Finsler Manifolds

• Finsler manifolds are smooth spaces endowed with a positively homogeneous, strongly convex fundamental function that generalizes Riemannian metrics.
• They incorporate nonlinear connections and anisotropic curvature invariants, leading to a non-symmetric distance function and novel geometric and analytic phenomena.
• Modern applications extend Finsler geometry into data science and manifold learning by exploiting direction-dependent metrics and curvature flows for robust analysis.

A Finsler manifold is a smooth manifold equipped with a fundamental function—a Finsler metric—on its tangent bundle that is positively homogeneous of degree one and exhibits strong convexity in the fiber variables. This structure generalizes Riemannian geometry by allowing the metric to depend smoothly on both position and direction, providing a framework for analyzing geometric, analytic, topological, and algebraic phenomena arising from the interplay of smoothness, convexity, and non-quadratic norms. The study of Finsler manifolds encompasses advanced connection theory, curvature invariants, global analysis, algebraic characterizations, and modern data science applications, with deep interconnections to classical differential geometry and extensions far beyond Riemannian paradigms.

1. Fundamental Structure and Metric Properties

A Finsler manifold is a pair $(M, F)$, where $M$ is a smooth $n$-dimensional manifold and $F: TM \to [0, \infty)$ is a $C^\infty$ function on the slit tangent bundle $TM\setminus \{0\}$, positively 1-homogeneous in the fiber variables and such that the fundamental tensor

$$g_{ij}(x,y) = \frac{1}{2} \frac{\partial^2 [F^2]}{\partial y^i \partial y^j}$$

is positive-definite for every $(x, y)\in TM\setminus\{0\}$ (Ootsuka et al., 2018, Shiohama et al., 2018, Vincze et al., 2022). This ensures that at each $x$, $F(x,\cdot)$ is a Minkowski norm on $M$0 whose indicatrix $M$1 is a smooth, strictly convex hypersurface (Vincze et al., 2022). In contrast to Riemannian metrics, the dependence of $M$2 (and hence local geometry) on the direction $M$3 renders Finsler geometry inherently anisotropic.

Distance between points is defined as

$M$4

which is generally asymmetric: $M$5 for generic $M$6 (Shiohama et al., 2018, Dagès et al., 12 Mar 2026). This loss of symmetry unlocks new geometric and analytic phenomena unavailable to Riemannian geometry.

2. Connection Theory and Curvature Invariants

Finsler geometry necessitates refined notions of connection, parallel transport, and curvature, given the direction-dependence of its basic structures.

Nonlinear and metric-preserving connections: Classical linear connections (e.g., Levi-Civita) have no direct analogue due to non-quadraticity. Instead, nonlinear connections, such as the Cartan (or Chern–Rund) connection, are defined on the double tangent bundle $M$7, splitting $M$8 into horizontal and vertical subbundles (Ootsuka et al., 2018, Attarchi et al., 2011). Admissible connections must satisfy metric-compatibility: $M$9 and additional symmetry constraints (Cartan conditions) on their coefficients. In moving frame (vielbein) formalism, the nonlinear connection coefficients $n$0, the spray coefficients $n$1, and the curvature two-forms $n$2 are algebraically expressible in terms of the frame commutators, the Hessian $n$3, and the Cartan tensor (Ootsuka et al., 2018).

Curvature: The Cartan connection yields the $n$4-curvature ($n$5), Berwald and Landsberg curvatures, and the central scalar invariant—flag curvature: $n$6 where $n$7 and $n$8 is a 2-plane containing $n$9 (Shiohama et al., 2018, Cheng, 2019). This generalizes sectional curvature and governs comparison theory, rigidity, and global topology in the Finsler setting.

Cartan tensor and deviation from Riemannian structure: The (0,3)-tensor

$F: TM \to [0, \infty)$0

measures the failure of $F: TM \to [0, \infty)$1 to be Riemannian: $F: TM \to [0, \infty)$2 if and only if $F: TM \to [0, \infty)$3 is quadratic in $F: TM \to [0, \infty)$4 (Shiohama et al., 2018, Shojaee et al., 2014, Cheng, 2019).

3. Indicatrix, Adapted Frames, and Sasakian Geometry

The unit sphere bundle (indicatrix bundle) $F: TM \to [0, \infty)$5 is a central construction, providing a natural stage for the study of intrinsic and extrinsic submanifold geometry (Stojanov, 2010, Attarchi et al., 2011). Each indicatrix $F: TM \to [0, \infty)$6 is a strictly convex, totally umbilic hypersurface in $F: TM \to [0, \infty)$7 with constant mean curvature $F: TM \to [0, \infty)$8 with respect to the induced Riemannian metric $F: TM \to [0, \infty)$9 for fixed $C^\infty$0 (Stojanov, 2010).

Natural structures on $C^\infty$1 include the adapted frame $C^\infty$2 corresponding to the horizontal and vertical decompositions induced by the nonlinear connection (Attarchi et al., 2011). The Sasaki metric, built from $C^\infty$3, and the almost complex structure $C^\infty$4 make $C^\infty$5 almost Hermitian.

However, on the indicatrix bundle, the induced contact metric structure fails to be Sasakian except in the Riemannian case, due to the non-vanishing Cartan tensor (Attarchi et al., 2011). The geometric obstruction stems from the direction-dependence of the metric, reflected in the “varying shape” of indicatrices.

4. Global Topology, Cut Loci, and Rigidity Phenomena

The global analysis of Finsler manifolds diverges from the Riemannian case along several axes (Shiohama et al., 2018, Tanaka et al., 2012):

• Distance asymmetry and reversibility: The Finsler distance is typically non-symmetric, leading to the necessity of separate treatment for forward and backward geodesics, balls, and cut loci. The reversibility constant

$C^\infty$6

quantifies this asymmetry.

• Cut locus: In two dimensions, the cut locus of a closed subset is a local tree, comprised of countably many rectifiable Jordan arcs; this structure breaks down in higher dimensions (Tanaka et al., 2012). Points where the distance to $C^\infty$7 is non-differentiable correspond to locations joined to $C^\infty$8 by at least two geodesic segments.
• Axiom of spheres and constant flag curvature: The presence of sufficiently many extrinsic spheres (totally umbilic submanifolds with parallel mean curvature in every tangent subspace) rigidly forces the manifold to have constant flag curvature (Sedaghat et al., 2019). This is a direct generalization of the Leung–Nomizu theorem, revealing the interplay of extrinsic and intrinsic geometry in Finsler settings.

5. Connections to Algebraic and Functional Structures

In both finite and infinite-dimensional settings, Finsler geometry is deeply intertwined with the algebraic structure of smooth function spaces (Jaramillo et al., 2011). The normed algebra of bounded, smooth functions with bounded derivatives, $C^\infty$9, completely encodes the weak Finsler geometry for large classes of Finsler manifolds, especially those modeled on Banach spaces. In particular:

• Complete finite-dimensional Finsler manifolds are uniquely determined (up to isometry) by the algebraic-normed structure of $TM\setminus \{0\}$0.
• For Banach–Finsler manifolds with additional bump function and convexity properties, the Banach–Stone and Myers–Nakai theorems extend, establishing a duality between geometric and function-algebraic data.

The Finsler norm itself is reconstructible from $TM\setminus \{0\}$1 via

$TM\setminus \{0\}$2

with the associated distance given by the Kantorovich–Rubinstein duality.

6. Applications and Contemporary Directions

Manifold learning and data science: Recent developments exploit the asymmetry intrinsic to Finsler structures in data analysis, particularly for embedding high-dimensional data in low-dimensional asymmetric spaces (Dagès et al., 12 Mar 2026). Traditional manifold learning methods (t-SNE, UMAP) rely on symmetric geometry; Finslerian generalizations (Finsler t-SNE, Finsler UMAP) retain and reveal density hierarchy and directionality present in real datasets through the use of Randers metrics as embedding spaces.

Gradient flows and curvature functionals: Flows driven by curvature-type functionals, extending Ricci flow to the Finsler context, have been constructed and studied (Shojaee et al., 2014). These flows, such as the conformal Finsler–Ricci flow, select canonical Finsler metrics (e.g., constant flag curvature models) within a conformal class and provide analytical tools paralleling the Ricci flow’s role in Riemannian geometry.

Compatible and semi-symmetric connections: Finsler manifolds admitting compatible linear connections (generalized Berwald manifolds) are of particular interest (Vincze et al., 2022). The rigidity and classification of such connections, especially semi-symmetric ones, admit algebraic treatment and often reduce to intrinsic conditions on the Finsler function.

Symmetries and hidden conserved quantities: The generalization of Killing vector fields to Finsler manifolds via the spray operator formalism unifies symmetry analysis and the algorithmic extraction of higher-order conservation laws, such as those underlying integrable systems in classical and quantum settings (Ootsuka et al., 2016).

7. Significance and Distinctive Features

Key features distinguishing Finsler from Riemannian geometry include:

• Direction-dependent metric tensor and curvature invariants, and the central role of the Cartan tensor;
• General lack of symmetry in distance, connection, and curvature, leading to refined notions of convexity, geodesics, and cut/conjugate loci;
• Broadened scope for natural geometric flows, rigidity theorems, and the study of complex topological structures;
• Algebraic characterizations of geometry via function spaces and Banach-module techniques.

Collectively, the theory of Finsler manifolds extends the reach of classical differential geometry, providing a unifying and flexible framework for problems in geometric analysis, topology, mathematical physics, and modern data science (Ootsuka et al., 2018, Shiohama et al., 2018, Attarchi et al., 2011, Stojanov, 2010, Cheng, 2019, Sedaghat et al., 2019, Shojaee et al., 2014, Tanaka et al., 2012, Dagès et al., 12 Mar 2026, Vincze et al., 2022, Jaramillo et al., 2011, Ootsuka et al., 2016).