Professional Documents
Culture Documents
A Fine Resolution Frequency Estimator Based On
A Fine Resolution Frequency Estimator Based On
A Fine Resolution Frequency Estimator Based On
Abstract—A DFT-based frequency estimator with low compu- An iterative technique can reduce a direct estimator’s accu-
tational complexity is proposed, which is derived by extracting the racy dependency on the frequency offset . It carries out itera-
phase difference from the peak DFT bins of two sub-segments from tions by treating the refined estimate of an earlier iteration as the
input samples. Some of its statistical properties, such as asymptot- coarse estimate of the next one until the convergence [9], [10].
ical unbiasedness and error variance, are derived. Furthermore, an For most of the iterative estimators, only 2 iterations can lead
iterative procedure of frequency estimate is also presented. Simu-
the MSE close to the CRB. The MSE of the iterative estimator
lation results show that the proposed estimator’s accuracy is close
to those of AM estimator and CO estimator. proposed by Aboutanios and Mulgrew [9] is only 1.0147 times
more than the asymptotic CRB for almost any frequency offset
Index Terms—DFT, frequency estimator, phase difference. case. The MSE of the iterative estimator proposed by Candan
and Orguner [8] also approximates the CRB closely.
From the above discussions, it is evident that it becomes hard
I. INTRODUCTION to improve the estimation accuracy of the above estimators. The
1070-9908 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1056 IEEE SIGNAL PROCESSING LETTERS, VOL. 22, NO. 8, AUGUST 2015
too low, and are almost surely to be identical. Assume this TABLE I
common peak bin is at and the corresponding frequency FREQUENCY ESTIMATE PROCEDURE
offset is ( ). Then, we have
(3)
In (3), since the normalized resolutions of the -point and
-point DFTs are and , respectively, their fractional
frequency offsets are and
accordingly, where represents the rounding operation.
Thus, the focus is on the estimation of the offset from the
phase measurements of these two peak DFT bins.
(12)
(5)
This method can be summarized in Table I.
Similarly, one can also derive the DFT result of as
B. Analysis of Computational Complexity
We now compare the detailed computational complexities be-
(6) tween the proposed estimator, Aboutanios and Mulgrew (AM)
estimator [9], and Candan and Orguner (CO) estimator [8].
Now we investigate the peak DFT bins at . Since As to AM estimator for the 1st iteration [9], the -point DFT
, combining (5) with (6) yields is implemented first. Then, the peak bin is searched and
two series summations for calculating the 0.5-translated peak
DFT bins and are conducted. Finally, and
are substituted to one formula to calculate the frequency offset .
(7)
As to CO estimator [8], the -point DFT is implemented first.
Taking the phase terms from (7), we have
Then, the peak index is searched and
pairs of DFT bins around the peak bin are utilized to calculate
(8) the offsets . Lastly, are substituted
to one formula to calculate the frequency offset .
As [8] pointed out that the computational complexity for cal-
Then, replacing in (8) by and subtracting culating the estimation formulas (for the proposed estimator,
from yields they are (10)–(12)) for a direct estimator is not significant, we
only need to compare the DFT calculations, DFT bins consid-
(9) ered and the extra operations listed in Table II.
Note that the number of the complex multiplications of the
To uniquely solve the fractional number in (9), the phase am-
-point DFT is about more than that of the -point DFT.
biguity term should be removed. Note that
From Table II, one can see that our proposed estimator in this
), since . If we implement
paper has a lower computational complexity than the other two.
modulo operation on the observed phase difference
as
IV. STATISTICAL PROPERTIES AND IMPROVED
(10) ITERATIVE PROCEDURE
This section mainly proves some statistical properties for our
Then . To move the phase difference into , proposed estimator, from which an improved iterative procedure
we conduct the following adjustment can also be derived. To facilitate the derivations, we only discuss
the high SNR case when noise cannot submerge the two peak
(11) DFT bins.
HUANG AND XIA: FINE RESOLUTION FREQUENCY ESTIMATOR 1057
(19)
In (19), is the SNR. Then, using the symmetry property,
we have
(20)
Then, combining (12) and (20), the variance of the ultimate es-
timate can be deduced as
A. Unbiasedness Property var
(21)
Assume that is zero-mean Gaussian noise with variance
and its DFT at is . We know that its real and When is large enough, substituting with
imaginary parts are both real Gaussian random variables with yields
zero mean and variance .
Combining (7), one can derive the peak DFT bin as (22)
Fig. 2. SNR gaps of Candan [6], CO estimator [8], AM estimator [9] and the
proposed estimator: (a) First iterations of the estimators; (b) Second iterations
of the estimators.