Lecture 11: White and red noise

c Christopher S. Bretherton
Winter 2014

Reference: Hartmann Atm S 552 notes, Chapter 6.1-2.

11.1 White noise

A common way to statistically assess the significance of a broad spectral peak
as in the Nino3.4 example is to compare with a simple noise process. White
noise has zero mean, constant variance, and is uncorrelated in time. As its
name suggests, white noise has a power spectrum which is uniformly spread
across all allowable frequencies.
In Matlab, w = randn(N) generates a sequence of length N of n(0, 1)
‘Gaussian’ white noise (i.e. with a normal distribution of mean 0 and std 1).
The upper two panels of Fig. 1 show a white noise sequence of length N = 128
and its periodogram, which shows that the power spectrum is uniformly spread
across frequencies with a mean spectral power of 1/N per harmonic.

11.2 Red noise

Red noise has zero mean, constant variance, and is serially correlated in time,
such that the lag-1 autocorrelation between two successive time samples has
correlation coefficient 0 < r < 1. As we will show shortly, red noise has a
power spectrum weighted toward low frequencies, but has no single preferred
period. Its ‘redness’ depends on r, which can be tuned to match the observed
time series. For the Nino3.4 case, a reasonable statistical null hypothesis would
be that the observed power spectrum could have been generated purely by red
noise.
To sequentially generate a n(0, 1) red noise sequence xj from a white noise
sequence wj , we set

x1 = w1
xj+1 = rxj + (1 − r2 )1/2 wj+1 , j ≥ 1 (11.2.1)

Using properties of normal distributions, it is easily shown that xj+1 is n(0, 1)

(Gaussian) and that the lag-1 correlation coefficient of xj+1 and xj is r. It
is also easy to show by induction that the correlation coefficient of xj+p and

1
Amath 482/582 Lecture 11 Bretherton - Winter 2014 2

White noise
3 0.05

2 0.04

Spectral power
1
0.03
wj

0
0.02
−1

−2 0.01

−3 0
0 50 100 −50 0 50
j M

Red noise (r = 0.85)

3 0.2

2
0.15

Spectral power
1

0 0.1
xj

−1
0.05
−2

−3 0
0 50 100 −50 0 50
j M

Figure 1: White and red noise time series (left) and their periodograms (right)

xj for p > 1 is rp = exp(−p log r) = exp(−Rp∆t), where R = − log r/∆t is

the decorrelation rate. The autocovariance sequence of red noise thus decays
exponentially with lag. The lag at which the autocorrelation drops to 1/e is
τ = R−1 .
The function rednoise.m (class web page) implements this algorithm, It
was used with the white noise sequence on the upper left of Fig. 1 and r = 0.85
to generate the red noise time series on the lower left. The periodogram of this
sequence, shown in the lower right, now has a predominance of spectral power
in low harmonics M .

11.3 Theoretical power spectrum of red noise

The true power spectrum of n(0, 1) red noise is most easily deduced as the DFT
of its autocovariance sequence. Rather than grinding through discrete sums,
it is more helpful to interpret the DFT as a Riemann sum that approximates
the continuous integral for the complex Fourier coefficients of the continuous
Amath 482/582 Lecture 11 Bretherton - Winter 2014 3

function a(t) = e−R|t| , L-periodically extended for |t| > L/2):

Sm = N −1 DFT(a)
Z L/2
≈ cM [a(t)]] = L−1 e−R|t|−iωM t dt
−L/2

−1
So far, the approximation is good if R∆t ≪ 1 and ωM ∆t ≪ 1, so that
a(t) exp(−iωm t) is well resolved by the grid of spacing ∆t. If in addition
RL/2 ≫ 1, the integrand becomes very small for |t| > L/2. Then, with negligi-
ble error we can extend the range of integration to infinity:
Z ∞
−1
Sm ≈ L e−R|t|−iωM t dt
−∞
Z0 Z ∞ 
= L−1 et[R−iωM ] dt + e−t[R+iωM ]
−∞ 0
 
1 1
= L−1 +
R − iωM R + iωM
2R
= L−1 2 2 .
R + ωM
∆ω R 2π
= 2 (∆ω = ) (11.3.1)
π R 2 + ωM L

The power spectrum of red noise has a maximum value for low frequencies
ωM ≪ R, and decreases at high frequencies - a ’red’ spectrum, as claimed. If
we sum the power spectrum across the harmonics, and think of it as a Riemann
sum approximation to a continuous integral
N N/2−1
X X ∆ω R
Sm ≈ 2
m=1
π R 2 + ωM
M=−N/2
(N/2−1)∆ω
1 R
Z
≈ dω
π −N ∆ω/2 R2 + ω 2
1 ∞ R
Z
≈ dω 2
π −∞ R + ω2
1 ∞
= tan−1 (ω/R) −∞
π
= 1

Consistent with Parseval’s theorem, we have deduced that the power spectrum
sums to 1, the variance that we constructed our red noise to have. In this deriva-
tion, extending the limits of the integral to infinity is a good approximation if
N ∆ω/2 ≫ R. Since N ∆ω/2 = N π/L = π/∆t, this is equivalent to R∆t ≪ 1,
which was the assumption we made in deriving the discrete red noise spectrum
(that the red noise is well resolved).
Amath 482/582 Lecture 11 Bretherton - Winter 2014 4

11.4 Fitting red noise to data

One common way of fitting the autocorrelation sequence is a red noise fit, as an
exponentially decreasing function of lag. This fit is shown in plot as chain dash,
using an e-folding time of τ = 6.1 months. Roughly speaking, measurements
closer together than τ will be significantly correlated and those further apart
will be only weakly correlated. This can be cast in terms of a effective lag-
1 autocorrelation r = exp(∆t/τ ) (= 0.85 in our case). Because the actual
autocorrelation is not exactly an exponentially decreasing function of lag, r is
not exactly the same as the true lag-1 autocorrelation of 0.9.
Script nino2 adds a red-noise fit to the SSTA power spectrum based on
τ = 6.1 months and scaled to match the observed variance of SSTA. There are
four harmonics in the 0.2-0.4 yr−1 range that clearly stand above the red noise
spectrum.
To test how likely this is to be a chance occurrence, we now look for a less
noisy way to estimate the power spectrum.