ORDER DETECTION
AND BLIND IDENTIFICATION OF 2
x 1MISO CHANNELS
Carlos
Estevcio
R.
Fernandes,'
Pierre
Comon
and
Ge'rard
Favier
13S
Laboratory-CNRS,
University
of Nice
Sophia
Antipolis,
France.
{cfernand,
favier, pcomon}@i3s.unice.fr
ABSTRACT
Inthispaper, we investigate theuse ofoutput4th-order
cu-mulantstodetect the number ofsourcesignalson a
multiple-input single-output (MISO) communications channel and
blindlyidentify their respective channel coefficients. More
particularly,
we areinterested in thecaseoftwo sources. The proposed cumulant-based order detectionprinciple allowsusforrecovering the longest channel order and its coefficients.
A similarprocedure is applied fordetecting thepresence of
asecond source aswell as estimating its associated channel order and coefficients. Whenonly estimated cumulants are
available,we implementtwo hypothesistestsbasedonthree proposed test-statistics. Computer simulations illustrate the performances obtained with the methodsproposedfor order detection and channel identification.
Index Terms- Order
detection,
MISOchannels,
Blindidentification, Higher-order statistics.
1. INTRODUCTION AND PROBLEM DESCRIPTION
We are interested in detecting the number of source
sig-nals and identifying channel parameters in the context ofa
multiple-input single-output(MISO)communicationsystem.
The transmitters are supposed to use the same carrier
fre-quencyand are located far apartfrom each other, hence the physical channels are different. The following assumptions
hold:
Al The input signals
sp
[n], for everyintegerp C[1, P],
are mutually independent and temporally i.i.d. (inde-pendentlyandidenticallydistributed).
A2 The random variables
sp,
p C[1, P],
have unknownnon-Gaussian distributions withnon-zerokurtoses.
A3 The channellengthsareboundedbyaknown value L.
Theoutput signal is the result ofalinear combination of the discrete-time input signals, s
[n]
C Cp. After baud ratesampling,the receivedsignal
x[n]
canbe writtenasfollows:p Lp
x[n]
= , ,hp[l]sp[n
-1],
p=l 1=0
(1)
PhDscholarship grantedfrom CAPES agency, Braziliangovernment.
where
hp[l],
with I C [0,Lp]
andp C [1, P], arethe channel parameters associated with the userp andhp[l]
= 0, Vl f[0,
Lp],
whereLp
< L, Vp C [1, P]. Let usdefine the chan-nel vectorhp
=[hp(0),
....,hp(Lp)]T
with p C [1, P]. Fromthe observation ofx[n] only,ourgoalsare to: PI: Detectthe number ofsources
P;
P2: Determine the channellengths
Lp
for eachp e [1,P]; P3: Identify thevectorshp
for eachpC [1, P].The case of overdetermined mixtures (more sensors
than sources) has been exhaustively treated in the
litera-ture, including static mixtures [1] as well as dynamic ones
[2]. On the other hand, underdetermined mixtures have
onlyrecently received attention. Our contribution is based
on cumulant-matching as in [3]. Due to assumption Al, the 4th-order cumulants of x[n], defined as C(i,j,k) A
cum{x [n], x[n+i],x*[n+ j],x[n+k] }, canbeexpressed in terms of the marginal cumulant contributions of each
source as: p
C(i,j,k) =Z Cp(i,j, k),
p=1
(2)
inwhichCp(t,j,k)dependsonthe unknown channel param-eters
hp[l]
andcanbe writtenasfollows [3,4]:Lp
Cp(ij,k) oyp
,h*[I]hp[I
+Zi]h*[I
+']hp[I
+k],
1=0
wherep e [1, P]and
-ysp
stands for the kurtosis of thesourcesp[n].
Sincehp[l]
=0, VlI[0,
Lp],
wehaveCp(i,
j,k)
=0 V i,j, k> LP. The numberPofsources(users) is unknown and assumedto be boundedby two, i.e. P < 2, sothatweconsider a MISO channel with two inputs (P = 2) and a
singleoutput,assumingtransmission channelsareof different
orders,sothat L1 > L2. Ifasinglesourceispresent(P = 1) the model still holds withL2 0 andh2
(0)
=0. Therefore,weget:
C(i, j,
LI)
=C1 (i, ,LI)
= h1[0]
hi
[i] h1 [j]
hi [LI].
(3)
2. CUMULANT-BASED ORDER DETECTION
Fromequation (3),it isstraightforwardtoobtain:
C(i,
k,
LI)
h1
[j] -C(j,
k,
LI)
h1
[i]
=0,
(4)
with 0 < i < j < LI and0 < k < L1. From equation (4)
we can get a setof L1(L1+1)2/2equations with L1+ 1
un-knowns, whichcanbe rewritten inamatrix formasfollows:
Ch1 = O, (5)
where 0 is the null-vector of dimensionL1
(Li
+ 1)2/2 and C EC(LI(Li+1)2/2)x(Li1+)
A solution to (5) is obtained by computing the right singularvectorof C associated with the smallestsingular value [3]. That solution is optimal in the total least squares(TLS)sense. To avoid the trivial solutionwe canimposeaunit-norm constraint
(E,,
|h1[n] 2 = 1).4.5 4 -5 35
33 2
25 m 2 , 1.5
10
0.5-3 4
Tested orders (I) 5 6
Fig. 1. Profile of the largestSVor1(f) ofC(f) for f [1,L]
with L1 =4,L2 =2andL =6.
Let usconsiderachannel order testing variable fC [1, L],
L > L1 > L2, and build L cumulant matrices
C(f)
CC(f(f+1)2/2)x(f+1)
satisfying (4) with f in the place ofL1.Thisyields
C(f)h1
= d, where d = 0if f > L1, as in(5), and d 74 0otherwise. That is the main idea behindour chan-nel order detector. Denoting byor(f)
> ... >Orf+I(f)
the singular values (SV) ofC(f),
we notethat for f > L1,C(f)
is the null matrix andclearlywehave:
or1(f)
= .. =7f+1(f)
=0. (6)search a rupture point in the profile ofagiven SV of
C(f),
e.g. 7i(f) for a fixed i e [1, £]. Actually, it would be
inter-estingto includenotonlyone SV
cri(f)
butas many aspos-sible. Ideally, for each f C [1, L],the f largest singular val-ues or (f),. ..
.,
o+1
(f) of matrix C (f) should be taken intoaccountfor the computation ofasingletestvariabletj(f)
con-taining information about the whole dynamic ofC(f); a de-cision threshold 'thr mustbe establishedsothati7(f) > T/t1hr
only if f<L1.
3.1. Test variables
Three testvariables
TI(f)
areproposed in Table 1. Note that TIA(f)
andi/B(f)
consist ofgeometricalmeans involving theSVof C(f). Onthe otherhand,
rqC
(f)canbedirectly derived from C(f) without singular value decomposition (SVD)com-putations, since
r/c
(f) = C(?) F;this factrepresents a great advantage ofTIC (f) overT/A (f) and /B(i)-Table 1. Computation oftestvariables.
f
\1/f
TIA(rQ)= ?) i(tZ(- ) )
i=l J
n vi(f)
TIB(f)
v-
(1+1
ICMt
=A/E
Jvi(f)l2
Vi=l
InordertodetectL1,wesearcharuptureintheprofile of thetestvariables,asfollows:
1. Initialize thelength variableasf= L;
2. Estimate theoutputcumulantsC(i,k,
£)
for0 < i < f and0 < k < f and build matrixC(f)
(as explainedin section2);On the other hand, ifwe take f = L1 we get
ui(LI)
70, i = 1,...
,LI,
with71(LI)
> ... >ULi(LI)
and (L1±1(L1) = 0. Therefore,theflargestSV ofC(f) onlytakenon-zerovalues for f < L1 andwe candetect the order L1 of the longestchannelby lookingfor thenon-zerovalues of
oi
(f),
i C[1,e],fortestedorders
f C[1, L].
Toillustratethis reasoning, fig. 1 plots the largest SV, or1(f), of the cumulant matrices C(f) for values of f in [1, L], withL = 6, L1 = 4 and L2 = 2. Perfectknowledgeof the cumulant valueswas assumed. Thefigure clearlyshows that or1(f)
= O, V f > 5 and the firstnon-zerovalue is reached for f=L1 =4.3. LONGEST CHANNEL IDENTIFICATION
Inpractice, cumulantsareaffectedby estimationerrors sothat
(6) is no longer valid for f > L1. Our problem is thento
3. Compute the SVD of
C(f),
rI(f), ....,
O7+-i
(f) itssingular values;and denote
4. Denoteby
vf
eCf
+1 the singularvectorofC(f)asso-ciated with
oe+I
(f). This is theoptimal solution of (5) intheTLS sense[3];5. Determine the testvariable
T/A(f), T/B(f)
orTc/(1)
de-fined in Table 1;6. Decreasefby 1unit andrepeat steps 2 to 5until f= 1.
In step 4, we construct a collection ofvectors, one of whichcorrespondstothe actual longestchannelvectorh1 C
CL,+1. Decision about which is the good one should be taken basedontheprofileof thetestvariables
i/A(i), i/B(f)
or
TIC(f).
A decision thresholdT/thr
mustbe established sothat
TI(f)
>TtLhr
ifandonlyif f < L1. Areasonable choiceIII-754
r T
n)
for the decision threshold is thegeometricalmeanof thetest
variable for f C [1, L],sothat:
L 1/L
Ttthr =
fl1
Ti(7)
where
T(f)
takes the values of eitherT'A(C),
T1B(Q)
orqc()
3.2.HI/H2
hypothesis testBasedonthe variabletj(f),the proposed hypothesistest
con-siders thefollowing hypotheses:
* Hypothesis H1: f>
LI
* Hypothesis H2: f =LI
We assume that H1 holds truefor f = L andtest
TI(M)
withrespect to athresholdTlthr for f =L,. ..,1. We accept H1 ifrq(f)
<lqthr,
otherwisewereject it andacceptH2, thus determining the orderTO = L1 of the longest channel. After that, step 4of thepreviously described procedureyields:h1i
=vfO: TO=Li.
(8)
Inthenextsection,wepropose amethodtoestimate the order L2 of the shortest channel and its coefficientvectorh2, if it exists. We also discuss thecaseof channels ofsameorder.
4. SHORTEST CHANNEL IDENTIFICATION
Recalling from section (1) that
Cp(i,
j,k) =0,V k >Lp,
wenotethatC2(ij,
k)
= 0 for k > L2 and henceC(i,j,
k)
C(i,
j,k),Vk
> L2. Fork=L2,weget:C2(i, j, L2) =
ys2
h2[0]h2[i]h2[]h2[L2] (9)Fromdefinition(2) withP=2, we canestimate themarginal cumulant contribution of the source
s2[n]
as C2(i,i,k)
C(i, j,
k)-
C(i,
j,k),
where C1(i,
j,
k)
is obtainedas fol-lows from the channelparametersh1[1] estimated in(8):Li
C (i,
j,k)
='Ys
Eh[l]hi
[I
+ i]hl + I]hl +k].
1=0
Therefore, the marginal cumulant C2(i, j,
k)
appears tobeagoodmetrictodetect thepresence(orabsence)ofa
sec-ondsourcewith shorter channel. If sucha source ispresent, C2(i,j,
k)
will be non-zero for somevalues of k C[1, LI].
Precisely, it equals zero for k > L2 and it is non-zero for k < L2. Hence, a new test variable
p(ii;)
can be setup todetect thepresenceof the secondsource sothat:
{
p(Q)
=0,
E >L2
p(Kc)
+4
O, K <L2-Wewill define thetestvariableas
pQ(/)
=C2Q(K)
I2,
i.e. thesquaredFrobeniusnormof matrixC2
(/)
CC(±+1)
x(I+1) in100
o0 2 3 4
Tested orders (K) 5 6
Fig. 2. Testvariable
p(Qi)
for ic C [1, L] withL1 =4, L2 =2andL =6.
which the element inposition(i, j)equalsC2 (i-1, j-1,ii), i,jC
[1,
/iC+1]Figure2illustrates the detection of L2usingp(ii;)forlc C [1, L]. The channels considered herearethe same asthose of section2 (LI =4,L2 =2andL =6). Perfectknowledgeof the cumulant valueswas assumed. Thefigureclearly shows that the firstnon-zerovalue of
p(Qi)
is reachedat c =L2 =2.4.1. H3/H4 hypothesis test
Let us nowconsider themorerealisticcasewhere theoutput
cumulants are notknown and needto be estimated. Inthat
case,the estimationerrorscorruptingC(i,j,k) imply inaccu-racies in the estimation ofthelongest channelh1, which leads
to abad reconstruction of C1
(i, j, k).
This means thatcu-mulative errors areinvolved in the calculation ofC2(i, j,k). Hence, (10) is no longer valid and another hypothesis test mustbe performed in order to detect the length L2 of the shortest channel. Forthispurpose, adecision threshold pthr
mustbe defined sothat
p(Q)
> Pth, only if/-; < L2. Here again,thegeometricalmean seems tobeagoodchoice:/L II1L
Pthr
(ft
p(k)) (11)k=I
and thehypothesistestis definedsothat:
* Hypothesis H3: EC > L2
* Hypothesis H4: C =L2
Startingwith /; = L1 andsuccessively decreasing
Ki,
weac-ceptH3when
p(Q)
< Pthr, otherwiseweacceptH4 andget anestimate ofL2. Ifthetestvariableneverreaches thethresh-old,thenweconclude thatasinglesourceispresent. Once we have determined L2,we can useC2
(i, L2)
to estimate the parameter vector h2 of the shortest channelby solvingalinearsystemsimilarto(5)intheTLSsense.4.2. Channels with identical lengths
In the case where L1 = L2, our approach is no longer
valid. Equation (3) becomes
C(i,
j,
L1)
= Ci(i,
j, L1)
+C2(i, j,
LI),
sothat theoutputcumulant contains informationon both channels. Hence, the estimation method described in section 3 willnotresult in a fair estimate ofhi and (8) doesnotholdtrue. Nevertheless, the hypothesistest HI/H2 canstill be usedto determineLI. Inaddition,we candetect thepresenceof the second sourceand determine whether the channels have thesame ordifferent orders. Proceedingtothe estimation ofthelongest channelasdescribed in section 3,we can use
h,
tocomputep(Qc)
as described earlier in thissec-tion. Thenwe canimplement thetest onthe shortestchannel, which will fail forhypothesis H3 atic = LI, indicating that both channelsareofsameorder. Therefore, the proposedtests HI/H2 andH3/H4canalsobe used for thecaseof channels ofsamelength. However, we cannotidentify the channel pa-rametersinthiscase. Infact, there isno means to seethesum
oftwo linearprocesses as ascalar linearprocess [5].
5. COMPUTER SIMULATIONS
The results shown in this sectionwere obtained from com-puterexperiments simulating the MISO system (1)with the
two following channels: hi = [1.0, 0.61 -1.18i,
-0.59-0.53i,-0.82 + 0.21i1T and h2 [1.0,-0.29
-0.33i,-0.62-0.73i1T sothat LI = 3 and L2 = 2. We
assumedL = 5and the 4th-order cumulantswereestimated fromanoutput sequenceof50000datasamples. Noadditive noise ispresent.
Table2. Longestchannel detectiontest:
P(tj
<Tlthr)
T1 =TjA
T1
=TjBT1
=TjC£=5 100%o
100%7o
100%7o
f=4 luu% 100% luu%
£=3 2.7% 0.88% 6.26%
Success rate 97.3% 99.1% 93.7%
Figure 3 shows thehistogram plots for the testvariable
hqA,
built from thesample outputcumulants over 5000 MonteCarloruns. Similarplotsfor
riB
andn/c
were also obtained. Table2 shows theprobability ofnq(1)
<Tlthr
highlighting thesuccess rate of the LI detection. Forf < 2all the
probabili-tieswere zero.Notethat
n/B
gives the bestprobability of good decisionsonthe channel orderLI withonly0.88%
probabil-ity of false alarm. Onthe otherhand, although
n/c
yieldsthepoorestperformance among the three considered criteria, it could be worthwhileconsideringto useit since it doesnot
re-quireany SVDcomputationandcanbe deriveddirectlyfrom
C(f)
F. Inthis case, the threshold level could be revised inordertoadjust the sensitivity of the detector(e.g., reduc-ingn/thr
ofan amountof10%
increases the success rate to 98.2%.).6. CONCLUSION
Inthispaper, amethodusingtheoutput4th-order cumulants has beenproposedto detect the number ofsourcesand
iden-IL
50l150 150
100 100
0( 2 4 6 8 °( 2 4 6 8
men(A.)=3273 me (A )=7.0
Fig. 3. Histogram plots for
qA
(mean value Ofrlthr is 2.19).tify theirrespective transmission channels fora 2 x 1
com-municationsystem. Thesameprinciple applies forany P x 1 MISO channel. The proposed order detection principle is based on the search for the highest value ofatest-variable
sothat the cumulant matrix isnon-zero. Afterextracting the contributions of that channel from the output cumulants, a
similar procedure is applied for detecting thepresence ofa
secondsource aswellasits associated channel order and
co-efficients. Becausethere isno means to see thesum oftwo
linearprocesses as ascalar linearprocess, we cannotestimate
parametersof channels ofsameorder.However,it would also be possible to detect the number ofsources via a linearity
test [5]. Considering estimated cumulants, two hypothesis
testsbasedonthreeproposedtest-statistics have been
imple-mented. Someperspectives includea tensor approach based
on a testof thetensorrank. Theasymptotic distribution ofthe proposedtestvariables also needstobeinvestigated.
Acknowledgement
The authors would liketothank theanonymousreviewers for their relevant and useful remarks.
7. REFERENCES
[1] P. Comon, "Independent Component Analysis, a new con-cept?,"SignalProcessing,Elsevier,vol.36,no.3,pp.287-314,
apr. 1994, Specialissue onHigher-Order Statistics.
[2] E. Moulines, P. Duhamel, J. F. Cardoso, and S. Mayrargue,
"Subspace methods for the blind identification ofmultichannel
FIRfilters,"IEEETrans.SigProc., vol.43,no.2,pp.516-525,
feb. 1995.
[3] P. Comon,
"MA
identificationusingfourth order cumulants,"Signal Processing, Elsevier,vol.26,no. 3,pp. 381-388, 1992. [4] D. R.Brillinger,and M.Rosenblatt, "Computationand
interpre-tation of kth-orderspectra," Spectral AnalysisofTimeSeries, Wiley,B.Harris(ed.),NewYork, USA,pp. 189-232, 1967.
[5] M.Hinich, "TestingforGaussianityandlinearityof astationary
timeseries," Journal ofTimeSeriesAnalysis,vol.3, no.3,pp. 169-176, 1982.