Power Spectrum Estimation and Likelihood Function
Transcription
Power Spectrum Estimation and Likelihood Function
Power Spectrum Estimation and Likelihood Function Luca Pagano May 12, 2015 L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 1 / 30 Power Spectrum Estimation A scalar field ∆T (n) defined over the full sky can be decomposed in spherical harmonic coefficients Z ∗ a`m = dn∆T (n)Y`m (n), (1) with, ∆T (n) = ` X X a`m Y`m (n). (2) `>0 m=−` Assuming ∆T Gaussian distributed, and where hC` i ≡ C`th ha`m i = 0, (3) ha`m a`∗0 m0 i = δ``0 δmm0 hC` i, (4) is specified by the theory of primordial perturbations. (5) L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 2 / 30 Power Spectrum Estimation An unbiased estimator of C`th is given by C` = ` X 1 |a`m |2 . 2` + 1 m=−` (6) C` -s are χ2ν -distributed with the mean equal to C`th , ν = 2` + 1 degrees of freedom (dof), and a 2 variance of 2C`th /ν. In the case of non-fullsky CMB measurements a position dependent weighting W (n) must be applied to the measured data Z 1 dnW i (n) fsky wi = 4π 4π (7) is the i-th moment of the arbitrary weighting scheme. The window function can also be expanded in spherical harmonics with the coefficients R ∗ (n), and with a power spectrum w`m = dnW (n)Y`m 2 w 2 and for which W(` = 0) = 4πfsky 1 L UCA PAGANO W` = 1 X |w`m |2 , 2` + 1 m P W(`)(2` + 1) = 4πfsky w2 . `≥0 PSE AND L IKELIHOOD (8) M AY 12, 2015 3 / 30 Power Spectrum Estimation A sky temperature fluctuation map ∆T (n) on which a window W (n) is applied can be decomposed in spherical harmonics coefficients Z ∗ ã`m = dn∆T (n)W (n)Y`m (n) X ∗ ≈ Ωp ∆T (p)W (p)Y`m (p), (9) (10) p where the integral over the sky is approximated by a discrete sum over the pixels that make the map, with an individual surface area Ωp . e can be defined as The pseudo power spectrum C e` = C L UCA PAGANO ` X 1 |ã`m |2 . 2` + 1 m=−` PSE AND L IKELIHOOD (11) M AY 12, 2015 4 / 30 Power Spectrum Estimation e ` is clearly different from the full sky angular spectrum C` , but their The pseudo power spectrum C ensemble averages can be related by X e` i = M``0 hC`0 i, (12) hC `0 where M``0 describes the mode-mode coupling resulting from the cut sky. This kernel depends only on the geometry of the cut sky and can be expressed simply in terms of the power spectrum W` of the spatial window applied to the survey. Starting from: Z ã`m = ∗ dn∆T (n)W (n)Y`m (n) = X Z a`0 m0 ∗ dnY`0 m0 (n)W (n)Y`m (n) = (13) `0 m 0 = X a`0 m0 K`m`0 m0 [W ], (14) `0 m0 where the kernel K describes the mode-mode coupling resulting from the sky weighting. L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 5 / 30 Power Spectrum Estimation Note that ã is a linear combination of Gaussian variables and is therefore Gaussian as well, but the ã`m -s are not independent. P If we use the series representaion of the window function, W (n) = `m w`m Y`m (n), the coupling kernel reads Z K`1 m1 `2 m2 ≡ dnY`1 m1 (n)W (n)Y`∗2 m2 (n) (15) Z X = w`3 m3 dnY`1 m1 (n)Y`3 m3 (n)Y`∗2 m2 (n) `3 m3 = X w`3 m3 (−1)m2 `3 m3 × `1 0 `2 0 (2`1 + 1)(2`2 + 1)(2`3 + 1) 4π `3 `1 `2 `3 , 0 m1 −m2 m3 with M`1 `2 = 2`2 + 1 X `1 (2`3 + 1)W`3 0 4π ` `2 0 `3 0 1/2 (16) 2 (17) . 3 L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 6 / 30 Power Spectrum Estimation The effect of the instrumental beam, experimental noise, and filtering of the TOD stream can be included as follows X e` i = e` i, hC M``0 F`0 B`20 hC`0 i + hN (18) `0 where B` is a window function describing the combined smoothing effects of the beam and finite e` i is the average noise power spectrum, We seek the solution such that pixel size, hN hC` i = `(` + 1)hC` i/2π is a piece-wise constant. If we replace hC` i by Q`b Pbl 0 hC`0 i = Q`b hCb i then −1 e e hCb i = Kbb (19) 0 Pb 0 ` hC` i − hN` i , where Kbb0 L UCA PAGANO = Pb` K``0 Q`0 b0 , = Pb` M``0 F`0 B`20 Q`0 b0 . PSE AND L IKELIHOOD M AY 12, 2015 7 / 30 Likelihood function It’s possible to evaluate the likelihood of the data for a given theoretical model exactly from the temperature and polarization maps. The standard likelihood is given by h i ~ t (S + N)−1 m ~ exp − 12 m ~ dm ~ |S)d m ~ = L(m , (20) 1/2 3n |S + N| (2π) p /2 ~ is the data vector containing the data, S and N are the signal and noise covariance where m matrix respectively. Computational prohibitive for high resolution maps In case of full sky observation and uncorrelated noise we can re-write it as: n o ~ |C` ) = (2` + 1) Tr [Ĉ` C`−1 ] + ln |C` | L(m (21) Issue: likelihood in case of cut sky and inomogeneus noise L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 8 / 30 Likelihood function To approximate the likelihood on the cut-sky, the usual approach when analysing the CMB temperature is to develop a form for the log likelihood that is quadratic in some function of the C` , and hence can easily be generalized to the cut-sky using an estimate of the C` covariance matrix. At large `, Eq. (21) is approximated by a symmetric Gaussian distribution where the variance is determined by the estimators themselves: 2` + 1 2 " − 2 ln LS (C` |Ĉ` ) = 2` + 1 2 " −2 ln LQ (C` |Ĉ` ) = Ĉ` − C` #2 Ĉ` . (22) . (23) Or by the theory: Ĉ` − C` C` #2 This distribution is closer to the true likelihood. It is often somewhat misleadingly referred to as the ‘Gaussian approximation’, even though it does not have the determinant term required for P(Ĉl |C` ) to be a normalized Gaussian distribution. Another possibility is − 2 ln Lf (C` |Ĉ` ) = 2` + 1 2 " Ĉ` − C` Cf ` #2 = −2 ln LPlanck (C` |Ĉ` ), (24) where Cf ` is some fixed fiducial model assumed to be smooth and close to the model C` under consideration. This is more interesting as although the shape of the likelihood is wrong at any `, gives results consistent with the exact likelihood function. L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 9 / 30 Likelihood function Adding a C` -dependent determinant term to the quadratic approximation can also produce valid results. " #2 2` + 1 Ĉ` − C` −2 ln LD (C` |Ĉ` ) = + ln |C` |. (25) 2 C` Beyond these quadratic/Gaussian approximations, other approximations that have been used include the log-normal distribution where the log-likelihood is quadratic in the log of the power −2 ln LLN (C` |Ĉ` ) = " 2` + 1 ln 2 Ĉ` C` !#2 (26) . This distribution is also somewhat biased: it only matches the exact full-sky result to second order in Ĉl /C` − 1. A weighted combination of the quadratic and the log-normal distributions can be a more accurate approximation to the exact likelihood, being correct to third order in Ĉl /C` − 1. This approximation was adopted in the analysis of WMAP data at high `: ln LWMAP (C` |Ĉ` ) = L UCA PAGANO 2 1 ln LQ (C` |Ĉ` ) + ln LLN (C` |Ĉ` ). 3 3 PSE AND L IKELIHOOD (27) M AY 12, 2015 10 / 30 WMAP Likelihood function - high-` Why is this likelihood third order correct? Writing Ĉ` as C` (1 + ). We can expand the exact expression for the likelihood around its maximum. Then, for a single multipole ` " # 2 3 −2 ln LExact = (2` + 1) [ − ln(1 + )] ' (2` + 1) − + O(4 ) 2 3 We note that the Gaussian likelihood approximation is equivalent to the above expression truncated at 2 " # 2 −2 ln LGauss = (2` + 1) 2 The expression for the log-normal likelihood approximation is: " # 2 3 −2 ln LLN = (2` + 1) − 2 2 (28) (29) (30) Thus, the WMAP approximation of the likelihood function is given by: ln LWMAP (C` |Ĉ` ) = L UCA PAGANO 1 2 ln LGauss (C` |Ĉ` ) + ln LLN (C` |Ĉ` ). 3 3 PSE AND L IKELIHOOD (31) M AY 12, 2015 11 / 30 WMAP Likelihood function - low-` Page,L. et al. (2007). Three-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Polarization Analysis. The Astrophysical Journal Supplement Series, 170(2), 335–376. L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 12 / 30 Panck Likelihood function - high-` Again factorized in low-` and high-` parts: where Different masking for each channel. More aggressive for high frequencies. Planck Collaboration. (2013). Planck 2013 results. XV. CMB power spectra and likelihood. Astronomy & Astrophysics, 571(0), 62. L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 13 / 30 Panck Likelihood function - high-` Masks. Point Sources and galactic L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 14 / 30 Panck Likelihood function - high-` Residual Foregrounds fitted at power spectrum level. Taking into account: Unresolved point sources Clustered galaxies Thermal Sunyaev-Zeodovich Kinetic Sunyaev-Zeodovich Cross-correlation between tSZ and Cosmic Infrared Background Inter calibration parameters L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 15 / 30 Panck Likelihood function - high-` L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 16 / 30 Panck Likelihood function - high-` L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 17 / 30 Panck Likelihood function - high-` The Poisson point source contribution to the 143?217 spectrum Kinetic SZ: the kSZ template used Cross-spectra between frequencies νi and νj , the tSZ template Cosmic infrared background Thermal-SZ/CIB cross-correlation L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 18 / 30 Panck Likelihood function - high-` L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 19 / 30 Panck Likelihood function - high-` L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 20 / 30 Panck Likelihood function - low-` At low-` real space likelihood: M = C + N where C is: L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 21 / 30 Panck Spectrum L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 22 / 30 Theoretical Power Spectrum We are able now to assign to a given C` a likelihood. How can we connect the theoretical C` to the cosmological parameters? Using a Boltzmann code: Code for Anisotropies in the Microwave Background (CAMB) http://camb.info/readme.html L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 23 / 30 Esercizio Stimare l’ampiezza dello spettro di potenza AS Dati: http://irsa.ipac.caltech.edu/data/Planck/release_1/ancillary-data/ previews/COM_PowerSpect_CMB_R1.10/COM_PowerSpect_CMB_R1.10.txt Generare uno spettro di potenza: http://lambda.gsfc.nasa.gov/toolbox/tb_camb_form.cfm In realtà già tutto fatto, trovate dati e spettro qui: https://dl.dropboxusercontent.com/u/17494978/ Stimare il parametro in approssimazione di likelihood uncorrelated Gaussian Media e intervalli di confidenza Ampiezza parametro lineare L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 24 / 30 Metodi per il best-fit e Stima delle incertezze Diversi metodi Local Minimization: Partendo da un modello si utilizza un algoritmo di minimizzazione iterativa (Levenberg-Marquardt). Simulated Annealing: Migliora l’algoritmo di minimizzazione introducendo perturbazioni random che vengono gradualmente raffreddate Genetic Algoritms: Ottimizzazione di algoritmi di evoluzione biologica Dobbiamo descrivere le incertezze sulla stima dei parametri del modello Discuteremo nel dettaglio due metodi Fit dei dati simulati MCMC, algoritmo Metropolis-Hastings L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 25 / 30 Teorema di Bayes e applicazione Il teorema di Bayes ci dice che p(~ d|~x )p(~x ) p(~x |~ d) = R ~ p(d|~x )p(~x )d ~x (32) Scegliendo una prior piatta p(~x ) ∼ 1 si ottiene che la posterior coincide con la likelihood p(~ d|~x ) ∼ exp(−χ2 (~x )/2) (33) Si utlizza il Monte Carlo Markov Chain MCMC è un metodo efficiente per analisi Bayessiana E’ particolarmente indicato nel caso di problemi a molti parametri, in quanto è più performante del semplice campionamento a step E’ veloce e affidabile Si basa sulla costruzione di una catena di stati randomizzati che per un grande numero di step converge alla distribuzione di probabilità L’algoritmo più utilizzato è il Metropolis-Hastings L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 26 / 30 L’algoritmo Metropolis-Hastings L’algoritmo Metropolis-Hastings puù essere implementato nel seguente modo: 1 2 3 4 5 Inizializzare la catena in ~ x0 a n = 0 Proporre un nuovo stato ~ x 0 che avrà come probabilità di transizione q(~ x 0 |~ xn ) Calcolare f (~ x 0 )/f (~ xn ) ∼ exp(−[χ2 (~ x 0 ) − χ2 (~ xn )]/2) Generare un numero random, distribuzione uniforme tra 0 e 1 n 0u, da una o f (~ x )q(~ x |~ x0) Se u ≤ α(~ x 0 |~ xn )) ≡ min f (~x )q(~xn0 |~x ) , 1 allora ~ xn+1 = ~ x 0 altrimenti ~ xn+1 = ~ xn n 6 7 n n =n+1 Ricominciare dal passo 2 La principale complicazione pratica di questo metodo è la scelta di q(~x 0 |~x ) Una scelta sbagliata può portare ad una lenta convergenza del metodo La scelta più efficiente possibile sarebbe la posterior p(~x 0 |~ d) stessa La scelta più comune per q(~x 0 |~x ) è una Gaussiana centrata in ~x Per incrementare l’efficienza e l’esplorazione dello spazio dei parametri, si può modificare il codice in modo da cambiare solo uno o due parametri più correlati in uno step L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 27 / 30 Parameter Estimation - Real World Explore the parameter space Public code Cosmological MonteCarlo (CosmoMC) CosmoMC is a Fortran 2008 Markov-Chain Monte-Carlo (MCMC) engine for exploring cosmological parameter space http://cosmologist.info/cosmomc/ L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 28 / 30 Parameter Estimation - Real World Parametrization Convergence diagnostics using getdist Data: CMB SN BAO MPK HST L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 29 / 30 External Datasets CMB: Ground: BICEP, ACT, SPT... Balloon: Boomerang, MAXIMA, Spider, EBEX... All small scales observation Astrophysical: SN Ia: Several data collection Observable: Luminosity distance BAO: SDSS, BOSS, 2dF Observable: periodic fluctuations in the density of the visible baryonic matter of the universe as "standard ruler" for length scale in cosmology. Angular diameter distance MPK: SDSS, 2dF Observable: Matter power spectrum HST: Cepheids measurements Observable: Luminosity distance L UCA PAGANO PSE AND L IKELIHOOD M AY 12, 2015 30 / 30