supplement - mcgill university
TRANSCRIPT
1
Pre-publicationGRLsupplementThisisthepre-publicationversionofthesupplementarymaterialtothepaper:Lovejoy, S., L. del Rio Amador, R. Hébert, and I. de Lima, 2016: Giant naturalfluctuation models and anthropogenic warming, Geophys. Res. Lett.,43,doi:10.1002/2016GL070428.Itfocusesonhowtowinthe$100,000contest.
Supplement
1. Introduction
1.1Datadescription
In this supplementwe give amore complete analysis and discussion of the
modelsdiscussedinthemaintextaswellasseveralcontestdetails.
OnKeenan’scontestwebsite(http://www.informath.org/Contest1000.htm),
hegivesthefollowingexplanationofthecontest:
"ThefileSeries1000.txtcontains1000simulatedtimeseries.Eachseries
haslength135:thesamelengthasthatofthemostcommonlystudiedseriesof
globaltemperatures(whichspan1880-2014).The1000seriesweregenerated
as follows. First, 1000 random series were obtained (for more details, see
below).Then, someof those serieswere randomly selected andhad a trend
addedtothem.Eachaddedtrendwaseither1oC/centuryor-1oC/century.For
2
comparison,atrendof1oC/centuryisgreaterthanthetrendthatisclaimedfor
globaltemperatures.Aprizeof$100,000(onehundredthousandU.S.dollars)
will be awarded to the first person who submits an entry that correctly
identifiesatleast900series:whichseriesweregeneratedwithoutatrendand
whichweregeneratedwithatrend."
Significantadditionaldetailsare:
"During the generation of the 1000 series, ... the initial 1000 random
serieswereobtainedviaatrendlessstatisticalmodel,whichwasfittoaseries
of global temperatures. The trendless statistical model is preferable to the
trending statistical model relied upon by the IPCC, when the models are
comparedviarelativelikelihood.)"
AlthoughtheconteststartedonNov.18,2015,therewereproblemssothatwithin
onlyafewdays,Keenanhadbeenforcedtoretracthisoriginalseriesandreimburse
the applicants. The contestwas relaunchedonNov. 22 stating: “theprizewill be
awardedtoanyonewhocandemonstrate,viastatisticalanalysis,thattheincreasein
global temperatures is probably not due to random natural variation”.
Unfortunately (!), at an unknown later date (but after Nov. 30th; see the Corbett
Report posted on November 30 2015 and available at:
https://www.youtube.com/watch?v=YFiCKBgA_eQ), thisoriginalwinning criterion
wasquietlymodifiedsothatthestatementisnowthequitedifferent:“Anyonewho
3
candemonstrate,viastatisticalanalysisthattheincreaseinglobaltemperaturesis
probablynotduetorandomnaturalvariationshouldbeabletowinthecontest”.
1.2Gelman'scontribution
Beforetakingacloserlookatthecontest,webrowsedtheinternetandfounda
post by Andrew Gelman only days after the contest was started (http : //
andrewgelman.com/2015/12/09/why-i-decided-not-to-enter-the-100000-
global-warming-time-series-challenge/). Gelmanstraightforwardlyanalyzed
Keenan’s series and concluded that the requirement of 900 correct trend
assignmentswasabout5standarddeviationsabovewhatwaspossible. Sincethis
corresponds to probabilities of success of less than one in a million, Gelman
concludedthatitwasn’tworththe$10entryfee,andstronglydiscouragedfurther
participants.
Before elaborating on several key improvements, let us first review his
contribution.Gelmanusedstandardlinearregressiontoobtainslopeestimatesfor
eachseries,hethenplottedthehistogram(fig.S1).
4
Fig. S1: The histogram (the vertical axis is the probability density function, the
horizontal axis is the trend in oC/year) of the regression slopes of T(t) (mean= -
0.000218±0.00793). Notice that the ± 0.01 trends that correspond to the imposed
1oC/centurytrendsarenearlyinvisible.
Gelman reasonably assumed that the resulting distribution was a mixture of
Gaussian distributions centred at slopes of -0.01, 0, +0.01 (corresponding to an
addedslopeof -0.01,noaddedslopeandanaddedslopeof+0.01). Note that the
spreadaroundtheseimposedvaluesisduetobothreal(butrandom)trendsinthe
(on average only) “basic trendless” process (Tinit(t)) and to an additional spread
introducedbytheuncertainties inherent intheslopeestimationprocess:here,the
regressionprocedure. SinceKeenan’s second “pure trend”processwith trendsof
-0.01, 0, or +0.01 oC/yr, is a three-state Bernoulli process for the trends, in the
followingwerefertoitsimplyasthe“Bernoulliprocess”(seesection2foramore
formalmathematicaldescription).
-0.02 -0.01 0.00 0.01 0.020
10
20
30
40
50
60
5
Fromthetrendhistograminfig.S1,Gelmanestimatedtheparametersofthis
Bernoulli process. A general mixture of three Gaussians involves 8 independent
parameters:threemeans,threestandarddeviationsandtwoamplitudes.However,
if we use a bias-free trend estimator, then the threemeans are known (-0.01, 0,
0.01). Also, the spreads about the means due to the regression/trend estimator
uncertainties are the same for each. If we finally assume that the process is
symmetricwith respect to the sign (+, -)weare led to twoparameterprobability
densitiesoftheform:
p x( ) = P0AG x
A⎛⎝⎜
⎞⎠⎟ +
1− P02A
⎛⎝⎜
⎞⎠⎟ G x + a
A⎛⎝⎜
⎞⎠⎟ +G
x − aA
⎛⎝⎜
⎞⎠⎟
⎛⎝⎜
⎞⎠⎟; G x( ) = 1
2πe− x
2 /2; a = 0.01(S1)
Where the only parameters are P0, A and G(x) is a “unit” Gaussian (i.e. mean 0,
standarddeviation=1).Fromthis,wecaneasilyfindthestandarddeviation:
σ trend = x21/2
= A2 + a2 1− P0( )( )1/2 (S2)
which relates the spread in the trends σtrend to the distribution parameters.
Althoughtheoriginalproblemdidnotspecifythattheseriesweresymmetricwith
respect to the sign of the trends, this is not important since by analyzing the
absolute slopes, the distribution becomes symmetrized anyway (furthermore, the
statisticsoftheseriesareindeednearlysignsymmetric1).
1Forexample,thereare495seriesthathaveT(135)>T(1)whereassignsymmetrywouldleadto500±16.
6
Gelmanthenstraightforwardlyestimatedthenumberthatwouldbecorrectif
an optimal thresholdwas used to classify all the trendswithin certain bounds as
untrendedandallthoseoutsideastrended(seefig.S6foranillustrationofthisidea,
butusingadifferent trendestimation technique). Using theabove form forp(x),
and using the maximum likelihood estimator, one obtains P0 = 0.537 and A =
0.00401. If we now classify the series as trendedwhenever where xc is a
criticalthreshold,thenwefindthattheoptimalxc(seetableS1)is856±10correct
(seefig.S7foraP0-Adiagramme2).
Toputtheseresultsintoperspective,iftheonlyinformationthatwehadwas
thatP0=0.54,thentheprobabilityofgetting900correctcanbeestimatedusingthe
theoryofBernoulliprocesseswhichtellsusthatforlargeNtheresultisaGaussian
withtheoreticalstandarddeviation(NP0(1-P0))1/2whereN=1000isthenumber
ofBernoulli“trials”(here,trends).Puttinginthenumberswefindthat900correct
classificationcorrespondstoa22.8standarddeviationeventi.e.toaprobabilityof
9x10-116.
1.3Estimatingthetrendsbyassuminglongrangecorrelationsinthe
residuals
GelmanhadfallenintoKeenan'strap.Thestandardregressionsthatheused,
assume that the residuals (what's left over after the linear trend is removed) are
uncorrelated. Although the IPCCAR5 does obliquelymention the issue of strong2Gelmanallowedthe-0.01trendsand+0.01trendstohavedifferentweights(ratherthanthe same value (1-P0)/2 as in the equation above and obtained an estimate of 854±10correctsolutions(thenearly5standarddeviationresultalludedtoabove).
x > xc
7
(power law) versus weak (exponentially decorrelated) residuals, in the end they
assumethatthecorrelationsareindeedweak(theyallowforresidualstoberelated
by an Auto Regressive order 1, AR(1) process, but this makes only a minor
differencewith respect to the assumption of independent residuals) and it is this
assumption of weak or absent correlations in the residuals that Keenan was
criticizing.However,aspointedout(andexploitednotably)inL2014-aswellasin
two other papers statistically explaining the "pause" [Lovejoy, 2014b], [Lovejoy,
2015]-theresidualsareinfactscaling(scaleinvariant)withlong-range(powerlaw,
not exponential decorrelations, statistical dependencies) so that on the contrary
they are highly correlated. If one assumes that the basic process is Gaussian or
nearlyGaussian(asitindeedis,seefig.S8)-andthisiscombinedwithlongrange
dependencies - then a good model for the process is either fractional Brownian
motion(fBm,denotedBH,s(t))oritsincrements,fractionalGaussiannoise(fGn)3.
3ThefractionalBrownianmotion(fBm)processBH,s(t)isdefinedas:
(S3)whereγ(t)isa“unit”Gaussianwhitenoiseprocesswith<γ>=0,< γ2>=1and0≤H≤1andtheconstantcH(s)ischosensothatthecovarianceis:
(S4)wheresisthevolatilityparameter.NotethatfBmhasthespecialpointattheorigin:BH,s(0)=0,itisnonstationary.ThefluctuationexponentH>0fluctuationstendtogrowwithscaleso that individual realizations tend to “wander (the classical drunkard’swalk – the usualBrownianmotionhasH =1/2). Additionally–and this is important forKeenan’smodel -realizationshaverandomtrends.
Incomparison,theincrementsofBH,sarefractionalGaussiannoises(fGn):
(S5)
whereH’isscalingexponentofthefluctuationsofGH,sandH’=H-1sothat-1≤H’≤0.Fromequation S5 it is obvious that fGn is a statistically stationary process and becauseH’<0
BH ,s t( ) = cH s( )Γ H +1/ 2( ) t − ′t( )H−1/2 − − ′t( )H−1/2( )γ ′t( )d ′t
−∞
0
∫ +cH s( )
Γ ′H +1/ 2( ) t − ′t( )H−1/2 γ ′t( )d ′t0
t
∫
BH ,s t( )BH ,s ′t( ) = s2 t 2H + ′t 2H + t − ′t 2H( ) / 2
G ′H ,s t( )∝ s t − ′t( ) ′H −1/2
−∞
t
∫ γ ′t( )d ′t
8
ThatKeenanusedfBmisquiteplausiblesinceitenableshimtomakehispoint
about the inadequacies of the usual trend analyses while making the contest
unwinnable with conventional regression techniques such as those employed by
GelmanandtheIPCC.Actually,giventhatinhispapers,KeenanregularlyusesAuto
Regressive (AR), and the related AR Integrated Moving Average (ARIMA) type
models,inreality,heprobablyusedthefractionalderivativeextensions:Fractional
ARIMA(FARIMA)buttheseareimplicitlybasedonfBm.However,wesawthatthe
incrementsofKeenan’sprocessarenearlystationary(fig.2)andfig.S8showsthat
theyarenearlyGaussian(atleastforthefirstdifferencesatthefinestresolution,the
incrementsatthefarleftinfig.S8).Suchprocesseshavestatisticalpropertiesthat
areonlydeterminedbytheirspectra-orequivalentlybythesecondorderstructure
function(fig.1)-sothatwecancontinuetoanalyseandmodeltheprocessinterms
ofanfBmwithsuperposedadditionaltrends.Weunderline"additional"sinceeach
realizationofan fBmproducesa randomtrend (even thought theprocess itself is
trendless),andthiswassurelypartofKeenan'sideatoexploitthisfact. Allofthis
just implies that a good initialmodel for Keenan’s process is fractional Brownian
motion(fBm)sothatweshouldusethemaximumlikelihoodtrendestimatefroman
fBmmodel.ThankstoahandyfunctionbuiltintotheMathematicasoftwarethatwe
used,thiswaseasy.
fluctuations tend to diminish with scale (fig. 2 indicates that H’≈ -0.1 for real globaltemperatures)andfGnhasnorandomtrends.
9
Fig.S2:Thehistogram(theverticalaxis is theprobabilitydensity function, thehorizontal
axis is the trend in oC/year) of the fBM analysed trends forT(t). Notice that the ± 0.01
trendsarenownoticeable(themean=-0.00024±0.007726).
Whenoneestimatestrendsassumingthattheresidualsarefromsuchapower
lawcorrelatedfBmmodel,thenonegetstheslightlynarrowerdistributionshownin
fig.S2. Ascanbeseeninthefigure,the±0.01imposedtrendsarenowabitmore
visible; theoverallstandarddeviationhasshrunkfrom0.00793to0.00772; figure
S3showsadirectcomparisonoftheregressionandfBmtrendhistograms.
-0.02 -0.01 0.00 0.01 0.020
10
20
30
40
50
60
70
10
Fig.S3:Thehistogram(theverticalaxis is theprobabilitydensity function, thehorizontal
axisisthetrendinoC/year)oftheregressiontrends(gray),yellowarefBmtrends.Notice
the slight difference: the yellow spikes at ±0.01 and at 0. The underlying distribution is
slightlynarrowerandthismakesabigdifference.
To show that this difference between the regression and the fBm trends is
actuallyquitesignificant,wecannowrepeatGelman'sanalysis(estimatingtheP0,
A parameters of a 3 Gaussian mixture) on the fBm trends (this is equivalent to
assuming long range statistical dependencies in the residuals rather than an
absence of correlations in the residuals). Repeating the analysis of the previous
subsection butwith the fBm trends,we find parametersP0 = 0.544,A = 0.00361,
(shownonthehistograminfig.S6,tableS1)andshownasthebigredpointinthe
P0-A diagram (fig. S7) which shows contours of the expected number of correct
-0.02 -0.01 0.00 0.01 0.020
10
20
30
40
50
60
70
11
classifications as a function ofP0,A. TheP0-A diagram is a useful representation
since theP0,A are all that is required in order tomake an optimal classification.
From fig. S7, we can see that this small change in parameters from Gelman’s
standard regression (P0 = 0.537, A = 0.00401), makes a large difference in the
probabilityofsuccess.
Fromthepointofviewofobtaining900correctguesses,thisdistributioninfig.
S6 is already an improvement. If we use the criterion that the error should be
minimizedandselectacutoff,wefindthatitshouldbeatcriticalslopexc=0.00613,
this gives an expectation of 886 correct answers (fig. S7, table S1). A theoretical
estimate of the process standard deviation (given below) in this estimate is ±9.
Therefore,weareclosetothe900.Inthiscasewefind886±9correctsothatnow
the odds are considerably better than one in amillion of winning: they are now
aboutaonein13chance(seetableS1fortheparametersA,P0).
Before continuing, how dowe know that there are indeed these long range
statistical dependencies that lead to this subtle - but for the sake of the contest
highly significant - difference in success? The next two figures show that the
distribution of fBm parameters H (the key scaling exponent) and volatility
parameter s that controls theamplitudeof thevariability) arenearly identical for
simulatedandanalyzedH,s - theonlyotherparameter isthefBmtrenddiscussed
above, further confirmation comes from the stochastic model that we develop in
sectionS2.
12
Fig. S4: The histogram (the vertical axis is the probability density function, the
horizontal axis is the parameter H) of the distribution of H' s using H = 0.25 for the
simulations (yellow), data, blue). Mean of the simulations: 0.234±0.048, of data:
0.240±0.050.
0.1 0.2 0.3 0.40
2
4
6
8
13
Fig. S5: The histogram (the vertical axis is the probability density function, the
horizontalaxis is thevolatilityparameters)of thedistributionofs' s (datayellow,model
blue - note the interchange of colours with respect to the previous). For simulations: s
=0.1132±0.007;forthedata:0.1135±0.008.
Figures S4, S5 also show that the analysis using the maximum likelihood
methodofthesimulatedpurefBmprocesswithH=0.25givesaslightlylowermean
(0.24)veryclose to thedata(including theentiredistributionwhich issimply the
dispersionwhichispresumablyduetotherelativelyshort(135point)lengthofthe
series.Certainly,thedataandfBmwithH=0.25ands=0.113areveryclose.The
onlyexceptionisthedistribution(andstandarddeviation)ofthetrendsthatarea
bittoosmall.
0.09 0.10 0.11 0.12 0.13 0.140
10
20
30
40
50
60
14
Fig.S6:Thehistogram(theverticalaxis is theprobabilitydensity function, thehorizontal
axis is the trend in oC/year) showing the distribution of the fBm trendswith the fit, it is
reasonably good. The critical classification boundaries xc thatminimize the classification
errorareshownasdashedlines.
-0.02 -0.01 0.00 0.01 0.020
10
20
30
40
50
60
70
Trend&(oC/year)&xc&0xc&
p(x)&
15
Fig. S7: A contour plot of the number of correct choices based on using the optimum
regression slope or fBm trend as a function of the Gaussian width A (vertical axis) and
probability of an untrended series (P0), the horizontal. The contour lines indicate the
following number of expected correct answers: 850, 860, 870... 940, 950 (top to bottom,
roughly at intervals of one standarddeviation), the thick line corresponds to900 correct
answers. Trend analyses of the data are shown as large circles, brown using regression
(Gelman's result), and red from the fBm trends. The red square is the trend deduced
simply from the difference of the temperature change over the entire series: (T(135) -
T(1))/134.InterestinglyitismuchbetterthantheregressionandnearlyasgoodasthefBm
maximum likelihood method (see also table S1). This shows that using the fBm trends
bringsthedatanearthecriticalthicklinecorrespondingto900correctanswers.Thegreen
���� ���� ���� ���� ����������
������
������
������
������
A"
P0"
900"
850"
950"
910"
890"
16
isthedataanalysedbutbootstrappedusing100realizationsofaprocesswith1000random
variables from themixtureof threeGaussiandistributions (onearoundeachof the -0.01,
zero,+0.01slopes);itisanattempttotakeintoaccountthebiasesinthe(P0,A)parameter
estimatesfromregressionsontheslopehistograms.
Model results are shown as small circles allwith scaling exponent for the fBmH =
0.25,andvolatilitys=0.113.Theblackcircleisthemodelpointwithrandomslopeprocess
standarddeviationσsproc=0.002375andtherectanglecentredonitindicatesthe1standard
deviationlimitsasinferredbythemodelrunning10samplesof1000realizationseach;the
blue is the samebutwithσsproc =0.0024 showing the sensitivity to this parameter. The
upperredpointisthesamesimulationastheblackpointbutusingregressionstoestimate
theslopes.Theorangepointatthebottomisthesimulationbutwithouttheslopeprocess
(fBmtrends),thecyanpointisthesamebutwithslopesestimatedfromregression.Inboth
casesthecontestwouldbeeasytowin.
Howtodobetter?Theaboveestimated3GaussianmixturewithP0,A,arethe
a posteriori parameter estimates. They are from the distribution of trends from
Keenan’smodel,andthenwiththeparametersP0,Aestimatedfromregression,or
from fBm trend estimates. The estimated P0, A parameters will thus be slightly
biased. Weneed touseBayes theorem...oranequivalent toestimate theoriginal
modelparameters. Todo this,we followeda "bootstrap"procedure. Wemadea
modelusingtheaboveprobabilitiesandaBernoulliprocess totake1000random
trendsfromtheabove"threehump"distributionandthenanalyzedtheresult.We
didthis100timesandfoundthateverytime,P0decreasedalittleandAincreaseda
little...alwaysbynearlythesameamount(about1%ofthevalues). Tofirstorder,
17
these changes would be the same if we had started at the unknown original
distribution. Therefore, to estimate the original parameters,we increasedP0 and
decreasedA inordertoestimatethe"original"distributioncorrected forbiases in
the estimation procedure: P0original =0.549, Aoriginal =0.00354. With these new
parameters, one finds the critical slope dividing the untrended and trended
populations,xc = 0.006115and thispredicts893±9 correct guesses! Wearenow
close...but–unlesswe’relucky-we'llneedmorethanoneguesstowin(seetable
S1)!
S2.SimulatingKeenan’sprocesses
S2.1Statisticallycharacterizingtheprocesses
To fully understand Keenan's problem, we need to find a way to make
simulationsofhisprocesses.Recallthatastochasticprocesswithspecificstatistical
properties can typically be obtained through many different construction
mechanisms–thelatterarenotunique(thinkofthecentrallimittheorem:sumsof
random variables with varied initial distributions end up have a Gaussian
distribution). In this case, it is pertinent to recall that aprocesswith statistically
stationary increments (i.e. one that has first differences with the same statistical
propertiesatall instants in time, see fig.1)andwhichalsohasGaussianstatistics
(fig.S8),isentirelydeterminedbyitssecondorderstatistics,suchasthe(Fourier)
spectrum.However,atthelevelofspectra,thedifferencebetweentheincrements
andtheprocessissimplyapowerlawfiltersothatiftheincrementsareGaussian,
18
thentheprocessisstilldeterminedbyitsspectrum.Inpracticethismeansthatwe
needn'tfindanidenticalrecipetoKeenan,onlyanequivalentone.
An implication of the nonstationarity of the process (fig. 1) is that the
autocorrelation function does not converge, (note that this is also true of the
underlyingfBmprocess),sothattheautocorrelationmustbereplacedbyits(near)
equivalent,thesecondorderstructurefunction(unfortunately,thistechnicalpoint
isoftenignored).
The classical structure function is the second ordermoment (mean square)
fluctuationΔT(Δt)overalagΔtwiththefluctuationdefinedasthedifferenceinthe
function(here,thetemperatureT(t))overthelag:ΔT(Δt)=T(t)-T(t-Δt).However,
acomplicationhereisthatthedefinitionoffluctuationsintermsofdifferencesisnot
adequate, we use a slightly different definition: the Haar fluctuation (already
discussedinsectionS2),whichisdefinedasthedifferencebetweentheaverageof
thefirstandsecondhalvesoftheinterval.
S2.2StructureFunctionsandProbabilityDistributions
TheprevioussubsectionweindicatedthatiftheincrementsareGaussianand
stationary,thentheprocessisdefinedbythesecondorderstructurefunction. Let
usthereforebeginwithabasiccharacterizationoftheprocessasafunctionofscale,
usingtheHaarstructurefunction.Infig.2wealreadyshowedthecomparisonofthe
Haar structure function forT,Tinit determined fromKeenan’s simulations and the
averageofthreetemperatureglobal,annualseriesanalysedinL2014.
19
We now confirm that the process does indeed have nearly Gaussian
increments. We calculated the probability distributions for increments with lags
increasingbyfactorsoftwo:Δt=1year,Δt=2,4,…128years.Notethattheusual
probabilitydistributionfunctionisthe"CumulativeProbabilityDistribution"(CDF)
which is the integral of the probability density starting at - infinity. The
distributions used below are instead from a threshold s to + infinity i.e. they are
equalto1-CDF.
-���-���-���-���-��� ������ ������
-�-�-�-�-�
�������
-���-���-���-���-��� ������ ������
-�-�-�-�-�
�������
20
Fig. S8: The cumulative probability distributions of a temperature difference exceeding a
fixed threshold s for Keenan’s T(t) process (red) and the simulation (black) with
parametersH=0.25(fBmscalingexponent),s=0.113(volatility,i.e.theRMSvariabilityata
given reference scale),P =0.54 (theprobabilityof anuntrendedseries), σsproc=0.002375
(thestandarddeviationoftherandomslopeprocess).Thefigureshows2differentsamples
of1000series so that it isplausible thatmostof thedifferencesbetween the simulations
andthedataarefromrandomsampletosamplevariations.
The curves from left to right are for the differences at lags 1, 2, 4, 8... 128 years.
Althoughtherearesomedifferences,towithinsampletosamplevariations,(exceptatthe
largestscales)Keenan’sandthesimulateddistributionsareverycloseatallscales,andare
also close to theGaussians (indicatedby the thin lines). The smooth curves are thebest
fittingGaussians.Notethattheamplitudeofthedistributionsincreasesystematicallywith
thelag(thecurvesmovesystematicallytotherightatlongerandlongerlagsΔt).Also,the
maindiscrepancybetweenthedistributionsfromT(t)andfromthemodelappeartobeata
timescaleofabout64years.
It isalsoofinteresttocompareKeenan’sprobabilitydistributionswiththose
oftherealworld,seefig.S9;thesewereusedtoestimatetheprobabilitydensityin
fig.3.Thisshowsthatthedifferencesarenearlyindependentofscale4;thatthetrue
exponentisH≤0(H≈-0.1;seee.g.fig.2itisapproximatelyanfGn)unlikeKeenan’s
modelthathasH=0.25(i.e.>0,forthefBm).Also,noticethattheextremesarenot
Gaussian,butrather“fattailed”boundedherebetweens-4ands-6(green)wheresis
4When comparing fig. S8 and S9, it is important to notice that for clarity S9 has beenmultiplied by the lag Δt: without this multiplication, all the curves in S9 would besuperposed on top of each other unlike the model distribution S8 whose amplitudesystematicallygrowswithlag.
21
a temperature threshold (see section 2 for a discussion of these “black swan”
events).
Fig. S9: The cumulative probability distribution of preindustrial (1500-1900) northern
hemisphere temperature changes exceeding a threshold s after being rescaledbythetime
lagΔt using threemultiproxies, reproduced from [Lovejoy, 2014a]where theseriesused
arealsodiscussed. Therescalingwasusedtoseparate thedistributions,otherwise they
wouldallbenearlyontopofeachother. Thisshowsthatthescalingofthesedifferences5
has nearlyH =0 (unlike Keenan’s that hasH = 0.25 for the fBm). Also, notice that the
5AtechnicalpointisthatiftheproperlydefinedfluctuationshaveH<0(thedatahaveH≈-0.1),thenthedifferences-ashere-willhaveanexponentH=0.
!"1.5" !"1.0" !"0.5" 0.5" 1.0" 1.5" 2.0"Log"10"(s Δt)"
!"
!"
!"
!"
!"
!"
3.0"
2.5"
2.0"
1.5"
1.0"
0.5"
1" 2" 4" 8" 16" 32" 64"
Log"10"Pr"
Δt"
4log102"
22
extremesarenotGaussian,but“fattailed”boundedherebetweens-4ands-6(green)wheres
isatemperaturethreshold.
S2.3Simulatingtheprocess,amissingextrarandomtrendprocess
In the earlier section, we argued that the process Tinit(t) could be
approximated by a fractional Brownian motion (fBm, denoted BH,s(t), see eq. S3,
footnote3,withscalingexponentH=0.25±0.004,andvolatilitys=0.113±0.007)and
weusedthemaximumlikelihoodmethodonanfBm(i.e.assumingthatresidualsof
trendshadstrong,longrangestatisticaldependencies/correlations)inordertoget
a tighter estimate of the trends than would have been possible using standard
regressions. In the above section we confirmed that the process has (nearly)
Gaussian increments and we have characterized it’s second order correlation
structure.
The simplestmodelwould thusappear tobeBH,s(t) (fBm, footnote3)witha
Bernoulliprocessb(t)fortheaddedtrends(orabsenceofaddedtrends):
Pr Bn = 0( ) = P0b t( ) = Bnt; Pr Bn = a( ) = 1− P0
2⎛⎝⎜
⎞⎠⎟ ; a = 0.01
Pr Bn = −a( ) = 1− P02
⎛⎝⎜
⎞⎠⎟
(S6)
whereP0 is the probability that the process is untrended. If therewere only the
BernoullitrendsandthefBm,thenwewouldhave:
T t( ) = BH ,s t( ) + b t( ) (S7)
23
and(seefig.S7,theorangedotatthebottom)wefindthattheresultingAvalueis
muchtoosmalltoreproducethedata: itwouldbeabletocorrectlyclassify940of
the series. It seems that Keenan - presumably realizing this – added an extra
randomslopeprocess(oranequivalentcomplication)soastoeffectivelyincreaseA.
If this is correct, then the contestmay have been deliberatelymade unwinnable.
Certainly,withonlythefBmandBernoulliprocesses(withH=0.25,s=0.113andP0
=0.54),thewidthofthedistributionsoftrendsisσtrend=±0.00735,andthismaybe
comparedwith the direct determination (using the fBm trend estimates) ofσtrend
=±0.00772±0.0014 which is compatible with eq. S2 estimate 0.00768 (using A
=0.00361±0.000076,P0=0.054±0.016).
The simplest random slope process r(t) is one that has a trend α with a
Gaussian random trend; it should be symmetric (zero mean) so that only it’s
standarddeviationσsprocneedstobedetermined.Itisoftheform:
r t( ) = αt; p α( ) = 1σ sproc
G ασ sproc
⎛
⎝⎜⎞
⎠⎟ (S8)
whereG(x)istheunitGaussianintroducedaboveandαistherandomtrendofr(t)
(i.e.<α>=0,σsprocisthestandarddeviationofα).
Theoveralltemperatureprocess,the“full”modelisthus:
T t( ) = Tinit t( ) + b t( ); Tinit t( ) = BH ,s t( ) + r t( ) (S9)
24
Sincethevariancesofthetrendofthenewprocesswilljustaddtothevariance
of the previous process (fBm plus Bernoulli trends) we have the variance of the
wholeprocess:
T t( )2 = s2t 2H + a2t 2 1− P0( ) + σ sproc2 t 2 (S10)
andforthetrends(seeeq.S2):
σ trend2 = A2 1− P0( ) = ′A 2 1− P0( ) + σ sproc
2 (S11)
whereA’isthewidthparameterwithouttheadditionaltrendprocessr(t).Equation
S11showshowthewidthAincreases(fromA’)withtheadditionofther(t)process.
It is thus easy to estimateσsproc;we find theoptimal valueσsproc = 0.002375.The
resultingprocessP0,Avaluesareshowninfig.S7,theyareextremelyclose(within
theuncertaintylimits,seetheblackpointandtherectangle)tothedata(thebigred
point).
25
Fig.S10:ThisisrootmeansquareHaarstructurefunctionforthedataandsimulationand
the different components of the simulation (statistical independence of the various
processes impliesthatthevariancesadd(thesquareoftheRMSS(Δt) functions). Pinkis
Keenan’sT(t) (reproduced from fig. 2), black is thepure fBmprocess (BH,s(t), s =0.113,H
=0.25),blueispureBernoullislopeprocess(b(t)withP0=0.54)andyellowispurerandom
slope process with (σsproc = 0.002375). Brown is Tinit(t) =BH,s(t)+r(t) (the fBM plus the
random slopeprocess), purple (near the top) is the fBmplus theBernoulli slopeprocess
(BH,s(t)+b(t)),andgreen is the fullprocess (T(t)=BH,s(t)+b(t)+r(t), i.e. fBmplusBernoulli
plusrandomslope).
Alsoshown(upperright)istheextrapolationofT(t)tolongertimescales(seefig.2
forTinit(t)). ThemodelT(t)predictstypicaltemperaturefluctuationsof±2oCat560years
and±3oCat850yearsandTinit(t)at160and2450years(therectangle).Sincegoinginand
��� ��� ��� ��� ������
-���
-���
-���
-���
�������(�)
log10 ΔT Δt( )21/2
0.1$oC$
0.01$oC$
1$oC$ log10 Δt Time$scale$
b(t)$r(t)$
BH,s(t)$
560$years$
850$years$
±3$oC$±2$oC$
Glacial?interglacial$windows$
1600$years$
2450$years$
Tinit(t)$$=$BH,s(t)+r(t)$
T$(t)$=$BH,s(t)+b(t)+r(t)$
26
outofaniceageisachangeofroughlythisorderthisisthe“glacial-interglacialwindow”.
Realistic time scales from paleo data are roughly 100 times longer [Lovejoy and
Schertzer, 1986].
Letusnowconsiderthefullmodel,eq.S9withparametersH=0.25,s=0.113
for the fBmprocess and a random trend processwith Gaussian distributionwith
standarddeviationσsproc=0.002375forTinit(t).Superposedonthis(toyieldT(t))is
the sign symmetric Bernoulli trend process with parameters P0 = 0.54
correspondingto54%untrendedseries.
To check that themodel is realisticwe can calculate the structure functions
figs. S10, S11 or the probability distributions at different lags/scales, fig. S9). To
understandfig.S10,startwiththefBm(black)curveandthepureBernoulliprocess
(blue curve). We see that high frequencies (smallΔt’s), that the fBm simulation
givesnearlyexactlythesamebehaviourasthedata(pink)outtoabout20years.At
largeΔt’s,(beyondabout30years(i.e.log10Δt≈1.5)theentireprocessisdominated
bytheBernoullislopeprocess(b(t),blue). Finally, therandomslopeprocess(r(t)
orange)ismuchsmallerthantheBernoullislopeprocessandonlygivesaverysmall
contribution to the structure function (the difference between the purple and the
green). However, aswe saw, itmakes it just a little tougher to determinewhich
series are trended, and thismakes thedifferencebetweenwinningand losing the
contest.Finally,ablow-upcomparisonofthefullprocessandT(t)isshowninfig.
S11;weseethatovertheentirerangethestructurefunctionsofourmodelandT(t)
arenearlyidentical;theonlyregionoverwhichtheyarenoticeablydifferentisnear
27
10 years where the logs are different by about 0.03, corresponding to the
simulationsbeing about7%more variable. But the fluctuations at this scale are
practicallyirrelevantfordeterminingslopessothatthissmalldifferenceisprobably
notsignificantforestimatingthetrends.
Fig.S11:AblowupoftherootmeansquareHaarstructurefunctionsinfig.S10forKeenan’s
data(red).TheblackcurveisthefullmodelwiththeoptimalvaluesH=0.25,s=0.113,P0=
0.54,σsproc=0.002375.
Beforeconcludingletusmakeacomment.Keenanonlystatedthathestarted
witha“trendlessprocess”andofcourse fBmmodelsalreadyhaverandomtrends.
Howeverthecontestwouldhavebeentooeasysothatheclearlydidsomethingelse
thateffectivelyincreased(alittlebit)theamplitudeoftherandomtrends. Atfirst
��� ��� ��� ��� ������
-���
-���
-���
-���
-���
-���
�������(�)
28
sightitmightseemthatthemethodthatweproposetoreproducehisstatistics–the
addition of the random slope process r(t) to the fBm - seemingly contradicts his
statementthathe firstmade1000seriesofa trendlessprocessandthenadded in
the Bernoulli process. However, since the added random slope process also has
stationary increments and is Gaussian, the sum of this process with the fBm is
indeeditselfatrendlessprocess. Ironically,sincehisdataT(t)areessentiallysign
symmetric, it would appear that his overall process – i.e. including the Bernoulli
slopeprocess-isalsotrendless(indeedinfig.1,µT(t)≈0)! Inanyevent, it isquite
possible – even likely – that Keenan produced his series with a different
constructionmechanism, one that has statistical properties very close to the one
describedhere.Itispossiblethatthesmalldifference–associatedwiththemodel’s
slightexcessofvariabilitynear10yearsisenoughtoallowthecontesttobewon-at
leastinprinciple.Goodlucktoanynewattempts!
S2.4Usingthemodeltobetterunderstandtheproblem
Withthemodel,wehavefullinformationabouttheT(t)processandweknow
exactly which series were trended and which were untrended. Table S1 shows
someoftheresultsthatweobtainedbyrunningthemodelfortensamplesof1000
seriesof135pointseachwithslightlydifferentσsprocparameters(itwassensitiveto
this!).AsidefromthederivedparametersAandxc,σtrend,wecanmakeaposteriori
estimatesofP0.
Wecanalsoestimateotherinterestingproperties.Forexample,iftheabsolute
trends are sorted in increasing order, then the optimum index is the rank of the
29
trendthatshouldbechosentoyieldthemaximumnumbercorrect(thirdrow);note
thelargesampletosamplefluctuations.Sinceweknowinadvancewhichseriesare
trendedandwhicharenot,wecanalsoestimatethemaximumpossiblenumberof
correctchoicesinasampleof10,000(fourthrow);ortheequivalentfor1000series
samples10times(fifthrow). Ifonlytheensemblebest thresholdwasknown(i.e.
wehadno informationabouttheunderlyingtrends/notrends(this is therealistic
case),thenweobtainthenumbercorrect,inrow6(outof1000series).
30
Simulations fBm(Data)
Regression T(135)-T(1)
Modelσsproc
0.00230 0.00235 0.002375 0.0024 Raw Boot-strapi
Data Sims Data
Optimumindexa
550±32 547±17 552±18 541±29 551 550±17
No.correctin10,000b
8933 8886 8896 8842
No.correctoptimumc
897±9 893±8 892±9 888±10
No.correctensemblethresholdd
892±10 887±9 888±9 887±9 886±9 893±9 856±9 872±9 880±9
Criticalthresholde
0.00606±0.00059
0.00593±0.00032
0.00569±0.00031
0.00639±0.00055
0.00613 0.00612 0.00635 0.00630 0.00627
Αf 0.00354 0.000358,0.00356
0.00360,0.00356
0.00368 0.00361 0.00354 0.00401 0.00380 0.00373
P0g 0.535 0.527,0.543
0.540,0.541
0.538 0.5435 0.549 0.537 0.551 0.554
σtrendh 0.00772 0.00781,0.00763
0.00771±0.00036
0.00773±0.0037
0.00772 0.00793 0.00775 0.00772
TableS1: Themodelparametersare:H=0.25,s=0.113(H is the fBmexponent,s is thevolatility parameter of the fBm, see equation in footnote 3). P0 = 0.54 for the Bernoulliprocess and σsproc for the random slope process (with the exception of the third line (asinglesampleof10,000),allthenumbersarewithrespecttoasampleof1000series,each135 points long and the uncertainties are with respect to such samples). With theexception of those indicated “regression” (two far right columns) all the trends wereestimatedusing themaximumlikelihoodmethodassumingan fBmprocess. The far rightcolumnisfromthedatawhenslopesareestimatedbythedifferencebetweenthefirstandlastmemberoftheseries(seefig.3).a If theabsolutetrendsaresorted in increasingorder, thentheoptimumindex isrankof thetrendthat should be chosen for the maximum number correct. Note the large sample to samplefluctuations.bThis is themaximumpossiblenumberofcorrectchoices inasampleof10,000; itwascalculatedwithknowledgeofthecorrectclassifications.cSameasbexceptformean(andstandarddeviation)ofthetensamples.d If only the ensemble optimum threshold was known (this is the realistic case), this is numbercorrect,samplesof1000.Itisslightlylessthanthe“optimum”numberintherowabove.eThisisthemean(andstandarddeviation)oftheoptimumthreshold.fAisthewidthoftheGaussianoftheslopesaboutthethreevalues-0.01,0,0.01(seeeq.S1).Fromthesimulations,theuncertaintieswithrespecttotheprocessparametersAis±0.000076andforP0,±0.016.gThisisthefractionofuntrendedseries(Bernoullitrendprocessb(t)=0).hThisisthestandarddeviationoftherawtrends.iThebootstrapparameterswereestimatedasdiscussed in the text; theyareprobablyoptimal forKeenan’sseries.
31
As expected, the estimate from the ensemble trend threshold yields slightly less
than use of the “optimum” threshold indicated in the row above. Although the
modelparameterspaceisbig(H,s,P0,σsproc),asexplainedabove,theparametersare
in fact constrained by a number of statistics: recall that H, s were chosen to
reproduce the fBm maximum likelihood estimates, P0, the empirical trend
distribution,σsprocthewidthAandtheoverallstandarddeviationofthetrendsσtrend.
Overallitseemsthatwecangettantalizinglyclose,to900correct,butnotquiteget
there.
Appendix:Sometheoryonrandomlyclassifyingtrends
In the climate contestproblem there are trendedanduntrended series. We
sawthatwecouldget towithinonestandarddeviationofawinningtechniqueby
usingthefBm(strongcorrelation)basedtrendestimatesandabootstrapprocedure
fordeterminingtheoptimumclassificationthresholdxc.Itisthereforetemptingto
“shuffle” or randomize the classification method so as to obtain many possible
guessessothatatleastonewillbeawinner.Unfortunately,thetheorybelowshows
thatthisisnotaveryfruitfulapproach.
Denotethe(unnormalized,relative)probabilitydensityofanabsolutetrendx
fromtheuntrendedpopulationaspu(x),fromthetrendedpopulation,pt(x).Lethere
beatotalnumberofNdraws(i.e.series,hereN=1000).
Theexpectationforthenumberofuntrendedandtrendedseriesis:
32
(A1)
Ifwenowusetherulethatanytrendx’≤xisclassifiedasuntrended,andanywith
trendx’>xistrended,thenthemeannumberoferrorsis:
(A2)
Theconditiontominimizethetotalmeanerroris:
(A3)
Whichimpliesthatthecriticalxcsatisfies:
(A4)
Fluctuations:
For a large number of series, the probabilitieswill approach a Gaussian,we
needtocalculatethevarianceofthetotalerror<E2>.FromarandomsampleofN
trends,theprobabilitythattheaboveprocedurewillmisclassifyagivenseriesisPE:
(A5)
Nu = N pu ′x( )d ′x0
∞
∫ ; Nt = N pt ′x( )d ′x0
∞
∫
pu ′x( ) + pt ′x( )( )d ′x0
∞
∫ = 1; N = Nu + Nt
Eu x( ) = N pu ′x( )d ′xx
∞
∫ ; Et x( ) = N pt ′x( )d ′x0
x
∫
∂ E∂x
= 0; E = Eu + Eu
pu xc( ) = pt xc( )
PE x( ) = pu ′x( )d ′x + pt ′x( )d ′x0
x
∫x
∞
∫
33
We seek the variance of theBernoulli process forwhich for a single series,E = 1
withprobabilityPE,0otherwise. TheoverallvarianceforNseries is thenN times
larger:
(A6)
Thus,therandomvariableEshouldbeaGaussianwith:
(A7)
(thenotationindicatesaGaussianwithmeanNPEandstandarddeviationafterthe
“±”).Inourcase,E≈100,N=1000,PE≈0.1sothat:E≈100±9asconfirmedbythe
numerics.
RandomizationbyaFlippingprocess:
Wecannowseetheeffectofrandomizingthevariablebyflippingeachchoice
with a probability density pf(x); this means that we first choose as before: we
classify a series as untrendedwhen x<xc, and as trendedwhen x>xc, but thenwe
reversethechoicewithprobabilitypf(x).
Bysimilarreasoningtoabove(calculationofthemean)wefindthattherewillbea
shiftδEinthenumberoferrorsfromE(x)toE’(x):
(A8)
MeanshiftδE:
E2 = NPE x( ) 1− PE x( )( )
E = NPE x( ) ± NPE x( ) 1− PE x( )( )⎡⎣ ⎤⎦1/2
δE x( ) = ′E x( )− E x( )
34
(A9)
sincefromtheaboveconditionforthecriticalxcwehave:
(A10)
and from the fact that pf≥0, we can see that δE≥0, so that no choice of pf(x) can
improvetheinitialchoicexc.Butwhataboutfluctuations?
Varianceofshift:
RepeatingthesameBernoulliargumentforthevariances,weobtain:
(A11)
Thus,theoverallresultisthat“flipping”increasestheerrorby:
(A12)
SinceδE≈9,σδE≈3(alsoconfirmedbynumerics). Weseethatflippingwillonly
decreasetheerroratthe3standarddeviationleveli.e.only0.13%ofthetime,and
thatthemethodofflipping(pf(x))is(surprisingly)irrelevant!Ifweare1standard
deviation(9counts)below900,thenweneeda6standarddeviationeventinorder
towin,i.e.wewouldneedaboutabillionentriestothecontest.
δE = N pf ′x( ) pu ′x( )− pt ′x( )( )d ′x0
xc
∫ + pf ′x( ) pt ′x( )− pu ′x( )( )d ′xxc
∞
∫⎡
⎣⎢⎢
⎤
⎦⎥⎥
pu x( ) > pt x( ); x < xcpu x( ) < pt x( ); x > xc
σδE = δE1/2 1− δEN
⎛⎝⎜
⎞⎠⎟1/2
≈ δE1/2
δE ± δE1/2
35
References:
Lovejoy,S.(2014a),Scalingfluctuationanalysisandstatisticalhypothesistestingof
anthropogenic warming, Climate Dynamics, 42, 2339-2351 doi:
10.1007/s00382-014-2128-2.
Lovejoy, S. (2014b), Return periods of global climate fluctuations and the pause,
Geophys.Res.Lett.,41,4704-4710doi:doi:10.1002/2014GL060478.
Lovejoy,S.(2015),Usingscalingformacroweatherforecastingincludingthepause,
Geophys.Res.Lett.,42,7148–7155doi:DOI:10.1002/2015GL065665.
Lovejoy,S.,andD.Schertzer(1986),Scaleinvarianceinclimatologicaltemperatures
andthelocalspectralplateau,AnnalesGeophysicae,4B,401-410.
Lovejoy, S., L. delRioAmador, andR.Hébert ( 2015), The ScaLIngMacroweather
Model (SLIMM): using scaling to forecast global-scale macroweather from
months to Decades,Earth Syst.Dynam., 6, 1–22 doi: http://www.earth-syst-
dynam.net/6/1/2015/,doi:10.5194/esd-6-1-2015.