Twin's paradox experiment is a meassurement of the extra dimensions.pptx
Bayesian Statistics as a New Tool for Spectral Analysis
1. BAYESIAN STATISTICS AS A NEW
TOOL FOR SPECTRAL ANALYSIS:
Application for Massive Stars Fundamental
Parameters Determination
Jean-Michel Mugnes
2. CLASSICAL SPECTRAL ANALYSIS
Aim: obtain stellar parameters: Teff, log g, vsin i ,
microturbulence (), macroturbulence,
abundances…
Many technics used: curve of growth, FFT, model
fitting « by eye » or with χ² calculation…
Iterative methods with Free & fixed parameters
a few lines used depending on their sensitivities.
(e.g. : Balmer lines -> log g & Teff, Si lines -> Teff, etc…)
3. THE CLASSICAL APPROACH
Iterative with Free & fixed parameters:
Build a Model grid (here TLUSTY Lanz & Hubeny 2007)
Teff & log g free
vsin i = 0 km.s-1
= 0 km.s-1
4. THE CLASSICAL APPROACH
Chi square analysis on Hbeta (vsin i & =0 km.s-1)
And it is only for
one line…
But what happens for different values of vsin i ?
Red diamond = Best
solution for a given vsin i
And for different values of ?
5. THE CLASSICAL APPROACH
And each line has it’s own « opinion »
The final results depends on the selected lines
And on the values of the fixed parameters.
Simultaneity is the key.
6. THE SIMULTANEOUS APPROACH
From free & fixed parameters to only free parameters.
Most
probable
Less
probable
« Free & fixed » fit:
χ² calculated for a given
vsin i and separatly
Simultaneous fit:
χ² calculated over all
values of Teff, log g, vsin i
and .
« Likelyhood of H »
7. DIFFERENT LINES, DIFFERENT LIKELYHOODS
Likelyhood =
Cexp ( - χ²/2σ²)
(here σ= 10 X σ_real)
a wide variety of
shapes
8. FROM AN ITERATIVE TO A SIMULTANEOUS METHOD
The Bayes Theorem:
Prior probability (line 1) Likelyhood (line 2)
Posterior probability
= prior probability for
line 3 , etc…
X
=
Likelyhood (line 1)
10. GOING FURTHER
Final probability
distribution for all the
parameters given by all
the lines in the spectrum
simultaneously
New model grid
Refined final
probability
(Remember that
σ= 10 X σ_real)
11. TESTING THE METHOD
Applied the method
on a randomly noised
synthetical spectrum
SNR going from 25 to
350 with steps of 25.
10 runs where
performed for each
SNR value
SNR=25
SNR=350
12. TESTING THE METHOD
Overall success rate is over 86 %
Around 78% for a SNR < 150
Around 92% for a SNR > 150
13. TESTING THE METHOD ON REAL SPECTRA
52 spectra of field and cluster B stars collected
at the Mont-Mégantic Observatory.
« Normal » stars : no binaries, chemicaly peculiar,
pulsating…
Well studied nearby stars.
Visible spectra between 3600 Å and 6000 Å, with
moderate resolution (fwhm=2.3 Å)
16. CONCLUSIONS
We have developed a new spectral analysis
method that :
simultaneously constraints all the parameters and all the
available lines
is robust against noise and uncertainties
is generally more accurate than the classical methods
is also fast, automated and gives the results with their
associated uncertainties
Works also with any given model atmosphere
(TLUSTY, ATLAS, PHOENIX,…)