Nash's Embedding Theorems: Unveiling the Intrinsic Nature of Manifolds — Epoche B2
Nash's Embedding Theorems: Unveiling the Intrinsic Nature of Manifolds Riemann's definition of a curved space asks for no surrounding space at all. A metric is written down on a set of coordinate patches, lengths and angles are computed from it, and the geometry is complete. Whether that geometry can also be realised as a subset of some Euclidean space is a separate question, and for eighty years nobody knew the answer in general. Nash settled it twice [1] , in 1954 and 1956, and the two answers disagree in a way that is easy to garble: the rougher the map one is willing to accept, the less room one needs. This essay builds the machinery required to state both results, computes the dimension count that explains where the large numbers come from, and shows on one explicit example why a surface that fits in four dimensions smoothly can also be forced into three if one gives up curvature. The metric is the geometry An $n$-dimensional manifold $M$ is a space in which every point has a neighbourhood described by $n$ coordinates $(x^1,\dots,x^n)$, glued together smoothly. A Riemannian metric $g$ on $M$ assigns to each point a symmetric positive-definite array $g_{ij}$ of numbers, one for each pair of coordinate directions, and thereby fixes the squared length of an infinitesimal displacement $dx^i$: $$ ds^2=\sum_{i=1}^{n}\sum_{j=1}^{n} g_{ij}(x)\,dx^i\,dx^j . $$ Here $ds$ is the length of the displacement, $g_{ij}(x)$ are the metric components at the point with coordinates $x$, and positive-definite means $ds^2\gt 0$ for every non-zero displacement. Symmetry $g_{ij}=g_{ji}$ means only $n(n+1)/2$ of the $n^2$ components are independent — a number that will govern everything below. The length of a curve is the integral of $ds$ along it, so every metric quantity follows from $g_{ij}$ alone. A concrete case: the sphere of radius $a$, in colatitude $\theta$ and longitude $\varphi$, carries $$ ds^2=a^2\,d\theta^2+a^2\sin^2\!\theta\;d\varphi^2,\qquad g_{\theta\theta}=a^2,\quad g_{\varphi\varphi}=a^2\sin^2\!\theta,\quad g_{\theta\varphi}=0 . $$ Nothing in these three numbers mentions a surrounding space. An observer confined to the surface can still detect the curvature, because the metric fixes the angle sum of a triangle whose sides are shortest paths. On a sphere of radius $a$ a geodesic triangle of area $A$ has angle sum $\alpha+\beta+\gamma=\pi+A/a^2$. Take the Earth, $a=6371\,$km, and a triangle covering $10^{6}\,\text{km}^2$ — about the area of Egypt. The excess is $10^{6}/(6371)^{2}=0.0246$ radians, or $1.41^\circ$: small, but a theodolite of arc-second precision would see it without ever leaving the ground. That is what intrinsic means. By contrast, the statement that the surface bulges outwards into a third dimension is extrinsic : it describes a placement, not the geometry. What an isometric embedding is required to solve An embedding of $M$ into $\mathbb{R}^N$ is a smooth injective map $\phi$ whose inverse on the image is also smooth, so that $M$ is laid into $\mathbb{R}^N$ without self-intersections or creases. It is isometric when the Euclidean lengths of curves in the image equal their lengths measured by $g$ [2] . Writing the map in components as $\phi=(\phi^1,\dots,\phi^N)$, and using the fact that the Euclidean metric is the identity array, this condition is the system of partial differential equations $$ \sum_{A=1}^{N}\frac{\partial\phi^{A}}{\partial x^{i}}\,\frac{\partial\phi^{A}}{\partial x^{j}}\;=\;g_{ij}(x),\qquad 1\le i\le j\le n , $$ in which $\phi^{A}$ is the $A$-th Euclidean coordinate of the image point and the sum runs over all $N$ of them. The left-hand side is the dot product of two tangent vectors of the image; the right-hand side is their inner product as measured inside $M$. Requiring the two to agree for every pair of coordinate directions is exactly the requirement that no length is distorted. Now count. Because $g_{ij}$ is symmetric, the system contains $n(n+1)/2$ distinct equations at each point, while the unknowns are the $N$ functions $\phi^{A}$. The system is therefore determined when $$ N=\frac{n(n+1)}{2}\;:\qquad n=2\Rightarrow N=3,\quad n=3\Rightarrow N=6,\quad n=4\Rightarrow N=10, $$ and overdetermined below that. The count proves nothing by itself — flat $\mathbb{R}^n$ sits isometrically inside $\mathbb{R}^n$, far below the line — but it says which side of the ledger a general metric falls on, and it explains the familiar case: for surfaces $n(n+1)/2=3$ coincides with $n+1$, which is why the intuition that a surface belongs in $\mathbb{R}^3$ is not foolish. From $n=3$ upwards the two part company, and the required dimension climbs like $n^2/2$ rather than $n$. This is not a defect of any particular construction; it is the shape of the equations. Janet and Cartan turned the count into a theorem in the 1920s [3] : a real-analytic metric can always be embedded locally and analytically in $\mathbb{R}^{n(n+1)/2}$. Local, analytic, and no more — the global problem stayed open. Where three dimensions genuinely fail: the flat torus For surfaces there is a sharp obstruction, and it comes from Gauss. Given a surface sitting in $\mathbb{R}^3$, its principal curvatures $\kappa_1,\kappa_2$ at a point are the largest and smallest rates at which the surface bends away from its tangent plane; their product $K=\kappa_1\kappa_2$ is the Gaussian curvature . Both factors are extrinsic — they describe bending in the ambient space — yet the Theorema Egregium states that $K$ can be computed from $g_{ij}$ and its first two derivatives alone. For the sphere of radius $a$ both principal curvatures equal $1/a$, so $K=1/a^2$, which is precisely the coefficient appearing in the angle excess above. The obstruction follows in three lines. Let $S$ be a compact surface in $\mathbb{R}^3$, and let $R$ be the radius of the smallest sphere centred at the origin that contains $S$. At a point $p$ where $S$ touches that sphere, the surface curves at least as sharply as the sphere in every direct