limitation on the available
J. Huang is with the Dept. of Information Engineering, The Chinese
University of Hong Kong (e-mail: [email protected]). modulation and coding schemes; and 4) “self-noise” due to
V. G Subramanian is with the Hamilton Institute, National University of channel estimation errors (e.g., [13]) or phase noise [23].
Ireland (email: [email protected]).
R. Agrawal is with Motorola Inc. (e-mail: [email protected]).
In Section III, we consider a dual formulation for the
R. A. Berry is with the Dept. of EECS, Northwestern University (e-mail: resulting optimization problem, which enables us to exploit the
[email protected]). problem’s structure and develop both optimal and simple sub-
Part of this work was done while J. Huang and V. G. Subramanian were
at Motorola. This work has been supported by the Competitive Earmarked
optimal algorithms with low complexity. Simulation results
Research Grants (Project Number 412308) established under the University are given in Section IV for these algorithms with dynamically
Grant Committee of the Hong Kong Special Administrative Region, China, varying weights under different choices of utility functions,
the Direct Grant (Project Number C001-2050398) of The Chinese University
of Hong Kong, SFI grant IN3/03/I346, the Motorola-Northwestern Center
subchannelization schemes, self-noise and per tone rate con-
for Seamless Communications, NSF CAREER award CCR-0238382, and the straints. We conclude in Section V.
National Key Technology R&D Program (Project Number 2007BAH17B04) A number of related formulations without self-noise and per
established by the Ministry of Science and Technology of the People’s
Republic of China. tone rate constraints for downlink OFDM resource allocation
Digital Object Identifier 10.1109/T-WC.2009.071266 have been studied including [14]–[20]. In [15], the goal is
1536-1276/09$25.00 
c 2009 IEEE
HUANG et al.: DOWNLINK SCHEDULING AND RESOURCE ALLOCATION FOR OFDM SYSTEMS 289

to minimize the total transmit power given target bit-rates where wi,t ≥ 0 is a time-varying weight assigned to the ith
for each user. Sum-rate maximization is considered in [16], user at time t tied to the QoS requirements of the user [1]–[4],
[18], [19], where [18], [19] also enforce a minimum bit-rate [7]–[9]. We note that (3) must be re-solved at each scheduling
per user. Weighted sum-rate maximization (for a fixed set instance because of changes in both the channel state and
of weights) is studied in [14], [20]. In [14], a suboptimal the weights (e.g., the gradients of the utilities). While the
algorithm with constant power per tone was shown in sim- former changes are due to the time-varying nature of wireless
ulations to have little performance loss. Other heuristics that channels, the latter changes are due to new arrivals and past
use a constant power per tone are given in [16]–[18]. We also service decisions.
consider such a heuristic in Section III-D. In [20], a similar
dual-based algorithm to ours is considered and simulations are
given which show that the duality gap of this problem quickly
A. OFDM capacity regions
goes to zero as the number of tones increases; we will revisit
this in Section III-B. Finally, in [21], the capacity region of a The solution to (3) depends on the channel state dependent
downlink broadcast channel with frequency-selective fading rate region R(e), where for simplicity we suppress the depen-
using a TDM scheme is given that covers our rate region dence on time. We consider a model appropriate for downlink
without any maximum rate constraints or self-noise. OFDM systems; related models have been considered in [14],
The previous papers optimize a static objective function [21]. In this model, R(e) is parameterized by the allocation
while we are interested in the case where the objective changes of tones to users and the allocation of power across tones. In a
according to a gradient-based algorithm. It is not a priori traditional OFDM system, at most one user may be assigned
clear if a good heuristic for a static problem applied to each to any tone. Initially, as in [15], we make the simplifying
time-step, will be a good heuristic for the dynamic case, assumption that multiple users can share one tone using
since the optimality result in [1]–[4], [7]–[9] is predicated on some orthogonalization technique (e.g. TDM).1 In practice,
solving the optimization problem exactly in each time-slot. if a scheduling interval contained multiple OFDM symbols,
Our simulation results show that the heuristics continue to we can implement such sharing by giving a fraction of the
perform well, at least for the scenarios considered in this paper. symbols to each user. We discuss the case where only one
In a companion paper [25], we use similar methods to solve user can use a tone in Section III-C.
the corresponding uplink problem. A more general solution Let N = {1, . . . , N } denote the set of tones. For each
framework that encompasses both the uplink and downlink j ∈ N and user i ∈ K, let eij be the received signal-
cases is provided in [29]. to-noise ratio (SNR) per unit power. We denote the power
allocated to user i on tone j by pij and the fraction of that
II. P ROBLEM F ORMULATION tone allocatedto user i by xij . The total power allocation
We consider downlink transmissions in an OFDM cell from must satisfy i,j pij  ≤ P , and the total allocation for each
a base station to a set K = {1, . . . , K} of mobile users. In tone j must satisfy i xij ≤ 1. For a given allocation,
each time-slot, the scheduling and resource allocation decision with perfect channel estimation, user i’s feasible rate on tone
p e
can be viewed as selecting a rate vector r t = (r1,t , . . . , rK,t ) j is rij = xij B log(1 + ijxijij ), which corresponds to the
from the current feasible rate region R(et ) ⊆ RK + , where et Shannon capacity of a Gaussian noise channel with bandwidth
indicates the time-varying channel state information available xij B and received SNR pij eij /xij . This SNR arises from
at the scheduler at time t. Following the gradient-based viewing pij as the energy per time-slot user i uses on tone j;
scheduling framework in [1]–[4], an rt ∈ R(et ) is selected the corresponding transmission power becomes pij /xij when
that has the maximum projection
Konto the gradient of a system only a fraction xij of the tone is allocated. Without loss of
utility function U (W t ) := i=1 Ui (Wi,t ), where Ui (Wi,t ) generality we set B = 1 in the following.
is an increasing concave utility function of user i’s average In a realistic OFDM system, imperfect carrier synchro-
throughput, Wi,t , up to time t. In other words, the scheduling nization and channel estimation may result in “self-noise”
and resource allocation decision is the solution to (e.g. [23], [13]). We model this in a similar way as [13]. Let

max ∇U (W t )T · r t = max Ui (Wi,t )ri,t , (1) the received signal on the jth tone of user i be given by yij =
r t ∈R(et ) r t ∈R(et )
i hij sij + nij , where hij , sij and nij are the (complex) channel
gain, transmitted signal and additive noise, respectively, with
where Ui (·)
is the derivative of Ui (·). For example, one class
nij ∼ CN (0, σ 2 ). Assume that hij = h̃ij + hij,δ , where h̃ij
of utility functions given in [2], [5] is 2
 ci is receiver i’s estimate of hij and hij,δ ∼ CN (0, δij ). After
α
Ui (Wi,t ) = α (Wi,t ) , α ≤ 1, α = 0,
(2) matched-filtering, the received signal will be zij = h̃∗ij yij
ci log(Wi,t ), α = 0, resulting in an effective SNR of
where α ≤ 1 is a fairness parameter and ci is a QoS weight.
4
With equal class weights, α = 1 results in the scheduling rule h̃ij pij pij ẽij
Eff-SNR = = , (4)
that maximizes the sum-rate during each slot; α = 0 results 2
σij h̃ij 2 + 2
δij pij h̃ij 2 1 + βij pij ẽij
in the proportionally fair rule.
In general, we consider the problem of
 1 We focus on systems that do not use superposition coding and successive

max wi,t ri,t , (3) interference cancellation within a tone, as such techniques are generally
r t ∈R(et ) considered too complex for practical systems.
i
290 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009

δ2 h̃ 2
where pij = E( sij 2 ), βij = h̃ ij2 and ẽij = σij2 .2 Here, in the jth subchannel and ẽijl is the SNR per unit power
ij ij
βij pij ẽij is the self-noise term. As in the case without self- for tone jl ) for user i in this subchannel is approximately
pij ẽij pe
noise (βij = 0), the effective SNR is still increasing in pij . kxij log(1 + xij +βp ij ẽij
). Since log(1 + x+βpe ) is a concave
However, it now has a maximum of 1/βij . For the sake of function of e, using Jensen’s inequality the rate achieved
presentation, we assume that β = βij for all i and j. The over a subchannel is upper bounded by taking ẽij to be the
analysis is almost identical if users have different βij ’s. arithmetic average of the channel gains of tones in subchannel
With self-noise, user i’s feasible rate on tone j becomes j. The rate can be lower bounded using the strict convexity
pij ẽij of log(l + exp(y)) for y ∈ (with l > 0) and Jensen’s
rij = xij log(1 + xij +βp ij ẽij
), where again xij models time-
sharing of a tone. Under these assumptions, we have inequality. If β = 0, taking y = log pe x and l = 1 we
 lower bound the rate by setting ẽij equal to the geometric
  
R(e) = r : ri =
pij ẽij
xij log 1 + xij +βp , average ofthe subchannel  gains. When β > 0 we take
ij ẽij x
j y = − log 1 + βpe and l = β, apply Jensen’s inequality
   followed by the arithmetic-mean geometric-mean inequality
pij ≤ P, xij ≤ 1 ∀j, (x, p) ∈ X , (5) to lower bound the rate by setting ẽij equal to the harmonic
i,j i average of the subchannel gains. The gap between the upper
N and lower bounds is quite small for reasonable values of pe;
where X := j=1 Xj , and for all j ∈ N ,
for the SNRs achieved by scheduled users in our simulations,
Xj := (xj , pj ) ≥ 0 : xij ≤ 1, pij ≤
xij s̃ij
∀i , (6) we do not see much difference.3 From here onwards we
ẽij
will use the terms tone/carrier/subchannel to mean the basic
with xj := (xij , ∀i ∈ K) and pj := (pij , ∀i ∈ K). allocation unit; the specific distinctions will be clear from the
Γ context.
Here, s̃ij = 1−Γijij β , where Γij < 1/β is a maximum SNR
constraint on tone j for user i, e.g., to model a constraint We consider the following subchannelizations: (i) adjacent
on the maximum rate per tone due to limited availability of channelization, where adjacent tones are grouped together as
modulation and coding schemes. At the cost of additional in the optional “band AMC mode” in IEEE 802.16d/e [11];
complexity, we could also include minimum rate constraints (ii) interleaved channelization, where tones are (perfectly)
to model inelastic traffic, and maximum rate constraints to interleaved as in the interleaved channelization in IEEE
incorporate buffer sizes. 802.16d/e [11]; and (iii) random channelization, where tones
We assume that ẽij is known by the scheduler for all i and j are randomly assigned as in systems that employ frequency
2
as is β (or δij ). In a frequency division duplex (FDD) system, hopping as in the Flash OFDM system [24]. Adjacent chan-
this knowledge can be acquired by having the base station nelization enables the resource allocation to better exploit fre-
transmit pilot signals, from which the users can estimate their quency diversity. Interleaved or random channelization reduces
channel gains and feed them back to the base station. In a the variance of the effective SNR across subchannels for each
time division duplex (TDD) system, these gains can also be user; when the variance is small, user i can simply feed back a
acquired by having the users transmit uplink pilots; the base single ei value. Random channelizations also aid in managing
station can then exploit reciprocity to measure the channel inter-cell interference.
gains. In both cases, this feedback information would need to
be provided within the channel’s coherence time.

B. Subchannelization III. O PTIMAL AND S UBOPTIMAL A LGORITHMS

With many tones and users, providing pilots and/or feed-
back per tone can require excessive overhead; e.g., in IEEE
802.16e [11], a channel with bandwidth 1.25Mhz to 20Mhz From (3) and (5), the scheduling and resource allocation
is divided from 128 to 2048 tones. One way to reduce this problem can be stated as:
overhead is for feedback and resource allocation to be done
at the granularity of subchannels of disjoint sets of tones,   
pij ẽij
i.e., constant power is used and coding is done across the max V (x, p) := wi xij log 1 + xij +βpij ẽij
(x,p)∈X
tones in the same subchannel. Our model can be adapted to i,j
  (7)
this setting by viewing N as the set of subchannels and ẽij subject to: pij ≤ P, and xij ≤ 1, ∀j ∈ N ,
as the effective SNR per unit power for user i on the jth i,j i
subchannel. Specifically, assuming that k tones are bundled
into subchannel j, ẽij is chosen so that the total rate (given by
 pij ẽijl where we still assume that users can time-share subchannels.
xij jl ∈Nj log(1 + xij +βp ) where Nj is the set of tones
ij ẽij l Next we show how to solve (7) via a dual formulation.
2 This is slightly different from the Eff-SNR in [13] in which the signal
power is instead given by hij 4 pij ; the following analysis works for such
a model as well by a simple change of variables. For the problem at hand, 3 For example, in our simulations of the optimal algorithm with β = 0.01,
(4) seems more reasonable in that the resource allocation will depend only the differences between achieved utilities under arithmetic average and har-
on h̃ij and not on hij . We also note that (4) is shown in [22] to give an monic average approximations are 0.005%, 0.1%, and 0.4% under adjacent,
achievable lower bound on the capacity of this channel. interleaved and random subchannelizations, respectively.
HUANG et al.: DOWNLINK SCHEDULING AND RESOURCE ALLOCATION FOR OFDM SYSTEMS 291

0.07

A. Optimal Dual Solution 0.06 β=0

β=0.01
Consider the Lagrangian, L(x, p, λ, µ) := λP +
N j j
j=1 Lj (x , p , λ, μj ), where
0.05

Optimal power pij*(x, λ,μ)

K
  0.04

j j pij ẽij
Lj (x , p , λ, μj ) := wi xij log 1 +
i=1
xij + βpij ẽij 0.03

K
  K
0.02
β=10/e
+ μj 1 − xij − λ pij , (8)
β=0.1
i=1 i=1 0.01

and µ = (μj )N j=1 . The corresponding dual function 0

L(λ, µ) := max(p,x)∈X L(x, p, λ, µ) can then be written as 10 12 14 16 18 20 22
Channel condition e ij (dB)
24 26 28 30

N

L(λ, µ) = λP + max Lj (xj , pj , λ, μj ). (9) Fig. 1. Optimal power p̃ij (15) with wi = 1 versus channel condition eij .
(pj ,xj )∈Xj
j=1
p
By directly evaluating the Hessian of x log(1 + x+βp ) it can
Notice from (10) that the optimal value of p∗ij is always a
be seen that this is jointly concave in (x, p). It follows that
linear function of xij . Substituting (10) into Lj (xj , pj , λ, μj )
Problem (7) is convex and satisfies Slater’s condition. Hence,
also results in a linear function of xij , namely,
there is no duality gap and so V ∗ := minλ≥0,µ≥0 L(λ, µ) is 
the optimal objective value [26]. Lj (xj , pj,∗ , λ, μj ) = xij (μij (λ) − μj ) + μj ,
Next we give a closed-form representation of L(λ, µ) in (9). i
We then show that minimizing L(λ, µ) over µ only requires  
w ẽ
searching for the maximum of user dependent metrics for where μij (λ) := wi h β, iλ ij , s̃ij , and
each tone j. The only numerical search needed is for the 
q(β,(ω−1)+ )∧s̃ij
minimization over λ, which is a one-dimensional search. h (β, ω, s̃ij ) := log 1 + 1+β (q(β,(ω−1)+ )∧s̃ij )
1) Computing the Dual Function: For a given xj , μj and    
λ, the pj which obtains the maximum in (9) is given by − 1 +
q β, (ω − 1) ∧ s̃ij .
ω

p∗ij (x, λ, µ) = xij p̃ij (λ) with From this it follows that any choice
  +   ⎧
w ẽ ⎪
p̃ij (λ) := ẽ1ij q β, iλ ij − 1 ∧ s̃ij ,(10) ⎨{1}, if μij (λ) > μj ;
x∗ij (λ, µ) ∈ [0, 1], if μij (λ) = μj ; (11)
⎪
⎩
where (x)+ = max(x, 0), a ∧ b = min(a, b), and {0}, if μij (λ) < μj ,

z, if β = 0; will maximize Lj (xj , pj,∗ , λ, μj ). Hence, L(λ, µ) := λP +
q(β, z) :=  2β+1    N
j=1 Lj (λ, μj ), where
4β(β+1)
2β(β+1) 1 + (2β+1) 2 z − 1 , if β > 0.
 +
Figure 1 shows p∗ij in (10) as a function of ẽij for Lj (λ, μj ) = μij (λ) − μj + μj . (12)
i
β = 0, 0.01, and 0.1. When β = 0, (10) becomes a “water-
filling” solution in which p∗ij (x, λ, µ) is non-decreasing in 2) Optimizing the Dual Function over λ and µ: Lemma 1
ẽij . For a fixed β > 0, due to self-noise, less power may characterizes the optimization of L(λ, µ) over µ.
be allocated to “better” subchannels. The constant β case is Lemma 1: For all λ ≥ 0,

applicable when the self-noise is due to phase noise as in L(λ) := min L(λ, µ) = λP + μ∗j (λ), (13)
[23]. On the other hand, when self-noise arises primarily from µ≥0
j
estimation errors, β may not be constant but could depend
on the channel quality. The exact dependence will depend on where for every tone j, the minimizing value of μ∗j is
the details of channel estimation. As an example, we also μ∗j (λ) = max μij (λ). (14)
show a curve for when β(e) = 10/e, which is motivated i

by the analysis in [22, Section IV] for the estimation error Proof of Lemma 1 is similar to the proof in [10]. For each
of a Gauss-Markov channel from a pilot with known power. tone j, (14) computes the maximum of user metric μij .
For that model, when the pilot power is either constant or Since L(λ) is the minimum of a convex function over a
inversely proportional to channel quality subject to maximum convex set, it is a convex function of λ; hence, it can be
and minimum power constraints (modeling power control), β minimized using an iterated one dimensional search (e.g.,
will be inversely proportional to e. It can be seen that the curve the Golden Section method [26] for which the computation
has a different shape and amplitude compared to the β = 0 complexity is O(log(1/)), where  is the target relative error
case. For simplicity of presentation, we assume constant β’s bound). Since there is no duality gap, this minimization gives
in the remainder of the paper. the optimal objective value in (7).
292 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009

B. Optimal primal variables with time-sharing Choosing different extreme points only effects the last term
Now we find optimal values of the primal variables (x, p). on the right of (16). It follows that the maximum subgradient
For every λ ≥ 0, with µ∗ (λ) as in (14), let of L(λ) corresponds to the extreme points given by
fˆ(j) := arg min p̃ij , ∀j. (17)
(x∗ , p∗ ) := arg max L (x, p, λ, µ∗ (λ)) ; (15) i∈A(j)
(x,p)∈X
The minimum subgradient corresponds to the extreme points
note that these satisfy (10) and (11).
Given that λ = λ∗ , it follows from duality theory, that f¯(j) := arg max p̃ij , ∀j. (18)
i∈A(j)
if the (x∗ , p∗ ) satisfying (15) are primal feasible and satisfy
complimentary slackness, then they are primal optimal primal. At λ∗ , the maximum subgradient (using (17)) is always
In particular, if for each tone j there exists a unique user i that nonnegative, and the minimum subgradient (using (18)) is
achieves the maximum in (14), then since there is no duality always non-positive. If either is zero, an integer primal optimal
gap, allocating tone j only to that user must be primal optimal. solution is found. In general, we have the following:
In general, given λ ≥ 0, let Aj := {i|μ∗ij (λ) = maxı̂ μ∗ı̂j (λ)} Proposition 2: There exists an optimal primal solution
be the set of users who achieve the maximum on tone j, and (x∗ (λ∗ ), p∗ (λ∗ )), where x∗ (λ∗ ) is given by time-sharing
|Aj | be the size of Aj . From (11) it follows that all x∗ that between the two extreme points in (17) and (18) so that the
solve (15) are those that satisfy the following properties: (i) convex combination of the corresponding subgradients is equal
for i ∈ Aj , x∗ij = 0; (ii) if |Aj | = 1, then x∗ij = 1 for i ∈ Aj ; to zero, and p∗ (λ∗ ) is given by (10).
and (iii) if |Aj | > 1, then for all i ∈ Aj , x∗ij ∈ [0, 1] and
 Proposition 2 implies that each time-shared tone is shared
x∗ = 1. In case (iii), not all tone allocations satisfying in the same proportion.
i∈Aj ij ∗
 ∗
i∈Aj xij = 1 may be primal feasible (e.g., ij pij maybe The above steps give an algorithm for finding the optimal
larger than P ). Breaking these ties is necessary to find a primal solution to (7) in two stages. First, find λ∗ that minimizes L(λ)
optimal solution. A key point is that when ties occur at a given as in Section III-A. This involves evaluating L(λ) for a fixed
λ, L(λ) may not be not differentiable at that λ. However, since value of λ as an inner loop, and a one-dimensional search over
L(λ) is a convex function, subgradients exist [27]. λ as an outer loop. The outer loop has a complexity that is
Proposition 1: For any λ ≥ 0, d is a subgradientof L(λ) independent of N and K. The inner loop has a complexity of
if and only if there exists(x∗ , p∗ ) satisfying (15), i x∗ij ≤ O(N K) due to searching for the maximum of K metrics (14)
1 for all j, μ∗j (λ) 1 − i x∗ij = 0 for each j, and P −
 on each of the N tones. Thus the total complexity of this stage
∗
i,j pij = d. is O(N K). Second, given λ∗ , we compute the maximum and
The proof of Proposition 1 can be found in [28] and follows minimum extreme points and find the optimal primal variables
by observing that that dual function is the maximum of a set as in Proposition 2 which also has a complexity of O(N K).
of Lagrangian functions which are linear in λ and that the Hence, the overall complexity of the optimal algorithm is
gradient of each of the Lagrangian functions (with respect to
O(N K).
λ) is given by P − i,j pij . At any given λ, we need to In our simulations, the actual complexity of the second stage
restrict attention to the maximizing x∗ , p∗ to obtain the set is typically much smaller than O(N K) because “typically”
of subgradients of L(λ). The  rest follows by observing that only a few ties occur. 4 However, the number of extreme
the resulting subgradient P − i,j p̃ij (λ)x∗ij is linear in x∗ij , points can be very large under interleaved channelization. This
which takes values in a convex set (product of simplexes). is because if two users are tied on one subchannel, it is very
Thus, in order to find the dual optimal, we need to search likely that they will also be tied on other subchannels since all
for λ∗ which has a zero subgradient (if λ∗ > 0; and non- subchannels have roughly the same channel gain for the same
negative if λ∗ = 0). From Proposition 1, this will also be user. However, if all the ties are due to the same two users, we
the check for primal feasibility and complimentary slackness can just allocate all subchannels with a tie to the same user
for the power constraint. Next we provide a solution for this and this will lead to either the largest or smallest subgradient.
check. We refer to an allocation as an extreme point if it These observations are consistent with [20], which argues that
satisfies (i)-(iii) and x∗ij ∈ {0, 1} for all i and j; such an an OFDM system with β = 0 in which no time-sharing is
allocation can be represented by a function f : N → K, so allowed will have a certain “duality gap” that is small for
that f (j) ∈ Aj indicates the user who is allocated channel j, a reasonable number of sub-channels. Problem (7) can be
i.e., x∗f (j)j = 1. Let B = {j : |Aj | = 1} and B c = N \ B. viewed as the dual of the dual problem in [20, eqn. (9)] and
For each j ∈ B, there are no ties, and so f (j) is unique. For the duality gap in [20] can be viewed as a measure of the
each tone j ∈ B c , there are |Aj | users accuracy of approximating the OFDMA scheduling problem
 in the tie, and so the
total number of extreme points is j∈Bc |Aj |. Each extreme by the time-sharing version of it from (7). When there is
point satisfies Proposition 1 and so provides a subgradient for exactly one extreme point, the duality gap is clearly zero
L(λ). From Proposition 1 it follows that all the subgradients (since we have an integer solution). The arguments in [20]
of L(λ) can be obtained as a convex combination of the values for a vanishing duality gap roughly correspond to showing
at the extreme points. Given an extreme point f , from (10), 4 For example, extensive simulation results show that for a system of 64
it follows that the corresponding subgradient d(f ) is given by subchannels (grouped from 512 tones) and 40 users in a high mobility
  environment, there are on average only two extreme points typically on one
d(f ) = P − p̃f (j)j − p̃f (j)j . (16) subchannel involving two users, at each scheduling interval (averaged over
j∈B j∈Bc 3000 scheduling intervals) under either adjacent or random channelizations.
HUANG et al.: DOWNLINK SCHEDULING AND RESOURCE ALLOCATION FOR OFDM SYSTEMS 293

that the spread in the power consumption of different extreme than the chosen value of λ, then we decrease λ and recalculate,
points (i.e., the maximum difference in subgradient values) is until a fixed-point is found. It can be shown that the algorithm
typically small for a reasonable number of carriers. When this will stop [28] in at most 2N steps at the correct λ̃. This
spread is small, one expects that fewer ties occur which is algorithm sorts 2N values and thus, has a complexity of
consistent with the above discussion. Discussions above also O(N log N ) which is larger than the O(N ) complexity of the
argue that the conclusions in [20] extend to the β > 0 case. one-dimensional search, but yields the exact optimal solution
in finite time as opposed to an -optimal solution. However,
C. Single user per tone regardless of how the power is allocated, we first need to
find the optimal λ∗ . It follows that if the largest or smallest
We now consider the case where no time-sharing is allowed, subgradients are used to break ties, the overall algorithm
i.e., xij ∈ {0, 1} for all i and j. Suppose we still find the will have a complexity of O(N K) or O(N K + N log N )
optimal λ∗ as in Section III-A. If there are noties on any of the depending on how the power is re-optimized.
tones or if there is an extreme point with j∈N p̃f (j)j = P ,
the optimal primal solution given in Section III-B only has
one user per tone, and we are done. If not, Proposition 2 will D. Single sort suboptimal algorithm
no longer give a solution that satisfies the integer constraints. Now we introduce two sub-optimal algorithms that do not
In this case, a reasonable heuristic is to simply choose one require finding the optimal λ∗ iteratively. Instead, a carrier
extreme point allocation. In our simulations, we choose the allocation is determined by a single sort on each tone based
extreme point corresponding to the subgradient with the small- on some easily calculated metric. These heuristic algorithms
est
 non-negative value; i.e., the extreme point f , for which are much faster than the previous algorithms, although their
j∈N p̃f (j)j is closest to P , without exceeding it. Other rules complexity is again O(N K).
for choosing an extreme point can also be used. Note that this 1) HEURISTIC 1: Each subchannel j is allocated to the
requires searching over all extreme points, which has a worst- user with the largest value of wi R̄ij , where
case complexity of O(K N ) (if all users were tied on every   
tone). However, as discussed above, typically there are only ẽij P/N
R̄ij = log 1 + s̃ij ∧
two users tied on one tone and so this has almost constant 1 + βẽij P/N
complexity. If instead the largest or smallest subgradient was
is the rate user i could achieve on subchannel j under power
used, the worst-case complexity would again be O(N K).
allocation P/N . Any ties are broken arbitrarily, and power
For a given extreme point f , the total transmit power allocation P/N is used. This metric was motivated in part
j∈N p̃f (j)j will be either greater or less than the constraint by work in [14], [16] where a uniform power allocation (not
P (unless this point is optimal). We then need to re-optimize
necessarily over all tones) was shown to be nearly optimal.
the power allocation for the given fixed feasible tone allocation
2) HEURISTIC 2: Here subchannels are allocated as in
x (i.e., xij = 1 if i = f (j), otherwise xij =0), i.e., solve
HEURISTIC 1. However, after this procedure, an optimal

max V (x, p) s.t. pij ≤ P. (19) power allocation is performed as in Section III-C (instead
p:(p,x)∈X
i,j of power allocation P/N ). It may turn out that no power is
assigned to some subchannels.
Let Lx (λ) be the dual function for this problem. Given
λ̃ = arg minλ≥0 Lx (λ), the optimal power allocation to (19)
IV. S IMULATION S TUDY
is given by (10) with λ = λ̃ and the given tone allocation x. A
simple one-dimensional search once again yields the optimal We report simulation results based on a realistic OFDM
λ. This will have a complexity of O(N ) (to get within  of simulator with assumptions and parameters commonly used
the optimal) since each tone has at most one user. in IEEE 802.16 standards [11]. We focus on the following
When the self-noise term β = 0, we can actually find the algorithms: the OPTIMAL algorithm which finds the optimal
optimal λ̃ in finite steps based on the following alternative λ∗ and then chooses a tone-allocation with one user per
characterization of λ̃, the proof of which is based on a similar tone as described in Section III-C5 , and HEURISTIC 1 and
argument as in [10]. HEURISTIC 2 from Section III-D.
Proposition 3: For β = 0 a given λ̂ is the unique optimal We simulate a single OFDM cell with M = 40 users
solution to the dual problem minλ≥0 Lx (λ) if and only if and a total transmission power of P = 6W at the base
 station. The channel gains eij ’s are the product of a fixed
i,j xij wi 1{λ̂∈Wij } location-based term for each user i and a frequency-selective
λ̂ =  Γij  , (20)
P − i,j eij 1{λ̂∈Yij } + i,j e1ij 1{λ̂∈Wij } fast-fading term. The location-based components are picked
    using an empirically obtained distribution for many users in
xij wi eij xij wi eij
where Wij = 1+Γ , x ij wi e ij , and Yij = 0, 1+Γij . a large system. The fast-fading term is generated using a
ij
Proposition 3 suggests the following algorithm [28] for block-fading model based upon the Doppler frequency (for
finding λ̃. First check if the power constraint is violated when 5 We simulated both the algorithms in Section III-B and III-C, and found
all
 users use maximum power on the allocated tones, i.e., if that they have identical performance under all parameter choices. This could
xij
(i,j) eij Γij > P . If this is false, the problem is solved. be due to the fact that the gap in making the time-sharing assumption is small
owing to there being very few significantly different extreme points at each
Otherwise, we need to search for λ̃ by starting from the largest scheduling interval as discussed at the end of Section III-B. We thus refer to
λ, and calculating the right side of (20). If the result is less the algorithm in Section III-C simply as the OPTIMAL algorithm.
294 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009

TABLE I
P ERFORMANCE FOR DIFFERENT CHOICES OF α ( ADJACENT
the block-length in time) and a standard reference mobile CHANNELIZATION , NO - SELF - NOISE , NO SNR CONSTRAINTS ).
delay-spread model (for variation in frequency). For the fast-
α Algorithm Utility Log U Rate Num.
fading terms, each multi-path component is held fixed for
0 OPTIMAL 10.74 10.74 60.8 7.73
2msec (i.e., a fading block length), which corresponds to a 0 HEURISTIC 1 10.66 10.66 54.6 7.29
250Hz Doppler frequency. The delay-spread is set to 1μsec. 0 HEURISTIC 2 10.72 10.72 57.3 7.35
The users’ channel conditions are averaged over the applicable 0.5 OPTIMAL 545.2 10.83 105.9 7.32
subchannelization scheme and fed back to the scheduler. 0.5 HEURISTIC 1 528.8 10.73 99.3 7.20
We consider a system bandwidth of 5MHz consisting of 0.5 HEURISTIC 2 542.8 10.81 103.2 7.01
512 OFDM tones, grouped into 64 subchannels (8 tones per 1 OPTIMAL 261677 6.79 261.7 2.58
1 HEURISTIC 1 261676 6.79 261.7 2.58
subchannel). The symbol duration is 100μsec with a cyclic 1 HEURISTIC 2 261676 6.77 261.7 2.58
prefix of 10μsec, which roughly corresponds to 20 OFDM
symbols per fading block (i.e., 2msec). This is one of the
allowed configurations in the IEEE 802.16 standards [11].
TABLE II
Resource allocation (i.e. solving (1)) is done once per fading P ERFORMANCE OF DIFFERENT SUBCHANNELIZATION SCHEMES (α = 0.5,
block. All the results are averaged over the last 2000 OFDM NO SELF - NOISE , NO SNR CONSTRAINTS ).
symbols out of 60000 OFDM symbols (i.e., 3000 fading
blocks) by which time we can be reasonably confident that
the system has reached stationarity. All users are infinitely Channelization Algorithm Utility Log U Rate Num.
back-logged and assigned a throughput-based utility as in (2) Adjacent OPTIMAL 545.15 10.83 105.9 7.32
Adjacent HEURISTIC 1 528.83 10.73 99.3 7.20
with parameter ci = 1 and the same fairness parameters (α) Adjacent HEURISTIC 2 542.84 10.81 103.2 7.01
across users. Interleaved OPTIMAL 494.61 10.53 92.4 1.79
The rate of user i on subchannel j is calculated as Interleaved HEURISTIC 1 486.40 10.47 88.4 1.14
 Interleaved HEURISTIC 2 487.02 10.48 87.8 1.15
0.56pij ẽij
rij = 0.28Bxij log 1 + , Random OPTIMAL 487.53 10.53 89.2 4.89
xij + βpij ẽij Random HEURISTIC 1 479.07 10.46 84.2 4.39
Random HEURISTIC 2 485.63 10.51 86.5 4.34
where B is the subchannel bandwidth. Here 0.56 accounts for
the “SNR gap” due to limited modulation and coding choices
and 0.28 accounts for various factors such as hybrid ARQ
transmission scheme and the overhead due to guard tones and
control symbols, etc. While the scheduling is based on the
geometric average for β = 0 and harmonic average for β > 0,
the decoded rate is based on per tone channel conditions.
The first set of simulation results are for a system with
adjacent channelization, no self-noise (β = 0), and no per-user Indeed, for the channel model used here, in the interleaved
SNR constraints (i.e., Γij = ∞ for all i and j). Table I shows case all subchannels can be shown to be almost identical,
the results for all three algorithms under different choices of explaining why it typically schedules only one or two users.
the utility parameter α. The column “Utility” gives the average Next we consider the case when the self-noise coefficient
utility per user for each algorithm. The column “log U” shows β = 0.0056 in Table III. Here we assume α = 0.5, and no per-
the log utility per user; this gives an alternate indication of user SNR constraint. The performance gap between the three
the “fairness” of the resulting allocation (same as utility for algorithms is slightly larger compared to the case without self-
α = 0). The column “Rate” is the average throughput per user noise in Table II.
in Kbps, and the final column is the average number of users
scheduled per scheduling interval. For each choice of α, the Figure 2 shows the throughput CDFs for all three al-
three algorithms perform close to each other for each of these gorithms, with β = 0.0056 and β = 0. Here adjacent
metrics. HEURISTIC 2 performs better than HEURISTIC 1, channelization is used, α = 0.5, and s̃ij = ∞ for all i and j.
since the former re-optimizes the power allocation after tone It is clear that users achieve better throughput when there is
allocation, and the latter just uses constant power allocation. no self-noise (β = 0). For each β the OPTIMAL algorithm
When α = 1 (maximum throughput), all three algorithms have always achieves better rates compared to the HEURISTIC
almost identical performance. ones.
Next we consider the effect of different subchannelization Table IV illustrates the effect of SNR constraints. In par-
schemes. Table II shows the performance of the three algo- ticular, we choose the SNR constraint to be ∞, 32.5dB,
rithms for the adjacent, random, and interleaved channelization and 22.5dB, respectively, and the same across all users and
schemes from Section II-A. We set α = 0.5, β = 0, all tones. We choose adjacent channelization with utility
and Γij = ∞ for all i and j. Again, both HEURISTIC parameter α = 0.5 and no self-noise. Compared to the no
algorithms perform close to the OPTIMAL algorithm. In all SNR constraints case, a constraint of 32.5dB does not change
cases, interleaved and random channelizations result in lower the results significantly, while a constraint of 22.5dB substan-
utility than the adjacent channelization. This is likely due tially decreases the achievable rates (13% for the OPTIMAL
to higher frequency diversity with adjacent channelization. algorithm and 27% for HEURISTIC 1 algorithm).
HUANG et al.: DOWNLINK SCHEDULING AND RESOURCE ALLOCATION FOR OFDM SYSTEMS 295

TABLE III TABLE IV

P ERFORMANCE OF DIFFERENT SUBCHANNELIZATION SCHEMES (α = 0.5, P ERFORMANCE OF DIFFERENT SNR CONSTRAINTS ( ADJACENT
β = 0.0056, NO SNR CONSTRAINTS ). CHANNELIZATION , α = 0.5, NO SELF - NOISE ).

Channelization Algorithm Utility Log U Rate Num. SNR Max Algorithm Utility Log U Rate Num.
Adjacent OPTIMAL 512.20 10.82 82.5 7.52 ∞ OPTIMAL 545.15 10.83 105.9 7.32
Adjacent HEURISTIC 1 489.32 10.70 73.7 7.40 ∞ HEURISTIC 1 528.83 10.73 99.3 7.20
Adjacent HEURISTIC 2 504.00 10.78 77.2 7.22 ∞ HEURISTIC 2 542.84 10.81 103.2 7.01
Interleaved OPTIMAL 467.00 10.51 73.5 1.98 32.5dB OPTIMAL 542.78 10.83 102.97 7.33
Interleaved HEURISTIC 1 453.16 10.43 66.8 1.26 32.5dB HEURISTIC 1 519.81 10.72 91.87 7.25
Interleaved HEURISTIC 2 454.59 10.44 66.9 1.27 32.5dB HEURISTIC 2 535.89 10.81 96.35 7.10
Random OPTIMAL 460.53 10.51 71.6 5.60 22.5dB OPTIMAL 522.48 10.82 88.11 7.40
Random HEURISTIC 1 446.58 10.42 64.7 4.89 22.5dB HEURISTIC 1 483.50 10.66 72.60 7.09
Random HEURISTIC 2 453.51 10.48 66.1 4.85 22.5dB HEURISTIC 2 505.81 10.77 78.61 6.92

1
the performance gap widens when per user SNR constraints
or channel estimation errors are considered.
0.95
OPTIMAL (β=0)
HEURISTIC 2 (β=0) R EFERENCES
0.9 HEURISTIC 1 (β=0) [1] R. Agrawal and V. Subramanian, “Optimality of certain channel aware
scheduling policies,” in Proc. 2002 Allerton Conference, 2002.
OPTIMAL (β=0.0056) [2] R. Agrawal, A. Bedekar, R. La, and V. Subramanian, “A class and
0.85 channel-condition based weighted proportionally fair scheduler,” in
HEURISTIC 2 (β=0.0056) Proc. ITC 2001, Salvador, Brazil, Sept. 2001.
[3] A. L. Stolyar, “On the asymptotic optimality of the gradient scheduling
0.8 algorithm for multiuser throughput allocation,” Operations Research,
HEURISTIC 1 (β=0.0056) vol. 53, no. 1, pp. 12–25, 2005.
[4] H. Kushner and P. Whiting, “Asymptotic properties of proportional-fair
0.75 sharing algorithms,” in Proc. 40th Annual Allerton Conference, 2002.
[5] J. Mo and J. Walrand, “Fair end-to-end window-based congestion
control,” IEEE/ACM Trans. Networking, vol. 8, no. 5, pp. 556–567,
Oct. 2000.
0.7
0 200 400 600 800 1000 [6] A. Jalali, R. Padovani, and R. Pankaj, “Data throughput of CDMA-
Throughput in Kbps HDR a high efficiency-high data rate personal communication wireless
system,” in Proc. IEEE VTC, Spring, 2000.
[7] L. Tassiulas and A. Ephremides, “Dynamic server allocation to parallel
Fig. 2. Empirical CDF of users’ throughputs (adjacent channelization, α = queue with randomly varying connectivity,” IEEE Trans. Inform. Theory,
0.5, no per-user SNR constraints). vol. 39, pp. 466–478, 1993.
[8] M. Andrews, et al., “Providing quality of service over a shared wireless
link,” IEEE Commun. Mag., vol. 39, no. 2, pp. 150–154, Feb. 2001.
[9] A. L. Stolyar, “Maximizing queueing network utility subject to stability:
V. C ONCLUSIONS greedy primal-dual algorithm,” Queueing Syst., vol. 50, no. 4, pp. 401–
457, 2005.
We have considered the problem of gradient-based schedul- [10] R. Agrawal, V. Subramanian, and R. Berry, “Joint scheduling and
ing and resource allocation for a downlink OFDM system, resource allocation in CDMA systems,” in Proc. WiOpt, Cambridge,
which essentially reduces to solving a convex optimization UK, Mar 2004 (journal version under submission).
[11] “IEEE 802.16e-2005 and IEEE Std 802.16-2004/Cor1-2005”
problem in each time-slot. We studied this problem for a (http://www.ieee802.org/16/).
model that accommodates various choices for user utility [12] “Long Term Evolution of the 3GPP radio technology,”
functions, different subchannelization techniques, and self- http://www.3gpp.org/Highlights/LTE/LTE.htm.
[13] H. Jin, R. Laroia, and T. Richardson, “Superposition by position,” in
noise due to imperfect channel estimates or phase noise. Using Proc. IEEE ITW 2006, Mar. 2006.
duality theory we first gave an optimal algorithm for solving [14] L. Hoo, B. Halder, J. Tellado, and J. Cioffi, “Multiuser transmit
a relaxed version of this problem in which users can time- optimization for multicarrier broadcast channels: asymptotic FDMA
capacity region and algorithms,” IEEE Trans. Commun., vol. 52, no. 6,
share each subchannel. This involves finding a maximum of pp. 922–930, 2004.
a per user (closed-form) metric for each subchannel and a [15] C. Y. Wong, R. S. Cheng, K. B. Letaief, and R. D. Murch, “Multiuser
one-dimensional search of an optimal dual variable. More OFDM with adaptive subcarrier, bit, and power allocation,” IEEE J.
Select. Areas Commun., vol. 17, no. 10, pp. 1747–1758, 1999.
interestingly, this algorithm typically automatically yields an [16] J. Jang and K. Lee, “Transmit power adaptation for multiuser OFDM
integer carrier allocation (except on one or two tones). To systems,” IEEE J. Select. Areas Commun., vol. 21, no. 2, pp. 171–178,
enforce such a constraint on all tones, we further proposed 2003.
[17] Y. Zhang and K. Letaief, “Multiuser adaptive subcarrier-and-bit allo-
an algorithm that picks an integer carrier allocation and re- cation with adaptive cell selection for OFDM systems,” IEEE Trans.
optimizes the power allocation accordingly. The numerical Wireless Commun., vol. 3, no. 5, pp. 1566–1575, 2004.
performance of this algorithm is almost identical to the optimal [18] T. Chee, C. Lim, and J. Choi, “Adaptive power allocation with user pri-
oritization for downlink orthogonal frequency division multiple access
solution of the relaxed problem. Finally, we proposed two systems,” in Proc. ICCS 2004, pp. 210–214, 2004.
even simpler suboptimal algorithms that only perform a single [19] H. Yin and H. Liu, “An efficient multiuser loading algorithm for OFDM-
sort on each of the tones and avoid any iterative calculations. based broadband wireless systems,” in Proc. IEEE Globecom, 2000.
[20] K. Seong, M. Mohseni, and J. M. Cioffi, “Optimal resource allocation
Simulations show that the suboptimal algorithms achieve close for OFDMA downlink systems,” in Proc. IEEE ISIT, pp. 1394–1398,
to optimal performance under a wide range of scenarios, and 2006.
296 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009

[21] L. Li and A. Goldsmith, “Capacity and optimal resource allocation for Vijay G. Subramanian (M’01) received his Ph.D.
fading broadcast channels–part I: ergodic capacity,” IEEE Trans. Inform. degree in Electrical Engg. from the University of
Theory, vol. 47, no. 3, pp. 1083–1102, 2001. Illinois at Urbana-Champaign, in 1999. From 1999
[22] M. Medard, “The effect upon channel capacity in wireless communica- to 2006, he was with the Networks Business, Mo-
tions of perfect and imperfect knowledge of the channel,” IEEE Trans. torola, Arlington Heights, IL, USA. Since May 2006
Inform. Theory, vol. 46, no. 3, pp. 935-946, May 2000. he is a Research Fellow at the Hamilton Institute,
[23] J. Lee, H. Lou and D. Toumpakaris, “Analysis of phase noise effects NUIM, Ireland. His research interests include infor-
on time-direction differential OFDM receivers,” in Proc. IEEE GLOBE- mation theory, communication networks, queueing
COM, 2005. theory, and applied probability and stochastic pro-
[24] QUALCOMM Flarion Technologies, http://www.qualcomm.com/qft/. cesses.
[25] J. Huang, V. Subramanian, R. Berry, and R. Agrawal, “Joint scheduling
and resource allocation in uplink OFDM systems for broadband wireless
access networks,” IEEE J. Select. Areas Commun., to appear, 2009.
[26] D. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA: Athena
Scientific, 1999.
[27] R. Rockafellar, Convex Analysis. Princeton University Press, 1970.
[28] J. Huang, V. Subramanian, R. Berry, and R. Agrawal, “Downlink
scheduling and resource allocation for OFDM systems,” tech. report. Rajeev Agrawal is a Fellow of the Technical Staff
Available: http://personal.ie.cuhk.edu.hk/∼ jwhuang/publication/OFDM at Motorola where his responsibilities include the
DL Technical Report.pdf. architecture, design and optimization of Motorola’s
[29] J. Huang, V. G. Subramanian, R. Berry, and R. Agrawal, “Scheduling next generation wireless systems. Prior to joining
and Resource Allocation in OFDMA Wireless Systems,” book chapter, Motorola in 1999, Rajeev was Professor of Electrical
submitted. and Computer Engg. and Computer Science depart-
ments at the University of Wisconsin - Madison.
He also spent a sabbatical year at IBM TJ Watson
Research, British Telecom Labs, and INRIA-Sophia
Antipolis. Rajeev received his M.S. (1987) and
Jianwei Huang (S’01-M’06) is an Assistant Pro- Ph.D. (1988) degrees in Electrical Engg.-systems
fessor in Information Engineering Department at the from the University of Michigan, Ann Arbor and his B.Tech. (1985) degree
Chinese University of Hong Kong. He received the in Electrical Engg. from the Indian Institute of Technology, Kanpur.
M.S. and Ph.D. degrees in Electrical and Computer
Engineering from Northwestern University in 2003
and 2005, respectively. From 2005 to 2007, He
worked as a Postdoctoral Research Associate at Randall A. Berry (S’93-M’00) received the M.S.
Princeton University. His main research interests lie and PhD degrees in Electrical Engg. and Computer
in the area of modeling and performance analysis of Science from the Massachusetts Institute of Tech-
communication networks, including cognitive radio nology in 1996 and 2000, respectively. In September
networks, OFDM and CDMA systems, wireless 2000, he joined the faculty of Northwestern Univer-
medium access control, multimedia communications, network economics, sity, where he is currently an Associate Professor of
and applications of optimization theory and game theory. Dr. Huang is Electrical Engg. and Computer Science. Dr. Berry
an Associate Editor of J OURNAL OF C OMPUTER & E LECTRICAL E NGI - is the recipient of a 2003 NSF CAREER award.
NEERING , a Guest Editor for IEEE J OURNAL OF S ELECTED A REAS IN He is currently serving on the editorial boards of
C OMMUNICATIONS and J OURNAL OF A DVANCES IN M ULTIMEDIA, and a the IEEE T RANSACTIONS ON I NFORMATION T HE -
TPC Co-Chair of the International Conference on Game Theory for Networks ORY and the IEEE T RANSACTIONS ON W IRELESS
(GameNets’09). C OMMUNICATIONS.