Hessian Riemannian Structures

• Hessian Riemannian structures are defined by a flat affine connection and a Riemannian metric derived from the Hessians of local potential functions.
• They play a central role in differential, information, and convex geometry by linking dual connections, curvature, and global topological invariants.
• Their framework underpins practical applications in optimization, gradient flows, and statistical modeling through Legendre transforms and duality principles.

A Hessian Riemannian structure is a Riemannian geometric structure modeled on local potentials whose second derivatives (Hessians) define the metric, together with an underlying flat affine connection. These structures play a central role in differential geometry, information geometry, convex analysis, integrable systems, and the study of geometric flows. Their significance extends from global topological constraints on manifolds to local curvature and bundle-theoretic invariants, and they provide the geometric framework underlying many optimization algorithms, duality principles, and structures in statistical modeling.

1. Definition and Local Characterization

A Hessian Riemannian structure on a smooth manifold $M$ is defined by a pair $(D,g)$, where $D$ is a torsion-free, flat affine connection and $g$ is a Riemannian metric such that in every $D$-affine chart, $g$ arises as the Hessian of a smooth potential: $$g_{ij}(x) = \frac{\partial^2 \phi}{\partial x^i \partial x^j},\qquad x \in U\subset M,$$ for some local potential $\phi: U \to \mathbb{R}$ (Vollmer, 1 Dec 2025, Liu, 1 Sep 2025, Boyom et al., 2015).

The requirement that $Dg$ is totally symmetric (i.e., $(D_X g)(Y,Z)$ symmetric in all three arguments) is equivalent to the existence of such a potential in $(D,g)$0-affine coordinates. This property underpins the close relation between Hessian geometry and statistical manifolds: a statistical manifold is a triple $(D,g)$1 with $(D,g)$2 totally symmetric, and if $(D,g)$3 is flat, the structure is Hessian (Osipov, 2022).

Dual connections arise intrinsically: for a Hessian metric $(D,g)$4, the connection $(D,g)$5 defined by

$(D,g)$6

is also flat and torsion-free, and the Levi-Civita connection $(D,g)$7 sits halfway between $(D,g)$8 and $(D,g)$9 (Liu, 1 Sep 2025).

2. Topological and Bundle-Theoretic Constraints

Hessian manifolds exhibit stringent global topological restrictions:

• The Euler characteristic of any compact Hessian manifold vanishes $D$0 (Liu, 1 Sep 2025).
• The fundamental group $D$1 must be infinite and torsion-free; for compact $D$2, it is never finite (Liu, 1 Sep 2025).
• Any compact Hessian manifold $D$3 with abelian $D$4 is homeomorphic to the flat torus $D$5 (Liu, 1 Sep 2025).

Flat line bundles classify significant invariants:

Bundle	Description
Canonical	$D$6; trivial iff $D$7 admits a $D$8-parallel $D$9-form
Obstruction	Transition functions capture nontriviality of global Hessian potential
Positive flat line	Existence is equivalent to Koszul-type Hessian metrics

The obstruction bundle encodes the failure to globally realize the Hessian metric from a single potential: it is trivial if and only if such a global potential exists (Liu, 1 Sep 2025).

3. Curvature, Duality, and Integrability

Although $g$0 is flat, the Levi-Civita curvature need not vanish. The difference tensor $g$1 encodes information about the metric’s deviation from flatness and is totally symmetric. Important curvature formulas include (Vollmer, 1 Dec 2025): $g$2 for Hessian manifolds, where $g$3 is the Amari–Chentsov cubic tensor (Liu, 1 Sep 2025).

Duality theory relates Koszul-type and radiant manifolds: a Hessian metric of Koszul type ($g$4 with $g$5 closed) is dual to a radiant structure under $g$6, and vice versa. The Legendre transform relates potential functions and their duals, with geometric meaning in terms of Lagrangian immersions and symplectic geometry (Liu, 1 Sep 2025).

In the context of manifolds with additional structure, such as Frobenius manifolds, Hessian geometry provides the underlying metric and connection. On spaces with nonzero curvature, "curved Frobenius" structures compatible with a (generalized) Hessian structure arise precisely when a specific finite-type prolongation system holds; the Hessian and Frobenius potentials can then be identified (Vollmer, 1 Dec 2025).

4. Hessian Structures in Optimization, Flows, and Information Geometry

Hessian Riemannian structures are foundational for Riemannian gradient flows and manifold optimization:

• Given a convex, Legendre-type function $g$7, the Riemannian metric $g$8 defines dynamics underlying interior-point and proximal algorithms. The associated Bregman divergence acts as both a barrier and Lyapunov function, and the metric exhibits unique integration properties connected to variational inequalities (Alvarez et al., 2018).
• The natural flows

$g$9

correspond to geodesics in Legendre–dual coordinates for linear objectives, and their global analytic properties (existence, convergence) are dictated by the geometry of $D$0.

• Similar structures underpin the analysis of multipopulation Wardrop equilibria, where Hessian-metric-projected flows on constraint manifolds provide globally convergent methods for variational inequalities in traffic and network equilibrium (Bakaryan et al., 22 Apr 2025).

In information geometry, the Fisher information metric is Hessian when the underlying family is exponential: $D$1 for the log-partition $D$2 in affine parameters. When the metric becomes degenerate (rank-deficient), a canonical foliation arises, and the transverse geometry is transversely Hessian (Boyom et al., 2015).

5. Locally Conformally Hessian Structures and Extensions

Generalizations of Hessian geometry—locally conformally Hessian (l.c.H.) structures—arise as quotients of Hessian manifolds by groups acting by Hessian homotheties. An l.c.H. manifold admits a closed Lee form $D$3 such that $D$4 is totally symmetric, and in local charts $D$5 is Hessian. Radiant l.c.H. structures, where the Lee vector field $D$6 is Killing and satisfies $D$7, admit a fibration over the circle with fibers statistical manifolds of constant curvature (Osipov, 2022).

Key structural results include the density of rank 1 radiant l.c.H. metrics and a precise description of the fibration for compact manifolds. Applications of l.c.H. structures span statistical geometry, thermodynamic geometry, and convex optimization, where local Hessian forms up to conformal factors naturally appear.

6. Examples, Enriched Structures, and Interactions with Physics

Explicit families of Hessian manifolds include:

• Flat models: Euclidean $D$8 with $D$9, toroidal quotients, and convex domains with diagonal or separable potentials.
• Negative curvature: Manifolds modeled on convex cones $g$0 with invariant Hessian metrics of constant negative scalar curvature, constructed via group and quotient methods (Liu, 1 Sep 2025).
• Golden and metallic structures: Projection operators derived from rank-one Hessian tensors built from functions such as $g$1 allow construction of polynomial geometric structures (golden, metallic ratios) with explicit integration, parallelism, and scalar curvature properties (Washburn et al., 1 Jun 2026).

Superintegrable systems produce Hessian metrics whose natural coordinates arise from the integrability conditions corresponding to Killing tensors. This construction leads to dual flat connections and coordinates adapted to integrability, unifying information-geometric and classical geometric analysis (Armstrong et al., 2024, Vollmer, 1 Dec 2025).

Born structures on the tangent bundle of a Hessian manifold (in the sense of Dombrowski) reveal an equivalence: $g$2 is Hessian if and only if the induced almost Born structure on $g$3 is integrable. This connection places Hessian geometry in the context of para-quaternionic and para-Hermitian geometry and generalizations relevant in mathematical physics (Sakamoto, 31 Jul 2025).

7. Extensions: Foliations, Conformal and Product Structures

In general situations, the Hessian property extends to foliated settings. If a metric tensor is only positive semi-definite but parallelizable by a flat, torsion-free connection, the kernel directions integrate to a foliation, and the transverse structure is Hessian. This construction is crucial in information geometry for singular (non-full-rank) models, mixture families, and spaces with affine symmetry groups (Boyom et al., 2015).

Further, the Hessian context supports a rich interplay with polynomial and product structures, such as golden and metallic product operators and integrable distributions, expanding the catalogue of explicit geometric models with controlled curvature and symmetry (Washburn et al., 1 Jun 2026).

---

Hessian Riemannian structures, through their interplay of differential-geometric, topological, bundle-theoretic, and analytic properties, form a unifying geometric framework for convex optimization, superintegrable systems, information geometry, and a broad class of geometric flows and duality theories. Their rich structure supports extensions to conformal, product, and foliated geometries, with distinctive global obstructions and local analytic behavior dictated by the underlying affine and metric data.