Вы находитесь на странице: 1из 9

1594 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.

5, MAY 2008

Effects of Non-homogeneous SINR Estimation


Error Statistics on Scheduling Performance in
OFDM Downlink Systems
Ji-Woong Choi, Member, IEEE, and John M. Cioffi, Fellow, IEEE

Abstract— This paper investigates the effect of inaccurate derived for PF scheduled OFDM systems with the assumption
SINR estimation on the performance of multi-user scheduling of the same average SINR for each user [6,7]. Although these
in downlink orthogonal frequency division multiplexing (OFDM) results provide some intuition on the effect of inaccurate SINR
systems. Considering path loss, time and frequency selectivity of
the channel, the error statistics of SINR estimation is analyzed. estimation, they only consider the case when every user has
Then, the effect of SINR inaccuracy is evaluated in respect to the same error statistics of SINR estimation. However, it is
system throughput and fairness when each user has different required in practice to consider the situations where each user
error statistics according to channel environments. experiences different channel environments that include path
Index Terms— OFDM systems, downlink scheduling, SINR loss, maximum Doppler spread, and delay spread, leading to
estimation. non-homogeneous error statistics on SINR estimation.
This paper investigates the effect of non-homogeneous
I. I NTRODUCTION property of SINR estimation on the performance of a multi-
user OFDM downlink system; especially for PF and MS

I N recent years, multi-user orthogonal frequency division


multiplexing (OFDM) has attracted much attention as
an effective transceiver technique for high-speed multi-user
scheduling. To this end, the tendency of error statistics with
respect to variation of channel environment is presented,
which is followed by the evaluation of system throughput
wireless systems that can provide high spectral efficiency by
and fairness. The rest of the paper is organized as follows:
incorporating various advanced technologies [1]. Though most
Section II describes the system model, and the error statistics
of previous researches have assumed perfect channel state
of SINR estimation is analyzed in Section III. Section IV
information (CSI) in the transmitter and receiver [2], it is
evaluates the effect of non-homogeneous SINR estimation
hard in practice to obtain perfect CSI such as the signal to
error on the system performance. Finally, Section V provides
interference and noise power ratio (SINR).
the concluding remarks.
Recently, a few researchers have investigated the effect
of inaccurate SINR estimation on the performance of multi-
user scheduling [3-7]. When proportional fair (PF) and max- II. S YSTEM M ODEL
imum SINR (MS) scheduling schemes are employed, the The block diagram of an OFDM downlink system is de-
data rate can be maximized by optimizing SINR thresholds picted in Fig. 1. For simplicity of analysis, the model divides
for modulation level selection in consideration of the SINR mobile terminal users into two groups: Ms users near the
estimation error, which is caused by channel feedback delay base station and Ml users near the cell boundary, respectively.
for downlink systems [3]. Assuming that the channel state is M (= Ms + Ml ) users receive common pilot signals from
binary, i.e., either a bad or a good channel, [4] numerically the base station and estimate SINR for the next time of
evaluates the effect of user mobility on the SINR estimation scheduling. Every user then returns its SINR information
error and system throughput for downlink systems using PF γ̂m (= [γ̂m,1 γ̂m,2 · · · γ̂m,F ]) for F subchannels to the base
scheduling. Moreover, [5] investigates the effect of SINR station. Finally, in each subchannel, the base station chooses
estimation error on the data rate and proposes an improved one user according to a scheduling policy and sends the packet
SINR estimator for downlink scheduling. When the SINR to the selected user.
estimation error occurs because of the feedback delay and X[n, k] represents transmit signal at the n-th symbol time
noise in the pilot signal, the achievable rate is analytically on the k-th subcarrier where k ∈ {0, 1, · · · , K − 1}. The
frequency domain symbol is converted into a time domain
Manuscript received June 16, 2006; revised October 18, 2007; accepted
January 23, 2008. The associate editor coordinating the review of this letter signal using inverse fast Fourier transform (IFFT) and a cyclic
and approving it for publication was R. Berry. This work was supported in part prefix (CP) is inserted to preserve the orthogonality among the
by the Information and Telecommunication National Scholarship Program of subcarriers as well as to eliminate the interference from the
the Ministry of Information and Communication, Korea.
J.-W. Choi was with the Department of Electrical Engineering, Stanford adjacent OFDM symbols. It is assumed that each data packet
University, Stanford, CA 94305, USA. He is now with Marvell Semiconduc- comprises Nt OFDM symbols and Nf subcarriers, where F
tor, Inc., Santa Clara, CA 95054, USA (e-mail: cjiwoong@marvell.com). (=K/Nf ) packets are scheduled each scheduling time (F is
J. M. Cioffi is with the Department of Electrical Engineering, Stanford
University, Stanford, CA 94305, USA (e-mail: cioffi@stanford.edu). assumed to be an integer). The pilot symbol is periodically
Digital Object Identifier 10.1109/TWC.2008.060379. inserted with a rectangular pattern (i.e., apart by dt and df
1089-7798/08$25.00 
c 2008 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
CHOI and CIOFFI: EFFECTS OF NON-HOMOGENEOUS SINR ESTIMATION ERROR STATISTICS ON SCHEDULING PERFORMANCE 1595

Ms : Number of users near the base station


Ml : Number of user near the cell boundary At the receiver, every user estimates the CIR Ĝm [n, k] at
γˆ : Estimated SINR
any n and k using the received pilot symbols. The estimated
γˆM s +1 CIRs are then used for coherent detection as well as for
γˆ1 calculating the SINR. If the multi-user diversity (MUD) is
γˆm γˆM utilized by employing opportunistic packet scheduling policies
User Ms+1
s s
γˆM s +Ml
User 1
User Ms
such as the MS and PF scheduling, the system throughput can
User ms
User Ms+Ml be significantly increased [8-10]. The MS scheduling selects a
γˆM s + ml user with the highest instantaneous SINR, which is formulated
User Ms+ml as

mQ [n, k] = arg max{γm [n, k]} (5)


Fig. 1. System framework m∈{1,2,··· ,M }

2 2 2
where γm [n, k](= ηm |Hm [n, k]| σZ ) is the instantaneous
symbols in the time and frequency grid, respectively). No SINR of user m. Here, γm [n, k]’s are assumed to be the same
transmit power control is employed at the base station. for all n and k in a packet assuming negligible change of
The received signal of user m, m ∈ {1, 2, · · · , M }, can be the SINR in the packet. In the following, the index [n, k]
written in the frequency domain as represents a packet which the symbol at [n, k] belongs to for
simplicity of description. Though MS scheduling maximizes
Ym [n, k] = ηm Hm [n, k]X[n, k] + Zm [n, k] (1) the system throughput by utilizing MUD, it cannot guaran-
= Gm [n, k]X[n, k] + Zm [n, k] tee the fairness among users. This fairness problem can be
2 alleviated by employing the PF scheduling scheme [9].
where ηm and Hm [n, k] are the path gain (i.e., inverse of
path loss in large-scale fading) and small-scale frequency- Let Rm [n, k] be the possible transmission data rate of user
domain channel impulse response (CIR) of user m, respec- m at [n, k] and R̄m [n, k] be the average data rate up to time
tively. Zm [n, k] is the background noise plus interference n. Then, the PF scheduler for subcarrier k selects the user
term, which can be approximated as a zero mean white according to
 
circularly symmetric complex Gaussian noise (ZMWCSCGN) Rm [n, k]
2 mQ [n, k] = arg max (6)
with variance σZ,m . Each user is assumed to have identical R̄m [n, k]
2 2 m∈{1,2,···M }
noise variance (σZ,m = σZ ). Without loss of generality,
2
E{|X[n, k]| } = 1 is assumed for simplicity where E denotes which can be rewritten by [10]
an ensemble average. Ts and fc are used to denote the symbol  
γm [n, k]
duration and subcarrier spacing, respectively. Also, Hm [n, k] mQ [n, k] = arg max (7)
is the Fourier transform (FT) of the time domain CIR hm (t, τ ) m∈{1,2,··· ,M } γ̄m
 
at t=nTs and f = kfc , which is represented as = arg max |Hm [n, k]|2
m∈{1,2,··· ,M }

Lm

hm (t, τ ) = hm,l (t)δ(τ − τm,l ) (2) 2 2 2 2
2
where γ̄m = E{γm [n, k]}=ηm E{|Hm [n, k]| } σZ =ηm σZ .
l=1

where δ(·) is the Dirac delta function, Lm , τm,l and hm,l (t)
are the number of multipaths, delay, and complex-valued CIR III. SINR E STIMATION
at time t of the l-th path for user m, respectively. {hm,l (t)} is Since the perfect measurement of the SINR γm [n, k] is
modeled as an independent ZMWCSCGN with average power infeasible in practice, the estimated value γ̂m [n, k] is used
2

Lm
2
σm,l such that σm,l = 1. instead. The estimation of the average SINR γ̄m is generally
l=1 quite accurate since it can be measured over a long time
The correlation functions of the CIR in the time and
interval [11]. On the other hand, the instantaneous gain
frequency domains are given by
|Hm [n, k]| of the channel is more erroneous since it should
∗ be estimated during a short time period where the CIR of
rHt,m [nt ] = E{Hm [n, k]Hm [n + nt , k]} (3)

rHf,m [kf ] = E{Hm [n, k]Hm [n, k + kf ]}. the channel does not vary significantly. Moreover, the receiver
usually should predict Hm [n, k] at time (n − Δ) considering
Also, the corresponding channel spectra, i.e., Doppler spec- a feedback delay Δ for the base station to schedule at time
trum and power-delay profile, can be represented as n, resulting in an increase of the SINR estimation error. As a

 result, the average SINR is assumed to be perfectly estimated
SHt,m (w) = rHt,m [n]e−jwn (4) while the instantaneous SINR estimate is erroneous. Here, we
n=−∞ assume that the signal power is calculated by squaring the
∞
amplitude of the estimated frequency-domain CIR instead of
SHf,m (w) = rHf,m [k]e−jwk direct computation of the power since its operation can be
k=−∞
shared with channel estimation that is necessary for coherent
where SHt,m (w) and SHf,m (w) are the FT of rHt,m [n] and detection. The following part describes the estimation of the
rHf,m [n]. instantaneous channel gain.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
1596 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 5, MAY 2008

2 2
The CIR corresponding to the pilot symbol for user m can where σs,m and σi,m are the MSE of user m from the self-
be estimated as distortion and interference, respectively. Wm (wt , wf ) is the
2-D FT of wm [p, q], and SHs,m (wt , wf ) is a sampled version
G̃m [np , kp ] = Ym [np , kp ]/X[np , kp ] (8)
of SHm (wt , wf ) [11]
= ηm Hm [np , kp ] + Z̃m [np , kp ]
SHs,m (wt , wf )
= Gm [np , kp ] + Z̃m [np , kp ]
f −1
t −1 d
d
1 2π 2π
where np and kp denote the symbol and subcarrier index of the = 2 SHm (wt − n, wf − k) (14)
pilot symbol, respectively, and Z̃m [np , kp ] denotes the noise (dt df ) n=0 k=0
dt df
term. The CIR at the data symbol is usually estimated by where
interpolating or extrapolating the CIRs at the pilot symbol. In
this paper, the interpolation is assumed to mean both interpo- SHm (wt , wf ) = SHt,m (wt )SHf,m (wf ). (15)
lation and extrapolation. Letting |X[np , kp ]| = 1 without loss
The interpolation error coefficient occurred by using a non-
of generality, Z̃[np , kp ] can be assumed to be a ZMWCSCGN
2 ideal interpolator is defined as
with variance σZ . The channel for user m is assumed to be
band-limited with the maximum Doppler frequency fdm and we,m [p, q] = wm [p, q] − wid [p, q]. (16)
the maximum time delay of τLm . When Nyquist sampling
theorem holds, i.e., dt fdm Ts ≤ 1/2 and df τLm fc ≤ 1/2 for Then, the 2-D FT of we,m [p, q] can be given by (17). Exclud-
m=1, 2,· · · , M , the CIR at the data symbol can be perfectly ing the negligible MSE in the out-of-passband spectrum (i.e.,
obtained by interpolating the CIRs at the adjacent noise-free wt ∈/ [−π/dt π/dt ] and wf ∈ / [−π/df π/df ]) [11], the self-
2
pilot symbols as [11] distortion MSE σs,m can be approximated by (18), where χm
denotes an index for the amount of the self-distortion, which is

 ∞
 a function of channel spectra and interpolator in a normalized
Gm [n, k] = Gs,m [n + p, k + q]wid [p, q] (9)
sense. The previous work in [11] shows that this index can be
p=−∞ q=−∞
generally approximated by using a few low order moments of
where Gs,m [n, k] is equal to Gm [n, k] at the pilot symbol, and the channel spectra as follows.
zero, otherwise. Also, wid [p, q] is the impulse response of an 1
ideal 2-dimensional (2-D) non-causal brick-wall interpolator χm = 2
represented as (2π) (dt df )2

π
π
sin(πp/dt ) sin(πq/df ) 2
SHm (wt , wf ) |We,m (wt , wf )| dwt dwf
wid [p, q] = . (10)
(πp/dt )(πq/df )
−π −π
Impracticality of wid [p, q] suggests the use of a non-ideal (2) (2) (4) (4)
≈ cIFm ,0 w̄t,m w̄f,m + cIFm ,1 w̄t,m + cIFm ,2 w̄f,m (19)
interpolator as
(n) (n)

 ∞
 where w̄m,t and w̄m,f are respectively the n-th order moment
Ĝm [n, k] = G̃s,m [n + p, k + q]wm [p, q] (11) of the Doppler spectrum and power delay profile represented
p=−∞ q=−∞ as
where wm [p, q] is the interpolator’s coefficient of user m, and
π
(n) 1
G̃s,m [n, k] is equal to G̃m [n, k] at the pilot symbol and zero, w̄t,m = wn SHt,m (w)dw (20)

otherwise. −π
Using (9) and (11), Ĝm [n, k] can be rewritten as shown in
π
(n) 1
(12), where em [n, k] is the estimation error and Z̃s,m [n, k] is w̄f,m = wn SHf,m (w)dw.
equal to Z̃m [n, k] at the pilot symbol and zero, otherwise. The 2π
−π
first term of (12) is the desired CIR, the second term is the
self-distortion owing to the use of a non-ideal interpolator, and cIFm ,0 , cIFm ,1 and cIFm ,2 are the coefficients of the ap-
2
the third term is the interference from the background noise proximated polynomial of |We,m (w1 , w2 )| and depend only
(n)
and inter-cell interference. on the interpolator and pilot spacing. The quantities w̄t,m
(n)
The mean square error (MSE) of the CIR estimate can be and w̄f,m increase as the fading becomes faster and more
represented as frequency selective, respectively, i.e., they generally increase
2
 in proportion to the n-th square of the Doppler spread and
σe,m = E |Ĝm [n, k] − Gm [n, k]|2 (13)
delay spread of the channel [11].
2 π π The interference MSE σi,m2
can be represented as
ηm
= SHs,m (wt , wf ) |We,m (wt , wf )|2 dwt dwf
(2π)2 2
π
π
−π −π σZ
2 2
2  
π π σi,m = 2 |Wm (wt , wf )| dwt dwf (21)
σZ 2 (2π) dt df
+ 2 |Wm (wt , wf )| dwt dwf −π −π
(2π) dt df 2
−π −π = σZ Im .
2 2
= σs,m + σi,m 2
The self-distortion MSE σs,m is a function of the path gain
as well as channel spectra of user m. On the other hand,
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
CHOI and CIOFFI: EFFECTS OF NON-HOMOGENEOUS SINR ESTIMATION ERROR STATISTICS ON SCHEDULING PERFORMANCE 1597


 ∞

Ĝm [n, k] = Gs,m [n + p, k + q] + Z̃s,m [n + p, k + q] wm [p, q]
p=−∞ q=−∞
∞ ∞

 
+ Gm [n, k] − Gs,m [n + p, k + q]wid [p, q]
p=−∞ q=−∞

 ∞
= Gm [n, k] + ηm Hs,m [n + p, k + q] (wm [p, q] − wid [p, q])
p=−∞ q=−∞

 ∞

+ Z̃s,m [n + p, k + q]wm [p, q] = Gm [n, k] − em [n, k] (12)
p=−∞ q=−∞


Wm (wt , wf ) − dt df , |wt | ≤ π/dt and |wf | ≤ π/df
We,m (wt , wf ) = (17)
Wm (wt , wf ), otherwise.

2 ηm2 π π 2 2
σs,m = (2π)2 (dt df )2
SHm (wt , wf ) |We,m (wt , wf )| dwt dwf = ηm χm (18)
−π −π

the interference is independent of the both path gain and shown in (23), where Ĥm is the estimate of Hm which can be
channel spectra, and only depends on the interpolator and obtained through dividing Ĝm in (12) by the path gain. Based
2
σZ . It is because self-distortion occurs owing to imperfect on (12), the perfect Gm can be represented as
interpolation of its own faded signal while the interference
Gm = Ĝm + em . (24)
comes from the filtered output of the background noise and
inter-cell interference. The power of the desired signal for Assuming that Ĝm and em are independent with the use of the
2
user m increases in proportion to the path gain ηm . The self- MMSE interpolator [12], they are independent ZMWCSCGNs
2 2 2 2 2
distortion σs,m accordingly grows while the interference σi,m with the variance of (ηm −σe,m ) and σe,m , respectively. Then,
is unchanged. For a given path gain, self-distortion increases the perfect SINR can be written by
as the channel selectivity grows in the time and frequency 
2 2
domain, whereas there is no variation on the interference. γm = |Gm | σZ (25)
2 
On the other hand, as the power of the noise σZ increases, 2 2
2 2 = γ̂m + |em | σZ + (Gm e∗m + G∗m em ) σZ 2
the interference σi,m increases whereas self-distortion σs,m
is unchanged. Thus, it can be inferred that these two causes where
of imperfect SINR estimation, self-distortion and interference, 
2 2
differently affect the performance of multi-user scheduling E{γm } = E{γ̂m } + E{ |em | σZ } (26)
2 2
2 2
2
according to the path gain, channel spectra and noise power = (ηm − σe,m ) σZ + σe,m σZ
σZ2
. 2
2
= ηm σZ .

IV. E FFECT OF SINR E STIMATION E RROR 2 2
γ̂m and |em | σZ have independent exponential distribution
This paper assumes perfect CIR for coherent detection since 2 2
with the variances of γ̄m (1 − σen,m ) and γ̄m σen,m , respec-
we evaluate the effect of inaccurate SINR estimation on multi- 2 2 2 −1
tively, where σen,m (= σe,m ηm = χm + Im γ̄m ) is the
user scheduling. This assumption is also plausible when the
variance of the normalized SINR estimation error. Ignoring
effect of mis-scheduling from SINR estimation error is a
the last term of (25) which is negligible compared to others
dominant factor for performance degradation [7]. If M users
[7], it can be represented by
are uniformly allocated in the frequency domain, the subcarrier 
2 2
index k in (5) can be omitted without loss of generality. This γm ≈ γ̂m + |em | σZ . (27)
implies that the multi-carrier system with an opportunistic
scheduler can be treated as a simple parallel extension of a Similarly, the instantaneous channel power can be written by
 2 Æ 2 
single-carrier system [8]. Assuming a wide-sense stationary |Hm |2 = Ĥm  + |em |2 ηm + (Ĥm e∗m + Ĥm

em ) ηm
channel, the analysis can also disregard the time index n. Thus,  2 Æ
the sequel will omit the index k and n. ≈ Ĥm  + |em |2 η2m (28)
The expected system throughput can be represented as
 2 
Csch = E[log2 (1 + γmQ )] (22)   2 2
where Ĥm  and |em | ηm have independent exponential
2 2
where mQ = Q(1, 2, ..., M ). For the MS and PF scheduling distribution with the variances of (1 − σen,m ) and σen,m ,
policies, the scheduling function Q can be represented as respectively.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
1598 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 5, MAY 2008



⎨ arg max {γ̂1 , γ̂2 ,...,γ̂M }, MS
m∈{1,2,...,M }
 2    2
Q (1, 2, ..., M ) =
⎪    2   (23)
⎩ arg max {Ĥ1  , Ĥ2  ,..., ĤM  }, PF.
m∈{1,2,...,M }

4.0
A. PF scheduling
Using Jensen’s inequality, the expected system throughput
under PF scheduling can be represented as 3.5

Throughput (bps/Hz)
CP F = E{log2 (1 + γmQ )} (29)
≤ log2 (1 + E{γmQ })
3.0
= CP Fu
Perfect
where E{γmQ } is written by (30) at the top of the next page. Homogeneous ( σ en
2
=0.005)
In (30), p(A) is  the probability that an event A happens, Homogeneous ( σ 2 =0.5)
2.5 en
2 2 Non-homogeneous (Mg=Mp)
and γe,m = |em | σZ . fm (x) and Fm (x) are the probability Non-homogeneous (Mg=3Mp)
density function (pdf) and cumulative density function (cdf) Non-homogeneous (Mg=1/3Mp)
of γ̂m , respectively, which are given by 2.0
2 0 10 20 30 40 50
exp(−x γ̄m (1 − σen,m ))
fm (x) = 2
(31) Number of users (M)
γ̄m (1 − σen,m )
(a) System throughput in respect to M

x
Fm (x) = fm (h)dh. 4.0
0
Note that the upper bound in (29) is relatively tight es-
pecially for a large number of users and easier to calculate 3.5
Throughput (bps/Hz)

[6,7]. Thus, this upper bound is considered in this paper for


simplicity of analysis. Although the difference between the
upper bound and the exact value is noticeable for small M , 3.0
the ordering of the curves is unchanged. When each user has Perfect (M=Mg)
the identical average SINR (γ̄m = γ̄) and homogeneous SINR Homogeneous ( σ 2 =0.005, M=Mg)
2 2 en
estimation accuracy (σen,m = σen ), each user is selected with Homogeneous ( σ en2 =0.5, M=Mg)
2.5
equal probability, which simplifies the system throughput as Non-homogeneous (Mg=Mp, M=2Mg)
the following [7] Non-homogeneous (Mg=3Mp, M=4/3Mg)
  Non-homogeneous (Mg=1/3Mp, M=4Mg)
M
2 2.0
CP Fu = log2 1 + γ̄ 1 + (1 − σen ) 1/i (32) 0 10 20 30 40 50

  2
i=2
 Number of users in good SINR estimation (Mg)
≈ log2 1 + γ̄ 1 + (1 − σen ) (ln(1 + M ) + κ − 1)
(b) System throughput in respect to Mg
where κ(≈0.577)
 M is Euler’s
 constant. It can be seen that the
Fig. 2. System throughput for PF scheduling when γ̄m = γ̄ (self-distortion
 dominant).
MUD gain (= 1/i ) increases by log2 (ln(1 + M )) as the
i=2
number of users increases with the perfect SINR estimation.
2
However, the gain is reduced with the scaling factor of (1-σen )
as SINR estimation becomes more inaccurate since the best The evaluation of the effect of non-homogeneous SINR
user is less likely to be selected at scheduling. estimation on the system throughput considers two special
If the users are scattered in a cell with different values cases: The first case is when the effect of self-distortion is
2 −1
of the path loss, (32) is no longer satisfied even if all the dominant, i.e., σen,m ≈ χm since χm  Im γ̄m . This corre-
2
users have the same interference σi,m . Moreover, the channel sponds to a situation when the effect of channel selectivity is
for next generation wideband mobile systems is likely to be overwhelming, or the interference power is relatively small,
more selective in the time and frequency domains as the e.g., in a single cell system or in a hot spot. The system
carrier frequency and the data bandwidth increase, which throughput (CP Fu ,eq gain ) is plotted in Fig. 2, where there are
result in wide variation of self-distortion for each user. In this two user groups that have good and poor accuracy in SINR
2 2
general case where users experience non-homogeneous SINR estimation (σen,m g
=0.005 and σen,m p
=0.5, where mg and mp
estimation error statistics, it is difficult to obtain the system are user indices for good and poor estimation, respectively)
throughput in a closed form since the user selection probability with the same average SINR (γ̄m = γ̄, i.e., equal path gain).
is different for each user. Thus, we numerically calculate the The users with good SINR estimation correspond to ones in
system throughput in what follows. the less selective channel in the time and frequency domains,
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
CHOI and CIOFFI: EFFECTS OF NON-HOMOGENEOUS SINR ESTIMATION ERROR STATISTICS ON SCHEDULING PERFORMANCE 1599

M 
 ∞
E{γmQ } = γp(mQ = m, γm = γ)dγ (30)
0
m=1
γ
M 
 ∞
= γ  p(mQ = m, γ̂m = x, γm = γ)dx

m=1 γ=0
0γ
M 
 ∞
≈ γ  p(mQ = m, γ̂m = x)p(γe,m = γ − x)dx

m=1 γ=0
0γ
M 
 ∞
γ  p(γ̂m = x)p(γ̂1 ≤ x)p(γe,m = γ − x)dx

γ̄1 γ̄m−1 γ̄m+1 γ̄M
= x, · · · , γ̂m−1 ≤ x, γ̂m+1 ≤ x, · · · , γ̂M ≤
m=1 γ=0 γ̄m γ̄m γ̄m γ̄m
0γ Æ
M 
 ∞
M 2
exp(−(γ − x) (γ̄m σen,m ))
γ  fm (x) dx

γ̄k
= Fk x 2
m=1 γ=0 γ̄m γ̄m σen,m
0 k=1,k=m

4.0
0.7 Homogeneous
Non-homogeneous (Mg=Mp)
0.6 Non-homogeneous (Mg=3Mp)
Non-homogeneous (Mg=1/3Mp)
Throughput (bps/Hz) 3.5
0.5
Prob.

0.4
3.0
0.3

0.2
2.5
Perfect
0.1 Homogeneous (σ en2 =0.5)
- Ml=3Ms Non-homogeneous (Mg=Mp)
0.0
2.0
0 10 20 30 40
0 10 20 30 40 50
Number of users (M)
Number of users (M)

Fig. 3. Probability that users with poor SINR estimation are chosen when
γ̄m = γ̄ (self-distortion dominant). Fig. 4. System throughput for PF scheduling with different user distribution
(self-distortion dominant).

i.e., slow and near flat fading, and vice versa for users with
poor SINR estimation. The system throughput (32) in the that exclusion of poor users worsens the fairness problem.
homogeneous condition is also plotted for comparison when In order to evaluate the performance when users experience
2 different path gains, Fig. 4 depicts the system throughput
σen = 0.005 and 0.5, respectively. The system throughput
in the homogeneous condition is substantially degraded for (CP Fu ,neq gain ). For simplicity in numerical computation, we
2 consider two values of path gain, γ̄m =0 dB for Ml users
poor SINR estimation (σen = 0.5), while the degradation is
negligible when all the users are with good SINR estimation near the cell boundary and 10dB for Ms users near the base
2 station, respectively as seen in Fig. 1. Considering uniformly
(σen = 0.005). For non-homogeneous SINR estimation, the
throughput loss increases as the portion of the number of users distributed users in a cell, we assume Ml = 3Ms . Since there
(Mp ) in poor estimation over that (Mg ) in good estimation is no correlation between channel selectivity and user location
increases. The system throughput is also plotted in Fig. 2 (b) in general, it is assumed that each user experiences good and
with respect to Mg . It can be observed that as Mg increases, poor accuracy in SINR  estimation
2 with equal probability, i.e.,
  2
all the curves for the non-homogeneous condition converge to Mg = Mp . Since E{Ĥm  }(= 1 − σen,m = 1 − χm ) is not
2
that of homogeneous users with σen,m = 0.005 although the a function of the path gain, user selection is independent of
non-homogeneous cases have larger M than the homogeneous the path gain for the self-distortion dominant case, resulting
2
case with σen,m = 0.005. This implies that the MUD gain is in the same distribution of user selection probability as in Fig
reduced to that only
 with
2 users in good
2 estimation. In other 2. We can observe the same tendency as in Fig. 2 for the
    2
}CP F,eq gain .
words, since E{Ĥmg  } > E{Ĥmp  }, only the users with system throughput, i.e., CP F,neq gain = E{ηm
good estimation are selected at scheduling for the most time, Consequently, the tendency of PF scheduling performance is
especially for large M . The probability that users with poor not affected by the path gain for the self-distortion dominant
estimation are chosen is depicted in Fig. 3 in terms of M . The case.
probability is seen to decrease with larger M , which suggests Fig. 2 showed that users with poor SINR estimation con-
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
1600 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 5, MAY 2008

4.0 4.5

3.5 4.0
Throughput (bps/Hz)

Throughput (bps/Hz)
3.0 3.5

Perfect ( σ en2 =0) Perfect (I=0)


2.5 σen2 ,m =0.05
p
3.0 I=0.05
σen2 ,m =0.10
p
I=0.10
σ2en,m =0.25 I=0.25
- σen,m = 0.005, Ml=3Ms , Mg=Mp
2 p

g σen2 , m =0.5 - Ml=3Ms I=0.5


p

2.0 2.5
0 10 20 30 40 0 10 20 30 40
Number of users (M)
Number of users (M)

2 Fig. 7. System throughput for PF scheduling (interference dominant).


Fig. 5. System throughput for PF scheduling with different σen,mp
(self-
distortion dominant).

0.5 between users with different channel selectivity, fairness is


broken between users near the base station and users near
the cell boundary, providing priority to users near the base
0.4 station who can transmit with the higher date rate. As a result,
the throughput increases when there is an SINR estimation
M=8
0.3
error as seen in Fig. 7, and this unexpected increase becomes
Prob.

more evident with larger interference I at the cost of more


severe degradation of fairness. In other words, denoting ms
0.2 and ml by indices for users  near2 the base station and cell
 
boundary, respectively, E{Ĥms  }(= 1 − I/γ̄ms ) is larger
 2
0.1 M=32  
than E{Ĥml  }(= 1 − I/γ̄ml ), especially for large I. As a
σ en2 ,m = 0.005 , Ml=3Ms, Mg=Mp
g result, users near the base station, providing higher throughput
0.0 than users near the cell boundary, are selected more frequently,
1 10 100
gaining the system throughput with some loss of fairness
σ en2 , m σ en2 , m
p g between users with different locations. The probability that
users near the base station are selected is plotted in Fig. 8.
Fig. 6. Probability that users with poor SINR estimation are chosen with The users near the base station are more likely to be chosen,
2
different ratio of σen,m 2
and σen,m (self-distortion dominant).
p g
especially for large M and I, even though they are only a
fourth of the number of total users, i.e., Ml = 3Ms . On
the other hand, the unfairness problem for the self-distortion
tribute very little to the system throughput for PF scheduling. dominant case is dependent upon not the location but the
Fig. 5 plots the system throughput for different σen,m 2
with selectivity of the channel, thus decreasing throughput.
p
2
σen,mg = 0.005 fixed, in order to investigate the sensitivity The throughput tendency is summarized in Table 1 for
2 2 these two special cases. The system environment is usually
of the disparity between σen,m p
and σen,m g
. It can be seen
neither self-distortion only nor interference only in practice.
that the performance becomes much worse as this disparity in-
2 In this practical case, it can be easily expected that the system
creases (i.e., as σen,m p
increases) because of more pronounced
has unfairness problem and the system throughput is changed
unfairness as seen in Fig. 6, in particular for large M .
according to the dominant cause of the unfairness. That is,
The second example considers the case when the interfer-
2 when the self-distortion effect is more influencing than the
ence is dominant on the SINR estimation error, i.e., σen,m ≈
−1 −1 interference (i.e., for large selective fading channel), user
Im γ̄m since χm  Im γ̄m . This corresponds to interference-
selection is prioritized to the users in less selective fading
limited cellular systems in practice. The system throughput
channels that have more accurate SINR estimation irrespective
is depicted in Fig. 7 with respect to Im when the users are
of user location, decreasing the system throughput (i.e., MUD
distributed as in Fig. 5, assuming χm = 0 and Im = I.
gain) by the number of users with poor SINR estimation. On
Unlike the self-distortion dominant case, user selection is
the other hand, when the effect of self-distortion is marginal
dependent upon the path gain, i.e., user location, since the
(i.e., interference dominant), user selection is mainly affected
effect
 of interference I varies according to the path gain as
by the path gain (i.e., user location), resulting in unwanted
 2 −1
E{Ĥmg  } = 1 − I γ̄m . Contrary to the self-distortion domi- increase of the system throughput caused by frequent selection
nant case where the throughput decreases owing to unfairness of users near the base station.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
CHOI and CIOFFI: EFFECTS OF NON-HOMOGENEOUS SINR ESTIMATION ERROR STATISTICS ON SCHEDULING PERFORMANCE 1601

TABLE I
T ENDENCY OF SYSTEM THROUGHPUT IN THE PRESENCE OF SINR ESTIMATION ERROR

Self-distortion dominant Interference dominant


Main causes - Large Doppler spread and/or long delay spread - Small Doppler spread and short delay spread
- Small interference - Large interference
Typical environments - Single-cell or multi-cell systems with frequency reuse Multi-cell systems without frequency reuse
- Near highway or highly scattering outside area
- Fairness broken between low and high channel spread users - Fairness broken between cell center/boundary users
Effects on PF scheduling (users with low channel spread are chosen with higher prob.) (users near cell-center are chosen with higher prob.)
- Decreased throughput due to imperfect user selection - Unwanted throughput increase due to scheduling
priority to cell-center users
Effects on MS scheduling Severe throughput decrease Moderate throughput decrease

1.0 5.0

0.9
M=16
0.8
4.5

Throughput (bps/Hz)
0.7

0.6 M=4
Prob.

0.5 4.0

0.4

Perfect ( σ en =0)
2
0.3
σ en2 ,m =0.05
3.5 p

0.2 σen2 ,m =0.10


p

σen2 ,m =0.25
p
0.1 σen2 ,m =0.5
- σ en,mg = 0.005, Ml=3Ms, Mg=Mp
2
- Ml=3Ms p

0.0 3.0
0.01 0.1 1 0 10 20 30 40
I Number of users (M)

Fig. 8. Probability that users near the base station are chosen (interference Fig. 9. System throughput for MS scheduling (self-distortion dominant).
dominant).

5.0

B. MS scheduling
The upper bound of the system throughput for MS schedul- 4.5
Throughput (bps/Hz)

ing can be represented as

CM Su = log2 (1 + E{γmQ }) (33) 4.0

where E{γmQ } is given by (34) at the top of the next page.


In a similar way to PF scheduling, we evaluate the effect Perfect (I=0)
3.5 I=0.05
of inaccurate SINR estimation on the performance for two I=0.10
cases: self-distortion dominant and interference dominant. I=0.25
Since MS scheduling considers 2 instantaneous SINR γ̂m , not
I=0.5
 - Ml=3Ms, Mg=Mp
  3.0
the instantaneous gain Ĥm  , the SINR estimation error is 0 10 20 30 40
likely to cause a different tendency on the system throughput. Number of users (M)
Fig. 9 and 10 depict the system throughput for self-distortion
dominant and interference dominant cases, respectively. Un- Fig. 10. System throughput for MS scheduling (interference dominant).
like PF scheduling, the system throughput is always degraded
from non-optimal user selection, i.e., selection of a user whose
SINR is not the highest. This degradation is more evident for V. C ONCLUSION
the self-distortion dominant case (i.e., E{γ̂m } = γ̄m (1−χm ))
than the interference dominant case (i.e., E{γ̂m } = γ̄m − Im ) This paper evaluates the effect of non-homogeneous SINR
since the error of SINR estimation is more influencing with estimation error on the performance of PF and MS scheduling
the normalization by the path gain, resulting in more frequent policies in OFDM downlink systems. The error statistics of
erroneous user selection. However, the difference from the SINR estimation is determined by both self-distortion and
optimum error-free performance is decreased compared to PF interference according to channel selectivity, path loss, and
scheduling since users with a larger path gain are likely to be noise power, which result in different error statistics for each
chosen even if they experience high SINR estimation errors user. Numerical results show that these two factors differently
under the MS scheduling. affect the system throughput and fairness, in particular for PF
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.
1602 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 5, MAY 2008

 ∞
E{γmQ } = γp(mQ = m, γm = γ)dγ (34)
0
m=1
⎛ ⎞
M

 ∞
γ
= γ⎝ p(mQ = m, γ̂m = x, γm = γ)dx⎠ dγ
m=1 γ=0
⎛ 0γ ⎞
M

 ∞

≈ γ ⎝ p(mQ = m, γ̂m = x)p(γe,m = γ − x)dx⎠ dγ


m=1 γ=0
⎛ 0γ ⎞

M

= γ ⎝ p(γ̂m = x)p(γ̂1 ≤ x, · · · , γ̂m−1 ≤ x, γ̂m+1 ≤ x, · · · , γ̂M ≤ x)p(γe,m = γ − x)dx⎠ dγ


m=1 γ=0
⎛ 0γ ⎞

M

!
M 2
exp(−(γ − x) (γ̄m σen,m ))
= γ ⎝ fm (x) Fk (x) 2
dx⎠ dγ.
γ=0 γ̄m σen,m
m=1 0 k=1,k=m

scheduling. This phenomenon becomes more prominent as the Ji-Woong Choi (S’00-M’04) received the B.S., M.S.
number of users increases. The result can be easily applied to and Ph.D. degree in electrical engineering from
Seoul National University (SNU), Korea, in 1998,
single-carrier systems where the SINR is estimated through 2000 and 2004, respectively. From Sept. 2004 to
time-domain processing. It is for further study to design a Oct. 2005, he worked as a postdoctoral research
scheme that mitigates the performance degradation adaptively associate in Inter-University Semiconductor Center
(ISRC), SNU, Korea. From Nov. 2005 to July 2007,
according to the dominant cause of the SINR estimation error. he was with Department of Electrical Engineering,
Stanford University, Stanford, CA, as a postdoctoral
visiting scholar. He also served as a Consultant to
R EFERENCES GCT Semiconductor, San Jose, CA, for development
[1] R. van Nee and R. Prasad, OFDM Wireless Multimedia Communications. of handheld digital TV receivers from July 2006 to July 2007. Since Aug.
Artech House, 2000. 2007, Dr. Choi has been with Marvell Semiconductor, Inc., Santa Clara,
[2] A. Czylwik, “Adaptive OFDM for wideband radio channels,” in Proc. CA, as a Senior Engineer in the Signal Processing Department. His research
IEEE Globecom, pp. 713-718, Nov. 1996. areas are wireless transmission systems including OFDM and CDMA systems
[3] Q. Ma and C. Tepedelenlioglu, “Practical multiuser diversity with outdate and signal processing for communication systems. He received Silver Award
channel feedback,” IEEE Trans. Veh. Technol., vol. 54, no. 4, pp. 1334- at Samsung Humantech Paper Contest in Feb. 2005 and Information and
1345, July 2005. Telecommunication National Scholarship from the Ministry of Information
[4] D. Piazza and L. B. Milstein, “Multiuser diversity-mobility tradeoff: and Communication, Korea in Oct. 2005.
modeling and performance analysis of a proportional fair scheduling,”
in Proc. IEEE Globecom, pp. 906-910, Nov. 2002. John M. Cioffi (S’77-M’78-SM’90-F’96) received
[5] P. Schulz-Rittich, A. Senst, T. Bilke, and H. Meyr, “The effect of im- the B.S. degree from the University of Illinois,
perfect SNR knowledge on multiantenna multiuser systems with channel Urbana-Champaign, in 1978, and the Ph.D. degree
aware scheduling,” in Proc. IEEE Globecom, pp. 153-157, Dec. 2003. from Stanford University, Stanford, CA, in 1984,
[6] J. Moon, J.-Y. Ko, and Y.-H. Lee, “A framework design for the next- both in electrical engineering. He was with Bell
generation radio access system,” IEEE J. Select. Areas Commun., vol. Laboratories from 1978 to 1984, and IBM Research
24, no. 3, pp. 554-564, Mar. 2006. from 1984 to 1986. He has been a Professor of Elec-
[7] J.-H. Jeon and Y.-H. Lee, “Prediction of channel information in multi-user trical Engineering with Stanford University since
OFDM systems,” in Proc. IEEE MWSCAS, pp. 468-472, Aug. 2006. 1986. He founded Amati Communications Corpo-
[8] X. Liu, E. K. P. Chong, and N. B. Shroff, “A framework for opportunistic ration in 1991 (purchased by Texas Instruments in
scheduling in wireless networks,” Computer Networks, vol. 41, no. 4, pp. 1997), and was Officer/Director from 1991 to 1997.
451-474, Mar. 2003. He is currently on the Board of Directors of ASSIA (Chairman), Afond,
[9] J. Holtzman, “Asymptotic analysis of proportional fair algorithm,” in Teranetics, and ClariPhy. He is on the advisory boards of Focus Ventures,
Proc. IEEE PIMRC, pp. F33-F37, Sept. 2001. Portview Ventures, Wavion, Quantenna, and Amicus. His specific interests
[10] F. Berggren and R. Jannit, “Asymptotically fair scheduling on fading are in the area of high-performance digital transmission. Dr. Cioffi is a
channels,” in Proc. IEEE VTC, pp. 1934-1938, Sept. 2002. member of the National Academy of Engineering. He was the recipient of
[11] J.-W. Choi and Y.-H. Lee, “Optimum pilot pattern for channel estimation the Marconi Fellow Award (2006), the Hitachi America Professorship in
in OFDM systems,” IEEE Trans. Wireless Commun., vol. 4, no. 5, pp. Electrical Engineering at Stanford (2002), the IEEE Kobayashi Medal (2001),
2083-2088, Sept. 2005. the IEEE Millennium Medal (2000), the IEE J. J. Tomson Medal (2000),
[12] P. Hoeher, S. Kaiser, and P. Robertson, “Two-dimensional pilot-symbol- the 1999 University of Illinois Outstanding Alumnus Award, the 1991 IEEE
aided channel estimation by Wiener filtering,” in Proc. IEEE ASSPC, pp. Communications Magazine Best Paper Award, the 1995 ANSI T1 Outstanding
21-24, Apr. 1997. Achievement Award, the National Science Foundation Presidential Investi-
gator Award (1987-1992), the ISSLS 2004 Outstanding Paper Award, and
IEEE ICC Outstanding Paper awards 2006 and 2007. He has published over
250 papers and holds over 80 patents, of which many are heavily licensed
including key necessary patents for the international standards in ADSL,
VDSL, DSM, and WiMAX.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:08 from IEEE Xplore. Restrictions apply.

Вам также может понравиться