EFFICIENT ALGORITHM FOR ESTIMATING THE PARAMETERS OF CHIRP SIGNAL

Transcription

1 EFFICIENT ALGORITHM FOR ESTIMATING THE PARAMETERS OF CHIRP SIGNAL ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 Abstract. Chirp signals play an important role in the statistical signal processing. Recently Kundu and Nandi [8] derived the asymptotic properties of the least squares estimators of the unknown parameters of the chirp signals model in presence of stationary noise. Unfortunately they did not discuss any estimation procedures. In this article we propose a computationally efficient algorithm for estimating different parameters of a chirp signal in presence of stationary noise. From proper initial guesses, the proposed algorithm produces efficient estimators in a fixed number of iterations. We suggest how to obtain the proper initial guesses also. The proposed estimators are consistent and asymptotically equivalent to least squares estimators of the corresponding parameters. We perform some simulation experiments to see the effectiveness of the proposed method, and it is observed that the proposed estimators perform very well. For illustrative purposes, we have performed the data analysis of a simulated data set. Finally, we propose some generalization in the conclusions. Key Words and Phrases: Chirp signals; least squares estimators; strong consistency, asymptotic distribution; linear processes; iteration. Department of Mathematics and Statistics, Indian Institute of Technology Kanpur, Pin 0806, India. Part of this work has been supported by a grant from the Center for Scientific and Industrial Research (CSIR), Government of India. 3 Part of this work has been supported by a grant from the Department of Science and Technology, Government of India. 4 Corresponding author. kundu@iitk.ac.in, Fax:

2 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3. Introduction In this paper we consider the following chirp signal model in presence of additive noise: () y(n) A 0 cos(α 0 n β 0 n ) B 0 sin(α 0 n β 0 n ) X(n); n,,n, where A 0, B 0 are non zero amplitudes, with restriction A 0 B0 M, for some constant M. The frequency and frequency rate, α 0, β 0, respectively, lie strictly between 0 and π. X(n) is a stationary noise sequence, and it has the following form; () X(n) a(j)ε(n j), a(j) <. j j Here, {ε(n)} is a sequence of independent and identically distributed (i.i.d.) random variables with zero mean, variance σ and with finite fourth moment. Given a sample of size N, the problem is to estimate the unknown amplitudes A 0, B 0, the frequency α 0 and the frequency rate β 0. A chirp signal occurs quite naturally in different areas of science and technology. It is a signal where frequency changes over time and this property of the signal has been exploited quite effectively, to measure the distance of an object from a source. This model can be found in different sonar, radar, communications problems, see for example Abatzoglou [], Kumaresan and Verma [6], Djuric and Kay [3], Gini et al. [5], Nandi and Kundu [9], Kundu and Nandi [8] and the references cited therein. In practice, the parameters like amplitudes, frequency, frequency rate are unknown, and one tries to find efficient estimators of these unknown parameters, having some desired statistical properties. Recently, Kundu and Nandi [8] derived the asymptotic properties of the least squares estimators (LSEs), and it is observed that the LSEs are consistent and asymptotically normally distributed. It is further observed that the LSEs are efficient, with the convergence rates of the amplitude, frequency and frequency rate are O p (N / ), O p (N 3/ ) and O p (N 5/ ) respectively. Here O p (N δ )

3 CHIRP 3 means N δ O p (N δ ) is bounded in probability. Note that a sequence of random variables {X n } is said to be bounded in probability, if for any ǫ > 0, there exists m,m, such that P(m < X n < M) ǫ. Unfortunately, Kundu and Nandi [8] did not discuss any estimation procedure of the LSEs. It is clear that they have to be obtained by some iterative procedure like Newton-Raphson method or its variant, see for example Tjalling [4]. Although, the initial guesses and convergence of the iterative procedure seem to be important issues. It is well known, see Rice and Rosenblatt [], that even in the simple sinusoidal model, finding the LSEs is not an trivial issue. The problem is due to the fact that the least squares surface has several local minima, and therefore, if the initial estimators are not properly chosen, any iterative procedure will converge to a local minimum rather than the global minimum. Several methods have been suggested in the literature to find the efficient frequency estimators, see for example a very recent article by Kundu et al. [7] in this respect. The main aim of this paper is to find estimators of the amplitudes, frequency and frequency rate efficiently, which have the same rates of convergence as the corresponding LSEs. It may be observed that the model () can be seen as a non-linear regression model, with A 0, B 0 as linear parameters, and α 0, β 0 as non-linear parameters. If we can find efficient estimators of the non-linear parameters α 0 and β 0, then efficient estimators of the linear parameters can be obtained by simple linear regression technique, see for example Richards []. Because of this reason, in this paper we mainly concentrate on estimating the non-linear parameters efficiently. We propose an iterative procedure which has been applied to find efficient estimators of the frequency and frequency rate. It is observed that if we start the initial guesses of α 0 and β 0 with convergence rates O p (N ) and O p (N ) ) respectively, then after four iterations the algorithm produces an estimate of α 0 with convergence rate O p (N 3/ ), and an estimate of β 0 with convergence rate O p (N 5/ ). Therefore, it is

4 4 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 clear that the proposed algorithm produces estimates which have the same rates of convergence as the LSEs. Moreover, it is known that the algorithm stops after finite number of iterations. We perform some simulation experiments, to see the effectiveness of the proposed method for different sample sizes and for different error variances. It is observed that the algorithm works very well. The mean squared errors (MSEs) of the proposed estimators are very close to the corresponding MSEs of the LSEs, and both are very close to the corresponding asymptotic variance of the LSEs. Therefore, the proposed method can be used very effectively instead of the LSEs. For illustrative purposes, we have analyzed one simulated data, and the performance is very satisfactory. The rest of the paper is organized as follows. We provide the properties of the LSEs in Section. In Section 3, we present the proposed algorithm and provide the theoretical justification of the algorithm. The simulation results and the analysis of a simulated data have been presented in Section 4. Conclusions appear in Section 5. All the proofs are presented in the Appendix.. Existing Results In this section we present briefly the properties of the LSEs for ready reference. The LSEs of the unknown parameters of the model () can be obtained by minimizing S(Θ) with respect to Θ (A,B,α,β), where ( S(Θ) y(n) A cos(αn βn ) B sin(αn βn ) ) n [ [ T [ [ A A Y W(α,β) Y W(α,β). B]] B]]

5 CHIRP 5 Here Y (y(),,y(n)) T, is the N data vector and W(θ) is the N matrix of the following form; (3) W(θ) cos(α β) sin(α β) cos(α 4β) sin(α 4β).. cos(nα N β) sin(nα N β) Note that, if α and β are known the LSEs of A 0 and B 0 can be obtained as Â(θ) and B(θ) respectively, where θ (α,β),. (4) (Â(θ), B(θ) ) T ( W T (θ)w(θ) ) W T (θ)y, Therefore, the LSEs of α 0 and β 0 can be obtained first by minimizing Q(α,β) with respect to α and β, where (5) Q(α,β) S(Â(θ), B(θ),α,β) Y T W(θ) ( W T (θ)w(θ) ) W T (θ)y. Once the LSEs of α 0 and β 0, say α and β are obtained the LSEs of A 0 and B 0 can be easily obtained as Â( α, β) and B( α, β) respectively, see for example Richards []. Kundu and Nandi [8] derived the properties of the LSEs and it is as follows. The LSEs of the unknown parameters of the model () are consistent to the corresponding parameters and they have the following asymptotic distribution ( N / (Â A 0),N / ( B ) B 0 ),N 3/ ( α α 0 ),N 5/ d ( β β 0 ) N 4 (0, 4cσ Σ), where c j a(j). The symbol means convergence in distribution, N 4 (0, 4cσ Σ) means a 4-variate normal distribution with mean vector 0, the dispersion matrix 4cσ Σ and (A 0 9B0) 4A 0 B 0 8B 0 5B 0 (6) Σ 4A 0 B 0 A 0 B0 (9A 0 B0) 8A 0 5A 0 8B 0 8A B 0 5A In the next section we provide the algorithm which produces consistent estimators of α 0 and β 0 which have the same rates of convergence as the LSEs. d

6 6 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 3. Proposed Algorithm The aim of this algorithm is to find the estimates of the frequency and frequency rate with the same rates of convergence as the LSEs. We will first deduce a method for improving the speed of convergence of general consistent estimators of α 0 and β 0 : this method is the crucial tool for implementing the main algorithm devloped in this paper. If α is an estimator of α 0, such that α α 0 O p (N ( δ ) ), for some δ > 0, and β is an estimator of β 0, such that β β 0 O p (N ( δ ) ), for some δ > 0, then an improved estimator of α 0 can be obtained as (7) α α 48 ( ) P α N Im N with (8) P α N n ( y(n) n N ) e i( αn βn ) Q N and Q N y(n)e i( αn βn ). Here Im( ) denotes the imaginary part of the complex argument. Similarly, an improved estimator of β 0 can be obtained as ( ) (9) β β 45 P β N 4Im N, Q N with N (0) P β N ) y(n) (n N e i( αn βn ) 3 n and Q N is same as defined in (8). n The following two theorems provide the justification for the improved estimators, whose proofs will be provided in the appendix. Theorem : If α α 0 O p (N ( δ ) ) for δ > 0 then ( (a) α α 0 O ) p N ( δ ) if δ 4 (b) N 3 d ( α α 0 ) N(0,σ) if δ > 4, where σ 384cσ, the asymptotic variance of the LSE of α A 0 B0 0.

7 Theorem : If β β 0 O p ( N ( δ ) ) for δ > 0 then CHIRP 7 ( (a) β β0 O ) p N ( δ ) if δ 4 (b) N 5 d ( β β0 ) N(0,σ) if δ > 4 where σ 360cσ, the asymptotic variance of the LSE of β A 0 B0 0. Theorem 3: If α α 0 O p (N ( δ) ) and β ( β 0 O ) p N ( δ ), then for δ > 4 and for δ > 4, Cov(N 3 ( α α 0 ),N 5 ( β β0 )) 360cσ. A 0 B0 Now we will show that starting from initial guesses α, β, with convergence rates α α 0 O p (N ) and β β 0 O p (N ) respectively, how the above procedure can be used to obtain efficient estimators. It may be noted that finding initial guesses with the above convergence rates are not very difficult. From Rice and Rosenblatt [] it follows that if one can show, α, β, the minimizer of Q(α,β) are such that, N( α α 0 ) 0 a.s.,n ( β β 0 ) 0 a.s., then one has the existence of such an initial ( πj guess when searched over the grids N, πk ), j,,n, and k,,n. Let N Θ 0 (A 0,B 0,α 0,β 0 ) be the true parameter value and D is a diagonal matrix, such ( ) that D diag,, N N N N, N N,. Note that, see Kundu and Nandi [8], () S (Θ 0 )D ( Θ Θ 0 )D [DS ( Θ)D], where S ( ) and S ( ) are the derivative vector and the Hessian matrix of S( ), respectively, and Θ is a point on line joining Θ and Θ 0. It has been shown in Kundu and Nandi [8] that () DS ( Θ)D lim N DS (Θ 0 )D a.s., which is a positive definite matrix. Then () gives (3) ( Θ Θ 0 )( [ ND) [DS S (Θ 0 )D] ( Θ)D ]. N Also S (Θ 0 )D N 0 a.s. implies ( Θ Θ 0 )( ND) 0 a.s. which means N( α α 0 ) 0 a.s.,n ( β β 0 ) 0 a.s.

8 8 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 The main idea of the proposed algorithm is not to use the whole sample size at the beginning, as it was first suggested by Bai et al. []. We will use part of the sample at the beginning and gradually proceed towards the complete sample. The algorithm can be described as follows. We will denote the estimates of α 0 and β 0 obtained at the i-th iteration as α (i) and β (i) respectively. Algorithm: Step : Choose N N 8/9. Therefore, α (0) α 0 O p (N ) O p (N /8 ), and β (0) β 0 O p (N ) O p (N /4 ). Perform step (7) and (9). Therefore, after the -st iteration we have α () α 0 O p (N /4 ) O p (N 0/9 ), and β () β 0 O p (N ) O p (N 0/9 ). Step : Choose N N 80/8. Therefore, α () α 0 O p (N /8 ), and β() β 0 O p (N /4 ). Perform step (7) and (9). Therefore, after the -nd iteration we have α () α 0 O p (N /4 ) O p (N 00/8 ), and β () β 0 O p (N / ) O p (N 00/8 ). Step 3: Choose N 3 N. Therefore, α () α 0 O p (N 9/8 3 ), and β() β 0 O p (N 38/8 3 ). Perform step (7) and (9). Therefore, after the 3-rd iteration we have α (3) α 0 O p (N 38/8 ), and β (3) β 0 O p (N 76/8 ). Step 4: Choose N 4 N, and perform step (7) and (9). Now we obtain the required convergence rates, i.e. α (4) α 0 O p (N 3/ ), and β (4) β 0 O p (N 5/ ). 4. Simulation and data analysis 4.. Simulation Results: In this section we present some simulation results for different sample sizes and for different error variances to show how the proposed

9 CHIRP 9 Table. Result for model with i.i.d. error sample size σ 0.05 σ 0.5 N50 PARA ASYV ( E-04) ( E-08) ( E-03) ( E-07) MEAN (Algo) MSE (Algo) ( E-03) ( E-06) ( E-03) ( E-06) MEAN (LSE) MSE (LSE) ( E-04) ( E-08) ( 0.866E-03) ( E-07) N00 PARA ASYV ( E-05) ( E-09) ( E-04) ( E-08) MEAN (Algo) MSE (Algo) ( E-03) ( E-08) ( E-03) ( E-08) MEAN (LSE) MSE (LSE) ( E-05) ( E-09) ( E-04) ( 0.560E-08) N00 PARA ASYV ( E-06) ( E-) ( E-05) ( E-0) MEAN (Algo) MSE (Algo) ( E-05) ( E-09) ( E-04) ( E-09) MEAN (LSE) MSE (LSE) ( E-05) ( E-0) ( E-04) ( E-09) method behaves in practice. We consider the following model (4) y(n).0 cos(.75n.05n ).0 sin(.75n.05n ) X(n). We have considered (i) X(n) ε(n) (i.i.d. errors) and (ii) X(n) ε(n) 0.5ε(n ) (stationary errors), where ε(n) s are i.i.d. normal random variables with mean 0 and variance σ. We have taken N 50, 00 and 00, σ 0.05 and 0.5. In each case we have obtained the initial guesses as has been suggested in Section 3. For each N and σ, we compute the average estimates of α 0 and β 0, and the associate MSEs based on 000 replications. The results are reported in Table and Table. For comparison purposes we have also computed the LSEs and the corresponding asymptotic variances. From the results presented in Tables and, it is clear that the performances of the estimators obtained by the proposed algorithm are quite satisfactory in comparison to the corresponding performances of the LSEs. The average MSEs of the proposed estimators are very close to the asymptotic variance of the corresponding LSEs. The performances are quite good even with moderate sample sizes. It is clear that even with the four steps of iterations, we are able to achieve the same accuracy of the LSEs.

10 0 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 Table. Result for model with Stationary error sample size σ 0.05 σ 0.5 N50 PARA ASYV ( E-04) ( E-08) ( E-03) ( E-07) MEAN (Algo) MSE (Algo) ( E-03) ( E-07) ( E-03) ( E-06) MEAN (LSE) MSE (LSE) ( E-04) ( E-08) ( E-03) ( E-07) N00 PARA ASYV ( E-05) ( E-09) ( E-04) ( E-08) MEAN (Algo) MSE (Algo) ( E-03) ( E-08) ( E-03) ( E-08) MEAN (LSE) MSE (LSE) ( E-05) ( E-09) ( E-04) ( E-08) N00 PARA ASYV ( E-06) ( E-) ( E-05) ( E-0) MEAN (Algo) MSE (Algo) ( E-05) ( E-09) ( E-04) ( E-09) MEAN (LSE) MSE (LSE) ( E-05) ( E-0) ( E-04) ( E-09) 4.. Data Analysis. In this subsection we present the analysis of a simulated data set mainly for illustrative purposes. We have generated a sample of size 00 from the following chirp model (5) y(n).5 cos(.0n.0n ).5 sin(.0n.0n ) X(n) here X(n) is a sequence of i.i.d. Gaussian random variables with mean zero and variance 0.5. It is presented in Figure. We also present the least squares surface Q(θ) in Figure as defined in (5). Since Q(θ) has only one trough, the initial guesses can be obtained as suggested in Section 3. Using the algorithm, we obtained the estimates of A 0, B 0, α 0, β 0 as ,.79, and respectively. The associated 95% confidence intervals are obtained using bootstrap approach, and they are (0.037,.840), (.369,.3475), (0.9998,.000) and (0.9999,.000) respectively. The residuals are reported in Figure 3. From the run test it follows that the residuals are random, as expected. 5. Conclusion In this paper we propose an efficient algorithm to estimate the nonlinear parameters of one dimensional chirp signal in presence of stationary noise. The main advantage

11 CHIRP Figure. Simulated data from the Chirp Model (5) of size Figure. Least squares surface plot of the proposed algorithm is that starting from a reasonable starting values for both the frequency and frequency rate, in four steps only, it produces estimates which are asymptotically equivalent to the LSEs. We have also proposed how to obtain the reasonable initial guesses from the least squares surfaces. We have performed an extensive simulation experiments and it is observed that the proposed algorithm is working very well, and it is producing estimates which are quite close to the LSEs. Although, in this paper we have considered the estimation procedure of the ordinary chirp model in presence of stationary noise, the method may be extended for

12 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA, Figure 3. Residuals plot generalized chirp signal model also as it was proposed by Djuric and Kay [3]. The work is in progress, it will be reported later. Acknowledgement The authors would like to thank the associate editor and a referee for their vauable comments. Appendix We need the following lemma for proving the theorems. Using Vinogradov s [5] result one can prove it. Lemma. If (θ,θ ) in (0,π) (0,π), t 0,, then except for countable number of points the followings are true. (i) (6) lim N N n cos(θ n θ n ) lim N N sin(θ n θ n ) 0. n

13 (ii) (7) lim N N t (8) lim N N t (9) lim N N t CHIRP 3 n t cos (θ n θ n ) n n t sin (θ n θ n ) n (t ) (t ). n t sin(θ n θ n ) cos(θ n θ n ) 0. n N h (0) lim n t cos(θ N N t nθ n ) cos(θ (nh)θ (nh) ) n { 0 for h 0 and for, (t) h0 Proof of Theorem : () and () Note that, Q N P α N y(t)e i( αt βt ) ( A 0 ib 0 )e i[(α 0 α)t(β 0 β)t ] X(t)e i( αt βt ) y(t)(t N )e i( αt βt ) X(t)(t N )e i( αt βt ) ( A 0 ib 0 ( A 0 ib 0 ( A 0 ib 0 )(t N )e i[(α 0 α)t(β 0 β)t ] To get nd term in equation () we use the lemma and get e i[(α 0 α)t(β 0 β)t ] o p (N) )e i[(α 0 α)t(β 0 β)t ] )(t N )ei[(α 0 α)t(β 0 β)t ] Using lemma and bivariate Taylor series expansion the st term in equation () becomes

14 4 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 e i[(α 0 α)t(β 0 β)t ] [ e i[t(α 0 α 0 )t (β 0 β 0 i(α 0 α)t i(β 0 β)t i (α 0 α) t! e i[t(α 0 α )t (β 0 β i (β 0 β) t 4 i (α 0 α)(β 0 β)t 3 e i[t(α 0 α )t (β 0 β! ]! e i[t(α 0 α )t (β 0 β O(N) O p (N δ )O(N ) O p (N δ )O(N 3 ) O p(n δ )O p (N 3 ) O p (N δ )O p (N δ )O p (N 4 ) O p(n 4 δ )O p (N 5 ) O p (N) O p (N δ ) O p (N δ ) O p(n δ ) O p (N δ δ ) O p(n δ ) O p (N) O p (N min(δ,δ ) ) O p (N min(δ,δ ) ) where (α,β ) is a point on the line joining (α 0,β 0 ) and ( α, β) To calculate the 3rd term in equation () we need the following two observations. (i) N n t n [sin(α 0 t β 0 t X(t) & N n t n [cos(α 0 t β 0 t X(t) satisfy conditions for Central Limit Theorem (CLT), see Fuller [4], when α 0,β 0 are the true value as in the model. (ii) sup t n X(t)e i[αtβt ] 0 a.s.. It can be proved along the same line as α,β N n Kundu and Nandi [8]. Now we choose L large such that L min(δ,δ ) < 0. Then using again bivariate Taylor series expansion the 3rd term in equation () becomes [ X(t)e i( αt βt ) X(t) e i(tα 0t β 0 ) L l L l ( i( α α 0 )t) l e i(tα 0t β 0 ) l! L l l (i( α α 0 )t) k (i( β β 0 )t ) l k k k!(l k)! L (i( α α 0 )t) k (i( β β 0 )t ) L k k0 k!(l k)! ( i( β β 0 )t ) l l! e i(tα 0t β 0 ) e i(tα t β ) ] e i(tα 0t β 0 )

15 L O p (N ) L l L k0 k l l! O p(n l lδ )O p (N l CHIRP 5 L ) l l! O p(n l lδ )O p (N l l k!(l k)! O p(n k kδ )O p (N (l k) (l k)δ )O p (N k(l k) k!(l k)! O p(n k kδ )O p (N (L k) (L k)δ )O p (N k(l k) ) L O p (N ) l! [O L p(n lδ ) O p (N lδ L k0 l L O p (N ) k!(l k)! [O p(n kδ (L k)δ l l l k0 l k!(l k)! [O p(n kδ (l k)δ l k L k0 ) ) k!(l k)! [O p(n kδ (l k)δ L O p (N ) O p (N ) l! (N δ N δ ) l O p (N) L! (N δ N δ ) L L O p (N ) l! O p(n l min(δ,δ ) ) L! O p(n Lmin(δ,δ ) ) l L O p (N ) Op (N ) l! O p(n l min(δ,δ ) ) O p () O p (N ) l Then equation () becomes Q N ( A 0 ib 0 O p (N)( A 0 ib 0 ) k!(l k)! [O p(n kδ (L k)δ )[O p (N) O p (N δ ) O p (N δ ( A 0 ib 0 )o p (N) O p (N ) Bivariate Taylor series gives st term in equation () as X(t)(t N )e i( αt βt ) L l L l ( i( α α 0 )t) l e i(tα 0t β 0 ) l! X(t)(t N ) [ e i(tα 0t β 0 ) L l l (i( α α 0 )t) k (i( β β 0 )t ) l k k k!(l k)! L (i( α α 0 )t) k (i( β β 0 )t ) L k k0 k!(l k)! ( i( β β 0 )t ) l l! e i(tα 0t β 0 ) e i(tα t β ) ] e i(tα 0t β 0 )

16 6 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 X(t)(t N )e i(tα 0t β 0 ) L l L l L k0 l! O p(n l lδ )O p (N l k L 3 ) l l! O p(n l lδ )O p (N l l k!(l k)! O p(n k kδ )O p (N (l k) (l k)δ )O p (N k(l k) k!(l k)! O p(n k kδ )O p (N (L k) (L k)δ )O p (N k(l k) ) X(t)(t N )e i(tα 0t β 0 ) L l l k L l k!(l k)! [O p(n 3 kδ (l k)δ X(t)(t N )e i(tα 0t β 0 ) L k0 L l 3 ) l! [O p(n 3 lδ ) O p (N 3 lδ l k0 k!(l k)! [O p(n kδ (L k)δ l L k0 3 ) k!(l k)! [O p(n kδ (L k)δ k!(l k)! [O p(n 3 kδ (l k)δ X(t)(t N L )e i(tα 0t β 0 ) O p (N 3 ) l! (N δ N δ ) l O p (N ) L! (N δ N δ ) L X(t)(t N L )e i(tα 0t β 0 ) l! O p(n 3 l min(δ,δ ) ) L! O p(n Lmin(δ,δ ) ) l X(t)(t N L )e i(tα 0t β 0 ) O p (N 3 ) l! O p(n l min(δ,δ ) ) O p (N) X(t)(t N )e i(tα 0t β 0 ) o p (N 3 ) Again using Taylor series we get nd term in equation () as (t N )ei[(α 0 α)t(β 0 β)t ] (t N [ ) e i[t(α 0 α 0 )t (β 0 β 0 i(α 0 α)t i(β 0 β)t i (α 0 α) t! l e i[t(α 0 α )t (β 0 β i (β 0 β) t 4 i (α 0 α)(β 0 β)t 3 e i[t(α 0 α )t (β 0 β! ]! e i[t(α 0 α )t (β 0 β

17 CHIRP 7 O(N) i(α 0 α)o(n 3 ) i(α 0 α)o p (N δ )O p (N 4 ) i(α 0 α)(β 0 β)o(n 5 ) i(β 0 β)o(n 4 ) i(β 0 β)o p (N δ )O p (N 6 ) O p (N) i(α 0 α)[o p (N 3 ) O p (N 3 δ ) O p (N 3 δ O p (N δ ) O p (N δ ) o p (N ) i(α 0 α)[o p (N 3 ) O p (N 3 δ ) O p (N 3 δ For 3rd term in equation () we calculate (t N )e i[(α 0 α)t(β 0 β)t ] Then equation () becomes (t N )O p() O p (N) P α N ( A 0 ib 0 )i(α 0 α)[o p (N 3 ) O p (N 3 δ ) O p (N 3 δ ( A 0 ib 0 )O p (N) N o p (N ) o p (N 3 ) X(t)(t N )e i(α 0tβ 0 t ) X(t)(t N )e i(α 0tβ 0 t ) ( A 0 ib 0 )i(α 0 α)[o p (N 3 ) O p (N 3 δ α α 48 P α N Im[ N ] Q N α α 48 N Im ( A 0 ib 0 )i(α 0 α)[o p (N 3 ) O p (N 3 δ O p (N)( A 0 ib 0 ) X(t)(t N )e i(α 0tβ 0 t ) α α (α 0 α) (α 0 α)o p (N δ ) 48 X(t)(t N )e i(α 0tβ 0 t ) N Im O p (N)( A 0 ib 0 ) Now using Lemma the variance of Z 48 X(t)(t N )e i(α 0tβ 0 t ) Im N O p (N)( A 0 ib 0 )

18 8 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 is asymptotically same as asymptotic variance of LSE of α 0. So α α 0 O p (N ( δ ) ) where δ (0, ) then α α 0 O p (N ( δ ) ) if δ 4 N 3 ( α α 0 ) N(0,σ ) if δ > by CLT of stochastic process in Fuller [4] where 4 σ 384σ c A 0 B 0 variance of LSE of α 0 where c a(j). j is asymptotic Proof of Theorem : (3) N P β N y(t)(t N )e i( αt βt ) 3 X(t)(t N 3 )e i( αt βt) ( A 0 ib 0 ( A 0 ib 0 )(t N 3 )e i[(α 0 α)t(β 0 β)t ] Taylor series expansion gives nd term in equation (3) as (t N 3 )ei[(α 0 α)t(β 0 β)t ] [ (t N 3 ) e i[t(α 0 α 0 )t (β 0 β 0 i(α 0 α)t i(β 0 β)t i (α 0 α) t! e i[t(α 0 α )t (β 0 β i (β 0 β) t 4 i (α 0 α)(β 0 β)t 3 e i[t(α 0 α )t (β 0 β! ]! )(t N 3 )ei[(α 0 α)t(β 0 β)t ] e i[t(α 0 α )t (β 0 β O(N ) i(β 0 β)o(n 5 ) i(β 0 β)o p (N δ )O p (N 7 ) i(α 0 α)(β 0 β)o(n 6 ) i(α 0 α)o(n 4 ) i(α 0 α)o p (N δ )O p (N 5 ) O p (N ) i(β 0 β)[o p (N 5 ) O p (N 5 δ ) O p (N 5 δ O p (N 3 δ ) O p (N 3 δ ) o p (N 3 ) i(β 0 β)[o p (N 5 ) O p (N 5 δ ) O p (N 5 δ

19 CHIRP 9 Expanding using bivariate Taylor series st term in equation (3) becomes X(t)(t N 3 )e i( αt βt) L l L l ( i( α α 0 )t) l e i(tα 0t β 0 ) l! X(t)(t N 3 ) [ e i(tα 0t β 0 ) L l l (i( α α 0 )t) k (i( β β 0 )t ) l k k k!(l k)! L (i( α α 0 )t) k (i( β β 0 )t ) L k k0 k!(l k)! X(t)(t N 3 )e i(tα 0t β 0 ) L l L l L k0 l! O p(n l lδ )O p (N l k L 5 ) l ( i( β β 0 )t ) l l! e i(tα 0t β 0 ) e i(tα t β ) ] l! O p(n l lδ )O p (N l e i(tα 0t β 0 ) l k!(l k)! O p(n k kδ )O p (N (l k) (l k)δ )O p (N k(l k) k!(l k)! O p(n k kδ )O p (N (L k) (L k)δ )O p (N k(l k)3 ) X(t)(t N L 3 )e i(tα 0t β 0 ) l! [O p(n 5 lδ ) O p (N 5 lδ L l l k l k!(l k)! [O p(n 5 kδ (l k)δ X(t)(t N L 3 )e i(tα 0t β 0 ) L k0 l l k0 k!(l k)! [O p(n 3 kδ (L k)δ l L k0 5 ) 5 ) k!(l k)! [O p(n 3 kδ (L k)δ k!(l k)! [O p(n 5 kδ (l k)δ X(t)(t N L 3 )e i(tα 0t β 0 ) O p (N 5 ) l! (N δ N δ ) l O p (N ) L! (N δ N δ ) L X(t)(t N L 3 )e i(tα 0t β 0 ) l! O p(n 5 l min(δ,δ ) ) L! O p(n Lmin(δ,δ ) ) l X(t)(t N L 3 )e i(tα 0t β 0 ) O p (N 5 ) l! O p(n l min(δ,δ ) ) O p (N) X(t)(t N 3 )e i(tα 0t β 0 ) o p (N 5 ) l

20 0 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 To get 3rd term in equation (3) we calculate (t N 3 )e i[(α 0 α)t(β 0 β)t] Then equation (3) becomes P β N (A 0 ib 0 (t N 3 )O p() O p (N ) )i(β 0 β)[o p (N 5 ) O p (N 5 δ ) O p (N 5 δ ( A 0 ib 0 )O p (N ) o p (N 3 ) o p (N 5 ) N X(t)(t N 3 )e i(α 0tβ 0 t ) X(t)(t N 3 )e i(α 0tβ 0 t) ( A 0 ib 0 )i(β 0 β)[o p (N 5 ) O p (N 5 δ β β 45 P β N 4Im[ N ] Q N β β ( A 0 ib 0 )i(β 0 β)[o p (N 5 ) O p (N 5 δ X(t)(t N 45 3 )e i(α 0tβ 0 t ) N 4Im O p (N)( A 0 ib 0 ) β β (β 0 β) (β 0 β)o p (N δ ) 45 X(t)(t N N 4Im O p (N)( A 0 ib 0 ) 3 )e i(α 0tβ 0 t ) Now using Lemma variance of Z 45 X(t)(t N Im N 3 O p (N)( A 0 ib 0 ) 3 )e i(α 0tβ 0 t ) is asymptotically same as asymptotic variance of LSE of β 0.. So β β0 O p (N ( δ ) ) where δ (0, ) then β β0 O p (N ( δ ) ) if δ 4 N 5 ( β β0 ) N(0,σ ) if δ > 4 by CLT of stochastic process in Fuller [4] where σ 360σ c A 0 B 0 variance of the LSE of β 0. is the asymptotic

21 CHIRP Proof of Theorem 3: For covariance of Z,Z it is enough to find the quantities g k or g k g k lim N and lim N g k lim N lim N which are given below. N k N 4 n N k N 5 n N k N 4 n N k N 5 n (A 0 B 0 ) (n N 3 )(n k N ) (B 0 cos(α 0 n β 0 n ) A 0 sin(α 0 n β 0 n )) (B 0 cos(α 0 (n k) β 0 (n k) ) A 0 sin(α 0 (n k) β 0 (n k) )) 90 (A 0 B 0 ) (B0 cos(α 0 n β 0 n ) A 0 sin(α 0 n β 0 n )) (B 0 cos(α 0 (n k) β 0 (n k) ) A 0 sin(α 0 (n k) β 0 (n k) )) [ n 3 n ( N k) n( N 3 ) (N3 6 N k (A 0 B 0 ) (n N )((n k) N 3 ) (B 0 cos(α 0 n β 0 n ) A 0 sin(α 0 n β 0 n )) (B 0 cos(α 0 (n k) β 0 (n k) ) A 0 sin(α 0 (n k) β 0 (n k) )) 90 (A 0 B 0 ) (B0 cos(α 0 n β 0 n ) A 0 sin(α 0 n β 0 n )) (B 0 cos(α 0 (n k) β 0 (n k) ) A 0 sin(α 0 (n k) β 0 (n k) )) [ n 3 n ( N k) n(k Nk N 3 ) (N3 6 Nk 360cσ and both are for k 0 using (0) and 0 otherwise. So, the asymptotic A 0 B 0 covariance of Z,Z has the same order as the asymptotic covariance of the LSEs, see Kundu and Nandi [8]. References [] Abatzoglou, T. (986), Fast maximum likelihood joint estimation of frequency and frequency rate, IEEE Transactions on Aerospace and Electronic Systems, vol., [] Bai, Z.D., Rao, C.R., Chow, M., Kundu, D. (003), Efficient algorithm for estimating the parameters of superimposed complex exponential model, Journal of Statistical Planning and Inference, vol. 0, [3] Djuric, P.M. and Kay, S.M. (990), Parameter estimation of chirp signals, IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 38, 8-6. [4] Fuller, W. (996), Introduction to statistical time series, -nd edition, John Wiley and Sons, New York. [5] Gini, F., Montanari, M. and Verrazzani, L. (000), Estimation of chirp signals in compound Gaussian clutter: A cyclostationary approach, IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 48,

22 ANANYA LAHIRI, & DEBASIS KUNDU,3,4 & AMIT MITRA,3 [6] Kumaresan, R. and Verma, S. (987), On estimating the parameters of chirp signals using rank reduction techniques, Proceedings of st Asilomar Conference, , Pacific Grove, California. [7] Kundu, D., Bai, Z.D., Nandi, S., Bai, L. (0), Super efficient frequency estimation, Journal of the Statistical Planning and Inference, vol. 4, [8] Kundu, D. and Nandi, S. (008), Parameter estimation of chirp signals in presence of stationary noise, Statistica Sinica, vol. 8, [9] Nandi, S. and Kundu, D. (004), Asymptotic properties of the least squares estimators of the parameters of the chirp signals, Annals of the Institute of Statistical Mathematics, vol. 56, [0] Prasad, A., Nandi, S., Kundu, D. (00), An efficient and fast algorithm for estimating the parameters of sinusoidal signals, Sankhya, vol. 55,no., [] Rice, J.A. and Rosenblatt, M. (988), On frequency estimation, Biometrika, vol. 75, [] Richards, F.S.G. (96), A method of maximum likelihood estimation, Journal of the Royal Statistical Society, Ser. B, vol. 3, [3] Saha, S. and Kay, S. (00), Maximum likelihood parameter estimation of superimposed chirps using Monte Carlo importance sampling, IEEE Transactions on Signal Processing, vol. 50, [4] Tjalling, J.Y. (995), Historical development of the Newton-Raphson method, SIAM Review, vol. 37, [5] Vinogradov, I.M. (954), The method of trigonometrical sums in the theory of numbers, Interscience, Translated from Russian. Revised and annotated by K. F. Roth and Anne Davenport. Reprint of the 954 translation. Dover Publications, Inc., Mineola, NY, 004. [6] Weyl, H. (96), Ueber die Gleichverteilung von Zahlen mod. Eins, Math. Ann., vol. 77,