Anda di halaman 1dari 9

www.seipub.

org/ijc

International Journal of Communications (IJC) Volume 2 Issue 2, June 2013

SamplingBasedEstimationofDQPSK
TransmissionBitErrorRateswithNakagami
mFadingChannels
GrahamV.Weinberg*1
DefenceScienceandTechnologyOrganisation(DSTO),ElectronicWarfareandRadarDivision,
P.O.Box1500,Edinburgh,SouthAustralia,5111,Australia
*1

Graham.Weinberg@dsto.defence.gov.au

estimators,usingaPoissonsamplingdistribution,can
beconstructedtoestimateit(seeWeinberg2006fora
series of such estimators). Additionally, it has also
been demonstrated that the work of (Weinberg 2006)
canbeappliedtoBERestimation.In(Weinberg2008),
aMonteCarloestimatorisproducedforasimplecase
of differential quadrature phasedshift keying
(DQPSK). It is shown that this can be extended to
provideaMonteCarlosolutiontoamoregeneralBER
estimationproblem.

Abstract
AgeneralMonteCarloestimatorforcalculationofbiterror
rates for a class of noncoherent and differentially coherent
communication systems is introduced. This estimator, by
exploiting the connection between the Generalised Marcum
QFunctionandPoissondistributions,isappliedtothecase
of differential quadrature phasedshift keying over a
frequencyselective Nakagamim channel with an arbitrary
power delay profile. It is confirmed that this estimator is
efficient,andprovidesgoodresultsinpractice.

MonteCarlomethodsutilisethestatisticalStrongLaw
of Large Numbers (SLLN), which permits the
estimation of expectations through simulation of
randomnumbers(RobertandCasella2004).TheSLLN
guaranteesthataMonteCarloestimatorwillconverge,
buttheissueiswhetheranaccurateapproximationis
obtained within a reasonable sample size. This is the
tradeoffinusingaverysimpleschemetoapproximate
anintractableresult.

Keywords
Communication System Performance; Bit Error Rates; Fading
Channels;Estimation;MonteCarloMethods

Bit Error Rates and Marcums Q-Function


Bit error rates (BERs) are an important performance
measure for communication systems, and
consequently, their estimation is of considerable
interest(Proakis2001,SimonandAlouini2000a).

TheBERunderconsiderationisgivenby

This paper examines the Monte Carlo estimation of


BERs for a class of noncoherent and differentially
coherent communications systems, operated over
frequency selective channels and Lfold diversity
reception (Simon and Alouini 2000b). These BERs
depend on the wellknown Marcum QFunction
(Marcum 1960), which provides an estimation
challenge.Ithasprovenusefultoapplyboundsonthe
Marcum QFunction to approximate these BERs,
enabling the determination of minimum and
maximum performance levels (Simon and Alouini
2000b, Chiani 1999, Ferrari and Corazza 2004).
Recently, it has been shown that the Marcum Q
Function can be related to discrete Poisson
distributions (Weinberg 2006). This connection
facilitates the samplingbased estimation of the
Marcum QFunction. Consequently, Monte Carlo

40

whereconstantsaandbaregivenby

and Qn is the generalized Marcum QFunction of


ordern,givenby

In(x)isthemodifiedBesselfunctionofordern,and

InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc

estimator, in the scenario of Section 3 through


simulations.

isthetotalinstantaneoussignaltonoiseratio(TISNR)
per bit of all L paths (see Proakis 2001). Under the
Nakagamimchannel model, with L =1, the TISNR is
modelled as a Gamma random variable with shape
parametermandmeantheaveragereceivedsignalto
noise ratio. Similarly, the case where L=1, and the
TISNRisadeterministicconstant,hasbeenaddressed
inanumberofpapers,including(FerrariandCorazza
2004andWeinberg2008).ThiscorrespondstoDQPSK
with Gray coding over an additive white noise
channel(AWGN).

Throughout the following we use P to denote


probability, E to be statistical mean with respect toP,
V is statistical variance, Cov is covariance, N =
{0,1,2,}andN+=N{0}.Theindicatorsetfunctionis
I[xA]=1ifandonlyifxAandiszerootherwise.
General Estimator
TheconstructionofageneralMonteCarloestimatorof
(1)requirestheapplicationofthePoissonMarcumQ
Functionassociationderivedfrom(Weinberg2006).

In (Ferrari and Corazza 2004) bounds are used to


estimate the BER, while in (Weinberg 2008) a Monte
Carlo estimator is used. In the case where TISNR is
random,suchasinDQPSKoverafrequencyselective
Nakagami mchannel, with an arbitrary power delay
profile, the situation is more complex. In these
scenarios, TISNR is a sum of random variables, and
weareinterestedinthemeanof(1).

The key result in the latter is that if X1(1) and X2(2)


are a pair of independent Poisson random variables
with means 1 and 2 respectively, then (2) can be
writtenas

Without loss of generality, these Poisson random


variablesareassumedtobeindependentofTISNRby
construction. This allows us to work directly with
Poisson distributions with randomised means, free
from writing down expectations with respect to the
TISNRinthefollowing.

Bounds on the BER in this case have been used


extensively to determine performance levels, as in
(SimonandAlouini2000bandChiani1999).
In(Weinberg2008)asimpleMonteCarloestimatoris
usedtoestimate(1)forthecaseofL=1.However,this
estimator is inefficient. This is because it requires
sampling from a correlated pair of random variables,
andusesnocompressionofdatathroughconditioning
(Bucklew2005).

Equation (3) can be applied to (1) to construct a


Monte Carlo estimator. For brevity, two random
variablesaredefinedas

In particular, it is shownin (Weinberg2008) that the


BER for this case can be written as a function of two
correlated negative binomial distributions, and the
sampling is performed through simulating this pair.
This can be done using an algorithm based upon
Poissonstopped sums of independent logarithmic
series random variables (Johnson, Kemp and Kotz
2005). However, the performance of such algorithms
canbeaffectedbythecorrelation,andthisnegatively
affected the behaviour of the estimator in (Weinberg
2008).

Notethat

where in the second probability in (4), we have


switched the two Poisson random variables without
lossofgenerality.ByconditioningonZ2,andallowing
fordependencebetweenZ1andZ2,wehave

Itisshownhowthissituationcanbeimprovedusing
conditional distributions. Specifically, we will be
interested in the application of the approach of
(Bucklew 2005). A general purpose Monte Carlo
estimator of (1) will be proposed, and it will be
analysedthroughaspecificexample.

ForafixednN+,afunctiongn:N>[0,2]isdefined
by

Thepaperisorganisedasfollows.Section2introduces
the general estimator. Section 3 provides the math
ematical tools to apply this estimator in a simple
scenario. Section 4 analyses the performance of this

41

www.seipub.org/ijc

International Journal of Communications (IJC) Volume 2 Issue 2, June 2013

However,inviewoftheanalysisin(Weinberg2006),it
wouldbeexpectedthatonlyminorimprovementsare
possible.

This is a sum of cumulative distribution functions of


Z1conditionedonZ2.Thenbyapplicationof(6)to(5),
and substitution of the result into (1), the mean bit
errorrateis

In the practical implementation of (8), the knowledge


of the distribution of Z2 and the cumulative
distribution function of Z1 | Z2 are required. In the
next section we derive these in the case of DPQSK
underNakagamiFading.

Note thatsince TISNR is considered tobe random its


mean value is of interest. The expression (7) is in a
useful form from a Monte Carlo simulation point of
view (Robert and Casella 2004). Additionally it also
providesaninterestingprobabilisticformofthemean
BER.Inparticular,itisobservedthat

Application to DQPSK under Nakagami


Fading
To illustrate the application of (8) DQPSK over a
frequency selective Nakagami mchannel model is
examined,asdescribedin(SimonandAlouini2000b).
All the necessary tools are derived to apply (8) to
estimate BERs in this context. For simplicity, our
attention will be restricted to the case where L=1. In
thisscenario,theestimator(8)becomes

Consequently, (7) is the mean of a function of a


Poisson random variable with a randomised mean
value.Itisremarkedthatthismayprovideameansof
constructing bounds on (1). This will be examined in
futureresearchdirections.

whereeachYk:=Yk,1hasthesamedistributionasZ2.
UndertheNakagamimchannelmodel,theTISNRhas
aGammadistributionwithdensity

Inviewof(7),aMonteCarloestimatorisdefined

where {Yk,n, k {1,2,,N}, n {1,2, , L}} consist of


independentandidenticallydistributedcopiesofZ2.

whereavistheaveragereceivedsignaltonoiseratio
and is the Gamma function (Simon and Alouini
2000b). Since Z1 and Z2 are Poisson random variables
with means a constant times a Gamma variable, we
canconditiononthiscommondistribution
to extract the necessary distributions for (10). In
particular, (Z1, Z2) can be shown to have jointly
dependent Negative Binomial marginal distributions
(see Johnson, Kemp and Kotz 2005), with covariance
matrix

Note that it is mathematically convenient to ensure


thatthesesamplingrandomvariablesareindependent
over all pairs of indicies (k,n). This is an unbiased
estimator of (1), which can be shown by taking the
mean of (8), noting that each Yk, n has the same
distributionasZ2,andusing(4).Bymeansofthefact
that variance is additive over independent random
variables, and V(aX) = a2V(X) for constant a, the
varianceof(8)is

Further expansion of (9), through evaluation of the


varianceofgnunderZ2,doesnotprovideinsightsinto
theperformanceof(8)ingeneral.Itisobservedthat(9)
isoforder1/N,butinpracticalapplications,itmaybe
possiblethatthevarianceisofsmallerorder.However,
it is expected that (8) is much more efficient than the
corresponding estimator in (Weinberg 2008). This is
due to the compression of data through conditioning
(Bucklew 2005). This claim will be investigated in
Section4foraparticularexample.

where1=0.5b2and2=0.5a2.Notethatinmatrix(11),
theentryaij=Cov(Zi,Zj)=E(ZiZj)E(Zi)E(Zj)foralli,
j {1, 2}. The main diagonal entries in (11) are calcu
latedusingthevariancesofthemarginaldistributions.
The term a12 = a21 is obtained based the fact that the
randomvariables

areindependentPoissonswithmeaniand

Additionally, it is noted that Importance Sampling


could be used to improve the convergence of (8).

42

InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc

analysis will be twofold. Firstly, we compare (10) to


the corresponding estimator proposed in (Weinberg
2008) and will explain why the estimator (10) has
better performance. Secondly, we will compare
simulations of (10) to results obtained via numerical
integration.

and then application of the corresponding moment


expressions for a Gamma distribution. It is observed
thatsincea120in(11),themarginalrandomvariables
Z1andZ2arepositivelycorrelated.
The marginal distribution Z2 can be shown to be
NegativeBinomialwithparametersmand

EstimatorComparison
Theestimatorproposedin(Weinberg2008)is

andpointprobabilitiesgivenby

whereh(i,j)=I[ij]0.5I[i=j],andforeachk,(X1k,X2k)
has the same joint distribution as (Z1, Z2). This
estimatorrequiressamplingfromthejointdistribution
(13)directly.Inordertosimulatethebivariatepair(Z1,
Z2), an overlapping sums (OS) algorithm is in use
developed by Oregon State University. This
algorithms performance is heavily dependent on the
correlation structure (11). In particular, it was found
that in cases of interest, the algorithm could not
produce simulations, and consequently, no Monte
Carloestimatecouldbeobtained.Thistendedtooccur
as av increased. For smaller av, and especially with
largerm,thejointdistributionisdecoupled,duetothe
fact that (11) becomes approximately diagonal. Here,
thealgorithmworkswell.However,ingeneral,cases
where the algorithm could be applied take a
significant number of simulations (N >> 106) to
produce results consistent with bounds on the BER.
ThereasonforthefailureoftheOSalgorithmisdueto
the fact that it cannot tolerate cases where the means
and variances of the marginal distributions are
significantlydifferent(MadsenandDalthorp2005).

forkN(seeJohnson,KempandKotz2005).Thisis
our sampling distribution. By outlining how to
construct the conditional distribution of Z1 | Z2, we
begin to construct the joint probability mass function
ofZ1andZ2.
It is recalled that when conditioned on the TISNR,
these variables become independent Poissons. Hence
foranyappropriateindiciesiandj,

Consequently,bydividing(13)into(12),wearriveat
theconditionaldistribution

A comparison is made between the variances of (10)


and (15). Using the definition of variance, it can be
shownthat

This implies the conditional distribution is a power


series distribution (Johnson, Kemp and Kotz 2005).
The estimator (10) requires us to sample (12) and
evaluate cumulative distribution functions of (14)
through the function (6) with n=1. This can be
implemented in a simple numerical algorithm. The
following Section provides an analysis of (10) in the
currentcontext.

where

(sinceitisanunbiasedestimatorofthisaverageBER)
and

Numerical Analysis
We now examine the performance of (10) for the
example outlined in the previous Section. This

43

www.seipub.org/ijc

International Journal of Communications (IJC) Volume 2 Issue 2, June 2013

Byastraightforwardcalculation,itcanbeshownthat

considered. These estimates are compared to a


corresponding result based upon adaptive Simpson
Quadrature. In particular, the algorithmic results in
(Tanda 1993) have been used for this purpose.
Throughout,theabsolutevalueoftherelativeerroris
used, and defined as |ASQMC|/ASQ, where ASQ is
the numerical estimate from (Tanda 1993) and MC is
the Monte Carlo estimate produced through the
estimator(10).

Sinceprobabilitiesaresmallerthan1,thesumofthree
termsinsidethebracketsin(17)isboundedaboveby

andconsequentlybyapplicationofthisboundto(17)
and(16),itcanbeshownthat

Thefirstcaseconsideredisthatm=1.Table1showsthe
result of applying a Monte Carlo estimator with a
sample size of N=50,000, while Table 2 shows the
effect of increasing N to 100,000. An interesting
observation is that the relative errors are roughly of
the same order for the two sample sizes, which is
indicative of a slow rate of convergence of (10).
Likewise,alsoitisnotedthattheerrorsarefairlysmall
for the first half of av values, then they tend to
increase.Bothtablesshowanestimateofthevariance
of(10),whichisofthesameorderforbothtables.

implying(10)isamoreefficientestimatorthan(15).
Hence, from a mathematical point of view, it is clear
that(10)isabetteroptionforMonteCarloestimation
of BERs. For this reason, the estimator (15) will be
considered redundant in the numerical analysis to
follow.
Simulations
Wenowexaminetheperformanceoftheestimator(10)
through simulation. The case of m {1, 3.5, 5, 10} is

TABLE 1 m=1, N = 50,000

44

InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc

TABLE 2 m=1, N=100,000

Figure1isaplotofthelogarithmicBERasafunction
ofav(indB),forincreasingsamplesizesN.
In this case we have set m=1. What can be observed
fromFigure1isthefastconvergenceoftheestimator
(10). Figure 2 shows a magnification of Figure 1 to
further illustrate this, which indicates that in practice
theestimator(10)willnotrequireasignificantlylarge,
sample size, and hence excluding from a heavy
computationalload.

FIG. 2 MAGNIFICATION OF THE PLOTS IN FIGURE 1

Next,anexampleofnonintegralmisinconsideration.
Table 3 contains Monte Carlo simulations with the
same sample size of N=50,000. We see behaviour
similartothecaseofm=1,exceptthattheresultstend
to be much better, in terms of maintaining a smaller
relative error. It is interesting to note that the Monte
Carlo variance becomes very small as av increases.
The table shows estimates only up to av = 17 dB
because of difficulty in computing an accurate
varianceforlargerav.

FIG. 1 AN EXAMPLE OF LOGARITHMIC BER FOR VARYING SAMPLE SIZES

45

www.seipub.org/ijc

International Journal of Communications (IJC) Volume 2 Issue 2, June 2013

TABLE 3 m=3.5, N=100,000

TABLE 4 m=5, N=50,000

46

InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc

TABLE 5 m=10, N=50,000

Thelasttwosetsofsimulationsareforthecasewhere
m=5 (Table 4) and m=10 (Table 5). In both cases,
N=50,000. It is seen that the relative errors behave in
muchthesamewayasinprevioussimulations.

superior performance to that proposed in (Weinberg


2008). In particular, the new estimator is free from
simulation difficulties due to sampling from a
problematic joint distribution and yielding more
consistent results. It was found that BERs with small
av could be estimated for sample sizes of order N =
50,000,butforlargeravamuchgreatersamplesizeis
required. These simulation results are consistent with
thoseobtainedthroughnumericalintegration.

In all the simulations considered, it has been found


thatincreasingNbroughtanimmediateimprovement
in the performance of (10) for smaller av but as av
increased,theperformancereduced.
Finally, it is noted that the results generated through
theestimator(10)aresubstantialimprovementsonthe
corresponding estimates in (Weinberg 2008), and the
estimator emulates the performance of the AWGN
efficientestimator(10)inthesamepaper.

REFERENCES

Bucklew,

A.

Conditional

Importance Sampling

Estimators. IEEE Transactions on Information Theory,


Vol.51,pp.143153,2005.

Conclusion

Chiani, M. Integral representation and bounds for the

By means of the PoissonMarcum QFunction


association from (Weinberg 2006), a general Monte
Carlo estimator has been derived, enabling the
estimation of DQPSK BERs under Nakagami fading
channels. This estimator uses knowledge of
conditional distributions associated with the Poisson
Marcum QFunction relationship. In the context of a
simple example, it is demonstrated mathematically
through simulations that the new estimator has

J.

MarcumQFunction.IEEElectronicsLetters,Vol.35,pp.
445446,1999.
Ferrari, G. and Corazza, G. E. Tight bounds and accurate
approximationsforDQPSKtransmissionbiterrorrates.
IEEElectronicsLetters,Vol.40,pp.12841285,2004.
http://oregonstate.edu/dept/statistics/epa_program/brief_003
.html,datelastaccessedJune2007.

47

www.seipub.org/ijc

International Journal of Communications (IJC) Volume 2 Issue 2, June 2013

Johnson, N. L., Kemp, A. W. and Kotz, S. Univariate

IEEETransactionsonCommunications,Vol.48,pp.359

DiscreteDistributions(Wiley,NewJersey,USA,2005).

366,2000.

Madsen, L. and Dalthorp, D. Simulating correlated count

Tanda, M. Bit error rate of DQPSK signals in slow

data. Environmental and Ecological Statistics, Vol. 14,

Nakagami fading. IEE Electronics Letters, Vol. 29, pp.

pp.129148,2005.

431432,1993.

Marcum, J. L. A statistical theory of target detection by

Weinberg, G. V. Poisson Representation, and Monte Carlo

pulsed radar. IRE Transactions on Information Theory,

Estimation, of the Generalised Marcum QFunction.

Vol.IT6,pp.59144,1960.

IEEETransactionsonAerospaceandElectronicSystems,

Proakis, J. G. Digital Communications. (McGrawHill,

Vol.42,pp.15201531,2006.

NewYork,USA,2001,4 Ed.)

Weinberg, G. V. Bit Error Rate Approximations using

th

Robert, C. P. and Casella, G. Monte Carlo Statistical

PoissonandNegativeBinomialSamplingDistributions.

Methods.(Springer,NewYork,USA,2004).

IETElectronicsLetters,Vol.44,pp.217219,2008.

Simon, M. K. and Alouini, M.S. Digital Communications

Graham V. Weinberg a mathematician, has completed a


BachelorofScienceandDoctorofPhilosophydegreesatthe
University of Melbourne, Australia. He has spent the last
decade working in the area of radar signal processing at
Australias Defence Science and Technology Organisation,
specializing in detection theory. He has also completed a
Masters degree in Signal and Information Processing,
throughtheUniversityofAdelaide.Hisinterestinbiterror
rate approximations stemmed from his earlier work with
fundamentaldetectionprobabilitiesinradar.

over Fading Channels: A Unified Approach to


Performance

Analysis.

(Wiley

Series

in

Telecommunications and Signal Processing, New York,


USA,2000).
Simon,M.K.andAlouini,M.S.ExponentialTypeBounds
on

the

Generalised

Marcum

QFunction

with

ApplicationtoErrorProbabilityoverFadingChannels.

48