FINITE-LENGTH MIMO ADAPTIVE EQUALIZATION USING

CANONICAL CORRELATIONS

Aleksandar Dogandžić and Arye Nehorai

EECS Department
University of Illinois at Chicago, Chicago, Illinois 60607
email: adogandz, nehorai
@eecs.uic.edu

ABSTRACT linear combination of basis functions, which includes var-

ious wireless channel models as special cases, see [1], [14].
We propose finite-length multi-input multi-output adaptive
However, unlike [13], where the measurements are real and
equalization methods for “smart” antenna arrays using the
basis functions are known, here we consider the measure-
statistical theory of canonical correlations. We show that
ment model with complex data and parametric basis func-
the proposed methods are related to maximum likelihood
tions. The proposed parametric basis function model is use-
reduced-rank channel and noise estimation algorithms in
ful in blind equalization and symbol detection, i.e. when
unknown spatially correlated noise, and to several recently
training data is not available, see Section 5.
proposed adaptive equalization schemes.
Denote by  an 
 data vector received by an array
of antennas at time  and assume that we have collected
1. INTRODUCTION
 temporal data vectors. Then, we consider the following
Multi-input multi-output (MIMO) channel equalization has
recently attracted much attention due to recent popularity measurement model:

"#$&%'%'%(  
of antenna arrays applied at the receiver [1] and transmitter
[2]. Adaptive and non-adaptive decision-feedback (DFE)  ! (2.1)
MIMO equalizers have been recently proposed in [3] and
[4], respectively (see also references therein). In this pa- where  is an )
+* channel response matrix of rank ,.-
per, we present methods for finite-length MIMO adaptive /1032  4*5 , 6 is a *
7 vector of basis functions,
spatial and temporal equalization based on canonical cor-
relation analysis [5], [6]. These methods are multivariate
extensions of the adaptive equalization algorithms in [7],
[8], [9], classical finite-length adaptive equalization in [10],
and blind adaptive beamforming methods which use finite
alphabet [11] and constant modulus [12] properties of the
received signal. We show a relationship between the pro-
posed methods and reduced-rank multivariate linear regres-
and ! is zero-mean Gaussian, temporally white and spa-
tially correlated noise with unknown positive definite co-
variance 8 . The basis functions 9" are chosen to de-
scribe the signal of interest, and  is a vector of unknown
basis-function parameters, which may be the unknown sym-
bols or phases of the received signal in constant-modulus
scenario (see Section 5.1.1).
To present the ML estimates of  and 8 , it is useful to
 
sion problem solved in [13]. GHII :;# < =?>&>'>@H
define  G QQ BA , C" D< 6$?>'>&>E H9G IEQ "FA ,
First, in Section 2, we briefly review the maximum like- JLK  M>N:O:1P ,
H G QUP I ; R@SK >TC"CEP , 
lihood (ML) channel and noise estimation in [13]. Then, @SK V>=:WCMEP , and
we describe the proposed adaptive equalization algorithm
X G I Q W G G G Y(Z\
 H I Y!I Z[E\ H IQ H QQ 
in Section 3, and discuss its application when training data
is available (see Section 4) or not available (i.e. blind sce- (2.2)
nario, see Section 5).
H G IY(I Z\ isthe
which G Y!Z[E\ cross-correlation between the vectors
 and H QQ 6 or the estimated coherence ma-
estimated
2. REDUCED-RANK ML ESTIMATION
We review the ML estimation in [13] for a reduced-rank ] Z\ between  and " , see [5, Section VII]. Also,
trix
channel. As is [13], we model the received signal as a ] denotes] Y(aZHermitian
[E\ ^ ] Z[square
\  Y!Z ; thisroot of a Hermitian ma-

This work was supported by the Air Force Office of Scientific Re-
trix , and
9
H G IEQ notation

H G QQ will be used
throughout the paper. Note that and are functions
of  . To simplify the notation, we omit these dependen-
search under Grants F49620-97-1-0481 and F49620-99-1-0067, the Na-
tional Science Foundation under Grant MIP-9615590, and the Office of
Naval Research under Grant N00014-98-1-0542. cies throughout this paper. Consider now the singular value
 X G IQ Ÿ
decomposition (SVD) of :
X G IUQ _ G `G a G P  –  ƒ – ” M – /˜uS™›šœ  œ L K(ž ”"">d”MEP&K  ž –
 : Ž —  ž ž
_ G P _ G  _ G _ G P bdce a G P a G  a G a G P b'fT
(2.3a)
Ÿ
g `G (2.3b)

`G < `G  hiNAjk mln* 
 (2.3c) —
< *5iNApom mqn*
`G Gy Gy Gy C"
 rs t 0puNvxw @=d {zT'%&%'%!  r|T (2.3d)
Gy
Gy
where “ P ” denotes Gy a conjugate transpose and h} =1}
jz$r}~%&%'%}  /1032  4*5r}€i . Again, for notational
simplicity we omit the dependence of the above quantities Fig. 1. Proposed MIMO adaptive equalization scheme.
on  .
We now present the ML estimate of the reduced-rank • •
 . First, we adopt the following notation: Here, > denotes the determinant. The normalizing con-
G_ G 
,N and a _ G ,N are the
channel matrix
straint prevents trivial solution (in which Ž and
a G matrices containing the first ,
equal
zero), and imposes the estimated beamformed signals Ž :
columns of and , respectively. For the model in (2.1)
with known  , the ML estimates of  and 8 are
to be uncorrelated.
It can be shown that, under Ž
H9G II Ž P “b , all the eigen-
G H G IZ I[E\ _ G  ,S `G  ,N a G ,N P H G QY!Q Z[E\  values of ”’>”EP=K
 are simultaneously minimized for
"‚ (2.4a) …
G H9G IIƒ HOG IZI[E\ _ G  ,N ` G \ ,N _ G ,N P O H G @IZI [\  G G¡ G¡ G¡
8"‚ (2.4b) Ž.‚ < G Z  "G \ M?>'>&> MBA P
G _ P H I Y!I Z[E\ …
If  is unknown, its ML estimate   ,S  (3.3a)
see [13], [14]. is ob- ¢ G¤ Z G¤ \ G¤
tained by maximizing the concentrated likelihood £ < G "G  G "?>'>&G > G "FA P G
` Y!Z[E\
„ ‡  Ž. H9IEQ HWQY!Q Z  ,N…  a ,N P H QQ  (3.3b)
"M†… ƒ Gy \ 
_G a G ,S are the matrices containing the first ,
(2.5)
ˆŠ‰ Z   ‹@ where ,N and G_ aG
see [14, eq. (4.1)] and [13, eq. (35)]. ToG find the ML esti- columnsG of
Ž¥  Œ
Ž 
 "
    ¢ " maximize
and , respectively (see [15]). Therefore,
mates of  and 8 , replace  in (2.4) by  . ¢
„ G and
„ (3.1), yielding
In the following, we propose an alternative criterion,  Ž.d  MO V , which isG the concentrated G ¢ like-
which is maximized for the same estimate of  as the con- in (2.5). Note that Ž."> M  " ,
lihood function G
centrated likelihood function (2.5). This criterion is moti- where G " is theG ML estimate G of the G
channel in (2.4a).
vated by the MIMO equalization scheme in Figure 1.
G
Also, §¦U¨G ©< ª ¦ œ Z "¬ G  ª ¦ œ \ M¬ G '%&%'%( ª ¦ œ "¬ GFA«o;
ŽŒ¢"B and ¦"MD< ¦ œ Z " ¦ œ \ d'… %&%'%( ¦ œ BApo
  9" can be viewed as estimated canonical … co-
3. MIMO ADAPTIVE EQUALIZATION
We analyze the adaptive MIMO equalization scheme de-
picted in Figure 1. We wish to find an ,Œ
 beamform- Gy
ordinates of the data and basis functions, respectively, whereas
ing matrix Ž and an ,h
* basis-function filtering matrix ‹@ are the estimated canonical correlations, see [5]. This
 that minimize the error between the beamformed data allows for an elegant interpretation of the proposed equal-
and filtered basis functions ‘’M“ŽO
ƒ 6" in ization scheme in the context of canonical correlation analy-
the mean-square sense. Define the error matrix as ”M sis, see e.g. [6, ch. G ¡ 12]. The first estimated G
G Z MEP and ¬ ¦ œ Z Gy "M ¤ G Z M P 
canonical coordi-
< ‘’$?>&>'>E‘’  BA . In the following, we show that this nates ª ¦ œ Z M

problem is related to canonical correlation analysis. have the largest estimated correlation @L among allG possi-
ª œ
We propose to estimate Ž , , and  by maximizing ble G ¡ linear combinations
¬ G œ of  and¤ G " . Further, ¦ \ 
the inverse of estimated geometric mean-squared error of  \ "Pd and ¦ \ " Gy  \ "P= have the
‘’ : largest estimated correlation jz$ among all possible linear
„  and " that are uncorrelated with
Ž.  M • LK  V>'” >'”M • ªG ¦ œ Z M and ¬ G ¦ œ Z " , and so on.
combinations of
(3.1)
P ƒ ƒ
For a single sensor with #< ª  ª  =d'%'%&%( ª 
subject to the normalizing constraint ­=BApo , rank ,OD , and basis functions chosen to model
H G II Ž P “b %
Ž  ƒ
the multipath effect by uniformly discretizing the time-delay
ƒ
(3.2) spread (i.e. 6M;< ®¯E®¯ =d'%'%&%(E®¯ *x˜LFA o ), the
…
 G G¡ X ¹!º X ¿ÁÀ º
¢ in (3.3) become rowG ¡ vectors, i.e. Ž.M° EP i.e. ² ¦ M ² " , see [15]. If  is the unknown
equalizers
G
and "± ¤ "P , where " can be interpreted as a sequence to be detected, maximum likelihood sequence es-
feedforward filter, G which shapes the channel to the desired timation (MLSE) can be used to minimize the above cost
impulse response ¤ " ; this is a classical adaptive equal- functions with respect to  , along the lines of [16].
ization scheme in [10]. For rank-1 channels (i.e. ,ÂD ) and basis functions cho-
For unknown  , the maximization of (2.5) can be per- sen to model the multipath effect by uniformly discretizing ƒ =d'%&%'%(
the time-delay spread (i.e.  ² #< ® ² dE® ² 
formed by iteration, as described in Section 5. ƒ
® ²  *° =FA«o and similarly for " ), the above equal-
4. MIMO EQUALIZATION AND SYMBOL ization and detection algorithms become very similar those
DETECTION USING TRAINING DATA in [7], [8], [9] (where the differences arise because the nor-
If training data is available, we can separate the equalization malizing constraints in [7], [8], [9] differ from (3.2)).

Ž and  [see (3.3)] and then detect the unknown sequence

and detection tasks as follows: use training data to estimate
5. BLIND MIMO EQUALIZATION
by applying metric combining (MC) [16, sec. IV.A] to the Two iterative procedures for blind MIMO equalization and
equalized data and basis functions. We show that this proce- symbol detection followfrom the results of Sections2 and3.
dure is equivalent to estimating 8 and  from the training The first iteration is based onG the ML resultsG in Section
2: first fix  and compute à " and 8n 8" using
data [using (2.4)] and detecting the unknown sequence us-
¾
(2.4), then fix  and 8 and minimize the interference re-
ing interference rejection combining (IRC) [16, eq. (8)] (see
jection combining cost function Ä ½ ‰ Z < Å
ƒ BAP>
also [17]).
 (
Y Z ƒ
8 >T<  9MBA with respect to  . Iterate between
Let :³²¨D<  ² =?>&>'>@ ²  ²xFA be the data set containing
the known (or training) sequence, described by known basis
functions C ² )<  ² @L?>'>&> G ²  ² BA . G Further,
 the above two steps as long as there is a significant increase
H Y(I@ZI [E\ H G ² IQ H G Y(QZQ [E\ 
by analogy in (2.5).
X IQ
with
_ G `G a G P
(2.2) and G
(2.3), define ²  G ² ² An alternative iterative ¢ on (3.1): first
G method is based
² H G ² I² Q , where H ² II :x²³:x² P K  ² , H ² QUQ “C²?C² P K  ² , fix  and compute  ­
Ž  .
Ž 
 "
  and  " using (3.3),
P then fix Ž and and maximize (3.1) with respect to  ; it-
and
X G IQ ² Gy :x²?C² K Gy ² . Define also the Gy singular values of
/10Š2  E*¯EW}µi
² ` G : r} ² @LW} Gy ² {zTW Gy ´
} '
% &
% "
% } Gy ²  erate as long as there is a significant increase in (3.1). In the
and ² ,S9¶t 0puNvxw ² @L ² jz$d'%'%&%( ² ,S| . Then, from
following section, we consider the full-rank channel with
 ,O* co-channel signals, which allows for further simplifi-
(3.3),G the estimates
_ G H G ofY(IZŽ I [\ and ¢ based`G on thea G training H G Y(data
Z\
²,N ²,NP ² QQ ,
cations of this iteration.
are Ž ²“ ²,SEP ² and ²“
G the G channelZ[E\ G and 5.1. Full-rank Channel with ,O* Co-channel Signals
`G noiseG estimates G Y(Z\ followG from (2.4) G
 G Z² [E\  G H ² I`G I _ ² ,NG ² ,N a G ² Z,N\ P H ² QQ and 8 ²  H ² G II9ƒ
and as

H II _ ²,N \ ,N _ ²V,SEP H II  ƒ Consider now an important special case of a full-rank chan-
¢ ² G ² ƒ ¢ ² . Note that @LK ²x'< Ž9²³:x² nel with ,#Æ* narrowband co-channel signals ® Z " ,
ƒ `Galso
\ ® \ "&%'%'%(U®  impinging on the array. Then, the ba-
² C ² A³>¯< Ž ² : ² ² C ² A P b G ² ,N .
Now, apply the beamformer … Ž ² to the data : contain- sis function vector … becomes  M;< ® Z "U® \ d
ing ¢ the unknown sequence and and the basis-function filter %'%&%!E® "FA«o . Since in (3.1) reduces to a square (and
… non-singular) matrix, we can recover the
G ² to the G hypothesized basis functions "G  , yielding
CM by computing Ƕ  Y!Z ŽW: , once  signals
generally
G G
 G ² ¦ £< ª ² ¦ œ GZ  ª ² ¦ œ \ G&%'%&%! ª ² ¦ œ FA«o·G  Ž9²› and
¬ œ ¬ œ ¬ œ `G Ž `G
and are es-

¢ ² ¦ Œ¸< ² ¦ Z " ² ¦ \ "'… %&%'%! ² ¦ "FA«o­ a G ,N± a G

timated. Note that (2.3c)
X G IEQ
simplifies
 _ G ,N `G ,N a G P ¢ o ", ˜
to £
 < 
 N
, d
 
 S
i «
A and
²?" . To find the unknown sequence  ,… we use the `G a G H G QY!Q Z[E\ G , implying that
H G IQ H G QY(Q Z G ,
H G II ƒ
,N P ,  "· , and 8n
G
metric combining G of the estimated canonical coordinates HG IQ H G QY(Q Z H G IP Q . Also, (2.5) reduces to „ M • HG QQ • K • HG QQƒ
 ² ¦ and  ² ¦" , i.e. minimize the following cost func- H G IP Q H G IY(I Z HG IQ •  • H9G II • K • H9G I@IŃ HG IQ H G QY(Q Z H G IP Q • , which can be
tion:
viewed as estimated geometric signal-to-noise ratio, see [18].
X¹!º » ƒ Gy \ Y(Z€»¾ ½ ž ªG ² ¦ œ ˆ ƒ ¬ G ² ¦ œ ˆ \
² ¦"M ˆŠ‰ … <¼ ²  ‹@FA    " žž  Define
Z ‰ Z žž ž G G
Ç6" Ž È"ÉdÊS: (5.1)
see [15]. Interestingly, the above cost function is equal to G ¢ Y(Z G D<  G "P 8 G Y!Z  G
BA Y!Z
the following IRC cost function: whereG Ž GÈ"ÉdÊ Y(TZ  G  Ž. GH IY(I Z G (
Y Z G 
 "
 H  G @
I (
Y I Z
>  G "P 8 " D< MG P "FA "P
X6¿ÁÀ º »¾ ½ ƒ G G (Y Z ƒ G H 
Q Z[EQ \ a G `G Y(Z _ G H 
I (
Y I 
Z 
[ \
² "$ ‰ <   ² 6BA P 8 ² <   ² 9"FAF  ,N ,NP G
is the (estimated) weighted
Z least squares (WLS) beamformer. Thus, Ç6" can be viewed
as G an WLSG estimate of G the basis function matrix CG  . Since 6. REFERENCES
Ž È"ÉdÊS "b , Ž È"ÉdÊS is a left inverse of  . [1] A.J. Paulraj and C.B. Papadias, “Space-time process-
… procedure described in the previous
The second iterative ing for wireless communications,” IEEE Signal Pro-
G
section simplifies as follows: first fix  and compute ÇË cessing Mag., vol. 14, no. 6, pp. 49–83, Nov. 1997.
Ç6 • Ì " •using (5.1).Ì Then, fix Ç and find  that maximizes [2] G.J. Foschini, “Layered space-time architecture for
M Y(Z , where MM;@LK  x>d< Ç ƒ C"BAN>d< Ç ƒ C"BAP . wireless communication in a fading environment when
using multi-element antennas,” Bell Labs Tech. J., vol.
Iterate as long as there is a significant increase in (3.1) be- 1, no. 2, pp. 41–59, 1996.
tween consecutive steps. [3] A. Maleki-Tehrani, B. Hassibi, and J.M. Cioffi,
A sub-optimal second step may be to simply project Ç “Adaptive equalization of multiple-input multiple-
output (MIMO) channels,” in Proc. ICC, New Or-
onto finite alphabet to demodulate the unknown symbols Ì  ; leans, LA, June 2000, pp. 1670–1674.
this would effectively minimize the diagonal entries of " [4] N. Al-Dhahir and A.H. Sayed, “The finite-length
and therefore its trace, but not necessarily the determinant multi-input multi-output MMSE-DFE,” IEEE Trans.
(for ,“*WD this is optimal, see the following section). Signal Processing, vol. 48, pp. 2921–2936, Oct. 2000.
[5] L.L. Scharf and J.K. Thomas, “Wiener filters in canon-
ical coordinates for transform coding, filtering, and
5.1.1. Single Source quantizing,” IEEE Trans. Signal Processing, vol. 46,
In the case of a single source, we have ,¨µ*+´ and the pp. 647–654, Mar. 1998.
[6] T.W. Anderson, An Introduction to Multivariate Sta-
basis function matrix degenerates to aG row vector C 
< ®¯¾ $"U®¯{z’Vd'%'%&%!E®¯HG  QQ"FA .G Then, H IQ  GÍ I¾ Q ;@SK  d> tistical Analysis, Wiley, New York, 2nd edition, 1984.
[7] F. Pipon, P. Chevalier, P. Vila, and J.J. Monot, “Joint
QQ  • •
Ä ½ ‰ Z ®¯EP and  , D@LK >ÎÄ ½ ‰ Z ®T" \ , spatial and temporal equalization for channels with ISI
and CCI,” in Proc. 1st IEEE SP Workshop SPAWC,
„ the G concentrated G G GÍ function in (2.5) becomes
GÍ likelihood
M , QQ K’ƒ , QQ ƒ „ IP Q H IY(I Z IQ  . After the monotonic trans-
and Paris, France, Apr. 1997, pp. 309–312.
[8] M.A. Lagunas, A.I. Perez-Neira, and J. Vidal, “Opti-
formation  LK § „ , the concentrated GÍ IEP Q H G IY!I Z GÍ likelihood
mal array combiner for sequence detectors,” in Proc.
Z E
I Q G QQ function
further simplifies to r K , . This con- ICASSP, Seattle, WA, May 1998, pp. 3341–3344.
[9] D. Giancola et al., “Space-time processing for time
centrated likelihood function can be maximized using the varying co-channel interference rejection and channel
iterative procedure from the previous section. For fixed  , estimation in GSM/DCS systems,” in Proc. 49th Veh.
the first step consists of computing [see (5.1)] Technol. Conf., Houston, TX, May 1999, pp. 1317–
1323.
Ï  Ï G  "MD G, QdQ K GÍ IP Q H G IY(I Z GÍ IQ > GÍ IP Q H G IY(I Z Åd (5.2) [10] D.D. Falconer and F.R. Magee Jr., “Adaptive channel
memory truncation for maximum likelihood sequence
estimation,” Bell Syst. Tech. J., vol. 52, pp. 1541–
 ¾ second step, fix Ï 
for Ð$Ez›'%&%'%( . Then, in the
1562, Nov. 1973.
Ì  • ƒ ®¯ • \
and minimize MMDLK §> Ä ½ ‰ Z Ï 
[11] A. Ranheim, “A decoupled approach to adaptive sig-
with nal separation using an antenna array,” IEEE Trans.
respect to  . If  contains unknown symbols and each time Veh. Technol., vol. 48, pp. 676–682, May 1999.
[12] B.G. Agee, “The least-squares CMA: a new technique
snapshot corresponds to a different symbol, each term in the for rapid correction of constant modulus signals,” in
above summation can be minimized separately; we can view Proc. ICASSP, Tokyo, Japan, Apr. 1986, pp. 953–956.
the second step as projection onto finite alphabet. In this [13] P. Stoica and M. Viberg, “Maximum likelihood pa-
case, the above iteration is identical to the recently proposed rameter and rank estimation in reduced-rank multivari-
ate linear regressions,” IEEE Trans. Signal Process-
decoupled weighted iterative least squares with projection ing, vol. 44, no. 12, pp. 3069–3078, Dec. 1996.
(DW-ILSP) [11]. [14] A. Dogandžić and A. Nehorai, “Space-time fading
In the case when the signal ®¯" is modeled only by channel estimation in unknown spatially correlated
using a constant modulus property, we can choose C noise,” in Proc. 37th Allerton Conf. Commun., Contr.,

< Ñ ™’Ò < ÓTÔ"@LFAFÑ ™5Ò < ÓTÔ"{zTFAF' %'%&%!Ñ ™’Ò < ÓTÔ" BA3A andG thus
Comput., Monticello, IL, Sept. 1999, pp. 948–957.
µ< Ô"@=dÔ"{zT&%'%'%Ô" BA o . Note that here , QQ D , and
[15] A. Dogandžić and A. Nehorai, “Finite-length multi-
input multi-output adaptive equalization using canon-
the first step of G the iteration consists of computing Ï 6
ÏG . GÍ IP Q H IY(I Z EK GÍ IEP Q H G IY!I Z GÍ IQ for ¨ÕTEz›'%'%&%?  .
ical correlation analysis,” submitted, 2000.
[16] G.E. Bottomley and K. Jamal, “Adaptive arrays and
Then,G in the second step, fix Ï E"D$Uz’&%'%'%? and com-
 MLSE equalization,” in Proc. 45th IEEE Veh. Technol.

pute R<pÖ Ï @LLÖ ¾ Ï {zT&%'%'%(SÖ Ï  FA«o , which minimizes
Conf., Chicago, IL, July 1995, pp. 50–54.
[17] M. Stojanovic, J. Catipovic, and J.G. Proakis, “Adap-
Ì • •
MM;@L¾K  T> Ä ½ ‰ Z Ï  G ƒ Ñ ™’Ò < Ô"BA \ , yielding ¾ Ì  "G M tive multichannel combining and equalization for un-
• • •
@SK  ? • > Ä ½ •Š‰ • \ Z Ï  ƒ Ñ ™’Ò < ÔFA \ DLK  ?> Ä ½ ‰ Z Ï  ƒ
derwater acoustic communications,” J. Acoust. Soc.
Am., vol. 94, pp. 1621–1631, Sept. 1993.
Ï EK Ï   , which is an estimated mean-squared ampli- [18] J.M. Cioffi, P.H. Algoet, and P.S. Chow, “Combined
tude fluctuation of the beamformer’s output Ï  . The ob- equalization and coding with finite-length decision
feedback equalization,” in Proc. Globecom Conf., San
tained algorithm is identical to the least-squares constant Diego, CA, Dec. 1990, pp. 1664–1668.
modulus algorithm (LSCMA) in [12].