Bayesian Statistics: hierarchical models. Sanjib Sharma (University of Sydney)

Size: px

Start display at page:

Download "Bayesian Statistics: hierarchical models. Sanjib Sharma (University of Sydney)"

Geoffrey Asher Parks
6 years ago
Views:

1 Bayesian Statistics: hierarchical models Sanjib Sharma (University of Sydney)

2 History of Bayesian statistics

3 Drop a ball on table. Predict its location. Put another ball ask the question is it left or right to the first. Keep doing this. Ultimately, you can narrow down the location to a small area. Initial Belief+ New data Improved Belief Likelihood +prior posterior Posterior becomes prior in the next round.

4 Laplace 1774 Independely rediscoverded. In words rather than Eq, Probability of a cause given an event is proportional to the probability of the event given its cause. p(c E) ~ p(e C) His friend Bouvard used his method to calculate the masses of Saturn and Jupiter. Laplace offered bets of to 1 odd and 1million to 1 that they were right to 1% for Saturn and Jupiter. Even now Laplace would have won both bets.

5 Largely ignored after Laplace till Theory of probability, 1939 by Harold Jeffrey Main reference. In WW-II, used at Bletchley Park to decode German Enigma cipher. There were conceptual difficulties Role of prior Data is random or model parameter is random

6 1950 onwards Tide had started to turn in favor of Bayesian methods. Lack of proper tools and computational power main hindrance. Frequentist methods were simpler which made them popular.

7 Cox 1946 showed that Bayesian prob theory can be derived from two basic rules. p(θ x) ~ p(x θ)p(θ)

9 1950 onwards Metropolis algorithm.

10 N interacting particles. A single configuration ω, can be completely specified by giving position and velocity of all the particles. A point in R2N space. E(ω), total energy of the system For system in equilibrium p(ω) ~ exp (- E(ω) / kt ) Computing any thermodynamic property, pressure, energy etc, requires integrals,which are analytically intractable Start with arbitrary config N particles. Move each by a random walk and compute ΔE the change in energy between old and new config If: ΔE < 0, always accept. Else: accept stochastically with probability exp (- ΔE / kt ) Immediate hit in statistical physics.

11 Hastings 1970 The same method can be used to sample an arbitrary pdf p(ω) by replacing E(ω)/kT ln p(ω) Had to wait till Hastings Generalized the algorithm and derived the essential condition that a Markov chain out to satisfy to sample the target distribution. Acceptance ratio not uniquely specified, other forms exist. His student Peskun 1973 showed that Metropolis gives the fastest mixing rate of the chain

12 MH algorithm q(y xt) f(x) xt y

13 1980 Simulated annealing Kirkpatrick 1983 To solve combinatorial optimization problems using MH algorithm using ideas of annealing from solid state physics. Useful when we have multiple maxima Minimize an objective function C(ω) by sampling from exp(-c(ω)/t) with progressively decreasing T. Allowing selection of globally optimum solution

14 1984 Expectation Maximization (EM) algorithm Dempster 1977 Provided a way to deal with missing data and hidden variables. Hierachical Bayesian models. Vastly increased the range of problems that can addressed by Bayesian methods. Deterministic and sensitive to initial condition. Stochastic versions were developed Data augmentation, Tanner and Wong 1987 Geman and Geman 1984 Introduced Gibbs sampling in the context of image restoration. First proper use of MCMC to solve a problem setup in Bayesian framework.

15 Image: Ryan Adams

16 1990 Gelfand and Smith 1990 Largely credited with revolution in statistics, Unified the ideas of Gibbs sampling, DA algorithm and EM algorithm. It firmly established that Gibbs samling and MH based MCMC algorithms can be used to solve a wide class of problems that fall in the category of hierarchical bayesian models.

17 Citation history of Metropolis et al/ 1953 Physics: well known from Statistics: only 1990 onwards Astronomy: 2002 onwards

18 Astronomy's conversion- 2002

19 Astronomy: Saha & Williams 1994 Christensen & Meyer 1998 Gravitational wave radiation Christensen et al and Knox et al Galaxy kinematics from absorption line spectra. Comsological parameter estimation using CMB data Lewis & Bridle 2002 Galvanized the astronomy community more than any other paper.

20 Lewis & Bridle 2002 Laid out in detail the Bayesian MCMC framework Applied it to one of the most important data sets of the time. Used it to address a significant scientific question- fundamanetal parameters of the universe. Made the code publicly available Making it easier for new entrants.

21 Bayesian hierarchical models p(θ {xi} ) ~ p(θ) i p( xi θ ) θ x0 x1 xn p(θ, {xi} {yi}) ~ p(θ) i p( xi θ ) p(yi xi, σyi) θ Level-0: Population Level-1: Individual Object-intrinsic x0 x1 xn Level-1: Individual Object-observable y0 y1 yn N

22 Test score of students from various schools J schools, nj students in j-th school. yij the score of i-th student in j-th school Y = {yij 0<j<J, 0<i<nj} The dispersion σ in score is known a priori. What we want to know Average score of a school and its uncertainty p(αj). Mean of students in a school (no-pool) Overall average of schools and its dispersion (μ,ω). Global mean without segregating into schools.

23 Average score of a school and its uncertainty p(αj). Mean of students in a school (no-pool) If nj very small, [10,20], estimate is very uncertain Overall average of schools and its dispersion (μ,ω). Global mean without segregating into schools. If one school has too many students, the mean and spread will be dominated by it.

24 BHM Some schools have very few students so uncertainty very high. p(αj y.j) ~ p(y.j αj,σ) p(αj,σ) There is more information in data from other schools. p(αj μ,ω) ~ p(y.j αj,σ) p(αj μ,ω) But, population statistics depends on schools, they are interrelated. We get joint info about population of schools as well as for individual schools. p(α,μ,ω Y) ~ p(μ,ω) j p(αj μ,ω) i p(yij αj,σ)

25 Global mean μ=0,ω=1, J=40, σ =1 and 2<nj<20. Blue - standard group mean (based on data from group) Green - BHM based group mean Green points are systematically closer to the global mean than the blue points. The shift is more for cases where error bars are larger The BHM estimates have smaller error bars. BHM also makes use of information from other groups.

26 Handling uncertainties p(θ, {xti} {xi}, {σxi} ) ~ p(θ) i p( xti θ ) p(xi xti, σx,i) p(xi xti, σx,i) ~ Normal( xi xti, σyi) Level-0: Population θ Level-1: Individual Object-intrinsic xt0 xt1 xtn Level-1: Individual Object-observable x0 x1 xn N

27 Kinematic modelling of MW p(θ, vl',vb',vr', s' vl,vb,vr, l,b,s) ~ p(vl',vb',vr' θ,l,b,s') * p(s s',σs) p(vr vr') p(vl vl')p(vb vb') 6d data from GAIA

28 Missing variables p(θ, {xti} {xi}, {σxi} ) ~ p(θ) i p( xti θ ) p(xi xti, σx,i) p(xi xti, σx,i) ~ Normal( xi xti, σxi) Certain σxi Level-0: Population θ Level-1: Individual Object-intrinsic xt0 xt1 xtn Level-1: Individual Object-observable x0 x1 xn N

29 Kinematic modelling of MW p(θ, vl',vb',vr', s' vr, l,b,s) ~ p(vl',vb',vr' θ,l,b,s') * p(s s',σs) p(vr vr') Spectroscopic data no proper motion.

30 Hidden variables p(θ, {xi} {yi}, {σyi} ) ~ p(θ) i p( xi θ ) p(yi xi, σyi) A function y(x) exists for mapping x y p(yi xi, σyi) ~ Normal( yi y(xi), σyi) Level-0: Population θ Level-1: Individual Object-intrinsic x0 x1 xn Level-1: Individual Object-observable y0 y1 yn N

31 Exoplanets

32 Mean velocity of center of mass v0 Semi-amplitude Time period T Eccentricity e Angle of pericenter from the ascending node Time of passage through the pericenter t

34 Hogg et al 2010

35 Hogg et al 2010

36 3d Extinction- EB-V(s) Pan-STARRS 1 and 2MASS Green et al. 2015

37 3d Extinction maps Obsevables: y=(j,h,k,teff,log g, [M/H],l,b) Intrinsic params: x=([m/h],t,m,s,l,b,e) Distance extinction relationship: p(e s,θ) Posterior: p(x y,σy, α) ~ p(y x, σy) p(x θ) p(θ) Likelihood: p(y x, σy) ~ N(y-y(x),σy) Prior: p(x θ) ~ p(t)p(m)p([m/h])p(s) p(e s,θ) p(θ)

38 How to solve BHM models Two step: Hogg et al p( θ {yi}, σy ) ~ p(α) i dxi p( yi xi, σy,i) p( xi θ ) ~ p(α) i dxi p( yi xi, σy,i ) p(xi) [p(xi θ ) / p( xi )] ~ p(α) i k [ p( xik θ ) / p( xik ) ] xik sampled from p( yi xi, σy,i ) p( xi ) MWG: p( θ, {xi} {yi}, {σyi} ) ~ p(θ) i p( yi xi, σy,i ) p(xi θ)

39 MH algorithm q(y xt) f(x) xt y

40 Image: Ryan Adams

41 Metropolis Within Gibbs Gibbs sampler requires sampling from conditional distribution. Replace this with a MH step. Rather than updating all at one time, one can A compliacated distribution can be broken up into sequence of smaller of easier to samplings is the main strength of this.

42 BMCMC- a python package pip install bmcmc Ability to solve hierarchical Bayesian models. Documentation:

44 Summary Hierarchical Bayesian models allow you to tackle a wide range of problems in astronomy. For more info and Monte Carlo based algorithms to solve Bayesian inference problems see, Sharma 2017 Annual Reviews of Astronomy and Astrophysics,

Principles of Bayesian Inference

Principles of Bayesian Inference Sudipto Banerjee University of Minnesota July 20th, 2008 1 Bayesian Principles Classical statistics: model parameters are fixed and unknown. A Bayesian thinks of parameters