ATMO 632: Week 4
Time Series Analysis:
Spectral Peaks, Resampling, and Monte Carlo methods
Course Home Page: http://www.cgd.ucar.edu/~svn/atmo632
Reading Material (overlaps with week 3)
- Wilks:
Chapter 8.4-8.5 (Harmonic and Spectral Analysis), 5.2.3 (Serial
correlations), Chapter 5.3.2 (Resampling methods)
- North:
Chapter 6 (Spectral Analysis of Time Series)
- Hartmann:
Chapter 6.2 (Harmonic Analysis)
Concepts
Power
spectra
How
to smooth spectra
Different
ways to plot spectra
Sums of
uncorrelated time series
Choosing
the right null hypothesis
AR(0)
AR(1)
AR(2)
Sampling
errors in power spectra
a
priori vs. a posteriori confidence intervals (see Hartmann’s notes:
6.2.6)
Serial
correlations and degrees of freedom (Wilks 5.2.3; North 4.10)
Resampling/Bootstrapping
(Wilks: 5.3.2)
Monte Carlo
methods
Generating
synthetic data
Lag-1
autocorrelation
Variance
of AR(1) process
Spatial
correlations
EOFs
ATMO 632: Homework for week 4 (due Sep. 28)
NOTE: You can obtain the datasets
used below by visiting the course home page
http://www.cgd.ucar.edu/~svn/atmo632
and clicking the appropriate links.
You do not need to type in the
URLs given below, which are provided for reference.
- Reference
red-noise spectrum: Consider 100 years (1904-2003) of the NAO time
series analyzed in Week 2 homework.
- Compute
the standard deviation sx
of the time series and the lag-1 autocorrelation r1,
i.e., the correlation between NAO index values for successive years.
- Use
the formula below to plot the reference red-noise power spectrum,
for the frequency range f=0,..,0.5 yr-1, with
increments Df = 0.01
(=1/N), where N=100:
(Wilks:
Sec. 8.5.5)
Note that S(fj)
satisfies the normalization condition
, with
. This ensures that for uncorrelated (r1=0)
data, we have
.
- Compute
95% confidence limits associated with the above red-noise spectrum,
assuming a
distribution with degrees of freedom
. Recall that the
statistic is
used to compare the sample variance
to the population variance
. For each
, assume the red-noise spectrum S(fj)
represents the population variance
. The sample variance
can therefore be expressed as

Look up the
values defining the 95% confidence limit, and use it to
compute
using the above
formula to obtain the confidence limits on the power spectrum. Overlay the
confidence limits on the red-noise spectrum plot (using a different line
style).
- Synthetic
red noise and “spectral peaks”:
- Generate
a synthetic red-noise time series, with 100 data points starting
from
, using the formula
,
where
is a gaussian random
variable with zero mean and unit variance. Compute the frequency power spectrum
S(fj) by applying a discrete fourier transform on the time
series. Use the same normalization condition for S(fj) as in
Problem 1. Overlay power spectrum on the previous plot of the reference red
noise spectrum. Note how many “spectral peaks” lie above the 95% limit.
- Generate
two more synthetic time series, using different realizations of random
variable
. Plot their power spectrum, along with the reference
red noise spectrum and the 95% confidence limits. Note how many “spectral
peaks” lie above the 95% limit for each of the two power spectra.
- What
do you learn from the power spectra of the three synthetic red-noise time
series?
- NAO
power spectrum: Consider 100 years (1904-2003) of the NAO time series
analyzed in Week 2 homework.
- Compute
the frequency power spectrum S(fj) by applying a
discrete fourier transform on the time series. Remember to detrend the
data before computing the transform. Use the same normalization condition
for S(fj) as in Problem 1. Plot the power spectrum and
overlay the reference red noise spectrum and the 95% confidence limits.
Are there any spectral peaks above the 95% limit?
- Average
together neighboring power spectral estimates in groups of five
(f=0.01-0.05, 0.06-0.10, ..., 0.46-0.50). Plot the 10 binned power
spectral estimates as a function of the middle frequency value for each
bin. Again plot the reference red-noise spectrum and 95% confidence
limits using the appropriate value of
. Are there any significant spectral peaks?
- Based
upon the above two plots, would you accept the null hypothesis that NAO
is just red noise? How do these results relate to what you learn from
Problem 2?
“Extra credit”
Spectral confidence limits
using Monte Carlo approach: Generate 1000 time synthetic red-noise series,
each with 100 data points, as in Problem 2. Compute the frequency-power
spectrum for each of the time series, binning frequencies in groups of five as
in Problem 3(b), resulting in 10 frequency values fj. Compute
the PDF of the power spectrum and the 2.5 percentile and 97.5 percentile power
for each of the 10 frequency values fj, and plot them as a
function of fj. Compare to the 95% confidence limits in
Problem 3(b). Discuss the advantages and disadvantages of using the Monte Carlo
approach to computing confidence limits, as opposed to using the traditional chi-square
test.