Atlases, Transition Functions and Differentiable Structure

A manifold does not arrive with coordinates stamped on it. General relativity insists that spacetime is a smooth four-dimensional manifold, yet the raw object underneath that phrase is only a topological space — a set of points glued together with a notion of nearness, locally indistinguishable from $\mathbb{R}^n$ but carrying no rulers, no axes, and, crucially, no calculus. You cannot differentiate a function on a bare topological space, because differentiation is a limiting comparison of nearby values and the space offers no arithmetic to take the limit in. Every derivative in physics — a velocity, a field gradient, the Riemann curvature — presupposes machinery that a topological manifold simply lacks. The atlas is the device that supplies it. A chart photographs a patch of the manifold as an open region of $\mathbb{R}^n$ ; a collection of charts covering the whole manifold is an atlas; and wherever two charts overlap they hand the same points two coordinate descriptions, related by a transition function that lives entirely inside Euclidean space. Because that overlap map is a map from open sets of $\mathbb{R}^n$ to open sets of $\mathbb{R}^n$ , ordinary multivariable calculus can interrogate it: is it continuous, is it $C^k$ , is it $C^\infty$ ? Declaring every transition function smooth promotes the atlas to a smooth atlas, and this one demand is what lets the word "smooth" mean anything for a function on the manifold itself. The payoff cascades through the whole subject. The maximal smooth atlas is the differentiable structure — genuine data, not a topological accident, as the exotic spheres attest. The Jacobians of the transition functions turn out to be the exact matrices in the tensor transformation law, so this bookkeeping is the birthplace of general covariance: physics written as tensor equations holds in every chart because both sides transform the same way under the transition Jacobians.

Key takeaways

A chart is a homeomorphism $\varphi:U\to\varphi(U)\subset\mathbb{R}^n$ that assigns $n$ coordinates to an open patch $U$ of the manifold $M$ ; an atlas is a family of charts whose domains cover $M$ .
On any overlap the transition function $\psi_{\beta\alpha}=\varphi_\beta\circ\varphi_\alpha^{-1}$ is a map between open subsets of $\mathbb{R}^n$ ; requiring every such map to be $C^\infty$ is the definition of a smooth atlas.
Smoothness of a function $f:M\to\mathbb{R}$ is defined by pulling it back through charts, $f\circ\varphi_\alpha^{-1}\in C^\infty$ ; the chain rule and the smooth transition maps make this test chart-independent.
The maximal smooth atlas is the differentiable structure — it is what upgrades a topological manifold to a smooth manifold, and it is extra input: exotic $\mathbb{R}^4$ and Milnor's 28 smooth 7-spheres show the same topology can carry inequivalent smooth structures.
The transition Jacobian $\partial x'^\mu/\partial x^\nu$ is invertible ( $\det\neq0$ ) and is precisely the matrix that appears in the tensor transformation law, making tensor equations coordinate-independent.
Key idea: calculus on a manifold is defined by pulling everything back to $\mathbb{R}^n$ through charts; the transition functions, being smooth maps of Euclidean space, are what make the definition independent of the chart, and their Jacobians are the universal transformation matrices for every tensor.

The core insight

Calculus on a manifold is defined by pulling everything back to $\mathbb{R}^n$ through charts; the transition functions, being smooth maps of Euclidean space, are what make that definition independent of the chosen chart, and their Jacobians are the universal transformation matrices for all tensors.

Where this lives today

The atlas is not a museum piece — it is running code. Numerical relativity, the field that computes gravitational waveforms for LIGO and Virgo, is atlas engineering at industrial scale. A single coordinate chart cannot cleanly cover an inspiralling black-hole binary: near each horizon the natural coordinates go bad, and far away a wave-friendly grid is wanted. So modern codes use multiple overlapping patches with transition maps between them — multiblock and multipatch methods in the Einstein Toolkit, and spectral multi-domain decompositions in the SpEC and SpECTRE codes — precisely the abstract atlas made computational, with the numerical transition operators playing the role of $\psi_{\beta\alpha}$ . The moving-puncture and horizon-excision techniques that made binary-black-hole simulations tractable in 2005 are, at heart, careful choices of chart that keep the transition functions smooth where the physics is violent. Meanwhile the purely mathematical frontier the portrait opened remains wide open: it is still unknown whether the smooth Poincaré conjecture holds in four dimensions — whether $S^4$ admits any exotic smooth structure at all — and $\mathbb{R}^4$ is the one Euclidean space known to carry uncountably many. Since spacetime is four-dimensional, the question of which differentiable structures nature could in principle use is not fully settled, a reminder that the atlas is genuine physical input.

The physical picture

Think of an atlas the way a cartographer thinks of a world map. No single flat sheet can faithfully depict the round Earth, so the mapmaker uses overlapping sheets — a Paris sheet, a Europe sheet, a world sheet — and prints, in the margins, the rules that say "this street on the Paris page is that grid square on the Europe page." Those marginal rules are the transition functions. As long as they vary smoothly, you can plan a route that runs off one sheet and onto the next without a kink at the seam: the joined-up description is as good as if a single impossible flat map had existed all along. The manifold plays the Earth; the charts play the sheets. And the reason we go through this trouble is that calculus only knows how to operate in $\mathbb{R}^n$ . The manifold itself is too abstract to differentiate on directly — it has points and neighbourhoods but no coordinates to form difference quotients with. So we borrow Euclidean calculus through a chart, do the computation there, and then check that the answer would have been the same through any other chart covering the same region. The transition function is that consistency check made precise, and the chain rule is the theorem that guarantees the check passes: a property verified in one chart automatically holds in every compatible one.

Figure

Two charts φ_α and φ_β cover overlapping patches of the manifold M and carry them down into ℝⁿ; the transition function ψ_βα = φ_β ∘ φ_α⁻¹ relabels one chart's coordinates as the other's and lives entirely in Euclidean space, where ordinary calculus decides whether it is C∞

The differentiable structure that emerges is genuinely extra data, not something the topology hands you for free. The same "shape" — the same topological manifold — can support inequivalent notions of smoothness. Milnor's exotic 7-spheres and the uncountable family of exotic $\mathbb{R}^4$ 's make this vivid: homeomorphic spaces that are not diffeomorphic, identical as rubber sheets but distinct as arenas for calculus. For the manifolds of physical spacetime the standard structure is the obvious and universal choice, but the mere existence of alternatives tells us that choosing the smooth atlas is a modeling decision, not a logical inevitability. The same lesson has a very practical face in GR, and it is the second figure's subject: features that look like catastrophes in one chart can be pure artifacts of that chart. The Schwarzschild radial metric component diverges at the horizon $r=2M$ , which for decades was misread as the edge of the world. It is nothing of the sort — a change of atlas sails smoothly across it — whereas the divergence at $r=0$ survives every chart because a coordinate-independent invariant blows up there. The atlas viewpoint is exactly what lets us tell the two apart.

Figure

The Schwarzschild radial metric component g_rr = 1 divided by (1 − 2M/r) blows up at r = 2M, yet this is a chart artifact: the Kretschmann invariant K = 48M²/r⁶ equals 0.75/M⁴ there, perfectly finite, and diverges only at the genuine curvature singularity r = 0

The core equations

A chart is the basic instrument. It selects an open patch $U_\alpha$ of the manifold and photographs it as an open region of $\mathbb{R}^n$ , assigning $n$ real numbers to every point the patch contains. Nothing about the manifold is altered by this — the chart is a reversible labelling we can compute with, and the requirement that it be a homeomorphism keeps the labelling faithful to the topology.

\varphi_\alpha:U_\alpha\subset M\ \longrightarrow\ \varphi_\alpha(U_\alpha)\subset\mathbb{R}^n \qquad (1)

where: $M$ — the $n$ -dimensional topological manifold; $U_\alpha$ — an open coordinate patch (the chart's domain); $\varphi_\alpha$ — a homeomorphism onto an open subset of $\mathbb{R}^n$ ; $n$ — the dimension ( $n=4$ for spacetime). A single chart rarely covers all of $M$ : no one chart covers a sphere. Because one chart is not enough, we collect many into an atlas whose domains together blanket the manifold. The atlas is the object that actually carries the calculus; a lone chart is only one page of it.

\mathcal{A}=\{(U_\alpha,\varphi_\alpha)\}_{\alpha\in I},\qquad \bigcup_{\alpha\in I}U_\alpha=M \qquad (2)

where: $\mathcal{A}$ — the atlas, a covering collection of charts; $I$ — the index set labelling them; the union condition says every point of $M$ lies in at least one chart. Wherever two charts overlap, a point receives two coordinate names, and the dictionary translating one into the other is the transition function. Its decisive property is that it never mentions $M$ : it maps an open set of $\mathbb{R}^n$ to an open set of $\mathbb{R}^n$ , so ordinary calculus applies to it verbatim.

\psi_{\beta\alpha}=\varphi_\beta\circ\varphi_\alpha^{-1}:\ \varphi_\alpha(U_\alpha\cap U_\beta)\longrightarrow\varphi_\beta(U_\alpha\cap U_\beta) \qquad (3)

where: $\psi_{\beta\alpha}$ — the overlap or change-of-coordinate map; $U_\alpha\cap U_\beta$ — the region where both charts apply; the domain and codomain are open subsets of Euclidean space. The single condition that makes an atlas smooth is that every one of these overlap maps is infinitely differentiable. This is where "smooth" enters the story — not through the charts individually, which have nothing on $M$ to be differentiated against, but through the maps between them.

\psi_{\beta\alpha}\in C^\infty\quad\text{for all }\alpha,\beta\ \Longleftrightarrow\ \mathcal{A}\ \text{is a smooth atlas} \qquad (4)

where: $C^\infty$ — infinitely differentiable in the ordinary multivariable sense; the biconditional is a definition, not a theorem — a smooth atlas is exactly one whose transition functions all pass this test. With a smooth atlas fixed, a function on the manifold is declared smooth if its coordinate representative is smooth in every chart. This is the pullback definition of smoothness, and its whole content is that the answer does not depend on which chart you use to compute it.

f:M\to\mathbb{R}\ \text{smooth}\iff f\circ\varphi_\alpha^{-1}\in C^\infty\big(\varphi_\alpha(U_\alpha)\big)\ \ \forall\alpha \qquad (5)

where: $f\circ\varphi_\alpha^{-1}$ — the coordinate representative of $f$ , an honest function on an open set of $\mathbb{R}^n$ ; the "for all $\alpha$ " is what Derivation 1 shows is automatic once it holds in one overlapping chart. The derivative of a transition function is its Jacobian, and it is the protagonist of every coordinate change in GR. Because a transition function is a diffeomorphism — smooth with a smooth inverse — its Jacobian is invertible everywhere on the overlap.

J^{\mu}{}_{\nu}=\frac{\partial x'^{\mu}}{\partial x^{\nu}}=\frac{\partial(\psi_{\beta\alpha})^{\mu}}{\partial x^{\nu}},\qquad \det J\neq 0 \qquad (6)

where: $x^{\nu}$ — coordinates in chart $\varphi_\alpha$ ; $x'^{\mu}$ — coordinates in chart $\varphi_\beta$ ; $J^{\mu}{}_{\nu}$ — the transition Jacobian; $\det J\neq0$ by the inverse-function theorem, since $\psi_{\beta\alpha}$ is a diffeomorphism. That same Jacobian, and its inverse, are exactly the matrices in the tensor transformation law. A tensor is, operationally, an object whose components transform this way under a change of chart — which is why tensor equations are chart-independent statements about geometry.

T'^{\mu_1\cdots}{}_{\nu_1\cdots}=\frac{\partial x'^{\mu_1}}{\partial x^{\rho_1}}\cdots\frac{\partial x^{\sigma_1}}{\partial x'^{\nu_1}}\cdots\,T^{\rho_1\cdots}{}_{\sigma_1\cdots} \qquad (7)

where: $T^{\rho\cdots}{}_{\sigma\cdots}$ — the tensor's components in chart $\varphi_\alpha$ ; $T'$ — its components in chart $\varphi_\beta$ ; each upper index carries a factor of $J$ and each lower index a factor of $J^{-1}$ . Finally, the differentiable structure is the maximal smooth atlas: the union of all charts compatible with a given smooth atlas. Fixing it is precisely what turns a topological manifold into a smooth manifold, and Derivation 2 proves it exists and is unique.

\mathcal{A}_{\max}=\bigcup\{\mathcal{A}'\ :\ \mathcal{A}'\ \text{compatible with}\ \mathcal{A}\} \qquad (8)

where: $\mathcal{A}_{\max}$ — the unique maximal smooth atlas containing $\mathcal{A}$ , i.e. the differentiable structure; "compatible" means the union of the two atlases is itself a smooth atlas.

Derivation AWhy the pullback definition of smoothness is chart-independent

The goal is to justify the definition that quietly underpins all of differential geometry: that a function $f:M\to\mathbb{R}$ counts as smooth if its coordinate representative $f\circ\varphi_\alpha^{-1}$ is smooth. That would be worthless if the verdict flipped from chart to chart — one observer calling $f$ smooth and another calling it kinked at the same point. We must show the property is intrinsic to $f$ and to the smooth structure, not to the accident of which chart we picked. The proof is short and rests entirely on two facts: transition functions are $C^\infty$ , and a composition of $C^\infty$ maps is $C^\infty$ .

Step 1. Start from a point $p$ in the overlap $U_\alpha\cap U_\beta$ and suppose $f\circ\varphi_\alpha^{-1}$ is smooth near $\varphi_\alpha(p)$ ; we want the same for the $\beta$ -representative. Insert the identity $\varphi_\alpha^{-1}\circ\varphi_\alpha$ into $f\circ\varphi_\beta^{-1}$ and regroup: $f\circ\varphi_\beta^{-1}=\big(f\circ\varphi_\alpha^{-1}\big)\circ\big(\varphi_\alpha\circ\varphi_\beta^{-1}\big)$ . Why: this is the standard trick of routing the computation through the chart we already control; nothing has changed except the grouping of the composition.

Step 2. Recognize the second factor. The map $\varphi_\alpha\circ\varphi_\beta^{-1}=\psi_{\alpha\beta}$ is a transition function of the smooth atlas, hence $C^\infty$ by the atlas condition (4). Why: this is the only place the smooth-atlas hypothesis is used — the whole point of demanding smooth overlap maps is to make exactly this factor differentiable.

Step 3. Compose. The first factor $f\circ\varphi_\alpha^{-1}$ is smooth by assumption and the second is smooth by Step 2, so their composition $f\circ\varphi_\beta^{-1}$ is smooth by the chain rule. Since $\beta$ was an arbitrary chart around $p$ , smoothness of $f$ does not depend on the chart, and the definition (5) is well posed. Why: the chain rule is doing the real work — it transplants the Euclidean notion of smoothness onto $M$ consistently, which is the sole reason the whole edifice of manifold calculus holds together.

Go deeper · Derivation B · 2 minSmooth compatibility is an equivalence relation with a unique maximal atlas

Derivation BSmooth compatibility is an equivalence relation with a unique maximal atlas

The goal is to pin down the object physicists call the differentiable structure. We have many smooth atlases for a given manifold — infinitely many, since we can always add a chart or shrink a domain — and we want to know that they organize into classes, with each class containing one canonical, all-encompassing representative. Concretely, call two smooth atlases $\mathcal{A}$ and $\mathcal{B}$ compatible when $\mathcal{A}\cup\mathcal{B}$ is again a smooth atlas, meaning every cross transition $\varphi_b\circ\varphi_a^{-1}$ with one chart drawn from each is $C^\infty$ . We show compatibility is reflexive, symmetric and transitive, and then build the maximal atlas explicitly.

Step 1. Reflexivity: $\mathcal{A}\sim\mathcal{A}$ . By hypothesis $\mathcal{A}$ is already a smooth atlas, so all of its internal transition functions are $C^\infty$ and $\mathcal{A}\cup\mathcal{A}=\mathcal{A}$ is smooth. Why: an atlas must be compatible with itself or the relation could never partition anything.

Step 2. Symmetry: if $\psi_{ab}\in C^\infty$ then $\psi_{ba}=\psi_{ab}^{-1}\in C^\infty$ . A transition function is a diffeomorphism, so its inverse is smooth, and $\mathcal{A}\cup\mathcal{B}=\mathcal{B}\cup\mathcal{A}$ . Why: this is where "diffeomorphism", not merely "smooth map", earns its keep — the inverse must also be smooth for the relation to be symmetric.

Step 3. Transitivity: if $\mathcal{A}\sim\mathcal{B}$ and $\mathcal{B}\sim\mathcal{C}$ , then on a triple overlap any cross transition factors as $\varphi_c\circ\varphi_a^{-1}=\big(\varphi_c\circ\varphi_b^{-1}\big)\circ\big(\varphi_b\circ\varphi_a^{-1}\big)$ , a composition of two $C^\infty$ maps and hence $C^\infty$ . Because $\mathcal{B}$ covers $M$ , every point of the $\mathcal{A}$ – $\mathcal{C}$ overlap lies in some chart of $\mathcal{B}$ , so the factorization is available everywhere. Why: transitivity is what lets us speak of an equivalence class at all, and the covering property of the intermediate atlas is the fine print that makes it airtight.

Step 4. Build the maximal atlas: $\mathcal{A}_{\max}=\{(U,\varphi):\varphi\ \text{is smoothly compatible with every chart of}\ \mathcal{A}\}$ . Given two members $(U,\varphi)$ and $(V,\chi)$ , write $\chi\circ\varphi^{-1}$ near any point as $\big(\chi\circ\varphi_a^{-1}\big)\circ\big(\varphi_a\circ\varphi^{-1}\big)$ using a chart $\varphi_a\in\mathcal{A}$ that covers the point; both factors are $C^\infty$ by membership, so $\mathcal{A}_{\max}$ is a smooth atlas. It contains every chart compatible with $\mathcal{A}$ , so nothing can be added — it is maximal — and any maximal atlas containing $\mathcal{A}$ must contain exactly these charts, so it is unique. Why: the boxed conclusion below is the definition of the differentiable structure, and its uniqueness is what makes "the smooth structure" a meaningful phrase rather than a choice among many.

\mathcal{A}\sim\mathcal{B}\ \Longrightarrow\ \exists!\ \mathcal{A}_{\max}\supseteq\mathcal{A}\ \ \text{maximal smooth (the differentiable structure)}

Go deeper · Derivation C · 2 minThe transition Jacobian is the transformation matrix of a tangent vector

Derivation CThe transition Jacobian is the transformation matrix of a tangent vector

The goal is to demonstrate that the atlas machinery is not passive bookkeeping but the actual engine of tensor transformations. We take the cleanest object built on the smooth structure — a tangent vector, defined as a directional-derivative operator on smooth functions — and show that its components change between charts by exactly the transition Jacobian of equation (6). Once this is established for a vector, the one-form law (with the inverse Jacobian) and the full tensor law (7) follow by taking tensor products, so the whole transformation calculus is downstream of the transition functions. Let $x^\nu$ and $x'^\mu$ be two charts with coordinate bases $\partial_\nu=\partial/\partial x^\nu$ and $\partial'_\mu=\partial/\partial x'^\mu$ .

Step 1. Expand the same abstract vector in both bases: $V=V^\nu\partial_\nu=V'^\mu\partial'_\mu$ . The vector $V$ is a chart-independent geometric object; only its components carry a chart label. Why: the equality of the two expansions is the entire physical input — the arrow is real, its numerical description is convention — and everything else is algebra to extract how the numbers must change.

Step 2. Relate the two coordinate bases by the chain rule. Acting on any smooth $f$ , $\partial_\nu f=\dfrac{\partial x'^\mu}{\partial x^\nu}\dfrac{\partial f}{\partial x'^\mu}$ , so as operators $\partial_\nu=\dfrac{\partial x'^\mu}{\partial x^\nu}\,\partial'_\mu$ . The coefficient $\partial x'^\mu/\partial x^\nu$ is precisely the transition Jacobian $J^\mu{}_\nu$ of the map $x'(x)=\psi_{\beta\alpha}$ . Why: this is the crux — the multivariable chain rule applied to the transition function is what injects the Jacobian into the transformation law; no new physics, just calculus on the overlap.

Step 3. Substitute Step 2 into $V=V^\nu\partial_\nu$ to get $V=V^\nu\dfrac{\partial x'^\mu}{\partial x^\nu}\partial'_\mu$ , and compare with $V=V'^\mu\partial'_\mu$ . Because $\{\partial'_\mu\}$ is a basis, the coefficients of each $\partial'_\mu$ must match. Why: linear independence of the basis is what turns an equality of vectors into an equality of components — the standard move that lets us "read off" a transformation law.

Step 4. Read off the contravariant law: $V'^\mu=\dfrac{\partial x'^\mu}{\partial x^\nu}V^\nu=J^\mu{}_\nu V^\nu$ . A one-form transforms with the inverse $\partial x^\nu/\partial x'^\mu$ so that the contraction $\omega_\mu V^\mu$ is chart-independent, and a general tensor transforms with one Jacobian per upper index and one inverse per lower index — equation (7). Why: this closes the loop back to the overview's claim that the transition Jacobians are the universal transformation matrices, and it is invertibility of $J$ (guaranteed by smoothness of the atlas) that makes the covariant half of the law well defined.

V'^{\mu}=\frac{\partial x'^{\mu}}{\partial x^{\nu}}\,V^{\nu}\qquad(\text{contravariant law})

Limiting cases

The regularity we demand of the transition functions is a dial, and where we set it decides what kind of manifold — and what kind of physics — we get.

Case	Result	Meaning
$C^0$ atlas	$\varphi_\beta\circ\varphi_\alpha^{-1}$ merely continuous	Only a topological manifold — no derivatives, no tangent vectors, no tensors, no metric.
$C^k$ atlas	$\varphi_\beta\circ\varphi_\alpha^{-1}\in C^k$	Calculus up to order $k$ is well defined; enough for many PDE arguments but caps how many curvature derivatives exist.
$C^\infty$ (smooth) atlas	$\varphi_\beta\circ\varphi_\alpha^{-1}\in C^\infty$	The standard GR setting — arbitrarily high derivatives, so metric, connection and curvature are all defined.
$C^\omega$ (analytic) atlas	$\varphi_\beta\circ\varphi_\alpha^{-1}$ real-analytic	Stronger still — solutions fixed by their germ at a point; powers rigidity theorems but is too rigid for generic spacetimes.

Common trapA coordinate singularity is a real singularity of spacetime

This is the single most common atlas-related error in GR, and it has a precise cure. A divergence in a metric component is a property of the chart, not necessarily of the manifold. In Schwarzschild coordinates $g_{rr}=1/(1-2M/r)$ blows up at $r=2M$ , which looks like the edge of spacetime — but transforming to Eddington–Finkelstein or Kruskal–Szekeres coordinates crosses $r=2M$ perfectly smoothly, and the geometry there is regular. The horizon is a coordinate artifact. The genuine article at $r=0$ is diagnosed instead by a chart-independent invariant: the Kretschmann scalar $K=R_{abcd}R^{abcd}=48M^2/r^6$ , which is finite at the horizon ( $K=0.75/M^4$ there) but diverges as $r\to0$ . The rule to carry away: before calling any "singularity" physical, ask whether a scalar invariant diverges, or whether a change of atlas makes it disappear. Charts are labels; only invariants are physics.

Worked examples

1 · The two-chart stereographic atlas on the sphere (a fully worked construction)

The 2-sphere $S^2$ cannot be covered by a single chart, yet it is a smooth manifold. Build the minimal smooth atlas from two stereographic projections, compute their transition function explicitly, and verify that it is smooth with an invertible Jacobian at a concrete point.

Step 1. Why one chart cannot suffice. $S^2$ is compact and has no boundary, while any single chart image is an open subset of $\mathbb{R}^2$ , which if it were all of $S^2$ would have to be both compact and open in $\mathbb{R}^2$ — impossible for a nonempty proper subset. So at least two charts are needed; stereographic projection provides exactly two.

Step 2. The two charts. Let $\varphi_N$ project from the north pole onto the equatorial plane, covering $S^2$ minus the north pole, and $\varphi_S$ project from the south pole, covering $S^2$ minus the south pole. Writing the plane coordinates as a complex number, $\varphi_N$ sends a point to $\zeta=u+iv$ and $\varphi_S$ sends it to $w$ . Their domains overlap on $S^2$ minus both poles, which maps to the punctured plane $\mathbb{R}^2\setminus\{0\}$ .

Step 3. The transition function. With orientations chosen so both charts are positively oriented, the overlap map is the inversion $w=\psi(\zeta)=1/\zeta$ , or in real components $(u,v)\mapsto(u,-v)/(u^2+v^2)$ . This is manifestly $C^\infty$ — indeed real-analytic — on the punctured plane, since $u^2+v^2>0$ there. The atlas $\{(S^2\setminus N,\varphi_N),(S^2\setminus S,\varphi_S)\}$ with transition $w=1/\zeta$ is a smooth atlas, so $S^2$ is a smooth manifold.

Step 4. Check the Jacobian at a point. Take $\zeta=2$ , i.e. $(u,v)=(2,0)$ ; then $w=1/2$ , i.e. $(0.5,0)$ . For the holomorphic map $w=1/\zeta$ the Jacobian determinant equals $|\mathrm{d}w/\mathrm{d}\zeta|^2=|-1/\zeta^2|^2=1/|\zeta|^4$ . At $\zeta=2$ this is $1/16=\mathbf{0.0625}\neq0$ , so the transition is a local diffeomorphism there, exactly as the smooth-atlas condition requires. The apparent "singularities" of spherical-polar coordinates at $\theta=0,\pi$ are just points where you must switch to the other stereographic chart, which is perfectly regular there.

2 · Horizon versus true singularity in Schwarzschild spacetime (fully numeric)

In Schwarzschild coordinates $g_{rr}=1/(1-2M/r)$ diverges at $r=2M$ . Decide, using a chart-independent invariant, whether $r=2M$ and $r=0$ are physical singularities. Work in geometrized units where $M$ has units of length; evaluate at $r=2M$ and compare with $r=3M$ and $r=4M$ .

Step 1. The suspect quantity. The metric component $g_{rr}=1/(1-2M/r)$ is infinite at $r=2M$ (as $r\to2M^+$ , $g_{rr}\to+\infty$ : at $r=2.2M$ it is $11$ , at $r=2.5M$ it is $5$ , at $r=4M$ it is $2$ ). Taken at face value this looks like a catastrophe at the horizon.

Step 2. The invariant test. A metric component is chart-dependent and proves nothing. The Kretschmann scalar $K=R_{abcd}R^{abcd}$ is a scalar — identical in every chart of the differentiable structure — and for Schwarzschild it is $K=48M^2/r^6$ . Evaluate it at the horizon: $K(2M)=48M^2/(2M)^6=48M^2/64M^6=\mathbf{0.75/M^4}$ , a finite number. Compared with $K(3M)=48/729\,M^{-4}\approx0.0658/M^4$ and $K(4M)=48/4096\,M^{-4}\approx0.0117/M^4$ , the horizon value is finite and unremarkable.

Step 3. Verdict. Because $K$ stays finite at $r=2M$ , no coordinate-independent quantity diverges there: $r=2M$ is a coordinate singularity, an artifact of the Schwarzschild chart that Kruskal–Szekeres coordinates remove. As $r\to0$ , however, $K=48M^2/r^6\to\infty$ in every chart, so $r=0$ is a true curvature singularity of the manifold. The atlas viewpoint delivers the whole distinction: only what survives a change of chart — a scalar invariant — can be physical.

Physics of the portrait

The abstract, chart-based definition of a manifold that GR takes for granted was not always available. For most of the nineteenth and early twentieth centuries a "surface" or "space" meant something sitting inside a bigger Euclidean space; the idea that a manifold could be defined intrinsically, by an atlas of charts glued with smooth transition functions and nothing else, had to be invented and then proved to lose nothing. Hassler Whitney is the mathematician who sealed that bargain.

Origins · 1 minHassler Whitney and how this was discovered

Hassler Whitney

1907–1989 · American mathematician, founder of differential topology, who put abstract smooth manifolds on a rigorous footing

Hassler Whitney came to mathematics by an unusual road — he studied physics and music at Yale and only later turned to topology at Harvard, where he did foundational work on graph theory and the four-color problem before reshaping geometry. In a landmark 1936 paper in the Annals of Mathematics, "Differentiable Manifolds", he gave the modern definition of a smooth manifold as a set equipped with an atlas of charts whose transition functions are $C^\infty$ , and then proved the theorem that made the abstract definition safe to use: every abstract smooth $n$ -manifold can be embedded as a smooth submanifold of $\mathbb{R}^{2n}$ (later sharpened to $\mathbb{R}^{2n}$ for embeddings and $\mathbb{R}^{2n-1}$ for immersions). The Whitney embedding theorem settled a genuine worry — that intrinsic, chart-defined manifolds might be a strictly larger and stranger class than the concrete surfaces of classical geometry. They are not: the two pictures coincide, so a relativist may define spacetime by its atlas and lose nothing. Whitney also introduced the tools that let differential topology distinguish smooth structures — Whitney classes and the theory of singularities of smooth maps — the very machinery Milnor would later use to discover exotic spheres. His insistence that the transition functions, and not any ambient space, carry all the geometric information is exactly the viewpoint this article is built on.

History

1854

— Riemann's habilitation lecture introduces the notion of a manifold (Mannigfaltigkeit) as a space of many dimensions

1913

— Weyl's Die Idee der Riemannschen Fläche gives the first rigorous atlas-based definition of an abstract surface

1932

— Veblen and Whitehead codify the differentiable manifold via atlases of charts and smooth transition functions

1936

— Whitney proves every abstract smooth manifold embeds in Euclidean space, unifying the intrinsic and embedded views

1963

— Kervaire and Milnor show the 7-sphere carries 28 inequivalent smooth structures, proving the differentiable structure is extra data

Going further

References. Misner, Thorne & Wheeler, Gravitation (1973), §2–§3 · Wald, General Relativity (1984), Ch. 2 · Carroll, Spacetime and Geometry (2004), §2.2 · Schutz, A First Course in General Relativity, Ch. 6 · Lee, Introduction to Smooth Manifolds (2nd ed.), Ch. 1 · Milnor, "On manifolds homeomorphic to the 7-sphere", Ann. Math. 64 (1956) 399 · Whitney, "Differentiable Manifolds", Ann. Math. 37 (1936) 645.

Provenance. The definitions of charts, atlases, transition functions, smooth compatibility and the maximal-atlas theorem follow Lee, Introduction to Smooth Manifolds (Ch. 1) and Wald §2.1. A verifier should reproduce: (i) the equivalence-relation proof for atlas compatibility and the unique maximal atlas (Lee Prop. 1.17); (ii) the chain-rule derivation of the contravariant law $V'^\mu=(\partial x'^\mu/\partial x^\nu)V^\nu$ (Carroll §2.3, Schutz Ch. 5); (iii) the two-chart stereographic atlas on $S^2$ with transition $w=1/\zeta$ and Jacobian determinant $1/|\zeta|^4$ (equal to $0.0625$ at $\zeta=2$ ); (iv) the Kretschmann scalar $K=48M^2/r^6$ , giving $K=0.75/M^4$ at $r=2M$ and $\to\infty$ at $r=0$ . The count of 28 smooth structures on $S^7$ is from Kervaire–Milnor (Ann. Math. 1963), Milnor's original exotic-sphere construction from 1956, and the existence of uncountably many exotic $\mathbb{R}^4$ from Donaldson–Freedman theory. No fabricated constants appear.

Builds on: Smooth Manifolds and Coordinate Charts. Leads to: Tangent Vectors and the Tangent Space · The Tensor Transformation Law · General Tensors and the Tensor Product · Tensor Densities and Weights · The Invariant Volume Element.

Ask the tutor: Want to see exactly how the transition Jacobian threads through to the metric, the Christoffel symbols, and the Riemann tensor — or how numerical relativity stitches real coordinate patches together across a black-hole horizon?

Why can't a single chart cover the whole manifold in general?

For many manifolds it simply cannot, for topological reasons. A chart image is an open subset of $\mathbb{R}^n$ , so a manifold covered by one chart must be homeomorphic to an open subset of $\mathbb{R}^n$ . The 2-sphere is compact and boundaryless, which no nonempty proper open subset of $\mathbb{R}^2$ is, so $S^2$ needs at least two charts — the stereographic pair does it. Spacetimes with nontrivial topology, black-hole exteriors and interiors, and even the cylinder all fail the single-chart test. This is not a defect: it is exactly why the atlas, with its transition functions, is the right primitive object rather than a global coordinate system.

What actually goes wrong if the transition functions are only continuous, not smooth?

Everything that requires a derivative collapses. With merely continuous ( $C^0$ ) overlap maps you have a topological manifold: you can speak of open sets, continuity and connectedness, but not of tangent vectors, because a tangent vector is a directional-derivative operator and there is nothing to differentiate. No tangent spaces means no vector fields, no one-forms, no metric tensor, no connection, and no curvature — so no Einstein equations. A $C^k$ atlas lets you take $k$ derivatives, which is enough for some existence proofs in the initial-value formulation of GR, but the metric, connection and curvature of standard GR all live comfortably only in the $C^\infty$ world, which is why smoothness is the default demand.

Is the smooth structure really extra data, or does it follow from the shape?

It is genuinely extra data. Topology fixes the "shape" — what is connected to what — but not the notion of differentiability, and the same topological manifold can carry inequivalent maximal smooth atlases. The famous examples are Milnor's exotic 7-spheres (the 7-sphere admits 28 distinct smooth structures) and exotic $\mathbb{R}^4$ (uncountably many, unique to four dimensions). Two such structures are homeomorphic but not diffeomorphic: continuous maps identify them, but no smooth map with smooth inverse does. For physical spacetime one adopts the standard structure without hesitation, but the existence of alternatives proves that choosing the differentiable structure is an act of modeling, not a theorem handed down by the topology.

How do I tell a coordinate singularity from a real one?

Compute a scalar invariant and see whether it diverges. A metric component like $g_{rr}$ is chart-dependent, so its blowing up may only mean the chart has failed. The clean test is a coordinate-independent scalar — most commonly the Kretschmann scalar $K=R_{abcd}R^{abcd}$ , but the Ricci scalar or other curvature invariants work too. In Schwarzschild, $K=48M^2/r^6$ is finite at the horizon $r=2M$ (so it is a coordinate singularity, removable by Kruskal coordinates) but infinite at $r=0$ (a true curvature singularity present in every chart). If some invariant diverges, the singularity is physical; if none does and a change of atlas restores regularity, it was an artifact of the coordinates.

Why are the transition Jacobians the same matrices that appear in the tensor law?

Because a tensor is defined by how it responds to a change of chart, and the only thing that changes between charts is the transition function. When you re-express a coordinate basis $\partial_\nu$ in a new chart, the chain rule produces the factor $\partial x'^\mu/\partial x^\nu$ — the transition Jacobian — and that factor is what a vector's components inherit. Stacking one Jacobian per upper index and one inverse Jacobian per lower index gives the general tensor transformation law. So the tensor law is not an independent postulate: it is the direct consequence of demanding that geometric objects be chart-independent while their components are pulled back through the atlas's transition functions.

Does general covariance say the laws of physics are the same in all coordinates?

It says physical laws, written as tensor equations, hold in every chart of the differentiable structure — and that is a consequence of the atlas formalism, not a separate mystical principle. If a tensor equation $T_{\mu\nu}=0$ holds in one chart, it holds in all, because both sides transform identically under the transition Jacobians: multiply each side by the same $J$ 's and zero stays zero. Coordinates are labels, and physics must not depend on the labelling. The subtle point is that general covariance in this bare form is almost automatic once you insist on tensors; its physical bite in GR comes from also demanding that there be no fixed background fields, only the dynamical metric.

Why demand C∞ rather than the even stronger analytic (C^ω) condition?

Analytic transition functions would make spacetime too rigid. A real-analytic function is determined everywhere by its behavior in any small neighborhood (its germ), so an analytic solution of the field equations in one patch would be forced into a unique continuation across the whole manifold. That is wonderful for rigidity theorems but wrong for physics: gravitational fields must be able to be locally different — a wave here, flatness there, a star somewhere else — with support that does not dictate the whole spacetime. Smooth ( $C^\infty$ ) functions allow this local independence (one can build smooth bump functions that are analytic nowhere), while still permitting arbitrarily many derivatives. So $C^\infty$ is the Goldilocks regularity: enough for all of GR's tensors, not so much that it over-constrains the geometry.