Manifolds and Tangent Spaces

The manifold-plus-tangent-space structure is the working substrate of modern theoretical physics, not decorative rigour. General relativity models spacetime as a four-dimensional pseudo-Riemannian manifold on which four-velocities are tangent vectors and gradients are covectors; the LIGO/Virgo gravitational-wave detections and the Event Horizon Telescope's black-hole images are quantitative tests of curvature — the path-dependence of comparing tangent vectors. In gauge theory the tangent bundle is the prototype for the fibre bundles whose connections are the electroweak and strong forces. Robotics plans motion directly on configuration manifolds — a rigid body's pose lives on the 6-manifold $SE(3)$ — because treating joint angles as flat vectors commits exactly the error the warning forbids. Geometric integrators for molecular dynamics and spacecraft attitude respect the manifold structure to conserve invariants over billions of steps, and topological data analysis assumes data clouds sample a low-dimensional manifold, using tangent-space estimates to learn its shape. The bare arena Riemann sketched in 1854 is now standard equipment across the exact sciences.

Hold a globe and lay a city street-map beside it: the globe is curved and closed, the map a flat sheet of $\mathbb{R}^2$, yet within one city the map is faithful. That fidelity within a patch, together with the fact that adjacent maps agree up to a smooth stretch where they overlap, is precisely the manifold compatibility condition. No single flat map covers the whole globe without a tear, which is why cartographers bind an atlas of overlapping charts. The first figure makes this concrete: the sphere carries a two-chart atlas, and at each point a flat tangent plane — the best linear approximation to the surface there.

A tangent vector is nothing more mysterious than the velocity of a curve threading through $p$: two curves sharing an instantaneous velocity define the same vector, so a tangent vector packages only first-order data. Equivalently, and more powerfully, it is the operation "take the rate of change of any function along me" — an idea needing no surrounding space, only the smooth functions on $M$. The subtlety that drives everything downstream is that a vector's components are chart-dependent while the vector is not: change coordinates and the numbers change, but in exactly the compensating way — with the Jacobian — that keeps the geometric arrow fixed. The second figure pins this down with real numbers: one tangent vector, read in two coordinate systems, wears two different lists of components.

The manifold is defined by how its charts fit together. Two charts are smoothly compatible when the map that translates coordinates from one to the other, on the region where they overlap, is infinitely differentiable. An atlas of pairwise-compatible charts is what promotes a topological space to a smooth manifold.

where: $U_\alpha$ — an open set of $M$ carried by chart $\varphi_\alpha$; $\varphi_\alpha$ — a homeomorphism $U_\alpha\to\varphi_\alpha(U_\alpha)\subset\mathbb{R}^n$; $\varphi_\beta\circ\varphi_\alpha^{-1}$ — the transition (overlap) map; $C^\infty$ — the class of infinitely differentiable maps. The transition maps live entirely in $\mathbb{R}^n$, so smoothness is checked with ordinary multivariable calculus — the whole content of "$M$ is smooth." The coordinate-free definition of a tangent vector asks nothing of an ambient space. A tangent vector at $p$ is a derivation: a real-linear operator on smooth functions that obeys the Leibniz product rule at $p$. These two properties alone will turn out to force it to be a directional derivative.

where: $v$ — the tangent vector, acting as a derivation; $f,g$ — smooth real functions near $p$; $f(p),g(p)$ — their values at $p$ (real numbers); $a,b$ — real constants. Leibniz plus linearity is a surprisingly rigid pair of conditions: the first derivation card below shows they pin $v$ down to exactly $n$ numbers. Choosing a chart $x^\mu$ turns the abstract derivation into a concrete list of components. The partial-derivative operators $\partial/\partial x^\mu$ form a basis of $T_pM$, and every tangent vector is a linear combination of them.

where: $v^\mu$ — the $n$ components of $v$ in the chart $x^\mu$; $\partial/\partial x^\mu|_p$ — the coordinate basis vector at $p$; repeated indices are summed (Einstein convention). The dimension of $T_pM$ equals the dimension of $M$ — a fact, not an assumption, proved from the derivation axioms. Because the vector $v$ is a single geometric object, its components in a new chart $\tilde x^\nu$ are fixed by the old ones. Applying the chain rule to the basis and demanding that $v$ itself not change yields the contravariant transformation law.

where: $\tilde v^{\nu}$ — components in the new chart; $\partial\tilde x^{\nu}/\partial x^{\mu}$ — the Jacobian of the coordinate change. The components carry the Jacobian while the basis vectors carry its inverse; they pull oppositely so the arrow stays put — the meaning of contravariant. The cotangent space $T_p^*M$ is the dual: the space of linear maps from $T_pM$ to $\mathbb{R}$. The differentials $\mathrm{d}x^\mu$ form the basis dual to $\{\partial_\nu\}$, and a general covector is a combination of them.

where: $\mathrm{d}x^\mu$ — the coordinate basis covector; $\delta^{\mu}{}_{\nu}$ — the Kronecker delta; $\omega_\mu$ — the components of the covector $\omega$; $T_p^*M$ — the cotangent space. The bracket $\langle\cdot,\cdot\rangle$ is the evaluation of a covector on a vector, a pure number requiring no metric. A covector is a geometric object too, so its components also transform — but with the inverse Jacobian, the opposite of a vector. This is the covariant law.

where: $\tilde\omega_\nu$ — covector components in the new chart; $\partial x^{\mu}/\partial\tilde x^{\nu}$ — the inverse Jacobian. Vectors up, covectors down; the two laws are inverse-transpose to each other, which is exactly what makes their contraction survive a coordinate change. The pairing of a covector with a vector is a single real number that every observer agrees on, because the two opposing Jacobians cancel. This is the algebraic root of why contracted indices produce scalars.

where: $\omega_\mu v^\mu$ — the contraction of covector and vector components. The prototypical covector is the differential of a function, $\mathrm{d}f=(\partial f/\partial x^\mu)\,\mathrm{d}x^\mu$, whose pairing with $v$ returns the directional derivative $v(f)=v^\mu\partial_\mu f$ — closing the loop back to the derivation definition. Assembling every tangent space over every point gives a new space in its own right, the tangent bundle. It is itself a smooth manifold, of twice the dimension, with a projection back down to $M$.

where: $TM$ — the tangent bundle, the disjoint union of all tangent spaces; $\pi$ — the projection sending a vector at $p$ to its base point $p$; $\bigsqcup$ — disjoint union. A point of $TM$ is a pair of a base point and a vector, needing $n+n$ coordinates. A vector field $X$ is a smooth section, $X:M\to TM$ with $\pi\circ X=\mathrm{id}_M$, written $X=X^\mu(x)\,\partial_\mu$.

1 · Stereographic charts on the 2-sphere (fully worked)

Cover $S^2$ (the unit sphere in $\mathbb{R}^3$) with two stereographic charts, one from the north pole and one from the south. Take the point $P=(0,\tfrac35,\tfrac45)$ on the sphere, find its coordinates in both charts, and verify the transition map is the smooth inversion $u\mapsto u/|u|^2$.

Step 1. Stereographic projection from the north pole $N=(0,0,1)$ sends a sphere point $(x,y,z)$ to $\varphi_N(x,y,z)=\big(\tfrac{x}{1-z},\tfrac{y}{1-z}\big)$, and it is a chart on $S^2\setminus\{N\}$. For $P$, with $z=\tfrac45$ so $1-z=\tfrac15$, this gives $\varphi_N(P)=\big(0,\ \tfrac{3/5}{1/5}\big)=\mathbf{(0,\ 3)}$. Check: $P\neq N$, so $P$ lies in this chart's domain.

Step 2. Projection from the south pole $S=(0,0,-1)$ sends $(x,y,z)$ to $\varphi_S(x,y,z)=\big(\tfrac{x}{1+z},\tfrac{y}{1+z}\big)$, a chart on $S^2\setminus\{S\}$. With $1+z=\tfrac95$, $\varphi_S(P)=\big(0,\ \tfrac{3/5}{9/5}\big)=\mathbf{(0,\ 1/3)}$. The two charts together omit no point — $N$ is covered by the south chart and $S$ by the north — so two charts form a complete atlas.

Step 3. On the overlap $S^2\setminus\{N,S\}$ the transition map is $\varphi_S\circ\varphi_N^{-1}(u,v)=\dfrac{(u,v)}{u^2+v^2}$, inversion in the unit circle. Test it on the north coordinates $(0,3)$: $u^2+v^2=0+9=9$, so the image is $(0,3)/9=\mathbf{(0,\ 1/3)}$ — exactly $\varphi_S(P)$. The two independent computations agree.

Step 4. Is the transition smooth? Inversion $F(u,v)=(u,v)/(u^2+v^2)$ is $C^\infty$ wherever $u^2+v^2\neq0$, i.e. everywhere on the overlap (the poles, where the denominator would vanish, are excluded by construction). Its Jacobian determinant is $\det\mathrm{D}F=-1/(u^2+v^2)^2$, which at our point equals $-1/9^2=\mathbf{-1/81}\neq0$, confirming a local diffeomorphism. So $S^2$ is a smooth 2-manifold with a 2-chart atlas, $P$ reads as $(0,3)$ and $(0,1/3)$, and the overlap map is the smooth inversion.

2 · One tangent vector, two coordinate systems (fully worked)

At the plane point $(x,y)=(1,1)$, take the tangent vector $\hat e_x=\partial/\partial x$. Express it in polar coordinates $(r,\theta)$ and confirm the contravariant transformation law.

Step 1. Set up the coordinate change $r=\sqrt{x^2+y^2}$, $\theta=\arctan(y/x)$. At $(1,1)$ we have $r=\sqrt2\approx1.414$ and $\theta=45^\circ=\pi/4$, so $\cos\theta=\sin\theta=1/\sqrt2$. These are the values feeding the Jacobian.

Step 2. The contravariant law (equation 4) reads $\tilde v^\nu=(\partial\tilde x^\nu/\partial x^\mu)v^\mu$ with $\tilde x=(r,\theta)$, $x=(x,y)$, and $v=(1,0)$ for $\hat e_x$. The needed Jacobian entries are $\partial r/\partial x=x/r=\cos\theta$ and $\partial\theta/\partial x=-y/r^2=-\sin\theta/r$.

Step 3. Read off the polar components: $v^r=\dfrac{\partial r}{\partial x}\cdot1=\cos\theta=\dfrac{1}{\sqrt2}\approx\mathbf{0.707}$ and $v^\theta=\dfrac{\partial\theta}{\partial x}\cdot1=-\dfrac{\sin\theta}{r}=-\dfrac{1/\sqrt2}{\sqrt2}=\mathbf{-\tfrac12}$. So the same arrow that was $(1,0)$ in Cartesian components is $(1/\sqrt2,\,-1/2)$ in polar components.

Step 4. Sanity check by rebuilding the vector: $v^r\,\partial_r+v^\theta\,\partial_\theta$ should equal $\partial_x$. Using $\partial_r=\cos\theta\,\partial_x+\sin\theta\,\partial_y$ and $\partial_\theta=-r\sin\theta\,\partial_x+r\cos\theta\,\partial_y$, the $\partial_y$ coefficient is $v^r\sin\theta+v^\theta r\cos\theta=\tfrac1{\sqrt2}\cdot\tfrac1{\sqrt2}+(-\tfrac12)\sqrt2\cdot\tfrac1{\sqrt2}=\tfrac12-\tfrac12=0$, and the $\partial_x$ coefficient is $\tfrac1{\sqrt2}\cdot\tfrac1{\sqrt2}+(-\tfrac12)(-\sqrt2\cdot\tfrac1{\sqrt2})=\tfrac12+\tfrac12=1$. The components change from (1, 0) to (1/√2, −1/2), yet the geometric vector is unchanged — the signature of contravariance.

3 · Counting dimensions: the double pendulum

A planar double pendulum has two angles $(\theta_1,\theta_2)$; identify its configuration manifold and the spaces mechanics builds on it. Each angle lives on a circle $S^1$, so the configuration space is the 2-torus $Q=T^2$, with $\dim Q=\mathbf{2}$. Velocities live in $T_qQ$, so the Lagrangian $L(q,\dot q)$ is a function on $TQ$ with $\dim TQ=2n=\mathbf{4}$; the momenta $p_i=\partial L/\partial\dot\theta^i$ are covectors in $T_q^*Q$, so phase space is $T^*Q$, also dimension $\mathbf{4}$. The velocity/momentum split is exactly the vector/covector split.

The manifold is one of the rare mathematical ideas whose birth can be dated to a single afternoon. Its lineage runs from Gauss's intrinsic curvature, through Riemann's leap to $n$ dimensions, to the topologists who made the definition airtight — and then straight into general relativity and gauge theory, where the manifold is not a metaphor but the stage on which physics is written.

Texts. Stone & Goldbart, Mathematics for Physics, Ch. 11 (Differential Calculus on Manifolds) · Nakahara, Geometry, Topology and Physics, 2nd ed., Ch. 5 · Lee, Introduction to Smooth Manifolds, 2nd ed., Ch. 1–3 and 10 · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 4 · Frankel, The Geometry of Physics, 3rd ed., Ch. 1–2 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 26. Key numerical checks: Hadamard's lemma gives $\dim T_pM=n$ with basis $\partial_\mu$; the contravariant and covariant laws are inverse-transpose, so $\omega_\mu v^\mu$ is invariant; the stereographic transition $u\mapsto u/|u|^2$ is smooth with $\det\mathrm{D}F=-1/|u|^4=-1/81$ at the worked point, and $P=(0,\tfrac35,\tfrac45)$ maps to $(0,3)$ and $(0,\tfrac13)$; $\hat e_x$ at $(1,1)$ has polar components $(1/\sqrt2,-1/2)$; and $\dim TM=\dim T^*Q=2n$.

Prerequisites. Multivariable calculus and the Jacobian · Linear algebra and dual spaces · Point-set topology. Leads to. Covariant and Contravariant Tensors · Tensor Algebra and the Tensor Product · The Metric Tensor · Christoffel Symbols and Affine Connections · Differential Forms.

Ask the tutor: Ask me to work any coordinate change as an explicit Jacobian — polar, spherical, or your own — and watch the same tangent vector keep its identity while its components rearrange.