why do wouter (and atlas) put asymmetric errors on data points ?

50
Why do Wouter (and ATLAS) put asymmetric errors on data points ? What is involved in the CLs exclusion method and what do the colours/lines mean ? ATLAS J/Ψ peak (muons) Excluding SM Higgs masses LEP exclusion Tevatron exclusion

Upload: lassie

Post on 08-Jan-2016

13 views

Category:

Documents


2 download

DESCRIPTION

What is involved in the CLs exclusion method and what do the colours/lines mean ?. Why do Wouter (and ATLAS) put asymmetric errors on data points ?. ATLAS J/ Ψ peak (muons). Excluding SM Higgs masses. LEP exclusion. Tevatron exclusion. Why do you put an error on a data-point anyway ?. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Why do Wouter (and ATLAS) put asymmetric errors on data points ?

What is involved in the CLs exclusion method and what do the colours/lines mean ?

ATLAS J/Ψ peak (muons)Excluding SM Higgs masses

LEP exclusion Tevatron exclusion

Page 2: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Why do you put an error on a data-point anyway ?

ATLAS J/Ψ peak (muons)

Estimate of underlying truth (model value)

Page 3: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Poisson distribution

Poisson distribution

Probability to observe n eventswhen λ are expected

λ=4.90

Number of observed events

#observed #observed Lambda hypothesisLambda hypothesis

fixedfixedvaryingvarying

P(n | λ ) =λne−λ

n!

P(0 | 4.9) = 0.00745

P(2 | 4.9) = 0.08940

P(3 | 4.9) = 0.14601

P(4 | 4.9) = 0.17887

Page 4: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Poisson distribution: properties

Poisson distribution

properties

the famous √N

(1) Mean:

(2) Variance:

(3) Most likely value: first integer ≤ λ

http://www.nikhef.nl/~ivov/Statistics/Poisson.pdf

P(n | λ ) =λne−λ

n!

⟨x⟩= λ

⟨(x − ⟨x⟩)2⟩= λ

Page 5: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Lambda known expected # events

λ=0.00 λ=1.00

λ=4.90λ=5.00

Page 6: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Large number of events

λ=40.0

Unfortunately this is not what you wanted to know …

What you have: What you want:

P(Nobs | λ )

P(λ |Nobs)

Page 7: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

From data to theory

Likelihood: Poisson distribution“what can I say about the measurement (Number of observed events) given an expectation from an underlying theory ?”

This is what you want to know: “what can I say about the underlying theory given my observation of a given number of events ?”

P(λ |Nobs) = P(Nobs | λ )P(λ )

Page 8: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

λ (hypothesis)

Nobs known (4) information on lambda

“Given a number of observed events (4): what is the most likely / average / mean underlying true vanue of λ ?”

#observed #observed Lambda hypothesisLambda hypothesis

fixedfixed

P(N

ob

s=

4|λ

)

varyingvarying

Normally you plot -2log(Likelihood)

Likelihood:

P(4 | 0) = 0.00000

P(4 | 2) = 0.09022

P(4 | 4) = 0.19537

P(4 | 6) = 0.13385

Page 9: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Properties of P(λ|N) for flat P(λ)

properties

(1) Mean:

(2) Variance:

(3) Most likely value: λmost likely = x

http://www.nikhef.nl/~ivov/Statistics/Poisson.pdf

P(λ |Nobs) = P(Nobs | λ )P(λ )

⟨λ⟩ =x +1

⟨(λ − ⟨λ ⟩)2⟩= x +1

Assuming P(λ) is flat

Page 10: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

This is normally presented as likelihood curve

λ (hypothesis)

P(N

ob

s=

4|λ

)-2

Log

(P(N

ob

s=

4|λ

))

Likelihood

-2Log(Prob)

4.002.32

-1.68

6.35

+2.35

sigma: ΔL=+1 ΔL=+1

68.4%

Pdf for λ

Page 11: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

ATLAS J/Ψ peak (muons)

4−1.68+2.35

So, if you have observed 4 eventsyour best estimate for λ is … :

Page 12: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

CLS method

http://www.nikhef.nl/~ivov/Statistics/thesis_I_v_Vulpen.pdf

Chapter 7.4

Page 13: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Your Higgs analysis

Discriminant variable Discriminant variable

Higgs

SM

Hebben we nou de Higgs gezien of niet ?

Higgs

SM

SM+Higgs

Scaled to correct cross-sections and 100 pb-1

Can also be an invariant mass plot

Page 14: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Approach 1: counting

Discriminant variable Discriminant variable

tellen tellen

Experiment 1 Experiment 2

Origin # events

SM 12.2

Higgs 5.1

MC total 17.3

Data 11

Origin # events

SM 12.2

Higgs 5.1

MC total 17.3

Data 17

Page 15: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Expectations

If the Higgs is NOT there:On average 12.2 events

If the Higgs is there:On average 17.2 events

Experiment 2:17 events observed

Experiment 1:11 events observed

SM SM + Higgs

Page 16: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Discovery

- Only look at what you expect from Standard Model background- Given the SM expectation: if probability to observe as many events you have observed (or more) is smaller than 5.7 10-7

SM hypothesis is very unlikely reject SM discovery !

- Only look at what you expect from Standard Model background- Given the SM expectation: if probability to observe as many events you have observed (or more) is smaller than 5.7 10-7

SM hypothesis is very unlikely reject SM discovery !€

Ppoisson (N |NSM )dN < 5.7Nobs

∫ 10−7

Page 17: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Test hypotheses: rules for discovery

In the hypothesis that there is NO Higgs (SM hypothesis):

What is the probability to observe as many events as I have observed …OR EVEN MORE

If P < 5.7 10-7 reject SM

P(N≥33|12.2) = 6.35 10-7

P(N≥34|12.2) = 2.24 10-7

P(N≥33|12.2) = 6.35 10-7

P(N≥34|12.2) = 2.24 10-7

Integrate this plot

SM + HiggsSM

Page 18: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Question 1: did you make a discovery ?

See previous slide:

Yes

Discovery No discovery

No€

Ppoisson (N |NSM )dN < 5.7Nobs

∫ 10−7€

Ppoisson (11 (or17) |12.2)dN > 5.711 (or 17)

∫ 10−7

Page 19: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Question 2: did you expect to make a discovery:

If the Higgs is NOT there:On average 12.2 events

If the Higgs is there:On average 17.2 events

If you observe exactly the number of events you expect (assuming the Higgs is there), it is not unlikely enough to be explained by the SM

NO discovery expected

If you observe exactly the number of events you expect (assuming the Higgs is there), it is not unlikely enough to be explained by the SM

NO discovery expected

SM SM + Higgs

Ppoisson (N |12.2)dN = 0.0717

Page 20: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Question 3: At what luminosity do you expect to make a discovery ?

Lumi x 1Lumi x 1

NSM = 12.2

NHiggs = 5.1

NSM = 122.0

NHiggs = 51.0

Lumi x 10Lumi x 10

Lumi x 12.5Lumi x 12.5

NSM = 152.5

NHiggs = 63.75

no

no

yes

SM + Higgs

SM + HiggsSM

SM

Ppoisson (N |12.2)dN = 0.0717

Ppoisson (N |122)dN = 5.5 10−6

173

Ppoisson (N |152.5)dN = 5.2 10−7

216

Page 21: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Discovery or not

It is not likely you get exactly the number of events you expect.

You can be lucky … or unlucky.

Page 22: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

From simple counting to the real thing in 3 steps

1) Introduce X (Likelihood ratio) test statistic

2) From simple counting to weighted counting (a real analysis)

3) Toy Monte-Carlo (fake experiments)

Page 23: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

From simple counting to the real thing in 3 steps

1) Introduce X (Likelihood ratio) test statistic

2) From simple counting to weighted counting (a real analysis)

3) Toy Monte-Carlo (fake experiments)

Page 24: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Hypothesis testing: likelihood ratio

frequently used: X=-2ln(Q)

Hypothesis 1: the Standard Model without the Higgs boson

Hypothesis 2: the Standard Model with the Higgs boson

Definieer een statistic (= variabele) die onderscheid maakt tussen de 2 hypotheses.Note: kan vanalles zijn: # events of Neural net output.

Ex: counting experiment

Q =Ls+b

Lb

Q =Ppoisson (n | λ s+b)

Ppoisson (n | λ b)

Likelihood ratio

Page 25: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Likelihood ratio: counting

Counting experiment

N events left after some a selection of cut on discriminant

Note: X = 0 means hypoteses equally likely

Used in plots:

More SM+Higgs like More SM like

100.000 SM experiments

100.000 SM + Higgs experiments

Q

Q =P(N | s +b)

P(N | b)

=e−(s+b)(s +b)n /n!

e−bbn /n!

=e−s (s +b)n

e−bbn

X = −2ln(Q)

14 events observed

Variabele transformatie

Page 26: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Likelihood ratio: counting

Counting experiment

N events left after some a selection of cut on discriminant

Note: X = 0 means hypoteses equally likely

Used in plots:

More SM+Higgs like More SM like

100.000 SM experiments

100.000 SM + Higgs experiments

Q =P(N | s +b)

P(N | b)

=e−(s+b)(s +b)n /n!

e−bbn /n!

=e−s (s +b)n

e−bbn

X = −2ln(Q)

P(14 |12.2) = 0.093

P(14 |17.3) = 0.076

X = 0.420

14 events observed

P(15 |12.2) = 0.076

P(15 |17.3) = 0.087

X = −0.278

15 events observed

Page 27: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

From simple counting to the real thing in 3 steps

1) Introduce X (Likelihood ratio) test statistic

2) From simple counting to weighted counting (a real analysis)

3) Toy Monte-Carlo (fake experiments)

Page 28: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Likelihood ratio

Counting experiment Weighted counting experiment

Eveny event has a weight according to a NN output or discriminant called pi : Signal: S(pi) and Background B(pi)

B(pi)

S(pi)+B(pi)

N events left after some a selection of cut on discriminant

tellen

Q =e−(s+b)(s +b)n /n!

e−bbn /n!

Q =e−(s+b)(s +b)n /n!

e−bbn /n!⋅

sS(pi) +bB(pi)

s+bi=1

n

∏B(pi)i=1

n

Page 29: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

From simple counting to the real thing in 3 steps

1) Introduce X (Likelihood ratio) test statistic

2) From simple counting to weighted counting (a real analysis)

3) Toy Monte-Carlo (fake experiments)

Page 30: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Many possible experiments

Discriminant variable Discriminant variable

tellen tellen

Experiment 1 Experiment 2

1) Experiment condensed in 1 variable Note: Each experiment (read ATLAS) yields only ONE value of Q see 2 slides ago for counting example 2) Do Toy-MC experiments to study distribution of Q Note: Two distributions: for SM and SM+Higgs hypothesis

Page 31: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Toy Monte Carlo experiment

SM toy experiment: Draw for each bin i a random number from Poisson with μ= λSM (i)

SM+Higgs toy experiment: Draw for each bin i a random number from Poisson with μ= λSM(i)+ λSM+Higgs(i)

λSM(i)+ λSM+Higgs(i)

λSM(i)

Page 32: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

The Higgs does not exist: 100,000 toy-experiments (SM)The Higgs exists: 100,000 toy-experiments (SM+Higgs)

Page 33: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

With 1 and 2 sigma bands for SM hypothesis

Note (again): each experiment will produce 1 (one) number in this plot

Page 34: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Different masses … different cross-sections

Small Higgs cross-section Large Higgs cross-section

Two hypotheses are more apart if: 1) cross-section of Higgs is larger 2) Higgs is more different from SM

Page 35: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

LEP plots

LEP paper Fig 1

Cross-section drops as function of mass

dummy

dummy

dummy

Page 36: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Expectation for Q or -2ln (Q): toy experiments

Probability that background resultsin the numer observed or (even) more

If 1-CLb < 5.7 10-7 we can say we reject the SM hypothesis discovery !

The famous 5 sigma

1- CL b = Pb(X ≥ Xobs) = Pb(X)dXX obs

Probability that background results in the numer observed or less€

CL b = Pb(X ≥ Xobs) = Pb(X)dXX obs

∫Clb = confidence level in the background

SMSM

SM+HiggsSM+Higgs

Page 37: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Discovery

Ppoisson (N |NSM )dN < 5.7Nobs

∫ 10−7

1 − CL b < 5.7 10−7

Page 38: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Do you expect to discover Higgs with at this mass ?

Average SM+Higgs experiment: 1-CLb = 2 10^-7So yes, you expect to make a discovery IF 10xSM

Page 39: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

The one 2-sigma is not the other 2-sigma

2.X sigma discrepancy at mh ~ 97 GeV Far away form what you expect from Higgs1.X sigma away at mh = 114 GeV Exactly what you expect from Higgs

No 5 sigma discovery what Higgs hypotheses can we reject

Page 40: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

No discovery

No 5 sigma deviation found … what now ?

Trying to say something on the hypothesis that the Higgs exists exclusion

Page 41: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Exclusion

CL s =CL s+b

CL b

< 0.05

- Look at what you expect from Standard Model +Higgs - Given the SM + Higgs expectation: if probability to observe as many events you have observed (or less) is smaller than 5% SM+Higgs hypothesis is not very likely reject SM+Higgs

- Look at what you expect from Standard Model +Higgs - Given the SM + Higgs expectation: if probability to observe as many events you have observed (or less) is smaller than 5% SM+Higgs hypothesis is not very likely reject SM+Higgs

Page 42: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Expectation for Q or -2ln (Q): toy experiments

If CLs < 0.05 we are allowed to rejectthe SM+Higgs at 95% confidence level

The famous 95% confidence level

Probability that signal hypothesis results in the numer observed or less

CL b = Pb(X ≥ Xobs) = Pb(X)dXX obs

Cls = confidence level in the signal

SMSM

SM+HiggsSM+Higgs

CL s+b = Ps+b(X ≥ Xobs) = Ps+b(X)dXX obs

CL s =CL s+b

CL b

Extra Normalisation:

This is why it is called modified frequentist

Page 43: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

CLs mean SM-only expeciment is 0.13 > 0.05 so NO !

Question 2: did you expect to be able to exclude ?

Page 44: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Question 3: At what luminosity do you expect to make a discovery ?

Lumi = 1x normal lumi

CLs = 0.13 no exclusion for average SM-only experiment

Lumi = 2x normal lumi

CLs = 0.034 exclusion for average SM-only experiment

#SM = 100 #H = 10

#SM = 200 #H = 20

Page 45: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

A scan:

Luminosity / nominal luminosity

CLs

CLs = 0.05

CLs = 0.13

CLs = 0.66

CLs = 0.046

2 sigma up

1 sigma down

Si: If you would have a 1 sigma downward fluctuation, i.e. you see less events than you expect there is less room for a SM+Higgs hypothesis. In this case you would have been able to exclude it.

You expect to be able to exclude at Lumi / Lumi nominal = 1.70

Page 46: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Question 4: At what Higgs xs do you expect to make a discovery ?

Higgs XS = 1x normal Higgs XS

CLs = 0.13 no exclusion for average SM-only experiment

Higgs XS = 2x normal Higgs XS

CLs = 0.006 exclusion for average SM-only experiment

#SM = 100 #H = 10

#SM = 100 #H = 20

Page 47: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

A scan:

Higgs XS / nominal Higgs XS

CLs

CLs = 0.05

CLs = 0.13

CLs = 0.66

CLs = 0.046

2 sigma up

1 sigma down

You expect to be able to exclude at Higgs XS / Higgs XS nominal = 1.40

Page 48: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

A projection along the CLs = 0.05 line

Hig

gs

XS

/ n

om

inal H

igg

s X

S

Nominal luminosity

SM only (mean)

At what Higgs XS scale factordo you expect to be able to exclude the Higgs hypothesis ?

SM only (1 sigma up)

SM only (2 sigma up)

SM only (2 sigma down)

SM only (1 sigma down)

1.4

Page 49: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Hig

gs

XS

/ n

om

inal H

igg

s X

S

1.4

You can now scan over Higgs masses

The important thing is of course what you actually measured

Page 50: Why do Wouter (and ATLAS) put asymmetric  errors on data points ?

Finito!