The MMSE Channel Estimation Based On DFT For

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

The MMSE Channel Estimation Based on DFT for

OFDM System
Jie Ma, Hua Yu and Shouyin Liu
Dept. of Electronics and Information Engineering
Huazhong Normal University
Wuhan, China
liu_china@hanmail.com

Abstract—The minimum mean-square error (MMSE) channel The channel frequency response (CFR) for the known pilot
estimation has well performance but higher complexity than subcarriers can be estimated by least-square (LS) algorithm [3]
least-square (LS) channel estimation, specially it requires the [4]. The LS estimation is relatively easy since it doesn’t need
channel statistical properties including the channel any channel apriority probability. However, the performance of
autocorrelation matrix and the noise variance. However, it is LS is not very well due to the presence of noise. Minimum
difficult to obtain the channel statistical properties in practice. In mean-square error (MMSE) estimation has well performance
this paper, through the DFT based estimator, in time domain, the by using the channel statistical properties including the channel
noise variance can be estimated, and the channel autocorrelation autocorrelation matrix and the noise variance [5-7]. Moreover,
matrix is also obtained by using the noise suppressed channel
since the MMSE estimation requires a large amount of
impulse response. Then the MMSE channel estimation for
computation, such as matrix inversion, there are many attempts
OFDM system is realized. Simulation results demonstrate that
the performance is better than the DFT based estimator and to reduce the complexity of the MMSE. The computational
closed to the ideal MMSE estimator. complexity of the MMSE channel estimation can be reduced
by means of singular value decomposition in [5]. In [6], a low-
Keywords- orthogonal frequency division multiplexing(OFDM), complexity MMSE estimation is proposed by partitioning off
channel estimation, Discrete Fourier Transform (DFT),minimum channel covariance matrix into some small matrices on basis of
mean-square error(MMSE) coherent bandwidth. However, the channel autocorrelation
matrix and the noise variance are always unknown in practice.
I. INTRODUCTION The Discrete Fourier Transform (DFT) based channel
Orthogonal frequency division multiplexing (OFDM) has estimation (DFT-CE) is a good choice for OFDM system with
attracted a lot of attention because of its high data rate, high the tradeoff between complexity and performance. It has the
spectrum efficiency and robustness against frequency-selective better performance than LS estimator and lower complexity
fading channels [1]. In OFDM system, the entire channel is than MMSE estimator. In DFT-CE, CFR is first estimated by
divided into many parallel narrow subchannels, thereby LS estimator and then transformed through Inverse Discrete
increasing the symbol duration and reducing the Inter-Symbol Fourier Transform (IDFT) to get channel impulse response
Interference (ISI) [2]. It has been adopted in several wireless (CIR) in time domain. Since the signal energy is usually
communication standards such as IEEE 802.11a wireless LAN, concentrated in limited paths having short delay, so only the
IEEE 802.16 wireless MAN, DAB systems. It is also limited paths are reserved and the rest is set to zero. After the
considered as one of the most promising techniques for future appropriate process the CIR can be transformed back to
wireless mobile communication. frequency domain by DFT. Then the performance is improved.
The LS and the MMSE estimator are modified in [8, 9] by
Channel estimation is required for the coherent OFDM limiting the number of channel paths in the time domain. For
system, which has 3-4 dB performance gain more than non- more accurate results, some algorithms are also developed to
coherent OFDM system. Channel estimation methods mainly estimate the number of significant channel paths by estimating
include blind channel estimation and pilot-based channel the noise power in the time domain [10, 11].
estimation. The blind channel estimation methods require a
large amount of date and the convergence rate is very slow. Though the MMSE algorithm has a better performance, it
Hence, it goes against the real-time channel estimation. In the needs autocorrelation matrix of the channel and noise variance.
pilot-based channel estimation methods, the transmitted signal In most researches, it is assumed that the channel statistical
is known at the receiver. There are two modes: the block pilot properties are known. But generally they are unknown in
mode and the comb pilot mode. In the block pilot mode, all the practice. In this paper, the noise variance is estimated in the
subcarriers of an OFDM symbol are dedicated to the known time domain by using the characteristics that signal energy in
pilots. In the comb pilot mode, only a few subcarriers are used time domain is more concentrated, and the channel
for the initial estimation process. In this paper, only the pilot- autocorrelation matrix is estimated by using the denoised
based channel estimation technique is investigated. channel impulse response. Then the noise variance and channel
autocorrelation matrix can be used in the MMSE algorithm.

This work was supported in part of the Project Sponsored by the


Scientific Research Foundation for the Returned Overseas Chinese Scholars,
State Education Ministry

978-1-4244-3693-4/09/$25.00 ©2009 IEEE


X(k)
Data
Pilot CP
S/P Mapper IFFT P/S D/A
Insertion Insertion

Channel
Y(k)
Data Channel CP n(t)
P/S Demapper FFT S/P A/D +
Estimator Removal

Fig.1 OFDM System Baseband Model

II. SYSTEM MODEL III. CHANNEL ESTIMATION

A. OFDM Model A. LS Channel Estimation


Fig.1 shows the OFDM baseband model used in this paper. The LS estimate of the channel can be obtained as
It is assumed that the cyclic prefix (CP) is longer than the
maximum propagation delay. So the orthogonality between
subcarriers and non-intersymbol-interference (ISI) can be H LS = arg{min{(Y − XH LS ) H (Y − XH LS )}} (3)
preserved. The number of subcarriers and multipaths is K and
L in the system, respectively. set
The received signal can be expressed:
∂{(Y − XH LS ) H (Y − XH LS )}
=0 (4)
Y = XH + N (1) ∂H LS

where X is a matrix of size K × K with the elements of the consequently


transmitted signals on its diagonal, Y is the received vector of
size K ×1 , H is a channel frequency response of size K ×1 ,
and N is a vector of independent identically distributed (i.i.d.) Y N
Hˆ LS = =H+ (5)
complex Gaussian noise with zero-mean and variance σ n2 . The X X
noise N is assumed to be uncorrelated with the channel H .
ignore the impact of noise, then
B. Channel Model
It is assumed that the signal is transmitted over a multipath
Hˆ LS = { H ( 0 ) H (1)" H ( K − 1)}
T

Rayleigh fading channel characterized by


T
(6)
⎪⎧ Y ( 0 ) Y (1) Y ( K − 1) ⎪⎫
=⎨ " ⎬
⎪⎩ X ( 0 ) X (1) X ( K − 1) ⎭⎪
L −1
h(τ ) = ∑ α i δ (τ − τ i ) (2)
i =0
Although, LS estimation algorithm is very simple, its
where {α i } are zero-mean complex Gaussian random variables performance is sensitive to the noise. The veracity of the
estimation is reduced at the low SNR.
with a power-delay profile θ (τ i ) = Ceτ i / τ rms , {τ i } are the time
delays of the different path and L is the number of multipaths. B. MMSE Channel Estimation
At the receiver, the synchronization is perfect. It is The MMSE estimate of the channel can be obtained as [8]
assumed that the CP is longer than the channel maximum
excess delay, and that the channel is quasi-stationary (i.e., the
channel does not change within one OFDM symbol duration Hˆ MMSE = FhˆMMSE = FRhY RYY −1Y (7)
but varies from symbol to symbol).
where
RhY = E[hY H ] = Rhh F H X H (8)
RYY = E[YY H ] = XFRhh F H X H + σ n2 I N (9) D. Modified Channel Estimation
substituting (8) (9) into (7), then As we know, the MMSE estimator needs the channel
statistical properties which are unknown in practice. The DFT-
CE can improve the performance through DFT and IDFT
Hˆ MMSE = FRhh F H X H ( XFRhh F H X H + σ n2 I N ) Y (10) based LS algorithm. Meanwhile, in the DFT algorithm, both
−1

the channel autocorrelation matrix and the noise variance can


be estimated. On one hand, the channel energy is concentrated
where Rhh = E[hh H ] is the channel autocorrelation matrix, σ n2 in the time domain, so the hˆ(n) should be noise
is the variance of noise, F = [WKnk ] is the DFT matrix with when n > Lg − 1 , the noise variance can be estimated. On the
nk
1 − j 2π other hand, we can get the channel autocorrelation matrix from
WKnk = e K
.
K the hˆ ' (n) estimated in DFT algorithm. Then the MMSE
estimator can be realized and the performance is improved.
The MMSE channel estimation has well performance but
The whole process can be concluded as follows:
higher complexity. It requires the inversion of a K × K matrix,
which implies a high complexity when K is large. Notice that Step 1, the channel frequency response Hˆ (k ) is estimated
it requires the channel statistical properties including the
by LS estimator.
channel autocorrelation matrix ( Rhh = E[hh H ] ) and noise
variance ( σ n2 ) which is always unknown in the practice Step 2, from (11), the hˆ(n) can be obtained
systems. Step 3, noise estimation
C. DFT-CE
K −1
1 2
DFT-CE is based on the property that the energy of the
channel is very concentrated in the time domain. The channel
σˆ n2 =
K − Lg

n = Lg
hˆ(n) (14)

estimate Hˆ (k ) can be derived in frequency domain based on


LS estimation firstly. Step 4, the channel autocorrelation matrix Rhh is estimated
Then it can be changed to the time domain with DFT ˆ
by h ' (n) in (12).

( )
H

K −1
Rhh = E[hˆ′(n) hˆ′(n) ] (15)
1 ⎡ nk ⎤
hˆ(n) =
K
∑ Hˆ (k ) exp ⎢ j 2π
⎣ K ⎥⎦ (11)
k =0 Step 5, Rhh and σˆ n2 is used in (10) with the MMSE
0 ≤ n ≤ K −1 algorithm which needs the channel autocorrelation matrix and
the noise variance.
The cyclic prefix Lg is always larger than the channel
impulse response L in OFDM systems. The energy is IV. SIMULATION
concentrated in the limited paths, while only noise energy exist In this section, we compare the MSE (mean square error)
in the other paths. In order to restrain the effect of the noise, and SER (symbol error rate) performance of the proposed
the channel impulse response are set to zeros when n ≥ Lg . The methods with the LS estimation, DFT-CE and MMSE
channel impulse response in the paths within the cyclic prefix estimation. In our simulation, the modulation is BPSK, pilot is
is remained. the training mode and the number of the subcarriers is 64. The
multi-path is assumed to L = 9 , each path is independent and
the energy is subjected to exponential distributions. In addition,
⎧⎪ hˆ(n) 0 ≤ n ≤ Lg − 1 the perfect synchronization is assumed. Considering that the
hˆ ' (n) = ⎨ (12)
⎪⎩ 0 otherwise time delay of the path is sampling interval of the OFDM
system, so there is no energy leakage in the simulation.
Then it is transformed to the frequency domain
A. MSE Comparion
The MSE performance of the proposed methods is
1 K −1 ⎡ nk ⎤ compared with the LS estimation, DFT-CE and MMSE
Hˆ ' (k ) = ∑ hˆ' (n) exp ⎢ − j 2π
K k =0 ⎣ K ⎥⎦ (13) estimation is shown in Fig.2. The performance of the LS
estimator is the worst due to the noise. The DFT-CE has the
0 ≤ n ≤ K −1 better performance since it eliminates the effect of the noise in
time domain. The ideal MMSE estimator has the best
The estimate accuracy can be improved and the complexity performance since it has the accurate channel statistical
is not increased, because the IDFT/DFT transforms can be properties. However, the computational complexity of MMSE
implemented with the fast algorithms IFFT/FFT. is high and the channel statistical properties are always
unknown in the practice systems. The performance of the 0
SER Compare
proposed method is closed to the ideal MMSE estimation and 10

better than the LS and DFT-CE, because it uses the channel


autocorrelation matrix and the noise variance which we
estimated in the MMSE algorithm. At high SNR, the
performance is much closer to the ideal MMSE estimator.
Since the effect of the noise is also small to the path within Lg

Symbol Error Rate


at the high SNR, the proposed method is much more accurate. -1
10
Above all, the simulation result is matched with the theory
analysis.

MSE Compare
2 Ls
10
DFT
Ls
mmse
DFT
proposed
1 mmse -2
10 10
proposed 0 5 10 15 20 25 30
SNR in dB
Mean Squared Error

0
10 Fig.3 The channel estimatin SER of the proposed method, the LS estimator,
DFT based estimator and the optimal MMSE estimator.

-1
10
REFERENCES

-2
10 [1] A. C. Bingham, “Multicarrier modulation for data transmission: An idea
whose time has come,” IEEE Comun. Mag., pp. 5–14, April 1990.
[2] M. Faulkner, “The effect of filtering on the performance of OFDM
-3
10 systems,” IEEE Trans. on Vehicular Technology, vol. 49, no. 5,
0 5 10 15 20 25 30
SNR in dB
September 2000.
[3] Y. Qiao, S. Yu, P. Su, and L. Zhang, “Research on an interative
Fig.2 The channel estimatin MSE of the proposed method, the LS estimator, algorithm of ls channel estimation in MIMO OFDM systems,” IEEE
DFT based estimator and the optimal MMSE estimator. Trans., Broadcast., vol. 51, no. 1, pp. 149-153, March 2005.
[4] W. Su and Z. Pan, “Iterative LS channel estimation for OFDM systems
based on transform-domain processing,” in Proc. WICOM. conf., pp.
B. SER Comparion 416-419, September 2007.
The SER performance of the proposed methods, LS [5] O. Edfors, M. Sandell, J.-J. van de Beek, S. Wilson, and P. Börjesson,
“OFDM channel estimation by singular value decomposition,” IEEE
estimator, DFT-CE and MMSE estimator is shown in Fig.3. Trans. Commun., vol. 46, no. 7, pp. 931-939, July 1998.
Because of the effect of the noise, the SER performance of LS
[6] M. Noh, Y. Lee, and Park, “Low complexity LMMSE channel
estimator is the worst. The DFT-CE is better than LS estimator. estimation for OFDM,” IEEE Proc. Commun., vol. 153, no.5, pp.645-
The ideal MMSE estimator has the best performance with the 650, October 2006.
assumption of the channel statistical properties are available. [7] Y. Li, J. Cimini, and N. Sollenberger, “Robust channel estimation for
However, the channel statistical properties are unknown in the OFDM systems with rapid dispersive fading channels,” IEEE Trans.
practice system. The performance of the proposed method is Commun., vol. 46, no. 7, pp. 902-915, July 1998.
better than the DFT based estimator and close to the ideal [8] J.-J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P.O.
MMSE estimator. Börjesson, “On channel estimation in OFDM systems,” in Proc. 45th
IEEE Vehicular Technology Conf., Chicago, pp. 815-819, July 1995.
[9] O. Edfors, M. Sandell, J.-J. van de Beek, S. K. Wilson, and P.O.
V. CONCLUTION Börjesson, “Analysis of DFT-based channel estimation for OFDM,”
Wireless Personal Commun., vol. 12, no. 1, pp. 55-70, January 2000.
In this paper, based on DFT-CE, an improved channel
[10] Y. Kang, K. Kim and H. Park, “Efficient DFT-based channel estimation
estimation method was proposed. The channel autocorrelation for OFDM systems on multipath channels,” IEEE ITE Commun, vol. 1,
matrix and noise variance were estimated, which could be used Issue 2, pp. 197-202, April 2007.
in the MMSE estimator. The simulation results showed that the [11] H. Jian, Y. Shi, “An efficient iterative DFT-based channel estimation
performance of MSE in the proposed method was close to the for MIMO-OFDM systems on multipath channels,” in Proc. ChinaCom.
ideal MMSE estimator and better than the DFT-CE. The SER Conf., pp. 45-49, August 2008.
performance was also improved effectively.

You might also like