PerformancecomparisonofthreedifferentestimatorsfortheNakagamim
parameterusingMonteCarlosimulation
AliAbdiandMostafaKaveh
ABSTRACT
Nakagamidistributionhasprovenusefulformodelingmultipathfadedenvelopeinwirelesschannels.Theshape
parameteroftheNakagamidistribution,knownasthemparameter,canbeestimatedindifferentways.Inthis
contribution,theperformanceoftheinversenormalizedvariance,Tolparev-Polyakov,andtheLorenzestimators
havebeencompared throughMonteCarlo simulation,andit hasbeenobservedthat theinversenormalized
varianceestimatorissuperiortotheothersoverabroadrangeofmvalues.
I.INTRODUCTION
The Nakagami distribution is a good candidate for modeling the fluctuations of the received signal
envelope )(tR ,propagatedthroughmultipathfadingwirelesschannels[1].TheNakagamiprobabilitydensity
function(PDF)hasthefollowingform:
0,
2
1
,0,exp
)(
2
)( 212 ≥Ω≥≥
Ω
−
ΩΓ
=−mr
mr
mrm
rf m
mm
R, (1)
where (.)
Γ
isthegammafunctionandmand
Ω
aretheshapeandscaleparametersgivenby[2,p.4]:
(
)
][,
][ ][ 2
2
2
2RE
RVar
RE
m=Ω= . (2)
Intheaboveformulas,E[.]andVar[.]denoteexpectationandvariance,respectively.Notethatinasense,mis
theinverseofthenormalizedvarianceof 2
R
[2,p.4].
Asisgenerallywellaccepted,ausefulandreasonableestimatorfor
Ω
is =
−
=Ω N
ii
RN 1
2
1
ˆ,whereNisthe
number of available samples i
R of the envelope. On the other hand, several different methods have been
proposedintheliteratureforestimatingm.Inthispaperwestudytheperformanceofthoseestimatorsusing
MonteCarlosimulationandshowthatonlyoneofthemhasthebestperformanceoverabroadrangeofvalues
ofm.
II.THREEESTIMATORSFORTHENAKAGAMImPARAMETER
Letusrepresentthekthsamplemomentby k
µ,i.e. =
−
=µ N
i
k
ik RN 1
1.Withthisnotation 2
ˆµ=Ω .Forthe
kthordermomentoftheNakagamirandomvariableRwehave[2,p.10]:
2
2
)(
)2(
][ k
k
kmm
km
RE Ω
Γ
+
Γ
=. (3)
So,ageneralmoment-basedestimator[3,pp.272-274], k
m
~
,formmaybewrittenas:
2
2
2
~
)
~
(
)2
~
(
k
k
k
kk
k
mm
km µ
µ
=
Γ
+
Γ
. (4)
Foroddk,wemustsolveatranscendentalequation(involvinggammafunction)toobtain k
m
~
,whileforevenk
wehaveanalgebricequationwhichisgenerallypreferred.Theoretically, k
m
~
mustbeclosetothetruevalueof
mforanyk.Butsinceintheprocessofmodelingempiricaldata(whichinevitablyhavefiniterange)usingan
infiniterangePDF,higherordersamplemomentsdeviatefromthetheoreticalmomentssignificantly[4],i.e. k
µ
differsfrom ][ k
RE drasticallyforlargekwhenRisarandomvariablewithinfiniterange,itisbettertousethe
lowestpossibleevenordersamplemoment.For 2
=
k,formula(4)reducestotheidentity
1
1
=
,whilefor 4
=
k
weobtaintheinversenormalizedvariance(INV)estimator, INV
m
ˆ:
2
24
2
2
ˆµ−µ µ
=
INV
m. (5)
Thenamecomesfromthefact thatbyreplacingthemomentsinthedefinitionofmin (2)withthesample
moments,wearriveat INV
m
ˆ.
TheTolparev-Polyakov(TP)estimator, TP
m
ˆ,hasbeenproposedin[5]:
)ln(4
)ln()3/4(11
ˆ
2
2
B
B
mTP µµ++
=, (6)
whereln(.)isthenaturallogarithmand
(
)
N
N
ii
RB 1
1
2
∏=
=.
ThethirdestimatoristheLorenz(L)estimator, L
m
ˆ[6]:
29.12
12
2
12 ])([ 4.17
)(
4.4
ˆdBdB
dBdB
L
mµ−µ
+
µ−µ
=, (7)
where dB
k
µisthekthsampledBmoment,i.e. =
−
=µ N
i
k
i
dB
kRN 1
1)log20( ,and log(.) isthelogarithmtothebase
10(noteaminortypoin[1, p. 154,eqn.(5.81)]whereformula(7)abovehasbeenreportedwith1.29asa
coefficientandnotintheexponent).
III.COMPARINGTHEPERFORMANCEOFESTIMATORSVIAMONTECARLOSIMULATION
InordertogenerateNakagamidistributedsamples,it is usefultonotethatifR isaNakagamirandom
variable,then 2
R
P
=
isagammarandomvariable[7,p.49]:
0,exp
)(
)( 1≥
Ω
−
Γ
Ω
=−p
mp
m
pm
pf m
m
P. (8)
Forgeneratinggamma distributedsamples,we haveusedthe StatisticsToolboxof Matlab.Withoutloss of
generality,throughoutthesimulationswehavegenerateddatawith
1
=
Ω
.Foranyfixedmfromtheset{0.5,1,
1.5,…,19,19.5,20},broadenoughtocoverthepracticalrangeofmfordifferentpropagationenvironments,
andanyfixedNfromtheset{100,1000,10000},500sequencesoflengthNweregenerated.Letm
ˆdenoteany
ofthethreepreviouslydefinedestimators.Thesamplemeanof m
ˆ, =
−500
1
1ˆ
500 jj
m,isplottedinFigs.1-3versus
m,togetherwiththesampleconfidenceregion,definedby
×
±
2
(samplestandarddeviationof m
ˆ),wherethe
samplestandarddeviationof m
ˆwascalculatedaccordingto 2
500
1
1
500
1
2
1)
ˆ
500(
ˆ
500 =
−
=
−−jj
jjmm .Thesample
confidenceregiondefinedhereisusefulforexaminingthevariationsofdifferentestimatorsintermsofmandN.
AccordingtoFig.1,forsmallN( 100
=
N), INV
m
ˆisunbiased,while L
m
ˆhasapositivebiaswhichincreases
asmincreases.Thesampleconfidenceregionsofboth INV
m
ˆand L
m
ˆaresimilar,whilethesampleconfidence
regionof TP
m
ˆistoobroadtomakethisestimatorusefulandreliableforsmallN.FormoderateN( 1000
=
N)in
Fig. 2, L
m
ˆ still has the same amount of increasing bias. The sample confidence regions of all the three
estimatorshavetightened.However,thesampleconfidenceregionfor TP
m
ˆismuchwiderthanthoseof INV
m
ˆ
and L
m
ˆ.ThesamebehaviorcanbeobservedinFig.3forlargeN( 10000
=
N).
IV.CONCLUSION
Accordingtotheabovediscussion,itisobviousthatforestimatingtheNakagamimparameter,theinverse
normalizedvarianceestimatorissuperiortotheothertwoestimators,becausecontrarytotheLorenzestimatorit
doesnotshowapositiveincreasingbiasversusm,andincontrastwiththeTolparev-Polyakovestimator,hasa
narrowsampleconfidenceregion.
REFERENCES
[1] J.D.Parsons,TheMobileRadioPropagationChannel.NewYork:Wiley,1992.
[2] M. Nakagami, “The m-distribution: A general formula of intensity distribution of rapid fading,” in
StatisticalMethodsinRadioWavePropagation.W.C.Hoffman,Ed.,NewYork:Pergamon,1960,pp.3-
36.
[3] B.W.Lindgren,StatisticalTheory,4thed.,NewYork:Chapman&Hall,1993.
[4] V.AnastassopoulosandG.A.Lampropoulos,“RadarcluttermodellingusingfinitePDFtail,”Electron.
Lett.,vol.32,pp.256-258,1996.
[5] R.G.TolparevandV.A.Polyakov,“EstimationoftheNakagamiprobabilitydensityparametersina
detectoremployingfalse-alarmprobabilitystabilization,”Telecommun.andRadioEng.,vol.43,pp.113-
115,1988.
[6] R. W. Lorenz, “Theoretical distribution functions of multipath propagationand their parameters for
mobileradiocommunicationinquasi-smoothterrain,”inNATOAGARDConf.Proc.TerrainProfilesand
ContoursinElectromagneticWavePropagation,no.269,Spatind,Norway,1979,pp.17.1-17.16.
[7] G.L.Stuber,PrinciplesofMobileCommunication.Boston,MA:Kluwer,1996.