Tangent Vectors and the Tangent Space

Every modern formulation of general relativity begins in the tangent space, and the abstraction earns its keep the moment fields become dynamical. Numerical-relativity codes that simulate binary black-hole mergers — the sources LIGO and Virgo detect — evolve tensor fields component by component in a chart, and the whole scheme rests on knowing that those components are contravariant and covariant objects living in $T_p\mathcal{M}$ and its dual, transforming by the Jacobian across the patched grids that tile the domain. Where the physics demands fermions — any spinor field on curved spacetime — the coordinate basis $\partial_\mu$ is no longer enough, and one switches to Cartan's orthonormal frame $e_a = e_a{}^\mu\partial_\mu$, an explicit basis of the tangent space at each event that lets the flat-space Dirac equation be glued on locally. And the deepest open problem of the field lives here too: quantizing gravity means quantizing the geometry that gives each point its tangent space, and rival programmes disagree about whether $T_p\mathcal{M}$ survives to the Planck scale at all, or dissolves into something discrete or noncommutative.

Stand at a point $p$ on a curved surface and ask the only first-order question available: "which way could I set off, and how fast?" Every answer is a curve $\gamma(t)$ with $\gamma(0)=p$, but the curve carries far more information than the question needs — its curvature, its acceleration, the whole path. Strip all of that away and keep only the initial velocity, and what remains is a tangent vector. Two curves that leave $p$ with the same velocity are the same vector; the equivalence class is the arrow. Crucially, that arrow does not lie in the surface — it lives in a fresh vector space $T_pM$, the tangent plane pressed against $M$ at $p$, and it is built without ever leaving the manifold.

There is a second, equivalent way to hold the same idea, and it is the one that scales to relativity. A tangent vector is not a thing you draw but a thing you do: an instruction for differentiating any scalar field. Feed it a function $f$ — a temperature, a potential, the curvature scalar — and it returns the rate of change of $f$ in its direction. Written this way, tangent vectors and directional derivatives are literally the same objects, adding and rescaling like the operators they are, which is exactly why $T_pM$ is a vector space. Every chart is a different scaffold laid over the same room: the basis vector $\partial_\mu$ means "increase $x^\mu$, hold the rest fixed," different charts give different scaffolds, and the Jacobian is the dictionary between them while the geometric vector stays put.

The cleanest definition never mentions coordinates. A tangent vector at $p$ is a derivation: a real-linear map on smooth functions that obeys the Leibniz product rule evaluated at $p$. These two axioms are the entire content of "directional derivative," abstracted away from any particular direction or chart, and everything else in this article is squeezed out of them.

where: $v$ — a tangent vector at $p$, viewed as a linear map $C^\infty(M)\to\mathbb{R}$; $f,g$ — smooth real-valued functions on $M$; $f(p)$ — the value of $f$ at $p$ (a number, which is why Leibniz produces $f(p)$ and $g(p)$, not $f$ and $g$); $a,b$ — real constants. Reading: $v$ is linear and behaves like a first derivative at the single point $p$. The kinematic definition looks entirely different but is the same object. Take a smooth curve through $p$ and let $v$ act on $f$ by differentiating $f$ along the curve at the moment it passes $p$. Linearity and Leibniz are then inherited from the ordinary derivative in $t$.

where: $\gamma$ — a smooth curve with $\gamma(0)=p$; $t$ — the curve parameter near $0$; $v[f]$ — the rate of change of $f$ along $\gamma$ at $p$. Reading: only the behaviour of $\gamma$ near $t=0$ can matter, and Derivation A shows only its first derivative survives. Pick a chart and the abstract space becomes computational. The coordinate basis vector $\partial_\mu|_p$ differentiates a function with respect to $x^\mu$ holding the other coordinates fixed, and every tangent vector is a unique combination of them (proved in Derivation B). The Einstein summation convention is in force: a repeated upper–lower index pair is summed over $1,\dots,n$.

where: $\partial_\mu|_p$ — the coordinate basis vector (rate of change along $x^\mu$); $V^\mu$ — the $n$ real components of $V$ in this chart; $x^{-1}$ — the inverse chart map, so $f\circ x^{-1}$ is an ordinary function on $\mathbb{R}^n$ that calculus can differentiate. Reading: the chart converts abstract derivations into partial derivatives. Feeding the coordinate functions to the basis vectors gives the Kronecker delta, and feeding them to an arbitrary $V$ reads off its components directly — an identity used constantly.

where: $x^\nu$ — the $\nu$-th coordinate regarded as a smooth scalar; $\delta^\nu{}_\mu$ — the Kronecker delta ($1$ if $\mu=\nu$, else $0$). Reading: to extract $V^\nu$, apply $V$ to the coordinate function $x^\nu$. Now change charts. Because a geometric vector cannot know which chart you chose, its components must transform to compensate. Applying $V$ to the new coordinate functions and using $(4)$ gives the contravariant law — the Jacobian acting on the components.

where: $V'^\mu$ — components in the primed chart $\{x'^\mu\}$; $\partial x'^\mu/\partial x^\nu$ — the Jacobian of the coordinate change, evaluated at $p$; $V^\nu$ — components in the unprimed chart. Reading: upper indices ride the Jacobian; this rule is the definition of a contravariant vector. The basis must then transform by the inverse Jacobian, oppositely to the components, so that the two motions cancel and $V=V^\mu\partial_\mu$ names the same object in every chart (Derivation C). Their common dimension count closes the circle.

where: $\partial'_\mu$ — the coordinate basis of the primed chart; $\partial x^\nu/\partial x'^\mu$ — the inverse Jacobian; $n = \dim M$. Reading: components and basis are contragredient, and there are exactly $n$ of each.

Special and degenerate situations sharpen what the tangent space is — and, just as usefully, what it is not.

1 · From a curve to its tangent vector, in two charts (confidence builder)

A particle moves on the plane along $\gamma(t) = (\cos t,\ \sin t)$, the unit circle. At $t=0$ it sits at $p=(1,0)$. Find its tangent vector there in Cartesian coordinates, then re-express it in polar coordinates $(r,\theta)$, and confirm the two descriptions agree.

Step 1. In Cartesian coordinates the components are the coordinate velocities: $\dot{x}(0) = -\sin 0 = 0$ and $\dot{y}(0) = \cos 0 = 1$. So $V = \dot{x}\,\partial_x + \dot{y}\,\partial_y = \mathbf{\partial_y}$ at $p$ — the particle is heading straight "up" the $y$-axis, as a circle crossing $(1,0)$ should.

Step 2. In polar coordinates the same curve is $r(t)=1,\ \theta(t)=t$, so $\dot{r}(0)=0$ and $\dot{\theta}(0)=1$, giving $V = \dot{r}\,\partial_r + \dot{\theta}\,\partial_\theta = \mathbf{\partial_\theta}$. Two different coordinate stories, one geometric vector.

Step 3. Check consistency through the basis. At $p=(1,0)$ we have $r=1,\ \theta=0$, and $\partial_\theta = \dfrac{\partial x}{\partial\theta}\partial_x + \dfrac{\partial y}{\partial\theta}\partial_y = (-r\sin\theta)\,\partial_x + (r\cos\theta)\,\partial_y = 0\cdot\partial_x + 1\cdot\partial_y = \partial_y$. The two expressions are the identical vector: $V = \partial_y = \partial_\theta$. Note $\partial_\theta$ here happens to have length $r=1$; move the same construction out to $r=5$ and $\partial_\theta$ would be five times longer — the coordinate basis is not normalized.

2 · The contravariant law and invariance, with integer arithmetic

Introduce new coordinates $x' = 2x + y,\ y' = x + y$ on the plane. A vector at a point has Cartesian components $V^x = 3,\ V^y = 1$. Compute its primed components with the transformation law, then verify explicitly that $V = V'^\mu\partial'_\mu$ recovers the original vector.

Step 1. The Jacobian is $\dfrac{\partial x'^\mu}{\partial x^\nu} = \begin{pmatrix} 2 & 1\\ 1 & 1\end{pmatrix}$ (rows $\mu = x',y'$; columns $\nu = x,y$). Apply the contravariant law: $V'^{x'} = 2\cdot 3 + 1\cdot 1 = \mathbf{7}$ and $V'^{y'} = 1\cdot 3 + 1\cdot 1 = \mathbf{4}$. The components changed completely, from $(3,1)$ to $(7,4)$.

Step 2. The basis transforms by the inverse Jacobian. Solving the coordinate relations gives $x = x' - y',\ y = 2y' - x'$, so $\partial'_{x'} = \partial_x - \partial_y$ and $\partial'_{y'} = -\partial_x + 2\partial_y$. (The determinant of the change is $2\cdot1 - 1\cdot1 = 1$, so it is invertible — a genuine chart change.)

Step 3. Reassemble the vector in the new basis: $V = 7\,\partial'_{x'} + 4\,\partial'_{y'} = 7(\partial_x - \partial_y) + 4(-\partial_x + 2\partial_y) = (7-4)\partial_x + (-7+8)\partial_y = 3\partial_x + \partial_y$. This is exactly the original $V$: the component change $(3,1)\to(7,4)$ and the basis change cancel, and the geometric vector is invariant. That cancellation is the entire reason upper and lower indices transform by reciprocal matrices.

The tangent space is the point where the loose arrows of nineteenth-century vector analysis were forced to grow up. Once geometry was recognized as intrinsic — a manifold with no ambient space — the naive arrow had to be rebuilt from the smooth structure alone, and the figure who did more than anyone to make "a linear space at each point" the organizing idea of geometry, and to carry it directly into general relativity, was Élie Cartan.

Texts. Misner, Thorne & Wheeler, Gravitation, §2.5–2.6 and §8 (the "machine" picture of vectors) · Wald, General Relativity, §2.2–2.3 (the derivation definition and dimension theorem) · Carroll, Spacetime and Geometry, §2.2–2.3 · Schutz, A First Course in General Relativity, ch. 5 · Hartle, Gravity, ch. 7 · Lee, Introduction to Smooth Manifolds, ch. 3 (the cleanest modern treatment of derivations and the Hadamard lemma). The key checkpoints of this article are: the constants-killed argument $v(\mathbf{1}) = 2v(\mathbf{1}) \Rightarrow v(\mathbf{1})=0$; the Hadamard-lemma span proof yielding $\dim T_pM = n$; the contravariant law $V'^\mu = (\partial x'^\mu/\partial x^\nu)V^\nu$ together with the inverse-Jacobian basis law and their cancellation; and the worked-example arithmetic $V'=(7,4)$ reconstructing to $V=(3,1)$, and the circle velocity $\partial_y = \partial_\theta$ at $(1,0)$. No dimensionful constants are asserted beyond the standard normalization $g(u,u)=-1$ in geometrized units.

Builds on: Smooth Manifolds and Coordinate Charts · Atlases, Transition Functions and the Differentiable Structure. Leads to: The Cotangent Space and One-Forms · Vector Fields and Integral Curves · General Tensors and the Tensor Product · The Tensor Transformation Law · The Metric Tensor and the Line Element.

Ask the tutor: ask me to work any coordinate change you like — polar, rotated, or sheared — and watch the components transform while the vector itself stands perfectly still.