Smooth Manifolds and Charts

Smooth Manifolds: Geometry Without an Ambient Space

Why Go Beyond Surfaces?

A surface in $\mathbb{R}^3$ is an intuitive object: we “see” it from the outside. But many natural geometric objects are not embedded in familiar space. The phase space of a mechanical system with $n$ degrees of freedom is a $2n$-dimensional manifold. The space of quantum states is an infinite-dimensional Hilbert space. The configuration space of a robotic arm is the manifold $SO(3) \times \mathbb{R}^3$.

A smooth manifold is a way to work with such spaces without referencing an external embedding.

Topological Manifold

Definition: A Hausdorff topological space $M$ in which every point has a neighborhood homeomorphic (topologically equivalent) to an open ball in $\mathbb{R}^n$ is called an $n$-dimensional topological manifold.

Simply put: a manifold is “locally flat”—near every point it looks like a piece of $\mathbb{R}^n$. A surface is “flat” like $\mathbb{R}^2$, space is “flat” like $\mathbb{R}^3$, although globally it may be more complicated.

A chart $(u, \varphi)$: an open set $u \subset M$ with a homeomorphism $\varphi: u \to \mathbb{R}^n$. This is a “coordinate system” on $u$. An atlas is a collection of charts covering all of $M$.

Transition maps: when two charts $(u_\alpha, \varphi_\alpha)$ and $(u_\beta, \varphi_\beta)$ overlap: $\psi_{\alpha\beta} = \varphi_\beta \circ \varphi_\alpha^{-1} : \varphi_\alpha(u_\alpha \cap u_\beta) \to \varphi_\beta(u_\alpha \cap u_\beta)$—a map between subsets of $\mathbb{R}^n$.

Smooth Manifold

An atlas is smooth if all transition functions $\psi_{\alpha\beta}$ are smooth ($C^\infty$) maps. A smooth manifold is a pair $(M, [A])$, where $[A]$ is a maximal smooth atlas (an equivalence class of compatible atlases).

Examples:

• $\mathbb{R}^n$: one chart, the identity atlas.
• Sphere $S^n$: two stereographic projection atlases. North-atlas: $\varphi_N(x_1, ..., x_{n+1}) = (x_1, ..., x_n) / (1 - x_{n+1})$. South-atlas: $\varphi_S$ is similar with $+x_{n+1}$. Transition map: $\varphi_S \circ \varphi_N^{-1}(u) = u/|u|^2$ (inversion)—smooth for $u \neq 0$ ✓.
• Torus $T^2 = \mathbb{R}^2/\mathbb{Z}^2$: quotient space. Atlas consists of four charts (with overlaps at the edges). A compact two-dimensional manifold.
• Projective space $\mathbb{R}P^n$: sphere $S^n$ with identification of antipodes $x \sim -x$. $\mathbb{R}P^2$ is the first example of a “non-orientable” closed manifold.

Smooth Maps and Diffeomorphisms

A function $f: M \to \mathbb{R}$ is smooth if for every chart $(u, \varphi)$ the function $f \circ \varphi^{-1}: \mathbb{R}^n \to \mathbb{R}$ is smooth.

A map $F: M \to N$ between manifolds is smooth if in local coordinates it is expressed by a smooth function: $\psi \circ F \circ \varphi^{-1}$ is smooth.

Diffeomorphism: A smooth bijection with a smooth inverse. This is an “equivalence” in the category of smooth manifolds. Diffeomorphic manifolds are geometrically indistinguishable.

Concrete Example: SO(3) and Rotations

The group $SO(3)$ is the set of all orthogonal matrices with $\det = 1$ (rotations in $\mathbb{R}^3$). This is a 3-dimensional manifold (compact, connected). Parameterizations: Euler angles (3 angles), axis–angle (axis + angle = 4 parameters with normalization).

$SO(3)$ is not the sphere $S^3$, but admits a double cover: $SU(2) \cong S^3 \to SO(3)$ (2-to-1). Unit quaternions ($|q| = 1$) $\cong S^3$—each rotation corresponds to two antipodal quaternions $q$ and $-q$.

In robotics: the configuration of a manipulator arm with 6 degrees of freedom is an element of $SO(3) \times \mathbb{R}^3$ (a 6-dimensional manifold). Control algorithms use the geometry of this manifold.

In quantum mechanics: spinor representations (half-integer angular momentum) are representations of $SU(2)$, not $SO(3)$. The electron’s spin-$\tfrac{1}{2}$ is a consequence of the double covering.

Manifolds in Machine Learning

The manifold hypothesis states: real high-dimensional data (images, audio, text) lie near a low-dimensional manifold in feature space. For example, the set of face images is not a “uniform” subspace of $\mathbb{R}^{1000000}$, but a manifold of dimension $\sim 50$ (pose, illumination, expression). UMAP and t-SNE algorithms build “maps” of this manifold, using ideas of preserving topology and metric. Autoencoders in neural networks essentially learn a parametrization of the latent data manifold—an “encoder” sets a chart, the “decoder” its inverse. Principal Component Analysis (PCA) seeks linear subspaces; nonlinear analogues (UMAP, diffusion maps) build diffeomorphisms between the data manifold and a flat low-dimensional space. Differential geometry provides the mathematical language to describe these structures.

Manifolds in Physics and Engineering

The concept of a smooth manifold permeates theoretical physics. The phase space of a mechanical system with $n$ degrees of freedom is a $2n$-dimensional symplectic manifold: $n$ generalized coordinates and $n$ momenta. The system's trajectory is an integral curve of a Hamiltonian vector field on this manifold. In relativity, spacetime is a four-dimensional Lorentzian manifold with an indefinite metric, and free fall is geodesic motion on it. In gauge field theory, which describes electromagnetism and nuclear forces, fields are sections of fiber bundles over spacetime—an extension of manifolds with additional structure. In engineering, the orientation of a rigid body is described by a point of the $SO(3)$ group—a three-dimensional compact Lie manifold. Control algorithms for quadcopters and satellites work directly on this manifold: equations of motion are integrated using the exponential map of the Lie group, allowing one to exactly preserve the geometry of rotations and avoid “gimbal lock”, characteristic of Euler angle parameterization.

Smooth Manifolds in Data Analysis and Machine Learning

Manifold-based data analysis has become one of the key areas of modern machine learning. The Isomap algorithm builds a low-dimensional representation of data preserving geodesic distances: instead of Euclidean distances, it uses the lengths of shortest paths in a nearest-neighbor graph approximating the geodesic distances on the manifold. This makes it possible to discover nonlinear structures in high-dimensional data. In face recognition, the set of images of one person under different lighting angles lies on a low-dimensional manifold in pixel space, and manifold learning methods reveal this manifold. In neuroscience, the geometry of the space of neural activations is described as a manifold: UMAP and t-SNE construct its low-dimensional embedding, revealing the structure of neural codes. The geometry of the Riemannian manifold of positive-definite matrices (covariance matrices) is used in computer vision for classification tasks that account for the structure of the space. Differential geometry provides the language for normalizing flows—a class of deep neural networks that transform arbitrary probability distributions via a sequence of diffeomorphisms.

Exercise: (a) Show that stereographic projection $\varphi_N: S^2 \setminus {N} \to \mathbb{R}^2$ is conformal (angle-preserving). (b) Why are the torus $T^2$ and sphere $S^2$ not diffeomorphic? Name a topological invariant distinguishing them. (c) $SO(2) \cong S^1$: prove this via an explicit parameterization (rotation matrix $= \exp(\theta J)$, $J = \begin{bmatrix} 0 & -1 \ 1 & 0 \end{bmatrix}$).