The Levi-Civita Connection: The One Compatible with the Metric

The precise statement

Let (M, g) be a smooth manifold equipped with a Riemannian metric g (a smooth, symmetric, positive-definite 2-tensor) — or more generally a nondegenerate pseudo-Riemannian metric of any signature. An affine connection ∇ is an ℝ-bilinear map (X, Y) ↦ ∇_XY on vector fields that is C^∞(M)-linear in X and satisfies the Leibniz rule ∇_X(fY) = (Xf)Y + f∇_XY.

The fundamental theorem of Riemannian geometry states: there exists a unique affine connection ∇ on M such that

• Metric compatibility: X⟨Y, Z⟩ = ⟨∇_XY, Z⟩ + ⟨Y, ∇_XZ⟩ for all vector fields X, Y, Z, equivalently ∇g = 0;
• Torsion-free: the torsion T(X, Y) = ∇_XY − ∇_YX − [X, Y] vanishes identically.

This ∇ is the Levi-Civita connection. Note nothing here requires completeness, compactness, or orientability — only smoothness and a nondegenerate metric.

The picture: transporting vectors without cheating

Imagine walking on a sphere carrying an arrow, trying to keep it 'as parallel as possible' as you go. There is no ambient notion of parallel intrinsic to the surface — a connection is precisely the rule that tells you how a vector changes as you move it. Many such rules exist; the metric selects the honest one.

Metric compatibility is the demand that this transport be an isometry: if you parallel-transport two vectors along any curve, their lengths and the angle between them never change. That is what makes 'parallel' deserve the name.

Torsion-free is subtler. Torsion measures the failure of infinitesimal parallelograms to close, or equivalently the antisymmetric part of ∇. Setting it to zero says the connection introduces no intrinsic 'twist': the symmetry Γᵏᵢⱼ = Γᵏⱼᵢ holds, and mixed covariant second derivatives of a function commute. Together these two natural, geometrically minimal demands leave no freedom at all.

The key idea: the Koszul formula forces the answer

The engine of the proof is a single algebraic identity. Suppose ∇ exists with both properties. Write metric compatibility three times, for the cyclic permutations of X, Y, Z:

• X⟨Y,Z⟩ = ⟨∇_XY,Z⟩ + ⟨Y,∇_XZ⟩
• Y⟨Z,X⟩ = ⟨∇_YZ,X⟩ + ⟨Z,∇_YX⟩
• Z⟨X,Y⟩ = ⟨∇_ZX,Y⟩ + ⟨X,∇_ZY⟩

Add the first two, subtract the third, and use the torsion-free identity ∇_XY − ∇_YX = [X, Y] to eliminate all terms except one. You obtain the Koszul formula:

2⟨∇_XY, Z⟩ = X⟨Y,Z⟩ + Y⟨Z,X⟩ − Z⟨X,Y⟩ + ⟨[X,Y],Z⟩ − ⟨[Y,Z],X⟩ + ⟨[Z,X],Y⟩.

The right side involves only g and Lie brackets — no ∇. Since g is nondegenerate, this determines ⟨∇_XY, Z⟩ for all Z, hence ∇_XY itself: uniqueness. For existence, define ∇ by this formula and check bilinearity, the Leibniz rule, compatibility, and vanishing torsion directly. Nondegeneracy is exactly what lets you invert ⟨·, Z⟩.

Worked example: Christoffel symbols and the sphere

In a coordinate chart with basis fields ∂ᵢ, the brackets [∂ᵢ, ∂ⱼ] vanish, so Koszul collapses to the classical Christoffel symbols:

Γᵏᵢⱼ = ½ gᵏˡ (∂ᵢ gⱼˡ + ∂ⱼ gᵢˡ − ∂ˡ gᵢⱼ),

where gᵏˡ is the inverse metric and ∇_∂ᵢ∂ⱼ = Γᵏᵢⱼ ∂ₖ (summation over repeated indices). Their symmetry Γᵏᵢⱼ = Γᵏⱼᵢ is exactly torsion-freeness.

Round sphere S² of radius 1, coordinates (θ, φ), metric ds² = dθ² + sin²θ dφ². The nonzero symbols are Γᶿ_φφ = −sinθ cosθ and Γᶠ_θφ = Γᶠ_φθ = cotθ. Feeding these into the geodesic equation d²xᵏ/dt² + Γᵏᵢⱼ (dxⁱ/dt)(dxʲ/dt) = 0 yields great circles as the geodesics — the shortest paths — confirming the connection encodes the geometry you expect from a globe.

Why nondegeneracy is the load-bearing hypothesis

The one hypothesis you cannot drop is nondegeneracy of g. The Koszul formula determines only the pairing ⟨∇_XY, Z⟩; recovering the vector ∇_XY requires the map v ↦ ⟨v, ·⟩ to be invertible, which is precisely nondegeneracy. On a degenerate metric (e.g. the induced 'metric' on a null hypersurface in Lorentzian geometry) no canonical Levi-Civita connection exists — this is a genuine obstruction, studied via Koszul connections and rigging techniques, not a mere technicality.

By contrast, positivity is not needed: the theorem holds verbatim for any signature, which is why general relativity uses the Levi-Civita connection of a Lorentzian metric. Dropping torsion-freeness instead gives the freedom to add any (1,2)-tensor with the right antisymmetry — this is the world of Einstein–Cartan theory and connections with torsion. Relatedly, the difference of any two connections is a tensor, so the space of metric-compatible connections is an affine space; torsion-free-ness cuts it down to a single point.

Why it matters: the foundation everything sits on

The Levi-Civita connection is the hinge that turns a metric into all of Riemannian geometry. Without a canonical ∇ you cannot even define the objects downstream:

• Geodesics and the exponential map: curves with ∇_γ′γ′ = 0 are the metric's straightest lines; they are locally length-minimizing precisely because ∇ is metric-compatible.
• Curvature: the Riemann tensor R(X,Y)Z = ∇_X∇_YZ − ∇_Y∇_XZ − ∇_[X,Y]Z is built from ∇; from it come sectional, Ricci, and scalar curvature.
• Physics: the geodesic equation is the equation of free fall, and the Ricci tensor of the Levi-Civita connection appears in Einstein's field equations.
• Theorems it unlocks: Gauss–Bonnet, the Bonnet–Myers diameter bound, Hopf–Rinow, comparison geometry, and Chern's intrinsic proof of Gauss–Bonnet all presuppose this connection.

Historically, Levi-Civita's 1917 notion of parallel transport (building on Ricci-Curbastro's calculus) gave Einstein the language for general relativity — a rare case where a purely geometric theorem reshaped physics.

Which connection properties pin down Levi-Civita, and what each rules out

Property	What it says	What fails without it
Metric compatibility (∇g = 0)	Parallel transport preserves inner products; ∂ₓ⟨Y,Z⟩ = ⟨∇ₓY,Z⟩ + ⟨Y,∇ₓZ⟩	Lengths and angles distort under parallel transport; geodesics need not be locally shortest
Torsion-free (T = 0)	∇ₓY − ∇ᵧX = [X,Y]; Γᵏᵢⱼ symmetric in i,j	Straightest curves can spiral; second covariant derivatives of functions aren't symmetric
Both together	Unique ∇ (Levi-Civita); Koszul formula determines it	—
Neither imposed	Infinitely many affine connections on M	No canonical differentiation of vector fields