5th International OFDM-Workshop 2000, Hamburg 18-1

MIMO Systems in the Subscriber-Line Network

Georg Tauböck and Werner Henkel
Telecommunications Research Center Vienna (FTW)
E-mail: [email protected] [email protected]

Abstract— We consider multi-tone modulation (DMT, Central Office Cabinet

B 9 ? 9
Discrete Multitone, OFDM, Orthogonal Frequency Divi- B ; ? ;
sion Multiplex) over cable bundles and present a compact B4C FEXT ? C
mathematical description of the transmission line. Both,
near- and far-end crosstalk (NEXT, FEXT) are taken into B@ > 9 ?A@ > 9
B4@ A? @
account and a new method for removing far-end crosstalk
is derived. NEXT 8:9
Keywords— MIMO, Multiple-Input-Multiple-Output, 8;

FEXT-Cancellation, NEXT-Cancellation, DMT, Discrete

Multitone, xDSL
8=< > 9
=8 <
I. M ULTI -I NPUT /M ULTI -O UTPUT S YSTEMS AND
THEIR MATHEMATICAL DESCRIPTION Fig. 1. FEXT and NEXT in a MIMO system without self-
NEXT; self-FEXT would mean the DE and F!E to be collo-
In this paper, we consider a transmission system,
cated.
which consists of several loops from one dedicated point
to another (e.g., from the central office to the cabinet).
We assume that these loops are used for bidirectional above operations can be described using matrices1 :
data transmission. Let denote the number of loops G (IH J K LMJON JP LQJSR J P TJON U J&V T
=W (2)
which are used for transmission from point A to point B
and
the number of loops which are used in the oppo- Here, W denotes the 6 -dimensional symbol vector to be
site direction (from B to A). This MIMO channel can be transmitted, G the 6 -dimensional symbol vector at the
T
described using a matrix       and a matrix input of the decision device, V the non-linear operation
      of time-discrete impulse responses. of conjugate complex extension (it transforms an 6 -
Let  and ! denote the transmitted symbol sequences dimensional vector into a X6 -dimensional vector) and
N
from point A to point B on loop " and from point B to the AX6Y[Z\AX6] DFT matrix; the AX6 , 7^_Z\AX6]
point A on loop # . Then the received symbol sequence
$&%
matrix P T describes the addition of the guard interval,
on loop ' is given by (c.f., Fig. 1) the AX6 , 7^Z`AX6 , 7^ matrix2 R represents the chan-
.

.
 nel impulse response (note that transmit filter, receive
$&%)( %&%+* %-, % * , % * 5 filter, and further processing are considered as part of
         (1)
%
0
 /
12 3 0  142 3 the channel), the AX6][Z\AX6 , 7^ matrix P L removes
FEXT NEXT the guard interval, the 6aZbAX6Y matrix K L removes
the second 6 components of a X6 -dimensional vector
For a moment, however, let us concentrate on only (inversion of the conjugate complex extension), and the
one loop with a DMT transmission over it. The data diagonal 6cZd6 matrix H performs frequency-domain
is divided into blocks of length 6 and each block equalization. It is also assumed that 7 is chosen greater
(conjugate-complex extended) is passed through an than the length of the channel impulse response (i.e.,
IFFT. Some redundancy is added, i.e., a guard interval where it significantly differs from zero). Then it can be
of length 7 is prepended; through a transmit filter and shown [1] that P L J&R JeP T represents a cyclic convolu-
the channel, the data reaches the receiver, where after a tion and therefore N JAP L JR J:P T JN U is a AX6YZ`AX6]
receive filter and further processing, the redundancy is diagonal matrix. Furthermore, it can be easily seen that
eliminated. An FFT and a frequency domain equaliza- f
This equation describes the transmission of one block which is
tion (which simply performs scalar multiplications with part
gih of a sequence of transmitted blocks
complex factors) is carried out and a decision device re- is a convolution matrix: it consists of the elements of the chan-
covers the original data. As shown in [1], many of the nel impulse response
5th International OFDM-Workshop 2000, Hamburg 18-2

K LjJkN JlP LmJkR JlP TnJkN U J:V T (Io J

=W W with an 6pZm6 II. E QUALIZATION OF DMT-MIMO S YSTEMS
o
diagonal matrix . In this section, we derive a new DMT-MIMO system
Now, we return to our MIMO system with a DMT equalization method, or, to formulate it in another way,
modulation scheme on each loop, all having the same we present a method for removing FEXT.
parameters (especially, 6 and 7 ). We further neglect First of all, observe that there does not occur FEXT if
NEXT, taking resort to various techniques to remove and only if the system matrices | =t are diagonal for all
it (e.g., frequency division for the multiplexing of up- (–z 5e5e5
t y y46 . This becomes obvious from comparing
stream and downstream data). Applying3  % (qP T J Eqn. (5) with Eqn. (1). Some processing is necessary
N U JiV T
=W
%
and r % (IK L JN J!P L J$ % to (1) and using to obtain such diagonal system matrices.
the linearity of the operations, we get4 Up to now, there exist proposals to remove FEXT
in the literature [1], which are based on the inversion
( K L JON J P L JSR %% JP T JSN U JOV T % ,
r % =W of the matrices | =t , either in the transmitter or in
 the receiver (this leads to the desired diagonal unit ma-
.
, K L JON J P L J R % JP T JSN U JeV T
 =W  trix). However, these methods require that all matrices
%
s / |
=t are invertible, which is not guaranteed in general.

. Bridge taps, e.g., lead to zeros in the transmit func-
( o % J
 W  5 (3) tion at some frequencies. Inversion may then be prob-
s
lematic. Furthermore, inversion at the transmitter may
cause stronger disturbances of loops outside the MIMO
Let r % =t and Wuv=t denote the t -th component (the
bundle and inversion at the receiver may enhance noise.
t -th carrier) of the 6 -dimensional vectors r % and W  ,
We, instead, follow another strategy of computing
respectively, and w %  =t the t -th diagonal element of
singular value decompositions (SVDs) [2] of the ma-
the 6xZ-6 diagonal matrix o %  . Then (3) can be equiv-
trices | =t , i.e.5 ,
alently written as
| ( — J ˜ J ™mš ({z
 =t =t =t =tsy t y 5e5e5 y46 (6)
.
% ( % J ({z
r =t w  =t W  =tsy t y 5e5e5 y46 (4)
s with unitary6 Z matrices — =t and ™ =t and real
diagonal Z matrices ˜ =t , whose elements are
or in a more compact way as matrix-vector multiplica- greater than or equal to zero.
tions Let ›e=t denote the -dimensional symbol vector to
be transmitted on the t -th carrier over the whole ca-
( | J (~z ble bundle. We perform multiplications of these vec-
rn=t =t W}=tsy t y 5e5e5 y46 (5)
tors with the matrices ™ =t , the resulting symbol vec-
with tors W}=tœ ( ™ =t J ›&=t are used as input of the
DMT-MIMO system and the output symbol vectors
( JeJeJ ( | J
rn=t€ =rS=t rv‚=tlƒ-y rn=t =t W}=t are then multiplied with the ma-
|
=t€
(
Aw
%
 =t %     y and trices — š =t , which yields symbol vectors
( JeJeJ 5
W„=t€ =W =t W  =t ƒ  ( —jš J
=t€ =t rn=t
( —jš J | J
=t =t W}=t
Note that overlined letters specify matrices and vec-
( — š J | J ™ J
tors that contain transmit functions and signals at one =t =t =t ›e=t
frequency, only, for all loops, whereas underlining spec- ( — š J — J ˜ J ™ š J ™ J
=t =t =t =t =t ›&=t
ifies matrices and vectors with multiple frequencies at ( ˜ J
=t ›e=tžy (7)
one loop.
( š U
In fact, we do not only transmit one block but a sequence of where we used Eqn. (6) and the property Ÿ Ÿ of
blocks, but under certain assumptions (all impulse responses †ˆ‡=‰ ,
ŠŒ‹]sŽŽA‘
a unitary matrix Ÿ . Note that the resulting system matri-
, have approximately the same delay and their lengths ces are the diagonal matrices ˜ =t , eliminating FEXT
are smaller than the length of the guard interval ’ ) there is no inter- ¡[¢[£¥¤u¦
block
“” interference and it suffices to consider only one block denotes the complex conjugated
¡£¥¤§¦ and transposed (also
‡ denotes the symbol vector before frequency-domain equaliza- called
¨ Hermitian transposed) matrix
tion of the • -th loop also called orthonormal
5th International OFDM-Workshop 2000, Hamburg 18-3

completely. After frequency domain equalization (for R EFERENCES

all loops and frequencies, separately), we finally obtain [1] G. Ginis, J.M. Cioffi, “Vectored-DMT: A FEXT cancelling
the transmitted symbol vectors ›=t . DSL System”, submitted to Globecom 2000.
Now we are able to state the algorithm for removing [2] G.H. Golub, C.F. Van Loan, “Matrix Computations”, North
Oxford Academic Publishers Ltd, a subsidiary of Kogan Page Ltd,
FEXT. The following steps have to be carried out:
p. 16, 1986.
1. only once in the startup-phase, as long as the trans- [3] V. Tarokh, N. Seshadri, A.R. Calderbank, “Space-time codes
for high data rate wireless communication: performance criterion
mission properties do not change:
and code construction”, IEEE Transactions on Information The-
(a) determination of the system matrices | =t ; ory, Vol. 44, No. 2, pp. 744-765, March 1998.
(b) calculation of the SVDs of the matrices | =t ,
i.e., | =t ( — =t J ˜ =t J ™jš =t ;
2. during transmission, for each symbol block of the ca-
ble bundle:
(a) in the transmitter multiplications of the symbol
vectors to be transmitted (›e=t ) with the matri-
ces ™ =t , i.e., W}=t ( ™ =t J ›e=t ;
(b) transmission of the resulting symbol vectors
(W_=t ) over the DMT-MIMO system, i.e.,
( | J
rn=t =t W„=t ;
(c) in the receiver multiplications of the received
symbol vectors (rn=t ) with the matrices — š =t ,
i.e.,  =t ( —mš =t J ›&=t ;
(d) frequency domain equalization (for all loops and
frequencies, separately).
Remarks: Compared to other proposals, this method
has the advantage that a singular value decomposition
of a matrix is always well defined. Furthermore, be-
cause of the unitarity of the matrices ™ =t and — š =t ,
the average power of all loops stays the same and there
is neither an increased disturbance of other loops nor
an average noise enhancement. We also want to empha-
size that this concept for FEXT removal can be extended
using a two-dimensional (over frequencies and loops)
bit-loading and power-distribution algorithm and addi-
tional coding in space (loops) / time / frequency may be
applied, too (c.f., space-time coding [3]).

III. C ONCLUSIONS
We presented a general description of symbol trans-
mission over cable bundles including FEXT and NEXT.
We applied it to a DMT transmission scheme; the result
was a compact mathematical formulation which was
based on simple matrix-vector multiplications. Singu-
lar value decompositions of the system matrices were
the key to equalize the FEXT channel. Finally, we mo-
tivated that two-dimensional bit-loading can be applied
and code constructions similar to so-called space-time
codes may be implemented.