Στατιστική ανάλυση και βελτιστοποίηση βιβλιοθηκών συνθετικών φασμάτων γαλαξιών και εφαρμογή στην Ευρωπαϊκή διαστημική αποστολή Gaia

Abstract

During the five years of the Gaia satellite operations, a huge database of a billion objects will be obtained, including stars of the Milky Way and the Local Group, galaxies, quasars, exoplanets, supernova and Solar System objects. This work was performed in the context of the preparation of the ESA Gaia mission, specifically for the needs of the Unresolved Galaxy Classifier (UGC), an algorithm which is going to classify and parametrize the millions of galaxy spectra Gaia will observe (Bellas-Velidis et al., 2010, 2012; Kontizas et al., 2011). In order to develop and operate UGC, libraries of synthetic galaxy spectra are required. These theoretical templates were created with PEGASE code ´ (in cooperation with IAP). The aim of this Thesis is to investigate the range of the values of some key PEGASE parameters, for which realistic synthetic ´ spectra are produced. SDSS observed spectra were also used to validate the synthetic ones and test the performance of UGC.The need to truncate the less realistic synthetic spectra from the library of Tsalmantza, Kontizas et al. 2009, resulted to the optimization of this library (Karampelas et al., 2010, 2011, 2012). The optimum synthetic galaxy spectra were catalogued in CDS (Centre de Donn´ees astronomiques de Strasbourg). This catalogue (Karampelas et al., 2011) is available through VizieR (http://cdsarc.u-strasbg.fr/viz-bin/qcat?J/A+A/538/A38) and through anonymous ftp (cdsarc.u-strasbg.fr - 30.79.128.5). In the context of the further optimization of the libraries of theoretical galaxy spectra, a more realistic library was produced for the Gaia BP/RP photometers wavelength range (330 - 1000 nm). Additionally, an analysis has been initiated in order to create both synthetic galaxy spectra with episodic star formation for BP/RP and the first set of high-resolution synthetic galaxyspectra (Rocca-Volmerange et al., 2011) for the Gaia RVS spectrograph wavelength range (847 - 874 nm). To accomplish the above, artificial intelligence and pattern recognition techniques were used, namely Support Vector Machines (SVM) and Principal Components Analysis (PCA).The structure of the Thesis is the following: The first chapter is introductory, the second chapter presents the analysis and the evaluation of the parameters of the theoretical models and the relevant data processing and results, the third chapter lists the conclusions and the future plans, while the last chapter contains the references and an Annex. Data analysis and results included in this Thesis are summarized below, classified per library of galaxy spectra (BP/RP semi-empirical, synthetic, new synthetic and spectra with episodic star formation and RVS synthetic). Semi-empirical library (BP/RP) Semi-empirical galaxy spectra were created by matching synthetic galaxyspectra with SDSS galaxy spectra. A χ 2 goodness of fit was implemented to the spectra and the left and right wings of the synthetic spectra were adjusted to the corresponding parts of the matching observed spectra, producing the final semi-empirical spectra. For each of these semi-empirical spectra the spectral type of the corresponding synthetic spectrum was adopted(Tsalmantza, Karampelas et al. 2012). The semi-empirical spectra were simulated for the Gaia instrument characteristics for three magnitudes (G = 15, 18 and 20) with the addition of noise. The simulated spectra were classified and parametrized with the SVM method. The percentage of the correct classifications was higher for brighter G-magnitude, ranging from 80 % (G = 20) to 92.2 % (G = 15). The same trend stands for each spectral type separately. However, Early-type and Irregular galaxies were better classified than Spirals and QSFG. Furthermore, the erroneously classified Early-type galaxies were classified mainly as Spirals and vice versa, while the erroneously classified irregular galaxies and QSFG were classified mainly as spirals. From the comparison of the above results with the corresponding results from the library of synthetic galaxy spectra (Tsalmantza, Kontizas et al. 2009) arises that the Early-type and Irregular galaxies were better classified in the semi-empirical case, while the contrary stands for Spirals and QSFG. The spectral type classification was in total more successful in the case of synthetic spectra, partially explained by the higher percentage of QSFG galaxies among them. The errors of the prediction of the redshift (parametrization) was lower for brighter G-magnitude, ranging from 3.2 % (G = 15) to 9.0 % (G = 20). The same trend stands independently of the spectral type and the classification correctness. However, Early-type and Spiral galaxies were better parametrized than Irregulars and QSFG. Furthermore, the prediction of the redshift of the correctly classified galaxies was better than the prediction of the erroneously classified ones by 1 % to 2%. Additionally, the mean absolute percentage errors of the redshift prediction were lower for z ≤ 0.2 galaxies (by 0.1 % for the correctly classified galaxies, by 0.2 % for all galaxies and by 0.8 % for QSFG with G = 20).Synthetic library (BP/RP) The PCA method was implemented to the synthetic spectra to perform the investigation of how realistic these theoretical templates are. The diversity of the synthetic spectra originates from the complexity induced by the theoretical assumptions under which they have been created. More specifically, it is due to (a) the different combinations of parameter values of the same star-formation law which may correspond to similar spectra, (b) the different combinations of parameter values of different star-formation laws which may correspond to similar spectra, and (c) the combinations of parameter values which may correspond to unrealistic or non-typical spectra. It was found that the library of synthetic galaxy spectra can be efficiently analyzed by only using the first two Principal Components - eigenspectra PC1 and PC2, which correspond to a very low mean reconstruction error of 2% and contain the 99 % of the total variance, thus allowing for a drastic dimensionality reduction without a significant loss of information. The stability of the most significant Components was confirmed after repeatedly implementing the PCA method to subsamples of synthetic spectra. PC1 was found to differ among the various spectral types, since it is correlated with the higher variance fraction, hence reflecting the major differences among these types. PC2 practically remains unchanged. Subsequently, the synthetic spectra were projected in the diagram of the first two Principal Components PC1 and PC2. Extended overlaps were discovered in this diagram, with Spiral galaxies been highly-overlapped with Early-type galaxies and QSFG, and Irregular galaxies been highly-overlapped with QSFG. At the same time, the star formation scenarios of each spectral type were investigated in detail and, together with the results from the PCA implementation, they led to specific criteria to truncate the non-desirablesynthetic galaxy spectra from the synthetic library. The optimization of the library was done by applying these criteria to the synthetic spectra (Karampelas et al., 2010, 2012), truncating (a) Early-type galaxies with emission lines (similarity with Spirals) or with very weak star formationrate, (b) Spiral galaxies with high p2 values (similarity with Irregulars and QSFG), (c) Irregular galaxies and QSFG with extremely high values of p2 and tinfall (similar to each other) and (d) QSFG without emission lines (sim-ilarity with Spirals). The optimum synthetic galaxy spectra are available at the CDS (Centre de Donn´ees astronomiques de Strasbourg) via anonymous ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via http://cdsarc.u-strasbg.fr/vizbin/qcat?J/A+A/538/A38.As a result of the optimization, the Early-type galaxies form a distinct group in the PC1 - PC2 diagram, no longer overlapping with Spiral galaxies and lacking of galaxies with non-desirable weak star formation rates. Additionally, Spiral - QSFG overlaps have been significantly limited, together with the existence of similar to each other late type galaxies. The overlaps in the (g − r) − (r − i) two-color diagram have also been limited, as a consequence of the optimization. More specifically, the Early-type galaxies form a distinct group at (g − r) > 0.8, while the extended overlap between Spirals and QSFG at 0.5 < (g − r) < 0.6 no longer exists, with the satisfactorycoverage of the SDSS observation having been retained. The simulated galaxy spectra of the synthetic library (Tsalmantza, Kontizas et al. 2009) and the optimized synthetic library (Karampelas et al., 2012) were classified and parametrized by using the SVM method in two different G = 15 samples: spectra with and without noise, extinction and redshift. In the first case, UGC performance was approximately 100 %. In the second case, the optimized Early-type galaxies were better classified by 25 % and Spirals and Irregulars by 3 %, while the classification performance in QSFG spectra remained unchanged, being already high. Regarding parametrization, most of the values of the various parameter were predicted more accurately in the optimized spectra. Nevertheless, the parametrization errors cannot be easily interpreted, since the different values of the parameters do not linearly affect the shape of the final spectra. Conclusively, the optimization of the library of synthetic galaxy spectra achieved through the investigation of the star formation histories of the various spectral types and the application of the PCA method (Karampelas et al., 2012), led to a better understanding of PEGASE modeling and of ´ some statistical methods which are able to efficiently analyze and process huge databases. New synthetic library (BP/RP)The existence of criteria to create optimal synthetic galaxy spectra and the constant need of even more realistic spectra for UGC led to the creation of a new synthetic library with PEGASE.2 code. These spectra (a) correspond to ´ ranges of star formation rate parameters values which take into account the optimization criteria, (b) include the special case of Luminous Red Galaxies (LRG), which have been modeled as high-redshift Early-type galaxies, (c) have been constructed for various values of disk inclination, and (d) have been evolutionary corrected. Early-type galaxies have been created both for zero redshift values and for random values in the range 0.0 - 0.6, while Spiral galaxies, Irregulars and QSFG have been created for zero redshift, for random redshift values in the range 0.0 - 0.2 and for five different inclination values. The statistical analysis of these spectra is ongoing. Spectra with episodic star formation (BP/RP) In episodic star formation, stars are born during bursts of star formation. An investigation of such scenarios, being modeled as an alternative version to scenarios of Irregular galaxies, has been started, with the corresponding star formation rate parameters having been defined. These parameters control thenumber, the starting time, the duration and the intensity of each burst, and appropriate constraints have been applied to them, resulting in the creation of 240 scenarios. The creation of the corresponding spectra with PEGASE ´ is ongoing. Synthetic library (RVS) The creation of the first library of high-resolution synthetic galaxy spectra for the RVS spectrograph of the Gaia satellite requires the inclusion of an appropriate library of high-resolution synthetic stellar spectra to PEGASE ´ code. The stellar library of Munari et al. (2005), which has been created specifically for RVS, was selected among the relevant available libraries. A number of preliminary tests have been implemented using BaSeL, the stellar library included in PEGASE.2. Synthetic galaxy spectra of Early-type, Spiral ´ and Irregular galaxies have been created and the RVS wavelength range hasbeen investigated, where the dominant spectral characteristic is the calcium triplet. The inclusion of the library of Munari et al. (2005) to PEGASE is ´ongoing.

Η παρούσα διδακτορική διατριβή πραγματοποιήθηκε στα πλαίσια της προετοιμασίας της Ευρωπαϊκής διαστημικής αποστολής Gaia για τις ανάγκες του συστήματος Unresolved Galaxy Classifier (UGC), το οποίο πρόκειται να ταξινομήσει και να παραμετροποιήσει τα εκατομμύρια των φασμάτων γαλαξιών που πρόκειται να παρατηρηθούν (Bellas-Velidis et al. 2010, 2012; Kontizas et al. 2011), χρησιμοποιώντας συνθετικά φάσματα κατασκευασμένα με τον κώδικα PEGASE. Στόχος αυτής της διατριβής είναι η διερεύνηση της σημασίας των θεωρητικών παραμέτρων του κώδικα PEGASE και τα εύρη των τιμών για τα oποία δημιουργούνται φάσματα που ανταποκρίνονται στην πραγματικότητα. Η βιβλιοθήκη συνθετικών φασμάτων γαλαξιών των Tsalmantza, Kontizas et al.2009 βελτιστοποιήθηκε (Karampelas et al. 2010, 2011, 2012) και τα βέλτιστα συνθετικά φάσματα γαλαξιών καταλογοποιήθηκαν και αναρτήθηκαν στο Κέντρο Αστρονομικών Δεδομένων CDS Karampelas et al. 2011), όντας διαθέσιμα μέσω της υπηρεσίας VizieR (http://cdsarc.u-strasbg.fr/viz-bin/qcat?J/A+A/538/A38). Δημιουργήθηκε επίσης μία νέα περισσότερο ρεαλιστική βιβλιοθήκη συνθετικών φασμάτων γαλαξιών για τα φωτόμετρα BP/RP (λλ 330 - 1000 nm) και διερευνήθηκε η δημιουργία φασμάτων επεισοδιακής αστρογέννεσης και φασμάτων για τον φασματογράφο RVS (λλ 847 - 874 nm) του δορυφόρου Gaia (Rocca-Volmerange et al. 2011). Στην παραπάνω διερεύνηση, και πέραν των απλών στατιστικών αναλύσεων, χρησιμοποιήθηκαν οι στατιστικές μέθοδοι τεχνητής νοημοσύνης και εύρεσης προτύπων SVM και PCA.