Академический Документы
Профессиональный Документы
Культура Документы
1:N
M:1 1:M
m(i)
r(i)
F
H
MF
M
h) is a diagonal
matrix with
MF
M
h = [H(0) H(2/M) H(2(M
1)/M)]
T
on its diagonal, i.e., the channel M-length FFT. The
pre-coding M N matrix is V = F
M
[I
NN
0]
T
F
H
N
. No-
tice that (3) is similar to a CP-OFDM system where M-length
pre-coded blocks of data symbols v(i) = Vs(i) are trans-
mitted over M narrowband sub-channels. Since the channel
is of order L, matrix may have, at most, L zero-diagonal
entries. However, the remaining (at least M-L=N) nonzero
entries guarantee the recovery of the symbols in s(i).
The mixing problem in (3) is much simpler than that in
(2). The number of unknowns is reduced to the M elements
on the diagonal of , while there were M N unknown el-
ements on A. This is the distinct advantage of the model in
(3). On the other hand, the search for a diagonal demixing
matrix will provide the original arrangement of the sources.
Also, the recovered symbols will differ from the true ones by
a complex scalar only (the same for all the sources).
3. BLIND ADAPTIVE CHANNEL ESTIMATION
In this section we develop an efcient blind adaptive algo-
rithm based on Blind Source Separation (BBS) to solve the
channel identication problem. The objective of BSS is to
obtain an output vector
y(i) = [y
0
y
N
]
T
= Bx(i) = BVs(i) +BF
M
n(i),
(4)
that yield the sources s(i). Let us dene B = V
D = V
H
D,
where D is M M diagonal matrix and ()
denotes matrix
pseudoinverse. Notice that V
= V
H
, therefore y(i) yields
the sources if D =
1
. Hence, by solving the BSS problem,
we will nd the inverse of the the channel M-length FFT, i.e.,
D = diag(
h
M
), where
h
M
= [1/H(0) 1/H(2/M)
1/H(2(M 1)/M)]
T
.
Many different methods have been proposed in the liter-
ature to solve the BBS problem [6]. We consider here ap-
proaches based on Independent Component Analysis (ICA)
and gradient descent for minimizing a suitable cost function.
The natural gradient (NG) [10] represents the steepest descent
direction of the cost function dened on a Riemannian space:
L(B) = G
1
L(B), (5)
where L(B) is the conventional Euclidean gradient,
L(B)
is the natural gradient, and G
1
is the inverse of the metric
tensor. We are interested in extending this approach to the
overdetermined BSS problem in (4). First, we introduce a Lie
group structure in our parameter space and the Riemannian
metric. Then, we derive the ICA based channel estimation
algorithm.
3.1. Lie group
We construct a Lie group structure on an subset Gl
(N, M)
B Mat(N, M)[B = V
D of Gl(N, M) B
Mat(N, M)[rank(B) = min(N, M), where matrix V is a
xed known matrix, and D is a diagonal matrix. The group
multiplication and inverse operations are given by
B
1
B
2
= V
(D
1
D
2
),
B
= V
D
1
, (6)
where B
i
= V
D
i
, is the multiplication operator, and ()
is
the inverse operator. The neutral element of the group is E =
V
I
M
. The inverse operation satises B
B = BB
= E.
Now, we have to dene a Riemannian or natural metric
in our space of interest, i.e., the inner product on the tangent
space T
B
of Gl
(N, M)
denes an onto-mapping: B B W. The multiplica-
tion transformation maps a tangent vector Xat Bto a tangent
vector X Wat B W. Therefore, we can dene a natural
metric on Gl
, Y B
)
euc
BB
= X B
, Y B
)
euc
E
= X B
, Y B
)
euc
B
(8)
3.2. Natural gradient
For a risk function l(B) dened on Gl
, SG
H
B
)
euc
B
, (10)
where
X B
, SG
H
B
)
euc
B
= tr((X B
)(SG
H
B
)
H
)
= tr(V
H
(D
X
D
1
)(V
H
(D
S
G
H
D
1
))
H
)
= tr(V
H
D
X
D
1
D
H
G
1
D
H
S
(V
H
)
H
), (11)
and taking denition (9) into account it follows
X, l(B))
euc
B
= X, S)
euc
B
= tr(XS
H
)
= tr(V
H
D
X
(V
H
D
S
)
H
)
= tr(V
H
D
X
D
H
S
(V
H
)
H
). (12)
Hence, we have D
1
D
H
G
1
= I
M
, and
l(B) = l(B)G
H
= l(B)D
H
D. (13)
3.3. Adaptive solution
Our target is to make the components of y(i) in (4) as mutu-
ally independent as possible. To this end, we employ a cost
function L(B) = El(B) based on the Kullback-Liebler di-
vergence [10]:
l(B) = H(y, B) +
N
k=1
H(y
k
, B), (14)
where H(y, B) =
p
k
(y
k
, B) log p
k
(y
k
, B)dy
k
, and p(y, B), p
k
(y
k
, B) are
the joint probability density function (pdf) of y and marginal
pdf of y
k
(k = 1, , M), respectively. The output signals y
are mutually independent iff L(B) = 0.
The risk function l(B) can be simplied for the overde-
termined BSS problem to [12]
l(B) = log [ det (BE
H
)[
N
k=1
log p
k
(y
k
, B), (15)
where det (BE
H
) is the determinant of matrix BE
H
, and E
is the neutral element of the Lie group Gl
k=1
log p
k
(y
k
, V
H
D).
(16)
We want to obtain the conventional gradient for this function
with respect to D. For the rst term, we have
d
dD
log [ det (V
H
D(V
H
)
H
)[
= D
H
. (17)
For the second term, let f =
N
k=1
log p
k
(y
k
, V
H
D), then
df
dD
l,l
=
N
k=1
k
(y
k
)(V
H
k,l
)
l
, (18)
where D
l,l
is the (l, l)-th entry of the diagonal matrix D,
(V
H
k,l
)
k
(y
k
,V
H
D)
p
k
(y
k
,V
H
D)
y
k
|y
k
|
is a function of
the sources estimates. p
k
(y
k
, V
H
D) denotes differentiation
of p
k
(y
k
, V
H
D) with respect to y
k
. We can now write the
complex gradient of (15) with respect to Das
l(B) = D
H
+, (19)
where is the M M diagonal matrix that results from
taking the main diagonal of matrix V(y)x
H
, and (y) =
[
1
(y
1
)
N
(y
N
)]
T
. Application of (13) and (19) leads to
the natural gradient in [7]
l(B) = I +D
H
D. (20)
This natural gradient can be further simplied as follows. First,
x = Vs, neglecting the effect of the noise in (3). Also, the
outputs y are intended to be an estimation of the sources s.
And third, D
1
is an estimation of . Therefore, we can nd
taking the main diagonal of matrix V(y)y
H
V
H
D
H
,
and the term D(i)
H
in (20) equals the diagonal matrix that
results fromtaking the main diagonal of matrix V(y)y
H
V
H
.
We propose to use this natural gradient for the adaptive es-
timation of the inverse of the channels transfer function
h
M
,
i.e., the diagonal of D. The algorithm takes the following
form
h
M
(i + 1) =
h
M
(i) E1 p(i)
h
M
(i). (21)
where p(i) is a vector equal to the main diagonal of matrix
V(y(i))y
H
(i)V
H
, stands for the Hadarmard (point-wise)
product, and controls the learning rate.
In the BSS problem in (4) it is easy to show that simple
second order statistics are useful to ensure separation, there-
fore we propose to use (y(i)) = y(i). Then, each entry
of p(i) is the squared absolute value of each entry of vector
Vy(i). Also, between two consecutive steps of the algorithm,
we propose to perform a denoising process [13] by taking ad-
vantage of the fact that the channel is FIR of order at most
L. This process involves the application of the IFFT to the
estimated channels transfer function to remove the spurious
taps located beyond tap number L + 1 before switching back
to the frequency domain to continue with next step of the al-
gorithm. Finally, the proposed selection (y(i)) = y(i) in-
volves that the algorithm will only accurately estimate min-
imum phase FIR channels. If the channel is non-minimum
phase, only the absolute value of the channel will be accu-
rately blindly estimated. Pilot symbols may be used to (at
least partially) provide the algorithm with information about
the channel phase. The algorithm is able to estimate accu-
rately non-minimum phase channel if it has some initial infor-
mation about the phase. Exploring our method more deeply
and improving the performance for non-minimumphase chan-
nels are major directions for future work.
0 500 1000 1500 2000
25
20
15
10
5
0
Number of symbols
R
e
l
a
t
i
v
e
c
h
a
n
n
e
l
e
s
t
i
m
a
t
i
o
n
e
r
r
o
r
p
o
w
e
r
(
d
B
)
NG
Pilots
SVD
RLS
Fig. 2. Performance of the proposed NG based algorithm,
RLS [4], direct SVD and the estimated channel using the rst
2 blocks as pilots. SNR=15 dB. No fading.
4. EXPERIMENTAL RESULTS
We present now several simulation examples. We consider a
4-QAM transmitter in the practical context of HIPERLAN/2
(HL2), a multicarrier wireless system operating in the 5-GHz
band with N = 64 carriers and a zero padding of length
L = 16. In HL2, each frame contains 500 blocks, and among
them the rst two are pilots. The channel selected complies
with channel model A in HL2 [14] with high frequency se-
lectivity. In addition to the proposed natural gradient (NG)
based adaptive channel estimation, we also simulate the RLS
and direct SVD approaches, as in [4]. Once the estimated
channel impulse response
h is available, the complex scalar
ambiguity is solved by = min
|h
h|
2
. For the pro-
posed algorithm we select a normalized value = 1/(p
H
p).
The following results are an average of 100 simulation runs.
In every run the channel, noise and signals are generated ran-
domly, and every entry of
h
M
(0) in (21) is set to one.
Fig. 2 presents the performance of the algorithms under a
non-fading channel for SNR=15 dB. The relative channel esti-
mation error power E|h
h(i)|
2
/|h|
2
in dB is plotted
versus the number of OFDM symbols received. The parame-
ters for the RLS and direct SVD methods in [4] are selected to
assure the same steady state error for all schemes. We observe
that the proposed algorithm performs the same or slightly bet-
ter than the direct SVD method but at a much lower computa-
tional complexity. The convergence speed of the RLS method
is remarkably slower.
Depicted in Fig. 3 we present the bit error rate (BER)
of the proposed NG based algorithm, the RLS and the di-
rect SVD technique. We also include a scheme with perfect
channel knowledge and the estimated channel using the rst
5 10 15 20 25 30
10
4
10
3
10
2
10
1
10
0
SNR (dB)
B
E
R
NG
Exact channel
SVD
Pilots
RLS
Fig. 3. Averaged BERperformance of the proposed NGbased
algorithm, RLS [4], direct SVD, the estimated channel using
the rst 2 blocks as pilots and exact channel.
2 blocks as pilots. A zero forcing detector was implemented
after 1000 blocks of symbols. We observe an indistinguish-
able performance of the adaptive schemes as compared to the
one with perfect channel knowledge.
We examine now a more realistic scenario involving a
fading channel (model A, receiver speed 3 m/sec). Fig. 4
presents the performance of the algorithms for SNR=15 dB
along the HL2 frame. As we can see, the proposed NG based
algorithmis able to reduce the estimation error along the frame
with a performance comparable as the direct SVD method,
while the convergence speed of the RLS method is lower.
5. REFERENCES
[1] A. Scaglione, G.B. Giannakis, and S. Barbarossa, Redun-
dant lterbank precoders and equalizers - Part I: Unication
and optimal designs - Part II: Blind channel estimation, syn-
chronization and direct equalization, IEEE Trans. on Signal
Processing, vol. 47, pp. 19882022, 1999.
[2] B. Muquet, A. Wang, G.B. Giannakis, M. Courville, and Pierre
Duhamel, Cyclic prexing or zero padding for wireless mul-
ticarrier transmissions?, IEEE Transactions on Communica-
tions, vol. 50, no. 12, pp. 21362148, 2002.
[3] Bertrand Muquet, Marc de Courville, and Pierre Duhamel,
Subspace-based blind and semi-blind channel estimation for
OFDM systems, IEEE Trans. on Signal Processing, vol. 50,
no. 7, pp. 16991712, July 2002.
[4] X.G. Doukopoulos and G.V. Moustakides, Blind adaptive
channel estimation in OFDM systems, IEEE Transactions on
Wireless Communications, vol. 5, no. 7, pp. 17161725, July
2006.
[5] A. Hyv arinen, J. Karhunen, and E. Oja, Independent compo-
nent analysis, John Willey and Sons, 2001.
0 100 200 300 400 500
25
20
15
10
5
0
Number of symbols
R
e
l
a
t
i
v
e
c
h
a
n
n
e
l
e
s
t
i
m
a
t
i
o
n
e
r
r
o
r
p
r
o
w
e
r
(
d
B
)
NG
Pilots
SVD
RLS
Fig. 4. Performance of the proposed NG based algorithm,
RLS [4], direct SVD and the estimated channel using the rst
2 blocks as pilots. SNR=15 dB. Fading channel.
[6] S. Haykin, Ed., Unsupervised Adaptive Filtering, vol. I, Blind
Separation of sources, Wiley Interscience, NY, 2000.
[7] R. Boloix-Tortosa, F.J. Payan-Somet, and J.J. Murillo-Fuentes,
Reduced complexity blind equalization schemes for ZP-
OFDM systems, Signal Processing Advances in Wireless
Communications, 2007. SPAWC 2007. IEEE 8th Workshop on,
pp. 15, June 2007.
[8] Rafael Boloix-Tortosa and Juan Jos e Murillo-Fuentes, Blind
source separation in the adaptive reduction of inter-channel in-
terference for OFDM, Lecture Notes on Computer Sciences
3195: Independent Component Analysis and Blind Signal Sep-
aration. ICA 2004., pp. 11421149, Granada, Sept. 2004.
[9] Jonathan H. Manton, A channel coded CP-OFDM interpreta-
tion of TZ-OFDM systems, in 11th IEEE signal processing
workshop on statistical signal processing, Singapore, august
2001.
[10] S. I. Amari, A. Cichocki, and H. H. Yang, A new learning al-
gorithm for blind signal separation, in Advances in Neural In-
formation Processing Systems, David S. Touretzky, Michael C.
Mozer, and Michael E. Hasselmo, Eds. 1996, vol. 8, pp. 757
763, The MIT Press.
[11] S. I. Amari, Natural gradient works efciently in learning,
Neural Computation, vol. 10, no. 2, pp. 251276, Jan 1998.
[12] L. Zhang, A. Cichocki, and S. Amari, Natural gradient al-
gorithm for blind separation of overdetermined mixture with
additive noise, IEEE Signal Processing Lett., vol. 6, no. 11,
pp. 293295, Nov. 1999.
[13] Jan-Jaap van de Beek, Ove Edfors, and Magnus Sandell, On
channel estimation in OFDM system, in VTC, Chicago, IL,
July 1995, vol. 2, pp. 815819.
[14] ETSI Normalization Commitee, Channel models for HIPER-
LAN/2 in different indoor scenarios, Norme ETSI, Sophia-
Antipolis, France., 1998.