Complications Flashcards
smoothing the periodogram, inconsistency, truncating the autocovariance function, smoothing the raw periodogram, spectral window, spectral linkage, harmonics
Distribution of DFT for a White Noise Process
-if {εt} is a white noise process then:
Re[εj^] ~ N(0,σε²/2)
-and
Im[εj^] ~ N(0,σε²/2)
I(fj) for a White Noise Process
I(fj) = |εj^|² = Re(εj^)² + Im(εj^)²
Distribution of I(fj) for a White Noise Process
-since I(fj) is the sum of the squares of two random variables:
2/σε² I(fj) ~ χ2²
-for 1≤j∞
Var[I(fj)] for an AR(1) Process
-if {Xt} is an AR(1) process then:
Var[I(fj)] = κ σε^4
-for some constant κ
Var[I(f)] in general
-in general, Var[I(f)] is constant as n increases
Inconsistency
Definition
- in general, Var[I(f)] is constant as n increases
- I(f) is an inconsistent estimate for the power spectrum 𝓘(f)
- as n increases we still attempt to estimate I(f) at n frequencies
- therefore, although we have more data, the amount of data per frequency remains the same
Preferred Alternative to
Inconsistency
- we would prefer an estimate which gets more reliable as we gather more data
- there are a number of ways to do this e.g. truncating the autocovariance function and smoothing the raw periodogram
Spectral Density in Terms of the Lag-k Sample Autocovariance
I(fj) = g0 + Σ gk cos(2πfj k)
(+ gn/2 (-1)^j if n is even)
-sum from 1≤k
Truncating the Autocovariance Function
Definition
-if we do not assume {Xt} to be periodic then the number of pairs{Xt, Xt+k} decreases as k increases so gk is less reliable for larger lags
-hence we might use:
𝓘t(fj) = λ0g0 + 2 Σ λk gk cos(2πfj k)
-where subscript t denotes that this is the truncation method
-sum from k=1 to m with m
Truncating the Autocovariance Function
Tukey Window
λk = 1/2 [1 + cos(πk/m)]
k=0,1,…,m
Truncating the Autocovariance Function
Parzen Window
λk = 1 - 6(k/m)² + 6(k/m)³ for k=0,1,…,m/2
λk = 2 (1 - k/m)³ for k = m/2 +1, …, m
Truncating the Autocovariance Function
Choosing m
- increasing m leads to the raw periodogram which is too noisy
- choosing m too small over-smooths, low variance but high bias
- a rule of thumb is to take m=√n/2
Smoothing the Raw Periodogram
Outline
- this technique is called Daniel smoothing, it is the default method in R
- it is based on the assumption that the power spectrum, 𝓘(f), will be ‘similar’ for values of f ‘close’ to each other
Smoothing the Raw Periodogram
Definition
-let w = 2w* + 1 be an odd, positive integer and define:
𝓘s(f) = 1/w Σ I(fj+k)
-where subscript s denotes that this is the smoothing method
-sum from k=-w* to k=w*
Smoothing the Raw Periodogram
Var[𝓘s(f)]
- recall that I(f) is asymptotically unbias but has constant variance
- also, I(fj) is asymptotically independent of I(fk) if j≠k
- therefore Var[𝓘s(f)] is of order 1/w
Smoothing the Raw Periodogram
E[𝓘s(f)]
E[𝓘s(f)] = 1/w Σ E[I(fj+k)]
≈ 1/w Σ 𝓘s(f)
-sum from k=-w* to k=w*
Smoothing the Raw Periodogram
Bias
-the smoothing estimate 𝓘s(f) is bias since its expectation only equals 𝓘s(f) if the power spectrum is linear on the interval:
[fj-w,fj+w]
-therefore we need to balance large w (small variance) with small w (small bias)
Smoothing the Raw Periodogram
Choosing w
- rule of thumb w≈√n
- or alternatively, use repeated application of the Daniel smoother with small values of w
Equivalence of Smoothing the Periodogram and Truncating the Autocovariance Function
-the Daniel smoothing method is equivalent to the lag window:
λk = 1 if k=0
λk = sin(wkπ/n)/msin(kπ/n) if k=1,2,…,n-1
Spectral Window
Definition
-the spectral window, K(f), is the DFT of the lag window
K(fj) = 1/√n Σ λk e^(-2πifj k)
-sum from k=0 to k=n-1
-it can be used to compare truncating the autocovariance function and smoothing the periodogram
Spectral Window
Calculating λk
λk = 1/√n Σ K(fj) e^(2πifj k)
-sum from j=0 to j=n-1
Spectral Window
Truncated Estimate of the Power Spectrum
-write out the definition of 𝓘t(f)
-sub in λk in terms of the spectral window
=>
𝓘t(f) = 1/n Σ K(fj’) I(fj-fj’)
-sum from j’=0 to j’=n-1
Spectral Window
Equivalence of Procedures
- from the truncated estimate of the power spectrum in terms of K:
- both our estimation procedures are equivalent to smoothing the periodogram using a kernel function K
- there can be a sharp cut-off in the frequency domain (Daniel smoothing) or the time domain (truncation) but not both
Spectral Leakage
Outline
- so far we have assumed that all periodic components are sine waves existing at one of the Fourier frequencies, fj=j/n
- but this need not be the case
Spectral Leakage
Xj^ for Xt=sin(2πft) for f=fk
Xj^ = √n/2i if j=k Xj^ = -√n/2i if j=-k Xj^ = 0, else
Spectral Leakage
Xj^ for Xt=e^(2πift)
Xj^ = √n if f=fj Xj^ = 0 if f=fk, k≠j Xj^ = 1/√n 1-e^[2π(f-fj)n]/1-e^[2πi(f-fj)], else
Harmonics
Outline
-what if there is a periodic component that is not a since wave
Harmonics
Definition
-let δjt be the contribution to Xj^ from time t:
Xj^ = Σ δj,t
-sum from t=0 to t=n-1
-from the definition of the DFT:
δj,t = 1/√n Xt e^(-2πifj t)
-the contribution to X2j^ from time t is:
δ2j,t = 1/√n Xt e^(-2πi2fj t)
= δj,t e^(-2πifj t)
…
δmj,t = δj,t [e^(-2πifj t)]^(m-1)
-where Xj^ is the fundamental and X2j^, X3j^, …are harmonics
Sine vs. Non-Sine Periodic Components
- if Xt contains a component, say sin(2πifk t) for some Fourier frequency fk, then it contributes only to Xk^
- if there is a non-trigonometric periodic component at frequency fj, its effects will also be felt at f2j, f3j, … - the harmonics of the fundamental frequency fj