aiwiki.page
English
Mathematics / Geometry & Topology / riemannian-manifold

Riemannian Manifold

A Riemannian manifold is a smooth manifold equipped with smoothly varying inner products that define lengths, angles, distances, curvature, and geometric differential operators.

46 keywords1 linked from13 not yet writtenWritten by AI
ManifoldInner productDifferential Geo…Tangent SpaceMetric TensorEinstein Summati…Positive-definit…IsometryRiemannian…

A Riemannian manifold is a smooth manifold equipped with a positive-definite inner product on each of its tangent spaces, varying smoothly from point to point. This additional structure, called a Riemannian metric, makes it possible to measure the lengths of curves and the angles between tangent directions. It also determines an intrinsic distance, a notion of straightest possible motion, and curvature. Unlike a surface described only as a subset of three-dimensional space, a Riemannian manifold can be studied without reference to any surrounding space and can have any finite dimension.

Riemannian manifolds are the basic objects of Riemannian geometry, a central branch of differential geometry. They connect local measurements with global shape: curvature can constrain a space’s topology, while completeness determines whether geodesics can continue indefinitely. Their metric structure also supports integration, differential equations, and geometric optimization. The subject developed from the intrinsic geometry of surfaces and Bernhard Riemann’s nineteenth-century generalization of geometry, and now provides a common framework for substantial parts of geometry, analysis, mathematical physics, and statistics.

Smooth structure and the Riemannian metric

An nn-dimensional Riemannian manifold is a pair (M,g)(M,g), where MM is a smooth nn-dimensional manifold and gg assigns to each point p∈Mp\in M an inner product

gp:TpM×TpM⟶Rg_p:T_pM\times T_pM\longrightarrow\mathbb R

on the tangent space TpMT_pM. The usual definition assumes that MM is Hausdorff and second countable; unless stated otherwise, it has no boundary. Smoothness of gg means that its coefficients in every smooth coordinate chart are smooth functions. Its positive definiteness means gp(v,v)>0g_p(v,v)>0 for every nonzero tangent vector vv.

In local coordinates x1,…,xnx^1,\ldots,x^n, the metric tensor has the form

g=gij dxi⊗dxj,gij=g ⁣(∂∂xi,∂∂xj).g=g_{ij}\,dx^i\otimes dx^j, \qquad g_{ij}=g\!\left(\frac{\partial}{\partial x^i}, \frac{\partial}{\partial x^j}\right).

Repeated upper and lower indices are summed, following the Einstein summation convention. At each point the coefficient array (gij)(g_{ij}) is a symmetric positive-definite matrix. For tangent vectors v=vi∂iv=v^i\partial_i and w=wj∂jw=w^j\partial_j, their inner product is gijviwjg_{ij}v^iw^j. The norm is ∥v∥g=g(v,v)\lVert v\rVert_g=\sqrt{g(v,v)}, and the angle between nonzero vectors is defined by the usual cosine formula.

Metric coefficients depend on coordinates, but the geometric measurement does not. Under a change from coordinates xx to coordinates yy,

g~ab=∂xi∂ya∂xj∂ybgij.\widetilde g_{ab} = \frac{\partial x^i}{\partial y^a} \frac{\partial x^j}{\partial y^b}g_{ij}.

This transformation law distinguishes a metric tensor from an arbitrary collection of functions. The expression ds2=gij dxidxjds^2=g_{ij}\,dx^i dx^j is traditional shorthand for the resulting infinitesimal squared-length rule, rather than an ordinary finite distance formula.

Every smooth manifold satisfying the standard countability assumptions admits a Riemannian metric. One constructs local Euclidean metrics in coordinate charts and combines them using a smooth partition of unity. Positive definiteness survives this weighted combination. A manifold therefore does not need to possess special topology in order to admit some Riemannian metric, although requirements on its curvature can impose strong restrictions.

The metric is additional structure, not something determined by the smooth manifold. Two metrics on the same manifold may have very different distances, volumes, and curvature. A diffeomorphism F:(M,g)→(N,h)F:(M,g)\to(N,h) is an isometry when F∗h=gF^*h=g; this preserves the full Riemannian geometry. By contrast, a homeomorphism preserves only topology, while a diffeomorphism preserves smooth structure without necessarily preserving lengths.

Model spaces and geometric constructions

The simplest example is Euclidean space Rn\mathbb R^n, with

g=(dx1)2+⋯+(dxn)2.g=(dx^1)^2+\cdots+(dx^n)^2.

Its metric coefficients are constant in Cartesian coordinates, its geodesics are straight lines, and its curvature vanishes. Nonconstant coefficients do not necessarily indicate curvature: the Euclidean plane in polar coordinates has metric dr2+r2dθ2dr^2+r^2d\theta^2. The apparent complication comes from coordinates, and the polar chart itself breaks down at the origin.

An embedded smooth submanifold of Euclidean space inherits a metric by restricting the ambient inner product to each tangent space. For a local parametrization F:U⊂Rn→RNF:U\subset\mathbb R^n\to\mathbb R^N, the induced metric is

gij=⟨∂F∂xi,∂F∂xj⟩.g_{ij}=\left\langle\frac{\partial F}{\partial x^i}, \frac{\partial F}{\partial x^j}\right\rangle.

The round sphere of radius rr is a principal example. In dimension two its metric in spherical coordinates is

g=r2(dθ2+sin⁡2θ dϕ2),g=r^2\bigl(d\theta^2+\sin^2\theta\,d\phi^2\bigr),

and its Gaussian curvature is 1/r21/r^2. The coordinate singularities at the poles are not singularities of the sphere’s metric.

Hyperbolic space supplies the complete, simply connected model of constant negative curvature. In the upper half-space description,

Hn={(x1,…,xn−1,y):y>0},g=(dx1)2+⋯+(dxn−1)2+dy2y2.\mathbb H^n=\{(x^1,\ldots,x^{n-1},y):y>0\}, \qquad g=\frac{(dx^1)^2+\cdots+(dx^{n-1})^2+dy^2}{y^2}.

This normalization gives sectional curvature −1-1. Although the boundary y=0y=0 lies at finite Euclidean coordinate distance, it is infinitely far away in the hyperbolic metric. Together with Euclidean space and round spheres, hyperbolic space provides the basic constant-curvature models underlying non-Euclidean geometry.

New examples arise through products and quotients. For manifolds (M,g)(M,g) and (N,h)(N,h), the product metric on M×NM\times N measures tangent vectors by

(g⊕h)((v,w),(v,w))=g(v,v)+h(w,w).(g\oplus h)((v,w),(v,w))=g(v,v)+h(w,w).

Quotienting Euclidean space by a lattice of translations produces a flat torus. Its topology differs from Euclidean space, even though every point has a neighborhood with Euclidean geometry. A freely and properly discontinuously acting group of isometries similarly produces a smooth Riemannian quotient.

Another important construction is a conformal change,

g~=e2fg,\widetilde g=e^{2f}g,

where ff is smooth. This preserves angles but generally changes lengths, geodesics, and curvature. Submanifold metrics, product metrics, quotient metrics, and conformal changes show how the same underlying geometric principles produce very different spaces. They also distinguish intrinsic geometry from embedding: a flat torus is not the same metric surface as the familiar curved doughnut-shaped torus embedded in R3\mathbb R^3.

Length, distance, geodesics, and completeness

For a piecewise smooth curve γ:[a,b]→M\gamma:[a,b]\to M, its Riemannian length is

Lg(γ)=∫abgγ(t)(γ˙(t),γ˙(t)) dt.L_g(\gamma)=\int_a^b \sqrt{g_{\gamma(t)}(\dot\gamma(t),\dot\gamma(t))}\,dt.

On a connected manifold, the distance dg(p,q)d_g(p,q) is the infimum of these lengths over curves joining pp to qq. Connected smooth manifolds are connected by piecewise smooth paths, so this infimum is finite. It defines a metric space structure whose topology is exactly the original manifold topology. On a disconnected manifold, this construction is normally applied separately to each connected component.

A geodesic is a curve with zero covariant acceleration. Equivalently, with a suitable affine parameter, it is a stationary point of the energy

E(γ)=12∫abg(γ˙,γ˙) dtE(\gamma)=\frac12\int_a^b g(\dot\gamma,\dot\gamma)\,dt

under variations with fixed endpoints. The calculus of variations leads to the coordinate equation

x¨k+Γijkx˙ix˙j=0.\ddot x^k+\Gamma^k_{ij}\dot x^i\dot x^j=0.

Geodesics have constant speed and locally minimize length, but need not minimize globally. A great-circle arc on a round sphere ceases to be minimizing after passing the antipodal point, although its continuation remains a geodesic.

For v∈TpMv\in T_pM, let γv\gamma_v be the geodesic with initial position pp and initial velocity vv. The exponential map is

exp⁡p(v)=γv(1),\exp_p(v)=\gamma_v(1),

where this geodesic exists up to time 11. Near the zero vector it is a diffeomorphism and defines normal coordinates. Radial geodesics in a sufficiently small normal neighborhood minimize distance from the center. The cut locus records where this minimizing behavior first fails. On a round sphere, the cut locus of a point is its antipode, where many minimizing geodesics meet.

Metric completeness means completeness of (M,dg)(M,d_g): every Cauchy sequence converges. Geodesic completeness means every maximal geodesic is defined for all real parameter values. The Hopf–Rinow theorem equates these properties for a connected finite-dimensional Riemannian manifold. It also implies that any two points of such a complete manifold are joined by a minimizing geodesic, and that closed bounded subsets are compact.

Completeness is stronger than the mere availability of local geodesics. The punctured Euclidean plane is incomplete because a straight line can reach the omitted origin in finite time. Every compact Riemannian manifold without boundary is complete. Completeness need not imply compactness: both Euclidean and hyperbolic spaces are complete and noncompact. On a manifold with boundary, the unqualified equivalences require modification because geodesics may reach the boundary.

Covariant differentiation and parallel transport

Tangent vectors at different points belong to different vector spaces, so ordinary subtraction does not intrinsically compare them. A connection supplies a rule for differentiation of vector fields. For Riemannian geometry, the distinguished rule is the Levi-Civita connection, characterized uniquely by metric compatibility and zero torsion. Metric compatibility requires

X(g(Y,Z))=g(∇XY,Z)+g(Y,∇XZ),X\bigl(g(Y,Z)\bigr) = g(\nabla_XY,Z)+g(Y,\nabla_XZ),

while zero torsion means

∇XY−∇YX=[X,Y].\nabla_XY-\nabla_YX=[X,Y].

Here [X,Y][X,Y] is the Lie bracket of vector fields.

In coordinates, the connection is represented by Christoffel symbols:

∇∂i∂j=Γijk∂k,Γijk=12gkℓ(∂igjℓ+∂jgiℓ−∂ℓgij),\nabla_{\partial_i}\partial_j = \Gamma^k_{ij}\partial_k, \qquad \Gamma^k_{ij} = \frac12g^{k\ell} \left( \partial_i g_{j\ell} +\partial_j g_{i\ell} -\partial_\ell g_{ij} \right),

where (gkℓ)(g^{k\ell}) is the inverse metric matrix. Thus the metric determines the connection through its first derivatives. The symbols themselves are not tensor components; they acquire extra terms under a coordinate change. At any chosen point, normal coordinates make all Christoffel symbols vanish there, but this does not generally make them vanish throughout a neighborhood.

The connection extends to a covariant derivative of tensor fields and to differentiation along curves. A vector field V(t)V(t) along γ\gamma is parallel when

DVdt=0.\frac{DV}{dt}=0.

Parallel transport preserves lengths and inner products. It supplies a linear isometry between the tangent spaces at the endpoints of a curve, but the result can depend on the path. For example, transporting a vector around a suitable closed triangle on a sphere rotates it relative to its initial direction. This path dependence is a manifestation of curvature.

The transformations obtained by parallel transport around loops based at a point form the holonomy group. Its structure can reveal geometric features not apparent in a coordinate formula. Certain reductions of holonomy correspond to parallel complex structures, special differential forms, or product decompositions. The de Rham decomposition theorem makes one such relation precise: for complete simply connected Riemannian manifolds, reducibility of the holonomy representation leads to an isometric product decomposition.

The connection also separates intrinsic and extrinsic descriptions. For a submanifold of Euclidean space, the Levi-Civita derivative is the tangent component of the ordinary ambient derivative. The normal component is encoded by the second fundamental form. This decomposition explains why a curve may have zero intrinsic acceleration while bending in the ambient space. A great circle is a geodesic of a sphere even though, viewed in R3\mathbb R^3, it is an accelerated circular trajectory.

Curvature and the behavior of nearby geodesics

The Riemann curvature tensor measures the failure of covariant derivatives to commute. With one standard sign convention,

R(X,Y)Z=∇X∇YZ−∇Y∇XZ−∇[X,Y]Z.R(X,Y)Z = \nabla_X\nabla_YZ-\nabla_Y\nabla_XZ-\nabla_{[X,Y]}Z.

For linearly independent tangent vectors u,vu,v, the sectional curvature of their spanned plane is

K(u,v)=g(R(u,v)v,u)g(u,u)g(v,v)−g(u,v)2.K(u,v)= \frac{g(R(u,v)v,u)} {g(u,u)g(v,v)-g(u,v)^2}.

Under these conventions round spheres have positive curvature. Some references use the opposite sign for the curvature tensor, requiring corresponding changes in formulas.

Sectional curvature describes curvature in individual two-dimensional tangent directions. In dimension two there is only one tangent plane at each point, and sectional curvature is the Gaussian curvature. In higher dimensions it can vary with the plane. Tracing the curvature tensor gives the Ricci tensor,

Ric⁡(u,v)=∑a=1ng(R(ea,u)v,ea),\operatorname{Ric}(u,v) = \sum_{a=1}^n g(R(e_a,u)v,e_a),

for an orthonormal basis e1,…,ene_1,\ldots,e_n. Tracing again gives scalar curvature,

Scal⁡=∑a=1nRic⁡(ea,ea).\operatorname{Scal}=\sum_{a=1}^n\operatorname{Ric}(e_a,e_a).

Constant sectional curvature KK implies Ric⁡=(n−1)Kg\operatorname{Ric}=(n-1)Kg and Scal⁡=n(n−1)K\operatorname{Scal}=n(n-1)K.

These contractions retain different amounts of information. In dimension two, the scalar curvature is twice the Gaussian curvature. In dimension three, the Ricci tensor determines the full Riemann tensor. In dimensions four and higher, an additional component, the Weyl tensor, can remain nonzero even when Ricci curvature vanishes. Thus “Ricci-flat” does not generally mean flat. A metric is locally Euclidean precisely when its entire Riemann curvature tensor vanishes.

Curvature governs the separation of neighboring geodesics. A variation through geodesics produces a Jacobi field JJ, satisfying

D2Jdt2+R(J,γ˙)γ˙=0.\frac{D^2J}{dt^2}+R(J,\dot\gamma)\dot\gamma=0.

Positive sectional curvature tends to focus nearby geodesics, whereas negative curvature tends to separate them, with precise statements expressed through comparison theorems. A nonzero Jacobi field vanishing at two points makes those points conjugate along the geodesic. Beyond the first conjugate point, that geodesic cannot remain minimizing.

Curvature also appears in small-scale volume measurements. In normal coordinates centered at pp,

det⁡(gij(x))=1−16Ric⁡ij(p)xixj+O(∣x∣3).\sqrt{\det(g_{ij}(x))} = 1-\frac16\operatorname{Ric}_{ij}(p)x^ix^j +O(|x|^3).

Consequently, in dimensions at least two, the volume of a small geodesic ball satisfies

Vol⁡(Br(p))=ωnrn(1−Scal⁡(p)6(n+2)r2+O(r4)),\operatorname{Vol}(B_r(p)) = \omega_n r^n \left( 1-\frac{\operatorname{Scal}(p)}{6(n+2)}r^2 +O(r^4) \right),

where ωn\omega_n is the volume of the Euclidean unit ball. Scalar curvature therefore measures the leading averaged departure of small-ball volume from its Euclidean value.

Global restrictions from curvature and topology

One of the defining achievements of Riemannian geometry is the derivation of global conclusions from local curvature conditions. The Gauss–Bonnet theorem provides a fundamental example. For a compact oriented surface without boundary,

∫MK dA=2πχ(M),\int_M K\,dA=2\pi\chi(M),

where χ(M)\chi(M) is the Euler characteristic. For an orientable surface of genus hh, this equals 2−2h2-2h. A sphere therefore has total curvature 4π4\pi, a torus has total curvature zero, and a surface of genus greater than one has negative total curvature.

The theorem constrains which curvature signs can occur everywhere on a given surface. A torus cannot carry a metric of strictly positive Gaussian curvature everywhere, even though it can carry metrics whose curvature changes sign or vanishes identically. In higher even dimensions, the Chern–Gauss–Bonnet theorem expresses the Euler characteristic as the integral of a polynomial in the curvature tensor. The integrand is more elaborate than scalar curvature alone.

The Cartan–Hadamard theorem states that a complete simply connected Riemannian manifold of nonpositive sectional curvature has an exponential map at every point that is a global diffeomorphism. It is therefore diffeomorphic to Rn\mathbb R^n, and any two points are joined by a unique geodesic. Without simple connectedness, uniqueness can fail: a flat torus has many geodesics between suitable pairs of points. Its universal cover, rather than the torus itself, has the global Euclidean-type structure described by the theorem.

Positive Ricci curvature gives a contrasting restriction. The Bonnet–Myers theorem states that a complete connected nn-dimensional manifold, n≥2n\geq2, satisfying

Ric⁡≥(n−1)k g,k>0,\operatorname{Ric}\geq(n-1)k\,g,\qquad k>0,

has diameter at most π/k\pi/\sqrt{k}. It follows that the manifold is compact and its fundamental group is finite. The uniform positive lower bound is essential; pointwise positivity without such a bound does not yield the same conclusion on a noncompact manifold.

Comparison geometry refines these relationships quantitatively. The Bishop–Gromov theorem compares volumes of geodesic balls under Ricci curvature lower bounds. In the nonnegative-Ricci case,

r⟼Vol⁡(Br(p))rnr\longmapsto \frac{\operatorname{Vol}(B_r(p))}{r^n}

is nonincreasing. Toponogov’s theorem compares geodesic triangles under sectional curvature lower bounds with triangles in constant-curvature model spaces. Such results do not merely attach a sign to curvature: they control distances, angles, volume growth, and the possible large-scale shapes of complete manifolds.

Integration, differential operators, and spectral geometry

A Riemannian metric determines a canonical volume density. In local coordinates it is

dvol⁡g=det⁡(gij) ∣dx1⋯dxn∣.d\operatorname{vol}_g = \sqrt{\det(g_{ij})}\, |dx^1\cdots dx^n|.

This permits integration even on nonorientable manifolds. On an oriented manifold it becomes the Riemannian volume form, with the absolute-value notation replaced by the oriented wedge product. The density transformation law ensures that integration is independent of coordinates. Surface area and ordinary Euclidean volume are special cases of this construction.

The metric converts differentials into vector fields. For a smooth function ff, its gradient is characterized by

g(grad⁡f,X)=df(X)g(\operatorname{grad}f,X)=df(X)

for every vector field XX. In coordinates,

(grad⁡f)i=gij∂jf.(\operatorname{grad}f)^i=g^{ij}\partial_jf.

The divergence of a vector field describes its infinitesimal volume expansion. Combining divergence and gradient gives the Laplace–Beltrami operator, the manifold version of the Laplacian:

Δgf=1det⁡g ∂i ⁣(det⁡g gij∂jf).\Delta_g f = \frac{1}{\sqrt{\det g}}\, \partial_i\!\left( \sqrt{\det g}\,g^{ij}\partial_j f \right).

With this convention, Δg\Delta_g has nonpositive eigenvalues on compact manifolds without boundary; many authors instead define the nonnegative operator −Δg-\Delta_g. The heat equation

∂tu=Δgu\partial_tu=\Delta_g u

describes diffusion governed by the metric. Integration by parts relates this operator to the energy of a function:

−∫MfΔgf dvol⁡g=∫M∣grad⁡f∣g2 dvol⁡g,-\int_M f\Delta_g f\,d\operatorname{vol}_g = \int_M|\operatorname{grad}f|_g^2\,d\operatorname{vol}_g,

for compact manifolds without boundary, or under appropriate support and boundary conditions.

On a compact connected manifold, the spectrum of −Δg-\Delta_g is a discrete sequence

0=λ0<λ1≤λ2≤⋯ ,0=\lambda_0<\lambda_1\leq\lambda_2\leq\cdots,

with multiplicities included, and its eigenfunctions give an orthonormal basis of the space of square-integrable functions. Weyl’s law states that the eigenvalue counting function satisfies

N(λ)∼ωn(2π)nVol⁡g(M)λn/2.N(\lambda)\sim \frac{\omega_n}{(2\pi)^n} \operatorname{Vol}_g(M)\lambda^{n/2}.

The spectrum therefore determines dimension and volume. It does not generally determine the manifold up to isometry: distinct compact Riemannian manifolds can be isospectral.

The metric also defines adjoints of differential operators on forms. Hodge theory identifies harmonic forms on a compact oriented Riemannian manifold with de Rham cohomology classes. Each class has a unique harmonic representative, even though that representative depends on the chosen metric. This is a precise link between analysis and topology: solutions of elliptic differential equations encode invariants that do not depend on the metric. Related methods extend to manifolds with boundary, where suitable boundary conditions become part of the problem.

Historical development and geometric evolution equations

The immediate background of Riemannian manifolds lies in the geometry of surfaces. In 1827, Carl Friedrich Gauss published the Theorema Egregium, establishing that Gaussian curvature is determined by the surface’s first fundamental form. Curvature could therefore be intrinsic: bending a surface without stretching it need not change its Gaussian curvature. A cylinder is locally isometric to the plane despite having a different appearance in three-dimensional space.

Bernhard Riemann extended this viewpoint in his 1854 habilitation lecture, On the Hypotheses Which Lie at the Foundations of Geometry, published posthumously in 1868. He considered spaces of arbitrary dimension and local squared-length expressions with coefficients varying from point to point. This replaced the assumption of a single fixed background geometry with the study of metrics themselves. Elwin Bruno Christoffel’s work in 1869 and the subsequent tensor calculus of Gregorio Ricci-Curbastro and Tullio Levi-Civita provided tools for coordinate-independent computation.

During the twentieth century, local differential geometry became increasingly connected to global topology and analysis. The Hopf–Rinow theorem of 1931 clarified completeness. Developments in comparison geometry, Morse theory, and Hodge theory related geodesics and differential operators to global structure. John Nash’s smooth isometric embedding theorem of 1956 established that abstract smooth Riemannian manifolds can be realized isometrically in a Euclidean space of sufficiently high dimension. This validates an extrinsic realization without making that realization necessary for the intrinsic definition.

A major later development was Ricci flow, introduced by Richard Hamilton in 1982:

∂g∂t=−2Ric⁡(g).\frac{\partial g}{\partial t}=-2\operatorname{Ric}(g).

Here the metric evolves according to its own curvature. Subject to gauge-fixing, the equation has a parabolic character and often smooths geometric irregularities, but singularities can form. A round sphere, for example, shrinks under the unnormalized flow. Techniques for understanding and continuing through singular behavior became central to applications in three-dimensional topology.

Grigori Perelman’s work of 2002–2003 supplied decisive tools for the Ricci-flow approach to three-dimensional geometrization, including monotonicity formulas and noncollapsing results. Together with the broader program developed by Hamilton, these advances led to proofs of the Poincaré and geometrization conjectures. The geometric object in this approach is not a single fixed metric but a time-dependent family on an underlying manifold, with controlled surgery used in the treatment of singularities. This perspective has made metric evolution an important method for investigating topological questions.

Relations to physics, statistics, and computation

Riemannian geometry provides useful language for physics, but the distinction between positive-definite and indefinite metrics is essential. Spacetime in general relativity is modeled by a Lorentzian pseudo-Riemannian manifold, not a Riemannian manifold: nonzero vectors can have positive, negative, or zero squared norm. Many constructions, including covariant differentiation and curvature, have analogues in both settings, but Riemannian distance and completeness theorems do not transfer unchanged. Riemannian metrics nevertheless occur on spacelike hypersurfaces and in Euclidean formulations of field theories.

In classical mechanics, configuration spaces often carry Riemannian metrics through kinetic energy. If generalized coordinates are q1,…,qnq^1,\ldots,q^n, an expression

T=12gij(q)q˙iq˙jT=\frac12g_{ij}(q)\dot q^i\dot q^j

defines a metric when the quadratic form is positive definite. Free motion with this kinetic energy follows geodesics. Constraints can make the configuration space curved even when the constituent particles move in ordinary Euclidean space. Potential energy introduces additional forces, so motion is not generally geodesic for the original kinetic metric.

In statistics, a regular identifiable family of probability densities p(x;θ)p(x;\theta) may carry the Fisher information metric:

gij(θ)=Eθ[∂ilog⁡p(X;θ) ∂jlog⁡p(X;θ)].g_{ij}(\theta) = \mathbb E_\theta \left[ \partial_i\log p(X;\theta)\, \partial_j\log p(X;\theta) \right].

Positive definiteness requires suitable nondegeneracy assumptions. This metric measures local statistical distinguishability. Under standard regularity conditions, the Kullback–Leibler divergence has the local expansion

DKL(pθ∥pθ+dθ)=12gij(θ) dθidθj+o(∥dθ∥2).D_{\mathrm{KL}}(p_\theta\Vert p_{\theta+d\theta}) = \frac12g_{ij}(\theta)\,d\theta^i d\theta^j +o(\|d\theta\|^2).

The divergence is not itself a symmetric distance, but its quadratic term yields Riemannian geometry.

Riemannian optimization treats constraints by regarding the feasible set as a manifold. Examples include fixed-length vectors on spheres, orthonormal frames, and symmetric positive-definite matrices. A metric determines the gradient direction, while geodesics or computationally cheaper retractions move an iterate along the manifold. Different metrics can give different algorithms on the same feasible set. With the Fisher metric, the resulting steepest-descent direction is the natural gradient, which accounts for the statistical geometry of the parameterization.

Numerical geometry also approximates lengths, curvature, diffusion, and eigenfunctions from meshes or sampled data. These computations require more than an abstract topology: they depend on a chosen metric and on how faithfully the discretization represents it. Finally, Riemannian metrics form only one class of length structures. A Finsler metric permits tangent-vector norms that need not arise from inner products, while sub-Riemannian geometry allows motion only in selected tangent directions. Their different distance and geodesic behavior marks the limits of the Riemannian framework.