HAL Id: hal-02572620
https://hal.archives-ouvertes.fr/hal-02572620
Submitted on 13 May 2020
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Properties of the partial Cholesky factorization and application to reduced-rank adaptive beamforming
Olivier Besson, François Vincent
To cite this version:
Olivier Besson, François Vincent. Properties of the partial Cholesky factorization and application to reduced-rank adaptive beamforming. Signal Processing, Elsevier, 2020, 167, pp.107300-107309.
�10.1016/j.sigpro.2019.107300�. �hal-02572620�
an author's https://oatao.univ-toulouse.fr/25983
https://doi.org/10.1016/j.sigpro.2019.107300
Besson, Olivier and Vincent, François Properties of the partial Cholesky factorization and application to reduced- rank adaptive beamforming. (2020) Signal Processing, 167. 107300-107309. ISSN 0165-1684
Properties of the partial Cholesky factorization and application to re duce d-rank adaptive beamforming
Olivier Besson∗,François Vincent
ISAE-SUPAERO, 10 Avenue Edouard Belin, Toulouse 31055, France
Keywords:
Adaptive beamforming Cholesky factorization Reduced rank Wishart matrices
a b s t r a c t
Reduced-rankadaptivebeamformingisawellestablishedandefficientmethodology,notablyfordistur- bancecovariancematriceswhicharethesumofastronglow-rankcomponent(interference)andascaled identitymatrix(thermalnoise).Eigenvalue orsingulardecompositionisoftenusedtoachieve rankre- duction.Inthispaper,westudyandanalyzeanalternative,namelyapartialCholeskyfactorization,asa meanstoretrieveinterferencesubspaceandtocomputereduced-rankbeamformers.First,westudythe anglesbetweenthetruesubspace andthat obtainedfrom partialCholeskyfactorizationofthe covari- ancematrix.Then,astatisticalanalysisiscarriedoutinfinitesamples.Usingpropertiesofpartitioned Wishartmatrices, weprovideastochasticrepresentationofthebeamformerbasedonpartialCholesky factorizationandofthecorrespondingsignaltointerferenceandnoiseratioloss.Weshowthatthelatter followsapproximatelyabetadistribution, similarlytothe beamformerbasedoneigenvalue decompo- sition.Finally,numericalsimulationsarepresented whichindicatethatareduced-rankadaptivebeam- formerbased onpartial Choleskyfactorizationincursalmostnoloss, and caneven performbetterin somescenariosthanitseigenvalueorsingularvalue-basedcounterpart.
1. Motivationofthework
Enhancing retrieval of a signal of interest (SoI) buried in noise and interference by means of adaptive filtering is a very widespread problemin many engineering applications, including radar sonar andcommunications [26], aswell asin finance with theselectionofamean-varianceefficientportfolio[23].The most widelyusedapproachconsistsindesigningalinearfilterw,which preserves the SoI through some constraints, and minimizes the output power, hence tends to cancel or at least attenuate inter- ference. In other words, one tries to minimize wHw, under a unit-gainconstraintwHv=1,wherestandsforthep×pcovari- ancematrixofthemeasurementsandvdenotestheSoIsignature.
The solution is wopt=−1v/(vH−1v), which provides the opti- malsignaltointerferenceandnoiseratio(SINR),whichwedenote asSINRopt=SINR(wopt).
In array processing applications,when =RN contains noise only(i.e.,thermalnoiseandinterference),thisfilterisreferredto as the minimum variance distortionlessresponse (MVDR) beam- former. If the SoI is presentin the measurement, in which case =RS+N, one speaks ofminimum power distortionlessresponse (MPDR) [26].When is known, there is no difference between
∗ Corresponding author.
E-mail address: olivier.besson@isae-supaero.fr (O. Besson).
the two filters. However, in practice is unknown and must be estimated from a set of n independent and identically dis- tributedsamplesxi,with(supposedlycommon)covariancematrix
=E
xixHi
.Useofthesamplecovariancematrixˆ inlieuofthe trueonemakesabigdifferencebetweentheMVDRandMPDRsce- narios. Indeed,while only 2p−3 samples are neededto achieve anaverageSINRlossequalto−3dBintheMVDRcase,thisfigure risestoabout(p+1)(1+SINRopt)intheMPDRcase,whichcanbe muchlarger.Inmostpracticalcases,thenumberofavailablesnap- shotsismuchlower, whichresultsina significantdegradation in theMPDRcontext.
In some cases one might even face a situation where n<p, which makes ˆ singular, and therefore its inverse does not ex- ist. Inorder to address thissituation andto improve the rateof convergence,twomainapproacheshaveemergedintheliterature.
The firstis based on diagonalloading [1,5,6], a simpletechnique which was shown to achieve, provided that the loading level is properlychosen,a fastconvergencerate, typicallyofthe orderof twicethenumberofinterferingsignals[2,6,7,10].Thesecondmain categoryisthat ofreduced-rank adaptivefilterswheretheweight vectorw isconstrainedto liein alow-dimensional subspace,see e.g.,[11,14,15,17,22].Thisapproachisparticularlysuitablewhenthe interferenceplusnoisecovariancematrixisthesumofalow-rank term and a scaled identity matrix, i.e., when RN=UUH+σ2I
where U is a p×r matrix of the eigenvectors of RN and is
https://doi.org/10.1016/j.sigpro.2019.107300
thediagonalmatrixofeigenvalues. Actually,forlarge interference to noise (INR) ratio wopt is approximately, up to scaling factor, theprojection ofv on thesubspace orthogonal to U, i.e., wopt
α(I−UUH)v.Inpractice,anestimate ofUismadeavailable from the eigenvalue decomposition (EVD) of ˆ or the singular value decomposition(SVD) ofthedatamatrixX=
x1 x2 ... xn
. AnalysisoftheSINRlosswasconductedin[21]whereitwasstated thatitapproximatelyfollowsaBetadistribution.In[16],theSINR lossisanalyzed startingfromthe asymptoticpropertiesofeigen- valuesandeigenvectors.Whiletheproofisrigorous,itholdsonly forn→∞,whilereduced-rankadaptivefilteringisespeciallyinter- estinginlowsamplesupport.
While EVD or SVD are the usual tools forretrieval ofprinci- pal subspace, one might investigate alternative, possibly simpler factorizationsthat could yield a similar performance in terms of adaptivebeamforming:thisis theobjective ofthe presentpaper.
Our goal is not to propose a new reduced-rank adaptive beam- former,ratherto suggest anotherimplementationandto validate its performance through a theoretical analysis. Towards this end weproposetouseapartial(truncated)Choleskyfactorization,that iswe suggest to use a conventional (with no pivoting) Cholesky factorizationalgorithm,applied toˆ, andtostop it afterr itera- tions.We are particularlyinterestedin thecasewhere ˆ isesti- mated fromn<p snapshots andis thus of rankn. Also, we will focus on matrices =UUH+σ2I (MVDR case) or =PvvH+ UUH+σ2I(MPDR case),whereUformsabasisfortheinterfer-
encesubspace.
The paperisorganizedasfollows.InSection 2,we willdefine thepartialCholeskyfactorizationandwillshowthatitisarather accuratemethodtoretrievetheprincipalsubspaceofwhenthe latteristhe sumofa low-rankmatrix anda scaled identityma- trix.InSection3wewillintroduceareduced-rankadaptivebeam- formerbased on thepartial Cholesky factorizationandderive its statistical properties. The simulations of Section 4 will compare thisbeamformer toits counterpart usingSVDandfinally conclu- sionswillbedrawninSection5.
2. PartialCholeskyfactorization
In thissection,we introducethe partialCholeskyfactorization andwestudytheanglesbetweenthesubspaceobtainedfromthe partialCholeskyfactorizationofandtheinterferencesubspace, intheMVDRaswellasintheMPDRcase.
2.1.Definition
Let us start fromthe (full) Choleskyfactorization [12,19] ofa positive definite matrix A which yields a p×p lower triangular matrixL,withpositivediagonalelements, suchthat A=LLH.The principleofthepartialCholeskyfactorizationissimplytostopthe factorizationafterrsteps(columns),yieldingap×rlowertriangu- larmatrix,whosepropertieswillbestudiedbelow.IfAisfull-rank, andnopivoting is used,then stoppingafter riterationsprovides thesameresultasselectingthefirstrcolumnsofa fullCholesky factorizationofA.IfAhasrankn<p,thenthefullCholeskyfactor- izationdoesnotexist,andstoppingafterr≤nsteps yieldsalow- rankapproximation, which will be used to obtain a basis of the interferencesubspace. Note that, when A has rank n, symmetric pivotingcanbeusedtoproduceap×nlower triangularmatrixL withpositivediagonalelements,andapermutationmatrixsuch that of A=LLH, see e.g., [12], Algorithm 4.2.4. Now, with a positivesemi-definitematrixA,Choleskyfactorizationwithpivot- ingcanalsobestoppedafterrsteps,wherer≤rank(A),toproduce alow-rankapproximationofA.Suchatechniquehasbeenusedin machinelearningwhereitservesthepurposeoffindingalow-rank approximationoftheKernelmatrixseee.g.,[3,9],anddifferences
amongmethods concernthestrategy forpivoting. In[18]conver- genceresultswereobtainedwhichprovetheeffectivenessofsuch amethod.
The partial Cholesky factor of , which will be denoted as pchol(, r), is thus the p×r (with r≤rank()) lower triangular matrixwithpositivediagonalelementsG=(GG−1
−2),whereG−1 isa r×r lowertriangular matrixwithpositive diagonalelements andG−2isa(p−r)×rmatrix,definedfrom
=
11 12
21 22
=
11 12
21 21−11112
+
0 0 0 2.1
=
G−1
G−2
GH−1 GH−2 +
0 0 0 2.1
(1)
where 2.1=22−21−11112, G−1GH−1=11 and G−2GH−1=21.Fromapracticalpointofview,G=pchol(,r) can be obtained, e.g., by usingonly rsteps of Algorithm4.2.2 of GolubandLoan[12].
2.2. Angleswiththeinterferencesubspace
Letusnow examinethe abilityof thepartial Choleskyfactor- ization to retrieve the principal subspace of when the latter isoftheform=UUH+σ2Iwhere=diag(λ1,...,λr)isthe diagonal matrix of eigenvalues and U is the matrix of eigenvec- tors. Towardsthis end, let usstudythe angles betweenthe sub- spacespannedbypchol(,r) andthesubspacespannedbyU.Let us partition U as U=[U1
U2] where U1 is r×r. From (1), one has 11=U1UH1 +σ2Irand21=U2UH1,sothat
range(G)=range 11
21
=range
U1UH1+σ2Ir
U2UH1
=range
U1
U2
UH1+
σ2Ir
0
=range
U1 U2
+
σ2
UH1 −1−1 0
=range
U1+1
U2
. (2)
ForlargeINR,i.e.,forλkσ2,1isamatrixwithsmallelements, hencerange(G)should beclosetorange(U).The anglesbetween thetwosubspacesareobtainedfromthesingularvaluesofthefol- lowingmatrix[12]
M=UH
U1+1
U2
(U1+1)H(U1+1)+UH2U2
−H/2
=
Ir+σ2−1 Ir+H1U1+UH11+H11 −H/2
=
Ir+σ2−1 Ir+2σ2−1+σ4−1 UH1U1
−1−1−H/2
Ir+σ2−1 Ir−σ2−1−1 2σ4−1
UH1U1 −1
−1+3 2σ4−2 Ir−σ4
2−1 UH1U1 −1
−1+1 2σ4−2
=Ir−σ4 2−1UH2
Is+U2UH2 −1U2−1. (3) Note that one should go up tosecond orderin the expansion.If we let θk,k=1,...,r denote the angles between the subspaces,
Fig. 1. Distance between range pchol
UU H+ σ2I , r
and range( U ) using the par- tial Cholesky decomposition. r = 2 interference with varying INR in the field of view of a p = 16 element uniform linear array.
one has Tr
MMH
=r
k=1cos2θk. Now, the square distance be- tween the two subspaces is d2=r
k=1θk2r
k=1sin2θk and, for largeINR,
r
k=1
sin2θkσ4Tr −1UH2
Is+U2UH2 −1U2−1
. (4)
Although this is not zero, as would be the case with the sin- gular value decomposition whoser principalleft singularvectors sharethesamesubspaceasU,thedistancebetweenrange(G)and range(U)goestozeroasINR−1 whenINRgrowslarge.
This is illustrated in Fig. 1 where we display the distance betweenrange(G) andrange(U) inthecaseofa p=16element uniform linear array with inter-distance half a wavelength. Two narrowband interference are present in the field of view with respective directions of arrival −10◦ and10◦,and a varying INR.
As canbe observed, thedistancegoesto zeroasboth INRsgrow large. Note that this figure concern the asymptotic case where is known. Of more practical interest is the case where is estimated from n snapshots with n possibly smaller than p. In this case, a different analysis should be conducted. In the next section, we provide some results about the statistical properties of pcholˆ
,r while numerical results about its application to adaptivebeamformingisthesubjectofSection4.
Let us now study what happens when =PvvH+UUH+
σ2I, i.e., when the SoI is present in the measurements. With no loss of generality, one can assume that v=ep where ep=
0 0 ... 0 1T
.Then,onlythe(p,p)elementofismod- ified, compared tothe MVDRcasewhere =UUH+σ2I.Since
thisaffectsonly22 itmeansthatthepartialCholeskyfactorization isleftunchanged,thatis
pchol
PvvH+UUH+σ2I,r =pchol
UUH+σ2I,r (5)
which implies that the angles between
range
pchol
PvvH+UUH+σ2I,r
and range(U) are still given by the analysis above. Note that this holds true provided that the first r components of v are zero. In contrast, the angles be- tweenrange(U) andthefirstreigenvectorsofUUH+σ2Iareno
longer zero.Indeed,it isonlyknownthat thefirstr+1principal eigenvectors of PvvH+UUH+σ2I share the same subspace as U v
. Accordingly, the subspace spanned by the r principal
Fig. 2. Distance between range eig
Pvv H+ UU H+ σ2I , r
and range( U ). The dis- tance between range
pchol
Pvv H+ UU H+ σ2I , r
and range( U ) is given by Fig. 1 . r = 2 interference with varying INR in the field of view of a p = 16 element uniform linear array. Signal to noise ratio is 0 dB.
eigenvectors contains a contribution from v and one cannot recover exactly the interference subspace, in contrast with the MVDR case. Therefore, when one wants to retrieve the interfer- encesubspaceinaMPDRcontext,apartialCholeskyfactorization, stopped afterr steps withrthe number ofinterfering signals,is a meaningful approach, because it brings the SoI component in a part of the matrix which is not used in the partial Cholesky factorization.
Fig. 2 displaysthe angles obtainedwith an EVD in the same scenario as before but with the SoI present. Note that the an- glesobtained witha partial Choleskyfactorization are still given by Fig. 1,due to (5). Comparingthe two figures, it appears that thedifference in termsof subspace proximity isnot very impor- tant. Therefore,fromthe point ofview of retrievingthe interfer- encesubspaceinaMPDRcontext,apartialCholeskyfactorization possessesinterestingproperties.
3. Reduced-rankadaptivebeamformingusingthepartial Choleskyfactorization
Inthissection, we consider theuse ofa partial Choleskyfac- torization for reduced-rank adaptive beamforming purposes. Let us assume that n independent and identically distributed com- plexGaussianvectorsxiwithzeromeanandcovariancematrix areavailableandgatheredinthematrixX=
x1 x2 ... xn
. We denote the distribution of X as X=d CNp,n(0,,In). We are especially interestedin the casewhere the numberof snapshots is less than the size of the observation space, i.e., n<p. Let S=XXH be the samplecovariance matrix andlet G=pchol(S,r) withr≤min(p,n)denoteitspartialCholeskyfactorization.Assaid above,Gconstitutesausefultooltoretrievetheinterferencesub- space, which leads naturally to consider the following reduced- rankadaptivebeamformer
w=P⊥Gv (6)
where P⊥G=I−G(GHG)−1GH is the orthogonal projector on the subspaceorthogonal torange(G). Inthesequel, weprovide asta- tisticalanalysisoftheSINRlossoftheadaptivebeamformerin(6). Priortothat,wealsoprovidesomestatisticalpropertiesofG.The mainresultsare statedbelowwhiletechnicalproofsare deferred