II. Nonparametric Spectrum Estimation for Stationary Random Signals - Non-parametric Methods -

Similar documents
Practical Spectral Estimation

Laboratory Project 2: Spectral Analysis and Optimal Filtering

Signal processing Frequency analysis

SPECTRUM. Deterministic Signals with Finite Energy (l 2 ) Deterministic Signals with Infinite Energy N 1. n=0. N N X N(f) 2

Massachusetts Institute of Technology Department of Electrical Engineering and Computer Science : Discrete-Time Signal Processing

1. Calculation of the DFT

Discrete Fourier Transform

Fourier Analysis of Signals Using the DFT

INTRODUCTION TO DELTA-SIGMA ADCS

ADSP ADSP ADSP ADSP. Advanced Digital Signal Processing (18-792) Spring Fall Semester, Department of Electrical and Computer Engineering

Complement on Digital Spectral Analysis and Optimal Filtering: Theory and Exercises

Periodogram and Correlogram Methods. Lecture 2

Parametric Method Based PSD Estimation using Gaussian Window

ECE 636: Systems identification

Chirp Transform for FFT

Complement on Digital Spectral Analysis and Optimal Filtering: Theory and Exercises

SNR Calculation and Spectral Estimation [S&T Appendix A]

Summary notes for EQ2300 Digital Signal Processing

L6: Short-time Fourier analysis and synthesis

EEO 401 Digital Signal Processing Prof. Mark Fowler

Centre for Mathematical Sciences HT 2017 Mathematical Statistics

System Identification

Multirate Digital Signal Processing

Design Criteria for the Quadratically Interpolated FFT Method (I): Bias due to Interpolation

6.435, System Identification

UNIVERSITY OF OSLO. Faculty of mathematics and natural sciences. Forslag til fasit, versjon-01: Problem 1 Signals and systems.

LAB 6: FIR Filter Design Summer 2011

convenient means to determine response to a sum of clear evidence of signal properties that are obscured in the original signal

Contents. Digital Signal Processing, Part II: Power Spectrum Estimation

DFT & Fast Fourier Transform PART-A. 7. Calculate the number of multiplications needed in the calculation of DFT and FFT with 64 point sequence.

Bispectral resolution and leakage effect of the indirect bispectrum estimate for different types of 2D window functions

Data Processing and Analysis

Computer Engineering 4TL4: Digital Signal Processing

-Digital Signal Processing- FIR Filter Design. Lecture May-16

Multimedia Signals and Systems - Audio and Video. Signal, Image, Video Processing Review-Introduction, MP3 and MPEG2

EEM 409. Random Signals. Problem Set-2: (Power Spectral Density, LTI Systems with Random Inputs) Problem 1: Problem 2:

EDISP (NWL3) (English) Digital Signal Processing DFT Windowing, FFT. October 19, 2016

IMPROVEMENTS IN MODAL PARAMETER EXTRACTION THROUGH POST-PROCESSING FREQUENCY RESPONSE FUNCTION ESTIMATES

INTRODUCTION TO DELTA-SIGMA ADCS

! Spectral Analysis with DFT. ! Windowing. ! Effect of zero-padding. ! Time-dependent Fourier transform. " Aka short-time Fourier transform

III.C - Linear Transformations: Optimal Filtering

Parametric Signal Modeling and Linear Prediction Theory 1. Discrete-time Stochastic Processes (cont d)

Statistical and Adaptive Signal Processing

Advanced Digital Signal Processing -Introduction

Autoregressive tracking of vortex shedding. 2. Autoregression versus dual phase-locked loop

Signals and Spectra - Review

E2.5 Signals & Linear Systems. Tutorial Sheet 1 Introduction to Signals & Systems (Lectures 1 & 2)

LINEAR-PHASE FIR FILTERS DESIGN

Basic Design Approaches

Random signals II. ÚPGM FIT VUT Brno,

Correlator I. Basics. Chapter Introduction. 8.2 Digitization Sampling. D. Anish Roshi

Introduction. Spectral Estimation Overview Periodogram Bias, variance, and distribution Blackman-Tukey Method Welch-Bartlett Method Others

EE 505 Lecture 10. Spectral Characterization. Part 2 of 2

FGN 1.0 PPL 1.0 FD

Review of spectral analysis methods applied to sea level anomaly signals

Part III Spectrum Estimation

The Discrete Fourier Transform

Distortion Analysis T

Parametric Signal Modeling and Linear Prediction Theory 1. Discrete-time Stochastic Processes

Time series models in the Frequency domain. The power spectrum, Spectral analysis

2. SPECTRAL ANALYSIS APPLIED TO STOCHASTIC PROCESSES

EE123 Digital Signal Processing

Lesson 1. Optimal signalbehandling LTH. September Statistical Digital Signal Processing and Modeling, Hayes, M:

1. FIR Filter Design

Experimental Fourier Transforms

covariance function, 174 probability structure of; Yule-Walker equations, 174 Moving average process, fluctuations, 5-6, 175 probability structure of

Class of waveform coders can be represented in this manner

Radar Systems Engineering Lecture 3 Review of Signals, Systems and Digital Signal Processing

Chapter 6: Nonparametric Time- and Frequency-Domain Methods. Problems presented by Uwe

Digital Speech Processing Lecture 10. Short-Time Fourier Analysis Methods - Filter Bank Design

! Introduction. ! Discrete Time Signals & Systems. ! Z-Transform. ! Inverse Z-Transform. ! Sampling of Continuous Time Signals

ELEN E4810: Digital Signal Processing Topic 11: Continuous Signals. 1. Sampling and Reconstruction 2. Quantization

sine wave fit algorithm

I. Signals & Sinusoids

Basics on 2-D 2 D Random Signal

Filter Analysis and Design

DSP. Chapter-3 : Filter Design. Marc Moonen. Dept. E.E./ESAT-STADIUS, KU Leuven

EE 435. Lecture 32. Spectral Performance Windowing

L29: Fourier analysis

Introduction to Biomedical Engineering

CONTENTS NOTATIONAL CONVENTIONS GLOSSARY OF KEY SYMBOLS 1 INTRODUCTION 1

The Discrete Fourier Transform (DFT) Properties of the DFT DFT-Specic Properties Power spectrum estimate. Alex Sheremet.

ELEG 305: Digital Signal Processing

Representation of Digital Signals

representation of speech

Review of Fundamentals of Digital Signal Processing

Chapter 7: Filter Design 7.1 Practical Filter Terminology

Problem Sheet 1 Examples of Random Processes

Lecture 4 - Spectral Estimation

EE482: Digital Signal Processing Applications

8 The Discrete Fourier Transform (DFT)

LABORATORY 3 FINITE IMPULSE RESPONSE FILTERS

Lecture 17: variance in a band = log(s xx (f)) df (2) If we want to plot something that is more directly representative of variance, we can try this:

Speech Signal Representations

The Selection of Weighting Functions For Linear Arrays Using Different Techniques

Filter Design Problem

Spectral Analysis. 1.1 Introduction ANDERS I. ERIKSSON

Frequency estimation by DFT interpolation: A comparison of methods

Department of Electrical and Computer Engineering Digital Speech Processing Homework No. 6 Solutions

Nonparametric and Parametric Defined This text distinguishes between systems and the sequences (processes) that result when a WN input is applied

Transcription:

II. onparametric Spectrum Estimation for Stationary Random Signals - on-parametric Methods - - [p. 3] Periodogram - [p. 12] Periodogram properties - [p. 23] Modified periodogram - [p. 25] Bartlett s method - [p. 30] Blackman-Tukey procedure - [p. 33] Welch s procedure - [p. 36] Minimum Variance MV) Method - [p. 42] MultiTaper Method MTM) 04/01/09 EC4440.SpFY09/MPF -Section II 1

-Goal: To estimate the Power Spectral Density PSD) of a wss process -Two main types of approaches eist: -on parametric methods: do not rely on data model -based on FT and FFT, or -Parametric methods: -based on different parametric models of the data: -Moving Average: MA, -AutoRegressive: AR, -AutoRegressive & Moving Average: ARMA -based on subspace methods 04/01/09 EC4440.SpFY09/MPF -Section II 2

I. Classical methods use FT of the data sequence or its correlation sequence 1) Periodogram Easy to compute Limited ability to produce accurate PS estimate Recall PSD defined as: jω ) ) P e = k = R k e jωk In practice: R k) is approimated with estimated correlation sequence R k ) 04/01/09 EC4440.SpFY09/MPF -Section II 3

Recall: Biased correlation estimate of n), n=0, -1, defined as 1 k 1 * Rˆ k) = + k) ) 0 k< 1 k= 0 ˆ j jωk P e = R k e L ω ) ˆ ) L k = L L 10% PSD estimate sometimes known as correlogram Assume L=-1 1 j jωk Pˆ e = R k e ω ) ˆ ) k= + 1 periodogram 04/01/09 EC4440.SpFY09/MPF -Section II 4

Recall 1 * Rˆ k) = k )* k) for k ) 0 0 k< 1 Define: DFT ) = ) ) ) n n w n X k ˆ 1 P e = FT k * k 1 jω) jω) * = X e X e jω ) * ) ) Periodogram defined in terms of data directly 1 P e X e ˆ j j ω) ω = ) 2 04/01/09 EC4440.SpFY09/MPF -Section II 5

How to Compute the Periodogram Rectangular window of length DFT ) = ) ) ) n n w n X k j πk ) ) ) 2 1 2 ˆ =, = 0,... 1 P e X k k Zero padding n) in the time-domain increases the number of frequency PSD points available 04/01/09 EC4440.SpFY09/MPF -Section II 6

Eample: n) white noise ) Rk = j ) Pe ω = [Hayes] 04/01/09 EC4440.SpFY09/MPF -Section II 7

Average behavior of the periodogram: 1 { ˆ jω )} ˆ jkω E P e = E R k) e + 1 = 1 k= + 1 { ˆ )} E R k e 1 k 1 ER k = E + k { )} * ) ) = 0 jω k k = R ˆ k ) k k = R k) wb k) wb k) = 0 ow 04/01/09 EC4440.SpFY09/MPF -Section II 8

1 j jω k { ˆ ω )} ˆ ) k= + 1 { } E P e = E R k e k { )} k ER k = R k) wb k), wb k) = 0 ow 1 { ˆ jω )} ) ) B k= + 1 ) E P e = R k w k e jωk 1 E P e P e W e 2π { ˆ j ω )} j )* j ) = ω ω B 1 sin ω /2) sin ω/ 2) 2 04/01/09 EC4440.SpFY09/MPF -Section II 9

Eample: n) 1 tone in white noise ) = sin ω + φ) + ) n A n v n 0 { jω } 1 jω jω EP e ) = Pe ) We B ) 2π 2 2 A j 0) j 0) v W B e ωω + ) W B e ωω = σ + + ) 4 [Hayes] 04/01/09 EC4440.SpFY09/MPF -Section II 10

Eample: set of tones in white noise n) n) = sin2π*150*t) + 2*sin2*π*140*t) + 0.1*randnsizet)); f s =1KHz Obtained using the MATLAB function periodogram.m periodogramn,[],'twosided',1024,fs); plots the two sided periodogram note: better to stick to one-sided when dealing with real-data) [Mathworks] 04/01/09 EC4440.SpFY09/MPF -Section II 11

Properties of the Periodogram 1) Periodogram is biased when dealing with finite windows 2) Periodogram is asymptotically unbiased i.e., lim j j E ω P e ) ω + = P e ) 3) Periodogram has limited ability to resolve closely spaced narrowband components limiting factor due to the rectangular window, the shorter the data length, the worse off) not good) 4) Values of the periodogram spaced in frequency by integer multiples of 2π/ are approimately uncorrelated. As data length increases, rate of fluctuation in the periodogram increases not good). 5) Variance of the periodogram does not decrease as data length increases not good). In statistical terms, it means the periodogram is not a consistent estimator of the PSD. evertheless, the periodogram can be a useful tool for spectral estimation in situations where SR is high, and especially if the data is long. 04/01/09 EC4440.SpFY09/MPF -Section II 12

n = sin2*pi*150*t) + 2*sin2*pi*140*t) + 0.1*randnsizet));; fs=1000; nfft=1024; 100 samples Comments Periodogram has limited ability to resolve closely spaced narrowband components limiting factor due to the rectangular window, the shorter the data length, the worse off) not good) 67 samples [MathWorks] 04/01/09 EC4440.SpFY09/MPF -Section II 13

Comments: Variance of the periodogram does not decrease as the data length increases not good). For white noise n) one can show [Therrien, p.590] jω P e ) lim var 0 + jω 4 var P e ) = σ [Hayes] 04/01/09 EC4440.SpFY09/MPF -Section II 14

Comments: As data length increases, rate of fluctuation in the periodogram increases not good). [Kay] 04/01/09 EC4440.SpFY09/MPF -Section II 15

Connections between Periodogram and filterbank Periodogram can be viewed as obtained by using a filterbank structure. Recall periodogram defined for n), n=0, -1, as: 1 ) ) ) DFT ) j2 πkp/ = = ) p= 0 n n w n X k pe j 2πk ) ) 1 P e X ) k k 2 ˆ =, = 0,... 1 04/01/09 EC4440.SpFY09/MPF -Section II 16

For n) defined over a finite time window {n-+1), n)} 1 ) ) ) DFT ) = = j2 πkp/ ) p= 0 n n w n X k n pe 1 p= 0 h p) ) ) ) DFT n = n w n X k) = n ph ) p) k=0, -1 04/01/09 EC4440.SpFY09/MPF -Section II 17

For n) defined over a finite time window {n-+1), n)} 1 ) ) ) DFT n = n w n X k) = n ph ) p) p= 0 j πk ) ) ) 2 Pˆ e = X k 2 1 k=0, -1 n) Filter H z) 1 hp ) = e - j2 π kp / 2 2 π )) ˆ j k P e Periodogram DFT filterbank 04/01/09 EC4440.SpFY09/MPF -Section II 18

Filter H z) 1 h0 p) =, k = 0 2 2 π 0 )) ˆ j P e n) Filter H z) 1 h p = e k = - j2 π p/ 1 ), 1 2 2 π )) ˆ j P e Filter H z) 1 h p = e k = - j2 π p 1)/ 1 ), 1 04/01/09 EC4440.SpFY09/MPF -Section II 19 2 2 1) )) ˆ j P e π j πk ) ) ) 2 Pˆ e = X k 2 1

n) Filter H z) 1 h0 p) =, k = 0 Filter H z) 1 h p = e k = - j2 π p/ 1 ), 1 Filter H z) 1 h p = e k = 2 2 π 0 )) ˆ j P e - j2 π p 1)/ 1 ), 1 04/01/09 EC4440.SpFY09/MPF -Section II 20 2 Lowpass filter 2 π )) ˆ j P e Modulation term 2 2 1) )) ˆ j P e π j πk ) ) ) 2 Pˆ e = X k 2 1

Filter H z) 1 h0 p) =, k = 0 2 2 π 0 )) ˆ j P e Lowpass filter n) Filter H z) h p h p e k - j2 π p/ 1 ) = 0 ), = 1 2 2 π )) ˆ j P e Filter H z) h p = h p e k = - j2 π p 1)/ 1 ) 0 ), 1 Modulation term 2 2 1) )) ˆ j P e π j πk ) ) ) 2 Pˆ e = X k 2 1 04/01/09 EC4440.SpFY09/MPF -Section II 21

Periodogram DFT matri formulation X 1 1 1 0) j2 / j2 1)/ X 1) π π 1 e e = 2 j2 π 1)/ j2 π 1) / X 1) 1 e e 1/ n ) 1/ n 1) 1/ n + 1) j πk ) ) ) 2 Pˆ e = X k 2 1 May be viewed as a rectangular window applied to data before FT operation 04/01/09 EC4440.SpFY09/MPF -Section II 22

2. Modified Periodogram can apply a general window to n) prior to computing the periodogram periodogram etension) + ω 1 P e ) = n) w R n) e ˆ j jnω n = 2 ˆ 1 E P e = P e * WR e 2π jω ) jω ) jω ) 2 ote: amount of smoothing is determined by the window type Matlab implementation: Hs = spectrum.periodogram'hamming'); create modified periodogram object psdhs,n,'fs',fs,'fft',1024,'spectrumtype', onesided') plot 04/01/09 EC4440.SpFY09/MPF -Section II 23

Eample: n = sin2*pi*150*t) + 2*sin2*pi*140*t) + 0.1*randnsizet));; fs=1000; nfft=1024; data length=100. Hamming window Rectangular window ote: lower Hamming window sidelobes, but wider mainlobes with the Hamming window than the rect. window [MathWorks] 04/01/09 EC4440.SpFY09/MPF -Section II 24

3. Bartlett s Method average several periodograms rectangular window applied to each) leads to consistent estimate of the PSD Define: K ˆ 1 P ˆ B e = P e K jω ) k ) jω ) k = 1 where K: number of segments in the data. K L 1 jω 1 jnω PB e ) = n+ k 1) L) e KL k= 1 n= 0 2 n) L points L points L points n 1 n) 2 n) K n) n n n 04/01/09 EC4440.SpFY09/MPF -Section II 25

var ote: ˆ jω ) 1 ˆ k var ) jω P ) B e P e K 1 jω P ) e K Eamples: white noise [Hayes] two sinusoids in noise 04/01/09 EC4440.SpFY09/MPF -Section II 26

Compare: basic periodograms Periodogram of white noise with variance 1. a) Overlay 50 plots, =64 & b) average c) Overlay 50 plots, =128 & d) average e) Overlay 50 plots, =256 & f) average [Hayes, Fig. 8.9] 04/01/09 EC4440.SpFY09/MPF -Section II 27

Bartlett estimates Spectrum estimation of white noise with unit variance a) 50-plot overlay, =512, b) ensemble average. c) Overlay of 50 Bartlett estimates with K=4 & L=128, =512, d) ensemble average. e) Overlay of 50 Bartlett estimates with K=8 & L=64, =512, f) ensemble average. K: number of segments L: segment length [Hayes, Fig. 8.14] 04/01/09 EC4440.SpFY09/MPF -Section II 28

Spectrum estimation of 2 tones in white noise a) 50-plot overlay, =512, b) ensemble average. c) Overlay of 50 Bartlett estimates with K=4 & L=128, =512, d) ensemble average. e) Overlay of 50 Bartlett estimates with K=8 & L=64, =512, f) ensemble average. ote: Spectral peaks broadening, why? π φ) π φ ) n) = Asin 0.2 n+ + sin 0.25 n+ + vn ) 1 2 [Hayes, Fig. 8.15] 04/01/09 EC4440.SpFY09/MPF -Section II 29

4. Blackman-Tukey BT) Procedure apply windowing to correlation samples prior to transformation to reduce variance of the estimator M ˆ jω P ) ) ˆ ) BT e = w k R k e k= M jωk where the window wk) length is much smaller than data length, wk) = 0 for k > M where M << Maimum recommended window length M</5 04/01/09 EC4440.SpFY09/MPF -Section II 30

n ) = 10 ep2 π j0.15) n) + 20 ep2 π j0.2) n) + wn ) wn )~ 0,0.1) Data length=20 Periodogram BT PSD estimate Triangular window, length=4 [Kay, Fig.4.9] 04/01/09 EC4440.SpFY09/MPF -Section II 31

n ) = 10 ep2 π j0.15) n) + 20 ep2 π j0.2) n) + wn ) Periodogram Data length=100 wn )~ 0,0.1) BT PSD estimate Triangular window, length=20 [Kay, Fig.4.10] 04/01/09 EC4440.SpFY09/MPF -Section II 32

5. Welch s Procedure Modification to the Bartlett procedure: Averaging modified periodograms Dataset split into K possibly overlapping segments of length L window applied to each segment all K periodograms are averaged where K ˆ 1 P ˆ w e = P e K jω ) k ) jω ) k = 1 ˆ 1 P w n n e L 1 ) ) ) 2 k k jωn = ) n= 0 window data n) located in K th segment MATLAB Hs = spectrum.welch'rectangular', segmentlength, overlap %); define Welch spectrum object psdhs,n,'fs',fs,'fft',512); plot 04/01/09 EC4440.SpFY09/MPF -Section II 33

[Hayes] ote: what do we gain with the Welch s method? Reduction in spectral leakage that takes place thru the sidelobes of the data window 04/01/09 EC4440.SpFY09/MPF -Section II 34

Performance comparison variability resolution Figure of merit Periodogram 1 0.89 2π/) 0.89 2π/) Bartlett 1/K 0.89 K2π/) 0.89 2π/) Welch 50% overlap, triang. window) [Table 8.7, Hayes] 9/8K 1.28 2π/L) 0.72 2π/) BT 2M/2 0.64 2π/M) 0.43 2π/) K: number of data segments : overall data length L: data section length M: BT window length All schemes limited by data length Definitions: { jω var P )} e variability normalized variance) 2 jω E P e ) { } figure of merit variability resolution resolution should be as small as possible. mainlobe width at its 1/2 power 6dB) down level 04/01/09 EC4440.SpFY09/MPF -Section II 35

6. Minimum Variance MV) Spectrum Estimate Capon PSD PSD estimated by a filterbank where each filter of length P<= data length) adapts itself to data characteristics, i.e., - has center frequency f i spread between 0 and f s ) - has bandwidth Δ P filters spread over [0,f s ] Δ 1/f s ) - depends on R - designed so components at f i are passed without distortion and rejects maimum amount of out-of-band power ˆ jω ) P PMV e =, H 1 e R e e= e e jω j P 1) ω T [1,,..., ] [Manolakis, p.474] 04/01/09 EC4440.SpFY09/MPF -Section II 36

MATLAB implementation ˆ ω) P, [1, ω,..., 1) ω P ] MV e = e= e e H 1 e R e ote: j j j P T H 1 H H 1 H 1 H e R e= e UΛ U ) e= e UΛ U e H 1 H H = ) ) e U Λ e U H jω j P 1) ω eu= [1, e,..., e ] u1 u [ u u ] = FT ),, FT ) 1 P P 04/01/09 EC4440.SpFY09/MPF -Section II 37

H 1 H 1 H H = ) Λ ) e R e e U e U [ FT u ),, FT u )] 1 1 [ FT u ),, FT u )] 1 = 1/ λ ) FT u ) + +1/ λ ) FT u ) H = Λ e R e 2 2 1 1 P P P 1 2 = 1/ λk ) FT uk ) k=1 P P H Pˆ MV jω e ) = P k=1 1/ λ ) FT u ) k P k 2 04/01/09 EC4440.SpFY09/MPF -Section II 38

Application: n e e wn wn j2 0.1) 2 0.12 ) 10 π n j π n = + ) + ), )~ 0,1) Filter order M=P Data length =500 100 averaged trials MV: solid line Bartlett method: small dashed line All-pole: large dashed line Δ f = 0.02 Wind. Length ~ 50 [Manolakis, e. 9.5.1] 04/01/09 EC4440.SpFY09/MPF -Section II 39

MATLAB Implementation R=covar,P); [v,d]=eig ; UI=diaginvd)+eps); VI=absfftv,nfft)).^2; P=10log10P)-10log10VI*UI); In practice, - pick P < - zeropad FFT to insure accurate measurements 04/01/09 EC4440.SpFY09/MPF -Section II 40

7. MultiTaper Method MTM) Spectrum Estimate Thompson) Periodogram can be viewed as obtained by using a filterbank structure, using rectangular windows). Multitaper method MTM) can be viewed the same way, with different filters, i.e., different windows. These optimal FIR filters are derived from a set of sequences known as discrete prolate spheroidal sequences. A set of independent estimates of the power spectrum is computed, by premultiplying the data by orthogonal tapered windows designed to minimize spectral leakage. [Victorin] [Mathworks, statistical signal processing, spectral analysis, non parametric methods] 04/01/09 EC4440.SpFY09/MPF -Section II 41

7. MultiTaper Method MTM) cont PSD variance is reduced by averaging over different tapered windows using the full data. Since the data length is not shortened, the overall PSD bias is smaller than that obtained when splitting data in segments. MTM method provides a time-bandwidth parameter, W, with which to balance the variance and resolution. W is directly related to the number of tapers used to compute the spectrum. There are always 2*W-1 tapers used to form the estimate. As W increases, - there are more estimates of the power spectrum, and the variance of the estimate decreases, - each estimate ehibits more spectral leakage i.e., wider peaks) and the overall spectral estimate is more biased. For each data set, there is usually a value for W that allows an optimal trade-off between bias and variance. [Mathworks, statistical signal processing, spectral analysis, non parametric methods] 04/01/09 EC4440.SpFY09/MPF -Section II 42

Periodogram fs = 1000Hz; =1000; 2 sines, Amplitude = [1 2]; Sinusoid frequencies f = [150;140]; Hs1 = spectrum.mtm4,'adapt'); psdhs1,n,'fs',fs,'fft',1024) MTM, W=4 MTM, W=3/2 04/01/09 EC4440.SpFY09/MPF -Section II 43

References: [Manolakis] [Hayes] Statistical digital signal processing and modeling, M. M. Hayes, 1996, Wiley. [Kay] Modern spectral estimation, S. M. Kay, 1988, Prentice Hall. [Victorin] A. Victorin, Multi-Taper Method for Spectral Analysis and Signal Reconstruction of Solar Wind Data, MS Plasma Physics, Sept. 2007, Royal Institute of Technology, Sweeden, http://www.ee.kth.se/php/modules/publications/reports/2007/trita- EE_2007_054.pdf 04/01/09 EC4440.SpFY09/MPF -Section II 44