Affine calculus for constrained minima of the Kullback-Leibler divergence

Giovanni Pistone

REVIEW 2 major objections 4 minor 27 references

Affine calculus for constrained minima of the Kullback-Leibler divergence

T0 review · 2 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The gradient of the KL divergence is a pair of dual affine coordinates

desk verdict A clean, mostly expository KL-gradient calculus whose variational-Bayes example is a Euclidean gradient flow, not the advertised natural-gradient flow, and needs fixing before review. read the letter →

arxiv 2502.02177 v3 pith:L2IHKH2S submitted 2025-02-04 math.ST stat.TH

Giovanni Pistone This is my paper · ORCID

classification math.STstat.TH MSC 62B1062B05

keywords affinecalculusKullback-Leiblerdivergencenaturalgradientstatisticalbundlemean-fieldapproximationvariationalBayesduallygeometryJensen-Shannon

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

The paper's aim is to turn the constrained minimization of the Kullback-Leibler divergence into a problem of affine calculus on the open probability simplex. Its central claim is Proposition 1: the total natural gradient of the KL divergence between two probability functions $q$ and $r$ is the pair $(-\operatorname{s}_q(r), -\eta_r(q))$, where $\operatorname{s}_q(r)$ is the exponential-chart coordinate (a centered log-likelihood ratio) and $\eta_r(q)$ is the mixture-chart coordinate (a centered density ratio). From this one identity, the paper derives gradient flows, entropy and cross-entropy gradients, the Jensen\u2013Shannon gradient, mean-field approximation gradients on product spaces, and a variational Bayes gradient flow. A sympathetic reader would care because the formula packages familiar machine-learning computations in a coordinate-free language where Fisher score, natural gradient, and conditional expectation all have exact affine expressions.

What carries the argument

The statistical bundle $S\mathcal{E}(\Omega)$ carries two dual affine atlases: the exponential chart $s_p(q) = \log(q/p) - E_p[\log(q/p)]$ and the mixture chart $\eta_p(q) = q/p - 1$. These charts turn densities and density ratios into centered random variables, and parallel transports $eU$ and $mU$ relate the fibers. The cumulant function $K_p(v) = \log E_p[e^v]$, whose first derivative is an expectation and whose second derivative is a covariance, is the computational tool that converts the KL divergence into readable gradient formulas. Proposition 1's identity $\mathrm{grad}\,D(q\|r) = (-s_q(r), -\eta_r(q))$ is the load-bearing result of the paper.

What would settle it

Take two Gaussians with different means and variances on the real line, compute the natural gradient of $D(\mathcal{N}(\mu_1,\sigma_1)\|\mathcal{N}(\mu_2,\sigma_2))$ in the usual Euclidean parameterization, and compare it with the pair $(-s_q(r), -\eta_r(q))$ expressed in continuous densities; a mismatch, or an undefined $\eta_r(q)$ where the supports do not overlap, would show the total gradient formula does not transfer to continuous models as written.

Watch

Extended reading notes

Core claim

In the dually affine geometry of the statistical bundle, the KL divergence has a total natural gradient with two components, each expressed in the chart dual to its variable: for the first variable, minus the exponential score of $r$ seen from $q$; for the second, minus the mixture coordinate of $q$ seen from $r$. The proof follows from differentiating the cumulant function representation $D(p \| \exp_p(v)) = K_p(v)$ and the mixture representation, respectively. The paper then shows that this gradient, together with the bundle derivative of marginalization, produces the natural gradients of the mean-field divergence and the variational lower bound in a few lines of affine computation.

Load-bearing premise

The gradient identity is proven for fully supported finite probability functions on an open simplex, and the paper assumes without proof that the same affine calculus carries over to the parameterized, sampled, or continuous models used in the machine-learning applications.

Editorial extensions

If this is right

The gradient flow toward a fixed $r$ is an exponential-family path $q(t) = e^{e^{-t}v_0 - K_r(e^{-t}v_0)}\cdot r$, and the flow toward a fixed $q$ is a mixture path $r(t) = e^{-t}r_0 + (1-e^{-t})q$, both with exponentially decaying temperature.
The mean-field divergence gradients of Proposition 5 are expressed through conditional expectations, linking the calculus to ANOVA-type decompositions and interaction terms.
In variational Bayes, the lower-bound gradient flow becomes $\dot{\theta} = -\mathrm{Hess}\,\psi(\theta)\theta + \mathrm{Cov}_{e_{q_2}(\theta\cdot u)}(u, \log q_{1|2}(x|\cdot))$, giving a geometric derivation of the Bayesian learning rule.
The Jensen\u2013Shannon divergence gradient reduces to $-\tfrac12 s_q((q+r)/2)$, showing that the affine score alone carries the divergence's full gradient.
In the product setting, the same identity yields the generator and discriminator gradients for adversarial conditional models through the transpose of the bundle derivative.

Reading between the lines

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

If the affine calculus transfers to continuous sample spaces with an appropriate functional-analytic setup, it would give exact natural-gradient formulas for variational inference in exponential families, potentially bypassing black-box gradient estimators.
The mixture component $\eta_r(q) = q/r - 1$ requires $q$ absolutely continuous with respect to $r$; in machine-learning settings with support mismatch, this signals a singularity that sampling approximations must handle, suggesting a need for smoothed or entropy-regularized variants.
The derived variational Bayes flow suggests a concrete test: discretize the continuous-time flow and compare it against standard stochastic gradient variational Bayes on a small conjugate model, which would validate the sampling versions the paper leaves open.
The bundle calculus with joint marginalization likely extends to multi-marginal optimal transport and Schr\"odinger bridge iterations, where the same affinity should yield clean interaction-only gradients.

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

Short version: the core calculus is correct and cleanly presented, but the paper's own variational-Bayes application is not the natural-gradient flow it claims to illustrate. Fix that, plus a sign typo in (17), and it's a useful theory note.

What's actually new: not much scientifically. Proposition 1 repackages known KL derivatives in centered score and mixture coordinates. That's okay—the package is neat. The bundle language gives a coherent way to derive mean-field and conditional-probability gradients, and Proposition 5 is a useful explicit formula. The derivations are self-contained finite-dimensional calculus and check out against standard directional derivatives. No code or experiments, but the paper says upfront that implementation is out of scope; that's fine for a math.ST note.

Soft spots. The reader flagged a sign issue in (17); I agree it's a typo (the D(p||r) term should be plus, and 'sp(v)' should be 's_p(r)'), and it doesn't affect Proposition 1 because the term is constant in the relevant derivative. More important is the variational Bayes section. Equation (43) gives d/dt L = dotθᵀ(Cov_r(u, log q1|2) − Hessψ(θ)θ). The paper then writes the gradient flow as (44): dotθ = −Hessψ(θ)θ + Cov(...). That is the Euclidean gradient in the θ chart, not the natural gradient under the Fisher metric G = Hessψ. The natural-gradient flow would be dotθ = −θ + G⁻¹Cov. The exact-model sanity check makes this concrete: (44) becomes dotθ = −G(θ − θ̄), whereas the natural flow is dotθ = −(θ − θ̄). So the main advertised application does not demonstrate the affine natural-gradient calculus. If (44) is meant only as a coordinate gradient flow, the text should say so; if it's meant as natural gradient, it's missing a factor G⁻¹. Either way, Section 3.5 needs rewriting or re-labelling. This is substantive, not just notation.

The citation pattern is fine. The author leans on his own statistical-bundle papers, but the main identities are verified against elementary calculus, so it's not circular.

Bottom line: this is a careful expository paper for readers who want the statistical-bundle language for KL-based gradients. It deserves a serious referee, but only after the variational-Bayes section is straightened out. I'd send it to review, with the expectation of major-minor revision.

Referee Report

2 major / 4 minor

Summary. The paper develops an affine calculus on the open probability simplex using the statistical bundle framework, and applies it to compute natural gradients of the Kullback-Leibler divergence and related quantities. The main result, Proposition 1, gives the total natural gradient of the KL divergence as (q,r) ↦ (-s_q(r), -η_r(q)). This formula is then applied to product sample spaces to derive gradient expressions for mean-field approximation, adversarial generative models, and variational Bayes. The paper claims that this dually affine formalism yields neat and principled versions of known computations in statistical machine learning.

Significance. If the calculations are correct, the paper offers a compact and elegant differential-geometric calculus for natural-gradient computations on the probability simplex, with potential value for unifying diverse computations in machine learning. The finite-dimensional proofs are largely coherent, and the formulas reduce correctly to standard directional derivatives in the examples that are fully worked. The paper is self-contained in its finite-dimensional setting and provides explicit expressions that could serve as a useful reference. However, the variational Bayes application contains a conceptual error that affects the advertised claim, and the constrained-minimization step in the Schrödinger section is delegated to an external reference. These issues currently weaken the paper's central message.

major comments (2)

[3.5, Eq. (44)] The equation presented as 'the gradient flow equation' for the variational lower bound is not the natural-gradient flow under the paper's own metric. From Eq. (43), d/dt L(r(t);x) = dot-theta(t)^T (Cov_{r(t)}(u, log q_{1|2}(x|·)) - Hess-psi(theta(t)) theta(t)). If the statistical-bundle gradient on the model is written as g = gamma^T eU^r_{q2} u, then d/dt L = dot-theta^T G gamma, where G = Hess-psi(theta). Hence the natural-gradient ascent is dot-theta = -theta + G^{-1} Cov, not dot-theta = -G theta + Cov as in Eq. (44). The paper's own sanity check confirms the discrepancy: in the exact-model case Eq. (44) becomes dot-theta = -G(theta - bar-theta), whereas the natural-gradient flow would be dot-theta = -(theta - bar-theta). Thus Eq. (44) is either an ordinary coordinate gradient flow (not a demonstration of the affine calculus) or a natural-gradient flow with a missing factor G^{-1}. This undermines one of the three headline applications.
[3.3 (Kantorovich and Schrödinger)] The paper does not actually carry out the constrained minimization that gives the paper its title. After computing the gradient of S_epsilon(q), the key step of isolating the interaction component and projecting onto the constraint Gamma(q1,q2) is left to reference [24], with no formula or proof in the present manuscript. Since the paper's stated purpose is an affine calculus for constrained minima of the KL divergence, this omission leaves the Schrödinger application incomplete and the claimed 'calculus' without a demonstrated mechanism for enforcing constraints.

minor comments (4)

[Eq. (17)] The sign of the additive D(p||r) term appears incorrect: since log(r/p) = s_p(r) - D(p||r), the last term should be +D(p||r) rather than -D(p||r). This typo does not affect the proof of Proposition 1 because the term is constant in v, but it should be corrected.
[3.5, notation] The notation Cov_{eq2(theta(t)·u)} is not defined; it should be Cov_{r(t)} or Cov_{e^{theta·u-psi(theta)} q2}.
[3.2 and 2.3, typos] In Section 3.2, 'velocity velocityr' should be 'velocity star-r'; in Section 2.3, 'KH-divergence' should be 'KL-divergence'.
[3.4] The connection to generative adversarial networks is sketched but not developed into a training algorithm or a minimax formulation; the section computes only the gradient of the KL divergence through the generator map. A sentence clarifying the intended use would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the KL total natural gradient is derived from first principles within the paper, and self-citations are background rather than load-bearing.

full rationale

The central result, Proposition 1, is self-contained: the paper defines the statistical bundle, the affine charts, and the cumulant function, then differentiates D(q||r) directly in equations (17)-(18) to obtain the gradient (-s_q(r), -eta_r(q)). The proof uses only the definitions in (6)-(11), which are standard exponential-family facts that the paper itself shows, and it does not presuppose the claimed gradient. The mean-field and conditional-probability applications are chain-rule computations from Proposition 1 together with the derivative of marginalization computed in Proposition 3; these are new derivations rather than renamed inputs. No fitted parameter is later called a prediction, and no uniqueness theorem is invoked to force a choice. The author's prior work is cited for background details of the statistical bundle and for an ANOVA interaction computation, but the main gradient identities are proven in-line and do not reduce to those citations. The Discussion explicitly limits the paper to formal gradient-flow equations and says that sampling implementations are not considered, which is an honest limitation rather than a circular step. The variational-Bayes flow equation (44) is obtained by reading off a coordinate gradient from the time-derivative identity in (43); whether that equation is the Fisher-natural gradient (it appears to omit the inverse Fisher information) is a correctness concern, not a circularity, because the equation is not assumed as the conclusion of the derivation. Therefore no significant circularity is present.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No empirical constants are fitted and no new entities are postulated. The framework's building blocks (statistical bundle, transports) are mathematical constructions from prior work. The central computations rest on exponential-family cumulant identities and smoothness assumptions on the open simplex.

assumptions (4)

standard math The cumulant function K_p(v)=log E_p[e^v] satisfies K_p(v)=D(p||e_p(v)), dK_p(v)[h]=E_{e_p(v)}[h], and d^2K_p(v)[h,k]=Cov_{e_p(v)}(h,k), as stated in equations (8)-(11).
Used throughout Section 2 as the engine for all gradient computations; it is a standard exponential-family result from Brown 1986 and is not re-derived.
domain assumption The statistical bundle S E(Omega) with exponential transport eU, mixture transport mU, and covariance duality is a valid dually affine structure on the open simplex.
Inherited from the author's prior work [9,23,25]; the paper cites it and does not prove the affine-chart claims here. Self-citation, but the finite-dimensional claims are elementary.
domain assumption All probability functions are strictly positive on a finite state space and all curves and scalar fields are smooth.
Logarithms, Fisher scores, KL gradients, and exponential charts are defined only in the interior of the simplex, as stated in Section 1.1. Boundary and non-smooth cases are outside scope.
domain assumption The constraint sets used in Sections 3.2 to 3.5 (product measures, transport plans, conditional families, exponential tilts) are smooth submanifolds, so ambient gradients plus the computed derivatives of the constraint maps describe constrained dynamics.
The paper computes derivatives of maps such as marginalization and conditioning but does not prove that the constrained optimization dynamics stay on the submanifold or converge.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Affine calculus for constrained minima of the Kullback-Leibler divergence." pith.science (2026). https://pith.science/paper/L2IHKH2S

@misc{pith2026250202177,
  author       = {Pith},
  title        = {Pith review of: Affine calculus for constrained minima of the Kullback-Leibler divergence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2IHKH2S}},
  note         = {Machine review of arXiv:2502.02177}
}

read the original abstract

The non-parametric version of Amari's dually affine Information Geometry provides a practical calculus to perform computations of interest in statistical machine learning. The method uses the notion of a statistical bundle, a mathematical structure that includes both probability densities and random variables to capture the spirit of Fisherian statistics. We focus on computations involving a constrained minimization of the Kullback-Leibler divergence. We show how to obtain neat and principled versions of known computation in applications such as mean-field approximation, adversarial generative models, and variational Bayes.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

[24]

752–759, Springer International Publishing, 2021

, Statistical bundle of the transport model , p. 752–759, Springer International Publishing, 2021

work page 2021
[1]

Organizer Preben F

Shun-ichi Amari, Geometry of semiparametric models and applications , Invited Papers Meeting IP64 “Likelihood and Geometry. Organizer Preben F. Blaesild. The 51st Se ssion of the International Statistical Institute Istanbul, August 18-26, 1997, 1997

work page 1997
[2]

2, 251–276

Shun-Ichi Amari, Natural gradient works eﬃciently in learning , Neural Computation 10 (1998), no. 2, 251–276

work page 1998
[3]

194, Springer, [Tokyo], 2016

Shun-ichi Amari, Information geometry and its applications , Applied Mathematical Sciences, vol. 194, Springer, [Tokyo], 2016. MR 3495836

work page 2016
[4]

Shun-ichi Amari, Ryo Karakida, and Masafumi Oizumi, Information geometry connecting Wasser- stein distance and Kullback-Leibler divergence via the ent ropy-relaxed transportation problem , Inf. Geom. 1 (2018), no. 1, 13–37. MR 3974671

work page 2018
[5]

MR 1 800 071

Shun-ichi Amari and Hiroshi Nagaoka, Methods of information geometry , American Mathematical Society, 2000, Translated from the 1993 Japanese original by Dais hi Harada. MR 1 800 071

work page 2000
[6]

Nihat Ay, Information geometry of the otto metric , Information Geometry (2024)

work page 2024
[7]

Brown, Fundamentals of statistical exponential families with app lications in statistical decision theory, IMS Lecture Notes

Lawrence D. Brown, Fundamentals of statistical exponential families with app lications in statistical decision theory, IMS Lecture Notes. Monograph Series, no. 9, Institute of Math ematical Statistics,

work page

Show all 27 references

[8]

13, 2250214.1–46

Goﬀredo Chirco, Luigi Malag` o, and Giovanni Pistone, Lagrangian and Hamiltonian dynamics for probabilities on the statistical bundle , International Journal of Geometric Methods in Modern Physics 19 (2022), no. 13, 2250214.1–46

2022
[9]

Goﬀredo Chirco and Giovanni Pistone, Dually aﬃne Information Geometry modeled on a Banach space, 2022, arXiv:2204.00917

2022 arXiv
[10]

Cover and Joy A

Thomas M. Cover and Joy A. Thomas, Elements of information theory , second ed., Wiley- Interscience [John Wiley & Sons], 2006. MR MR2239987 (2007h:00002 )

2006
[11]

MR 1138207 (92i:53001)

Manfredo Perdig˜ ao do Carmo, Riemannian geometry , Mathematics: Theory & Applications, Birkh¨ auser Boston Inc., 1992, Translated from the second Port uguese edition by Francis Flaherty. MR 1138207 (92i:53001)

1992
[12]

5, Cambridge University Press, New York, 2016, Algorithms, evidence, and data science

Bradley Efron and Trevor Hastie, Computer age statistical inference , Institute of Mathematical Statistics (IMS) Monographs, vol. 5, Cambridge University Press, New York, 2016, Algorithms, evidence, and data science. MR 3523956

2016
[13]

Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, D avid Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, Generative adversarial networks , 2014

Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, D avid Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, Generative adversarial networks , 2014

2014
[14]

Mohammad Emtiyaz Khan and H ˚ a vard Rue, The Bayesian learning rule , J. Mach. Learn. Res. 24 (2023), Paper No. [281], 46. MR 4664718

2023
[15]

Diederik P Kingma and Max Welling, Auto-encoding variational Bayes , 2022

2022
[16]

Landau and Eugenij M

Lev D. Landau and Eugenij M. Lifshits, Course of theoretical physics. statistical physics. , 3rd ed., vol. V, Butterworth-Heinemann, 1980

1980
[17]

160, Springer-Verlag, 1995

Serge Lang, Diﬀerential and Riemannian manifolds , third ed., Graduate Texts in Mathematics, vol. 160, Springer-Verlag, 1995. MR 96d:53001

1995
[18]

Lin, Divergence measures based on the Shannon entropy, IEEE Transactions on Information Theory 37 (1991), no

J. Lin, Divergence measures based on the Shannon entropy, IEEE Transactions on Information Theory 37 (1991), no. 1, 145–151

1991
[19]

2, 137–179

Luigi Malag` o, Luigi Montrucchio, and Giovanni Pistone, Wasserstein riemannian geometry of gauss- ian densities , Information Geometry 1 (2018), no. 2, 137–179

2018
[20]

1034, Springer-Verlag, 1983

Julian Musielak, Orlicz spaces and modular spaces , Lecture Notes in Mathematics, vol. 1034, Springer-Verlag, 1983. MR 724434 (85m:46028)

1983
[21]

5–6, 355–607, arXiv:1803.00567v2

Gabriel Peyr´ e and Marco Cuturi, Computational optimal transport , Foundations and Trends in Ma- chine Learning 11 (2019), no. 5–6, 355–607, arXiv:1803.00567v2

2019 arXiv
[22]

Giovanni Pistone, Information geometry of the Gaussian space , Information geometry and its appli- cations, Springer Proc. Math. Stat., vol. 252, Springer, Cham, 20 18, pp. 119–155. MR 3876116
[23]

2, 221–242

, Information geometry of the probability simplex: A short co urse, Nonlinear Phenomena in Complex Systems 23 (2020), no. 2, 221–242

2020
[25]

Giovanni Pistone and Carlo Sempi, An inﬁnite-dimensional geometric structure on the space of all the probability measures equivalent to a given one , Ann. Statist. 23 (1995), no. 5, 1543–1561. MR 97j:62006 18

1995
[26]

Hermann Weyl, Space-time-matter, Dover, New York, 1952 (eng), translation of the 1921 RAUM ZEIT MATERIE. De Castro Statistics, Collegio Carlo Alberto Email address : giovanni.pistone@carloalberto.org Nuovo SEFIR, c/o Coworld, Centro Direzionale Milano Due, Pa lazzo Canova, I-2...

1952
[1986]

MR MR882001 (88h:62018)

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [24]

752–759, Springer International Publishing, 2021

, Statistical bundle of the transport model , p. 752–759, Springer International Publishing, 2021

work page 2021

[2] [1]

Organizer Preben F

Shun-ichi Amari, Geometry of semiparametric models and applications , Invited Papers Meeting IP64 “Likelihood and Geometry. Organizer Preben F. Blaesild. The 51st Se ssion of the International Statistical Institute Istanbul, August 18-26, 1997, 1997

work page 1997

[3] [2]

2, 251–276

Shun-Ichi Amari, Natural gradient works eﬃciently in learning , Neural Computation 10 (1998), no. 2, 251–276

work page 1998

[4] [3]

194, Springer, [Tokyo], 2016

Shun-ichi Amari, Information geometry and its applications , Applied Mathematical Sciences, vol. 194, Springer, [Tokyo], 2016. MR 3495836

work page 2016

[5] [4]

Shun-ichi Amari, Ryo Karakida, and Masafumi Oizumi, Information geometry connecting Wasser- stein distance and Kullback-Leibler divergence via the ent ropy-relaxed transportation problem , Inf. Geom. 1 (2018), no. 1, 13–37. MR 3974671

work page 2018

[6] [5]

MR 1 800 071

Shun-ichi Amari and Hiroshi Nagaoka, Methods of information geometry , American Mathematical Society, 2000, Translated from the 1993 Japanese original by Dais hi Harada. MR 1 800 071

work page 2000

[7] [6]

Nihat Ay, Information geometry of the otto metric , Information Geometry (2024)

work page 2024

[8] [7]

Brown, Fundamentals of statistical exponential families with app lications in statistical decision theory, IMS Lecture Notes

Lawrence D. Brown, Fundamentals of statistical exponential families with app lications in statistical decision theory, IMS Lecture Notes. Monograph Series, no. 9, Institute of Math ematical Statistics,

work page