210 likes | 408 Views
Use of neural networks for the identification of new z ≥ 3.6 radio QSOs from FIRST-SDSS DR5 R. Carballo Dpto. Matemática Aplicada y Ciencias de la Computación, UC J.I. González-Serrano Instituto de Física de Cantabria (CSIC-UC) C.R. Benn Isaac Newton Group, La Palma
E N D
Use of neural networks for the identification of new z ≥ 3.6 radio QSOs from FIRST-SDSS DR5 R. Carballo Dpto. Matemática Aplicada y Ciencias de la Computación, UC J.I. González-Serrano Instituto de Física de Cantabria (CSIC-UC) C.R. Benn Isaac Newton Group, La Palma F. Jiménez-Luján Dpto. Física Moderna, UC ; IFCA MNRAS submitted
Scientific aims for the selection of a sample of radio-loud QSOs at z ≥ 3.6 • Radio QSOs represent only 10% of whole population. Compared to RQ, the initial photometric samples of QSO candidates are less contaminated by stars. • RL QSOs have its own interest. What is their situation with respect to the whole QSO population ? • Is the radio-loudness distribution (radio to optical flux ratio) bimodal? • Which processes mark the radio-loud character? Black-hole mass, mass of the host galaxy • RL QSOs and BALs (radio emission as a test for orientation scenarios) • How does the fraction of RL QSOs relates to redshift, optical luminosity?. • Optical luminosity function of RL QSOs at z ≥ 3.6 (bright end)
Initial sample: FIRST - SDSS Phot DR5 pairs (DR5 Julio 2006) S(1.4 GHz) ≥ 1 mJy SDSS r ≤ 20.2 photometric quality SDSS r unresolved radio-optical separation ≤ 1.5” Overlapping area FIRST-SDSS 7391 deg2 Number of FIRST-SDSS matches 8663 Data 8 variables r, u-g,g-r,r-i,i-z, sep, Stot, Speak / Stot Spectroscopic classification from SDSS Spec DR5 with SDSS spectra L 4248 52 QSOs z≥ 3.6 3754 QSOs z< 3.6 230 stars 59 galaxies 153 unknown without SDSS spectra U 4415
redshift-colour relation for QSOs r i u g z
Classifier: separation in two classes • FIRST-SDSS pairs with SDSS-DR5 spectra L 4248 • 52 QSOs z ≥ 3.6 • 3754 QSOs z < 3.6 • 230 stars • 59 galaxies • 153 unknown • FIRST-SDSS pairs without SDSS-DR5 spectra U 4415 • What QSO types?: • BALs and absorbed at Lyα included • “Narrow-line” (Type II) QSOs excluded • Lobe-dominated QSOs excluded (4% of all radio morphologies) Supervised learning: classifier Application of the classifier
Logistic Discriminant • Search for the best hyperplane in the d-dimensional space of input variables separating the sample in two classes: C1 high-z QSO and C2 remaining sources • Input vector x • Output y(x) Є [0,1] • Target values if xЄ C1 theny(x)=1 • else y(x)=0 w are the parameters of the model
The error function is the mean square error Software:Neural Network Matlab Toolbox
Train-Test approach: leave-one-out Training set 4247 Test 1 The 4248 sources are used as test → 4248 trained NNs Class assignment: x belongs to class C1 if y(x) ≥ yc
Efficiency versus completeness for labelled sample yc : 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7,0.8, 0.9 from right to left Results r, u-g, g-r, r-i, i-z, sep Efficiency 50 / 81 62 ± 9 %contaminants: 1 star + 2 galaxies + 28 QSOs ( 19 with 3.2 ≤ z ≤ 3.6) Completeness50 / 52 96 ± 4 %
Application of the classifier to the 4415 unlabelled sources The sample of 58 high-z QSO candidates 31 Outside DR5 spectr. area 27 Inside DR5 spectr. area
Spectroscopic check: candidates Efficiency 24/40 = 60 ± 12 % (sample with fainter mags) a mags Previous surveys of RL QSOs at these redshifts (literature) eff < 20 % • NED NASA Extragalactic Database # 4: 3 High-z QSOs + QSO z = 3.3 • Follow-up spectroscopy at WHT ISIS, La Palma. 2007 April and July # 27: 17 High-z QSOs + 5 QSO 3.2 < z < 3.6 + 5 other • SDSS DR6 spectroscopy. 2007 July # 9: 4 High-z QSOs + QSO z = 3.4 + 4 other (+5 ISIS+1NED) (+1 ISIS+1NED)
Spectroscopic check: non candidates 4415 U – 58 candidates = 4357 non candidates NED: 382 QSOs but none with z ≥ 3.6 DR6: 898 with spectrum (some already included NED) most of them QSOs (765 / 898 = 85% ) and none with z ≥ 3.6 No evidence of having missed high-z QSOs →high completeness
The high-z QSO sample. How did the NNs helped us? 8663 FIRST-SDSS DR5 phot 4248 yes SDSS DR5 spec 52 Hz QSO 4415 no SDSS DR5 spec 58 Hz QSO candidates 24 confirmed (out of 40) 7 expected (out of 18 unobserved) From52to76 (52+24) → Sample increased by a factor 1.46
SDSS spectroscopic completeness for the selection of high-z QSOs • SDSS DR5 photometric area covered by DR5 and DR6 spec • 52 high-z QSO s DR5 spec + 10 DR6 spec • 7 high-z QSOs (NED, this work) • 15 without spectrum → 4 expected high-z QSOs • (62) / (62+11) = 62 / 73 = 85% for r ≤ 20.2 and 3.6 ≤z ≤4.6
Conclusions and future work • A logistic discriminant trained on FIRST-SDSS DR5 sources with available DR5 spectra allowed an efficient and highly complete selection of 3.6 ≤z ≤ 4.6 QSOs. Within this survey, efficiency and completeness were 62% and 96% respectively. • The application of the classifier to 4415 DR5 photometric sources without DR5 spectrum yielded 58 candidates. Spectroscopic identifications of 40 of them (NED, observations at the WHT, DR6) yielded efficiency 24/40 = 60%. More than 1000 sources classed as non high-z QSOs have spectrum from NED or DR6 and none of them is a high-z QSO, implying a high completeness. • We increased the high-z QSO sample from 52 to 76, i.e. a factor 1.46. 17 spectra from this work, 11 not included at DR6. The optical luminosity function of RL QSOs (at the bright end) will be computed from this sample. • From the new identifications we estimate the incompleteness of SDSS for the spectroscopic selection of 3.6 ≤z≤ 4.6 FIRST-QSOs to be around 15%, for r≤ 20.2.
Neural Networks: Schematics of a Multilayer Perceptron with a hidden layer and a single output d:p:1 x1 x2 x3 x4 x5 x6 xd y d 1
Results () WHT and DR6 apart
Efficiency versus completeness for labelled sample and 8 sets of variablesyc : 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7,0.8, 0.9 from right to leftInput variables:r, u-g, g-r, r-i, i-z sep radio flux peak-to-total ratio Results comp 96% eff 62% r,u-g,g-r,r-i,i-z, separation