Skip to content

MultiNest is a Bayesian inference tool which calculates the evidence and explores the parameter space which may contain multiple posterior modes and pronounced (curving) degeneracies in moderately high dimensions.

farhanferoz/MultiNest

Repository files navigation

MultiNest

MultiNest Farhan Feroz, Mike Hobson f.feroz@mrao.cam.ac.uk arXiv:0704.3704, arXiv:0809.3437 & arXiv:1306.2144 Released Nov 2019


MultiNest Licence

Users are required to accept to the licence agreement given in LICENCE file.

Users are also required to cite the MultiNest papers (arXiv:0704.3704, arXiv:0809.3437 & arXiv:1306.2144) in their publications.


Required Libraries:

MultiNest requires lapack. To use MPI support, some MPI library must also be installed.


MPI Support:

The code is MPI compatible. In order to disable the MPI parallelization, remove -DMPI compilation flag.


gfortran compiler:

You might need to use the following flag while compiling MultiNest with gfortran compiler to remove the restriction imposed by gfortran on line length.

-ffree-line-length-none


The subtoutine to begin MultiNest are as follows:

subroutine nestRun(IS, mmodal, ceff, nlive, tol, efr, ndims, nPar, nCdims, maxModes, updInt, Ztol, root, seed, pWrap, feedback, resume, outfile, initMPI, logZero, maxiter, loglike, dumper, context)

logical IS !do Importance Nested Sampling (INS)?

logical mmodal !do mode separation?

integer nlive !number of live points

logical ceff !run in constant efficiency mode

double precision tol !evidence tolerance factor

double precision efr !sampling efficiency

integer ndims !number of dimensions

integer nPar !total no. of parameters

integer nCdims !no. of parameters on which clustering should be performed (read below)

integer maxModes !maximum no. of modes (for memory allocation)

integer updInt !iterations after which the output files should be written

double precision Ztol !null log-evidence (read below)

character(LEN=1000) root !root for MultiNest output files

integer seed !random no. generator seed, -ve value for seed from the sys clock

integer pWrap[ndims] !wraparound parameters?

logical feedback !need update on sampling progress?

logical resume !resume from a previous run?

logical outfile !write output files?

logical initMPI !initialize MPI routines?, relevant only if compiling with MPI. Set it to F if you want your main program to handle MPI initialization

double precision logZero !points with loglike < logZero will be ignored by MultiNest

integer maxiter !max no. of iterations, a non-positive value means infinity. MultiNest will terminate if either it has done max no. of iterations or convergence criterion (defined through tol) has been satisfied

loglike(Cube,ndims,nPar,lnew) !subroutine which gives lnew=loglike(Cube(ndims))

dumper(nSamples,nlive,nPar,physLive,posterior, paramConstr,maxloglike,logZ,INSlogZ,logZerr,c) !subroutine called after every updInt*10 iterations with the posterior distribution, parameter constraints, max loglike & log evidence values integer context. Not required by MultiNest, any additional information user wants to pass


likelihood routine: slikelihood(Cube,ndims,nPar,lnew,context)

Cube(1:nPar) has nonphysical parameters.

Scale Cube(1:n_dim) & return the scaled parameters in Cube(1:n_dim) & additional parameters that you want to be returned by MultiNest along with the actual parameters in Cube(n_dim+1:nPar).

Return the log-likelihood in lnew.


Dumper routine: dumper(nSamples,nlive,nPar,physLive,posterior,paramConstr,maxloglike,logZ,INSlogZ,logZerr,context)

This routine is called after every updInt*10 iterations & at the end of the sampling allowing the posterior distribution & parameter constraints to be passed on to the user in the memory. The argument are as follows:

nSamples = total number of samples in posterior distribution

nlive = total number of live points

nPar = total number of parameters (free + derived)

physLive(nlive, nPar+1) = 2D array containing the last set of live points (physical parameters plus derived parameters) along with their loglikelihood values

posterior(nSamples, nPar+2) = posterior distribution containing nSamples points. Each sample has nPar parameters (physical + derived) along with the their loglike value & posterior probability

paramConstr(1, 4nPar): paramConstr(1, 1) to paramConstr(1, nPar) = mean values of the parameters paramConstr(1, nPar+1) to paramConstr(1, 2nPar) = standard deviation of the parameters paramConstr(1, nPar2+1) to paramConstr(1, 3nPar) = best-fit (maxlike) parameters paramConstr(1, nPar4+1) to paramConstr(1, 4nPar) = MAP (maximum-a-posteriori) parameters

maxLogLike = maximum loglikelihood value

logZ = log evidence value from the default (non-INS) mode

INSlogZ = log evidence value from the INS mode

logZerr = error on log evidence value

context= not required by MultiNest, any additional information user wants to pass

The 2D arrays are Fortran arrays which are different to C/C++ arrays. In the example dumper routine provided with C & C++ eggbox examples, the Fortran arrays are copied on to C/C++ arrays.


Tranformation from hypercube to physical parameters:

MultiNest native space is unit hyper-cube in which all the parameter are uniformly distributed in [0, 1]. User is required to transform the hypercube parameters to physical parameters. This transformation is described in Sec 5.1 of arXiv:0809.3437. The routines to tranform hypercube parameters to most commonly used priors are provided in module priors (in file priors.f90).


Checkpointing:

MultiNest is able to checkpoint. It creates [root]resume.dat file & stores information in it after every updInt iterations to checkpoint, where updInt is set by the user. If you don't want to resume your program from the last run run then make sure that you either delete [root]resume.dat file or set the parameter resume to F before starting the sampling.


Periodic Boundary Conditions:

In order to sample from parameters with periodic boundary conditions (or wraparound parameters), set pWrap[i], where i is the index of the parameter to be wraparound, to a non-zero value. If pWrap[i] = 0, then the ith parameter is not wraparound.


Constant Efficiency Mode:

If ceff is set to T, then the enlargement factor of the bounding ellipsoids are tuned so that the sampling efficiency is as close to the target efficiency (set by efr) as possible. This does mean however, that the evidence value may not be accurate.


Sampling Parameters:

The recommended paramter values to be used with MultiNest are described below. For detailed description please refer to the paper arXiv:0809.3437

nPar:

Total no. of parameters, should be equal to ndims in most cases but if you need to store some additional parameters with the actual parameters then you need to pass them through the likelihood routine.

efr:

defines the sampling efficiency. 0.8 and 0.3 are recommended for parameter estimation & evidence evalutaion respectively.

tol:

A value of 0.5 should give good enough accuracy.

nCdims:

If mmodal is T, MultiNest will attempt to separate out the modes. Mode separation is done through a clustering algorithm. Mode separation can be done on all the parameters (in which case nCdims should be set to ndims) & it can also be done on a subset of parameters (in which case nCdims < ndims) which might be advantageous as clustering is less accurate as the dimensionality increases. If nCdims < ndims then mode separation is done on the first nCdims parameters.

Ztol:

If mmodal is T, MultiNest can find multiple modes & also specify which samples belong to which mode. It might be desirable to have separate samples & mode statistics for modes with local log-evidence value greater than a particular value in which case Ztol should be set to that value. If there isn't any particularly interesting Ztol value, then Ztol should be set to a very large negative number (e.g. -1.d90).


Progress Monitoring:

MultiNest produces [root]phys_live.dat & [root]ev.dat files after every updInt iterations which can be used to monitor the progress. The format & contents of these two files are as follows:

[root]phys_live.dat:

This file contains the current set of live points. It has nPar+2 columns. The first nPar columns are the ndim parameter values along with the (nPar-ndim) additional parameters that are being passed by the likelihood routine for MultiNest to save along with the ndim parameters. The nPar+1 column is the log-likelihood value & the last column is the node no. (used for clustering).

[root]ev.dat:

This file contains the set of rejected points. It has nPar+3 columns. The first nPar columns are the ndim parameter values along with the (nPar-ndim) additional parameters that are being passed by the likelihood routine for MultiNest to save along with the ndim parameters. The nPar+1 column is the log-likelihood value, nPar+2 column is the log(prior mass) & the last column is the node no. (used for clustering).


Posterior Files:

These files are created after every updInt*10 iterations of the algorithm & at the end of sampling.

MultiNest will produce five posterior sample files in the root, given by the user, as following

[root].txt

Compatable with getdist with 2+nPar columns. Columns have sample probability, -2*loglikehood, parameter values. Sample probability is the sample prior mass multiplied by its likelihood & normalized by the evidence.

[root]post_separate.dat

This file is only created if mmodal is set to T. Posterior samples for modes with local log-evidence value greater than Ztol, separated by 2 blank lines. Format is the same as [root].txt file.

[root]stats.dat

Contains the global log-evidence, its error & local log-evidence with error & parameter means & standard deviations as well as the best fit & MAP parameters of each of the mode found with local log-evidence > Ztol.

[root]post_equal_weights.dat

Contains the equally weighted posterior samples. Columns have parameter values followed by loglike value.

[root]summary.txt

There are nmode+1 (nmode = number of modes) rows in this file. First row has the statistics for the global posterior. After the first line there is one row per mode with nPar*4+2 values in each line in this file. Each row has the following values in its column mean parameter values, standard deviations of the parameters, bestfit (maxlike) parameter values, MAP (maximum-a-posteriori) parameter values, local log-evidence, maximum loglike value. If IS = T (i.e. INS being used), first row has an additional value right at the end, INS log-evidence estimate.


INS Output Files:

In INS mode (when IS = T), MultiNest will produce 3 additional binary files: [root]IS.iterinfo, IS.points, IS.ptprob These files are used for resuming job with IS (Importance Nested sampling) mode set to T. They can be quite large and can be deleted once the job has finished.


Birth contour files:

MultiNest produces [root]phys_live-birth.txt & [root]dead-birth.txt files after every updInt iterations which can be used to reconstruct a full nested sampling run, as well as simulate dynamic nested sampling. The format & contents of these two files are as follows:

[root]phys_live-birth.txt:

This file contains the current set of live points. It has nPar+3 columns. The first nPar columns are the ndim parameter values along with the (nPar-ndim) additional parameters that are being passed by the likelihood routine for MultiNest to save along with the ndim parameters. The nPar+1 column is the log-likelihood value. The nPar+2 column is the log-likelihood value that the point was born at & the last column is the node no. (used for clustering). This is identical to the [root]phys_live-birth.dat file, except for an additional column including the birth contours

[root]dead-birth.txt:

This file contains the set of rejected points. It has nPar+4 columns. The first nPar columns are the ndim parameter values along with the (nPar-ndim) additional parameters that are being passed by the likelihood routine for MultiNest to save along with the ndim parameters. The nPar+1 column is the log-likelihood value, the nPar+2 column is the log-likelihood value that the point was born at and the nPar+2 column is the log(prior mass) & the last column is the node no. (used for clustering). This is identical to the [root]ev.dat file, except for an additional column including the birth contours.


C/C++ Interface:

Starting in MultiNest v2.18 there is a common C/C++ interface to MultiNest. The file is located in the "includes" directory. Examples of how it may be used are found in the example_eggboxC and example_eggboxC++ directories.

You may need to check the symbol table for your platform (nm libmulitnest.a | grep nestrun) and edit the multinest.h file to define NESTRUN. Please let us know of any modifications made so they may be included in future releases.


Python Interface:

Johannes Buchner has written an easy-to-use Python interface to MultiNest called PyMultiNest which provides integration with existing scientific Python code (numpy, scipy). It allows you to write Prior & likelihood functions in Python. It is available from:

https://github.com/JohannesBuchner/PyMultiNest


R Interface:

Johannes Buchner has written an R bridge for MultiNest called RMultiNest. It allows likelihood functions written in R (http://www.r-project.org) to be used by MultiNest. It is available from:

https://github.com/JohannesBuchner/RMultiNest


Matlab version of MultiNest:

Matthew Pitkin and Joe Romano have created a simple Matlab version of MultiNest. It doesn't have all the bells-and-whistles of the full implementation and just uses the basic Algorithm 1 from arXiv:0809.3437. It is available for download the MultiNest website:

http://ccpforge.cse.rl.ac.uk/gf/project/multinest/


Visualization of MultiNest Output:

[root].txt file created by MultiNest is compatable with the format required by getdist package which is part of CosmoMC package. Refer to the following website in order to download or get more information about getdist: http://cosmologist.info/cosmomc/readme.html#Analysing

Johannes Buchner's PyMultiNest can also be used on exisiting MultiNest output to plot & visualize results. https://github.com/JohannesBuchner/PyMultiNest


Toy Problems

There are 8 toy programs included with MultiNest.

example_obj_detect:

The object detection problem discussed in arXiv:0704.3704. The positions, amplitudes & widths of the Gaussian objects can be modified through params.f90 file. Sampling parameters are also set in params.f90.

example_gauss_shell:

The Gaussian shells problem discussed in arXiv:0809.3437. The dimensionality, positions and thickness of these shells can be modified through params.f90 file. Sampling parameters are also set in params.f90.

example_gaussian:

The Gaussian shells problem discussed in arXiv:1001.0719. The dimensionality of the problem can be modified through params.f90 file. Sampling parameters are also set in params.f90.

example_eggboxC/example_eggboxC++:

The C/C++ interface includes the egg box problem discussed in arXiv:0809.3437. The toy problem and sampling parameters are set in eggbox.c & eggbox.cc files.

example_ackley:

The Ackley mimimization problem (see T. Bäck, Evolutionary Algorithms in Theory and Practice, Oxford University Press, New York (1996).)

example_himmelblau:

The Himmelblau's minimization problem. (see http://en.wikipedia.org/wiki/Himmelblau's_function)

example_rosenbrock:

Rosenbrock minimization problem. (see http://en.wikipedia.org/wiki/Rosenbrock_function)

example_gaussian:

Multivariate Gaussian with uncorrelated paramters.


Common Problems & FAQs:

  1. MultiNest crashes after segmentation fault.

Try increasing the stack size (ulimit -s unlimited on Linux) & resume your job.

  1. Output files (.txt & post_equal_weights.dat) files have very few (of order tens) points.

If tol is set to a reasonable value (tol <= 1) & the job hasn't finished then it is possible for these files to have very few points in them. If even after the completion of job these files have very few points then increase the stack size (ulimit -s unlimited on Linux) & resume the job again.

  1. Not all modes are being reported in the stats file.

stats file reports all the modes with local log-evidence value greater than Ztol. Set Ztol to a very large negative number (e.g. -1.d90) in order for the stats file to report all the modes found.

  1. Compilation fails with error something like 'can not find nestrun function'.

Check the symbol table for your platform (nm libnest3.a | grep nestrun) & edit multinest.h file to define NESTRUN appropriately.

  1. When to use the constant efficiency mode?

If the sampling efficiency with the standard MultiNest (when ceff = F) is very low & no. of free parameters is relatively high (roughly greater than 30) then constant efficiency mode can be used to get higher sampling efficiency. Users should use ask for around 5% or lower targer efficiency (efr <= 0.05) in constant efficiency mode & ensure that posteriors are stable when a slightly lower value is used for target efficiency. Evidence values obtained with constant efficiency mode will not be accurate in the default mode (IS = F). With INS (IS = T) accurate evidences can be calculated (read arXiv:1306.2144) for more details.


Importance Nested Sampling:

Sampling Parameters:

INS can calculate an order of magnitude more accurate evidence values than the vanilla nested sampling, for the same number of live points & target efficiency (efr). INS can also calculate reasonably accurate evidence values in constant efficiency mode (when ceff = T). However, users should always check that the posterior distributions & evidence values are stable for reasonable change in these sampling parameters. Potentially users can use fewer live points &/or higher target efficiency (efr) with INS to get the same level of accuracy on evidence & therefore analysis can be sped-up.

Quoted Error on Log-Evidence:

The quoted error on log-evidence in INS mode is generally an overestimate. Users should use the quoted error on log-evidence from vanilla nested sampling as an upper bound on the error on INS log-evidence.

Output Files:

In INS mode, MultiNest will produce 3 additional binary output files (see output files section) in which it will record information about all the points it has collected. These files can be quite large. It is recommended that the user sets updint parameter (which sets the number of iterations after which posterior files are updated) to 1000 or more otherwise a lot of time might be wasted in writing these large files. These binary files are only used while resuming jobs & can be deleted once the job has finished.

Memory Requirements:

INS requires a lot more memory than the defult mode (when IS = F). With nlive <= 1000, memory requirements are feasible on modern computers but with nlive >= 4000, segfaults may results due to not enough memory available.

Multi-Modal Mode:

Multi-modal mode is not supported yet when IS = T. If mmodal = T along with IS = T, MultiNest will set mmodal = F.


About

MultiNest is a Bayesian inference tool which calculates the evidence and explores the parameter space which may contain multiple posterior modes and pronounced (curving) degeneracies in moderately high dimensions.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published