The Analysis of Non-Stationary Time Series with Time Varying Frequencies using Time Deformation The Analysis of Non-Stationary Time Series with Time Varying Frequencies using Time Deformation Southern Methodist University Research Group headed by Henry Gray and Wayne Woodward
Weakly Stationary Process
Weakly Stationary Process Models for Stationary Processes Models for Stationary Processes - AR(p), ARMA(p,q), etc - AR(p), ARMA(p,q), etc
Power Spectrum Power Spectrum F
Power Spectrum Power Spectrum Most important tool for analyzing frequency content of stationary time series F
Power Spectrum Power Spectrum Most important tool for analyzing frequency content of stationary time series - Fourier transform of autocovariance F
Note: Spectrum not really appropriate for series with time- varying frequencies
Data with Time-Varying Frequency Behavior Data with Time-Varying Frequency Behavior
Cyclical behavior is continuously changing across time, and the autocovariance will depend on both lag and time.
What Happens if We Ignore the Fact that Frequencies are Changing in Time? What Happens if We Ignore the Fact that Frequencies are Changing in Time?
traditional spectral analysis is not really a suitable tool to detect these cycles What Happens if We Ignore the Fact that Frequencies are Changing in Time? What Happens if We Ignore the Fact that Frequencies are Changing in Time?
traditional spectral analysis is not really a suitable tool to detect these cycles ARMA(p,q) models not appropriate What Happens if We Ignore the Fact that Frequencies are Changing in Time? What Happens if We Ignore the Fact that Frequencies are Changing in Time?
traditional spectral analysis is not really a suitable tool to detect these cycles ARMA(p,q) models not appropriate - forecasts from such models are poor What Happens if We Ignore the Fact that Frequencies are Changing in Time? What Happens if We Ignore the Fact that Frequencies are Changing in Time?
traditional spectral analysis is not really a suitable tool to detect these cycles ARMA(p,q) models not appropriate - forecasts from such models are poor whitening filters may not whiten What Happens if We Ignore the Fact that Frequencies are Changing in Time? What Happens if We Ignore the Fact that Frequencies are Changing in Time?
traditional spectral analysis is not really a suitable tool to detect these cycles ARMA(p,q) models not appropriate - forecasts from such models are poor whitening filters may not whiten standard filtering methods may be ineffective What Happens if We Ignore the Fact that Frequencies are Changing in Time? What Happens if We Ignore the Fact that Frequencies are Changing in Time?
Current Methods for Analyzing Data with Time-Varying Frequencies Current Methods for Analyzing Data with Time-Varying Frequencies windowed Fourier transforms (Gabor) smoothing spline ANOVA (Guo, et al., JASA, Sept. 2003) wavelet analysis autoregression with time-varying parameters
An Alternative Approach: TIME DEFORMATION (Gray and Zhang, 1988)
An Alternative Approach: TIME DEFORMATION (Gray and Zhang, 1988) -transform time ( instead of X ( t ))
An Alternative Approach: TIME DEFORMATION (Gray and Zhang, 1988) -transform time ( instead of X ( t )) -continuous case
M-Stationary Process (Gray and Zhang, 1988) M-Stationary Process (Gray and Zhang, 1988)
M-Stationary Process (Gray and Zhang, 1988) M-Stationary Process (Gray and Zhang, 1988) Note: is called the M-autocorrelation
Note M-stationary processes can be viewed as stationary processes by time deformation
Note M-stationary processes can be viewed as stationary processes by time deformation
Note M-stationary processes can be viewed as stationary processes by time deformation
Continuous Euler Process (EAR) (Gray and Zhang, 1988) The p th order continuous Euler process is given by
Continuous Euler Process (EAR) (Gray and Zhang, 1988) The p th order continuous Euler process is given by DUAL PROCESS:
M-Spectrum M-Spectrum Let X(t) be an M-stationary process. Then the M- spectrum is given by
Discrete M-Stationary Process
Discrete Euler Process – EAR(p) (Gray, Vijverberg, Woodward, 2003) Discrete Euler Process – EAR(p) (Gray, Vijverberg, Woodward, 2003)
Dual Process -- Y k Dual Process -- Y k Notes: this says we can induce stationarity by sampling properly i.e. we should sample at h k, then index on k the resulting process, Y k, is AR( p ) estimation of coefficients, i, M-autocorrelation, and M-spectrum accomplished using Y k
Discrete EARMA(p,q) Processes -- Dual is ARMA(p,q)
EARMA M- Spectral Estimator
Practical Issues most data sets are observed at equally-spaced time points methods are used to obtain data at “Euler time points”, i.e. at t = h k -scientific sampling -interpolation -Kalman filter “realization offset” of the observed realization needs to be estimated
Realizations with Different Realization Offsets ( h j Realizations with Different Realization Offsets ( h j )
M-Periodic Function
A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all t
M-Periodic Function A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all tExample:
M-Periodic Function A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all tExample:
M-Periodic Function A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all t
M-Periodic Function A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all t Example:
M-Periodic Function A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all t
M-Periodic Function A function g(t) is M-periodic with M-period if is the minimum value such that g(t) = g(t ) for all t
Instantaneous Period: Instantaneous Period:
Instantaneous Period Instantaneous Period Terminology
Instantaneous Frequency Instantaneous Frequency Terminology
Instantaneous Period Instantaneous Period Instantaneous Frequency Instantaneous Frequency Discrete Case Discrete Case Terminology
Note: Note: ip(t;f*) is in “calendar” or “regular” time and therefore f(t;f*) is in cycles per sampling unit based on the units of the original equally-spaced data set.
A Discrete Euler Model A Discrete Euler Model
Dual Process:
Plots for Original Data Original Data Sample ACF Spectral Estimators
Plots for Dual Data Plots for Dual Data M-AutocorrelationDual Data M-Spectrum
Forecasts Forecasts
Instantaneous Spectrum
Instantaneous Autocorrelation: X (l, t)
- depends on l and t
Instantaneous Autocorrelation: X (l, t) M-Autocorrelation
Instantaneous Autocorrelation: X (l, t) M-Autocorrelation
Instantaneous Autocorrelation: X (l, t) M-Autocorrelation
Large Brown Bat Echolocation Signal
First 100 and Last 60 Points First 100 and Last 60 Points
Sample Autocorrelation and Spectrum
Plots for Dual Data M-frequency
Forecasts of Points 40–80 Forecasts of Points 40–80
Forecasts of Last 24 Points Forecasts of Last 24 Points
Instantaneous Spectrum
First 100 and Last 60 Points First 100 and Last 60 Points
Animation
“Snapshots” “Snapshots” t = 1 t = 14 t = 115 t = 381
Gabor Transform Gabor Transform
Continuous Wavelet Transform
Filtering Example Filtering Example Original Data Butterworth Filtered Data
Filtering Example Filtering Example Original Data Butterworth Filtered Data
Final Filtering Results Final Filtering Results
Extension: G( )-Stationary Process
Effect of Varying Effect of Varying = - 1 = 0 = 1 = 1.5
Applications Applications Seismic dispersion curves -known to be a compacting signal Arterial blood flow data -Doppler signal -lesions create a second (higher frequency) Doppler signal Bat identification by echolocation signal Musical transients …
Instantaneous Autocorrelation Instantaneous Autocorrelation
Filtering Example Filtering Example
Instantaneous Autocorrelation Instantaneous Autocorrelation
G( )-Stationary Process G( )-Stationary Process Let X ( t ) be a stochastic process defined for - We call { X(t); t (0, )} a G( )-stationary process Dual Process : Y ( u) = X(t), where Y ( u ) is a stationary process.
G( )-Stationary Process G( )-Stationary Process Let X ( t ) be a stochastic process defined for - We call { X(t); t (0, )} a G( )-stationary process Dual Process : Y ( u) = X(t), where Y ( u ) is a stationary process.
Filtering Example Filtering Example
Simulated Euler(12) Data
Simulated Data - Dual
Simulated Data – Instantaneous Spectrum
Simulated Data – Gabor Transform
M-Spectrum (i.e. Spectrum of Dual)
Discrete M-Spectrum