Orientability and Covering Maps
Proving that a manifold is orientable is usually a matter of exhibiting an orientation, but proving
that a manifold is not orientable is harder, because one must rule out every possible
orientation at once. The theory of covering spaces provides one of the most effective tools for the
negative direction, by tying orientability to the behaviour of deck transformations. The first
observation is that orientability always lifts along a cover.
Proposition (Orientability Lifts Along Covers)
If \(\pi : E \to M\) is a smooth covering map and \(M\) is orientable, then \(E\) is also
orientable.
The converse direction—descending an orientation from \(E\) to \(M\)—is the interesting one, and it
fails in general. It succeeds exactly when the symmetries of the cover respect the orientation
chosen upstairs. To state this precisely we name the relevant condition on a group action.
Definition (Orientation-Preserving Action)
Suppose a Lie group \(G\) acts smoothly on an oriented smooth manifold \(E\), say on the left.
The action is an orientation-preserving action if for each \(g \in G\) the
diffeomorphism \(x \mapsto g \cdot x\) is
orientation-preserving.
For a normal covering the deck group acts transitively on each fiber, and this transitivity is
exactly what allows a fiberwise condition to assemble into a global orientation downstairs. The
result is a clean criterion: the base is orientable precisely when the deck group preserves the
orientation of the total space.
Theorem (Orientability of the Base of a Normal Cover)
Suppose \(E\) is a connected, oriented, smooth manifold, with or without boundary, and
\(\pi : E \to M\) is a smooth
normal covering map.
Then \(M\) is orientable if and only if the action of
\(\mathrm{Aut}_\pi(E)\)
on \(E\) is orientation-preserving.
Proof.
Let \(\mathcal{O}_E\) denote the given orientation on \(E\).
Suppose first that \(M\) is orientable, and let \(q\) be an arbitrary point of \(E\). Because
\(M\) is connected it has
exactly two orientations,
and one of them, call it \(\mathcal{O}_M\), has the property that
\(d\pi_q : T_q E \to T_{\pi(q)} M\) is orientation-preserving. The pullback orientation
\(\pi^{*}\mathcal{O}_M\) then agrees with the given orientation at \(q\), so by connectedness it
equals \(\mathcal{O}_E\) everywhere. Now let \(\varphi \in \mathrm{Aut}_\pi(E)\). The identity
\(\pi \circ \varphi = \pi\) gives
\[
\varphi^{*}\mathcal{O}_E
= \varphi^{*}\bigl(\pi^{*}\mathcal{O}_M\bigr)
= (\pi \circ \varphi)^{*}\mathcal{O}_M
= \pi^{*}\mathcal{O}_M
= \mathcal{O}_E,
\]
so \(\varphi\) is orientation-preserving. Hence the action of \(\mathrm{Aut}_\pi(E)\) preserves
orientation.
Conversely, suppose the action of \(\mathrm{Aut}_\pi(E)\) is orientation-preserving, and let
\(p \in M\). Choose a connected evenly covered neighbourhood \(U\) of \(p\); there is a smooth
local section
\(\sigma : U \to E\), which induces the pullback orientation \(\sigma^{*}\mathcal{O}_E\) on \(U\).
Suppose \(\sigma_1 : U \to E\) is any other smooth local section over \(U\). Because \(\pi\) is a
normal cover, \(\mathrm{Aut}_\pi(E)\) acts transitively on each fiber, so there is a deck
transformation \(\varphi\) with \(\sigma_1(p) = \varphi\bigl(\sigma(p)\bigr)\). Then
\(\varphi \circ \sigma\) is a local section agreeing with \(\sigma_1\) at \(p\), and since local
sections of a covering are determined by one value on a connected domain, \(\sigma_1 = \varphi
\circ \sigma\) on all of \(U\). Because \(\varphi\) is orientation-preserving,
\[
\sigma_1^{*}\mathcal{O}_E
= \sigma^{*}\varphi^{*}\mathcal{O}_E
= \sigma^{*}\mathcal{O}_E,
\]
so the orientations induced by \(\sigma\) and \(\sigma_1\) coincide. We may therefore define an
orientation \(\mathcal{O}_M\) on \(M\) by declaring it on each evenly covered open set to be the
pullback induced by any local section; the computation just made shows these agree on overlaps,
so \(\mathcal{O}_M\) is a well-defined orientation. Hence \(M\) is orientable.
The theorem converts a question about the global topology of \(M\) into a question about the signs
of finitely many diffeomorphisms of \(E\), which is often decidable by direct computation. The two
examples that follow carry it out in the cases that recur most often in applications.
Projective Spaces and the Möbius Band
Real projective space is the quotient of a sphere by the antipodal identification, and the quotient
map is the prototypical normal cover. Whether the projective space is orientable therefore reduces,
through the criterion just proved, to whether the antipodal map preserves the orientation of the
sphere. The answer alternates with dimension.
Proposition (Orientability of Projective Spaces)
For \(n \geq 1\), real projective space
\(\mathbb{RP}^n\)
is orientable if and only if \(n\) is odd.
Proof.
The smooth covering map \(q : S^n \to \mathbb{RP}^n\) is a two-sheeted normal cover whose only
nontrivial deck transformation is the antipodal map \(\alpha(x) = -x\). Give \(S^n\) its
standard orientation, determined by the outward unit normal \(N_x = x\): an ordered basis
\((E_1, \dots, E_n)\) of \(T_x S^n\) is positively oriented exactly when
\((x, E_1, \dots, E_n)\) is positively oriented in \(\mathbb{R}^{n+1}\). The antipodal map has
differential \(-\,\mathrm{Id}\) on \(\mathbb{R}^{n+1}\), so it sends such a basis to
\((-E_1, \dots, -E_n)\) at the point \(-x\), where the outward normal is
\(N_{-x} = -x\). Comparing orientations,
\[
\bigl(N_{-x}, -E_1, \dots, -E_n\bigr) = \bigl(-x, -E_1, \dots, -E_n\bigr)
= (-1)^{n+1}\bigl(x, E_1, \dots, E_n\bigr)
\]
as a relation between signed bases of \(\mathbb{R}^{n+1}\), since negating all \(n+1\) vectors
contributes the factor \((-1)^{n+1}\). Thus \(\alpha\) carries the induced orientation of
\(T_x S^n\) to \((-1)^{n+1}\) times the induced orientation of \(T_{-x} S^n\), so \(\alpha\) is
orientation-preserving precisely when \((-1)^{n+1} = 1\), that is, when \(n\) is odd. By the
criterion for the base of a normal cover, \(\mathbb{RP}^n\) is orientable if and only if the
deck action is orientation-preserving, which happens if and only if \(n\) is odd.
The case \(n = 2\) is the one that surfaces in applications: \(\mathbb{RP}^2\) is nonorientable,
and its orientation double cover is the two-sphere with the antipodal identification. A signal
defined on directions in three-dimensional space, invariant under reversal of a director, lives
naturally on \(\mathbb{RP}^2\) rather than on \(S^2\) — this is precisely the domain of the
diffusion signal in diffusion-weighted magnetic resonance imaging, where the measurement at each
voxel depends only on an unoriented axis. One way to bring orientation-dependent machinery to bear
is to pass to the orientable cover \(S^2\); another, pursued in gauge-equivariant networks for
such data, is to work directly on the nonorientable \(\mathbb{RP}^2\) using a gauge formulation
that does not presuppose a global orientation. Either way the orientation double cover is the
structure that organizes the choices, and the covering theorem of the next section guarantees it
exists and is essentially unique.
The same mechanism diagnoses the most familiar nonorientable surface directly, without reference
to a sphere.
Proposition (Nonorientability of the Möbius Band)
The Möbius band is not orientable.
Proof.
Realize the open Möbius band as the quotient of \(\mathbb{R}^2\) by the smooth, free, and
properly discontinuous action of \(\mathbb{Z}\) given by
\[
n \cdot (x, y) = \bigl(x + n,\, (-1)^n y\bigr).
\]
The quotient map \(q : \mathbb{R}^2 \to E\) onto the quotient \(E\) is a smooth normal covering,
and the deck group is the image of this \(\mathbb{Z}\)-action. Give \(\mathbb{R}^2\) its standard
orientation, with orientation form \(dx \wedge dy\). The generator \((x, y) \mapsto (x + 1, -y)\)
pulls this form back to \(dx \wedge d(-y) = -\,dx \wedge dy\), so it reverses orientation. The
deck action is therefore not orientation-preserving, and by the criterion for the base of a
normal cover, \(E\) is not orientable. The closed Möbius band, being the image of a strip
\(\mathbb{R} \times [-r, r]\) under the same covering, fails to be orientable by the identical
argument applied to the restricted cover.
These computations exhibit nonorientability concretely, but they presuppose a convenient covering
in each case. The final section removes that dependence by constructing, for every manifold, a
canonical orientable cover whose connectedness encodes orientability.
The Orientation Covering
Every smooth manifold, orientable or not, admits a canonical two-sheeted cover that is itself
orientable, built so that the fiber over a point records the two orientations of the tangent space
there. To treat the orientable and nonorientable cases uniformly, it is convenient to allow the
cover to be disconnected; a surjective local diffeomorphism that is evenly covered but whose total
space need not be connected will be called a generalized smooth covering map. The construction
itself is uniform, and the topology of the result is what distinguishes the two cases.
Definition (The Orientation Covering)
Let \(M\) be a connected smooth positive-dimensional manifold, with or without boundary, and let
\[
\widehat M = \bigl\{ (p, \mathcal{O}_p) : p \in M,\ \mathcal{O}_p \text{ an orientation of } T_p M \bigr\}.
\]
The projection \(\widehat\pi : \widehat M \to M\) sending \((p, \mathcal{O}_p)\) to \(p\) has
every fiber of cardinality two, since each tangent space has exactly two orientations. The set
\(\widehat M\) carries a canonical smooth structure making \(\widehat\pi\) a generalized smooth
covering map, with \(\widehat M\) oriented so that the tautological pointwise orientation
\((p, \mathcal{O}_p) \mapsto \mathcal{O}_p\) is continuous; the map \(\widehat\pi\) is called
the orientation covering, and \(\widehat M\) the oriented double
covering of \(M\). A connected open subset \(U \subseteq M\) is evenly covered by
\(\widehat\pi\) if and only if it is orientable, and over such a set every orientation of \(U\)
is the pullback induced by a local section of \(\widehat\pi\).
The defining feature of \(\widehat M\) is that orientations of \(M\) correspond to global sections
of \(\widehat\pi\). Whether such a section exists is a question about how the two sheets are
connected, and that is precisely what the next theorem resolves: orientability of \(M\) is the
same as disconnectedness of \(\widehat M\).
Theorem (Orientation Covering Theorem)
Let \(M\) be a connected smooth manifold, with or without boundary, and let
\(\widehat\pi : \widehat M \to M\) be its orientation covering.
If \(M\) is orientable, then \(\widehat M\) has exactly two components, and the restriction of
\(\widehat\pi\) to each component is a diffeomorphism onto \(M\). If \(M\) is nonorientable,
then \(\widehat M\) is connected, and \(\widehat\pi\) is a two-sheeted smooth covering map.
Proof.
If \(M\) is orientable, then by the characterization of evenly covered sets, \(M\) itself is
evenly covered by \(\widehat\pi\); hence \(\widehat M\) is the disjoint union of two open sets,
each mapped homeomorphically onto \(M\), and each restriction of \(\widehat\pi\) is a
diffeomorphism. The two components correspond to the two orientations of \(M\).
Now suppose \(M\) is nonorientable. Let \(W\) be a component of \(\widehat M\). The restriction
of \(\widehat\pi\) to \(W\) is a covering map onto \(M\), so all its fibers have the same
cardinality, which must be \(1\) or \(2\) since the fibers of \(\widehat\pi\) have cardinality
\(2\). If the cardinality were \(1\), then \(\widehat\pi|_W\) would be an injective smooth
covering map, hence a diffeomorphism, and its inverse would be a global smooth section of
\(\widehat\pi\), inducing an orientation on \(M\)—contradicting nonorientability. Therefore the
cardinality is \(2\), which forces \(W = \widehat M\), so \(\widehat M\) is connected; and since
\(\widehat\pi\) is a local diffeomorphism with two-element fibers, it is a two-sheeted smooth
covering map.
The orientation covering is determined by \(M\) up to isomorphism, independent of the particular
construction used to build it.
Theorem (Uniqueness of the Orientation Covering)
Let \(M\) be a nonorientable connected smooth manifold, with or without boundary, and let
\(\widehat\pi : \widehat M \to M\) be its orientation covering. If \(\widetilde M\) is an
oriented smooth manifold admitting a two-sheeted smooth covering map
\(\widetilde\pi : \widetilde M \to M\), then there is a unique orientation-preserving
diffeomorphism \(\varphi : \widetilde M \to \widehat M\) such that
\(\widehat\pi \circ \varphi = \widetilde\pi\).
Proof Sketch.
Define \(\varphi\) by sending \(\tilde q \in \widetilde M\) to the pair
\(\bigl(\widetilde\pi(\tilde q),\ (d\widetilde\pi_{\tilde q})_{*}\mathcal{O}^{\widetilde M}_{\tilde q}\bigr)\),
where \(\mathcal{O}^{\widetilde M}\) is the given orientation of \(\widetilde M\) and
\((d\widetilde\pi_{\tilde q})_{*}\) pushes it forward through the isomorphism
\(d\widetilde\pi_{\tilde q}\) (an isomorphism because \(\widetilde\pi\), being a covering map, is
a local diffeomorphism). By construction \(\widehat\pi \circ \varphi = \widetilde\pi\), so
\(\varphi\) maps each fiber of \(\widetilde\pi\) into the corresponding fiber of
\(\widehat\pi\). Since both \(\widetilde\pi\) and \(\widehat\pi\) are local diffeomorphisms,
locally \(\varphi = \widehat\pi^{-1} \circ \widetilde\pi\) for suitable local sections, so
\(\varphi\) is smooth and is itself a local diffeomorphism.
It remains to see that \(\varphi\) is bijective on each two-element fiber. Fix \(p \in M\);
both \(\widetilde\pi^{-1}(p)\) and \(\widehat\pi^{-1}(p)\) have two elements, the latter being
the two orientations of \(T_pM\). Since \(M\) is nonorientable and connected, its orientation
covering \(\widehat M\) is connected by the Orientation Covering Theorem, and the same argument
shows \(\widetilde M\) is connected: a disconnected oriented two-sheeted cover would split into
two sheets each mapping diffeomorphically onto \(M\), whose inverse would be a global section
and would orient \(M\). A connected two-sheeted cover is normal, with deck group of order two
generated by the nontrivial automorphism \(\tau\) interchanging the two points of each fiber.
By the
criterion for the base of a normal cover,
\(M\) nonorientable forces the deck action to be orientation-reversing, so \(\tau\) reverses
the orientation of \(\widetilde M\). The two points \(\tilde q\) and \(\tau(\tilde q)\) of a
fiber therefore push their orientations forward to the two opposite orientations of
\(T_pM\); hence \(\varphi\) sends the two points of \(\widetilde\pi^{-1}(p)\) to the two
distinct points of \(\widehat\pi^{-1}(p)\), so it is fiberwise bijective. A bijective local
diffeomorphism is a diffeomorphism.
That \(\varphi\) is orientation-preserving is immediate from its definition: it carries
\(\mathcal{O}^{\widetilde M}_{\tilde q}\) to the tautological orientation
\((d\widetilde\pi_{\tilde q})_{*}\mathcal{O}^{\widetilde M}_{\tilde q}\) at
\(\varphi(\tilde q)\), which is precisely how \(\widehat M\) is oriented. For uniqueness,
suppose \(\varphi'\) is another orientation-preserving diffeomorphism over \(M\). At any point
\(\tilde q\), both \(\varphi(\tilde q)\) and \(\varphi'(\tilde q)\) lie in the same two-element
fiber \(\widehat\pi^{-1}(\widetilde\pi(\tilde q))\), and both must carry the orientation of
\(\widetilde M\) to the same orientation of \(T_{\widetilde\pi(\tilde q)}M\), since each is
orientation-preserving over \(M\); the two fiber points being distinguished exactly by their
orientation, \(\varphi(\tilde q) = \varphi'(\tilde q)\). Thus \(\varphi\) is unique.
A little covering-space theory now turns the connectedness criterion into a purely
group-theoretic sufficient condition for orientability. Recall that the index of a subgroup is the
number of its cosets.
Theorem (Simply Connected Manifolds Are Orientable)
Let \(M\) be a connected smooth manifold, with or without boundary, and suppose the
fundamental group
of \(M\) has no subgroup of
index two.
Then \(M\) is orientable. In particular, if \(M\) is
simply connected,
then it is orientable.
Proof.
Suppose \(M\) is not orientable. By the Orientation Covering Theorem its orientation covering
\(\widehat\pi : \widehat M \to M\) is then a connected two-sheeted covering. A connected
two-sheeted covering of \(M\) corresponds, under the covering-space dictionary, to an index-two
subgroup of the fundamental group of \(M\), namely the image of the fundamental group of the
cover. Thus nonorientability produces a subgroup of index two. Contrapositively, if the
fundamental group of \(M\) has no subgroup of index two, then \(M\) is orientable. A simply
connected manifold has trivial fundamental group, which has no subgroup of index two, so it is
orientable.
Orientability across the application domains
The index-two criterion settles several manifolds at once that recur as domains for learning on
curved spaces. Spheres \(S^n\) with \(n \geq 2\) are simply connected, hence orientable, as are
compact connected Lie groups that happen to be simply connected; orientability there is not an
assumption to be checked case by case but a consequence of topology. The nonorientable case is
equally consequential: when data lives on a nonorientable space such as the projective plane,
the oriented double cover is the canonical orientable space on which orientation-dependent
constructions become available, and its uniqueness means that this passage is forced rather
than chosen. The volume form assembled earlier, transported to such a cover, supplies the
invariant measure on which integration and averaging on these domains rely.