A uniqueness theorem for the variational free energy decomposition

Michael P. Rubin

REVIEW 1 major objections 3 minor 15 references

A uniqueness theorem for the variational free energy decomposition

T0 review · 1 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper proves that any exact decomposition of log Z into a separable objective minus a nonnegative gap vanishing only at the Gibbs measure must be the variational free energy with the Kullback-Leibler divergence.

desk verdict A genuinely new characterization of the ELBO/Gibbs-Bogoliubov identity, with a clean proof; the only real blemish is an abstract that overstates what the necessity examples show about hypothesis (R). read the letter →

arxiv 2607.22710 v1 pith:IACRBKOC submitted 2026-07-20 cs.IT math-phmath.ITmath.MPmath.PR

Michael P. Rubin This is my paper · ORCID

classification cs.ITmath-phmath.ITmath.MPmath.PR MSC 82B0394A1762F1539B22

keywords variationalfreeenergyGibbs-Bogoliubovinequalityevidencelowerboundrelativeentropyuniquenesstheoremfunctionalequationsalpha-divergencedecomposition

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks how much freedom exists in the identity log Z = -F(Q) + D(Q||π) that underlies the Gibbs–Bogoliubov inequality and variational inference. It proves that if a functional G is additively separable into a reference-measure term and a weight term, and the residual Δ is nonnegative and vanishes exactly at the Gibbs measure π, then G must be the variational free energy F(Q;p,ℓ)=D(Q||p)-E_Q[logℓ] and Δ must be the Kullback-Leibler divergence D(Q||π). The proof reduces the assumptions to a measurable group homomorphism and shows only the standard split survives. The significance is that objectives built from alpha- or Rényi divergences can provide bounds on log Z but can never yield an exact identity of this form; the ELBO identity is structural.

What carries the argument

The central device is the group homomorphism χ_Q: (P,·)→(R,+) defined by χ_Q(v)=B'(Q,ℓv)-B'(Q,ℓ), which the regularity assumption (R) promotes to a linear functional <m_Q, log v>. The proof then uses scale invariance (giving Σ m_Q = 0), the fact that the residual depends only on the Gibbs measure, and the second-order vanishing of relative entropy near Q = π, to force m_Q = 0 and hence G = F, Δ = D(·||π).

What would settle it

Construct a decomposition satisfying (D), (S), and (P) using a nonmeasurable additive map φ: R^n → R with φ(1)=0, giving G ≠ F and Δ ≠ D(·||π); that would refute the conclusion if (R) is dropped. Alternatively, a proof that any such φ must be unbounded below on the hypersurface {h: ⟨Q,e^h⟩=1} would show (R) is redundant and strengthen the theorem.

Watch

Extended reading notes

Core claim

Theorem 1.1: Let (G,Δ) satisfy log Z = -G + Δ for every trial Q and model (p,ℓ), with G additively separable (G = A(Q,p)+B(Q,ℓ), B measurable in ℓ) and Δ ≥ 0 with equality iff Q = π. Then G must be the variational free energy F(Q;p,ℓ) = D(Q||p) - E_Q[logℓ] and Δ(Q,π) = D(Q||π). The split A/B is unique up to a Q-dependent constant a(Q) that cancels in G. Equivalently, no divergence other than the Kullback-Leibler divergence can serve as the residual in an exact, additively separable decomposition of log Z.

Load-bearing premise

The most fragile premise is the regularity assumption (R), that the weight-dependent term B(Q,ℓ) is Lebesgue measurable in ℓ; without it, unmeasurable additive functions of the log-weights could in principle break the conclusion, and the paper leaves open whether strictness alone would exclude them.

Editorial extensions

If this is right

Objectives based on α- or Rényi divergences cannot be arranged into an exact identity of the form log Z = -G + Δ with Δ a strict divergence and G separable; the absence of an identity is structural.
In mean-field theory and variational inference, the functional to optimize is forced once one demands an exact decomposition; the only modelling freedom is the trial class.
The additive split A(Q,p)+B(Q,ℓ) is ambiguous by a Q-dependent gauge that cancels in G; only the sum is pinned down.
The bound and the identity are not two grades of the same thing: KL divergence delivers both, alternative divergences deliver only bounds.
The proof mechanism suggests the result extends beyond finite state spaces with bounded densities and a suitable topology, though this is not carried out.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the open regularity question is resolved negatively (i.e., nonmeasurable homomorphisms can satisfy all other hypotheses), then the theorem's uniqueness would fail only in nonconstructive, Hamel-basis territory, leaving all computable applications untouched.
One can read the theorem as a foundational justification for why the ELBO, rather than any alpha-divergence objective, is the canonical exact variational identity; this could motivate reconsidering variational methods that abandon exact decompositions.
The gauge freedom suggests that constant offsets in log-likelihoods or baseline energies are absorbed by the Q-dependent split, which may clarify why additive constants in energy functions are harmless in practice.
Because the theorem characterizes decompositions rather than divergences, it could pair with axiomatic characterizations of relative entropy to give a two-sided picture: KL is the unique divergence and the unique exact-decomposition residual.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: this paper proves something real and new. The prior literature characterizes divergences or inference rules; this is the first result I know that characterizes which functionals admit an exact decomposition of log Z as a computable part plus a strict divergence in the Gibbs measure. Theorem 1.1 and Corollary 4.2 are the payload, and they are worth having. I traced the proof in Section 5 and it is sound. Step 3's ratio-realization argument is valid, the measurable-homomorphism step is standard Aczel, and Step 5's perturbation correctly forces m_Q = 0. The gauge freedom for A and B is handled honestly, and the examples in Section 6 do real work. I also give credit for explicitly flagging the open questions: Question 6.6 and Remarks 6.7 and 7.3 are not buried.

The soft spot is the status of (R), and the abstract oversells it. The paper says "counterexamples show that each hypothesis is needed," but that is only true for (S) and (P). For (R), Section 6 does not give a counterexample; it shows that dropping (R) lets Hamel-basis homomorphisms survive, and then asks in Question 6.6 whether (P) would kill them automatically. The theorem as stated is proved correctly, so this is not a flaw in the main result. But the minimality claim is stronger than what is established. If (R) turns out to be redundant, the theorem is a three-hypothesis theorem; if not, the extra hypothesis is doing essential work and the necessity story should say so. Either way, the abstract needs adjustment, and the body should acknowledge that the necessity of (R) is unresolved, not demonstrated by counterexample.

Everything else checks out. The distinction between characterizing divergences and characterizing decompositions is well drawn, and the Discussion's separation from convex duality is fair. The finite-state restriction is reasonable for a mechanism proof, and the authors say they expect no obstruction in the infinite case without claiming to have checked it.

Who should read this: anyone working on variational inference, mean-field theory, or the foundations of the ELBO. It is not a computational paper; it is a structural theorem. It deserves a serious referee, and with a minor revision on the abstract and a clearer statement about (R) I would accept it. Send it to peer review.

Referee Report

1 major / 3 minor

Summary. The paper proves a uniqueness theorem for exact decompositions of log Z(p, ℓ) into a computable variational objective and a residual divergence. On a finite alphabet, it shows that if (G, Δ) satisfies log Z = -G + Δ, with G additively separable as A(Q,p)+B(Q,ℓ), with ℓ ↦ B(Q,ℓ) Lebesgue measurable for each Q, and with Δ ≥ 0 vanishing exactly at Q = π_{p,ℓ}, then G must be the variational free energy F(Q;p,ℓ)=D(Q∥p)-E_Q[log ℓ] and Δ must be D(Q∥π), up to a Q-dependent gauge a(Q) that cancels in G. The proof reduces the hypotheses to a measurable group homomorphism χ_Q:(P,·)->(R,+) and eliminates it by combining linear representation with a second-order expansion of D(Q∥π_ε) and the nonnegativity of Δ. The paper also gives examples showing that the separability and strictness hypotheses are needed and that (D) alone is vacuous; the status of the measurability hypothesis (R) is left explicitly open.

Significance. If accepted, this is a satisfying structural result: the ELBO/Gibbs–Bogoliubov identity is the unique exact separable decomposition with a strict residual depending on the model only through the Gibbs measure. The proof is short, self-contained, and I find it correct conditional on (R). The corollary that α- and Rényi-divergence objectives cannot be arranged to satisfy such an exact identity under (S) and (P) is a useful, falsifiable consequence. The paper's main advertised caveat is the unresolved status of (R), which affects the sharpness claim but not the truth of Theorem 1.1 as stated.

major comments (1)

[Abstract; Section 6; Remark 6.5; Question 6.6] The abstract and the opening of Section 6 claim that counterexamples show each hypothesis is needed, but no counterexample for (R) is provided. Example 6.3 shows (S) is needed and Example 6.4 shows (P) is needed; Remark 6.5 and Question 6.6 explicitly leave open whether a nonmeasurable Hamel-basis homomorphism can be completed to a decomposition satisfying (P). Thus the necessity/redundancy of (R) is unresolved, and the advertised statement that all hypotheses are necessary is stronger than what is proved. Please revise the abstract and Section 6 to say, e.g., that (S) and (P) are shown necessary, while (R) is used essentially in the proof but its optimality remains open.

minor comments (3)

[Section 7, Remark 7.3] The extension beyond finite X is described as expected but 'not carried out'. This is an honest limitation, but the sentence 'We expect no obstruction' should be labeled as a conjecture to avoid being read as a result.
[Section 5, Step 3] The range assertion for p' ↦ Σ_x tv(x)p'(x) is correct for finite X, but the proof is terse; adding one sentence on concentrating mass near the coordinates attaining the min and max would improve readability.
[Section 6, Example 6.4] In the concrete case X={1,2}, m=(1,-1), the displayed expression Δ(Q,π)=D(Q∥π)+log(π_1/π_2) is negative whenever π_1<π_2 and Q=π; consider saying 'for Q=π with π_1<π_2' explicitly, since otherwise the sentence is a bit compressed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the proof expands around the known identity and forces the residual to zero; the unresolved status of (R) is a completeness concern, not circularity.

full rationale

The paper's derivation is not circular. Proposition 3.1 is a direct, verified identity for the KL-based decomposition. Step 1 defines A′ and B′ by subtracting exactly that known solution, so the proof's task is to show the deviation must vanish under hypotheses (D), (S), (R), and (P). Nothing in (D), (S), (R), or (P) states G=F or Δ=D(·∥π); indeed Examples 6.3 and 6.4 exhibit pairs satisfying (D) and some hypotheses while violating the conclusion, showing the hypotheses do not bake in the target. The proof's use of the classical theorem that measurable additive maps are linear is an explicit, external functional-equations result, not a self-citation or an imported uniqueness theorem from the authors. The only significant caveat is Remark 6.5/Question 6.6: the paper neither proves that (R) is necessary nor that it is redundant, so the abstract's claim that counterexamples show every hypothesis is needed is stronger than what is demonstrated. That is a hypothesis-optimality or completeness issue, not a circularity issue: the conditional statement of Theorem 1.1 remains independent of its conclusion. There is no fitted parameter presented as a prediction, no self-definitional construction, and no ansatz smuggled in via prior work. Honest non-finding: score 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard functional-equation facts (measurable additive maps are linear; log is a group isomorphism (P, .) -> (R^X, +)) and on the three explicitly stated hypotheses (S), (R), (P), each shown necessary in Section 6. No numbers are fitted to data; the only freedom in the conclusion is the Q-dependent gauge a(Q) in the intermediate split A/B, which the theorem shows cancels in G and is therefore not a free parameter of the result. There are no invented entities. The abstract's claim that counterexamples show 'each hypothesis is needed' is precise for (S) and (P); for (R) Section 6 demonstrates only proof-mechanism failure, leaving necessity open (Question 6.6).

assumptions (6)

standard math Measurable additive maps R^X -> R are R-linear (Aczel, Lectures on Functional Equations, Ch. 2)
Invoked in Step 3 of the proof (Section 5) to convert the measurable group homomorphism chi_Q : (P, .) -> (R,+) into the linear functional <m_Q, log .> in Eq. (7); imported background, not proven in the paper.
standard math Gibbs/de Finetti inequality: D(Q||R) >= 0 with equality iff Q = R (Cover-Thomas)
Used to state the variational principle after Proposition 3.1 and to make hypothesis (P) coherent; standard textbook result.
domain assumption Hypothesis (S): G(Q;p,ell) = A(Q,p) + B(Q,ell), additive separability
The substantive structural hypothesis of Theorem 1.1 (Section 4); the paper shows it is necessary by Example 6.3 (G = 2D(Q||p) - 2E_Q[log ell] + log Z violates (S)).
domain assumption Hypothesis (R): ell -> B(Q,ell) is Lebesgue measurable for each fixed Q
Regularity excluding Hamel-basis pathologies; necessity is unresolved (Remark 6.5, Question 6.6): no counterexample is given, only failure of the proof mechanism.
domain assumption Hypothesis (P): Delta(Q,pi) >= 0 with equality iff Q = pi
Strictness assumption; shown necessary by Example 6.4 (Delta = D + <m, log pi> with sum m = 0 satisfies (D)+(S)+(R) but violates (P)).
standard math Every strictly positive probability vector pi is realizable as a Gibbs measure (take p = pi, ell = 1)
Used in Step 4 (Section 5) to extend Eq. (11) from Gibbs measures to all pairs (Q, pi) in Delta^o x Delta^o.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A uniqueness theorem for the variational free energy decomposition." pith.science (2026). https://pith.science/paper/IACRBKOC

@misc{pith2026260722710,
  author       = {Pith},
  title        = {Pith review of: A uniqueness theorem for the variational free energy decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IACRBKOC}},
  note         = {Machine review of arXiv:2607.22710}
}

abstract

For a finite system with reference measure $p$ and positive weight $\ell$, the variational free energy $F(Q;p,\ell)=D(Q\,\|\,p)-E_Q[\log\ell]$ satisfies the exact identity $\log Z(p,\ell)=-F(Q;p,\ell)+D(Q\,\|\,\pi)$, where $Z$ is the partition function and $\pi$ the associated Gibbs measure. For a uniform reference measure this is the Gibbs-Bogoliubov inequality of mean-field theory; for a Bayesian model it is the evidence decomposition of variational inference. Variational objectives built from $\alpha$- or R\'enyi divergences retain useful bounds on $\log Z$ but not identities of this form, which raises the question of which functionals admit an exact decomposition. We prove the following characterization. Suppose a pair $(G,\Delta)$ satisfies $\log Z=-G+\Delta$ with $\Delta$ a function of $Q$ and $\pi$ alone; suppose $G$ is additively separable into a term depending on the reference measure and a term depending on the weight, with mild regularity; and suppose $\Delta$ is nonnegative and vanishes precisely at $Q=\pi$. Then $G=F$ and $\Delta=D(\cdot\,\|\,\pi)$. The proof reduces the hypotheses to a homomorphism from the multiplicative group of positive functions into $(\mathbb{R},+)$ and removes it using the second-order vanishing of relative entropy at its minimum. Counterexamples show that each hypothesis is needed. The result characterizes the decomposition rather than the divergence, and is therefore complementary to the axiomatic characterizations of relative entropy due to Shore-Johnson, Csisz\'ar and Amari, which it neither uses nor extends.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references

[1]

Aczél,Lectures on Functional Equations and Their Applications, Academic Press, New York, 1966

J. Aczél,Lectures on Functional Equations and Their Applications, Academic Press, New York, 1966

1966
[2]

S. M. Ali and S. D. Silvey,A general class of coefficients of divergence of one distri- bution from another, J. Roy. Statist. Soc. Ser. B28(1966), 131–142

1966
[3]

Amari,α-divergence is unique, belonging to bothf-divergence and Bregman di- vergence classes, IEEE Trans

S.-i. Amari,α-divergence is unique, belonging to bothf-divergence and Bregman di- vergence classes, IEEE Trans. Inform. Theory55(2009), no. 11, 4925–4931

2009
[4]

Amari and H

S.-i. Amari and H. Nagaoka,Methods of Information Geometry, Transl. Math. Monogr., vol. 191, Amer. Math. Soc. and Oxford Univ. Press, 2000

2000
[5]

D. M. Blei, A. Kucukelbir and J. D. McAuliffe,Variational inference: a review for statisticians, J. Amer. Statist. Assoc.112(2017), 859–877

2017
[6]

L. M. Bregman,The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming, USSR Comput. Math. Math. Phys.7(1967), 200–217

1967
[7]

T. M. Cover and J. A. Thomas,Elements of Information Theory, 2nd ed., Wiley, 2006

2006
[8]

Csiszár,Information-type measures of difference of probability distributions and indirect observations, Studia Sci

I. Csiszár,Information-type measures of difference of probability distributions and indirect observations, Studia Sci. Math. Hungar.2(1967), 299–318

1967

Show all 15 references

[9]

M. I. Jordan, Z. Ghahramani, T. S. Jaakkola and L. K. Saul,An introduction to variational methods for graphical models, Machine Learning37(1999), 183–233

1999
[10]

Li and R

Y. Li and R. E. Turner,Rényi divergence variational inference, Advances in Neural Information Processing Systems29(2016), 1073–1081

2016
[11]

Minka,Divergence measures and message passing, Microsoft Research Technical Report MSR-TR-2005-173, 2005

T. Minka,Divergence measures and message passing, Microsoft Research Technical Report MSR-TR-2005-173, 2005

2005
[12]

Opper and D

M. Opper and D. Saad (eds.),Advanced Mean Field Methods: Theory and Practice, MIT Press, Cambridge, MA, 2001

2001
[13]

J. E. Shore and R. W. Johnson,Axiomatic derivation of the principle of maximum entropy and the principle of minimum cross-entropy, IEEE Trans. Inform. Theory26 (1980), 26–37

1980
[14]

Uffink,Can the maximum entropy principle be explained as a consistency require- ment?, Stud

J. Uffink,Can the maximum entropy principle be explained as a consistency require- ment?, Stud. Hist. Philos. Modern Phys.26(1995), 223–261

1995
[15]

M. J. Wainwright and M. I. Jordan,Graphical models, exponential families, and variational inference, Found. Trends Mach. Learn.1(2008), 1–305. Wyss Institute for Biologically Inspired Engineering, Har v ard Univer- sity, Boston, MA 02115, USA Email address:Michael.Rubin@Wyss.H...

2008

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Aczél,Lectures on Functional Equations and Their Applications, Academic Press, New York, 1966

J. Aczél,Lectures on Functional Equations and Their Applications, Academic Press, New York, 1966

1966

[2] [2]

S. M. Ali and S. D. Silvey,A general class of coefficients of divergence of one distri- bution from another, J. Roy. Statist. Soc. Ser. B28(1966), 131–142

1966

[3] [3]

Amari,α-divergence is unique, belonging to bothf-divergence and Bregman di- vergence classes, IEEE Trans

S.-i. Amari,α-divergence is unique, belonging to bothf-divergence and Bregman di- vergence classes, IEEE Trans. Inform. Theory55(2009), no. 11, 4925–4931

2009

[4] [4]

Amari and H

S.-i. Amari and H. Nagaoka,Methods of Information Geometry, Transl. Math. Monogr., vol. 191, Amer. Math. Soc. and Oxford Univ. Press, 2000

2000

[5] [5]

D. M. Blei, A. Kucukelbir and J. D. McAuliffe,Variational inference: a review for statisticians, J. Amer. Statist. Assoc.112(2017), 859–877

2017

[6] [6]

L. M. Bregman,The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming, USSR Comput. Math. Math. Phys.7(1967), 200–217

1967

[7] [7]

T. M. Cover and J. A. Thomas,Elements of Information Theory, 2nd ed., Wiley, 2006

2006

[8] [8]

Csiszár,Information-type measures of difference of probability distributions and indirect observations, Studia Sci

I. Csiszár,Information-type measures of difference of probability distributions and indirect observations, Studia Sci. Math. Hungar.2(1967), 299–318

1967