Revised Belief Propagation by jack-dunham · Pull Request #26 · ITensor/ITensorNetworksNext.jl

jack-dunham · 2025-11-25T18:15:58Z

This PR express belief propagation in terms of the new interface based on AlgorithmsInterface.jl and the included AlgorithmsInterfaceExtensions.jl library.

Note, the tolerance for the second belief propagation test has been increased from 1.0e-14 to 1.0e-12, however this test was not passing reliable before the changes in this PR anyway.

Introduce `BeliefPropagationProblem` wrapper to hold the cache and the error `diff` field. Also simplifies some kwargs wrangling.

…be set from another cache

Also includes some fixes to the way `TensorNetwork` types are constructed based on index structure.

for more information, see https://pre-commit.ci

…instead of trying to operate on existing graphs The reason for this is: - One only cares about the edges of the input graph - A simple graph cannot be used as it "forgets" its edge names resulting in recursion - As shown with `TensorNetwork`, removing edges may not always be defined.

…s from an array.

…a order-0 array

This was caused by the change to the `cache` being backed by a directed graph.

github-actions · 2026-01-09T19:46:45Z

Your PR no longer requires formatting changes. Thank you for your contribution!

…ble when constructing nested algorithms

Fix broken merge

codecov · 2026-02-13T23:15:13Z

Codecov Report

❌ Patch coverage is 50.27778% with 179 lines in your changes missing coverage. Please review.
✅ Project coverage is 57.88%. Comparing base (6f3521f) to head (fef588d).

Files with missing lines	Patch %	Lines
...eliefpropagation/abstractbeliefpropagationcache.jl	36.78%	55 Missing ⚠️
src/tensornetwork.jl	14.06%	55 Missing ⚠️
src/beliefpropagation/beliefpropagationcache.jl	52.50%	38 Missing ⚠️
src/abstracttensornetwork.jl	22.72%	17 Missing ⚠️
src/beliefpropagation/beliefpropagationproblem.jl	89.21%	11 Missing ⚠️
src/LazyNamedDimsArrays/symbolicnameddimsarray.jl	0.00%	2 Missing ⚠️
src/LazyNamedDimsArrays/lazynameddimsarray.jl	0.00%	1 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (6f3521f) and HEAD (fef588d). Click for more details.

HEAD has 5 uploads less than BASE

Flag BASE (6f3521f) HEAD (fef588d)

6 1

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #26      +/-   ##
==========================================
- Coverage   63.58%   57.88%   -5.71%     
==========================================
  Files          21       21              
  Lines         736     1040     +304     
==========================================
+ Hits          468      602     +134     
- Misses        268      438     +170

Flag	Coverage Δ
docs	`0.00% <0.00%> (ø)`

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

JoeyT1994 · 2026-03-06T15:04:58Z

src/beliefpropagation/abstractbeliefpropagationcache.jl

+    messages = incoming_messages(bp_cache, vertex)
+    state = factors(bp_cache, vertex)
+
+    return (reduce(*, messages) * reduce(*, state))[]


This should be done with a contraction sequence finder. Probably with a default like "greedy" but the ability to overload the sequence kwargs would be ideal.

JoeyT1994 · 2026-03-06T15:06:59Z

src/beliefpropagation/abstractbeliefpropagationcache.jl

+message_type(bpc::AbstractBeliefPropagationCache) = message_type(typeof(bpc))
+message_type(::Type{<:AbstractBeliefPropagationCache{<:Any, <:Any, ED}}) where {ED} = ED
+
+function free_energy(bp_cache::AbstractBeliefPropagationCache)


Maybe this should be called logscalar ? Technically it returns log(scalar(bp_cache)) where scalar(bp_cache) is the BP approximation of the underlying network. The free energy is really -kBTlog(Z) so free_energy might be a misnomer.

Agreed, math terminology would be preferred over physics terminology.

JoeyT1994 · 2026-03-06T15:08:13Z

src/beliefpropagation/beliefpropagationcache.jl

+    return rv
+end
+
+function induced_subgraph_bpcache(graph, subvertices)


Is there a reason this can't ne called induced_subgraph and dispatch based on the type?

It is a convention we are using where induced_subgraph_bpcache is a function defining the implementation of induced_subgraph for bpcache (perhaps this suffix should be renamed to something more explicit). Then if one defines a type NotAnAbstractBeliefPropagationCache that doesn't subtype AbstractBeliefPropagationCache but should behave like an AbstractBeliefPropagationCache (at least for subgraph purposes) one can define:

function induced_subgraph_from_vertices(cache::NotAnAbstractBeliefPropagationCache, subvertices) return induced_subgraph_bpcache(cache, subvertices) end

Note the appropriate function to overload is usually induced_subgraph_from_vertices, which assumes the subvertices argument as already been canonized via to_vertices. In any case, overloading induced_subgraph directly can be lead to annoying method ambiguities as induced_subgraph is a function in Graphs.jl that we have no control over. Matt and I have discussed refining this aspect of the code to move away from induced_subgraph entirely and just having the call stack be:

subgraph(graph, vertices) = subgraph_<implementation>(graph, to_vertices(graph, vertices))

which would simplify things greatly.

JoeyT1994 · 2026-03-06T15:09:57Z

src/beliefpropagation/beliefpropagationproblem.jl

+using NamedGraphs.PartitionedGraphs: quotientvertices
+
+@kwdef struct StopWhenConverged <: AI.StoppingCriterion
+    tol::Float64 = 0.0


Can we use Number instead of Float64. It will be good to be able to use Float32 (and other precisions) arithmetic if we wish.

We can use AbstractFloat.

Just a note that probably we should parametrize the type (we can constrain the parameter to AbstractFloat or Real).

JoeyT1994 · 2026-03-06T15:11:00Z

src/beliefpropagation/beliefpropagationproblem.jl

+
+    if algorithm.normalize
+        # TODO: use `sum` not `norm`
+        message_norm = LinearAlgebra.norm(state.iterate)


message_norm = sum(state.iterate) .

This is important for stability in the presence of complex numbers (avoids blow up of non-PSDness of messages).

JoeyT1994 · 2026-03-06T15:12:56Z

src/beliefpropagation/beliefpropagationproblem.jl

+        maxiter = is_tree(cache) ? 1 : nothing,
+        tol = -Inf,
+        message_diff_function = if tol > -Inf
+            (m1, m2) -> norm(m1 / norm(m1) - m2 / norm(m2))


This message diff is currently dependent on phase, i.e. m1 = -m2 will count as non-converged. 1 - abs2(dot(m1, m2)) might be better, with m1 and m2 normalized.

JoeyT1994 · 2026-03-06T15:15:46Z

test/test_beliefpropagation.jl

+        @assert U * LinearAlgebra.diagm(S) * V ≈ W
+        id = [1.0 0.0; 0.0 1.0]
+        set!(sqrt_Ws, e, id)
+        set!(sqrt_Ws, reverse(e), U * LinearAlgebra.diagm(S) * V)


Currently the SVD is not really playing a role here.
set!(sqrt_Ws, reverse(e), W) would work. Or, probably best, take the square root via the SVD and set U*sqrt(S) on e and sqrt(S)*V on reverse(e).

JoeyT1994 · 2026-03-06T15:16:39Z

test/test_beliefpropagation.jl

+    end
+
+    bpc = BeliefPropagationCache(tn)
+    bpc = ITensorNetworksNext.beliefpropagation(bpc; maxiter = 1)


It would be nice if the maxiter default could be set to 1 if graph(bpc) ==1 and unset otherwise.

@JoeyT1994 what do you mean by graph(bpc) == 1? Shouldn't the check be that it is a tree?

Yes sorry I mean is_tree(graph(bpc))

jack-dunham force-pushed the bp branch from 407983f to 3573d95 Compare November 25, 2025 21:48

JoeyT1994 and others added 29 commits January 6, 2026 09:55

Working BP Commit

b4ae4d7

BP Code

d77d063

Express BP in terms of SweepIterator interface

b80e36e

Introduce `BeliefPropagationProblem` wrapper to hold the cache and the error `diff` field. Also simplifies some kwargs wrangling.

Add method for setmessages! that allows messages from one cache to …

fe44b80

…be set from another cache

Network is now passed to forest_cover_edge_sequence directly.

3ce0898

test file formatting

f6e4fd0

Add DataGraphsPartitionedGraphsExt glue for TensorNetwork type

63840a9

Also includes some fixes to the way `TensorNetwork` types are constructed based on index structure.

Make abstract tensor network interface more generic.

ba22ab5

BP Caching overhauls

49b0870

Remove dead deps

db46c04

Fix merge

400e373

Fix type inference in TensorNetwork construction

b9aafe8

[pre-commit.ci] auto fixes from pre-commit.com hooks

4090e61

for more information, see https://pre-commit.ci

Remove ITensorBase dep

be0750e

[LazyNamedDimsArrays] Fix parenttype method

9ebf031

BP Cache now uses new DataGraphsinterface

16fe303

Adjust default_message to take a message type as its first argument

24a4335

Remove unnecessary code and fix ambiguities in AbstractTensorNetwork

c43884e

TensorNetwork type now uses new DataGraphs interface

dd6f645

Sweeping algorithms based on AlgorithmsInterface.jl (ITensor#30)

7bb579c

Upgrade to NamedDimsArrays.jl v0.11 (ITensor#38)

032447a

[LazyNamedDimsArrays] New symnameddims method that pulls out indice…

b256d79

…s from an array.

The function region_scalar should now return a scalar, rather than …

b2da9d8

…a order-0 array

Fix double counting in edge_scalars function

8506e26

This was caused by the change to the `cache` being backed by a directed graph.

Minor code formatting

938180a

Expressed belief propagation in terms of AlgorithmsInterface

4461967

Fixes to TensorNetwork construction from tensor list

d68860a

Minor simplifications to contract_network interface.

2f5c783

jack-dunham force-pushed the bp branch from 3573d95 to 2f5c783 Compare January 9, 2026 19:41

jack-dunham force-pushed the bp branch 2 times, most recently from a20ddcb to ac74a7f Compare February 11, 2026 14:44

jack-dunham added 8 commits February 13, 2026 17:24

Merge branch 'main' into bp

9a45a5b

Upgrade DataGraphs and NamedGraphs dependencies

4eec9b6

[AlgorithmsInterfaceExtensions] Allowing mapping over a generic itera…

202724c

…ble when constructing nested algorithms

Upgrade serial BP to use own <:Algorithm structs.

69542e3

Simplify BP cache to only store factors

9925069

Upgrade to DataGraphs v0.3.1 and NamedGraphs v0.10

292f2fa

Fix compat

9d937aa

Fix broken merge

5432fe2

Fix broken merge

jack-dunham force-pushed the bp branch from 729052b to 5432fe2 Compare February 13, 2026 23:09

jack-dunham added 4 commits February 19, 2026 17:44

Bug fix; upgrade tests

c916c84

Add 2D TN test

4a511a1

Formatting

5b97af3

Merge branch 'main' into bp

fef588d

JoeyT1994 requested changes Mar 6, 2026

View reviewed changes

Conversation

jack-dunham commented Nov 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Jan 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Feb 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

jack-dunham commented Nov 25, 2025 •

edited

Loading

github-actions bot commented Jan 9, 2026 •

edited

Loading

codecov bot commented Feb 13, 2026 •

edited

Loading