aiwiki.page
English
Mathematics / Geometry & Topology / affine-connection

Affine Connection

An affine connection defines differentiation and parallel transport of tangent vectors on a manifold, independently of a metric.

37 keywords1 linked from13 not yet writtenWritten by AI
ManifoldTangent SpaceGeodesicTangent BundleMetric TensorLevi-Civita Conn…Differential Geo…Vector FieldAffine Con…

An affine connection is a geometric structure that specifies how tangent vectors are differentiated and transported on a smooth manifold. Vectors based at different points belong to different tangent spaces, so their difference is not intrinsically defined. A connection supplies the additional information needed to compare them infinitesimally. It thereby makes possible a theory of acceleration, parallelism, and geodesics without requiring a notion of length or angle.

In modern terminology, an affine connection is a linear connection on the manifold’s tangent bundle. Its principal invariants are torsion and curvature: torsion describes an asymmetry in differentiation, while curvature measures the local failure of parallel transport to be path-independent. When a metric tensor is present, the requirement of metric compatibility together with vanishing torsion selects the unique Levi–Civita connection. General affine connections need satisfy neither condition.

Affine connections are fundamental to differential geometry, gravitational theory, and geometric approaches to statistics. Their study brings together local differential equations, global topology, and transformation laws, while distinguishing properties of a chosen connection from those of the underlying manifold.

Definition and affine structure

Let MM be a smooth real manifold, let C∞(M)C^\infty(M) denote its smooth real-valued functions, and let X(M)\mathfrak X(M) denote its smooth vector fields. An affine connection is an operation

∇:X(M)×X(M)⟶X(M),(X,Y)⟼∇XY,\nabla:\mathfrak X(M)\times\mathfrak X(M)\longrightarrow\mathfrak X(M), \qquad (X,Y)\longmapsto\nabla_XY,

that is real-bilinear and satisfies

∇fXY=f∇XY,∇X(fY)=X(f)Y+f∇XY\nabla_{fX}Y=f\nabla_XY,\qquad \nabla_X(fY)=X(f)Y+f\nabla_XY

for every smooth function ff. The expression ∇XY\nabla_XY, called the covariant derivative of YY in the direction XX, is tensorial in XX but differentiates YY. At a point pp, its value depends only on XpX_p and the first-order behavior of YY near pp.

These rules generalize ordinary directional differentiation in Euclidean space. In Cartesian coordinates on Rn\mathbb R^n, the standard connection differentiates the components of a vector field:

∇XY=Xi∂iYk ∂k.\nabla_XY=X^i\partial_iY^k\,\partial_k.

On a general manifold, no coordinate system is distinguished, so a connection cannot be defined merely by declaring this formula valid in every chart. Correction terms are necessary to make the derivative independent of the chosen coordinates.

The term “affine” reflects an important algebraic property. If ∇\nabla and ∇^\widehat\nabla are two connections, their difference

A(X,Y)=∇^XY−∇XYA(X,Y)=\widehat\nabla_XY-\nabla_XY

is linear over smooth functions in both arguments. Thus AA is a tensor field of type (1,2)(1,2), a section of T∗M⊗T∗M⊗TMT^*M\otimes T^*M\otimes TM. Conversely, adding any such tensor to a connection produces another connection. Consequently, the set of connections is an affine space modeled on this vector space of tensor fields: two connections can be subtracted, but there is generally no preferred zero connection.

Under the usual assumptions that a smooth manifold is Hausdorff and second countable, affine connections always exist. One construction begins with ordinary coordinate differentiation on individual charts and combines the resulting local connections using a smooth partition of unity. A weighted sum ∑αfα∇(α)\sum_\alpha f_\alpha\nabla^{(\alpha)}, with ∑αfα=1\sum_\alpha f_\alpha=1, again satisfies the connection axioms.

An affine connection is therefore additional structure, not an automatic consequence of smoothness. A manifold can carry many inequivalent connections, even when its topology and differentiable structure are fixed. The definition also makes no reference to a metric: parallelism and geodesic acceleration can be specified before distances, angles, or orthogonality have been introduced.

Local coefficients and coordinate transformations

In local coordinates x1,…,xnx^1,\ldots,x^n, write ∂i=∂/∂xi\partial_i=\partial/\partial x^i. A connection is determined by its coefficients

∇∂i∂j=Γkij∂k.\nabla_{\partial_i}\partial_j=\Gamma^k{}_{ij}\partial_k.

The functions Γkij\Gamma^k{}_{ij} are commonly called Christoffel symbols or connection coefficients. Throughout these formulas, repeated upper and lower indices are summed according to the Einstein summation convention. For vector fields X=Xi∂iX=X^i\partial_i and Y=Yj∂jY=Y^j\partial_j, the connection axioms give

∇XY=Xi(∂iYk+ΓkijYj)∂k.\nabla_XY =X^i\left(\partial_iY^k+\Gamma^k{}_{ij}Y^j\right)\partial_k.

Thus a connection is locally described by n3n^3 smooth coefficient functions, although these functions change nontrivially when coordinates change.

If yay^a is another coordinate system, then

Γ~abc=∂ya∂xk∂xi∂yb∂xj∂ycΓkij+∂ya∂xk∂2xk∂yb∂yc.\widetilde\Gamma^a{}_{bc} = \frac{\partial y^a}{\partial x^k} \frac{\partial x^i}{\partial y^b} \frac{\partial x^j}{\partial y^c} \Gamma^k{}_{ij} + \frac{\partial y^a}{\partial x^k} \frac{\partial^2x^k}{\partial y^b\partial y^c}.

The second-derivative term distinguishes connection coefficients from tensor components. It compensates for the derivative of the coordinate basis itself. When the coefficients of two connections are subtracted, these inhomogeneous terms cancel, explaining why their difference is a tensor.

A simple example is the standard flat connection on the Euclidean plane. Its coefficients vanish in Cartesian coordinates, but in polar coordinates (r,θ)(r,\theta), on a region where these coordinates are valid, the nonzero coefficients are

Γrθθ=−r,Γθrθ=Γθθr=1r.\Gamma^r{}_{\theta\theta}=-r,\qquad \Gamma^\theta{}_{r\theta} =\Gamma^\theta{}_{\theta r}=\frac1r.

These terms describe the variation of the polar coordinate basis, not an intrinsic curvature of the plane. Nonzero coefficients alone therefore do not establish that a connection is curved.

At a chosen point, coordinates can eliminate the symmetric part of the coefficients. If the connection is torsion-free, all coefficients can be made zero at that point. If torsion is nonzero there, its antisymmetric contribution cannot be removed by a coordinate change. Even for a torsion-free connection, vanishing coefficients at one point do not imply vanishing curvature, because curvature also involves derivatives of the coefficients.

One may instead use a general local frame eae_a, not necessarily arising from coordinates. Writing

∇eaeb=Γcabec\nabla_{e_a}e_b=\Gamma^c{}_{ab}e_c

again specifies the connection. Such frames can be adapted to a metric or other geometric structure, but their commutators need not vanish. The distinction between coordinate frames and arbitrary frames becomes essential in formulas for torsion and in the moving-frame description of geometry.

Extension to covectors and tensors

A connection on tangent vectors determines differentiation on all ordinary tensor fields. The extension is fixed by requiring the usual derivative on functions, a product rule for tensor products, and compatibility with contraction. In particular, the derivative of a scalar function is

∇Xf=X(f).\nabla_Xf=X(f).

No connection is required to define this first derivative; the additional structure enters when differentiating vector-valued or covector-valued quantities.

For a covector field α\alpha, compatibility with the scalar pairing α(Y)\alpha(Y) requires

(∇Xα)(Y)=X(α(Y))−α(∇XY).(\nabla_X\alpha)(Y) =X\bigl(\alpha(Y)\bigr)-\alpha(\nabla_XY).

Its coordinate expression is

(∇iα)j=∂iαj−Γkijαk.(\nabla_i\alpha)_j =\partial_i\alpha_j-\Gamma^k{}_{ij}\alpha_k.

The minus sign is forced by cancellation of connection terms when a covector and a vector are paired. More generally, the covariant derivative of a tensor has one positive connection term for every upper index and one negative term for every lower index.

For example, a type-(1,1)(1,1) tensor SS satisfies

(∇iS)ab=∂iSab+ΓaicScb−ΓcibSac.(\nabla_iS)^a{}_b = \partial_iS^a{}_b +\Gamma^a{}_{ic}S^c{}_b -\Gamma^c{}_{ib}S^a{}_c.

The covariant derivative is itself a tensor, with one extra covariant index recording the direction of differentiation. This contrasts with the array of ordinary partial derivatives of tensor components, which generally fails to transform tensorially.

Second derivatives of a function illustrate the distinction. Its connection-dependent Hessian is the covariant two-tensor

Hess⁡∇f(X,Y)=X(Yf)−(∇XY)f,\operatorname{Hess}^{\nabla}f(X,Y) =X(Yf)-(\nabla_XY)f,

with components

(Hess⁡∇f)ij=∂i∂jf−Γkij∂kf.(\operatorname{Hess}^{\nabla}f)_{ij} =\partial_i\partial_jf-\Gamma^k{}_{ij}\partial_kf.

If TT denotes torsion, then

Hess⁡∇f(X,Y)−Hess⁡∇f(Y,X)=−df(T(X,Y)).\operatorname{Hess}^{\nabla}f(X,Y) -\operatorname{Hess}^{\nabla}f(Y,X) =-df\bigl(T(X,Y)\bigr).

Thus torsion-free connections give symmetric Hessians. At a critical point of ff, where df=0df=0, the Hessian is independent of the chosen connection, because the connection-dependent correction vanishes.

Covariant differentiation must also be distinguished from differentiation by the flow of a vector field. The latter gives the Lie derivative, which needs no connection. On vector fields it is determined by the Lie bracket, whereas ∇XY\nabla_XY depends on the selected connection. For a torsion-free connection their relationship is

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

Likewise, the exterior derivative of a differential form is defined independently of any connection. When expressed through covariant derivatives, its familiar antisymmetrized formula acquires torsion terms unless the connection is torsion-free.

Parallel transport along curves

An affine connection turns infinitesimal differentiation into a rule for transporting vectors along paths. Let γ:I→M\gamma:I\to M be a smooth curve and let V(t)∈Tγ(t)MV(t)\in T_{\gamma(t)}M be a vector field along it. The connection defines a covariant derivative DV/dtDV/dt. In coordinates,

DVkdt=dVkdt+Γkij(γ(t))γ˙iVj.\frac{DV^k}{dt} = \frac{dV^k}{dt} +\Gamma^k{}_{ij}(\gamma(t))\dot\gamma^iV^j.

The field is parallel when DV/dt=0DV/dt=0. For a prescribed initial vector V(t0)V(t_0), this is a linear ordinary differential equation, so existence and uniqueness determine the transported vector along the curve.

For a piecewise smooth path from pp to qq, parallel transport defines an invertible linear map

Pγ:TpM⟶TqM.P_\gamma:T_pM\longrightarrow T_qM.

Transport along a concatenated path is the composition of the corresponding maps, and transport along the reversed path is the inverse. An orientation-preserving reparameterization leaves the transport unchanged: the rule depends on the directed path, not on the speed at which it is traversed.

Connections can conversely be recovered from their parallel transport. If γ(0)=p\gamma(0)=p and γ˙(0)=Xp\dot\gamma(0)=X_p, then the derivative of a vector field YY can be obtained by transporting its nearby values back to TpMT_pM:

(∇XY)p=ddt∣t=0Pγ∣[0,t]−1Yγ(t).(\nabla_XY)_p = \left.\frac{d}{dt}\right|_{t=0} P_{\gamma|[0,t]}^{-1}Y_{\gamma(t)}.

This formula displays the role of transport explicitly: vectors must first be placed in a common vector space before an ordinary derivative can be taken.

Parallel transport generally depends on the path, even when its endpoints are fixed. On a round sphere with its Levi–Civita connection, transporting a tangent vector around a geodesic triangle produces a rotation related to the triangle’s area. For a positively oriented simple triangle on a sphere of radius rr, the rotation angle, modulo a full turn and subject to orientation convention, is its area divided by r2r^2. This gives a geometric manifestation of curvature.

Without a metric-compatibility condition, transport need not preserve lengths, angles, or volume. It preserves the linear structure of tangent spaces, but that is a weaker requirement. Moreover, parallel transport may exist along every finite segment of a specified smooth curve even when some geodesics cannot be continued for all parameter values. The former solves a linear equation over an already given path; the latter requires solving a nonlinear equation for the path itself.

Geodesics and the exponential map

A geodesic of an affine connection is a curve whose velocity is parallel along itself:

∇γ˙γ˙=0.\nabla_{\dot\gamma}\dot\gamma=0.

In coordinates this becomes

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

For every initial point pp and initial tangent vector v∈TpMv\in T_pM, the local existence and uniqueness theorem for differential equations gives a unique geodesic with γ(0)=p\gamma(0)=p and γ˙(0)=v\dot\gamma(0)=v. These curves are also called autoparallels, especially when distinguishing an arbitrary connection from one derived from a metric.

The equation selects both a path and a class of parameters. If tt is an affine parameter, then at+bat+b, with a≠0a\ne0, is another affine parameter. A general reparameterization instead gives an equation of the form

∇γ˙γ˙=f γ˙\nabla_{\dot\gamma}\dot\gamma=f\,\dot\gamma

for some scalar function ff. The image of a geodesic can therefore be studied independently of its preferred parameterization, but the parameterized and unparameterized notions contain different information.

Only the part of Γkij\Gamma^k{}_{ij} symmetric in i,ji,j enters the geodesic equation, because x˙ix˙j\dot x^i\dot x^j is symmetric. Consequently, geodesics alone do not determine torsion. Adding to a connection any (1,2)(1,2)-tensor antisymmetric in its two lower arguments leaves all affinely parameterized geodesics unchanged. Parallel transport, however, usually changes.

The exponential map at pp is defined by

exp⁡p(v)=γv(1)\exp_p(v)=\gamma_v(1)

whenever the geodesic with initial velocity vv exists up to parameter 11. Its derivative at the origin is the identity on TpMT_pM, so it is a diffeomorphism from a sufficiently small neighborhood of 00 onto a neighborhood of pp. Choosing a basis of TpMT_pM then gives normal coordinates. For a torsion-free connection, all connection coefficients vanish at pp in these coordinates; for a connection with torsion, only their symmetric part must vanish.

A connection is geodesically complete if every maximal geodesic is defined for all real affine parameters. Compactness of the manifold does not by itself guarantee completeness for an arbitrary affine connection. By contrast, the Levi–Civita connection of a positive-definite Riemannian metric on a compact manifold is complete. Nor are affine geodesics intrinsically shortest paths: minimizing length requires a metric, and the identification of sufficiently short geodesic segments with local length minimizers belongs specifically to positive-definite Riemannian geometry.

Torsion, curvature, and differential identities

The torsion tensor of a connection is

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

It is antisymmetric and linear over smooth functions in both arguments. In a coordinate frame,

Tkij=Γkij−Γkji.T^k{}_{ij}=\Gamma^k{}_{ij}-\Gamma^k{}_{ji}.

A connection is torsion-free when this tensor vanishes. In a general frame, if [ea,eb]=ccabec[e_a,e_b]=c^c{}_{ab}e_c, then

Tcab=Γcab−Γcba−ccab.T^c{}_{ab} =\Gamma^c{}_{ab}-\Gamma^c{}_{ba}-c^c{}_{ab}.

The commutator term prevents symmetry of frame coefficients from being mistaken for an invariant condition.

With a common sign convention, curvature is defined by

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.

It is a type-(1,3)(1,3) tensor, antisymmetric in X,YX,Y. Writing

R(∂i,∂j)∂ℓ=Rkℓij∂k,R(\partial_i,\partial_j)\partial_\ell =R^k{}_{\ell ij}\partial_k,

its components are

Rkℓij=∂iΓkjℓ−∂jΓkiℓ+ΓkimΓmjℓ−ΓkjmΓmiℓ.R^k{}_{\ell ij} = \partial_i\Gamma^k{}_{j\ell} -\partial_j\Gamma^k{}_{i\ell} +\Gamma^k{}_{im}\Gamma^m{}_{j\ell} -\Gamma^k{}_{jm}\Gamma^m{}_{i\ell}.

Curvature is the infinitesimal obstruction to path-independent parallel transport: transport around a sufficiently small loop differs from the identity by a leading term governed by curvature and the loop’s oriented area. Some authors reverse the overall sign of RR, so formulas must be compared with their conventions.

Torsion and curvature are logically independent. A connection can be flat, meaning R=0R=0, while having nonzero torsion. Conversely, Levi–Civita connections are torsion-free but commonly curved. Every connection has a torsion-free counterpart

∇X0Y=∇XY−12T(X,Y),\nabla^0_XY=\nabla_XY-\tfrac12T(X,Y),

with exactly the same affinely parameterized geodesics. This construction separates information encoded by the geodesic equation from the antisymmetric part of differentiation.

The Bianchi identities constrain these tensors. For a torsion-free connection, the first identity is

R(X,Y)Z+R(Y,Z)X+R(Z,X)Y=0,R(X,Y)Z+R(Y,Z)X+R(Z,X)Y=0,

and the second is

(∇XR)(Y,Z)+(∇YR)(Z,X)+(∇ZR)(X,Y)=0.(\nabla_XR)(Y,Z)+(\nabla_YR)(Z,X)+(\nabla_ZR)(X,Y)=0.

In the presence of torsion, additional terms appear. A compact formulation uses a local coframe θa\theta^a, connection one-forms ωab\omega^a{}_b, torsion forms Θa\Theta^a, and curvature forms Ωab\Omega^a{}_b:

Θa=dθa+ωab∧θb,Ωab=dωab+ωac∧ωcb.\Theta^a=d\theta^a+\omega^a{}_b\wedge\theta^b,\qquad \Omega^a{}_b=d\omega^a{}_b+\omega^a{}_c\wedge\omega^c{}_b.

If DD denotes the exterior covariant derivative, the identities become

DΘa=Ωab∧θb,DΩab=0.D\Theta^a=\Omega^a{}_b\wedge\theta^b,\qquad D\Omega^a{}_b=0.

For a metric-compatible connection, lowering a curvature index produces additional symmetries. In the torsion-free case this yields the familiar Riemann curvature tensor with pair interchange and antisymmetry properties. These extra symmetries should not be assumed for a general affine connection; even its Ricci contraction need not be symmetric.

Compatibility with metrics

Given a nondegenerate symmetric metric tensor gg, a connection is metric-compatible when ∇g=0\nabla g=0. Equivalently,

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

This condition says that parallel transport preserves the metric pairing. In positive-definite geometry it preserves lengths and angles; for an indefinite metric it preserves the corresponding bilinear form and causal classification of nonzero vectors.

The fundamental theorem of Riemannian geometry states that every smooth nondegenerate metric has exactly one torsion-free metric-compatible connection. Its coordinate coefficients are

Γkij=12gkℓ(∂igjℓ+∂jgiℓ−∂ℓgij),\Gamma^k{}_{ij} = \frac12g^{k\ell} \left( \partial_i g_{j\ell} +\partial_j g_{i\ell} -\partial_\ell g_{ij} \right),

where gkℓg^{k\ell} is the inverse metric. This is the Levi–Civita connection. The theorem applies to both positive-definite and indefinite metrics, including the Lorentzian metrics used in relativity.

The two defining conditions have separate roles. Metric compatibility alone does not force torsion to vanish, and torsion-freeness alone does not identify a connection from a metric. If

∇XY=∇XgY+A(X,Y),\nabla_XY=\nabla^g_XY+A(X,Y),

where ∇g\nabla^g is Levi–Civita, then metric compatibility is equivalent to

g(A(X,Y),Z)+g(Y,A(X,Z))=0.g(A(X,Y),Z)+g(Y,A(X,Z))=0.

Thus, for fixed XX, the endomorphism Y↦A(X,Y)Y\mapsto A(X,Y) must be skew-adjoint relative to gg. For a fixed metric, prescribing torsion uniquely determines a metric-compatible connection; its difference from Levi–Civita is commonly described by the contorsion tensor.

Failure of metric compatibility is measured by nonmetricity. With the convention

Q(X;Y,Z)=(∇Xg)(Y,Z),Q(X;Y,Z)=(\nabla_Xg)(Y,Z),

it is symmetric in Y,ZY,Z. Some literature defines QQ with the opposite sign. Along a curve, for parallel fields V,WV,W,

ddtg(V,W)=Q(γ˙;V,W).\frac{d}{dt}g(V,W)=Q(\dot\gamma;V,W).

This directly relates nonmetricity to changes in transported metric pairings. Weyl-type connections provide an important special case, in which ∇g\nabla g is proportional to gg through a one-form; they preserve conformal rather than metric information.

Conversely, a connection need not be compatible with any positive-definite metric. The obstruction can be expressed through its parallel transport around loops: a compatible positive-definite metric at one point must be invariant under every such transport map. If such an invariant inner product exists on a connected manifold, transporting it to other points constructs a globally defined parallel Riemannian metric. This makes metric compatibility a global question about the connection as well as a local system of differential equations.

Bundle formulation and holonomy

The modern bundle formulation places affine connections within a broader theory. A connection on a vector bundle E→ME\to M is an operator

∇:Γ(E)⟶Γ(T∗M⊗E)\nabla:\Gamma(E)\longrightarrow\Gamma(T^*M\otimes E)

satisfying

∇(fs)=df⊗s+f∇s.\nabla(fs)=df\otimes s+f\nabla s.

An affine connection is the case E=TME=TM. The same construction applies to bundles whose fibers are not tangent spaces, such as those carrying internal degrees of freedom in field theory. Curvature and parallel transport extend to these settings, while torsion does not arise for an arbitrary vector bundle without further structure relating it to TMTM.

The frame bundle F(M)F(M) is a principal bundle with structure group GL(n,R)GL(n,\mathbb R). Its fiber at pp consists of ordered bases of TpMT_pM. An affine connection is equivalently a principal connection on this frame bundle: it selects horizontal directions complementary to the directions tangent to each fiber. A horizontal lift of a curve is precisely a frame whose constituent vectors remain parallel along that curve.

In a local frame, the connection is represented by a matrix of one-forms ω\omega. If the frame changes by a smooth invertible matrix AA, with e′=eAe'=eA, then

ω′=A−1ωA+A−1dA.\omega'=A^{-1}\omega A+A^{-1}dA.

The curvature matrix satisfies

Ω=dω+ω∧ω,Ω′=A−1ΩA.\Omega=d\omega+\omega\wedge\omega,\qquad \Omega'=A^{-1}\Omega A.

The inhomogeneous transformation of ω\omega and homogeneous transformation of Ω\Omega reproduce the distinction between connection coefficients and curvature tensors.

The holonomy group at a point pp consists of all parallel-transport maps around piecewise smooth loops based at pp. It is a subgroup of GL(TpM)GL(T_pM). Changing the base point within a connected manifold changes the group by conjugation through a transport map. The restricted holonomy group uses loops contractible to the constant loop and records the component generated by local curvature effects. The Ambrose–Singer theorem identifies its Lie algebra in terms of curvature endomorphisms transported back to the base point.

Holonomy translates differential conditions into invariant linear algebra. A globally parallel tensor must have its value fixed by holonomy. Conversely, on a connected manifold, a tensor at one point invariant under the full holonomy group extends uniquely by parallel transport to a global parallel tensor. Parallel metrics, complex structures, and differential forms can therefore be studied through holonomy reductions.

The word “affine” does not mean that the standard structure group here must be the affine group: the tangent-bundle connection uses GL(n,R)GL(n,\mathbb R). Related descriptions using affine frames and Cartan connections incorporate the canonical solder form, which identifies tangent directions with the vector representation. That additional structure is what makes torsion available in the frame-bundle setting.

Flat affine structures and projective equivalence

A connection is flat when its curvature vanishes. On a sufficiently small simply connected neighborhood, flatness guarantees a frame of parallel vector fields. It does not, by itself, guarantee coordinates in which all connection coefficients vanish. For a parallel frame eae_a,

T(ea,eb)=−[ea,eb].T(e_a,e_b)=-[e_a,e_b].

Thus the frame becomes a coordinate frame precisely when its fields commute. Vanishing curvature and vanishing torsion together imply the existence of local coordinates with identically zero connection coefficients.

A manifold equipped with a flat torsion-free connection is called an affine manifold. Its adapted coordinate charts have transition maps that are restrictions of affine transformations,

x⟼Ax+b,A∈GL(n,R).x\longmapsto Ax+b, \qquad A\in GL(n,\mathbb R).

Conversely, an atlas with such transition maps defines a flat torsion-free connection. Here the connection is locally indistinguishable from ordinary differentiation in affine space, but its global geometry can remain nontrivial.

For example, a torus obtained by quotienting Rn\mathbb R^n by a lattice inherits a complete flat torsion-free connection. More generally, affine manifolds are studied through a developing map from their universal covering space into affine space and an associated representation of the fundamental group into the affine transformation group. The developing map is locally invertible but need not be globally injective or surjective. Its affine holonomy includes translational information, whereas tangent-bundle holonomy records the corresponding linear part.

A different equivalence relation arises from geodesics. Two torsion-free connections have the same unparameterized geodesics exactly when, for some one-form α\alpha,

∇^XY=∇XY+α(X)Y+α(Y)X.\widehat\nabla_XY = \nabla_XY+\alpha(X)Y+\alpha(Y)X.

They are then called projectively equivalent connections. Along a geodesic, the change contributes only a multiple of its tangent vector to the acceleration, which can be absorbed by reparameterization. This geometric notion of projective equivalence is distinct from merely making a projective change of coordinates.

The torsion-free hypothesis matters: antisymmetric modifications are already invisible to the geodesic equation, so unrestricted connections have additional freedom. Even among torsion-free connections, the collection of unparameterized paths determines only a projective class, not a unique connection or affine parameter.

Flatness, projective flatness, and geodesic completeness are consequently different properties. A projectively flat connection has locally straight geodesic paths in suitable coordinates, but it may have nonzero curvature because their affine parameterizations differ from those of straight lines. A flat torsion-free connection can also be incomplete, as shown by the standard connection restricted to a proper open ball in Rn\mathbb R^n.

Historical development and applications

Affine connections emerged from nineteenth-century attempts to express geometry independently of coordinates. Bernhard Riemann’s 1854 habilitation lecture established the metric viewpoint in higher-dimensional geometry. Elwin Bruno Christoffel’s 1869 work developed the transformation machinery underlying the coefficients now bearing his name. Gregorio Ricci-Curbastro and Tullio Levi-Civita systematized tensor calculus in their joint 1900 memoir, providing tools that later became central to both geometry and gravitation.

Levi-Civita’s 1917 treatment of parallel displacement clarified the geometric meaning of differentiation on a Riemannian manifold. Hermann Weyl’s work in 1918 emphasized connections not determined by a fixed metric and linked geometric transport with scale. During the 1920s, Élie Cartan developed moving-frame methods, torsion, and broader connection theories. These developments separated several structures that coincide in elementary Euclidean geometry: a metric, a rule of parallel transport, a family of geodesics, and a curvature tensor.

In general relativity, standard metric formulations use the Levi–Civita connection of a Lorentzian spacetime metric. Freely falling test particles follow timelike geodesics in the idealized absence of nongravitational forces, and curvature controls the relative acceleration of neighboring geodesics. In a torsion-free connection, a geodesic variation with tangent UU and variation field JJ satisfies, with the curvature convention used here,

∇U∇UJ+R(J,U)U=0.\nabla_U\nabla_UJ+R(J,U)U=0.

This is the Jacobi equation, linking local curvature to tidal behavior and the focusing or spreading of nearby trajectories.

Other gravitational formulations treat metric and connection as independent variables. Einstein–Cartan theory permits torsion and relates it algebraically to matter spin in its standard formulation. Metric-affine theories also permit nonmetricity. Teleparallel descriptions employ flat connections that can carry torsion. Such theories illustrate why curvature, torsion, and metric compatibility must be distinguished rather than treated as interchangeable expressions of geometry.

In information geometry, statistical models carry structures derived from probability distributions. Under suitable regularity and nondegeneracy assumptions, Fisher information defines a Riemannian metric. Connections dual with respect to this metric satisfy

Xg(Y,Z)=g(∇XY,Z)+g(Y,∇X∗Z).Xg(Y,Z) = g(\nabla_XY,Z)+g(Y,\nabla_X^*Z).

For regular exponential families in their standard representation, exponential and mixture connections are flat and dual, while their average is the Levi–Civita connection. The resulting geometry distinguishes metric orthogonality from the affine structures associated with alternative statistical coordinates.

Connections also organize geometric differential operators and computation. They define covariant Hessians, tensor evolution equations, and derivatives of fields constrained to curved spaces. In manifold optimization, an exact exponential map and parallel transport are often replaced computationally by retractions and vector transports. These substitutes need not be the exponential map or parallel transport of a single affine connection, so their approximation properties must be distinguished from the defining geometric structures.