Toponogov 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.
Rauch comparison controls infinitesimal separation along geodesics. Toponogov’s theorem gives a finite comparison: it relates the side lengths and angles of a geodesic triangle to those of a triangle in a constant-curvature surface. It extends the local estimates supplied by the Rauch comparison theorem (Theorem 3). We begin with a definition.
Definition 1. A geodesic triangle on a differentiable manifold \(M\) is a set of three minimizing geodesic segments, each parametrized by arclength, denoted
such that \(\gamma_i (\ell_i) = \gamma_{i+1} (0)\) for \(i=1,2\), \(\gamma_3 (\ell_3) = \gamma_1 (0)\), and \(\ell_i + \ell_{i+1} \geq \ell_{i+2}\). The endpoints of the geodesic segments are called the vertices of the geodesic triangle. Let \(\sphericalangle (-\gamma'_i(\ell_i), \gamma'_{i+1} (0))\) denote the angle of corresponding vertices \(i=1,2\), and \(\sphericalangle (-\gamma'_3(\ell_3), \gamma'_1(0))\) denote the angle of vertex \(i=3\).
We use two formulations. The triangle formulation fixes three side lengths and compares angles. The hinge formulation fixes two side lengths and their included angle, then compares the distance between the remaining endpoints.
Theorem 2 (Toponogov). Let \(M\) be a complete connected Riemannian manifold of dimension at least two, with \(K\geq H\). Let \(M_H^2\) be the complete simply connected surface of constant curvature \(H\), and put \(D_H=\pi/\sqrt H\) if \(H>0\), \(D_H=\infty\) otherwise.
(a) Triangle comparison. A triangle made of three minimizing segments has a comparison triangle in \(M_H^2\) with the same side lengths, whose corresponding angles are no greater. For positive curvature, its sides are at most \(D_H\). The comparison is unique up to a model isometry when no side has length \(D_H\); antipodal and zero-side degeneracies are understood by limits.
(b) Hinge comparison. Let a minimizing segment of length \(a\) join \(x\) to \(p\), and let a unit-speed geodesic \(\gamma:[0,b]\to M\) start at \(x\), making angle \(\alpha\) with that segment. If \(H>0\), assume \(b\leq D_H\). Construct the model hinge with lengths \(a,b\) and angle \(\alpha\). Then
The second leg need not minimize. If \(H>0\), Bonnet–Myers already ensures \(a\leq D_H\).
Remark 3. Part (a) fixes three side lengths and compares angles; part (b) fixes two lengths and their included angle and compares the opposite distance. These formulations are related by the model cosine laws. In part (a), all three sides minimize, consistently with Definition 1. A statement allowing a nonminimizing third side requires separate triangle-inequality hypotheses.
Remark 4. The two segments and their angle in part (b) form a hinge. The proof below first records the source's valid local comparison, then supplies a global argument. The archival thin-hinge discussion does not justify a uniform subdivision, and its proposed parallel-normal representation of a fixed endpoint is not generally valid; those steps cannot serve as the globalization proof.
Proof of Theorem 2. Local comparison. In a sufficiently small normal neighborhood of a hinge vertex, map the corresponding model configuration into \(M\) by
where the linear isometric embedding \(i\) identifies the two initial directions. The length form of Rauch comparison gives \(L(F\bar c)\leq L(\bar c)\) for the model closing segment \(\bar c\). Since \(F\bar c\) joins the required endpoints,
The normal neighborhoods must contain the entire closing segment, which is ensured by taking the configuration sufficiently small. This proves the local hinge statement. We now prove its global version without assuming a large configuration lies in one such neighborhood.
A scalar form of Hessian comparison. For a curvature parameter \(k\), define
Thus \(h_k'=s_k,\ h_k''=c_k=1-kh_k\). It is strictly increasing for \(0<r<D_k\). Assume for the moment that \(K\geq k\) and, when \(k>0\), that all distances under consideration are strictly less than \(D_k\). At a point where \(r=d(p,\cdot)\) is smooth, Hessian comparison gives
Indeed, the chain rule contributes \(c_k(r)\,dr\otimes dr\), exactly completing the radial term. Along a unit-speed geodesic, \(u(t)=h_k(d(p,\gamma(t)))\) therefore satisfies \(u''+ku\leq1\) at smooth points.
Cut points and upper supports. Fix \(z\neq p\) and choose a minimizing segment from \(p\) to \(z\), of length \(r\). Let \(p_\varepsilon\) be its point at distance \(\varepsilon\) from \(p\). The function
is an upper support for \(d(p,y)\) at \(z\), by the triangle inequality. It is smooth near \(z\): the reversed segment from \(z\) to \(p\) minimizes beyond \(p_\varepsilon\), so \(p_\varepsilon\) is before its cut time; symmetry of the cut relation then places \(z\) outside the cut locus of \(p_\varepsilon\). Hessian comparison at \(z\), now with radial distance \(r-\varepsilon\), gives for a unit vector \(W\), writing \(v=dr_\varepsilon(W)\),
Here \(|v|\leq1\) and the error tends to zero as \(\varepsilon\downarrow0\). The increasing function \(h_k\) preserves the upper-support property. At \(p\) itself, \(h_k(d(p,\cdot))\) is smooth in normal coordinates: its power series is a smooth function of the squared radial norm, with Hessian \(g\) at the origin. Thus every point of the parameter interval has smooth upper supports with second derivative at most \(1-ku(t)+\epsilon\), for every \(\epsilon>0\).
This support construction also appears in Eschenburg, Comparison Theorems in Riemannian Geometry, §6, pp. 24–25. The following one-dimensional argument spells out why it suffices here.
From supports to the scalar inequality. On a compact parameter interval choose \(G\in C^2\) with \(G''=1-ku\). Then \(w=u-G\) has upper supports with second derivative at most any prescribed \(\epsilon>0\). This implies that \(w\) is concave. Otherwise, on some \([a,b]\), it lies strictly below its endpoint chord at an interior point. For a small \(\eta>0\), the quadratic \(q(t)=\text{chord}(t)+\eta(t-a)(t-b)\) still lies above \(w\) somewhere. The negative minimum of \(w-q\) occurs inside \((a,b)\). An upper support \(w_\epsilon\) there also makes \(w_\epsilon-q\) attain a local minimum, forcing \(w_\epsilon''\geq q''=2\eta\), contrary to \(\epsilon<2\eta\).
Consequently \(w''\) is a nonpositive measure, and
in the distributional sense. The same concavity gives one-sided derivatives; they are finite here because the distance and \(h_k\) are locally Lipschitz. Variation of constants, valid also for this measure by integrating the scalar equation, yields
One can verify the formula by differentiating distributions and its initial values; uniqueness of the homogeneous scalar equation then gives it. Since \(s_k\geq0\) before \(D_k\), the measure term is nonpositive.
Apply the initial data. Let \(a=d(p,x)>0\). Use the chosen minimizing leg from \(x\) to \(p\) in the upper-support construction at \(x\). The first variation formula gives
The model function \(\bar u(t)=h_k(d(\bar p,\bar\gamma(t)))\) solves
These equations follow either from the constant-curvature Hessian identity or the model cosine laws; the modified distance is smooth even at its zero and, on the sphere, at the antipode. Therefore
Monotonicity of \(h_k\) on the distance range gives the hinge inequality. The case \(a=0\) is simply \(d(x,\gamma(t))\leq t\).
For \(H\leq0\), take \(k=H\). For \(H>0\), take \(0<k<H\) and let \(k\uparrow H\). Bonnet–Myers gives \(\operatorname{diam}M\leq D_H<D_k\), while \(b\leq D_H<D_k\), so every preceding step applies, including the endpoint \(b=D_H\). Model distances vary continuously with \(k\), proving (b) at \(H\).
Recover the triangle statement. Write \(a,b\) for the sides adjacent to a vertex, \(c\) for the opposite side, and \(\alpha\) for its actual angle. In a model hinge, the opposite distance \(F_k(a,b,\theta)\) is continuous and strictly increasing in \(0<\theta<\pi\) whenever the adjacent sides are nonzero and shorter than \(D_k\). This follows directly from
The triangle inequality and (b) give \(|a-b|=F_k(a,b,0)\leq c\leq F_k(a,b,\alpha)\). Thus there is a model angle \(\bar\alpha\leq\alpha\) giving the same third side. The three side lengths determine this model triangle up to isometry. Repeating the hinge comparison at each vertex gives all three angle inequalities for that one triangle.
For positive \(H\), perform this construction at \(k<H\). Normalize the first vertex and initial direction in the spherical models and let \(k\uparrow H\). Compactness of the remaining initial directions gives a limiting triangle with the same side lengths and all the angle inequalities. This also handles antipodal or degenerate configurations, for which uniqueness may fail. No assertion of smooth dependence of ordinary distance at a coincident or antipodal endpoint is needed. ◻
Reading: Carmo (1992), Riemannian geometry.