You are on page 1of 6

This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society prior to the

acceptance and publication.

Accuracy of Sine-wave Frequency Estimation


by an Iterative Interpolated DFT Algorithm
Daniel Belega Dario Petri Dominique Dallet
Department of Measurements and Department of Industrial Engineering, IMS Laboratory
Optical Electronics, University of Trento, University of Bordeaux-IPB ENSEIRB
”Politehnica” University of Timişoara, Trento, Italy MATMECA,
Timişoara, Romania dario.petri@unitn.it Talence, France
daniel.belega@upt.ro dominique.dallet@ims-bordeaux.fr

Abstract — In this paper the accuracy of the sine-wave frequency been provided [8]. The procedure proposed in [8] has been
estimator returned by an iterative Interpolated Discrete Fourier generalized in [9] and analytical solutions for both frequency
Transform (IpDFT) algorithm based on a Maximum Sidelobe estimators have been derived in [10]. Recently, the iterative
Decay (MSD) window is analyzed. The expressions for the procedure proposed in [8] has been generalized to MSD
contribution to the frequency estimation error of either the windows in order to reduce the effect on the estimated
spectral interference from the sine-wave image component and frequency of the image component when real-valued sine-
wideband noise are derived. It is shown that two algorithm waves are analyzed and non-coherent sampling occurs [11].
iterations ensure the minimum noise sensitivity achievable with However, algorithm sensitivities to the spectral interference
the adopted window. The accuracy of the derived expressions is
from the image component and wideband noise have not been
verified by means of computer simulations and validated by
yet analyzed in the scientific literature. This is the aim of this
experimental results.
paper when the algorithm described in [11] is based on DTFT
Keywords – Uncertainty analysis, frequency estimation, spectrum samples.
interpolation algorithms, sine-wave, windowing. The paper is organized as follows. In Section II the
considered iterative IpDFT algorithm is shortly described. The
expressions for the frequency estimation error due to the
I. INTRODUCTION spectral interference from the image component and the
One of the most common procedures employed for frequency estimation variance due to wideband noise are
estimating sine-wave parameters is the so called Interpolated derived in Section III. The accuracy of the derived expressions
Discrete Fourier Transform (IpDFT) algorithm [1]-[8]. This are then verified by means of both computer simulations and
algorithm compensates the detrimental contributions on the experimental results in Section IV. Conclusions drawn from
estimated parameters of both spectral leakage and picket-fence the performed analysis are finally reported in Section V.
effects by using windowing and interpolating spectrum
samples, respectively. Maximum Sidelobe Decay (MSD) II. THE ITERATIVE IPDFT ALGORITHM
cosine-class windows [4]-[7] are often adopted because related The analyzed discrete-time noisy sine-wave is modeled as:
parameter estimators are expressed by very simple equations,
thus simplifying algorithm implementation. In addition, the
MSD windows effectively reduce the effect on the estimated x(m ) = A sin (2πfm + φ ) + e n (m), m = 0, 1, 2, … (1)
parameters of possible narrow-band disturbances [4]-[7]. The
IpDFT algorithm firstly estimates the sine-wave frequency by where A, f, and φ are its amplitude, discrete frequency, and
interpolating two or more Discrete Fourier Transform (DFT) initial phase and en(⋅) is an additive wideband noise, assumed
spectrum samples. Then, the sine-wave amplitude and phase
are estimated using the obtained frequency estimate. Therefore, to be white, Gaussian with zero mean and variance σ n2 . The
the achievement of accurate sine-wave frequency estimates is discrete frequency f, equal to the ratio between the frequency fin
of major significance. of the original continuous-time sine-wave and the sampling
To reduce the noise sensitivity when estimating the rate fs, is assumed to be smaller than 0.5 to satisfy the Nyquist
frequency of a complex-valued sine-wave, an iterative IpDFT theorem. If M samples are acquired, this parameter can be
algorithm based on the rectangular window has been proposed expressed as:
in [8]. It has been shown that the variance of the proposed
frequency estimator almost achieves the related Cramér-Rao
Lower Bound (CRLB) for unbiased estimators after only two f in ν l +δ
f = = = , (2)
iterations. The algorithm interpolates two Discrete Time fs M M
Fourier Transform (DTFT) samples located half-bin apart from
the largest DFT spectrum sample. Interpolation formulas, where l and δ (−0.5 ≤ δ < 0.5) are the integer and the fractional
based on either DTFT complex values or DTFT spectrum, have parts of the number of acquired sine-wave cycles ν, which

978-1-4799-6144-6/15/$31.00 ©2015 IEEE


represents also the signal discrete frequency expressed in bins. step 1: determine the integer part l of the number of acquired
It is worth remembering that δ = 0 in the case of coherent sine-wave cycles by applying a maximum search
sampling, while δ ≠ 0 in most practical applications [12]. procedure to the DFT spectrum samples |Xw(k)|, k = 1,
The frequency-domain interpolation algorithm analyzed in 2,…, M/2 – 1.
this paper adopts the MSD cosine windows. These windows
have the highest sidelobe decay rate for a given number H ≥ 2 step 2: i: = 0 and δˆ0 = 0
of terms [13]. They are defined as:
step 3: i: = i + 1
M −1 l +δˆi −1 + r
− j 2π m

w(m ) =
H −1
∑ (− 1)h ah cos⎛⎜ 2πh m ⎞⎟, m = 0,1,…, M − 1 (3)
step 4: X w (l + δˆi −1 + r ) = ∑x w (m)e M
, r = ±0.5
h =0 ⎝ M⎠ m =0

in which the window coefficients are [6]: a 0 = C 2HH−−12 / 2 2 H −2 ,


ˆ( ˆ )
2 H − 1 X w l + δ i −1 + 0.5 − X w l + δ i −1 − 0.5 ( )
h(δˆi −1 ) = .
ah = C2HH−−h2−1 / 22H −3 , h = 1, 2,.., M −1, where C qp = p!/[( p − q)!q!] .
2 ( )
X w l + δˆi −1 + 0.5 + X w l + δˆi −1 − 0.5 ( )
The DTFT of the windowed signal xw(m) = x(m)·w(m), m = 0, δˆi = δˆi −1 + h(δˆi −1 )
1, 2,.., M −1, is:
step 5: repeat steps 3 and 4 once.

X w (λ ) =
A
2j
[ ]
W (λ − ν )e jφ − W (λ + ν )e − jφ + E wn (λ ),
(4)
step 6: return the estimated values for l and δˆ2 .
Observe that, as compared with the iterative algorithm
λ ∈ [0, M ) proposed in [8], the procedure above employs an H-term MSD
window instead of the rectangular window in order to reduce
where W(·) and Ewn(⋅) are the DTFT of the window w(·) and the the contribution on the estimated frequency of the spectral
weighted wideband noise w(m)⋅en(m), respectively. For M >> 1 interference from the sine-wave image component.
and |λ| << M , W(·) can be expressed as [6]:
III. INFLUENCE OF THE SPECTRAL INTERFERENCE DUE TO
THE IMAGE COMPONENT AND WIDEBAND NOISE
W (λ ) =
M sin (πλ ) (2 H − 2 )! e − jπλ , λ << M . (5) The iterative algorithm presented in Section II does not
2 2 H −2 πλ H −1

∏ (h
h =1
2
−λ 2
) consider the effects of both the spectral interference from the
image component and wideband noise. The related contributions
on the fractional frequency estimator returned by each
It is worth noticing that the second term in the square procedure iteration are separately analyzed in the following.
brackets of (4) represents the image component of the sine- A. Contribution of the spectral interference from the sine-
wave spectrum. wave image component
Since the number of acquired samples M is known the
estimation of the discrete frequency f requires the estimation of The contribution of the sine-wave image component on the
frequency estimate returned by the classical IpDFT algorithm
the number of acquired sine-wave cycles ν = l + δ. The integer
applied in the first pass of iteration is already known [14]:
part l can be estimated, for instance, by a maximum search
procedure applied to the DFT spectrum samples |Xw(k)|, k = 1,
2,…, M/2 – 1 [5]. Conversely, the fractional part δ can be
2( H + δ − 0.5)(l + δ ) W (2l + δ + 0.5)
estimated by employing the values of the two DTFT samples Δδ 1 = δˆ1 − δ ≅ p1 , (7)
|Xw (l – 0.5)| and |Xw (l + 0.5)| and neglecting the contributions due 2l + δ − H + 0.5 W (0.5 − δ )
to both spectral interference from the sine-wave image
component and wideband noise [11]. We have: where p1 = (–1)Hcos(2πδ + 2φ). In particular, the maximum
frequency estimation error associated to the classical IpDFT
algorithm is:
2 H − 1 X w (l + 0.5) − X w (l − 0.5)
δˆ = . (6)
2 X w (l + 0.5) + X w (l − 0.5) 2( H + δ − 0.5)(l + δ ) W (2l + δ + 0.5)
Δδ 1 max ≅ . (8)
2l + δ − H + 0.5 W (0.5 − δ )
To increase the accuracy of the estimator δˆ the following
iterative algorithm has been suggested [8], [11]: Similarly, the error due to the sine-wave image component
on the frequency estimate returned after two iterations is given
by (see (A.14) in the Appendix):
(2 H − 1)(l + δ ) W (2l + 2δ + 0.5) where SNR = A 2 /(2σ n2 ) , SL(-0.5) and ENBW are the Signal-
Δδ 2 = δˆ2 − δ ≅ p 2 , (9)
2l + 2δ − H + 0.5 W (0.5) to-Noise Ratio, the window worst-case Scalloping Loss, and
the window Equivalent Noise BandWidth, respectively [15].
For the H-term MSD window we have [6]:
where p2 = (–1)H + 1sgn(|δ|-0.25)cos(2πδ + 2φ), in which sgn(⋅)
is the sign function. The maximum value of (9) is: SL(−0.5) = 2 4 H −3 [( H − 1) !]4 /[π (2 H − 1) !(2 H − 2) !] and
ENBW = C42HH−−42 /(C2HH−−12 ) 2 .
(2 H − 1)(l + δ ) W (2l + 2δ + 0.5)
Δδ 2 ≅ . (10) IV. SIMULATION AND EXPERIMENTAL RESULTS
max 2l + 2δ − H + 0.5 W (0.5)
In this Section the accuracy of the expressions derived in
Using (5), from (7) and (9) it follows that the errors Δδ1 and Section III for the estimators δˆ1 and δˆ2 is verified through
Δδ2 are null when δ = -0.5 and |δ| = 0.25, respectively, due to computer simulations. Moreover, their behaviors are validated
the behavior of the window spectrum sidelobes. Also, (8) and by means of experimental data.
(10) return Δδ 1 max = Δδ 2 max when δ = 0, i.e. when coherent The synthesized sine-wave amplitude is A = 1, the number
of analyzed samples is M = 1024 and the two-term MSD (or
sampling occurs. This condition almost coincides, as shown Hann window [13]) is adopted.
also by computer simulations reported in Section IV, with the Fig. 1 shows the errors |Δδ1|max and |Δδ2|max returned by
worst-case values for Δδ 1 max and Δδ 2 max for a given value simulations, (8) and (10), respectively, as a function of the
of l. As a consequence, the curves expressing the estimation number of acquired sine-wave cycles ν when pure sine-waves
error due to the image component exhibit almost the same are considered. During simulations the value of ν was varied
envelope and no significant accuracy improvement can be in the range [2.01, 10) with a step of 0.04. For each value of ν
generally ensured by iterating the IpDFT algorithm twice. the maximum error was determined by varying the sine-wave
Moreover, expressions (8) and (10) show that the contribution phase φ in the range [0, 2π) rad with a step of π/50 rad.
of the image component tends to decrease as the integer part l
of the acquired number of cycles increases, and quickly
become negligible due to the rapid spectrum sidelobe decaying
rate of the H-term MSD windows.
B. Contribution of wideband noise
Let’s assume that the number of acquired cycles ν is high
enough that the effect of the spectral interference from the
image component on the estimated frequency is negligible.
The variance of the estimator δˆ1 returned by the first pass
of iteration has been already analyzed in the scientific
literature. It is given by [14]:
H −1
2π 2 ( H − 0.5 − δ ) 2 (δ + 0.5) 2
σ δ2ˆ ≅
1
[(2 H − 1) !] 2
∏ [h
h =1
2
− (δ + 0.5)2 ] 2
Figure 1. Frequency estimation errors |Δδ1|max and |Δδ2|max due to the sine-
(11)
wave image component returned by simulations, (8) and (10), respectively,
2
2(4 H − 3)(δ − 0.25) + 2 H 2
− 1 C 42HH−−42
× 2 2
σ n2 , versus ν. The Hann window is adopted and M = 1024 samples are analyzed.
(2 H − 1) cos (πδ ) A M
Fig. 1 shows that there exists a very good agreement
As for the second pass of iteration the variance of the between simulation and theoretical results. Moreover, it can be
related estimator δˆ2 is close to the value returned by (11) seen that the worst-case errors returned after one or two
when δ = 0 [8], i.e., iterations are achieved when δ ≅ 0, and they are very close each
other. Conversely, errors exhibit a minimum at |δ| ≅ 0.5 and |δ|
H −1 ≅ 0.25, respectively, as expected.
(2 H − 1)π 2 C 42HH−−42
σ δ2ˆ ≅ σ δ2ˆ
2 1 |δ = 0

[(2 H − 2) !] 2
∏ (h 2 − 0.25) 2
16 A 2 M
σ n2 . Fig. 2 shows the variances of the fractional frequency
h =1 estimators δˆ1 and δˆ2 obtained by simulations, (11) and (12),
(12)
respectively, versus δ when SNR = 50 dB and l = 73 cycles.
which can be written as: The related CRLB, equal to 3 /(π 2 ⋅ SNR ⋅ M ) [5], [16], is also
reported in Fig. 2. During simulations the value of δ was varied
2 H − 1 ENBW 1 in the range [-0.5, 0.5) with a step of 0.04. For each value of δ,
σ δ2ˆ ≅ , (13) 10000 runs were considered by choosing the phase φ at random
2 8 SL2 (−0.5) M SNR
in the range [0, 2π) rad.
Figure 3. Frequency estimation errors |Δδ1|max and |Δδ2|max versus δ. The
Figure 2. Frequency estimation variances σ δ2ˆ and σ δ2ˆ due to wideband
frequency of the analyzed sine-wave varies in the range [250, 340] Hz and the
1 2
noise returned by simulations, (11) and (12), respectively, and the related related SINAD is about 60 dB. The reference frequency was achieved by the
CRLB, versus δ when SNR = 50 dB and l = 73. The Hann window is adopted 4PSF algorithm. The Hann window is adopted and 1000 runs of M = 1024
and 10000 runs of M = 1024 samples are analyzed at each the frequency value. samples are analyzed at each frequency value.

According to simulations, the minimum of the worst-case


Fig. 2 shows that the agreement between simulated and error magnitudes are achieved in the first and second iterations
theoretical results is very good. Also, the variance σ δ2ˆ is at |δ| ≅ 0.5 and |δ| ≅ 0.25, respectively. Also, |Δδ1|max and
2

almost independent of δ and coincides with the minimum value |Δδ2|max are almost equal when δ ≅ 0.
of the first iteration variance σ δ2ˆ , which is achieved for δ = 0.
1 Fig. 4 shows the variances of the obtained fractional
However, the variance σ δ2ˆ is about 2.6-times higher than the frequency estimators δˆ1 and δˆ2 as a function of δ, which has
2

CRLB, as expected due to windowing [5], [16]. Fig. 2 also been determined as the mean value of the estimates obtained
shows that only two iterations suffice to compensate the effect after two algorithm iterations.
of δ on the frequency estimation variance.
The behaviors of the derived expressions for the frequency
estimation errors and variances have also been validated by
means of experimental results. Sine-waves were provided by
an Agilent 33220A signal generator and acquired by means of
an NI6023E acquisition board. The full scale range and the
sampling rate of the acquisition board were set to 10 V and 100
kHz, respectively. The sine-waves amplitude was 2 V and the
frequencies were varied in the ranges [250, 340] Hz and [6.89,
6.98] kHz with a step of 10 Hz. Acquiring M = 1024 samples at
the considered sampling rate these frequency values correspond
to the acquisition of l = 3 and l = 71 integer sine-wave cycles,
respectively. The first and the second frequency ranges have
been used to validate the behaviors of (8), (10), (11), and (12),
respectively. For each signal frequency 1000 runs of M = 1024
samples each were considered and the Hann window was Figure 4. Frequency estimation variances σ δ2ˆ and σ δ2ˆ due to wideband
1 2
adopted. The SIgnal-to-Noise And Distortion ratio (SINAD) noise returned by experimental data versus δ. The frequency of the analyzed
estimated by the four-parameter sine-fit algorithm described in sine-wave varies in the range [6.89, 6.98] kHz and the related SINAD is about
the IEEE Standard 1241 [17] was about 60 dB. The errors 60 dB. The Hann window is adopted and 1000 runs of M = 1024 samples are
|Δδ1|max and |Δδ2|max returned by the analyzed algorithm are analyzed at each frequency value.
shown in Fig. 3 as a function of δ, which has been determined A comparison between Fig. 2 and Fig. 4 shows that estimator
as the mean value of the estimates obtained after two algorithm variances obtained by simulation and experimental data exhibit
iterations. The reference value for δ was achieved by means of very similar behaviors. Also, reminding that the estimation
six iterations of the four-parameter sine-fit (4PSF) algorithm variance expected for the Hann window is about 2.6-times the
[17], in which the initial values for the signal parameters were CRLB, a SNR value of about 65 dB can be derived from Fig. 4.
estimated by applying the three-parameter sine-fit (3PSF)
algorithm. The signal frequency employed in the 3PSF V. CONCLUSIONS
algorithm was determined by using the classical IpDFT
algorithm based on the Hann window. In this paper the contributions of both the spectral
interference from the image component and wideband noise on
the frequency estimators returned by an iterative IpDFT
algorithm when applied to real-valued noisy sine-waves have [17] IEEE Std. 1241, Standard for Terminology and Test Methods for
been analyzed. The expressions for the contributions of these Analog-to-Digital Converters, 2010.
two detrimental phenomena on returned frequency estimator
have been derived. It has been shown that iterating the IpDFT APPENDIX
algorithm does not affect the worst-case frequency estimation
error due to the spectral interference from the image DERIVATION OF THE EXPRESSION FOR THE ERROR Δδ2
component. Conversely, iterating twice allows a reduction of
the frequency estimation sensitivity to wideband noise, so From (7), δˆ1 = δ + Δδ 1 . Thus from (4) we have:
ensuring the achievement of the minimum value returned by
the classical IpDFT algorithm in the case when coherent
sampling occurs. The accuracy of the derived expressions has X w (l + δ + Δδ 1 + r ) =
been confirmed through both simulations and experimental
results.
A
2j
[ ]
W (Δδ 1 + r )e jφ − W (2l + 2δ + Δδ 1 + r )e − jφ , r = ±0.5.
(A.1)
REFERENCES
2 2 2
[1] D.C. Rife and G.A. Vincent, “Use of the discrete Fourier transform in Using the identity z1 + z 2 = z1 + z 2 + 2 Re{z1 z 2* } ,
the measurement of frequencies and levels of tones,” Bell Syst. Tech. J,
vol. 49, 197-228, 1970. where z1 and z2 are complex-valued variables, Re{⋅} denotes
[2] B.G. Quinn, Estimation of frequency, amplitude, and phase from the the real part of its argument, and (⋅)* the conjugation operator,
DFT of a time series, IEEE Trans. Signal Process. vol. 45, no. 3, pp. we have:
814-817, 1997.
[3] J. Schoukens, R. Pintelon, and H. Van hamme: “The interpolated fast
Fourier transform: a comparative study,” IEEE Trans. Instrum. and
X w (l + δ + Δδ 1 + r ) =
2
Meas., vol. 41, no. 2, pp. 226-232, Apr. 1992.
[4] T. Grandke, “Interpolation algorithms for discrete Fourier transforms of
weighted signals,” IEEE Trans. Instrum. Meas., vol. IM-32, no. 2, pp.
350 – 355, Jun. 1983.
A2
4
[W (Δδ 1 + r ) + W (2l + 2δ + Δδ 1 + r )
2 2
(A.2)
[5] C. Offelli and D. Petri, “The influence of windowing on the accuracy of
multifrequency signal parameter estimation,” IEEE Trans. Instrum.
{ }]
− 2 Re W (Δδ 1 + r )W * (2l + 2δ + Δδ 1 + r )e j 2φ .
Meas., vol. 41, no. 2, pp. 256-261, Apr.1992.
[6] D. Belega and D. Dallet, “Multifrequency signal analysis by interpolated Using (5) we obtain:
DFT method with maximum sidelobe decay windows,” Measurement,
vol. 42, no. 3, pp. 420-426, Apr. 2009.
[7] D. Belega, D. Dallet, and D. Petri, “Statistical description of the sine-
wave frequency estimator provided by the interpolated DFT method,“
{
Re W (Δδ 1 + 0.5)W * (2l + 2δ + Δδ 1 + 0.5)e j 2φ =}
Measurement, vol. 45, no. 1, pp. 109-117, Jan. 2012. (−1) H sgn(| δ | −0.25) W (Δδ 1 + 0.5) (Α.3)
[8] E. Aboutanios and B. Mulgrew, “Iterative frequency estimation by
interpolation on Fourier coefficients,” IEEE Trans. Signal Process. vol. × W (2l + 2δ + Δδ 1 + 0.5) cos(2πδ + 2φ ) ,
53, no. 4, pp. 1237-1241, Apr. 2005.
[9] Y. Liu, Z. Nie, Z. Zhao, andQ.H. Liu, Generalization of iterative Fourier and
interpolation algorithm for single frequency estimation, Digital Signal
Process., vol. 21, no. 1, pp. 141-149, 2011.
[10] J-R Liao and S. Lo, Analytical solutions for frequency estimators by
interpolation of DFT coefficients, Signal Process., vol. 100, pp. 93-100, {
Re W (Δδ 1 − 0.5)W * (2l + 2δ + Δδ 1 − 0.5)e j 2φ =}
(−1) H +1 sgn(| δ | −0.25) W (Δδ 1 − 0.5)
2014.
(Α.4)
[11] D. Belega, D. Petri, and D. Dallet, “Iterative sine-wave frequency
estimation by generalized Fourier interpolation algorithms,” in Proc. × W (2l + 2δ + Δδ 1 − 0.5) cos(2πδ + 2φ ) ,
11th IEEE ISETC 2014 symposium, pp. 105-108, November 14-15,
2014, Timişoara, Romania.
[12] A. Ferrero and R. Ottoboni, “High-accuracy Fourier analysis based on where sgn(⋅) is the sign function.
synchronous sampling techniques,” IEEE Trans. Instrum. Meas., vol. 41
no. 6, pp. 780-785, Dec. 1992. Since W (2l + 2δ + Δδ 1 + r ) << W (Δδ 1 + r ) when l is high
[13] A.H. Nuttall, “Some windows with very good sidelobe behavior,” IEEE enough, and considering the first-order Taylor’s series
Trans. Acoust., Speech, Signal Processing, vol. ASSP-29, no.1, pp 84– expansion of (1 + x)1/2 we have:
91, Feb. 1981.
[14] D. Belega, D. Petri, and D. Dallet, “Accurate sine-wave frequency
estimation by means of the interpolated discrete time Fourier transform
approach,” submitted to Meas. Sci. Technol. journal X w (l + δ + Δδ 1 + 0.5) ≅
(A.5)
[ ]
[15] F.J. Harris, “On the use of windows for harmonic analysis with the A
discrete Fourier transform,” Proc. of the IEEE, vol. 66, no. 1, pp. 51–83, W (Δδ 1 + 0.5) + p 2 W (2l + 2δ + Δδ 1 + 0.5) ,
Jan. 1978. 2
[16] C. Offelli and D. Petri, “Weigthing effect on the dicrete time Fourier
transform of noisy signals,” IEEE Trans. Instrum. Meas., vol. 40, no. 6, and
pp. 972-981, Dec. 1991.
X w (l + δ + Δδ 1 − 0.5) ≅ W (Δδ 1 − 0.5) =
2 H − 1 + 2Δδ 1
W (Δδ 1 + 0.5) , (Α.10)
(Α.6) 2 H − 1 − 2Δδ 1
A
2
[
W (Δδ 1 − 0.5) − p 2 W (2l + 2δ + Δδ 1 − 0.5) , ]
and
H+1
where p2 = (-1) sgn(|δ| - 0.25)cos(2πδ+ 2φ).
2l + 2δ + Δδ 1 + H − 0.5
Using (A.5) and (A.6), the expression for h(δˆ1 ) used in the W (2l + 2δ + Δδ 1 − 0.5) =
2l + 2δ + Δδ 1 − H + 0.5 (Α.11)
iterative procedure becomes:
× W (2l + 2δ + Δδ 1 + 0.5) .

2 H − 1 W (Δδ 1 + 0.5) − W (Δδ 1 − 0.5) Using (A.10) and (A.11), after some calculations (A.8)
h(δˆ1 ) ≅ provides:
2 W (Δδ 1 + 0.5) + W (Δδ 1 − 0.5)
⎡ W (2l + 2δ + Δδ 1 + 0.5) + W (2l + 2δ + Δδ 1 − 0.5) ⎤
⎢1 + p 2 ⎥ h(δˆ1 ) ≅ −Δδ 1
⎢ W (Δδ 1 + 0.5) − W (Δδ 1 − 0.5) ⎥.
×⎢
W (2l + 2δ + Δδ 1 + 0.5) − W (2l + 2δ + Δδ 1 − 0.5) ⎥ (2 H − 1 − 2Δδ 1 )(l + δ ) W (2l + 2δ + Δδ 1 + 0.5) (Α.12)
⎢1 + p 2 ⎥ + p2 .
⎢⎣ W (Δδ 1 + 0.5) + W (Δδ 1 − 0.5) ⎥⎦ 2l + 2δ + Δδ 1 − H + 0.5 W (Δδ 1 + 0.5)

(Α.7) Since |Δδ1| << 0.5, its contribution can be neglected and it
follows that:
Since

(2 H − 1)(l + δ ) W (2l + 2δ + 0.5)


W (2l + 2δ + Δδ 1 + 0.5) − W (2l + 2δ + Δδ 1 − 0.5) δˆ2 = δˆ1 + h(δˆ1 ) ≅ δ + p 2 ,
2l + 2δ − H + 0.5 W (0.5)
<< W (Δδ 1 + 0.5) + W (Δδ 1 − 0.5) (Α.13)

it follows: which leads to:

2 H − 1 W (Δδ 1 + 0.5) − W (Δδ 1 − 0.5) (2 H − 1)(l + δ ) W (2l + 2δ + 0.5)


h(δˆ1 ) ≅ (1 + p 2ε ), (Α.8) Δδ 2 ≅ p 2 . (Α.14)
2 W (Δδ 1 + 0.5) + W (Δδ 1 − 0.5) 2l + 2δ − H + 0.5 W (0.5)

where:

W (2l + 2δ + Δδ 1 + 0.5) + W (2l + 2δ + Δδ 1 − 0.5)


ε=
W (Δδ 1 + 0.5) − W (Δδ 1 − 0.5)
(Α.9)
W (2l + 2δ + Δδ 1 + 0.5) − W (2l + 2δ + Δδ 1 − 0.5)
− .
W (Δδ 1 + 0.5) + W (Δδ 1 − 0.5)

From (5) after some algebra we have:

You might also like