User’s Guide

Bumps is designed to determine the ideal model parameters for a given set of measurements, and provide uncertainty on the parameter values. This is an inverse problem, where measured data can be predicted from theory, but theory cannot be directly inferred from measured data. This means that bumps must search through parameter space, calling the theory function many times to find the parameter values that are most consistent with the data.

Unlike traditional Levenburg-Marquardt fitting programs, Bumps does not require normally distributed measurement uncertainty. If a measurement comes from counting statistics, for example, you can define your model with poisson probability rather than gaussian probability. Parameter values can have constraints. For example, if the size of a sample is known to within 5%, the size parameter in the model can set to a gaussian distribution with a standard deviation of 5%. Simple bounds are also supported. Parameter expressions allow you to set the value of a parameter based on other parameters, which allows simultaneous fitting of multiple datasets to different models without having to define a specialized fit function.

Bumps includes Markov chain Monte Carlo (MCMC) methods to compute the joint distribution of parameter probabilities. These methods require hundreds of thousand function calls to explore the search space, so for moderately complex problems, you need to run in parallel. Bumps can fully utilize multiple cores on one computer, or through MPI, it runs on supercomputing clusters.

# Data handling has been removed so that we can ship a pure python package. In addition to inverse problem solving, bumps has acquired code for theory building and data handling. For example, many problems have measurements in which the instrument resolution plays a role, and the theory function must be convolved with a data dependent resolution function.

Using Bumps

Model scripts associate a sample description with data and fitting options to define the system you wish to refine.

Data Representation

Data is loaded from instrument specific file formats into a generic Probe. The probe object manages the data view and by extension, the view of the theory. The probe object also knows the measurement resolution, and controls the set of theory points that must be evaluated in order to computed the expected value at each point.

Experiment

Sample descriptions and data sets are combined into an Experiment object, allowing the program to compute the expected reflectivity from the sample and the probability that reflectivity measured could have come from that sample. For complex cases, where the sample varies on a length scale larger than the coherence length of the probe, you may need to model your measurement with a CompositeExperiment.

Parameters

The adjustable values in each component of the system are defined by Parameter objects. When you set the range on a parameter, the system will be able to automatically adjust the value in order to find the best match between theory and data.

Fitting

One or more experiments can be combined into a FitProblem. This is then given to one of the many fitters, such as DEFit, which adjust the fitting parameters, trying to find the best fit. See Optimizer Selection for a description of available optimizers and Bumps Options for a description of the bumps options. Entropy can be calculated when the fit is complete. See Calculating Entropy.