vincecr0ft / pynfold Goto Github PK

Iterative ("Bayesian"; as proposed by D'Agostini)

• notebook on base algorithm
• backend implementation
• verification of algorithm
• documentation of implementation

The priority of this issue depends on how well the numpy linear algebra solutions perform. If calculations are particularly slow then this might need bumping up.

• expand pynfold base class to use theano for matrix algebra.

Expand base functionality to include features not directly classified as 'unfolding' but that uses the same data stored within the base class.

• identify any expansions

A simple inversion of the response matrix without regularisation.

• A notebook to detail the base algorithm
• Implementation in pynfold backend
• description and examples of implementation.

Implement Tikonov regularisation algorithm in pynfold. First iteration might use a notebook depending on the status of the pynfold backend milestone.

• Implement algorithm

PyFold should either be able to take in ROOT Histograms AND numpy arrays or use root_numpy to convert algorithms such that the histograms can be compared bin to bin.

• Devise method for comparing results
  • Bin counting (if algorithm is deterministic)
  • χ² test (otherwise)
• Compare results

Singular value decomposition (SVD; as proposed by Höcker and Kartvelishvili and implemented in TSVDUnfold)

• Notebook describing base algorithm
• implementation in pynfold backend
• documentation and example of implementation

A clone of the TUnfold method developed by Stefan Schmitt

• A notebook detailing the base algorithm
• Implementation in PynFold backend
• description and examples of implementation

Should be initialisable and contain:

• relevant Histogram objects

• response/migration matrix

• error handling in case object defined without necessary properties.

• set priors (nominally flat)

• plotting functionality (separate issue?)

Avoid ROOT dependencies inside, let's have it be pure python. If people want to use ROOT, they should use something else to convert ROOT histograms into numpy format.

OneHist should just be a numpy array.

TwoHist should just be a numpy matrix (don't call it a 2-d histogram).

The bin-by-bin unfolding method aka simple correction factors.

• Notebook detailing the base algorithm
• implementation in pynfold backend
• documentation and examples of implementation

The nomenclature of these fields might be misleading. This is a regularised unfolding, but also iterative. The precise algorithm used should be exactly the same

• if there is no RooUnfold equivalent find another package that is.
• write a notebook to describe algorithm with rigour.

make folders for:

• backend
• examples
• notebooks for work in progress

As described here

• make new issue to document/implement

IDS (iterative dynamically stabilized) unfolding found here.

• make new issue to document/implement

To enable comparisons between PynFold and RooUnfold it should be possible to read several input formats and provide a custom Histogram object.

• check if numpy array (else)

• check if numpy histogram (else)

• check if ROOT installed and if so is THF

• apply histogram operations (subtract, multiply, divide etc.)

• ensure that operations propagate errors correctly.

• Basic support for 1 dimensional histograms of equal size (truth and measured)

• 1 dimensional histograms of different size

• Basic 2 dimensional histograms (response matrix)

• multidimensional hists (truth and measured)