Topology

Smooth manifolds

Spaces built from an atlas $\{(U_\alpha, \varphi_\alpha)\}_{\alpha \in A}$ that locally look like $\mathbb{R}^n$ and support calculus, culminating in Stokes' theorem $\int_M d\omega = \int_{\partial M} \omega$ , the setting for modern geometry, physics, and robotics.

IntuitionCurved spaces that look flat up close

Standing on Earth's surface, the ground looks flat even though the planet is a sphere. A smooth manifold formalizes this: a space that, zoomed in near any point, looks like ordinary Euclidean space $\mathbb{R}^n$ , even though globally it can be curved, twisted, or have interesting topology (like a sphere, torus, or the space of rotations). This "locally flat, globally curved" idea lets us do calculus — derivatives, integrals, vector fields — on shapes that are not flat. The parametric surface below (sphere, torus, helicoid) shows several manifolds whose local patches are each a distorted copy of the flat plane.

A torus built by gluing local coordinate patches — a smooth

2

-manifold embedded in

\mathbb{R}^3

UndergraduateAtlases, tangent spaces, and differentials

Definition: Smooth atlas and manifold

An $n$ -dimensional smooth manifold $M$ is a topological space covered by charts $\{(U_\alpha, \varphi_\alpha)\}_{\alpha \in A}$ , where each $\varphi_\alpha : U_\alpha \to \mathbb{R}^n$ is a homeomorphism onto an open subset, and on overlaps the transition maps $\varphi_\beta \circ \varphi_\alpha^{-1} : \varphi_\alpha(U_\alpha \cap U_\beta) \to \varphi_\beta(U_\alpha \cap U_\beta)$ are $C^\infty$ (infinitely differentiable). This smoothness of transitions is what lets us transfer calculus from $\mathbb{R}^n$ , done in any one chart, consistently to the whole manifold.

\varphi_\beta \circ \varphi_\alpha^{-1} : \varphi_\alpha(U_\alpha \cap U_\beta) \to \varphi_\beta(U_\alpha \cap U_\beta)

At each point $p \in M$ , the tangent space $T_p M$ is the $n$ -dimensional vector space of all velocity vectors of curves through $p$ — it is the best linear approximation to $M$ at $p$ . A smooth map $f: M \to N$ then has a differential (derivative) at each point, a linear map between tangent spaces:

df_p : T_p M \to T_{f(p)} N

Manifolds and their dimension / tangent space
Manifold	Dimension	$T_pM$ description
Sphere $S^{n-1} \subset \mathbb{R}^n$	$n-1$	Vectors orthogonal to $p$
Rotation group $SO(3)$	$3$	Skew-symmetric matrices (angular velocity)
Torus $T^2 = S^1\times S^1$	$2$	Plane spanned by two angle directions

AdvancedTwo central theorems

Theorem: The Regular Value (Preimage) Theorem

Let $f: M^m \to N^n$ be smooth and $q \in N$ a regular value (meaning $df_p$ is surjective for every $p \in f^{-1}(q)$ ). Then $f^{-1}(q)$ is a smooth submanifold of $M$ of dimension $m-n$ .

Why is it true?

This theorem is the primary tool for manufacturing manifolds: instead of exhibiting an atlas by hand, we describe a shape as the zero set of a map and just check a linear-algebra condition (surjectivity of the differential) at each solution point.

Proof

Fix $p \in f^{-1}(q)$ . Since $df_p: T_pM \to T_qN$ is surjective and $\dim T_pM = m \ge n = \dim T_qN$ , its kernel $K = \ker df_p$ has dimension $m-n$ . Choose a linear complement $W$ so $T_pM = K \oplus W$ with $\dim W = n$ ; then $df_p|_W : W \to T_qN$ is an isomorphism.

Work in local coordinates centered at $p$ and $q$ (via charts), so $f$ becomes a smooth map $\mathbb{R}^m \to \mathbb{R}^n$ with $f(0)=0$ and $df_0$ surjective. Reorder coordinates $(x,y) \in \mathbb{R}^{m-n}\times\mathbb{R}^n$ so that $\partial f/\partial y$ at $0$ is the invertible $n\times n$ block (possible since $df_0$ has rank $n$ ).

Define $\Phi(x,y) = (x, f(x,y))$ . Then $d\Phi_0 = \begin{pmatrix} I & 0 \\ \partial f/\partial x & \partial f/\partial y \end{pmatrix}$ has $\det d\Phi_0 = \det(\partial f/\partial y) \ne 0$ , so $\Phi$ is a local diffeomorphism by the Inverse Function Theorem. In the new coordinates $(x,y') = \Phi(x,y)$ , the equation $f=q$ (i.e. $f=0$ ) becomes exactly $y'=0$ .

So near $p$ , $f^{-1}(q)$ is the set $\{y'=0\}$ , which in these coordinates is literally an $(m-n)$ -dimensional coordinate slice — a smooth chart for $f^{-1}(q)$ . Since $p$ was arbitrary, every point of $f^{-1}(q)$ has such a chart, and the transition maps between these charts are restrictions of the (smooth) transition maps of $M$ , hence smooth. Therefore $f^{-1}(q)$ is a smooth $(m-n)$ -dimensional submanifold.

Theorem: Stokes' Theorem

For a compact oriented $n$ -manifold with boundary $M$ and a smooth $(n-1)$ -form $\omega$ on $M$ , $\int_M d\omega = \int_{\partial M} \omega$ .

Why is it true?

This single identity unifies the fundamental theorem of calculus, Green's theorem, the divergence theorem, and the classical Stokes' theorem from vector calculus into one statement about differential forms, and is the analytic engine behind de Rham cohomology.

Proof

Step 1 (local case, half-space). First suppose $M = \mathbb{H}^n = \{x_n \ge 0\}$ and $\omega$ has compact support in a single chart. Write $\omega = \sum_i f_i\, dx_1\wedge\cdots\widehat{dx_i}\cdots\wedge dx_n$ . Then $d\omega = \sum_i (-1)^{i-1}\frac{\partial f_i}{\partial x_i}dx_1\wedge\cdots\wedge dx_n$ , and $\int_{\mathbb{H}^n} d\omega = \sum_i (-1)^{i-1}\int \frac{\partial f_i}{\partial x_i}\,dx_1\cdots dx_n$ .

For $i < n$ , integrating $\partial f_i/\partial x_i$ over $x_i \in \mathbb{R}$ first and using compact support gives $0$ by the ordinary Fundamental Theorem of Calculus ( $f_i \to 0$ at $x_i=\pm\infty$ ). For $i=n$ , integrating over $x_n \in [0,\infty)$ gives $\int \frac{\partial f_n}{\partial x_n}dx_n = [f_n]_0^\infty = -f_n(x_1,\dots,x_{n-1},0)$ (again using compact support at $x_n=\infty$ ), so only the $i=n$ term survives: $\int_{\mathbb{H}^n}d\omega = (-1)^{n-1}\int_{\mathbb{R}^{n-1}} \left(-f_n(x_1,\dots,x_{n-1},0)\right)dx_1\cdots dx_{n-1}$ .

On the boundary $\partial\mathbb{H}^n = \{x_n=0\}$ (oriented so the outward normal $-\partial_n$ comes last, giving orientation sign $(-1)^n$ ), the restriction of $\omega$ is $\omega|_{\partial} = f_n\, dx_1\wedge\cdots\wedge dx_{n-1}$ (all other terms restrict to $0$ since they contain $dx_n$ or vanish on the slice). A direct sign check using the standard boundary orientation convention shows $\int_{\partial \mathbb{H}^n}\omega = (-1)^n \int f_n\,dx_1\cdots dx_{n-1}$ , matching the formula above exactly. So $\int_{\mathbb{H}^n}d\omega = \int_{\partial\mathbb{H}^n}\omega$ in this local model.

Step 2 (partition of unity, globalize). For general $M$ and general $\omega$ , cover $M$ by finitely many charts $\{(U_\alpha,\varphi_\alpha)\}$ (using compactness) and choose a smooth partition of unity $\{\rho_\alpha\}$ subordinate to this cover, i.e. $\sum_\alpha \rho_\alpha = 1$ with $\operatorname{supp}\rho_\alpha \subset U_\alpha$ . Write $\omega = \sum_\alpha \rho_\alpha\omega$ ; each $\rho_\alpha\omega$ has compact support inside a single chart, where the chart is either entirely interior (in which case $\int_{\partial}\rho_\alpha\omega = 0$ trivially and $\int_M d(\rho_\alpha\omega)=0$ by Step 1 applied to $\mathbb{R}^n$ with no boundary) or meets $\partial M$ (Step 1 applies directly after transporting via $\varphi_\alpha$ , which preserves both $d$ and orientation).

Since $d$ is linear, $d\omega = \sum_\alpha d(\rho_\alpha\omega)$ (using $\sum_\alpha d\rho_\alpha = d(\sum_\alpha \rho_\alpha) = d(1) = 0$ to handle the cross terms $d\rho_\alpha \wedge \omega$ correctly when summed). Integrating and summing the local identities from Step 1 over all $\alpha$ : $\int_M d\omega = \sum_\alpha \int_M d(\rho_\alpha\omega) = \sum_\alpha \int_{\partial M} \rho_\alpha\omega = \int_{\partial M}\omega$ , which is exactly $\int_M d\omega = \int_{\partial M} \omega$ .

AdvancedManufacturing manifolds: the sphere and $O(n)$

Apply the Regular Value Theorem to $f:\mathbb{R}^n \to \mathbb{R}$ , $f(x)=|x|^2$ : since $df_x(v) = 2\langle x,v\rangle$ is surjective (nonzero) for every $x\ne 0$ , $q=1$ is a regular value, so $S^{n-1} = f^{-1}(1)$ is a smooth $(n-1)$ -manifold — no explicit atlas needed! Similarly, for $f:\mathbb{R}^{n\times n}\to \operatorname{Sym}(n)$ , $f(A)=A^TA$ (landing in the $\binom{n+1}{2}$ -dimensional space of symmetric matrices), one checks $df_A$ is surjective at every $A$ with $A^TA=I$ , so $O(n) = f^{-1}(I)$ is a smooth manifold of dimension $n^2 - \binom{n+1}{2} = \binom{n}{2}$ — this single computation instantly gives $SO(3)$ , the configuration space of a rigid body's rotations, dimension $3$ .

UndergraduateReal-World Applications and Worked Examples

Robotics models a rigid body's orientation as a point on the manifold $SO(3)$ , not as three Euler angles (which have coordinate singularities — gimbal lock). Motion planning, control, and state estimation (e.g. the Kalman filter used on satellites and drones) are all done using the tangent space $\mathfrak{so}(3)$ (angular velocities) and the exponential map, exploiting the smooth manifold structure to avoid singularities. In geometric mechanics, a mechanical system's configuration space is a manifold $M$ (e.g. a double pendulum's is $T^2$ ), and Lagrangian/Hamiltonian dynamics live naturally on $TM$ and $T^*M$ ; conserved quantities correspond to symmetries of $M$ via Noether's theorem, a manifold-theoretic upgrade of classical mechanics.

Example: Dimension of the special orthogonal group $SO(3)$

Using $f(A)=A^TA$ on $3\times 3$ real matrices, confirm the dimension of $O(3)$ (hence $SO(3)$ , its identity component) predicted by the Regular Value Theorem.

Solution

The domain $\mathbb{R}^{3\times 3}$ has dimension $n^2=9$ . The codomain $\operatorname{Sym}(3)$ (symmetric $3\times 3$ matrices) has dimension $\binom{n+1}{2} = \binom{4}{2}=6$ .

One checks $df_A$ is surjective at every $A\in O(3)$ : differentiating $f(A)=A^TA$ gives $df_A(H) = H^TA + A^TH$ ; for any symmetric $S$ , take $H = \tfrac12 AS$ , then $df_A(H) = \tfrac12 S^TA^TA + \tfrac12 A^TAS = \tfrac12 S + \tfrac12 S = S$ (using $A^TA=I$ ), showing surjectivity.

By the Regular Value Theorem, $O(3) = f^{-1}(I)$ has dimension $9 - 6 = 3$ . Since $SO(3)$ is the connected component of the identity, it has the same dimension, $3$ — matching the intuitive count of $3$ independent rotation axes/angles (e.g. roll, pitch, yaw).

Example: Verifying Stokes' theorem on a disk

Let $M$ be the unit disk $\{x^2+y^2\le 1\}\subset\mathbb{R}^2$ with boundary the unit circle, and $\omega = x\,dy$ . Verify $\int_M d\omega = \int_{\partial M}\omega$ directly.

Solution

Compute the left side: $d\omega = dx\wedge dy$ , so $\int_M d\omega = \int_M dx\,dy = \pi(1)^2 = \pi$ (the area of the unit disk).

Compute the right side: parametrize $\partial M$ by $x=\cos t, y=\sin t$ , $t\in[0,2\pi]$ , so $dy = \cos t\,dt$ . Then $\int_{\partial M}\omega = \int_0^{2\pi} \cos t \cdot \cos t\,dt = \int_0^{2\pi}\cos^2t\,dt$ .

Using $\cos^2 t = \tfrac{1+\cos 2t}{2}$ , $\int_0^{2\pi}\cos^2t\,dt = \tfrac12\left[t+\tfrac{\sin 2t}{2}\right]_0^{2\pi} = \tfrac12(2\pi) = \pi$ . Both sides equal $\pi$ , confirming Stokes' theorem in this case.

Common mistake. A frequent mistake is thinking a "manifold" must be embedded in some

\mathbb{R}^N

and thinking of the atlas as extra decoration — actually the atlas (or an intrinsic equivalent) is the definition; abstract manifolds like

SO(3)

or the space of rotations need not come with a preferred embedding. Another common error in applying the Regular Value Theorem: forgetting to check that

q

is regular at every point of

f^{-1}(q)

, not just "generically" — a single point where

df_p

fails to be surjective (a critical point) can make

f^{-1}(q)

fail to be a manifold there (e.g. it can have a cone singularity). And in Stokes' theorem, orientation conventions on

\partial M

matter: reversing the boundary orientation flips the sign of

\int_{\partial M}\omega

Research frontier as of 2026

Classifying smooth manifolds up to diffeomorphism remains fundamentally open in dimension $4$ : **exotic $\mathbb{R}^4$ 's** (smooth structures on $\mathbb{R}^4$ not diffeomorphic to the standard one, discovered via Donaldson and Freedman's work in the 1980s) exist in uncountably many varieties, and whether the smooth $4$ -dimensional Poincaré conjecture holds is still open. In applied directions, manifold learning and geometric deep learning treat high-dimensional data as sampled from a low-dimensional manifold embedded in a larger space, using tools like the tangent space and Riemannian metrics to design neural networks equivariant to symmetries of $M$ (e.g. spherical CNNs on $S^2$ , graph neural networks that respect manifold structure). Symplectic and contact topology (Gromov, Eliashberg school) continue to reveal exotic phenomena — like exotic symplectic structures on $\mathbb{R}^{2n}$ — that have no analogue in the smooth (non-symplectic) category.

What smoothness condition must the transition maps $\varphi_\beta \circ \varphi_\alpha^{-1} : \varphi_\alpha(U_\alpha \cap U_\beta) \to \varphi_\beta(U_\alpha \cap U_\beta)$ satisfy for an atlas to define a smooth manifold?

By the Regular Value Theorem applied to $f(A)=A^TA$ on $3\times 3$ matrices, what is $\dim SO(3)$ ?

In robotics, why is $SO(3)$ preferred over three Euler angles to represent orientation?

Stokes' theorem $\int_M d\omega = \int_{\partial M} \omega$ generalizes which classical result when $M$ is a $1$ -dimensional interval $[a,b]$ ?

References

John M. Lee (2012). Introduction to Smooth Manifolds
Victor Guillemin, Alan Pollack (1974). Differential Topology
F. Bullo, R. M. Murray (1999). Riemannian Manifolds in Robot Motion Planning and Control