Estimation
of
Wiener
Phase
Noise
by
the
Autocorrelation
of
the
ICI
Weighting
Function
in
OFDM
Systems
Yi-Ching Liao and
Kwang-Cheng Chen
Institute of Communication
Engineering, College of Electrical Engineering
National Taiwan
University, Taipei, Taiwan,
10617, R.O.C.
E-mail: ycliao@
santos.ee.ntu.edu.tw, chenkc@cc.ee.ntu.edu.tw
Abstract- PerformancedegradationduetoWienerphasenoise which causesbothcommonphaseerror(CPE)andinter-carrier interference (ICI) is a crucial challenge to the implementation of OFDM systems. In this paper, we theoretically employ the Lorentzian model to investigate the autocorrelation function of the ICI weighting function which is the discrete Fourier trans-form oftheexponentialphase noise process. This autocorrelation function canbe shown tobe the kernel of the covariance of the ICI. Based on this kernel, a pilot-aided decision-directed CPE estimatorisproposedaccordingtomaximum-likelihoodcriterion. Different from conventionalmaximum likelihoodapproachwhich ideally assumes the ICI observed on different subcarriers to be independentidentically distributed, we systematically derive the covariances among carriers and practically utilize them to enhance the estimation. Finally, three conventional CPE estimatorsarecomparedwith theproposedschemebycomputer simulation, the numerical results illustrate the effectiveness of the proposed algorithm.
Index Terms-OFDM, Phase Noise and Maximum Likelihood Estimation.
I. INTRODUCTION
Orthogonal frequency division multiplexing (OFDM) has been attracting considerable research interests as apromising candidate tohigh dataratecommunications since its resistance
toimpairments such as frequency selective fadingand
impul-sive noise. However, OFDM is tremendously more sensitive tocarrier frequency offset and phase noise than single carrier systems [1]. Mainly resulting from the instability of local
oscillator, phase noise can be classified into two categories. When the system is onlyfrequency-locked, the resulting phase noise is modeled as a zero-mean Wiener process. When the system is phase-locked, the resulting phase noise is modeled as a zero-mean stationary random process [2]. Here, we concentrate onthesuppression of Wiener phase noise.
Methodsfor thecompensation of the effects of phase noise
have been proposed by several authors. The conventional
ap-proachescanbecategorizedinto decision directed approaches and pilot-aided approaches. The decision directed approaches
[3], [4], estimate the CPE using the averaged phase rotation of the observed symbols from the ideal constellation points. Asymbol by symbol pre-compensation is necessarytoensure this rotationnotexceedingthedecisionboundary. Asfor pilot-aidedapproach[5]-[7], theaverageorweighted average of the
phase differences between the transmitted and received pilot
symbols are used toestimate the CPE.
In this paper, by using the Lorentzian model, the
autocor-relation function of the ICI weighting function which can
be shown to be the kernel of the second order statistics of
the ICI is systematicallyderived. Different fromconventional
maximum likelihood (ML) schemes [4], [7] which ideally
assume the ICI observed on different subcarriers to be
inde-pendentidentically
distributed,
wepractically
investigate theircovariances bytheautocorrelation function of theICI
weight-ing function and combine pilot-aided and decision-directed approaches toyieldageneralized maximum likelihood estima-tion scheme. Simulaestima-tion results demonstrate that the proposed algorithm outperforms the conventional approaches
The rest of this paper is organized as follows: Section II presents the phase noise corrupted OFDM signal model and thephase noise model respectively. Based on that, Section III
first investigate the statistical characteristics of the sufficient
statistics, then, it proceed to
derive
the maximum likelihood estimator for the CPE. Performance evaluation via computersimulation is addressed in Section IV. Finally, Section V discusses and concludes thispaper.
II. BACKGROUND
A. Phase Noise CorruptedOFDMSignalModel
ConsideringageneralOFDM systemusing N-point inverse
fast Fourier transform
(IFFT)
for modulation. Assume thefrequency domain subcarrier index set is composed of three
mutually exclusive subsets defined by
D-{d1,d2,
,dNd}
P
{pl1,P2
..*PNP
}(1)
X-={N,
n2, ---',nN,}
I
where D denotes the set of indices for Nd data-conveying
subcarriers,Pis thesetof indices for
Np
pilot subcarriers and X stands forNn virtual subcarriers. Then, the set of indices forNu useful subcarriers U can be defined asA-{Ul,U2,* * * UN,,,,
I
=DUP,(2)
whereNu = (Nd+
Np).
LetXm(k)bethe modulatedsymbol on the kth subcarrierof themth OFDM symbol. For k EU,Xm(k)
is taken from someconstellation with zero mean and average power4
2E{IXm(k)12}.
The output of the IFFThas aduration of T seconds which is equivalent to N samples. A
Ng-sample
cyclic prefix longer than the channel impulse response is preceded to eliminate theinter-symbol interference (ISI).At the receiver, timing and frequency recovery is assumed tobe accomplished.Considering the multiplicative phase noise
and the additive white noise, the received nth sample of the
mth OFDM symbol can bewritten as
rm(n)
=[xm(nl)
0hm(n)]
ej(km(n)+O)
+Tm(ln)
(3)in which 0 is the circular convolution and
q.(n)
= q(m(N
+Ng)
+Ng
+n)
(4)
where
xrm(n), hm(n)
andq,
(n)
represent the transmitted signal, the channel impulse response and the phase noise respectively, while(m(n)
denotes the AWGN noise and 0 isthe initialphase of the phase noiseprocess. After removing the cyclic
prefix
and performing the FFT, thefrequency
domain symbolcan beexpressed byRm(k)
=4)m(0)Hm(k)Xm(k)+ E
4'm(k
-I)Hm(l)Xm(l)
+Zm(k)
IEUi$k
I,, (k)
where
Hm(k)
is thechannelfrequency
response andZm(k)
denotes the frequencydomain
expression
of(m(n). 4.m(h)
isthe discrete Fourier transform of the
phase
noiseprocessgiven
by
N-1
'Jm(h)
- I Eej(em(n)+G)e-j2
(6)n=O
And itcan be viewed as a
weighting
function on the trans-mittedfrequency domainsymbols.
Inparticular,
whenh =0,
4.m(0)
isthe time average of thephase
noise process withinone OFDM
symbol
duration. This term isusually
known asthe commonphase error
(CPE)
whichcauses the samephase
rotation andamplitude distortiontoeachtransmittedfrequency
domainsymbol. On the other
hand,
when h=0,
the secondtermin
(5)
is theinter-carrier interference(ICI) resulting
fromthe contributions of other subcarriers
by
theweighting
of4.m(h)
due to the loss oforthogonality.
In the rest of thispaper, weshall call
4m(h)
the ICIweighting
function. Based on(5),
thereceivedfrequency
domainvector can be given by r. =,4m(0)HmXm
+im
+Cm
=4bm(O)HmXm
+ cmI where cm = tm +(ml
Hm=diag(Hm(0),Hm(l),m.
,H(N -1)) Xm =[Xm(0)
Xm(1)
...Xm(N- 1)]T
,bm-
[Im(O)
In(l)
..Im(N
-1)]
(m -
[(mn(°)
(m(l)
(mn(N-
1)]
Xand
diag(.)
is adiagonal matrix. From now on, we shall adda second subscript to one of the vector or matrix variables
definedin(7) and (8) to indicate its sub-vector or sub-matrix which is taken according to one of the subcarrier index sets
in (1) and (2). The second subscript may be chosen from
{p, d,n,u} which relates to {P, D,
K,
U} respectively. For example,rm,p
=[Rm(pi)
Rm(P2)
...Rm(pNP)]T
(9)
stands for thereceived pilot vector.
Conventionally,
rm,p is utilized to obtain the channelre-sponse and common phase error to carry out equalization on
ri,d,
then the equalized results are sent to the detection block to get the decisions. Sinceaccurate channel estimation in OFDM systems can be obtained by either preambles or pilot symbols[81,
in the following sections we assume that the channel frequency response is acquired perfectly at the receiver.B.
Phase
NoiseModel
Accuratemodeling of oscillator phase noise isakey factor to the analysis and simulation of the distortion caused by phase noise. For a classic model of phase noise,
On(n)
can be modeled as adiscrete-time Wiener process[1]
withE[qm(n)]
= 0E[(qm(n
+An)
-Om(n))2]
=4ir/3TIAnl/N,
(10)where
,3
(Hz) denotes the one-sided 3 dB linewidth of theLorentzian power
density
spectrum ofthefree-running
oscil-lator. The Lorentzian spectrum is the squared
magnitude
of afirst order lowpass filtertransfer function
[8].
Thesingle-sided
spectrum So
(f)
isgiven by
2/7r/3
S,I(f)
=
1+f2/z2'
and the Lorentzian spectra with different linewidth are shown in
Figure 1.
-(7)
a(8)
10° 10 104 f(Hz)(1
1) one-sided 3 dB 106As for the initial phase 0, it can be modeled as a random variableuniformly distributed in [0, 27r)and isindependentof
bm(ln).
Lateranalysis
and simulation will be based on this model.111. PILOT-AIDED DECISION-DIRECTED CPE ESTIMATION Most OFDM systems employ pilots to facilitate receiver
synchronization since data-aided estimation gives better and steadier estimate.However,sincepilotscostsystemutilization,
the number of pilots should bekeptaslow aspossible which
confines the performance ofpilot-aided CPE estimation algo-rithms. Comparatively, decision-directed approaches
enjoy
a largerobservation space.Nevertheless, ensuringthe acceptablecorrectness of the decisions is also critical.
Therefore,
usepilotsto acquiretheCPEandprovideaninitial compensation,
thenenlarge the observation space by
including
the tentativedecisions as the sufficient statistics to perform the final esti-mate canbenefitfromtheadvantages ofbothapproaches.We
refer thismethod to the
pilot-aided-decision-directed
(PADD)approach. In the following, we shall
investigate
the statisticalcharacteristics of the sufficient statistics
rmn,
andsystemati-cally derive the MLestimator in accordance with the PADD approach.
A. Statistical Characteristics of the
Sufficient
Statistics Considering the PADD approach, since the data and thechannel frequency response are acquired, the statistical
char-acteristics of rm,u depend on that of Lm,u and 'm,u The
AWGN noise on each subcarrier can be modeled as a zero mean complex Gaussian random variable with variance az.
Since the ICI on each subcarrier is composed of the data
symbols on other subcarriers, we can apply the central limit theorem to model it as a complexGaussian random variable.
In the following, we first show that the ICI has a zero mean,
then,thesecondorderstatisticsof the ICI will beinvestigated. 1) The Mean of the ICI: Since
Xm(k)
andHm(k)
are assumed tobe known in thePADDcircumstance, the mean of theICI depends on that of the ICI weighting function4m(h) which can beexpressed asN-i
E[4Fm(h)]
= + E[ej(O-(n)+O)J
eJ2IT. (12)n=O
Since 0 is independentof
qm(n),
the expectationin (12) can be decomposed into the product ofE[ejO-(n)]
andE[ej0].
Let
Io(w)
denotes thecharacteristic function of the uniformrandom variable, we have
E[ej0]
=To(w)
l2
27wwjr
=2
sin(
7)ei
(13)
=0.
Therefore, the mean ofthe ICIbecomes zero.
2)
The Second Order Statisticsof
the ICI: In the PADDscenario,
the sufficient statistics are thereceived useful sym-bolsgiven
by
rm,u =
4Pm(0)Hm,uXm,u
+tm,u+(m,u- (14) We maydenote thecovariance matrix of tm,uby
C,nl
u,thenthecovariance matrix of
rm,u
can beexpressed
asCrr u =CtLnu + hzI (15)
Letthe elementof
C,L
bedenotedby cir(k1, k2),
itcanbeexpressed
aso1(kl,k2) = 5 S
Hm(1I)H(192)Xm(11)Xm(l2)
I1EU 12Eull=A-kl 12:A-k2
E
[4Dm(Ukj
-l)
>mI (Uk2 -12)].
(16)
Wecanobservethatoj(k1,
k2) depends
ontheautocorrelation function ofthe ICIweighting
function4cm(h). Therefore,
we have thefollowing
proposition.
Proposition
1(The
autocorrelationfuinction of
J>m(h)):
For a Wienerphase
noise with the Lorentzian spectrum,given
the one-sided 3 dB linewidth,B,
the autocorrelation function ofthe ICIweighting
function41m(h)
can bedefinedby
RD(h1,h2)
E[4Pm(hi)4P*
(h2)]
andR+(h1,
h2) =6(h,-h2)N
2N
+-Nez
- Nu N2 2-ezi -e-z11-N
+Nez2
_eNu1 + -Iez2 (-6(h,-h2)N)
Nu ( N21-e(z1-Z2)
1ezi 1- + 1e22 1eZ2J
(17)
where
Z1-u+
jVl
--[/3T
+jhl]
Z2--U + 2=--
[/3T
+jh2]
(18)
and
(-)N
denotes the moduloby N, z-represents thecomplexconjugate
ofz.Proof:
cf.Appendix.
UAs for
3,
since itcanbeobtained by preamble signal and has been proposed in the literature[91,
we may assume that az is knownby
the receiverhenceforth.B. MaximumLikelihood CPE Estimator
From the above
discussion,
the sufficient statistics 7mucan be modeled as a complex Gaussian random vector with meanvector l.m(O)Hm,uXm,u andcovariance matrixCr,. Hence, the log-likelihood function shall be
(4Pm
()) 2R{XmHHHC
rmn,u(D(0)}
-xmumuC H
Tofindthe ML estimation of m(0),wemay first let4Dm(0) =
Ame4O-
andlet thederivatives of (19) w.r.t. Am and4,m
equal to zero to getXr
m_
X$,,uHH{,uC-uHmn.uxm,u
HumurrnHmuz,
Then, the ML estimation of the CPE will bexH HH C-1 rm,
41m(O)
= m,u m,urmc
m,u
H% Hm umH",
C21
rm,vHm,uXm,u
T° a a A a (20) Io,,
E10(21)
10IV. PERFORMANCE SIMULATION
Theproposed maximum-likelihoodCPE estimator are
eval-uated in frequency selective slowly fading channels with 50 nsand75nsrmsdelay spread [10].Channelimpulse response
remains static within a frame containing 16 symbols, but varies independently from frame to frame. Transmitted data is constructed according to the IEEE 802.1 la standard [11]. 16QAM and64 QAM which are more sensitive to phase noise than M-PSK, are used in the simulation. The phase noise is
simulated using theLorentzianmodelwith 3equals to 1 kHz and 2 kHz. Pilot-aided approaches based on averaged-phase [5], weighted-averaged-phase
[6]
andleast-square criterion[7]are also simulated as a comparison. Each simulation point is conducted by 3-105 OFDM symbols. The probability of
symbol error (SER) with 16 QAM and 64 QAMin different
channels are shown in Figure 2 andFigure 3 respectively.
It is easy to observe that Wiener phase noise causes an
irreducible errorfloorto OFDMreceiverperformance, which
isunacceptable in practice. Abouttheeffect ofdifferent chan-nels, comparing Fig. 2(a)with Fig. 2(b),we canfind that the shorterrmsdelayspreadgivesthebetterperformancewhich is
also evident in
Fig
3. Observing Fig. 2 which corresponds to 16QAM modulation, in general, theproposed
ML estimator and the LS estimatorgreatly outperform
theaveraged-phase
and
weighted-averaged-phase
approach.
Theperformance
gapsbetween the
proposed
ML estimator and the LS estimator atmoderate SNRare1 dBunder1 kHzlinewidth and2 dBunder
2 kHz linewidth
respectively.
In Fig. 3 which corresponds to 64
QAM
modulation, we can observe that theperformance
of the three conventional approaches get closer and theperformance
gap between thetwoaveragebased
approaches
and theproposed
MLestimatorbecome smaller
compared
to 16QAM.
And theperformance
gapsbetween the
proposed
MLestimatorand theLSestimatorat moderate SNR are 2.5 dB under 1 kHz linewidth and 2 dB under 2 kHz linewidth respectively. This phenomenon is mainly caused by the shrink of the decision
boundary
whenemploying highordermodulation. V. CONCLUSIONS
In this paper, a CPE estimator to effectively remove the complexgain caused by Wienerphasenoiseonthe frequency
domain transmitted symbols is proposed. We systematically
derive the autocorrelation function of the ICIweighting func-tionbasedontheLorentzian model. The secondorder statistics
\sus A,... ~ ~5, ~ X
I.A,
a . Ew s \s,-o 1015 20 25. 10 is 20 25 SNR(Eb/No,dB)(a)rmsdelayspread=50 ns 10° 2 u a a. 10 0-II .I
,1
I1 5., \,, 1012 10 is 20 SNR(Eb/No,dB) Nocorrection --2k Hz, avg. 3=2kHz,w-avg. 3=2kHz,LS - -2kHz,ML 3=1kHz, aVg. - I=1kHz,w-avg. + f3=lkHz,LS * 3=1kHz,ML * Ideal I -I,.2 30 3 -a-Nocorrection 032kHz, avg. 1-32kHz,w-avg. ..=2kHz.LS -12kHz, ML [3=lkHz.avg. a 1=1kHz,w-avg. 33=1kHz, LS I =lkHz, ML Ideal (b) rmsdelay spread=75 nsFig. 2. SER Performance of the CPE correction schemes with 16 QAM of the ICI can be obtained by this autocorrelation function. Different from theconventional ML
approaches
whichdo notconsiderthestatistic of Wiener
phase noise,
weinvestigate
thecovariances between theICIobservedondifferent
subcarriers
to yield a generalized maximum likelihood estimator. The
effectiveness oftheproposed
algorithm
ismanifestedby
simu-lations and is shown to
outperform
the conventional schemes.APPENDIX PROOF OF PROPOSITION 1
By definition, considering (6), the autocorrelation function of
4)(h)
canbeexpressed asN-1 N E
[eJ(0.(ni)-O.(n2))]
h2n2-N2
ns=Ofl30
(22) By(10),
q.m(ni)
-4)m(n2)
canbe treatedasaGaussianran-dom variable withzeromeanand variance
4r3TInj
-n2l
/N.
Therefore,
E[e
j(e
(nh)-O.¢(n2))1
=e-27r3T
N (23)6 aG
,5
D -9 DQ 10 10l E '.: 'o 10 5 10 15 20 25 30 35 SNR(Eb/No,dB)
(a) rmsdelayspread= 50ns
V.,5 ^--b-0- ^Noc c {correction 332kHz, avg. ..-2kHz, w-avg. .=2kHz, LS - =2kHz,ML .=1kHz, avg. "':4--, * 33l~~~~~~~~~-kHz, w-avg. P-=1kHz,LS \, 4--33=lkHz, ML Ideal 5 1 20 25 30 35 _* 5 10 15 20 25 30 35 SNR(Eb/N,.dB) (b)rmsdelay spread=75ns
Fig. 3. SER Performance of the CPE correctionschemes with 64QAM
Substitute (23) into (22), we have
N-1 N-1 2 3T 'N- 272 h2n2-hl
R4)(hl,h2)= -27r
N2i=
e (24)nl=On2=o
Let
N2RD(hl,
h2) bedenoted by SI, then divide the double summation intotwo parts as followsN-1 N-1nl-i
SI = ei N + 5e-2r3T
n=O nl=In2=0 (25)
(e N +2e& Ne),
It is easy to show that the first summation in (25) is
N6(hl-h2)N-
Let the double summation in (25) be denoted by S2 and use changeofvariableby letting t =ni - n2, wehave
N-iN-1-t
S2 N e-j2 +ej2 e N
t=i n2=O
(26)
Carry
outthe
summationof
n2, itbecomes-N-1
S2
=6(hl-h2)N
E
(N
-t)
(e-j2ir4
+ei2-
4)
t=1 (1-6(hl-h2)
-ej2i
e) i +ej27rha.
h; L e322rT ±ej2~~e27r4
) 7 -27r)3Tt e-N-(27)
In (27),we maydenote the first andthe second summation by
Q,
and Q2 respectively. Since They are composed ofgeometric series, by the definition in (18), after some
ma-t
nipulation,L5
1-N+Nez -eNu I- N+Nez2- Nu
Qi- 2-ez1 -e-z 2-ez2 - e-z2 (28)
and
Q2= (1-Nu) ( 1 -=fezi + 1-eZ2 1 (29)
Finally,
Rp
(hi, h2) can beiteratively evaluatedandgiven byS1
N`(hl-h2)N
+S2 Ro(=N2
= N26(hs-h2)N(N
+Qi) (1-6(ho-h2)N)Q2
N2 N2(1ez -Z2) (30) Q.E.D. REFERENCES[1] T. Pollet, M. Van Bladel, and M. Moeneclaey, "BER sensitivity of
OFDM systems to carrier frequency offset and wiener phase noise;'
IEEE Trans. Commun., vol.43, pp. 191-193, Feb./March/April 1995.
[2] L. PiazzoandP. Mandarini, "Analysisofphase noise effects inofdm
modems," IEEE Trans.onComm.,vol.50,no. 10,pp. 1696-1705,Oct
2002.
[3] V.AbhayawardhanaandI.Wassell,"Commonphaseerrorcorrectionfor
OFDM inwirelesscommunication,"inProc. IEEE Global Telecommun.
Conf(Globecom'02),vol. 1, Taipei,ROC,Nov. 2002,pp. 17-21. [4] K. Nikitopoulos and A. Polydoros, "Compensation schemesforphase
noise and residual frequencyoffset in OFDMsystems,"inProc. IEEE
Global Telecommun. Conf (Globecom'01), vol. 1, San Antonio, TX,
Nov.2001,pp.25-29.
[5] T.Onizawa,M.Mizoguchi,T.Sakata, and M.Morikura, "Anewsimple adaptive phase tracking schemeemploying phase noise estimation for OFDMsignals,"inProc. IEEEVTC'02Spring,vol.3, Birmingham, AL, May 2002,pp. 1252-1256.
[6] P. Robertson and S. Kaiser, "Analysis of the effects ofphase-noisein orthogonal frequency division multiplex (OFDM) systems," in Proc. IEEE Int. Conf Commun. (ICC'95), vol. 3, Seattle, WA, June 1995,
pp. 1652-1657.
[7] S. Wu and Y. Bar-Ness, "A phase noise suppression algorithm for OFDM-basedWLANs,"IEEECommun.Lett.,vol.6,pp.535-537,Dec. 2002.
[8] R. VanNeeand R. Prasad, OFDMforMultimedia Wireless Communi-cations. Boston. MA:ArtechHouse,2000.
[9] S. He and M.Torkelson, "EffectiveSNRestimation in OFDMsystem simulation," inProc. IEEE GlobalTelecommun. Conf (Globecom'98), vol.2,Nov. 1998, pp.945-950.
[10] B. O'Hara and A.Petrick, The IEEE 802.11 handbook:A designer's companion. Standard InformationNetworkIEEEPress, 1999. [11] Wireless LANMedium Access Control(MAC)and Physical Layer(PHY)
Specifications: High-Speed Phvsical Laverin 5GHzBand, IEEE Std.
802.1la,Sep 1999.
10 e---s--- 4--o---4---- 2 -iaEa-- No correction 1 =2kHz, avg. . -- ,B=2k Hz,w-avg -=2k Hz, LS -33=2kHz, ML X.=1kHz, avg. 2\ ^>\>,<=* 3=1kHz,w-avg. a-4-.,=1kHz, LS ; 3:=lkHz, ML Ideal X4 *X*,* 2 I -1 i