Sphere Theorem
Throughout, manifolds and maps are smooth unless stated otherwise. On a Riemannian manifold, \(\nabla\) denotes the Levi–Civita connection. The needed local constructions start with coordinate charts.
The sphere theorem asks how tightly positive sectional curvature must be pinched to determine the topology of a manifold. Under the hypotheses below, the curvature bounds force the manifold to be homeomorphic to a sphere. The proof combines an injectivity-radius estimate with comparison geometry to construct two regions that fit together as hemispheres do.
Theorem 1 (Sphere Theorem). Let \(M\) be a compact, simply connected differentiable manifold of \(\dim M = n\). Let \(K\) denote its sectional curvature, satisfying
If \(h = \frac14\), then \(M\) is homeomorphic to a sphere \(\mathbb S^n\).
Assume \(M\) is connected and without boundary, \(n\geq2\), and \(K_{\max}\) is the global maximum over all tangent two-planes. Rescale the metric by \(K_{\max}\); then sectional curvatures divide by \(K_{\max}\). Compactness turns the strict lower bound into
for some constant \(\delta\). All subsequent lengths refer to this rescaled metric.
Definition 2. The number \(h\) in (1) is called the pinching number.
More precisely, \(h\) is a lower pinching threshold in the displayed strict inequality; the actual global pinching ratio is \(K_{\min}/K_{\max}\). A single metric may satisfy the inequality for many smaller values of \(h\).
Remark 3. Rauch initiated the pinching approach, and Berger and Klingenberg established the strict quarter-pinched topological sphere theorem. The threshold cannot be lowered in a theorem covering all dimensions: complex projective spaces \(\mathbb{CP}^m\), \(m\geq2\), have their standard metric with sectional curvatures in \([1,4]\), but are not spheres (for instance, their second integral cohomology is nonzero). Thus equality at the quarter-pinching threshold allows additional spaces. See the introduction of Brendle–Schoen for this historical context and the rank-one symmetric models.
We first obtain the geometric estimates needed for this two-region description. The first result is the estimate of injectivity radius in manifolds of even dimension, which complements Proposition 12 (whose proof requires using some elements of Morse theory).
Proposition 4 (Klingenberg). Let \(M\) be a simply connected, compact, differentiable manifold of \(\dim M = n \geq 2\), and \(n\) is even. Furthermore, suppose \(0 < K \leq 1\). Then, \(\iota (M) \geq \pi\).
Proof. Suppose \(\ell=\operatorname{inj}(M)<\pi\). Compactness gives a pair \(p,q\) realizing this nearest cut distance. Since \(K\leq1\), the first conjugate time is at least \(\pi\); hence no minimizing segment of length \(\ell\) from \(p\) to \(q\) is conjugate. The nearest-cut-point lemma gives exactly two such segments. Its global injectivity-radius version joins them smoothly into a closed unit-speed geodesic \(\gamma:[0,2\ell]\to M\), with \(p=\gamma(0)\), \(q=\gamma(\ell)\).
A simply connected manifold is orientable, since its orientation double cover is trivial. Parallel transport round \(\gamma\) therefore preserves orientation. It fixes \(T=\gamma'\) and acts on the odd-dimensional normal space by an orientation-preserving orthogonal map. The nonreal eigenvalues occur in conjugate pairs, and the product of all eigenvalues is one; in odd dimension there must be an eigenvalue \(+1\). Its unit eigenvector extends to a nonzero periodic parallel normal field \(V\).
For the periodic variation \(\gamma_s(t)=\exp_{\gamma(t)}(sV(t))\), define \(E(s)=\frac12\int_0^{2\ell}|\partial_t\gamma_s|^2dt\). There are no endpoint terms, and
Thus \(E(s)<E(0)=\ell\) for all sufficiently small nonzero \(s\). Cauchy–Schwarz gives \(L(\gamma_s)^2\leq4\ell E(s)<4\ell^2\), so these loops have length less than \(2\ell\).
Choose \(t_s\) maximizing \(d(\gamma_s(0),\gamma_s(t))\), and put \(q_s=\gamma_s(t_s)\). The two arcs of the loop give
At \(s=0\), the distance along the original loop is exactly \(\min(t,2\ell-t)\): both half-segments minimize and every shorter segment is inside the injectivity radius. Hence its unique maximizing parameter is \(\ell\). Uniform convergence of the loops and distances implies \(t_s\to\ell\), \(q_s\to q\).
There is a unique minimizing segment from \(q_s\) to \(\gamma_s(0)\); let \(u_s\) be its initial unit tangent. Its length is less than the injectivity radius, so the distance is smooth near \(q_s\). At the maximum along \(\gamma_s\), first variation gives \(\langle u_s,\gamma_s'(t_s)\rangle=0\). Pass to a convergent subsequence of these unit vectors in a local tangent trivialization. Smooth dependence of geodesics and convergence of their lengths show that the limit \(u\) starts a minimizing segment from \(q\) to \(p\), and \(u\perp T(\ell)\). But the only two minimizing segments between these nearest cut points start at \(q\) in the directions \(T(\ell)\) and \(-T(\ell)\). This is a contradiction. Therefore \(\operatorname{inj}(M)\geq\pi\). ◻
Lemma 5 (Berger). Let \(M\) be a compact differentiable manifold, and \(p, q \in M\) such that \(d(p,q) = \operatorname{diam}(M)\). Then, for all \(w \in T_p M\), there exists a minimizing geodesic \(\gamma\), where \(\gamma(0) = p\) connects to \(q\), such that \(\langle \gamma'(0), w \rangle \geq 0\).
Proof. This is the critical-point property of a maximum of \(d_q\), including possible cut points. Here is a direct first-variation argument. If the conclusion failed for a nonzero \(w\), compactness of the set of minimizing initial directions from \(p\) to \(q\) would give a uniform negative bound on \(\langle u,w\rangle\). Put \(p_t=\exp_p(tw)\). Every sequence of minimizing segments from \(p_t\) to \(q\), as \(t\downarrow0\), has a subsequence converging to a minimizing segment from \(p\) to \(q\). Thus their unit initial directions \(u_t\), compared in local coordinates, satisfy \(\langle u_t,w_t\rangle\leq-c<0\) for small \(t\), where \(w_t\) is the velocity of \(p_t\).
Choose a fixed small \(\eta>0\), and let \(z_t\) be the point at distance \(\eta\) along such a segment from \(p_t\) to \(q\). Uniform normal neighborhoods give the Taylor estimate
with an error uniform over these directions. Since \(d(p_t,q)=\eta+d(z_t,q)\), the triangle inequality implies
This contradicts the diameter-maximizing property of \(p,q\). For \(w=0\), any minimizing segment works. ◻
Lemma 6 (Berger–Tsukamoto covering). Suppose \(M\) is compact and simply connected, with \(\frac14<\delta\leq K\leq1\), and \(d(p,q)=D=\operatorname{diam}M\). For every
one has \(M=B_r(p)\cup B_r(q)\). Both balls are normal coordinate balls, by the even- and odd-dimensional injectivity estimates. The lower endpoint of this interval must be strict: on the unit round sphere, two open balls of radius \(\pi/2\) centered at antipodes omit the equator.
Proof. The interval is nonempty because \(\delta>1/4\). Suppose a point \(u\) lies outside both balls. Put \(a=d(p,u)\), \(c=d(q,u)\), so \(r\leq a,c\leq D\). Choose a minimizing segment from \(p\) to \(u\), and use Lemma 5 to choose a minimizing segment from \(p\) to \(q\) making angle \(\alpha\leq\pi/2\) with it. Bonnet–Myers gives \(D\leq\pi/\sqrt\delta\). Toponogov hinge comparison and the spherical cosine law yield
Both cosine factors in the product are strictly negative, because \(a,D\geq r>\pi/(2\sqrt\delta)\), while the sine term is nonnegative. This remains valid at an antipodal model endpoint by continuity. But \(c\geq r>\pi/(2\sqrt\delta)\) implies \(\cos(\sqrt\delta c)<0\), a contradiction. The injectivity estimates give \(r<\operatorname{inj}(M)\), so the two balls are diffeomorphic to Euclidean balls. ◻
Remark 7. The covering proof uses the global hinge comparison. Intersections of metric spheres do not themselves provide totally geodesic normal neighborhoods, so the archival localization claim cannot justify replacing this step by a local Rauch comparison. The diameter bound used here is \(\pi/\sqrt\delta\).
Lemma 8. Fix \(r\) in the interval of Lemma 6. For each unit \(v\in T_pM\), the segment \(\gamma_v(t)=\exp_p(tv)\), \(0\leq t\leq r\), contains exactly one point \(m(v)=\gamma_v(\rho(v))\) with
The functions \(\rho\) and \(m\) are continuous in \(v\). The analogous assertion holds with \(p,q\) exchanged.
Proof. Since \(r<\operatorname{inj}(M)\), \(d(p,\gamma_v(t))=t\). The function
is continuous and nonincreasing: for \(s<t\), the triangle inequality gives \(d(q,\gamma_v(t))\leq d(q,\gamma_v(s))+t-s\). It is positive at zero and negative at \(r\), since \(\gamma_v(r)\notin B_r(p)\) and the covering lemma puts it in \(B_r(q)\). Thus a zero exists in \((0,r)\).
If there were two zeros \(0<s<t<r\), then
A minimizing segment from \(q\) to \(\gamma_v(s)\), followed by \(\gamma_v|_{[s,t]}\), would consequently be minimizing. A minimizing broken geodesic has no corner, so it agrees by uniqueness of the geodesic equation with the segment from \(p\) to \(\gamma_v(t)\). Its initial point at distance \(t\) from the terminal point must then be \(p\), giving \(q=p\), a contradiction. Therefore there is one zero, with strict opposite signs on its two sides.
If \(v_j\to v\), every convergent subsequence of \(\rho(v_j)\in[0,r]\) has limit a zero of \(\varphi_v\), by continuity of the exponential map and distance. Uniqueness forces that limit to be \(\rho(v)\); hence the full sequence converges. This proves continuity, including uniform separation of the root from zero and \(r\) by compactness of the unit sphere. ◻
Proof of Theorem 1. Normalize the curvature as above and choose \(r\in(\pi/(2\sqrt\delta),\pi)\). Let \(p,q\) realize the diameter. The injectivity estimate is Proposition 4 in even dimensions and the preceding note's odd-dimensional Klingenberg estimate in odd dimensions. Define
These are closed, \(D_p\cup D_q=M\), and \(D_p\cap D_q=\Sigma\). The covering lemma implies \(D_p\subset B_r(p)\): if a point lies instead in \(B_r(q)\), its distance to \(p\) is no larger and hence is still less than \(r\). Likewise \(D_q\subset B_r(q)\).
For \(v\in S_pM\), Lemma 8 shows that \(\varphi_v\) is positive before \(\rho(v)\) and negative after it. Thus
In particular, the continuous radial map
is bijective. Continuity at zero follows from the uniform bound \(\rho(v)<r\). Compactness of the domain and the Hausdorff property make it a homeomorphism. Its unit-sphere image is exactly \(\Sigma=\partial D_p\), since the strict sign change occurs on every radial segment. The same construction gives a homeomorphism \(F_q\) from a closed unit ball onto \(D_q\), with \(\partial D_q=\Sigma\).
To make the gluing explicit, the two boundary parameterizations determine a homeomorphism
Extend it radially by \(H(sv)=s\,h(v)\), \(H(0)=0\). This is a homeomorphism of closed balls; the continuity of the inverse follows from the same radial formula. View \(\mathbb S^n\) as two closed \(n\)-balls with equal boundary points identified. Map the northern ball by \(F_p\) and the southern ball by \(F_q\circ H\). The two maps agree on the common boundary, so they define a continuous map onto \(M\). They are injective on the separate interiors, whose images are disjoint by the strict distance inequalities, and on the common boundary. Hence the map is a continuous bijection from the compact sphere onto the Hausdorff manifold \(M\), and is a homeomorphism.
The two hemisphere images meet on their common boundary; they are not disjoint when that boundary is included. Also, the equidistant distance is \(\rho(v)<r\), generally depending on direction, rather than the constant \(r\). ◻
Remark 9. The preceding radial gluing proves a homeomorphism; it does not supply smoothness across the interface or at the two centers. The stronger conclusion that \(M\) is diffeomorphic to a sphere follows from the differentiable sphere theorem. Brendle–Schoen, Theorem 1, proves the more general result that a compact manifold of dimension at least four with strictly pointwise quarter-pinched curvature is diffeomorphic to a spherical space form; simple connectivity makes that space form the sphere. The low-dimensional cases use the corresponding surface and three-dimensional results.
Petersen–Tao, Theorem 1.1, gives a dimension-dependent \(\varepsilon(n)>0\) for which a closed simply connected manifold with sectional curvatures in \([1/4-\varepsilon(n),1]\) is diffeomorphic to a sphere or a compact rank-one symmetric space. These are cited extensions; their Ricci-flow proofs are not part of the elementary proof above.
Remark 10. Hamilton's 1982 theorem, Theorem 1.1, says that a closed three-manifold with positive Ricci curvature admits a metric of constant positive sectional curvature. Thus a connected such manifold is diffeomorphic to a spherical space form \(\mathbb S^3/\Gamma\), and is a sphere if it is simply connected. Without that hypothesis, \(\mathbb{RP}^3\) and lens spaces are counterexamples to the sphere-only conclusion.
In dimension four, Chang–Gursky–Yang, Theorem A and its following remarks, assumes a smooth closed manifold with positive Yamabe constant \(Y(M,[g])\) of its conformal class and
Here \(W_g\) is the Weyl tensor, distinct from \(Y(M,[g])\), and the norm is the full \((0,4)\)-tensor norm \(|W|^2=\sum_{i,j,k,l}W_{ijkl}^2\) in an orthonormal frame. The conclusion is that \(M\) is diffeomorphic to \(\mathbb S^4\) or \(\mathbb{RP}^4\). The norm convention matters: viewing Weyl curvature as an operator on two-forms changes its squared norm by a factor of four. These external theorems have different hypotheses from sectional quarter-pinching.
Reading: Carmo (1992), Riemannian geometry.