Matrix Product Belief Propagation

This repository contains the code relative to the paper Matrix Product Belief Propagation for reweighted stochastic dynamics over graphs (PNAS, arxiv). When possible, variable names match the notation used there. An errata file is also found here.

Usage requires Julia version >= 1.8, although 1.9 or 1.10 are recommended. You can download Julia here.

To install the package:

using Pkg; Pkg.add("MatrixProductBP")

The package is based on TensorTrains.jl for the operations on tensor trains (a.k.a. matrix product states).

The directory is organized as follows:

• The main module MatrixProductBP implements the BP equations. Essential ingredients are:

  • An IndexedBiDiGraph on which the system is defined (for more details, see the documentation of IndexedGraphs.jl)
  • BPFactors living on the factor nodes of the graph. Any concrete subtype (e.g. SISFactor) must implement a method (fᵢ::SISFactor)(xᵢᵗ⁺¹::Integer, xₙᵢᵗ::Vector{<:Integer}, xᵢᵗ::Integer) that computes the contribution of that factor to the joint probability as a function of $x_i^{t+1},\boldsymbol{x}_{\partial i}^{t}, x_i^{t}$.
  • Optionally, single-node reweightings ϕ and edge reweightings ψ

  Some allowed operations:

  • iterate!(bp::MPBP; kw...)
  • pair_beliefs(bp::MPBP)
  • beliefs(bp::MPBP)
  • bethe_free_energy(bp::MPBP)

  There is some montecarlo (Soft Margin) utility in src/sampling/:

  • sample(bp::MPBP, nsamples::Integer): draw samples from the prior weighted with their likelihood
  • draw_node_observations!(bp::MPBP, nobs::Integer): draw samples from the prior and add them to the MPBP object as reweighting

  Code for computing probabilities exactly by exhaustive enumeration in src/exact/ (only for small system size and final time):

  • exact_prob(bp::MPBP)

• A submodule Models that can be included with

using MatrixProductBP.Models

that implements two models: Glauber dynamics and the SIS model of epidemic spreading. It contains model definitions and specialized versions of the BP equations that are more efficient than the generic one.

• A test folder with small examples

• A notebooks folder where results and plots are produced. The notebooks generating the results contained in the article output generate some data in .jld format that is then read by notebooks/plots_for_paper.ipynb which produces the figures in the article.

To get a hang of how the package works, take a look at, e.g. test/glauber_small_tree.jl, notebooks/glauber_infinite_graph_large_degree_small_beta.ipynb, or notebooks/sis_compare_softmargin.ipynb.