Nuclear Outbursts in the Centers of Galaxies

Home , Galaxy Zoo, IPTF14hls, List of stellar streams, Phillips relationship, Seyfert galaxy, SN 393

A dissertation presented to the faculty of the College of Arts and Sciences of Ohio University

In partial fulﬁllment of the requirements for the degree Doctor of Philosophy

Reza Katebi December 2019

This dissertation titled Nuclear Outbursts in the Centers of Galaxies

by REZA KATEBI

has been approved for the Department of Physics and Astronomy and the College of Arts and Sciences by

Ryan Chornock Assistant Professor of Physics and Astronomy

Florenz Plassmann Dean, College of Arts and Sciences 3 Abstract

KATEBI, REZA, Ph.D., December 2019, Physics Nuclear Outbursts in the Centers of Galaxies (182 pp.) Director of Dissertation: Ryan Chornock This dissertation consists of two parts. In the ﬁrst part, we focus on studying the nuclear outbursts in the centers of galaxies and their nature in order to better understand the behavior of central Super Massive Black Holes (SMBHs) and their interaction with the surrounding environment, and to better understand the accretion disk structure. Nuclear outbursts can be better understood by studying the changes in the broad emission lines and the underlying continuum. We quantify the properties of these nuclear outbursts using multi-wavelength observations including optical, ultraviolet, and X-rays from MDM Observatory, the Sloan Digital Sky Survey, Swift, and Magellan. Some of these nuclear outbursts are linked to Tidal Disruption Events (TDEs) and nuclear supernovae (SNs), while a number of these events are proposed to be a rare phenomenon called “changing-look” Active Galactic Nuclei (AGN). These types of AGNs have been observed to optically transition from type 1 to type 2 and vice versa on timescales of months to years, where broad emission lines such as Hα and Hβ appeared or disappeared followed by an increase or decrease in the continuum light. We investigate two transient events called PS1-13cbe and PS1-10cdq that were observed during outburst by the PS1 survey in 2013 and 2010, respectively. We investigate TDE, SN, and AGN activity as the three possible scenarios for the nature of these events. In the case of PS1-13cbe, we conclude that “changing-look” behavior caused by thermal accretion disk instabilities is the most plausible explanation for the outburst. However, in the case of PS1-10cdq, we favor the tidal disruption scenario because of the structure of the lightcurve and spectral evolution. In the second part of this dissertation, we focus on galaxy morphology prediction using a newly designed neural network called “Capsule Networks”. We 4 automate the process of morphology prediction and eliminate the need for feature engineering and heavy data prepossessing prior to classiﬁcation. We also reconstruct the galaxy images while preserving the brightness structure of the galaxies. This study provides one of the possible solutions for classifying objects for the era of large sky surveys where the amount of data for galaxies will dramatically increase and automated methods will play a very crucial role. 5 Dedication

To my father who taught to me to question everything To my mother who has always encouraged me to pursue my dreams To my sisters who have always supported me To my wife Volha who truly inspires me to be the best of myself and To everyone who is brave enough to wander in the realm of unknown and seek answers. 6 Acknowledgments

During my PhD, I had the privilege of meeting so many kind and smart people that I want to thank for their help and support. But at first I want to take a moment and express my deepest gratitude toward my advisor Professor Ryan Chornock who taught me the real meaning of science. Whenever I was struggling with a problem he was there to help me to move toward the right direction with his exceptionally vast knowledge. He taught me the correct path of scientific research and made me the physicist I am today. I would like to acknowledge him for his endless support throughout my PhD; no matter where I am or whoever I become, I will always be his lifelong student. I would like to thank physics and astronomy staff members for their effort to provide an efficient and proper atmosphere for research and learning. I would like to express my gratitude toward the members of my dissertation committee Dr. Razvan Bunescu, Dr. Madappa Prakash, and Dr. Joseph Shields for kindly accepting to be involved in my PhD. I would like to specially thank my wife Volha who has always supported me with love and inspired me to be better and stronger. At last but not least, I would like to thank my mother Sakineh, and my sisters Soheyla, Soraya, and Romina for always being there for me and their continuous support. 7 Table of Contents

Page

Abstract...... 3

Dedication...... 5

Acknowledgments...... 6

List of Tables...... 10

List of Figures...... 11

List of Acronyms...... 15

List of Sorted Acronyms...... 17

1 Introduction...... 19 1.1 Active Galactic Nuclei...... 19 1.1.1 AGN Classiﬁcation...... 20 1.1.2 LINERs...... 24 1.1.3 AGN Uniﬁcation...... 25 1.1.4 Accretion Process and Variability in AGNs...... 26 1.2 Changing Look AGNs...... 30 1.3 Tidal Disruption Events...... 33 1.4 Supernovae...... 38 1.4.1 Superluminous Supernovae...... 38 1.4.2 Sources that Power Superluminous Supernovae...... 39 1.4.2.1 Circumstellar Interaction...... 39 1.5 The Modular Open Source Fitter for Transients...... 41 1.6 Galaxy Morphology Prediction...... 42 1.6.1 Deep Learning...... 44 1.6.1.1 Feed-Forward Neural Networks...... 44 1.6.1.2 Convolutional Neural Networks...... 46

2 PS1-13cbe: The Rapid Transition of a Seyfert 2 to a Seyfert 1...... 49 2.1 Introduction...... 49 2.2 Observations of PS1-13cbe...... 51 2.2.1 Optical Photometry...... 52 2.2.2 Observations of the Host Galaxy...... 54 2.2.3 X-ray Photometry...... 55 2.2.4 Optical Spectroscopy...... 55 8

2.3 Observational Features of PS1-13cbe...... 57 2.3.1 Host galaxy of PS1-13cbe...... 57 2.3.2 Astrometry...... 60 2.3.3 Multi-band Light Curves of PS1-13cbe...... 61 2.3.4 Spectral Features of PS1-13cbe...... 65 2.4 Interpretation of the Features of PS1-13cbe...... 69 2.4.1 Type IIn Supernovae Interpretation of PS1-13cbe...... 70 2.4.2 PS1-13cbe as a TDE...... 72 2.4.3 PS1-13cbe as a “Changing Look” AGN...... 74 2.4.3.1 Obscuration of the AGN...... 75 2.4.3.2 Tidal Disruption Events...... 77 2.4.3.3 Accretion disk instabilities...... 77 2.4.4 Comparison to other Changing Look AGNs...... 79 2.5 Conclusions...... 80

3 PS1-10cdq...... 89 3.1 Observations of PS1-10cdq...... 89 3.1.1 Optical Photometry...... 89 3.1.2 Observations of the Host Galaxy...... 90 3.1.3 X-ray Photometry...... 90 3.1.4 Optical Spectroscopy...... 91 3.2 Observational Features of PS1-10cdq...... 92 3.2.1 Host Galaxy of PS1-10cdq...... 92 3.2.2 Multi-band Light Curves of PS1-10cdq...... 94 3.2.3 Spectral Features of PS1-10cdq...... 96 3.3 Interpretation of Features of PS1-10cdq...... 107 3.3.1 PS1-10cdq as a Variable AGN...... 108 3.3.2 PS1-10cdq as SLSN...... 111 3.3.3 PS1-10cdq as a TDE...... 113 3.4 Conclusions and Future Remarks...... 118

4 Galaxy morphology prediction using capsule networks...... 122 4.1 Introduction...... 122 4.2 Galaxy Zoo 2...... 125 4.3 Related Work...... 126 4.4 Approach...... 127 4.4.1 Experimental Setup...... 128 4.4.2 Data Preprocessing...... 128 4.4.3 Capsule Network...... 129 4.4.4 Network Architecture...... 131 4.4.4.1 Capsule Network...... 131 4.4.4.2 Baseline Network...... 132 4.4.5 Implementation and Resources...... 133 9

4.5 Results...... 134 4.5.1 Regression...... 134 4.5.2 Classiﬁcation Based on Answers to Question 1 and Reconstruction of Galaxies...... 135 4.6 Conclusions...... 137

5 Conclusions and Outlook...... 142 5.1 Conclusions...... 142 5.2 Outlook...... 145

References...... 147

Appendix: Supplementary Material...... 178 10 List of Tables

Table Page

2.1 Photometry of PS1-13cbe...... 82 2.1 continued ...... 83 2.1 continued ...... 84 2.1 continued ...... 85 2.1 continued ...... 86 2.1 continued ...... 87 2.2 Luminosity of Broad Lines...... 88

3.1 Photometry of the host of PS1-10cdq...... 92 3.2 Late time spectra of the host of PS1-10cdq (Walter Baade (WB), and center (CN))...... 93 3.3 Luminosity of Broad Lines...... 105 3.4 Model parameters from ﬁtting the CSM model from MOSFiT to the light-curve of PS10cdq...... 115 3.5 Model parameters from ﬁtting the TDE model from MOSFiT to the light-curve of PS10cdq...... 120

4.1 Computed RMSE between the predictions and true crowd-sourced probabilities. A relative error reduction of 8.8% was achieved...... 134 4.2 Training and testing accuracy vs number of epochs for classiﬁcation based on the answers to question one. A relative error reduction of 36.5% was achieved. 137 11 List of Figures

Figure Page

1.1 A representation of a feed forward neural network where HL stands for hidden layer. All of the layers are fully connected to each other...... 46 1.2 A schematic representation of a convolutional layer followed by a pooling layer. Each neuron in layer l + 1 is only connected to a local region of feature maps from layer l. The pooling layer applies a mean or max function on the local region of all feature maps from the convolutional layer l + 1...... 48

2.1 The observed transient luminosities of PS1-13cbe from the PS1 survey in grizyP1 filters after correction for Galactic extinction. S: marks the epoch of the LDSS spectrum (MJD 56570)...... 53 2.2 Light curves of PS1-13cbe in grizyP1 bands, corrected for Galactic extinction. Note the small-scale fluctuations in the light curve (such as near −20 d)..... 54 2.3 Optical spectra for PS1-13cbe, from bottom to top: host galaxy of PS1-13cbe from SDSS (grey), model of the host galaxy generated by FAST 1.0 (black), spectrum during outburst observed using LDSS3 (purple), spectra obtained with OSMOS in red and blue for the red and blue setups, respectively, and the most recent OSMOS spectrum (olive green)...... 57 2.4 The excitation diagrams using the [N II]λ6583/Hα, [O I]λ6300/Hα, [O III]λ5007/Hβ and [O III]λ5007/[O II]λ3726 line ratios (Baldwin et al., 1981; Kewley et al., 2006). The navy blue squares display the position of the host galaxy of PS1-13cbe. The shaded area represents the location of the SDSS galaxies calculated using MPA-JHU where darker regions represent higher number density of the galaxies. The extreme star formation line (solid red; Kewley et al., 2001, 2006), the revised star formation line (dotted black; Kauffmann et al., 2003) and the Seyfert-LINER classification line (dashed black; Kewley et al., 2006; Fernandes et al., 2010) are also plotted. Comp: AGN/star forming composites...... 59 2.5 SED of SDSS J2221+0030 in quiescence (purple circles), scaled SED templates of an Sb spiral galaxy (dotted orange line) and a Seyfert 2 galaxy (dashed blue lines) from the SWIRE template library (Polletta et al., 2007). The host exhibits a mid-infrared (MIR) excess relative to the star-forming template due to an AGN...... 61 2.6 The gP1 PS1 template image of SDSS J2221+0030 showing the position of the centroid of the galaxy (green “X”), the position of PS1-13cbe (red dot), and its position uncertainty is shown by a circle with 5σ radius (red circle). The position uncertainty is dominated by the outburst...... 62 12

2.7 Total optical luminosity light curve of PS1-13cbe integrated over the grizyP1 filters and relative to the baseline flux in the template images. The total optical luminosity was estimated using the spectral distribution at each epoch and the trapezoidal rule (blue triangles). The estimated luminosity at early and late times using yP1 (red circles) and zP1 (purple squares) bands assumed the same colour correction as measured from those epochs with all filters...... 63 2.8 Top: Rest-frame blackbody temperature from fitting the optical photometry. α Middle: Rest-frame spectral index from fitting a power law fν ∝ ν to the optical photometry, where α is the spectral index. Bottom: observed g-r colour diagram. All three panels show little or no evolution during the outburst..... 64 2.9 Continuum subtracted Hα line profiles. Top: Multiple-component Gaussian fit to the Hα+[NII] emission lines (blue), individual components (dashed green) and broad component of the Hα (dashed black). Bottom: The fit residuals. Top-left (on each panel): Observation date and numbers of days before/after the peak (navy blue)...... 67 2.10 Continuum subtracted Hβ line profiles. Top: Multiple-component Gaussian fit to the Hβ+[O III] emission lines (blue), individual components (dashed green) and broad component of the Hβ (dashed black). Bottom: The fit residuals. Top-left (on each panel): Observation date and numbers of days before/after the peak (navy blue). Strong residuals are present in the red wing of λ4959 due to poor subtraction of the 5577Å sky line...... 68 2.11 Optical spectrum of PS1-13cbe taken with LDSS3 during the outburst after subtracting the stellar continuum (blue). The spectra of the comparison objects include a QSO (SDSS QSO; Vanden Berk et al., 2001) template, a SN Type IIn (SN1994Y; Filippenko, 1997) and a TDE (ASASSN-14li; Holoien et al., 2015) from bottom to top (black), the emission lines are labeled (purple)...... 71

3.1 The observed Galactic extinction corrected luminosities of PS1-10cdq from the PS1 survey in grizyP1 ﬁlters. S: marks the epoch of the LDSS spectrum (MJD 55626)...... 90 3.2 Galactic extinction corrected light curves of PS1-10cdq in grizyP1 and Galex NUV bands...... 91 3.3 Optical spectra of PS1-10cdq. Spectrum ∼ 84 days after the peak of outburst observed using LDSS3 in dark red. Spectrum taken ∼ +352 days after the peak of outburst in purple. For detailed information about late time spectra shown in navy blue see Table 3.2. All the spectra are binned for the purpose of better visualization...... 94 3.4 Spectrum of the host galaxy generated by combining three late time spectra (spectra taken by IMACS on January, March, and May of 2017; see Table 3.2) (blue), the model of the host galaxy simulated with the FAST 1.0 code (Kriek et al., 2009) (orange)...... 95 13

3.5 SED of SDSS J100229+0130 in quiescence using observed and archival data (purple circles) (see Section 3.1.2), SED of PS1-13cbe at outburst from PS1 survey (green stars), scaled SED templates of an Sb spiral galaxy (dotted orange line) and a Seyfert 2 galaxy (dashed blue lines) from the SWIRE template library (Polletta et al., 2007). The observed MIR excess in the host relative to the star-forming template is due to an AGN...... 96 3.6 Total optical luminosity light curve of PS1-10cdq including the baseline that was calculated by integrating over grizp1 bands using trapezoidal method at each epoch...... 97 3.7 Top: Rest-frame blackbody temperature from fitting the optical photometry. α Middle: Rest-frame spectral index from fitting a power law fν ∝ ν to the optical photometry, where α is the spectral index. Red squares: represent fits where in addition to PS1 optical data, NUV data points from Galex are available. Bottom: observed g-r colour diagram. Strong evolution is evident in all panels...... 98 3.8 Continuum subtracted Hα+[N II] line profiles. Top: Multiple-component Gaussian fit to the Hα+[N II] emission lines (blue), individual components (dashed green) and broad component of the Hα (dashed black). Bottom: The fit residuals. Top-right (on each panel): Observation date and numbers of days after the peak (black)...... 101 3.9 Continuum subtracted Hβ line profile. Top: Multiple-component Gaussian fit to the Hβ emission line (blue), individual components (dashed green) and broad component of the Hβ (dashed black). Bottom: The fit residuals. Top-right (on each panel): Observation date and numbers of days after the peak (black).... 104 3.10 Continuum subtracted Mg II λλ2800. Top: Single Gaussian fit to the Mg II (blue). Bottom: The fit residuals. Top-right: Observation date and numbers of days after the peak (black)...... 106 3.11 Light-curve of PS1-10cdq (blue). Light curves of the comparison objects extremely luminous SLSN Type IIn/TDE candidate CSS100217 (Drake et al., 2011)(red), and TDE candidate PS16dtm (Blanchard et al., 2017)(green).... 109 3.12 Optical spectrum of PS1-10cdq taken with IMACS +84 days after the outburst after subtracting the stellar continuum (navy). Optical spectrum of TDE candidate PS16dtm +87 days after the outburst (Blanchard et al., 2017)(red). Optical spectrum of extremely luminous SLSN Type IIn/TDE candidate CSS100217 +84 days after the outburst (Drake et al., 2011)(purple). The emission lines are labeled (black)...... 110 3.13 light curves of PS1-10cdq in grizyPs and NUV GALAX bands (colored circles). Fits generated by MOSFiT’s CSM interaction model (colored lines)...... 113 3.14 Corner plot showing the posterior distributions of parameter realizations for MOSFiT’s CSM model. The 16th, 50th, and 84th quantiles of the posterior distribution of the model parameters (dash black lines). Median values with lower and upper error values of the model parameters are reported in the title of histograms (red)...... 114 14

3.15 Light curves of PS1-10cdq in grizyP1 and NUV Galex bands (colored circles). Fits generated by MOSFiT’s TDE model (colored lines)...... 118 3.16 Corner plot showing the posterior distributions of parameter realizations for MOSFiT’s CSM model. The 16th, 50th, and 84th quantiles the posterior distribution of the model parameters (dash black lines). Median values with lower and upper error values of the model parameters are reported in the title of histograms (red)...... 119

4.1 The architecture of the model used in this work. Top: represents architecture of the capsule layers where the GalaxCap layer has 2 or 37 capsules based on the different setups discussed in Section 4.4.4. Bottom: represents the structure of the decoder that acts as regularization during the training...... 132 4.2 Training and testing RMSE vs number of epochs for the regression scenario... 135 4.3 Training and testing accuracy vs number of epochs for classification based on the answers to question one...... 138 4.4 Original and reconstructed images of the galaxies...... 139 4.5 Estimated Sersic´ index of original images versus reconstructed images...... 140 4.6 Mean and 95% confidence interval of the difference between the estimated Sersic´ index for the original and reconstructed images (nOriginal − nReconstructed).. 141

A.1 A convolution on an 7×7 input image (green squares) with zero-padding P = 1 (white squares) with receptive field of size 3 × 3 (yellow squares) and a stride of S = 1 where the receptive field moved from left to right by one pixel..... 180 A.2 A max-pooling operation over a 3 × 3 convolutional matrix (blue squares) with receptive field of size 2 × 2 (red squares) and stride of 1...... 181 15 List of Acronyms In order of appearance.

AGN Active Galactic Nuclei

SMBH Super Massive Black Hole

BH Black Hole

UV ultraviolet

BLR Broad Line Region

NLR Narrow Line Region

GR General Theory of Relativity

Quasar Quasi Stellar Radio Source

QSO Quasar

RLQ Radio Loud Quasar

RQQ Radio Quiet Quasar

SSRQ Steep Spectrum Radio Quasars

FSRQ Flat Spectrum Radio Quasars

BLRG Broad-Line Radio Galaxies

NELG Narrow-Emission-Line X-ray Galaxy

NLRG Narrow-Line Radio Galaxies

FR Fanaroﬀ-Riley

BL Lac BL Lacertae

LINER Low Ionization Nuclear Emission-Line Region

ISCO Innermost Stable Circular Orbit

CL Changing Look

TDE Tidal Disruption Event

SN Supernovae

CSI Circumstellar Interaction 16

CCSN Core Collapse Supernova

ZAMS zero-age Main Sequence

CSM Circumstellar Material

SLSN Super Luminous Supernova

FS Forward Shock

RS Backward Shock

CDS Cool Dense Shell

IC Inverse Compton

MCMC Markov Chain Monte Carlo

WAIC Watanabe-Akaike Information Criterion

PSRF Potential Scale Reduction Factor

FFNN Feed Forward Nerual Network

CNN Convolutional Neural Network

ReLU Rectiﬁed Linear Function

CapsNet Capsule Network 17 List of Sorted Acronyms Sorted in alphabetical order.

AGN Active Galactic Nuclei

BH Black Hole

BL Lac BL Lacertae

BLR Broad Line Region

BLRG Broad-Line Radio Galaxies

CapsNet Capsule Network

CCSN Core Collapse Supernova

CDS Cool Dense Shell

CL Changing Look

CNN Convolutional Neural Network

CSI Circumstellar Interaction

CSM Circumstellar Material

FFNN Feed Forward Nerual Network

FR Fanaroﬀ-Riley

FS Forward Shock

FSRQ Flat Spectrum Radio Quasars

GR General Theory of Relativity

IC Inverse Compton

ISCO Innermost Stable Circular Orbit

LINER Low Ionization Nuclear Emission-Line Region

MCMC Markov Chain Monte Carlo

NELG Narrow-Emission-Line X-ray Galaxy

NLR Narrow Line Region

NLRG Narrow-Line Radio Galaxies 18

PSRF Potential Scale Reduction Factor

QSO Quasar

Quasar Quasi Stellar Radio Source

ReLU Rectiﬁed Linear Function

RLQ Radio Loud Quasar

RQQ Radio Quiet Quasar

RS Backward Shock

SLSN Super Luminous Supernova

SMBH Super Massive Black Hole

SN Supernovae

SSRQ Steep Spectrum Radio Quasars

TDE Tidal Disruption Event

UV ultraviolet

WAIC Watanabe-Akaike Information Criterion

ZAMS zero-age Main Sequence 19 1 Introduction

1.1 Active Galactic Nuclei

Active Galactic Nuclei are among the brightest objects in the Universe. Many galaxies have very bright central regions that can be more luminous than the light from the remainder of the galaxy. This bright central region leads to their naming as Active Galactic Nuclei (AGN). The origin of this emission is thought to be the accretion of the matter falling into a

6 9 Super Massive Black Hole (SMBH) with the mass of ≈ 10 − 10 M (Lynden-Bell, 1969). Historically, there were alternative suggestions for AGNs rather than SMBHs in the center of the galaxies. Some examples are multiple supernovae explosions caused by star formation bursts (Colgate and Cameron, 1963) or the presence of pulsars (Hoyle and Fowler, 1967). But shortly after that, it was shown that nuclear reactions would not be able to generate the observed energy and since all of the energy is coming from a very compact region (because of the observed rapid time variations), then accretion onto a SMBH at the center of the host galaxy would be a more suitable explanation (accretion is known to be one of the most eﬃcient processes)(Lynden-Bell, 1969). Nowadays, it is an accepted idea that most of the galaxies contain a central Black Hole (BH) (Magorrian et al., 1998; Ho, 2004). One of the strongest pieces of evidence supporting this scenario is our own Milky Way which contains a BH at center; the presence of this BH was conﬁrmed by tracing the orbital motions of the stars closest to Sgr. A* (a compact radio source which was a BH candidate) (e.g., Gillessen et al., 2009). As mentioned, accretion of the falling matter in SMBH is believed to be the origin of the emission from the center of the galaxies (Lynden-Bell, 1969). In the accretion process, matter emits ultraviolet (UV) and soft X-rays by losing angular momentum through viscous and turbulent processes. Hard X-rays are also produced very close to the BH because of the possible connection with a sea of hot electrons above the disk (Urry and Padovani, 20

1995). By realizing that the BH mass and the properties of the host galaxy have a strong correlation, researchers started linking BH mass growth with the formation of the host galaxy. Black hole growth that results from accretion and the energy feedback has affected most of the recent structure formation studies (Granato et al., 2004; Springel et al., 2005; Hopkins et al., 2006; Ho, 2008). All of the AGNs emit in a wide spectrum of wavelengths from γ-rays and X-rays to radio that are known to originate from different regions. Strong UV and optical emission lines are believed to be produced in the rapidly moving gas clouds in the gravitational potential of the BH known as Broad Line Region (BLR) clouds. UV and optical radiation experience obscuration by a torus, which is a warped disk of gas and dust particles outside of the accretion disk. Narrow emission lines, however, are produced in slowly moving gas clouds beyond the torus that are called the Narrow Line Region (NLR). Along the poles of the disk, the outflow of energetic articles generates jets that are the main source of radio emission. Outward high velocity streams of plasma in jets beams radiate relativistically in the forward direction (Urry and Padovani, 1995). Because of their high luminosity, AGNs are observable up to very high redshifts. For that reason they can provide very useful information about the structure of the Universe at early times and its evolution. Also, understanding more about characteristics of the AGNs such as BH mass, BH spin, the type of host galaxy and accretion rate will help us to find out more about the accretion process and interaction of the BH with its environment (Urry and Padovani, 1995). On the other hand, the SMBHs in their center can be a good probe to investigate the General Theory of Relativity (GR).

1.1.1 AGN Classiﬁcation

The axisymmetric model of AGNs suggests that there can be diﬀerent types of AGNs based on the angle toward the line of sight. Based on their optical spectra, luminosity, and 21 radio loudness, a brief summary of the classiﬁcation of AGNs is provided in the following paragraphs (Urry and Padovani, 1995).

• Type 1: This type of AGN is known to have bright continua and broad emission lines that originate from fast moving clouds close to the BH in the BLR region. They are subdivisions of Type 1 galaxies which are signiﬁcantly diﬀerent only in luminosity and radio loudness.

– Seyfert 1 galaxies: The main classiﬁcation of the Seyfert galaxies was done by Carl Seyfert (Seyfert, 1943) and they are divided to Seyfert 1 and Seyfert 2 sub- categories (Khachikian and Weedman, 1974). Seyfert 1 are radio-quiet Type 1 AGNs and they are seen nearby at low redshifts because of their low luminosity. Their spectra show both broad and narrow lines. As mentioned before, broad lines originate from high velocity, dense, and highly ionized gas clouds in the BLR which is close to the BH. In contrast, narrow lines are emitted from low- ionization, low velocity and less dense gas clouds in the NLR, which is beyond the torus and the accretion disk (Ho, 2008).

– Quasi Stellar Radio Source (Quasar): Quasars (QSOs) (Schmidt, 1963) are highly luminous and for that reason they can be observed up to high redshifts. Because of their high luminosity, only the star-like nucleus is observed, whereas a surrounding galaxy can be rarely observed. That is why they are called “Quasi Stellar Radio Sources”. QSOs can be both radio-loud (RLQ) and radio-quiet (RQQ). Based on the shape of their continuum, RLQs can be divided into a subgroup of Steep Spectrum Radio Quasars (SSRQ) or Flat Spectrum Radio Quasars (FSRQ) (Urry and Padovani, 1995) 22

– Broad-Line Radio Galaxies (BLRG): They are low luminosity Type 1 AGNs that have strong radio emission that probably comes from radio lobes which are powered by jets (Blandford and Rees, 1974; Urry and Padovani, 1995)

– Narrow Line Seyfert 1 (NLS1) galaxies: The most important characteristics

of NLS1s are the widths of the broad Balmer lines (specially, FWHMHβ < 2000

−1 km s ) and weakness of [O III]λ5007 emission relative to Hβ of [O III]/Hβtotl < 3 (Osterbrock and Pogge, 1985; Goodrich, 1989). Furthermore, they typically show strong Fe II features which anti-correlate with [O III] emission and broad Balmer emission lines. Also, the presence of Fe II features is suggested as an

extra measure for classiﬁcation of NLS1s (Fe II/Hβtotl < 0.5;V eron-Cetty´ et al., 2001).

• Type 2: This type of the AGNs has weak continua and narrow emission lines only. It is believed that the broad line region is obscured by a torus of absorbing gas and dust around it. Like the Type 1 AGNs, the Type 2s can be also divided into subdivisions based on their luminosity and radio loudness (see Figure 1 in (Urry and Padovani, 1995)). The high luminosity Type 2 AGNs have been studied using samples selected from the SDSS survey in optical, radio, and X-rays (Zakamska et al., 2003, 2004). Using spectropolarimetry, Zakamska et al.(2005) revealed the presence of a hidden Type 1 nuclei in these high luminosity Type 2 AGNs.

– Seyfert 2 galaxies: Seyfert 2 galaxies are low luminosity and radio-quiet Type 2 AGNs. Only narrow emission lines can be seen in their spectra. Seyfert 2 galaxies have [OIII]λ5007/Hβ ratio of > 3, which can be useful for their identiﬁcation (Shuder and Osterbrock, 1981). There are also other known low luminosity counterpart Type 2 AGNs that are called narrow-emission-line X- ray galaxies (NELGs) (Mushotzky, 1982). 23

– Narrow-Line Radio Galaxies (NLRGs): NLRGs are radio loud Type 2 AGNs and like BLRGs their radio emission comes from radio lobes which are powered by a jet of particles (Blandford and Rees, 1974; Urry and Padovani, 1995). The NLRGs are divided into different morphological categories based on the appearance of the radio jets (Fanaroff and Riley, 1974): The low luminosity Fanaroff-Riley Type 1 radio galaxies (FRI) are the first category, where the center is the brightest, the radio-jets are symmetric, and the intensity decreases by distance from the nucleus. The second category objects are high luminosity FRII radio galaxies, which have very well defined radio-lobes with hot spots and the brightness increasing by distance from the core. FRIIs are known to contain both NLRGs and BLRGs (Blandford and Rees, 1974; Urry and Padovani, 1995).

• Type 0: A small number of AGNs are called Type 0 AGNs because of their unusual characteristics (Urry and Padovani, 1995). These objects have close to zero degrees angles toward the line of sight.

– BL Lacertae (BL Lac): These objects are radio-loud AGNs with an unusual characteristic of not having strong emission or absorption features. Some other characteristics of them include surprisingly variable and high polarization, fast variability and high brightness temperatures (Urry and Padovani, 1995).

In summary, the AGNs are divided into the three main sub-classes of Type 1, Type 2 and Type 0, based on their optical spectra properties. Each of these types can have sub- classes based on radio-loudness and luminosity. As a last note, Low Ionization Nuclear Emission-Line Region galaxies (LINERs) will be discussed as a source of weak AGNs in the next section. 24

1.1.2 LINERs

Heckman(1980) studied nuclear activity of a class of objects that were di fferent from known AGNs and HII regions. These so called Low Ionization Nuclear Emission-Line Region galaxies, are known to have low-ionized narrow emission lines. The objects which had [OII]λ3727/[OIII]λ5007 ≥ 1 and [OI]λ6300/[OIII]λ5007 ≥ 1/3 were classified as LINERs (Filippenko, 1996). Numerous spectroscopic surveys concluded that LINERs exist in very large numbers and are more common than it was thought before (Heckman, 1980; Heckman et al., 1980; Filippenko, 1996). Moreover, there is a possibility that more than half of spiral galaxies contain LINERs (Ho et al., 1995). There is not too much activity observed in their nucleus and most of the emission comes from the stars in the galaxy. A number of other studies have shown that some of the LINERs contain broad Hα line similar to Seyfert 1 galaxies (Keel, 1983a,b; Filippenko and Sargent, 1985; Stauffer, 1982; Veron-Cetty´ and Veron´ , 1986; Filippenko, 1996). The LINERs are believed to have a continuum with non-stellar origin (Netzer, 1990). The main candidate for the excitation of the LINERs is photoionization by a central Low Luminosity Active Galactic Nuclei (LLAGN). Additionally, LINERs can be a radio (Wrobel and Heeschen, 1991) or an X-ray (Halpern and Steiner, 1983) emission source. ∼ 25% of the LINERs have UV emission from an unresolved core (Maoz et al., 1996) that can be the sign of accretion activity in their center. However, dust obscuration prevents UV detection in the other LINERs (Ho, 2008; Pogge et al., 2000). The LINERs can also fall into Type 1 and Type 2 categories, based on the observed obscuration. Understanding the nature of the LINERs is very important, since they make a large proportion of the nearby galaxies. Possible power sources for ionization of the LINERs are: shock heating created by the accretion process; winds generated by star-bursts or galaxy mergers and interactions; cooling gas flow and photoionization by evolved or main sequence very hot stars (Filippenko, 1996). 25

1.1.3 AGN Uniﬁcation

The simple model of AGN suggests that there is a SMBH with an accretion disk around it. The BLR is located in a greater distance with the fast moving gas clumps. There is a warped torus of gas which surrounds the BLR and the NLR is located at larger distances. As mentioned in previous section, AGNs come with a great number of different sub-classes, but the classification of them between Type 1 and Type 2 is based on obscuration of the nucleus and the classification of them based on radio-loudness depends on the angle of the jets with the line of sight (Urry and Padovani, 1995). The main scenario for the classification of the AGNs is the orientation to the line of sight. By observing NGC 1068, Antonucci and Miller(1985) realized that this classic Seyfert 2 galaxy has very broad lines similar to the Seyfert 1 broad Balmer lines that are observed in polarized light. Later discovery of the other similar Seyfert 2s with Seyfert 1 like broad Balmer lines (Miller and Goodrich, 1990) inspired the model that states that Seyfert 2 and Seyfert 1 galaxies are the same objects which are observed from different angles. This model proposes that there is a warped torus of the absorbing gas and dust particles surrounding the BLR in a way that the BLR is hidden from some orientations which is classified as a Type 2 AGN. Based on this model, emission from the BLR that is scattered by high energy electrons outside the nucleus that permits the observation of the broad lines in the polarized light (Antonucci and Miller, 1985; Antonucci, 1993). As a part of this unification scheme, it is important to discuss broad and narrow lines and their origins in more details. Broad lines are permitted lines with Doppler-broadened

3 4 −1 10 widths of 10 −10 km s that originate from the BLR. The BLR contains dense (nH ≥ 10 atoms cm−3), fast moving and highly ionized gas clouds. The observed continuum radiation with wavelength λ < 912Å in most Seyfert galaxies (these photons did not pass through the BLR because if they did, they would be absorbed by hydrogen gas in the galaxy and along the line of sight to us) and Doppler-broadened emission lines and highly ionized 26 gas, suggest that the BLR is a compact region located deep in the gravitational well of the central BH. In contrast, narrow emission lines, such as [OIII]λ5007, are coming from the transitions which are mainly forbidden with widths of 102−103 km s−1 that are emitted from

8 the NLR. The forbidden lines are observed only where the density of the gas is nH ≤ 10 , which suggests that the NLR should contain low density gas clouds. These forbidden lines have not shown any variations in short timescales, demonstrating that the NLR is located far from the central BH (Sparke and Gallagher III, 2007; Bennert et al., 2002, chapter 9). However, in one of the recent studies Peterson et al. (2013) show that the fluxes of the narrow [OIII]λλ4959, 5007 emission lines in the spectra of NGC 5548 (a well-known Seyfert 1 galaxy) vary in the timescale of decades. Therefore, from this variability they show that the NLR is more compact and denser than previously known (Peterson et al., 2013). Although it is not complete, the simple unification model mentioned above that is only based on orientation effects has been the base model for the unification of Type 1 and Type 2 AGNs. However, multiple AGN types in some objects have been recently reported in different epochs, which challenges this model. These interesting objects and the possible explanations for this type of changes will be discussed in section 1.2.

1.1.4 Accretion Process and Variability in AGNs

The material that is far away from the SMBH in the central regions of the galaxies loses angular momentum while flowing inward and toward the central engine. The angular momentum is lost through the viscous processes and is converted to heat and the magnetic field. This rotationally supported flow forms what is known as an accretion disk (Koratkar and Blaes, 1999). Accretion disks are among the most efficient energy providers from a compact volume, which makes them one of the main components of the AGN models and a possible source of the power in them (Salpeter, 1964; Zeldovich and Novikov, 1964). 27

However, there is a limit to their luminosity which is known as Eddington luminosity, which happens when the radiative force is equal to gravitational force (Eddington, 1925)

4πGMmpc 38 −1 M Ledd = ≈ 1.26 × 10 ergs (1.1) σT M

where mp is the mass of the electron and σT is the Thompson scattering cross section. The accretion flow of the matter in the accretion disk is believed to be the source of X-ray, UV and optical emission which powers AGN by ionizing the surrounding gas in both the BLR and NLR and may also be the birth place of the jets and winds (Koratkar and Blaes, 1999). Observation of the central engine of AGN, which contains the accretion flow, is hard and it is difficult to resolve the small volume at the galaxy centers where it is located. However, one strong piece of observational evidence that shows that there is gas located deep in the potential well of the SMBH in AGNs is the detection of the very broad Fe Kα line (Tanaka et al., 1995; Mushotzky et al., 1995; Nandra et al., 1997). The simplest model of the accretion flow is that the plasma orbits the central BH in “Keplerian” orbits and the flow of gas is geometrically thin and optically thick. The vertical structure of the disk; however, results from balancing the gas pressure and the gravitational force provided by the BH (Koratkar and Blaes, 1999). The standard theoretical models proposed for this geometrically thin disk structure, (Shakura and Sunyaev, 1973; Novikov and Thorne, 1973; Koratkar and Blaes, 1999) make a couple of important assumptions: (1) the disk is stationary and axisymmetric and flows all the way to the Innermost Stable Circular Orbit (ISCO) where the shear stress is assumed to be zero; (2) the gravitational field of the hole is assumed to be Newtonian and matter orbits around the hole with Keplerian orbits; (3) the mass of the accretion disk is negligibly small compared to the mass of the BH; (4) the viscosity makes some material fall inward and some move outward as it orbits because of conservation of angular momentum. The lost gravitational energy during this process is converted to radiation due to conservation 28 of energy; (5) the disk is thin, so heat transfers vertically and at the same radius; therefore, the emission is a local blackbody (Pereyra et al., 2006; Koratkar and Blaes, 1999). Despite the fact that the above model is the base model used for accretion flow, there exist a number of limitations for it. The first one is that the model does not work for high accretion luminosities where the radiation pressure plays an important role. Also, the SED

1/3 of Fν ∝ ν resulting from this model is bluer than the observed UV and optical spectra of the AGNs. Furthermore, a local blackbody is not a good approximation even with an optically thick disk, which may not be always the case (Koratkar and Blaes, 1999). AGN luminosities have been observed to vary over time. Variability has been detected in a wide range of wavelengths with timescales from hours to years with the shortest in X-rays (Koratkar and Blaes, 1999). X-ray observations with variation time-scales of a few hours are observed in Seyfert galaxies (Green et al., 1993; Nandra et al., 1997) and the observed average variability time-scale in Seyfert galaxies is on the order of weeks to months (Kaspi et al., 1996; Giveon et al., 1999). A number of models have been proposed for the source of this variability, such as supernovae (Aretxaga et al., 1997) and microlensing (Hawkins, 2000), but accretion disk instability is the most promising one (Kawaguchi et al., 1998; Kelly et al., 2009). Reverberation mapping studies (Blandford and McKee, 1982) have shown that the broad emission lines respond to the variations with time lag which is strong evidence that the continuum variations are derived from accretion instabilities (Peterson et al., 2004; Kelly et al., 2009). Optical and UV emission are believed to be thermal emission originating from the accretion disk (Shields, 1978) and if this scenario is valid then thermal ﬂuctuations would be the source of the variations (Kelly et al., 2009). One observational piece of evidence in favor of this scenario is that quasars have been observed to be bluer when brighter (Giveon et al., 1999; Trevese et al., 2001; Geha et al., 2003). 29

The model developed by Shakura and Sunyaev (1973) is also known as a standard α model. The key assumption in their model is that the viscosity is the source of thermal emission and angular momentum lost in accretion disks, which is proportional to the total pressure of the disk. The viscosity is

ν = αcsh (1.2)

where h is gas scale height and cs is speed of sound. A number of studies on quasar variability estimated the value of α to be α ≈ 0.001 (Siemiginowska and Czerny, 1989; Collier and Peterson, 2001; Starling et al., 2004; Kelly et al., 2009). Three important timescales in accretion disks are :

tdyn = 1/Ω (1.3)

−2 −2 −1 H H −1H tvisc ≈ α ≈ α tdyn (1.4) cs R R −1 tth = α tdyn (1.5) q GM where Ω = R3 is angular velocity, tdyn is dynamical timescale, tvisc is the viscous H timescale and tth is the thermal timescale and since in geometrically thin disks ( R ) 1 then the ordering of the timescales is tvisc tth > tdyn (Lasota, 2016). In radiation-dominated disks, which are also believed to be the characteristic of the inner disk environment, the α-disk is unstable both thermally and viscously (Shakura and Sunyaev, 1976; Lightman and Eardley, 1989). This thermal instability is believed to grow exponentially on thermal timescales which are on the order of months to years for AGNs and instability timescales in optical light curves of AGNs are not usually detected to be ≥ tth (Kelly et al., 2009), except some specific sources (Czerny et al., 2003; Lub and de Ruiter, 1992). In an optically-thick accretion disk, different wavelength ranges of the spectrum come from different regions of the flow that are spatially separated. Thus, time lags in variability can happen because of the finite speed of the propagating signal between different regions (Koratkar and Blaes, 1999). A possible explanation would be that optical 30 and UV variations are caused by reprocessing of X-ray or UV radiation from the inner regions of the disk (Krolik et al., 1991; Collin-Souffrin, 1991). This scenario has been proven to be correct by multiwavelength observations of the Seyfert galaxies NGC 4151 and NGC 5548 that showed time lags between X-ray and UV/optical which suggest that the UV/optical variability is driven by reprocessing of X-rays (Edelson et al., 1996; Clavel et al., 1992; McHardy et al., 2014).

1.2 Changing Look AGNs

As mentioned in section 1.1, there is a SMBH at the center of most galaxies that grows by accreting matter. AGNs are identified by the observed properties of their optical spectra, broad and narrow lines in Type 1s and just narrow lines in Type 2s. Also, intermediate AGNs such as Type 1.8 and 1.9, are classified by weak or absent broad Hβ while showing broad Hα in their spectra (Osterbrock, 1981). As discussed in section 1.1.3, the standard unification model suggests that the Type 1 and the Type 2 AGNs are the result of different viewing angles to the line of sight of an obscuring torus of gas and dust particles surrounding the BLR and the intermediate type AGNs are suggested to be the result of partial obscuration or reddening by optically thin dust (Stern and Laor, 2012). It is believed that the simple orientation-based model cannot explain the Type 1 and Type 2 classification scheme to the full extent. This model is challenged when different AGN Types are observed in the same object at different epochs of time. One alternative suggestion is that Type 2s were Type 1s and their engine is now not active anymore or basically turned off; therefore, the Type 2s are evolved version of the Type 1s (Penston and Perez´ , 1984; Runnoe et al., 2016). This alternative view was first suggested based on the observations of broad emission lines disappearance in NGC 4151 (Lyutyj et al., 1984; Penston and Perez´ , 1984) and 3C 390.3 (Penston and Perez´ , 1984). 31

The objects that change type at different epochs of time can help us to advance our understanding of the evolutionary side of unification model. Although rare, these objects have been observed in the past. The term “changing-look” was first used for the AGNs that showed X-ray absorption variations, which in simple terms means that they were observed to appear both “Compton-thin” or “reflection-dominated” at different epochs of time in timescale of years (Matt et al., 2003; Bianchi et al., 2009; Puccetti et al., 2007; Risaliti et al., 2009; Marchese et al., 2012). Additionally, the term “changing-look” (CL) AGNs has been also used to describe the objects that have been observed to optically transition from Type 1 to Type 1.8, 1.9 and 2 or vice versa, by sudden appearance or disappearance of broad emission lines such as the broad Hβ emission line (Runnoe et al., 2016). Such objects have been detected in the past four decades in a number of the AGNs. In some objects the broad emission lines disappeared and the continuum faded completely (Collin-Souffrin et al., 1973; Tohline and Osterbrock, 1976; Sanmartim et al., 2014; Denney et al., 2014; Barth et al., 2015; Runnoe et al., 2016), while in others broad emission lines appeared or, in other words, the AGN “turned-on” (Cohen et al., 1986; Storchi-Bergmann et al., 1993; Aretxaga et al., 1999; Eracleous and Halpern, 2001; Shappee et al., 2014). These objects are at low redshift with low luminosities. The “changing-look” behavior that is observed in some of the AGNs can possibly be explained by three suggested mechanisms. The first mechanism is by variation of obscuration where the obscuring material such as dust clouds moves in or out of the line of sight and obscures or gives a clear view to the BLR (Elitzur, 2012). Another possible explanation can be a change in the accretion rate. The change in the accretion rate transforms the structure of the BLR and the objects will transition from the high accretion rate and luminous Type 1 to a low accretion state Type 2 (Elitzur et al., 2014), because multiple models have been suggested that a radiatively efficient BLR cannot exist in low accretion rates. Some of the possible explanations are: the lack of sufficient ionizing 32 photons to power the BLR (Korista and Goad, 2004), the size of critical radius (at this radius disk changes from gas pressure dominated to radiation pressure dominated) becoming smaller than ISCO (Nicastro, 2000; Nicastro et al., 2003), a minimum required bolometric luminosity where the BLR cannot exist if it goes lower (Laor, 2003), and other reasons (Denney et al., 2014; Elitzur et al., 2014; Trump et al., 2011). It has also been argued that transient events, such as Tidal Disruption Events (TDEs), can cause type change (Eracleous et al., 1995). There are observational results that suggest that each of the three scenarios mentioned above is a possible physical process behind the CL behavior in different objects. For example, Alexander et al.(2013) observed a source at redshift z = 0.510 and argued that obscuring material moved into the line of the sight and obscured the nucleus of the object that changed the class of the source from a Type 1 to a Type 2. In another observation, Shappee et al.(2013) reported an interesting transition from Sy 1.8 to Sy 1 in NGC 2617 and they argued that X-ray irradiation of the disk, likely resulting from coronal activity, is causing this variability because the variability was first detected in X-rays and then with time lags in UV and NIR (Shappee et al., 2014). LaMassa et al.(2015) also favor changes in the accretion power as an explanation for the observed transition from a Type 1 quasar to a Type 1.9 AGN in J0159+0033 because they estimated the crossing time for an possible obscuring object outside the BLR with a Keplerian orbit to be tcross = 20 years that is too long to explain this transition which happened in shorter timescales. They are also skeptical about the existence of obscuring material with the required physical properties to obscure the BLR and whole continuum at such radii. However, Merloni et al.(2015) proposed an alternative scenario that the nature of flare in J0159+0033 is a tidal disruption event (TDE; will be discussed in more detail in section 1.3) by interpreting that this object was first in a low accretion state and suddenly showed an outburst and brightened dramatically. 33

Additionally, they showed that the very rapid rise and decay time of the ﬂare is consistent with the t−5/3 behavior expected in TDEs. The number of objects that show CL behavior is increasing and there already have been a number of bulk searches in the SDSS and Pan-STARRS surveys that found new CL candidates (Ruan et al., 2016; MacLeod et al., 2016). As we can see, the reason behind CL behavior is not very well known yet and is still under debate. Thus, understanding the nature of these objects will help us to further study the structure of the accretion disk and the accretion process of BHs and consequently shed more light on the structure of the BLR. Furthermore, it will give us an opportunity to investigate the uniﬁcation models of AGNs in more detail.

1.3 Tidal Disruption Events

When a star orbits a BH with a suﬃciently close pericenter and the BH’s gravitational force exceeds the self-gravity of the star, the star is disrupted (Hills, 1975). This happens at a distance which is known as tidal radius:

M 1/3 M −1/3 r r ≈ × 12 BH ∗ ∗ cm t 7 10 6 (1.6) 10 M M r

In a Tidal Disruption Event (TDE), debris from the disrupted star is divided into almost two equal parts. One of the parts becomes unbound on hyperbolic orbits while the other part is bound to the hole with elliptical trajectories and a pericenter the same as original star’s pericenter. Eventually, bound debris will be accreted to the BH and produce a luminous transient event (Rees, 1988; Komossa, 2015). However, the tidal radius scales diﬀerently

2GMBH with the mass of the BH (as seen in eq.1.6) than the Schwarzschild radius (rs = c2 ), so solar type stars will pass the horizon without getting disrupted for SMBHs with masses

8 higher than ∼ 10 M . This limit is higher when spin is included (Beloborodov et al., 1992). 34

7 In TDEs involving a solar-type star and SMBHs with the mass of ≤ 10 M which accretes with Eddington limit, the spectrum is a blackbody with the characteristic blackbody temperatures of ∼ 2.5 × 105 K at the tidal radius (Ulmer, 1999; Strubbe and Quataert, 2009). The radiation from these events peaks in the UV and soft X-rays, declining on the timescale of months to years. One fundamental puzzle is that the UV/optical flux predicted by a standard BH accretion theory is orders of magnitude less than the observed values and has a blue constant color (Roth et al., 2016). A possible solution to this problem is the presence of gas which absorbs and re-emits the radiation with temperatures around ∼ 20000 − 30000 K. The source of this gas is thought to be the formation of an envelope around the BH (Loeb and Ulmer, 1997; Guillochon et al., 2014; Coughlin and Begelman, 2014); or supper-Eddington mass outflow (Strubbe and Quataert, 2009; Lodato and Rossi, 2010; Stone and Metzger, 2015; Vinko´ et al., 2014; Miller, 2015) or the circularization of the material at much larger distances than the tidal radius of the BH (Shiokawa et al., 2015; Piran et al., 2015; Hayasaki et al., 2016; Bonnerot et al., 2015; Guillochon and Ramirez- Ruiz, 2013, 2015; Dai et al., 2015). The light curves of TDEs are predicted to follow a rapid rise to a maximum peak and then decline by the fall back rate of t−5/3 (Rees, 1988; Evans and Kochanek, 1989). This fall-back rate was derived by assuming that the the spread of the specific energy for debris is constant with mass. However, recent detailed modeling has shown that the internal structure of the disrupted star can affect this assumption and predicts a faster decline rate than of t−5/3 for the light-curve of the TDEs after the peak (Ramirez-Ruiz and Rosswog, 2009; Lodato et al., 2008; Stone et al., 2013; Guillochon and Ramirez-Ruiz, 2013). Furthermore, the process which turns accreted mass into observable radiation is not simple to simulate. The simplest model to simulate the Spectral Energy Distribution (SED) of TDEs and their evolution is the thin disk model but the new and more complex models 35 have added a thick disk and a wind or outflow to represent super-Eddington accretion (Guillochon and Ramirez-Ruiz, 2013). The number of observed TDEs is rapidly increasing and the rate of these events are estimated to be ∼ 10−4 to 10−5 events per year per galaxy (e.g., Magorrian and Tremaine, 1999; Donley et al., 2002; Wang and Merritt, 2004; Kesden, 2012; Stone and Metzger, 2015; Van Velzen and Farrar, 2014). A large portion of TDE candidates are soft X-ray flares with large amplitudes from the galactic centers (Komossa and Bade, 1999; Donley et al., 2002; Komossa et al., 2004; Halpern et al., 2004; Esquej et al., 2007; Cappelluti et al., 2009a; Maksym et al., 2010; Saxton et al., 2012; Hryniewicz and Walter, 2016; Lin et al., 2015; Komossa, 2015). In these events, the luminosity peaks at 1044 erg s−1 and the SED peaks at soft X-rays with energies ≤ 0.1 kev. Additionally, the decay rate of the light curve is consistent with the predicted fall-back rate of the debris t−5/3 (Rees, 1988; Evans and Kochanek, 1989; Lodato et al., 2008; Guillochon and Ramirez-Ruiz, 2013; Roth et al., 2016). Van Velzen et al.(2011) reported the first optical TDEs (SDSS TDE1, SDSS TDE2) from multi-epoch imaging of Stripe 82 in SDSS. They selected these candidates because of their unusually blue colors with slow evolution which was very different from supernovae (SNe) and AGN variability. Gezari et al.(2012) reported the TDE PS1-10jh from PAN- STARRS1 (PS1) with optical light curves from the time of the rise and decline from the peak which became the subject of studies for the process that derives TDEs (Guillochon and Ramirez-Ruiz, 2013; Bogdanovic´ et al., 2014). There have been TDE candidates discovered in ultraviolet (UV) (Gezari et al., 2006, 2009) and other optical TDE candidates reported from PAN-STARRS1 (Chornock et al., 2013), ASASSN (Holoien et al., 2014, 2015), PTF (Cenko et al., 2012a; Arcavi et al., 2014) and ROTSE (Vinko´ et al., 2014). The light curves of these TDEs rise and peak at R-band luminosity 2 × 1043 erg s−1 and decline consistent with t−5/3(Arcavi et al., 2014). Additionally, TDEs like PTF10iya and 36

ASASSN-14li are observed both in optical and X-ray (Holoien et al., 2015; Miller et al., 2015). Swift J164449.3+573451 (Sw 1644+57; Bloom et al., 2011; Levan et al., 2011; Burrows et al., 2011; Zauderer et al., 2011), Swift J2058.4+0516 (Sw 2058+05; Cenko et al., 2012a,b) and Swift J1112.2-8238 (Brown et al., 2015) are three long lived (∼ 107 s) relativistic TDE candidates that first discovered by γ-ray triggers and then observed in X-rays. The decline rate of the X-ray light curves in Sw 1644+57 and Sw 2058+05 are approximately consistent with t−5/3. All of these three candidates are believed to have launched relativistic jets (Giannios and Metzger, 2011; Brown et al., 2015) which are very rare in TDEs and have not been observed in any other candidates yet. Recently, broad emission lines have been observed in some TDE candidates. SDSS TDE2 showed a broad Hα line (Van Velzen et al., 2011) and PS1-10jh exhibited broad He II lines with H-lines lacking (Gezari et al., 2012). These were the first objects to show spectral features among TDE candidates. This low ratio of H:He lines has been the matter of debate in recent studies. A number of explanations have been proposed, including the disruption of H-poor star with He-rich core (Gezari et al., 2012; Bogdanovic´ et al., 2014; Strubbe and Murray, 2015); the effects of the optical depth (Gaskell and Rojas Lobos, 2013); the presence of a H-rich BLR which is truncated (Guillochon et al., 2014); the presence of a H-rich envelope surrounding the accretion disk which absorbs and re-emits the radiation

from the central parts (Roth et al., 2016); H-burning in the stellar core of a M ≥ M star (Kochanek, 2016). Arcavi et al (2014) discovered new TDEs with wide range of H:He ratio whereas other candidates show no sign of emission lines (e.g., Chornock et al., 2013; Cenko et al., 2012a). Transients that are detected in the galactic nucleus can be related to the AGN activity. As mentioned in section 1.1.4 AGNs are known to show variability in their continuum and emission lines (Peterson, 2001) that is believed to be because of change in accretion rate 37 related to disk instability or thermal ﬂuctuations (Kawaguchi et al., 1998; Pereyra et al., 2006; Kelly et al., 2009). Also, as discussed in section 1.2, there are newly discovered class of objects called CL AGNs that show appearing or disappearing broad emission lines that is usually followed by almost an order magnitude change (increase or decrease) in the continuum brightness (Storchi-Bergmann et al., 1993; Shappee et al., 2014; LaMassa et al., 2015; MacLeod et al., 2016; Gezari et al., 2017). The timescales of these changes are on the order of years. However, recently iPTF16bco exhibited a change from a LINER to a broad line quasar in a shorter timescale of . 1 yr and is thought to be due to intrinsic accretion changes (Gezari et al., 2017). These recently discovered class of objects can be interpreted as TDEs or vice versa. For example, Blanchard et al.(2017) favor a TDE as the source of the transient PS16dtm over AGN variability. They argue that the rise time for PS16dtm is 50 days, which is much shorter than viscous timescale (∼ 100 years). Also, rapid two order of magnitude change in UV/Optical ﬂares in not typically seen in AGNs, since the order of magnitude variability in CL AGNs happens in timescale of years (Grupe et al., 2010; Drake et al., 2011). On the other hand, they argue that the dimming in the X-rays cannot be explained by CL AGN scenario, since as observed in NG 2617 and iPTF16bco, as a result of AGN accretion stage changes an increase in both X-ray and UV/optical is expected to be seen. The upcoming sky surveys will discover hundreds or thousands of TDEs. These surveys include SKA in radio (Donnarumma et al., 2015), Large Synoptic Survey Telescope (LSST) in optical (Van Velzen et al., 2011), LOFT in hard X-ray (Rossi et al., 2015) and possibly China’s Einstein Probe in soft X-ray (Yuan et al., 2014). These detections will help us to better understand the nature of TDEs and study SMBHs in the galaxies that are rather quiescent. 38

1.4 Supernovae

A supernova (SN) is the explosion of a star at the end if its life when the equilibrium between the gravitational pull from the mass of the star and radiation pressure caused by fusion is lost. SNe are classified to two major classes: Type I or hydrogen poor and Type II or Hydrogen rich based on the presence of the hydrogen features in the observed spectra of these events. Each of these general classes are again classified to different sub-classes based on the spectral features in the case of Type Is and light curve decay behaviour in the case of Type IIs. Type I is classified to Type Ia with strong silicon features, Type Ib with strong helium features, and Type Ic with no helium or silicon detected in their spectra. While, Type II is classified to Type IIL with linear decay, and Type IIP for a plateau in the light-curve after the peak of the outburst. Type II contains another subclass called Type IIn where strong narrow hydrogen emission and absorption lines are present in the spectra. Most of the Type II SNe are powered by circumstellar interaction (CSI) or hydrogen recombination while Type Is are powered by radioactivity of 56 Ni. Furthermore, majority of these events are caused by core collapse supernovae (CCSNe) as a part of the evolution of a star with zero-age main sequence (ZAMS) mass of > 8M where a compact object such as s neutron star or black hole is produced. On the other hand, Type Ia is the only exception that is caused by thermonuclear explosion of a white dwarf that gains mass and passes

Chandrasekhar limit (∼ 1.4M ) through binary interaction and accretion and explodes. Studying SNe can help us to better understand stellar evolution, properties of the circumstellar material (CSM), and the environment of the galaxy that hosts them.

1.4.1 Superluminous Supernovae

A number of outbursts have been observed that were 10-100 times brighter than classical SNe. Because of these intense brightness diﬀerence they are called superluminous supernovae (SLSNe). Their luminosity at the peak of the outburst is ∼ 1044 − 1045 erg s−1 39 that makes them detectable at redshifts z > 4 (Berger et al., 2012; Cooke et al., 2012; Quimby et al., 2011). These rare outbursts are observed in faint host galaxies (Angus et al., 2016; Lunnan et al., 2014; Schulze et al., 2018) with low metallicity and with the rate of 10−4 − 10−5 of the CCSN rate (McCrum et al., 2015). The classiﬁcation scheme of SLSNe is similar to the SNe. Type I or hydrogen poor SLSNe are very similar to SN Ic with strong oxygen lines in their spectra (e.g., SN 2015bn; Jerkstrand et al., 2017; Nicholl et al., 2016b,a; Bhirombhakdi et al., 2018). Type II or hydrogen rich SLSNe typically show narrow emission and absorption with prominent Hα emission lines similar to SNe IIn that are similarly called SLSNe IIn (e.g., SN 2006gy; Quimby, 2006). However, in some rare cases the narrow lines are not detected that are called SLSNe II (e.g., SN 2008es; Gezari et al., 2008; Miller et al., 2008; Inserra et al., 2018; Bhirombhakdi et al., 2019).

1.4.2 Sources that Power Superluminous Supernovae

Based on their types SLSNe can be powered by diﬀerent sources. The SLSNe IIn are thought to be powered by CSI very similar to SNe IIn while the power sources of the SLSNe I and II are not well known and under debate. Radioactive 56Ni, and magnetar spindown are other possible sources for both SLSNe I and II. In this section we will mainly discuss CSI as the power source of the SLSNe IIn, since we discuss an object detected by PS1 survey in chapter3 where SLSN IIn can be one of the possible interpretations for this outburst.

1.4.2.1 Circumstellar Interaction

The Type II SNe are caused by core-collapse of massive stars and come in diﬀerent luminosity ranges. The more luminous ones are typically being interpreted as Type IIn SNe because of the presence of the relatively narrow Balmer emission lines that are believed to be originated from the strong interaction of the shocks with the CSM (Schlegel, 1990). 40

These shocks are generated from the interaction of the accelerated ejecta (with velocities of ∼ 104 km s−1) with CSM and consist of a forward (FS) and backward component (RS). The emission lines in the spectra of SNe IIn typically include multiple components: a narrow component with a width of ∼ 102 km s−1, an intermediate component with a width of ∼ 103 km s−1, and a broad component with a width of 104 km s−1. The narrow components of the Balmer emission lines are believed to be originated from the pre-shock CSM that was shed by a progenitor star while intermediate components and broad components of the emission lines are originated from the post-shock dense gas known as cool dense shell (CDS) and ejecta, respectively (Chugai and Danziger, 1994). The evolution of a limited number of SLSNe IIn candidates has been studied from the early time of the explosion to years after. They show blue spectra with few emission lines with simple profiles at early times. At the same time, later time spectra tend to have more complex continuum shape and emission line profiles and after several hundred days the spectra show very little continuum dominated by residual Hα emission lines (Smith et al., 2008a,b, 2010). When the CSM is sufficiently dense, it decelerates the shock waves thus converting their bulk kinetic energy to X-rays and then optical radiation that significantly increases the bolometric luminosity of the system (Fransson, 1984). The presence of such dense CSM suggests a more massive progenitor or a progenitor with a slow wind (e.g., Chugai et al., 2004; Smith et al., 2008a). Another source of radiation in the CSI is from synchrotron and inverse Compton (IC) caused by electrons that are accelerated by shocks to relativistic speeds. The radiation generated by IC is more significant at early times since the density of high energy photons is higher (Margutti et al., 2018). The radiation caused by upscattering of optical photons in the SN photosphere is in UV, X-rays and low energy γ rays. The peak of synchrotron radiation can range from X-rays to radio which depends on the time of generation after the explosion. It is notable that the synchrotron radiation is mostly self- 41 absorbed at early times and is observed to peak in radio only at late times (Pooley et al., 2002). The CSI has been studied through analytical models (Chatzopoulos et al., 2012; Chevalier, 1982; Chevalier and Fransson, 1994, 2017; Chevalier and Liang, 1989; Fransson, 1984; Nadyozhin, 1985) where the CSI is typically represented by self-similar solutions with a power-law density profile; however, these models’ assumption of symmetric CSM structure is questionable since there is evidence that in some cases CSM is in the shape of a disk or a torus with clumpy distribution (Andrews et al., 2010, 2011, 2016; Chugai et al., 1995; Fransson et al., 2002; Smith et al., 2009). Also, most of the parameters (e.g, energy, mass, velocity, and density profiles of the ejecta) in these models are uncertain and model sensitive (Chatzopoulos et al., 2012). The CSI is more accurately simulated through 3D hydrodynamic simulations (Balberg et al., 2000; Bersten et al., 2011; Blinnikov et al., 1998; Chieffi et al., 2003; Dessart and Hillier, 2010; Kasen and Woosley, 2009; Pumo and Zampieri, 2011; Utrobin, 2007; Woosley and Weaver, 1995; Zampieri et al., 1998).

1.5 The Modular Open Source Fitter for Transients

The Modular Open Source Fitter for Transients (MOSFiT) is a Python 2.7/3.x package that is installed in the Anaconda environment. This package is used for ﬁtting, estimating, and sharing the parameters of transients such as but not limited to SLSNe, SN Ia and Ic, TDEs, and Kilonovas (Guillochon et al., 2018). It is designed to work with datasets with the data format of Astrocats: Open Astronomy Catalogs 1. The data can be downloaded directly by running MOSFiT on the name of a speciﬁc object. In order to use private data, the data should be transformed to a JSON 2 format that matches the data format of Astrocats. 1 https://github.com/astrocatalogs/astrocats 2 https://www.json.org/ 42

The MOSFiT package gives a lot of flexibility to the users. The package runs through the command line and users can specify the transient model, model parameters, and extra outputs such as luminosity and temperature. After running the package, each of the modules take the inputs and process them to return an output such as extinction correction, and SED that is going to be used by other modules. The MOSFiT contains an ensemble of Markov Chain Monte Carlo (MCMC) walkers that follow a Bayesian approach. The objective function is a Gaussian Process Likelihood where the Watanabe- Akaike information criterion (WAIC) is being applied to calculate this objective from the ensemble. The model converges when the potential scale reduction factor (PSRF) goes under a desired value. The standard and extra outputs of the model are then saved in JSON format and can be visualized through a provided Jupyter Notebook. The MOSFiT tries to fit each band pass separately. However, MOSFiT heavily relies on the SED to calculate fluxes through luminosity at each epoch in each band-pass because all of the models are represented as bolometric light curves. In all of the models a single blackbody SED is being used except the model for SLSNe where a modified blackbody SED is being applied. We use MOSFiT to fit for TDE and SLSN IIn scenarios in Chapter3.

1.6 Galaxy Morphology Prediction

Galaxies have a variety of shapes, colors, and sizes that are used to indicate their evolution and interactions during the course of cosmic history. Galaxy morphology studies can help us to better understand the evolution of galaxies with redshift and provide us information about the dynamical history of galaxies without the need of expensive spectroscopy. In order to carry out these studies, all sky surveys and an accurate method for morphological classification is crucial. Recently, large sky surveys such as SDSS made a large amount of data for objects in our universe available. Traditionally, the morphological 43 classification of the galaxies has been done by visual inspection by experts that is inefficient and impractical for the large data available from current surveys and even larger upcoming datasets from surveys such as LSST (Ivezic et al., 2008). Different automated methods have been proposed to solve this issue; however, historically these methods did not reach required reliability level for scientific studies (Clery, 2011). The Galaxy Zoo project was launched to help to accelerate finding a reliable classification method by introducing a crowd-sourcing method (Lintott et al., 2008). During the initial Galaxy Zoo project ∼ 900, 000 galaxies were annotated by online participants in the short time span of months. Since the successful initial project, there have been other iterations of the Galaxy Zoo projects with more fine grained classification scheme (e.g., Galaxy Zoo 2; Willett et al., 2013). For example, during the Galaxy Zoo 2 project participants were asked questions from a decision tree where the answer to each question would determine the next question to be asked from this decision tree (Figure 1; Willett et al., 2013). Then, the answers of the participants on the same image were transformed to vote fractions which were used to determine the confidence levels of the answers. The data from Galaxy Zoo projects have already been used for studies regarding galaxy structure, formation, and evolution (Skibba et al., 2009; Bamford et al., 2009; Schawinski et al., 2009; Lintott et al., 2009; Darg et al., 2010; Masters et al., 2010, 2011; Simmons et al., 2014; Melvin et al., 2014; Willett et al., 2015). Recent improvements in computer vision through deep neural networks (e.g., Krizhevsky et al., 2012), available large annotated datasets through Galaxy Zoo projects, and increased computing power through the GPUs have made the possibility of providing a more accurate automated approach more promising. Different machine learning methods such as Principal Component Analysis (PCA) (Naim et al., 1995; Lahav et al., 1995; De La Calleja and Fuentes, 2004), Support Vector Machines (SVM) (Tasca et al., 2009), feed forward neural network (Storrie-Lombardi et al., 1992), and convolutional neural 44 networks (Gravet et al., 2015; Dieleman et al., 2015; Dom´ınguez Sanchez´ et al., 2018) have been used for galaxy morphological predictions. In the the following section we will provide an overview of deep neural networks.

1.6.1 Deep Learning

Deep neural networks try to build models to extract hierarchical and abstract features and discover accurate symmetries directly from the data. These networks consist of multiple layers where each layer extracts more abstract features based on the information available from lower layers that represent the input data. These abstract features are typically calculated by a non-linear transformation function. Next, the parameters of this multi-layer structure are optimized by training the model over a large dataset. Although neural networks have been discovered decades ago (McCulloch and Pitts, 1943; Fukushima, 1980), only recent the advancements in GPU computing and advances in the underlying methods enabled training them on large-scale datasets.

1.6.1.1 Feed-Forward Neural Networks

One simple example of deep neural networks is a structure called Feed Forward Neural Network (FFFNN) that consists of three kinds of layers: input layer, hidden layer, and output layer (a representation of FFFNN can be found in Figure 1.1). Each layer is connected to the next layer via a set of initialized weight parameters. The input layer is just the layer that feeds the data to the network. Next, the network has multiple layers of what are called hidden layers where each of these layers is made of a number of neurons that compute a weighted summation of the incoming inputs followed by a non- linear transformation. Let al be the input of the layer l + 1, Wl to be matrix of weights that connect layer l to layer l + 1, and bl to be a vector of biases. Therefore, the output of layer l + 1 (al+1 ) can be 45 represented as, sl l+1 l+1 l+1 X l l l ai = f (Zi ), Zi = Wi ja j + bi (1.7) j=1

where i and j represent number of neurons in layer l and l + 1 respectively, sl represents number of layers being used in FFNN, and f is the activation non-linear function. The most common choices for activation function are Rectiﬁed Linear Unit (ReLU, f (x) = max(x, 0);

1 Nair and Hinton, 2010), sigmoidal function ( f (x) = 1+e−x ), and hyperbolic tangent ( f (x) = tanh(x)). During the training process, the objective is to optimize parameters of the network in a way that the output generated by network (anl ) closely resembles the true values of the desired target output (y). This prediction error is quantiﬁed by a cost function J(anl , y). Optimization of this error function is achieved by a process called gradient descent where the weights and biases are repeatedly updated by taking inﬁnitesimal steps in the opposite direction of the gradient of the cost function with respect to updating parameters,

∂J(anl , y) Wl = Wl − η , (1.8) ∂Wl ∂J(anl , y) bl = bl − η (1.9) ∂bl where η is the learning rate which is a hyperparameter that determines the step size at each epoch of training. It is worth noting that models with many hidden layers were not commonly used, because the gradient information would vanish during the back propagation process which made lower level parameters a difficult task (Hochreiter et al., 2001). However, the introduction of the ReLU activation function (Nair and Hinton, 2010) significantly reduced the gradient vanishing problem and made training large networks computationally efficient. Also, the introduction of the dropout method (Hinton et al., 2012; Srivastava et al., 2014) made training even deeper and larger networks with many more parameters possible. The dropout is a regularizer that when applied on a layer l+1 randomly removes (or sets to zero) the output values of the previous layer (l) with probability p. Therefore, at each epoch of 46 training a different subset of output values from layer l is being removed. This will force each neuron in layer l+1 to learn useful features without heavily relying on other neurons in the same layer since they can be randomly removed (see Section A.1 for more information about FFFNNs).

HL2 (al+1) Output layer Input layer HL1 (al)

X3 anl

Figure 1.1: A representation of a feed forward neural network where HL stands for hidden layer. All of the layers are fully connected to each other.

1.6.1.2 Convolutional Neural Networks

Convolutional neural networks (CNNs; Fukushima, 1980; LeCun et al., 1998) are another class of neural networks that extract hierarchical features from the input like FFNNs but with reduced connectivity and shared weights that decrease the number of weights substantially. They consist of three diﬀerent types of layers: convolutional layers, pooling layers with shared weights, and fully connected layers. The convolutional layer extracts a local combination of features from a set of feature maps from the previous layer and convolves them using a set of ﬁlters to a new set of feature maps. Convolution operation 47 can be written as, K l+1 l+1 l+1 X l l l an = f (Zn ), Zn = Wmn ∗ am + bn (1.10) m=1 where K is the number of input feature maps in the form of matrices, ∗ represents the

l l linear convolution operation, Wmn represent the filters of layer l+1, and bn represent bias of feature map n. Each neuron in layer l + 1 is only connected to a local region (receptive field with size of F × F) in the input volume from layer l that is replicated to cover the entire input by shifting the receptive field (see Section A.1 for more information.). As we can see, the number of parameters reduces significantly when we replace the matrix multiplication in a fully connected layer shown in equation 1.7 with summation of convolutions of features map from previous layer in equation 1.10. This parameter reduction significantly improves the generalization performance of the network where the network performs better on unseen samples. The pooling layer is located after each convolutional layer that reduces the spatial dimensions of the input feature maps by applying a max or mean function (Boureau et al., 2010). This spatial reduction enables the convolutional layer to exploit larger regions of the input while reducing the number of parameters since pooling layers do not have trainable parameters. Therefore, they increase computation speed and reduce the chances of overfitting. State of the art CNNs are constructed of multiple convolutional layers followed by pooling layers which allow the deeper layers to extract more abstract representations of the input by allowing them to exploit a larger part of the input. We show a schematic representation of a convolutional layer followed by a pooling layer in Figure 1.2. The fully connected layers are added after multiple convolutional and pooling layers and the procedure is the same as mentioned in the Section 1.6.1.1. Because of the mentioned advantages and demonstrated strength for classification of large scale datasets (Krizhevsky et al., 2012), they have been used in many different computer vision problems. 48

Figure 1.2: A schematic representation of a convolutional layer followed by a pooling layer. Each neuron in layer l + 1 is only connected to a local region of feature maps from layer l. The pooling layer applies a mean or max function on the local region of all feature maps from the convolutional layer l + 1. 49 2 PS1-13cbe:The Rapid Transition of a Seyfert 2 to a

Seyfert 1

Originally published in (Katebi et al., 2019a):

Monthly Notices of the Royal Astronomical Society, Katebi, R., Chornock, R., Berger, E., Jones, D.O., Lunnan, R., Margutti, R., Rest, A., Scolnic, D.M., Burgett, W.S., Kaiser, N. and Kudritzki, R.P., 2019. PS1-13cbe: the rapid transition of a Seyfert 2 to a Seyfert 1. Monthly Notices of the Royal Astronomical Society, 487(3), 4057-4070. In this chapter, we present a nuclear transient event, PS1-13cbe, that was ﬁrst discovered in the Pan-STARRS1 survey in 2013. The outburst occurred in the nucleus of the galaxy SDSS J222153.87+003054.2 at z = 0.12355, which was classiﬁed as a Seyfert 2 in a pre-outburst archival SDSS spectrum. PS1-13cbe showed the appearance of strong broad Hα and Hβ emission lines and a non-stellar continuum in a Magellan spectrum taken 57 days after the peak of the outburst that resembled the characteristics of a Seyfert 1. These broad lines were not present in the SDSS spectrum taken a decade earlier and faded away within two years, as observed in several late-time MDM spectra. We argue that the dramatic appearance and disappearance of the broad lines and factor of ∼ 8 increase in the optical continuum is most likely caused by variability in the pre-existing accretion disk than a tidal disruption event, supernova, or variable obscuration. The timescale for the turn-on of the optical emission of ∼ 70 days observed in this transient is among the shortest observed in a “changing look” active galactic nucleus.

2.1 Introduction

The axisymmetric unification model of active galactic nuclei (AGN) suggests that there can be different types of AGNs based on the angle toward the line of sight. However, 50 it is believed that the simple orientation-based model cannot fully explain the Type 1 (both narrow and broad emission lines are present in the spectra) and Type 2 (only narrow emission lines are present in the spectra) classification scheme because this model is challenged when different AGN types are observed in the same object at different epochs of time. One alternative suggestion is that some Type 2s were Type 1s and their engine is now not active anymore or basically turned off; therefore, the Type 2s are evolved version of the Type 1s (Penston and Perez´ , 1984; Runnoe et al., 2016). This alternative view was first suggested based on the observations of broad emission lines disappearance in the Type 2 NGC 4151 (Lyutyj et al., 1984; Penston and Perez´ , 1984) and 3C 390.3 (Penston and Perez´ , 1984). The term “changing-look” (CL) was first used for the AGNs that showed X-ray absorption variations (Matt et al., 2003; Bianchi et al., 2009; Puccetti et al., 2007; Risaliti et al., 2009; Marchese et al., 2012). Recently, this term has also been used to describe the type of AGNs that have been observed to optically transition from Type 1 to Type 1.8, 1.9 and 2 or vice versa, by the sudden appearance or disappearance of broad emission lines such as the broad Hβ emission line (MacLeod et al., 2016; Runnoe et al., 2016). In some objects, the broad emission lines disappeared and the continuum faded completely (Collin- Souffrin et al., 1973; Tohline and Osterbrock, 1976; Sanmartim et al., 2014; Denney et al., 2014; Barth et al., 2015; Runnoe et al., 2016), while in others broad emission lines appeared or, in other words, the AGN “turned-on” (Cohen et al., 1986; Storchi-Bergmann et al., 1993; Aretxaga et al., 1999; Eracleous and Halpern, 2001; Shappee et al., 2014). These objects are mostly at low redshift with low absolute luminosities. However, recently LaMassa et al.(2015) discovered a luminous changing look AGN with a redshift of z = 0.31 that transitioned from Type 1 quasar to a Type 1.9 AGN in ∼ 9 years. Subsequent studies (MacLeod et al., 2016; Runnoe et al., 2016; MacLeod et al., 2018) have started finding larger samples in the Sloan Digital Sky Survey (SDSS). Most recent, Gezari et al.(2017) 51 discovered a quasar with a rapid “turn-on” timescale of < 1 year that demonstrated one of the fastest changes of state to date. The luminosities of Seyfert galaxies have been observed to vary over time. Variability has been detected in a wide range of wavelengths with timescales from hours to years with the shortest in X-rays that show variation time-scales of a few hours in Seyfert galaxies (Green et al., 1993; Nandra et al., 1997). The observed average variability time-scale in the optical in Seyfert galaxies is on the order of weeks to months (Kaspi et al., 1996; Giveon et al., 1999). Accretion disk instabilities are the most promising model as the source of this variability (Kawaguchi et al., 1998; Kelly et al., 2009). Here, we report the rapid turn-on (∼ 70 days) of a nuclear transient, PS1-13cbe, from a galaxy at redshift z = 0.12355 classified as a Seyfert 2 in a pre-event SDSS spectrum which was accompanied by the appearance of broad Balmer lines and hence a transition to a Seyfert 1. This represents the most rapid “turn-on” in a changing look AGN to date. We discuss the observations in Sections 2.2 and 2.3. We discuss the possible scenarios for the origin of the variations observed in PS1-13cbe in Section 2.4. We summarize and present our conclusions in Section 2.5. Throughout this work we are assuming a standard

−1 −1 ΛCDM cosmology with H0 = 69.6 km s Mpc , Ωm = 0.286 and ΩΛ = 0.714 parameters

(Bennett et al., 2014) that yields a luminosity distance of dL = 582 Mpc. All magnitudes are in the AB system and all dates are UT. We assume a Galactic extinction value of E(B − V) = 0.06 (Schlaﬂy and Finkbeiner, 2011).

2.2 Observations of PS1-13cbe

The Medium Deep Survey (MDS) of the Pan-STARRS1 (PS1) sky survey performed daily (in-season) deep monitoring of ten ∼ 7 sq. deg. ﬁelds over the years 2010–2014 to ﬁnd transient and variable sources. The typical observation sequence was composed of

gP1 and rP1 bands on the ﬁrst night, iP1 on the next night, and then zP1 on the third night. 52

This pattern was repeated during the ∼ 6 months of the observing season and was only interrupted by the weather and times near full moon, when observations in the yP1 filter were taken. A more complete description of the survey and photometric system were given by Tonry et al.(2012) and Chambers et al.(2016). On 2013 July 9, we detected a transient event, PS1-13cbe, at coordinates α = 22h21m53.86s, δ = +00◦30054.5600 (J2000) coincident with the nucleus of a galaxy using the photpipe transient discovery pipeline, described by Rest et al.(2014) and Scolnic et al.(2018). The host galaxy was observed as part of SDSS and given the name SDSS J222153.87+003054.2 (hereafter SDSS J2221+0030), with a spectroscopic redshift of z = 0.12355 (Ahn et al., 2014). We constructed a difference image light curve for PS1-13cbe using the PS1 transient pipeline (Scolnic et al., 2018). The template images were created from a stack of high-quality observations excluding the observing season containing the outburst (the year 2013) and were then subtracted from all observations of the transient. It is important to note that this photometry represents a flux difference relative to the host contribution present in the template (which is consistent with the SDSS photometry; Ahn et al. 2014). We include the PS1 photometry in Table 2.1.

2.2.1 Optical Photometry

As shown in Figure 2.1, the light curves of PS1-13cbe were constant and consistent with zero change in ﬂux relative to the baseline in the template for three observational seasons and then showed a rise peaking at MJD 56512.6, decline, and a second rise

(gP1 = 19.5 mag at peak; see Figure 2.2). The gP1 band luminosity showed a rise of about ∼ 1.5 × 1043 erg s−1 from the base luminosity in the course of ∼ 70 days and then declined to ∼ 0.7 × 1043 erg s−1 in ∼ 50 days and rose back up again to ∼ 1 × 1043 erg s−1 in the course of next ∼ 50 days, at which point the MDS ended. 53

15 y

) z −1 10 i r g erg s

42 5 (10 ν L

ν 0 S −5 −1200 −1000 −800 −600 −400 −200 0 200 Time (MJD − 56512.6) Figure 2.1: The observed transient luminosities of PS1-13cbe from the PS1 survey in

grizyP1 ﬁlters after correction for Galactic extinction. S: marks the epoch of the LDSS spectrum (MJD 56570).

With the SDSS value for the quiescent host ﬂux, g0 = 19.21 ± 0.01 mag (Ahn et al.,

2014), and gP1 = 19.5 mag for PS1-13cbe at the peak of the outburst, we can see that

the total luminosity of the galaxy increased by & 75%. However, the quiescent host flux value is dominated by star light. We estimate that the central AGN contributes .10% of the continuum flux in the quiescent spectrum taken by SDSS because the absorption lines from star light are not noticeably diluted by a non-stellar continuum, so the amplitude of the outburst from the AGN must be significantly larger, as discussed below. After the MDS ended, we obtained late-time photometry in g0 using the MDM4K and Templeton detectors on the 2.4 m Hiltner and 1.3 m McGraw-Hill telescopes at MDM Observatory on 2015 November 18 and 2017 June 18, respectively. The total magnitude of 54

15 16

18 y − 4 z − 3 i − 2 19 r − 1 g 20 Magnitude + constant 21 −100 −50 0 50 100 Time (MJD − 56512.6)

Figure 2.2: Light curves of PS1-13cbe in grizyP1 bands, corrected for Galactic extinction. Note the small-scale ﬂuctuations in the light curve (such as near −20 d).

the host galaxy (including any possible transient contribution) was g0 = 19.23 ± 0.05 mag (g0 = 19.22 ± 0.13 mag) at 831 (1408) days after the peak, which are consistent with the SDSS pre-outburst photometry. Therefore, the system returned to the baseline ﬂux value in

. 2 yr after the outburst.

2.2.2 Observations of the Host Galaxy

In addition to the SDSS observation of the host (∼ 10 years before the outburst), the source was detected by the AllWISE survey in the W1, W2, W3, and W4 bands with cataloged values of 17.82 ± 0.037, 18.20 ± 0.34, 16.65 ± 0.23 and 15.13 ± 0.37 magnitudes, respectively (∼ 3 years before the outburst; Chang et al., 2015). We also obtained photometry using the Neil Gehrels Swift Observatory (Gehrels et al., 2004) with 55 the UV Optical Telescope (UVOT; Roming et al., 2005) in the u, uvw1, and uvw2 ﬁlters on 2016 November 27–29 (∼ 3 years after the outburst) and we measure values of 21.10±0.13, 21.76 ± 0.21 and 22.41 ± 0.19 magnitudes, respectively.

2.2.3 X-ray Photometry

We obtained an X-ray observation of SDSS J2221+0030 using the X-ray telescope (XRT; Burrows et al., 2003) on Swift with a total exposure time of 7.1 ks between 2016 November 27 and November 29. The host was not detected with a 3σ upper limit of 2.13 × 10−3 cts s−1 (0.3 − 10 keV). Using the Galactic neutral hydrogen column density in

20 -2 the direction of PS1-13cbe of NH = 4.49 × 10 cm (Kalberla et al., 2005) and assuming no intrinsic absorption, and a typical photon index of Γ = 2, we calculate the unabsorbed X-

−14 −2 −1 42 ray ﬂux to be fx(2 − 10 keV) ≤ 4.0 × 10 erg cm s , which translates to LX ≤ 1.6 × 10 erg s−1. The absorbing column densities of Seyfert 2 galaxies range from 1022 to 1025 cm-2 (e.g., Risaliti et al. 1999) and therefore, assuming a minimum intrinsic column density of 1022 cm-2 for the host galaxy in addition to the Galactic value, we recalculate the

−13 −2 −1 42 unabsorbed X-ray ﬂux to be fx(2 − 10 keV) ≤ 1.0 × 10 erg cm s or LX ≤ 4.1 × 10

−1 erg s . Additionally, using LX and the empirical bolometric correction from Marconi et al. (2004), this corresponds to a minimal upper limit on the bolometric luminosity of the object

44 −1 Lbol ≤ 0.6 × 10 erg s 3 years after the outburst, although this value can increase if the

45 −1 intrinsic absorption is higher (e.g., Lbol ≤ 9.0 × 10 erg s for an intrinsic column density of 1024 cm-2). We did not ﬁnd any archival X-ray observations of the object before or at the time of the outburst.

2.2.4 Optical Spectroscopy

A pre-outburst spectrum of SDSS J2221+0030 was obtained on 2003 May 26 by SDSS (Ahn et al., 2014). After the detection by PS1, we observed PS1-13cbe on 2013 56

October 5 (+57 days after peak) for 1200 s using the Low Dispersion Survey Spectrograph- 3 (LDSS3) on the 6.5 m Magellan Clay telescope. We used a 100-wide long slit with the VPH-all grism to cover the observed wavelength range 3700 − 10000 Å with a resolution of ∼ 9 Å. At late times, we obtained five epochs of spectroscopy using the Ohio State Multiple Object Spectrograph on the 2.4 m Hiltner telescope at MDM Observatory (OSMOS; Martini et al., 2011). The first two spectra were taken on 2015 October 3 and 2015 November 12 using the 1.200 center slit with a VPH-red grism and an OG530 filter that covered 5350 − 10200 Å (resolution = 5 Å) in the observed frame. No significant spectral differences were present, so we combined these two observations for all subsequent analysis. The next two spectra were obtained on 2015 December 1 and 2016 November 16 using the same 1.200 center slit and a VPH-blue grism that covered 3675−5945 Å (resolution = 2 Å) in the observed frame. The final spectrum was taken on 2017 June 18 using a 1.000 outer slit and VPH-red grism that covered 3930−9050 Å (resolution = 4 Å) in the observed frame. We preprocess our data using standard procedures such as flat-fielding, bias subtraction, and wavelength calibration using arc lamps in IRAF 3. Additionally, we removed cosmic rays using the L.A.Cosmic (Van Dokkum, 2001) task. Finally, we calibrate our data using our own IDL procedures and observations of the standard stars BD+174708 for red spectra and Feige110 for blue spectra. All of the spectra for PS1-13cbe are shown in Figure 2.3.

3 IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc. (AURA) under cooperative agreement with the National Science Foundation. 57

3 2017/06/18 ) −1 Å −1

s 2 2016/11/16 −2

2015/12/01 /

erg cm 2015/10/03

−16 1 (10 λ 2013/10/05 Logf 0 2003/06/26

−1 3000 4000 5000 6000 7000 8000 Rest frame wavelength (Å)

Figure 2.3: Optical spectra for PS1-13cbe, from bottom to top: host galaxy of PS1-13cbe from SDSS (grey), model of the host galaxy generated by FAST 1.0 (black), spectrum during outburst observed using LDSS3 (purple), spectra obtained with OSMOS in red and blue for the red and blue setups, respectively, and the most recent OSMOS spectrum (olive green).

2.3 Observational Features of PS1-13cbe

2.3.1 Host galaxy of PS1-13cbe

To isolate the emission lines of SDSS J2221+0030 and correct for stellar absorption lines, we simulated a galaxy model with the FAST 1.0 code (Kriek et al., 2009) using the archival SDSS spectrum and optical photometry in ugriz bands along with the Swift u, uvw1, and uvw2 photometry. We experimented with the initial parameters and generated the best fit with the stellar age of 3 × 109 years and e-folding timescale of τ ≈ 109 years, by 58 assuming the star-formation history to be exponentially declining, the stellar initial mass function from Chabrier(2003), the Bruzual and Charlot(2003) spectral library, a Milky Way dust law (Cardelli et al., 1989), and solar-like metallicity of Z = 0.02. The FAST model of the host galaxy is over-plotted on the original SDSS spectrum (grey) in the bottom of the Figure 2.3 (black). After subtracting this model from the pre-outburst optical spectrum of the host galaxy of PS1-13cbe (bottom of Figure 2.3 in grey), we fit the profiles of the narrow emission lines [S II]λλ6717, 6731, [O III]λλ4959, 5007, [O II]λ3727, [O I]λ6300, [N II]λλ6549, 6583, Hα, and Hβ. It has been shown that a model for narrow emission lines can be obtained from [S II] lines in most cases (Ho et al., 1997). By using this fact, we constrained parameters of the Gaussian profiles of the other narrow lines. The lines were not well modeled with single Gaussians, so we used double Gaussian profiles to fit the line profile of all the narrow lines except [O II]λ3726 and [O I]λ6300. In Figure 2.4, we plot the line ratios in excitation diagrams (Baldwin et al., 1981). We also show the extreme star formation line (Kewley et al., 2001, 2006), the pure star formation line (Kauffmann et al., 2003) and the Seyfert- LINER classification line (Kewley et al., 2006; Fernandes et al., 2010). Additionally, we show 30000 randomly selected galaxies (shaded area in Figure 2.4) with emission-line fluxes from the MPA-JHU DR7 catalog 4 (Aihara et al., 2011). The automatic Portsmouth pipeline from SDSS classified the host galaxy of PS1-13cbe as a LINER (Sarzi et al., 2006); however, our emission line ratios calculated from the SDSS spectrum after subtraction of the stellar continuum (navy squares) classify the host galaxy as a clear Seyfert (Figure 2.4). Next, we constructed the spectral energy distribution (SED) of the host galaxy using the observed and archival photometry data. We also scaled and plotted SEDs of an Sb spiral galaxy and a Seyfert 2 galaxy from the SWIRE template library (Polletta et al., 2007) alongside the SED of the host (see Figure 2.5). The SED of the Seyfert 2 galaxy

4 http://wwwmpa.mpa-garching.mpg.de/SDSS/DR7/ 59

1.0 AGN Seyfert 0.4 Seyfert

0.2 0.5

0.0 HII ) β

−0.2 0.0 HII LINER LOG([OIII/OII]) LOG([OIII]/H −0.4

HII −0.5 −0.6

−0.8 LINER Comp −1.0 −1.0 −1.0 −0.5 0.0 0.5 −2.0 −1.5 −1.0 −0.5 0.0 −2.0 −1.5 −1.0 −0.5 0.0 LOG([NII/Hα]) LOG([OI/Hα]) LOG([OI/Hα])

Figure 2.4: The excitation diagrams using the [N II]λ6583/Hα, [O I]λ6300/Hα, [O III]λ5007/Hβ and [O III]λ5007/[O II]λ3726 line ratios (Baldwin et al., 1981; Kewley et al., 2006). The navy blue squares display the position of the host galaxy of PS1-13cbe. The shaded area represents the location of the SDSS galaxies calculated using MPA-JHU where darker regions represent higher number density of the galaxies. The extreme star formation line (solid red; Kewley et al., 2001, 2006), the revised star formation line (dotted black; Kauﬀmann et al., 2003) and the Seyfert-LINER classiﬁcation line (dashed black; Kewley et al., 2006; Fernandes et al., 2010) are also plotted. Comp: AGN/star forming composites.

ﬁts well from the mid-IR to UV, including the contribution from star light. As we can see in Figure 2.5, the SED of the spiral galaxy ﬁts the optical and UV part of the host SED as well; however, there is a broad excess in the mid-IR (W3 and W4). This excess is because of the contribution of dust heated by the AGN in the host galaxy. After subtracting the host contribution (using the Sb template) from the observed values, we estimate the luminosity 60

43 −1 of the AGN component to be νLν(W4) ≈ (1.63 ± 0.65) × 10 erg s . Then, using a bolometric correction factor of 10.1 ± 1.4 for the W4 band from Runnoe et al.(2012), we

44 estimate the total bolometric luminosity of the AGN to be Lbol(W4) ≈ (1.65 ± 0.65) × 10 erg s−1. Furthermore, narrow emission lines are possible indicators of the intrinsic bolometric luminosity (Netzer, 2009). We estimate the bolometric luminosity of the AGN from the luminosity of the narrow [O III] λ5007 emission line in the SDSS spectrum and used the conversion that Lbol = 3500L(λ5007), which has a variance of 0.38 dex by assuming a

44 standard AGN SED (Heckman et al., 2004). This results in an estimate of Lbol = 1.6 × 10 erg s−1, which is consistent with the bolometric luminosity calculated using the W4 band above. This luminosity can be an overestimate if some of the [OIII] emission originates from star formation. However, as shown in Figure 2.4, SDSS J2221+0030 is classiﬁed as a clear Seyfert galaxy and therefore narrow line excitation is dominated by the AGN. Moreover, as we discussed in Section 2.2.3, we estimate the upper limit on the bolometric luminosity in quiescence from the X-ray non-detection to be ≤ 0.6 × 1044 erg s−1, which is consistent with the bolometric luminosity estimated before the outburst in quiescence. This is only an estimate because the X-ray ﬂux is a non-detection, whereas having an X- ray detection would help us to have a better estimate of the intrinsic absorption and column density.

2.3.2 Astrometry

To find the location of PS1-13cbe in its host galaxy, we performed relative astrometry between the PS1/MDS template images of the host galaxy and the position of the transient reported by photpipe. First, we fit a 2-dimensional Gaussian function to the templates in all filters to find the centroid of SDSS J2221+0030 and then, using the weighted average centroid coordinates of the transient, we calculate the offset of PS1-13cbe from the center 61

44 10 Sy2 Sb

) data − 43 10 (ergs ν ν

42 Log 10

12 13 14 15 16 10 10 10 10 10 Log ν (Hz)

Figure 2.5: SED of SDSS J2221+0030 in quiescence (purple circles), scaled SED templates of an Sb spiral galaxy (dotted orange line) and a Seyfert 2 galaxy (dashed blue lines) from the SWIRE template library (Polletta et al., 2007). The host exhibits a mid-infrared (MIR) excess relative to the star-forming template due to an AGN.

of its host galaxy to be 0.03600 ± 0.03500 (101 ± 100 pc), consistent with the nucleus. We used a systematic astrometric error ﬂoor of 0.1 pixel to calculate the uncertainty in the oﬀset (Scolnic et al., 2018). We show the position of PS1-13cbe relative to the center of its host in gP1 band in Figure 2.6.

2.3.3 Multi-band Light Curves of PS1-13cbe

We calculate the total optical luminosity light curve of PS1-13cbe using multi-band observations shown in Figure 2.1. This has been done by ﬁrst ﬁnding the epochs where 62

Nucleus Outburst

1′′ = 2.8 kpc

Figure 2.6: The gP1 PS1 template image of SDSS J2221+0030 showing the position of the centroid of the galaxy (green “X”), the position of PS1-13cbe (red dot), and its position uncertainty is shown by a circle with 5σ radius (red circle). The position uncertainty is dominated by the outburst.

gP1 and rP1 bands were observed simultaneously and then interpolating the iP1, zP1, and yP1 bands at these epochs using Legendre polynomials. Next, by integrating the spectral 63 distribution at each epoch and using the trapezoidal rule, we calculate the total ﬂux and thereafter optical luminosity over 3685 − 8910 Å in the rest frame. At late and early times, we had data only from the yP1 and zP1 bands, so we calculate the luminosity assuming the same colour correction as measured from those epochs with all ﬁlters. As seen in Figure 2.7, the total optical luminosity of PS1-13cbe rises to a peak value of (1.06 ± 0.01) × 1043 erg s−1 in the course of ∼ 70 days, declines in next ∼ 50 days, and then rises back up again.

12 )

−1 10 z y 8 g−r−i−z−y Trapezoidal erg s 42 6 4 2

Luminosity (10 0 −300 −200 −100 0 100 Time (Rest frame days after peak)

Figure 2.7: Total optical luminosity light curve of PS1-13cbe integrated over the grizyP1 filters and relative to the baseline flux in the template images. The total optical luminosity was estimated using the spectral distribution at each epoch and the trapezoidal rule (blue triangles). The estimated luminosity at early and late times using yP1 (red circles) and zP1 (purple squares) bands assumed the same colour correction as measured from those epochs with all filters. 64

Furthermore, to study the evolution in continuum colour and temperature while remaining agnostic about the overall SED of the transient, we ﬁt both power law and blackbody models. We also calculate the g − r colour and show it alongside the spectral index and blackbody temperatures for all epochs in Figure 2.8. The reason we chose gP1 and rP1 bands as a proxy for the colour is that they were taken on the same night in the PS1 survey and do not need to be interpolated. No strong colour evolution during the outburst is evident in Figure 2.8.

9500 9000 8500 8000 7500 7000 Temperature (K) 6500 0.0 −0.2 −0.4

index −0.6 Spectral −0.8 −1.0 0.3 0.2 0.1 0.0 g−r (mag) −0.1 −0.2 −50 0 50 100 Time (Rest frame days after peak)

Figure 2.8: Top: Rest-frame blackbody temperature from ﬁtting the optical photometry.

α Middle: Rest-frame spectral index from ﬁtting a power law fν ∝ ν to the optical photometry, where α is the spectral index. Bottom: observed g-r colour diagram. All three panels show little or no evolution during the outburst. 65

2.3.4 Spectral Features of PS1-13cbe

In Figure 2.3, we show the optical spectra for PS1-13cbe. The spectra of PS1-13cbe show remarkable evolution over the course of fourteen years. The spectra contain narrow and broad emission line profiles, including hydrogen Balmer lines, [S II]λλ6717, 6731, [O III]λλ4959, 5007, [O II]λ3726, [O I]λ6300 and [N II]λλ6549, 6583. We note that the spectra were taken with different effective apertures (slit and seeing) and different spectral resolutions. We fit the Hα+[N II] and Hβ+[O III] complexes in order to investigate the presence of broad Hα and Hβ lines. First, we scaled our OSMOS (blue setup) and LDSS3 spectra using the flux of [O III] λ5007 from the SDSS spectrum and scaled the OSMOS red setup using the [N II] flux from SDSS. We did these scalings with the assumptions that these narrow lines are centrally concentrated in the host and that the fluxes do not change in such a short timescale because it has been shown that narrow emission lines only slowly vary over decades (Peterson et al., 2013). In addition, the LDSS3 spectrum was taken during the outburst and contains

α transient flux. In that case, first we fit a power law fν ∝ ν to the optical photometry from PS1 at the epoch when the spectrum was taken, where α = −0.58. After that, we constructed a model from a linear combination of the power-law continuum and the host galaxy model shown in Figure 2.3. Additionally, we smoothed this model by a Gaussian with full width at half maximum (FWHM) = 5 Å to better match the resolution of our data and subtracted it from the LDSS3 spectrum, isolating the emission-line spectrum. Next, we fit the [S II] lines using double Gaussian profiles and use this model to constrain the multi-component Gaussian profiles that were used to fit the Hα+[N II] complex to reduce the number of free parameters (Ho et al., 1997). Specifically, we model emission lines using two Gaussian profiles for each narrow line and one broad component for Hα and simultaneously fit for parameters of the broad and and narrow lines. We fix the narrow components to the wavelengths of Hα and [N II] λλ6549, 6583. Also, we fix the 66 widths and relative amplitudes of the two Gaussian components of the narrow lines using the values from the [S II] model, leaving only the overall amplitudes of the narrow Hα and [NII] lines as free parameters. The parameters of the broad component of Hα were allowed to vary freely. We fit the [O III] λλ4959, 5007 lines using two Gaussian profiles for each narrow line without constraining the parameters. For Hβ, however, we use two Gaussian profiles, a single one for the narrow component with just the centroid fixed and allowing the other parameters to vary, and one for the broad component with no constraints. The SDSS and OSMOS spectra lack transient flux, so we only subtract a scaled galaxy model and perform the same procedure as for the LDSS3 spectrum to fit the narrow and broad emission lines, with the exception of fixing the width and the the centroid of the broad component of both Hα and Hβ to the values derived from LDSS3 spectra, allowing only the normalization to vary. We show the resultant Gaussian fits to the Hα+[NII] and Hβ+[O III] emission lines in Figure 2.9 and 2.10 respectively. As we show in Figure 2.9 and Figure 2.10, the existence of the broad Hα and Hβ components is clear in the LDSS3 spectrum (a weak broad Hγ emission line was also detected). Ho et al.(1997) found that they could reliably extract a weak broad H α component that comprised ≥ 20% of the Hα + [NII] blend from spectra with spectral resolution close to ours. As shown in Figure 2.9, there is marginal evidence for a broad component of Hα in the SDSS and OSMOS data, which is 21% and 17% of the flux in the Hα + [NII] complex, respectively. However, it is notable that there are visible wiggles in the residuals of the fit from the spectra that suggest the broad emission line profiles are not purely Gaussian. For this reason, we are not confident that broad Hβ line in the SDSS data and broad lines in the OSMOS are real and are not result of deviations from a Gaussian profile. We report the measured luminosities of the broad components of the Hα and Hβ emission lines in Table 2.2. 67

SDSS Data LDSS3 Data OSMOS Data OSMOS Data Best Fit Best Fit Best Fit Best Fit 4 Individual Individual Individual Individual

) Gaussians Gaussians Gaussians Gaussians

−1 Broad Hα Broad Hα Broad Hα Broad Hα

Å 2003/06/26 2013/10/05 2015/10/03 2017/06/18 −1 3 − 3697 d + 57 days + 785 d + 1409 d s −2

2 erg cm −16 1 (10 λ f

0.4 0.2 0.0

Residual −0.2 −0.4 6500 6550 6600 6500 6550 6600 6500 6550 6600 6500 6550 6600 Rest wavelength (Å)

Figure 2.9: Continuum subtracted Hα line profiles. Top: Multiple-component Gaussian fit to the Hα+[NII] emission lines (blue), individual components (dashed green) and broad component of the Hα (dashed black). Bottom: The fit residuals. Top-left (on each panel): Observation date and numbers of days before/after the peak (navy blue).

As mentioned in Section 2.3.1, we classiﬁed the SDSS spectrum to be that of a Seyfert. Moreover, the galaxy can be classiﬁed as a Seyfert 1.9/2 galaxy since it just shows narrow emission lines and possible weak broad Hα without any sign of a broad Hβ line, as shown in the left panels of Figures 2.9 and 2.10. At the time of the outburst, the spectrum of PS1- 13cbe taken with LDSS3 showed broad Hα and Hβ components with FWHM velocities of 3385 ± 32 km s−1 and 3277 ± 110 km s−1, respectively. These high values of FWHM are the sign of the high velocity, dense, and highly ionised gas clouds in the Broad Line Region (BLR) close to the central black hole (BH) where the broad emission lines originate from. The presence of these broad lines suggests that the galaxy transformed from a Seyfert 1.9/2 68

SDSS Data LDSS3 Data OSMOS Data OSMOS Data Best Fit Best Fit Best Fit Best Fit Individual Individual Individual Individual 1.5 Gaussians ) Gaussians Gaussians Gaussians

−1 Broad Hβ Broad Hβ Broad Hβ Broad Hβ

Å 2003/06/26 2013/10/05 2015/10/03 2017/06/18 −1 − 3697 days + 57 days + 785 days + 1409 d s

−2 1.0 erg cm

−16 0.5 (10 λ f

0.0

0.3 0.2 0.1 0.0

Residual −0.1 −0.2 4800 4900 5000 4800 4900 5000 4800 4900 5000 4800 4900 5000 Rest wavelength (Å)

Figure 2.10: Continuum subtracted Hβ line profiles. Top: Multiple-component Gaussian fit to the Hβ+[O III] emission lines (blue), individual components (dashed green) and broad component of the Hβ (dashed black). Bottom: The fit residuals. Top-left (on each panel): Observation date and numbers of days before/after the peak (navy blue). Strong residuals are present in the red wing of λ4959 due to poor subtraction of the 5577Å sky line.

to a Seyfert 1 galaxy. On the other hand, we did not detect broad emission lines in the spectra taken with OSMOS and that means the host galaxy of PS1-13cbe transformed back from a Seyfert 1 to a Seyfert 2 galaxy in less than 2 years and continues in that state as there are no signs of broad lines in spectra taken with OSMOS ∼ 4 years after the peak of the outburst. We estimate the mass of the central BH using two methods. The ﬁrst method uses the stellar kinematics and the revised scaling relation between the SMBH mass and stellar velocity dispersion (McConnell and Ma, 2013), where σ = 93.52 km s−1 was provided 69

+3.5 6 from the SDSS spectrum of the host galaxy, and results in an estimate of MBH = 2.9−1.7×10

M . However, this estimation is subject to uncertainties because of the large scatter and lack

of constraints at the low σ and MBH region of the MBH − σ scaling relation, particularly in a case where we do not have a decomposition of the galaxy that separates the bulge component (McConnell and Ma, 2013; Kormendy and Ho, 2013). The mass of the SMBH can also be estimated using photoionisation equilibrium, by applying a mass-scaling relationship based on the FWHM of the broad Hβ emission line and continuum luminosity. Therefore, using the measured FWHM ≈ 3277 km s−1, intrinsic

42 −1 luminosity at 5100Å (λL5100) of (6.2 ± 0.2) × 10 erg s at the time of the spectrum, and the mass-scaling relationship of Vestergaard and Peterson(2006), we calculate the

7 mass of the central BH to be (2.2 ± 0.1) × 10 M . We note that the values used in this calculation are measured during the outburst and can be aﬀected by the changes in the accretion structure and also rely on the assumption that the (unknown) shape of the ionising continuum is similar to those objects used to calibrate the scaling relationship. We prefer this mass estimate in our calculations below because we believe the assumptions behind this photoionisation calculation to be more robust.

45 −1 The Eddington luminosity for this BH mass is LEdd = (2.7±0.1)×10 erg s . We also

43 calculate the intrinsic luminosity at 5100Å (λL5100) for PS1-13cbe that is (1.16±0.01)×10

−1 erg s at the time of the peak and then we estimate the bolometric luminosity using λLλ and a conversion factor of 8.1 (Runnoe et al., 2012) to convert from monochromatic to bolometric luminosity to be (9.4±0.1)×1043 erg s−1. This results in an Eddington parameter of λEdd = Lbol/LEdd ≈ 0.03 at the peak of the outburst.

2.4 Interpretation of the Features of PS1-13cbe

In Figure 2.11, we show the spectrum of PS1-13cbe at the time of the outburst (after subtraction of the star light component from the host galaxy model) alongside the 70 comparison objects, including a QSO (SDSS QSO template), a Type IIn SN (SN1994Y; Filippenko, 1997), and a tidal disruption event (TDE: ASASSN-14li; Holoien et al. 2015). All of these objects show the presence of broad Hα and Hβ lines with emission line proﬁles similar to the PS1-13cbe spectrum. It is notable that the PS1-13cbe spectrum closely resembles that of the QSO template, which is consistent with AGN activity. However, we estimate the spectral index at the epoch of the LDSS3 spectrum to be α = −0.58, which is redder than the highly variable QSOs studied in Wilhite et al.(2005) that had spectral index of α = −2 for the average diﬀerence spectrum (bright phase minus faint phase). In this section, we discuss SNe, TDE, and AGN variability as three possible interpretations for PS1-13cbe. However, before discussing the details about these scenarios we have summarized the key features of PS1-13cbe:

• PS1-13cbe occurred in the nucleus of a Seyfert 2 galaxy with a central SMBH with

7 mass of ∼ 2 × 10 M .

• PS1-13cbe brightened in the course of ∼ 70 days and reached a peak total optical luminosity of (1.06 ± 0.01) × 1043 erg s−1.

• The temperature of PS1-13cbe roughly stayed constant and g − r (transient component) did not show any colour evolution.

• The spectra of PS1-13cbe show signiﬁcant evolution over the course of ∼ 12 years where broad Hα,Hβ, and Hγ lines appear and disappear from the spectra.

2.4.1 Type IIn Supernovae Interpretation of PS1-13cbe

Type IIn SNe show broad H Balmer lines and blue continua in their optical spectra that can resemble the spectra of AGNs in speciﬁc phases of their evolution (Filippenko, 1989). As we show in Figure 2.11, the lack of P Cygni proﬁles in the spectrum of PS1-13cbe is similar to the spectral features of Type IIn SNe such as SN1994Y (Filippenko, 1997) 71

5 γ Hβ H He II Hα He I 4 ASASSN−14li He I )

−1 β Hα Hγ H Å

−1 3 He I s SN1994Y He I −2

2 Hα β Hγ H erg cm Fe II Fe II He I He I

−16 QSO 1 SDSS template

(10 Hα λ Hβ

Logf γ 0 H Fe II PS1−13cbe Fe II He I 2013/10/05 He I

−1

3000 4000 5000 6000 7000 8000 Rest Frame Wavelength (Å)

Figure 2.11: Optical spectrum of PS1-13cbe taken with LDSS3 during the outburst after subtracting the stellar continuum (blue). The spectra of the comparison objects include a QSO (SDSS QSO; Vanden Berk et al., 2001) template, a SN Type IIn (SN1994Y; Filippenko, 1997) and a TDE (ASASSN-14li; Holoien et al., 2015) from bottom to top (black), the emission lines are labeled (purple).

that is used as an example here for the purpose of the comparison. Additionally, the peak

luminosity of PS1-13cbe, Mr ≈ −19.4, is not substantially brighter than the typical Type

IIn SNe, which have a broad peak luminosity range of −18.5 ≤ Mr ≤ −17 (e.g., Kiewe et al. 2011). The Type-IIn SN hypothesis is however disfavored for the following reasons. First, the narrow emission line ratios in Figure 2.4 clearly identify the galaxy as a Seyfert, which increases the chances that this variability was caused by the existing AGN rather than a SN. Furthermore, the temperature of PS1-13cbe is roughly constant and it does not show 72 any colour evolution, which is not consistent with the cooling of the ejecta typically seen in SNe. Additionally, the double hump behavior seen in the light curves (Figure 2.2 and Figure 2.7) of PS1-13cbe is not commonly observed in Type IIn SNe (e.g., Kiewe et al., 2011; Taddia et al., 2013). While Type II SNe have been very rarely observed to exhibit multiple rebrightening bumps after the ﬁrst peak (e.g., iPTF14hls and iPTF13z; Arcavi et al., 2017; Nyholm et al., 2017), the small ﬂuctuations observed in the light curve of PS1- 13cbe are not present in their light curves and the rise and rebrightening timescales in the case of PS1-13cbe are much shorter in comparison. Based on the reasons provided in this section, we disfavor the Type IIn SN origin of PS1-13cbe, but this possibility cannot be completely ruled out.

2.4.2 PS1-13cbe as a TDE

Most of the optically detected TDEs show a lack of colour evolution and constant blackbody temperature, consistent with our observations of PS1-13cbe (Gezari et al., 2012; Chornock et al., 2013). Furthermore, broad emission lines such as Hα,Hβ and He II have been detected in the spectra of the TDEs and disappeared at later times (Arcavi et al., 2014; Van Velzen et al., 2011). Additionally, broad Hα and Hβ lines are detected in the spectrum close to the time of the peak and either were not present (Hβ) or only weakly present (Hα) in earlier spectra (SDSS) and disappear in the later spectra (OSMOS). Therefore, it is plausible that PS1-13cbe has a TDE origin. However, we do not have optical photometry data on the decline to check whether the light-curve decays at the predicted rate of t−5/3 for TDEs (Rees, 1988; Evans and Kochanek, 1989). The lack of any X-ray observations at the time of the outburst is another limitation we face to investigate the TDE and SN scenarios. As seen in Figure 2.4, the host of PS1-13cbe is classiﬁed as a Seyfert galaxy, which points to the presence of a pre-existing AGN. However, TDEs can happen in AGN galaxies, and in fact it has been suggested that TDEs may prefer galaxies with pre-existing steady 73 accretion to their central BH (Hills, 1975; Blanchard et al., 2017) where the dense star- formation clouds and the pre-existing accretion disk can increase the chance of the tidal encounter with the stars and reduce the relaxation time (Perets et al., 2006; Blanchard et al., 2017). For example, the spectrum of the hosts of the optical/UV TDE SDSS J0748 (Wang et al., 2011) and optical TDE candidate PS16dtm (Blanchard et al., 2017) show weak and narrow emission lines that might be the result of AGN presence in the core. Similar to them, the host of the TDE ASASSN-14li (Holoien et al., 2015) also shows traces of ongoing weak AGN activity at the center (van Velzen et al., 2015; Alexander et al., 2016). Another possibility for brightening and rebrightening of nuclear transients could be repeat tidal stripping of stars (Campana et al., 2015; Ivanov and Chernyakova, 2006; Komossa et al., 2016), although the timescale for rebrightening in PS1-13cbe is very short. However, we disfavor the TDE interpretation of PS1-13cbe because of the following reasons. The blackbody temperatures in optically-selected TDEs typically range from ∼ 20000 K (e.g., PS1-11af; Chornock et al., 2013) to ∼ 35000 K and higher (e.g., ASASSN-14li; Holoien et al., 2015), while the continuum of PS1-13cbe at the time of the outburst is not as blue as a typical TDE and, as we show in Figure 2.8, the inferred blackbody temperature of PS1-13cbe is no more than ∼ 9000 K. Moreover, the He II λ4686 line that is frequently seen in the spectra of TDEs (e.g., ASASSN-14li in Figure 2.11), is not visible in the spectrum of PS1-13cbe that was taken close to the time of outburst. However, it is worth mentioning that there are TDE candidates that lack the presence of the He II λ4686. Additionally, the rebrightening of PS1-13cbe by almost 75% is not typically seen in TDEs. However, it is notable that the TDE candidate ASASSN-15lh showed rebrightening in the UV +60 days after the peak (Dong et al., 2016; Leloudas et al., 2016; Margutti et al., 2017). The optical TDE candidate PS16dtm also showed dimming and rebrightening ∼ 150 days after the beginning of the rise (Blanchard et al., 2017). 74

Furthermore, the light curves of the PS1-13cbe have noticeable ﬂuctuations on few days timescales (such as the one visible in griP1 near −20 days in Figure 2.2) that are inconsistent with the observed smooth optical light curves of most of the other known optically-selected TDEs (e.g., Van Velzen et al., 2011; Gezari et al., 2012; Chornock et al., 2013; Arcavi et al., 2014; Holoien et al., 2014). However, noticeable ﬂuctuations have been observed in the light curves of optical TDE candidate PS16dtm (Blanchard et al., 2017). The optical and spectral features of PS1-13cbe and the reasons provided in this section show that the TDE origin of PS1-13cbe is a possible but not a likely scenario.

2.4.3 PS1-13cbe as a “Changing Look” AGN

Recently, a new type of AGN variability was discovered in objects called “changing look” AGNs that show the appearance or disappearance of broad emission lines followed by an order of magnitude increase or decrease in the continuum and change type from Type 1 to Type 1.8, 1.9 or vice versa (e.g., Runnoe et al., 2016; Shappee et al., 2014; LaMassa et al., 2015; Gezari et al., 2017; MacLeod et al., 2016). This CL behavior observed in some AGNs can be caused by at least three mechanisms. In the ﬁrst scenario, variation of the obscuration when material such as dust clouds outside of the BLR move in or out of the line of sight that can obscure or clear the view to the BLR (Elitzur, 2012). Another mechanism can be variations of accretion rate that transforms the structure of the BLR. The AGN will transfer from Type 1 when the accretion rate is high and the broad lines are visible to a Type 2 when the accretion rate is low and the broad lines cannot exist or vice versa (Elitzur et al., 2014). Furthermore, it has been suggested that transient events such as TDEs can cause this CL behavior (Eracleous et al., 1995). As shown in Figure 2.9 and 2.10, we only detected the presence of weak broad Hα and no broad Hβ emission line in the spectrum taken by SDSS. However, the strong broad Hα line and Hβ lines appeared at the time of the outburst that followed an observed increase 75 in the ﬂux and disappeared again in the spectra taken at later times. This shows that the AGN changed type from a Type 1.9/2 to a Type 1 and then to a Type 2 because no broad emission lines were detected in later observations.

We estimate the luminosity at 5100 Å (λL5100) based on our estimate that the non- stellar continuum contributes ≤ 10% in the SDSS spectrum to be ≤ 0.15 × 1043 erg s−1 in quiescence. Additionally, we measured the intrinsic 5100 Å (λL5100) luminosity at the

43 −1 peak of the outburst to be 1.16 × 10 erg s , which shows a factor & 8 increase in optical luminosity. Also, during this time the broad Hα emission varied by a factor of ∼ 4 in ﬂux. The observed luminosity changes and appearance and disappearance of the Hα and Hβ lines can be caused by one of the three mentioned mechanisms that are discussed in more detail in the following sections.

2.4.3.1 Obscuration of the AGN

The CL behavior seen in PS1-13cbe can be caused by intervening material which is located outside of the BLR and orbits on a Keplerian orbit that can obscure or give a clear view to the BLR by moving in or out of the line of sight. The bolometric luminosity of ∼ 9.4 × 1043 erg s−1 estimated at the time of outburst using bolometric corrections for unobscured AGNs is less than the bolometric luminosity of ∼ 1.6×1044 erg s−1 at the time of quiescence estimated from the narrow λ5007 and mid-IR excess, which possibly indicates that the AGN is not fully unobscured at the peak of the outburst. In addition, the redder spectrum compared to the QSO template suggests such a scenario to be possible. However, we disfavor the changing obscuration scenario in the case of PS1-13cbe for the following reasons. First, we estimate the characteristic radius of the BLR using the R-L relation calibrated presented by (Bentz et al., 2013) to be RBLR ∼ 11 light days. Then, using the relation for crossing time presented by LaMassa et al.(2015), we estimate the crossing time for an obscuring object orbiting on a Keplerian orbit outside of BLR. In the most ideal case 76 that minimizes the crossing time, the obscuring object should be at rorb ≥ RBLR = 11 light days so that it can intercept a substantial amount of the broad Balmer ﬂux from the BLR.

Even in the extreme case of assuming rorb = RBLR, the crossing time for the obscuring object is ∼ 23 years. In a more realistic scenario where rorb ≥ 3RBLR the crossing time is even higher. Not only are these timescales too long to explain the CL behavior of PS1-13cbe, but also the existence of intervening material with physical properties that can obscure the whole region of the continuum and the BLR at such radii is not obvious. One of the possible candidates for obscuration is the torus that lies just outside of the BLR beyond the dust sublimation radius (e.g., Suganuma et al., 2006; Koshida et al., 2014; LaMassa et al., 2015). However, the sublimation radius is itself 4-5 times larger than the RBLR where the crossing timescale is ∼ 29 years, which is again too long to explain the observed event. Another possible obscuring scenario is that the dimming of the continuum itself promotes the formation of the dust that is able to cover the BLR. However, the timescale for such dust formation with the gas density of ∼ 105 cm−3 in the torus (e.g., Nenkova et al., 2002) is ∼ 103 years (Draine, 2009; LaMassa et al., 2015) that is far too long for this scenario to be true in the case of PS1-13cbe. In addition, if the dust obscuration scenario is true, then the change in luminosity should follow the colour evolution and thus the colour evolution and luminosity change should be correlated. By contrast, there is no colour evolution while the luminosity is changing in the case of PS1-13cbe. We fit a linear regression model to luminosity versus g − r colour index and found a slope of −0.01 ± 0.01 with a coefficient of determination R2 ∼ 0.056. These results confirm the lack of correlation between luminosity and g − r colour index. 77

2.4.3.2 Tidal Disruption Events

As we mentioned in the the Section 2.4.2, the lack of the colour evolution, constant blackbody temperature, and appearance/disappearance of the broad Hα and Hβ lines suggest that the optical outburst and the appearance of the apparent blue continuum may have been powered by a TDE in the pre-existing AGN. Based on the reasons that we provided in the Section 2.4.2, we conclude that TDE origin of the outburst in PS1-13cbe is not favored, but it cannot be ruled out completely.

2.4.3.3 Accretion disk instabilities

In the light curves of PS1-13cbe shown in Figure 2.2, there are small undulations that can be seen in all of the filters (e.g., near −20 days). The amplitudes of these changes are consistent with the variability observed in AGNs (e.g., MacLeod et al., 2012) and point to fluctuations in ongoing accretion activity. It has been shown that in AGNs the optical/UV emission is generated in the accretion disk, with two possible classes of models for the propagation of fluctuations. One is that the locally generated viscous perturbations can produce local blackbody emission (e.g., Krolik, 1999; Liu et al., 2008). In the case of these so-called “outside-in” variations that are produced by changes in accretion rate, the accretion flow fluctuations propagating inward and across the accretion disk first affect the optical region located at outer radii and then affect the UV and X-ray emitting regions. Another origin can be the reprocessing of the UV or X-rays (e.g., Krolik et al., 1991; Cackett et al., 2007; Liu et al., 2008). In this case, the X-rays from the central source irradiate the disk and produce “inside-out” variations from short to long wavelengths (LaMassa et al., 2015; Shappee et al., 2013). The timescale over which the accretion changes happen that might produce “outside- in” variations is known as the inflow timescale. More accurately, changes in the accretion responsible for the “outside-in” variations happen on the inflow time scale, which is the 78 time it takes a parcel of gas in a given radius in the accretion disk to radially move to the center. Assuming the optical continuum emission radius of R ≈ 200 rS (e.g., Morgan et al., 2010; Fausnaugh et al., 2016) and using the relationship between radius and inflow timescale presented by LaMassa et al.(2015), we calculate the inflow timescale to be

6 tinfl ∼ 10 years, which is much longer than the observed change in the continuum flux of PS1-13cbe. However, it is notable that the optical continuum originates in a part of the disk where the radius is an order of magnitude larger than the UV-emitting region, which results in a several thousand times larger inflow timescale than for the UV-emitting region (LaMassa et al., 2015). Thus, we conclude that the rapid continuum flux change in PS1-13cbe is too short to be generated by outside-in variations (perturbations in a given radius of the disk that propagate radially inward), but might be more characteristic of a disturbance in the inner accretion flow that propagates outward. Assuming a standard thin disk model (Shakura and Sunyaev, 1973) with an optically thick and geometrically thin accretion disk, we calculate the dynamical (orbital) time-scale q GM tdyn = 1/Ω of the SMBH where Ω = R3 . We can rewrite the dynamical time-scale to be 3/2 3/2 GM R 2GM tdyn = 2 3 where c is the speed of light and rS = 2 is the Schwarzschild radius c rS c 7 of the central BH. The dynamical timescale around a SMBH with mass of 2.2 × 10 M 3/2 R can be written as tdyn ≈ 310 s. Assuming an optical emission distance of R ≈ 200 rS rS (e.g., Morgan et al., 2010; Fausnaugh et al., 2016), we calculate the dynamical timescales

−1 of ≈ 10 days. Then, using tth = α tdyn we calculate the thermal timescale to be ≈ 99 days. We note that these values are not strongly dependent on the uncertainties in the BH mass for this object, so the ordering of timescales is robust. Among all of the calculated timescales, only the thermal timescale of ∼ 99 days is reasonably similar to the observed timescale of ∼ 70 days in the case of PS1-13cbe and suggests another possible scenario where reprocessing of the UV or X-rays can produce 79 the optical variations. In this case, an increase in the X-rays that originate in the smaller hot corona that is closer to the central BH can heat the inner part of the accretion disk first, move outward, and generate inside-out variations by irradiating the disk and driving an increase in the blue and then red emission. This scenario has been observed and well studied in NGC 2617 (Shappee et al., 2013; Oknyansky et al., 2017). Shappee et al. (2013) detected the variability first in X-rays and then with time lags in UV and NIR and concluded that the observed continuum flux variability resulted from inside-out variations. In another example, NGC 4151, it was also observed that X-ray variability drove UV- optical variability (Edelson et al., 2017).

2.4.4 Comparison to other Changing Look AGNs

PS1-13cbe is one of the few CL AGNs that has been observed during the turn-on phase (Cohen et al., 1986; Storchi-Bergmann et al., 1993; Aretxaga et al., 1999; Eracleous and Halpern, 2001; Shappee et al., 2014; Gezari et al., 2017) by suddenly demonstrating the appearance of broad Hα and Hβ emission lines. The CL behavior has been observed in other candidates where the broad lines appeared or disappeared in spectra taken a couple of years to decades apart (e.g., LaMassa et al., 2015; Runnoe et al., 2016; MacLeod et al., 2018). By contrast, in PS1-13cbe the outburst timescale is very short. PS1-13cbe “turned- on” in only 70 days and the broad lines were observed in a spectrum taken 57 days after the peak of the outburst. Furthermore, the broad lines disappeared before two years after the time of the peak, which is again very short, and they never re-appeared in later spectra (see Figure 2.3). However, we should note that by “turn-on” we refer to brightening of a pre-existing AGN, but the presence of the strong narrow emission lines with a Seyfert 2 spectrum implies that narrow-line region was exposed to AGN accretion activity in the recent past. 80

Gezari et al.(2017) also presented the CL quasar iPTF 16bco that had a “turn-on”

8 timescale of ≤ 1 year which is very short for a BH with MBH ∼ 10 M compared to others. PS1-13cbe and iPTF 16bco are the only CLs that demonstrate extremely short turn- on timescales which push the limit of the accretion disk theory. One more interesting fact about PS1-13cbe is that the light curves showed a double peak behavior observed in all of the optical bands (see Figure 2.2). This behavior was also observed in the X-rays/UV in the case of NGC 2617 (Shappee et al., 2014), in the UV/optical in the case of the ASASSN- 15lh TDE candidate (Dong et al., 2016; Leloudas et al., 2016; Margutti et al., 2017), and in the optical band in the case of the PS16dtm TDE candidate (Blanchard et al., 2017).

2.5 Conclusions

We present a transient event that was discovered in the PS1/MDS survey, PS1-13cbe, at redshift z = 0.12355. The outburst happened in the nucleus of a galaxy that is classified as a Seyfert 2 (see Figure 2.4) using the SDSS archival data that was taken a decade before the outburst. At the time of the outburst, the galaxy changed type to a Seyfert 1 as broad Hα and Hβ appeared and the continuum brightened in the spectrum taken with LDSS3 +57 days after the peak and then changed its type back to a Seyfert 2 as the broad Hα and Hβ disappeared in the spectrum taken with OSMOS 2 years later and did not reappear in spectra taken 3 and 4 years after the outburst. The optical photometry shows that the continuum flux increased by a factor of ∼8 on a timescale of ∼ 70 days and declined for next ∼ 50 days and then rose again over the course of the next ∼ 50 days. Observational evidence presented in this work argues against the Type IIn SN and TDE interpretations. The constant colour evolution and blackbody temperature during the outburst and also the presence of a pre-existing AGN disfavour the SN Type IIn scenario. As mentioned, TDEs have been observed in the galaxies with pre-existing AGNs (e.g., SDSS J0748; Wang et al., 2011); however, the lack of a broad He IIλ4686 emission line, 81 low blackbody temperature at the time of peak, a light curve that has small fluctuations, and unusual re-brightening by 75% are inconsistent with properties of known TDEs. We conclude that PS1-13cbe is a changing-look AGN that has been powered by instabilities in the accretion disk. We argued against the obscuration scenario and TDE origin of these accretion disk instabilities by showing that the crossing and viscous timescales are longer than the timescale observed in the case of PS1-13cbe. Furthermore, we also argued against outside-in variations by calculating the inflow timescale which is too long to explain the observed timescale here. We also conclude that the thermal instabilities in the accretion disk are most likely the source of the outburst and CL behavior. These thermal instabilities may have caused inside-out variations that have generated the observed optical variability in the light curves of PS1-13cbe. One very interesting point about PS1-13cbe is that the observed turn-on timescale pushes the limits of viscous accretion disk theory which predicts much longer timescales and it might be one of the CLs that have shown the most rapid change of the state (iPTF 16bco; Gezari et al., 2017) compared to the other CLs. Other CLs have been observed over a timespan of years and sometimes decades apart; however, the short timescale observed here suggests that other candidates may have been through these short timescale outbursts. Therefore, more frequent observations with a higher cadence and multiwavelength coverage to overcome limitations such as the one we faced here with the lack of X-ray observations at the time of outburst, will provide us more insight to better understand the changing look behavior of AGNs. 82

Table 2.1: Photometry of PS1-13cbe.

Data of Observation Epoch Filter Mag Mag (MJD) (days) (observed) Uncertainty

56217.3 -295.3 gp1 (23.5)

56475.5 -37.1 gp1 19.89 0.03

56480.5 -32.1 gp1 19.79 0.03

56486.6 -26.0 gp1 19.76 0.03

56489.5 -23.1 gp1 19.91 0.03

56501.4 -11.2 gp1 19.79 0.03

56507.6 -5.0 gp1 19.76 0.03

56512.5 -0.1 gp1 19.66 0.03

56518.5 5.9 gp1 19.85 0.03

56531.5 18.9 gp1 20.16 0.04

56534.5 21.9 gp1 20.20 0.04

56539.4 26.8 gp1 20.30 0.04

56545.3 32.7 gp1 20.38 0.04

56548.3 35.7 gp1 20.42 0.06

56563.3 50.7 gp1 20.49 0.05

56571.2 58.6 gp1 20.50 0.04

56574.2 61.6 gp1 20.43 0.06

56590.3 77.7 gp1 20.35 0.04

56593.3 80.7 gp1 20.24 0.04

56596.2 83.6 gp1 20.22 0.04

56601.2 88.6 gp1 20.34 0.07

56628.3 115.7 gp1 20.11 0.17

56285.2 -227.4 rp1 (23.5) 83

Table 2.1: continued

Data of Observation Epoch Filter Mag Mag (MJD) (days) (observed) Uncertainty

56475.6 -37.0 rp1 19.85 0.04

56479.5 -33.1 rp1 19.75 0.04

56486.6 -26.0 rp1 19.66 0.03

56489.5 -23.1 rp1 19.80 0.03

56501.4 -11.2 rp1 19.70 0.03

56507.6 -5.0 rp1 19.64 0.03

56508.5 -4.1 rp1 19.63 0.03

56512.6 0.0 rp1 19.60 0.03

56518.5 5.9 rp1 19.72 0.03

56531.5 18.9 rp1 20.04 0.04

56534.5 21.9 rp1 20.08 0.04

56538.3 25.7 rp1 20.03 0.05

56545.3 32.7 rp1 20.11 0.04

56571.2 58.6 rp1 20.44 0.06

56574.3 61.7 rp1 19.81 0.37

56590.3 77.7 rp1 20.29 0.05

56593.3 80.7 rp1 20.11 0.05

56596.3 83.7 rp1 20.05 0.04

56597.3 84.7 rp1 20.08 0.04

56626.2 113.6 rp1 20.11 0.07

56284.2 -228.4 ip1 (23.5)

56473.5 -39.1 ip1 19.65 0.04

56476.6 -36.0 ip1 19.60 0.04 84

Table 2.1: continued

Data of Observation Epoch Filter Mag Mag (MJD) (days) (observed) Uncertainty

56481.6 -31.0 ip1 19.51 0.05

56484.5 -28.1 ip1 19.42 0.03

56491.5 -21.1 ip1 19.54 0.03

56505.4 -7.2 ip1 19.46 0.03

56508.6 -4.0 ip1 19.44 0.03

56513.6 1.0 ip1 19.44 0.03

56516.5 3.9 ip1 19.46 0.03

56520.5 7.9 ip1 19.54 0.03

56532.4 19.8 ip1 19.79 0.04

56538.4 25.8 ip1 19.72 0.04

56540.4 27.8 ip1 19.70 0.04

56550.5 37.9 ip1 20.18 0.06

56558.4 45.8 ip1 20.07 0.05

56564.4 51.8 ip1 19.96 0.05

56567.2 54.6 ip1 20.10 0.06

56569.3 56.7 ip1 20.02 0.05

56575.3 62.7 ip1 20.01 0.05

56588.4 75.8 ip1 19.79 0.04

56597.2 84.6 ip1 19.88 0.04

56599.3 86.7 ip1 19.89 0.04

56626.2 113.6 ip1 19.85 0.06

56216.2 -296.4 zp1 (23.5)

56453.5 -59.1 zp1 19.86 0.05 85

Table 2.1: continued

Data of Observation Epoch Filter Mag Mag (MJD) (days) (observed) Uncertainty

56456.6 -56.0 zp1 19.71 0.04

56474.6 -38.0 zp1 19.58 0.05

56477.5 -35.1 zp1 19.67 0.04

56480.5 -32.1 zp1 19.44 0.04

56482.5 -30.1 zp1 19.41 0.03

56485.6 -27.0 zp1 19.54 0.04

56488.4 -24.2 zp1 19.55 0.03

56490.5 -22.1 zp1 19.48 0.05

56492.5 -20.1 zp1 19.54 0.04

56506.6 -6.0 zp1 19.53 0.04

56511.6 -1.0 zp1 19.47 0.04

56517.4 4.8 zp1 19.49 0.04

56519.6 7.0 zp1 19.60 0.04

56521.6 9.0 zp1 19.65 0.04

56530.4 17.8 zp1 19.81 0.04

56536.5 23.9 zp1 19.84 0.05

56539.4 26.8 zp1 19.91 0.05

56541.3 28.7 zp1 19.79 0.05

56544.3 31.7 zp1 19.87 0.06

56549.3 36.7 zp1 19.90 0.06

56557.5 44.9 zp1 20.07 0.06

56568.3 55.7 zp1 20.35 0.14

56570.3 57.7 zp1 20.03 0.06 86

Table 2.1: continued

Data of Observation Epoch Filter Mag Mag (MJD) (days) (observed) Uncertainty

56584.3 71.7 zp1 19.92 0.05

56589.3 76.7 zp1 19.94 0.05

56592.3 79.7 zp1 19.87 0.05

56595.2 82.6 zp1 19.84 0.05

56600.3 87.7 zp1 19.81 0.05

56613.3 100.7 zp1 19.87 0.05

56628.2 115.6 zp1 20.03 0.06

56266.3 -246.0 yp1 (23.5)

56434.6 -78.0 yp1 20.63 0.17

56436.6 -76.0 yp1 20.42 0.10

56437.6 -75.0 yp1 20.73 0.29

56454.6 -58.0 yp1 19.70 0.11

56467.5 -45.0 yp1 19.82 0.04

56484.5 -28.0 yp1 19.34 0.05

56493.6 -19.0 yp1 19.50 0.03

56494.6 -18.0 yp1 19.58 0.03

56511.5 -1.0 yp1 19.37 0.03

56517.4 5.0 yp1 19.32 0.06

56522.6 10.0 yp1 19.66 0.04

56532.4 20.0 yp1 19.66 0.04

56556.4 44.0 yp1 19.73 0.03

56559.3 47.0 yp1 19.94 0.09

56562.3 50.0 yp1 19.88 0.11 87

Table 2.1: continued

Data of Observation Epoch Filter Mag Mag (MJD) (days) (observed) Uncertainty

56563.2 51.0 yp1 19.89 0.04

56566.5 54.0 yp1 20.04 0.17

56567.3 55.0 yp1 19.77 0.14

56569.2 57.0 yp1 20.03 0.19

56573.3 61.0 yp1 20.00 0.08

56574.2 62.0 yp1 19.89 0.19

56584.4 72.0 yp1 19.72 0.04

56585.3 73.0 yp1 19.96 0.07

56586.3 74.0 yp1 20.15 0.05

56613.3 101.0 yp1 19.79 0.04

56616.2 104.0 yp1 19.88 0.08

56620.2 108.0 yp1 19.80 0.04

56626.2 114.0 yp1 19.70 0.17

56638.2 126.0 yp1 19.77 0.04

Note: Magnitudes provided here are not reddening corrected and the epochs are relative to the peak in the rest frame. 3σ upper limit values are represented in parentheses. 88

Table 2.2: Luminosity of Broad Lines

Date (UT) Instrument Broad Hα Broad Hβ

2003 June 26 SDSS 4.87 ± 0.37 1.25 ± 0.39 2013 Oct 05 LDSS3 16 ± 0.28 4.39 ± 0.22 2015 Oct 03 OSMOS 3.92 ± 0.45 < 1.2 2016 Nov 16 OSMOS – < 0.85 2017 June 18 OSMOS < 0.86 < 1.14

Note: Luminosity of broad Hα and Hβ lines in SDSS, LDSS3, and OSMOS spectra reported in units of 1040 erg s−1. Non-detections are reported as 3σ upper limits. 89 3 PS1-10cdq

In this chapter, we discuss a transient that was discovered in PS1 survey in 2010. We discuss TDE, SLSN, and AGN activity as the possible sources of this nuclear outburst.

3.1 Observations of PS1-10cdq

On 2010 December 10, we discovered a transient event consistent with the nucleus of a galaxy, PS1-10cdq, at coordinates α = 10h02m29.92s, δ = +01◦30006.100 (J2000) using the Photpipe transient discovery pipeline, described by Rest et al.(2014) and Scolnic et al.(2018). The host galaxy was observed by SDSS and was given the name SDSS J100229.92+013006.1 (hereafter SDSS J1002+013) with a redshift of z = 0.373. The difference image light curve of PS1-10cdq is constructed by the PS1 transient pipeline. A stack of high quality images (excluding the outburst observing season; year 2010) was used to create template images that were subtracted from all of the observations of the transient PS1-10cdq. This flux difference is relative to the host galaxy contribution in the templates.

3.1.1 Optical Photometry

As we show in Figure 3.1, the light curves of PS1-10cdq were constant with zero change relative to baseline in one season before the outburst. Then, the light curves showed a peak at MJD 55545.5 and declined. The luminosity in gp1 band (νLν) increased from a base luminosity of 0.25 × 1044 erg s−1 to a peak luminosity of 2.01 × 1044 erg s−1 in the

44 course of ∼ 200 days (gp1 = 19.50 at peak; see Figure 3.2) and then declined to 0.35 × 10 erg s−1 in the course of ∼ 135 days. The magnitude of these variations are ∼ 100 times larger than the AGN variability of RMS ∼ 0.02 × 1044 erg s−1 observed in other seasons excluding the one that contains the outburst. Also, the transient was observed by Galex in the Near Ultra Violet (NUV) band shown by olive bars in Figure 3.1. 90

2.0 y z i r

) g

−1 1.5 NUV erg s 44 1.0 (10 ν L ν 0.5

0.0 S −1000 −500 0 500 1000 1500 Time (MJD − 55545.5)

Figure 3.1: The observed Galactic extinction corrected luminosities of PS1-10cdq from the

PS1 survey in grizyP1 ﬁlters. S: marks the epoch of the LDSS spectrum (MJD 55626).

3.1.2 Observations of the Host Galaxy

In addition to the PS1 observations of the host in gp1, rp1, ip1, zp1, and yp1 with values of 21.61 ± 0.03, 20.85 ± 0.01, 20.67 ± 0.03, 20.22 ± 0.02, and 19.21 ± 0.22 magnitudes, we found archival observations from XMM, GALEX, SDSS, IRAC, and MIPS that are summarized in Table 3.1.

3.1.3 X-ray Photometry

As we show in Table 3.1, the host galaxy of the PS1-10cdq has been observed with XMM Newton in Cosmic Evolution Survey (COSMOS) survey with total exposure time of 14.8 ks in November 24 2008 (∼ 2 years before the outburst; Cappelluti et al., 2009b).

−14 −2 −1 The host galaxy was detected with fx(0.5 − 2 keV) = (0.46 ± 0.07) × 10 erg cm s . 91

20 z − 4 i − 3

Magnitude + constant r − 2 22 g − 1 NUV

0 20 40 60 80 100 120 140 Time (MJD − 55545.5)

Figure 3.2: Galactic extinction corrected light curves of PS1-10cdq in grizyP1 and Galex NUV bands.

20 −2 Using galactic neutral hydrogen column density of nH ≈ 2.68 × 10 cm (Kalberla et al., 2005) in the direction of PS1-10cdq, and a photon index of Γ = 2, we calculated

−14 −2 −1 the X-ray ﬂux to be fx(2 − 10 keV) = (0.58 ± 0.09) × 10 erg cm s which gives

42 −1 Lx(2 − 10keV) = (2.83 ± 0.44) × 10 erg s . Additionally, using the empirical bolometric

correction from Marconi et al.(2004) and LX, we calculated a bolometric luminosity of

43 −1 Lbol ≈ (3.71 ± 0.41) × 10 erg s .

3.1.4 Optical Spectroscopy

We observed PS1-10cdq using the Inamori Magellan Areal Camera and Spectrograph (IMACS) on the 6.5 m Magellan Baade (MC) telescope for 5400 s on 2011 March 9 (∼ +84 days after the peak of the outburst). We used a 0.900 long slit with the 200-15.0 grism to 92

Table 3.1: Photometry of the host of PS1-10cdq.

Observed Passband Photometry Uncertainty Units No. Measurement

1 2-10 keV (XMM) – < 1.75E-14 erg cm−2 s−1 2 0.5-2 keV (XMM) 0.46E-14 ±0.07E-14 erg cm−2 s−1 3 FUV (GALEX) AB 23.03 ±0.22 mag 4 NUV (GALEX) AB 22.32 ±0.13 mag 5 u (SDSS Model) AB 21.80 ±0.25 asinh mag 6 g (SDSS Model) AB 21.10 ±0.05 asinh mag 7 r (SDSS Model) AB 20.18 ±0.037 asinh mag 8 i (SDSS Model) AB 20.01 ±0.05 asinh mag 9 z (SDSS Model) AB 19.55 ±0.12 asinh mag 10 3.6 microns (IRAC) 69.79 ±0.30 microJy 11 4.5 microns (IRAC) 73.41 ±0.41 microJy 12 5.8 microns (IRAC) 74.67 ±1.44 microJy 13 8.0 microns (IRAC) 136.10 ±2.55 microJy 14 24 microns (MIPS) 522.88 ±11.78 microJy

cover wavelength range of 3900 − 10150 Å with a resolution of ∼ 8Å. We have many late-time observations that we summarize in Table 3.2 and show in Figure 3.3.

3.2 Observational Features of PS1-10cdq

3.2.1 Host Galaxy of PS1-10cdq

We generated a quiescent spectrum by combining three post-outburst spectra of SDSS J100229.92+013006.1 (spectra taken with IMACS on January, March, and May of 2017; 93

Table 3.2: Late time spectra of the host of PS1-10cdq (Walter Baade (WB), and center (CN)).

Date Observed INST EXP Telescope Grism Slit Wavelength Resolution Seconds Range (Å) (Å)

2013 Dec 30 LDSS3 1800 6.5 m MC VPH-all 100 CN 3580 − 10700 ∼ 9 2014 Dec 19 IMACS 1800 6.5 m WB 300-17.5 0.900 top 3700 − 9450 ∼ 5 2015 Feb 18 IMACS 3000 6.5 m WB 300-17.5 0.900 3700 − 9450 ∼ 5 2015 Apr 23 IMACS 3000 6.5 m WB 300-17.5 0.900 3700 − 9450 ∼ 5 2017 Jan 21 LDDS3-C 900 6.5 m MC VPH-all 100 CN 3580 − 10700 ∼ 9 2017 May 1 LDSS3-C 2400 6.5 m MC VPH-all 100 CN 3580 − 10700 ∼ 9 2018 Mar 14 LDSS3-C 3600 6.5 m MC VPH-all 100 CN 3580 − 10700 ∼ 9

see Table 3.2). In order to isolate emission lines and subtract stellar absorption lines, we used this combined spectrum, optical photometry in SDSS griz bands to simulate a model of the host galaxy using the FAST 1.0 code (Kriek et al., 2009). We experimented with diﬀerent initial parameters and were able to generate the best ﬁt using a stellar age of ∼ 109 years and e-folding timescale of τ ≈ 109 years, by assuming the star-formation history to be exponentially declining, the stellar initial mass function from Chabrier(2003), the Bruzual and Charlot(2003) spectral library, a Milky Way dust law (Cardelli et al., 1989), and solar- like metallicity of Z = 0.02. The model of the host galaxy simulated by FAST is plotted alongside the combined spectrum of the host in the Figure 3.4. As we see in Table 3.1, the host galaxy of PS1-10cdq was detected in X-rays that points to presence of an AGN. Additionally, as we show in Figure 3.5, we constructed the SED of the host galaxy using observed values from PS1 and archival photometry data (see Table 3.1 for more details). 94

8 Mar 2018

6 May 2017

Jan 2017

Apr 2015 ) + constant 4 −1 Feb 2015 Å −1

s 2 Dec 2014 −2 Dec 2014

0 Dec 2013 (erg cm λ Dec 2011

Logf −2 Mar 2011 Hβ α Mg II H 3000 4000 5000 6000 7000 8000 Rest frame wavelength (Å)

Figure 3.3: Optical spectra of PS1-10cdq. Spectrum ∼ 84 days after the peak of outburst observed using LDSS3 in dark red. Spectrum taken ∼ +352 days after the peak of outburst in purple. For detailed information about late time spectra shown in navy blue see Table 3.2. All the spectra are binned for the purpose of better visualization.

The MIR, UV, and NUV excess of the host relative to the star-forming template observed in the SED is another evidence of AGN activity.

3.2.2 Multi-band Light Curves of PS1-10cdq

We calculate the total optical luminosity of the PS1-10cdq using multi-band observations (see Figure 3.1). In order to do this, first, we find the epochs where gp1 and rp1 bands were observed simultaneously. Next, using Legendre polynomials, we interpolated for ip1 and zp1 bands at these epochs. Finally, we calculate the total flux which translates to total optical luminosity over 3685 − 8910 Å in the rest frame, by integrating 95

Combined Host Spectrum

0.05 Scaled Host Model Å

0.04 0.03 0.02 ( 0.01

4000 4500 5000 5500 6000 6500 Rest frame wavelength (Å)

Figure 3.4: Spectrum of the host galaxy generated by combining three late time spectra (spectra taken by IMACS on January, March, and May of 2017; see Table 3.2) (blue), the model of the host galaxy simulated with the FAST 1.0 code (Kriek et al., 2009) (orange)

the SED at each epoch using the trapezoidal rule. As we show in Figure 3.6, the total optical luminosity of PS1-10cdq increased from the baseline value of (0.37 ± 0.03) × 1044 erg s−1 to (1.65 ± 0.07) × 1044 erg s−1 in the course of <∼ 200 days and then declined to (0.43 ± 0.01) × 1044 erg s−1 in the course of ∼ 135 days. Furthermore, we ﬁt power-law and black body models in order to study the temperature and color evolution while staying agnostic on the SED of the transient. We also calculated g − r color and show it alongside the temperature and spectral index in

Figure 3.7. We use gp1 and rp1 ﬁlters because they are taken on the same nights in PS1; 96

Sb 44 10 Sy2 data )

− Outburst

(ergs 43

ν 10 ν Log

42 10

12 13 14 15 16 17 18 10 10 10 10 10 10 10 Log ν (Hz)

Figure 3.5: SED of SDSS J100229+0130 in quiescence using observed and archival data (purple circles) (see Section 3.1.2), SED of PS1-13cbe at outburst from PS1 survey (green stars), scaled SED templates of an Sb spiral galaxy (dotted orange line) and a Seyfert 2 galaxy (dashed blue lines) from the SWIRE template library (Polletta et al., 2007). The observed MIR excess in the host relative to the star-forming template is due to an AGN.

therefore no interpolation is needed. As we show in Figure 3.7, there is a clear color and temperature evolution during the outburst.

3.2.3 Spectral Features of PS1-10cdq

In Figure 3.3, we show the optical spectra of PS1-10cdq that demonstrate signiﬁcant spectral evolution over the course of nine years. The spectra include narrow and broad emission line proﬁles, such as hydrogen Balmer lines, Mg II λλ2800, [S II]λλ6717, 6731, 97

2.0

1.5 erg/s) 44

1.0

Luminosity (10 0.5

0.0 −400 −200 0 200 400 Time (Rest frame days after peak)

Figure 3.6: Total optical luminosity light curve of PS1-10cdq including the baseline that was calculated by integrating over grizp1 bands using trapezoidal method at each epoch.

[O III]λλ4959, 5007, [O II]λ3726, [O I]λ6300 and [N II]λλ6549, 6583. Our spectra were taken with different spectral resolutions and effective apertures (slit and seeing). In order to investigate the presence of broad Mg II λλ2800 doublet, Hα and Hβ lines, we fit the Hα+[N II] and Hβ+[O III] complexes and Mg II λλ2800 doublet. First, we scaled our spectra using the flux of [O III] λ5007 from a spectrum that we generated by combining the latest three spectra taken by LDSS3 as described in Table 3.2. We did these scalings using the assumption that these narrow lines are concentrated in the central region of the host and that the fluxes of these narrow lines do not change in these short timescale because it has been shown that narrow emission lines vary slowly over decades (Peterson et al., 2013). Additionally, we subtracted the host galaxy model that we discussed in Section 3.2.1. 98

K) 12 3 10 8 griz 6 griz+NUV 4 Temperature (10 0 −1 −2 index Spectral −3 −4 0.6 0.5 0.4 0.3 0.2

g−r (mag) 0.1 0.0 −0.1 0 20 40 60 80 100 Time (Rest frame days after peak)

Figure 3.7: Top: Rest-frame blackbody temperature from ﬁtting the optical photometry.

α Middle: Rest-frame spectral index from ﬁtting a power law fν ∝ ν to the optical photometry, where α is the spectral index. Red squares: represent ﬁts where in addition to PS1 optical data, NUV data points from Galex are available. Bottom: observed g-r colour diagram. Strong evolution is evident in all panels.

Next, in order to reduce the number of free parameters, we fit the [S II] lines using single Gaussian profiles and use this model to constrain the multi-component Gaussian profiles that we used to fit the Hα+[N II], and Hβ+[O III] (Ho et al., 1997). Specifically, in the case of the spectrum taken closest to outburst in March 2011 (shown in red in Figure 3.3), we model emission lines using one Gaussian profile for each narrow line and one broad component for Hα and simultaneously fit for parameters of the broad and and narrow lines. We fix the narrow components to the wavelengths of Hα,Hβ, [N II] λλ6549, 6583, and [O III]λλ4959, 5007. Also, we fix the widths and relative amplitudes 99 of the Gaussian component of the narrow lines using the values from the [S II] model, leaving only the overall amplitudes of the narrow Hα, and [N II] lines as free parameters. However, in the case of [O III] and Mg II lines, we let all the parameters be free. The parameters of the broad component of Hα were allowed to vary freely. For Hβ, however, we fix the centroid of the broad line and only let the width component vary. Also, in the case [S II] profile was not available, we let the centroid of both narrow and broad line to vary, while in the later spectra we fix the centroid and width of the broad components of Hα and Hβ lines and only let the amplitude vary. In the case of Mg II λλ2800 doublet we used a single Gaussian. 100

0.06 Data Fit Broad Line Line Individual Individual

0.05 Gaussians Gaussians 2011-03-09 2013-12-30

0.04 +84 d +1110 d

0.03

0.02 ( ( 0.01 f f 0.00 −

6400 6450 6500 6550 6600 6650 6700 6400 6450 6500 6550 6600 6650 6700 Rest wavelength () Rest wavelength ()

(a) (b)

Line Individual

Gaussians

2015-02-18 +1526 d

( ( f f − −

− 6400 6450 6500 6550 6600 6650 6700 Rest wavelength ()

Data Fit 0.04 Broad Line Individual Gaussians 2017-01-21 0.03 +2228 d

0.02 0.01 ( ( f f 0.00 −

− 6400 6450 6500 6550 6600 6650 6700 Rest wavelength ()

(e) (f)

0.04 Data Fit Broad Line

Individual Gaussians

0.03 2017-05-01 +2328 d 0.02

0.01 ( ( f f 0.00 −

−

− − 6400 6450 6500 6550 6600 6650 6700 Rest wavelength ()

(g) (h)

Figure 3.8: Continuum subtracted Hα+[N II] line profiles. Top: Multiple-component Gaussian fit to the Hα+[N II] emission lines (blue), individual components (dashed green) and broad component of the Hα (dashed black). Bottom: The fit residuals. Top-right (on each panel): Observation date and numbers of days after the peak (black). 102

0.035 Data Fit 0.030 Individual Gaussians

Broad Line Line

0.025 2011-03-09 2011-12-30

+84 d +380 d 0.020 0.015

0.010

0.005 ( (

f f 0.000 −

4800 4825 4850 4875 4900 4925 4800 4825 4850 4875 4900 4925 Rest wavelength () Rest wavelength ()

(a) (b)

Line Line 2013-12-30 2014-12-19

+1110 d +1464 d ( ( f f

− −

4800 4825 4850 4875 4900 4925 4800 4820 4840 4860 4880 4900 4920 4940 Rest wavelength () Rest wavelength ()

0.020 Data Fit

Individual Gaussians

0.015 Line Broad Line 2015-02-18 2015-04-23

+1526 d +1590 d

0.010 0.005 ( ( f − f 0.000

−

0.01 0.00

Residual

4800 4820 4840 4860 4880 4900 4920 4940 4800 4825 4850 4875 4900 4925 Rest wavelength () Rest wavelength ()

(e) (f)

Line Line 2017-01-21 2017-05-01 +2228 d +2328 d ( ( f f − −

4800 4825 4850 4875 4900 4925 4800 4825 4850 4875 4900 4925 Rest wavelength () Rest wavelength ()

(g) (h) 104

Data 0.020 Fit Individual Gaussians

Broad Line 0.015 2018-03-14

+2646 d 0.010 0.005 ( f 0.000

4800 4825 4850 4875 4900 4925 Rest wavelength ()

(i)

Figure 3.9: Continuum subtracted Hβ line profile. Top: Multiple-component Gaussian fit to the Hβ emission line (blue), individual components (dashed green) and broad component of the Hβ (dashed black). Bottom: The fit residuals. Top-right (on each panel): Observation date and numbers of days after the peak (black). 105

Table 3.3: Luminosity of Broad Lines

Date (UT) Instrument Broad Hα Broad Hβ

2011 Mar 09 IMACS 3.37 ± 0.43 1.91 ± 0.26 2011 Dec 30 IMACS – 0.63 ± 0.10 2013 Dec 30 LDSS3 2.02 ± 0.43 0.16 ± 0.07 2014 Dec 19 IMACS 2.01 ± 0.48 0.49 ± 0.21 2015 Feb 18 IMACS 1.43 ± 0.45 < 0.72 2015 Apr 23 IMACS 1.29 ± 0.16 0.30 ± 0.06 2017 Jan 21 LDSS3-C 0.51 ± 0.17 < 0.30 2017 May 01 LDSS3-C 1.27 ± 0.18 0.21 ± 0.05 2018 Mar 14 LDSS3-C 0.91 ± 0.39 0.19 ± 0.08

Note: Luminosity of broad Hα and Hβ lines in the spectra of PS1-10cdq reported in units of 1041 erg s−1. Non-detections are reported as 3σ upper limits. 106

0.008 Data Fit 2011-12-30 +380 d

0.006

0.004

0.002 ( f 0.000

2760 2780 2800 2820 2840 2860 Rest wavelength ()

Figure 3.10: Continuum subtracted Mg II λλ2800. Top: Single Gaussian ﬁt to the Mg II (blue). Bottom: The ﬁt residuals. Top-right: Observation date and numbers of days after the peak (black).

As we show in Figure 3.8 and Figure 3.9, broad components of Hα and Hβ emission lines with FWHM of 3609 ± 262 and 3516 ± 286 km s−1, respectively, clearly exist in the spectra taken close to the time of outburst (strong broad lines are shown in panels (a) and 107

(b)). We also measured a full proﬁle FWHM of 686±19 km s−1 for Hα emission line at this epoch. Furthermore, as we show in Figure 3.10, we detected a broad Mg II doublet in the spectrum taken +380 days after the outburst (shown in purple in Figure 3.3) with FWHM of 3702 ± 216 km s−1. Also, as we show in Figure 3.8 and Figure 3.9, weak and broad components of Hα and Hβ emission lines exist in later spectra. The full proﬁle FWHM of 686 ± 19 km s−1 for Hα emission line, presence of Fe II features in the spectra (see Figure 3.11), and the fact that the host galaxy was detected in X-rays by the COSMOS survey (see Section 3.1.3) suggests that the host is a NLS1 galaxy that hosts an AGN. We report the measured luminosities of broad components of Hα and Hβ emission lines in Table 3.3.

3.3 Interpretation of Features of PS1-10cdq

Distinguishing the nature of the outburst in the cases where the nuclear outburst occurs in a galaxy that host an AGN is very diﬃcult because the emission from the pre- existing AGN is always present and can contaminate the observations. The variability of the AGN itself could also be the origin of the observed nuclear outburst. The TDE candidate PS16dtm (Blanchard et al., 2017) and extremely luminous Type IIn/TDE candidate CSS100217 (Drake et al., 2011; Blanchard et al., 2017) are examples of such outbursts that happened in NLS1 galaxies that host AGN very similar to PS1-10cdq. In Figure 3.11, we show the light-curve of PS1-10cdq alongside CSS100217 and PS16dtm. Both of these objects show broad components of Hα,Hβ, and Mg II (only in the case of PS16dtm) in their spectra. In Figure 3.12 we also show the host subtracted spectrum of the PS1-10cdq at the time of outburst alongside the spectra of the CSS100217 and PS16dtm close to the peak of their outbursts. 108

In this section, we discuss AGN variability, TDE, and SNe as three possible scenario for the outburst observed in PS1-10cdq. However, we provide details about these scenarios, we discuss summarized key features of PS1-10cdq:

• PS1-10cdq occurred in the center of a NLS1 galaxy.

• PS1-10cdq brightened from a baseline total optical luminosity of (0.37 ± 0.03) × 1044 erg s−1 to (1.65 ± 0.07) × 1044 erg s−1 in the course of ∼ 200 days and then declined to (0.43 ± 0.01) × 1044 erg s−1 in the course of ∼ 135 days.

• The host galaxy of PS1-10cdq was detected in X-rays by XMM-Newton COSMOS

43 −1 survey with a bolometric luminosity of Lbol ≈ (3.71 ± 0.41) × 10 erg s .

• PS1-10cdq showed clear temperature and color evolution where the temperature declined while the g − r colour became redder.

• Spectra of PS1-10cdq show signiﬁcant evolution over the course of ∼ 9 years, where broad components of Hα, and Hβ disappear or become weaker.

3.3.1 PS1-10cdq as a Variable AGN

The host galaxy of the PS1-10cdq is a NLS1 galaxy that contains an AGN. Therefore, it is possible that the PS1-10cdq outburst is caused by an accretion event surrounding the central SMBH. However, we disfavor the AGN variability scenario because of the following reasons. The magnitude of the variations in the outburst season are ∼ 100 times larger than the AGN variability of RMS ∼ 0.02 × 1044 erg s−1 observed in other seasons excluding the one that contains the outburst. Furthermore, assuming that the Hα and Hβ lines are generated at BLR, using the measured Hβ FWHM∼ 3520 ± 290 km s−1, intrinsic

42 −1 luminosity at 5100Å (λL5100) of (6.2±0.2)×10 erg s at the time of the spectrum, and the mass-scaling relationship of Woo et al.(2015), we calculate the mass of central SMBH to be 109

− Mr MV

−

− −

−

Figure 3.11: Light-curve of PS1-10cdq (blue). Light curves of the comparison objects extremely luminous SLSN Type IIn/TDE candidate CSS100217 (Drake et al., 2011)(red), and TDE candidate PS16dtm (Blanchard et al., 2017)(green).

7 46 (8.3±1.35)×10 M . The Eddington luminosity for this BH mass is LEdd = (1.±0.2)×10

−1 erg s . We also calculate the intrinsic luminosity at 5100Å (λL5100) for PS1-10cdq that is (1.56 ± 0.02) × 1044 erg s−1 at the time of the peak and then we estimate the bolometric

luminosity using λLλ and a conversion factor of 8.1 (Runnoe et al., 2012) to convert from monochromatic to bolometric luminosity to be (1.26 ± 0.02) × 1045 erg s−1. This results in an Eddington parameter of λEdd = Lbol/LEdd ≈ 0.1 at the peak of the outburst. Assuming a standard thin disk model (Shakura and Sunyaev, 1973) with an optically thick and geometrically thin accretion disk, we calculate the dynamical (orbital) time- 3/2 3/2 GM R 2GM scale to be tdyn = 2 3 where c is the speed of light and rS = 2 is the c rS c 110

CSS100217 +84d PS16dtm H +87d H Fe II Fe II He I

( -10cdq +84d H 4000 4500 5000 5500 6000 6500 7000 7500 RestÅ

Figure 3.12: Optical spectrum of PS1-10cdq taken with IMACS +84 days after the outburst after subtracting the stellar continuum (navy). Optical spectrum of TDE candidate PS16dtm +87 days after the outburst (Blanchard et al., 2017)(red). Optical spectrum of extremely luminous SLSN Type IIn/TDE candidate CSS100217 +84 days after the outburst (Drake et al., 2011)(purple). The emission lines are labeled (black)

Schwarzschild radius of the central BH. The dynamical timescale around a SMBH with 3/2 7 R mass of 8.3 × 10 M can be written as tdyn ≈ 1520 s. Assuming an optical emission rS distance of R ≈ 53 rS (e.g., Morgan et al., 2010; Fausnaugh et al., 2016), we calculate

−1 the dynamical timescales of ≈ 5 days. Then, using using tth = α tdyn we calculate the thermal timescale to be ≈ 52 days. Next, we calculate the viscous timescale to be

−3 tvisc = 0.1 tdyn ≈ 14 years. These timescales are either much shorter or longer than the observed timescale in PS1-10cdq. 111

The observed variability can be also caused by obscuring material outside of the BLR moving in or out of the line of sight. However, using the R-L relation calibrated presented by (Bentz et al., 2013), we estimate characteristic radius of BLR to be RBLR ∼ 43 light days. Then, using the relation for crossing time presented by LaMassa et al.(2015), we estimate the crossing time for an obscuring object orbiting on a Keplerian orbit outside of BLR with the extreme case of assuming rorb = RBLR where the crossing time for the obscuring object is ∼ 33 years which is much longer than observed timescale for the outburst of the PS1-

10cdq. In more realistic cases where the obscuring material are at rorb ≥ 3RBLR the crossing timescales are even longer. Additionally, Graham et al.(2017) show that the rate of these extreme AGN variability in the AGN population is ∼ 10−5 yr−1 sr−1 which is statistically very low. Based on the optical and spectral features of the PS1-10cdq and the reasons provided in this section we disfavor the AGN variability scenario.

3.3.2 PS1-10cdq as SLSN

As we show in Figure 3.12, the presence of the broad H balmer lines, blue continua, and lack of P Cygni proﬁles in the spectrum is similar to Type IIn SNe (Filippenko, 1989). Additionally, in Figure 3.7, we show that the temperature of the PS1-10cdq is cooling and there is a color evolution that is consistent with the cooling of the ejecta commonly observed in SNe. The SLSN Type IIn origin of PS1-10cdq is however disfavoured based on the following reasons. First, the fact that the host galaxy is NLS1 and that the outburst happened close to the nucleus increases the chances of PS1-10dq being related to activity near the central SMBH. Additionally, the outburst dimmed to the baseline in the course of ∼ 100 days which is very rapid for SLSN IIns that last several hundred days. Furthermore, the FWHM of broad components of the Hα and Hβ emission lines are of the order of 103 112

km s−1, while the broad components reported in the case of Type IIn SNe are of the order of 104 km s−1 and stronger than Balmer lines in the PS1-10cdq (Chugai and Danziger, 1994). On the other hand, we used the MOSFiT CSM interaction model (Guillochon et al., 2018,

and references therein) to ﬁt the grizyP1, and Galex NUV bands. We only limited explosion

time to texp > −190 days before the ﬁrst detected datapoint because at ∼ −220 days the event was not detected. While the model converges with PS RF ∼ 1.99 which is less than the assumed convergence criteria PS RF ≤ 1.20, it has a hard time ﬁtting the observed light-curves especially the NUV band as seen in Figure 3.13. As we see Figure 3.13, the estimated peak of the outburst has high variance and some models estimate reasonable

peak values while others estimate values like Mr ∼> −24 which is brighter than the most luminous SLSN candidates ever observed (Dong et al., 2016; Drake et al., 2011; Benetti et al., 2014; Miller et al., 2008; Vreeswijk et al., 2014). In Figure 3.14, we show the corner plot of the posterior distributions of the model parameters which suggests that the MCMC process of MOSFiT needs to be initialized using more walkers and run for more iterations to better reconstruct the posterior distributions of the parameters. For our simulation, we only used 100 walkers and only ran for 5000 iterations. The results in Table 3.4 are comparable to other CSI powered SNe. The model

4 5 estimates the explosion generated ejecta mass 0.3 − 18 M with ejecta velocity of 10 − 10

−1 km s that interacts with CSM mass of 3 − 11 M which is in the typical CSM mass range

of 0.1 − 10 M (Branch and Wheeler, 2017). The CSM model estimates that explosion happened somewhere between −154 and −42 days before the ﬁrst observed data-point in the light-curve which leads to the large uncertainty in estimating the peak luminosity and time of the outburst. These high uncertainties in the estimated values of the parameters indicates that the CSM model cannot reconstruct the light-curve behaviour of PS1-10cdq which reduces the possibility of a SLSN IIn origin. 113

Based on the reasons provided in this section, we disfavour SLSN IIn origin of PS1- 10cdq; however, this possibility cannot be completely ruled out.

Figure 3.13: light curves of PS1-10cdq in grizyPs and NUV GALAX bands (colored circles). Fits generated by MOSFiT’s CSM interaction model (colored lines).

3.3.3 PS1-10cdq as a TDE

As we mentioned before, the host galaxy of PS1-10cdq in a NLS1 galaxy with an AGN present and it has been suggested that TDEs may prefer galaxies with a pre-existing accretion disk around their SMBH (Hills, 1975; Blanchard et al., 2017) because the dense star-formation clouds and the pre-existing accretion disk around the central SMBH of these galaxies can increase the chance of the tidal encounter of the stars with the holes and also reduce the relaxation time (Perets et al., 2006; Blanchard et al., 2017). For example, weak 114

+0.19 log MCSM = 0.85 0.30

+0.69 log Mej (M ) = 0.56 1.00

) 1.0 M (

j 0.5 e M

g 0.0 o l

0.5 +0.95 log nH, host = 17.56 0.92

19.2 t s o h ,

H 18.4 n g

o 17.6 l

16.8 +0.60 log = 13.21 0.36

12.0

g 12.5 o l

13.0

13.5 +0.00 log Tmin (K) = 4.00 0.00 ) K (

n 3.9985 i m T

g 3.9970 o l

3.9955 +71.37 texp (days) = 114.72 37.53

3.9940

) 40 s y a 80 d ( p x e

t 120

160 +0.05 log = 0.44 0.04

0.35

0.40 g o l

0.45

0.50 1 +0.24 log vej(km s ) = 4.56 0.41 ) 1

s 4.75 m k (

j 4.50 e v

g 4.25 o l

4.00

0.5 0.0 0.5 1.0 80 40 0.25 0.50 0.75 1.00 16.8 17.6 18.4 19.2 13.5 13.0 12.5 12.0 160 120 0.50 0.45 0.40 0.35 4.00 4.25 4.50 4.75 3.9940 3.9955 3.9970 3.9985 1 log MCSM log Mej (M ) log nH, host log log Tmin (K) texp (days) log log vej(km s )

Figure 3.14: Corner plot showing the posterior distributions of parameter realizations for MOSFiT’s CSM model. The 16th, 50th, and 84th quantiles of the posterior distribution of the model parameters (dash black lines). Median values with lower and upper error values of the model parameters are reported in the title of histograms (red).

and narrow emission lines have been detected in the spectra of the hosts of the optical/UV TDE SDSS J0748 (Wang et al., 2011) and optical TDE candidate PS16dtm (Blanchard 115

Table 3.4: Model parameters from ﬁtting the CSM model from MOSFiT to the light-curve of PS10cdq.

Parameters Values

+0.19 log MCSM (M ) 0.85−0.30 +0.69 log Me j (M ) 0.56−1.00 −2 +0.95 log nH,host cm 17.56−0.93 −3 +0.6 log ρCSM (g cm ) −13.2−0.36

log Tmin (K) 4.00

+71.37 texp (days) −114.72−37.53 −1 +0.24 log ve j (km s ) 4.56−0.41

et al., 2017) that are associated with the pre-existing AGN in the central environment. Therefore, PS1-10cdq can result from interaction of incoming stellar debris with pre- existing AGN accretion disk. Similarly, the accretion of the stellar debris could increase the UV and optical emission and provide the energy to excite the BLR region. Additionally, the suppression of the NUV compared to optical light indicates an absorption mechanism that is likely wavelength dependent which suggests the presence of an optically thick absorbing layer. This absorption is likely related to the circularized stellar debris of a TDE which can have a simple solution in the optimal scenario (Loeb and Ulmer, 1997). However, the exact properties of this absorption depend on the parameters of the disruption and can be difficult to extract (Guillochon et al., 2014; Roth et al., 2016). Despite all of these complications, we fitted the multi-band light curve of PS1-10cdq using the TDE model of the MOSFiT software (Guillochon et al., 2018). We only limited the first fall-back time to be t f irst f all−back = −190 days before the first observed outburst 116

datapoint and ∼ 20 days after the last non-detection season datapoint. We let the other parameters vary and set the convergence property to be PS RF ≤ 1.2 with 100 walkers initialized and run the model for 5000 iterations. The MOSFiT TDE model converged with PS RF = 1.2 and as we show in Figure 3.15, it generates a good ﬁt for all of the bands including Galex NUV which means that it also accounts for the NUV suppression. The TDE model of MOSFiT package assumes that the photosphere radius has a power law

l dependence on luminosity Rph = Rph0 ap (L/Ledd) where ap is is the semi-major axis of the accreting mass at the peak of outburst. The model also estimates a viscous timescale which includes the time delay eﬀects of the circulization of the stellar debris and the accretion disk surrounding the SMBH (Mockler et al., 2019). Additionally, we show the best parameters of the model after the convergence in Table 3.5 that have less uncertainty compared to the parameters generated by CSM model (see Table 3.4), and the corner plot of the posterior distribution of the model parameters in Figure 3.16. These posteriors are very dispersed and indicate the need to initialize MCMC with higher number of walkers and run the simulation for more iterations. The parameters

+3.15 in Table 3.5, correspond to tidal disruption of a 11.12−3.01 M star by a SMBH with mass of 7 +0.632 ∼ 4×10 M with Tviscous = 0.060−0.004 days. The estimated SMBH mass is consistent with 7 the SMBH mass of (8.3 ± 1.35) × 10 M that is estimated by the mass-scaling relationship of Woo et al.(2015) in Section 3.3.1. This mass is on the heavy side of the samples studied in Mockler et al.(2019) and is similar to the BH mass estimated for PS1-11af (Chornock

+9.68 et al., 2013). The disruption of the star happened at t f irst f all−back = −185.15−3.57 days before +0.12 the ﬁrst detection. The model estimates an impact parameter of β = Rt/Rp = 0.77−0.13 which is deﬁned as the ratio of the tidal radius and pericenter radius of the star. The value of the β indicates a partial disruption of the encountering star. Furthermore, the spectra of some known TDEs show broad Hα,Hβ, and He II emission lines (Arcavi et al., 2014; Van Velzen et al., 2011). As we show in Figure 3.8 117

and Figure 3.9, broad Hα and Hβ emission lines have been detected in the spectra taken close to the time of the peak (see panels (a) and (b)) that either were weakly present or not present in the late spectra. Also, a Mg II line is present in the spectrum taken +380 days after the outburst (shown in purple in Figure 3.3) that is present in the spectra of TDE candidate PS16dtm (Blanchard et al., 2017) and the broad components of the Balmer emission lines have similar FWHM of the order of 103 km s−1 to optical TDE candidate PS16dtm. Additionally, as we show in Figure 3.12, the spectrum of the PS1-10cdq taken +84 days after the ﬁrst detection has similar spectral features to the spectra of PS1-16dtm and CSS100217 taken at similar times which are both interpreted as TDEs by Blanchard et al.(2017). However, there are reasons that can be against TDE scenario that we discuss here. The blackbody temperature of optically-selected TDEs typically range between ∼ 20000 K (e.g., PS1-11af; Chornock et al., 2013) and ∼ 35000 K and even higher (e.g., ASASSN- 14li; Holoien et al., 2015), while in the case of PS-10cdq the continuum at the time of the outburst is not as blue as an optical TDE, and as we show in Figure 3.7, the measured blackbody temperature of PS1-10cdq is < 12000 K. However, as we see in in Figure 3.15 the NUV is suppressed, which indicates that some of the optical radiation may also be absorbed by the circularized optically thick stellar debris. Additionally, there is a clear color evolution and temperature cooling shown in Figure 3.7 which is not typically seen in optical TDEs that lack color evolution and show constant blackbody temperature. However, we should mention that there are TDE candidates that show color evolution (e.g., Holoien et al., 2019). Furthermore, the He II λ4686 line which is typically detected in the spectra of TDEs (e.g., ASASSN-14li; Holoien et al., 2015) is not detected in the spectra of PS1- 10cdq. However, the He II λ4686 line could be blended with Fe II lines and also some TDEs have been detected that lack the He II λ4686 line in their spectra. Therefore, reasons 118

provided here cannot strongly rule out TDE scenario and we believe that TDE is a possible origin of the outburst in PS1-10cdq.

Figure 3.15: Light curves of PS1-10cdq in grizyP1 and NUV Galex bands (colored circles). Fits generated by MOSFiT’s TDE model (colored lines).

3.4 Conclusions and Future Remarks

In this chapter, we present a transient event discovered by PS1/MDS survey, PS1- 10cdq, at the center of a NLS1 galaxy at redshift z = 0.373. The light-curve of PS1-1cdq brightened to total optical luminosity of (0.37 ± 0.03) × 1044 erg s−1 to (1.65 ± 0.07) × 1044 erg s−1 in the course of ∼ 200 days and then declined to (0.43 ± 0.01) × 1044 erg s−1 in the course of ∼ 135 days. The spectra of PS1-10cdq shows strong evolution where strong 119

+0.34 log Rph0 = 2.09 0.24

+1.06 log Tviscous = 1.22 1.23 1

s 0 u o c s i v 1 T g o l 2

+0.12 b (scaled ) = 0.77 0.13

1.05 )

d 0.90 e l a c

s 0.75 ( b

0.60 +0.07 log Mh(M ) = 7.58 0.07

7.7 )

M 7.6 ( h M

g 7.5 o l

7.4 +0.23 log = 1.85 0.22

1.50

g 1.75 o l

2.00

+0.31 2.25 phot. exp. 0l0 = 3.60 0.45 0 l 0

. 3.6 p x e

. 3.2 t o h p 2.8 +0.03 log nH, host = 21.43 0.04 t

s 21.45 o h , H

n 21.40 g o l 21.35

+3.15 21.30 starmass = 11.12 3.01

15.0 s

s 12.5 a m r

a 10.0 t s 7.5

+9.68 5.0 texp(days) = 185.15 3.57

160 ) s y

a 168 d ( p x

e 176 t

184 +0.04 log = 0.66 0.04

0.60 g

o 0.64 l

0.68

0.72 2 1 0 1 1.6 2.0 2.4 2.8 7.4 7.5 7.6 7.7 2.8 3.2 3.6 5.0 7.5 0.60 0.75 0.90 1.05 2.25 2.00 1.75 1.50 10.0 12.5 15.0 184 176 168 160 0.72 0.68 0.64 0.60 21.30 21.35 21.40 21.45

log Rph0 log Tviscous b (scaled ) log Mh(M ) log phot. exp. 0l0 log nH, host starmass texp(days) log

Figure 3.16: Corner plot showing the posterior distributions of parameter realizations for MOSFiT’s CSM model. The 16th, 50th, and 84th quantiles the posterior distribution of the model parameters (dash black lines). Median values with lower and upper error values of the model parameters are reported in the title of histograms (red).

broad Hα emission lines are detected +84 days after the ﬁrst detection and disappear or are weakly present in the later obtained spectra (see Tabel3.3 and Figure 3.3). 120

Table 3.5: Model parameters from ﬁtting the TDE model from MOSFiT to the light-curve of PS10cdq.

Parameters Values

+0.34 log Rph0 (photosphere power law constant) 2.09−0.24 −1.23 log Tviscous (days) −1.22+1.06 +0.12 b (scaled β) 0.77−0.13 +0.07 log MBH (M ) 7.58−0.07 +0.23 log (eﬃciency) −1.85−0.22 +0.31 l (photosphere power law exponent) 3.60−0.45 −2 +0.03 log nH,host (cm ) 21.43−0.04 +3.15 M∗ (M ) 11.12−3.01 +9.68 t f irst f all−back (days) −185.15−3.57

The observational evidence presented in this chapter is against the SLSN Type IIn and AGN variability. The presence of broad Balmer emission lines, temperature cooling, and presence of color evolution in the light-curve of PS1-10cdq are in favor of SLSN Type IIn origin. However, the broad components of emission lines are not as strong as typically observed in SLSN IIn candidates. Furthermore, we used the CSM model of MOSFiT package to ﬁt the light-curve of the PS1-10cdq. As we presented in Section 3.3.2, the model does not reconstruct the light-curve behavior of the PS1-10cdq very well and estimated parameters have large uncertainties (see Table 3.4). Finally, the fact that the outburst occurred in the center of a NLS1 galaxy with pre-existing AGN increases the chances of outburst being related to activities in the environment of central SMBH. However, in Section 3.3.1, we argue against the AGN variability as the source of the outburst by 121 measuring the timescales related to accretion instabilities that are too long or too short to explain the timescale observed in PS1-1cdq. We conclude that PS1-1cdq is powered by the tidal disruption of a star. The pre- existing AGN increases the chances of the encounter of pre-exisiting AGN with stars in the vicinity of the BH. Also, the observed appearance and disappearance of Balmer emission lines is typically observed in optical TDEs. Furthermore, the suppression of the NUV compared to optical light points to presence of an optically thick absorbing layer that is also observed in the TDE candidate PS16dtm. Moreover, we reasonably reconstructed the light-curve behaviour of PS1-10cdq using TDE model from MOSFiT package with lower uncertainties compared to CSM model (for more details see Section 3.3.3). While there are some reasons that could argue against TDE, they are not strong enough to rule TDE origin out; therefore, we prefer a TDE as the origin of the PS1-10cdq. However, as we show in Figure 3.14 and Figure 3.16, the posterior distributions of the estimated parameters by CSM and TDE models from MOSFiT are highly dispersed. This high dispersion suggests that a larger number of walkers are needed and that the model should be run for a higher number of iterations since we only ran the model with 100 walkers and only for 5000 iterations. In the future, we would like to run the models for higher number of iterations with more walkers initialized to be able to better assess the SLSN Type IIn and TDE origins of PS1-10cdq. 122 4 Galaxy morphology prediction using capsule networks

Originally published in (Katebi et al., 2019b):

Monthly Notices of the Royal Astronomical Society, Katebi, R., Zhou, Y., Chornock, R., and Bunescu, R. (2019). Galaxy morphology prediction using capsule networks. Monthly Notices of the Royal Astronomical Society, 486(2), 1539-1547. In this chapter, we studied the performance of Capsule Network, a recently introduced neural network architecture that is viewpoint invariant and spatially aware (see Section A.2 for more details.), on the task of galaxy morphology classification. We designed two evaluation scenarios based on the answers from the question tree in the Galaxy Zoo project. In the first scenario, we used Capsule Network for regression and predicted probabilities for all of the questions. In the second scenario, we chose the answer to the first morphology question that had the highest user agreement as the class of the object and trained a Capsule Network classifier, where we also reconstructed galaxy images. We achieved promising results in both of these scenarios. Automated approaches such as the one introduced here will play a critical role in the upcoming large sky surveys.

4.1 Introduction

Morphological classiﬁcations have been used by astronomers to classify galaxies based on their visual aspects such as size and shape. Studying morphological classiﬁcations is crucial to understand the evolution of galaxies by redshift (e.g., Maddox et al., 1990) and their properties such as interaction with other galaxies. Also, morphological studies provide valuable information about the dynamical history of the galaxies without the need of expensive spectroscopy. All-sky surveys are the key solutions to probe galaxy formation and evolution. 123

In order to conduct these studies, observation of a large number of galaxies and determination of their morphological classification is crucial. Large sky surveys such as the Sloan Digital Sky Survey (SDSS) (e.g., Blanton et al., 2017) provided a large amount of data for the objects in our universe including galaxies. The morphological classification of galaxies has been traditionally done by experts, which is both inefficient and impractical for the large datasets available from current sky surveys and even larger upcoming ones such as the Large Synoptic Survey Telescope (LSST) (Ivezic et al., 2008). The Galaxy Zoo project (Lintott et al., 2008) started with the hope of partially solving this problem by a crowdsourcing method. The project was very successful and ∼ 900, 000 galaxies were classified by online participants in a time span of months. Since then, other iterations of the Galaxy Zoo project have annotated other datasets with more complex classification schemes (e.g., Willett et al., 2016). However, even this approach is not feasible for the available and upcoming large datasets. The amount of data is increasing as modern telescopes continue to take data, and projects like LSST will significantly increase the number of galaxies observed. Therefore, classifying these galaxies by crowdsourcing and visual inspection is next to impossible and developing an automated classification tool is necessary. Recently, improvement in computer vision techniques primarily through deep neural networks (e.g., Krizhevsky et al., 2012) and available computing power through GPUs have made this automated approach more promising. In an attempt to find an automated classification approach, an international competition was launched by Galaxy Zoo on Kaggle 5 using the images from the Galaxy Zoo 2 project (Willett et al., 2013) and the winning team provided a CNN (see Section 1.6.1.2 and Section A.1 for more information about CNNs) that exploits both translational and rotational symmetry in the images. This method can reproduce human classifications with near

5 https://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge 124 perfect accuracy of > 99% for the images with a high agreement among the Galaxy Zoo participants (Dieleman et al., 2015). However, in order to reduce overfitting and improve rotational invariance, Dieleman et al.(2015) flipped, rotated and cropped images to extract 16 viewpoints for each image. Next, they trained a convolutional neural network using augmented data such that the network learns features irrespective of viewpoint (Dieleman et al., 2015). However, the problem with this approach is that it cannot cover all of the possible rotations, orientations and their combinations; therefore, it still heavily depends on different training setups. Another problem is that this method is computationally very expensive because of the large augmented data. Moreover, it has been known that CNNs lose valuable information such as spatial hierarchies between features in the image. They also lack rotational invariance, which causes CNNs to incorrectly assign labels to objects as long as a set of features is present during the test-time while disregarding the spatial relationship of these features to each other (Sabour et al., 2017). Recently, Sabour et al.(2017) introduced a new type of network structure called Capsule Network (CapsNet) to address these issues in CNNs. This new structure contains capsules that are a nested set of layers. In contrast to traditional CNNs, this network is spatially aware and rotationally and transitionally invariant with the use of dynamic routing and reconstruction as regularization. Sabour et al.(2017) achieved state of the art result of 0.25% test error on the Modified National Institute of Standards and Technology (MNIST) dataset of handwritten digits with shifting the images only by two pixels without applying any other data augmentation methods (e.g., rotation, flipping, scaling, etc.). In this work, we are proposing the use of CapsNet for the task of galaxy morphology prediction as a better alternative for CNNs. 125

4.2 Galaxy Zoo 2

The Galaxy Zoo is an online project where participants perform galaxy morphology classification by answering a series of questions on the coloured images of the galaxies. In this work, we used data from the Galaxy Zoo 2 project that was labeled using 11 questions with 37 possible answers in total (Willett et al., 2013). The questions were asked in a hierarchical manner, with the selection of the next question determined by the answer to the previous question. The answers provided by users for one image transformed to a set of weighted vote fractions. These results have been used to study structure, formation and evolution of the galaxies (e.g., Melvin et al., 2014; Smethurst et al., 2015). Also, the accuracy of the results from the Galaxy Zoo projects was confirmed by comparing them with smaller samples classified by experts and automated pipelines (Bamford et al., 2009; Willett et al., 2013). Here, we used the dataset provided by Galaxy Zoo 2 for an international contest. The galaxies were selected with a variety of colours, sizes, and morphological classes. The goal of the project was to find an algorithm that could be applied to many different types of galaxies in the upcoming surveys. The total number of objects was limited by the depth of imaging in SDSS and the morphological categories that were over-represented as a function of the colour. This approach ensured that the colour does not play a role in the morphological classification and the models are purely based on the structures of the galaxies observed in the images. We only used the training set of the provided dataset because we did not have access to the labels of the validation dataset. The training set consisted of 61,578 JPEG coloured images of the galaxies with the size of 424 × 424 pixels. The morphological data was in the form of cumulative probabilities that gave higher weights to the questions that were asked higher in the question tree and determined a more fundamental morphological structure. The goal of the contest was to predict probabilities for each of the 37 answers in the question tree; therefore, the task was a regression as opposed to classification. However, 126 in this work we also reconstructed galaxy images based on the answers to question 1. This classification scheme is discussed in Section 4.5.2 in more detail.

4.3 Related Work

Machine learning techniques such as neural networks have been used in astronomy research in the past few decades (e.g., Firth et al., 2003; Collister and Lahav, 2004; Charnock and Moss, 2017; Schawinski et al., 2018). Galaxy morphology classification is traditionally done by manually extracting a number of features that are known to discriminate different classes. Examples of these features are: surface brightness, ellipticity, concentration, radii, and log-likelihood values measured from different types of radial profiles (e.g., Storrie-Lombardi et al., 1992). Storrie-Lombardi et al.(1992) used feed forward neural networks and 13 extracted parameters as input for training a classifier. Subsequent works used other machine learning methods such as kernel support vector machines (SVMs) (Tasca et al., 2009), and principal component analysis (PCA) (Naim et al., 1995; Lahav et al., 1995; De La Calleja and Fuentes, 2004) to extract features from the images. Others used predefined features (eg., Ball et al., 2004; Banerji et al., 2010). Next, they trained feed-forward neural networks using these features. These methods still heavily rely on feature extraction. In another approach, researchers used general purpose image features rather than galaxy-specific ones to perform galaxy morphology classification combined with nearest-neighbor classifiers (e.g., Kuminski et al., 2014). Significant research has been done where deep neural networks were applied to study galaxy morphology (eg., CNN; Huertas-Company et al., 2015; Dom´ınguez Sanchez´ et al., 2018). Recently, Dieleman et al.(2015) used CNNs for this task. Their approach is di fferent from the ones introduced before in two ways. First, the morphological classification scheme provided by Galaxy Zoo 2 was a much more fine-grained task compared to the 127

past work (mentioned above) where the task was classifying galaxies into a limited number of morphological classes (except Kuminski et al., 2014). Second, they did not use any prior handcrafted features or features that were extracted using machine learning algorithms such as PCA and SVM, which typically need many hours to develop. Instead, their proposed deep neural network learns hierarchies of features that allow the network to detect more abstract and complex features in the augmented images. They have used one CNN with 4 convolutional plus pooling layers for each of these viewpoints and connected it to two fully connected layers that were regularized using dropout method (Hinton et al., 2012; Srivastava et al., 2014). However, besides the problems that were mentioned in the Section 4.1, CNNs are known to lose important information about the spatial hierarchies between features in the image during the pooling process (usually max pooling) (Sabour et al., 2017). In our approach, we used CapsNet, which was proposed to solve the problems of CNNs that were discussed above. CapsNet uses capsules that are a group of neurons where their activity vector carries information about the object. The length of the vector represents the probability of the existence of the object, and its orientation represents the instantiation parameters of the object. CapsNet is positionally equivariant since it applies aﬃne transformation to the previous layers. In addition, it applies reconstruction as regularization where it minimizes the Euclidean distance between reconstructed and the original images. Also, reconstruction can be used to show that CapsNet preserved the part- whole relationship in the input image. We will discuss the structure of CapsNet in the Section 4.4.3 in more details.

4.4 Approach

In this section, we discuss our approach for galaxy morphology classiﬁcation that is quite diﬀerent from other ones proposed earlier. First, we discuss our experimental setup. 128

Next, we talk about the preprocessing that we did in order to prepare the data for training. Last, we discuss the network structure, training process and our implementation.

4.4.1 Experimental Setup

The dataset that we used contains 65,578 images with associated probabilities of 37 answers of the questions asked during Galaxy Zoo 2 project. The task on the competition was to predict the probabilities for each of these 37 answers and calculate a root-mean- square-error (RMSE). We took two approaches here. In the first approach that is discussed in more details in Section 4.5.1, we calculated RMSE, which was the goal of competition. In the second scenario that is discussed in more details in Section 4.5.2, we took only the answers to the first question as the ground truth and chose objects where annotators had more than 0.8 agreement on choosing one of the two answers to the first question. Therefore, we assigned two classes to the training examples based on the answer with the highest probability. In both evaluation scenarios, we divided the dataset to 80% training and 20% testing subsets.

4.4.2 Data Preprocessing

We first cropped the images to reduce the dimensions of the input to the network. The majority of the objects were in the center of the images that fit in a square smaller than the size of the image; therefore, we cropped images from 424 × 424 pixels to 216 × 216 pixels and then down-sampled them 3 times to 72 × 72 pixels. We shifted images 2 pixels in each dimension with zero padding. We did not do any other data preprocessing and augmentation because CapsNet performs well with small datasets (Sabour et al., 2017). We trained using both coloured and grayscale images and we achieved better results with coloured images. Also, it is known that there is colour difference between the different parts of the galaxy such as bulge and the disk components (Kennedy et al., 2016). Therefore, since we did not normalize each image independently and normalized images by 129 subtracting mean value and dividing standard deviation for each filter on the whole dataset, colour has a positive impact on the performance of our networks. However, we should note that the coloured images were used for both our baseline and CapsNet architecture.

4.4.3 Capsule Network

Capsules in CapsNet (Sabour et al., 2017) are groups of neurons that output vectors that represent diﬀerent poses of the input. One of the disadvantages of the CNNs as mentioned before comes from pooling layers. In order to overcome this, after a number of convolutional layers without any pooling, CapsNet reshapes the results of the last convolution layer to Primary Capsules that are just multidimensional vectors that represent the existence and the pose information (such as rotation, translation, or any other transformation) of the features. Next, CapsNet applies an algorithm called “routing by agreement” to Primary Capsules (PrCap; layer l). In this algorithm, the capsules in layer PrCap predict the outputs of the capsules in the “Galax Capsules” layer (GalaxCap; l + 1). The ﬁnal layer (GalaxCap) has a 16-dimensional capsule for each galaxy class. Each Primary Capsule sends its vector to all of the possible capsules in GalaxCap while generating a prediction vector for all possible capsules in GalaxCap. The routing weights that connect capsules of these two layers get stronger if the predictions coming from capsules in PrCap layer have a strong agreement with the actual outputs of the capsules in GalaxCap layer and weaker if they disagree during the routing iterations.

Taking ui as the activation value for the capsule i in the PrCap layer, the predicted outputu ˆ j|i of the capsules in the GalaxCap layer is represented by,

uˆ j|i = Wi jui (4.1)

where Wi j is learned by the network during the backward propagation. Next, the coupling coeﬃcients of the Primary and Galax Capsules (ci j) are calculated by applying a Softmax function on the initial logits bi j that are set to zero at the initial stage of the routing by 130

agreement process, exp(bi j) ci j = P (4.2) k exp(bik) where k is the number of capsules in the GalaxCap layer. After that, the input layer of capsule j in GalaxCap layer is calculated as follows: X s j = ci juˆ j|i (4.3) i Then, a non-linear squashing function represented in eq. 4.4 is applied on the output vectors to keep their length between 0 and 1 because the length of these vectors represent the probability of the presence of the object in the image, ||s ||2 s v j j j = 2 (4.4) 1 + ||s j|| ||s j||

Next, the log probabilities are updated by the actual outputs of the v j capsules j in GalaxCap

layer and the predicted outputsu ˆ j|i as following,

bi j ← bi j + v j · uˆ j|i (4.5)

Each of the capsules k in the GalaxCap layer is associated with a margin loss function lk that has the following form,

2 2 + − lk = Tkmax 0, m − ||vk|| + λ(1 − Tk)max 0, ||vk|| − m (4.6) where Tk is one when class k is present and zero otherwise. In this work, we chose λ = 0.5, m+ = 0.9 and m− = 0.1 for consistency with previous work (Sabour et al., 2017). In the case of regression, we used mean-square error (MSE) between the predictions and true crowd-sourced probabilities as the loss function that is as following,

37 0 X 0 2 MSE(pk, pk) = (pk − pk) (4.7) k=1 6 0 where pk is the answer probabilities associated with an image and pk are probabilities predicted by the network.

6 These probabilities are post-processed vote fractions of the answers of the Galaxy Zoo participants. 131

4.4.4 Network Architecture

4.4.4.1 Capsule Network

Our network structure was based on the original CapsNet with the same architecture introduced by Sabour et al.(2017) that is shown in Figure 4.1. The structure of the network was as following (background and terminology are mentioned in Section A.1 and Section A.2):

• Inputs: 72 × 72 downsampled images of the galaxies.

• Layer 1: a convolutional layer with 256 ﬁlters with a receptive ﬁeld of 9 × 9 and a stride of 1 with no zero padding that leads to the 256 feature maps with the size of 64 × 64.

• Layer 2: second convolutional layer with 256 ﬁlters with a receptive ﬁeld of 9 × 9 and a stride of 2 applied and then reshaped to 32 primary capsules with 8 dimensions where each dimension is a feature map with the size of 28 × 28.

• Last layer: 2 or 37 capsules based on the training scheme studied in this work where each of them represented one class.

• Decoder: the decoder part of the network was composed of three fully connected layers with 512, 1024 and 15,552 neurons respectively where the neurons in the ﬁrst two had ReLU as their activation function and the neurons of the last layer had a Sigmoid activation function. The number of neurons in the last layer were equal to the number of pixels in the input image. In fact, the reconstruction loss is the squared diﬀerence between the reconstructed image and the input image and it was scaled to 0.0005, so it would not dominate during the training process.

The decoder part of the network forces the capsules to learn features during the training that are useful for the reconstruction of the image; therefore, it acts like a 132 regularization for the network and controls the overﬁtting. For the regression task, we removed the decoder part of the network and computed RMSE as discussed in Section 4.4.3.

Figure 4.1: The architecture of the model used in this work. Top: represents architecture of the capsule layers where the GalaxCap layer has 2 or 37 capsules based on the diﬀerent setups discussed in Section 4.4.4. Bottom: represents the structure of the decoder that acts as regularization during the training

4.4.4.2 Baseline Network

The architecture that we have used here for our baseline CNN matches the depth of the CapsNet that was also used by Sabour et al.(2017). We tested several networks with different number of filters in each layer to find a network with optimal performance. For the baseline network we used cross entropy as the loss function for the classification 133

scheme and RMSE in the regression case. The architecture of the baseline network with best performance is as described in the following:

• 72 × 72 downsampled images of the galaxies as input.

• A convolution layer with 512 ﬁlters with a receptive ﬁeld of 9 × 9 and a stride of 1.

• Max pooling with a receptive ﬁeld of 2 × 2 and a stride of 2.

• Rectiﬁed Linear Function (ReLU) (Glorot et al., 2011; Nair and Hinton, 2010) as activation function

• A convolution layer with 256 ﬁlters with a receptive ﬁeld of 5 × 5 and a stride of 1.

• Max pooling with a receptive ﬁeld of 2 × 2 and a stride of 2.

• ReLU as activation function

• A fully connected layer with 1024 neurons with ReLU as their activation function with dropout rate of 0.5.

• A fully connected layer where the number of neurons is assigned based on the number of classes in the classiﬁcation scheme.

4.4.5 Implementation and Resources

We implemented our model 7 in Python using the Pytorch library based on the code provided in gram ai(2018) that enabled us to use GPU acceleration. Moreover, the Pytorch library carried out the diﬀerentiations with the autograd method. We used one NVIDIA Tesla P100 GPU unit along with 4 CPUs on the Owen cluster at the Ohio Supercomputer Center (OSC) with 16Gb of memory (Center, 1987). For training our networks, we used an Adam optimizer.

7 https://github.com/RezaKatebi/Galaxy-Morphology-CapsNet 134

4.5 Results

4.5.1 Regression

In this section, we removed the decoder part of the CapsNet and computed the RMSE between the predictions and true crowd-sourced probabilities as explained in Section 4.4.3. We also removed the last fully connected layer, dropouts and Log-Softmax layer in the baseline model. We ran both models for 30 epochs. The baseline took 6 hours while CapsNet took 30 hours of real-time computing. One reason behind this was that our pilot study was only allocated one GPU on the cluster and we had to choose a batch size of 5 because of limited memory. We should note that in terms of the number of parameters, the baseline model has 208,961,829 while CapsNet has 124,209,845 in this training scheme. We reported the computed RMSEs in Table 4.1. We also show RMSE vs number of epochs in Figure 4.2 for both training and testing. As we can see in the results, CapsNet outperformed our baseline model. The result achieved by our CapsNet (RMSE∼ 0.101) during the test time did not outperform the result achieved by Dieleman et al.(2015)(RMSE ∼ 0.077). However, we should note that the network used by Dieleman et al.(2015) was trained by a much larger dataset with heavy augmentation and the CNN they used is much deeper than networks represented here.

Model Training Testing

Baseline 0.106 0.114 CapsNet 0.082 0.103

Table 4.1: Computed RMSE between the predictions and true crowd-sourced probabilities. A relative error reduction of 8.8% was achieved. 135

CapsNet 0.130 CapsNet 0.15 Baseline Baseline 0.14 0.125

0.13 0.120 0.12 0.115 RMSE 0.11 RMSE

0.10 0.110

0.09 0.105

0.08 0 5 10 15 20 25 30 0 5 10 15 20 25 30 Epochs Epochs

(a) Training (b) Testing

Figure 4.2: Training and testing RMSE vs number of epochs for the regression scenario.

4.5.2 Classiﬁcation Based on Answers to Question 1 and Reconstruction of Galaxies

In this setup we only chose question one from the question tree, because this question is the most fundamental. Speciﬁcally, the question asked “Is the galaxy simply smooth and rounded, with no sign of a disk?”. There are three answers to this question that determined whether the object is round and smooth (elliptical galaxies), has a disk (spiral galaxies), or is an artifact or a star. We ﬁrst calculated the measure of agreement using equation 7 in Dieleman et al.(2015) that reads, H(p) a(p) = 1 − (4.8) log(n)

Pn where H(p) = − i=1 p(xi) log(p(xi)) is the entropy of the discrete probability distribution

p(xi) over n options. The value of a(p) is between 0 and 1 where 0 stands for minimal agreement and 1 stands for maximal agreement. Next, we chose the images where the measure of agreement of participants was a(p) > 0.8 where participants only chose between the ﬁrst two answers (1.1 and 1.2; see Figure 1 and Table 2 in Willett et al.(2013)). For this task, we picked the answer with the highest probability as the correct answer to question one. On 988 images the participants chose 1.1 and on 5,094 images they chose 136

1.2 as an answer to question 1 with more than 0.8 measure of agreement. By doing this task, we want to know whether CapsNet can learn and preserve the pose information of the galaxies. We trained the baseline model and CapsNet for this scheme for 200 epochs with a batch size of 20. The training took 1 hour and 3 hours for the baseline model and CapsNet, respectively. In terms of the number of parameters, the best performing baseline model had 208,926,978 while CapsNet had 28,276,672 for this training scheme. We show the accuracy curves versus number of epochs for both training and testing in Figure 4.3. As we show in Table 4.2, CapsNet outperforms the baseline CNN by 36.5% error reduction. Furthermore, we show the reconstructed images at 10, 100 and 200 epochs generated by CapsNet versus the original images in Figure 4.4. These reconstructed images are very detailed and they show that CapsNet correctly learned the posing of the input galaxy. However, the reconstruction has no obvious impact on the test accuracy (Figure 4.3). In order to check whether reconstructed images preserved physical properties of the original images, we used brightness profiles of the galaxies to indicate the Sersic´ index (Sersic´ , 1963) for each galaxy. The Sersic´ profile or the Sersic´ law shows how intensity I of a galaxy changes with the distance R from its center. The Sersic´ profile has the following form,

1/n log(I(R)) = log (I0) − kR (4.9)

where I0 is the intensity at R = 0 and n is the Sersic´ index that controls the curvature of the proﬁle. We used the GALFIT software (Peng et al., 2002) to estimate the Sersic´ index for a subset of reconstructed and original images (116 samples of each) and the results can be found in Figure 4.5. We should note that we used a Gaussian Point Spread Function (PSF) with an average FWHM of 6 pixels that was estimated using the stars present in the ﬁeld. However, Willett et al.(2013) mention that each Galaxy Zoo image was re-scaled to a variable number of arcseconds per pixel during image creation, which causes slight 137

changes in the PSF and therefore GALFIT slightly underestimates or overestimates the Sersic´ index. Furthermore, we calculated the mean (−0.41) and 95% confidence interval ([−1.33, 0.51]) of the difference between the Sersic´ index estimated for our sample of the original and reconstructed images (nOriginal −nReconstructed) and the results can be found in Figure 4.6. These results indicate that the reconstructed images fairly preserved the Sersic´ profile of the original images. However, the estimated Sersic´ index for reconstructed images are mostly larger than the original counterparts. The reason behind this is that the reconstructed images have stronger and spatially larger central light sources than the original images; therefore, the estimated Sersic´ indexes are larger for them.

Model Training Testing

Baseline 99.87 % 96.63% CapsNet 100% 97.86%

Table 4.2: Training and testing accuracy vs number of epochs for classiﬁcation based on the answers to question one. A relative error reduction of 36.5% was achieved.

4.6 Conclusions

In this work, we presented a new method for performing morphological classification of the galaxies. We used a recently introduced neural network structure called Capsule Network in two different scenarios. In the first scenario, we trained models to predict the true crowd-sourced probabilities using both our baseline model and CapsNet. As shown in Table 4.1, CapsNet clearly outperforms the best performing baseline CNN by 8.8%. The RMSE∼ 0.101 achieved our CapsNet did not outperform RMSE∼ 0.077 achieved by Dieleman et al.(2015). 138

100 98 98 96 96 94 94 92 92 90

Accuracy 90 Accuracy

88 88 CapsNet CapsNet 86 CapsNet (No Reconstruction) 86 CapsNet (No Reconstruction) 84 Baseline 84 Baseline

0 25 50 75 100 125 150 175 200 0 25 50 75 100 125 150 175 200 Epochs Epochs

(a) Training (b) Testing

Figure 4.3: Training and testing accuracy vs number of epochs for classiﬁcation based on the answers to question one.

In the second scenario, we chose objects where the participants had more than 0.8 agreement when answering question 1 from the question tree in the Galaxy Zoo project. Next, we chose the answer with the highest probability to be the class of the object. As we can see in Table 4.2, CapsNet outperforms the baseline CNN by 36.5% error reduction. We also reconstructed galaxy images using the decoder part of CapsNet that were very detailed and very close to their original counterparts while preserving their pose information. Furthermore, the Sersic´ index of the galaxies shows that the reconstructed images preserve the physical properties of the original images. However, the estimated Sersic´ index for the reconstructed images is higher than the estimated Sersic´ index of their original counterparts. This can be explained by a larger central light source in the reconstructed images. Thus, training the network on larger datasets with more resolution will be a possible solution to improve this result. CapsNet worked really well despite the fact that we did not do any data augmentation and view point extraction similar to Dieleman et al.(2015) and our network is much 139

(a) Reconstruction after 10 epochs (b) Reconstruction after 100 epochs

Figure 4.4: Original and reconstructed images of the galaxies.

shallower compared to the one presented in their work. Another thing to note is that the CapsNet proposed here has fewer parameters compared to the baseline CNN. Also, CapsNet learns features until the last layer that can be used to reconstruct galaxy 140

Figure 4.5: Estimated Sersic´ index of original images versus reconstructed images.

images, which shows that CapsNet learned pose information of the original galaxies (see Figure 4.4). Therefore, we believe that CapsNet is more suitable for the task of galaxy morphological classiﬁcation. Furthermore, CapsNet can be used without use of heavy data augmentation with high performance. However, we should note that the current implementation of the routing by agreement is slow and more work is needed to reduce the computational complexity. We should note that in our work we used the same number of capsules and the same values for λ, m+ and m− as in Sabour et al.(2017). In the future, we would like to tune the depth of the network and the number of capsules used in the network along with the diﬀerent values of the parameters. Additionally, it would be interesting to apply CapsNet 141

− − − − nOriginal − nReconstructed

Figure 4.6: Mean and 95% conﬁdence interval of the diﬀerence between the estimated

Sersic´ index for the original and reconstructed images (nOriginal − nReconstructed).

on larger and more recent datasets generated by the Galaxy Zoo project. Furthermore, extending to multiple GPUs will help to overcome the limitations of our pilot study. Upcoming large sky surveys such as LSST will increase the amount of data on galaxies dramatically and an automated method for tasks like morphological classiﬁcation is highly needed. The method presented here is a possible solution for such tasks. 142 5 Conclusions and Outlook

5.1 Conclusions

This dissertation reports results on two different areas of research. In the first part of the dissertation, we studied two nuclear outbursts and discussed the power sources of these events. The two outburst events named PS1-13cbe and PS1-10cdq were first observed in the PS1 survey during 2013 and 2010 observing runs, respectively. We utilize multiple-epoch and multi-wavelength photometric and spectroscopic observations to extract properties of the light curve such as the evolution of the luminosity, temperature, and color. We also discuss the evolution of the spectra by fitting and studying the strength and variability of the broad and weak emission lines and changes of the continuum. Using these extracted properties we then assess SN, TDE, and AGN activity as three possible power sources of these events. In the second part of the dissertation, we focused on using a modern deep learning technique to tackle the long posed problem of galaxy morphology prediction. To improve the prediction process we made use of a newly introduced neural network structure called “Capsule Networks”. In the following paragraphs, we summarize the key results from this dissertation and outline the future directions of research that can be taken to improve our understanding of nuclear outbursts and the environment surrounding them. In the case of PS1-13cbe, we favor AGN activity and argue against the SN and TDE scenarios. The presence of a pre-existing AGN increases the chances of AGN activity. Also, the constant blackbody temperature and color evolution is not consistent with the cooling ejecta of a SN. Using these observational features, we disfavor the SN Type IIn scenario. Although TDEs have been observed in host galaxies with pre-existing AGN (e.g., SDSS J0748; Wang et al., 2011), the lack of a broad He IIλ4686 emission line, low blackbody temperature at the time of peak, a light curve that has small fluctuations, and unusual re-brightening by 75% observed in PS1-13cbe is not consistent with known 143

TDE candidates; therefore, we disfavored TDE origin as well. Finally, we concluded that PS1-13cbe is a CL AGN that is powered by accretion disk instabilities. However, we rejected the TDE and obscuration as the the origin of these instabilities by calculating the crossing and viscous timescales and showing that they are much longer than the observed timescale in PS1-13cbe. Additionally, we showed that the inﬂow timescale is also much longer than the observed timescale which eliminates the possibility of outside-in variations. We then concluded that the thermal instabilities in the accretion disk are the source of this CL behavior and outburst. In the case of PS1-10cdq, we favor a TDE as the power source of the outburst and argue against the AGN activity and SLSN Type IIn. The outburst happened in the center of a NLS1 galaxy. The temperature cooling and strong color evolution is consistent with the cooling ejecta of SNe. Also, the presence of strong narrow and broad Balmer emission lines close to the outburst can be caused by CSI. However, the broad components of these emission lines are not as strong as typically observed in SLSN Type IIn candidates. Moreover, we tried to ﬁt the light-curve of PS1-10cdq using the CSM model of MOSFiT but we could not reconstruct the light-curve behavior especially in the NUV band and the parameters estimated by the model have high uncertainties. Finally, the presence of the pre-existing AGN increases the chances of the outburst having originated in the vicinity of the central SMBH. However, we argue against AGN variability as the source of the outburst by calculating the accretion disk instabilities timescales and show that they are too long or too short to explain the observed timescale in PS1-10cdq. Furthermore, we also reject the obscuration of the engine by obscuring material outside of the BLR moving in and out of the line of sight. We calculate the crossing timescale of these obscuring materials to be ≥ 33 years which is much longer than observed timescale here. Additionally, the rate of extreme AGN variability is estimated to be ∼ 10−5 yr−1 sr−1 which is very low and reduces the chances of PS1-10cdq to be caused by AGN variability (Graham et al., 2017). 144

It has been shown that a pre-existing AGN can increase the chances of a TDE. Also, broad Hα and Hβ emission lines detected in the spectrum taken close to the outburst of PS1-10cdq have been detected in other optical TDEs. Moreover, the suppression of NUV compared to the optical light suggests presence of an optically thick absorbing layer that is likely related to the circularized stellar debris of a TDE. Furthermore, we were able to reconstruct the light-curve behavior of PS1-10cdq using the TDE model of MOSFiT software that also accounted for the observed NUV suppression. Additionally, the spectrum of the PS1-10cdq taken close to the outburst shows a lot of similarities to the spectra of PS16dtm and CSS100217 taken close to the first detection, which are both suggested to be caused by a TDE (Blanchard et al., 2017). Based on these reasons, we believe that the outburst observed in PS1-10cdq is powered by a TDE which is one of the brightest TDEs observed with a very short decline timescale to the baseline. However, we would like to note the posterior distributions of the parameters of both CSM and TDE models from MOSFiT package are highly dispersed and in order to overcome this we need to run them with higher number of initialized walkers and for more iterations. In the second part of this dissertation, we used CapsNets for the task of galaxy morphological predictions and compared to the best performing baseline CNN. Using the Galaxy Zoo 2 dataset we designed two experiments. In the first experiment, we trained both of the neural networks to predict crowd-sourced probabilities. The CapsNet outperformed the baseline CNN by 8.8% in error reduction (see Table 4.1). In the second scenario, we selected objects where participants had more than 0.8 agreement on answering question 1 in the Galaxy Zoo 2 decision tree. After that we assigned the answer with the highest probability as the class of the object. In this case, CapsNet also outperforms the baseline by 36.5% error reduction (see Table 4.2). On the other hand, we were able to reconstruct galaxy images using the decoder part of the CapsNet. As we show in Figure 4.4, these reconstructed images are very detailed and very close to original counterparts that also 145 preserve their pose information. Next, we estimated the Sersic´ index for both original and reconstructed galaxy images which shows the reconstructed galaxy images also preserve the physical properties of their original counterparts. However, as shown in Figure 4.5, the estimated Sersic´ index is higher for reconstructed galaxies, which can be explained by the presence of the larger central light source in the reconstructed images. This can be solved by training the network on a larger dataset with higher resolution. We should note that the CapsNet used in our work has fewer parameters compared to our baseline CNN and is shallower than previously used neural networks for this task. Additionally, one of the strengths of CapsNets is that they learn features up to the last layer that can be used for reconstruction of the original galaxies while preserving their pose information. Furthermore, we did not use heavy augmentation and the CapsNet used here is much shallower compared to the previous works (e.g., Dieleman et al., 2015). Despite all of these differences, CapsNet performed well. However, we should note that the current implementation of the routing by agreement algorithm is slow, which increases computation time and further work is needed to reduce the computational complexity of this algorithm (see Section A.2 for more information.)

5.2 Outlook

The research presented in this thesis opens possible research areas to be explored in the future. In the first part of this thesis we explored the nature of outbursts that occurred in the center of their host galaxies. One of the interesting observations was in the case of PS1-13cbe where the turn-on timescale is much shorter than the values predicted by viscous accretion theory. The simple axisymmetric model classifies AGNs based on their angle of sight and this short-time outburst observed here challenges this viewpoint and suggests a more complex structure. PS1-13cbe might be one of the most rapid change of state observed in CLs. Most of the CLs have been observed in the timespan of years to decades 146 apart, but the very rapid change of state observed in the case of PS1-13cbe suggests that these CLs may have been through these short timescale outbursts. Therefore, more frequent multi-wavelength observations can catch more CLs during these short timescale outbursts and give us a better understanding of CL behavior. In the case of PS1-10cdq, the outburst timescale is very short that is similar to PS1-13cbe that further supports the necessity of more frequent multi-wavelength observations that will enable us to catch more of these short timescale outbursts. In the second part of the thesis, we used a CapsNet for the task of galaxy morphology predictions that outperformed a baseline CNN. We should note that we used the exact structure and hyperparameters introduced in Sabour et al.(2017) and in the future we would like to tune the number of capsules, the depth of the network, and values of the hyperparameters. However, we should note that currently work is being done to reduce the computational complexity of capsules and “routing by agreement” algorithm. In the future research it is important to consider using more efficient implementations of CapsNets and test them on larger datasets and more complex tasks. Upcoming large sky surveys such as LSST will increase the amount of data dramatically and rapid growth in the field of deep learning is presenting viable tools like CapsNets that can be used to automate tasks such as object classification, triggering, image reconstruction, object generation, and noise correction. 147 References

Ahn, C. P., Alexandroﬀ, R., Prieto, C. A., Anders, F., Anderson, S. F., Anderton, T., Andrews, B. H., Aubourg, E.,´ Bailey, S., Bastien, F. A., et al. (2014). The tenth data release of the sloan digital sky survey: First spectroscopic data from the sdss-iii apache point observatory galactic evolution experiment. The Astrophysical Journal Supplement Series, 211(2):17.

Aihara, H., Prieto, C. A., An, D., Anderson, S. F., Aubourg, E.,´ Balbinot, E., Beers, T. C., Berlind, A. A., Bickerton, S. J., Bizyaev, D., et al. (2011). The eighth data release of the sloan digital sky survey: ﬁrst data from sdss-iii. The Astrophysical Journal Supplement Series, 193(2):29.

Alexander, D., Stern, D., Del Moro, A., Lansbury, G., Assef, R., Aird, J., Ajello, M., Ballantyne, D., Bauer, F., Boggs, S., et al. (2013). The nustar extragalactic survey: A ﬁrst sensitive look at the high-energy cosmic x-ray background population. The Astrophysical Journal, 773(2):125.

Alexander, K. D., Berger, E., Guillochon, J., Zauderer, B. A., and Williams, P. K. (2016). Discovery of an outﬂow from radio observations of the tidal disruption event asassn-14li. The Astrophysical Journal Letters, 819(2):L25.

Andrews, J., Clayton, G. C., Wesson, R., Sugerman, B., Barlow, M., Clem, J., Ercolano, B., Fabbri, J., Gallagher, J., Landolt, A., et al. (2011). Evidence for pre-existing dust in the bright type iin sn 2010jl. The Astronomical Journal, 142(2):45.

Andrews, J., Gallagher, J., Clayton, G. C., Sugerman, B., Chatelain, J., Clem, J., Welch, D., Barlow, M., Ercolano, B., Fabbri, J., et al. (2010). Sn 2007od: a type iip supernova with circumstellar interaction. The Astrophysical Journal, 715(1):541.

Andrews, J., Krafton, K. M., Clayton, G. C., Montiel, E., Wesson, R., Sugerman, B. E., Barlow, M., Matsuura, M., and Drass, H. (2016). Early dust formation and a massive progenitor for sn 2011ja? Monthly Notices of the Royal Astronomical Society, 457(3):3241–3253.

Angus, C., Levan, A., Perley, D., Tanvir, N. R., Lyman, J., Stanway, E., and Fruchter, A. (2016). A hubble space telescope survey of the host galaxies of superluminous supernovae. Monthly Notices of the Royal Astronomical Society, 458(1):84–104.

Antonucci, R. (1993). Uniﬁed models for active galactic nuclei and quasars. Annual review of astronomy and astrophysics, 31:473–521.

Antonucci, R. and Miller, J. (1985). Spectropolarimetry and the nature of ngc 1068. The Astrophysical Journal, 297:621–632. 148

Arcavi, I., Gal-Yam, A., Sullivan, M., Pan, Y.-C., Cenko, S. B., Horesh, A., Ofek, E. O., De Cia, A., Yan, L., Yang, C.-W., et al. (2014). A continuum of h-to he-rich tidal disruption candidates with a preference for e+ a galaxies. The Astrophysical Journal, 793(1):38.

Arcavi, I., Howell, D. A., Kasen, D., Bildsten, L., Hosseinzadeh, G., McCully, C., Wong, Z. C., Katz, S. R., Gal-Yam, A., Sollerman, J., Taddia, F., Leloudas, G., Fremling, C., Nugent, P. E., Horesh, A., Mooley, K., Rumsey, C., Cenko, S. B., Graham, M. L., Perley, D. A., Nakar, E., Shaviv, N. J., Bromberg, O., Shen, K. J., Ofek, E. O., Cao, Y., Wang, X., Huang, F., Rui, L., Zhang, T., Li, W., Li, Z., Zhang, J., Valenti, S., Guevel, D., Shappee, B., Kochanek, C. S., Holoien, T. W. S., Filippenko, A. V., Fender, R., Nyholm, A., Yaron, O., Kasliwal, M. M., Sullivan, M., Blagorodnova, N., Walters, R. S., Lunnan, R., Khazov, D., Andreoni, I., Laher, R. R., Konidaris, N., Wozniak, P., and Bue, B. (2017). Energetic eruptions leading to a peculiar hydrogen-rich explosion of a massive star. , 551:210–213.

Aretxaga, I., Femandes, R. C., and Terlevich, R. J. (1997). Qso variability: probing the starburst model. Monthly Notices of the Royal Astronomical Society, 286(2):271–283.

Aretxaga, I., Joguet, B., Kunth, D., Melnick, J., and Terlevich, R. (1999). Seyfert 1 mutation of the classical seyfert 2 nucleus ngc 7582. The Astrophysical Journal Letters, 519(2):L123.

Balberg, S., Zampieri, L., and Shapiro, S. L. (2000). Black hole emergence in supernovae. The Astrophysical Journal, 541(2):860.

Baldwin, J., Phillips, M., and Terlevich, R. (1981). Classiﬁcation parameters for the emission-line spectra of extragalactic objects. Publications of the Astronomical Society of the Paciﬁc, 93(551):5.

Ball, N. M., Loveday, J., Fukugita, M., Nakamura, O., Okamura, S., Brinkmann, J., and Brunner, R. J. (2004). Galaxy types in the Sloan Digital Sky Survey using supervised artiﬁcial neural networks. , 348:1038–1046.

Bamford, S. P., Nichol, R. C., Baldry, I. K., Land, K., Lintott, C. J., Schawinski, K., Slosar, A., Szalay, A. S., Thomas, D., Torki, M., et al. (2009). Galaxy zoo: the dependence of morphology and colour on environment. Monthly Notices of the Royal Astronomical Society, 393(4):1324–1352.

Banerji, M., Lahav, O., Lintott, C. J., Abdalla, F. B., Schawinski, K., Bamford, S. P., Andreescu, D., Murray, P., Raddick, M. J., Slosar, A., Szalay, A., Thomas, D., and Vandenberg, J. (2010). Galaxy Zoo: reproducing galaxy morphologies via machine learning. , 406:342–353.

Barth, A. J., Bennert, V. N., Canalizo, G., Filippenko, A. V., Gates, E. L., Greene, J. E., Li, W., Malkan, M. A., Pancoast, A., Sand, D. J., et al. (2015). The lick agn monitoring 149

project 2011: Spectroscopic campaign and emission-line light curves. The Astrophysical Journal Supplement Series, 217(2):26.

Beloborodov, A., Illarionov, A., Ivanov, P., and Polnarev, A. (1992). Angular momentum of a supermassive black hole in a dense star cluster. Monthly Notices of the Royal Astronomical Society, 259(2):209–217.

Benetti, S., Nicholl, M., Cappellaro, E., Pastorello, A., Smartt, S., Elias-Rosa, N., Drake, A., Tomasella, L., Turatto, M., Harutyunyan, A., et al. (2014). The supernova css121015: 004244+ 132827: a clue for understanding superluminous supernovae. Monthly Notices of the Royal Astronomical Society, 441(1):289–303.

Bennert, N., Falcke, H., Schulz, H., Wilson, A. S., and Wills, B. J. (2002). Size and structure of the narrow-line region of quasars. The Astrophysical Journal Letters, 574(2):L105.

Bennett, C., Larson, D., Weiland, J., and Hinshaw, G. (2014). The 1% concordance hubble constant. The Astrophysical Journal, 794(2):135.

Bentz, M. C., Denney, K. D., Grier, C. J., Barth, A. J., Peterson, B. M., Vestergaard, M., Bennert, V. N., Canalizo, G., De Rosa, G., Filippenko, A. V., et al. (2013). The low-luminosity end of the radius-luminosity relationship for active galactic nuclei. The Astrophysical Journal, 767(2):149.

Berger, E., Chornock, R., Lunnan, R., Foley, R., Czekala, I., Rest, A., Leibler, C., Soderberg, A. M., Roth, K., Narayan, G., et al. (2012). Ultraluminous supernovae as a new probe of the interstellar medium in distant galaxies. The Astrophysical Journal Letters, 755(2):L29.

Bersten, M. C., Benvenuto, O., and Hamuy, M. (2011). Hydrodynamical models of type ii plateau supernovae. The Astrophysical Journal, 729(1):61.

Bhirombhakdi, K., Chornock, R., Margutti, R., Nicholl, M., Metzger, B. D., Berger, E., Margalit, B., and Milisavljevic, D. (2018). Where is the engine hiding its missing energy? constraints from a deep x-ray non-detection of the superluminous sn 2015bn. The Astrophysical Journal Letters, 868(2):L32.

Bhirombhakdi, K., Chornock, R., Miller, A. A., Filippenko, A. V., Cenko, S. B., and Smith, N. (2019). The type ii superluminous sn 2008es at late times: near-infrared excess and circumstellar interaction. Monthly Notices of the Royal Astronomical Society, 488(3):3783–3793.

Bianchi, S., Piconcelli, E., Chiaberge, M., Bailon,´ E. J., Matt, G., and Fiore, F. (2009). How complex is the obscuration in active galactic nuclei? new clues from the suzaku monitoring of the x-ray absorbers in ngc 7582. The Astrophysical Journal, 695(1):781. 150

Blanchard, P., Nicholl, M., Berger, E., Guillochon, J., Margutti, R., Chornock, R., Alexander, K., Leja, J., and Drout, M. (2017). Ps16dtm: a tidal disruption event in a narrow-line seyfert 1 galaxy. The Astrophysical Journal, 843(2):106.

Blandford, R. and McKee, C. (1982). Reverberation mapping of the emission line regions of seyfert galaxies and quasars. The Astrophysical Journal, 255:419–439.

Blandford, R. and Rees, M. (1974). A twin-exhaustmodel for double radio sources. Monthly Notices of the Royal Astronomical Society, 169(3):395–415.

Blanton, M. R., Bershady, M. A., Abolfathi, B., Albareti, F. D., Prieto, C. A., Almeida, A., Alonso-Garc´ıa, J., Anders, F., Anderson, S. F., Andrews, B., et al. (2017). Sloan digital sky survey iv: mapping the milky way, nearby galaxies, and the distant universe. The Astronomical Journal, 154(1):28.

Blinnikov, S., Eastman, R., Bartunov, O., Popolitov, V., and Woosley, S. (1998). A comparative modeling of supernova 1993j. The Astrophysical Journal, 496(1):454.

Bloom, J. S., Giannios, D., Metzger, B. D., Cenko, S. B., Perley, D. A., Butler, N. R., Tanvir, N. R., Levan, A. J., O’Brien, P. T., Strubbe, L. E., et al. (2011). A possible relativistic jetted outburst from a massive black hole fed by a tidally disrupted star. Science, 333(6039):203–206.

Bogdanovic,´ T., Cheng, R. M., and Amaro-Seoane, P. (2014). Disruption of a red giant star by a supermassive black hole and the case of ps1-10jh. The Astrophysical Journal, 788(2):99.

Bonnerot, C., Rossi, E. M., Lodato, G., and Price, D. J. (2015). Disc formation from tidal disruptions of stars on eccentric orbits by schwarzschild black holes. Monthly Notices of the Royal Astronomical Society, 455(2):2253–2266.

Boureau, Y.-L., Ponce, J., and LeCun, Y. (2010). A theoretical analysis of feature pooling in visual recognition. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 111–118.

Branch, D. and Wheeler, J. C. (2017). Supernova explosions. Springer.

Brown, G. C., Levan, A. J., Stanway, E. R., Tanvir, N., Cenko, S., Berger, E., Chornock, R., and Cucchiaria, A. (2015). Swift j1112. 2- 8238: a candidate relativistic tidal disruption ﬂare. Monthly Notices of the Royal Astronomical Society, 452(4):4297–4306.

Bruzual, G. and Charlot, S. (2003). Stellar population synthesis at the resolution of 2003. Monthly Notices of the Royal Astronomical Society, 344(4):1000–1028.

Burrows, D., Kennea, J., Ghisellini, G., Mangano, V., Zhang, B., Page, K. L., Eracleous, M., Romano, P., Sakamoto, T., Falcone, A., et al. (2011). Relativistic jet activity from the tidal disruption of a star by a massive black hole. Nature, 476(7361):421–424. 151

Burrows, D. N., Hill, J. E., Nousek, J. A., Wells, A. A., Short, A. T., Ambrosi, R. M., Chincarini, G., Citterio, O., and Tagliaferri, G. (2003). Swift x-ray telescope (xrt). In X-Ray and Gamma-Ray Telescopes and Instruments for Astronomy, volume 4851, pages 1320–1326. International Society for Optics and Photonics. Cackett, E. M., Horne, K., and Winkler, H. (2007). Testing thermal reprocessing in active galactic nuclei accretion discs. Monthly Notices of the Royal Astronomical Society, 380(2):669–682. Campana, S., Mainetti, D., Colpi, M., Lodato, G., DAvanzo, P., Evans, P. A., and Moretti, A. (2015). Multiple tidal disruption flares in the active galaxy ic 3599. Astronomy & Astrophysics, 581:A17. Cappelluti, N., Ajello, M., Rebusco, P., Komossa, S., Bongiorno, A., Clemens, C., Salvato, M., Esquej, P., Aldcroft, T., Greiner, J., et al. (2009a). A candidate tidal disruption event in the galaxy cluster abell 3571. Astronomy & Astrophysics, 495(2):L9–L12. Cappelluti, N., Brusa, M., Hasinger, G., Comastri, A., Zamorani, G., Finoguenov, A., Gilli, R., Puccetti, S., Miyaji, T., Salvato, M., et al. (2009b). The xmm-newton wide- field survey in the cosmos field-the point-like x-ray source catalogue. Astronomy & Astrophysics, 497(2):635–648. Cardelli, J. A., Clayton, G. C., and Mathis, J. S. (1989). The relationship between infrared, optical, and ultraviolet extinction. The Astrophysical Journal, 345:245–256. Cenko, S. B., Bloom, J. S., Kulkarni, S., Strubbe, L. E., Miller, A. A., Butler, N. R., Quimby, R. M., Gal-Yam, A., Ofek, E. O., Quataert, E., et al. (2012a). Ptf10iya: a short- lived, luminous flare from the nuclear region of a star-forming galaxy. Monthly Notices of the Royal Astronomical Society, 420(3):2684–2699. Cenko, S. B., Krimm, H. A., Horesh, A., Rau, A., Frail, D. A., Kennea, J. A., Levan, A. J., Holland, S. T., Butler, N. R., Quimby, R. M., et al. (2012b). Swift j2058. 4+ 0516: Discovery of a possible second relativistic tidal disruption flare? The Astrophysical Journal, 753(1):77. Center, O. S. (1987). Ohio supercomputer center. http://osc.edu/ark:/19495/f5s1ph73. Chabrier, G. (2003). Galactic stellar and substellar initial mass function. Publications of the Astronomical Society of the Pacific, 115(809):763. Chambers, K. C., Magnier, E., Metcalfe, N., Flewelling, H., Huber, M., Waters, C., Denneau, L., Draper, P., Farrow, D., Finkbeiner, D., et al. (2016). The pan-starrs1 surveys. arXiv preprint arXiv:1612.05560. Chang, Y.-Y., van der Wel, A., da Cunha, E., and Rix, H.-W. (2015). Stellar masses and star formation rates for 1 m galaxies from sdss+ wise. The Astrophysical Journal Supplement Series, 219(1):8. 152

Charnock, T. and Moss, A. (2017). Deep Recurrent Neural Networks for Supernovae Classiﬁcation. , 837:L28.

Chatzopoulos, E., Wheeler, J. C., and Vinko, J. (2012). Generalized semi-analytical models of supernova light curves. The Astrophysical Journal, 746(2):121.

Chevalier, R. (1982). The radio and x-ray emission from type ii supernovae. The Astrophysical Journal, 259:302–310.

Chevalier, R. A. and Fransson, C. (1994). Emission from circumstellar interaction in normal type ii supernovae. The Astrophysical Journal, 420:268–285.

Chevalier, R. A. and Fransson, C. (2017). Thermal and non-thermal emission from circumstellar interaction. Handbook of Supernovae, pages 875–937.

Chevalier, R. A. and Liang, E. P. (1989). The interaction of supernovae with circumstellar bubbles. The Astrophysical Journal, 344:332–340.

Chieﬃ, A., Dominguez, I., Hoﬂich,¨ P., Limongi, M., and Straniero, O. (2003). Theoretical light curves of type ii-p supernovae and applications to cosmology. Monthly Notices of the Royal Astronomical Society, 345(1):111–122.

Chornock, R., Berger, E., Gezari, S., Zauderer, B. A., Rest, A., Chomiuk, L., Kamble, A., Soderberg, A. M., Czekala, I., Dittmann, J., et al. (2013). The ultraviolet-bright, slowly declining transient ps1-11af as a partial tidal disruption event. The Astrophysical Journal, 780(1):44.

Chugai, N. and Danziger, I. (1994). Sn 1988z: low-mass ejecta colliding with the clumpy wind? Monthly Notices of the Royal Astronomical Society, 268(1):173–180.

Chugai, N., Danziger, I., and Della Valle, M. (1995). Optical spectrum of sn 1978k: emission from shocked clouds in the circumstellar wind. Monthly Notices of the Royal Astronomical Society, 276(2):530–536.

Chugai, N. N., Blinnikov, S. I., Cumming, R. J., Lundqvist, P., Bragaglia, A., Filippenko, A. V., Leonard, D. C., Matheson, T., and Sollerman, J. (2004). The type iin supernova 1994w: evidence for the explosive ejection of a circumstellar envelope. Monthly Notices of the Royal Astronomical Society, 352(4):1213–1231.

Clavel, J., Nandra, K., Makino, F., Pounds, K., Reichert, G., Urry, C., Wamsteker, W., Peracaula-Bosch, M., Stewart, G., and Otani, C. (1992). Correlated hard x-ray and ultraviolet variability in ngc 5548. The Astrophysical Journal, 393:113–125.

Clery, D. (2011). Galaxy zoo volunteers share pain and glory of research.

Cohen, R. D., Puetter, R., Rudy, R. J., Ake, T., and Foltz, C. B. (1986). Variability of markarian 1018-seyfert 1.9 to seyfert 1. The Astrophysical Journal, 311:135–141. 153

Colgate, S. A. and Cameron, A. G. W. (1963). Enhancement of light output from a supernova. Nature, 200(UCRL-7502).

Collier, S. and Peterson, B. M. (2001). Characteristic ultraviolet/optical timescales in active galactic nuclei. The Astrophysical Journal, 555(2):775.

Collin-Souﬀrin, S. (1991). On the origin of the optical and uv continuum in active galactic nuclei. Astronomy and Astrophysics, 249:344–350.

Collin-Souﬀrin, S., Alloin, D., and Andrillat, Y. (1973). A discussion of the new variations observed in the nucleus of the seyfert galaxy ngc 3516. Astronomy and Astrophysics, 22:343–354.

Collister, A. A. and Lahav, O. (2004). Annz: estimating photometric redshifts using artiﬁcial neural networks. Publications of the Astronomical Society of the Paciﬁc, 116(818):345.

Cooke, J., Sullivan, M., Gal-Yam, A., Barton, E. J., Carlberg, R. G., Ryan-Weber, E. V., Horst, C., Omori, Y., and D´ıaz, C. G. (2012). Superluminous supernovae at redshifts of 2.05 and 3.90. Nature, 491(7423):228.

Coughlin, E. R. and Begelman, M. C. (2014). Hyperaccretion during tidal disruption events: weakly bound debris envelopes and jets. The Astrophysical Journal, 781(2):82.

Czerny, B., Doroshenko, V., Nikołajuk, M., Schwarzenberg-Czerny, A., Loska, Z., and Madejski, G. (2003). Variability of accretion ﬂow in the core of the seyfert galaxy ngc 4151. Monthly Notices of the Royal Astronomical Society, 342(4):1222–1240.

Dai, L., McKinney, J. C., and Miller, M. C. (2015). Soft x-ray temperature tidal disruption events from stars on deep plunging orbits. The Astrophysical Journal Letters, 812(2):L39.

Darg, D. W., Kaviraj, S., Lintott, C. J., Schawinski, K., Sarzi, M., Bamford, S., Silk, J., Andreescu, D., Murray, P., Nichol, R., et al. (2010). Galaxy zoo: the properties of merging galaxies in the nearby universe–local environments, colours, masses, star formation rates and agn activity. Monthly Notices of the Royal Astronomical Society, 401(3):1552–1563.

De La Calleja, J. and Fuentes, O. (2004). Machine learning and image analysis for morphological galaxy classiﬁcation. Monthly Notices of the Royal Astronomical Society, 349(1):87–93.

Denney, K., De Rosa, G., Croxall, K., Gupta, A., Bentz, M., Fausnaugh, M., Grier, C., Martini, P., Mathur, S., Peterson, B., et al. (2014). The typecasting of active galactic nuclei: Mrk 590 no longer ﬁts the role. The Astrophysical Journal, 796(2):134. 154

Dessart, L. and Hillier, D. J. (2010). Supernova radiative-transfer modelling: a new approach using non-local thermodynamic equilibrium and full time dependence. Monthly Notices of the Royal Astronomical Society, 405(4):2141–2160.

Dieleman, S., Willett, K. W., and Dambre, J. (2015). Rotation-invariant convolutional neural networks for galaxy morphology prediction. Monthly notices of the royal astronomical society, 450(2):1441–1459.

Dom´ınguez Sanchez,´ H., Huertas-Company, M., Bernardi, M., Tuccillo, D., and Fischer, J. (2018). Improving galaxy morphologies for sdss with deep learning. Monthly Notices of the Royal Astronomical Society, 476(3):3661–3676.

Dom´ınguez Sanchez,´ H., Huertas-Company, M., Bernardi, M., Tuccillo, D., and Fischer, J. L. (2018). Improving galaxy morphologies for SDSS with Deep Learning. , 476:3661– 3676.

Dong, S., Shappee, B., Prieto, J., Jha, S., Stanek, K., Holoien, T.-S., Kochanek, C., Thompson, T., Morrell, N., Thompson, I., et al. (2016). Asassn-15lh: A highly superluminous supernova. Science, 351(6270):257–260.

Donley, J. L., Brandt, W., Eracleous, M., and Boller, T. (2002). Large-amplitude x- ray outbursts from galactic nuclei: a systematic survey using rosat archival data. The Astronomical Journal, 124(3):1308.

Donnarumma, I., Rossi, E., Fender, R., Komossa, S., Paragi, Z., Van Velzen, S., and Prandoni, I. (2015). Ska as a powerful hunter of jetted tidal disruption events. arXiv preprint arXiv:1501.04640.

Draine, B. T. (2009). Interstellar dust models and evolutionary implications. arXiv preprint arXiv:0903.1658.

Drake, A., Djorgovski, S., Mahabal, A., Anderson, J., Roy, R., Mohan, V., Ravindranath, S., Frail, D., Gezari, S., Neill, J. D., et al. (2011). The discovery and nature of the optical transient css100217: 102913+ 404220. The Astrophysical Journal, 735(2):106.

Eddington, A. (1925). Circulating currents in rotating stars. The Observatory, 48:73–75.

Edelson, R., Alexander, T., Crenshaw, D., Kaspi, S., Malkan, M., Peterson, B., Warwick, R., Clavel, J., Filippenko, A., Horne, K., et al. (1996). Multiwavelength observations of short-timescale variability in ngc 4151. iv. analysis of multiwavelength continuum variability. The Astrophysical Journal, 470:364.

Edelson, R., Gelbord, J., Cackett, E., Connolly, S., Done, C., Fausnaugh, M., Gardner, E., Gehrels, N., Goad, M., Horne, K., et al. (2017). Swift monitoring of ngc 4151: evidence for a second x-ray/uv reprocessing. The Astrophysical Journal, 840(1):41. 155

Elitzur, M. (2012). On the uniﬁcation of active galactic nuclei. The Astrophysical Journal Letters, 747(2):L33.

Elitzur, M., Ho, L. C., and Trump, J. R. (2014). Evolution of broad-line emission from active galactic nuclei. Monthly Notices of the Royal Astronomical Society, page stt2445.

Eracleous, M. and Halpern, J. P. (2001). Ngc 3065: a certiﬁed liner with broad, variable balmer lines. The Astrophysical Journal, 554(1):240.

Eracleous, M., Livio, M., Halpern, J. P., and Storchi-Bergmann, T. (1995). Elliptical accretion disks in active galactic nuclei. The Astrophysical Journal, 438:610–622.

Esquej, P., Saxton, R., Freyberg, M., Read, A., Altieri, B., Sanchez-Portal, M., and Hasinger, G. (2007). Candidate tidal disruption events from the xmm-newton slew survey. Astronomy & Astrophysics, 462(3):L49–L52.

Evans, C. R. and Kochanek, C. S. (1989). The tidal disruption of a star by a massive black hole. The Astrophysical Journal, 346:L13–L16.

Fanaroﬀ, B. and Riley, J. (1974). The morphology of extragalactic radio sources of high and low luminosity. Monthly Notices of the Royal Astronomical Society, 167(1):31P–36P.

Fausnaugh, M., Denney, K. D., Barth, A. J., Bentz, M. C., Bottorﬀ, M., Carini, M., Croxall, K. V., De Rosa, G., Goad, M. R., Horne, K., et al. (2016). Space telescope and optical reverberation mapping project. iii. optical continuum emission and broadband time delays in ngc 5548. The Astrophysical Journal, 821(1):56.

Fernandes, R. C., Stasinska,´ G., Schlickmann, M., Mateus, A., Asari, N. V., Schoenell, W., and Sodre Jr, L. (2010). Alternative diagnostic diagrams and the forgottenpopulation of weak line galaxies in the sdss. Monthly Notices of the Royal Astronomical Society, 403(2):1036–1053.

Filippenko, A. (1996). An introduction to liners. In The Physics of Liners in View of Recent Observations, volume 103, page 17.

Filippenko, A. and Sargent, W. (1985). A search for’dwarf’seyfert 1 nuclei. i-the initial data and results. The Astrophysical Journal Supplement Series, 57:503–522.

Filippenko, A. V. (1989). The’seyfert 1’optical spectra of the type ii supernovae 1987f and 1988i. The Astronomical Journal, 97:726–734.

Filippenko, A. V. (1997). Optical spectra of supernovae. Annual Review of Astronomy and Astrophysics, 35(1):309–355.

Firth, A. E., Lahav, O., and Somerville, R. S. (2003). Estimating photometric redshifts with artiﬁcial neural networks. Monthly Notices of the Royal Astronomical Society, 339(4):1195–1202. 156

Fransson, C. (1984). Comptonization and uv emission lines from type ii supernovae. Astronomy and Astrophysics, 133:264–284.

Fransson, C., Chevalier, R. A., Filippenko, A. V., Leibundgut, B., Barth, A. J., Fesen, R. A., Kirshner, R. P., Leonard, D. C., Li, W., Lundqvist, P., et al. (2002). Optical and ultraviolet spectroscopy of sn 1995n: evidence for strong circumstellar interaction. The Astrophysical Journal, 572(1):350.

Fukushima, K. (1980). Neocognitron: a self organizing neural network model for a mechanism of pattern recognition unaﬀected by shift in position. Biological cybernetics, 36(4):193–202.

Gaskell, C. M. and Rojas Lobos, P. A. (2013). The production of strong, broad he ii emission after the tidal disruption of a main-sequence star by a supermassive black hole. Monthly Notices of the Royal Astronomical Society: Letters, 438(1):L36–L40.

Geha, M., Alcock, C., Allsman, R. A., Alves, D. R., Axelrod, T. S., Becker, A. C., Bennett, D., Cook, K. H., Drake, A. J., Freeman, K., et al. (2003). Variability-selected quasars in macho project magellanic cloud ﬁelds. The Astronomical Journal, 125(1):1.

Gehrels, N., Chincarini, G., Giommi, P., Mason, K., Nousek, J., Wells, A., White, N., Barthelmy, S., Burrows, D., Cominsky, L., et al. (2004). The swift gamma-ray burst mission. The Astrophysical Journal, 611(2):1005.

Gezari, S., Chornock, R., Rest, A., Huber, M., Forster, K., Berger, E., Challis, P. J., Neill, J., Martin, D., Heckman, T., et al. (2012). An ultraviolet-optical ﬂare from the tidal disruption of a helium-rich stellar core. Nature, 485(7397):217–220.

Gezari, S., Halpern, J., Grupe, D., Yuan, F., Quimby, R., McKay, T., Chamarro, D., Sisson, M., Akerlof, C., Wheeler, J. C., et al. (2008). Discovery of the ultra-bright type ii-l supernova 2008es. The Astrophysical Journal, 690(2):1313.

Gezari, S., Heckman, T., Cenko, S. B., Eracleous, M., Forster, K., Gonçalves, T. S., Martin, D. C., Morrissey, P., Neff, S. G., Seibert, M., et al. (2009). Luminous thermal flares from quiescent supermassive black holes. The Astrophysical Journal, 698(2):1367.

Gezari, S., Hung, T., Cenko, S., Blagorodnova, N., Yan, L., Kulkarni, S., Mooley, K., Kong, A., Cantwell, T., Yu, P., et al. (2017). iptf discovery of the rapid turn-on of a luminous quasar. The Astrophysical Journal, 835(2):144.

Gezari, S., Martin, D., Milliard, B., Basa, S., Halpern, J., Forster, K., Friedman, P., Morrissey, P., Neﬀ, S., Schiminovich, D., et al. (2006). Ultraviolet detection of the tidal disruption of a star by a supermassive black hole. The Astrophysical Journal Letters, 653(1):L25. 157

Giannios, D. and Metzger, B. D. (2011). Radio transients from stellar tidal disruption by massive black holes. Monthly Notices of the Royal Astronomical Society, 416(3):2102– 2107.

Gillessen, S., Eisenhauer, F., Trippe, S., Alexander, T., Genzel, R., Martins, F., and Ott, T. (2009). Monitoring stellar orbits around the massive black hole in the galactic center. The Astrophysical Journal, 692(2):1075.

Giveon, U., Maoz, D., Kaspi, S., Netzer, H., and Smith, P. S. (1999). Long-term optical variability properties of the palomargreen quasars. Monthly Notices of the Royal Astronomical Society, 306(3):637–654.

Glorot, X., Bordes, A., and Bengio, Y. (2011). Deep sparse rectiﬁer neural networks. In Proceedings of the Fourteenth International Conference on Artiﬁcial Intelligence and Statistics, pages 315–323.

Goodrich, R. W. (1989). Spectropolarimetry of’narrow-line’seyfert 1 galaxies. The Astrophysical Journal, 342:224–234.

Graham, M. J., Djorgovski, S., Drake, A. J., Stern, D., Mahabal, A. A., Glikman, E., Larson, S., and Christensen, E. (2017). Understanding extreme quasar optical variability with crts–i. major agn ﬂares. Monthly Notices of the Royal Astronomical Society, 470(4):4112–4132. gram ai (2018). gram-ai/capsule-networks.

Granato, G. L., De Zotti, G., Silva, L., Bressan, A., and Danese, L. (2004). A physical model for the coevolution of qsos and their spheroidal hosts. The Astrophysical Journal, 600(2):580.

Gravet, R., Cabrera-Vives, G., Perez-Gonz´ alez,´ P. G., Kartaltepe, J., Barro, G., Bernardi, M., Mei, S., Shankar, F., Dimauro, P., Bell, E., et al. (2015). A catalog of visual-like morphologies in the 5 candels ﬁelds using deep learning. The Astrophysical Journal Supplement Series, 221(1):8.

Green, A. R., McHardy, I., and Lehto, H. (1993). On the nature of rapid x-ray variability in active galactic nuclei. Monthly Notices of the Royal Astronomical Society, 265(3):664– 680.

Grupe, D., Komossa, S., Leighly, K. M., and Page, K. L. (2010). The simultaneous optical- to-x-ray spectral energy distribution of soft x-ray selected active galactic nuclei observed by swift. The Astrophysical Journal Supplement Series, 187(1):64.

Guillochon, J., Manukian, H., and Ramirez-Ruiz, E. (2014). Ps1-10jh: The disruption of a main-sequence star of near-solar composition. The Astrophysical Journal, 783(1):23. 158

Guillochon, J., Nicholl, M., Villar, V. A., Mockler, B., Narayan, G., Mandel, K. S., Berger, E., and Williams, P. K. (2018). Mosﬁt: Modular open source ﬁtter for transients. The Astrophysical Journal Supplement Series, 236(1):6.

Guillochon, J. and Ramirez-Ruiz, E. (2013). Hydrodynamical simulations to determine the feeding rate of black holes by the tidal disruption of stars: the importance of the impact parameter and stellar structure. The Astrophysical Journal, 767(1):25.

Guillochon, J. and Ramirez-Ruiz, E. (2015). A dark year for tidal disruption events. The Astrophysical Journal, 809(2):166.

Halpern, J. and Steiner, J. (1983). Low-ionization active galactic nuclei-x-ray or shock heated? The Astrophysical Journal, 269:L37–L41.

Halpern, J. P., Gezari, S., and Komossa, S. (2004). Follow-up chandra observations of three candidate tidal disruption events. The Astrophysical Journal, 604(2):572.

Hawkins, M. (2000). Quasar variability: Correlations with amplitude. Astronomy and Astrophysics Supplement Series, 143(3):465–481.

Hayasaki, K., Stone, N., and Loeb, A. (2016). Circularization of tidally disrupted stars around spinning supermassive black holes. Monthly Notices of the Royal Astronomical Society, 461(4):3760–3780.

Heckman, T., Crane, P., and Balick, B. (1980). An optical and radio survey of the nuclei of bright galaxies-sample selection and observations. Astronomy and Astrophysics Supplement Series, 40:295–305.

Heckman, T. M. (1980). An optical and radio survey of the nuclei of bright galaxies-activity in normal galactic nuclei. Astronomy and Astrophysics, 87:152–164.

Heckman, T. M., Kauﬀmann, G., Brinchmann, J., Charlot, S., Tremonti, C., and White, S. D. (2004). Present-day growth of black holes and bulges: The sloan digital sky survey perspective. The Astrophysical Journal, 613(1):109.

Hills, J. (1975). Possible power source of seyfert galaxies and qsos. Nature, 254(5498):295–298.

Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. R. (2012). Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580.

Ho, L. C. (2004). Coevolution of Black Holes and Galaxies: Volume 1, Carnegie Observatories Astrophysics Series, volume 1. Cambridge University Press.

Ho, L. C. (2008). Nuclear activity in nearby galaxies. Annu. Rev. Astron. Astrophys., 46:475–539. 159

Ho, L. C., Filippenko, A. V., and Sargent, W. (1995). A search for’dwarf’seyfert nuclei. 2: an optical spectral atlas of the nuclei of nearby galaxies. The Astrophysical Journal Supplement Series, 98:477–593.

Ho, L. C., Filippenko, A. V., Sargent, W. L., and Peng, C. Y. (1997). A search for dwarf seyfert nuclei. iv. nuclei with broad hα emission. The Astrophysical Journal Supplement Series, 112(2):391.

Hochreiter, S., Bengio, Y., Frasconi, P., Schmidhuber, J., et al. (2001). Gradient ﬂow in recurrent nets: the diﬃculty of learning long-term dependencies.

Holoien, T.-S., Kochanek, C., Prieto, J., Stanek, K., Dong, S., Shappee, B., Grupe, D., Brown, J., Basu, U., Beacom, J., et al. (2015). Six months of multiwavelength follow-up of the tidal disruption candidate asassn-14li and implied tde rates from asas-sn. Monthly Notices of the Royal Astronomical Society, 455(3):2918–2935.

Holoien, T.-S., Prieto, J., Bersier, D., Kochanek, C., Stanek, K., Shappee, B., Grupe, D., Basu, U., Beacom, J., Brimacombe, J., et al. (2014). Asassn-14ae: a tidal disruption event at 200 mpc. Monthly Notices of the Royal Astronomical Society, 445(3):3263– 3277.

Holoien, T. W.-S., Vallely, P. J., Auchettl, K., Stanek, K., Kochanek, C. S., French, K. D., Prieto, J. L., Shappee, B. J., Brown, J. S., Fausnaugh, M. M., et al. (2019). Discovery and early evolution of asassn-19bt, the ﬁrst tde detected by tess. arXiv preprint arXiv:1904.09293.

Hopkins, P. F., Hernquist, L., Cox, T. J., Di Matteo, T., Robertson, B., and Springel, V. (2006). A uniﬁed, merger-driven model of the origin of starbursts, quasars, the cosmic x-ray background, supermassive black holes, and galaxy spheroids. The Astrophysical Journal Supplement Series, 163(1):1.

Hoyle, F. and Fowler, W. A. (1967). Gravitational red-shifts in quasi-stellar objects. Nature, 213(5074):373–374.

Hryniewicz, K. and Walter, R. (2016). Unbeamed tidal disruption events at hard x-rays. Astronomy & Astrophysics, 586:A9.

Huertas-Company, M., Gravet, R., Cabrera-Vives, G., Perez-Gonz´ alez,´ P. G., Kartaltepe, J. S., Barro, G., Bernardi, M., Mei, S., Shankar, F., Dimauro, P., Bell, E. F., Kocevski, D., Koo, D. C., Faber, S. M., and Mcintosh, D. H. (2015). A Catalog of Visual-like Morphologies in the 5 CANDELS Fields Using Deep Learning. , 221:8.

Inserra, C., Prajs, S., Gutierrez, C., Angus, C., Smith, M., and Sullivan, M. (2018). A statistical approach to identify superluminous supernovae and probe their diversity. The Astrophysical Journal, 854(2):175. 160

Ivanov, P. B. and Chernyakova, M. (2006). Relativistic cross sections of mass stripping and tidal disruption of a star by a super-massive rotating black hole. Astronomy & Astrophysics, 448(3):843–852. Ivezic, Z., Tyson, J., Abel, B., Acosta, E., Allsman, R., AlSayyad, Y., Anderson, S., Andrew, J., Angel, R., Angeli, G., et al. (2008). Lsst: from science drivers to reference design and anticipated data products. arXiv preprint arXiv:0805.2366. Jerkstrand, A., Smartt, S., Inserra, C., Nicholl, M., Chen, T.-W., Kruhler,¨ T., Sollerman, J., Taubenberger, S., Gal-Yam, A., Kankare, E., et al. (2017). Long-duration superluminous supernovae at late times. The Astrophysical Journal, 835(1):13. Kalberla, P. M., Burton, W., Hartmann, D., Arnal, E., Bajaja, E., Morras, R., and Poppel,¨ W. (2005). The leiden/argentine/bonn (lab) survey of galactic hi-ﬁnal data release of the combined lds and iar surveys with improved stray-radiation corrections. Astronomy & Astrophysics, 440(2):775–782. Kasen, D. and Woosley, S. (2009). Type ii supernovae: model light curves and standard candle relationships. The Astrophysical Journal, 703(2):2205. Kaspi, S., Maoz, D., Netzer, H., Peterson, B., Alexander, T., Barth, A., Bertram, R., Cheng, F., Chuvaev, K., Edelson, R., et al. (1996). Multiwavelength observations of short timescale variability in ngc 4151. ii. optical observations1. arXiv preprint astro-ph/9605080. Katebi, R., Chornock, R., Berger, E., Jones, D. O., Lunnan, R., Margutti, R., Rest, A., Scolnic, D. M., Burgett, W. S., Kaiser, N., et al. (2019a). Ps1-13cbe: the rapid transition of a seyfert 2 to a seyfert 1. Monthly Notices of the Royal Astronomical Society, 487(3):4057–4070. Katebi, R., Zhou, Y., Chornock, R., and Bunescu, R. (2019b). Galaxy morphology prediction using capsule networks. Monthly Notices of the Royal Astronomical Society, 486(2):1539–1547. Kauﬀmann, G., Heckman, T. M., Tremonti, C., Brinchmann, J., Charlot, S., White, S. D., Ridgway, S. E., Brinkmann, J., Fukugita, M., Hall, P. B., et al. (2003). The host galaxies of active galactic nuclei. Monthly Notices of the Royal Astronomical Society, 346(4):1055–1077. Kawaguchi, T., Mineshige, S., Umemura, M., and Turner, E. L. (1998). Optical variability in active galactic nuclei: starbursts or disk instabilities? The Astrophysical Journal, 504(2):671. Keel, W. (1983a). Spectroscopic evidence for activity in the nuclei of normal spiral galaxies. The Astrophysical Journal, 269:466–486. Keel, W. C. (1983b). An h-alpha forbidden n ii survey of the nuclei of a complete sample of spiral galaxies. The Astrophysical Journal Supplement Series, 52:229–239. 161

Kelly, B. C., Bechtold, J., and Siemiginowska, A. (2009). Are the variations in quasar optical ﬂux driven by thermal ﬂuctuations? The Astrophysical Journal, 698(1):895.

Kennedy, R., Bamford, S. P., Haußler,¨ B., Baldry, I., Bremer, M., Brough, S., Brown, M. J. I., Driver, S., Duncan, K., Graham, A. W., Holwerda, B. W., Hopkins, A. M., Kelvin, L. S., Lange, R., Phillipps, S., Vika, M., and Vulcani, B. (2016). Galaxy And Mass Assembly (GAMA): understanding the wavelength dependence of galaxy structure with bulge-disc decompositions. , 460:3458–3471.

Kesden, M. (2012). Tidal-disruption rate of stars by spinning supermassive black holes. Physical Review D, 85(2):024037.

Kewley, L. J., Dopita, M., Sutherland, R., Heisler, C., and Trevena, J. (2001). Theoretical modeling of starburst galaxies. The Astrophysical Journal, 556(1):121.

Kewley, L. J., Groves, B., Kauﬀmann, G., and Heckman, T. (2006). The host galaxies and classiﬁcation of active galactic nuclei. Monthly Notices of the Royal Astronomical Society, 372(3):961–976.

Khachikian, E. Y. and Weedman, D. W. (1974). An atlas of seyfert galaxies. The Astrophysical Journal, 192:581–589.

Kiewe, M., Gal-Yam, A., Arcavi, I., Leonard, D. C., Enriquez, J. E., Cenko, S. B., Fox, D. B., Moon, D.-S., Sand, D. J., Soderberg, A. M., et al. (2011). Caltech core-collapse project (cccp) observations of type iin supernovae: typical properties and implications for their progenitor stars. The Astrophysical Journal, 744(1):10.

Kochanek, C. (2016). Abundance anomalies in tidal disruption events. Monthly Notices of the Royal Astronomical Society, 458(1):127–134.

Komossa, S. (2015). Tidal disruption of stars by supermassive black holes: Status of observations. Journal of High Energy Astrophysics, 7:148–157.

Komossa, S. and Bade, N. (1999). The giant x-ray outbursts in ngc 5905 and ic 3599: Follow-up observations and outburst scenarios. Arxiv preprint astro-ph/9901141.

Komossa, S., Grupe, D., Schartel, N., Gallo, L., Gomez, J., Kollatschny, W., Kriss, G., Leighly, K., Longinotti, A., Parker, M., et al. (2016). The extremes of agn variability. Proceedings of the International Astronomical Union, 12(S324):168–171.

Komossa, S., Halpern, J., Schartel, N., Hasinger, G., Santos-Lleo, M., and Predehl, P. (2004). A huge drop in the x-ray luminosity of the nonactive galaxy rx j1242. 6–1119a, and the ﬁrst postﬂare spectrum: testing the tidal disruption scenario. The Astrophysical Journal Letters, 603(1):L17. 162

Koratkar, A. and Blaes, O. (1999). The ultraviolet and optical continuum emission in active galactic nuclei: the status of accretion disks. Publications of the Astronomical Society of the Paciﬁc, 111(755):1.

Korista, K. T. and Goad, M. R. (2004). What the optical recombination lines can tell us about the broad-line regions of active galactic nuclei. The Astrophysical Journal, 606(2):749.

Kormendy, J. and Ho, L. C. (2013). Coevolution (or not) of supermassive black holes and host galaxies. Annual Review of Astronomy and Astrophysics, 51:511–653.

Koshida, S., Minezaki, T., Yoshii, Y., Kobayashi, Y., Sakata, Y., Sugawara, S., Enya, K., Suganuma, M., Tomita, H., Aoki, T., et al. (2014). Reverberation measurements of the inner radius of the dust torus in 17 seyfert galaxies. The Astrophysical Journal, 788(2):159.

Kriek, M., Van Dokkum, P. G., Labbe,´ I., Franx, M., Illingworth, G. D., Marchesini, D., and Quadri, R. F. (2009). An ultra-deep near-infrared spectrum of a compact quiescent galaxy at z= 2.2. The Astrophysical Journal, 700(1):221.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classiﬁcation with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105.

Krolik, J., Horne, K., Kallman, T., Malkan, M., Edelson, R., and Kriss, G. (1991). Ultraviolet variability of ngc 5548-dynamics of the continuum production region and geometry of the broad-line region. The Astrophysical Journal, 371:541–562.

Krolik, J. H. (1999). Book review: Active galactic nuclei: from the central black hole to the galactic environment/princeton u press, 1999. Nature, 398:678.

Kuminski, E., George, J., Wallin, J., and Shamir, L. (2014). Combining human and machine learning for morphological analysis of galaxy images. Publications of the Astronomical Society of the Paciﬁc, 126(944):959.

Lahav, O., Naim, A., Buta, R., Corwin, H., De Vaucouleurs, G., Dressler, A., Huchra, J., Van Den Bergh, S., Raychaudhury, S., Sodre,´ L., et al. (1995). Galaxies, human eyes, and artiﬁcial neural networks. Science, 267(5199):859–862.

LaMassa, S. M., Cales, S., Moran, E. C., Myers, A. D., Richards, G. T., Eracleous, M., Heckman, T. M., Gallo, L., and Urry, C. M. (2015). The discovery of the ﬁrst changing look quasar: New insights into the physics and phenomenology of active galactic nuclei. The Astrophysical Journal, 800(2):144.

Laor, A. (2003). On the nature of low-luminosity narrow-line active galactic nuclei. The Astrophysical Journal, 590(1):86. 163

Lasota, J.-P. (2016). Black hole accretion discs. In Astrophysics of Black Holes, pages 1–60. Springer.

LeCun, Y., Bottou, L., Bengio, Y., and Haﬀner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324.

Leloudas, G., Fraser, M., Stone, N., van Velzen, S., Jonker, P., Arcavi, I., Fremling, C., Maund, J., Smartt, S., Kr`ıhler, T., et al. (2016). The superluminous transient asassn-15lh as a tidal disruption event from a kerr black hole. Nature Astronomy, 1:0002.

Levan, A., Tanvir, N. R., Cenko, S. B., Perley, D., Wiersema, K., Bloom, J. S., Fruchter, A. S., de Ugarte Postigo, A., OBrien, P., Butler, N., et al. (2011). An extremely luminous panchromatic outburst from the nucleus of a distant galaxy. Science, 333(6039):199– 202.

Lightman, A. P. and Eardley, D. M. (1989). Black holes in binary systems: Instability of disk accretion. Accretion: A Collection of Inﬂuential Papers, page 197.

Lin, D., Maksym, P. W., Irwin, J. A., Komossa, S., Webb, N. A., Godet, O., Barret, D., Grupe, D., and Gwyn, S. D. (2015). An ultrasoft x-ray ﬂare from 3xmm j152130. 7+ 074916: A tidal disruption event candidate. The Astrophysical Journal, 811(1):43.

Lintott, C. J., Schawinski, K., Keel, W., Van Arkel, H., Bennert, N., Edmondson, E., Thomas, D., Smith, D. J., Herbert, P. D., Jarvis, M. J., et al. (2009). Galaxy zoo:hanny’s voorwerp, a quasar light echo? Monthly Notices of the Royal Astronomical Society, 399(1):129–140.

Lintott, C. J., Schawinski, K., Slosar, A., Land, K., Bamford, S., Thomas, D., Raddick, M. J., Nichol, R. C., Szalay, A., Andreescu, D., et al. (2008). Galaxy zoo: morphologies derived from visual inspection of galaxies from the sloan digital sky survey. Monthly Notices of the Royal Astronomical Society, 389(3):1179–1189.

Liu, H., Bai, J., Zhao, X., and Ma, L. (2008). Tests for standard accretion disk models by variability in active galactic nuclei. The Astrophysical Journal, 677(2):884.

Lodato, G., King, A., and Pringle, J. (2008). Stellar disruption by a supermassive black hole: is the light curve really proportional to t- 5/3? Monthly Notices of the Royal Astronomical Society, 392(1):332–340.

Lodato, G. and Rossi, E. M. (2010). Multiband light curves of tidal disruption events. Monthly Notices of the Royal Astronomical Society, 410(1):359–367.

Loeb, A. and Ulmer, A. (1997). Optical appearance of the debris of a star disrupted by a massive black hole. The Astrophysical Journal, 489(2):573.

Lub, J. and de Ruiter, H. (1992). Optical variability of seyfert nuclei. ii-fairall 9. Astronomy and Astrophysics, 256:33–44. 164

Lunnan, R., Chornock, R., Berger, E., Laskar, T., Fong, W., Rest, A., Sanders, N., Challis, P., Drout, M., Foley, R. J., et al. (2014). Hydrogen-poor superluminous supernovae and long-duration gamma-ray bursts have similar host galaxies. The Astrophysical Journal, 787(2):138.

Lynden-Bell, D. (1969). Galactic nuclei as collapsed old quasars. Nature, 223(5207):690– 694.

Lyutyj, V., Oknyanskij, V., and Chuvaev, K. (1984). Ngc 4151-a seyfert 2 in a deep photometric minimum. Soviet Astronomy Letters, 10:335.

MacLeod, C. L., Green, P. J., Anderson, S. F., Bruce, A., Eracleous, M., Graham, M., Homan, D., Lawrence, A., LeBleu, A., Ross, N. P., Ruan, J. J., Runnoe, J., Stern, D., Burgett, W., Chambers, K. C., Kaiser, N., Magnier, E., and Metcalfe, N. (2018). Changing-Look Quasar Candidates: First Results from Follow-up Spectroscopy. ArXiv e-prints, page arXiv:1810.00087.

MacLeod, C. L., Ivezic,´ Z.,ˇ Sesar, B., de Vries, W., Kochanek, C. S., Kelly, B. C., Becker, A. C., Lupton, R. H., Hall, P. B., Richards, G. T., et al. (2012). A description of quasar variability measured using repeated sdss and poss imaging. The Astrophysical Journal, 753(2):106.

MacLeod, C. L., Ross, N. P., Lawrence, A., Goad, M., Horne, K., Burgett, W., Chambers, K. C., Flewelling, H., Hodapp, K., Kaiser, N., et al. (2016). A systematic search for changing-look quasars in sdss. Monthly Notices of the Royal Astronomical Society, 457(1):389–404.

Maddox, S., Sutherland, W., Efstathiou, G., Loveday, J., and Peterson, B. (1990). Galaxy evolution at low redshift. Monthly Notices of the Royal Astronomical Society, 247:1P.

Magorrian, J. and Tremaine, S. (1999). Rates of tidal disruption of stars by massive central black holes. Monthly Notices of the Royal Astronomical Society, 309(2):447–460.

Magorrian, J., Tremaine, S., Richstone, D., Bender, R., Bower, G., Dressler, A., Faber, S., Gebhardt, K., Green, R., Grillmair, C., et al. (1998). The demography of massive dark objects in galaxy centers. The Astronomical Journal, 115(6):2285.

Maksym, W. P., Ulmer, M. P., and Eracleous, M. (2010). A tidal disruption ﬂare in a1689 from an archival x-ray survey of galaxy clusters. The Astrophysical Journal, 722(2):1035.

Maoz, D., Barth, A. J., Sternberg, A., Filippenko, A. V., Ho, L. C., Macchetto, F. D., Rix, H.-W., and Schneider, D. P. (1996). Hubble space telescope ultraviolet images of ﬁve circumnuclear star-forming rings. arXiv preprint astro-ph/9604012. 165

Marchese, E., Braito, V., Della Ceca, R., Caccianiga, A., and Severgnini, P. (2012). Ngc 454: unveiling a new changing lookactive galactic nucleus. Monthly Notices of the Royal Astronomical Society, 421(2):1803–1812.

Marconi, A., Risaliti, G., Gilli, R., Hunt, L. K., Maiolino, R., and Salvati, M. (2004). Local supermassive black holes, relics of active galactic nuclei and the x-ray background. Monthly Notices of the Royal Astronomical Society, 351(1):169–185.

Margutti, R., Chornock, R., Metzger, B., Coppejans, D., Guidorzi, C., Migliori, G., Milisavljevic, D., Berger, E., Nicholl, M., Zauderer, B., et al. (2018). Results from a systematic survey of x-ray emission from hydrogen-poor superluminous sne. The Astrophysical Journal, 864(1):45.

Margutti, R., Metzger, B., Chornock, R., Milisavljevic, D., Berger, E., Blanchard, P. K., Guidorzi, C., Migliori, G., Kamble, A., Lunnan, R., et al. (2017). X-rays from the location of the double-humped transient asassn-15lh. The Astrophysical Journal, 836(1):25.

Martini, P., Stoll, R., Derwent, M., Zhelem, R., Atwood, B., Gonzalez, R., Mason, J., OBrien, T., Pappalardo, D., Pogge, R. W., et al. (2011). The ohio state multi-object spectrograph. Publications of the Astronomical Society of the Paciﬁc, 123(900):187.

Masters, K. L., Mosleh, M., Romer, A. K., Nichol, R. C., Bamford, S. P., Schawinski, K., Lintott, C. J., Andreescu, D., Campbell, H. C., Crowcroft, B., et al. (2010). Galaxy zoo: passive red spirals. Monthly Notices of the Royal Astronomical Society, 405(2):783–799.

Masters, K. L., Nichol, R. C., Hoyle, B., Lintott, C., Bamford, S. P., Edmondson, E. M., Fortson, L., Keel, W. C., Schawinski, K., Smith, A. M., et al. (2011). Galaxy zoo: bars in disc galaxies. Monthly Notices of the Royal Astronomical Society, 411(3):2026–2034.

Matt, G., Guainazzi, M., and Maiolino, R. (2003). Changing look: from compton-thick to compton-thin, or the rebirth of fossil active galactic nuclei. Monthly Notices of the Royal Astronomical Society, 342(2):422–426.

McConnell, N. J. and Ma, C.-P. (2013). Revisiting the scaling relations of black hole masses and host galaxy properties. The Astrophysical Journal, 764(2):184.

McCrum, M., Smartt, S., Rest, A., Smith, K., Kotak, R., Rodney, S., Young, D., Chornock, R., Berger, E., Foley, R. J., et al. (2015). Selecting superluminous supernovae in faint galaxies from the ﬁrst year of the pan-starrs1 medium deep survey. Monthly Notices of the Royal Astronomical Society, 448(2):1206–1231.

McCulloch, W. S. and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. The bulletin of mathematical biophysics, 5(4):115–133. 166

McHardy, I., Cameron, D., Dwelly, T., Connolly, S., Lira, P., Emmanoulopoulos, D., Gelbord, J., Breedt, E., Arevalo, P., and Uttley, P. (2014). Swift monitoring of ngc 5548: X-ray reprocessing and short-term uv/optical variability. Monthly Notices of the Royal Astronomical Society, 444(2):1469–1474.

Melvin, T., Masters, K., Lintott, C., Nichol, R. C., Simmons, B., Bamford, S. P., Casteels, K. R., Cheung, E., Edmondson, E. M., Fortson, L., et al. (2014). Galaxy zoo: an independent look at the evolution of the bar fraction over the last eight billion years from hst-cosmos. Monthly Notices of the Royal Astronomical Society, 438(4):2882–2897.

Merloni, A., Dwelly, T., Salvato, M., Georgakakis, A., Greiner, J., Krumpe, M., Nandra, K., Ponti, G., and Rau, A. (2015). A tidal disruption ﬂare in a massive galaxy? implications for the fuelling mechanisms of nuclear black holes. Monthly Notices of the Royal Astronomical Society, 452(1):69–87.

Miller, A., Chornock, R., Perley, D., Ganeshalingam, M., Li, W., Butler, N., Bloom, J., Smith, N., Modjaz, M., Poznanski, D., et al. (2008). The exceptionally luminous type ii-linear supernova 2008es. The Astrophysical Journal, 690(2):1303.

Miller, J. and Goodrich, R. (1990). Spectropolarimetry of high-polarization seyfert 2 galaxies and uniﬁed seyfert theories. The Astrophysical Journal, 355:456–467.

Miller, J. M., Kaastra, J. S., Miller, M. C., Reynolds, M. T., Brown, G., Cenko, S. B., Drake, J. J., Gezari, S., Guillochon, J., Gultekin, K., et al. (2015). Flows of x-ray gas reveal the disruption of a star by a massive black hole. Nature, 526(7574):542–545.

Miller, M. C. (2015). Disk winds as an explanation for slowly evolving temperatures in tidal disruption events. The Astrophysical Journal, 805(1):83.

Mockler, B., Guillochon, J., and Ramirez-Ruiz, E. (2019). Weighing black holes using tidal disruption events. The Astrophysical Journal, 872(2):151.

Morgan, C. W., Kochanek, C., Morgan, N. D., and Falco, E. E. (2010). The quasar accretion disk size-black hole mass relation. The Astrophysical Journal, 712(2):1129.

Mushotzky, R. (1982). The x-ray spectrum and time variability of narrow emission line galaxies. The Astrophysical Journal, 256:92–102.

Mushotzky, R., Fabian, A., Iwasawa, K., Kunieda, H., Matsuoka, M., Nandra, K., and Tanaka, Y. (1995). Detection of broad iron k lines in active galaxies. Monthly Notices of the Royal Astronomical Society, 272(1):L9–L12.

Nadyozhin, D. (1985). On the initial phase of interaction between expanding stellar envelopes and surrounding medium. Astrophysics and space science, 112(2):225–249. 167

Naim, A., Lahav, O., Sodre Jr, L., and Storrie-Lombardi, M. (1995). Automated morphological classification of apm galaxies by supervised artificial neural networks. Monthly Notices of the Royal Astronomical Society, 275(3):567–590. Nair, V. and Hinton, G. E. (2010). Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814. Nandra, K., George, I., Mushotzky, R., Turner, T., and Yaqoob, T. (1997). Asca observations of seyfert 1 galaxies. i. data analysis, imaging, and timing. The Astrophysical Journal, 476(1):70. Nenkova, M., Ivezic,´ Z.,ˇ and Elitzur, M. (2002). Dust emission from active galactic nuclei. The Astrophysical Journal Letters, 570(1):L9. Netzer, H. (1990). Agn emission lines. Active Galactic Nuclei, pages 57–158. Netzer, H. (2009). Accretion and star formation rates in low-redshift type ii active galactic nuclei. Monthly Notices of the Royal Astronomical Society, 399(4):1907–1920. Nicastro, F. (2000). Broad emission line regions in active galactic nuclei: the link with the accretion power. The Astrophysical Journal Letters, 530(2):L65. Nicastro, F., Martocchia, A., and Matt, G. (2003). The lack of broad-line regions in low accretion rate active galactic nuclei as evidence of their origin in the accretion disk. The Astrophysical Journal Letters, 589(1):L13. Nicholl, M., Berger, E., Margutti, R., Chornock, R., Blanchard, P., Jerkstrand, A., Smartt, S., Arcavi, I., Challis, P., Chambers, K., et al. (2016a). Superluminous supernova sn 2015bn in the nebular phase: evidence for the engine-powered explosion of a stripped massive star. The Astrophysical Journal Letters, 828(2):L18. Nicholl, M., Berger, E., Smartt, S., Margutti, R., Kamble, A., Alexander, K. D., Chen, T.- W., Inserra, C., Arcavi, I., Blanchard, P. K., et al. (2016b). Sn 2015bn: a detailed multiwavelength view of a nearby superluminous supernova. The Astrophysical Journal, 826(1):39. Novikov, I. and Thorne, K. (1973). Black holes, eds. C. DeWitt and BS DeWitt, Gordon and Breach, New York. Nyholm, A., Sollerman, J., Taddia, F., Fremling, C., Moriya, T., Ofek, E., Gal-Yam, A., De Cia, A., Roy, R., Kasliwal, M., et al. (2017). The bumpy light curve of type iin supernova iptf13z over 3 years. Astronomy & Astrophysics, 605:A6. Oknyansky, V., Gaskell, C., Huseynov, N., Lipunov, V., Shatsky, N., Tsygankov, S., Gorbovskoy, E., Mikailov, K. M., Tatarnikov, A., Buckley, D., et al. (2017). The curtain remains open: Ngc 2617 continues in a high state. Monthly Notices of the Royal Astronomical Society, 467(2):1496–1504. 168

Osterbrock, D. (1981). Seyfert galaxies with weak broad h alpha emission lines. The Astrophysical Journal, 249:462–470.

Osterbrock, D. E. and Pogge, R. W. (1985). The spectra of narrow-line seyfert 1 galaxies. The Astrophysical Journal, 297:166–176.

Peng, C. Y., Ho, L. C., Impey, C. D., and Rix, H.-W. (2002). Detailed structural decomposition of galaxy images. The Astronomical Journal, 124(1):266.

Penston, M. and Perez,´ E. (1984). An evolutionary link between seyfert i and ii galaxies? Monthly Notices of the Royal Astronomical Society, 211(1):33P–39P.

Perets, H. B., Hopman, C., and Alexander, T. (2006). Massive perturbers in the galactic center. In Journal of Physics: Conference Series, volume 54, page 293. IOP Publishing.

Pereyra, N. A., Berk, D. E. V., Turnshek, D. A., Hillier, D. J., Wilhite, B. C., Kron, R. G., Schneider, D. P., and Brinkmann, J. (2006). Characteristic qso accretion disk temperatures from spectroscopic continuum variability. The Astrophysical Journal, 642(1):87.

Peterson, B. (2001). Advanced lectures on the starburst-agn connection, ed. i. aretxaga, d. kunth, & r. mujica.´

Peterson, B., Denney, K. D., De Rosa, G., Grier, C., Pogge, R., Bentz, M., Kochanek, C., Vestergaard, M., Kilerci-Eser, E., Dalla Bonta,` E., et al. (2013). The size of the narrow- line-emitting region in the seyfert 1 galaxy ngc 5548 from emission-line variability. The Astrophysical Journal, 779(2):109.

Peterson, B. M., Ferrarese, L., Gilbert, K., Kaspi, S., Malkan, M., Maoz, D., Merritt, D., Netzer, H., Onken, C., Pogge, R., et al. (2004). Central masses and broad-line region sizes of active galactic nuclei. ii. a homogeneous analysis of a large reverberation- mapping database. The Astrophysical Journal, 613(2):682.

Piran, T., Svirski, G., Krolik, J., Cheng, R. M., and Shiokawa, H. (2015). Disk formation versus disk accretionwhat powers tidal disruption events? The Astrophysical Journal, 806(2):164.

Pogge, R. W., Maoz, D., Ho, L. C., and Eracleous, M. (2000). The narrow-line regions of liners as resolved with the hubble space telescope. The Astrophysical Journal, 532(1):323.

Polletta, M., Tajer, M., Maraschi, L., Trinchieri, G., Lonsdale, C., Chiappetti, L., Andreon, S., Pierre, M., Le Fevre, O., Zamorani, G., et al. (2007). Spectral energy distributions of hard x-ray selected active galactic nuclei in the xmm-newton medium deep survey. The Astrophysical Journal, 663(1):81. 169

Pooley, D., Lewin, W. H., Fox, D. W., Miller, J. M., Lacey, C. K., Van Dyk, S. D., Weiler, K. W., Sramek, R. A., Filippenko, A. V., Leonard, D. C., et al. (2002). X-ray, optical, and radio observations of the type ii supernovae 1999em and 1998s. The Astrophysical Journal, 572(2):932.

Puccetti, S., Fiore, F., Risaliti, G., Capalbi, M., Elvis, M., and Nicastro, F. (2007). Rapid nh changes in ngc 4151. Monthly Notices of the Royal Astronomical Society, 377(2):607– 616.

Pumo, M. and Zampieri, L. (2011). Radiation-hydrodynamical modeling of core-collapse supernovae: light curves and the evolution of photospheric velocity and temperature. The Astrophysical Journal, 741(1):41.

Quimby, R. (2006). Supernova 2006gy in ngc 1260. Central Bureau Electronic Telegrams, 644.

Quimby, R., Kulkarni, S., Kasliwal, M., Gal-Yam, A., Arcavi, I., Sullivan, M., Nugent, P., Thomas, R., Howell, D., Nakar, E., et al. (2011). Hydrogen-poor superluminous stellar explosions. Nature, 474(7352):487.

Ramirez-Ruiz, E. and Rosswog, S. (2009). The star ingesting luminosity of intermediate- mass black holes in globular clusters. The Astrophysical Journal Letters, 697(2):L77.

Rees, M. J. (1988). Tidal disruption of stars by black holes of 106–108 solar masses in nearby galaxies. Nature, 333(6173):523–528.

Rest, A., Scolnic, D., Foley, R., Huber, M., Chornock, R., Narayan, G., Tonry, J., Berger, E., Soderberg, A. M., Stubbs, C. W., et al. (2014). Cosmological constraints from measurements of type ia supernovae discovered during the ﬁrst 1.5 yr of the pan-starrs1 survey. The Astrophysical Journal, 795(1):44.

Risaliti, G., Maiolino, R., and Salvati, M. (1999). The distribution of absorbing column densities among seyfert 2 galaxies. The Astrophysical Journal, 522(1):157.

Risaliti, G., Miniutti, G., Elvis, M., Fabbiano, G., Salvati, M., Baldi, A., Braito, V., Bianchi, S., Matt, G., Reeves, J., et al. (2009). Variable partial covering and a relativistic iron line in ngc 1365. The Astrophysical Journal, 696(1):160.

Roming, P. W., Kennedy, T. E., Mason, K. O., Nousek, J. A., Ahr, L., Bingham, R. E., Broos, P. S., Carter, M. J., Hancock, B. K., Huckle, H. E., et al. (2005). The swift ultra-violet/optical telescope. Space Science Reviews, 120(3):95–142.

Rossi, E., Donnarumma, I., Fender, R., Jonker, P., Komossa, S., Paragi, Z., Prandoni, I., and Zampieri, L. (2015). Loft as a discovery machine for jetted tidal disruption events. arXiv preprint arXiv:1501.02774. 170

Roth, N., Kasen, D., Guillochon, J., and Ramirez-Ruiz, E. (2016). The x-ray through optical ﬂuxes and line strengths of tidal disruption events. The Astrophysical Journal, 827(1):3.

Ruan, J. J., Anderson, S. F., Cales, S. L., Eracleous, M., Green, P. J., Morganson, E., Runnoe, J. C., Shen, Y., Wilkinson, T. D., Blanton, M. R., et al. (2016). Toward an understanding of changing-look quasars: an archival spectroscopic search in sdss. The Astrophysical Journal, 826(2):188.

Runnoe, J. C., Brotherton, M. S., and Shang, Z. (2012). Updating quasar bolometric luminosity corrections. Monthly Notices of the Royal Astronomical Society, 422(1):478– 493.

Runnoe, J. C., Cales, S., Ruan, J. J., Eracleous, M., Anderson, S. F., Shen, Y., Green, P. J., Morganson, E., LaMassa, S., Greene, J. E., et al. (2016). Now you see it, now you don’t: the disappearing central engine of the quasar j1011+ 5442. Monthly Notices of the Royal Astronomical Society, 455(2):1691–1701.

Sabour, S., Frosst, N., and Hinton, G. E. (2017). Dynamic routing between capsules. In Advances in Neural Information Processing Systems, pages 3859–3869.

Salpeter, E. (1964). Accretion of interstellar matter by massive objects. The Astrophysical Journal, 140:796–800.

Sanmartim, D., Storchi-Bergmann, T., and Brotherton, M. S. (2014). 2d stellar population and gas kinematics of the inner kiloparsec of the post-starburst quasar sdss j0330- 0532. Monthly Notices of the Royal Astronomical Society, 443(1):584–597.

Sarzi, M., Falcon-Barroso,´ J., Davies, R. L., Bacon, R., Bureau, M., Cappellari, M., Tim de Zeeuw, P., Emsellem, E., Fathi, K., Krajnovic, D., et al. (2006). The sauron projectv. integral-ﬁeld emission-line kinematics of 48 elliptical and lenticular galaxies. Monthly Notices of the Royal Astronomical Society, 366(4):1151–1200.

Saxton, R., Read, A., Esquej, P., Komossa, S., Dougherty, S., Rodriguez-Pascual, P., and Barrado, D. (2012). A tidal disruption-like x-ray ﬂare from the quiescent galaxy sdss j120136. 02+ 300305.5. Astronomy & Astrophysics, 541:A106.

Schawinski, K., Turp, M. D., and Zhang, C. (2018). Exploring galaxy evolution with generative models. , 616:L16.

Schawinski, K., Virani, S., Simmons, B., Urry, C. M., Treister, E., Kaviraj, S., and Kushkuley, B. (2009). Do moderate-luminosity active galactic nuclei suppress star formation? The Astrophysical Journal Letters, 692(1):L19.

Schlaﬂy, E. F. and Finkbeiner, D. P. (2011). Measuring reddening with sloan digital sky survey stellar spectra and recalibrating sfd. The Astrophysical Journal, 737(2):103. 171

Schlegel, E. M. (1990). A new subclass of type ii supernovae? Monthly Notices of the Royal Astronomical Society, 244:269–271.

Schmidt, M. (1963). 3c 273: a star-like object with large red-shift. Nature, 197(4872):1040–1040.

Schulze, S., Kruhler,¨ T., Leloudas, G., Gorosabel, J., Mehner, A., Buchner, J., Kim, S., Ibar, E., Amor´ın, R., Herrero-Illana, R., et al. (2018). Cosmic evolution and metal aversion in superluminous supernova host galaxies. Monthly Notices of the Royal Astronomical Society, 473(1):1258–1285.

Scolnic, D., Jones, D., Rest, A., Pan, Y., Chornock, R., Foley, R., Huber, M., Kessler, R., Narayan, G., Riess, A., et al. (2018). The complete light-curve sample of spectroscopically conﬁrmed sne ia from pan-starrs1 and cosmological constraints from the combined pantheon sample. The Astrophysical Journal, 859(2):101.

Sersic,´ J. (1963). Inﬂuence of the atmospheric and instrumental dispersion on the brightness distribution in a galaxy. Boletin de la Asociacion Argentina de Astronomia La Plata Argentina, 6:41.

Seyfert, C. K. (1943). Nuclear emission in spiral nebulae. The Astrophysical Journal, 97:28.

Shakura, N. and Sunyaev, R. (1976). A theory of the instability of disk accretion on to black holes and the variability of binary x-ray sources, galactic nuclei and quasars. Monthly Notices of the Royal Astronomical Society, 175(3):613–632.

Shakura, N. I. and Sunyaev, R. A. (1973). Black holes in binary systems. observational appearance. Astronomy and Astrophysics, 24:337–355.

Shappee, B., Grupe, D., Mathur, S., Prieto, J., Kochanek, C., Stanek, K., Jencson, J., Talabere, A., Beacom, J., Szczygiel, D., et al. (2013). Swift monitoring of the agn outburst in ngc2617: Signiﬁcant x-ray and uv brightening. The Astronomer’s Telegram, 5059:1.

Shappee, B., Prieto, J., Grupe, D., Kochanek, C., Stanek, K., De Rosa, G., Mathur, S., Zu, Y., Peterson, B., Pogge, R., et al. (2014). The man behind the curtain: X-rays drive the uv through nir variability in the 2013 active galactic nucleus outburst in ngc 2617. The Astrophysical Journal, 788(1):48.

Shields, G. (1978). Thermal continuum from acretion disks in quasars. Nature, 272:706– 708.

Shiokawa, H., Krolik, J. H., Cheng, R. M., Piran, T., and Noble, S. C. (2015). General relativistic hydrodynamic simulation of accretion ﬂow from a stellar tidal disruption. The Astrophysical Journal, 804(2):85. 172

Shuder, J. and Osterbrock, D. (1981). Empirical results from a study of active galactic nuclei. The Astrophysical Journal, 250:55–65.

Siemiginowska, A. and Czerny, B. (1989). Observational constraints on viscosity in agn accretion discs. Monthly Notices of the Royal Astronomical Society, 239(1):289–295.

Simmons, B. D., Melvin, T., Lintott, C., Masters, K. L., Willett, K. W., Keel, W. C., Smethurst, R., Cheung, E., Nichol, R. C., Schawinski, K., et al. (2014). Galaxy zoo: Candels barred discs and bar fractions. Monthly Notices of the Royal Astronomical Society, 445(4):3466–3474.

Skibba, R. A., Bamford, S. P., Nichol, R. C., Lintott, C. J., Andreescu, D., Edmondson, E. M., Murray, P., Raddick, M. J., Schawinski, K., Slosar, A., et al. (2009). Galaxy zoo: disentangling the environmental dependence of morphology and colour. Monthly Notices of the Royal Astronomical Society, 399(2):966–982.

Smethurst, R., Lintott, C., Simmons, B., Schawinski, K., Marshall, P., Bamford, S., Fortson, L., Kaviraj, S., Masters, K., Melvin, T., et al. (2015). Galaxy zoo: evidence for diverse star formation histories through the green valley. Monthly Notices of the Royal Astronomical Society, 450(1):435–453.

Smith, N., Chornock, R., Li, W., Ganeshalingam, M., Silverman, J. M., Foley, R. J., Filippenko, A. V., and Barth, A. J. (2008a). Sn 2006tf: precursor eruptions and the optically thick regime of extremely luminous type iin supernovae. The Astrophysical Journal, 686(1):467.

Smith, N., Chornock, R., Silverman, J. M., Filippenko, A. V., and Foley, R. J. (2010). Spectral evolution of the extraordinary type iin supernova 2006gy. The Astrophysical Journal, 709(2):856.

Smith, N., Foley, R. J., Bloom, J. S., Li, W., Filippenko, A. V., Gavazzi, R., Ghez, A., Konopacky, Q., Malkan, M. A., Marshall, P. J., et al. (2008b). Late-time observations of sn 2006gy: still going strong. The Astrophysical Journal, 686(1):485.

Smith, N., Silverman, J. M., Chornock, R., Filippenko, A. V., Wang, X., Li, W., Ganeshalingam, M., Foley, R. J., Rex, J., and Steele, T. N. (2009). Coronal lines and dust formation in sn 2005ip: not the brightest, but the hottest type iin supernova. The Astrophysical Journal, 695(2):1334.

Sparke, L. S. and Gallagher III, J. S. (2007). Galaxies in the universe: an introduction. Cambridge University Press.

Springel, V., Di Matteo, T., and Hernquist, L. (2005). Modelling feedback from stars and black holes in galaxy mergers. Monthly Notices of the Royal Astronomical Society, 361(3):776–794. 173

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overﬁtting. The Journal of Machine Learning Research, 15(1):1929–1958.

Starling, R. L., Siemiginowska, A., Uttley, P., and Soria, R. (2004). Constraints on active galactic nucleus accretion disc viscosity derived from continuum variability. Monthly Notices of the Royal Astronomical Society, 347(1):67–73.

Stauﬀer, J. (1982). A nuclear spectroscopic survey of disk galaxies. ii-galaxies with emission lines not excited by stellar photoionization. The Astrophysical Journal, 262:66– 80.

Stern, J. and Laor, A. (2012). Type 1 agn at low z–ii. the relative strength of narrow lines and the nature of intermediate type agn. Monthly Notices of the Royal Astronomical Society, 426(4):2703–2718.

Stone, N., Sari, R., and Loeb, A. (2013). Consequences of strong compression in tidal disruption events. Monthly Notices of the Royal Astronomical Society, 435(3):1809– 1824.

Stone, N. C. and Metzger, B. D. (2015). Rates of stellar tidal disruption as probes of the supermassive black hole mass function. Monthly Notices of the Royal Astronomical Society, 455(1):859–883.

Storchi-Bergmann, T., Baldwin, J. A., and Wilson, A. S. (1993). Double-peaked broad line emission from the liner nucleus of ngc 1097. The astrophysical journal. Chicago. Vol. 410, no. 1, pt. 2 (June 1993), p. L11-L14.

Storrie-Lombardi, M., Lahav, O., Sodre Jr, L., and Storrie-Lombardi, L. (1992). Morphological classiﬁcation of galaxies by artiﬁcial neural networks. Monthly Notices of the Royal Astronomical Society, 259(1):8P–12P.

Strubbe, L. E. and Murray, N. (2015). Insights into tidal disruption of stars from ps1-10jh. Monthly Notices of the Royal Astronomical Society, 454(3):2321–2343.

Strubbe, L. E. and Quataert, E. (2009). Optical ﬂares from the tidal disruption of stars by massive black holes. Monthly Notices of the Royal Astronomical Society, 400(4):2070– 2084.

Suganuma, M., Yoshii, Y., Kobayashi, Y., Minezaki, T., Enya, K., Tomita, H., Aoki, T., Koshida, S., and Peterson, B. A. (2006). Reverberation measurements of the inner radius of the dust torus in nearby seyfert 1 galaxies. The Astrophysical Journal, 639(1):46.

Taddia, F., Stritzinger, M., Sollerman, J., Phillips, M., Anderson, J., Boldt, L., Campillay, A., Castellon,´ S., Contreras, C., Folatelli, G., et al. (2013). Carnegie supernova project: observations of type iin supernovae. Astronomy & Astrophysics, 555:A10. 174

Tanaka, Y. a., Nandra, K., Fabian, A., Inoue, H., et al. (1995). Gravitationally redshifted emission implying an accretion disk and massive black hole in the active galaxy mcg-6- 30-15. Nature, 375(6533):659.

Tasca, L., Rouan, D., Pelat, D., Kneib, J., Le Fevre,` O., Capak, P., Kartaltepe, J., Koekemoer, A., McCracken, H., Salvato, M., et al. (2009). A robust morphological classiﬁcation of high-redshift galaxies using support vector machines on seeing limited images-ii. quantifying morphological k-correction in the cosmos ﬁeld at 1¡ z¡ 2: Ks band vs. i band. Astronomy & Astrophysics, 497(3):743–753.

Tohline, J. and Osterbrock, D. (1976). Variation of the spectrum of the seyfert galaxy ngc 7603. The Astrophysical Journal, 210:L117–L120.

Tonry, J., Stubbs, C., Lykke, K., Doherty, P., Shivvers, I., Burgett, W., Chambers, K., Hodapp, K., Kaiser, N., Kudritzki, R.-P., et al. (2012). The pan-starrs1 photometric system. The Astrophysical Journal, 750(2):99.

Trevese, D., Kron, R. G., and Bunone, A. (2001). Continuum variability of active galactic nuclei in the optical-ultraviolet range. The Astrophysical Journal, 551(1):103.

Trump, J. R., Impey, C. D., Kelly, B. C., Civano, F., Gabor, J. M., Diamond-Stanic, A. M., Merloni, A., Urry, C. M., Hao, H., Jahnke, K., et al. (2011). Accretion rate and the physical nature of unobscured active galaxies. The Astrophysical Journal, 733(1):60.

Ulmer, A. (1999). Flares from the tidal disruption of stars by massive black holes. The Astrophysical Journal, 514(1):180.

Urry, C. M. and Padovani, P. (1995). Uniﬁed schemes for radio-loud active galactic nuclei. Publications of the Astronomical Society of the Paciﬁc, pages 803–845.

Utrobin, V. P. (2007). An optimal hydrodynamic model for the normal type iip supernova 1999em. Astronomy & Astrophysics, 461(1):233–251.

Van Dokkum, P. G. (2001). Cosmic-ray rejection by laplacian edge detection. Publications of the Astronomical Society of the Paciﬁc, 113(789):1420.

van Velzen, S., Anderson, G. E., Stone, N. C., Fraser, M., Wevers, T., Metzger, B. D., Jonker, P. G., van der Horst, A. J., Staley, T. D., Mendez, A. J., Miller-Jones, J. C. A., Hodgkin, S. T., Campbell, H. C., and Fender, R. P. (2015). A radio jet from the optical and x-ray bright stellar tidal disruption ﬂare asassn-14li. Science, 351(6268):62–65.

Van Velzen, S. and Farrar, G. R. (2014). Measurement of the rate of stellar tidal disruption ﬂares. The Astrophysical Journal, 792(1):53.

Van Velzen, S., Farrar, G. R., Gezari, S., Morrell, N., Zaritsky, D., Ostman,¨ L., Smith, M., Gelfand, J., and Drake, A. J. (2011). Optical discovery of probable stellar tidal disruption ﬂares. The Astrophysical Journal, 741(2):73. 175

Vanden Berk, D. E., Richards, G. T., Bauer, A., Strauss, M. A., Schneider, D. P., Heckman, T. M., York, D. G., Hall, P. B., Fan, X., Knapp, G. R., Anderson, S. F., Annis, J., Bahcall, N. A., Bernardi, M., Briggs, J. W., Brinkmann, J., Brunner, R., Burles, S., Carey, L., Castander, F. J., Connolly, A. J., Crocker, J. H., Csabai, I., Doi, M., Finkbeiner, D., Friedman, S., Frieman, J. A., Fukugita, M., Gunn, J. E., Hennessy, G. S., Ivezic,´ Z.,ˇ Kent, S., Kunszt, P. Z., Lamb, D. Q., Leger, R. F., Long, D. C., Loveday, J., Lupton, R. H., Meiksin, A., Merelli, A., Munn, J. A., Newberg, H. J., Newcomb, M., Nichol, R. C., Owen, R., Pier, J. R., Pope, A., Rockosi, C. M., Schlegel, D. J., Siegmund, W. A., Smee, S., Snir, Y., Stoughton, C., Stubbs, C., SubbaRao, M., Szalay, A. S., Szokoly, G. P., Tremonti, C., Uomoto, A., Waddell, P., Yanny, B., and Zheng, W. (2001). Composite Quasar Spectra from the Sloan Digital Sky Survey. , 122:549–564.

Veron-Cetty,´ M.-P. and Veron,´ P. (1986). Study of a complete sample of galaxies. ii- spectroscopy of the nuclei. Astronomy and Astrophysics Supplement Series, 66:335–372.

Veron-Cetty,´ M.-P., Veron,´ P., and Gonc¸alves, A. (2001). A spectrophotometric atlas of narrow-line seyfert 1 galaxies. Astronomy & Astrophysics, 372(3):730–754.

Vestergaard, M. and Peterson, B. M. (2006). Determining central black hole masses in distant active galaxies and quasars. ii. improved optical and uv scaling relationships. The Astrophysical Journal, 641(2):689.

Vinko,´ J., Yuan, F., Quimby, R. M., Wheeler, J. C., Ramirez-Ruiz, E., Guillochon, J., Chatzopoulos, E., Marion, G., and Akerlof, C. (2014). A luminous, fast rising uv- transient discovered by rotse: A tidal disruption event? The Astrophysical Journal, 798(1):12.

Vreeswijk, P. M., Savaglio, S., Gal-Yam, A., De Cia, A., Quimby, R. M., Sullivan, M., Cenko, S. B., Perley, D. A., Filippenko, A. V., Clubb, K. I., et al. (2014). The hydrogen- poor superluminous supernova iptf 13ajg and its host galaxy in absorption and emission. The Astrophysical Journal, 797(1):24.

Wang, J. and Merritt, D. (2004). Revised rates of stellar disruption in galactic nuclei. The Astrophysical Journal, 600(1):149.

Wang, T.-G., Zhou, H.-Y., Wang, L.-F., Lu, H.-L., and Xu, D. (2011). Transient superstrong coronal lines and broad bumps in the galaxy sdss j074820. 67+ 471214.3. The Astrophysical Journal, 740(2):85.

Wilhite, B. C., Berk, D. E. V., Kron, R. G., Schneider, D. P., Pereyra, N., Brunner, R. J., Richards, G. T., and Brinkmann, J. V. (2005). Spectral variability of quasars in the sloan digital sky survey. i. wavelength dependence. The Astrophysical Journal, 633(2):638.

Willett, K. W., Galloway, M. A., Bamford, S. P., Lintott, C. J., Masters, K. L., Scarlata, C., Simmons, B. D., Beck, M., Cardamone, C. N., Cheung, E., et al. (2016). Galaxy 176

zoo: morphological classiﬁcations for 120 000 galaxies in hst legacy imaging. Monthly Notices of the Royal Astronomical Society, 464(4):4176–4203.

Willett, K. W., Lintott, C. J., Bamford, S. P., Masters, K. L., Simmons, B. D., Casteels, K. R., Edmondson, E. M., Fortson, L. F., Kaviraj, S., Keel, W. C., et al. (2013). Galaxy zoo 2: detailed morphological classiﬁcations for 304 122 galaxies from the sloan digital sky survey. Monthly Notices of the Royal Astronomical Society, 435(4):2835–2860.

Willett, K. W., Schawinski, K., Simmons, B. D., Masters, K. L., Skibba, R. A., Kaviraj, S., Melvin, T., Wong, O. I., Nichol, R. C., Cheung, E., et al. (2015). Galaxy zoo: the dependence of the star formation–stellar mass relation on spiral disc morphology. Monthly Notices of the Royal Astronomical Society, 449(1):820–827.

Woo, J.-H., Yoon, Y., Park, S., Park, D., and Kim, S. C. (2015). The black hole mass–stellar velocity dispersion relation of narrow-line seyfert 1 galaxies. The Astrophysical Journal, 801(1):38.

Woosley, S. and Weaver, T. A. (1995). The evolution and explosion of massive stars ii: Explosive hydrodynamics and nucleosynthesis. Technical report, Lawrence Livermore National Lab., CA (United States).

Wrobel, J. and Heeschen, D. (1991). Radio-continuum sources in nearby and bright e/s0 galaxies-active nuclei versus star formation. The Astronomical Journal, 101:148–169.

Yuan, W., Komossa, S., Zhang, C., Feng, H., Ling, Z.-X., Zhao, D., Zhang, S.-N., Osborne, J., O’Brien, P., Willingale, R., et al. (2014). Detecting tidal disruption events of massive black holes in normal galaxies with the einstein probe. Proceedings of the International Astronomical Union, 10(S312):68–70.

Zakamska, N. L., Schmidt, G. D., Smith, P. S., Strauss, M. A., Krolik, J. H., Hall, P. B., Richards, G. T., Schneider, D. P., Brinkmann, J., and Szokoly, G. P. (2005). Candidate type ii quasars from the sloan digital sky survey. iii. spectropolarimetry reveals hidden type i nuclei. The Astronomical Journal, 129(3):1212.

Zakamska, N. L., Strauss, M. A., Heckman, T. M., Ivezic,´ Z.,ˇ and Krolik, J. H. (2004). Candidate type ii quasars from the sloan digital sky survey. ii. from radio to x-rays. The Astronomical Journal, 128(3):1002.

Zakamska, N. L., Strauss, M. A., Krolik, J. H., Collinge, M. J., Hall, P. B., Hao, L., Heckman, T. M., Ivezic,´ Z.,ˇ Richards, G. T., Schlegel, D. J., et al. (2003). Candidate type ii quasars from the sloan digital sky survey. i. selection and optical properties of a sample at 0.3¡ z¡ 0.83. The Astronomical Journal, 126(5):2125.

Zampieri, L., Colpi, M., Shapiro, S. L., and Wasserman, I. (1998). Supernova fallback and the emergence of a black hole. The Astrophysical Journal, 505(2):876. 177

Zauderer, B., Berger, E., Soderberg, A., Loeb, A., Narayan, R., Frail, D., Petitpas, G., Brunthaler, A., Chornock, R., Carpenter, J., et al. (2011). Birth of a relativistic outﬂow in the unusual [ggr]-ray transient swift j164449. 3+ 573451. Nature, 476(7361):425– 428.

Zeldovich, Y. B. and Novikov, I. (1964). Estimating mass of super-star. DOKLADY AKADEMII NAUK SSSR, 158(4):811. 178 Appendix:Supplementary Material

A.1 Neural Networks

We explained the inner workings of FFNNs and CNNs in Section 1.6.1.1 and Section 1.6.1.2; however, in this section we explain the terminologies used in more details. Batch: A set of training examples used during each update of model parameters. Epoch: One pass through the entire training dataset where each of the training example has been seen once. Activation Function: A function that applies a nonlinear transformations on a weighted sum of the inputs from the previous layer and passes the computed value to the next layer. Neuron: It is a node in the computational graph of neural networks that takes a batch of inputs and calculates an output value for each input by applying the activation function on a linear transformation of the input. Hidden Layer: A group of neurons that are not connected to each other and that can be computed in parallel by applying the same activation to a linear combination of the activation values of the neurons on the previous layer. Hidden layers reside between input and output layers. Objective Function (Loss Function): A measure of the quality of the model, which captures how well the model ﬁts the training examples and sometimes also contains a term that corresponds to the model complexity. Gradient Descent: An algorithm that is used to update the current model parameters using the gradient of the objective function (loss function) with respect to the parameters. Backpropagation: In the gradient descent algorithm, ﬁrst, the outputs of each neuron in each layer is calculated in a process called forward pass and then the gradient of the objective function with respect to each parameter is calculated in a backward process in 179

the neural network. Hence, the name backpropagation is given since the gradient of the objective function with respect to parameters in layer l is computed based on the gradient with respect to the parameters in the layer l + 1. Learning Rate: During the training using gradient descent algorithm, the gradient of parameters are multiplied by a scalar typically shown by η which is called learning rate. The learning rate is a hyperparameter and needs to be tuned during the training process. Convolutional Layer: As we can see in Figure 1.2, each neuron in layer l + 1 is only connected to a local receptive field with a size of F × F in the input volume from layer l that is replicated to cover the entire input by shifting the receptive field by a given stride value. For example, a stride of 1 means moving the receptive field by one pixel at a time. Sometimes the images or convolved features are padded by zeros (adding P pixels with zero values on each side of the image) to keep the image size the same or to be able to go deeper since the image size decreases when increasing the depth of the CNN (see Figure A.1). The output size of a convolutional layer is calculated using,

W0 = (W − F + 2P)/S + 1 (A.1)

L0 = (L − F + 2P)/S + 1 (A.2)

where W and L are the width and the length of the input image, W0 and L0 are the width and the length of the output image, and S represents strides to shift the receptive ﬁeld with the size F on the input image. Pooling: Reducing the size of the output matrix of a convolutional layer by taking the maximum or average value across the pooling area (see Figure A.2). Dropout: A form of regularization where a random selection of a ﬁxed number of neurons in a neural network layer is removed during a single gradient step. This method is analogous to training a large network to approximate a large ensemble of smaller networks. 180

Figure A.1: A convolution on an 7 × 7 input image (green squares) with zero-padding P = 1 (white squares) with receptive ﬁeld of size 3 × 3 (yellow squares) and a stride of S = 1 where the receptive ﬁeld moved from left to right by one pixel.

A.2 Capsule Networks

Convolutional neural networks are known to be transitionally equivariant because the convolution operator is transition equivariant itself. To show this, let’s assume an image I with the size of N × N pixels and a kernel (receptive ﬁeld) with the size of F × F. The P P convolution of the image at position (i, j) is given by Ci, j = m n Ii−m, j−nKm,n. Next, let’s translate each pixel on image I by (i + p, j + q). Recomputing the convolution of

0 0 P P the translated image at position (i , j ), we get Ci0, j0 = m n Ii+p−m, j+q−nKm,n. Comparing

0 0 Ci0, j0 and Ci, j we get i = i + p and j = j + q which shows that the convolution of the image at position (i, j) is also shifted by (p, q), hence proves that the convolution operator is transition equivariant. However, the kernel K is learned during the training of the CNN where the kernel attempts to match the given object. For CNNs to become invariant to transformations such as rotation of the input, either multiple kernels should be used where 181

Figure A.2: A max-pooling operation over a 3 × 3 convolutional matrix (blue squares) with receptive ﬁeld of size 2 × 2 (red squares) and stride of 1.

each kernel is attempting to learn a slight orientation of the input object or augmentation of the object by rotating it with different angles. Both of these methods are computationally expensive and cannot cover all of the possible transformations. On the other hand, during the max-pooling layers valuable spatial information is lost and as the CNNs grow deeper the receptive fields cover larger portion of the image that makes spatial information loss even more severe where only local and temporal information is learned by network. Therefore, CNNs fail to learn part to whole relationship in the input. Capsule networks were introduced as a solution to the mentioned problems in CNNs. Each capsule in CapsNet learns to recognize an implicitly defined fragment and outputs both the probability of the presence of this fragment and a set of “instant parameters” that include the precise pose (rotation, translation, and etc), lighting, and other deformations of the visual fragment relative to an implicitly defined version of the fragment. The instantiation parameters of the larger fragments are predicted by applying transformation 182

matrices on the smaller recognized fragments. These transformations learn to encode the intrinsic spatial relationship between the small and larger fragments. Furthermore, no spatial information is being lost because no max-pooling layers are being applied. Therefore, CapsNets are viewpoint invariant and spatially aware. One drawback of CapsNets is that they are computationally expensive. For example, the total time complexity of all convolutional layers in a typical CNN is,   Xd  O  n · s2 · n · m   l−1 l l l (A.3) l=1 where l is the index of the convolutional layer, d is the number of convolutional layers

(depth), nl−1 is the number of input channels form layer l−1, nl is the number of ﬁlters in the layer l, sl is the spatial size of each ﬁlter, and ml is the spatial size of output feature maps. However, in CapsNets, in addition to convolutional layers, “routing by agreement” adds another O(N) complexity where N is the number of routing iterations which empirically is set to 3 (Sabour et al., 2017). For example, in our training scenario in Section 4.5.1 our CapsNet took ∼ 30 hours for training 30 epochs while our baseline CNN only took ∼ 5 hours of training time for the same number of epochs. However, CapsNets are a newly introduced structures and currently work is being done to speed up capsules and especially the “routing by agreement” algorithm. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

Thesis and Dissertation Services ! !