Äîêóìåíò âçÿò èç êýøà ïîèñêîâîé ìàøèíû. Àäðåñ îðèãèíàëüíîãî äîêóìåíòà : http://www.adass.org/adass/proceedings/adass94/ballesterp.ps Äàòà èçìåíåíèÿ: Tue Jun 13 20:42:47 1995 Äàòà èíäåêñèðîâàíèÿ: Tue Oct 2 02:53:15 2012 Êîäèðîâêà: Ïîèñêîâûå ñëîâà: ýòà êèëÿ

Astronomical Data Analysis Software and Systems IV
ASP Conference Series, Vol. 77, 1995
R. A. Shaw, H. E. Payne, and J. J. E. Hayes, eds.
Robust Data Analysis Methods for Spectroscopy
P. Ballester
European Southern Observatory, Karl­Schwarzschild­Str. 2, D­85748
Garching, Germany
Abstract. This paper describes various methods for the analysis of
spectroscopic data, particularly as they relate to wavelength calibration
of echelle­format spectra. Methods that are robust to outlier values are
highlighted.
1. Introduction
Classical statistical techniques usually lack the ability to reject outliers which
appear in real data, resulting from measurement errors, mixed distributions,
noise, etc. Different methods offer inherent robust properties in the presence
of contamination. The application of robust methods to the domain of spectral
data analysis is reviewed in this paper. A short introduction to robust regression
by Least­Median of Squares is presented in Section 2. This algorithm is applied
to the determination of echelle dispersion relations. Section 3 is an excerpt from
a recently published article (Ballester 1994) related to the Hough transform and
its application to echelle order detection and automated arc line identification.
Wavelet transform can be useful in complementing robust techniques for mul­
tiresolution analysis, and Section 4 presents an application of the Mexican hat
transform to the detection of spectral features.
2. Robust Regression
2.1. LS versus LMS
The sensitivity of least­squares (LS) regression to outliers is a traditional prob­
lem in data analysis, and as an alternative Rousseeuw (1987) introduced the
Least Median of Squares (LMS) which consists of minimizing the term:
med
n
(y i \Gamma ffx i \Gamma fi) 2
o
; i = 1; 2; ::; N (1)
This minimization cannot be obtained analytically and therefore requires re­
peated evaluations for different subsamples of size p drawn from the n observa­
tions. A complete trial would require m = C p
n evaluations which would rapidly
become impracticable for large n and p values. Rousseeuw & Leroy (1987) de­
termined the minimum number m of subsamples required to obtain a given
probability ff of drawing at least one subsample containing only good observa­
tions from a sample containing a fraction ffl of outliers. By requiring ff to be
1

2
sufficiently close to 1 (e.g 0.95 or 0.99), m can be determined for given values of
p and ffl: ff = 1 \Gamma (1 \Gamma (1 \Gamma ffl) p ) m
The smallest fraction ffl of contamination that can cause the estimator to
deviate arbitrarily from the estimate performed on the uncontaminated sample
is called the breakdown point of the estimator. In the case of the LS estimator,
only one outlier is sufficient to make the fitted parameters deviate arbitrarily
from the expected value, therefore the breakdown point of LS is 0%. The LMS­
based Reweighted Least Square (RLS) corresponds geometrically to finding the
narrowest strip covering half of the observations. The breakdown point of the
RLS is 50%. Most the other robust estimators do not attain a breakdown point
of 30% (Rousseeuw & Leroy 1987). In the following sections of this paper, the
applications have been realized using the PROGRESS algorithm available from the
Statlib statistical library (e­mail: statlib@lib.stat.cmu.edu ).
2.2. Iterative Dispersion Relation Determination
Wavelength calibration using an arc lamp exposure is a basic step in spectral
calibration, usually requiring a careful analysis of errors and selection of lines,
especially in echelle spectroscopy where arc spectra containing several hundreds
of lines are commonly found. Several problems occur when using standard poly­
nomial regression:
ffl LS regression assumes no error on the independent variables, whereas er­
rors can occur both on the wavelength, due to line blending, misiden­
tifications and quantization, and on the position, due to pixelation and
centering errors.
ffl The lack of robustness to contamination will cause any outlier or misiden­
tified line to affect the complete solution and introduce unnecessary errors
into the residuals.
ffl In order to minimize the number of initial identifications, low­order rela­
tions are usually involved at the beginning of the calibration process (e.g.,
the echelle relation), introducing model errors.
Robust regression is an adapted method to take care of these problems. By
combining Hough transform based automated arc line identification and robust
regression, it is possible to perform an automated calibration, providing as in­
formation the central wavelength and the average dispersion in a single order.
This first estimate is refined by Hough transform cross­matching and extended
to the complete spectrum using the echelle relation (m– = f(x)). However the
accuracy of this relation is limited by several factors: in particular, optical mis­
alignments occurring between echelle grating, cross­disperser, and detector. In
the first iteration the unknown rotation angle is introduced into the error model
of the echelle relation. Lines are then identified by a robust regression based
iterative loop.
2.3. Results
Figure 1 shows the calibration of a Th­Ar exposure taken with the EMMI spec­
trograph at La Silla observatory. The wavelength range covered is 380--940nm

3
Figure 1. Th­Ar exposure obtained with the EMMI spectrograph at
La Silla observatory and figure of residuals.
over 24 orders. The combination of arc line identification by Hough transform
and robust regression allows to perform the calibration with the only information
of central wavelength (679 nm \Sigma 50 pixels), average dispersion (0.03nm pixel \Gamma1 \Sigma
30%) and absolute order number (24) for the relative order 18. The method
requires no instrument dependent knowledge and allows for a relatively large
inaccuracy of the initial solution. The figure of residuals shows a final accuracy
of about 1=15 pixel rms. The Th­Ar line catalog was provided by H. Hensberge
and corrected for line blending (De Cuyper & Hensberge 1995).
3. Hough Transform
A description of the Hough transform and its astrophysical applications can be
found in Ballester (1994) and Ragazonni & Barbieri (1994). For the application
of the HT to the detection of echelle orders, we use a representation of the
HT assuming no preliminary segmentation of the image, since the sought­for
features are brighter than the background. Accumulating a function of the
intensity makes it possible to detect the orders by order of brightness.
The identification of arc lines consists of associating a list of line positions
in pixel space with a list of reference wavelengths. The principle of the method
is to perform all possible associations in a pixel­wavelength space. A three­
dimensional HT allows us to detect, within a given range of central wavelength
and average dispersion, the non­linear dispersion relation maximizing the num­
ber of associations. In the general case, the maximum is searched for in a cube of
the Hough space, providing the parameters (– c ; ff; fi) of the dispersion relation:
– = – c + ffx(1 + fix). Usually the simplification fi = 0 makes it possible to use
a two­dimensional HT.

4
Figure 2. Spectrum of a standard star and associated binary mask
(after scaling and translation), and dyadic Mexican hat transform of
the above spectrum decomposed on 7 scales. The coefficients of the
sharp positive features are concentrated at small scales. Coefficients
of the absorption lines are maximized at different scales, depending on
their spatial extension.
4. Wavelet Transform
The wavelet transform has been widely described (e.g., Starck, Murtagh, &
Bijaoui 1995) and consists of the convolution product of a function with an
analyzing wavelet. By choosing a wavelet which is the second derivative of a
smoothing function, the wavelet coefficients become proportional to the second
derivative of the smoothed signal. The Mexican hat transform involves the
wavelet: /(x) = (1 \Gamma x 2 ) exp (\Gammax 2 =2) which is the second derivative of a Gaus­
sian. Since the continuum of a spectrum varies smoothly, its second derivative
will show increased values at the position of sharper spectral features. Figure 2
shows the Mexican hat transform of a spectrum presenting emission features,
as well as absorption lines, of different widths. The wavelet coefficients become
maximum (in absolute value) at different scales depending on the spatial ex­
tension of the features. The Mexican hat transform can be used to generate
a multiresolution mask of the spectral features. The transformation applied to
the wavelet coefficients at the different scales includes segmentation, generation
of windows, and multiresolution recombination using a coarse­to­fine approach.
After segmentation at each scale, the coefficients are compared between suc­
cessive scales. The process starts from the largest scale and a feature will be
retained at the next scale if its associated wavelet coefficients are larger. After
recombination the mask is binarized and provides a schematic representation of
the spectrum, providing the positions of the continuum and of the features.
Acknowledgments. I would like to thank P. GrosbÜl, H. Hensberge, F.
Murtagh, and J. L. Starck for helpful discussions.
References
Ballester P. 1994, A&A, 1011
De Cuyper J. P., & Hensberge H. 1995, this volume, p. ??

5
Ragazzoni R., & Barbieri C. 1994, PASP, 106, 683
Rousseeuw P. J., & Leroy A. M. 1987, Robust Regression and Outlier Detection
(New York, Wiley)
Starck J. L., Murtagh, F., & Bijaoui, A. 1995, this volume, p. ??