Professional Documents
Culture Documents
6, JUNE 2009
493
AbstractThere are a number of approaches to estimating carrier frequency offset (CFO) that causes inter-carrier interference
(ICI) in orthogonal frequency division multiplexing (OFDM) systems: self-cancelation method, the extended Kalman filter (EKF),
filter (HF), etc. In particular, the HF is of inparticle filter,
terest because prior statistical noise information is not necessarily
required in its application. Cost reference particle filter (CRPF),
newly developed in the particle filtering framework, has the same
feature as HF; it also does not require the prior noise information
of the state and the measurement equation. In this letter, we compare and analyze the performances of two similar methods. The
simulation results show that CRPF outperforms HF, particularly
when the bit energy to noise ratio of the measurement is low. Therefore, CRPF is very effective and robust, especially when the noise
statistics are unknown with a low bit energy to noise ratio.
information of the state and the measurement equations. References, [4] and [5] show that HF outperforms the EKF when
the noise distributions are unknown. The cost reference particle
filter (CRPF), recently developed in the particle filtering framework [6], has the same feature as HF; the prior noise distributions are not necessarily required when we apply it. In this letter,
we compare and analyze the performances of two methods that
have the common feature, for estimating CFO in OFDM systems. Simulation results show that CRPF outperforms HF, eswhere
is
pecially when the bit energy to noise ratio (
the bit energy, and
is the noise power spectral density.) is
low. Therefore, CRPF is effective and robust, particularly when
.
the noise information is not available with a low level of
I. INTRODUCTION
HE ORTHOGONAL frequency-division multiplexing
(OFDM) scheme is popularly employed for modern
wide-band digital communications such as digital television,
wireless networking, broadband internet access, etc. due to its
various advantages. However, it also has a few disadvantages,
e.g., carrier frequency offset (CFO), sensitivity to frequency
synchronization, etc. In this letter, we focus on the first one,
i.e., CFO which causes inter-carrier interference (ICI) in the
system.
For many problems in engineering or statistical science, we
can model and describe them by the dynamic state system (DSS)
model where the states of interest are correlated in time or space.
Based on the DSS model, there are numerous approaches to estimating the states of interest dynamically with time or space. The
CFO estimation problem also can be well described by the DSS
model, and a number of approaches are proposed to estimate
CFO to combat ICI [1][3]. The Kalman filtering, specifically
the extended Kalman filtering (EKF) is an effective and popular
method to combat ICI in the literature. Particle filtering and
filter (HF) are also employed besides the EFK [2], [3]. In particular, HF is of interest because we can apply it without the noise
Manuscript received January 21, 2009; revised February 19, 2009. First published March 16, 2009. Current version published April 24, 2009. The associate
editor coordinating the review of this manuscript and approving it for publication was Dr. Z. Jane Wang.
The authors are with the Department of Electronic Engineering, Sogang University, Seoul, Korea (e-mail: jaechan@gmail.com; dhong@sogang.ac.kr).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LSP.2009.2017571
(1)
is a symbol, and is the number of sub-carriers.
where
The cyclic prefix is added up to the signal to mitigate the intersymbol interference, and removed at the receiver. Then the
received signal is expressed in the time domain as follows:
(2)
where is a normalized CFO (meaning the relative offset from a
is the -tap channel
carrier frequency);
is unknown additive noise. We want
impulse response; and
to estimate based on the received signals, and then decode the
symbols.
The dynamic state system (DSS) that describes the hidden
state and observed measurement with additive noise processes of and at time is expressed as follows:
(3)
(4)
where and are the state transition and the observation function, respectively, which are known. Then, the corresponding
DSS and the measurement equations for the problem can be expressed as follows:
(5)
(6)
494
Filtering
(8)
,
,
, and
are the weight parameters that
where
is the number of total time
are positive definite matrices;
denotes the vector norm; and
implies
steps;
. Because the state equation does not have the additive process noise in (5), (8) is modified for the problem as
follows:
(9)
(11)
(12)
The cost function in CRPF, which is recursive additive
structure and corresponds to the weight in standard
particle filtering (SPF) is defined as:
where
is the forgetting factor which makes it possible to adaptively change the
amount of contributions of past particles in evaluating cost
is the incremental cost function which
function, and
given . The cost
indicates the accuracy of the estimate of
function is a measure of estimate quality like the weight
as in SPF. Similarly to SPF, the cost-based random measure
is represented by a set of particles and associated costs as:
where
stands for
,
is the particle index, and
is the number of particles. Besides the cost function, the risk function is defined in CRPF
as:
where
; a good choice of the risk func. The risk function measures
tion is
the adequacy of the estimate
given the observation ,
;
and is also a prediction of the cost increment
.
the cost increment can be computed by
Based on these definitions, the sequential algorithm proceeds
with time, recursively repeating the steps of risk evaluation,
resampling, particle propagation, and updating the cost.
The steps of CRPF algorithm are summarized in Table I.
CRPF can be easily adopted to the estimation of CFO based
is an identity function in the problem; and
on (5) and (6);
. The rest of the
steps are straightforward following Table I.
2) For
Compute
Compute
where denotes complex conjugate, and
.
Update
the
estimate
.
495
TABLE I
COST REFERENCE PARTICLE FILTER ALGORITHM
does not vary with time, which makes it easily tractable to compute that minimizes the squared errors even though it does not
guarantee a satisfactory performance. We define the squared errors up to the time as
(14)
is a weighting factor. If we express the complexwhere
number measurement as
, then the that minican be computed from the following equation:
mizes
(15)
Fig. 2. BER performances in linear scale.
E. Particle Filtering
We may also apply standard particle filter (SPF) [8] to the
problem, and compare its performance with those of previously
mentioned methods even though SPF requires noise information. However, there is a technical problem to apply SPF if we
use the prior density (
, where denotes the particle
index) as the proposal distribution. In the DSS equation of the
problem, the state is assumed to be a deterministic constant,
and does not vary with time; consequently, the prior density is
the dirac delta function of
. Therefore, identical particles are always generated, and do not converge to the true state
once particles initially start from a wrong state (e.g., 0.1 as in
the simulations of the other methods in later section.) Therefore, in order to apply SPF, a proposal distribution that depends
on not only the previous state but also the current measurement
has to be used, other than the prior density. However, it is often
not easily tractable to use a proposal distribution other than the
prior density. Therefore, we do not consider SPF in this letter.
IV. SIMULATIONS
By computer simulations, we illustrate the performance comparison of the methods, especially CRPF and HF. We also compare the performances of the EKF and RLS methods against
those of the previously mentioned techniques. We consider a
single antenna, 64-subcarrier OFDM system; the binary phase
shift keying (BPSK) scheme is employed for the simulation.
Various additive Gaussian noise power levels are applied to the
. The normeasurement equation for different ratios of
malized CFO is set to 0.35 that we estimate. The parameters
;
;
;
for HF are: the initial state estimate
; and
. The parameters for CRPF are: the number
;
;
;
; and the
of particles is 500;
initial variance for the Gaussian propagation density
is
0.0053. The 500 identical, initial estimates are generated for
CRPF. A similar result is obtained when we apply 200 particles for CRPF, and it may not considerably affect the perfor-
496
30 dB in the realization.