On the multiple breakpoint problem and the number of significant breaks
in homogenisation of climate records
Separation of true from spurious breaks
Ralf Lindau & Victor Venema University of Bonn
Internal and External Variance
Consider the differences of one station compared to a neighbour or a
reference.
Breaks are defined by abrupt changes in the station-reference time series.
Internal variance within the subperiods External variance
between the means of different subperiods
Criterion:
Maximum external variance attained by a minimum number of breaks
Decomposition of Variance
n total number of years N subperiods
ni years within a subperiod
The sum of external and internal variance is constant.
First Question
How do random data behave?
Needed as stop criterion for the number of significant breaks.
Random Time Series
with stddev = 1
Segment averages xi scatter randomly mean : 0
stddev: 1/
Because any deviation from zero can be seen as inaccuracy due to the limited number of members.
ni
2 -distribution
The external variance
is equal to the mean square sum
of a random standard normal distributed variable.
Weighted measure for the variability of the subperiods‘
means
From 2 to distribution
n = 21 years k = 7 breaks
As the total variance is normalized to 1, a kind of normalized
chi2-distribution is expected:
This is the -distribution.
data
2 , 1 2 ) 1
(
2 1 1 1
2
k n
B k
v v v
p
k k n
The exceeding probability P gives the best (maximum) solution for v
Incomplete Beta Function
v
pdv v
P
0
1 ) ( 7 breaks in 21 years
Added variance per break
5 ln 1 2
2 ln 1 1
1
*
*
*
*
k k k
dk dv v k
1
0
1 )
( i
l
l
l v m
l v v m
P
Incomplete -function:
2
3
n m
2 i k
Transformation to dv/dk:
mean 90%
95%
The extisting algorithm Prodige
Original formulation of Caussinus and Mestre for the penalty term in Prodige
Translation into terms used by us.
Normalisation by k* = k / (n -1)
Derivation to get the minimum
In Prodige it is postulated that the relative gain of external variance is a constant for given n.
1 2 ln min
ln v k* n
0
ln 1 2
1
*
n
dk dv v
n
dk dv
v 2ln
1 1
*
ln min
1 1 2
ln
n
n v k
min )
1 ln(
2 )
(
) (
1 ln )
(
1
2 1
1
2
n
n l k Y
Y Y Y n Y
C n
i i
k
j j j k
Shorter length, less certainty
n = 21 years n = 101 years
Exceeding probability 1/128
1/64 1/32 1/16 1/8 1/4
Second Question
How do true breaks behave?
True Breaks
Identical Behaviour
True breaks behave identical to random data.
But the abscissa-scale is now:
k / nk instead of k / n.
Compared to random time series the external variance grows faster by the factor n / nk
data theory
nk = 19 true breaks within n = 100 years time series
Assumed / True Break Number k / nk
Break vs Scatter Regime
Simulated data with 19 breaks interfered by scatter
The internal variance decrease as a function of break number.
In the break regime the variance decrease faster by the factor:
15 breaks are detectable,
depending on signal to noise ratio.
Time series length Number of true breaks
Conclusions
• The analysis of random data shows that the external variance is -distributed, which leads to a new formulation for the
penalty term.
• True breaks are also -distributed. Their external variance increases faster by a factor of n/nk compared to random scatter.