Status: closed (2026-10-26), option 2. Found while checking
AlgebraicFactors.ToricFactor(L-SPEC-004 Algebraic Factors Toric and Grassmannian). Related: free_energy, toric, L-SPEC-001 The Factor Interface. Code:lib/Mycelium.jl/src/free_energy.jl(messages_into,factor_free_energies),lib/Mycelium.jl/src/factors.jl(PriorFactor).
Sources: original analysis; code:
lib/Mycelium.jl/src/factors.jl,free_energy.jl,lib/Mycelium.jl/test/runtests.jl,lib/AlgebraicFactors.jl/test/runtests.jl.Theory (CT-ML wiki): Variational Free Energy
Closure (2026-10-26)
Option 2 is implemented with the public trait Mycelium.uses_cavity(factor), default false.
Factors whose local_free_energy reads incoming messages override it as true in their own
package. factor_free_energies checks their edges with absorbs(edge.direction) and throws
ArgumentError for an Emitting() edge, naming the factor, channel and variable and asking
for Bidirectional(). Both graph scalar_free_energy and bethe_free_energy use that path.
A stored nothing on a permitted edge is unchanged. Message passing and schedules are unchanged.
Mycelium’s “LP-014: cavity messages and edge directions” checks the trait values, all three
graph entry points, diagnostic, permitted missing messages, and a directed data/relay/loss
graph’s unchanged free energy. AlgebraicFactors’ “LP-014: a prior on an Emitting edge” replaces
the original @test_broken with @test_throws ArgumentError, keeps exact marginals checked
against enumeration in both directions, and checks the Bidirectional free energy against
. Its “tree BP and Bethe partition function” retains the larger tree checks.
History: original report (2026-10-25)
Problem. The Bethe factor belief is . messages_into reads
from the store, and an Emitting() edge carries no variable-to-factor message, so
it returns nothing for that channel. PriorFactor.local_free_energy then uses its own belief
alone as the factor belief, instead of the prior combined with the cavity message. The marginals
are unaffected (the outgoing message is right), but the Bethe free energy is wrong, silently.
On a tree of ToricFactors with a categorical PriorFactor on one leaf, the free energy
matches by enumeration when the prior’s edge is Bidirectional(), and differs by an
amount of order one when it is Emitting(); the testset “tree BP and Bethe partition function”
in lib/AlgebraicFactors.jl/test/runtests.jl records both (the second as @test_broken). A
DataFactor on an Emitting() edge is not affected: its belief is a Dirac either way.
Connecting a prior with Emitting() is the natural way to write it, and nothing warns.
Options.
- Let Mycelium send variable-to-factor messages along every edge for the free-energy computation, whatever its direction (the direction then constrains scheduling only).
- Throw (or warn) in
factor_free_energieswhen a factor whose free energy depends on the cavity sits on an edge without a variable-to-factor message. - Document that free energies need
Bidirectional()edges for every non-Dirac factor.
Leaning. Option 2 now, so the error is not silent; option 1 if direction is meant to be a scheduling hint only.