Smooth Functions and \(C^\infty(M)\)
The previous pages equipped a topological manifold with a
smooth structure:
a maximal collection of charts whose transition maps are smooth in the ordinary Euclidean sense. That structure was
introduced precisely so that differentiation could be transported from \(\mathbb{R}^n\) to the manifold without
depending on which chart performs the transport. We now collect the dividend. With a smooth structure fixed, we can
say what it means for a real-valued function — and, in the next section, a map between manifolds — to be smooth, and
the definition will be chart-independent by construction.
Throughout, \(M\) denotes a smooth manifold of dimension \(n\), possibly with boundary; when boundary plays no role we
say only "smooth manifold." We begin with the simplest case: functions valued in a Euclidean space.
Smoothness of a Function
Definition: Smooth Function on a Manifold
Let \(M\) be a smooth \(n\)-manifold (with or without boundary) and let \(f : M \to \mathbb{R}^k\) be any function.
We say \(f\) is smooth if for every point \(p \in M\) there exists a
smooth chart
\((U, \varphi)\) with \(p \in U\) such that the composite
\[
f \circ \varphi^{-1} : \varphi(U) \to \mathbb{R}^k
\]
is smooth in the ordinary sense on the open set \(\varphi(U) \subseteq \mathbb{R}^n\). When \(M\) has nonempty
boundary and \(\varphi(U) \subseteq \mathbb{H}^n\) is not open in \(\mathbb{R}^n\), smoothness of
\(f \circ \varphi^{-1}\) is understood in the sense of
smoothness on a subset of a half-space:
the composite admits a smooth extension to an open neighborhood of each point.
The definition asks only for one such chart at each point, yet it is independent of the choice. This is the
payoff of maximality. If \((U, \varphi)\) and \((V, \psi)\) are both smooth charts of \(M\) containing \(p\), they are
smoothly compatible,
so the transition map \(\psi \circ \varphi^{-1}\) is a diffeomorphism between open subsets of \(\mathbb{R}^n\) (or
\(\mathbb{H}^n\)). On the overlap we may write
\[
f \circ \psi^{-1} = \bigl(f \circ \varphi^{-1}\bigr) \circ \bigl(\varphi \circ \psi^{-1}\bigr),
\]
a composition of an ordinary-smooth function with a smooth transition map; hence if \(f\) reads smoothly through one
smooth chart at \(p\), it reads smoothly through every smooth chart at \(p\). Smoothness of a function is therefore a
well-defined property of \(f\) alone, owing nothing to a coordinate choice.
The Coordinate Representation
Definition: Coordinate Representation of a Function
With \(f : M \to \mathbb{R}^k\) and a smooth chart \((U, \varphi)\) as above, the composite
\[
\widehat{f} := f \circ \varphi^{-1} : \varphi(U) \to \mathbb{R}^k
\]
is called the coordinate representation of \(f\) relative to \((U, \varphi)\). Writing the
coordinates of a point as \(\varphi(q) = (x^1, \dots, x^n)\), the coordinate representation is the honest
Euclidean function \(\widehat{f}(x^1, \dots, x^n)\) on which all ordinary calculus may be performed.
The coordinate representation is the device through which every manifold computation is ultimately carried out: one
chooses a chart, passes to \(\widehat{f}\), differentiates or integrates in \(\mathbb{R}^n\) as usual, and then
interprets the result back on \(M\). The chart-independence established above guarantees that statements which are
invariant under transition maps — vanishing of \(f\), smoothness of \(f\), the value \(f(p)\) — do not depend on the
chart used to compute them, even though the explicit formula for \(\widehat{f}\) certainly does.
The Algebra \(C^\infty(M)\)
Among all smooth maps into Euclidean space, the real-valued ones, \(k = 1\), are singled out by an extra layer of
structure: they can be multiplied. We record the resulting object, which will be a constant companion in everything
that follows — vector fields act on it, tangent vectors are derivations of it, and the entire algebraic approach to
differential geometry is organized around it.
Definition: The Algebra \(C^\infty(M)\)
The set of all smooth real-valued functions \(f : M \to \mathbb{R}\) is denoted \(C^\infty(M)\). It is a real
vector space under pointwise addition and scalar multiplication, and a commutative ring under pointwise
multiplication; for \(f, g \in C^\infty(M)\), \(c \in \mathbb{R}\), and \(p \in M\),
\[
(f + g)(p) = f(p) + g(p), \qquad (cf)(p) = c\,f(p), \qquad (fg)(p) = f(p)\,g(p).
\]
Equipped with all three operations, \(C^\infty(M)\) is a commutative algebra over \(\mathbb{R}\).
Proof that the operations preserve smoothness:
The vector space axioms hold pointwise in \(\mathbb{R}\); the only substantive point is closure — that the
results of these operations are again smooth. Fix \(p \in M\) and a smooth chart \((U, \varphi)\) containing
\(p\). For \(f, g \in C^\infty(M)\), the coordinate representations \(\widehat{f} = f \circ \varphi^{-1}\) and
\(\widehat{g} = g \circ \varphi^{-1}\) are smooth functions on \(\varphi(U) \subseteq \mathbb{R}^n\). Composition
with \(\varphi^{-1}\) distributes over the pointwise operations:
\[
(f + g) \circ \varphi^{-1} = \widehat{f} + \widehat{g}, \qquad
(cf) \circ \varphi^{-1} = c\,\widehat{f}, \qquad
(fg) \circ \varphi^{-1} = \widehat{f}\,\widehat{g}.
\]
Sums, scalar multiples, and products of smooth functions on an open subset of \(\mathbb{R}^n\) are smooth, by the
elementary differentiation rules of multivariable calculus. Hence each of \(f + g\), \(cf\), and \(fg\) reads
smoothly through the chart \((U, \varphi)\) at \(p\). As \(p\) was arbitrary, all three are smooth on \(M\), so
\(C^\infty(M)\) is closed under the three operations. Commutativity and the ring and module axioms are inherited
verbatim from \(\mathbb{R}\), completing the verification that \(C^\infty(M)\) is a commutative \(\mathbb{R}\)-algebra.
Two remarks fix the perspective. First, the constant functions form a copy of \(\mathbb{R}\) inside \(C^\infty(M)\),
so \(\mathbb{R} \subseteq C^\infty(M)\) as a subalgebra; the unit of the algebra is the constant function \(1\).
Second, the entire argument used the chart only to certify smoothness — the operations themselves are defined directly
on \(M\), pointwise, with no reference to coordinates. This is the recurring pattern of the subject: structure is
defined intrinsically on the manifold, while smoothness is the one property that must be checked in a chart.
Smooth Maps Between Manifolds
Smoothness of a function valued in \(\mathbb{R}^k\) needed only a chart on the domain, because the codomain
\(\mathbb{R}^k\) is already a Euclidean space in which ordinary calculus lives. For a map between two manifolds, both
ends require coordinates: we must read the map through a chart on the domain and a chart on the codomain. The
definition that follows arranges these charts so that the resulting notion is again chart-independent, and so that an
apparently separate property — continuity — comes for free.
The Definition
Definition: Smooth Map Between Manifolds
Let \(M\) and \(N\) be smooth manifolds (with or without boundary) and let \(F : M \to N\) be any map. We say
\(F\) is smooth if for every point \(p \in M\) there exist smooth charts \((U, \varphi)\)
containing \(p\) and \((V, \psi)\) containing \(F(p)\) such that
\[
F(U) \subseteq V
\qquad\text{and}\qquad
\psi \circ F \circ \varphi^{-1} : \varphi(U) \to \psi(V)
\]
is smooth in the ordinary sense, as a map between open subsets of \(\mathbb{R}^n\) and \(\mathbb{R}^m\) (or
half-spaces, when boundary is present).
One feature of this definition is deliberate and easy to overlook: the requirement \(F(U) \subseteq V\). Without it,
the composite \(\psi \circ F \circ \varphi^{-1}\) would not even be defined — \(F\) might carry points of \(U\) outside
the domain \(V\) of \(\psi\). The inclusion is imposed precisely so that smoothness, as written, automatically forces
\(F\) to be continuous. We could instead have demanded continuity of \(F\) up front and then asked for smooth
coordinate representations on overlaps; the two formulations are equivalent, but the present one is leaner, because it
extracts continuity as a consequence rather than positing it as a hypothesis. We make that consequence precise now.
Smoothness Implies Continuity
Proposition: Smooth Maps Are Continuous
If \(F : M \to N\) is a smooth map between smooth manifolds, then \(F\) is continuous.
Proof:
Continuity is a local property, so it suffices to show that each point of \(M\) has a neighborhood on which \(F\)
is continuous. Fix \(p \in M\) and choose smooth charts \((U, \varphi)\) about \(p\) and \((V, \psi)\) about
\(F(p)\) as in the definition of smoothness, so that \(F(U) \subseteq V\) and the coordinate representation
\(\widehat{F} := \psi \circ F \circ \varphi^{-1}\) is smooth, hence continuous, on \(\varphi(U)\). Because
\(F(U) \subseteq V\), the restriction \(F|_U\) maps into \(V\), and on \(U\) we may solve for \(F\) by composing
with the coordinate homeomorphisms in reverse:
\[
F|_U = \psi^{-1} \circ \widehat{F} \circ \varphi.
\]
Here \(\varphi : U \to \varphi(U)\) and \(\psi^{-1} : \psi(V) \to V\) are homeomorphisms — this is part of what
it means to be a chart — and \(\widehat{F}\) is continuous as just noted. The right-hand side is therefore a
composition of three continuous maps, so \(F|_U\) is continuous. Since \(U\) is an open neighborhood of \(p\) and
\(p\) was arbitrary, \(F\) is continuous on all of \(M\).
The proof exposes exactly where \(F(U) \subseteq V\) is spent: it is what allows the equation
\(F|_U = \psi^{-1} \circ \widehat{F} \circ \varphi\) to be written at all, since \(\widehat{F}\) is only defined where
\(F\) lands inside \(V\). Drop the inclusion and the factorization collapses; with it, continuity is a one-line
corollary of the continuity of smooth Euclidean maps.
Equivalent Characterizations and Locality
The pointwise definition is convenient for verification but awkward for theory, since it quantifies over a chosen pair
of charts at each point. Two standard reformulations remove the awkwardness: one replaces "some chart" by "every
chart," the other isolates the local nature of smoothness.
Proposition: Equivalent Characterizations of Smoothness
For a map \(F : M \to N\) between smooth manifolds, the following are equivalent.
(a) \(F\) is smooth.
(b) \(F\) is continuous, and for every pair of smooth charts \((U, \varphi)\) for \(M\) and
\((V, \psi)\) for \(N\), the coordinate representation
\(\psi \circ F \circ \varphi^{-1}\) is smooth on the open set
\(\varphi\bigl(U \cap F^{-1}(V)\bigr) \subseteq \mathbb{R}^n\).
Proof sketch:
That (b) implies (a) is immediate, since (b) supplies a smooth coordinate representation about every point. For
the converse, continuity is the proposition that smooth maps are continuous, and given arbitrary charts \((U, \varphi)\), \((V, \psi)\), one
compares them at each point of \(U \cap F^{-1}(V)\) with the charts furnished by smoothness of \(F\); the two
coordinate representations differ by transition maps, which are smooth, so \(\psi \circ F \circ \varphi^{-1}\)
is smooth on the overlap. The continuity of \(F\) is what makes \(U \cap F^{-1}(V)\) open, so that "smooth on it"
is meaningful.
Proposition: Smoothness Is Local
Let \(F : M \to N\) be a map between smooth manifolds.
(a) If every point of \(M\) has a neighborhood \(U\) such that the restriction \(F|_U\) is
smooth, then \(F\) is smooth.
(b) Conversely, if \(F\) is smooth, then its restriction to every open subset of \(M\) is smooth.
Proof sketch:
Both directions follow from the fact that the defining condition is itself phrased pointwise: smoothness at
\(p\) depends only on the behavior of \(F\) on an arbitrarily small neighborhood of \(p\), since charts may be
shrunk to open subsets. Restricting \(F\) to an open set does not disturb the charts witnessing smoothness at
interior points, giving (b); and a witness on each \(U\) is a witness for the whole, giving (a).
Locality has an indispensable practical consequence: maps that are defined piecewise, by separate smooth formulas on
the members of an open cover, are smooth provided the formulas agree on overlaps. This is the manifold version of the
familiar principle that a function pieced together from smooth pieces on open sets is smooth.
Corollary: Gluing Lemma for Smooth Maps
Let \(M\) and \(N\) be smooth manifolds and let \(\{U_\alpha\}\) be an open cover of \(M\). Suppose that for each
\(\alpha\) we are given a smooth map \(F_\alpha : U_\alpha \to N\), and that these agree on overlaps:
\(F_\alpha = F_\beta\) on \(U_\alpha \cap U_\beta\) for all \(\alpha, \beta\). Then the map \(F : M \to N\)
defined by \(F|_{U_\alpha} = F_\alpha\) is well-defined and smooth.
Proof sketch:
The agreement on overlaps makes \(F\) well-defined as a function. Each point \(p \in M\) lies in some
\(U_\alpha\), an open set on which \(F\) coincides with the smooth map \(F_\alpha\); thus \(F\) is smooth on a
neighborhood of every point, and the locality of smoothness promotes this to smoothness on \(M\).
The Reach of Gluing, and Its Limit
The gluing lemma is stated for an open cover, and the openness is essential — not a technical convenience.
When the pieces are glued along a set that is closed rather than open, agreement on the seam no longer guarantees
smoothness of the result, even when each piece is itself perfectly smooth and the pieces match in value where they
meet. A single example on the line makes the failure visible.
Take the two closed half-lines \(A_+ = [0, \infty)\) and \(A_- = (-\infty, 0]\), which cover \(\mathbb{R}\) but
overlap only in the single point \(\{0\}\), and define
\[
f_+ : A_+ \to \mathbb{R}, \quad f_+(x) = x,
\qquad
f_- : A_- \to \mathbb{R}, \quad f_-(x) = -x.
\]
Each is the restriction of a smooth function on all of \(\mathbb{R}\), hence smooth on its (closed) domain, and they
agree where the domains meet, since \(f_+(0) = 0 = f_-(0)\). Yet the glued function is
\[
f(x) =
\begin{cases}
x, & x \ge 0,\\
-x, & x \le 0,
\end{cases}
\qquad\text{that is,}\qquad
f(x) = |x|,
\]
which is not differentiable at the origin, let alone smooth. The hypothesis of the gluing lemma is not met here
because \(A_+\) and \(A_-\) are closed, not open: neither contains an open neighborhood of \(0\) on which its formula
is the only one in force, so the smoothness of the pieces says nothing about the smoothness of \(f\) across the seam.
This is a genuine limitation, and overcoming it requires a tool we have not yet built: a way to interpolate smoothly
between formulas across a closed boundary, blending one into the other so that all derivatives match. That tool is the
smooth bump function, and the systematic device built from it — a smooth partition of unity subordinate to a cover —
is precisely what converts locally defined smooth data into a single global smooth map even when the cover is not
cooperative. We construct partitions of unity in the next page of this series, where the paracompactness of manifolds
secured earlier becomes the engine of the construction; the failure of \(|x|\) above is the obstruction they are
designed to defeat.
The Coordinate Representation of a Map
Definition: Coordinate Representation of a Map
Let \(F : M \to N\) be a map, \((U, \varphi)\) a smooth chart for \(M\), and \((V, \psi)\) a smooth chart for
\(N\) with \(F(U \cap F^{-1}(V)) \subseteq V\). The composite
\[
\widehat{F} := \psi \circ F \circ \varphi^{-1}
\]
is the coordinate representation of \(F\) with respect to the chosen charts. In coordinates
\(\varphi = (x^1, \dots, x^n)\) on the domain and \(\psi = (y^1, \dots, y^m)\) on the codomain, \(\widehat{F}\) is
an ordinary map between open subsets of Euclidean space, written \(\bigl(y^1(x), \dots, y^m(x)\bigr)\), and
smoothness of \(F\) means precisely that each such \(\widehat{F}\) is smooth.
By the equivalence established above, this representation is smooth for every admissible pair of charts once
\(F\) is smooth, not merely for the pair witnessing smoothness at a point. The coordinate representation \(\widehat{F}\)
is the object on which all explicit computation is performed, and it will return in sharper form when we differentiate
maps between manifolds: the derivative of \(F\) at a point will itself be read through these same coordinates.
Constructing Smooth Maps
The definition certifies smoothness one map at a time, but mathematics is built by assembling maps from simpler ones.
We now record the closure properties that let us recognize smoothness without returning to charts: the elementary maps
are smooth, and smoothness survives composition. These are the rules invoked, usually silently, in every later
construction — and the composition rule in particular is where the continuity we extracted in the previous section
earns its keep.
Proposition: Elementary Smooth Maps and Composition
Let \(M\), \(N\), and \(P\) be smooth manifolds, with or without boundary.
(a) Every constant map \(c : M \to N\) is smooth.
(b) The identity map \(\mathrm{id}_M : M \to M\) is smooth.
(c) If \(U \subseteq M\) is an
open submanifold,
then the inclusion map \(\iota : U \hookrightarrow M\) is smooth.
(d) If \(F : M \to N\) and \(G : N \to P\) are smooth, then the composition
\(G \circ F : M \to P\) is smooth.
Proof:
(a) Let \(c \equiv q_0\) be constant with value \(q_0 \in N\), and fix \(p \in M\). Choose any
smooth chart \((V, \psi)\) about \(q_0\) and any smooth chart \((U, \varphi)\) about \(p\) small enough that
\(c(U) = \{q_0\} \subseteq V\). The coordinate representation \(\psi \circ c \circ \varphi^{-1}\) is the constant
map \(\varphi(U) \to \{\psi(q_0)\}\), which is smooth. Hence \(c\) is smooth.
(b) Fix \(p \in M\) and a smooth chart \((U, \varphi)\) about \(p\); use \((U, \varphi)\) on both
domain and codomain. The coordinate representation of \(\mathrm{id}_M\) is
\(\varphi \circ \mathrm{id}_M \circ \varphi^{-1} = \mathrm{id}_{\varphi(U)}\), the identity on an open subset of
\(\mathbb{R}^n\), which is smooth.
(c) Fix \(p \in U\). Since \(U\) is an open submanifold, any smooth chart \((W, \varphi)\) for
\(M\) with \(p \in W\) restricts to a smooth chart \((W \cap U, \varphi|_{W \cap U})\) for \(U\), and these charts
belong to the smooth structures of \(U\) and \(M\) respectively. Using the restricted chart on the domain \(U\)
and \((W, \varphi)\) on the codomain \(M\), the coordinate representation of the inclusion is
\(\varphi \circ \iota \circ (\varphi|_{W \cap U})^{-1} = \mathrm{id}_{\varphi(W \cap U)}\), again an identity on an
open set, hence smooth.
(d) Fix \(p \in M\). Applying the definition of smoothness to \(G\) at the point \(F(p)\), there
exist smooth charts \((V, \theta)\) about \(F(p)\) and \((W, \psi)\) about \(G(F(p))\) with
\(G(V) \subseteq W\) such that the coordinate representation
\(\psi \circ G \circ \theta^{-1} : \theta(V) \to \psi(W)\) is smooth. Now we use that \(F\) is
continuous — established for every
smooth map in the previous section — so that \(F^{-1}(V)\) is open and contains \(p\). We may therefore choose a
smooth chart \((U, \varphi)\) about \(p\) with \(U \subseteq F^{-1}(V)\), so that \(F(U) \subseteq V\) and the
coordinate representation \(\theta \circ F \circ \varphi^{-1} : \varphi(U) \to \theta(V)\) is smooth.
With these charts in hand, \(G \circ F\) carries \(U\) into \(W\), since
\(G\bigl(F(U)\bigr) \subseteq G(V) \subseteq W\), so the coordinate representation of \(G \circ F\) relative to
\((U, \varphi)\) and \((W, \psi)\) is defined on all of \(\varphi(U)\), and there it factors as
\[
\psi \circ (G \circ F) \circ \varphi^{-1}
\;=\;
\bigl(\psi \circ G \circ \theta^{-1}\bigr) \circ \bigl(\theta \circ F \circ \varphi^{-1}\bigr).
\]
The right-hand side is a composition of two smooth maps between open subsets of Euclidean spaces, hence smooth by
the chain rule for ordinary smooth maps. Therefore \(G \circ F\) reads smoothly through \((U, \varphi)\) and
\((W, \psi)\) at \(p\); as \(p\) was arbitrary, \(G \circ F\) is smooth.
Part (d) is the structural heart of the proposition, and its proof is worth pausing over, because it shows the
previous section's labor being repaid. The factorization
\(\psi \circ (G \circ F) \circ \varphi^{-1} = (\psi \circ G \circ \theta^{-1}) \circ (\theta \circ F \circ \varphi^{-1})\)
is the entire content of the argument, but writing it requires that all three coordinate representations be defined on
compatible domains — and securing \(p \in U \subseteq F^{-1}(V)\) is exactly what the continuity of \(F\) provides.
Smoothness alone, without its continuity corollary, would not let the charts be aligned; this is why we proved
continuity first and in full, rather than assuming it.
Smoothness of the Standard Constructions
These closure rules are not abstract bookkeeping; they certify the smoothness of the maps that populate the examples
built earlier in the series. When we constructed
the smooth sphere \(\mathbb{S}^n\),
real projective space \(\mathbb{RP}^n\),
and
product manifolds,
we equipped each with a smooth structure but had not yet defined what a smooth map out of or into it should be. We can
now name the maps those constructions were silently arranging to make smooth.
On the product manifold \(M \times N\), the two projections
\[
\pi_M : M \times N \to M, \qquad \pi_N : M \times N \to N
\]
are smooth: in the product charts used to build \(M \times N\), each projection has coordinate representation a linear
projection of Euclidean coordinates, manifestly smooth. Dually, given smooth maps \(F : Z \to M\) and \(G : Z \to N\)
from any smooth manifold \(Z\), the map \((F, G) : Z \to M \times N\) is smooth, and these two facts together are the
characteristic property of the product. For the sphere, the inclusion
\(\iota : \mathbb{S}^n \hookrightarrow \mathbb{R}^{n+1}\) is smooth, as is the antipodal map
\(x \mapsto -x\); for projective space, the quotient map
\(\pi : \mathbb{S}^n \to \mathbb{RP}^n\) identifying antipodes is smooth. Each of these is verified by passing to the
charts fixed when the space was constructed and observing that the coordinate representation is an ordinary smooth map —
the closure rules above then assemble these primitives into every map we shall need.
The pattern is uniform and worth stating plainly: once a space is given its smooth structure, the smoothness of maps
involving it is decided entirely in the charts that defined the structure, and the elementary maps together with
composition generate the rest. With this vocabulary in place, one notion remains before the theory of smooth maps is
complete — the notion of when two smooth manifolds are to be regarded as the same.
Diffeomorphisms and Smooth Invariants
A smooth structure is extra data laid over a topological manifold, and the maps that respect it in both directions are
the structure-preserving isomorphisms of the smooth category. We isolate them now, and then extract the first
invariant they protect.
Definition: Diffeomorphism
Let \(M\) and \(N\) be smooth manifolds, with or without boundary. A diffeomorphism from \(M\) to
\(N\) is a smooth bijective map \(F : M \to N\) whose inverse \(F^{-1} : N \to M\) is also smooth. If such a map
exists, \(M\) and \(N\) are said to be diffeomorphic, written \(M \approx N\).
Because the identity is smooth, the composition of smooth maps is smooth, and the inverse of a diffeomorphism is by
definition a diffeomorphism, the relation "is diffeomorphic to" is an equivalence relation on smooth manifolds. A
diffeomorphism is in particular a homeomorphism — it is a continuous bijection with continuous inverse, since smooth
maps are continuous — so diffeomorphic manifolds are homeomorphic. The converse is famously false, but that is a story
for a later stage; here we draw out what diffeomorphism does guarantee.
The equivalence relation immediately settles a question left open earlier in the series. We saw that the line
\(\mathbb{R}\) carries
distinct smooth structures
— the standard one and the one declared by the chart \(x \mapsto x^3\), whose maximal atlases do not overlap and so
are genuinely different as smooth structures. Yet the two are diffeomorphic: the map \(x \mapsto x^{1/3}\) is smooth
from the standard structure to the cube-root structure with smooth inverse, so as smooth manifolds they are the same.
The lesson is that "distinct smooth structure" and "distinct smooth manifold" are not the same notion; the first counts
atlases, the second counts diffeomorphism classes, and the cube-root example separates them. Whether a manifold can
carry smooth structures that are not diffeomorphic — genuinely exotic ones — is a far deeper question, first
answered affirmatively for \(\mathbb{R}^4\) and for spheres in high dimensions, and one we are not yet equipped to
address.
The Pattern of "Sameness" Across Mathematics
The definition of diffeomorphism instantiates a pattern the curriculum has met before. In topology, two spaces
count as the same when related by a
homeomorphism
— a continuous bijection with continuous inverse. In algebra, two groups count as the same when related by a
group isomorphism
— a bijective homomorphism whose inverse is again a homomorphism. The diffeomorphism is the differential-geometric
member of the same family: a bijection that preserves the relevant structure in both directions. Each category
fixes a notion of admissible map, and "sameness" is always invertibility within that notion. Reading the three
definitions side by side reveals that they are one idea wearing three coats — continuity, the group operation, and
smoothness, respectively.
Diffeomorphism Invariance of Dimension
A topological manifold has a well-defined dimension because the
topological invariance of dimension
forbids open subsets of \(\mathbb{R}^m\) and \(\mathbb{R}^n\) from being homeomorphic unless \(m = n\). That theorem
is genuinely hard: its honest proofs rest on algebraic topology, on machinery such as homology that detects the global
shape obstructions distinguishing Euclidean spaces of different dimensions. In the smooth category there is a parallel
statement, and — remarkably — it is comparatively elementary, because differentiation linearizes the problem and hands
it to linear algebra.
Theorem: Diffeomorphism Invariance of Dimension
A nonempty smooth manifold of dimension \(m\) cannot be diffeomorphic to a smooth manifold of dimension \(n\)
unless \(m = n\).
Proof:
Suppose \(M\) is a nonempty smooth \(m\)-manifold, \(N\) is a nonempty smooth \(n\)-manifold, and
\(F : M \to N\) is a diffeomorphism. Choose any point \(p \in M\), and let \((U, \varphi)\) and \((V, \psi)\) be
smooth charts containing \(p\) and \(F(p)\) respectively, shrunk if necessary so that \(F(U) \subseteq V\). Consider
the coordinate representation \(\widehat{F} = \psi \circ F \circ \varphi^{-1}\), defined on the open set
\(\varphi(U) \subseteq \mathbb{R}^m\). It is smooth because \(F\) is; it is injective because \(F\), \(\varphi\),
and \(\psi\) are; and, regarded as a map onto its image
\[
\widehat{W} := \widehat{F}\bigl(\varphi(U)\bigr) = \psi\bigl(F(U)\bigr) \subseteq \mathbb{R}^n,
\]
it is bijective by construction, with inverse the restriction of \(\varphi \circ F^{-1} \circ \psi^{-1}\) to
\(\widehat{W}\), which is smooth because \(F^{-1}\) is. We are careful not to assert
\(\widehat{W} = \psi(V)\): the inclusion \(F(U) \subseteq V\) may be strict, so the image may be a proper subset of
\(\psi(V)\). No such equality is needed. What we have produced is a smooth injective map
\(\widehat{F} : \varphi(U) \to \mathbb{R}^n\), defined on a nonempty open subset of \(\mathbb{R}^m\), with a smooth
inverse defined on its image — and that alone, as we show next, forces \(m = n\).
This last statement is where smoothness pays off, and it is settled by linearization. Write \(a = \varphi(p)\) and
\(b = \widehat{F}(a)\). Since \(\widehat{F}\) and \(\widehat{F}^{-1}\) are smooth mutual inverses, differentiating
the two identities \(\widehat{F}^{-1} \circ \widehat{F} = \mathrm{id}\) and
\(\widehat{F} \circ \widehat{F}^{-1} = \mathrm{id}\) by the
chain rule
— with the roles \(\mathbf{h} = \widehat{F}\), \(\mathbf{g} = \widehat{F}^{-1}\) and the identity having
differential the identity matrix — yields
\[
D(\widehat{F}^{-1})(b)\,\circ\,D\widehat{F}(a) = \mathrm{Id}_{\mathbb{R}^m},
\qquad
D\widehat{F}(a)\,\circ\,D(\widehat{F}^{-1})(b) = \mathrm{Id}_{\mathbb{R}^n}.
\]
Thus the linear map \(D\widehat{F}(a) : \mathbb{R}^m \to \mathbb{R}^n\) admits a two-sided inverse, namely
\(D(\widehat{F}^{-1})(b)\), and is therefore a linear isomorphism. But a linear isomorphism between
\(\mathbb{R}^m\) and \(\mathbb{R}^n\) exists only when the dimensions agree, so \(m = n\). This proves the
Euclidean statement, and with it the theorem.
The contrast with the topological case is the moral of the proof, and it is worth making explicit. Both theorems
assert that dimension is an invariant — topological invariance under homeomorphism, smooth invariance under
diffeomorphism. Yet the topological version must reach for the heavy apparatus of algebraic topology, because a mere
homeomorphism carries no derivative to exploit and the obstruction lives in the global shape of the spaces. A
diffeomorphism, by contrast, comes with a derivative at every point, and that derivative is a linear isomorphism whose
very existence pins down the dimension on the spot. The same conclusion is reached by incomparably lighter means; the
smooth structure converts a deep topological fact into a one-line consequence of the invertibility of linear maps.
This is the first concrete dividend of working in the smooth category rather than the merely topological one, and the
theme — that smoothness linearizes, and linear algebra then decides — will recur throughout differential geometry.
Diffeomorphism Invariance of the Boundary
Dimension is not the only structure a diffeomorphism protects. For manifolds with boundary there is a companion
statement, asserting that a diffeomorphism cannot mix boundary points with interior points — it carries boundary to
boundary and interior to interior. This is the natural counterpart to the dimension theorem, and together the two say
that a diffeomorphism preserves every piece of the coarse structure a smooth manifold-with-boundary carries.
Theorem: Diffeomorphism Invariance of the Boundary
Let \(M\) and \(N\) be smooth manifolds with boundary and let \(F : M \to N\) be a diffeomorphism. Then \(F\)
restricts to a diffeomorphism \(\partial M \to \partial N\) of the boundaries and to a diffeomorphism
\(\operatorname{Int} M \to \operatorname{Int} N\) of the interiors; in particular \(F(\partial M) = \partial N\).
Proof:
Whether a point of a smooth manifold-with-boundary is a boundary point or an interior point is an intrinsic
property, independent of any chart: this is the content of the
smooth invariance of the boundary,
which gives the well-defined partition of any such manifold into
interior and boundary.
We use it to show \(F\) preserves this partition. Suppose, for contradiction, that some interior point
\(p \in \operatorname{Int} M\) is carried to a boundary point \(F(p) \in \partial N\). Choose an interior chart
\((U, \varphi)\) about \(p\), so that \(\varphi(U)\) is open in \(\mathbb{R}^n\), and a boundary chart
\((V, \psi)\) about \(F(p)\), so that \(\psi(V)\) is open in the half-space \(\mathbb{H}^n\) with \(\psi(F(p))\)
lying on \(\partial\mathbb{H}^n\). Shrinking as necessary, the coordinate representation
\(\psi \circ F \circ \varphi^{-1}\) is a diffeomorphism from an open subset of \(\mathbb{R}^n\) onto a relatively
open subset of \(\mathbb{H}^n\) that contains the boundary point \(\psi(F(p))\). But smooth invariance of the
boundary forbids exactly this: a point lying in the open subset of \(\mathbb{R}^n\) is an interior point of the
domain, while its image \(\psi(F(p)) \in \partial\mathbb{H}^n\) is a boundary point of the target, and no
diffeomorphism between subsets of half-spaces can carry an interior point to a boundary point. This contradiction
shows \(F(\operatorname{Int} M) \subseteq \operatorname{Int} N\). Applying the same argument to the diffeomorphism
\(F^{-1}\) gives \(F^{-1}(\operatorname{Int} N) \subseteq \operatorname{Int} M\), equivalently
\(\operatorname{Int} N \subseteq F(\operatorname{Int} M)\). The two inclusions together force
\(F(\operatorname{Int} M) = \operatorname{Int} N\), so \(F\) restricts to a bijection
\(\operatorname{Int} M \to \operatorname{Int} N\); it is smooth with smooth inverse, hence a diffeomorphism. Taking
complements, \(F\) restricts to a diffeomorphism \(\partial M \to \partial N\) as well.
The two invariance theorems are the boundary and dimension faces of one principle: a diffeomorphism is an isomorphism
of the smooth category, and an isomorphism preserves every invariant the category recognizes. Dimension and the
boundary are the first two such invariants we can name; the tangent spaces, vector fields, and tensors of later
chapters will join them, each preserved for the same structural reason.
Why Diffeomorphism Is the Right Notion for Geometry and Its Applications
Diffeomorphism is the equivalence under which every differential-geometric quantity is preserved: tangent spaces
correspond, smooth functions pull back to smooth functions, and — once they are defined — vector fields, tensors,
and integrals all transform consistently. To regard two manifolds as "the same smooth space" is precisely to
possess a diffeomorphism between them. This is also where the subject meets its computational applications. In the
configuration spaces of robotics, in the smooth latent manifolds posited by geometric approaches to machine
learning, and in the reparametrizations that leave a model's intrinsic geometry unchanged, the maps that are
allowed to count as "the same situation, differently described" are exactly the diffeomorphisms. The invariance of
dimension proved above is the most basic of the quantities they protect: it is what makes the dimension of a
configuration space, or of a data manifold, a meaningful number rather than an artifact of a chosen description.