Kevin Stevenson AST 4762/5765

What is MCMC?  Random sampling algorithm  Estimates model parameters and their uncertainty  Only samples regions of high probability rather than uniform sampling  Faster  More efficient  Region is called “phase space”

Phase Space  Space in which all possible states of a system are represented  Each space corresponds to one unique point  Every parameter (or DoF) is represented by an axis  Eg. 3 position vectors (x, y, z) require a 3- dimensional phase space  Eg. Add time to produce a 4-D phase space  Can be represented very easily in Python using arrays

Markov Chain  A stochastic (or random) process having the Markov property  Indeterminate future, evolution is described by probability distributions  “Given the present state, future states are independent of the past states”  In other words…  At a given step, the system has a set of parameters that define its state  At the next step, the system might change states or it might remain in the same state according to a certain probability  Each prospective step is determined ONLY by its current state (no past memory)

Example: Random Walk  Consider a drunk standing under a lamppost trying to get home  He takes a step in a random direction (N, E, S, W), each having equal probability  Having forgotten his previous step, he again takes a step in a random direction  Forms a Markov chain

Random Walk Methods  Metropolis-Hastings algorithm  Vary all parameters simultaneously  Accept step with a certain probability  Gibbs sampling  Special (usually faster) case of M-H  Hold all parameters constant, except one  Vary parameter to find best fit  Choose next parameter and repeat  Slice sampling  Multi-try Metropolis

Avoiding Random Walk  May want stepper to avoid doubling back  Faster convergence  Harder to implement  Methods  Successive over-relaxation  Variation on Gibbs sampling  Hybrid Monte Carlo  Introduces momentum

Metropolis-Hastings Algorithm  Goal: want to estimate model parameters and their uncertainty  M-H algorithm generates a sequence of samples from a probability distribution that is difficult to sample from directly  Distribution may not be Gaussian  May not know distribution at all  How does it generate this set?

Preferential Probability  Want to visit a point x with a probability proportional to some given distribution functions, π(x)  “Probability distribution” or “target density”  Preferentially samples where π(x) is large  Probability distribution:  Probability of x falling within a particular interval  Ergodic  Must, in principle, be able to reach every point in the region of interest

Let Me Propose…  Proposal distribution/density:  Depends on current state, x 1  Generates a new proposed sample, x 2  Must also be ergodic  Can be approximated by a Gaussian centered around x 1  May be symmetric:

Target & Proposal Densities  P(x) = target density  Q(x,x t ) = proposal density

Don’t We All Want To Feel Accepted?  Acceptance probability:  If α ≥ 1:  Accept the proposed step  Current state becomes x 2  If α < 1:  Accept step with probability α  Reject step with probability 1 – α  State remains at x 1

Not Too Hot, Not Too Cold  Acceptance rate: fraction of accepted steps  Want an acceptance rate of 30 – 70%  Too high => slow convergence  Too low => small sample size  Must tune the proposal density, Q, to obtain an acceptable acceptance rate  If Q is Gaussian the we tune the standard deviation, σ  Think of σ as a step size

What Is π?

Where to start  Some starting positions are better than others  The equilibrium distribution is rapidly approached from any starting position, x 0  Proof: Due to ergodicity, choosing any point as the starting point is equivalent to jumping into the equilibrium distribution chain at that particular point in time  Suggestions for choosing a starting point:  Best or mean parameters from previous run  Least squares fit (scipy.optimize)  Several starting locations from corners of phase space

Infinite Iterations  How long do you run MCMC?  As # iterations -> ∞, algorithm converges to a precise value  This is NOT the true value, but the best apparent value for the dataset  Run MCMC long enough to:  Forget initial conditions (burn-in)  Characterize your distribution  Error in your parameter mean is smaller than observed dispersion in you Markov Chain

Burn-in  Need burn-in to “forget” the starting position  Remove AT LEAST the first 2% of the total run length  Better yet, look at your data!

Through The Fire And Flames  Remaining set of states represent a sample from the distribution π(x)  Compute the mean (or median) and error of each parameter in your set  Use every m th step for computations and histograms where m should be longer than the correlation time scale between steps  m ~ 10 – 100  Relation between apparent and true values is indicated by the width of the distribution  Plot histogram to see shape  Fit Gaussian to determine width and, hence, error

Now Here We Stand  Recap  Chosen our proposal distribution, initial parameters and number of iterations  Ran MCMC and removed burn-in portion  Determined the mean/median of the apparent values  Computed their errors  What’s next?  Plug those parameters into the model  Analyze your results (do science!!!)