MCMC Class Reference

Markov Chain Monte Carlo simulator. More...

#include <mcmc.h>

List of all members.

Public Member Functions
	MCMC (const std::vector< Signal > &signals, TNtuple *data)
	~MCMC ()
TNtuple *	operator() (unsigned nsteps, float burnin_fraction, unsigned sync_interval=1000)
hemi::Array< float > *	build_lut (const std::vector< Signal > &signals, TNtuple *data)
Protected Member Functions
void	nll (const float *v, double *nll, double *event_partial_sums, double *event_total_sum)
Private Attributes
unsigned	nsignals
	number of signals
unsigned	nevents
	number of events
unsigned	nnllblocks
	number of cuda blocks for nll partial sums
unsigned	nllblocksize
	size of cuda blocks for nll partial sums
unsigned	nnllthreads
	number of threads for nll partial sums
unsigned	nreducethreads
	number of threads to use in partial sum reduction kernel
unsigned	blocksize
	size of blocks for per-signal kernels
unsigned	nblocks
	number of blocks for per-signal kernels
std::string	varlist
	string identifier list for ntuple indexing
hemi::Array< float > *	expectations
	signal rate expectation values
hemi::Array< float > *	constraints
	signal rate gaussian constraints
hemi::Array< float > *	lut
	Event/PDF probability lookup table.
hemi::Array< RNGState > *	rngs
	CURAND RNGs, ignored in CPU mode.

---

Detailed Description

Markov Chain Monte Carlo simulator.

Given a set of signal PDFs and a dataset, random walk to map out the likelihood space.

---

Constructor & Destructor Documentation

MCMC::MCMC	(	const std::vector< Signal > &	signals,
		TNtuple *	data
	)

Constructor

Parameters:

	signals	List of Signals defining the PDFs and expectations
	data	The dataset as a TNtuple with fields "e:r"

MCMC::~MCMC

(

)

Destructor

Free HEMI arrays

---

Member Function Documentation

hemi::Array< float > * MCMC::build_lut	(	const std::vector< Signal > &	signals,
		TNtuple *	data
	)

Build Pj(xi) lookup table

Build a table with nevents rows and nsignals columns, containing the value of normalized PDF j evaluted at event i.

Parameters:

	signals	Signals defining the PDFs
	data	TNtuple defining the dataset

Returns:

HEMI array of probabilities, indexed as noted above

void MCMC::nll	(	const float *	v,
		double *	nll,
		double *	event_partial_sums,
		double *	event_total_sum
	)			[protected]

Evaluate the NLL function

-logL = sum(Nj) + 1/2*sum((r-r')^2/s^2) - sum(log(sum(Nj*Pj(xi))))

Nothing is returned -- the output stays in the nll array, so it can stay on the device.

This is done in three steps, to support running on the GPU:

1. Compute partial sums of chunks of events for the last term 2. Total up partial sums from step 1 3. Add normalization and other constraints with sum from step 2

Parameters:

	v	Parameter vector at which to evaluate
	nll	Container for output NLL value
	event_partial_sums	Pre-allocated buffer for event term calculation
	event_total_sum	Pre-allocated buffer for event term total

TNtuple * MCMC::operator()	(	unsigned	nsteps,
		float	burnin_fraction,
		unsigned	sync_interval = 1000
	)

Perform walk.

Parameters:

	nsteps	Number of random-walk steps to take
	burnin_fraction	Fraction of initial steps to throw out
	sync_interval	How often to copy accepted from GPU to storage

Returns:

TNtuple containing accepted points and their likelihoods

---

The documentation for this class was generated from the following files:

• mcmc.h
• mcmc.cpp