Stochastic modeling of AGN light curves
Transcription
Stochastic modeling of AGN light curves
Credit: ESO/Kornmesser Stochastic Modeling of Quasar Variability Brandon C. Kelly (UCSB, CGE Fellow, bckelly@physics.ucsb.edu) Aneta Siemiginowska (CfA), Malgosia Sobolewska (CfA), Tommaso Treu (UCSB), Matt Malkan (UCLA), Anna Pancoast (UCSB), Jong-Hak Woo (Seoul), Jill Bechtold (Arizona) Friday, May 24, 13 Flux Flux The Data Analysis Challenge: Aperiodic Lightcurves Time Friday, May 24, 13 Time The Data Analysis Challenge: Aperiodic Lightcurves Mrk 766 Vaughan & Fabian (2003) What variability ‘features’ can we measure for quasar lightcurves? Friday, May 24, 13 Variance / Frequency Quantifying Variability with the Power Spectral Density (PSD) Frequency [arbitrary] Friday, May 24, 13 Time [arbitrary] Variance / Frequency Quantifying Variability with the Power Spectral Density (PSD) Frequency [arbitrary] Friday, May 24, 13 Time [arbitrary] Variance / Frequency Quantifying Variability with the Power Spectral Density (PSD) Frequency [arbitrary] Friday, May 24, 13 Time [arbitrary] Variance / Frequency Quantifying Variability with the Power Spectral Density (PSD) Frequency [arbitrary] Friday, May 24, 13 Time [arbitrary] Disadvantages of Traditional Non-parameteric Tools for Quantifying Aperiodic Variability Lomb-Scargle Periodogram Measurements True Mock Data True Empirical Structure Function Friday, May 24, 13 Disadvantages of Traditional Non-parameteric Tools for Quantifying Aperiodic Variability Lomb-Scargle Periodogram Measurements True Mock Data True Empirical Structure Function Friday, May 24, 13 Tools for Characterizing Aperiodic (Quasar) Variability: What should they do? Friday, May 24, 13 Tools for Characterizing Aperiodic (Quasar) Variability: What should they do? • Handle irregular/arbitrary sampling patterns and measurement errors Friday, May 24, 13 Tools for Characterizing Aperiodic (Quasar) Variability: What should they do? • Handle irregular/arbitrary sampling patterns and measurement errors • Produce interpretable results: • Connection to physical models • Connection to features in the power spectrum Friday, May 24, 13 Tools for Characterizing Aperiodic (Quasar) Variability: What should they do? • Handle irregular/arbitrary sampling patterns and measurement errors • Produce interpretable results: • Connection to physical models • Connection to features in the power spectrum • Fast & scalable to massive time domain surveys • By the end of LSST we will have millions to billions of multiwavelength lightcurves with ~ 50-1000+ epochs Friday, May 24, 13 Tools for Characterizing Aperiodic (Quasar) Variability: What should they do? • Handle irregular/arbitrary sampling patterns and measurement errors • Produce interpretable results: • Connection to physical models • Connection to features in the power spectrum • Fast & scalable to massive time domain surveys • By the end of LSST we will have millions to billions of multiwavelength lightcurves with ~ 50-1000+ epochs • Handle multiwavelength/multivariate time series • Account for correlations/time lags among lightcurves in different bands Friday, May 24, 13 Two approaches to (stochastic) modeling of real lightcurves: Frequency Domain and Time Domain Friday, May 24, 13 Monte-Carlo Methods (Done+1992,Uttley +2002,Emmanaloupolous 2010,2013) • Extremely flexible, limited only by ability to do simulation • Can be computationally expensive • Reliance on χ2 may not provide optimal use of information in lightcurve Friday, May 24, 13 Monte-Carlo Methods (Done+1992,Uttley +2002,Emmanaloupolous 2010,2013) • Extremely flexible, limited only by ability to do simulation • Can be computationally expensive • Reliance on χ2 may not provide optimal use of information in lightcurve Friday, May 24, 13 Monte-Carlo Methods (Done+1992,Uttley +2002,Emmanaloupolous 2010,2013) • Extremely flexible, limited only by ability to do simulation • Can be computationally expensive • Reliance on χ2 may not provide optimal use of information in lightcurve Friday, May 24, 13 Monte-Carlo Methods (Done+1992,Uttley +2002,Emmanaloupolous 2010,2013) • Extremely flexible, limited only by ability to do simulation • Can be computationally expensive • Reliance on χ2 may not provide optimal use of information in lightcurve Friday, May 24, 13 Gaussian Processes (Rybicki & Press 1992, Kelly+ 2009,2011, Miller+2010) loglik = log |⌃| 1 (y 2 T µ) ⌃ 1 (y • Likelihood-based approach, enables Bayesian inference • Statistically powerful, but limited by Gaussian assumption • In general, computationally expensive (O(n3)) Friday, May 24, 13 µ) Gaussian Processes (Rybicki & Press 1992, Kelly+ 2009,2011, Miller+2010) loglik = 0 log |⌃| 1 ⌃=@ n⇥n A 1 (y 2 T µ) ⌃ 1 (y ⌃ij = Cov(y(ti ), y(tj )) Z 1 2⇡if |ti = P SD(f )e 1 • Likelihood-based approach, enables Bayesian inference • Statistically powerful, but limited by Gaussian assumption • In general, computationally expensive (O(n3)) Friday, May 24, 13 µ) tj | df Simple and fast tool: First order continuous autoregressive process (CAR(1), Kelly+2009) Lightcurve dL(t) = dt (L(t) ⌧ Characteristic Time Scale White Noise Power Spectrum µ) + dW (t) 1/τ σ2 LC Mean Frequency • Solution provides likelihood function, enables maximum-likelihood or Bayesian inference • Fitting is fast! Only O(n) operations to evaluate likelihood function (e.g., Kelly+2009, Kozlowski +2010) or do interpolation Friday, May 24, 13 Fitting the CAR(1) model: Illustration τ=5 days σ=1 τ=1 day σ=1 Friday, May 24, 13 Fitting the CAR(1) model: Illustration τ=5 days σ=1 τ=1 day σ=1 Friday, May 24, 13 CAR(1) does well on optical lightcurves with typical sampling of current surveys MacLeod+(2010), ~10,000 quasars from stripe 82 Kozlowski+(2010), Ogle-III Kelly+(2009), AGN Watch Friday, May 24, 13 total variability Trends involving the CAR(1) process parameters Kelly+(in prep), data from reverberation mapping MacLeod+(2010), Stripe82 Friday, May 24, 13 Using the CAR(1) model to find quasars MacLeod+(2011) Based on Stripe82 variable point sources Quasars Stars Quasars Butler & Bloom (2011) Stars Works because quasars have more correlated variability on longer time scales compared to stars Friday, May 24, 13 Using the CAR(1) model as a tool for simulating/ interpolating quasar lightcurves Reverberation Mapping (Zu+2011, see also Brewer+, Pancoast+2011) Toy Models of Accretion Disks (Dexter & Argol 2011) See also Emmanaloupolous+(2013) for more flexible simulation methods Friday, May 24, 13 Is there evidence for deviations from the CAR(1) model in the optical? Stripe82: No (Andrae+2013) Friday, May 24, 13 Ogle-III: Maybe (Zu+2013) Kepler: Yes (Mushotzky+2011) Stochastic Modeling in the X-rays: Mixtures of CAR(1) processes (Kelly+2011) Friday, May 24, 13 Stochastic Modeling in the X-rays: Mixtures of CAR(1) processes (Kelly+2011) Process is solution of stochastic diffusion equation Diffusion Time Scale @u(x, t) @ 2 u(x, t) @W (x, t) =D + 2 @t @x @t u(x, t) = 1 X ck (x)CAR(t; ⌧k ) k=1 Potential for astrophysicallymotivated lightcurve models? Friday, May 24, 13 Diffusion time scale across noise field Mixture of CAR(1) processes does well for X-rays too Gonzalez-Martin & Vaughan (2012), see also McHardy+(2006) Kelly+(2011) Friday, May 24, 13 log Power Spectrum Estimating Black Hole Mass from Mixed CAR(1) process model McHardy (2013) log Frequency Friday, May 24, 13 See also, e.g., Yu & Lu (2001), Nikolajuk+(2004), Papadakis (2004), O’Neill+(2005), Miniutti+(2009), Caballero-Garcia+(2012) Estimating Black Hole Mass from Mixed CAR(1) process model log Power Spectrum Excess Variance Zhou+(2010) log Frequency Friday, May 24, 13 See also, e.g., Yu & Lu (2001), Nikolajuk+(2004), Papadakis (2004), O’Neill+(2005), Miniutti+(2009), Caballero-Garcia+(2012) Estimating Black Hole Mass from Mixed CAR(1) process model log Power Spectrum Excess Variance Ponti+(2012) log Frequency Friday, May 24, 13 See also, e.g., Yu & Lu (2001), Nikolajuk+(2004), Papadakis (2004), O’Neill+(2005), Miniutti+(2009), Caballero-Garcia+(2012) Estimating Black Hole Mass from Mixed CAR(1) process model log Power Spectrum New method works with X-ray counts. Mass estimates with ~0.3 dex precision. Kelly+(in prep) log Frequency Friday, May 24, 13 See also, e.g., Yu & Lu (2001), Nikolajuk+(2004), Papadakis (2004), O’Neill+(2005), Miniutti+(2009), Caballero-Garcia+(2012) Current Work: More Flexible Stochastic Models dLp (t) dLp 1 (t) + ↵p + . . . + ↵1 L(t) = p 1 dt dt dq ✏(t) + q q dt q dq 1 ✏(t) + . . . + ✏(t) 1 q 1 dt • Continuous-time autoregressive moving average models (CARMA(p,q)) provide flexible modeling of variability • Power spectrum is a sum of Lorentzian functions (cf. X-ray binaries) • Can still derive likelihood function of the data, enables maximum-likelihood or Bayesian inference • Still very fast, computational complexity scales as O(n), but more complex algorithms needed to perform do the ‘fit’ Friday, May 24, 13 Example: Quasar vs Variable Stars Friday, May 24, 13 Example: Quasar vs Variable Stars Quasar Kelly+(in prep) LPV, AGB LPV, RGB 68% Probability Intervals Friday, May 24, 13 Credit: ESO/Kornmesser Main Takeaway Point: Stochastic modeling provides a useful and powerful framework to quantify quasar variability that can be applied to lightcurves of arbitrary sampling and with measurement error. Friday, May 24, 13 Time Domain Stochastic Modeling: Outstanding issues and directions for future work Friday, May 24, 13 Time Domain Stochastic Modeling: Outstanding issues and directions for future work • Multivariate lightcurves: • Explicit modeling of time lags and correlation structure between different wavelengths • Vector-valued CARMA(p,q) processes may provide general framework Friday, May 24, 13 Time Domain Stochastic Modeling: Outstanding issues and directions for future work • Multivariate lightcurves: • Explicit modeling of time lags and correlation structure between different wavelengths • Vector-valued CARMA(p,q) processes may provide general framework • Moving beyond a single stationary Gaussian process: • Direct modeling of stochastic process + flares, other ‘state’ changes (Sobolewska+, in prep) • Using alternatives to Gaussian noise (Emmanaloupolous+2013) • Non-stationary and non-linear models Friday, May 24, 13 Time Domain Stochastic Modeling: Outstanding issues and directions for future work • Multivariate lightcurves: • Explicit modeling of time lags and correlation structure between different wavelengths • Vector-valued CARMA(p,q) processes may provide general framework • Moving beyond a single stationary Gaussian process: • Direct modeling of stochastic process + flares, other ‘state’ changes (Sobolewska+, in prep) • Using alternatives to Gaussian noise (Emmanaloupolous+2013) • Non-stationary and non-linear models Uttley+(2005) Friday, May 24, 13 Time Domain Stochastic Modeling: Outstanding issues and directions for future work • Multivariate lightcurves: • Explicit modeling of time lags and correlation structure between different wavelengths • Vector-valued CARMA(p,q) processes may provide general framework • Moving beyond a single stationary Gaussian process: • Direct modeling of stochastic process + flares, other ‘state’ changes (Sobolewska+, in prep) • Using alternatives to Gaussian noise (Emmanaloupolous+2013) • Non-stationary and non-linear models Friday, May 24, 13 Time Domain Stochastic Modeling: Outstanding issues and directions for future work • Multivariate lightcurves: • Explicit modeling of time lags and correlation structure between different wavelengths • Vector-valued CARMA(p,q) processes may provide general framework • Moving beyond a single stationary Gaussian process: • Direct modeling of stochastic process + flares, other ‘state’ changes (Sobolewska+, in prep) • Using alternatives to Gaussian noise (Emmanaloupolous+2013) • Non-stationary and non-linear models • Building astrophysically-motivated stochastic models • Stochastic partial differential calculations + accretion flow models? Friday, May 24, 13