Status: resolved (2026-11-01), by option 1 below. Found while building Mycelium’s sampler; measured and now checked in
lib/Lenticulum.jl/test/lp020.jl. Related: free_energy, gradient_couplings, sampling, Extending Mycelium, LP-002 The Jacobian side is not implemented.
Sources: original repository analysis; code:
lib/Mycelium.jl/src/free_energy.jl(free_energy_gradient,parameter_step),lib/Mycelium.jl/src/factors.jl(factor_belief,belief_free_energy),lib/Lenticulum.jl/src/gaussian.jl(FrozenGaussianBelief),lib/AlgebraicFactors.jl/src/toric.jl(ToricBelief),lib/AlgebraicFactors.jl/src/polynomial.jl(PolynomialBelief),lib/Mycelium.jl/src/sampling.jl(sample_gradient); tests:lib/Lenticulum.jl/test/lp020.jl,lib/AlgebraicFactors.jl/test/lp020.jlTheory (CT-ML wiki): Variational Free Energy
Problem. With every edge on the default diagonal coupling, free_energy_gradient (and so
parameter_step) differentiates each factor’s local free energy with its incoming messages
held fixed. The local free energy recomputes the factor’s belief from those
messages and the parameters, so the belief moves with . Writing
:
- the gradient of the graph’s free energy at a belief-propagation fixed point is (the envelope theorem: the Bethe free energy is stationary in the beliefs, so they are held fixed, not the messages);
- the fixed-message derivative adds , which is not zero in general.
Measured in lib/Lenticulum.jl/test/lp020.jl, on the smallest example: , ,
, observed. On this tree the Bethe free energy equals exactly. At
, :
| () | ||
|---|---|---|
| exact gradient of | 0.679 | −1.383 |
through inference (sample_gradient) | 0.679 | −1.383 |
fixed messages (free_energy_gradient) | −0.146 | −0.441 |
The discrepancy changes the fixed point, not only the step. At the maximum-likelihood
the fixed-message derivative in is −0.574, not 0. Alternating inference with fixed-message
steps in stops at . (In it happens to reach the maximum-likelihood value.) So
parameter_step with diagonal couplings does not, in general, train to maximum likelihood, even on
a tree with exact inference.
Options.
- Differentiate at fixed beliefs. A factor hook for the expected energy of a given belief, , differentiated with frozen. That is the envelope gradient, exact at a fixed point, and as cheap as now. It needs one method per factor family. Closed form for the Gaussian and switching factors (mean and covariance terms); an expectation of the pointwise energy in general.
- Differentiate locally. minus the cross-entropy term . That is generic over factors, but needs that expectation for every belief type (closed form for Gaussian and categorical messages).
- Differentiate through inference. Already available, as
ExactCouplingwith a schedule replay (gradient_couplings.md) or assample_gradient’s re-run of inference per sample. Exact on trees, but costs an inference per gradient evaluation (per coordinate with reference differences).
Resolution (2026-11-01). Option 1, decided by Daniel (“fix LP-20”). Two hooks in Mycelium:
factor_belief(factor, msgs, ps, st) returns the factor’s belief, evaluated once outside AD, and
belief_free_energy(factor, b, ps, st) is at that fixed belief. The graph
gradient differentiates the latter. Lenticulum implements them for GaussianFactor,
GaussianPrior, LinearConstraintFactor, SwitchingConstraintFactor and the collapsed Gaussian
composite, through one frozen joint Gaussian of each factor’s non-point channels
(FrozenGaussianBelief). Theory and limits: free_energy, section “Fixed beliefs, not fixed
messages”.
Checked in lib/Lenticulum.jl/test/lp020.jl:
- on the example above the gradient is now , and alternating inference with it converges to the maximum-likelihood ;
- on four trees (this one; a Gaussian factor and prior with learnable noise; the composite; a switching constraint, including its log-weights) the graph gradient matches central differences of the converged Bethe free energy to , with reference differences, ForwardDiff and Zygote.
What remains.
- Factors without
factor_belief(the neural and diffusion factors) keep the fixed-message derivative. That is exact with point messages, and they warn once when their messages are not points.ToricFactorand the degree-onePolynomialFactorgained the hooks on 2026-11-02, with the same check against central differences (lib/AlgebraicFactors.jl/test/lp020.jl).PriorFactorneeds none, since its energy does not read the parameters. sample_gradientnow uses the fixed-belief gradient per sample by default (one inference per sample); differentiating through inference stays asthrough_inference = true, and the two agree to inlib/Lenticulum.jl/test/switching_constraint.jl(sampling).