The Covariant Derivative

Differentiation is comparison: to know how a field changes you must compare its value at a point with its value an instant away. For a scalar this is harmless — the partial derivative $\partial_\mu\phi$ is already a well-behaved covector. For a vector field it is a genuine problem. A vector $V=V^\nu e_\nu$ carries components $V^\nu$ and basis vectors $e_\nu$ , and in curvilinear coordinates or on a curved manifold those basis vectors point in different directions at different points. Two nearby vectors live in two different tangent spaces, and there is no coordinate-free way to subtract them until you say how the basis itself turns. Differentiate only the components and you throw away the turning of the frame; the result $\partial_\mu V^\nu$ picks up an inhomogeneous, non-tensorial piece and stops being geometric. The covariant derivative $\nabla_\mu$ is the repair. It differentiates the whole object $V^\nu e_\nu$ , packages the derivatives of the basis into connection coefficients — the Christoffel symbols $\Gamma^\lambda{}_{\mu\nu}$ defined by $\partial_\mu e_\nu=\Gamma^\lambda{}_{\mu\nu}e_\lambda$ — and returns a true tensor. On a manifold with a metric there is one canonical connection, the Levi-Civita connection, singled out by being torsion-free and metric-compatible, which fixes the $\Gamma$ 's uniquely from the metric. This single construction defines parallel transport and geodesics, produces the Riemann tensor through the failure of $\nabla$ to commute with itself, and reappears verbatim in gauge theory as the gauge covariant derivative — the bridge from algebra at a point to calculus across a manifold.

Key takeaways

The partial derivative $\partial_\mu V^\nu$ of a vector field is not a tensor: under a coordinate change it picks up a second-derivative term because the basis vectors themselves vary from point to point.
The covariant derivative repairs this with a connection term, $\nabla_\mu V^\nu=\partial_\mu V^\nu+\Gamma^\nu{}_{\mu\lambda}V^\lambda$ , whose inhomogeneous transformation cancels the offending piece, leaving a $(1,1)$ tensor.
Lower indices carry a minus sign and every index gets its own $\Gamma$ : $\nabla_\mu V_\nu=\partial_\mu V_\nu-\Gamma^\lambda{}_{\mu\nu}V_\lambda$ .
On a metric manifold the Levi-Civita connection is the unique torsion-free, metric-compatible one ( $\nabla_\mu g_{\nu\rho}=0$ ), fixing $\Gamma^\lambda{}_{\mu\nu}=\tfrac12 g^{\lambda\sigma}(\partial_\mu g_{\sigma\nu}+\partial_\nu g_{\sigma\mu}-\partial_\sigma g_{\mu\nu})$ .
Nonzero $\Gamma$ does not mean curvature: plane polar coordinates give $\Gamma^r{}_{\theta\theta}=-r$ in flat space; only the commutator $[\nabla_\mu,\nabla_\nu]$ — the Riemann tensor — detects real curvature.
Geometrically $\nabla$ is parallel transport: around a spherical octant a transported vector returns rotated by exactly 90°, the enclosed solid angle.
Key idea: to differentiate a tensor covariantly, differentiate each component the ordinary way, then add (for each upper index) and subtract (for each lower index) a Christoffel term for the position-dependence of the basis; the connection term exactly cancels the non-tensorial part of the partial derivative.

The core insight

To differentiate a tensor field covariantly, differentiate each component the ordinary way, then add (for each upper index) and subtract (for each lower index) a Christoffel-symbol term that accounts for the position-dependence of the basis. The connection term exactly cancels the non-tensorial part of the partial derivative, so the output is again a tensor.

Where this lives today

The covariant derivative is the operating system of modern gravitational physics. Numerical-relativity codes that simulate binary black-hole and neutron-star mergers — the templates matched against LIGO–Virgo–KAGRA detections since GW150914 in 2015 — evolve Einstein's equations in formulations (BSSN, Z4c) whose every spatial derivative is a covariant derivative built from a dynamically updated connection; a sign error in a single $\Gamma$ term wrecks the constraint evolution. Event Horizon Telescope images of the M87 and Sgr A\* shadows are matched against null geodesics in the Kerr metric, and the GPS constellation corrects its clocks via geodesic motion and $\nabla_\mu T^{\mu\nu}=0$ . Beyond gravity the identical construction runs the Standard Model: the gauge covariant derivative $D_\mu=\partial_\mu-igA_\mu$ is a connection on an internal fibre bundle, and lattice-QCD codes transport quark fields with discrete "link variables" — the direct descendants of Levi-Civita's 1917 idea.

The physical picture

Walk across the surface of a sphere carrying an arrow, at every step keeping it "as parallel as possible" — never actively twisting it, only sliding it along. This is parallel transport, the operational meaning of "no covariant change," $\nabla_{\dot\gamma}V=0$ . The surprise is that after a closed loop the arrow returns rotated — and the rotation is not sloppiness but the curvature of the surface, read off directly. The covariant derivative is the instrument that separates the arrow's genuine turning from the apparent turning forced on it by a coordinate grid that twists underfoot.

Figure

Parallel transport of a vector around a spherical octant — three right-angle geodesic corners. The vector slides without local twisting yet returns rotated by exactly 90°, equal to the enclosed area π/2 of the unit sphere, which is the integrated Gaussian curvature ∫K dA.

In flat space with Cartesian axes the basis vectors are the same everywhere, all $\Gamma$ vanish, and $\nabla_\mu$ collapses to $\partial_\mu$ — nothing turns, so no correction is needed. Switch to plane polar coordinates and the story changes even though the space stays perfectly flat: the basis $\hat r,\hat\theta$ swings around as you circle the origin, so $\partial_\theta\hat r=\hat\theta$ and the connection springs to life with $\Gamma^\theta{}_{r\theta}=1/r$ and $\Gamma^r{}_{\theta\theta}=-r$ . This is the lesson of the second figure: a nonzero connection can be a pure coordinate artefact, and the $\Gamma V$ term is exactly the bookkeeping that subtracts this apparent, grid-induced change so $\nabla$ measures only the genuine, basis-independent change of the object.

Figure

Plane polar basis vectors r̂ (blue, radial) and θ̂ (amber, tangential) at eight points on a circle. The frame rotates from point to point in flat space, so ∂θ r̂ = θ̂ and the connection is nonzero: Γθ_rθ = 1/r and Γr_θθ = −r, even though the Riemann curvature is identically zero.

The core equations

The whole subject follows from one demand: the derivative of a tensor must again be a tensor. For a contravariant vector, differentiate the full object $V=V^\lambda e_\lambda$ rather than its bare components, and the derivative of the basis supplies the connection term.

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

where: $\nabla_\mu$ — covariant derivative along $x^\mu$ ; $V^\nu$ — contravariant components; $\Gamma^\nu{}_{\mu\lambda}$ — Christoffel (connection) coefficients; $\partial_\mu=\partial/\partial x^\mu$ . Reading: the first term is the change in the numbers, the second corrects for the turning of the basis; together they form a $(1,1)$ tensor. For a covariant (lower-index) vector the correction reverses sign, forced by demanding that $\nabla$ commute with contraction so that $\nabla_\mu(V_\nu W^\nu)$ is the ordinary gradient of the scalar $V_\nu W^\nu$ .

\nabla_\mu V_\nu = \partial_\mu V_\nu - \Gamma^\lambda{}_{\mu\nu} V_\lambda \qquad (2)

where: $V_\nu$ — covariant components; the minus sign is the signature of a lower index. Reading: each index contributes its own $\Gamma$ — plus for up, minus for down. The general mixed rule is just this pattern applied index by index: one $+\Gamma$ per upper slot, one $-\Gamma$ per lower slot, all riding a single $\partial_\mu$ .

\nabla_\mu T^{\nu}{}_{\rho} = \partial_\mu T^{\nu}{}_{\rho} + \Gamma^{\nu}{}_{\mu\lambda}T^{\lambda}{}_{\rho} - \Gamma^{\lambda}{}_{\mu\rho}T^{\nu}{}_{\lambda} \qquad (3)

where: $T^{\nu}{}_{\rho}$ — components of a $(1,1)$ tensor field. Reading: for $p$ upper and $q$ lower indices there are $p+q+1$ terms; the rule never gets harder than one $\Gamma$ per index. On a manifold with a metric the connection is not free: requiring it to be symmetric (torsion-free) and to annihilate the metric ( $\nabla_\mu g_{\nu\rho}=0$ ) fixes the Christoffel symbols uniquely in terms of the metric and its first derivatives.

\Gamma^\lambda{}_{\mu\nu} = \tfrac{1}{2}\, g^{\lambda\sigma}\!\left(\partial_\mu g_{\sigma\nu} + \partial_\nu g_{\sigma\mu} - \partial_\sigma g_{\mu\nu}\right) \qquad (4)

where: $g_{\mu\nu}$ — metric components; $g^{\lambda\sigma}$ — inverse metric, $g^{\lambda\sigma}g_{\sigma\nu}=\delta^\lambda{}_\nu$ . Reading: this is the Levi-Civita connection — what makes general relativity a metric theory with no extra structure to postulate. The connection is not a tensor — under a coordinate change it picks up an inhomogeneous second-derivative term, and this is a feature, not a bug: that term is precisely what cancels the non-tensorial part of $\partial_\mu V^\nu$ .

\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 (5)

where: the first term is the tensorial part (three Jacobians); the second, carrying a second derivative of the coordinate map, is the inhomogeneous offender that spoils tensoriality of $\Gamma$ alone. Finally, the failure of two covariant derivatives to commute is not zero — it is curvature, packaged algebraically into the Riemann tensor.

[\nabla_\mu,\nabla_\nu]V^\rho = R^\rho{}_{\sigma\mu\nu}\,V^\sigma \qquad (6)

where: $R^\rho{}_{\sigma\mu\nu}=\partial_\mu\Gamma^\rho{}_{\nu\sigma}-\partial_\nu\Gamma^\rho{}_{\mu\sigma}+\Gamma^\rho{}_{\mu\lambda}\Gamma^\lambda{}_{\nu\sigma}-\Gamma^\rho{}_{\nu\lambda}\Gamma^\lambda{}_{\mu\sigma}$ — Riemann curvature tensor. Reading: the right side multiplies $V^\sigma$ with no derivative of $V$ , so curvature is a genuine tensor measuring the path-dependence of parallel transport. A frequently-used corollary collapses the covariant divergence to a single ordinary derivative, using the contracted identity $\Gamma^\mu{}_{\mu\nu}=\partial_\nu\ln\sqrt{|g|}$ :

\nabla_\mu V^\mu = \frac{1}{\sqrt{|g|}}\,\partial_\mu\!\left(\sqrt{|g|}\,V^\mu\right) \qquad (7)

where: $g=\det g_{\mu\nu}$ ; $\sqrt{|g|}$ — invariant volume element. Reading: this one line reproduces the textbook divergence in cylindrical and spherical coordinates with no unit-vector bookkeeping.

Derivation AWhy the partial derivative is not a tensor

The goal is to show explicitly, from the transformation law alone, that the ordinary partial derivative of a vector's components fails to transform as a tensor, and to isolate the exact term responsible. This pins down what the connection must do: its inhomogeneous transformation is reverse-engineered to cancel precisely this leftover, which is why $\nabla$ and $\Gamma$ come as a matched pair rather than independent objects.

Step 1. Start from the contravariant transformation law $V'^\nu=\dfrac{\partial x'^\nu}{\partial x^\beta}V^\beta$ . Why: this is the defining property of a contravariant vector, and any test of tensoriality must begin from how the object is declared to transform.

Step 2. Differentiate with respect to the primed coordinate using the chain rule $\partial'_\mu=\dfrac{\partial x^\alpha}{\partial x'^\mu}\partial_\alpha$ , giving $\partial'_\mu V'^\nu=\dfrac{\partial x^\alpha}{\partial x'^\mu}\partial_\alpha\!\left(\dfrac{\partial x'^\nu}{\partial x^\beta}V^\beta\right)$ . Why: we want the derivative in the new frame, and the operator itself transforms with a Jacobian, so it must be carried inside consistently.

Step 3. Apply the product rule to the bracket. One factor, $\partial x'^\nu/\partial x^\beta$ , depends on position, so it is differentiated too:

\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 exactly the tensor transformation we hoped for; the second, containing a second derivative of the coordinate map, is pure contamination.

Step 4. Diagnose the offending term. It vanishes only for linear (affine) coordinate changes; for any curvilinear chart or curved manifold it survives. Comparing with equation (5), it is the negative of the inhomogeneous piece of the $\Gamma$ transformation, so adding $\Gamma^\nu{}_{\mu\lambda}V^\lambda$ cancels it identically and restores a tensor. Why: this closes the logic — the connection is defined by the requirement that this exact term disappear.

Go deeper · Derivation B · 1 minThe Levi-Civita connection from the metric

Derivation BThe Levi-Civita connection from the metric

The goal is to derive the closed-form expression (4) for the Christoffel symbols by imposing the two conditions that define the Levi-Civita connection — vanishing torsion (symmetry of $\Gamma$ in its lower indices) and metric compatibility ( $\nabla g=0$ ) — and to see why they leave no freedom. The payoff is that on any Riemannian manifold the derivative is completely determined by the metric, which is what makes gravity geometry.

Step 1. Write out metric compatibility with rule (3) applied to the two lower indices of $g_{\nu\rho}$ :

\partial_\mu g_{\nu\rho} = \Gamma^\sigma{}_{\mu\nu}g_{\sigma\rho} + \Gamma^\sigma{}_{\mu\rho}g_{\nu\sigma}

Why: setting $\nabla_\mu g_{\nu\rho}=0$ trades the unknown $\Gamma$ 's for known derivatives of the metric, giving us equations to solve.

Step 2. Form the cyclic combination — this expression with indices $(\mu\nu\rho)$ , plus the same with $(\nu\rho\mu)$ , minus the one with $(\rho\mu\nu)$ . Using the symmetry $\Gamma^\sigma{}_{\mu\nu}=\Gamma^\sigma{}_{\nu\mu}$ , four of the six $\Gamma g$ terms cancel in pairs and the remaining two coincide:

\partial_\mu g_{\nu\rho} + \partial_\nu g_{\rho\mu} - \partial_\rho g_{\mu\nu} = 2\,\Gamma^\sigma{}_{\mu\nu}g_{\sigma\rho}

Why: the cyclic-sum trick is engineered so that torsion-freeness collapses six terms to a single clean one, which can then be solved.

Step 3. Free the upper index by contracting with the inverse metric $g^{\lambda\rho}$ , using $g^{\lambda\rho}g_{\sigma\rho}=\delta^\lambda{}_\sigma$ :

\Gamma^\lambda{}_{\mu\nu} = \tfrac12 g^{\lambda\rho}\!\left(\partial_\mu g_{\nu\rho} + \partial_\nu g_{\rho\mu} - \partial_\rho g_{\mu\nu}\right)

Why: this is equation (4); the two conditions have determined every component of the connection with no residual choice, proving the Levi-Civita connection is unique. *The result is symmetric in its lower indices, $\Gamma^\lambda{}_{\mu\nu}=\Gamma^\lambda{}_{\nu\mu}$ , confirming that the Levi-Civita connection is torsion-free.*

Go deeper · Derivation C · 1 minThe commutator of covariant derivatives is curvature

Derivation CThe commutator of covariant derivatives is curvature

The goal is to compute $[\nabla_\mu,\nabla_\nu]V^\rho$ directly and show that, although each $\nabla$ is a differential operator, their commutator is purely algebraic in $V$ — every derivative-of- $V$ term cancels, leaving a coefficient multiplying $V^\sigma$ undifferentiated. That coefficient is the Riemann tensor, and this is the cleanest proof that curvature is a genuine tensor field measuring the path-dependence of parallel transport.

Step 1. Expand $\nabla_\nu V^\rho=\partial_\nu V^\rho+\Gamma^\rho{}_{\nu\sigma}V^\sigma$ and then apply $\nabla_\mu$ to the result, treating $\nabla_\nu V^\rho$ as a $(1,1)$ tensor so it collects a $+\Gamma$ for its upper index $\rho$ and a $-\Gamma$ for its lower index $\nu$ . Why: the second derivative must be taken covariantly on the whole tensor, not naively, or the index structure is wrong.

Step 2. Antisymmetrize in $\mu\leftrightarrow\nu$ to form the commutator. The terms $\partial_\mu\partial_\nu V^\rho$ are symmetric and cancel; the terms $\Gamma^\rho{}_{\nu\sigma}\partial_\mu V^\sigma$ pair off against their $\mu\leftrightarrow\nu$ partners and cancel; and the $-\Gamma^\lambda{}_{\mu\nu}(\cdots)$ terms cancel because $\Gamma^\lambda{}_{\mu\nu}$ is symmetric (torsion-free). Why: this is where torsion-freeness earns its keep — it is exactly what removes the surviving derivative-of- $V$ pieces so the result is algebraic.

Step 3. Collect what remains, all of it proportional to $V^\sigma$ :

[\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 \equiv R^\rho{}_{\sigma\mu\nu}V^\sigma

Why: because the bracket multiplies $V^\sigma$ with no derivative, it is a tensor by the quotient theorem — the Riemann tensor — and its nonvanishing is the coordinate-free signature of curvature. *For a torsion-free connection the commutator on a vector is purely algebraic in $V$ — no derivatives of $V$ survive — confirming that curvature is a $(1,3)$ tensor field.*

Limiting cases

The covariant derivative degrades gracefully to familiar operations whenever the geometry allows:

Case	Result	Meaning
Flat space, Cartesian coordinates	$\Gamma^\lambda{}_{\mu\nu}=0\ \Rightarrow\ \nabla_\mu=\partial_\mu$	Constant basis, no correction; tensor calculus becomes ordinary calculus.
Scalar field	$\nabla_\mu\phi=\partial_\mu\phi$	No index to correct; the gradient covector is already tensorial.
Metric tensor	$\nabla_\mu g_{\nu\rho}=0$	Metric compatibility — the connection is built precisely so $g$ is covariantly constant.
Antisymmetrized lower indices	$\nabla_{[\mu}\omega_{\nu]}=\partial_{[\mu}\omega_{\nu]}$	Symmetric $\Gamma$ 's cancel; the exterior derivative needs no connection.
Covariant divergence	$\nabla_\mu V^\mu=\tfrac{1}{\sqrt{\|g\|}}\partial_\mu(\sqrt{\|g\|}V^\mu)$	The single contracted- $\Gamma$ identity absorbs all connection terms.

Common trapNonzero Christoffel symbols mean the space is curved.

They do not. The Christoffel symbols are coordinate-dependent and transform inhomogeneously — they are not the components of any tensor. In plane polar coordinates on the flat plane $\Gamma^r{}_{\theta\theta}=-r$ and $\Gamma^\theta{}_{r\theta}=1/r$ are nonzero, yet the plane is as flat as ever; conversely you can always make all $\Gamma$ vanish at any one chosen point by moving to a local inertial frame, even in genuinely curved spacetime. The coordinate-free test is not $\Gamma\neq0$ but the Riemann tensor $R^\rho{}_{\sigma\mu\nu}=\partial_\mu\Gamma^\rho{}_{\nu\sigma}-\partial_\nu\Gamma^\rho{}_{\mu\sigma}+\Gamma^\rho{}_{\mu\lambda}\Gamma^\lambda{}_{\nu\sigma}-\Gamma^\rho{}_{\nu\lambda}\Gamma^\lambda{}_{\mu\sigma}$ , a genuine tensor that vanishes identically for the flat plane in every coordinate system. Two traps ride along: never raise or lower indices on $\Gamma$ , and never take " $\Gamma=0$ here" to mean "flat here" — the derivatives of $\Gamma$ still carry curvature.

Worked examples

1 · Christoffel symbols and geodesics on the 2-sphere (confidence builder)

The round 2-sphere of radius $a$ has metric $\mathrm{d}s^2=a^2(\mathrm{d}\theta^2+\sin^2\theta\,\mathrm{d}\varphi^2)$ , so $g_{\theta\theta}=a^2$ , $g_{\varphi\varphi}=a^2\sin^2\theta$ , with inverses $g^{\theta\theta}=1/a^2$ , $g^{\varphi\varphi}=1/(a^2\sin^2\theta)$ . Find the Christoffel symbols, evaluate them at $\theta=60^\circ$ , and confirm that the equator is a geodesic.

Step 1. The only nonconstant metric component is $g_{\varphi\varphi}$ , with $\partial_\theta g_{\varphi\varphi}=2a^2\sin\theta\cos\theta$ ; every other derivative vanishes, and this one derivative feeds all the surviving symbols.

Step 2. Apply equation (4). For $\Gamma^\theta{}_{\varphi\varphi}$ only the $-\partial_\theta g_{\varphi\varphi}$ term survives: $\Gamma^\theta{}_{\varphi\varphi}=\tfrac12 g^{\theta\theta}(-\partial_\theta g_{\varphi\varphi})=\tfrac12\tfrac{1}{a^2}(-2a^2\sin\theta\cos\theta)=\mathbf{-\sin\theta\cos\theta}$ .

Step 3. Similarly $\Gamma^\varphi{}_{\theta\varphi}=\Gamma^\varphi{}_{\varphi\theta}=\tfrac12 g^{\varphi\varphi}\,\partial_\theta g_{\varphi\varphi}=\tfrac12\tfrac{1}{a^2\sin^2\theta}(2a^2\sin\theta\cos\theta)=\mathbf{\cot\theta}$ ; all other components vanish, and both are independent of the radius $a$ .

Step 4. Evaluate at $\theta=60^\circ$ : $\Gamma^\theta{}_{\varphi\varphi}=-\sin60^\circ\cos60^\circ=\mathbf{-0.4330}$ and $\Gamma^\varphi{}_{\theta\varphi}=\cot60^\circ=1/\sqrt3=\mathbf{0.5774}$ .

Step 5. Check the equator $\theta=\pi/2$ : there $\Gamma^\theta{}_{\varphi\varphi}=0$ , so the geodesic equation $\ddot\theta+\Gamma^\theta{}_{\varphi\varphi}\dot\varphi^2=0$ is solved by $\theta\equiv\pi/2$ with $\varphi$ advancing uniformly. The equator is a geodesic, and by symmetry so is every great circle — the straightest lines a sphere allows.

2 · Parallel-transport holonomy around a latitude circle (fully worked)

Parallel-transport a tangent vector once around a circle of constant colatitude $\theta_0$ on the unit sphere. By how much does it come back rotated, and what is the number for a circle at $\theta_0=30^\circ$ (i.e. latitude $60^\circ$ N)?

Step 1. The rotation after a closed loop — the holonomy angle — equals the integral of the Gaussian curvature over the enclosed cap (Gauss–Bonnet applied to the transport). For the unit sphere $K=1$ , so the holonomy is simply the enclosed area, i.e. the solid angle of the polar cap: $\alpha=\displaystyle\iint_{\text{cap}}K\,\mathrm{d}A=\Omega_{\text{cap}}$ .

Step 2. The cap area from the pole down to colatitude $\theta_0$ is $\Omega_{\text{cap}}=\int_0^{2\pi}\!\!\int_0^{\theta_0}\sin\theta\,\mathrm{d}\theta\,\mathrm{d}\varphi=2\pi\big(1-\cos\theta_0\big)$ — the closed-form holonomy for any latitude circle.

Step 3. Insert $\theta_0=30^\circ$ , where $\cos30^\circ=0.8660$ : $\alpha=2\pi(1-0.8660)=\mathbf{0.8418\ \text{rad}=48.2^\circ}$ . A vector carried once around latitude $60^\circ$ N returns tilted by about $48^\circ$ — a large, measurable effect of curvature with no local twisting anywhere along the path.

Step 4. Sanity checks: a tiny loop near the pole, $\theta_0\to0$ , gives $\alpha\to0$ (flat patch); transport around the equator, $\theta_0=\pi/2$ , gives $\alpha=2\pi\equiv0\pmod{2\pi}$ , no net rotation — consistent with the equator being a geodesic. Holonomy $=2\pi(1-\cos\theta_0)$ , the enclosed solid angle; $48.2^\circ$ at colatitude $30^\circ$ .

Physics of the portrait

The covariant derivative was built by three people across half a century: Christoffel supplied the symbols, Ricci-Curbastro forged them into a calculus, and Levi-Civita gave the calculus its geometric soul — parallel transport — just in time for Einstein to build gravity on it.

Origins · 1 minTullio Levi-Civita and how this was discovered

Tullio Levi-Civita

1873–1941 · differential geometry and the absolute differential calculus

Tullio Levi-Civita was the student and then collaborator of Gregorio Ricci-Curbastro at Padua; together in 1900 they published Méthodes de calcul différentiel absolu et leurs applications, the memoir that systematized the covariant derivative into the "absolute differential calculus" we now call tensor analysis. The machinery drew on Elwin Bruno Christoffel's 1869 symbols — invented to study the equivalence of quadratic differential forms — but Ricci and Levi-Civita turned a technical device into a general apparatus for coordinate-free calculus on manifolds. When Einstein needed exactly such a language for general relativity he found it ready-made, and corresponded warmly with its author. Levi-Civita's own decisive step came in 1917 with parallel transport: he showed the connection has a purely geometric meaning as the rule for sliding a vector along a curve while keeping it "parallel," so the Christoffel symbols become the generators of transport and curvature the failure of transport to be path-independent — the picture that still dominates differential geometry. A gentle, generous man, he was dismissed from his Rome chair in 1938 under the Fascist racial laws and died isolated in 1941; the connection that bears his name outlived the regime that silenced him.

History

1869

— Christoffel introduces the symbols and covariant differentiation of quadratic forms

1887

— Ricci-Curbastro begins developing the absolute differential calculus (tensor calculus)

1900

— Ricci and Levi-Civita publish the founding memoir on absolute differential calculus

1915

— Einstein casts general relativity using the metric covariant derivative

1917

— Levi-Civita introduces parallel transport, giving the connection its geometric meaning

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. on differential geometry and covariant derivatives · Arfken & Weber, Mathematical Methods for Physicists, tensor-analysis chapter · Hassani, Mathematical Physics, chapters on manifolds and connections · Nakahara, Geometry, Topology and Physics, §7 (connections and curvature) · Byron & Fuller, Mathematics of Classical and Quantum Physics, tensor chapters · Misner, Thorne & Wheeler, Gravitation, ch. 8–10 · Wald, General Relativity, ch. 3. Key numerical anchors: the mixed-tensor sign pattern follows from requiring $\nabla$ to commute with contraction; the metric formula (4) is recovered by the cyclic-sum trick; the sphere symbols are $\Gamma^\theta{}_{\varphi\varphi}=-\sin\theta\cos\theta$ and $\Gamma^\varphi{}_{\theta\varphi}=\cot\theta$ , giving $-0.4330$ and $0.5774$ at $60^\circ$ ; the holonomy is $2\pi(1-\cos\theta_0)=0.8418$ rad $=48.2^\circ$ at $\theta_0=30^\circ$ ; the contracted symbol satisfies $\Gamma^\mu{}_{\mu\nu}=\partial_\nu\ln\sqrt{|g|}$ , yielding the compact divergence (7); and the commutator (6) is algebraic in $V$ for a torsion-free connection. The polar values $\Gamma^r{}_{\theta\theta}=-r$ and $\Gamma^\theta{}_{r\theta}=1/r$ are the standard flat-space benchmarks.

Prerequisites. Covariant and Contravariant Tensors · Manifolds and Tangent Spaces · Christoffel Symbols and Affine Connections · The Metric Tensor. Leads to. Parallel Transport and Geodesics · The Riemann Curvature Tensor · Raising and Lowering Indices · Exterior Calculus and the Exterior Derivative.

Ask the tutor: Ask me to hand-compute the Christoffel symbols and Riemann tensor for any metric you name — Schwarzschild, FLRW, or the torus — and watch which coordinate artefacts cancel and which curvature survives.

Why does a scalar field need no Christoffel correction but a vector field does?

A scalar carries no indices and transforms trivially — the same number in every coordinate system — so $\partial_\mu\phi$ already transforms as a covector, a legitimate tensor with no repair needed. A vector carries an index whose transformation involves position-dependent Jacobians; when you differentiate its components you also, by the product rule, differentiate those Jacobians, producing a second-derivative term with no tensorial meaning. The $\Gamma V$ correction is built to cancel exactly this term, so the deficiency is specific to objects that carry indices, and it grows by one $\Gamma$ term for every additional index.

Is the Christoffel symbol a tensor?

No, and this is one of the most important points in the subject. Under a coordinate change $\Gamma$ transforms with the usual three-Jacobian tensor part plus an extra inhomogeneous term containing a second derivative of the coordinate map. Because of that term, if $\Gamma$ vanishes in one coordinate system it need not vanish in another — which no tensor would ever do. So you must never raise or lower an index on $\Gamma$ , and you cannot read anything coordinate-independent from its value at a point. The non-tensorial term is not a flaw — it is precisely what makes the full combination $\nabla_\mu V^\nu$ tensorial.

What is the difference between an affine connection and the Levi-Civita connection?

An affine connection is any rule assigning coefficients $\Gamma^\lambda{}_{\mu\nu}$ that transform inhomogeneously in the right way so that $\nabla$ produces tensors; there are infinitely many, and in general they may carry torsion (an antisymmetric part $\Gamma^\lambda{}_{[\mu\nu]}$ ) and respect no metric. The Levi-Civita connection is the unique affine connection that is simultaneously torsion-free and metric-compatible, singled out on any metric manifold with no extra input — which is why general relativity comes with a preferred covariant derivative for free. Theories with torsion (Einstein–Cartan gravity) exist, but the Levi-Civita choice is the default whenever a metric is present.

How is the covariant divergence related to the familiar spherical and cylindrical formulas?

They are the same thing. The compact form $\nabla_\mu V^\mu=\tfrac{1}{\sqrt{|g|}}\partial_\mu(\sqrt{|g|}V^\mu)$ uses the invariant volume factor $\sqrt{|g|}$ , which is $r$ in plane polar, $\rho$ in cylindrical, and $r^2\sin\theta$ in spherical coordinates. Substituting $\sqrt{|g|}=r^2\sin\theta$ reproduces the standard spherical divergence $\tfrac{1}{r^2}\partial_r(r^2 V^r)+\cdots$ instantly, with no need to differentiate unit vectors by hand. This single identity replaces the error-prone bookkeeping in the back of every methods textbook, and it is exactly why the covariant form is preferred for conservative numerical schemes on curved domains.

What does the covariant derivative have to do with parallel transport?

They are two descriptions of one object. Parallel transport slides a vector along a curve while keeping it "as parallel as possible," and the condition for that is precisely that its covariant derivative along the curve vanish: $\tfrac{\mathrm{d}V^\mu}{\mathrm{d}\tau}+\Gamma^\mu{}_{\nu\rho}\tfrac{\mathrm{d}x^\nu}{\mathrm{d}\tau}V^\rho=0$ . So the connection generates transport, and $\nabla$ measures how far a field departs from being parallel-transported. Because transport around a closed loop generally returns a vector rotated — the holonomy — the covariant derivative also encodes curvature: the commutator $[\nabla_\mu,\nabla_\nu]$ is the infinitesimal version of loop transport, giving the Riemann tensor.

Why is the covariant derivative the same idea as the gauge covariant derivative?

Because both cure the same disease: a local freedom that makes naive differentiation non-covariant. In geometry the freedom is the choice of frame at each point, and the cure is $\nabla_\mu=\partial_\mu+\Gamma$ . In gauge theory the freedom is the choice of internal phase at each point, and the cure is $D_\mu=\partial_\mu-igA_\mu$ , where the gauge potential $A_\mu$ plays the role of the connection. In both cases the connection transforms inhomogeneously precisely to absorb the offending term, its curvature is the physical object — the Riemann tensor there, the field strength $F_{\mu\nu}$ here — and parallel transport becomes minimal coupling. This shared structure is why Yang–Mills theory and general relativity look so alike.