Proc. IEEE Benelux Signal Processing Symposium (SPS-2002), Leuven, Belgium, March 21–22, 2002
COMBINING PER TONE EQUALIZATION AND WINDOWING IN DMT RECEIVERS Gert Cuypers, Koen Vanbleu, Geert Ysebaert, Marc Moonen
K.U.Leuven, Dept. of Electrical Engineering (ESAT), SISTA Kasteelpark Arenberg 10, 3001 Heverlee, Belgium
gert.cuypers@esat.kuleuven.ac.be
ABSTRACT
A novel equalization technique for Discrete multitone (DMT) based modems is proposed which incorporates receiver win- dowing operations. The method is especially useful in an environment dominated by crosstalk and narrowband inter- ference, such as digital subscriber line (DSL). The high side- lobe level of the DFT filters present in classical DMT re- ceivers leads to considerable performance degradation due to intercarrier interference and broadband susceptibility to narrowband disturbers. Window functions can be used to alleviate this, but they are difficult to combine with equal- ization techniques, e.g. per tone equalization [1]. The pre- sented method offers a way to do so without unduly increas- ing the complexity.
1. INTRODUCTION
Discrete multitone (DMT) modulation divides the available bandwidth over several carriers that are individually QAM modulated by means of an IFFT. A cyclic prefix (CP) is added to each symbol. If the channel impulse response length does not exceed the CP length, recovery can be done easily by means of an FFT and equalizing the channel by means of a 1-tap frequency domain equalizer (FEQ) for each tone. In case the channel length exceeds the cyclic prefix, intersymbol interference (ISI) and intercarrier inter- ference (ICI) result ([2], [3]). To avoid the overhead result- ing from long prefixes, this structure is usually preceded by a
-taps time domain equalizer (TEQ) such that the convo- lution of channel and TEQ is sufficiently short.
Alternatively, a per tone equalizer (PTEQ) can be used, which has a
-taps FEQ for each tone individually, hence maximizing the SNR on each tone separately [1]. In the presence of radio frequency interference (RFI) or bridged taps the PTEQ shows a significant performance improve- ment compared to the existing TEQ solution.
Indeed, apart from the channel distortion, DMT transmis- sion is also impaired by white noise, RFI (emerging from AM broadcast and HAM radio) and crosstalk ([4], [5]). It is well-known that window functions can have a beneficial effect on the sidelobe behaviour. However, if the window is
limited to the DMT symbol, the windowing itself leads to the loss of orthogonality [6]. One can either take this into account [7] or avoid it alltogether by extending the win- dow beyond the boundary of the symbol, but still within a cyclic extension. Because these window functions ren- der the demodulator more spectrally contained, they help counter NEXT and RFI.
A PTEQ can behave as a window function, but only to counter disturbances present during equalizer training. To combat RFI emerging after training, a combination of PTEQ and a fixed time window is needed. A previous method to imple- ment this was successful, but had a rather high increase in computational complexity [8].
In section 2 we introduce a new receiver structure, where per tone equalization and windowing operations are com- bined in a computational efficient scheme. Section 3 shows simulation results. Finally, in section 4 conclusions will be given.
2. PER TONE EQUALIZATION WITH WINDOWING
Assume a DMT system with DFT size
, and tones
, carrying the frequency-domain information
at time
. The data model is adopted from [1]. A window of length
! #"is applied to the incoming samples, and a
-taps equalizer w
$is used for each tone
individually.
The received time domain samples are denoted as
%and are grouped in a column vector y
&of length
"' () (*
:
+
&-,/.
% 000
%
&1
132413576
98
2
.
For the derivation assume a window
.:<;=:>
00?0
:
576
>
0?00
: 5 000
:
576
> 8
, with
:576
> ,
, and
: 576> ,
, and for which the following symmetry condition applies:
@
:&A
:&A
6
>&B-C
D@
:A
135
:&A
13576
>&B
, for
EFGH
"
*
. (1) For each tone
, the
-taps per tone equalizer w
$is deter- mined from the following cost function:
IKJL
M
w
NPO@ $
w
B-,IQJL
M
w
N RSUT
TT $
w T
F
y
&
VW&
TTT
X
, with (2)
S02-1
Proc. IEEE Benelux Signal Processing Symposium (SPS-2002), Leuven, Belgium, March 21–22, 2002
F
,
@
B
0?00
. . . . . .
000
@
B
,
@
B-,
:<; :?>
0?00
:
576
> 576
>
5
0?00
6 >
: 5
0?00 :
576
> 5 1 6 >
,
,
N
!
The vector
"@
B
corresponds to a combination of a windowing operation, folding (to length
) and pointwise multiplication with the
$#%row of the DFT matrix. (The bar on w
$has been kept to keep the notation compatible to [1]). From (2), it looks as if a
taps equalizer requires the computation of
complete (windowed) DFTs at the sym- bol rate. However, we will show how the matrix
&can be rewritten in a simpler form. Indeed, apart from the part cor- responding to the window taper, there is a simple relation- ship between the elements in matrix
&, that was exploited in the PTEQ [1]. We will extend this for the tapered parts, such that only one DFT needs to be calculated, and the oth- ers can be derived from this one and some correction terms.
Define the differences of the window as:
'(
, .*);+) >-,,,.)
576
> 8
,
)7A
, S
:<;
E ,
:&A
:&A
6 >
0/
E / "
and, for tone
, the modulated differences as
'(
N , . ) ; ) >
00?0
)
576
>1
576
>
8
. If we define
2 ,
,,,
G2 6 >
. . . . . .
.. . . . . . . .
,,,
, and
M
,'(
N
0?00
-
'( N0?00
. . . . . .
00?0
'
N
000
-
'
N
00?0
@
B
,
then the matrix
&can be rewritten as
& , 2
M
,(3)
Now we can substitute w
$T
by v
$T
, defined as v
$T
,w
$T
0 2
, such that (2) can be rewritten as a function of v
$IQJL
M
v
N O@ $
v
B-,IQJL
M
v
NRS TTT$
v T
M
y
&
V
TTT X
(4)
b
0 µ/2−1b
b
µ−1b
2µ−1b
µN
µ
1
µ
Figure 1: Trapezoidal window
To simplify v
$T
M
y
&, consider the part related to the first
F
elements of v
$T
, and define u
$ ,v
$@3
F
B
. Hence we can write:
$
u T
'
N
00?0
. . . . . .
000
'
N
.*4 5 4 8
y
&6 78 9
:
, (5)
which also defines the matrix
; , < ,,,<
&1 2 13576>=
2
of difference terms
< A , @% A % A 1 B
.
For some windows, (5) can be further simplified, e.g. in case of a trapezoidal window, with:
: A ,@?
A 1 >
5
(/
E / @" B
576 A 6 >
5 " / E / @A
" B
,
which is depicted in figure 1. Note that, because of the con- stant slope at the window edge, all differences
) Aare equal to
" 6 >, and thus the modulated differences are equal to a scaled version of the first
"elements of the
#%row of the DFT-matrix. We can substitute this explicitely, such that (5) becomes:
$
u T
"
@ ?0
" B
000
. . . . . .
00?0
@ 3
" B
<
<
&1
>
.. .
<
&132413576>=
(6)
Instead of computing all these DFTs in full, one can also perform the DFT on the last row only, and derive the others by making use of this one DFT and some correction terms, similar to, but slightly different from (3). For a row
Eand a tone
, we can write that:
BCD"EGF0HJI1KMLNPOQBCD*EGFRKMLS>TULNWVYX0Z\[^]`_^abHcX0Z\[^d`[^]`_^aLS dT
e fMg h
correction term
By recursing this formula, all DFT-outputs at tone
can be replaced by the DFT output of the last row only and
S02-2
Proc. IEEE Benelux Signal Processing Symposium (SPS-2002), Leuven, Belgium, March 21–22, 2002
c a µ
N
N/2−1,1 N/2−1,2T−2
W
N/2−1,2T−1
W
N/2−1,2T−4
W W
0,2T−4
W
N/2−1,2T−2
W W
N/2−1,2T−30,2T−3 0,2T−2
W W
T−2 T−2 µ
y
(k)delay downsample
...
...
...
...
...
...
...
...
... ...
0,T−2 0,T−1
W
W
a+b.c b
0,1
b
W
Ν + ν Ν + ν Ν + ν Ν + ν Ν + ν
Ν + ν
Ν + ν ∆
∆
N−point FFT
∆ ∆
∆
∆
∆
N−point FFT Ν + ν
Ν + ν
Window
0 0
0 Ν + ν
0 0 Ν + ν
∆
∆
∆
∆ Ν + ν
Figure 2: Signal flow graph. Note that
$denotes the
#%element of the equalizator
$for tone
all needed correction terms. Unfortunately, these correc- tion terms are tone-dependent, such that they cannot be pro- cessed for all tones together, as could be done with the PTEQ.
It might be computationally more interesting to break them into common terms (
< A
’s). In matrix form, the new opti- mization criterion can be represented as follows:
IKJL
M
w
N O@ $
w
B-,IQJL
M
w
N RS TTT $
w T
F
y
&
VW&
TTT X
, with (7)
T
I
_^aO
_Ra _ [ a-I
_^aO
_^a dO
_RaO
_^a dI
_RaO
_Ra _ [ a-I
_RaO
_Ra
T "!$#&%('*)
-
T+, "!$#-%.'*)
T/, 0!$%) 12222
3
.
The rows (from bottom to top) correspond to the original windowed DFT output, that has been omitted in the tran- sition from (4) to (5), the DFT corresponding to the last row in (6), and the
A @
WA
B < A
’s needed to compute the other (sliding) DFT’s from (6),
<
,...
<
&132 6 =
,
<
&1 5
,...
<
&132413576 =
. The indices
4are added because the original equation (2) can be viewed as a constrained version of the unconstrained optimization problem of (7). If one denotes the number of rows in matrix
&as
, the following re- lationship holds:
, A @ A B A , A A
.
This implies that the amount of taps needed roughly dou- bles. Nonetheless, the complexity is still lower compared to
the original problem from (2), because the calculation of the
DFTs can be avoided. In comparison to a previous imple- mentation [8], the method is attractive in the most common case that
65 ".
A signal flow graph of this scheme is shown in fig. 2. In summary, for each tone
, the optimal equalization is deter- mined by
7
the
#%output of the windowed DFT of the received symbol
7
the
#%output of the DFT of a vector of difference terms
. <,,,
<
&1 576
>
,,,
8
(completed with zeroes)
7
difference terms
<
,,,
<
13236>=
,
<
&135 ,,,
<
&132413576>=
. As stated before, (7) corresponds to the unconstrained ver- sion of (2). The latter can be derived from the former by premultiplication of the matrix
&by a constraint matrix
8
:
8 , 9
I
2 6 5 ,
I
2 6
O
O
2 6
O
236
I
;:,
i.e.
&and
8 &span the same row space, and optimisa-
tion using either one of these matrices will yield the same result. Obviously, enforcing the constraint will reduce the number of taps needed, at the cost of being stuck with tone- dependent correction terms.
S02-3
Proc. IEEE Benelux Signal Processing Symposium (SPS-2002), Leuven, Belgium, March 21–22, 2002
3 4 5 6 7 8 9 10 11 12 13
0 0.5 1 1.5 2 2.5 x 10
6Number of taps
bit rate [bps]
Trapezoidal window, no RFI PTEQ, no RFI Trapezoidal window, RFI present PTEQ, RFI present
Figure 3: Comparison between PTEQ and the new method in absence and presence of RFI
3. SIMULATION RESULTS
Simulations have been done for an Asymmetric DSL (ADSL) system, using a standard loop T1.601#13 with NEXT from 24 DSL sources, and RFI present as indicated in table 1, ren- dering the energy in a 4312.5Hz ADSL band. Notice that the last RFI interferer was originally located at 1900kHz, but appears at 308kHz due to aliasing. The used tones are from tone 38 up to tone 256. The symbol size
is 512 and the prefix length is
A
.
620 740 800 980 1100 1160 308
-92.2 -90.5 -59.9 -59.6 -95.5 -79.8 -112.7 Table 1: Nominal RFI frequencies [kHz] and power levels [dBm/4312.5Hz band]
The presented method, making use of the compression with the constraint matrix and a window extension
" ,
, is compared with the PTEQ of [1]. Figure 3 shows the bit rate for both equalization techniques, in the absence and pres- ence of RFI interference. As could be expected, in the ab- sence of RFI, the gain is negligible. If RFI is present how- ever, the proposed method performs significantly better than the PTEQ. As the number of taps increases, the difference becomes smaller.
4. CONCLUSIONS
A new method for combining windowing and per tone equal- ization has been presented. The windowing operation re- mains unaltered and is executed prior to the FFT at the re-
ceiver side. A
-taps per tone equalizer is eventually trans- formed into an equivalent
@A A B
-taps equalizer. The inputs for this equalizer are determined from two
-point FFT’s and
A @ A B