Departement Elektrotechniek ESAT-SISTA/TR 2000-9
Fullband Error Adaptation of Subband Adaptive Filters 1
Koen Eneman, Marc Moonen 2
April 2000
Published inthe Proceedings of the IEEE International Symposium
2000 on AdaptiveSystems for Signal Processing, Communications
and Control(AS{SPCC), pp. 293{298
1
This report is available by anonymous ftpfrom ftp.esat.kuleuven.ac.be in the
directorypub/SISTA/eneman/reports/00-9.ps.gz
2
ESAT (SISTA) - Katholieke Universiteit Leuven, Kardinaal Mercier-
laan 94, 3001 Leuven (Heverlee), Belgium, Tel. 32/16/321809,
Fax 32/16/321970, WWW: http://www.esat.kuleuven.ac.be/sista. E-mail:
koen.eneman@esat.kuleuven.ac.be. MarcMoonenisaResearchAssociatewith
theF.W.O.(FundforScienticResearch{Flanders). Thisresearchworkwas
carriedoutat theESAT laboratoryof theKatholiekeUniversiteit Leuven,in
theframeoftheBelgianState,PrimeMinister'sOÆce{FederalOÆceforSci-
entic,TechnicalandCulturalAairs{InteruniversityPolesofAttractionPro-
gramme{IUAPP4{02(1997{2001): Modeling,Identication,Simulationand
Control ofComplexSystems, theConcertedResearch ActionMIPS(`Model{
basedInformationProcessingSystems')andGOA{MEFISTO{666(Mathemat-
ical Engineeringfor Informationand CommunicationSystemsTechnology)of
theFlemishGovernment,ResearchProjectF.W.O.nr. G.0295.97(`Designand
implementationofadaptivedigitalsignalprocessingalgorithmsforbroadband
applications'). Thescienticresponsibilityisassumedbyitsauthors.
Koen Eneman Marc Moonen
ESAT { KatholiekeUniversiteit Leuven
KardinaalMercierlaan 94,B{3001Heverlee {Belgium
email : koen.eneman@esat.kuleuven.ac.be
marc.moonen@esat.kuleuven.ac.be
Abstract
For many years now, subband and frequency{domain
adaptive ltering techniqueshave beenproposedfor the
identicationofhigh{orderFIRsystems. Standardfull-
band algorithms are less attractiveas the implementa-
tion cost is higher and their convergence behaviour is
typically worse.
Subband processing has many desirable properties.
However,whenusedtoimplementadaptivelters,vari-
oussideeectsoccurwhichreduceperformance. Onthe
other hand, frequency{domain adaptive lters, suchas
thePBFDAF,donotsuerfromtheseproblemsdespite
being(nearly)equivalent tosubbandadaptive lters,be
itwith\poor"lterbanks.
In this paper an alternativefullband error basedadap-
tationschemeforsubbandadaptivesystemswillbepro-
posed. Itwillbeshownthatthe weightupdatingmecha-
nismoftheso{calledunconstrainedPBFDAFisclosely
related to the proposed fullband error adaptation algo-
rithm.
1 Introduction
Subbandandfrequency{domainadaptiveschemeshave
beenatopic ofinterestfor manyyearsnow. Theyare
employed to identify high{order FIR systems and are
a promising alternative for standard fullband adapta-
tion algorithms such as LMS. Still, with the available
multirate techniques, it is diÆcult to meet all the re-
quirements.
Frequency{domain techniques are well understood [6]
andhencetheirperformanceisverytractable. Onthe
other hand, subband adaptive lters |at rst sight|
mayhavealowercomplexityandabetterperformance.
Unfortunately,thispictureofthesubbandapproachis
certainlytoooptimistic.
In this paper we will focus on an alternative fullband
error adaptation algorithm for subband adaptive sys-
tems and show that the unconstrainedPBFDAFuses
ter weights. This is an attempt to generalise and ex-
tendthefrequency{domainaliasingcompensationtech-
niquesto subbandadaptivesystems.
2 Subband Adaptive Filtering
2.1 General setup
The general setup for a subband adaptive system is
showningure1. Theloudspeakerandmicrophoneare
...
...
...
... ...
+ -
-
+ +
- near-end signal
far-end signal
error signal
adaptive filters
analysis filter bank synthesis filter bank
+ +
+ +
e
0
1
M
1 d0
d1
dM1 y0
y1
yM1 L
L
L
L
L L
L L L
d=s+w?x s
x
f
f
f i=0
i=1
i=M 1
H
0 H0
H1 H1
HM
1 HM
1
G
0
G1
GM
1 F0
F1
FM
1 w[k]
Figure 1: Subband adaptive lter with ideal lter
banks: echocancellation setup
addedforconvenience,indicatinghowtheadaptivel-
terstructuremaybeemployedinanacousticechocan-
cellation setup. While acoustic echo cancellation has
beenadrivingapplicationformanyresearchersinthis
eld, all results obviously apply to other applications
too.
TheinputsignalsxanddarefedintoidenticalM{band
analysis lter banks. After subsamplingwith afactor
+
... ... ...
+
... ...
... ...
+
F0 H (z)
H (z)
JG T
(z) L
L
L
L
L
L
L L L
z 1
z 1
z 1 z
1
z 1
z 1 j=0
x d
e j=L 1
j=L 1
F1
FM
1
Figure 2: Subband adaptive lter : polyphase imple-
mentation
each subband. The outputs of the subband adaptive
lters are recombined in the synthesislter bank and
fedtotheoutput. Theidealcharacteristicsoftheana-
lysis bank lters H
i
and synthesisbank lters G
i are
shown (idealbandpasslters). Dueto aliasing eects,
thissetupwillonlyworkforM>L.
2.2 Polyphase implementation
Theoutputsoftheanalysisbankltersareimmediately
downsampled. Henceitischeapertodoalllteropera-
tionsatthedownsampledrate. Byre{arranginggure
1weobtain gure2. H(z) and G (z) are respectively
called theanalysis andsynthesispolyphase matrix. J
istheanti{diagonalmatrix. Element(i;j)ofH(z)is
[H(z)]
ij
=H
ij:L (z)
i=0!M 1
j=0!L 1
(1)
H
i
j:L
(z)is the j{th outof L polyphase componentof
the i{th subband lter h
i
[k], in other words the z{
transformofh
i
[j+Lk]. Similarly,
[G(z)]
ij
=G
i
j:L (z)
i=0!M 1
j =0!L 1
(2)
2.3 DFT modulatedsubband adaptive lters
Subband adaptive systems are often based on DFT
modulated lterbanks. M subbandltersare derived
fromasingleprototypelterh
0 [k] :
h
i [k]=h
0 [k]e
j 2 k i
M
; i=0!M 1
,
H
i
(z)=H
0 (e
j 2 i
M
z) (3)
Itappearsthattheanalysisltersarefrequencyshifted
versions of each other and each ltercoversa partof
thefrequencyspectrum.
+
... ...
...
... ...
... ...
+ +
F
F
F 1 F0
B(z)
B(z)
C(z) L
L
L
L
L
L
L L L
z 1 z
1
z 1
z 1
z 1
z 1 j=0
j=0
x d
e j=L 1
j=L 1
F1
FM
1
Figure3: DFTmodulatedsubbandadaptivelter
DFTmodulatedlterbanksmaybe implementedeÆ-
cientlyusing polyphase decomposition and fast signal
transforms. In [1] ageneralframework forDFT mod-
ulatedsubbandsystemswasproposed. ADFTmodu-
latedlterbankwithL{folddownsamplingcanbeim-
plementedasatappeddelay lineof sizeL followedby
astructuredMLmatrixB(z),containingpolyphase
componentsoftheprototypeh
0
,and anMM DFT
matrixF. ForDFTmodulatedlterbanksgure2can
beredrawnresultingingure3. Thesynthesisbankis
constructedinasimilarfashionwithanLM matrix
C(z).
Bysplittingsignalsintosubbandsandsubsequentsub-
sampling faster (initial)convergenceand bettertrack-
ingpropertiesarehopedfor. Astheadaptivecomputa-
tionsaswellasthelterbankconvolutionscanbedone
at a reduced sampling rate, the subband approach is
supposed togiveabetterperformanceat alowercost.
Unfortunately,thispictureofthesubbandapproachis
certainlytoooptimistic[2].
3 Frequency{domain Adaptive Filters
AsacheaperalternativetoLMS,thefrequency{domain
adaptive lter(FDAF)wasintroduced, which is adi-
recttranslationofBlockLMSinfrequencydomain[6].
TheimplementationcostfortheFDAFislow,but the
input{outputdelayintroducedbythealgorithmistyp-
icallytoohigh.
The FDAF can be extended by splitting the acoustic
impulse response in equal parts. In this way a kind
of mixed time{ and frequency{domain adaptive lter
is obtained, called the Partitioned Block Frequency{
Domain AdaptiveFilter (PBFDAF)[5][7]. Here block
lengthscanbeadjusted, resultingin acheapadaptive
lterwithacceptableprocessingdelay.
TheL
FB
{tapsfullband adaptivelterw (n)
[k]atblock
indexnispartitionedin LFB
equalpartsw
p
n
oflength
P each:
w
pn 8p
= 2
6
4 w
(n)
[pP]
.
.
.
w (n)
[(p+1)P 1]
3
7
5
; p=0: L
FB
P 1
W
p
n 8p
= F
w
p
n
0
lP
lL 1
(4)
TheequationsdeningthePBFDAFare 2
:
X
pn 8p
= diag 8
<
: F
2
6
4
x[(n+1)L pP M+1]
.
.
.
x[(n+1)L pP]
3
7
5 9
=
; (5)
y =
0
P 1 0
0 I
L
F 1
L
FB
P 1
X
p=0 X
pn W
pn (6)
d =
0
d
n
lP 1
lL
; d
n
= 2
6
4
d[nL+1]
.
.
.
d[(n+1)L]
3
7
5 (7)
e = d y (8)
W
pn+1 8p
= W
pn +F
I
P 0
0 0
L 1
F 1
X H
p
n Fe(9)
IneachiterationLnewx{samplesaretakenin,andL
newlteroutputsamplesareproduced. Liscalledthe
blocklength, the corresponding input/output delayis
2L 1. Fis anMM DFTmatrix, =2diag(
m )
contains the subband dependent step sizes and M =
P+L 1. IfP isdivisiblebyL(whichistypicallythe
case), X
pn
=X
0
n pP=L
and henceequation 5requires
only1DFToperation,whichcorrespondstop=0.The
other X
pn
can be recoveredfrom previous iterations.
VeryoftenMischosenequaltoM=P+L 1+s=2 r
withs>0andraninteger. Inmostpracticalapplica-
tionsP=L.
There exists twovariantsof this algorithm, called the
constrained and the unconstrained PBFDAF. For the
unconstrainedversiontheweightupdatecompensation
isleftoutresultinginthefollowingweightupdateequa-
tion:
W
p
n+1 8p
=W
p
n +X
H
pn
Fe (10)
The unconstrained updating requires 3 DFTs per
iteration whereas the constrained PBFDAF is more
expensive, having an extra 2LFB
P
DFTs to compute.
The latter on the other hand has better convergence
properties.
The PBFDAF turns out to be a special subband
adaptive lter having interesting convergence prop-
erties at a low implementation cost [2]. Despite the
low frequency{selective lter banks upon which the
1
Weassumethat L
FB
P
isinteger.
2
PBFDAF is based the algorithm doesn't seem to
suer from aliasing eects and hence the convergence
propertiesare comparabletothoseof LMS.
4 Alternative adaptation scheme
Inanattempttogeneraliseandextendthefrequency{
domain errorcorrection to subbandadaptive systems,
wenowfocuson analternativeadaptation schemefor
subbandadaptivesystems, which adjuststhe subband
lters F
m
(see gure4) based on thefullband errore
insteadof using thesubband errorsignals
m
=d
m
y
m
directlyasisnormallydonein astandardsubband
scheme(seegure1).
- - -
... ...
... ...
... ...
+ +
+
+
e z
1
z 1
L L L
e0 eL
2 eL
1 d0
d1
dM
1
x0
x1
xM
1
y
0
y
1
yM
1 F
0
F
1
F
M 1
JG T
(z)
JG T
(z)
Figure 4: Oversampled subband adaptivelter : syn-
thesispart
4.1 Fullbanderror adaptation scheme
Dene:
2
6
4 Y
k
0
.
.
.
Y k
M 1 3
7
5
| {z }
Y
k
= 2
6
4 X
k
0
::: 0
.
.
. .
.
. .
.
.
0 ::: X k
M 1 3
7
5
| {z }
X
k
2
6
4 F
k
0
.
.
.
F k
M 1 3
7
5
| {z }
wk
;
(11)
in which(seegure4)
X k
m
= 2
6
4 x
m [k L
f
+1] ::: x
m [k L
f L
SB +2]
.
.
.
.
.
.
.
.
.
x
m
[k] ::: x
m [k L
SB +1]
3
7
5
(12)
F k
m
= 2
6
4 f
k
m [0]
.
.
.
f k
m [L
SB 1]
3
7
5
; Y k
m
= 2
6
4 y
m [k L
f +1]
.
.
.
y
m [k]
3
7
5
w
k
is theadaptiveltervector. L
SB
isthelength of
the subband adaptive lters, L
f
is the length of the
synthesisbankpolyphaselters. Furtherdene
2
6
4 e
0 [k]
.
.
.
e
L 1 [k]
3
7
5
| {z }
E FB
k
= 2
6
4 S
00:L
::: S
M 10:L
.
.
. .
.
.
.
.
.
S
0
L 1:L
::: S
M 1
L 1:L 3
7
5
| {z }
S T
2
6
4
k
0
.
.
.
k
M 1 3
7
5
| {z }
E SB
k
(14)
inwhich
2
6
4
k
0
.
.
.
k
M 1 3
7
5
| {z }
E SB
k
= 2
6
4 D
k
0
.
.
.
D k
M 1 3
7
5
| {z }
Dk
2
6
4 Y
k
0
.
.
.
Y k
M 1 3
7
5
| {z }
Yk
; (15)
D k
m
= 2
6
4 d
m [k L
f +1]
.
.
.
d
m [k]
3
7
5
; k
m
= 2
6
4
m [k L
f +1]
.
.
.
m [k]
3
7
5
(16)
S
ml:L
=
g
ml:L [L
f
1] ::: g
ml:L [0]
(17)
E FB
k
isthefullbanderrorvector,E SB
k
containsthesub-
band errors,Srepresentsthesynthesislterbankand
g
ml:L [k] = g
m
[kL+l] is the l{th out of L polyphase
componentsofthem{thsynthesislterg
m [k].
Optimal errorsuppression isobtained whenEq. 14 is
assmallas possible:
min
w
k E
n
jjE FB
k jj
2
2 o
(18)
Theoptimalw is
w
opt
=argmin
w
k jjS
T
D
k S
T
X
k w
k jj
2
2
: (19)
In practice the optimal subband lters are estimated
adaptivelyusingasteepestdescentalgorithm. Thegra-
dientwithrespecttow
k is
rw
k
=2X H
k S
S T
(X
k w
k D
k
): (20)
Therefore,w
opt
canbefoundinaniterativewayusing
afullband errorblockadaptationalgorithm:
w
k +1
= w
k
+2X H
k S
S T
(D
k X
k w
k ) (21)
= w
k
+2X H
k S
S T
E SB
k
(22)
with = diag(
m
) a diagonal matrix with subband
dependent stepsizes. The subband errors are passed
throughthesynthesisbanktwice before theyarecom-
binedwiththefar{endsignalx.
3
Inastandardsubbandadaptationschemethesubband
errorsareminimised,leadingto
min
w
k E
n
jjE SB
k jj
2
2 o
: (23)
Thegradientwithrespecttow
k is
rw
k
=2X H
k (X
k w
k D
k
): (24)
Therefore, theoptimal w canbefound in aniterative
waybyoptimisingthesubbanderrors:
w
k +1
=w
k
+2X H
k E
SB
k
: (25)
Equation25correspondstoastandardsubbandweights
updating algorithm using a block{LMS algorithm in
each band. The compensation matrix S
S T
is absent
here. It is hoped that employing the fullband error
adaptation scheme i.o. subband error adaptationwill
leadtoimprovedperformance.
4.2 PBFDAFweightupdating
The alternative adaptation scheme studied in para-
graph4.1,leadingtoaweightupdatingequationbased
on fullband errors (Eq. 22) is now applied to the
PBFDAF.
For DFT modulated lter banks the synthesis bank
polyphasematrixG(z)canbewrittenas
F T
C T
(z)J: (26)
ForthePBFDAFitcanbeproven[2]thatwhenP isa
multiple ofL
C(z)=
I
L 0
L
::: 0
L 0
| {z }
M
=C (27)
isindependentofzandhencealsoG(z)isindependent
of z. This means the synthesis polyphase lters have
length L
f
= 1. S as dened in Eq. 14 then equals
G(z)= S=F T
C T
J. For thePBFDAFthe update
equation22thenbecomes
w
k +1
= w
k +2
M X
H
k F C
T
CF 1
(D
k X
k w
k ) (28)
= w
k +2
M X
H
k F
I
L 0
0 0
M L
F 1
E SB
k (29)
E SB
k
arethesubbanderrors,FC T
CF 1
doestheerror
correction. It can be proven (see [4]) that the weight
updateequationofaso{calledunconstrainedPBFDAF
for which P is a multiple of L (Eq. 10) is closely re-
lated to equation 29, derived here based on fullband
errorfeedback.