The Affine Connection

Differentiation is the one operation that ordinary calculus hands you for free and curved geometry takes away. On the flat plane you subtract a vector at one point from a vector at a neighbouring point, divide by the step, and call the limit a derivative — no questions asked. But the two vectors do not live in the same place. A vector at $p$ belongs to the tangent space $T_pM$ built in Manifolds and Tangent Spaces, and a vector at a nearby $q$ belongs to a different vector space $T_qM$ . Nothing intrinsic to a bare smooth manifold tells you how to line the two spaces up, so the subtraction that defines a derivative is, literally, meaningless. The affine connection is the extra structure that repairs this: a rule for carrying a vector from $T_qM$ back to $T_pM$ so that "how much did it change" finally has an answer. That rule is packaged into a derivative operator $\nabla$ acting on the tensor fields of Tensor Fields, and in a coordinate basis its whole content collapses onto $n^3$ numbers, the connection coefficients $\Gamma^{\lambda}{}_{\mu\nu}$ , which record how the basis vectors themselves twist and stretch from point to point. The decisive and often-missed fact is that a connection needs no metric. Parallel transport, geodesics, affine parameters, and the entire Riemann curvature tensor all follow from a bare connection with no notion of length in sight — a metric is a logically separate object, and the fundamental theorem of Riemannian geometry says that demanding compatibility with one, plus zero torsion, singles out a unique connection: the Levi-Civita connection of general relativity. In physics the coefficients $\Gamma$ play the role Newton gave to the gravitational field, and their derivatives carry the tidal gravity no free-falling observer can erase.

Key takeaways

An affine connection is the structure on a smooth manifold that says how to compare vectors in neighbouring tangent spaces, promoting the non-tensorial partial derivative to a genuine covariant derivative $\nabla$ of tensor fields.
In a coordinate basis it reduces to the connection coefficients $\Gamma^{\lambda}{}_{\mu\nu}$ , which are NOT tensor components: they transform with an inhomogeneous second-derivative term that exactly cancels the non-tensorial part of $\partial_\mu V^\nu$ .
A connection is logically independent of a metric — it defines parallel transport, geodesics, and curvature on its own; the metric is a separate input.
The antisymmetric part $\Gamma^{\lambda}{}_{[\mu\nu]}$ IS a tensor, the torsion; zero torsion plus metric compatibility uniquely fixes $\Gamma$ to the Levi-Civita form used in general relativity.
Curvature is the failure of two covariant derivatives to commute, $[\nabla_\mu,\nabla_\nu]V^\rho=R^{\rho}{}_{\sigma\mu\nu}V^\sigma$ , built purely from $\Gamma$ and its derivatives — this is where tidal gravity lives.
Key idea: a connection is a derivative operator $\nabla$ that supplies coefficients $\Gamma^{\lambda}{}_{\mu\nu}$ to correct the partial derivative, letting you differentiate tensors covariantly and parallel-transport vectors between neighbouring tangent spaces — all without ever invoking a metric.

The core insight

A connection is a derivative operator $\nabla$ that, by supplying connection coefficients $\Gamma^{\lambda}{}_{\mu\nu}$ to correct partial derivatives, lets you differentiate tensor fields covariantly and parallel-transport vectors between neighbouring tangent spaces — all without ever invoking a metric.

Where this lives today

The connection is not a historical curiosity but the working object of every modern gravitational calculation. When the LIGO and Virgo detectors match a black-hole merger against theory, the templates are numerical solutions of Einstein's equations in which the Christoffel symbols are recomputed millions of times per simulation, propagating the spacetime that carries the gravitational wave. Gravity Probe B, flying four fused-quartz gyroscopes in polar orbit from 2004 to 2005, measured the parallel transport of a spin vector directly: it confirmed the geodetic precession of 6601.8 milliarcseconds per year and the far smaller frame-dragging (Lense–Thirring) precession of 37.2 milliarcseconds per year, both computed by solving the parallel-transport equation (7) in the Earth's Kerr-like field — an experiment whose entire meaning is "what does the connection say 'no rotation' means near a spinning mass." The same mathematics, with torsion switched back on, defines Einstein–Cartan gravity, where the torsion tensor is sourced by intrinsic spin and may matter at the extreme densities of the early universe or a neutron-star core. And the recognition that $\Gamma$ is the gravitational cousin of the Yang–Mills gauge potential $A_\mu$ — both non-tensorial connections whose curvature is gauge-covariant — is why fibre-bundle language now unifies gravity with the strong and electroweak forces in the mathematical physics of today.

The physical picture

Picture an ant walking on a curved surface carrying a tiny spear and trying to keep it "pointing the same way." On a sphere there is no global notion of the same direction, so the ant does the only local thing available: at each small step it slides the spear forward without turning it relative to the surface. That step-by-step rule of no-turning is exactly the connection. Follow it around a closed loop and the spear can come back rotated — the leftover rotation is curvature — but the rule that generated the transport, applied one infinitesimal step at a time, is the connection itself. The connection is the local law; curvature is its global echo. The reason a plain partial derivative fails is that $\partial_\mu V^\nu$ blends two different things: the true change of the vector and the spurious change of the basis vectors used to describe it. The coefficients $\Gamma$ are the bookkeeping for that second effect. Nowhere is this clearer than on the perfectly flat plane written in polar coordinates: the basis vectors $\partial_r$ and $\partial_\theta$ swing around as you move, so several $\Gamma$ are nonzero even though the space has no curvature at all. This is the single most important thing to internalise about the connection, and the first figure makes it concrete with numbers.

Figure

The flat plane in polar coordinates: the basis vectors ∂_r and ∂_θ rotate from point to point, so the connection coefficients are nonzero — Γ^r_θθ = −r and Γ^θ_rθ = 1/r — even though the curvature R vanishes everywhere. Nonzero Γ does not mean curved space

Think of $\Gamma^{\lambda}{}_{\mu\nu}$ , then, as a position-dependent "fictitious force" field — the same mathematical species as the centrifugal and Coriolis terms that appear when you use rotating axes. In a freely falling, locally inertial frame it can be made to vanish at a point, exactly reproducing special relativity there; that is Einstein's equivalence principle in geometric dress. What cannot be transformed away is the derivative of $\Gamma$ , and that is where genuine gravity hides. The second figure shows the payoff of following the connection's no-turning rule all the way around a closed loop on a curved surface: the transported vector returns rotated, and the rotation angle is a hard, computable number set by the curvature enclosed.

Figure

Parallel transport of a vector around the spherical octant bounded by three geodesic arcs meeting at right angles: the connection keeps the vector unturned relative to each arc, yet it returns rotated by exactly 90° — the rotation equals the enclosed area divided by R², here π/2 steradian, a direct measure of curvature

The last thing worth fixing in mind is a freedom: connection and metric are independent inputs. You can lay many different connections on the same manifold, each defining its own geodesics and its own parallel transport. General relativity commits to one specific choice by imposing two natural conditions, but the connection concept is broader than that choice and reappears, essentially unchanged, in gauge theory and in Einstein–Cartan gravity where torsion is switched back on.

The core equations

The connection is fixed the moment you say what happens to the basis vectors. Differentiating one coordinate basis field along another must produce some vector, and its components define the coefficients — everything else in this article is a consequence of this single declaration.

\nabla_{\partial_\mu}\partial_\nu = \Gamma^{\lambda}{}_{\mu\nu}\,\partial_\lambda \qquad (1)

where: $\nabla$ — the affine connection, a covariant derivative operator; $\partial_\mu$ — coordinate basis vector fields, $\mu=0,\dots,n-1$ ; $\Gamma^{\lambda}{}_{\mu\nu}$ — the connection coefficients, $n^3$ functions of position and not tensor components. Physical reading: $\Gamma$ tells you how much of each basis direction the field $\partial_\nu$ acquires when you nudge it along $\partial_\mu$ . Applying the Leibniz rule to an arbitrary field $V=V^\nu\partial_\nu$ then turns the definition into a working formula. The result is the covariant derivative: the partial derivative plus a correction that restores the basis rotation the partial derivative silently discarded.

\nabla_\mu V^\nu = \partial_\mu V^\nu + \Gamma^{\nu}{}_{\mu\lambda}\,V^\lambda \qquad (2)

where: $V^\nu$ — components of a contravariant vector field; $\partial_\mu V^\nu$ — the ordinary partial, by itself not a tensor; $\Gamma^{\nu}{}_{\mu\lambda}V^\lambda$ — the correction restoring basis rotation. Physical reading: only the sum is a genuine $(1,1)$ tensor; each piece alone transforms badly, and their bad parts are engineered to cancel. For a one-form the sign of the correction flips. This is forced, not chosen: the covariant derivative must agree with the partial derivative on scalars, and a contracted pair $\omega_\nu V^\nu$ is a scalar.

\nabla_\mu \omega_\nu = \partial_\mu \omega_\nu - \Gamma^{\lambda}{}_{\mu\nu}\,\omega_\lambda \qquad (3)

where: $\omega_\nu$ — components of a covariant vector (one-form). Physical reading: upper indices get $+\Gamma$ , lower indices get $-\Gamma$ , and the opposite signs are exactly what makes $\nabla$ commute with contraction. The proof that $\Gamma$ is not a tensor is written into its own transformation law, which carries an extra inhomogeneous term built from second derivatives of the coordinate change.

\Gamma'^{\lambda}{}_{\mu\nu} = \frac{\partial x'^{\lambda}}{\partial x^{\rho}}\frac{\partial x^{\sigma}}{\partial x'^{\mu}}\frac{\partial x^{\tau}}{\partial x'^{\nu}}\,\Gamma^{\rho}{}_{\sigma\tau} + \frac{\partial x'^{\lambda}}{\partial x^{\rho}}\frac{\partial^2 x^{\rho}}{\partial x'^{\mu}\partial x'^{\nu}} \qquad (4)

where: $x^\mu,\,x'^\mu$ — old and new coordinate charts; the first term is the tensorial part; the second, $\dfrac{\partial x'^{\lambda}}{\partial x^{\rho}}\dfrac{\partial^2 x^{\rho}}{\partial x'^{\mu}\partial x'^{\nu}}$ , is the inhomogeneous piece that no tensor possesses. Physical reading: it is this extra term that lets $\Gamma$ be zeroed at a point by a clever choice of coordinates. The pattern of one $+\Gamma$ per upper index and one $-\Gamma$ per lower index extends to any tensor, giving the master rule that makes $\nabla$ a derivation on the whole tensor algebra.

\nabla_\mu T^{\nu_1\cdots}{}_{\rho_1\cdots} = \partial_\mu T^{\nu_1\cdots}{}_{\rho_1\cdots} + \sum_i \Gamma^{\nu_i}{}_{\mu\sigma}T^{\cdots\sigma\cdots}{}_{\rho_1\cdots} - \sum_j \Gamma^{\sigma}{}_{\mu\rho_j}T^{\nu_1\cdots}{}_{\cdots\sigma\cdots} \qquad (5)

where: $T^{\nu_1\cdots}{}_{\rho_1\cdots}$ — an arbitrary tensor field; one $+\Gamma$ term per upper index, one $-\Gamma$ term per lower index. Physical reading: differentiate the components normally, then add one correction for every slot to account for its basis rotating. Although $\Gamma$ is not a tensor, its antisymmetric part is — the inhomogeneous term is symmetric in $\mu\nu$ and drops out under antisymmetrization. What survives is the torsion.

T^{\lambda}{}_{\mu\nu} = \Gamma^{\lambda}{}_{\mu\nu} - \Gamma^{\lambda}{}_{\nu\mu} = 2\,\Gamma^{\lambda}{}_{[\mu\nu]} \qquad (6)

where: $T^{\lambda}{}_{\mu\nu}$ — the torsion, a genuine $(1,2)$ tensor; $[\mu\nu]$ — antisymmetrization over the lower pair. Physical reading: torsion measures the failure of infinitesimal parallelograms to close; general relativity sets it to zero. Given a curve, the connection also says what it means to move a vector "without turning": kill its covariant derivative along the tangent. That is parallel transport.

\frac{\mathrm{d}}{\mathrm{d}\lambda}V^\mu + \Gamma^{\mu}{}_{\nu\rho}\,\frac{\mathrm{d}x^\nu}{\mathrm{d}\lambda}\,V^\rho = 0 \qquad (7)

where: $V^\mu$ — the transported vector; $\lambda$ — curve parameter; $\mathrm{d}x^\nu/\mathrm{d}\lambda$ — tangent to the curve. Physical reading: this is the ant's no-turning rule, one differential equation per step. A geodesic is the curve that parallel-transports its own tangent vector — the "straightest possible" line the connection allows.

\frac{\mathrm{d}^2 x^\mu}{\mathrm{d}\lambda^2} + \Gamma^{\mu}{}_{\nu\rho}\,\frac{\mathrm{d}x^\nu}{\mathrm{d}\lambda}\frac{\mathrm{d}x^\rho}{\mathrm{d}\lambda} = 0 \qquad (8)

where: $\lambda$ — an affine parameter; $\Gamma^{\mu}{}_{\nu\rho}$ — the only place geometry enters. Physical reading: in general relativity this is the equation of free fall, with proper time as the affine parameter. Finally, curvature is assembled entirely from $\Gamma$ and its first derivatives, as the failure of two covariant derivatives to commute (Derivation 2).

R^{\rho}{}_{\sigma\mu\nu} = \partial_\mu \Gamma^{\rho}{}_{\nu\sigma} - \partial_\nu \Gamma^{\rho}{}_{\mu\sigma} + \Gamma^{\rho}{}_{\mu\lambda}\Gamma^{\lambda}{}_{\nu\sigma} - \Gamma^{\rho}{}_{\nu\lambda}\Gamma^{\lambda}{}_{\mu\sigma} \qquad (9)

where: $R^{\rho}{}_{\sigma\mu\nu}$ — the Riemann curvature tensor. Physical reading: it is the coordinate-independent content of the connection and the true seat of tidal gravity; unlike $\Gamma$ , it cannot be transformed away.

Derivation AThe covariant derivative as the unique tensorial correction to the partial derivative

The goal here is to prove that the connection is not an arbitrary embellishment but the forced repair of a broken object. We start from the observed fact that the partial derivative of a vector's components does not transform as a tensor — it picks up an unwanted second-derivative term — and we show that adding a term of the form $\Gamma^{\nu}{}_{\mu\lambda}V^\lambda$ supplies precisely the opposite inhomogeneous piece, cancelling it. Turned around, demanding that $\nabla_\mu V^\nu$ be a tensor forces the inhomogeneous transformation law (4) on $\Gamma$ : its non-tensorial behaviour is a requirement, not a defect.

Step 1. Differentiate the vector transformation $V'^\nu=(\partial x'^\nu/\partial x^\beta)V^\beta$ with respect to $x'^\mu$ . The product rule hits both factors, and the derivative of the Jacobian produces a second-derivative term:

\partial'_\mu V'^\nu = \frac{\partial x^\alpha}{\partial x'^\mu}\frac{\partial x'^\nu}{\partial x^\beta}\,\partial_\alpha V^\beta + \frac{\partial x^\alpha}{\partial x'^\mu}\frac{\partial^2 x'^\nu}{\partial x^\alpha \partial x^\beta}\,V^\beta

Why: the first term is what a genuine $(1,1)$ tensor would give; the second is pure contamination, and it is what we must cancel.

Step 2. Transform the correction term $\Gamma^{\nu}{}_{\mu\lambda}V^\lambda$ using the connection law (4). After contracting with $V'^\lambda$ , its tensorial part combines with Step 1's first term, while its inhomogeneous part evaluates to the exact negative of Step 1's second term:

\Gamma'^{\nu}{}_{\mu\lambda}V'^\lambda = (\text{tensor part}) - \frac{\partial x^\alpha}{\partial x'^\mu}\frac{\partial^2 x'^\nu}{\partial x^\alpha\partial x^\beta}\,V^\beta

Why: the identity $\partial'_\mu(\partial x'^\nu/\partial x^\beta)+\dots=0$ relating the two ways of writing the mixed second derivative is what flips the sign, so the two contaminations are equal and opposite.

Step 3. Add (2) together. The two second-derivative terms annihilate, leaving a purely homogeneous law:

\nabla'_\mu V'^\nu = \frac{\partial x^\alpha}{\partial x'^\mu}\frac{\partial x'^\nu}{\partial x^\beta}\,\nabla_\alpha V^\beta

Why: this is the definition of transforming as a $(1,1)$ tensor — the corrected derivative is coordinate-independent geometry, unlike either piece alone.

Step 4. Read the logic backwards. If we demand the cancellation in Step 3 for every vector field $V$ , the transformation law of $\Gamma$ is uniquely determined to be (4), inhomogeneous term and all. Why: this is why " $\Gamma$ is not a tensor" is a theorem rather than an accident — a tensorial $\Gamma$ could never repair a non-tensorial $\partial$ .

Go deeper · Derivation B · 1 minCurvature from covariant derivatives that refuse to commute

Derivation BCurvature from covariant derivatives that refuse to commute

The aim is to extract the Riemann tensor from the connection by a purely algebraic route. On flat space partial derivatives commute, $[\partial_\mu,\partial_\nu]=0$ ; on a manifold with a connection, covariant derivatives generally do not, and the amount by which they fail turns out to contain no derivatives of the vector field at all. That "no surviving derivatives" fact is what guarantees the commutator acts as a tensor — the Riemann tensor — times the vector, plus a torsion term. Computing $[\nabla_\mu,\nabla_\nu]V^\rho$ directly is the cleanest way to see both objects appear at once.

Step 1. Take one covariant derivative, treating the result as a $(1,1)$ tensor with a free upper $\rho$ and a free lower $\nu$ :

\nabla_\nu V^\rho = \partial_\nu V^\rho + \Gamma^{\rho}{}_{\nu\sigma}V^\sigma

Why: to differentiate again we must use the full tensor rule (5), which needs this object correctly identified as $(1,1)$ .

Step 2. Apply (5) once more, adding a $+\Gamma$ for the upper index and a $-\Gamma$ for the lower one:

\nabla_\mu \nabla_\nu V^\rho = \partial_\mu(\nabla_\nu V^\rho) + \Gamma^{\rho}{}_{\mu\lambda}\nabla_\nu V^\lambda - \Gamma^{\lambda}{}_{\mu\nu}\nabla_\lambda V^\rho

Why: the last term, carrying the lower-index correction $-\Gamma^{\lambda}{}_{\mu\nu}$ , is the one that survives antisymmetrization as torsion.

Step 3. Subtract the same expression with $\mu\leftrightarrow\nu$ . Every term containing a derivative of $V$ either cancels or collects into $-(\Gamma^{\lambda}{}_{\mu\nu}-\Gamma^{\lambda}{}_{\nu\mu})\nabla_\lambda V^\rho$ , while the $\partial\Gamma$ and $\Gamma\Gamma$ pieces assemble into the coefficient of $V^\sigma$ :

[\nabla_\mu,\nabla_\nu]V^\rho = \big(\partial_\mu\Gamma^{\rho}{}_{\nu\sigma} - \partial_\nu\Gamma^{\rho}{}_{\mu\sigma} + \Gamma^{\rho}{}_{\mu\lambda}\Gamma^{\lambda}{}_{\nu\sigma} - \Gamma^{\rho}{}_{\nu\lambda}\Gamma^{\lambda}{}_{\mu\sigma}\big)V^\sigma - T^{\lambda}{}_{\mu\nu}\nabla_\lambda V^\rho

Why: the bracket is precisely $R^{\rho}{}_{\sigma\mu\nu}$ from (9), and $T^{\lambda}{}_{\mu\nu}$ is the torsion from (6).

Step 4. For a torsion-free connection the last term drops, leaving the Ricci identity $[\nabla_\mu,\nabla_\nu]V^\rho=R^{\rho}{}_{\sigma\mu\nu}V^\sigma$ . Why: with no derivative of $V$ on the right, the map $V\mapsto[\nabla_\mu,\nabla_\nu]V$ is algebraic, which is exactly the statement that $R$ is a tensor. *The absence of $\partial V$ terms on the right-hand side confirms that $R$ is a genuine tensor.*

Go deeper · Derivation C · 1 minThe fundamental theorem: torsion-free plus metric-compatible gives a unique connection

Derivation CThe fundamental theorem: torsion-free plus metric-compatible gives a unique connection

This derivation delivers the single most important connection in physics. So far a manifold admits infinitely many connections; general relativity needs exactly one, tied to the metric that measures spacetime intervals. We impose two conditions — vanishing torsion, so $\Gamma$ is symmetric in its lower indices, and metric compatibility $\nabla_\rho g_{\mu\nu}=0$ , so parallel transport preserves lengths and angles — and show they determine $\Gamma$ uniquely by the classic permute-and-add manoeuvre, yielding the Levi-Civita connection whose coefficients are the Christoffel symbols.

Step 1. Write metric compatibility out with the tensor rule (5) and lower an index by defining $\Gamma_{\nu\rho\mu}\equiv g_{\nu\lambda}\Gamma^{\lambda}{}_{\rho\mu}$ :

\nabla_\rho g_{\mu\nu} = \partial_\rho g_{\mu\nu} - \Gamma_{\mu\rho\nu} - \Gamma_{\nu\rho\mu} = 0

Why: metric compatibility is the physical demand that transporting two vectors does not change their inner product, so norms and angles ride along unchanged.

Step 2. Write the same statement three times, cyclically permuting $(\rho,\mu,\nu)$ , to obtain three equations relating $\partial g$ to sums of two lowered-index $\Gamma$ 's:

\partial_\rho g_{\mu\nu}=\Gamma_{\mu\rho\nu}+\Gamma_{\nu\rho\mu},\quad \partial_\mu g_{\nu\rho}=\Gamma_{\nu\mu\rho}+\Gamma_{\rho\mu\nu},\quad \partial_\nu g_{\rho\mu}=\Gamma_{\rho\nu\mu}+\Gamma_{\mu\nu\rho}

Why: three permutations give three equations for the six independent lowered coefficients — enough to solve once symmetry is used.

Step 3. Impose zero torsion, $\Gamma^{\lambda}{}_{\mu\nu}=\Gamma^{\lambda}{}_{\nu\mu}$ , i.e. symmetry in the last two lowered indices. Add the second and third equations and subtract the first; four of the six terms cancel pairwise:

\partial_\mu g_{\nu\rho}+\partial_\nu g_{\rho\mu}-\partial_\rho g_{\mu\nu} = 2\,\Gamma_{\rho\mu\nu}

Why: without the torsion-free symmetry the cancellation fails and the connection is not determined — this is where the second condition earns its keep.

Step 4. Raise the free index with $g^{\lambda\rho}$ to isolate $\Gamma$ :

\Gamma^{\lambda}{}_{\mu\nu} = \tfrac{1}{2}\,g^{\lambda\rho}\big(\partial_\mu g_{\rho\nu} + \partial_\nu g_{\rho\mu} - \partial_\rho g_{\mu\nu}\big)

Why: the right side is built entirely from the metric and its first derivatives, so the solution exists, is explicit, and is unique — the fundamental theorem of (pseudo-)Riemannian geometry. *The manifest symmetry of the result in $\mu\nu$ is consistent with the torsion-free assumption used to derive it.*

Limiting cases

Structural reductions of a general affine connection, from the most general down to the unique connection of general relativity:

Case	Result	Meaning
Symmetric (torsion-free)	$\Gamma^{\lambda}{}_{\mu\nu} = \Gamma^{\lambda}{}_{\nu\mu}$	The torsion vanishes and infinitesimal parallelograms close; the standard assumption in general relativity.
Metric-compatible	$\nabla_\rho g_{\mu\nu} = 0$	Parallel transport preserves inner products, so lengths and angles are conserved along every curve.
Levi-Civita (both)	$\Gamma^{\lambda}{}_{\mu\nu} = \tfrac{1}{2}g^{\lambda\sigma}(\partial_\mu g_{\sigma\nu}+\partial_\nu g_{\sigma\mu}-\partial_\sigma g_{\mu\nu})$	Uniquely fixed by torsion-free plus metric-compatible; the Christoffel symbols of general relativity.
Flat connection	$R^{\rho}{}_{\sigma\mu\nu} = 0$	Coordinates exist in which $\Gamma=0$ everywhere; parallel transport is path-independent. Minkowski and Euclidean space.
Locally inertial at $p$	$\Gamma^{\lambda}{}_{\mu\nu}\big\vert_p = 0,\ \ \partial_\sigma\Gamma^{\lambda}{}_{\mu\nu}\big\vert_p \neq 0$	$\Gamma$ can be killed at one point but not its derivatives — the geometric content of the equivalence principle.

Common trapIf the Christoffel symbols vanish at a point the space must be flat there

This is the single most common error with connections, and the polar-coordinate figure above already refutes half of it: the flat plane has nonzero $\Gamma$ , so $\Gamma\neq0$ certainly does not imply curvature. The converse trap is subtler. At any single point $p$ you can choose coordinates — Riemann normal coordinates, the mathematical form of a freely falling frame — in which every $\Gamma^{\lambda}{}_{\mu\nu}$ vanishes, no matter how curved the space is. So $\Gamma=0$ at a point tells you nothing about flatness. Flatness is the statement that the tensor $R^{\rho}{}_{\sigma\mu\nu}$ , built from the derivatives of $\Gamma$ , vanishes — and those derivatives generally cannot all be killed at once. This is exactly why gravity is real even though a falling observer feels weightless: the field $\Gamma$ transforms away locally, but its tidal derivative does not.

Worked examples

1 · Christoffel symbols of the flat plane in polar coordinates (fully numeric)

The Euclidean plane in polar coordinates has metric $ds^2=dr^2+r^2\,d\theta^2$ , so $g_{rr}=1$ , $g_{\theta\theta}=r^2$ , $g^{rr}=1$ , $g^{\theta\theta}=1/r^2$ , and all off-diagonal components vanish. Compute every Christoffel symbol, evaluate them at $r=2$ , and confirm the space is flat.

Step 1. Use the Levi-Civita formula from Derivation 3. Only $g_{\theta\theta}=r^2$ has a nonzero derivative, $\partial_r g_{\theta\theta}=2r$ ; every other $\partial g$ is zero. This immediately kills most of the $n^3=8$ candidate symbols.

Step 2. The surviving radial symbol: $\Gamma^{r}{}_{\theta\theta}=\tfrac12 g^{rr}(2\partial_\theta g_{r\theta}-\partial_r g_{\theta\theta})=\tfrac12(1)(-2r)=\mathbf{-r}$ . The angular symbols: $\Gamma^{\theta}{}_{r\theta}=\Gamma^{\theta}{}_{\theta r}=\tfrac12 g^{\theta\theta}\partial_r g_{\theta\theta}=\tfrac12\cdot\tfrac{1}{r^2}\cdot 2r=\mathbf{1/r}$ . All other components vanish.

Step 3. Evaluate at $r=2$ : $\Gamma^{r}{}_{\theta\theta}=\mathbf{-2}$ and $\Gamma^{\theta}{}_{r\theta}=\mathbf{0.5}$ . These are honestly nonzero, yet the plane is flat — the coefficients are recording the swing of the polar basis vectors, nothing more.

Step 4. Confirm flatness with (9). The one candidate independent component is $R^{r}{}_{\theta r\theta}=\partial_r\Gamma^{r}{}_{\theta\theta}-\partial_\theta\Gamma^{r}{}_{r\theta}+\Gamma^{r}{}_{r\lambda}\Gamma^{\lambda}{}_{\theta\theta}-\Gamma^{r}{}_{\theta\lambda}\Gamma^{\lambda}{}_{r\theta}$ . Substituting: $\partial_r(-r)-0+0-\Gamma^{r}{}_{\theta\theta}\Gamma^{\theta}{}_{r\theta}=-1-(-r)(1/r)=-1+1=\mathbf{0}$ . The Riemann tensor vanishes: the plane is flat despite nonzero $\Gamma$ . Radial lines ( $\theta$ fixed, $r=\lambda$ ) also satisfy the geodesic equation (8) trivially, since every term contains a $\Gamma$ paired with the vanishing $\dot\theta$ — straight lines through the origin are geodesics, as they must be.

2 · Holonomy around the spherical octant (fully numeric)

On the unit 2-sphere, parallel-transport a vector once around the closed geodesic triangle bounded by the equator and two meridians 90° apart — the octant of the second figure. By how much does it rotate, and how does that pin down the curvature?

Step 1. The triangle has three vertices where geodesics meet at right angles: the north pole (where the two meridians cross at 90°) and the two equatorial corners (where each meridian meets the equator at 90°). Its interior angles sum to $\tfrac{\pi}{2}+\tfrac{\pi}{2}+\tfrac{\pi}{2}=\tfrac{3\pi}{2}$ .

Step 2. For a geodesic triangle the parallel-transport holonomy — the angle by which a transported vector fails to return to itself — equals the angular excess, the amount the interior angles beat the flat-space value $\pi$ : $\Delta\phi=(\text{sum of angles})-\pi=\tfrac{3\pi}{2}-\pi=\tfrac{\pi}{2}=\mathbf{90°}$ .

Step 3. Check it against area. The octant is one eighth of the sphere's total area $4\pi R^2$ , so with $R=1$ its area is $\tfrac{1}{8}(4\pi)=\tfrac{\pi}{2}$ . The Gauss–Bonnet relation gives holonomy $=\displaystyle\int K\,dA=K\cdot\text{Area}$ , and the unit sphere has constant Gaussian curvature $K=1/R^2=1$ , so holonomy $=1\cdot\tfrac{\pi}{2}=\tfrac{\pi}{2}$ . Both routes agree: the vector returns rotated by exactly $\pi/2=90°$ .

Step 4. Interpret. The connection's no-turning rule was followed faithfully along every edge, yet the vector came back rotated — that rotation is a coordinate-free number reading out the enclosed curvature. Shrink the loop to a point and the holonomy scales with the enclosed area, recovering the infinitesimal statement $[\nabla_\mu,\nabla_\nu]V^\rho=R^{\rho}{}_{\sigma\mu\nu}V^\sigma$ of Derivation 2. On a flat surface the same construction gives 0° — holonomy is the operational signature of curvature.

Physics of the portrait

The affine connection is a young idea wearing an old name. The coefficients had been written down for metric geometry decades before anyone realised the connection was a structure in its own right — and the person who cut it loose from the metric was Hermann Weyl.

Origins · 1 minHermann Weyl and how this was discovered

Hermann Weyl

1885–1955 · the mathematician who abstracted the affine connection from the metric

Hermann Weyl was among the last universalists — a student of Hilbert at Göttingen who moved with equal authority through analysis, number theory, the foundations of mathematics, quantum mechanics, and relativity. His 1918 masterwork Raum-Zeit-Materie (Space-Time-Matter) and the same year's Reine Infinitesimalgeometrie did something Einstein's 1915 theory had not: they identified the connection — Weyl's affiner Zusammenhang, literally "affine connection" — as a geometric object logically prior to and independent of the metric. Where Levi-Civita's 1917 parallel transport had lived inside a metric world, Weyl saw that the rule for comparing vectors at neighbouring points could be posited on its own, with the metric added later as separate data. This is the exact conceptual content of this article, and it opened the door to the modern view of gravitation as a gauge theory. Weyl went further, proposing in 1918 a scale-gauge geometry in which lengths too could change under transport — a unified theory of gravity and electromagnetism that Einstein famously rebutted (atomic clocks would drift), but whose "gauge" idea, transplanted from real scale to complex phase, became the backbone of all of modern particle physics. A committed intuitionist and a graceful writer, Weyl left Göttingen for Princeton's Institute for Advanced Study in 1933; the affine connection, now standard in every relativity text, is his quiet monument.

History

1869

— Christoffel introduces the symbols now bearing his name while classifying quadratic differential forms

1900

— Ricci-Curbastro and Levi-Civita codify the absolute differential calculus, the tensor calculus Einstein would need

1916

— Einstein's general relativity uses the Christoffel symbols as the gravitational field without naming the connection

1917

— Levi-Civita defines parallel transport, giving the Christoffel symbols an intrinsic geometric meaning

1918

— Weyl abstracts the affine connection as a structure independent of the metric and coins the term

1923

— Cartan builds the general theory of affine connections, restoring torsion as an independent geometric datum

Going further

Texts. Misner, Thorne & Wheeler, Gravitation, §10 (covariant differentiation and connection coefficients) · Wald, General Relativity, §3.1 (Derivative Operators) · Carroll, Spacetime and Geometry, §3.2–3.3 (The Connection) · Schutz, A First Course in General Relativity, §5 (Preface to Curvature) · Hartle, Gravity, §8 (Geodesics) · Nakahara, Geometry, Topology and Physics, §7.2 (Connections on fibre bundles). Key checkpoints in the derivations: the cancellation of inhomogeneous terms in Derivation 1; the Ricci-identity commutator in Derivation 2 leaving no $\partial V$ terms; the Levi-Civita formula from Derivation 3 and its manifest $\mu\nu$ symmetry; the polar Christoffel symbols $\Gamma^{r}{}_{\theta\theta}=-r$ , $\Gamma^{\theta}{}_{r\theta}=1/r$ with $R^{r}{}_{\theta r\theta}=0$ ; and the spherical-octant holonomy of $\pi/2$ consistent with Gauss–Bonnet. The Gravity Probe B figures (geodetic $6601.8$ mas/yr, frame-dragging $37.2$ mas/yr) are the published 2011 results. Sign conventions follow the mostly-plus signature with $R^{\rho}{}_{\sigma\mu\nu}=\partial_\mu\Gamma^{\rho}{}_{\nu\sigma}-\dots$

Builds on: Manifolds and Tangent Spaces · Tensor Fields · Vector Fields and the Lie Bracket. Leads to: Covariant Derivative · Parallel Transport · Christoffel Symbols · Torsion Tensor · Metric Compatibility · The Levi-Civita Connection · The Geodesic Equation · Geodesic Deviation Equation.

Ask the tutor: Ask me to parallel-transport a vector around any loop you name — a polar square, a spherical cap, a small coordinate parallelogram — and I will show you exactly how much it rotates and why.

Why can't I just use the partial derivative of a vector field

Because the partial derivative of vector components is not a tensor. Under a coordinate change it picks up an extra term built from second derivatives of the transformation, so two observers computing $\partial_\mu V^\nu$ disagree in a way that has no geometric meaning. The trouble is that $\partial_\mu V^\nu$ secretly mixes the true change of the vector with the change of the basis vectors you happened to use. The connection term $\Gamma^{\nu}{}_{\mu\lambda}V^\lambda$ is engineered to carry exactly the opposite spurious term, so the sum $\nabla_\mu V^\nu$ transforms cleanly as a $(1,1)$ tensor. Only then does "the rate of change of the vector field" mean something both observers agree on.

Is the connection the same thing as the metric

No, and keeping them separate is the whole point of this article. A connection is a rule for parallel transport and differentiation; a metric is a rule for lengths and angles. You can put a connection on a manifold that has no metric at all, and you can put many different connections on a manifold that does have one. General relativity happens to use a very special connection — the Levi-Civita connection — that is fixed by the metric once you also demand zero torsion and metric compatibility (the fundamental theorem). But that is a choice general relativity makes, not a logical necessity. Gauge theories and Einstein–Cartan gravity use connections in settings where this metric link is loosened or absent.

What is torsion, and why does general relativity set it to zero

Torsion is the antisymmetric part of the connection, $T^{\lambda}{}_{\mu\nu}=\Gamma^{\lambda}{}_{\mu\nu}-\Gamma^{\lambda}{}_{\nu\mu}$ , and despite $\Gamma$ not being a tensor, torsion is — the non-tensorial part of $\Gamma$ is symmetric in the lower indices and cancels in the difference. Geometrically it measures whether infinitesimal parallelograms built by parallel transport fail to close. General relativity sets it to zero for two reasons: it is the simplest natural assumption, and it makes the connection uniquely determined by the metric. Theories that keep torsion, notably Einstein–Cartan gravity, source it with the intrinsic spin of matter, so torsion becomes physically relevant only at densities where spin alignment matters.

How can the flat plane have nonzero connection coefficients

Because the connection coefficients describe the coordinate basis vectors, not the intrinsic curvature. In polar coordinates the basis vectors $\partial_r$ and $\partial_\theta$ point in different directions at different places — $\partial_\theta$ even changes length as $r$ grows — so differentiating a constant vector field produces nonzero results that $\Gamma$ must account for. That is why $\Gamma^{r}{}_{\theta\theta}=-r$ and $\Gamma^{\theta}{}_{r\theta}=1/r$ are nonzero on the perfectly flat plane. Curvature is a statement about the derivatives of $\Gamma$ , combined into the Riemann tensor; for the plane those combine to exactly zero. Nonzero $\Gamma$ signals curved coordinates; nonzero $R$ signals curved space.

Where does curvature come from in terms of the connection

Curvature is the failure of two covariant derivatives to commute. On flat space with Cartesian coordinates $\partial_\mu$ and $\partial_\nu$ commute; once you covariantize, $[\nabla_\mu,\nabla_\nu]V^\rho$ is generally nonzero, and the remarkable fact is that it contains no derivatives of $V$ — it is purely $R^{\rho}{}_{\sigma\mu\nu}V^\sigma$ (plus a torsion term if torsion is present). Because no derivatives of $V$ survive, the operation is algebraic in $V$ , which is exactly the statement that $R$ is a tensor. Concretely $R^{\rho}{}_{\sigma\mu\nu}=\partial\Gamma-\partial\Gamma+\Gamma\Gamma-\Gamma\Gamma$ : curvature is the second-order, coordinate-independent content of the connection, and it is what parallel transport around a small loop reads out as a rotation.

What is the physical meaning of the connection in general relativity

The connection coefficients play the role that Newtonian physics assigned to the gravitational force field. A freely falling particle moves on a geodesic, $\ddot x^\mu+\Gamma^{\mu}{}_{\nu\rho}\dot x^\nu\dot x^\rho=0$ , so $\Gamma$ is what "bends" trajectories — it is the gravitational field. But unlike a Newtonian field, $\Gamma$ can be transformed to zero at any point by choosing a freely falling frame, which is the geometric statement of the equivalence principle: locally, gravity disappears. What cannot be removed is the derivative of $\Gamma$ , packaged as curvature, which drives the geodesic deviation between nearby free-fallers — the tidal forces that are the true, frame-independent signature of gravity.