This course develops comparison geometry, the part of geometric analysis that studies how curvature bounds control distance, volume, topology, and global shape. The central idea is to compare a manifold to explicit model spaces of constant or controlled curvature, then use that comparison to derive qualitative and quantitative consequences. Along the way, the course connects local differential geometry with global phenomena such as diameter bounds, fundamental group restrictions, volume growth, rigidity, and convergence of spaces. The chapters are arranged to build a toolkit in layers. After introducing model geometries and curvature bounds, the course studies Jacobi fields and the index form as the basic variational machinery behind comparison results. Rauch comparison, Hessian and Laplacian comparison, and triangle comparison then turn those local tools into geometric estimates, while Bishop-Gromov comparison and its analytic consequences extend them to volume and function theory. Later chapters focus on global structure under Ricci and sectional curvature assumptions, including splitting and soul theorems, before moving to Gromov-Hausdorff convergence, compactness, and rigidity. The closing chapter ties these themes together by showing how the same comparison estimates reappear in rigidity arguments, compactness statements, and limiting examples. # Introduction Comparison geometry asks how much global geometry is forced by local curvature inequalities. The guiding theme of this course is that bounds on sectional curvature, Ricci curvature, or scalar curvature become useful only after they are translated into statements about geodesics, distance functions, volumes, and compactness. This introduction fixes the language and the scale of the course: we are not trying to classify all Riemannian manifolds, but to build a toolkit for comparing an unknown manifold with a small list of model spaces. The prerequisites are the standard objects of Riemannian geometry: smooth manifolds, Riemannian metrics, the Levi-Civita connection, geodesics, the exponential map, curvature tensors, and basic topology. Some measure theory enters through volume comparison, and elementary functional analysis helps with compactness arguments, but the central arguments are geometric. Later courses in minimal surfaces, harmonic maps, Ricci flow, and metric geometry use these comparison tools as background. ## What Comparison Geometry Tries to Control What can a curvature inequality say about points that are far apart? A curvature tensor is defined infinitesimally, but the conclusions we want are global: whether geodesics stop minimizing, whether the manifold has finite diameter, how quickly balls grow, and what compactness properties a family of manifolds has. The bridge from local curvature to global shape is built from the variation theory of geodesics. [explanation: Local To Global Strategy] A typical comparison theorem begins with a local differential inequality along geodesics. Jacobi fields encode how nearby geodesics separate, and the index form records the second variation of length or energy. Once this infinitesimal information is compared with the corresponding calculation in a constant-curvature model space, it can be integrated along geodesics to control conjugate points, distance functions, triangles, or volume elements. This pattern appears throughout the course. Sectional curvature comparison controls the behaviour of geodesic triangles and distance functions; Ricci curvature comparison controls the Jacobian of the exponential map and hence the volume of balls; lower Ricci bounds combined with non-collapsing hypotheses give compactness statements for sequences of manifolds. [/explanation] This explanation also shows why comparison geometry is not merely a collection of estimates. The estimates are organized around geometric mechanisms: geodesic variation, curvature operators, and model spaces. We therefore begin by fixing the objects that will be compared. ## Standing Geometric Conventions Which version of the basic Riemannian objects will be used in these notes? Since the same symbols recur in every chapter, we set them once. Throughout the course, manifolds are smooth and connected unless a statement says otherwise. [definition: Riemannian Manifold] A Riemannian manifold is a smooth manifold $M$ equipped with a smooth Riemannian metric $g$, where each $g_p$ is an [inner product](/page/Inner%20Product) on $T_pM$ depending smoothly on $p \in M$. [/definition] The metric converts tangent vectors into lengths and makes it possible to define distance, volume, and curvature. The curves most sensitive to these structures are the locally length-minimizing paths determined by the Levi-Civita connection, so the next definition names the curves along which comparison arguments are performed. When several metrics are under discussion, the notation $d_g$ denotes the associated distance function and $d\operatorname{vol}_g$ denotes the Riemannian volume measure. [definition: Geodesic] Let $(M,g)$ be a Riemannian manifold with Levi-Civita connection $\nabla$. A smooth curve $\gamma:I\to M$ is a geodesic if \begin{align*} \nabla_{\dot\gamma}\dot\gamma = 0 \end{align*} on $I$. [/definition] Geodesics are the paths along which comparison arguments are performed. They are also the source of the first global obstructions: two geodesics may meet again, a geodesic may stop minimizing before it ceases to exist, and [normal coordinates](/theorems/2713) may fail at the cut locus. To make global statements rather than only local ones, we need an assumption that geodesics do not end because the parameter interval was too short. [definition: Complete Riemannian Manifold] A Riemannian manifold $(M,g)$ is complete if every geodesic $\gamma:(a,b)\to M$ extends to a geodesic defined on all of $\mathbb R$. [/definition] Completeness is the background assumption for many global theorems. It allows geodesics to extend for all time, and through the Hopf-Rinow theorem it connects the geodesic picture with metric compactness of closed bounded sets. [remark: Notation For Curvature] The Riemann curvature tensor is denoted by $R$, Ricci curvature by $\operatorname{Ric}$, scalar curvature by $S$, and sectional curvature by $K(\sigma)$ for a $2$-plane $\sigma \subseteq T_pM$. For a unit speed geodesic $\gamma$, curvature terms along $\gamma$ are often viewed as endomorphisms of $\dot\gamma(t)^\perp$. [/remark] These conventions are enough to state the types of hypotheses that drive the course. The next question is why different curvatures lead to different comparison theorems. ## Curvature Bounds As Data What information is contained in $K \ge k$, $K \le k$, or $\operatorname{Ric}\ge (n-1)k g$? These inequalities do not have the same strength. Sectional curvature sees every two-dimensional direction, Ricci curvature sees an averaged trace over directions orthogonal to a vector, and scalar curvature is still more averaged. [definition: Sectional Curvature Bound] Let $(M,g)$ be a Riemannian manifold. We write $K \ge k$ if $K(\sigma) \ge k$ for every point $p \in M$ and every $2$-plane $\sigma \subseteq T_pM$, and we write $K \le k$ if $K(\sigma) \le k$ for every such $\sigma$. [/definition] Sectional curvature bounds are the most rigid comparison hypotheses in this course. They control how geodesics spread in every two-dimensional direction and therefore support triangle comparison, Rauch comparison, and Toponogov-type results. Many analytic estimates need less directional information; for those, the relevant input is the trace of sectional curvature in directions orthogonal to a vector. [definition: Ricci Curvature Lower Bound] Let $(M,g)$ be an $n$-dimensional Riemannian manifold. We write $\operatorname{Ric}\ge (n-1)k g$ if \begin{align*} \operatorname{Ric}(v,v) \ge (n-1)k |v|^2 \end{align*} for every tangent vector $v \in TM$. [/definition] Ricci lower bounds are weaker than sectional lower bounds but still retain strong analytic consequences. They govern volume growth, mean curvature of geodesic spheres, heat kernel estimates in later analysis courses, and compactness phenomena under suitable normalizations. [example: Three Levels Of Curvature Information] On a surface, fix $p\in M$ and an [orthonormal basis](/page/Orthonormal%20Basis) $e_1,e_2$ of $T_pM$. There is only one tangent $2$-plane, so write \begin{align*} K(p)=K(\operatorname{span}\{e_1,e_2\}). \end{align*} By the definition of Ricci curvature as the trace of sectional curvatures through a unit vector, \begin{align*} \operatorname{Ric}(e_1,e_1)&=K(\operatorname{span}\{e_1,e_2\})=K(p),\\ \operatorname{Ric}(e_2,e_2)&=K(\operatorname{span}\{e_2,e_1\})=K(p). \end{align*} Also $\operatorname{Ric}(e_1,e_2)=0$ in this orthonormal basis, so for $v=ae_1+be_2$, \begin{align*} \operatorname{Ric}(v,v) &=a^2\operatorname{Ric}(e_1,e_1)+2ab\operatorname{Ric}(e_1,e_2)+b^2\operatorname{Ric}(e_2,e_2)\\ &=a^2K(p)+2ab\cdot 0+b^2K(p)\\ &=K(p)(a^2+b^2)\\ &=K(p)|v|^2. \end{align*} Thus, in dimension $2$, Ricci curvature is exactly sectional curvature times the metric. In dimension $n\ge 3$, the same trace contains an average over many planes. If $e_1,\dots,e_n$ is an orthonormal basis and $e_1$ is fixed, then \begin{align*} \operatorname{Ric}(e_1,e_1) = \sum_{j=2}^n K(\operatorname{span}\{e_1,e_j\}). \end{align*} For instance, in dimension $3$, the values \begin{align*} K(\operatorname{span}\{e_1,e_2\})=-1, \qquad K(\operatorname{span}\{e_1,e_3\})=2 \end{align*} give \begin{align*} \operatorname{Ric}(e_1,e_1) =K(\operatorname{span}\{e_1,e_2\})+K(\operatorname{span}\{e_1,e_3\}) =-1+2=1. \end{align*} So a Ricci lower bound can hold in the $e_1$-direction even while one sectional curvature through $e_1$ is negative, because the trace records the sum rather than each summand separately. Scalar curvature traces once more: \begin{align*} S(p)=\sum_{i=1}^n \operatorname{Ric}(e_i,e_i) =2\sum_{1\le i<j\le n}K(\operatorname{span}\{e_i,e_j\}), \end{align*} so it loses still more directional information. Sectional curvature can therefore control individual geodesic triangles, while Ricci and scalar curvature remember only averaged data. [/example] The example explains why the course treats sectional and Ricci comparison as related but distinct branches. The common reference objects for both branches are the constant-curvature model spaces. ## Model Spaces And Normalized Functions Which spaces provide the comparison yardsticks? The basic models are simply connected complete Riemannian manifolds of constant sectional curvature. Their geodesics, volumes, and triangles can be computed explicitly, so they serve as the calibrated cases against which other manifolds are measured. [definition: Space Form] A space form of curvature $k$ is a complete connected Riemannian manifold with constant sectional curvature $k$. [/definition] The simply [connected space](/page/Connected%20Space) forms are Euclidean space, the round sphere after scaling, and hyperbolic space after scaling. Quotients by suitable groups give further complete examples, such as flat tori and hyperbolic surfaces. To compare volumes, Hessians, and Jacobi fields without rewriting the three constant-curvature cases separately, the standard radial functions are introduced next. [definition: Model Functions] For $k \in \mathbb R$, define $\operatorname{sn}_k:[0,\infty)\to\mathbb R$ by \begin{align*} \operatorname{sn}_k:r \mapsto \begin{cases} \frac{1}{\sqrt{k}}\sin(\sqrt{k}r), & k>0,\\ r, & k=0,\\ \frac{1}{\sqrt{-k}}\sinh(\sqrt{-k}r), & k<0. \end{cases} \end{align*} Define $D_k:=\{r\in[0,\infty):\operatorname{sn}_k(r)\ne 0\}$ and $\operatorname{ct}_k:D_k\to\mathbb R$ by \begin{align*} \operatorname{ct}_k:r \mapsto \frac{\operatorname{sn}_k'(r)}{\operatorname{sn}_k(r)}. \end{align*} [/definition] These functions package the constant-curvature calculations. In geodesic polar coordinates, powers of $\operatorname{sn}_k(r)$ describe model volume densities, while $\operatorname{ct}_k(r)$ appears in Hessian and mean-curvature comparison. [example: Volume Density In The Models] Let $d\theta$ denote the standard volume element on the unit sphere $S^{n-1}$. In geodesic polar coordinates on a constant-curvature model space, the radial volume element has the form \begin{align*} \operatorname{sn}_k(r)^{n-1}\,dr\,d\theta. \end{align*} For Euclidean space, $k=0$ and $\operatorname{sn}_0(r)=r$, so \begin{align*} \operatorname{sn}_0(r)^{n-1}\,dr\,d\theta =r^{n-1}\,dr\,d\theta. \end{align*} For the round sphere of curvature $1$, before reaching the first antipodal point at $r=\pi$, \begin{align*} \operatorname{sn}_1(r) =\frac{1}{\sqrt{1}}\sin(\sqrt{1}r) =\sin r, \end{align*} and therefore the polar density factor is \begin{align*} \operatorname{sn}_1(r)^{n-1} =(\sin r)^{n-1}. \end{align*} For hyperbolic space of curvature $-1$, \begin{align*} \operatorname{sn}_{-1}(r) =\frac{1}{\sqrt{-(-1)}}\sinh(\sqrt{-(-1)}r) =\sinh r, \end{align*} so the density factor is \begin{align*} \operatorname{sn}_{-1}(r)^{n-1} =(\sinh r)^{n-1}. \end{align*} Since \begin{align*} \sinh r=\frac{e^r-e^{-r}}{2}, \end{align*} the hyperbolic density satisfies \begin{align*} (\sinh r)^{n-1} =\left(\frac{e^r-e^{-r}}{2}\right)^{n-1} =\left(\frac{e^r(1-e^{-2r})}{2}\right)^{n-1}. \end{align*} Thus, for large $r$, the density is a positive constant multiple of $e^{(n-1)r}$ up to bounded factors, and geodesic balls in $\mathbb H^n$ grow exponentially in radius. [/example] This model computation foreshadows the Bishop-Gromov theorem. The full theorem replaces exact equality in a model space by a monotonicity statement under a Ricci lower bound. ## The Main Mechanisms Of The Course How do the model calculations enter the geometry of an arbitrary manifold? The answer is that one studies differential equations and variational forms along geodesics, then compares their solutions with the model equations. The first half of the course develops these mechanisms, while the second half applies them to global structure and compactness. [explanation: Jacobi Fields And Index Forms] Jacobi fields arise from variations through geodesics and satisfy a second-order linear differential equation involving the curvature tensor. They measure infinitesimal separation of geodesics. The index form is the second variation quadratic form for energy, and its sign detects conjugate points and minimizing properties. Rauch comparison and related results are obtained by comparing this differential equation or quadratic form with its constant-curvature analogue. [/explanation] Once Jacobi field comparison is available, it becomes possible to compare Hessians of distance functions and the shape of geodesic spheres. These estimates are the analytic doorway into volume comparison. [explanation: Triangle And Volume Comparison] Triangle comparison asks whether geodesic triangles in $M$ are thinner or fatter than triangles with the same side lengths in a model space. Volume comparison asks whether the volume of $B(p,r)$ grows faster or slower than the corresponding model ball. Both theories depend on controlling the exponential map, but triangle comparison usually requires sectional curvature bounds, while volume comparison is governed by Ricci bounds. [/explanation] The compactness mechanism enters after distance and volume estimates are in place: one can then discuss convergence of metric spaces and families of manifolds with uniform geometric bounds. [explanation: Compactness And Limits] Gromov-Hausdorff convergence provides a way to take limits of compact metric spaces without choosing coordinates. In Riemannian geometry, curvature and diameter bounds can force precompactness of families of manifolds. The limits need not be smooth manifolds, which is why comparison geometry naturally points toward metric geometry. [/explanation] These mechanisms give the course its order. Each chapter introduces a tool only after identifying the geometric obstruction it is meant to control. ## Course Roadmap How will the course fit together? It begins with curvature bounds and model geometries, then develops Jacobi fields and the index form. It then uses those tools for comparison theorems and ends with compactness and examples of geometric limits. [explanation: Chapter Progression] The first chapter reviews curvature bounds, space forms, completeness, injectivity radius, conjugate radius, and cut loci. The next chapter derives the Jacobi equation and the index form, then applies Sturm-type comparison. Rauch comparison leads into Hessian and Laplacian comparison for distance functions. Toponogov comparison gives a global triangle theorem under sectional curvature lower bounds. The Ricci side of the course develops Bishop-Gromov volume comparison, Myers' diameter theorem, and Cheeger-Gromoll-type splitting phenomena. The compactness chapter introduces the Gromov-Hausdorff viewpoint and explains how comparison estimates survive under limits. The closing material connects the course to geometric analysis beyond comparison geometry, including minimal submanifolds, harmonic maps, and Ricci flow. [/explanation] This roadmap is not a list of independent topics. Each later theorem reuses earlier information about geodesics, curvature, and models, so the technical chapters should be read in sequence. ## First Guiding Examples Which examples should remain in mind while reading the course? The most useful examples are not exotic; they are the constant-curvature spaces and their quotients. They show the sharpness of comparison theorems and warn against strengthening hypotheses beyond what curvature bounds permit. [example: Sphere As Positive Curvature Model] View $S^n$ as the unit sphere in $\mathbb R^{n+1}$ with the round metric, and fix $p\in S^n$ and a unit vector $v\in T_pS^n$. Since $v\perp p$, the curve \begin{align*} \gamma(t)=\cos(t)p+\sin(t)v \end{align*} lies on $S^n$, because \begin{align*} |\gamma(t)|^2 &=\langle \cos(t)p+\sin(t)v,\cos(t)p+\sin(t)v\rangle\\ &=\cos^2(t)|p|^2+2\cos(t)\sin(t)\langle p,v\rangle+\sin^2(t)|v|^2\\ &=\cos^2(t)+2\cos(t)\sin(t)\cdot 0+\sin^2(t)\\ &=1. \end{align*} Its velocity has constant length $1$: \begin{align*} \dot\gamma(t)&=-\sin(t)p+\cos(t)v,\\ |\dot\gamma(t)|^2 &=\sin^2(t)|p|^2-2\sin(t)\cos(t)\langle p,v\rangle+\cos^2(t)|v|^2\\ &=\sin^2(t)+\cos^2(t) =1. \end{align*} The acceleration is \begin{align*} \ddot\gamma(t) =-\cos(t)p-\sin(t)v =-\gamma(t), \end{align*} which is normal to $S^n$ at $\gamma(t)$, since the normal line to the unit sphere at $\gamma(t)$ is spanned by $\gamma(t)$. Hence the tangential component of $\ddot\gamma(t)$ is zero, so $\gamma$ is a unit speed geodesic for the round metric. At time $t=\pi$, \begin{align*} \gamma(\pi) &=\cos(\pi)p+\sin(\pi)v\\ &=-p+0\cdot v\\ &=-p. \end{align*} Thus every unit speed geodesic starting at $p$ reaches the antipodal point after length $\pi$. This is the model phenomenon behind positive-curvature diameter bounds: in the curvature-$1$ sphere, geodesics refocus at distance $\pi$, so conjugate points and diameter estimates are two sides of the same comparison picture. [/example] The sphere gives the compact positive-curvature picture. The flat examples show that zero curvature can still carry global topology. [example: Flat Torus As A Quotient Model] Let $T^n=\mathbb R^n/\mathbb Z^n$, and write $[x]$ for the class of $x\in\mathbb R^n$. The metric is induced by the quotient map $\pi:\mathbb R^n\to T^n$, so on each sufficiently small coordinate patch $\pi$ is an isometry from an open subset of Euclidean space. In Euclidean coordinates the metric coefficients are constant, \begin{align*} g_{ij}=\delta_{ij}, \end{align*} so their first derivatives vanish: \begin{align*} \frac{\partial g_{ij}}{\partial x^k}=0 \end{align*} for all $i,j,k$. Hence the Christoffel symbols in these coordinates are \begin{align*} \Gamma^k_{ij} =\frac12\sum_{\ell=1}^n g^{k\ell}\left( \frac{\partial g_{j\ell}}{\partial x^i} +\frac{\partial g_{i\ell}}{\partial x^j} -\frac{\partial g_{ij}}{\partial x^\ell} \right) =0, \end{align*} and the curvature tensor components are \begin{align*} R^\ell{}_{ijk} =\frac{\partial \Gamma^\ell_{jk}}{\partial x^i} -\frac{\partial \Gamma^\ell_{ik}}{\partial x^j} +\sum_m\Gamma^m_{jk}\Gamma^\ell_{im} -\sum_m\Gamma^m_{ik}\Gamma^\ell_{jm} =0. \end{align*} Therefore every sectional curvature of $T^n$ is $0$. The torus is compact because the image of the compact cube $[0,1]^n\subset\mathbb R^n$ under $\pi$ is all of $T^n$. It also has many closed geodesics. If $m\in\mathbb Z^n$ and $x\in\mathbb R^n$, define \begin{align*} \gamma(t)=[x+tm]. \end{align*} Its lift $\widetilde\gamma(t)=x+tm$ satisfies \begin{align*} \dot{\widetilde\gamma}(t)=m, \qquad \ddot{\widetilde\gamma}(t)=0, \end{align*} so $\widetilde\gamma$ is a Euclidean geodesic and its projection $\gamma$ is a geodesic on the flat quotient. Since $m\in\mathbb Z^n$, \begin{align*} \gamma(t+1) &=[x+(t+1)m]\\ &=[x+tm+m]\\ &=[x+tm], \end{align*} so $\gamma$ is closed. The universal cover $\mathbb R^n$ has no cut locus: for $x,y\in\mathbb R^n$, the segment $\sigma(t)=x+t(y-x)$ has length $|y-x|$, while every piecewise smooth curve $c$ from $x$ to $y$ satisfies \begin{align*} L(c) =\int_0^1 |c'(t)|\,dt \ge \left|\int_0^1 c'(t)\,dt\right| =|c(1)-c(0)| =|y-x|. \end{align*} On the quotient, minimizing can fail for topological reasons. For example, the geodesic $\gamma(t)=[te_1]$ from $[0]$ has length $t$ on $[0,t]$, but when $t>\frac12$, \begin{align*} d_{T^n}([0],[te_1]) \le |te_1-e_1| =| (t-1)e_1 | =1-t <t. \end{align*} Thus the same locally geodesic path stops minimizing after passing halfway around the circle direction, even though the curvature remains identically zero. [/example] Hyperbolic examples point in the opposite direction from the sphere. Negative curvature tends to separate geodesics and creates rapid volume growth. [example: Hyperbolic Space As Negative Curvature Model] Let $\mathbb H^n$ have constant sectional curvature $-1$. In geodesic polar coordinates about a point $p$, the model radial volume element is \begin{align*} \operatorname{sn}_{-1}(r)^{n-1}\,dr\,d\theta. \end{align*} By the definition of $\operatorname{sn}_k$, \begin{align*} \operatorname{sn}_{-1}(r) &=\frac{1}{\sqrt{-(-1)}}\sinh(\sqrt{-(-1)}\,r)\\ &=\sinh r, \end{align*} so the density factor is \begin{align*} \operatorname{sn}_{-1}(r)^{n-1} =(\sinh r)^{n-1}. \end{align*} Using \begin{align*} \sinh r=\frac{e^r-e^{-r}}{2} =\frac{e^r(1-e^{-2r})}{2}, \end{align*} we get \begin{align*} (\sinh r)^{n-1} &=\left(\frac{e^r(1-e^{-2r})}{2}\right)^{n-1}\\ &=2^{-(n-1)}e^{(n-1)r}(1-e^{-2r})^{n-1}. \end{align*} For $r\ge 1$, the factor $(1-e^{-2r})^{n-1}$ lies between $(1-e^{-2})^{n-1}$ and $1$, so the density is bounded above and below by positive constant multiples of $e^{(n-1)r}$. Thus geodesic spheres in $\mathbb H^n$ have exponentially growing area density, and geodesic balls have exponentially growing volume in the radius. Negative curvature also affects triangles: in the constant curvature $-1$ model, geodesics separate faster than in Euclidean space, so comparison triangles are thinner than their Euclidean counterparts. This is the basic model behaviour behind comparison theorems for upper sectional curvature bounds. [/example] These examples will return in nearly every chapter. When a theorem is stated with equality cases, the equality cases are usually model spaces or quotients of model spaces; when a hypothesis cannot be weakened, one of these examples often explains the obstruction. ## What To Track While Reading What should the reader monitor from lecture to lecture? The main task is to keep separate the type of curvature bound, the object being compared, and the conclusion obtained. A bound on sectional curvature, a bound on Ricci curvature, and a bound on scalar curvature are different inputs, even when their notation looks similar. [remark: Three Recurrent Questions] For each theorem in the course, ask which curvature is bounded, whether the bound is above or below, and whether completeness or compactness is assumed. Then identify the object being compared: Jacobi fields, index forms, triangles, Hessians, Laplacians, volumes, or metric limits. This prevents confusion between results whose statements look parallel but whose hypotheses have different strength. [/remark] The course begins in earnest with curvature bounds and model geometries. The purpose of this introductory chapter is to make the later pattern visible: local curvature inequalities become global geometry only after they pass through geodesics, variational formulas, and comparison with constant-curvature models. The introduction has now prepared the main theme: curvature bounds only become global geometry after they are translated through geodesics and comparison with model spaces. The next chapter fixes that language and begins asking how sectional, Ricci, and scalar curvature inequalities constrain the shape of a manifold. # 1. Curvature Bounds and Model Geometries This chapter fixes the language in which comparison geometry is phrased. The guiding question is how much global geometry can be recovered from inequalities involving sectional, Ricci, or scalar curvature. We also introduce the constant-curvature model spaces and the first places where geodesic normal coordinates stop being globally faithful. ## Curvature as Comparison Data Comparison theorems begin with a choice of which curvature quantity is controlled. Sectional curvature records the curvature of each two-plane and is the finest of the three basic invariants used here; Ricci curvature averages sectional curvature over directions orthogonal to a given vector; scalar curvature averages Ricci curvature over an orthonormal basis. The strength of a comparison result depends on which of these data are bounded and whether the bound is from below or above. To compare the geometry seen by a two-dimensional geodesic surface, we need a number attached to each two-plane in a tangent space. This motivates the following definition, the curvature datum that later appears in the scalar Jacobi equation along a geodesic. [definition: Sectional Curvature] Let $(M,g)$ be a Riemannian manifold and let $p \in M$. The sectional curvature at $p$ is the function \begin{align*} K_p:\operatorname{Gr}_2(T_pM) \to \mathbb R,\qquad \sigma \mapsto K_p(\sigma), \end{align*} defined on the Grassmannian of two-dimensional subspaces of $T_pM$ by \begin{align*} K_p(\sigma)=\frac{g(R(X,Y)Y,X)}{g(X,X)g(Y,Y)-g(X,Y)^2}, \end{align*} where $X,Y \in \sigma$ are linearly independent. [/definition] The denominator is the squared area of the parallelogram spanned by $X$ and $Y$, so the definition is independent of the chosen basis of $\sigma$. Before passing to averages, it is useful to anchor the three constant signs that will become the comparison models. [example: Constant Curvature Plane Values] For the standard Euclidean metric on $\mathbb R^n$, the curvature tensor is $R=0$, so for linearly independent $X,Y \in \sigma$, \begin{align*} K_p(\sigma) &=\frac{g(R(X,Y)Y,X)}{g(X,X)g(Y,Y)-g(X,Y)^2} \\ &=\frac{g(0,Y,X)}{g(X,X)g(Y,Y)-g(X,Y)^2} \\ &=0. \end{align*} For a space form with constant sectional curvature $k$, the curvature tensor is normalized by \begin{align*} R(X,Y)Y=k\bigl(g(Y,Y)X-g(X,Y)Y\bigr). \end{align*} Substituting this into the definition of sectional curvature gives \begin{align*} g(R(X,Y)Y,X) &=g\bigl(k(g(Y,Y)X-g(X,Y)Y),X\bigr) \\ &=k g(Y,Y)g(X,X)-k g(X,Y)g(Y,X) \\ &=k\bigl(g(X,X)g(Y,Y)-g(X,Y)^2\bigr), \end{align*} and therefore \begin{align*} K_p(\sigma) &=\frac{k\bigl(g(X,X)g(Y,Y)-g(X,Y)^2\bigr)} {g(X,X)g(Y,Y)-g(X,Y)^2} \\ &=k. \end{align*} Thus the unit round sphere $S^n$ has $K_p(\sigma)=1$ for every two-plane, while hyperbolic space $\mathbb H^n$ with curvature normalization $-1$ has $K_p(\sigma)=-1$ for every two-plane. The same normalization gives the Ricci and scalar curvature constants. If $e_1,\dots,e_{n-1}$ is an orthonormal basis of $v^\perp$, then \begin{align*} \operatorname{Ric}_p(v,v) &=\sum_{i=1}^{n-1}K_p(\operatorname{span}(v,e_i))|v|^2 \\ &=\sum_{i=1}^{n-1}k|v|^2 \\ &=(n-1)k|v|^2 \\ &=(n-1)k\,g(v,v). \end{align*} Hence $\operatorname{Ric}=(n-1)kg$. For an orthonormal basis $E_1,\dots,E_n$ of $T_pM$, \begin{align*} S(p) &=\sum_{j=1}^{n}\operatorname{Ric}_p(E_j,E_j) \\ &=\sum_{j=1}^{n}(n-1)k\,g(E_j,E_j) \\ &=\sum_{j=1}^{n}(n-1)k \\ &=n(n-1)k. \end{align*} The three model signs $k=0,1,-1$ are therefore the same signs that later appear in $\operatorname{sn}_k$, $\operatorname{ct}_k$, and the normalizations of Ricci and scalar curvature. [/example] This example shows the reference signs, but many comparison theorems need only an average of the sectional curvatures through a fixed direction. A volume estimate for geodesic balls cannot track every two-plane separately: the relevant infinitesimal object is the trace of the radial Jacobi equation over the $n-1$ angular directions. This motivates the following definition, which supplies the right datum when the question concerns volume growth rather than individual two-planes. [definition: Ricci Curvature] Let $(M,g)$ be an $n$-dimensional Riemannian manifold and let $p \in M$. The Ricci curvature at $p$ is the unique symmetric bilinear map \begin{align*} \operatorname{Ric}_p:T_pM \times T_pM \to \mathbb R,\qquad (v,w) \mapsto \operatorname{Ric}_p(v,w), \end{align*} whose diagonal values are \begin{align*} \operatorname{Ric}_p(v,v)=\sum_{i=1}^{n-1} K_p(\operatorname{span}(v,e_i)) |v|^2 \end{align*} for $v \ne 0$, where $e_1,\dots,e_{n-1}$ is an orthonormal basis of $v^\perp$, and $\operatorname{Ric}_p(0,0)=0$. [/definition] Ricci curvature is an averaged radial curvature, so it is well suited to results about volumes and mean curvature of distance spheres. It still remembers a direction, however, and some global variational questions only retain the total trace of curvature at a point. Taking one more trace gives the curvature quantity that survives when even the direction $v$ is no longer specified, which motivates the following definition. [definition: Scalar Curvature] Let $(M,g)$ be an $n$-dimensional Riemannian manifold. The scalar curvature is the function \begin{align*} S:M \to \mathbb R,\qquad p \mapsto S(p), \end{align*} defined by \begin{align*} S(p)=\sum_{i=1}^{n} \operatorname{Ric}_p(e_i,e_i), \end{align*} where $e_1,\dots,e_n$ is any orthonormal basis of $T_pM$. [/definition] Scalar curvature is a still coarser average. To state comparison hypotheses in a way that matches the model spaces, we need a convention for writing lower and upper curvature bounds. [remark: Lower and Upper Bounds] A statement such as $K \ge k$ means $K_p(\sigma) \ge k$ for every $p \in M$ and every two-plane $\sigma \subseteq T_pM$. A statement such as $\operatorname{Ric} \ge (n-1)k$ means $\operatorname{Ric}_p(v,v) \ge (n-1)k|v|^2$ for every $p \in M$ and $v \in T_pM$. [/remark] This normalisation makes space forms of constant sectional curvature $k$ satisfy $\operatorname{Ric}=(n-1)k g$ and $S=n(n-1)k$. The factor $(n-1)$ aligns Ricci comparison with the radial volume density in the $n$-dimensional model, and symmetric examples show why the average can carry useful information. [example: Projective Spaces and Averaging] Use the Fubini--Study metric on $\mathbb{CP}^m$ normalized so that every holomorphic line has sectional curvature $4$. If $J$ is the complex structure and $X,Y$ are orthonormal tangent vectors, the corresponding curvature formula is \begin{align*} K(\operatorname{span}(X,Y))=1+3g(JX,Y)^2. \end{align*} Thus a totally real plane, where $g(JX,Y)=0$, has \begin{align*} K(\operatorname{span}(X,Y))=1+3\cdot 0^2=1, \end{align*} while a holomorphic plane, where $Y=JX$, has \begin{align*} g(JX,Y)=g(JX,JX)=g(X,X)=1 \end{align*} and hence \begin{align*} K(\operatorname{span}(X,JX))=1+3\cdot 1^2=4. \end{align*} So the sectional curvature is positive but is not constant across two-planes. For the Ricci curvature, let $v$ be a unit tangent vector and choose an orthonormal basis of $v^\perp$ of the form \begin{align*} Jv,\ e_1,Je_1,\dots,e_{m-1},Je_{m-1}, \end{align*} with each $e_a,Je_a$ orthogonal to both $v$ and $Jv$. Then \begin{align*} \operatorname{Ric}(v,v) &=K(\operatorname{span}(v,Jv))+\sum_{a=1}^{m-1}K(\operatorname{span}(v,e_a)) +\sum_{a=1}^{m-1}K(\operatorname{span}(v,Je_a))\\ &=4+\sum_{a=1}^{m-1}1+\sum_{a=1}^{m-1}1\\ &=4+(m-1)+(m-1)\\ &=2m+2\\ &=2(m+1). \end{align*} By homogeneity in $v$, this gives \begin{align*} \operatorname{Ric}=2(m+1)g \end{align*} for this scaling. If the metric is rescaled to $\lambda g$, the same statement becomes \begin{align*} \operatorname{Ric}=\frac{2(m+1)}{\lambda}(\lambda g). \end{align*} Thus the individual sectional curvatures remember the angle a two-plane makes with the complex structure, while the Ricci trace averages those values to a single constant multiple of the metric. [/example] ## Constant-Curvature Model Geometries The next problem is to choose the reference spaces against which other manifolds are compared. The model spaces should have the same dimension, complete geodesics, maximal symmetry, and constant curvature equal to the numerical bound in the theorem under study. This leads to Euclidean space, the round sphere, and hyperbolic space. The phrase "constant curvature model" needs a formal meaning before it can serve as comparison data. This motivates the following definition, which packages completeness, connectedness, and the requirement that every sectional curvature have the same value. [definition: Space Form] A Riemannian manifold $(M,g)$ is a space form of curvature $k \in \mathbb R$ if it is complete, connected, and has sectional curvature $K_p(\sigma)=k$ for every $p \in M$ and every two-plane $\sigma \subseteq T_pM$. [/definition] The definition allows quotients such as flat tori and hyperbolic surfaces, so it is not yet a unique model for a given value of $k$. This motivates the following theorem: the simply connected model is unique and every other complete space form is locally obtained from it. [quotetheorem:5346] [citeproof:5346] This result explains why comparison formulas reduce to a single parameter $k$, but the hypotheses are doing real work. The restriction $n \ge 2$ matters because sectional curvature is attached to two-planes; in dimension one there are no two-planes and every connected complete one-dimensional Riemannian manifold is locally flat. Connectedness is part of the space-form definition so that a single model is being classified rather than a disjoint union of unrelated components. Completeness rules out incomplete open pieces of model spaces, such as an open ball in $\mathbb R^n$, which has the same local curvature as Euclidean space but is not the global Euclidean model. Simple connectedness rules out quotients: flat tori, real projective spaces with the round metric, and closed hyperbolic manifolds are complete constant-curvature manifolds but are not the simply connected models. Thus the theorem classifies the universal model spaces, not all complete space forms themselves; the missing information in the general case is the discrete group of isometries used to form the quotient. To turn those model spaces into usable estimates, we need a single radial function that records how geodesic spheres expand in the model geometry. In Euclidean space this coefficient is $r$, on the sphere it oscillates because geodesics refocus, and in hyperbolic space it grows like $\sinh r$ because nearby radial geodesics separate exponentially. Isolating this common coefficient gives the scalar solution of the constant-curvature Jacobi equation and supplies the reference term in volume, mean-curvature, and Laplacian comparison. This motivates the following definition. [definition: Model Sine Function] For $k \in \mathbb R$, the model sine function $\operatorname{sn}_k:[0,\infty) \to \mathbb R$ is \begin{align*} \operatorname{sn}_k(r)= \begin{cases} \frac{1}{\sqrt{k}}\sin(\sqrt{k}r), & k>0,\\ r, & k=0,\\ \frac{1}{\sqrt{-k}}\sinh(\sqrt{-k}r), & k<0. \end{cases} \end{align*} [/definition] The function $\operatorname{sn}_k$ is the length of an angular Jacobi field in the $k$-space form, normalised to vanish at the origin with derivative $1$. For comparison estimates one usually needs not the length itself but the rate at which geodesic spheres expand relative to their current size. That rate is undefined at zeros of $\operatorname{sn}_k$, so the logarithmic derivative must be treated as a function on its natural punctured domain. [definition: Model Cotangent Function] For $k \in \mathbb R$, let \begin{align*} D_k=\{r \in (0,\infty):\operatorname{sn}_k(r) \ne 0\}. \end{align*} The model cotangent function is the function \begin{align*} \operatorname{ct}_k:D_k \to \mathbb R \end{align*} defined by \begin{align*} \operatorname{ct}_k(r)=\frac{\operatorname{sn}_k'(r)}{\operatorname{sn}_k(r)}. \end{align*} [/definition] In positive curvature, $\operatorname{sn}_k$ first vanishes again at the radius \begin{align*} r=\frac{\pi}{\sqrt{k}}, \end{align*} reflecting antipodal focusing on the sphere. Listing the three signs side by side gives the formulas that will be substituted into later comparison inequalities. [example: Model Functions in the Three Signs] Using the three cases in the definition of $\operatorname{sn}_k$, we compute the corresponding logarithmic derivatives $\operatorname{ct}_k=\operatorname{sn}_k'/\operatorname{sn}_k$. For $k=1$, since $\sqrt{1}=1$, \begin{align*} \operatorname{sn}_1(r) &=\frac{1}{\sqrt{1}}\sin(\sqrt{1}r) \\ &=\sin r. \end{align*} Differentiating gives \begin{align*} \operatorname{sn}_1'(r)=\cos r, \end{align*} so on the interval where $\sin r\ne 0$ before the first zero, namely $0<r<\pi$, \begin{align*} \operatorname{ct}_1(r) &=\frac{\operatorname{sn}_1'(r)}{\operatorname{sn}_1(r)} \\ &=\frac{\cos r}{\sin r} \\ &=\cot r. \end{align*} For $k=0$, the definition gives \begin{align*} \operatorname{sn}_0(r)=r. \end{align*} Therefore \begin{align*} \operatorname{sn}_0'(r)=1, \end{align*} and for $r>0$, \begin{align*} \operatorname{ct}_0(r) &=\frac{\operatorname{sn}_0'(r)}{\operatorname{sn}_0(r)} \\ &=\frac{1}{r}. \end{align*} For $k=-1$, since $\sqrt{-(-1)}=1$, \begin{align*} \operatorname{sn}_{-1}(r) &=\frac{1}{\sqrt{-(-1)}}\sinh(\sqrt{-(-1)}r) \\ &=\sinh r. \end{align*} Differentiating gives \begin{align*} \operatorname{sn}_{-1}'(r)=\cosh r, \end{align*} so for $r>0$, where $\sinh r\ne 0$, \begin{align*} \operatorname{ct}_{-1}(r) &=\frac{\operatorname{sn}_{-1}'(r)}{\operatorname{sn}_{-1}(r)} \\ &=\frac{\cosh r}{\sinh r} \\ &=\coth r. \end{align*} Thus the three model cotangent terms are $\cot r$, $1/r$, and $\coth r$, matching the spherical, Euclidean, and hyperbolic radial geometries respectively. [/example] ## Completeness and the Limits of Normal Coordinates Normal coordinates turn Riemannian geometry near a point into geometry in a tangent space. The global problem is that the exponential map may fail to be defined for all tangent vectors, or it may be defined but fail to be one-to-one, or it may lose maximal rank at a conjugate point. Completeness and the cut-locus language separate these obstructions. The first obstruction is finite-time escape of geodesics. This motivates the following definition, which records the condition that every geodesic can be continued for arbitrary positive and negative time. [definition: Geodesic Completeness] A Riemannian manifold $(M,g)$ is geodesically complete if every maximal geodesic $\gamma:I \to M$ is defined on all of $\mathbb R$. [/definition] Geodesic completeness is a differential condition, while metric completeness is phrased using Cauchy sequences for the Riemannian distance. Comparison arguments constantly move between these viewpoints: Jacobi-field estimates live along geodesics, but compactness and existence questions are metric. To use completeness in global comparison arguments, one needs a precise bridge from complete geodesic extension to metric consequences: existence of minimizing geodesics, compactness of closed bounded sets, and the equivalence of the natural completeness conditions. Hopf--Rinow is the theorem that supplies this bridge. [quotetheorem:2726] [citeproof:2726] Hopf--Rinow gives existence of global geodesics and minimizing segments, but its hypotheses and conclusion should not be overread. Connectedness is needed because, on a disconnected manifold, two points in different components cannot be joined by any path or geodesic, regardless of completeness inside each component. Completeness is also essential: an open Euclidean ball is not complete, and geodesics can leave it in finite time even though its local curvature is zero. Even under Hopf--Rinow, minimizing geodesics need not be unique and the theorem gives no description of where uniqueness fails; that failure is measured by the cut locus and by the injectivity radius. This motivates the following definition, which measures how far the exponential map remains a coordinate chart around a fixed base point. [definition: Injectivity Radius] Let $(M,g)$ be a Riemannian manifold. The pointwise injectivity radius is the function \begin{align*} \operatorname{inj}:M \to [0,\infty],\qquad p \mapsto \operatorname{inj}(p), \end{align*} defined by \begin{align*} \operatorname{inj}(p)=\sup\{r>0: \exp_p|_{B(0,r)} \text{ is a diffeomorphism onto its image}\}. \end{align*} The injectivity radius of $M$ is the number \begin{align*} \operatorname{inj}(M)=\inf_{p \in M}\operatorname{inj}(p). \end{align*} [/definition] The injectivity radius measures how far geodesic polar coordinates remain unique around a point. One way those coordinates can fail is infinitesimal: even before two different minimizing geodesics reach the same point, the differential of the exponential map may lose rank. To separate this infinitesimal obstruction from the metric obstruction of multiple shortest paths, we introduce a second radius. The conjugate radius records the first distance at which Jacobi fields can make the exponential map singular along some geodesic from the base point. [definition: Conjugate Radius] Let $(M,g)$ be a Riemannian manifold. The pointwise conjugate radius is the function \begin{align*} \operatorname{conj}:M \to [0,\infty],\qquad p \mapsto \operatorname{conj}(p), \end{align*} defined by declaring $\operatorname{conj}(p)$ to be the infimum of all $t>0$ for which there exists a unit-speed geodesic $\gamma$ with $\gamma(0)=p$ such that $p$ and $\gamma(t)$ are conjugate along $\gamma$; if no such $t$ exists, then $\operatorname{conj}(p)=\infty$. [/definition] Conjugate points detect the failure of $d(\exp_p)$ to be invertible. A different obstruction is metric rather than infinitesimal: a geodesic can remain nonsingular while another geodesic of the same length arrives at the same endpoint, or while it simply stops being shortest beyond a certain time. The cut locus records this first loss of minimizing behaviour in each unit direction from the base point. [definition: Cut Locus] Let $(M,g)$ be a complete connected Riemannian manifold and let $p \in M$. Let $S_pM=\{v \in T_pM:|v|=1\}$. The cut-time function based at $p$ is the function \begin{align*} c:S_pM \to (0,\infty] \end{align*} defined by \begin{align*} c(v)=\sup\{t>0: d(p,\exp_p(sv))=s \text{ for every } 0\le s\le t\}. \end{align*} The cut locus of $p$ is the set \begin{align*} \operatorname{Cut}(p)=\{\exp_p(c(v)v):v \in S_pM,\ c(v)<\infty\}. \end{align*} [/definition] The cut locus marks where polar coordinates cease to be globally single-valued. For comparison arguments, the useful region is the [open set](/page/Open%20Set) before the cut time in every radial direction, because there the distance from $p$ is smooth and radial geodesics give unique angular labels. This prepares the coordinate form needed for volume density, mean-curvature, and Laplacian comparison. [illustration:geometric-analysis-i-geodesic-polar-coordinates] With the regular radial region isolated, the next task is to record precisely what remains true before the cut time. This is the region where each point is reached by a unique minimizing radial geodesic from $p$, so the variables $r$ and $v$ genuinely label points rather than competing geodesic representatives. The coordinate theorem states that the exponential map gives the polar chart on this regular region and that the Gauss lemma separates the radial direction from the angular directions. This separation is the structural input that later lets the angular Jacobian become the volume density and lets its logarithmic radial derivative become $\Delta r$. [quotetheorem:5347] [citeproof:5347] The theorem gives the local structure used throughout the course, but the exclusions in its statement are essential. The base point $p$ is removed because polar coordinates have an angular indeterminacy at $r=0$: every direction $v \in S_pM$ represents the same point before exponentiation. The cut locus is removed because at a cut point either two minimizing radial geodesics arrive at the same point or the exponential map has lost rank through conjugacy, so the coordinates cease to be single-valued or smooth. Thus the theorem supplies polar coordinates on the maximal regular radial region based at $p$; it does not assert that a complete manifold admits global polar coordinates. In the next comparison arguments, the decomposition $g=dr^2+g_r$ is used to identify the Riemannian volume density as the angular Jacobian of $\exp_p$, and to express $\Delta r$ as the radial derivative of the logarithm of that density. Those are the quantities compared with $\operatorname{sn}_k^{n-1}$ and $(n-1)\operatorname{ct}_k$. The round sphere is the basic case where loss of uniqueness and conjugate focusing occur together. [example: Cut Locus on the Round Sphere] Identify the unit sphere as \begin{align*} S^2=\{x\in \mathbb R^3:|x|=1\} \end{align*} with the round metric, and fix $p\in S^2$. For each unit vector $v\in T_pS^2$, the meridian starting at $p$ in direction $v$ is \begin{align*} \gamma_v(t)=\cos t\,p+\sin t\,v. \end{align*} This curve has \begin{align*} \gamma_v(0)&=p,\\ \gamma_v(\pi)&=\cos \pi\,p+\sin \pi\,v=-p, \end{align*} so every meridian reaches the antipodal point at time $\pi$. For $0<t<\pi$, the spherical distance from $p$ to $\gamma_v(t)$ is determined by the angle between the two unit vectors: \begin{align*} \cos d(p,\gamma_v(t)) &=p\cdot \gamma_v(t)\\ &=p\cdot(\cos t\,p+\sin t\,v)\\ &=\cos t\,p\cdot p+\sin t\,p\cdot v\\ &=\cos t\cdot 1+\sin t\cdot 0\\ &=\cos t. \end{align*} Since $d(p,\gamma_v(t))\in[0,\pi]$ and $t\in(0,\pi)$, this gives \begin{align*} d(p,\gamma_v(t))=t. \end{align*} Thus each meridian is minimizing up to, but not beyond, length $\pi$. At $t=\pi$, the endpoint is independent of $v$: \begin{align*} \gamma_v(\pi)=-p \end{align*} for every unit tangent vector $v\in T_pS^2$. Hence infinitely many minimizing geodesics from $p$ arrive at the same point $-p$, so uniqueness fails there. Also, along a meridian the angular Jacobi fields have length factor $\sin t$, and \begin{align*} \sin 0=0,\qquad \sin \pi=0, \end{align*} so the first conjugate point to $p$ along the meridian occurs at distance $\pi$. Therefore the cut locus of $p$ is the single point $\{-p\}$, and on the unit round sphere the first conjugate point and the first loss of uniqueness occur at the same distance. [/example] The sphere can make the two obstructions appear inseparable, so the flat cylinder is a useful contrast. It has no curvature-driven focusing, but topology still creates more than one minimizing path. [example: Cut Locus on a Flat Cylinder] Write the flat cylinder as $\mathbb R/2\pi\mathbb Z \times \mathbb R$ with product metric \begin{align*} ds^2=d\theta^2+dz^2, \end{align*} and take $p=([0],z_0)$. A lift of $p$ to the universal cover $\mathbb R^2$ is $(0,z_0)$, and the lifts of a point $q=([\theta],z)$ are \begin{align*} (\theta+2\pi m,z),\qquad m\in \mathbb Z. \end{align*} Hence the squared lengths of lifted straight segments from $(0,z_0)$ to these lifts are \begin{align*} L_m^2=(\theta+2\pi m)^2+(z-z_0)^2. \end{align*} For $0<\theta<\pi$, compare $m=0$ with any $m\ne 0$: \begin{align*} (\theta+2\pi m)^2-\theta^2 &=4\pi m\theta+4\pi^2m^2 \\ &=4\pi m(\theta+\pi m). \end{align*} If $m>0$, then $m(\theta+\pi m)>0$; if $m<0$, then $\theta+\pi m<0$ and again $m(\theta+\pi m)>0$. Thus $L_m^2>L_0^2$, so the minimizing geodesic is unique. The same argument with $-\pi<\theta<0$ shows that the unique minimizing lift is the one with angular displacement in $(-\pi,\pi)$. At the opposite angular point, $q=([\pi],z)$, the two lifts $(\pi,z)$ and $(-\pi,z)$ give \begin{align*} L_0^2&=\pi^2+(z-z_0)^2,\\ L_{-1}^2&=(-\pi)^2+(z-z_0)^2=\pi^2+(z-z_0)^2. \end{align*} They project to two distinct minimizing geodesics on the cylinder, one going around the circle in the positive angular direction and one in the negative angular direction. Therefore the cut locus of $p$ is \begin{align*} \{([\pi],z):z\in \mathbb R\}, \end{align*} the vertical line opposite $p$ around the circular factor. There is no conjugate point along these geodesics. In the lifted flat coordinates, a Jacobi field $J$ orthogonal to a geodesic satisfies \begin{align*} J''(t)+R(J(t),\dot\gamma(t))\dot\gamma(t)=0. \end{align*} The product cylinder is locally Euclidean, so $R=0$ and the equation becomes \begin{align*} J''(t)=0. \end{align*} Thus \begin{align*} J(t)=At+B. \end{align*} If $J(0)=0$ and $J(T)=0$ for some $T>0$, then $B=0$ and $AT=0$, so $A=0$ and $J$ is the zero field. Hence no nonzero Jacobi field vanishes at two distinct times. The cylinder therefore shows that the cut locus can come from loss of uniqueness alone, without conjugate focusing. [/example] The cylinder suggests a broader lesson: quotients of model geometries can preserve local curvature while changing global minimizing behaviour. Flat tori and hyperbolic surfaces show the same distinction in compact and negatively curved settings. [example: Flat Tori and Hyperbolic Surfaces] Write the flat torus as $T^n=\mathbb R^n/\Lambda$, where $\Lambda$ is a lattice, and take $p=[0]$. Choose a shortest nonzero lattice vector $\lambda\in\Lambda$, so $|\lambda|\le |\mu|$ for every nonzero $\mu\in\Lambda$, and set \begin{align*} q=\left[\frac{\lambda}{2}\right]. \end{align*} The point $q$ has lifts \begin{align*} \frac{\lambda}{2}+\mu,\qquad \mu\in\Lambda. \end{align*} For every $\mu\in\Lambda$, \begin{align*} \left|\frac{\lambda}{2}+\mu\right| &=\frac{1}{2}|\lambda+2\mu|. \end{align*} The vector $\lambda+2\mu$ lies in $\Lambda$. It is nonzero, since $\lambda+2\mu=0$ would give $\lambda=-2\mu$, and then $\mu=-\lambda/2$ would be a nonzero lattice vector with $|\mu|=|\lambda|/2<|\lambda|$, contradicting the choice of $\lambda$. Hence \begin{align*} |\lambda+2\mu|\ge |\lambda|, \end{align*} so \begin{align*} \left|\frac{\lambda}{2}+\mu\right|\ge \frac{|\lambda|}{2}. \end{align*} Equality occurs for $\mu=0$ and for $\mu=-\lambda$, because \begin{align*} \left|\frac{\lambda}{2}+0\right|=\frac{|\lambda|}{2}, \qquad \left|\frac{\lambda}{2}-\lambda\right| =\left|-\frac{\lambda}{2}\right| =\frac{|\lambda|}{2}. \end{align*} Thus the two straight segments from $0$ to $\lambda/2$ and from $0$ to $-\lambda/2$ project to two distinct minimizing geodesics from $p$ to $q$. Therefore the cut locus of $p$ is nonempty even though the torus is locally Euclidean and has sectional curvature $0$. For a closed hyperbolic surface $M=\mathbb H^2/\Gamma$, the curvature is locally that of $\mathbb H^2$, so $K=-1$. Along a unit-speed geodesic, an orthogonal Jacobi field has the form $J(t)=f(t)E(t)$ with $E$ parallel and \begin{align*} f''(t)+Kf(t)=0. \end{align*} Since $K=-1$, this becomes \begin{align*} f''(t)-f(t)=0, \end{align*} so \begin{align*} f(t)=A\sinh t+B\cosh t. \end{align*} If $J(0)=0$, then $f(0)=B=0$. If also $J(T)=0$ for some $T>0$, then \begin{align*} 0=f(T)=A\sinh T. \end{align*} Because $\sinh T>0$, we get $A=0$, hence $J$ is the zero field. Thus there are no conjugate points. Nevertheless, a shortest nontrivial closed geodesic on the compact surface has two minimizing half-arcs between opposite points on that geodesic, so the cut locus is nonempty. These examples separate local curvature from global topology: curvature controls conjugate focusing, while topology can still create multiple minimizing representatives and hence positive restrictions on injectivity radius. [/example] The chapter ends with the basic dictionary now in place: curvature bounds supply the comparison data, space forms provide the models, and completeness allows geodesic methods to reach global questions. The next chapter studies Jacobi fields, which are the mechanism by which curvature controls the derivative of the exponential map. With curvature bounds and model spaces in place, the problem becomes one of tracking how nearby geodesics respond to curvature. Jacobi fields provide exactly that infinitesimal mechanism, and the next chapter turns them into the index-form machinery needed for comparison. # 2. Jacobi Fields and the Index Form Jacobi fields are the infinitesimal record of how a geodesic changes inside a family of nearby geodesics. In the previous chapter, curvature entered through the geometry of model spaces and through the behaviour of geodesic polar coordinates. This chapter develops the analytic mechanism behind that behaviour: a second-order linear equation along a geodesic whose coefficients are given by curvature. The index form then turns that equation into a variational statement about when geodesics minimise and when nearby geodesics begin to focus. ## Variations Through Geodesics The first question is how to differentiate a family of geodesics without losing the geometric content of the family. A variation has two directions: the geodesic direction and the parameter direction. The commutation properties of the Levi-Civita connection convert this two-variable picture into an ordinary differential equation along a single geodesic. [definition: Geodesic Variation] Let $(M,g)$ be a Riemannian manifold and let $I,J\subset \mathbb R$ be intervals. A smooth map $\alpha:I\times J\to M$ is a geodesic variation if, for each $s\in J$, the curve $\gamma_s:I\to M$ defined by $\gamma_s(t)=\alpha(t,s)$ is a geodesic. [/definition] A geodesic variation gives the two-variable object whose slices are geodesics. To linearise it along the central slice, we need a single vector field that records the parameter direction at $s=0$. [definition: Variation Field] Let $\alpha:I\times J\to M$ be a smooth variation and set $\gamma(t)=\alpha(t,0)$. The variation field of $\alpha$ along $\gamma$ is the smooth section $J:I\to TM$ of the pullback bundle $\gamma^*TM$, with $J(t)\in T_{\gamma(t)}M$, given by \begin{align*} J(t)=\frac{\partial \alpha}{\partial s}(t,0). \end{align*} [/definition] The variation field is the first-order displacement of the family, but it is not yet useful until we know the equation it satisfies. This motivates the Jacobi equation from geodesic variations: it is the calculation that turns the geodesic equation for every slice into a curvature equation along the central geodesic. [quotetheorem:2716] [citeproof:2716] The theorem shows that every geodesic variation produces a solution of a linear second-order equation, but its hypotheses matter. The curves $t\mapsto\alpha(t,s)$ must be geodesics; for a general variation, the same computation contains an additional acceleration term measuring the failure of the slices to solve the geodesic equation. The equality $[S,T]=0$ and the torsion-free property of the Levi-Civita connection are what allow the two partial directions to be exchanged without adding torsion terms. The theorem also does not say that every vector field along $\gamma$ arises from a geodesic variation; it isolates precisely those first-order fields governed by the curvature equation. This motivates naming the solutions themselves, since comparison arguments often begin with the equation rather than with an explicit variation. [definition: Jacobi Field] Let $\gamma:I\to M$ be a geodesic. A Jacobi field along $\gamma$ is a smooth section $J:I\to TM$ of $\gamma^*TM$, with $J(t)\in T_{\gamma(t)}M$, satisfying \begin{align*} \nabla_{\dot\gamma}\nabla_{\dot\gamma}J+R(J,\dot\gamma)\dot\gamma=0. \end{align*} [/definition] Jacobi fields form a finite-dimensional [vector space](/page/Vector%20Space) because the equation is a linear second-order ordinary differential equation along $\gamma$. Initial data $J(t_0)$ and $\nabla_{\dot\gamma}J(t_0)$ determine the solution, so the flat case provides the baseline model. [example: Jacobi Fields In Euclidean Space] On $M=\mathbb R^n$ with its Euclidean metric, the Levi-Civita connection is ordinary differentiation in the standard coordinates, and the curvature tensor is zero. Along the affine geodesic $\gamma(t)=x_0+tv$, a vector field has the form $J(t)=(J^1(t),\dots,J^n(t))$, so the Jacobi equation becomes \begin{align*} 0 &=\nabla_{\dot\gamma}\nabla_{\dot\gamma}J+R(J,\dot\gamma)\dot\gamma\\ &=\frac{d^2J}{dt^2}+0\\ &=(J^1{}''(t),\dots,J^n{}''(t)). \end{align*} Thus each component satisfies $J^i{}''(t)=0$. Integrating once gives $J^i{}'(t)=b^i$ for a constant $b^i$, and integrating again gives $J^i(t)=a^i+tb^i$ for a constant $a^i$. Hence \begin{align*} J(t) &=(a^1+tb^1,\dots,a^n+tb^n)\\ &=(a^1,\dots,a^n)+t(b^1,\dots,b^n)\\ &=a+tb \end{align*} for fixed vectors $a,b\in\mathbb R^n$. The constant term is produced by translating the initial point, while the linear term is produced by changing the initial velocity; in flat space, Jacobi fields record exactly those two affine ways of varying a straight-line geodesic. [/example] The Euclidean example shows that some Jacobi fields merely move along the same geodesic or reparametrise it, while others measure separation from it. This motivates the definition of transverse Jacobi fields, which discards the tangential part and keeps the perpendicular geometry relevant for focusing. [definition: Transverse Jacobi Field] Let $\gamma:I\to M$ be a geodesic. A Jacobi field $J:I\to TM$ along $\gamma$, viewed as a smooth section of $\gamma^*TM$, is transverse if \begin{align*} (J(t),\dot\gamma(t))_g=0 \end{align*} for every $t\in I$. [/definition] The transverse condition leaves us with an equation on the perpendicular bundle along $\gamma$. This motivates the curvature endomorphism along a geodesic, because we need to regard $R(J,\dot\gamma)\dot\gamma$ as a linear operator acting on transverse fields. [definition: Curvature Endomorphism Along A Geodesic] Let $\gamma:I\to M$ be a unit-speed geodesic. The curvature endomorphism along $\gamma$ is the family of linear maps $\mathcal R_t:\dot\gamma(t)^\perp\to \dot\gamma(t)^\perp$ defined by \begin{align*} \mathcal R_t(V)=R(V,\dot\gamma(t))\dot\gamma(t). \end{align*} [/definition] The curvature symmetries imply that each $\mathcal R_t$ is self-adjoint with respect to $g_{\gamma(t)}$. With this notation, a transverse Jacobi field satisfies $\nabla_{\dot\gamma}^2J+\mathcal R_tJ=0$. Curvature bounds become differential inequalities for this operator family, and the round sphere is the model where the operator is constant and positive. [example: Geodesic Variation On The Sphere] Let $S^n$ have sectional curvature $1$, let $\gamma$ be a unit-speed geodesic, and choose a parallel field $E$ along $\gamma$ with $(E(t),\dot\gamma(t))_g=0$. Since $E$ is parallel, \begin{align*} \nabla_{\dot\gamma}E=0. \end{align*} For $J(t)=\sin(t)E(t)$, the product rule gives \begin{align*} \nabla_{\dot\gamma}J &=\frac{d}{dt}(\sin t)E(t)+\sin(t)\nabla_{\dot\gamma}E\\ &=\cos(t)E(t)+\sin(t)\cdot 0\\ &=\cos(t)E(t), \end{align*} and applying the product rule once more gives \begin{align*} \nabla_{\dot\gamma}\nabla_{\dot\gamma}J &=\nabla_{\dot\gamma}(\cos(t)E(t))\\ &=\frac{d}{dt}(\cos t)E(t)+\cos(t)\nabla_{\dot\gamma}E\\ &=-\sin(t)E(t)+\cos(t)\cdot 0\\ &=-\sin(t)E(t)\\ &=-J(t). \end{align*} Because $S^n$ has sectional curvature $1$ and $E(t)\perp\dot\gamma(t)$ with $|\dot\gamma(t)|=1$, the curvature endomorphism in this transverse direction is \begin{align*} R(E(t),\dot\gamma(t))\dot\gamma(t)=E(t). \end{align*} Thus \begin{align*} R(J(t),\dot\gamma(t))\dot\gamma(t) &=R(\sin(t)E(t),\dot\gamma(t))\dot\gamma(t)\\ &=\sin(t)R(E(t),\dot\gamma(t))\dot\gamma(t)\\ &=\sin(t)E(t)\\ &=J(t), \end{align*} so \begin{align*} \nabla_{\dot\gamma}\nabla_{\dot\gamma}J+R(J,\dot\gamma)\dot\gamma &=-J+J\\ &=0. \end{align*} Also, \begin{align*} (J(t),\dot\gamma(t))_g &=(\sin(t)E(t),\dot\gamma(t))_g\\ &=\sin(t)(E(t),\dot\gamma(t))_g\\ &=0, \end{align*} so $J$ is transverse. Finally, \begin{align*} J(0)=\sin(0)E(0)=0,\qquad J(\pi)=\sin(\pi)E(\pi)=0. \end{align*} This Jacobi field records how great circles through the north pole separate initially and then meet again at the south pole. [/example] The sphere shows that positive curvature can force initially separated geodesics to meet again. The next section converts that focusing into the loss of local minimising behaviour. ## The Index Form And Conjugate Points The variational question is whether a geodesic segment minimises energy among nearby curves with the same endpoints. First variation says that geodesics are critical points. The second variation is the quadratic form whose sign decides stability. [definition: Index Form] Let $\gamma:[a,b]\to M$ be a geodesic, and let $\mathcal V_\gamma$ be the real vector space of piecewise smooth vector fields along $\gamma$. The index form of $\gamma$ is the bilinear map \begin{align*} I_\gamma:\mathcal V_\gamma\times\mathcal V_\gamma\to\mathbb R \end{align*} defined by \begin{align*} I_\gamma[V,W]=\int_a^b\left((\nabla_{\dot\gamma}V,\nabla_{\dot\gamma}W)_g-(R(V,\dot\gamma)\dot\gamma,W)_g\right)\,dt. \end{align*} Here $\nabla_{\dot\gamma}V$ and $\nabla_{\dot\gamma}W$ are taken on each smooth subinterval, and the integral is the sum of the corresponding integrals across the finitely many break points. [/definition] The index form is built to be the Hessian of energy at a geodesic after boundary terms are handled. To use it for stability, one must know that it is not merely a formal curvature expression: it must coincide with the second derivative of energy for genuine variations with fixed endpoints. The required statement converts variational calculus into the [bilinear form](/page/Bilinear%20Form) above. [quotetheorem:2729] [citeproof:2729] The formula makes curvature visible in the stability problem: positive curvature lowers the second variation in transverse directions, while negative curvature raises it. The fixed-endpoint hypothesis is essential because otherwise the first and second variation formulas contain boundary terms, so a variation may change the energy simply by sliding endpoints rather than by detecting instability of the geodesic segment. The statement is also local in the space of curves: positivity of this quadratic form gives infinitesimal stability, not a global distance-minimising conclusion without additional control. The borderline case occurs when a nonzero variation field vanishes at both endpoints and solves the Jacobi equation, so the next definition names that degeneracy. [definition: Conjugate Points Along A Geodesic] Let $\gamma:[a,b]\to M$ be a geodesic. The point $\gamma(b)$ is conjugate to $\gamma(a)$ along $\gamma$ if there exists a nonzero Jacobi field $J$ along $\gamma$ such that \begin{align*} J(a)=0,\qquad J(b)=0. \end{align*} The multiplicity is the dimension of the vector space of such Jacobi fields. [/definition] A conjugate point means that the exponential map has lost rank at a nonzero tangent vector based at the initial point. To connect this intrinsic definition with normal coordinates, we need to identify singular Jacobi fields with kernel vectors of the differential of $\exp_p$. [quotetheorem:5348] [citeproof:5348] The criterion links the analytic equation to normal coordinates: before the first conjugate point, the exponential map has no Jacobi degeneracy at the corresponding nonzero tangent vector. The nonzero hypothesis on $v$ separates genuine conjugacy from the base-point behaviour $d(\exp_p)_0=\operatorname{id}_{T_pM}$, where the exponential map has full rank and no endpoint geodesic segment has yet been selected. Radial variation itself is not discarded in the theorem: for $w=v$, the associated field is $J(t)=t\dot\gamma(t)$, so $J(1)=\dot\gamma(1)\ne0$ when $v\ne0$ and it does not give a kernel vector. Thus conjugacy is detected exactly by a nonzero kernel vector of $d(\exp_p)_v$, while the usual geometric intuition is that in many examples the kernel lies in angular directions. For instance, on the round sphere at $v=\pi u$ the angular directions orthogonal to $u$ lie in the kernel, whereas the radial direction still maps to the nonzero terminal velocity. This is the first place where Jacobi fields connect directly to the cut locus and to the failure of geodesic polar coordinates. To use this information in variational arguments, we need a [comparison principle](/theorems/4870) saying that Jacobi fields minimise the index form among fields with fixed endpoint values. [quotetheorem:5349] [citeproof:5349] The [index lemma](/theorems/5349) is the bridge from Jacobi fields to comparison geometry, since it turns a differential equation into a minimising property of a quadratic form. The endpoint and nonconjugacy hypotheses prevent hidden null directions from being mistaken for genuine minimisers of the quadratic form. If a conjugate point is present, a nonzero Jacobi field can vanish at both endpoints, and the equality statement is no longer rigid without quotienting out that degeneracy. The lemma is therefore a local comparison statement before the first conjugate point, not a global minimisation theorem for arbitrary endpoint pairs. The natural global question is then how many independent directions make the second variation negative once conjugate points occur. [quotetheorem:5350] [citeproof:5350] The Morse index theorem gives a precise count of how minimising behaviour fails. The field space consists of piecewise smooth fields with both endpoint values fixed to zero; the piecewise regularity is enough for [integration by parts](/theorems/2098) on each smooth subinterval, with endpoint and matching terms controlling the boundary contributions. The endpoint nonconjugacy hypothesis is needed because a conjugate endpoint produces a null direction of the index form rather than a strictly negative direction, so the index alone no longer records the full degeneracy. If the endpoint is conjugate, one must keep track of both the index and the nullity. A geodesic segment cannot remain locally energy-minimising past a conjugate point because the index form has acquired a negative direction, and the round sphere gives the standard example. [example: Loss Of Minimality On The Round Sphere] Let $\gamma:[0,L]\to S^n$ be a unit-speed meridian from the north pole, with $L>\pi$, and choose a parallel unit field $E$ along $\gamma$ satisfying $E(t)\perp \dot\gamma(t)$. On the round sphere of curvature $1$, \begin{align*} R(E(t),\dot\gamma(t))\dot\gamma(t)=E(t). \end{align*} Set \begin{align*} V(t)=\sin\left(\frac{\pi t}{L}\right)E(t). \end{align*} Then \begin{align*} V(0)=\sin(0)E(0)=0,\qquad V(L)=\sin(\pi)E(L)=0, \end{align*} so $V$ is an admissible fixed-endpoint variation field. Since $E$ is parallel, \begin{align*} \nabla_{\dot\gamma}V &=\frac{d}{dt}\left(\sin\left(\frac{\pi t}{L}\right)\right)E(t) +\sin\left(\frac{\pi t}{L}\right)\nabla_{\dot\gamma}E\\ &=\frac{\pi}{L}\cos\left(\frac{\pi t}{L}\right)E(t) +\sin\left(\frac{\pi t}{L}\right)\cdot 0\\ &=\frac{\pi}{L}\cos\left(\frac{\pi t}{L}\right)E(t). \end{align*} Also, by linearity of $R$ in the first slot, \begin{align*} R(V(t),\dot\gamma(t))\dot\gamma(t) &=R\left(\sin\left(\frac{\pi t}{L}\right)E(t),\dot\gamma(t)\right)\dot\gamma(t)\\ &=\sin\left(\frac{\pi t}{L}\right)R(E(t),\dot\gamma(t))\dot\gamma(t)\\ &=\sin\left(\frac{\pi t}{L}\right)E(t). \end{align*} Therefore the index form is \begin{align*} I_\gamma[V,V] &=\int_0^L\left(|\nabla_{\dot\gamma}V|^2-(R(V,\dot\gamma)\dot\gamma,V)_g\right)\,dt\\ &=\int_0^L\left(\frac{\pi^2}{L^2}\cos^2\left(\frac{\pi t}{L}\right) -\sin^2\left(\frac{\pi t}{L}\right)\right)\,dt\\ &=\frac{\pi^2}{L^2}\int_0^L\cos^2\left(\frac{\pi t}{L}\right)\,dt -\int_0^L\sin^2\left(\frac{\pi t}{L}\right)\,dt. \end{align*} With $u=\pi t/L$, so $dt=(L/\pi)\,du$, the two integrals are \begin{align*} \int_0^L\cos^2\left(\frac{\pi t}{L}\right)\,dt &=\frac{L}{\pi}\int_0^\pi\cos^2u\,du =\frac{L}{\pi}\cdot \frac{\pi}{2} =\frac{L}{2},\\ \int_0^L\sin^2\left(\frac{\pi t}{L}\right)\,dt &=\frac{L}{\pi}\int_0^\pi\sin^2u\,du =\frac{L}{\pi}\cdot \frac{\pi}{2} =\frac{L}{2}. \end{align*} Hence \begin{align*} I_\gamma[V,V] &=\frac{\pi^2}{L^2}\cdot \frac{L}{2}-\frac{L}{2}\\ &=\frac{\pi^2}{2L}-\frac{L}{2}\\ &=\frac{\pi^2-L^2}{2L}. \end{align*} Since $L>\pi$, we have $\pi^2-L^2<0$, and therefore \begin{align*} I_\gamma[V,V]<0. \end{align*} By the second variation formula for fixed endpoints, this negative index direction means the meridian segment of length $L>\pi$ is not locally energy-minimising beyond the south pole; the first conjugate point occurs at time $\pi$, where the transverse Jacobi field $\sin(t)E(t)$ vanishes again. [/example] Length has the same second variation as energy for normal variations of a unit-speed geodesic, up to the removal of tangential reparametrisation terms. Hence comparison arguments usually work with energy and the index form, while the geometric conclusion is phrased in terms of distance and length. ## Sturm Comparison And Geometric Focusing The remaining question is how curvature bounds force quantitative behaviour of Jacobi fields. Along a fixed transverse parallel field in constant curvature, the Jacobi equation becomes a scalar second-order equation. Sturm comparison is the principle that ordering the coefficients orders the zeros and growth of solutions. [definition: Scalar Jacobi Equation] Let $k:\mathbb R\to\mathbb R$ be a [continuous function](/page/Continuous%20Function). The scalar Jacobi equation with coefficient $k$ is the ordinary differential equation for a twice differentiable function $y:\mathbb R\to\mathbb R$ given by \begin{align*} y''(t)+k(t)y(t)=0. \end{align*} [/definition] The scalar Jacobi equation models a transverse Jacobi field when curvature acts by multiplication on the chosen direction. This motivates Sturm comparison, because zeros of scalar solutions represent focusing times and hence the possible appearance of conjugate points. [quotetheorem:3510] [citeproof:3510] The scalar theorem is the one-dimensional shadow of matrix comparison. Its hypotheses should be read exactly as scalar ODE hypotheses: the comparison is about how the coefficient in \begin{align*} y''(t)+k(t)y(t)=0 \end{align*} controls the possible zeros and signs of solutions on the interval where the theorem applies. In geometric applications, those zeros represent focusing of transverse variations and hence possible conjugate points. The important lesson for Jacobi fields is therefore not a separate quotient formula, but the direction of influence: larger positive curvature makes transverse solutions focus sooner, while lower or negative curvature delays focusing. In the vector-valued Jacobi equation, the scalar coefficient is replaced by the curvature endomorphism $\mathcal R_t$, and inequalities must be interpreted as quadratic-form inequalities on $\dot\gamma(t)^\perp$. [explanation: Matrix Interpretation Of Sturm Comparison] Choose a parallel orthonormal frame $E_1(t),\dots,E_{n-1}(t)$ for $\dot\gamma(t)^\perp$ along a unit-speed geodesic. A transverse Jacobi field can be written as $J(t)=\sum_i u_i(t)E_i(t)$, and the Jacobi equation becomes \begin{align*} u''(t)+A(t)u(t)=0, \end{align*} where $A(t)$ is the symmetric matrix representing $\mathcal R_t$. A lower sectional curvature bound $K\ge \kappa$ gives $A(t)\ge \kappa I$ as a quadratic-form inequality, while an upper bound gives the reverse inequality. The index form is the correct way to compare this system because matrix solutions need not commute at different times. Instead of comparing eigenvalues pointwise as if the system were diagonal for all $t$, one compares quadratic forms on vector fields. This is why the index lemma is paired with Sturm comparison in comparison geometry. [/explanation] The matrix interpretation explains why constant-curvature functions serve as comparison solutions even when the true curvature varies. This motivates returning to the model sine function introduced in Chapter 1. For comparison arguments, $\operatorname{sn}_\kappa$ is used through its ODE characterization \begin{align*} \operatorname{sn}_\kappa''(t)+\kappa\operatorname{sn}_\kappa(t)=0,\qquad \operatorname{sn}_\kappa(0)=0,\qquad \operatorname{sn}_\kappa'(0)=1, \end{align*} which records radial Jacobi fields in the space form of curvature $\kappa$. The first zero of $\operatorname{sn}_\kappa$ is finite exactly when $\kappa>0$. This is the model for positive-curvature focusing and motivates the comparison estimate in the next example. [example: Focusing Under Positive Curvature] [claim]Under the stated lower sectional curvature bound, $\gamma(0)$ has a conjugate point along $\gamma$ at some time $t\le \pi/\sqrt{\kappa}$.[/claim] [proof]Set $L_0=\pi/\sqrt{\kappa}$. Since $\dim M\ge2$, choose a unit vector $E(0)\perp \dot\gamma(0)$ and extend it by parallel transport along $\gamma$, so \begin{align*} \nabla_{\dot\gamma}E=0,\qquad |E(t)|=1,\qquad E(t)\perp \dot\gamma(t). \end{align*} For $\kappa>0$ the model function is \begin{align*} \operatorname{sn}_\kappa(t)=\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t), \end{align*} so define \begin{align*} V(t)=\operatorname{sn}_\kappa(t)E(t) =\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t)E(t). \end{align*} Then \begin{align*} V(0) &=\frac{1}{\sqrt{\kappa}}\sin(0)E(0)=0,\\ V(L_0) &=\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}\cdot \pi/\sqrt{\kappa})E(L_0)\\ &=\frac{1}{\sqrt{\kappa}}\sin(\pi)E(L_0)=0. \end{align*} Also, \begin{align*} \nabla_{\dot\gamma}V &=\frac{d}{dt}\left(\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t)\right)E(t) +\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t)\nabla_{\dot\gamma}E\\ &=\cos(\sqrt{\kappa}t)E(t) +\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t)\cdot 0\\ &=\cos(\sqrt{\kappa}t)E(t), \end{align*} hence \begin{align*} |\nabla_{\dot\gamma}V|^2 =\cos^2(\sqrt{\kappa}t)|E(t)|^2 =\cos^2(\sqrt{\kappa}t). \end{align*} Since $E(t)$ and $\dot\gamma(t)$ are orthonormal and $K(\dot\gamma,E)\ge\kappa$, \begin{align*} (R(E,\dot\gamma)\dot\gamma,E)_g=K(\dot\gamma,E)\ge \kappa. \end{align*} Therefore \begin{align*} (R(V,\dot\gamma)\dot\gamma,V)_g &=\left(R\left(\frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t)E,\dot\gamma\right)\dot\gamma, \frac{1}{\sqrt{\kappa}}\sin(\sqrt{\kappa}t)E\right)_g\\ &=\frac{1}{\kappa}\sin^2(\sqrt{\kappa}t)(R(E,\dot\gamma)\dot\gamma,E)_g\\ &\ge \frac{1}{\kappa}\sin^2(\sqrt{\kappa}t)\cdot \kappa\\ &=\sin^2(\sqrt{\kappa}t). \end{align*} Thus \begin{align*} I_\gamma[V,V] &=\int_0^{L_0}\left(|\nabla_{\dot\gamma}V|^2-(R(V,\dot\gamma)\dot\gamma,V)_g\right)\,dt\\ &\le \int_0^{L_0}\left(\cos^2(\sqrt{\kappa}t)-\sin^2(\sqrt{\kappa}t)\right)\,dt\\ &=\int_0^{L_0}\cos(2\sqrt{\kappa}t)\,dt\\ &=\left[\frac{1}{2\sqrt{\kappa}}\sin(2\sqrt{\kappa}t)\right]_{0}^{\pi/\sqrt{\kappa}}\\ &=\frac{1}{2\sqrt{\kappa}}\sin(2\pi)-\frac{1}{2\sqrt{\kappa}}\sin(0)\\ &=0. \end{align*} If $\gamma(L_0)$ is conjugate to $\gamma(0)$, the claim is proved. If it is not conjugate and there is no conjugate point in $(0,L_0)$, then the fixed-endpoint index form is positive definite by the *[Morse Index Theorem For Geodesic Segments](/theorems/5350)*. Since $V$ is nonzero and satisfies $V(0)=V(L_0)=0$, positive definiteness would give $I_\gamma[V,V]>0$, contradicting $I_\gamma[V,V]\le0$. Hence a conjugate point must occur at some time $t\le L_0=\pi/\sqrt{\kappa}$.[/proof] On the round sphere of radius $1/\sqrt{\kappa}$, every sectional curvature equals $\kappa$, and the transverse field $\operatorname{sn}_\kappa(t)E(t)$ vanishes exactly at $t=0$ and $t=\pi/\sqrt{\kappa}$ before its next zero. Thus the comparison time is sharp: geodesics from a point meet at the antipodal point precisely at distance $\pi/\sqrt{\kappa}$. [/example] The focusing example shows how a positive lower curvature bound leads toward zeros of Jacobi fields. This motivates the absence theorem in the opposite curvature regime, where nonpositive sectional curvature prevents a nonzero Jacobi field from vanishing twice. [quotetheorem:2738] [citeproof:2738] The absence theorem explains why negative curvature is associated with defocusing rather than focusing. The nonpositive curvature hypothesis is used through the sign of the curvature term; positive curvature reverses this mechanism, as on the round sphere where transverse Jacobi fields vanish again at time $\pi$. The conclusion rules out conjugate points, but it does not by itself say that every geodesic segment globally minimises, since topology and the cut locus can introduce multiple geodesics without conjugacy. The hyperbolic model shows the same phenomenon quantitatively, with exponential growth replacing oscillation. [example: Negative Curvature And Diverging Geodesics] Let $\mathbb H^n$ have constant sectional curvature $-1$, let $\gamma$ be a unit-speed geodesic, and let $E$ be a parallel unit field along $\gamma$ with $E(t)\perp\dot\gamma(t)$. Define \begin{align*} J(t)=\sinh(t)E(t). \end{align*} Then \begin{align*} J(0)=\sinh(0)E(0)=0. \end{align*} Since $E$ is parallel, \begin{align*} \nabla_{\dot\gamma}J &=\frac{d}{dt}(\sinh t)E(t)+\sinh(t)\nabla_{\dot\gamma}E\\ &=\cosh(t)E(t)+\sinh(t)\cdot 0\\ &=\cosh(t)E(t), \end{align*} so \begin{align*} \nabla_{\dot\gamma}J(0)=\cosh(0)E(0)=E(0). \end{align*} Applying the product rule again gives \begin{align*} \nabla_{\dot\gamma}\nabla_{\dot\gamma}J &=\nabla_{\dot\gamma}(\cosh(t)E(t))\\ &=\frac{d}{dt}(\cosh t)E(t)+\cosh(t)\nabla_{\dot\gamma}E\\ &=\sinh(t)E(t)+\cosh(t)\cdot 0\\ &=\sinh(t)E(t)\\ &=J(t). \end{align*} Because the sectional curvature is $-1$ and $E(t),\dot\gamma(t)$ are orthonormal, \begin{align*} R(E(t),\dot\gamma(t))\dot\gamma(t)=-E(t). \end{align*} By linearity in the first slot, \begin{align*} R(J(t),\dot\gamma(t))\dot\gamma(t) &=R(\sinh(t)E(t),\dot\gamma(t))\dot\gamma(t)\\ &=\sinh(t)R(E(t),\dot\gamma(t))\dot\gamma(t)\\ &=-\sinh(t)E(t)\\ &=-J(t). \end{align*} Hence \begin{align*} \nabla_{\dot\gamma}\nabla_{\dot\gamma}J+R(J,\dot\gamma)\dot\gamma &=J-J\\ &=0, \end{align*} so $J$ is a Jacobi field with the prescribed initial data. For $t>0$, \begin{align*} \sinh(t)=\frac{e^t-e^{-t}}{2}>0, \end{align*} because $e^t>e^{-t}$. Therefore $J(t)\ne0$ for every $t>0$. Also, \begin{align*} |J(t)|=|\sinh(t)E(t)|=\sinh(t)|E(t)|=\sinh(t)=\frac{e^t-e^{-t}}{2}, \end{align*} so the size of the field grows like $\frac12 e^t$ as $t\to\infty$. Thus negative curvature produces transverse Jacobi fields that do not refocus after their initial zero and instead record exponential separation of nearby geodesics. [/example] The chapter closes with the main mechanism now in place. Curvature controls Jacobi fields; Jacobi fields control the index form; the index form controls minimising behaviour and conjugate points. Rauch comparison in Chapter 3 will use this chain first to control the differential of the exponential map; later Bonnet--Myers and splitting arguments will reuse the same second-variation mechanism for diameter and topology. The Jacobi-field formalism has shown how curvature enters through second variation and the index form. Rauch comparison now converts that analytic control into geometric estimates on the exponential map, and those estimates will later drive diameter, topology, and rigidity results. # 3. Rauch Comparison and Applications Rauch comparison is the point where the Jacobi-field machinery from the previous chapter becomes geometric measurement. The guiding principle is that sectional curvature controls how fast initially neighbouring geodesics separate, and therefore controls the differential of the exponential map. This chapter turns that principle into length comparison, metric comparison in normal coordinates, and first global consequences for conjugate points and injectivity. ## Comparing the Growth of Jacobi Fields The basic question is how the length of a Jacobi field changes when the curvature along the reference geodesic is bounded above or below. In a constant-curvature model, the transverse Jacobi equation reduces to the scalar equation studied in Chapter 2, with the model functions $\operatorname{sn}_k$ and $\operatorname{ct}_k$ introduced in Chapter 1. Rauch comparison says that general manifolds lie on the appropriate side of these model behaviours whenever their sectional curvatures do. Let $k \in \mathbb R$. Recall the model function $\operatorname{sn}_k:[0,\infty) \to \mathbb R$ given by \begin{align*} \operatorname{sn}_k(t) &= \begin{cases} \frac{1}{\sqrt{k}}\sin(\sqrt{k}t), & k>0,\\ t, & k=0,\\ \frac{1}{\sqrt{-k}}\sinh(\sqrt{-k}t), & k<0, \end{cases} \end{align*} on its natural interval before the first zero when $k>0$. It solves $y''+ky=0$ with $y(0)=0$ and $y'(0)=1$. [definition: Transverse Jacobi Field] Let $(M,g)$ be a Riemannian manifold, let $\gamma:[0,a]\to M$ be a unit-speed geodesic, and let $J$ be a Jacobi field along $\gamma$. The field $J$ is transverse if \begin{align*} (J(t),\dot{\gamma}(t))_g=0 \end{align*} for every $t\in[0,a]$. [/definition] Transverse fields describe angular separation of geodesics issuing from the same point. The tangential part records only reparametrisation effects, so comparison geometry removes it before applying curvature bounds. The next question is whether a pointwise inequality between sectional curvatures forces a pointwise inequality between these angular separation fields; this is exactly the content of the Rauch comparison theorem. [remark: Quoted result: Rauch Comparison Theorem] Let $(M,g)$ and $(\widetilde M,\widetilde g)$ be Riemannian manifolds, and let $\gamma:[0,a]\to M$ and $\widetilde\gamma:[0,a]\to \widetilde M$ be unit-speed geodesics. Let $J$ and $\widetilde J$ be nonzero transverse Jacobi fields along $\gamma$ and $\widetilde\gamma$ with \begin{align*} J(0)=0,\qquad \widetilde J(0)=0,\qquad |D_tJ(0)|=|D_t\widetilde J(0)|. \end{align*} Assume that, after identifying the initial normal spaces by a linear isometry sending $D_tJ(0)$ to $D_t\widetilde J(0)$ and extending by parallel transport along the two geodesics, the radial sectional curvatures satisfy \begin{align*} K_M\bigl(\operatorname{span}(\dot\gamma(t),E(t))\bigr) \le K_{\widetilde M}\bigl(\operatorname{span}(\dot{\widetilde\gamma}(t),\widetilde E(t))\bigr) \end{align*} for every $t\in[0,a]$ and every corresponding pair of parallel unit normal fields $E,\widetilde E$. If $\widetilde J(t)\neq 0$ for $0<t\le a$, then \begin{align*} \frac{|J(t)|}{|\widetilde J(t)|} \end{align*} is nondecreasing on $(0,a]$. In particular, $|J(t)|\ge |\widetilde J(t)|$ for $0\le t\le a$. [/remark] The theorem has a useful contrapositive interpretation: large positive curvature focuses geodesics faster, while smaller curvature permits transverse Jacobi fields to grow faster. The hypotheses are doing real work: without matching initial normal speed the comparison would be changed by an arbitrary scale factor, and without excluding zeros of the comparison field the quotient may cease to be defined. Rauch is also only an infinitesimal statement along chosen geodesics; it does not by itself say that geodesics globally minimise or that the exponential map is injective. For computations and applications, comparing with an arbitrary second manifold is usually more information than is needed. The next step is to put the comparison manifold equal to the simply connected space form of curvature $k$, producing scalar estimates in terms of $\operatorname{sn}_k$. [quotetheorem:5351] [citeproof:5351] The space-form version turns the theorem into a usable ruler: $\sin t$, $t$, and $\sinh t$ are the three basic profiles for angular separation. The normalisation $|D_tJ(0)|=1$ fixes the unit of angular speed; for general initial speed the right-hand side is multiplied by $|D_tJ(0)|$. The no-conjugate-point condition in the lower-curvature case is not cosmetic: after a zero of a Jacobi field, length comparison cannot be read from the same quotient. Before introducing examples, it is worth remembering that this theorem measures radial spreading only, not global uniqueness of geodesics. [illustration:geometric-analysis-i-jacobi-separation-profiles] Before introducing further estimates, it is useful to see these profiles in the models where the inequalities become equalities. [example: Length Growth In Constant Curvature] Let $\gamma$ be a unit-speed geodesic in a space form of constant sectional curvature $k$, and let $E(t)$ be a parallel unit field along $\gamma$ with $E(t)\perp \dot\gamma(t)$. For a transverse Jacobi field of the form $J(t)=j(t)E(t)$, the Jacobi equation becomes \begin{align*} 0&=D_tD_tJ+R(J,\dot\gamma)\dot\gamma\\ &=j''(t)E(t)+j(t)R(E(t),\dot\gamma(t))\dot\gamma(t)\\ &=\bigl(j''(t)+k j(t)\bigr)E(t), \end{align*} so $j''+kj=0$. The conditions $J(0)=0$ and $|D_tJ(0)|=1$ mean \begin{align*} j(0)=0,\qquad |j'(0)|=1. \end{align*} Choosing the positive initial direction gives $j'(0)=1$. On the round sphere, $k=1$, so $j(t)=\sin t$ because \begin{align*} j''(t)+j(t)&=-\sin t+\sin t=0,\\ j(0)&=\sin 0=0,\\ j'(0)&=\cos 0=1. \end{align*} Thus $|J(t)|=|\sin t|$, and on $0\le t\le \pi$ this is $|J(t)|=\sin t$. In Euclidean space, $k=0$, so $j(t)=t$ since \begin{align*} j''(t)&=0,\\ j(0)&=0,\\ j'(0)&=1. \end{align*} In hyperbolic space of curvature $-1$, $j(t)=\sinh t$ because \begin{align*} j''(t)-j(t)&=\sinh t-\sinh t=0,\\ j(0)&=\sinh 0=0,\\ j'(0)&=\cosh 0=1. \end{align*} Thus the three model separation profiles are $\sin t$, $t$, and $\sinh t$: positive curvature makes neighbouring geodesics reconverge by time $\pi$, zero curvature separates them linearly, and negative curvature separates them faster than linearly. [/example] The model example shows the qualitative meaning of the comparison inequalities, but later compactness arguments require a rougher estimate that is stable under perturbing curvature and initial conditions. For that purpose, the course records Berger's comparison estimate as a quantitative bound on Jacobi fields over a fixed scale. [remark: Quoted result: Uniform Jacobi Field Estimate Under Bounded Curvature] Let $(M,g)$ be a Riemannian manifold, let $U\subset M$ be a geodesic ball on which $|K_M|\le \Lambda$, and let $r_0>0$ be fixed. Consider the family of unit-speed geodesics $\gamma:[0,r_0]\to M$ with $\gamma([0,r_0])\subset U$. There is a constant $C=C(n,\Lambda,r_0)>0$ such that every Jacobi field $J$ along one of these geodesics satisfies \begin{align*} |J(t)|+|D_tJ(t)| \le C\bigl(|J(0)|+|D_tJ(0)|\bigr) \end{align*} for $0\le t\le r_0$. [/remark] This estimate is less precise than Rauch in its conclusion but stronger as a stability statement: it prevents Jacobi fields from developing uncontrolled oscillation on a fixed scale when curvature is uniformly bounded. Each hypothesis has a concrete role. If the curvature bound is removed, the scalar equation $j''+K(t)j=0$ along a surface geodesic can be forced to oscillate arbitrarily fast by taking $K(t)$ very large and positive on a short interval, so no constant depending only on $r_0$ can control both $j$ and $j'$. If the requirement $\gamma([0,r_0])\subset U$ is dropped, the geodesic may leave the region where curvature is controlled and enter such a high-curvature region before time $r_0$. If the time scale is not fixed, even the equation $j''-\Lambda j=0$ permits growth like $e^{\sqrt{\Lambda}t}$, so a uniform bound independent of the interval length is false. Unlike Rauch, the estimate does not compare against a sharp model field and does not imply injectivity or absence of conjugate points by itself. It is the input behind many compactness arguments where exact model comparison is unavailable. ## Exponential Coordinates and Metric Distortion The next problem is to translate Jacobi-field comparison into statements about coordinates. Normal coordinates are built from the exponential map, and their metric coefficients are controlled by how $\exp_p$ stretches vectors orthogonal to the radial direction. The differential of the exponential map is therefore the bridge between the ODE comparison theorem and geometric comparison inside geodesic balls. [definition: Radial Jacobi Field] Let $(M,g)$ be a Riemannian manifold, let $p\in M$, let $v\in T_pM$, and let $w\in T_pM$. Define $\gamma(t)=\exp_p(tv)$ wherever this is defined. The radial Jacobi field determined by $v$ and $w$ is the Jacobi field $J$ along $\gamma$ satisfying \begin{align*} J(0)=0,\qquad D_tJ(0)=w. \end{align*} [/definition] The terminology reflects the fact that $J$ is obtained by varying the initial velocity of a radial geodesic. To turn this observation into metric estimates, we need the exact identity linking radial Jacobi fields with the derivative of $\exp_p$. [quotetheorem:2717] [citeproof:2717] This identity is exact, but it is not yet a comparison theorem. The hypotheses ensure that the whole radial variation is defined up to the time being evaluated. This cannot be omitted: on a non-complete manifold such as the open unit ball in $\mathbb R^n$, a radial geodesic can leave the manifold before time $1$, and then $\exp_p(v)$ and its derivative at $v$ are not defined. The smooth dependence on the initial velocity is also part of the statement; at a vector where the exponential map is outside its domain there is no variation to differentiate. Even when the formula is defined at a conjugate vector, it may identify a singular derivative rather than an invertible coordinate change, as on $S^n$ at a vector of length $\pi$ where angular Jacobi fields vanish. With this formula, Rauch comparison becomes a statement about the singular values of $(d\exp_p)_v$ on the subspace perpendicular to $v$. The radial direction is special: by the Gauss lemma, radial lengths are preserved by geodesic polar coordinates. The natural next question is how these radial and angular facts combine into a two-sided estimate for the metric tensor in normal coordinates. [quotetheorem:5352] [citeproof:5352] The metric comparison theorem is the local measurement form of Rauch comparison: curvature bounds become distortion bounds for polar-coordinate spheres. The factor $\operatorname{sn}_k(r)/r$ is forced by linearity of $(d\exp_p)_v$ on $T_v(T_pM)$; in Euclidean space it is exactly $1$, so arbitrary angular tangent vectors are unchanged. The nonsingularity and first-zero restrictions mark the differential obstruction: the comparison controls stretching only while normal coordinates are still valid, and it says nothing about two distinct radial geodesics meeting farther away. [illustration:geometric-analysis-i-normal-coordinate-angular-vector] In a small ball with bounded curvature, the model functions have matching first-order behaviour, so the theorem predicts quantitative almost-Euclidean geometry. [example: Metric Distortion In A Small Geodesic Ball] Assume $|K_M|\le \Lambda$ on $B(p,R)$, so $-\Lambda\le K_M\le \Lambda$. For an angular vector $u\perp e$ with $|e|=1$, the normal-coordinate comparison from *Metric Comparison In Normal Coordinates* gives \begin{align*} \operatorname{sn}_{\Lambda}(r)|u| \le |(d\exp_p)_{re}(ru)| \le \operatorname{sn}_{-\Lambda}(r)|u|. \end{align*} For $\Lambda>0$, the Taylor series of $\sin x$ and $\sinh x$ at $x=0$ give \begin{align*} \operatorname{sn}_{\Lambda}(r) &=\frac{\sin(\sqrt{\Lambda}r)}{\sqrt{\Lambda}}\\ &=\frac{\sqrt{\Lambda}r-\frac{(\sqrt{\Lambda}r)^3}{6}+O(r^5)}{\sqrt{\Lambda}}\\ &=r-\frac{\Lambda r^3}{6}+O(r^5), \end{align*} and \begin{align*} \operatorname{sn}_{-\Lambda}(r) &=\frac{\sinh(\sqrt{\Lambda}r)}{\sqrt{\Lambda}}\\ &=\frac{\sqrt{\Lambda}r+\frac{(\sqrt{\Lambda}r)^3}{6}+O(r^5)}{\sqrt{\Lambda}}\\ &=r+\frac{\Lambda r^3}{6}+O(r^5). \end{align*} When $\Lambda=0$, both model functions are exactly $\operatorname{sn}_0(r)=r$, so the same estimates hold with the cubic coefficient equal to $0$. Thus an angular Euclidean polar vector of length $r|u|$ is sent to a vector whose length lies between \begin{align*} \left(r-\frac{\Lambda r^3}{6}+O(r^5)\right)|u| \quad\text{and}\quad \left(r+\frac{\Lambda r^3}{6}+O(r^5)\right)|u|. \end{align*} For squared lengths, \begin{align*} \left(r-\frac{\Lambda r^3}{6}+O(r^5)\right)^2 &=r^2-\frac{\Lambda r^4}{3}+O(r^6),\\ \left(r+\frac{\Lambda r^3}{6}+O(r^5)\right)^2 &=r^2+\frac{\Lambda r^4}{3}+O(r^6). \end{align*} So normal coordinates differ from the Euclidean polar metric by order $r^3$ in angular lengths, and by order $r^4$ in angular squared lengths. [/example] The example explains why sufficiently small geodesic balls look almost Euclidean under bounded curvature. The comparison theorem gives a uniform version of this statement, which is essential when passing to limits of manifolds with common curvature bounds. [remark: Normal Coordinates Depend On Two Obstructions] Normal coordinates at $p$ can fail either because a conjugate point makes $(d\exp_p)_v$ singular or because two different radial geodesics reach the same point with the same length. Rauch comparison controls the first obstruction directly. The second obstruction is global and is measured by the cut locus and injectivity radius. [/remark] This distinction sets up the final section. Bounds on Jacobi fields give conjugacy information immediately, while injectivity information needs a local-to-global argument or an additional topological input. ## Injectivity and Conjugacy Under Curvature Bounds The final question is how far the comparison estimates force normal coordinates to remain valid. Curvature upper bounds prevent focusing, curvature lower bounds can force focusing by comparison with spheres, and nonpositive curvature interacts with simple connectedness to remove the cut-locus obstruction. These consequences are the first major applications of Rauch comparison. [definition: Conjugate Radius] Let $(M,g)$ be a Riemannian manifold. As in Chapter 1, the pointwise conjugate-radius function is the map \begin{align*} \operatorname{conj}:M\to[0,\infty] \end{align*} defined at $p\in M$ by \begin{align*} \operatorname{conj}(p)=\inf\{t>0: \text{there is a unit-speed geodesic }\gamma\text{ with }\gamma(0)=p\text{ and }\gamma(t)\text{ conjugate to }p\}. \end{align*} The conjugate radius of $M$ is \begin{align*} \operatorname{conj}(M)=\inf_{p\in M}\operatorname{conj}(p). \end{align*} [/definition] Conjugate radius isolates the differential obstruction to normal coordinates. The first comparison consequence asks how long Jacobi fields are forced to remain nonzero when sectional curvature is bounded above. [quotetheorem:5353] [citeproof:5353] The estimate is sharp in the basic positively curved model. Completeness ensures that geodesics exist for the required comparison time, while the upper curvature bound is the sign condition preventing faster-than-model focusing. The theorem concerns only conjugate points, so it does not rule out a cut point before the stated radius; for instance, global topology can make minimizing geodesics stop being unique even when no Jacobi field has vanished. The sphere fixes the scale $\pi/\sqrt{k}$ and shows that the comparison result is measuring an actual focusing phenomenon, not an artefact of the proof. [example: Conjugate Radius Of The Round Sphere] On $S^n$ with sectional curvature $1$, every unit-speed geodesic is a great circle. Let $J$ be a nonzero transverse Jacobi field along such a geodesic $\gamma$ with $J(0)=0$, and put $a=|D_tJ(0)|>0$. Choose the parallel unit field $E(t)$ along $\gamma$ with \begin{align*} E(0)=\frac{D_tJ(0)}{|D_tJ(0)|}. \end{align*} In constant curvature $1$, a field of the form $J(t)=j(t)E(t)$ satisfies \begin{align*} 0 &=D_tD_tJ+R(J,\dot\gamma)\dot\gamma\\ &=j''(t)E(t)+j(t)R(E(t),\dot\gamma(t))\dot\gamma(t)\\ &=j''(t)E(t)+j(t)E(t)\\ &=\bigl(j''(t)+j(t)\bigr)E(t), \end{align*} so $j''+j=0$. The initial conditions are \begin{align*} j(0)=0,\qquad j'(0)=a. \end{align*} The function $j(t)=a\sin t$ has these initial data, since \begin{align*} j''(t)+j(t)&=-a\sin t+a\sin t=0,\\ j(0)&=a\sin 0=0,\\ j'(0)&=a\cos 0=a. \end{align*} Hence \begin{align*} J(t)=a\sin(t)E(t), \qquad |J(t)|=a|\sin t|. \end{align*} For $0<t<\pi$, $\sin t>0$, so $J(t)\neq0$; at $t=\pi$, \begin{align*} J(\pi)=a\sin(\pi)E(\pi)=0. \end{align*} Thus the first conjugate point along each great circle occurs at time $\pi$, and therefore $\operatorname{conj}(S^n)=\pi$, exactly matching the upper-curvature comparison bound for $k=1$. [/example] The sharpness on the sphere shows that the estimate cannot be improved using only the upper curvature bound. The complementary problem is what a positive lower curvature bound can say before the first conjugate point; this leads to a Bonnet-type focusing consequence that foreshadows later Ricci comparison. [quotetheorem:5354] [citeproof:5354] This theorem is included mainly as a bridge to the Ricci comparison material. The lower curvature bound is essential: in Euclidean space, where $K=0$, the field $J(t)=tE$ violates any spherical upper bound $|J(t)|\le |D_tJ(0)|\sin(\sqrt{k}t)/\sqrt{k}$ for fixed $k>0$ and small positive $t$ beyond the first-order term. Hyperbolic space gives an even stronger failure, since transverse fields grow like $\sinh t$. Completeness is also needed for a statement over the full interval: an incomplete open subset of a positively curved sphere can have geodesics that leave the manifold before the comparison time, so the field is not available to estimate there. The no-conjugate-point clause is part of the conclusion's range, not an auxiliary technicality, because the comparison quotient stops encoding a single nonsingular exponential chart after conjugacy; on the round sphere, angular fields vanish at $t=\pi$ and the same polar chart cannot be continued through that time as a nonsingular chart. For sectional curvature, the most powerful global statement in this chapter is instead obtained when curvature is nonpositive and the manifold is simply connected. The question becomes whether the local absence of conjugate points can eliminate all failures of global normal coordinates. [quotetheorem:2737] [citeproof:2737] The hypotheses separate the two obstructions exactly. The curvature condition $K_M\le0$ removes conjugate points and makes $\exp_p$ locally invertible everywhere, while completeness supplies geodesics long enough to reach all points. Simple connectedness is necessary for the final injectivity conclusion: flat tori have $K_M=0$ and complete metrics, but their exponential maps from a tangent space wrap around with many sheets. Cartan-Hadamard answers both obstructions from the previous section: Rauch removes conjugate points, and simple connectedness removes multiple sheets of the exponential covering. The resulting manifolds behave, from any chosen base point, like global normal-coordinate spaces. [example: Hadamard Manifolds As Global Normal Coordinate Spaces] In $\mathbb R^n$ with its Euclidean metric, the geodesic starting at $p$ with initial velocity $v$ is \begin{align*} \gamma_v(t)=p+tv, \end{align*} because $\gamma_v''(t)=0$, $\gamma_v(0)=p$, and $\gamma_v'(0)=v$. Hence \begin{align*} \exp_p(v)=\gamma_v(1)=p+v. \end{align*} If $\exp_p(v)=\exp_p(w)$, then \begin{align*} p+v=p+w, \end{align*} and subtracting $p$ from both sides gives $v=w$. Thus $\exp_p:T_p\mathbb R^n\to \mathbb R^n$ is globally defined and globally injective; it is also onto because every $q\in\mathbb R^n$ satisfies \begin{align*} q=\exp_p(q-p). \end{align*} Hyperbolic space has constant sectional curvature $-1\le 0$, is complete, and is simply connected, so the same global conclusion follows from *Cartan Hadamard Theorem*: for every $p$, the map $\exp_p:T_p\mathbb H^n\to \mathbb H^n$ is a global diffeomorphism. More generally, if $(M,g)$ is complete, simply connected, and has $K_M\le0$, then *Cartan Hadamard Theorem* gives that $\exp_p:T_pM\to M$ is a global diffeomorphism for every $p\in M$. Therefore each $q\in M$ has a unique vector $v\in T_pM$ with \begin{align*} q=\exp_p(v). \end{align*} For $q\ne p$, this vector is nonzero, so it has the unique [polar decomposition](/theorems/3074) \begin{align*} v=r e,\qquad r=|v|>0,\qquad e=\frac{v}{|v|}\in S_pM. \end{align*} Thus every point of $M\setminus\{p\}$ is written uniquely as $\exp_p(r e)$ with $r>0$ and $e\in S_pM$. In this sense a Hadamard manifold has no conjugate locus and no cut locus: the exponential map gives one global normal-coordinate chart, and polar coordinates fail only at the chosen center $p$. [/example] Cartan-Hadamard is local-to-global in the strongest possible sense: an infinitesimal curvature inequality eliminates conjugate points, and simple connectedness eliminates multi-valued global geodesic coordinates. This is the prototype for cut-locus theory, where one separates failure of normal coordinates into conjugate points and multiple minimizing geodesics. Chapters 4 and 5 next weaken the conclusions in stages: Hessian and Laplacian comparison keep distance-function control, while Bonnet--Myers uses Ricci curvature to bound diameter. Bishop-Gromov volume comparison in Chapter 7 then turns the traced Jacobi-field estimates into volume monotonicity. Rauch comparison has supplied the basic control on how geodesics separate. The next chapter weakens this pointwise information into differential inequalities for the distance function, so that curvature can be read from the Hessian and Laplacian of distance. # 4. Hessian and Laplacian Comparison This chapter turns the Jacobi-field comparison results from the previous chapter into differential inequalities for distance functions. The main object is the distance $r(x)=d(p,x)$ from a fixed base point $p$, whose second derivatives encode the curvature of geodesic spheres centred at $p$. The new difficulty is that $r$ is smooth only away from $p$ and away from the cut locus, so the chapter also records the weak and barrier interpretations needed when comparison estimates are used under integrals or in maximum-principle arguments. ## Smooth Distance Away from the Cut Locus What can be differentiated when the distance function is not smooth everywhere? Before the cut locus the exponential map gives polar coordinates, and in that region the radial distance behaves like the coordinate $t$ along each unit-speed geodesic from $p$. We first name the function whose regularity will drive the rest of the chapter. [definition: Radial Distance Function] Let $(M,g)$ be a complete Riemannian manifold and let $p \in M$. The radial distance function from $p$ is the function $r_p:M\to [0,\infty)$ defined by \begin{align*} r_p(x)=d(p,x). \end{align*} [/definition] The radial distance function is globally meaningful, but it is not globally smooth. Hessian comparison requires ordinary second derivatives, so we need to separate the region where the minimizing radial geodesic is unique from the region where cut points intervene. This motivates the following definition of the smooth radial domain. [definition: Smooth Radial Domain] Let $(M,g)$ be complete and let $p\in M$. The smooth radial domain of $p$ is \begin{align*} \Omega_p=M\setminus (\{p\}\cup \operatorname{Cut}(p)), \end{align*} where $\operatorname{Cut}(p)$ is the cut locus of $p$. [/definition] The smooth radial domain is designed to remove exactly the known obstructions to polar coordinates. To use it, we need the analytic statement that $r_p$ is smooth there and that its gradient is the radial unit vector. This motivates the next theorem, which is the local foundation for every Hessian computation in the chapter. [quotetheorem:5355] [citeproof:5355] Completeness is used here to make minimizing geodesics and the cut-time description available globally from $p$; without it, geodesics can leave the manifold before the distance sphere is reached. Removing $p$ is also necessary, since even in $\mathbb R^n$ the function $|x|$ is not smooth at $0$, and removing the cut locus is necessary because on the round sphere the distance from the north pole is not differentiable at the south pole. The theorem does not say that $r_p$ has classical second derivatives everywhere; it only identifies the open region where ordinary Hessian computations are legitimate. The previous theorem turns the distance function into a differentiable object on $\Omega_p$. The next question is what its second derivative measures geometrically. Since level sets of $r$ are geodesic spheres, the Hessian should record how these spheres bend as their outward normal field changes. [definition: Shape Operator of a Geodesic Sphere] Let $x\in \Omega_p$ with $r(x)=t>0$, and let $S(p,t)$ be the geodesic sphere through $x$. Its outward unit normal at $x$ is $\nu=\nabla r$. The shape operator of $S(p,t)$ at $x$ is the [linear map](/page/Linear%20Map) \begin{align*} A_x:T_xS(p,t)\to T_xS(p,t),\qquad A_x(X)=\nabla_X\nu. \end{align*} [/definition] With this convention, the mean curvature of a Euclidean sphere of radius $t$ is \begin{align*} \frac{n-1}{t}. \end{align*} The shape operator is a hypersurface object, while Hessian comparison is stated for the function $r$. The following theorem identifies these two languages and also records that no second derivative appears in the radial direction. [quotetheorem:5356] [citeproof:5356] The tangency hypothesis matters: $A_x$ is defined only on the tangent space of the geodesic sphere, while the full tangent space also contains the radial direction. If the radial direction were not separated, the Euclidean example would already give a contradiction, since $\operatorname{Hess}|x|$ vanishes on the radial vector but equals $|X|^2/r$ on tangential vectors. The theorem is therefore an identification on $\Omega_p$, not a statement about nonsmooth cut points or about arbitrary extensions of $A_x$. The Euclidean case fixes the normalization and the sign convention. It is worth computing it before adding curvature, because every model comparison reduces to this behaviour near the base point. [example: Euclidean Geodesic Spheres] In $\mathbb R^n$ with $p=0$, fix $x\ne0$ and write $r=|x|$. The radial distance is $r(x)=|x|$, and for any vector $Y\in T_x\mathbb R^n\cong\mathbb R^n$, \begin{align*} dr_x(Y) &=\left.\frac{d}{ds}\right|_{s=0}|x+sY| \\ &=\left.\frac{d}{ds}\right|_{s=0}\bigl(\langle x+sY,x+sY\rangle\bigr)^{1/2} \\ &=\frac{1}{2}|x|^{-1}\cdot 2\langle x,Y\rangle \\ &=\left\langle \frac{x}{r},Y\right\rangle . \end{align*} Thus $\nabla r=x/r$. Since the Euclidean connection is ordinary differentiation, \begin{align*} \nabla_X\nabla r &=D_X\left(\frac{x}{r}\right) \\ &=\frac{X}{r}+x\,D_X(r^{-1}) \\ &=\frac{X}{r}-x\,r^{-2}D_Xr \\ &=\frac{X}{r}-x\,r^{-2}\frac{\langle x,X\rangle}{r} \\ &=\frac{X}{r}-\frac{\langle x,X\rangle}{r^3}x. \end{align*} Therefore \begin{align*} \operatorname{Hess}r(X,X) &=\langle \nabla_X\nabla r,X\rangle \\ &=\left\langle \frac{X}{r}-\frac{\langle x,X\rangle}{r^3}x,X\right\rangle \\ &=\frac{|X|^2}{r}-\frac{\langle x,X\rangle^2}{r^3}. \end{align*} If $X$ is tangent to the sphere of radius $r$, then $\langle x,X\rangle=0$, so \begin{align*} \operatorname{Hess}r(X,X)=\frac{|X|^2}{r}. \end{align*} For the radial unit vector $\partial_r=x/r$, \begin{align*} \operatorname{Hess}r(\partial_r,\partial_r) &=\frac{|\partial_r|^2}{r}-\frac{\langle x,\partial_r\rangle^2}{r^3} \\ &=\frac{1}{r}-\frac{r^2}{r^3} \\ &=0. \end{align*} The outward unit normal to the sphere is $\nu=\nabla r=x/r$, and for tangent $X$ the formula above gives $\nabla_X\nu=X/r$. Hence $A=(1/r)I$ on the tangent space of the sphere, so the trace over an orthonormal basis of the $(n-1)$-dimensional tangent space is $(n-1)/r$. [/example] This example is the model against which all comparison estimates should be read. Positive curvature focuses geodesics faster than Euclidean space, while negative curvature lets them spread faster. ## The Riccati Equation Along Radial Geodesics How does curvature enter the derivative of the shape operator? The Jacobi equation governs the angular variation fields along a radial geodesic, and differentiating the shape operator along the radial direction produces a matrix Riccati equation. Let $\gamma:[0,a)\to M$ be a unit-speed geodesic starting at $p$ and assume $\gamma(t)\in \Omega_p$ for $0<t<a$. Along $\gamma$, write $E(t)=\dot\gamma(t)^\perp\subset T_{\gamma(t)}M$ for the normal hyperplane. To state the evolution equation compactly, we name the curvature operator that acts on this normal hyperplane. [definition: Radial Curvature Endomorphism] Let $\gamma$ be a unit-speed geodesic. The radial curvature endomorphism along $\gamma$ is the family of self-adjoint maps $R_\gamma(t):\dot\gamma(t)^\perp\to \dot\gamma(t)^\perp$ defined by \begin{align*} R_\gamma(t)X=R(X,\dot\gamma(t))\dot\gamma(t). \end{align*} [/definition] The radial curvature endomorphism is the curvature term in the perpendicular Jacobi equation $J''+R_\gamma J=0$. To connect Jacobi fields with geodesic spheres, we need to rewrite this second-order equation in terms of the logarithmic derivative $A(t)J(t)=J'(t)$. This motivates the Riccati equation for the shape operator. [quotetheorem:5357] [citeproof:5357] The assumption that the geodesic remains in $\Omega_p$ is essential because the Jacobi fields with endpoint values on the geodesic sphere span the angular tangent space only before cut or conjugate behaviour intervenes. At a conjugate point the logarithmic derivative can blow up, while at a cut point the sphere may fail to be smooth as a hypersurface. The Riccati equation is therefore a smooth radial evolution equation, not a global equation for distance spheres across the cut locus. The matrix equation contains more information than many applications require. For volume growth and maximum principles, the relevant scalar is the trace of $A$, because this is both the mean curvature of the geodesic sphere and the Laplacian of the distance. We therefore introduce the trace before deriving its differential inequality. [definition: Radial Mean Curvature] Let $(M^n,g)$ be complete and let $p\in M$. The radial mean curvature of the geodesic spheres centred at $p$ is the function \begin{align*} m:\Omega_p\to\mathbb R,\qquad m(x)=\operatorname{tr}A_x. \end{align*} [/definition] Radial mean curvature converts the shape-operator equation into a scalar inequality. Since $\operatorname{Hess}r$ vanishes in the radial direction, $m=\Delta r$ on $\Omega_p$, so this scalar inequality is also an inequality for the Laplacian of distance. This motivates taking the trace of the Riccati equation and estimating the quadratic term by Cauchy--Schwarz. [quotetheorem:5358] [citeproof:5358] The inequality is one-sided because the estimate $\operatorname{tr}(A^2)\ge m^2/(n-1)$ is one-sided. Equality forces the shape operator to be a scalar multiple of the identity, as happens in space forms; without this umbilicity condition, the trace loses information about individual principal curvatures. This is why Ricci curvature is enough for an upper Laplacian comparison under a lower Ricci bound, but not enough by itself for the reverse comparison. The model spaces solve the Riccati equation with equality, so they give the comparison functions that appear later. Hyperbolic space is the most useful non-Euclidean model for geometric analysis, since its geodesic spheres approach a positive limiting mean curvature. [example: Hyperbolic Distance Spheres] In $\mathbb H^n_k$ with constant sectional curvature $k=-a^2<0$, the model function is \begin{align*} \operatorname{sn}_k(t)=a^{-1}\sinh(at). \end{align*} For a tangent direction to the geodesic sphere, the corresponding radial Jacobi field has length factor $\operatorname{sn}_k(t)$, so the shape operator is multiplication by the logarithmic derivative $\operatorname{sn}_k'(t)/\operatorname{sn}_k(t)$. Since \begin{align*} \operatorname{sn}_k'(t) &=\frac{d}{dt}\bigl(a^{-1}\sinh(at)\bigr) \\ &=a^{-1}\cdot a\cosh(at) \\ &=\cosh(at), \end{align*} we get \begin{align*} \frac{\operatorname{sn}_k'(t)}{\operatorname{sn}_k(t)} &=\frac{\cosh(at)}{a^{-1}\sinh(at)} \\ &=a\frac{\cosh(at)}{\sinh(at)} \\ &=a\coth(at). \end{align*} Thus on the tangent space of the radius-$t$ geodesic sphere, \begin{align*} A(t)=a\coth(at)I. \end{align*} Taking the trace over an orthonormal basis of the $(n-1)$-dimensional tangent space gives \begin{align*} \Delta r &=\operatorname{tr}A(t) \\ &=\sum_{i=1}^{n-1} a\coth(at) \\ &=(n-1)a\coth(at). \end{align*} For the large-radius limit, use \begin{align*} \coth(at) &=\frac{\cosh(at)}{\sinh(at)} \\ &=\frac{(e^{at}+e^{-at})/2}{(e^{at}-e^{-at})/2} \\ &=\frac{1+e^{-2at}}{1-e^{-2at}}, \end{align*} so $\coth(at)\to 1$ as $t\to\infty$, and therefore \begin{align*} \Delta r\to (n-1)a. \end{align*} Near the base point, the Taylor expansions $\sinh(at)=at+O(t^3)$ and $\cosh(at)=1+O(t^2)$ give \begin{align*} a\coth(at) &=a\frac{\cosh(at)}{\sinh(at)} \\ &=a\frac{1+O(t^2)}{at+O(t^3)} \\ &=\frac{1+O(t^2)}{t+O(t^3)} \\ &=\frac{1}{t}+O(t). \end{align*} Hence \begin{align*} \Delta r=(n-1)a\coth(at)=\frac{n-1}{t}+O(t), \end{align*} so hyperbolic geodesic spheres have the same leading mean-curvature singularity as Euclidean spheres near the centre, but their mean curvature approaches the positive constant $(n-1)a$ at infinity. [/example] The limiting value at infinity is one of the first signs that negative curvature creates exponential volume growth. The comparison theorems below package this calculation for any manifold with a sectional or Ricci curvature bound. ## Hessian Comparison Which curvature bound controls the full Hessian of $r$ rather than only its trace? A sectional curvature bound is the right hypothesis, because the Hessian in each tangential direction depends on the sectional curvature of the radial two-plane containing that direction. For a real number $k$, let $\operatorname{sn}_k$ be the model function from the space form of curvature $k$, and set \begin{align*} \operatorname{ct}_k(t)=\frac{\operatorname{sn}_k'(t)}{\operatorname{sn}_k(t)} \end{align*} on the interval where $\operatorname{sn}_k(t)>0$. Thus \begin{align*} \operatorname{ct}_0(t)=\frac{1}{t},\qquad \operatorname{ct}_{a^2}(t)=a\cot(at),\qquad \operatorname{ct}_{-a^2}(t)=a\coth(at). \end{align*} The function $\operatorname{ct}_k$ is the principal curvature of a radius-$t$ geodesic sphere in the simply connected space form of curvature $k$. Comparing a general geodesic sphere to this model is the full [Hessian comparison theorem](/theorems/5359). [quotetheorem:5359] [citeproof:5359] This theorem states that a lower bound on sectional curvature forces geodesic spheres to bend inward at least as fast as the model, which appears as an upper bound for the outward shape operator. The sign is often a source of mistakes: on the round sphere, $\operatorname{ct}_1(t)=\cot t$ decreases and becomes negative after $\pi/2$, matching the fact that distance spheres past the equator bend toward the base point. The sectional curvature hypothesis is stronger than a Ricci hypothesis because it controls each radial two-plane separately; a Ricci bound only controls the sum of these directions and cannot recover every Hessian component. The theorem also says nothing at cut points, where the distance sphere may have corners rather than a well-defined second fundamental form. [illustration:geometric-analysis-i-radial-jacobi-hessian-comparison] [example: Round Sphere and the Sign of Comparison] Represent the unit sphere as $S^n\subset\mathbb R^{n+1}$ and take the north pole to be $p=(1,0,\ldots,0)$. For $0<r<\pi$, a point away from the antipodal cut point can be written as \begin{align*} q=(\cos r,\sin r\,u), \end{align*} where $u\in S^{n-1}$. The outward radial unit vector at $q$ is \begin{align*} \nu=\nabla r=(-\sin r,\cos r\,u), \end{align*} because \begin{align*} |\nu|^2=\sin^2 r+\cos^2 r\,|u|^2=\sin^2 r+\cos^2 r=1 \end{align*} and $\nu$ is the velocity of the unit-speed meridian $r\mapsto(\cos r,\sin r\,u)$. Let $X$ be tangent to the geodesic sphere $r=\text{constant}$ at $q$. First take a curve $u(s)\subset S^{n-1}$ with $u(0)=u$ and $u'(0)=W$, so the corresponding curve in the geodesic sphere is \begin{align*} q(s)=(\cos r,\sin r\,u(s)). \end{align*} Its velocity at $s=0$ is \begin{align*} X=q'(0)=(0,\sin r\,W). \end{align*} Along the same curve, the radial normal field is \begin{align*} \nu(s)=(-\sin r,\cos r\,u(s)), \end{align*} so its ambient derivative is \begin{align*} \frac{d}{ds}\bigg|_{s=0}\nu(s)=(0,\cos r\,W). \end{align*} This vector is already tangent to $S^n$ at $q$, since \begin{align*} \langle (0,\cos r\,W),(\cos r,\sin r\,u)\rangle =\cos r\,\sin r\,\langle W,u\rangle =0, \end{align*} because $W\in T_uS^{n-1}$. Hence the Levi-Civita derivative on $S^n$ is \begin{align*} \nabla_X\nu=(0,\cos r\,W). \end{align*} Since \begin{align*} \cot r\,X =\cot r\,(0,\sin r\,W) =(0,\cos r\,W), \end{align*} we get \begin{align*} A X=\nabla_X\nu=\cot r\,X. \end{align*} Therefore, for every vector $X$ tangent to the geodesic sphere, \begin{align*} \operatorname{Hess}r(X,X) &=\langle A X,X\rangle \\ &=\langle \cot r\,X,X\rangle \\ &=\cot r\,|X|^2. \end{align*} Near the north pole, \begin{align*} \cot r &=\frac{\cos r}{\sin r} \\ &=\frac{1+O(r^2)}{r+O(r^3)} \\ &=\frac{1}{r}+O(r), \end{align*} so $\operatorname{Hess}r(X,X)=|X|^2/r+O(r)|X|^2$, matching the Euclidean leading term. At the equator $r=\pi/2$, $\cot r=0$, so the Hessian in tangential directions vanishes. For $\pi/2<r<\pi$, one has $\sin r>0$ and $\cos r<0$, hence $\cot r<0$; the negative sign records that the outward radial normals are converging toward the south pole. [/example] The Hessian theorem is powerful but demands sectional curvature control. Many geometric analysis arguments need only the scalar inequality for $\Delta r$, and Ricci curvature is the correct weaker hypothesis for that purpose. ## Laplacian Comparison What remains true if only Ricci curvature is bounded below? The [trace Riccati inequality](/theorems/5358) shows that the mean curvature $m=\Delta r$ can still be compared with the model mean curvature, because Ricci curvature is exactly the trace of the radial curvature operator. [quotetheorem:5360] [citeproof:5360] The lower Ricci bound is used exactly through the traced curvature term, and the conclusion is correspondingly only an upper bound for $\Delta r$. The theorem does not provide a lower bound for $\Delta r$ from an upper Ricci bound alone, because the term $\operatorname{tr}(A^2)$ was estimated in only one direction. For lower Hessian or lower Laplacian estimates one returns to sectional curvature comparison or imposes additional structure that controls the full shape operator. The leading singularity matters: both the actual and model mean curvatures behave like $(n-1)/t$ near the base point. This shared asymptotic anchors the Riccati comparison at $t=0$ even though neither side is finite there. [example: Maximum-Principle Consequence from Laplacian Comparison] Assume $\operatorname{Ric}\ge 0$ on a complete noncompact $n$-manifold, fix $p\in M$, and write $r(x)=d(p,x)$. On the smooth radial domain $\Omega_p$, the Laplacian comparison estimate for the case $k=0$ gives \begin{align*} \Delta r\le (n-1)\operatorname{ct}_0(r)=\frac{n-1}{r}, \end{align*} where $\operatorname{ct}_0(r)=1/r$. For $R>0$, set \begin{align*} \psi(x)=R^2-r(x)^2. \end{align*} Since $R^2$ is constant, $\Delta \psi=-\Delta(r^2)$. To compute $\Delta(r^2)$ on $\Omega_p$, apply the chain rule for the Laplacian: \begin{align*} \Delta(r^2) &=2r\,\Delta r+2|\nabla r|^2. \end{align*} On $\Omega_p$, the distance function satisfies $|\nabla r|=1$, so \begin{align*} \Delta(r^2) &=2r\,\Delta r+2. \end{align*} Using $\Delta r\le (n-1)/r$ and $r>0$ on $M\setminus\{p\}$, we get \begin{align*} 2r\,\Delta r &\le 2r\cdot \frac{n-1}{r} \\ &=2(n-1). \end{align*} Therefore \begin{align*} \Delta(r^2) &=2r\,\Delta r+2 \\ &\le 2(n-1)+2 \\ &=2n. \end{align*} Equivalently, \begin{align*} \Delta\psi=-\Delta(r^2)\ge -2n \end{align*} on the smooth radial domain. This is the estimate used in maximum-principle arguments for functions such as $u-\varepsilon r^2$; if the contact point lies on the cut locus, the same comparison input must be interpreted through the barrier form of Laplacian comparison rather than by differentiating $r$ classically there. [/example] The example also indicates why the cut locus cannot be ignored. A maximum point of an auxiliary function may lie on the cut locus, exactly where $r$ is not smooth. ## The Cut Locus, Barriers, and Weak Inequalities How should comparison inequalities be used at points where $r$ is not differentiable? The correct answer depends on the argument: maximum principles need barriers, while [integration by parts](/theorems/210) requires distributional or weak formulations. [definition: Upper Barrier] Let $u:M\to\mathbb R$ be continuous and let $x\in M$. An upper barrier for $u$ at $x$ is a smooth map $\phi:U\to\mathbb R$, where $U\subset M$ is an open neighbourhood of $x$, such that \begin{align*} \phi(x)=u(x),\qquad \phi(y)\ge u(y) \end{align*} for all $y\in U$. [/definition] An upper barrier is only a touching function; it does not yet encode any differential inequality. For comparison geometry we need the stronger statement that near every nonsmooth point there are touching smooth functions whose Laplacians satisfy the desired bound at the contact point. This motivates the barrier Laplacian inequality. [definition: Barrier Laplacian Inequality] Let $u:M\to\mathbb R$ be continuous and let $F:M\to\mathbb R$ be continuous. We say that $\Delta u\le F$ in the upper barrier sense if, for every $x\in M$ and every $\varepsilon>0$, there is an upper barrier $\phi_\varepsilon$ for $u$ at $x$ such that \begin{align*} \Delta \phi_\varepsilon(x)\le F(x)+\varepsilon. \end{align*} [/definition] The distance function has natural barriers obtained by moving the base point slightly along the minimizing geodesic. This construction is needed because a cut point may have several minimizing geodesics from $p$, but after shifting the base point along one chosen geodesic, the remaining segment becomes smooth for the local argument. This is the Calabi trick. [quotetheorem:5361] [citeproof:5361] The chosen minimizing geodesic is part of the data; at a cut point there may be several such geodesics, and the barrier is built from one of them rather than from a globally smooth choice. The conclusion is local and one-sided: it gives a smooth function touching from above, not a smooth replacement for $r_p$ in a whole neighbourhood of the cut locus. This distinction is what makes the construction suitable for maximum principles but insufficient by itself for integration by parts. [illustration:geometric-analysis-i-calabi-trick] This trick is safe for pointwise maximum-principle arguments because it replaces the nonsmooth function by a smooth function touching from above at the point under consideration. The next theorem records the precise barrier version of Laplacian comparison, which is the form to use when an auxiliary function reaches its extremum at a cut point. [quotetheorem:5362] [citeproof:5362] A barrier statement handles contact points but does not by itself justify integration by parts across the cut locus. For integral estimates we need a distributional formulation, phrased against nonnegative test functions. This form tracks the singular second-derivative contribution of $r$ without pretending that $r$ is smooth everywhere. [definition: Weak Laplacian Inequality] Let $u\in L^1_{\mathrm{loc}}(M)$ and let $F\in L^1_{\mathrm{loc}}(M)$. We say $\Delta u\le F$ weakly on an open set $U\subset M$ if for every nonnegative $\phi\in C_c^\infty(U)$, \begin{align*} \int_U u\,\Delta\phi\,d\operatorname{vol}_g\le \int_U F\phi\,d\operatorname{vol}_g. \end{align*} [/definition] The sign convention here matches the distributional identity $(\Delta u)(\phi)=\int u\Delta\phi\,d\operatorname{vol}_g$. Under this convention, upper bounds for $\Delta r$ pass to integral estimates. The point of the next theorem is that the cut locus contributes with the correct sign, so no extra positive error term appears. [quotetheorem:5363] [citeproof:5363] The weak theorem is the version to cite inside an integral. It prevents a common mistake: treating the cut locus as harmless merely because it has measure zero. The following warning states the safe rule for later geometric analysis arguments. [remark: Safe Use in Integration Arguments] When integrating Laplacian comparison, apply the weak formulation directly to a nonnegative [test function](/page/Test%20Function) or justify the passage by smoothing or exhaustion. Do not write an integral over $M$ as though $r$ were smooth across $\operatorname{Cut}(p)$. The cut locus has measure zero for first-order integration, but second-order integration by parts can still detect it through a singular distributional term. [/remark] This distinction is especially important in geometric analysis, where cutoff functions and maximum principles are applied repeatedly. The barrier form controls contact points, while the weak form controls integrals. [example: Cutoff Estimates from Distance Comparison] Assume $\operatorname{Ric}\ge0$, fix $R>0$, and let $\eta:[0,\infty)\to[0,1]$ be smooth, nonincreasing, supported in $[0,2R)$, and equal to $1$ on $[0,R]$. Put $r(x)=d(p,x)$ and $\chi(x)=\eta(r(x))$. On the smooth radial domain, the chain rule for gradients gives \begin{align*} \nabla\chi &=\eta'(r)\nabla r. \end{align*} Taking the divergence and using the product rule, \begin{align*} \Delta\chi &=\operatorname{div}\bigl(\eta'(r)\nabla r\bigr) \\ &=\langle \nabla(\eta'(r)),\nabla r\rangle+\eta'(r)\Delta r \\ &=\eta''(r)\langle \nabla r,\nabla r\rangle+\eta'(r)\Delta r. \end{align*} Since $|\nabla r|=1$ away from $p$ and the cut locus, this becomes \begin{align*} \Delta\chi=\eta''(r)+\eta'(r)\Delta r. \end{align*} By *Laplacian Comparison in Weak Form*, the Ricci lower bound gives \begin{align*} \Delta r\le \frac{n-1}{r} \end{align*} weakly on $M\setminus\{p\}$. Because $\eta'\le0$, multiplying this inequality by $\eta'(r)$ reverses the order: \begin{align*} \eta'(r)\Delta r\ge \eta'(r)\frac{n-1}{r}. \end{align*} Therefore the cutoff satisfies, in the weak sense, \begin{align*} \Delta\chi &\ge \eta''(r)+\eta'(r)\frac{n-1}{r}. \end{align*} If $\eta$ is chosen with scale bounds \begin{align*} |\eta'(t)|\le \frac{C_1}{R},\qquad |\eta''(t)|\le \frac{C_2}{R^2}, \end{align*} then on the annulus where $\eta'$ or $\eta''$ may be nonzero, \begin{align*} \Delta\chi &\ge -\frac{C_2}{R^2}-\frac{C_1(n-1)}{Rr}. \end{align*} Thus the radial cutoff has a distributional Laplacian lower bound controlled by the scale terms $R^{-2}$ and $R^{-1}(n-1)/r$, while the weak formulation accounts for the cut locus without differentiating $r$ there. [/example] The chapter's main lesson is that comparison geometry gives differential inequalities for distance functions, but the meaning of those inequalities changes at the cut locus. In the smooth radial domain, Hessian and Laplacian comparison are ordinary tensor inequalities. At nonsmooth points, the Calabi trick supplies barriers for maximum principles, and weak formulations supply the version needed for integration. Hessian and Laplacian comparison give local differential control of distance, but the consequences become global only when combined with completeness and the behavior of minimizing geodesics. The next chapter uses the same Ricci-based inequalities to bound diameter and to extract topological information from positive Ricci curvature. # 5. Diameter, Fundamental Group, and Positive Ricci Curvature This chapter turns the comparison theory of Jacobi fields into global restrictions on the size and topology of a complete Riemannian manifold. The previous chapter used Ricci curvature, through the traced Riccati inequality, to compare the Laplacian of distance; here the same lower bound is used along a single geodesic to force negative second variation, bound diameter, and prove compactness. The main theorem is Bonnet--Myers: a positive Ricci lower bound prevents geodesics from minimizing beyond the diameter of the corresponding round sphere. ## Geodesic Focusing from Ricci Lower Bounds The guiding question is how an averaged curvature inequality, such as a lower bound for Ricci curvature, can force every long geodesic to stop minimizing. Sectional curvature controls each two-plane separately, but Ricci curvature controls the trace of the curvature endomorphism transverse to a geodesic. The index form is the mechanism that converts this trace control into a negative second variation for sufficiently long geodesics. Recall that if $\gamma:[0,\ell]\to M$ is a unit-speed geodesic, the admissible variation fields form the vector space \begin{align*} \mathcal X_0(\gamma)=\{V\in C^\infty(\gamma^*TM): V(0)=0,\ V(\ell)=0\}. \end{align*} The index form is the symmetric bilinear map \begin{align*} I_\gamma:\mathcal X_0(\gamma)\times \mathcal X_0(\gamma)\to \mathbb R \end{align*} defined by \begin{align*} I_\gamma(V,W)=\int_0^\ell \left((D_tV,D_tW)- (R(V,\dot\gamma)\dot\gamma,W)\right)\,dt. \end{align*} For a minimizing geodesic segment without endpoint conjugacy, the second variation of energy gives $I_\gamma(V,V)\ge 0$ for all $V\in \mathcal X_0(\gamma)$. To contradict minimality, it is enough to build endpoint-vanishing test fields whose total index is negative. [definition: Ricci Lower Bound] Let $(M,g)$ be an $n$-dimensional Riemannian manifold and let $K\in\mathbb R$. We say that $M$ satisfies $\operatorname{Ric}\ge (n-1)K$ if \begin{align*} \operatorname{Ric}_p(v,v)\ge (n-1)K\,g_p(v,v) \end{align*} for every $p\in M$ and every $v\in T_pM$. [/definition] The normalization $(n-1)K$ is chosen so that the simply connected space form of constant sectional curvature $K$ has exactly this Ricci curvature. To use this lower bound in the index form, we need a test-field estimate that sums over all transverse directions and replaces individual sectional terms by their Ricci trace. The following focusing estimate is the local engine behind the diameter theorem. [quotetheorem:5364] [citeproof:5364] The estimate says that a geodesic longer than the model diameter has a direction in which length can be decreased through endpoint-fixed variations. This is the analytic form of geodesic focusing: positive Ricci curvature makes neighboring geodesics bend toward each other strongly enough to obstruct long minimizing segments. Each hypothesis has a specific role in the argument. The condition $K>0$ is what lets the curvature term dominate the derivative term in the sine test fields; if $K\le 0$, the same computation cannot force a negative index, as Euclidean space and flat tori have no positive focusing scale. The length assumption \begin{align*} \ell>\frac{\pi}{\sqrt K} \end{align*} is also sharp for this test: at the model length the summed index is zero, and below that scale the estimate has the wrong sign. Finally, Ricci control is used only after summing over a full transverse orthonormal frame; a lower bound on an unrelated average, or on only some transverse directions, would not control the trace term appearing in the summed index form. [example: Round Sphere Diameter] On the round sphere of sectional curvature $K>0$, the radius is $R=1/\sqrt K$. If $\gamma$ is a unit-speed geodesic and $E_1,\dots,E_{n-1}$ is an orthonormal frame perpendicular to $\dot\gamma$, then each two-plane $\operatorname{span}\{\dot\gamma,E_i\}$ has sectional curvature $K$, so \begin{align*} \operatorname{Ric}(\dot\gamma,\dot\gamma) &=\sum_{i=1}^{n-1} (R(E_i,\dot\gamma)\dot\gamma,E_i)\\ &=\sum_{i=1}^{n-1} K\\ &=(n-1)K. \end{align*} Thus $\operatorname{Ric}=(n-1)K g$, exactly matching the normalization in the Ricci lower bound. For two points on the same great circle with central angle $\theta\in[0,\pi]$, the shorter great-circle arc has length \begin{align*} R\theta=\frac{\theta}{\sqrt K}. \end{align*} If a unit-speed great circle segment has length $\ell\le \pi/\sqrt K$, then $\theta=\sqrt K\,\ell\le \pi$, and the spherical distance between its endpoints is \begin{align*} d(\gamma(0),\gamma(\ell)) =\frac{\theta}{\sqrt K} =\frac{\sqrt K\,\ell}{\sqrt K} =\ell. \end{align*} So such a segment minimizes between its endpoints. For antipodal points, $\theta=\pi$, hence \begin{align*} d(p,-p)=\frac{\pi}{\sqrt K}. \end{align*} The conjugate-point scale is the same. If $E$ is a parallel unit field perpendicular to $\dot\gamma$, then \begin{align*} J(t)=\frac{1}{\sqrt K}\sin(\sqrt K\,t)\,E(t) \end{align*} satisfies \begin{align*} D_t^2J+R(J,\dot\gamma)\dot\gamma &=-K\frac{1}{\sqrt K}\sin(\sqrt K\,t)E +K\frac{1}{\sqrt K}\sin(\sqrt K\,t)E\\ &=0, \end{align*} so $J$ is a Jacobi field. Its zeros occur when $\sin(\sqrt K\,t)=0$, and the first positive zero is \begin{align*} t=\frac{\pi}{\sqrt K}. \end{align*} Thus the focusing estimate detects exactly the model diameter: below that time great-circle segments still minimize, at that time antipodal points occur, and the first conjugate point appears at the same scale. [/example] The sphere example also explains why the theorem uses a strict inequality in the focusing estimate. At the model length, the sine test fields have zero total index on the constant-curvature model, so the argument detects loss of strict minimality only after passing the model diameter. ## Bonnet--Myers and Diameter Control The next problem is to pass from the infinitesimal second-variation obstruction to a global bound on all distances. Completeness enters at exactly this point: Hopf--Rinow guarantees that any two points can be joined by a minimizing geodesic. If the distance were too large, the preceding index estimate would contradict the minimizing property. [quotetheorem:2734] [citeproof:2734] Bonnet--Myers is a diameter theorem rather than a volume theorem. Unlike Bishop--Gromov comparison, it requires a strictly positive Ricci lower bound and produces a bound on the whole [metric space](/page/Metric%20Space) rather than only on ball growth. Completeness is needed twice: first to ensure that distances are realized by minimizing geodesics, and then to turn boundedness into compactness through Hopf--Rinow. A round open annulus in $S^2$ has positive Ricci curvature with the induced metric, but it is not complete and is not compact, so the compactness conclusion would fail without the completeness hypothesis. Connectedness ensures that all points lie in one metric component and that a single ball centered at $p$ covers the manifold; a disjoint union of two round spheres satisfies the same local Ricci inequality componentwise, but it is not a single connected metric space with one global diameter comparison. The positivity condition is also essential: flat tori have $\operatorname{Ric}=0$ and can be rescaled to have arbitrarily large diameter. [example: Normalized Diameter Estimates] Let $(M^n,g)$ be complete and connected. If \begin{align*} \operatorname{Ric}\ge (n-1)g, \end{align*} then this is the normalized Ricci lower bound with $K=1$, since \begin{align*} (n-1)K g=(n-1)\cdot 1\cdot g=(n-1)g. \end{align*} By *Bonnet--Myers Theorem*, \begin{align*} \operatorname{diam}(M)\le \frac{\pi}{\sqrt{1}}=\pi. \end{align*} More generally, suppose \begin{align*} \operatorname{Ric}\ge c g \end{align*} for some constant $c>0$. To put this into the form $\operatorname{Ric}\ge (n-1)K g$, choose $K$ so that \begin{align*} (n-1)K=c. \end{align*} Dividing by $n-1$ gives \begin{align*} K=\frac{c}{n-1}. \end{align*} Since $c>0$, this $K$ is positive, so *Bonnet--Myers Theorem* gives \begin{align*} \operatorname{diam}(M) &\le \frac{\pi}{\sqrt K}\\ &=\frac{\pi}{\sqrt{c/(n-1)}}\\ &=\pi\sqrt{\frac{n-1}{c}}. \end{align*} This estimate has the correct behavior under homothetic rescaling. Let $\tilde g=\lambda^2 g$ with $\lambda>0$. Distances scale by $\lambda$, so \begin{align*} \operatorname{diam}(M,\tilde g)=\lambda\,\operatorname{diam}(M,g). \end{align*} For a constant rescaling, the Levi-Civita connection is unchanged, hence the Ricci tensor as a $(0,2)$-tensor is unchanged: \begin{align*} \operatorname{Ric}_{\tilde g}=\operatorname{Ric}_g. \end{align*} If $\operatorname{Ric}_g\ge c g$, then \begin{align*} \operatorname{Ric}_{\tilde g} &=\operatorname{Ric}_g\\ &\ge c g\\ &=c\lambda^{-2}\tilde g. \end{align*} Thus the rescaled metric has lower bound constant $\tilde c=c/\lambda^2$, and the diameter estimate becomes \begin{align*} \operatorname{diam}(M,\tilde g) &\le \pi\sqrt{\frac{n-1}{\tilde c}}\\ &=\pi\sqrt{\frac{n-1}{c/\lambda^2}}\\ &=\lambda\pi\sqrt{\frac{n-1}{c}}, \end{align*} which is exactly the original bound multiplied by the distance-scaling factor $\lambda$. [/example] The compactness conclusion should not be confused with a statement about all compact manifolds. It says that the curvature lower bound forces compactness when completeness is already present; compact manifolds may have zero, negative, or sign-changing Ricci curvature. [example: Flat Tori and the Missing Positive Ricci Hypothesis] For $L>0$, let \begin{align*} T_L^n=\mathbb R^n/L\mathbb Z^n \end{align*} with the metric induced from the Euclidean metric on $\mathbb R^n$. The quotient map is a local isometry, so in the standard quotient charts the metric coefficients are constant: \begin{align*} g_{ij}=\delta_{ij}. \end{align*} Hence the Christoffel symbols are \begin{align*} \Gamma_{ij}^k=\frac12\sum_{m=1}^n g^{km} \left(\partial_i g_{jm}+\partial_j g_{im}-\partial_m g_{ij}\right) =\frac12\sum_{m=1}^n \delta^{km}(0+0-0) =0. \end{align*} Therefore each curvature component is \begin{align*} R^k{}_{\ell ij} =\partial_i\Gamma_{j\ell}^k-\partial_j\Gamma_{i\ell}^k +\sum_{m=1}^n\left(\Gamma_{im}^k\Gamma_{j\ell}^m-\Gamma_{jm}^k\Gamma_{i\ell}^m\right) =0-0+\sum_{m=1}^n(0-0) =0, \end{align*} and taking the trace gives \begin{align*} \operatorname{Ric}_{\ell j}=\sum_{k=1}^n R^k{}_{\ell kj}=0. \end{align*} Thus $\operatorname{Ric}\equiv 0$, so each $T_L^n$ has nonnegative Ricci curvature but no positive Ricci lower bound of the form $\operatorname{Ric}\ge (n-1)K g$ with $K>0$. The space $T_L^n$ is compact because it is the quotient of the compact cube $[0,L]^n$ after identifying opposite faces, and it is complete because every Euclidean line $t\mapsto x+tv$ projects to a geodesic $t\mapsto [x+tv]$ defined for all $t\in\mathbb R$. To see that the diameters are unbounded, compare the two points \begin{align*} p=[0],\qquad q=\left[\frac L2 e_1\right]. \end{align*} For any lattice vector $Lm$ with $m=(m_1,\dots,m_n)\in\mathbb Z^n$, the squared Euclidean length of a lift from $0$ to $\frac L2 e_1+Lm$ is \begin{align*} \left|\frac L2 e_1+Lm\right|^2 &=\left(L\left(m_1+\frac12\right)\right)^2+\sum_{j=2}^n (Lm_j)^2\\ &=L^2\left(\left(m_1+\frac12\right)^2+\sum_{j=2}^n m_j^2\right). \end{align*} Since $m_1$ is an integer, \begin{align*} \left|m_1+\frac12\right|\ge \frac12, \end{align*} and each $m_j^2\ge 0$, so \begin{align*} \left|\frac L2 e_1+Lm\right|^2\ge L^2\cdot \frac14. \end{align*} Taking square roots gives every such lift length at least $L/2$, while the lift with $m=0$ has length exactly $L/2$. Hence \begin{align*} d_{T_L^n}(p,q)=\frac L2, \end{align*} so \begin{align*} \operatorname{diam}(T_L^n)\ge \frac L2. \end{align*} As $L\to\infty$, these compact complete flat tori have $\operatorname{Ric}\ge 0$ and diameters tending to infinity, so nonnegative Ricci curvature alone cannot imply a uniform diameter bound. [/example] The flat torus example shows that positivity is needed for metric diameter control. The next question is whether the same hypothesis also restricts the fundamental group. The universal cover converts this into another Bonnet--Myers application, because Ricci curvature and completeness are preserved by local isometries. [quotetheorem:5365] [citeproof:5365] This corollary is often the first indication that curvature controls topology, not only metric size. It rules out many compact manifolds, including all flat tori, from carrying complete metrics with Ricci bounded below by a positive constant. The hypotheses should be read exactly as in Bonnet--Myers, because the proof applies that theorem to the universal cover. Completeness is needed so that the lifted metric on $\tilde M$ is complete; without it, a round open annulus in $S^2$ has positive Ricci curvature and fundamental group $\mathbb Z$, but it is incomplete. Connectedness fixes a single base component and identifies the deck group with one fundamental group; disconnected unions reduce to separate statements about components rather than one global conclusion. The strict positivity of the Ricci lower bound is necessary: every flat torus is complete and connected with infinite fundamental group, but its Ricci curvature is zero. The conclusion is also only finiteness of $\pi_1(M)$; it does not imply simple connectivity, since spherical space forms can have positive Ricci curvature and nontrivial finite fundamental group. ## Synge-Type Topological Consequences Bonnet--Myers uses Ricci curvature and gives finiteness of the fundamental group. The final question in this chapter asks what stronger topological conclusions follow from positive sectional curvature. [Synge's theorem](/theorems/2730) adds orientability and parity assumptions to convert curvature positivity into restrictions on geodesic loops and orientation-reversing isometries. [quotetheorem:2730] [citeproof:2730] The theorem is not a replacement for Bonnet--Myers. It assumes compactness and positive sectional curvature, while Bonnet--Myers derives compactness from positive Ricci curvature. Its conclusion is more topological and more sensitive to dimension and orientability. The parity and orientability assumptions are necessary, not cosmetic. In odd dimensions, spherical space forms such as lens spaces have positive sectional curvature and may have nontrivial finite fundamental group, so the even-dimensional simply connected conclusion cannot be extended to odd dimensions. In even dimensions, orientability cannot be dropped because $\mathbb{RP}^{2m}$ has positive sectional curvature but is nonorientable and has fundamental group $\mathbb Z/2\mathbb Z$. Compactness supplies the shortest geodesic or shortest deck-transformation segment used in the proof; on noncompact positively curved manifolds such minimizing representatives need not exist in the required class. Positive sectional curvature is stronger than positive Ricci curvature and is used pointwise in the selected two-plane; flat tori and many positively Ricci curved spherical space-form examples show that weaker curvature hypotheses do not force Synge's conclusions. [example: Lens Spaces and Orientability Caveats] For a concrete family, fix integers $p\ge 2$ and $q_1,\dots,q_m$ with $\gcd(p,q_j)=1$ for each $j$, and let $\mathbb Z_p$ act on the unit sphere $S^{2m-1}\subset \mathbb C^m$ by \begin{align*} [z]\cdot (z_1,\dots,z_m) = \left(e^{2\pi i q_1/p}z_1,\dots,e^{2\pi i q_m/p}z_m\right). \end{align*} If the $k$th power of this generator fixes $(z_1,\dots,z_m)$, then for each index $j$ with $z_j\ne 0$, \begin{align*} e^{2\pi i kq_j/p}z_j=z_j \end{align*} implies \begin{align*} e^{2\pi i kq_j/p}=1, \end{align*} so $p\mid kq_j$. Since $\gcd(p,q_j)=1$, this gives $p\mid k$. Thus no nonidentity element fixes a point of $S^{2m-1}$, because at least one coordinate $z_j$ is nonzero. The quotient \begin{align*} L(p;q_1,\dots,q_m)=S^{2m-1}/\mathbb Z_p \end{align*} is therefore a smooth spherical space form. The quotient map \begin{align*} \pi:S^{2m-1}\to L(p;q_1,\dots,q_m) \end{align*} is a local isometry for the quotient metric. Hence every two-plane in the quotient lifts locally to a two-plane in the round sphere, and the sectional curvature is the same as upstairs: \begin{align*} \sec_{L(p;q_1,\dots,q_m)}=1. \end{align*} Each generator acts on every complex coordinate by a planar rotation, whose real determinant is \begin{align*} \det \begin{pmatrix} \cos(2\pi q_j/p)&-\sin(2\pi q_j/p)\\ \sin(2\pi q_j/p)&\cos(2\pi q_j/p) \end{pmatrix} = \cos^2(2\pi q_j/p)+\sin^2(2\pi q_j/p) =1. \end{align*} Multiplying over the $m$ complex coordinates gives total real determinant $1$, so the action preserves orientation and the quotient is orientable. Since $S^{2m-1}$ is simply connected when $m\ge 2$, the deck group of the covering is exactly $\mathbb Z_p$, and therefore \begin{align*} \pi_1\left(L(p;q_1,\dots,q_m)\right)\cong \mathbb Z_p. \end{align*} Thus these odd-dimensional positively curved manifolds are orientable but need not be simply connected, so the even-dimensional simply connected conclusion cannot be extended to all dimensions. [/example] The parity assumptions also explain why projective spaces appear differently across dimensions. Real projective space $\mathbb{RP}^{2m}$ is nonorientable and has positive sectional curvature induced from the sphere, so it is excluded from the even-dimensional orientable hypothesis. Real projective space $\mathbb{RP}^{2m+1}$ is orientable and has fundamental group $\mathbb Z/2\mathbb Z$, which is compatible with the odd-dimensional conclusion. [remark: Comparison Philosophy] In Bonnet--Myers, averaging sectional curvatures through Ricci curvature supplies many test fields at once and gives a universal diameter bound. In Synge's theorem, positivity of each sectional curvature supplies a sharper index contradiction once topology produces a special geodesic and orientation supplies a compatible parallel normal field. Both arguments show how the second variation translates curvature positivity into global restrictions. [/remark] The three results of this chapter form a hierarchy. Positive Ricci curvature on a complete manifold forces bounded diameter, compactness, and finite fundamental group. Positive sectional curvature, together with dimension and orientability hypotheses, forces stronger conclusions about simple connectivity or orientability. The common engine is the same: long or topologically distinguished geodesics cannot remain minimizing when curvature makes the index form negative. The diameter and fundamental-group results show how index-form arguments force global restrictions when curvature is positive enough. The next chapter keeps the same comparison philosophy but shifts from geodesics and second variation to the geometry of triangles, where Toponogov theory captures curvature through angle and distance inequalities. # 6. Triangle Comparison and Toponogov Theory This chapter continues the comparison-geometry part of the course by turning the local estimates from Jacobi fields into global statements about metric triangles. The prerequisites are the preceding material on geodesics and cut loci from Chapter 1, conjugate points and the index form from Chapter 2, Rauch comparison from Chapter 3, and the constant-curvature model spaces. The new guiding question is: if every sectional curvature is bounded below by a constant $k$, how much thicker must geodesic triangles be than their counterparts in the simply connected surface of constant curvature $k$? The answer is Toponogov comparison, which becomes the main bridge from curvature bounds to convexity, diameter estimates, and rigidity. ## Model Triangles and Comparison Angles The first problem is to translate a triangle in an arbitrary Riemannian manifold into a triangle in a geometry where side lengths and angles can be computed exactly. The comparison model is the complete simply connected two-dimensional space form $M_k^2$ of constant sectional curvature $k$: the Euclidean plane for $k=0$, the sphere of radius \begin{align*} \frac{1}{\sqrt{k}} \end{align*} for $k>0$, and the hyperbolic plane of curvature $k$ for $k<0$. [definition: Model Triangle] Let $a,b,c>0$ lie in the admissible side-length region for $M_k^2$: the triangle inequalities hold, and when $k>0$ the perimeter and side lengths are restricted as specified below. A model triangle with side lengths $a,b,c$ in $M_k^2$ is a geodesic triangle $\bar{p}\bar{q}\bar{r}\subset M_k^2$ such that \begin{align*} d_{M_k^2}(\bar{q},\bar{r}) &= a, & d_{M_k^2}(\bar{p},\bar{r}) &= b, & d_{M_k^2}(\bar{p},\bar{q}) &= c. \end{align*} [/definition] For $k\le 0$, the usual triangle inequalities are enough for existence and uniqueness up to isometry. For $k>0$, the model is compact and geodesics stop being uniquely minimizing at the antipodal scale, so we impose \begin{align*} a+b+c<\frac{2\pi}{\sqrt{k}}. \end{align*} In the local comparison statements below the relevant side lengths are also kept below \begin{align*} \frac{\pi}{\sqrt{k}}, \end{align*} which keeps the comparison triangle away from the antipodal ambiguity. Under these restrictions, the model triangle is unique up to an isometry of $M_k^2$, so its angles are well defined functions of the original side lengths; this is why the next definition can turn metric data alone into the comparison angles used in Toponogov inequalities. [illustration:geometric-analysis-i-hinge-comparison] The existence and uniqueness of model triangles lets side-length data produce a model angle even when the original space has no linear tangent plane. We need the following definition to make this angle available as the local measuring device for Toponogov-style comparisons before stating the first triangle inequalities. [definition: Comparison Angle] Let $(X,d)$ be a metric space and let $k\in\mathbb R$. The $k$-comparison angle is the partially defined function \begin{align*} \widetilde{\angle}_k:(p,q,r)\longmapsto \widetilde{\angle}_k qpr \end{align*} on triples of distinct points $p,q,r\in X$ for which a model triangle $\bar{p}\bar{q}\bar{r}\subset M_k^2$ exists with the same side lengths as $pqr$. Its value is \begin{align*} \widetilde{\angle}_k qpr := \angle \bar{q}\bar{p}\bar{r}. \end{align*} [/definition] Thus $\widetilde{\angle}_k qpr\in[0,\pi]$ whenever it is defined. The comparison angle is a metric substitute for the angle between initial velocities of geodesics. Actual Riemannian angles require chosen geodesic segments and their initial tangent vectors, and in a general metric space neither object is available. Side lengths, however, make sense in every metric space, so the comparison angle extracts the angle that a model triangle with those side lengths would have. In a smooth manifold, if $q$ and $r$ approach $p$ along two specified geodesics, the comparison angle converges to the Riemannian angle between their initial tangent vectors. [example: Spherical Comparison Angle] On the unit sphere $S^2=M_1^2$, take two unit-speed minimizing geodesic segments of length $a=b=\pi/3$ issuing from the north pole and meeting there with included angle $\gamma=\pi/2$. The side $c$ opposite $\gamma$ is computed from the spherical cosine law: \begin{align*} \cos c &=\cos a\cos b+\sin a\sin b\cos\gamma\\ &=\cos\left(\frac{\pi}{3}\right)\cos\left(\frac{\pi}{3}\right) +\sin\left(\frac{\pi}{3}\right)\sin\left(\frac{\pi}{3}\right)\cos\left(\frac{\pi}{2}\right)\\ &=\frac{1}{2}\cdot\frac{1}{2}+\frac{\sqrt{3}}{2}\cdot\frac{\sqrt{3}}{2}\cdot 0\\ &=\frac{1}{4}. \end{align*} Since $c\in(0,\pi)$ and $\cos$ is strictly decreasing on $[0,\pi]$, this gives \begin{align*} c=\arccos\left(\frac{1}{4}\right). \end{align*} For the $1$-comparison angle, the model space is again $M_1^2=S^2$, so the model triangle has the same side lengths and the same included angle: \begin{align*} \widetilde{\angle}_1 qpr=\gamma=\frac{\pi}{2}. \end{align*} For the Euclidean comparison angle $\theta=\widetilde{\angle}_0 qpr$ with the same three side lengths, the Euclidean cosine law gives \begin{align*} c^2 &=a^2+b^2-2ab\cos\theta\\ &=\left(\frac{\pi}{3}\right)^2+\left(\frac{\pi}{3}\right)^2 -2\left(\frac{\pi}{3}\right)\left(\frac{\pi}{3}\right)\cos\theta, \end{align*} hence \begin{align*} \cos\theta &=\frac{a^2+b^2-c^2}{2ab}\\ &=\frac{2(\pi/3)^2-\arccos(1/4)^2}{2(\pi/3)^2}\\ &=1-\frac{9\arccos(1/4)^2}{2\pi^2}. \end{align*} Numerically, \begin{align*} \arccos\left(\frac{1}{4}\right)\approx 1.318 \qquad\text{and}\qquad \frac{\sqrt{2}\pi}{3}\approx 1.481, \end{align*} so $c^2<2(\pi/3)^2$ and therefore $\cos\theta>0$. Since $\theta\in[0,\pi]$, this implies \begin{align*} \widetilde{\angle}_0 qpr=\theta<\frac{\pi}{2}. \end{align*} Thus the same three side lengths produce a right comparison angle in the spherical model but an acute comparison angle in the Euclidean model, reflecting the positive angular excess of spherical triangles. [/example] The example shows that model angles are computable only once the model side-angle relations are known. In variable-curvature comparison the side lengths of a hinge are often known, while the angle information is hidden. The comparison machinery therefore needs an exact model-plane dictionary: given two adjacent sides and the included angle, it must determine the opposite side, and conversely it must recover a comparison angle from three side lengths. The constant-curvature cosine laws provide that dictionary in the Euclidean, spherical, and hyperbolic models. [quotetheorem:5366] [citeproof:5366] The signs in the spherical and hyperbolic formulas explain the qualitative picture. Positive curvature makes triangles fatter than Euclidean triangles with the same two adjacent sides and included angle, while negative curvature makes them thinner. The hypotheses in the theorem are also part of the statement, not bookkeeping: the three positive side lengths must be realised by a minimizing geodesic triangle in the chosen model, and the angle $\alpha$ must be the angle between the two sides adjacent to $b$ and $c$. If the triangle inequalities fail, as in side lengths $1,1,3$, no geodesic triangle with those side lengths exists in any metric model, so the formula has no geometric object to describe. The spherical restrictions are not cosmetic. Once a side reaches the antipodal length, a pair of points on the round sphere can be joined by many minimizing geodesics, so a triple of side lengths no longer determines a unique model triangle or a well-defined comparison angle. There are also global side-length restrictions: three arcs of total length $2\pi/\sqrt{k}$ can wrap around a great circle, and the intended minimizing triangle can degenerate or fail to be unique. The cosine laws do not assert that an arbitrary geodesic triangle in a variable-curvature manifold satisfies these identities; a small triangle on a surface with nonconstant curvature has curvature-error terms, and Toponogov will replace exact identities by inequalities. Their role in the next section is to convert a metric inequality for the opposite side of a hinge into an angle inequality for a triangle with prescribed side lengths. [example: Thin Triangles in Negative Curvature] In the hyperbolic plane of curvature $-1$, take a geodesic hinge with adjacent sides $R,R$ and included angle $\alpha\in(0,\pi]$, and let $c$ be the opposite side. By the hyperbolic case of the *[Constant Curvature Law of Cosines](/theorems/5366)*, \begin{align*} \cosh c &=\cosh R\cosh R-\sinh R\sinh R\cos\alpha\\ &=\cosh^2 R-\sinh^2 R\cos\alpha. \end{align*} Using \begin{align*} \cosh R=\frac{e^R+e^{-R}}{2}, \qquad \sinh R=\frac{e^R-e^{-R}}{2}, \end{align*} we get \begin{align*} \cosh^2 R &=\left(\frac{e^R+e^{-R}}{2}\right)^2\\ &=\frac{e^{2R}+2+e^{-2R}}{4}, \end{align*} and \begin{align*} \sinh^2 R &=\left(\frac{e^R-e^{-R}}{2}\right)^2\\ &=\frac{e^{2R}-2+e^{-2R}}{4}. \end{align*} Therefore \begin{align*} \cosh c &=\frac{e^{2R}+2+e^{-2R}}{4} -\frac{e^{2R}-2+e^{-2R}}{4}\cos\alpha\\ &=\frac{(1-\cos\alpha)e^{2R}+2(1+\cos\alpha)+(1-\cos\alpha)e^{-2R}}{4}\\ &=\frac{1-\cos\alpha}{4}e^{2R} +\frac{1+\cos\alpha}{2} +\frac{1-\cos\alpha}{4}e^{-2R}. \end{align*} Since $\alpha>0$, we have $1-\cos\alpha>0$, so \begin{align*} \cosh c =\frac{1-\cos\alpha}{4}e^{2R}\left(1+O(e^{-2R})\right) \end{align*} as $R\to\infty$. Because \begin{align*} \cosh c=\frac{e^c+e^{-c}}{2} =\frac{e^c}{2}\left(1+e^{-2c}\right), \end{align*} and the previous line implies $c\to\infty$, we have $e^{-2c}=O(e^{-4R})$. Hence \begin{align*} \frac{e^c}{2}\left(1+e^{-2c}\right) &=\frac{1-\cos\alpha}{4}e^{2R}\left(1+O(e^{-2R})\right), \end{align*} so \begin{align*} e^c &=\frac{1-\cos\alpha}{2}e^{2R}\left(1+O(e^{-2R})\right). \end{align*} Taking logarithms gives \begin{align*} c &=\log\left(\frac{1-\cos\alpha}{2}e^{2R}\left(1+O(e^{-2R})\right)\right)\\ &=2R+\log\left(\frac{1-\cos\alpha}{2}\right)+\log\left(1+O(e^{-2R})\right)\\ &=2R+\log\left(\frac{1-\cos\alpha}{2}\right)+o(1). \end{align*} Thus, for fixed positive opening angle, the opposite side is almost the sum of the two long sides, up to a constant error depending only on $\alpha$; this is the metric sense in which large hyperbolic triangles become thin and tripod-like. [/example] ## Alexandrov Inequalities from Lower Sectional Curvature Bounds The main problem is to replace constant curvature formulas by inequalities on a manifold whose sectional curvature is only bounded below. Lower curvature bounds should force geodesics starting from the same point to spread no faster than they do in the model space $M_k^2$. The result is an Alexandrov-type triangle comparison statement. [definition: Geodesic Hinge] Let $(M,g)$ be a Riemannian manifold. A geodesic hinge consists of two unit-speed geodesic segments $\gamma_1:[0,a]\to M$ and $\gamma_2:[0,b]\to M$ with common initial point $p=\gamma_1(0)=\gamma_2(0)$. Its angle is \begin{align*} \alpha = \angle(\dot{\gamma}_1(0),\dot{\gamma}_2(0)). \end{align*} [/definition] A hinge is the infinitesimal data from which a triangle is built. The hinge formulation of Toponogov answers the first comparison question: after fixing two side lengths and their included angle, how large can the opposite side be under a lower sectional curvature bound? [quotetheorem:5367] [citeproof:5367] The hinge theorem controls one side from two sides and an included angle. Completeness is needed to make the global comparison statement stable under the geodesic constructions used in the proof; an open hemisphere of the round sphere has the same local curvature as the sphere but is not complete, and minimizing segments between points approaching the deleted equator can leave the space. The lower sectional curvature bound is the comparison input: a surface with a narrow region of curvature far below $k$ can let initially close geodesics spread faster than the $M_k^2$ model, reversing the desired endpoint inequality. The minimizing hypothesis is essential because a non-minimizing side can wind around a flat cylinder or sphere and produce hinge data unrelated to the metric distance between its endpoints. The theorem also has built-in limitations. It does not compare arbitrary geodesic arcs with common initial point; it compares minimizing segments whose endpoint distance is measured in the ambient metric. The spherical smallness assumptions keep the model hinge inside the range where the opposite side is uniquely determined by $a,b,\alpha$ and remains below the antipodal scale; beyond that range, the same data can correspond to different model configurations or to no canonical minimizing triangle. In applications, a metric triangle is usually specified by its three side lengths, and the geometric information we want is how its angles compare with the model angles. This leads to the full triangle form of Toponogov. [quotetheorem:5368] [citeproof:5368] The theorem says that lower curvature bounds produce angle excess. A triangle in a manifold with $\sec_M\ge 0$ has angles no smaller than its Euclidean comparison angles, and the sum of its angles is at least $\pi$ whenever the triangle is inside a region where the comparison applies. Each hypothesis rules out a real failure mode. If a chosen side is not minimizing, its initial angle can describe a long geodesic loop rather than the metric triangle determined by $p,q,r$. If $k>0$ and the side lengths cross the antipodal scale, the comparison angle may depend on which model triangle is chosen, so the statement loses a canonical meaning. Nonunique minimizing geodesics can also produce several possible Riemannian angles at the same vertex; Toponogov applies to a fixed choice of minimizing sides, not to an unspecified angle. These limitations explain why later rigidity arguments spend considerable effort controlling cut loci and equality cases. [example: Geodesic Bigons and Uniqueness Phenomena] Assume $(M,g)$ is complete and satisfies $\sec_M\ge 1$. Let $\gamma_0,\gamma_1:[0,L]\to M$ be two unit-speed minimizing geodesics from $p$ to $q$, with \begin{align*} L=d(p,q)<\pi. \end{align*} They form a geodesic bigon. For $0<t<L$, set \begin{align*} x_t=\gamma_0(t), \qquad y_t=\gamma_1(t), \qquad \ell_t=d(x_t,y_t). \end{align*} The triangle with vertices $p,x_t,y_t$ has two sides of length $t$ from $p$ and third side $\ell_t$. Since $\sec_M\ge 1$, the triangle form of *Toponogov Comparison Theorem* gives \begin{align*} \angle x_t p y_t \ge \widetilde{\angle}_1 x_t p y_t, \end{align*} whenever the spherical comparison triangle is in the admissible range. In the unit spherical model, the comparison angle $\theta_t=\widetilde{\angle}_1 x_t p y_t$ is determined by the spherical cosine law: \begin{align*} \cos \ell_t &=\cos t\cos t+\sin t\sin t\cos\theta_t\\ &=\cos^2 t+\sin^2 t\cos\theta_t. \end{align*} Solving for $\cos\theta_t$ gives \begin{align*} \sin^2 t\cos\theta_t &=\cos \ell_t-\cos^2 t,\\ \cos\theta_t &=\frac{\cos \ell_t-\cos^2 t}{\sin^2 t}. \end{align*} Thus, if $\ell_t\to 0$, then $\cos\ell_t\to 1$, and hence \begin{align*} \cos\theta_t &\to \frac{1-\cos^2 t}{\sin^2 t}\\ &=\frac{\sin^2 t}{\sin^2 t}\\ &=1. \end{align*} Since $\theta_t\in[0,\pi]$ and $\cos$ is strictly decreasing on $[0,\pi]$, this implies $\theta_t\to 0$. This is the sense in which a geodesic bigon is tested by thin spherical comparison triangles: as the third side tends to zero, the corresponding spherical comparison angle collapses to zero. In a strongly convex normal ball, however, any two points are joined by a unique minimizing geodesic lying in the ball, so two distinct short minimizers from $p$ to $q$ cannot form a positive-width bigon there. Near the cut locus, uniqueness of minimizers fails, so the same comparison is an obstruction to uncontrolled opening angles rather than a global uniqueness theorem. [/example] ## Convexity, Diameter, and Rigidity Applications The final problem is to see what triangle comparison buys beyond a single triangle. Because Toponogov controls how geodesic segments sit relative to one another, it gives convexity statements, diameter bounds, and rigidity conclusions when equality occurs. [definition: Geodesically Convex Subset] Let $(M,g)$ be a Riemannian manifold. A subset $C\subset M$ is geodesically convex if for every $p,q\in C$ there exists a minimizing geodesic segment from $p$ to $q$ whose image is contained in $C$. [/definition] The definition isolates the metric outcome supplied by comparison arguments: minimizing chords should remain inside the region being studied. It does not by itself explain how convexity is verified, because a set may be defined analytically by a boundary condition or as a distance sublevel set rather than by listing its geodesics. The next criterion supplies that verification step. It converts a Hessian inequality for a signed distance function into local metric convexity, so later arguments can combine Hessian comparison or Toponogov control with a boundary barrier instead of checking every minimizing segment directly. This formulation keeps the normal convention explicit and avoids hiding the boundary sign inside a comparison argument. [quotetheorem:5369] [citeproof:5369] This convexity result is a template rather than a universal theorem about all closed sets. Closedness ensures that the boundary separates the region from its exterior in the limiting argument; for an open ball in Euclidean space, endpoints can approach the missing boundary and the corresponding closed-chord conclusion is not a statement about the original open set. The $C^2$ boundary assumption is used to make the signed distance smooth in a collar and to relate the boundary value of $\operatorname{Hess}\rho$ to the second fundamental form \begin{align*} \operatorname{II}_{\nu}(v,v)=g(\nabla_v\nu,v), \qquad v\in T(\partial C), \end{align*} for the outward normal. If the boundary is only Lipschitz, as at the corner of a square, the signed distance need not have a classical Hessian in any collar, so the barrier proof has no pointwise second-derivative inequality to use. The collar smoothness is separate from [boundary regularity](/theorems/99): at focal points of the boundary, even a smooth hypersurface can have a signed distance function that loses smoothness, as happens near the centre of a round annulus where closest boundary points are not unique. The geodesic hypotheses also have concrete roles. Uniqueness of the minimizing geodesic in $U$ keeps the proof attached to a single curve on which $\rho\circ\gamma$ is convex; on a round sphere near antipodal configurations, several minimizing segments may exist and some can leave a chosen neighbourhood or cross the wrong side of a barrier. Requiring the Hessian inequality throughout the collar, not only on $\partial C$, prevents a geodesic from slipping through a region where $\rho$ becomes concave after leaving the boundary; a hypersurface can have nonnegative second fundamental form at one boundary point while the parallel hypersurfaces immediately outside develop the opposite sign. With this convention, ordinary convex domains in Euclidean space have $\operatorname{II}_{\nu}\ge 0$ on the boundary, but local metric convexity requires the corresponding signed-distance Hessian inequality in the collar where the geodesic is tested. In comparison geometry that collar inequality is often obtained from Hessian comparison, itself a consequence of the same Jacobi-field and Toponogov machinery developed earlier. In rigidity arguments this is often the point where analytic control enters the metric proof: Toponogov controls triangles, Hessian comparison controls distance barriers, and the convexity criterion turns those controls into a statement that whole minimizing geodesics stay inside the comparison region. If the boundary bends the wrong way, as for the exterior of a Euclidean ball with its own outward normal, two nearby points in the region may be joined by a shorter chord passing outside the region. In applications, the set $C$ is usually a metric ball, a sublevel set of a distance function, or a domain bounded by a hypersurface with controlled second fundamental form. [example: Spherical Caps] On the unit sphere $S^n\subset \mathbb R^{n+1}$, fix a centre $o\in S^n$ and let \begin{align*} C=\{x\in S^n:d(o,x)\le r\}, \qquad 0<r<\frac{\pi}{2}. \end{align*} We show that the shorter great-circle segment between any two points of $C$ stays in $C$. Let $x,y\in C$, and set $\delta=d(x,y)$. Since \begin{align*} \delta\le d(x,o)+d(o,y)\le 2r<\pi, \end{align*} the shorter minimizing segment from $x$ to $y$ is unique unless $x=y$, and for $x\ne y$ it is \begin{align*} \sigma(t) = \frac{\sin((1-t)\delta)}{\sin\delta}x + \frac{\sin(t\delta)}{\sin\delta}y, \qquad 0\le t\le 1. \end{align*} Because $x,y\in C$, \begin{align*} \langle o,x\rangle=\cos d(o,x)\ge \cos r, \qquad \langle o,y\rangle=\cos d(o,y)\ge \cos r. \end{align*} Taking the Euclidean inner product with $o$ gives \begin{align*} \langle o,\sigma(t)\rangle &= \frac{\sin((1-t)\delta)}{\sin\delta}\langle o,x\rangle + \frac{\sin(t\delta)}{\sin\delta}\langle o,y\rangle\\ &\ge \cos r\, \frac{\sin((1-t)\delta)+\sin(t\delta)}{\sin\delta}. \end{align*} Now \begin{align*} \sin((1-t)\delta)+\sin(t\delta) &= 2\sin\left(\frac{\delta}{2}\right) \cos\left(\frac{(1-2t)\delta}{2}\right), \end{align*} while \begin{align*} \sin\delta = 2\sin\left(\frac{\delta}{2}\right)\cos\left(\frac{\delta}{2}\right). \end{align*} Since $0\le |1-2t|\le 1$ and $0<\delta/2<\pi/2$, \begin{align*} \cos\left(\frac{(1-2t)\delta}{2}\right) \ge \cos\left(\frac{\delta}{2}\right)>0, \end{align*} so \begin{align*} \frac{\sin((1-t)\delta)+\sin(t\delta)}{\sin\delta}\ge 1. \end{align*} Therefore \begin{align*} \langle o,\sigma(t)\rangle\ge \cos r. \end{align*} Since $d(o,z)=\arccos\langle o,z\rangle$ on the unit sphere and $\arccos$ is decreasing on $[-1,1]$, this implies \begin{align*} d(o,\sigma(t))\le r. \end{align*} Thus $\sigma([0,1])\subset C$. This agrees with the boundary-Hessian test. If $s=d(o,\cdot)$ and $\rho=s-r$, then $\rho\le 0$ on $C$ and the outward unit normal to $\partial C$ is $\nu=\partial_s$. In geodesic polar coordinates around $o$, \begin{align*} g=ds^2+\sin^2 s\,g_{S^{n-1}}, \end{align*} so for any vector $v$ tangent to a geodesic sphere $s=\text{constant}$, \begin{align*} \nabla_v\partial_s=\frac{\cos s}{\sin s}v=\cot s\,v. \end{align*} Hence on $\partial C$, where $s=r$, \begin{align*} \operatorname{II}_{\nu}(v,v) &=g(\nabla_v\nu,v)\\ =g(\nabla_v\partial_s,v)\\ &=\cot r\,|v|^2>0. \end{align*} More generally, if $w=a\partial_s+v$ with $v\perp\partial_s$, then \begin{align*} \operatorname{Hess}\rho(w,w) = \operatorname{Hess}s(w,w) = \cot s\,|v|^2, \end{align*} which is nonnegative in any collar with $s<\pi/2$. The condition $r<\pi/2$ is exactly what keeps the boundary bending inward toward the cap, so minimizing chords cannot leave the cap and return. [/example] [illustration:geometric-analysis-i-spherical-cap-convexity] The spherical cap example shows that positive curvature has a distinguished global scale: the antipodal distance on the model sphere. Bonnet-Myers gives only the estimate $\operatorname{diam}(M)\le \pi/\sqrt{k}$ under $\operatorname{Ric}\ge (n-1)k$, and the estimate alone does not identify the manifold. For instance, real projective space with the quotient of the round metric has $\sec\equiv 1$ but diameter $\pi/2$, while many positively curved metrics have diameter strictly below the spherical value. Equality at the full antipodal scale is therefore a much stronger condition than an ordinary upper bound: it says that two points are as far apart as the model permits. This raises the rigidity question for complete manifolds with $\sec_M\ge k>0$: if the diameter reaches the largest value allowed by the model, must the whole manifold be the model sphere? [remark: Quoted result: Maximal Diameter Theorem] Let $(M^n,g)$ be a complete connected $n$-dimensional Riemannian manifold with $\sec_M\ge k>0$. By Bonnet-Myers, these hypotheses make $M$ compact. If \begin{align*} \operatorname{diam}(M)=\frac{\pi}{\sqrt{k}}, \end{align*} then $(M,g)$ is isometric to the standard round $n$-sphere of radius $1/\sqrt{k}$, equivalently the simply connected space form of constant sectional curvature $k$. [/remark] The course uses this theorem as a capstone application of Toponogov rather than as a separate technical development; the full rigidity proof requires a detailed equality analysis for comparison triangles and cut-locus behaviour. The hypotheses include connectedness and dimension because the conclusion is the round sphere of the same dimension, not merely some space of constant curvature. Completeness and the positive lower curvature bound imply compactness by Bonnet-Myers, so the diameter is realised by a pair of points $p,q$ with $d(p,q)=\pi/\sqrt{k}$. The rigidity content is that every direction at $p$ behaves as if it points along a minimizing geodesic to the antipodal point $q$, and Toponogov forces the intervening geodesic triangles to match their spherical models. The lower sectional curvature bound is essential. A Ricci lower bound alone gives the same numerical diameter estimate but supports different equality and near-equality phenomena, while a nonpositive or mixed lower sectional curvature bound has no antipodal model scale. The theorem also does not say that a manifold with diameter close to $\pi/\sqrt{k}$ is itself round, nor does it classify spaces whose diameter is merely large. Quotients illustrate the point: $\mathbb{RP}^n$ with the round quotient metric has $\sec\equiv 1$ and is locally spherical, but its diameter is $\pi/2$, so the quotient obstruction is excluded by the maximal diameter equality rather than by a separate topological assumption. Below the maximal value, triangle comparison still gives strong inequalities, but the equality mechanism needed to unfold the whole manifold into a round sphere is absent. [explanation: Equality Cases in Toponogov] Equality in Toponogov is much stronger than equality in an ordinary metric triangle inequality. If a triangle has the same side lengths and angles as its model triangle, the Jacobi field comparison used in the proof must be sharp along the ruled surface swept out by the relevant geodesic variations. Sharpness forces the sectional curvatures of the planes tangent to that variation to equal the model curvature $k$ and prevents hidden cut-locus behaviour along the sides. This is the mechanism behind rigidity results. A comparison inequality first gives an extremal diameter, angle, or convexity conclusion; then equality pushes the local curvature and Jacobi field data to match the model. With enough overlapping triangles, the local model behaviour propagates across the manifold. [/explanation] The abstract equality principle becomes concrete on the round sphere, where every comparison inequality is realised by an actual model triangle. The following example identifies the extremal configuration that the maximal diameter theorem is designed to characterise. [example: Equality on a Round Hemisphere] Identify $S^n$ with the unit sphere in $\mathbb R^{n+1}$, take \begin{align*} p=e_{n+1}, \qquad q=-e_{n+1}, \end{align*} and let $x$ lie on the equator $x_{n+1}=0$. On the unit sphere, \begin{align*} d(u,v)=\arccos\langle u,v\rangle \end{align*} for $u,v\in S^n$. Since \begin{align*} \langle p,q\rangle &=\langle e_{n+1},-e_{n+1}\rangle\\ &=-1, \end{align*} we get \begin{align*} d(p,q)=\arccos(-1)=\pi. \end{align*} Also, \begin{align*} \langle p,x\rangle &=\langle e_{n+1},x\rangle\\ &=x_{n+1}\\ &=0, \end{align*} so \begin{align*} d(p,x)=\arccos(0)=\frac{\pi}{2}. \end{align*} Similarly, \begin{align*} \langle q,x\rangle &=\langle -e_{n+1},x\rangle\\ &=-x_{n+1}\\ &=0, \end{align*} and therefore \begin{align*} d(q,x)=\arccos(0)=\frac{\pi}{2}. \end{align*} Thus the three side lengths are \begin{align*} d(p,x)=\frac{\pi}{2},\qquad d(q,x)=\frac{\pi}{2},\qquad d(p,q)=\pi, \end{align*} and they satisfy \begin{align*} d(p,x)+d(x,q)=\frac{\pi}{2}+\frac{\pi}{2}=\pi=d(p,q). \end{align*} The corresponding comparison triangle in $S^2$ is therefore degenerate: its three vertices lie on one great circle, with the equatorial point between the two antipodal endpoints. Since every spherical distance lies in $[0,\pi]$ and the pair $p,q$ has distance $\pi$, the diameter of the unit round sphere is exactly $\pi$. This is the extremal configuration behind the maximal diameter statement: equality is realised by the round sphere itself, not merely by an abstract metric space with the same numerical diameter. [/example] The chapter leaves us with the central philosophy of comparison geometry: curvature bounds become metric inequalities, and equality in those inequalities remembers the model space. The next stage of the course uses the same principle for volume, where Jacobi field comparison controls not only distances and angles but also the measure distortion in geodesic polar coordinates. Triangle comparison has made the model spaces visible in a purely metric form. The next chapter applies the same Jacobi-field and Riccati machinery to geodesic polar coordinates, where the main object is no longer distance or angle but volume distortion. # 7. Bishop-Gromov Volume Comparison These notes are part of the comparison-geometry portion of Geometric Analysis I. The preceding chapters developed geodesic polar coordinates, Jacobi fields, Rauch comparison, and the traced Riccati inequality behind Laplacian comparison. This chapter turns those tools into estimates for the volume of geodesic balls. The guiding question is how a lower Ricci curvature bound controls the average spreading of geodesics from a point, even when sectional curvature is not bounded in every radial plane. The main prerequisites are the polar-coordinate volume formula, the cut locus convention, model space functions, and the Jacobi-field comparison results already proved in the course. The outcome is Bishop-Gromov monotonicity: relative ball volumes decrease when compared with the appropriate simply connected model space. ## Volume Density in Polar Coordinates How does curvature enter the volume form near a base point? In normal coordinates, the exponential map converts radial distance and angular direction into points of the manifold, and the Riemannian volume measure acquires a Jacobian factor measuring the infinitesimal spreading of geodesics. [definition: Radial Volume Density] Let $(M^n,g)$ be a complete Riemannian manifold and let $p \in M$. Let \begin{align*} c:S_pM\to(0,\infty] \end{align*} be the cut-time function, so $c(\theta)$ is the cut time in direction $\theta$, and set \begin{align*} \mathcal D_p=\{(r,\theta):\theta\in S_pM,\ 0<r<c(\theta)\}. \end{align*} The radial volume density is the function \begin{align*} J_p:\mathcal D_p\to(0,\infty) \end{align*} determined by \begin{align*} (\exp_p)^*(d\operatorname{vol}_g) = J_p(r,\theta)\,dr\,d\theta \end{align*} in geodesic polar coordinates $(r,\theta)$. [/definition] The factor $J_p(r,\theta)$ is the determinant of the transverse Jacobi field matrix along the radial geodesic. In Euclidean space it is $r^{n-1}$, so every deviation from $r^{n-1}$ records curvature-induced focusing or spreading. [example: Euclidean Volume Density] In $\mathbb R^n$ with its flat metric, geodesic polar coordinates at $0$ are \begin{align*} \Phi:(0,\infty)\times S^{n-1}\to \mathbb R^n\setminus\{0\}, \qquad \Phi(r,\theta)=r\theta. \end{align*} The radial derivative is $\partial_r\Phi=\theta$, and if $v\in T_\theta S^{n-1}$ is tangent to the unit sphere, then \begin{align*} d\Phi_{(r,\theta)}(v)=rv. \end{align*} Choose an orthonormal basis $e_1,\dots,e_{n-1}$ of $T_\theta S^{n-1}$. Since $\theta$ is orthogonal to $T_\theta S^{n-1}$, the vectors \begin{align*} \partial_r\Phi=\theta,\qquad d\Phi(e_1)=re_1,\dots,d\Phi(e_{n-1})=re_{n-1} \end{align*} are mutually orthogonal in $\mathbb R^n$, with lengths \begin{align*} |\theta|=1,\qquad |re_i|=r. \end{align*} Therefore the Gram matrix of the pulled-back Euclidean metric in the basis $(\partial_r,e_1,\dots,e_{n-1})$ is \begin{align*} \begin{pmatrix} 1 & 0 & \cdots & 0\\ 0 & r^2 & \cdots & 0\\ \vdots & \vdots & \ddots & \vdots\\ 0 & 0 & \cdots & r^2 \end{pmatrix}, \end{align*} so its determinant is \begin{align*} 1\cdot (r^2)^{n-1}=r^{2n-2}, \end{align*} and the volume density is the square root, \begin{align*} \sqrt{r^{2n-2}}=r^{n-1}. \end{align*} Thus \begin{align*} d\mathcal L^n=r^{n-1}\,dr\,d\theta, \end{align*} and hence $J_0(r,\theta)=r^{n-1}$ for every $\theta\in S^{n-1}$. Integrating this density over the Euclidean ball gives \begin{align*} \operatorname{vol}(B(0,R)) &=\int_{S^{n-1}}\int_0^R r^{n-1}\,dr\,d\theta\\ &=\int_{S^{n-1}}\left[\frac{r^n}{n}\right]_{r=0}^{r=R}\,d\theta\\ &=\int_{S^{n-1}}\frac{R^n}{n}\,d\theta\\ &=\frac{\operatorname{vol}(S^{n-1})}{n}R^n. \end{align*} Since the Euclidean unit ball volume satisfies $\omega_n=\operatorname{vol}(S^{n-1})/n$, this is \begin{align*} \operatorname{vol}(B(0,R))=\omega_n R^n. \end{align*} The flat model therefore has exactly the reference density $r^{n-1}$, so later comparison ratios measure how curved geodesic spheres deviate from this Euclidean spreading. [/example] To compare this density with curved model spaces, we use the model sine functions from the first chapter. For $k\in\mathbb R$, define the model radius interval \begin{align*} I_k= \begin{cases} (0,\pi/\sqrt{k}), & k>0,\\ (0,\infty), & k\le 0, \end{cases} \end{align*} and define \begin{align*} \operatorname{sn}_k:I_k\to(0,\infty),\qquad r\mapsto \begin{cases} \frac{1}{\sqrt{k}}\sin(\sqrt{k}r), & k>0,\\ r, & k=0,\\ \frac{1}{\sqrt{-k}}\sinh(\sqrt{-k}r), & k<0. \end{cases} \end{align*} We also write \begin{align*} \operatorname{ct}_k:I_k\to\mathbb R,\qquad \operatorname{ct}_k(r)=\frac{\operatorname{sn}_k'(r)}{\operatorname{sn}_k(r)} \end{align*} for $r\in I_k$. The simply connected space form of constant sectional curvature $k$ has polar density $\operatorname{sn}_k(r)^{n-1}$ before its first conjugate point. [illustration:geometric-analysis-i-volume-density-cut-time] The first comparison theorem is the infinitesimal version of Bishop-Gromov. Since the polar volume element is controlled by the determinant of transverse Jacobi fields, a Ricci lower bound should constrain the logarithmic derivative of $J_p(r,\theta)$ against the corresponding model logarithmic derivative. [quotetheorem:5370] [citeproof:5370] This comparison is pointwise in the angular variable only up to the cut time. The Ricci lower bound is exactly what turns the matrix Jacobi equation into a scalar trace inequality; without it, the average expansion of radial geodesics can exceed the model even if some sectional curvatures are controlled. Hyperbolic space of constant curvature $-1$ is the basic model: if compared against the Euclidean model $k=0$, then $J_p(r,\theta)=\sinh(r)^{n-1}$, and \begin{align*} \frac{J_p(r,\theta)}{r^{n-1}} \end{align*} increases for large $r$ because the missing hypothesis $\operatorname{Ric}\ge 0$ fails. The cut-time restriction is also essential, because after the cut locus polar coordinates can count the same point more than once and the density is no longer a smooth local volume factor. The theorem does not compare individual sectional curvatures or guarantee absence of conjugate points; it gives a one-directional upper control on radial volume density. The next step is to integrate it over all directions while accounting for the fact that polar coordinates cease to be injective after the cut locus. [remark: Cut Locus Convention] The polar integration formula is used with the domain \begin{align*} \mathcal D_p=\{(r,\theta):\theta\in S_pM, 0<r<c(\theta)\}. \end{align*} The cut locus has Riemannian measure zero, so integrating over $\mathcal D_p$ computes the volume of geodesic balls centred at $p$. [/remark] ## Relative Volume Monotonicity What survives after integrating the radial comparison? The density estimate becomes a monotonicity theorem for ratios of ball volumes, comparing $M$ with the simply connected model space $M_k^n$ of constant sectional curvature $k$. [definition: Model Ball Volume] For $k \in \mathbb R$ and $n\ge 2$, the model ball volume is the function \begin{align*} V_k^n:I_k\to\mathbb R_{>0},\qquad R\mapsto \omega_{n-1}\int_0^R \operatorname{sn}_k(t)^{n-1}\,dt, \end{align*} where $\omega_{n-1}=\operatorname{vol}(S^{n-1})$ is the volume of the Euclidean unit $(n-1)$-sphere. [/definition] When $k>0$, we also use the continuous extension of $V_k^n$ to $R=\pi/\sqrt{k}$; its endpoint value is the total volume of the simply connected round sphere of sectional curvature $k$. The normalisation by $V_k^n(R)$ removes the growth rate forced by the model geometry. When $k=0$, this asks whether $\operatorname{vol}(B(p,R))/R^n$ decreases with $R$ under nonnegative Ricci curvature. [quotetheorem:5371] [citeproof:5371] The theorem is strongest because it compares all scales at once, but its direction is important. Without the Ricci lower bound there is no reason for the relative volume ratio to decrease. For instance, hyperbolic space $\mathbb H^n$ does not satisfy $\operatorname{Ric}\ge 0$, and \begin{align*} \frac{\operatorname{vol}(B(p,R))}{V_0^n(R)} = \frac{\omega_{n-1}\int_0^R\sinh(t)^{n-1}\,dt}{\omega_nR^n} \end{align*} eventually grows exponentially rather than decreasing. The result is also an upper relative growth theorem, not a noncollapsing theorem: a manifold can satisfy the same Ricci lower bound while having very small ball volumes. Completeness ensures that geodesic balls and radial geodesics exist at the scales being compared, and the base point matters because the comparison is made from polar coordinates centred at that point. Letting $r\to 0$ gives the absolute Bishop comparison, while comparing $R$ and $2R$ gives volume doubling estimates. [quotetheorem:5372] [citeproof:5372] Bishop comparison gives the coarse upper bound, but it should not be read as a two-sided volume estimate. The lower Ricci bound is essential: if radial Ricci is allowed to become more negative than the model, geodesic balls can grow faster than $V_k^n(R)$. The nearby hyperbolic model makes this failure concrete: $\mathbb H^n$ has $\operatorname{Ric}=-(n-1)g$, so its ball volumes grow exponentially and violate the Euclidean comparison for $k=0$. More flexible warped-product examples can force rapid volume growth over a chosen annulus by making the radial Ricci curvature very negative there. The theorem also does not prevent collapse, since quotienting or shrinking directions can keep Ricci bounded below while making volumes small. What the model volume function records is the maximal growth permitted by the curvature lower bound, and the following examples identify the two reference behaviours used throughout the rest of the course. [example: Hyperbolic And Euclidean Growth] In the Euclidean model, $\operatorname{sn}_0(t)=t$, so the model ball volume is \begin{align*} V_0^n(R) &=\omega_{n-1}\int_0^R t^{n-1}\,dt\\ &=\omega_{n-1}\left[\frac{t^n}{n}\right]_{t=0}^{t=R}\\ &=\frac{\omega_{n-1}}{n}R^n. \end{align*} Since $\omega_n=\omega_{n-1}/n$ for the Euclidean unit ball volume, this gives \begin{align*} V_0^n(R)=\omega_nR^n. \end{align*} For the hyperbolic model of constant curvature $-1$, $\operatorname{sn}_{-1}(t)=\sinh t$, hence \begin{align*} V_{-1}^n(R)=\omega_{n-1}\int_0^R \sinh(t)^{n-1}\,dt. \end{align*} Using \begin{align*} \sinh t=\frac{e^t-e^{-t}}{2}, \end{align*} we have for all $t\ge 0$, \begin{align*} \sinh t\le \frac{e^t}{2}, \end{align*} and for $t\ge 1$, \begin{align*} \sinh t &=\frac{e^t}{2}(1-e^{-2t})\\ &\ge \frac{e^t}{2}(1-e^{-2})\\ &\ge \frac{e^t}{4}. \end{align*} Therefore, for $R\ge 1$, \begin{align*} V_{-1}^n(R) &\le \omega_{n-1}\int_0^R \left(\frac{e^t}{2}\right)^{n-1}\,dt\\ &=\frac{\omega_{n-1}}{2^{n-1}}\int_0^R e^{(n-1)t}\,dt\\ &=\frac{\omega_{n-1}}{2^{n-1}(n-1)} \left(e^{(n-1)R}-1\right), \end{align*} while for $R\ge 2$, \begin{align*} V_{-1}^n(R) &\ge \omega_{n-1}\int_1^R \left(\frac{e^t}{4}\right)^{n-1}\,dt\\ &=\frac{\omega_{n-1}}{4^{n-1}}\int_1^R e^{(n-1)t}\,dt\\ &=\frac{\omega_{n-1}}{4^{n-1}(n-1)} \left(e^{(n-1)R}-e^{n-1}\right). \end{align*} Thus the hyperbolic model volume has exponential order $e^{(n-1)R}$. By Bishop-Gromov monotonicity, if $\operatorname{Ric}\ge 0$, then \begin{align*} \operatorname{vol}(B(p,R))\le V_0^n(R)=\omega_nR^n, \end{align*} so ball growth is at most Euclidean. If $\operatorname{Ric}\ge -(n-1)g$, then \begin{align*} \operatorname{vol}(B(p,R))\le V_{-1}^n(R), \end{align*} so the corresponding model upper growth is hyperbolic, with exponential rate $n-1$. [/example] The examples show upper growth rates from the base point scale, but they do not yet give the scale-to-scale control needed in analysis. A growth estimate from $0$ to $R$ does not by itself say how many $R$-balls are needed to control a $2R$-ball, nor does it give uniform constants in local integral estimates. Covering arguments require a controlled number of smaller balls to cover a larger ball, and estimates for averages compare $B(p,R)$ with $B(p,2R)$. The next theorem is needed to turn the relative model-volume comparison into this nested-ball form: it applies Bishop-Gromov at the two radii $R$ and $2R$ and isolates the model ratio as the doubling constant. This is also the point where the comparison theorem connects with metric-measure analysis outside Riemannian geometry. We need the following theorem because volume doubling is one of the structural hypotheses behind maximal-function estimates, Poincare inequalities, heat-kernel bounds, and measured Gromov-Hausdorff compactness; Bishop-Gromov supplies it from curvature rather than assuming it as an axiom. Later analytic arguments use this in a concrete way: doubling controls covering multiplicities in local estimates, noncollapsing converts averaged $L^2$ bounds into scale-invariant bounds, and heat-kernel or elliptic estimates use comparable volumes of nested balls to keep constants uniform across the manifold. [quotetheorem:5373] [citeproof:5373] Doubling prevents volume from expanding too quickly, but its hypotheses and scope are limited. The Ricci lower bound is what makes the constant depend only on the model ratio; without it, even nearby concentric balls can have uncontrolled volume growth. A concrete model is a rotationally symmetric metric on $\mathbb R^n$ of the form \begin{align*} g=dr^2+f(r)^2g_{S^{n-1}}, \end{align*} where $f(r)=r$ for $r\le 1$ and $f(r)$ rises very rapidly on $1<r<2$. Then $\operatorname{vol}(B(0,2))/\operatorname{vol}(B(0,1))$ can be made arbitrarily large by choosing the rise of $f$ large enough, and the radial Ricci curvature becomes correspondingly negative in the transition region. The estimate is local in scale through the parameter $R_0$, and for negative lower bounds the constant worsens as larger radii are allowed. It is also still an upper comparison, so compactness arguments require lower bounds preventing collapse. Bishop-Gromov supplies such lower bounds at smaller radii once a single reference ball has positive volume. [example: Noncollapsing From One Ball] Assume $\operatorname{Ric}\ge -(n-1)K g$ with $K\ge 0$, and suppose \begin{align*} \operatorname{vol}(B(p,R_0))\ge v_0>0. \end{align*} For $0<r\le R_0$, *Bishop-Gromov monotonicity* applied with model curvature $-K$ gives \begin{align*} \frac{\operatorname{vol}(B(p,R_0))}{V_{-K}^n(R_0)} \le \frac{\operatorname{vol}(B(p,r))}{V_{-K}^n(r)}. \end{align*} Since $V_{-K}^n(r)>0$ and $V_{-K}^n(R_0)>0$, multiplying both sides by $V_{-K}^n(r)$ gives \begin{align*} \frac{V_{-K}^n(r)}{V_{-K}^n(R_0)} \operatorname{vol}(B(p,R_0)) \le \operatorname{vol}(B(p,r)). \end{align*} Using $\operatorname{vol}(B(p,R_0))\ge v_0$ in the left-hand side yields \begin{align*} \operatorname{vol}(B(p,r)) &\ge \frac{V_{-K}^n(r)}{V_{-K}^n(R_0)} \operatorname{vol}(B(p,R_0))\\ &\ge \frac{V_{-K}^n(r)}{V_{-K}^n(R_0)}v_0. \end{align*} Thus a lower volume bound at one scale propagates to every smaller concentric ball, with the explicit constant $V_{-K}^n(r)V_{-K}^n(R_0)^{-1}v_0$ determined by the comparison model. [/example] ## Rigidity and Almost-Rigidity What does equality in Bishop-Gromov mean geometrically? Since the theorem was obtained by integrating a Riccati comparison, equality forces all intermediate inequalities to be equalities along almost every radial geodesic, and this is a strong local statement about the metric in polar coordinates. [quotetheorem:5374] [citeproof:5374] The local conclusion explains why equality is rare: the comparison forgets most sectional information, but equality restores it in the radial directions. The extra hypotheses are not cosmetic. On the round sphere, a ball annulus that crosses the antipodal cut point cannot be described by a single smooth polar chart even though the pre-cut radial densities are model densities. For trace loss, a family of transverse Jacobi fields with eigenvalues whose product matches the model determinant can have non-model individual eigenvalues; the ball-volume ratio records the determinant, not the full transverse metric. For angular data, a metric cone or a warped annulus of the form \begin{align*} dr^2+\operatorname{sn}_k(r)^2 h \end{align*} over an angular metric $h$ with the same total angular volume as the round sphere has the same radial volume profile but not the same angular geometry unless $h$ is the round metric. Global rigidity therefore requires hypotheses controlling the cut locus, smooth equality in the matrix comparison, and the angular metric data. [example: Positive Ricci And Maximal Volume] Let $(M^n,g)$ be complete with $\operatorname{Ric}\ge (n-1)g$, and fix $p\in M$. By the *Bonnet-Myers theorem*, $\operatorname{diam}(M)\le \pi$, so every $q\in M$ satisfies $d(p,q)\le \pi$. Thus $M$ is contained in the closed ball $\overline B(p,\pi)$, and the endpoint form of *Bishop Volume Comparison* gives \begin{align*} \operatorname{vol}(M) &\le V_1^n(\pi)\\ &=\omega_{n-1}\int_0^\pi \sin(t)^{n-1}\,dt\\ &=\operatorname{vol}(S^n), \end{align*} where $S^n$ is the unit round sphere. Assume now that equality holds: \begin{align*} \operatorname{vol}(M)=\operatorname{vol}(S^n)=V_1^n(\pi). \end{align*} For $0<R<\pi$, define \begin{align*} F(R)=\frac{\operatorname{vol}(B(p,R))}{V_1^n(R)}. \end{align*} By *Bishop-Gromov Monotonicity*, $F$ is nonincreasing. The small-radius asymptotics give \begin{align*} \lim_{R\to 0}F(R)=1, \end{align*} so $F(R)\le 1$ for every $0<R<\pi$. On the other hand, as $R\uparrow \pi$, \begin{align*} \operatorname{vol}(B(p,R))\to \operatorname{vol}(M), \qquad V_1^n(R)\to V_1^n(\pi), \end{align*} and therefore \begin{align*} \lim_{R\uparrow\pi}F(R) = \frac{\operatorname{vol}(M)}{V_1^n(\pi)} = 1. \end{align*} Since $F$ is nonincreasing and has limiting value $1$ at both ends, for every $0<R<\pi$ we have \begin{align*} 1=\lim_{s\uparrow\pi}F(s)\le F(R)\le \lim_{r\downarrow 0}F(r)=1, \end{align*} hence \begin{align*} F(R)=1. \end{align*} Thus every ball centered at $p$ has the same volume as the corresponding round spherical ball. The equality case in *Bishop-Gromov Monotonicity* then forces the radial volume density and radial shape operator to agree with the round model before the cut time: in polar coordinates, \begin{align*} J_p(r,\theta)=\sin(r)^{n-1} \end{align*} and the transverse metric has the round warping $\sin^2(r)g_{S^{n-1}}$ wherever the polar chart is smooth. Hence the maximal-volume case is precisely the rigid case behind the usual theorem: a complete manifold with $\operatorname{Ric}\ge(n-1)g$ and $\operatorname{vol}(M)=\operatorname{vol}(S^n)$ is the standard round sphere. [/example] Almost-rigidity is the analytic version of the same idea. If the Bishop-Gromov ratio nearly stays constant across an annulus, then the average radial density defect is small, so the mean curvature and Jacobi-field behaviour are close to those of the model in an integral sense. [remark: Almost-Rigidity Principle] Under a uniform lower Ricci bound, near equality in Bishop-Gromov on a ball is used to deduce quantitative closeness to a model ball after adding noncollapsing and compactness input. The precise conclusion depends on the topology of the region and on the metric notion of closeness being used, such as Gromov-Hausdorff distance or measured Gromov-Hausdorff distance. The course uses this principle as motivation for later compactness theorems rather than proving a full quantitative stability theorem here. [/remark] The almost-rigidity principle also warns that volume comparison by itself is only an upper control. To see the missing ingredient, it is useful to examine a sequence with perfect Ricci lower bounds but shrinking volume. [example: Why Noncollapsing Is Needed] For $\varepsilon>0$, let \begin{align*} M_\varepsilon=S^1_\varepsilon\times (S^1)^{n-1}, \end{align*} where $S^1_\varepsilon$ is the circle of radius $\varepsilon$ and each remaining circle has radius $1$, all with the product flat metric. Each factor is flat, and the curvature tensor of a Riemannian product is the sum of the curvature tensors of the factors on the corresponding tangent components, so \begin{align*} \operatorname{Ric}_{M_\varepsilon}=0. \end{align*} Thus the hypothesis $\operatorname{Ric}\ge 0$ holds uniformly in $\varepsilon$, and *Bishop Volume Comparison* gives the Euclidean upper bound \begin{align*} \operatorname{vol}(B_{M_\varepsilon}(p,R))\le \omega_n R^n \end{align*} for every $p\in M_\varepsilon$ and every $R>0$. The total volume is the product of the factor volumes: \begin{align*} \operatorname{vol}(M_\varepsilon) &=\operatorname{vol}(S^1_\varepsilon)\operatorname{vol}((S^1)^{n-1})\\ &=(2\pi\varepsilon)(2\pi)^{n-1}\\ &=(2\pi)^n\varepsilon. \end{align*} Since \begin{align*} \lim_{\varepsilon\downarrow 0}(2\pi)^n\varepsilon=0, \end{align*} the sequence collapses in volume even though the Ricci lower bound is fixed. The metric collapse is also visible directly. The projection \begin{align*} \pi_\varepsilon:S^1_\varepsilon\times (S^1)^{n-1}\to (S^1)^{n-1} \end{align*} does not increase distances, and each fiber is a circle of diameter $\pi\varepsilon$. Hence every point of $M_\varepsilon$ lies within distance at most $\pi\varepsilon$ of the section $\{x_0\}\times (S^1)^{n-1}$, which is isometric to $(S^1)^{n-1}$. As $\varepsilon\downarrow 0$, these fibers shrink to points, so the metric limit is the lower-dimensional flat torus $(S^1)^{n-1}$. This shows that lower Ricci bounds and Bishop-type upper volume bounds do not prevent collapse; a lower volume bound at one fixed scale is the extra input that keeps the limiting geometry $n$-dimensional. [/example] The chapter’s main lesson is that Ricci curvature controls volume through the trace of the Jacobi equation. Bishop-Gromov monotonicity is therefore a bridge between curvature and analysis: it supplies doubling, noncollapsing criteria, and rigidity mechanisms that will be used in compactness and convergence arguments. Bishop-Gromov comparison turns curvature control into monotonicity of volume ratios, linking geometry to measure and analysis. The next chapter builds on that bridge by converting volume and distance estimates into isoperimetric and analytic inequalities for functions on the manifold. # 8. Isoperimetric and Analytic Consequences of Comparison This chapter turns the volume comparison of Chapter 7, together with the diameter and distance-function estimates from Chapters 4 and 5, into analytic estimates. The guiding theme is that a curvature lower bound controls how sets separate the manifold, and separation controls functions through their level sets. We move from distance spheres and annuli, to isoperimetric constants, to eigenvalue and Sobolev estimates that are used throughout geometric analysis. ## Coarea and Area Comparison for Distance Spheres The first problem is how to convert information about a function into information about the hypersurfaces on which it is constant. For comparison geometry the most important function is the distance from a basepoint, since its level sets are geodesic spheres and its sublevel sets are metric balls. [quotetheorem:5375] [citeproof:5375] The formula says that the volume of a region can be recovered by integrating the areas of its slices. The Lipschitz hypothesis is what makes this statement stable: it gives an a.e. gradient by [Rademacher's theorem](/page/Rademacher's%20Theorem), prevents uncontrolled oscillation on null sets, and ensures that level sets carry the expected [Hausdorff measure](/page/Hausdorff%20Measure) for almost every level. For a rough measurable function, the preimages $u^{-1}(t)$ may have no geometric relation to the ambient volume, and there is no gradient term capable of recording how the slices are stacked. When $u(x)=r(x)=d(p,x)$, the slices are distance spheres, so this becomes the bridge between Bishop-Gromov volume comparison and hypersurface area comparison. [example: Euclidean Coarea for Radial Functions] In $\mathbb R^n$, let $u(x)=|x|$ and let $f(x)=F(|x|)$, where $F$ is compactly supported. For $x\neq 0$, \begin{align*} \nabla u(x)=\left(\frac{x_1}{|x|},\ldots,\frac{x_n}{|x|}\right), \qquad |\nabla u(x)|^2 =\sum_{i=1}^n \frac{x_i^2}{|x|^2} =\frac{|x|^2}{|x|^2} =1. \end{align*} The point $0$ has $\mathcal L^n$-measure zero, so the coarea formula applied to $u$ gives \begin{align*} \int_{\mathbb R^n}F(|x|)\,d\mathcal L^n(x) &=\int_{\mathbb R^n}F(|x|)|\nabla u(x)|\,d\mathcal L^n(x)\\ &=\int_{\mathbb R}\left(\int_{u^{-1}(t)}F(|x|)\,d\mathcal H^{n-1}(x)\right)dt\\ &=\int_0^\infty\left(\int_{\partial B(0,t)}F(t)\,d\mathcal H^{n-1}\right)dt\\ &=\int_0^\infty F(t)\mathcal H^{n-1}(\partial B(0,t))\,dt. \end{align*} Write $\omega_n:=\mathcal L^n(B(0,1)\subset\mathbb R^n)$. By Euclidean scaling, \begin{align*} \mathcal L^n(B(0,t))=\omega_n t^n. \end{align*} Applying the same coarea identity to the indicator of $B(0,t)$ gives \begin{align*} \omega_n t^n =\mathcal L^n(B(0,t)) =\int_0^t \mathcal H^{n-1}(\partial B(0,s))\,ds. \end{align*} Differentiating in $t>0$ yields \begin{align*} \mathcal H^{n-1}(\partial B(0,t))=n\omega_n t^{n-1}. \end{align*} Therefore \begin{align*} \int_{\mathbb R^n}F(|x|)\,d\mathcal L^n(x) =n\omega_n\int_0^\infty F(t)t^{n-1}\,dt. \end{align*} The power $t^{n-1}$ is the hypersurface-area scaling of distance spheres; a power $t^n$ would be measuring accumulated ball volume instead. [/example] The radial example shows the model behaviour: volume grows by accumulating sphere areas. On a curved manifold, naive slicing by smooth hypersurfaces breaks at the cut locus: a distance sphere may have corners, self-overlap in geodesic polar coordinates, or singular pieces where several minimizing geodesics arrive at the same point. To compare this behaviour, we need a precise level-set object that remains meaningful at cut points; this motivates naming distance spheres, with Hausdorff measure used later for their area. [definition: Distance Sphere] Let $(M,g)$ be a Riemannian manifold, let $p\in M$, and let $r>0$. The distance sphere of radius $r$ about $p$ is \begin{align*} S(p,r):=\{x\in M:d(p,x)=r\}. \end{align*} [/definition] The area of $S(p,r)$ is measured by $\mathcal H^{n-1}(S(p,r))$. Distance spheres may fail to be smooth at cut points, but the Hausdorff measure is still the right object for integration. The next statement records the comparison consequence needed later for annuli and level-set estimates. [illustration:geometric-analysis-i-distance-sphere-cut-locus] [quotetheorem:5376] [citeproof:5376] This theorem is the infinitesimal form of Bishop-Gromov comparison. It does not assert that every distance sphere is smooth, nor that the ratio is pointwise monotone at cut radii; the correct statement is an almost-everywhere level-set statement. It also gives practical estimates for annuli, because an annulus is a union of distance spheres. [example: Annular Volume Estimate] Assume $\operatorname{Ric}\ge 0$ on a complete $n$-manifold, fix $p\in M$, and write \begin{align*} V(t):=\operatorname{vol}(B(p,t)). \end{align*} By *Bishop-Gromov volume comparison*, the function $t\mapsto V(t)/t^n$ is nonincreasing on $(0,\infty)$. Hence, for $0<r<R$, \begin{align*} \frac{V(r)}{r^n}\ge \frac{V(R)}{R^n}. \end{align*} Multiplying by $r^n>0$ gives \begin{align*} V(r)\ge \frac{r^n}{R^n}V(R). \end{align*} Therefore the annulus $A(p;r,R)=B(p,R)\setminus B(p,r)$ satisfies \begin{align*} \operatorname{vol}(A(p;r,R)) &=V(R)-V(r)\\ &\le V(R)-\frac{r^n}{R^n}V(R)\\ &=\left(1-\frac{r^n}{R^n}\right)V(R)\\ &=\frac{R^n-r^n}{R^n}V(R). \end{align*} With $\omega_n=\mathcal L^n(B(0,1)\subset\mathbb R^n)$, this is equivalently \begin{align*} \operatorname{vol}(A(p;r,R)) \le \omega_n(R^n-r^n)\frac{\operatorname{vol}(B(p,R))}{\omega_n R^n}. \end{align*} The factor $(R^n-r^n)/R^n$ is the Euclidean relative shell thickness, so the curved annulus is controlled by the outer ball volume times the corresponding Euclidean fraction. [/example] ## Isoperimetric Constants and Poincare Inequalities The next problem is to control oscillation of functions using geometry. A function with two large regions of different values must pass through many level sets; if every separating hypersurface has large area, the function cannot change value cheaply. The obstruction is a neck: two large chambers connected by a thin tube have functions that are almost constant on each chamber and change only across the tube, so their gradient energy can be small. [definition: Cheeger Constant] Let $(M,g)$ be a compact Riemannian manifold. The Cheeger constant is the assignment \begin{align*} h:\{\text{compact Riemannian manifolds}\}\to[0,\infty] \end{align*} given by \begin{align*} h(M):=\inf_E \frac{\mathcal H^{n-1}(\partial E)}{\min\{\operatorname{vol}(E),\operatorname{vol}(M\setminus E)\}}, \end{align*} where the infimum is taken over smooth domains $E\subset M$ with $0<\operatorname{vol}(E)<\operatorname{vol}(M)$. [/definition] The denominator prevents a tiny region from giving a misleading separation ratio. The constant measures the least boundary area needed to cut the manifold into two substantial pieces. [example: Cheeger Scale on a Round Sphere] Let $S^n\subset\mathbb R^{n+1}$ carry the unit round metric, and take the closed northern hemisphere \begin{align*} E:=\{x\in S^n:x_{n+1}\ge 0\}. \end{align*} Its boundary is the equator \begin{align*} \partial E=\{x\in S^n:x_{n+1}=0\}\cong S^{n-1}, \end{align*} so \begin{align*} \mathcal H^{n-1}(\partial E)=\operatorname{vol}(S^{n-1}). \end{align*} Reflection across the equator sends $E$ to $S^n\setminus E$ and preserves the round volume measure, hence \begin{align*} \operatorname{vol}(E)=\operatorname{vol}(S^n\setminus E) =\frac{1}{2}\operatorname{vol}(S^n). \end{align*} Therefore this one admissible competitor gives \begin{align*} h(S^n) &\le \frac{\mathcal H^{n-1}(\partial E)} {\min\{\operatorname{vol}(E),\operatorname{vol}(S^n\setminus E)\}}\\ &= \frac{\operatorname{vol}(S^{n-1})} {\min\left\{\frac{1}{2}\operatorname{vol}(S^n),\frac{1}{2}\operatorname{vol}(S^n)\right\}}\\ &= \frac{\operatorname{vol}(S^{n-1})}{\frac{1}{2}\operatorname{vol}(S^n)}\\ &= \frac{2\operatorname{vol}(S^{n-1})}{\operatorname{vol}(S^n)}. \end{align*} The *Levy-Gromov Isoperimetric Inequality* says that spherical caps give the sharp model isoperimetric profile under the corresponding positive Ricci lower bound, so the hemisphere quotient captures the correct spherical separation scale. [/example] The sphere example shows that separation geometry has a numerical scale. To turn that scale into an estimate for functions, we need the spectral quantity that measures the least possible energy of a nonconstant mean-zero function; this motivates the first nonzero eigenvalue. Without removing constants, the Rayleigh quotient has value $0$ on every compact manifold, so it cannot detect geometry. [definition: First Nonzero Laplace Eigenvalue] Let $(M,g)$ be compact and connected. The first nonzero eigenvalue is the assignment \begin{align*} \lambda_1:\{\text{compact connected Riemannian manifolds}\}\to[0,\infty) \end{align*} given by \begin{align*} \lambda_1(M):=\inf\left\{\frac{\int_M |\nabla f|^2\,d\operatorname{vol}_g}{\int_M f^2\,d\operatorname{vol}_g}: f\in C^\infty(M),\ f\not\equiv 0,\ \int_M f\,d\operatorname{vol}_g=0\right\}. \end{align*} [/definition] The mean-zero condition removes the constant eigenfunctions. The following theorem is one of the main ways that an isoperimetric lower bound becomes a spectral lower bound. The obstruction it rules out is concentration on one side of a narrow bottleneck: if a hypersurface of tiny area separates two large regions, a test function can switch sign across that hypersurface with small gradient energy. [quotetheorem:5377] [citeproof:5377] [Cheeger inequality](/theorems/5377) is robust because it does not require smooth minimisers for the isoperimetric problem. The compactness and connectedness hypotheses are part of the spectral meaning of the statement, not decorative assumptions. If $M$ is disconnected, then a function that is constant with different values on two components has zero gradient and mean zero after rescaling, so $\lambda_1(M)=0$ and the separating geometry is instead detecting the components. If $M$ is noncompact, the bottom of the spectrum may be continuous rather than an isolated first eigenvalue, and compactly supported test functions can escape to infinity; a separate formulation using the bottom of the $L^2$ spectrum is then required. It is only a lower bound: it does not determine the spectrum, and it can be far from sharp on highly symmetric spaces such as round spheres. Once a lower bound for $h(M)$ is known from geometry, the spectral conclusion follows directly. [example: First Eigenvalue of a Round Sphere] Let $S^n$ have the unit round metric, and let $f_i:S^n\to\mathbb R$ be the restriction $f_i(x)=x_i$. The linear function $x_i$ on $\mathbb R^{n+1}$ is a homogeneous harmonic polynomial of degree $1$, so by *Spherical Harmonic Spectral Decomposition* its restriction satisfies \begin{align*} -\Delta_{S^n}f_i=1(1+n-1)f_i=nf_i. \end{align*} Also $f_i$ is mean-zero because the antipodal map $x\mapsto -x$ preserves round volume and sends $f_i$ to $-f_i$: \begin{align*} \int_{S^n} f_i\,d\operatorname{vol} =\int_{S^n} f_i(-x)\,d\operatorname{vol}(x) =-\int_{S^n} f_i(x)\,d\operatorname{vol}(x), \end{align*} hence $\int_{S^n}f_i\,d\operatorname{vol}=0$. Since $f_i\not\equiv 0$, it is an admissible test function for $\lambda_1(S^n)$, and integration by parts gives \begin{align*} \frac{\int_{S^n}|\nabla f_i|^2\,d\operatorname{vol}} {\int_{S^n}f_i^2\,d\operatorname{vol}} &= \frac{\int_{S^n}f_i(-\Delta f_i)\,d\operatorname{vol}} {\int_{S^n}f_i^2\,d\operatorname{vol}}\\ &= \frac{\int_{S^n}f_i(nf_i)\,d\operatorname{vol}} {\int_{S^n}f_i^2\,d\operatorname{vol}}\\ &=n. \end{align*} Therefore $\lambda_1(S^n)\le n$. The same spherical harmonic decomposition gives the full spectrum \begin{align*} \ell(\ell+n-1),\qquad \ell=0,1,2,\ldots . \end{align*} For $\ell=0$ the eigenvalue is $0$, corresponding to constants. For $\ell=1$ the eigenvalue is $n$. For every $\ell\ge 2$, \begin{align*} \ell(\ell+n-1)\ge 2(2+n-1)=2(n+1)>n. \end{align*} Thus there is no nonzero eigenvalue below $n$, and the first nonzero eigenvalue is \begin{align*} \lambda_1(S^n)=n. \end{align*} Cheeger inequality gives only a lower bound from separation geometry, while the spherical harmonic calculation identifies the exact spectral value. [/example] The sphere calculation compares exact spectral information with the coarser Cheeger bound. For local analysis on balls, the corresponding target is not an eigenvalue but an inequality controlling deviation from the average; this motivates the Poincare inequality. Without subtracting the average, constant functions would violate any estimate by gradient energy alone, and without a scale factor the inequality would fail under rescaling of the metric. [definition: Local Poincare Inequality] Let $(M,g)$ be a Riemannian manifold. A local $L^2$ Poincare inequality holds on balls of radius at most $R_0$ if there exists $C_P>0$ such that for every geodesic ball $B=B(x,r)$ with $0<r\le R_0$ and every $f\in C^1(\overline{B})$, \begin{align*} \int_B |f-f_B|^2\,d\operatorname{vol}_g \le C_P r^2\int_B |\nabla f|^2\,d\operatorname{vol}_g, \end{align*} where \begin{align*} f_B:=\frac{1}{\operatorname{vol}(B)}\int_B f\,d\operatorname{vol}_g. \end{align*} [/definition] The factor $r^2$ is forced by scaling: gradients have one inverse length. The comparison question is whether curvature bounds provide such constants uniformly over all small balls, and the answer uses Ricci control of geodesic segments. Without a Ricci lower bound, thin negatively curved regions can distort geodesic averaging and destroy uniform constants even when each individual ball has a Poincare inequality. [quotetheorem:5378] [citeproof:5378] This estimate is local and does not require compactness. Completeness is used to guarantee the minimizing geodesics needed in the segment argument and to place balls in the global comparison framework; without it, geodesics may leave the manifold before the averaging argument is complete. The Ricci lower bound is the geometric input that controls how geodesic families spread, and without such a bound a sequence of balls can develop increasingly thin necks or rapid volume distortion, forcing the Poincare constants to diverge. The scale restriction $r\le R_0$ is also essential when $K>0$ is replaced by a lower bound $\operatorname{Ric}\ge -(n-1)K g$: negative curvature effects accumulate over long distances, so one cannot expect a single constant independent of the radius. Thus the theorem does not give a global spectral gap on a noncompact manifold, and its constant is only uniform on the prescribed scale. It is the analytic companion to volume doubling, which controls how measures of nested balls compare. [definition: Volume Doubling] A Riemannian manifold $(M,g)$ satisfies volume doubling on scale $R_0$ with constant $C_D$ if for every $x\in M$ and every $0<r\le R_0$, \begin{align*} \operatorname{vol}(B(x,2r))\le C_D\operatorname{vol}(B(x,r)). \end{align*} [/definition] Volume doubling prevents mass from escaping too rapidly as balls expand. Combined with Poincare inequalities, it forms the standard analytic package behind heat kernel estimates, Harnack inequalities, and compactness arguments. [example: Volume Doubling Plus Poincare] Suppose $\operatorname{Ric}\ge 0$, fix $x\in M$, and write \begin{align*} V_x(t):=\operatorname{vol}(B(x,t)). \end{align*} By *Bishop-Gromov volume comparison*, the function $t\mapsto V_x(t)/t^n$ is nonincreasing on $(0,\infty)$. Since $r<2r$, this gives \begin{align*} \frac{V_x(2r)}{(2r)^n}\le \frac{V_x(r)}{r^n}. \end{align*} Multiplying both sides by $(2r)^n>0$ yields \begin{align*} V_x(2r) &\le (2r)^n\frac{V_x(r)}{r^n}\\ &=2^n r^n\frac{V_x(r)}{r^n}\\ &=2^n V_x(r). \end{align*} Thus every ball satisfies the doubling estimate \begin{align*} \operatorname{vol}(B(x,2r))\le 2^n\operatorname{vol}(B(x,r)). \end{align*} The local Poincare estimate from *Poincare Inequality from Ricci Lower Bounds*, applied with $K=0$, gives a constant $C_P=C_P(n)$ such that for every $f\in C^1(\overline{B(x,r)})$, \begin{align*} \int_{B(x,r)} |f-f_{B(x,r)}|^2\,d\operatorname{vol}_g \le C_P r^2\int_{B(x,r)}|\nabla f|^2\,d\operatorname{vol}_g. \end{align*} The [first inequality](/theorems/2897) controls how volume changes when the radius is doubled, and the second controls oscillation on the same scale; together they give the standard local analytic package with constants depending only on the dimension. [/example] ## Levy-Gromov and Cheng Comparison The next question is how sharp comparison geometry can make the analytic estimates. Cheeger inequality gives a general bridge from isoperimetry to spectrum, but stronger curvature and diameter assumptions give model-space inequalities directly. The obstruction is that a lower bound for $h(M)$ alone forgets the volume at which the worst cut occurs; sharp comparison needs the whole isoperimetric profile. [definition: Isoperimetric Profile] Let $(M,g)$ be compact with volume $V$. The isoperimetric profile is the function $I_M:(0,V)\to[0,\infty)$ defined by \begin{align*} I_M(v):=\inf\{\mathcal H^{n-1}(\partial E): E\subset M\text{ smooth},\ \operatorname{vol}(E)=v\}. \end{align*} [/definition] The profile records all possible volume cuts, while the Cheeger constant keeps only the worst ratio. It is best visualised by sweeping through domains of increasing volume and recording the least boundary area needed at each volume. [illustration:geometric-analysis-i-isoperimetric-profile] The comparison theorem below is stated in the course as a sharp global result. Positive Ricci curvature is essential here: without it, long necks and warped products can have isoperimetric profiles much smaller than the spherical model at intermediate volumes. [remark: Quoted result: Levy-Gromov Isoperimetric Inequality] Let $(M,g)$ be a closed $n$-dimensional Riemannian manifold with $\operatorname{Ric}\ge (n-1)g$. Let $S^n$ denote the unit round sphere, and set $V_M=\operatorname{vol}(M)$ and $V_S=\operatorname{vol}(S^n)$. Then for every $v\in(0,V_M)$, \begin{align*} I_M(v)\ge \frac{V_M}{V_S} I_{S^n}\left(\frac{V_S}{V_M}v\right). \end{align*} [/remark] This theorem is quoted rather than proved in this course. Its proof uses symmetrisation and deep comparison of isoperimetric minimisers; the important point here is that the round sphere is the sharp model under the Ricci lower bound $\operatorname{Ric}\ge(n-1)g$. The theorem is global and normalised by total volume, so it does not by itself give local Sobolev constants on collapsed balls. [example: Cheeger Lower Bound from Levy-Gromov] Let $E\subset M$ be a smooth domain, and write \begin{align*} v:=\operatorname{vol}(E). \end{align*} First assume $0<v\le V_M/2$. By *Levy-Gromov Isoperimetric Inequality*, \begin{align*} \mathcal H^{n-1}(\partial E) \ge \frac{V_M}{V_S}I_{S^n}\left(\frac{V_S}{V_M}v\right). \end{align*} Dividing by $v>0$ gives \begin{align*} \frac{\mathcal H^{n-1}(\partial E)}{v} &\ge \frac{1}{v}\frac{V_M}{V_S}I_{S^n}\left(\frac{V_S}{V_M}v\right)\\ &= \frac{I_{S^n}\left(\frac{V_S}{V_M}v\right)} {\frac{V_S}{V_M}v}. \end{align*} Set \begin{align*} \theta:=\frac{v}{V_M}. \end{align*} Then $0<\theta\le 1/2$, and the last inequality becomes \begin{align*} \frac{\mathcal H^{n-1}(\partial E)}{\operatorname{vol}(E)} \ge \frac{I_{S^n}(V_S\theta)}{V_S\theta}. \end{align*} Hence \begin{align*} \frac{\mathcal H^{n-1}(\partial E)}{\operatorname{vol}(E)} \ge c_n, \qquad c_n:=\inf_{0<\theta\le 1/2} \frac{I_{S^n}(V_S\theta)}{V_S\theta}. \end{align*} If instead $\operatorname{vol}(E)>V_M/2$, apply the same argument to $M\setminus E$. Since \begin{align*} \partial(M\setminus E)=\partial E \end{align*} and \begin{align*} \operatorname{vol}(M\setminus E)<V_M/2, \end{align*} we get \begin{align*} \frac{\mathcal H^{n-1}(\partial E)} {\operatorname{vol}(M\setminus E)} \ge c_n. \end{align*} Combining the two cases, \begin{align*} \frac{\mathcal H^{n-1}(\partial E)} {\min\{\operatorname{vol}(E),\operatorname{vol}(M\setminus E)\}} \ge c_n. \end{align*} Taking the infimum over all admissible smooth domains $E$ gives \begin{align*} h(M)\ge c_n. \end{align*} Thus Levy-Gromov turns the full spherical isoperimetric profile into a uniform Cheeger lower bound; the constant depends only on the dimension through the unit round sphere. [/example] The previous example extracts Cheeger estimates from an isoperimetric profile. A parallel spectral comparison asks what happens when the test functions are chosen from model balls, which motivates Cheng's eigenvalue theorem. The obstruction is the cut locus: distance functions are not smooth everywhere, so any clean comparison statement must either avoid pathological radii or interpret the ball weakly as a Dirichlet domain. [remark: Quoted result: Cheng Eigenvalue Comparison] Let $(M,g)$ be complete and satisfy $\operatorname{Ric}\ge (n-1)k g$. Let $B(p,R)$ be a relatively compact geodesic ball, and assume either that $R$ is chosen so that $\partial B(p,R)$ is smooth, or that $\lambda_1(B(p,R))$ is interpreted as the bottom of the Dirichlet spectrum on the metric ball via the Rayleigh quotient over $C_c^\infty(B(p,R))$. Then \begin{align*} \lambda_1(B(p,R))\le \lambda_1(B_k(R)), \end{align*} where $B_k(R)$ is the radius-$R$ ball in the simply connected $n$-dimensional space form of sectional curvature $k$, with $R$ below the model diameter when $k>0$. [/remark] This theorem is stated in the course as a comparison result. The proof uses radial test functions from the model ball and Laplacian comparison for the distance function; equality cases require additional rigidity analysis. The inequality is one-sided: Ricci lower bounds give an upper bound for Dirichlet eigenvalues of balls, not a lower bound, because the [comparison test](/theorems/173) function may be transplanted to $M$. [example: Estimating Eigenvalues on Spherical Balls] For the unit round sphere $S^n$, the sectional curvature is $1$, so the model parameter in *Cheng Eigenvalue Comparison* is $k=1$. Let $B_M(p,R)\subset M$ be a relatively compact geodesic ball in an $n$-manifold satisfying $\operatorname{Ric}\ge(n-1)g$, and let $B_{S^n}(q,R)$ be the radius-$R$ ball in the unit round sphere, with $R<\pi$. Cheng comparison gives \begin{align*} \lambda_1(B_M(p,R))\le \lambda_1(B_{S^n}(q,R)). \end{align*} When $M=S^n$ itself, the comparison ball and the given ball are isometric: for any $p,q\in S^n$, a round isometry $\Phi:S^n\to S^n$ sends $p$ to $q$, hence \begin{align*} \Phi(B_{S^n}(p,R))=B_{S^n}(q,R). \end{align*} If $\varphi\in C_c^\infty(B_{S^n}(q,R))$ and $\psi=\varphi\circ\Phi$, then $\Phi$ preserves the metric, the gradient norm, and the volume form, so \begin{align*} \int_{B_{S^n}(p,R)}|\nabla\psi|^2\,d\operatorname{vol} &=\int_{B_{S^n}(q,R)}|\nabla\varphi|^2\,d\operatorname{vol},\\ \int_{B_{S^n}(p,R)}\psi^2\,d\operatorname{vol} &=\int_{B_{S^n}(q,R)}\varphi^2\,d\operatorname{vol}. \end{align*} Therefore the Rayleigh quotients agree: \begin{align*} \frac{\int_{B_{S^n}(p,R)}|\nabla\psi|^2\,d\operatorname{vol}} {\int_{B_{S^n}(p,R)}\psi^2\,d\operatorname{vol}} = \frac{\int_{B_{S^n}(q,R)}|\nabla\varphi|^2\,d\operatorname{vol}} {\int_{B_{S^n}(q,R)}\varphi^2\,d\operatorname{vol}}. \end{align*} Taking the infimum over admissible test functions gives \begin{align*} \lambda_1(B_{S^n}(p,R))=\lambda_1(B_{S^n}(q,R)). \end{align*} Thus equality is attained in Cheng comparison for the model sphere itself, with the first Dirichlet eigenfunction given by the corresponding radial model eigenfunction. This model value is the benchmark for how positive Ricci curvature and radius constrain low-frequency oscillation on geodesic balls. [/example] ## Sobolev Consequences Under Noncollapse The final problem is to pass from geometric comparison to function-space estimates that are stable under limits. Ricci lower bounds alone control relative volume growth, but Sobolev inequalities also need a lower bound preventing unit balls from collapsing to very small volume. Flat tori with one circle factor shrinking to length $\varepsilon$ show the obstruction: curvature can stay nonnegative while the local volume scale degenerates. [definition: Volume Noncollapse] Let $(M,g)$ be a complete Riemannian manifold. On a scale $R_0>0$, the manifold is volume noncollapsed with constant $v_0>0$ if \begin{align*} \operatorname{vol}(B(x,R_0))\ge v_0 \end{align*} for every $x\in M$. [/definition] Noncollapse supplies a lower measure scale, while Ricci comparison supplies upper control on relative growth. The analytic question is whether these two geometric inputs force a uniform Sobolev inequality on small balls. The standard local estimate has constants depending only on the comparison data. [remark: Quoted result: Local Sobolev Inequality under Ricci Lower Bound and Volume Noncollapsing] Let $(M,g)$ be a complete $n$-dimensional Riemannian manifold with $n \ge 3$. Let $K \ge 0$, $R_0 > 0$, and $v_0 > 0$. Suppose that \begin{align*} \operatorname{Ric}_g \ge -(n-1)K g \end{align*} on $M$, and that for every $x \in M$, \begin{align*} \operatorname{vol}_g(B_g(x,R_0)) \ge v_0. \end{align*} Then there exists a constant $C = C(n,K,R_0,v_0) > 0$ such that for every $x \in M$, every $r \in (0,R_0]$, and every $f \in C_c^1(B_g(x,r))$, \begin{align*} \left(\int_{B_g(x,r)} |f|^{\frac{2n}{n-2}} \, d\operatorname{vol}_g\right)^{\frac{n-2}{n}} \le C \int_{B_g(x,r)} \left(|\nabla f|_g^2 + r^{-2} f^2\right) \, d\operatorname{vol}_g. \end{align*} [/remark] The restriction $n\ge 3$ is tied to the displayed critical exponent $2n/(n-2)$; in dimension $2$ the corresponding endpoint is not a finite-power Sobolev embedding and is replaced by different estimates. Completeness again supports the comparison and covering arguments on all balls under consideration. Compact support keeps the estimate local and avoids boundary trace terms on rough metric balls, while the extra $r^{-2}f^2$ term makes the right-hand side insensitive to constants and to boundary behaviour. The theorem does not follow from Ricci lower bounds alone; collapse can make the Sobolev constant diverge even when volume doubling survives in a relative form, as shrinking flat tori show. If one imposes zero boundary values or works with a global compact manifold, this lower-order term can often be absorbed into the gradient term through Poincare. [example: Analytic Package on Noncollapsed Ricci Balls] Let $\operatorname{Ric}\ge -(n-1)K g$ and assume \begin{align*} \operatorname{vol}(B(x,1))\ge v_0 \end{align*} for every $x\in M$. Fix $x\in M$ and $0<r\le 1$, and write \begin{align*} V_x(t):=\operatorname{vol}(B(x,t)). \end{align*} Let $V_K(t)$ denote the volume of the radius-$t$ ball in the simply connected space form of sectional curvature $-K$. By *Bishop-Gromov volume comparison*, the ratio $t\mapsto V_x(t)/V_K(t)$ is nonincreasing, so for $0<r<2r$, \begin{align*} \frac{V_x(2r)}{V_K(2r)} \le \frac{V_x(r)}{V_K(r)}. \end{align*} Multiplying by $V_K(2r)>0$ gives \begin{align*} V_x(2r) &\le V_K(2r)\frac{V_x(r)}{V_K(r)}\\ &=\frac{V_K(2r)}{V_K(r)}V_x(r). \end{align*} Since $0<r\le 1$, the model ratio is bounded on this scale by \begin{align*} C_D(n,K):=\sup_{0<s\le 1}\frac{V_K(2s)}{V_K(s)}<\infty, \end{align*} and hence \begin{align*} \operatorname{vol}(B(x,2r))\le C_D(n,K)\operatorname{vol}(B(x,r)). \end{align*} The same Ricci lower bound gives the local Poincare estimate from *Poincare Inequality from Ricci Lower Bounds*: there is $C_P=C_P(n,K,1)$ such that every $f\in C^1(\overline{B(x,r)})$ satisfies \begin{align*} \int_{B(x,r)} |f-f_{B(x,r)}|^2\,d\operatorname{vol}_g \le C_P r^2\int_{B(x,r)}|\nabla f|^2\,d\operatorname{vol}_g. \end{align*} The noncollapse assumption supplies the missing lower volume scale, so *Local Sobolev Inequality from Ricci Lower Bound and Noncollapse* gives a constant $C_S=C_S(n,K,1,v_0)$ such that every $f\in C_c^1(B(x,r))$ satisfies \begin{align*} \left(\int_{B(x,r)} |f|^{\frac{2n}{n-2}}\,d\operatorname{vol}_g\right)^{\frac{n-2}{n}} \le C_S\int_{B(x,r)}\left(|\nabla f|^2+r^{-2}f^2\right)d\operatorname{vol}_g. \end{align*} Thus each ball of radius at most $1$ has doubling, Poincare, and Sobolev constants depending only on $n,K,$ and $v_0$; these uniform constants are the analytic input used for bounded-energy compactness and Moser iteration. [/example] This chapter completes the passage from comparison geometry to analysis. Coarea turns functions into level sets, isoperimetric inequalities turn level sets into lower bounds, and Poincare-Sobolev estimates turn those lower bounds into control of oscillation and integrability. The remaining comparison tools in the course use this package as a background language for compactness and regularity. Volume comparison has shown that lower Ricci curvature controls how sets expand and how measures distort. The next chapter uses those same estimates, together with the earlier distance-function bounds, to move from quantitative control to rigidity in the nonnegative Ricci setting. # 9. Nonnegative Ricci Curvature and Splitting These notes continue the course's comparison-geometry thread by moving from estimates to rigidity. The prerequisites are the Hopf--Rinow theorem and geodesic completeness from Chapter 1, Laplacian comparison from Chapter 4, Bishop-Gromov volume comparison from Chapter 7, and the Bochner formula. The chapter's main goal is to explain why nonnegative Ricci curvature, together with the existence of a globally minimizing geodesic in both directions, forces an isometric product splitting. Analytically, the proof converts distance comparison into weak subharmonicity of Busemann functions, then uses equality and Bochner's formula to produce a parallel vector field. ## Busemann Functions and Lines in Complete Manifolds The basic question is how to encode the geometry seen from infinity. On a noncompact complete manifold, ordinary distance functions depend on a chosen basepoint and develop cut loci. A ray or line gives a preferred direction to infinity, and the associated Busemann function records signed escape rate along that direction. [definition: Ray] Let $(M,g)$ be a complete Riemannian manifold. A unit-speed geodesic $\rho:[0,\infty)\to M$ is a ray if \begin{align*} d(\rho(s),\rho(t))=|s-t| \end{align*} for all $s,t\ge 0$. [/definition] A ray is a geodesic that remains globally minimizing between any two of its points. To obtain rigidity rather than just an escape direction, we need the same minimizing condition in both time directions. [definition: Line] Let $(M,g)$ be a complete Riemannian manifold. A unit-speed geodesic $\gamma:\mathbb R\to M$ is a line if \begin{align*} d(\gamma(s),\gamma(t))=|s-t| \end{align*} for all $s,t\in\mathbb R$. [/definition] Lines arise in product manifolds and in universal covers of manifolds with suitable topology. They should not be confused with long minimizing segments: a line is an infinite object with a global minimizing property across all pairs of times, and the next example supplies the model to keep in mind before defining the function attached to an end. [example: Product Lines] Let $(N,h)$ be complete and put $M=N\times\mathbb R$ with product metric $g=h+dt^2$. Fix $p\in N$ and define $\gamma:\mathbb R\to M$ by $\gamma(s)=(p,s)$. Its velocity is $\dot\gamma(s)=(0,1)$, so \begin{align*} |\dot\gamma(s)|_g^2 &= |0|_h^2 + 1^2\\ &=1. \end{align*} The product connection gives $\nabla^M_{\dot\gamma}\dot\gamma=(\nabla^N_0 0,0)=0$, hence $\gamma$ is a unit-speed geodesic. It remains to check that it is globally minimizing. Let $\alpha(u)=(\eta(u),\tau(u))$, $u\in[a,b]$, be any piecewise smooth curve from $(p,s)$ to $(p,t)$. Its length satisfies \begin{align*} L(\alpha) &=\int_a^b \sqrt{|\eta'(u)|_h^2+|\tau'(u)|^2}\,du\\ &\ge \int_a^b |\tau'(u)|\,du\\ &\ge \left|\int_a^b \tau'(u)\,du\right|\\ &=|\tau(b)-\tau(a)|\\ &=|t-s|. \end{align*} On the other hand, the vertical curve $u\mapsto(p,(1-u)s+ut)$ for $u\in[0,1]$ has length \begin{align*} \int_0^1 |t-s|\,du=|t-s|. \end{align*} Therefore \begin{align*} d_M((p,s),(p,t))=|s-t| \end{align*} for all $s,t\in\mathbb R$, so $\gamma$ is a line. This is the model situation for the splitting theorem: the $\mathbb R$-coordinate is already a globally minimizing direction. [/example] This product example shows that a line should carry a coordinate measuring signed displacement, but a general manifold has no given product coordinate. This motivates defining such a coordinate as a limiting renormalized distance to points escaping along a ray. [definition: Busemann Function] Let $(M,g)$ be a complete Riemannian manifold and let $\rho:[0,\infty)\to M$ be a ray. The Busemann function associated to $\rho$ is the map $b_\rho:M\to\mathbb R$ defined by \begin{align*} b_\rho &: M\to\mathbb R, & x&\mapsto \lim_{t\to\infty}\bigl(t-d(x,\rho(t))\bigr). \end{align*} [/definition] The limit exists because the triangle inequality gives monotonicity in $t$ and bounds the expression above on compact sets. Before using curvature, we need the metric properties that make $b_\rho$ a legitimate replacement for a coordinate function. [quotetheorem:5379] [citeproof:5379] Completeness is used here to ensure that the ray is defined for all forward time and that distance functions behave globally; on an incomplete manifold, a unit-speed geodesic may stop before it defines an end at infinity. The ray hypothesis is also essential: if a geodesic segment ceases to minimize after a cut point, the identity $b_\rho(\rho(s))=s$ has no analogue. The theorem supplies only coarse metric control, so the next step needs a line in order to compare two opposite ends and turn an inequality into equality. The preceding theorem gives control for one end of a ray, but splitting requires comparing two ends at once. When a line is available, the two associated functions satisfy a global inequality whose equality case will become the analytic core of the theorem. [definition: Opposite Busemann Functions of a Line] Let $(M,g)$ be complete and let $\gamma:\mathbb R\to M$ be a line. Define rays $\gamma_+(t)=\gamma(t)$ and $\gamma_-(t)=\gamma(-t)$ for $t\ge 0$. The associated Busemann functions are the maps $b_+,b_-:M\to\mathbb R$ defined by \begin{align*} b_+ &: M\to\mathbb R, & x&\mapsto \lim_{t\to\infty}\bigl(t-d(x,\gamma(t))\bigr),\\ b_- &: M\to\mathbb R, & x&\mapsto \lim_{t\to\infty}\bigl(t-d(x,\gamma(-t))\bigr). \end{align*} [/definition] For any $x\in M$, the inequality $d(x,\gamma(t))+d(x,\gamma(-t))\ge 2t$ gives $b_+(x)+b_-(x)\le 0$. The next example shows the equality pattern in a product, which is the pattern the comparison argument will force under $\operatorname{Ric}\ge 0$. [example: Busemann Functions on a Cylinder] Take $M=S^1\times\mathbb R$ with product metric $d\theta^2+dt^2$, where $\theta$ is the circle coordinate and $t$ is the real coordinate. Fix $\theta_0\in S^1$ and use the vertical line \begin{align*} \gamma(s)=(\theta_0,s) \end{align*} as the reference line. For a point $(\theta,t)\in S^1\times\mathbb R$, let $a=d_{S^1}(\theta,\theta_0)$. For $T>t$, the product distance to the forward point $\gamma(T)=(\theta_0,T)$ is \begin{align*} d\bigl((\theta,t),\gamma(T)\bigr) &=\sqrt{a^2+(T-t)^2}. \end{align*} Hence \begin{align*} T-d\bigl((\theta,t),\gamma(T)\bigr) &=T-\sqrt{a^2+(T-t)^2}\\ &=t+\bigl((T-t)-\sqrt{a^2+(T-t)^2}\bigr)\\ &=t+\frac{(T-t)^2-\bigl(a^2+(T-t)^2\bigr)} {(T-t)+\sqrt{a^2+(T-t)^2}}\\ &=t-\frac{a^2}{(T-t)+\sqrt{a^2+(T-t)^2}}. \end{align*} The denominator tends to $+\infty$ as $T\to\infty$, so $b_+(\theta,t)=t$. Similarly, \begin{align*} d\bigl((\theta,t),\gamma(-T)\bigr) &=\sqrt{a^2+(-T-t)^2}\\ &=\sqrt{a^2+(T+t)^2}, \end{align*} and therefore \begin{align*} T-d\bigl((\theta,t),\gamma(-T)\bigr) &=T-\sqrt{a^2+(T+t)^2}\\ &=-t+\bigl((T+t)-\sqrt{a^2+(T+t)^2}\bigr)\\ &=-t+\frac{(T+t)^2-\bigl(a^2+(T+t)^2\bigr)} {(T+t)+\sqrt{a^2+(T+t)^2}}\\ &=-t-\frac{a^2}{(T+t)+\sqrt{a^2+(T+t)^2}}. \end{align*} Letting $T\to\infty$ gives $b_-(\theta,t)=-t$. Thus $b_++b_-=0$, and the level sets of $b_+$ are exactly the compact circles $S^1\times\{t\}$. This shows that the splitting direction can be noncompact while the transverse cross-sections remain compact. [/example] The cylinder calculation exhibits equality, but so far equality is only a feature of a known product. To prove a product from curvature, we next need a weak convexity principle for Busemann functions. ## Convexity and Subharmonicity Behind Splitting The analytic problem is to pass from distance comparison to an equality statement for a nonsmooth limiting function. Distance functions are smooth away from cut loci, but Busemann functions are generally only Lipschitz. The correct statements are therefore distributional or barrier inequalities, and they are strong enough to combine with maximum principles. [definition: Distributional Subharmonicity] Let $(M,g)$ be a Riemannian manifold and let $u:M\to\mathbb R$ be a function in $L^1_{\mathrm{loc}}(M)$. The function $u$ is subharmonic in the distributional sense if \begin{align*} \int_M u\,\Delta \phi\,d\operatorname{vol}_g \ge 0 \end{align*} for every nonnegative $\phi\in C_c^\infty(M)$. [/definition] This convention corresponds to $\Delta u\ge 0$ distributionally, so it is the weak form needed for maximum principles. This motivates proving that Busemann functions inherit subharmonicity from Laplacian comparison when $\operatorname{Ric}\ge 0$. [quotetheorem:5380] [citeproof:5380] The curvature hypothesis is the source of the sign: if Ricci curvature is allowed to be negative, hyperbolic space gives Busemann functions with different Laplacian behaviour and no Cheeger-Gromoll splitting conclusion. Completeness again matters because comparison is applied to distance functions from points escaping to infinity along a ray. This subharmonicity is only a one-sided statement for each end separately; a line adds the triangle-inequality bound $b_++b_-\le 0$, which is what lets a maximum principle enter the proof. This subharmonicity gives an inequality for each end separately, but a line also supplies the triangle-inequality bound $b_++b_-\le 0$. This motivates applying the maximum principle to the sum, because equality at points of the line can then propagate across the manifold. [quotetheorem:5381] [citeproof:5381] The line hypothesis is indispensable here: a single ray gives one subharmonic Busemann function but no opposite function whose sum is globally bounded above. The Ricci hypothesis fixes the sign of the comparison inequality; in hyperbolic space, a geodesic line has opposite Busemann functions whose sum is strictly negative away from the line, so equality propagation fails. Connectedness is also being used at the last step: on a disjoint union of a flat cylinder containing the line and a second complete nonnegative-Ricci component, the same argument says nothing about the second component. The theorem does not yet say that either Busemann function is smooth, that its level sets are hypersurfaces, or that $M$ has a product metric; it only converts the metric inequality for the two ends into an equality. The equality just proved makes each Busemann function both subharmonic and superharmonic, so the next task is to extract differential geometry from this weak harmonicity, and the Bochner formula is the tool that turns harmonicity into parallelism. [quotetheorem:5382] [citeproof:5382] This is the analytic heart of the theorem, and it is also where nonnegative Ricci curvature is used in its sharp form. If $\operatorname{Ric}$ had negative directions, the Bochner identity would no longer make $|\nabla b|^2$ subharmonic, so the maximum-principle argument could fail. The line hypothesis is just as restrictive: a one-ended complete surface may have rays and Busemann functions, but without an opposite minimizing ray there is no equality $b_+ + b_-=0$ from which harmonicity follows. The conclusion is much stronger than harmonicity alone: on the flat plane, $u(x,y)=x^2-y^2$ is smooth and harmonic but has nonzero Hessian, so harmonicity by itself does not produce a parallel gradient. Here the equality case produces a parallel unit gradient field whose flow will become the product coordinate. [example: Why Infinite Diameter Is Not Enough] A ray already forces infinite diameter: if $\rho:[0,\infty)\to M$ is a ray, then for every $R>0$, \begin{align*} d(\rho(0),\rho(R))=R, \end{align*} so $\operatorname{diam}(M)\ge R$ for every $R$ and hence $\operatorname{diam}(M)=\infty$. This is only a one-sided minimizing condition. A line requires much more. It is a single geodesic $\gamma:\mathbb R\to M$ satisfying \begin{align*} d(\gamma(s),\gamma(t))=|s-t| \end{align*} for all $s,t\in\mathbb R$. Thus, for every $T>0$, \begin{align*} d(\gamma(-T),\gamma(T))=2T. \end{align*} A one-ended paraboloid-type surface may contain rays escaping to infinity, so it has infinite diameter, but its escaping directions all go out the same end; two long minimizing segments going far out need not fit together into one bi-infinite geodesic that remains minimizing between $\gamma(-T)$ and $\gamma(T)$ for every $T$. This distinction is exactly why the splitting argument assumes a line. A single ray gives one Busemann function, while a line gives two opposite Busemann functions $b_+$ and $b_-$. The triangle inequality then gives the global bound \begin{align*} d(x,\gamma(T))+d(x,\gamma(-T))\ge d(\gamma(T),\gamma(-T))=2T, \end{align*} so \begin{align*} \bigl(T-d(x,\gamma(T))\bigr)+\bigl(T-d(x,\gamma(-T))\bigr)\le 0, \end{align*} and passing to the limit gives $b_+(x)+b_-(x)\le 0$. Infinite diameter alone supplies far-apart points, but it does not supply this two-ended inequality along one globally minimizing geodesic. [/example] [illustration:geometric-analysis-i-busemann-product-coordinates] The example separates the coarse condition of being large from the rigid condition of containing a line. With the analytic equality statement established, we can now turn the parallel gradient field into a global product. ## The Cheeger-Gromoll Splitting Theorem The geometric question is now what a parallel unit vector field does to the global shape of the manifold. Locally, its orthogonal distribution is preserved by parallel transport and its integral curves are geodesics. Completeness turns this local product into a global isometry. [quotetheorem:2767] [citeproof:2767] The theorem is a rigidity statement for Ricci curvature. Nonnegative Ricci curvature permits many noncompact geometries, such as paraboloids or products with curved compact factors, but the presence of one line removes all twisting and warping in that direction. Completeness is needed for the gradient flow to exist for every $t\in\mathbb R$; without it, the local product determined by $\nabla b$ may stop before reaching all levels. In practice, the theorem is applied by first identifying a line, often from a product factor, a limiting argument in a universal cover, or a deck transformation acting by translation. [example: Flat Cylinders] Take $M=S^1\times\mathbb R$ with product metric $d\theta^2+dt^2$. Since $S^1$ is one-dimensional, its Ricci tensor is zero, and the product Ricci tensor is the sum of the Ricci tensors on the two factors, so \begin{align*} \operatorname{Ric}_{S^1\times\mathbb R} = \operatorname{Ric}_{S^1}+\operatorname{Ric}_{\mathbb R} = 0+0 = 0. \end{align*} Fix $\theta_0\in S^1$ and define $\gamma:\mathbb R\to S^1\times\mathbb R$ by $\gamma(s)=(\theta_0,s)$. Its velocity is $\dot\gamma(s)=(0,1)$, hence \begin{align*} |\dot\gamma(s)|^2 = |0|_{S^1}^2+|1|_{\mathbb R}^2 = 0+1 = 1. \end{align*} The product connection gives \begin{align*} \nabla_{\dot\gamma}\dot\gamma = (\nabla^{S^1}_{0}0,\nabla^{\mathbb R}_{1}1) = (0,0), \end{align*} so $\gamma$ is a unit-speed geodesic. For $s,t\in\mathbb R$, any piecewise smooth curve $\alpha(u)=(\eta(u),\tau(u))$ from $(\theta_0,s)$ to $(\theta_0,t)$ satisfies \begin{align*} L(\alpha) &=\int_a^b \sqrt{|\eta'(u)|_{S^1}^2+|\tau'(u)|^2}\,du\\ &\ge \int_a^b |\tau'(u)|\,du\\ &\ge \left|\int_a^b \tau'(u)\,du\right|\\ &=|\tau(b)-\tau(a)|\\ &=|t-s|. \end{align*} The vertical curve $u\mapsto(\theta_0,(1-u)s+ut)$ for $u\in[0,1]$ has length \begin{align*} \int_0^1 |t-s|\,du=|t-s|, \end{align*} so \begin{align*} d(\gamma(s),\gamma(t))=|s-t| \end{align*} for all $s,t\in\mathbb R$. Thus $\gamma$ is a line, and the *Cheeger-Gromoll Splitting Theorem* recovers the cylinder as the product $S^1\times\mathbb R$ with $N=S^1$. By contrast, a complete one-ended nonnegatively curved surface may have $\operatorname{Ric}\ge 0$ and rays, but without a bi-infinite globally minimizing geodesic the theorem has no line to split off. [/example] The cylinder has one Euclidean direction, while higher-rank examples may have several independent directions. Applying the same theorem to the remaining factor separates all Euclidean directions one at a time. [quotetheorem:5383] [citeproof:5383] The word independent matters: in $\mathbb R^2$, choosing the horizontal line twice does not produce two Euclidean factors, since the second choice lies in the span already extracted. Maximality matters separately: if we stop after one split of $\mathbb R^2\cong \mathbb R\times\mathbb R$, the residual factor is still a line, so the statement "the residual factor contains no line" would be false. The residual factor may still be noncompact, but by construction it contains no line, so the splitting theorem has no further direction to extract. A standard model to keep in mind is a complete one-ended paraboloid-type surface with nonnegative sectional curvature: it is noncompact and has rays, but it has no line because geodesics going through the end do not minimize between all opposite times. This form is especially useful on universal covers, where deck transformations can generate axes or quasi-axes; once a genuine line is found, the splitting separates the corresponding translation direction and reduces the remaining geometry to a lower-dimensional factor. This form is especially useful on universal covers. A deck transformation that translates along a line interacts with the splitting, so curvature begins to restrict the possible fundamental group. ## Structure Consequences for Nonnegative Ricci Curvature The final question of the chapter is topological: how much can nonnegative Ricci curvature say about the fundamental group and large-scale geometry? Splitting is the rigidity tool, while volume comparison supplies growth bounds. Together they explain why groups arising from nonnegative Ricci curvature resemble Euclidean lattices at large scale. [remark: Quoted result: Milnor Growth Theorem] Let $(M,g)$ be a compact Riemannian manifold with $\operatorname{Ric}\ge 0$. Then its fundamental group $\pi_1(M)$ has polynomial growth of degree at most $\dim M$ with respect to any finite generating set. [/remark] The standard argument combines the action of $\pi_1(M)$ by deck transformations on the universal cover $\widetilde M$ with Bishop-Gromov volume comparison, comparing word metric balls in $\pi_1(M)$ to metric balls in $\widetilde M$. A finite generating set gives a fundamental-domain packing estimate in $\widetilde M$, and Bishop-Gromov bounds the volume growth of the ambient balls by a degree-$\dim M$ polynomial. The same deck-transformation viewpoint also explains why splitting enters fundamental-group questions. If an infinite-order deck transformation $\tau$ has a displacement-minimizing axis, then the curve obtained by joining successive points $\tau^j(\tilde p)$ in the universal cover produces a bi-infinite minimizing geodesic. More generally, one often first obtains long minimizing segments from the orbit of $\tilde p$ under powers of $\tau$, then passes to a limiting line using completeness and compactness of directions. Once this line exists, Cheeger-Gromoll splits off the axis direction, and the action of $\tau$ becomes a translation on the resulting $\mathbb R$ factor, possibly combined with an isometry of the residual factor. [remark: Relation to Gromov's Theorem] Gromov's theorem says that every finitely generated group of polynomial growth is virtually nilpotent. Milnor's theorem therefore implies that the fundamental group of a compact manifold with $\operatorname{Ric}\ge 0$ is virtually nilpotent. The splitting theorem gives stronger information in cases where an infinite-order deck transformation produces a line in the universal cover. [/remark] Milnor's theorem gives a growth restriction, but equality-type geometry asks for more: what happens when the universal cover has as many Euclidean directions as its dimension allows? In that case the splitting factor has no room left for curved directions. [quotetheorem:2768] [citeproof:2768] The theorem is an iterated equality case of splitting: every independent line gives one Euclidean factor, and the residual factor is precisely the part where no line remains. The possibility $q=0$ is important, because a complete noncompact manifold with nonnegative Ricci curvature need not contain any line. The integer $q$ is also bounded by dimension, and when $q=\dim M$ there is no residual curved factor left, so the whole manifold is Euclidean. Products such as $S^2\times\mathbb R^q$ show why the theorem records the [maximal Euclidean factor](/theorems/2768) rather than claiming that all nonnegative-Ricci manifolds are flat. This also explains the equality mechanism behind splitting without quoting the same theorem again. If a smooth function has parallel unit gradient on a complete connected manifold, its gradient flow supplies the $\mathbb R$ coordinate and the level sets supply the transverse factor. A harmonic function whose gradient is not parallel, such as $x^2-y^2$ on $\mathbb R^2$, does not determine a product coordinate; the missing ingredient is the vanishing Hessian, which keeps the flow from shearing the metric. In later geometric analysis, the same pattern appears repeatedly: a curvature inequality gives a differential inequality, equality makes a quantity parallel, and parallel objects force product structure. The final example records the range of possible equality cases. [example: Product Manifolds with Nonnegative Ricci Curvature] Let $(N,h)$ be complete with $\operatorname{Ric}_N\ge 0$, and set \begin{align*} M=N\times\mathbb R^k, \qquad g=h+g_{\mathrm{Euc}}. \end{align*} For tangent vectors $(X,\xi),(Y,\eta)\in T_pN\oplus T_x\mathbb R^k$, the product metric is \begin{align*} g_{(p,x)}((X,\xi),(Y,\eta)) = h_p(X,Y)+\xi\cdot\eta. \end{align*} The Levi-Civita connection splits by factors, so the curvature tensor has no mixed terms, and \begin{align*} \operatorname{Ric}_M((X,\xi),(X,\xi)) &= \operatorname{Ric}_N(X,X)+\operatorname{Ric}_{\mathbb R^k}(\xi,\xi)\\ &= \operatorname{Ric}_N(X,X)+0\\ &\ge 0. \end{align*} Thus $M$ has nonnegative Ricci curvature. Fix $p_0\in N$, $x_0\in\mathbb R^k$, and a unit vector $v\in\mathbb R^k$. Define \begin{align*} \gamma(s)=(p_0,x_0+sv). \end{align*} Then \begin{align*} \dot\gamma(s)=(0,v), \qquad |\dot\gamma(s)|_g^2 = |0|_h^2+|v|^2 = 1, \end{align*} and the product connection gives \begin{align*} \nabla^M_{\dot\gamma}\dot\gamma = (\nabla^N_0 0,\nabla^{\mathbb R^k}_v v) = (0,0), \end{align*} since $v$ is a constant Euclidean vector field. Hence $\gamma$ is a unit-speed geodesic. For $s,t\in\mathbb R$, the product distance satisfies \begin{align*} d_M\bigl((p_0,x_0+sv),(p_0,x_0+tv)\bigr) &= \sqrt{d_N(p_0,p_0)^2+|x_0+sv-(x_0+tv)|^2}\\ &= \sqrt{0^2+|(s-t)v|^2}\\ &= \sqrt{(s-t)^2|v|^2}\\ &= |s-t|. \end{align*} Therefore every straight Euclidean-factor geodesic of this form is a line in $M$. Now define \begin{align*} b_v:N\times\mathbb R^k\to\mathbb R, \qquad b_v(p,x)=v\cdot x. \end{align*} For any tangent vector $(X,\xi)$, \begin{align*} db_v(X,\xi)=v\cdot \xi. \end{align*} The vector field $(0,v)$ satisfies \begin{align*} g((0,v),(X,\xi)) = h(0,X)+v\cdot \xi = v\cdot \xi = db_v(X,\xi), \end{align*} so \begin{align*} \nabla b_v=(0,v). \end{align*} Consequently \begin{align*} |\nabla b_v|_g^2 = |(0,v)|_g^2 = |0|_h^2+|v|^2 = 1. \end{align*} Since $(0,v)$ is parallel in the product connection, \begin{align*} \nabla^M_{(X,\xi)}\nabla b_v = \nabla^M_{(X,\xi)}(0,v) = (\nabla^N_X0,\nabla^{\mathbb R^k}_\xi v) = (0,0), \end{align*} and therefore, for all tangent vectors $(X,\xi),(Y,\eta)$, \begin{align*} \operatorname{Hess} b_v((X,\xi),(Y,\eta)) = g(\nabla^M_{(X,\xi)}\nabla b_v,(Y,\eta)) = 0. \end{align*} Thus $|\nabla b_v|=1$ and $\operatorname{Hess} b_v=0$. The equality case is therefore not the same as flatness of the whole manifold: the Euclidean factor supplies the parallel Busemann coordinates, while the factor $N$ may still have nonzero curvature subject to $\operatorname{Ric}_N\ge 0$. [/example] The chapter's conclusion is that nonnegative Ricci curvature is flexible until a line appears. A line supplies two Busemann functions, comparison makes them harmonic, Bochner's formula makes their gradient parallel, and the parallel field turns the manifold into a product. This chain is the prototype for many rigidity arguments in comparison geometry. The splitting theorem shows that a line can force an entire manifold to decompose, so lower Ricci curvature becomes a rigidity statement rather than just an estimate. The next chapter asks for an even stronger global picture under nonnegative sectional curvature, where the soul theorem describes the topology of complete noncompact manifolds. # 10. Nonnegative Sectional Curvature and Soul Phenomena This chapter studies the strongest global structure theorem in the noncompact part of comparison geometry. Earlier chapters used lower curvature bounds to control triangles in Chapter 6, volume in Chapter 7, and compactness in Chapter 11; here the special assumption $\operatorname{sec} \ge 0$ makes Busemann-type functions convex enough to find a compact core. The guiding question is how much topology can remain at infinity on a complete open manifold whose sectional curvature never becomes negative. The answer is the soul picture: the manifold retracts onto a compact totally geodesic submanifold and is diffeomorphic to its normal bundle. ## Totally Convex Exhaustions From Distance-Like Functions The first problem is to replace compactness by protected compact approximations to the manifold. On a complete noncompact manifold, distance from a point is not smooth across the cut locus and need not be convex in general. Nonnegative sectional curvature supplies a different source of convexity: limits of distance functions to points escaping to infinity. [definition: Ray] Let $(M,g)$ be a complete Riemannian manifold. A ray is a unit-speed geodesic $\rho:[0,\infty) \to M$ such that $d(\rho(s),\rho(t))=|s-t|$ for all $s,t \ge 0$. [/definition] Rays encode directions to infinity, but the ray itself is only a curve and cannot directly define compact approximations to the whole manifold. To use a ray in an exhaustion argument, one needs a scalar function whose superlevel sets record how far a point lies in that asymptotic direction. For a fixed direction to infinity, compare each point $x$ with the moving point $\rho(t)$ far out along the ray. The renormalized quantity $t-d(x,\rho(t))$ measures the asymptotic distance defect of $x$ relative to that direction. The following definition packages that limiting defect as a function on all of $M$, so that curvature comparison can later be applied to its level sets. [definition: Busemann Function] Let $(M,g)$ be a complete noncompact Riemannian manifold and let $\rho:[0,\infty)\to M$ be a ray. The Busemann function associated to $\rho$ is the map \begin{align*} b_\rho:M&\to \mathbb R, & x&\mapsto \lim_{t\to\infty}(t-d(x,\rho(t))). \end{align*} [/definition] The limit exists because the triangle inequality makes $t-d(x,\rho(t))$ monotone nondecreasing in $t$ and bounded above on compact sets. Its geometric value comes from curvature: on a complete manifold with $\operatorname{sec}\ge 0$, Busemann functions are concave in the usual weak comparison sense. That weak second derivative sign is the input that makes their superlevel sets convex enough for the soul construction. The curvature hypothesis is doing real work: on a complete surface of negative curvature, such as the hyperbolic plane, Busemann functions have the opposite convexity behaviour in horocyclic directions. Completeness is also needed so that rays and the comparison triangles used in the limiting argument exist for all time. The conclusion does not say that $b_\rho$ is smooth, strictly concave, or proper; in Euclidean space it is affine and its level sets are noncompact hyperplanes. This limitation is exactly why the soul construction uses families of distance-like functions rather than a single Busemann function. [example: Euclidean Busemann Functions] In $\mathbb R^n$ with the Euclidean metric, fix a unit vector $u\in \mathbb R^n$ and let $\rho(t)=tu$. For $x\in\mathbb R^n$, the Busemann function is \begin{align*} b_\rho(x) &=\lim_{t\to\infty}\left(t-d(x,tu)\right)\\ &=\lim_{t\to\infty}\left(t-|x-tu|\right). \end{align*} Since $|u|=1$, \begin{align*} |x-tu|^2 &=\langle x-tu,x-tu\rangle\\ &=\langle x,x\rangle-2t\langle x,u\rangle+t^2\langle u,u\rangle\\ &=|x|^2-2t(x\cdot u)+t^2. \end{align*} Thus \begin{align*} t-|x-tu| &=t-\sqrt{t^2-2t(x\cdot u)+|x|^2}\\ &=\frac{t^2-\left(t^2-2t(x\cdot u)+|x|^2\right)} {t+\sqrt{t^2-2t(x\cdot u)+|x|^2}}\\ &=\frac{2t(x\cdot u)-|x|^2} {t+\sqrt{t^2-2t(x\cdot u)+|x|^2}}\\ &=\frac{2(x\cdot u)-|x|^2/t} {1+\sqrt{1-2(x\cdot u)/t+|x|^2/t^2}}. \end{align*} Letting $t\to\infty$ gives \begin{align*} b_\rho(x) &=\frac{2(x\cdot u)-0}{1+\sqrt{1-0+0}}\\ &=x\cdot u. \end{align*} For $x,y\in\mathbb R^n$ and $s\in[0,1]$, \begin{align*} b_\rho((1-s)x+sy) &=((1-s)x+sy)\cdot u\\ &=(1-s)(x\cdot u)+s(y\cdot u)\\ &=(1-s)b_\rho(x)+s b_\rho(y). \end{align*} So the Busemann functions in flat space are affine, hence both convex and concave. This shows that nonnegative curvature allows flat directions at infinity rather than forcing strict convexity. [/example] The affine example exposes a limitation: a single direction to infinity does not trap compact sets. The soul construction needs a proper family of convex compact approximations, so we isolate the required object before proving its existence. [definition: Totally Convex Exhaustion] Let $(M,g)$ be a complete noncompact Riemannian manifold. A totally convex exhaustion is a nested family $(C_a)_{a\in A}$ of compact subsets of $M$, indexed by an unbounded subset $A\subset\mathbb R$, such that each $C_a$ is totally convex, $C_a\subset C_b$ whenever $a\le b$, and every compact subset of $M$ is contained in some $C_a$. [/definition] The definition records the geometric output of the distance-like functions while avoiding an unnecessary smoothness or uniqueness assumption on a single global function. What remains is to show that the curvature assumption supplies these protected compact sets without imposing them as extra structure. [quotetheorem:5384] [citeproof:5384] This theorem is the analytic entry point into the soul theorem. The sign convention is worth keeping straight: the Busemann-type limits used by Cheeger and Gromoll are concave, and the compact sets used in the construction are protected superlevel sets. Nonnegative sectional curvature is essential because Toponogov comparison is what prevents minimizing geodesics between points in one protected level set from escaping through the end of the manifold. In the hyperbolic plane, for example, the usual Busemann function has the opposite bending behaviour for the opposite curvature sign, and horoball level sets model the wrong behaviour for the Cheeger-Gromoll compact-core argument. Completeness is again not cosmetic: without it, large metric balls and limiting rays can fail to exist inside the manifold. The theorem does not produce a unique exhaustion, nor does it immediately identify the final compact core as smooth. Its role is to provide nested compact totally convex sets, and the remaining work is to pass from these nonsmooth convex sets to a smooth totally geodesic soul. ## Totally Convex Sets And Retractions The next problem is to make a convex core intrinsic. Ordinary geodesic convexity asks for at least one minimizing geodesic between two points to stay in the set. The soul construction needs a stronger condition because minimizing geodesics may not be unique. [definition: Totally Convex Subset] Let $(M,g)$ be a complete Riemannian manifold. A subset $C\subset M$ is totally convex if every geodesic segment $\gamma:[0,1]\to M$ with $\gamma(0),\gamma(1)\in C$ and length $L(\gamma)=d(\gamma(0),\gamma(1))$ satisfies $\gamma([0,1])\subset C$. [/definition] Total convexity is designed to survive cut locus phenomena. It implies that a geodesic cannot enter and exit the set through a shortcut outside it. For a candidate soul, the next issue is whether the distance from the candidate has the same convex behaviour as the exhaustion functions that produced it. In the Cheeger-Gromoll construction, the compact totally convex sets occur as protected levels of the exhaustion. If $C_a$ is one such protected compact set, then total convexity means that every minimizing geodesic with endpoints in $C_a$ remains in $C_a$. This is stronger than saying that $C_a$ is connected or geodesically convex for a preferred choice of minimizing segment; it is the property that lets the construction pass to smaller and smaller intrinsic cores without losing control of geodesics in the ambient manifold. [illustration:geometric-analysis-i-totally-convex-level-set] Once a compact core has been isolated, the retraction theory needs local control of the distance to that core. Total convexity is global and nonsmooth, while the normal bundle picture is local and differential. The bridge is the [second variation formula for squared distance to a totally geodesic submanifold](/theorems/5385): it identifies the normal directions where the distance function has product-like behaviour and measures the curvature error before the cut or focal locus appears. [quotetheorem:5385] [citeproof:5385] This result is a local second-variation statement, not a global convexity theorem for distance to an arbitrary totally geodesic submanifold. It says that normal-product directions are the model equality directions and that nearby behaviour is controlled by the index form until the cut or focal locus appears. The embedded tubular-neighbourhood hypothesis is necessary: without uniqueness of the nearest point, even the distance to a closed geodesic can fail to be smooth at points with two minimizing normal segments. Nonnegative curvature alone does not make $d(\cdot,S)$ globally convex in every direction, and the theorem is not a product-splitting statement. Its purpose here is local: it prepares the retraction that collapses the manifold back to its core while keeping control of distances. [definition: Sharafutdinov Retraction] Let $(M,g)$ be complete, open, and satisfy $\operatorname{sec}\ge 0$, and let $S\subset M$ be a soul obtained from a Cheeger-Gromoll protected exhaustion \begin{align*} S=C_0\subset C_1\subset C_2\subset \cdots \subset M. \end{align*} For each adjacent pair of protected levels, let \begin{align*} p_i:C_i\to C_{i-1} \end{align*} denote the projection obtained by the inward gradient flow of the corresponding concave barrier, interpreted by barrier approximation at nonsmooth levels. The Sharafutdinov retraction onto $S$ is the locally uniform limiting map \begin{align*} p:M\to S \end{align*} obtained from the finite compositions $p_1\circ p_2\circ\cdots\circ p_i$ on exhausting compact subsets. [/definition] This is not an arbitrary continuous retraction onto the soul. It is the particular map selected by the convex exhaustion and limiting gradient-flow procedure; it fixes $S$ because each finite-stage projection fixes the lower level already reached. The next theorem supplies the metric contraction property for this Cheeger-Gromoll-Sharafutdinov map. [quotetheorem:5386] [citeproof:5386] This retraction is stronger than a homotopy equivalence because it is compatible with the metric. The curvature hypothesis is essential: for a general open manifold retracting onto a compact submanifold, there is no reason for any retraction to be distance nonincreasing, and warped ends can force transverse distances to expand under projection. The theorem also does not say that $p$ is smooth everywhere or that it is a Riemannian submersion in full generality; additional regularity of the Sharafutdinov retraction is a separate and delicate question. What it gives for the soul theorem is enough: topology at infinity is organised by a metric collapse onto a compact core, leaving room for twisted normal bundle topology but not arbitrary end topology. The relation with the normal-bundle diffeomorphism is as follows. The soul theorem identifies each $x\in M$ with a normal vector based at $p(x)\in S$, obtained by following the normal geodesic from the soul through $x$ and recording its initial velocity. The Sharafutdinov map is the projection to the base of this bundle picture, while the normal exponential map reconstructs the point from the base point and normal vector. Thus the retraction records the homotopy and metric collapse, whereas the diffeomorphism $M\cong \nu(S)$ records the smooth topology of the open manifold. [illustration:geometric-analysis-i-sharafutdinov-retraction] [example: Product Retraction] Let $M=S\times \mathbb R^k$ with the product metric, where $S$ is compact and has $\operatorname{sec}\ge 0$, and write points as $(s,v)$ with $s\in S$ and $v\in\mathbb R^k$. The submanifold $S\times\{0\}$ is totally geodesic because the product connection splits: if a geodesic has initial velocity $(\xi,0)$ at $(s,0)$, then its $\mathbb R^k$ component is the Euclidean geodesic with initial position $0$ and initial velocity $0$, hence is identically $0$. It is also totally convex. Indeed, let \begin{align*} \gamma(t)=(\alpha(t),\beta(t)),\qquad 0\le t\le 1, \end{align*} be a minimizing geodesic in $S\times\mathbb R^k$ with endpoints $(s_0,0)$ and $(s_1,0)$. Its length is \begin{align*} L(\gamma) &=\int_0^1 \sqrt{|\alpha'(t)|^2+|\beta'(t)|^2}\,dt\\ &\ge \int_0^1 |\alpha'(t)|\,dt\\ &\ge d_S(s_0,s_1). \end{align*} On the other hand, the curve $(\sigma(t),0)$, where $\sigma$ is a minimizing geodesic from $s_0$ to $s_1$ in $S$, has length $d_S(s_0,s_1)$, so \begin{align*} d_{S\times\mathbb R^k}((s_0,0),(s_1,0))=d_S(s_0,s_1). \end{align*} Since $\gamma$ is minimizing, equality holds in the previous inequalities. Thus \begin{align*} \sqrt{|\alpha'(t)|^2+|\beta'(t)|^2}=|\alpha'(t)| \end{align*} for almost every $t$, so $|\beta'(t)|^2=0$ almost everywhere. With $\beta(0)=0$, this gives $\beta(t)=0$ for all $t$, and therefore $\gamma([0,1])\subset S\times\{0\}$. Now define \begin{align*} p:S\times\mathbb R^k&\to S\times\{0\},& p(s,v)&=(s,0). \end{align*} For $(s,v),(t,w)\in S\times\mathbb R^k$, the product distance is \begin{align*} d_{S\times\mathbb R^k}((s,v),(t,w))^2 =d_S(s,t)^2+|v-w|^2. \end{align*} Since $|v-w|^2\ge 0$, \begin{align*} d_{S\times\mathbb R^k}(p(s,v),p(t,w))^2 &=d_{S\times\mathbb R^k}((s,0),(t,0))^2\\ &=d_S(s,t)^2\\ &\le d_S(s,t)^2+|v-w|^2\\ &=d_{S\times\mathbb R^k}((s,v),(t,w))^2. \end{align*} Taking square roots gives \begin{align*} d_{S\times\mathbb R^k}(p(s,v),p(t,w)) \le d_{S\times\mathbb R^k}((s,v),(t,w)). \end{align*} Thus the projection onto $S\times\{0\}$ is distance nonincreasing; this is the product model for the Sharafutdinov retraction, with the noncompact directions exactly the normal Euclidean directions. [/example] The product example is too rigid, because normal bundles to souls need not split globally as products. Flat strips are still useful as equality models, but the standard flat strip theorem is a theorem of nonpositive curvature, not a consequence of the assumption $\operatorname{sec}\ge 0$ used in this chapter. In the soul setting, flat ruled surfaces arise only after extra equality information has been proved, for instance from a product splitting, a parallel Jacobi field calculation, or a rigidity statement for the Sharafutdinov retraction. Thus the correct lesson is not that nonnegative curvature automatically forces every pair of parallel geodesics to bound a flat strip; rather, whenever the soul construction produces a genuine parallel normal variation with vanishing index form, the resulting two-plane has zero sectional curvature and behaves locally like a product. [remark: Flat Strips As Equality Models] The classical flat strip theorem says that in a complete simply connected manifold with nonpositive sectional curvature, two parallel geodesic lines bound a totally geodesic flat strip. This sign is opposite to the standing hypothesis $\operatorname{sec}\ge 0$ in the soul theorem. In nonnegative curvature, flat strips should therefore be treated as rigidity phenomena coming from additional equality cases, not as a general comparison theorem. [/remark] This distinction matters for the rest of the chapter. The soul theorem uses nonnegative curvature to build convex cores and metric retractions, while flat strips only diagnose the special directions in which those constructions have equality. [example: Flat Strips In A Product] Let $a>0$ and let $\eta:\mathbb R\to N$ be a unit-speed geodesic. In $N\times\mathbb R$ with the product metric, define \begin{align*} F:\mathbb R\times[0,a]&\to N\times\mathbb R,& F(t,s)&=(\eta(t),s). \end{align*} The two boundary curves are \begin{align*} \gamma_0(t)&=F(t,0)=(\eta(t),0),& \gamma_1(t)&=F(t,a)=(\eta(t),a). \end{align*} Because the product connection splits, $\gamma_0$ and $\gamma_1$ are geodesics: their accelerations are \begin{align*} \nabla^{N\times\mathbb R}_{\gamma_i'}\gamma_i' = \left(\nabla^N_{\eta'}\eta',0\right) = (0,0), \end{align*} since $\eta$ is a geodesic in $N$. For fixed $t$, the connecting segment from $\gamma_0(t)$ to $\gamma_1(t)$ is \begin{align*} \sigma_t(s)=F(t,s)=(\eta(t),s),\qquad 0\le s\le a. \end{align*} Its velocity is $\sigma_t'(s)=(0,1)$, so its length is \begin{align*} L(\sigma_t) &=\int_0^a |(0,1)|\,ds\\ &=\int_0^a 1\,ds\\ &=a. \end{align*} The tangent vectors of the ruled surface are \begin{align*} F_t(t,s)&=(\eta'(t),0),& F_s(t,s)&=(0,1). \end{align*} Using the product metric and $|\eta'(t)|=1$, \begin{align*} \langle F_t,F_t\rangle &=\langle \eta'(t),\eta'(t)\rangle_N+0\\ &=1,\\ \langle F_s,F_s\rangle &=0+1\\ &=1,\\ \langle F_t,F_s\rangle &=\langle \eta'(t),0\rangle_N+0\\ &=0. \end{align*} Thus \begin{align*} F^*g_{N\times\mathbb R} = dt^2+ds^2, \end{align*} so $F$ is an isometric parametrisation of the strip $\mathbb R\times[0,a]$ with its Euclidean product metric. The connecting segments are perpendicular to both boundary geodesics because $\langle F_s,F_t\rangle=0$ along $s=0$ and $s=a$. Hence these two parallel product geodesics bound a genuine flat strip; the flatness comes from the product metric, while the assumption that $\eta$ is geodesic is what makes the two boundary curves geodesics. [/example] ## The Soul Theorem The final problem is topological: classify the possible complete open manifolds with nonnegative sectional curvature. Compact comparison theorems give diameter and volume consequences, but open manifolds require a compact replacement for the missing boundary. The soul is that replacement. [definition: Soul] Let $(M,g)$ be a complete noncompact Riemannian manifold with $\operatorname{sec}\ge 0$. A soul of $M$ is a compact, connected, totally geodesic, totally convex submanifold $S\subset M$ without boundary. [/definition] The definition names the geometric core produced by the convex-exhaustion procedure, but it does not yet say that such a core exists or that it controls the topology of the whole open manifold. The next theorem is needed to turn the nested totally convex sets into a smooth compact core and then to prove that the entire manifold is recovered from the normal directions to that core. This is the point where the analytic exhaustion becomes a topological classification statement. [quotetheorem:5387] [citeproof:5387] The theorem converts a curvature inequality into a classification principle. Completeness and noncompactness are both essential: deleting a point from a compact nonnegatively curved manifold creates an open manifold where geodesic completeness fails, while compact manifolds have no end from which to extract the convex exhaustion. Connectedness avoids the irrelevant ambiguity of choosing separate souls in different components. The theorem does not say that $M$ is a product, does not assert uniqueness of the soul as a subset, and does not classify which vector bundles admit nonnegatively curved total spaces. Its output is a diffeomorphism type controlled by the normal bundle of some compact totally geodesic core. Topologically, the theorem says that all characteristic-class information of $M$ at infinity is carried by the vector bundle $\nu(S)\to S$. When the normal bundle is twisted, the open manifold can have tangent-bundle or sphere-bundle topology even though it retracts to a compact core. Rigidity results such as Perelman's theorem below add curvature positivity to force this bundle picture to collapse to the point-soul case. [illustration:geometric-analysis-i-soul-theorem] [example: Tangent Bundles Of Spheres] For a concrete case, take one of the standard complete nonnegatively curved metrics on $TS^2$ coming from the Cheeger-Gromoll-type constructions on tangent bundles of compact rank-one symmetric spaces. In this metric, the zero section \begin{align*} S^2&\hookrightarrow TS^2,& p&\mapsto (p,0) \end{align*} is the soul. The normal bundle of the zero section inside the total space $TS^2$ is naturally identified with $TS^2$ itself, because at each point $(p,0)$ the vertical tangent space is \begin{align*} T_{(p,0)}(T_pS^2)\cong T_pS^2. \end{align*} This normal bundle is not the product bundle $S^2\times\mathbb R^2$. Indeed, the Euler class of $TS^2$ satisfies \begin{align*} \langle e(TS^2),[S^2]\rangle &=\chi(S^2)\\ &=2, \end{align*} by the Poincare-Hopf theorem. The product bundle has a nowhere-zero constant fibrewise section, so its Euler class is \begin{align*} e(S^2\times\mathbb R^2)=0. \end{align*} Since Euler class is invariant under vector-bundle isomorphism, $TS^2$ is not isomorphic to $S^2\times\mathbb R^2$ as a rank-$2$ real vector bundle. Thus the soul theorem has to say that $M$ is diffeomorphic to the normal bundle of the soul, not necessarily to a product $S\times\mathbb R^k$. Tangent bundles of spheres give exactly this phenomenon: the soul is the zero section, while the noncompact directions can be twisted by the topology of the tangent bundle. [/example] A different source of examples comes from convex hypersurfaces in Euclidean space. These show that souls may collapse to a point. [example: Paraboloids With Nonnegative Curvature] Parametrize the paraboloid as the graph \begin{align*} F:\mathbb R^n&\to \mathbb R^{n+1},& F(x)&=(x,|x|^2). \end{align*} For $v\in T_x\mathbb R^n\cong\mathbb R^n$, \begin{align*} dF_x(v)&=(v,2\langle x,v\rangle), \end{align*} so the induced metric is \begin{align*} g_x(v,w) &=\langle v,w\rangle_{\mathbb R^n}+4\langle x,v\rangle\langle x,w\rangle. \end{align*} In particular, \begin{align*} g_x(v,v) &=|v|^2+4\langle x,v\rangle^2\\ &\ge |v|^2. \end{align*} Thus every divergent curve in $\mathbb R^n$ has at least its Euclidean length in the induced metric, and since $\mathbb R^n$ is complete, the paraboloid is complete. The upward unit normal is \begin{align*} N(x)=\frac{(-2x,1)}{\sqrt{1+4|x|^2}}. \end{align*} Since the graphing function is $f(x)=|x|^2$, its Hessian satisfies \begin{align*} \operatorname{Hess} f(v,w)=2\langle v,w\rangle. \end{align*} The second fundamental form of a graph is therefore \begin{align*} \mathrm{II}_x(v,w) &=\frac{\operatorname{Hess} f(v,w)}{\sqrt{1+|\nabla f(x)|^2}}\\ &=\frac{2\langle v,w\rangle}{\sqrt{1+4|x|^2}}. \end{align*} For every nonzero $v$, \begin{align*} \mathrm{II}_x(v,v) &=\frac{2|v|^2}{\sqrt{1+4|x|^2}} >0, \end{align*} so the shape operator is positive definite. By the Gauss equation for a hypersurface in Euclidean space, if $\sigma=\operatorname{span}\{e_i,e_j\}$ is a $g$-orthonormal two-plane diagonalizing the shape operator with principal curvatures $\kappa_i,\kappa_j$, then \begin{align*} \operatorname{sec}_M(\sigma) &=\kappa_i\kappa_j \ge 0. \end{align*} Thus the paraboloid has nonnegative sectional curvature. The map $F:\mathbb R^n\to M$ is a global parametrization with inverse $(x,z)\mapsto x$, so $M$ is diffeomorphic to $\mathbb R^n$. Its soul is therefore a point: a positive-dimensional compact soul would have the homotopy type of $M$ by the *[Cheeger-Gromoll Soul Theorem](/theorems/5387)*, but $M\cong\mathbb R^n$ is contractible while a closed positive-dimensional manifold has nonzero top-dimensional $\mathbb Z_2$-homology. Hence this curved complete example has the same smooth topology as Euclidean space. [/example] The tangent-bundle examples show that nonnegative curvature allows positive-dimensional souls, while the paraboloid shows that a point soul can occur. The remaining rigidity question is whether positive-dimensional souls can survive once the curvature is strictly positive somewhere. A positive-dimensional soul would force flat normal directions near the soul, so strict positivity at even one point creates a global obstruction to that topology. [remark: Quoted result: Perelman Soul Conjecture Theorem] Let $(M,g)$ be complete, connected, noncompact, and satisfy $\operatorname{sec}\ge 0$. If there exists a point $p\in M$ at which every sectional curvature is positive, then every soul of $M$ is a point. Hence $M$ is diffeomorphic to $\mathbb R^n$. [/remark] Perelman's theorem is a deep strengthening of the Cheeger-Gromoll soul theorem and lies outside the elementary soul construction developed in these notes. The preceding flat-strip heuristics explain why positive-dimensional souls are suspicious under strict positivity, but they do not by themselves prove the global rigidity statement. The hypotheses are sharp in spirit: if the curvature is merely nonnegative everywhere with no point of strict positivity, products such as $S\times\mathbb R^k$ have positive-dimensional souls; if completeness is dropped, the conclusion need not reflect the global topology of the metric completion. The theorem does not say that the metric is rotationally symmetric or close to the Euclidean metric, only that the underlying smooth manifold is $\mathbb R^n$. This result is a useful warning about equality cases. Nonnegative curvature allows Euclidean directions, vector bundle topology, and flat product phenomena; strict positivity at one point removes these phenomena at the level of the soul and leaves only the point-soul case. The soul theorem completes the noncompact rigidity story by showing how nonnegative sectional curvature collapses the manifold onto a compact core. The next chapter shifts from individual manifolds to sequences of them, asking when the comparison-geometric estimates survive under Gromov-Hausdorff limits. # 11. Gromov-Hausdorff Convergence and Compactness This chapter turns comparison geometry into compactness theory. Earlier chapters used curvature bounds to control geodesics, volumes, and triangle geometry inside one manifold; now we ask when a whole sequence of manifolds has a geometric limit. The prerequisites are compact metric spaces, Hausdorff distance for compact subsets of a metric space, geodesic and length spaces, the [Bishop-Gromov volume comparison theorem](/theorems/5371), and the Ricci and sectional curvature comparison results from the preceding chapters. The point is that a Riemannian manifold can be treated as a metric space, so convergence must be formulated without choosing coordinates or embeddings into a fixed ambient space. The guiding tension is between control and degeneration. Lower curvature bounds and diameter bounds often force enough finite-dimensional behaviour to extract a convergent subsequence, but the limit may fail to be a smooth manifold if regions collapse or singularities form. Gromov-Hausdorff convergence is the language that records exactly what survives at the metric level. ## Comparing Metric Spaces Without an Ambient Space Suppose two compact metric spaces are meant to be close, but they are not subsets of the same space. The first problem is to replace pointwise comparison by finite sampling: if every point of a [compact space](/page/Compact%20Space) lies near a controlled finite set, then the whole metric can be approximated by finitely many distances. [definition: Epsilon Net] Let $(X,d)$ be a metric space and let $\varepsilon>0$. A subset $A \subset X$ is an $\varepsilon$-net in $X$ if \begin{align*} X = \bigcup_{a \in A} B(a,\varepsilon). \end{align*} [/definition] An $\varepsilon$-net is useful because it reduces the geometry of $X$ at scale $\varepsilon$ to a finite or countable list of centres. In compact spaces, finite nets exist at every scale, and this is the first compactness feature used in the Gromov argument. [example: Epsilon Nets On The Circle] Let $S^1$ have circumference $2\pi$ and intrinsic length distance. Fix $N\in\mathbb N$, and choose points \begin{align*} p_k=e^{2\pi i k/N},\qquad k=0,1,\dots,N-1. \end{align*} The intrinsic distance along $S^1$ between consecutive chosen points is the shorter arc length \begin{align*} d(p_k,p_{k+1})=\frac{2\pi}{N}, \end{align*} with indices taken modulo $N$. If $q\in S^1$, then $q$ lies on one of the arcs from $p_k$ to $p_{k+1}$ of length $2\pi/N$. Writing $t$ for the distance from $q$ to $p_k$ along that arc, the distance from $q$ to $p_{k+1}$ along the same arc is $2\pi/N-t$, so \begin{align*} \min\{d(q,p_k),d(q,p_{k+1})\} &\le \min\left\{t,\frac{2\pi}{N}-t\right\} \\ &\le \frac{\pi}{N}. \end{align*} Thus every point of $S^1$ is within distance $\pi/N$ of one of the selected points. With closed balls this is a $\pi/N$-net; with open balls it is an $\varepsilon$-net for every $\varepsilon>\pi/N$. This calculation shows the scale dependence of covering number: at scale $\varepsilon\approx \pi/N$, the circle is covered by $N\approx \pi/\varepsilon$ centers, so the number of centers grows like $\varepsilon^{-1}$. A two-dimensional surface with comparable local geometry instead has area-scale cells of size about $\varepsilon^2$, so its covering number grows like $\varepsilon^{-2}$. [/example] Finite nets compare one space to itself. To compare two spaces, we allow an approximate distance-preserving map whose image is dense at a chosen scale. [definition: Epsilon Approximation] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces. A map $f:X \to Y$ is an $\varepsilon$-approximation if \begin{align*} |d_Y(f(x),f(x'))-d_X(x,x')| &< \varepsilon \quad \text{for all } x,x'\in X,\\ Y &= \bigcup_{x\in X} B(f(x),\varepsilon). \end{align*} [/definition] The map in this definition need not be continuous. Its role is metric rather than topological: it says that all distances in $X$ are reproduced in $Y$ up to error $\varepsilon$, and every point of $Y$ is seen from the image. [definition: Hausdorff Distance] Let $(Z,d_Z)$ be a metric space and let $A,B\subset Z$ be nonempty compact subsets. The Hausdorff distance between $A$ and $B$ in $Z$ is \begin{align*} d_H^Z(A,B)=\inf\{r>0: A\subset \bigcup_{b\in B}B(b,r) \text{ and } B\subset \bigcup_{a\in A}B(a,r)\}. \end{align*} [/definition] Hausdorff distance solves the comparison problem once the two spaces have been embedded into a common ambient metric space. Gromov's idea is to minimize over all possible common ambient spaces. [illustration:geometric-analysis-i-gh-ambient-space] The previous definition still depends on the chosen ambient space $Z$. To compare intrinsic spaces, such as two manifolds with no preferred embedding into Euclidean space, we remove that choice by allowing every possible common metric realization and keeping only the best Hausdorff error. [definition: Gromov-Hausdorff Distance] Let $(X,d_X)$ and $(Y,d_Y)$ be compact metric spaces. Their Gromov-Hausdorff distance is \begin{align*} d_{GH}(X,Y)=\inf d_H^Z(\varphi(X),\psi(Y)), \end{align*} where the infimum is taken over all metric spaces $(Z,d_Z)$ and all isometric embeddings $\varphi:X\to Z$, $\psi:Y\to Z$. [/definition] The distance $d_{GH}$ identifies compact metric spaces up to isometry, but the definition hides the ambient space over which the infimum is taken. This is inconvenient for applications: comparison geometry usually supplies finite nets, almost distance-preserving maps, or a matching of points, not an explicit optimal ambient space. To use the definition in compactness proofs, we need a way to encode comparison data directly on $X\times Y$ and then recover the same Gromov-Hausdorff topology from that data. The next definition supplies this symmetric matching language. [illustration:geometric-analysis-i-epsilon-approximation] Approximate maps compare spaces in one direction, but limits often come from symmetric data: a point in one space may be represented by several nearby points in another, especially near collapse or branching. We need the following definition to record this matching relation without forcing a single-valued inverse, and its distortion measures how much the relation changes pairwise distances. [definition: Correspondence And Distortion] Let $(X,d_X)$ and $(Y,d_Y)$ be metric spaces. A correspondence between $X$ and $Y$ is a relation $R\subset X\times Y$ such that for every $x\in X$ there exists $y\in Y$ with $(x,y)\in R$, and for every $y\in Y$ there exists $x\in X$ with $(x,y)\in R$. Its distortion is \begin{align*} \operatorname{dis}(R)=\sup\{|d_X(x,x')-d_Y(y,y')|:(x,y),(x',y')\in R\}. \end{align*} [/definition] Correspondences are the symmetric replacement for approximate maps. They allow a point of one space to be matched with several points of the other, which is useful when no canonical direction of comparison is available. The point of introducing them is not a new topology, but a usable test for the same topology. The next theorem is needed to justify changing freely among three forms of comparison: small Hausdorff distance after a common embedding, small-error approximate maps, and small-distortion correspondences. Once these are equivalent up to constants, later compactness proofs can choose whichever form the geometry naturally supplies. [quotetheorem:5388] [citeproof:5388] This criterion is the working definition in many arguments, but each hypothesis has content. Compactness ensures that small-scale sampling controls the whole space; for noncompact spaces, an approximate model on one bounded region says little about distant geometry, which is why pointed Gromov-Hausdorff convergence is used instead. The almost dense image condition cannot be dropped: a point embedded isometrically into a circle has zero distortion as a map from the point, but it does not approximate the circle because most of the circle is unseen. The distortion condition cannot be dropped either: any compact space maps onto a point with dense image in the one-point target, but this destroys all distances unless the original diameter tends to zero. Thus the criterion is weaker than smooth convergence, and it records only metric information supplied simultaneously by distance control and approximate density. [example: Rescaled Spheres] Let $S^n_r\subset \mathbb R^{n+1}$ carry the intrinsic round metric of radius $r$, and let $P=\{*\}$ be the one-point metric space. A minimizing geodesic on $S^n_r$ is a great-circle arc, so if two points have central angle $\alpha\in[0,\pi]$, their intrinsic distance is \begin{align*} d_r(x,y)=r\alpha. \end{align*} The largest possible central angle is $\pi$, attained by antipodal points, hence \begin{align*} \operatorname{diam}(S^n_r) &=\sup_{x,y\in S^n_r} d_r(x,y) \\ &=\sup_{\alpha\in[0,\pi]} r\alpha \\ &=r\pi. \end{align*} To see the Gromov-Hausdorff convergence directly, form a metric space $Z_r=S^n_r\sqcup P$ by keeping the metric $d_r$ on $S^n_r$ and setting \begin{align*} d_{Z_r}(x,*)=\pi r \end{align*} for every $x\in S^n_r$. The triangle inequality involving two points $x,y\in S^n_r$ and $*$ follows from \begin{align*} d_r(x,y)\le \operatorname{diam}(S^n_r)=\pi r\le 2\pi r=d_{Z_r}(x,*)+d_{Z_r}(*,y). \end{align*} Thus $S^n_r$ and $P$ are embedded isometrically in $Z_r$, and every point of $S^n_r$ is within distance $\pi r$ of $*$, while $*$ lies in $P$. Therefore \begin{align*} d_{GH}(S^n_r,P)\le d_H^{Z_r}(S^n_r,P)\le \pi r\to 0. \end{align*} Under the standard dilation map $F:S^n_1\to S^n_r$, $F(u)=ru$, the round metric pulls back as \begin{align*} F^*g_r=r^2 g_1. \end{align*} After rescaling by $r^{-2}$, \begin{align*} F^*(r^{-2}g_r)=r^{-2}F^*g_r=r^{-2}r^2g_1=g_1. \end{align*} So the rescaled sphere is isometric to the fixed unit sphere $S^n_1$. The same underlying family therefore has two different metric limits depending on scale: it collapses to a point at its original scale, but remains the unit sphere after distances are divided by $r$. [/example] ## The Arzela-Ascoli Method For Compact Metric Spaces The [compactness theorem](/theorems/2748) for metric spaces asks for conditions that prevent infinitely many unrelated shapes from appearing either at smaller and smaller scales or by drifting farther and farther apart at large scale. The answer is uniform [total boundedness](/page/Total%20Boundedness) plus a uniform diameter bound: at every scale, the spaces admit nets with a uniformly bounded number of points, and their total size is bounded. [definition: Uniform Covering Bound] A class $\mathcal C$ of compact metric spaces has a uniform covering bound if for every $\varepsilon>0$ there exists $N(\varepsilon)\in \mathbb N$ such that every $X\in \mathcal C$ admits an $\varepsilon$-net with at most $N(\varepsilon)$ points. [/definition] This condition is the metric-space analogue of equicontinuity in Arzela-Ascoli, while the diameter bound keeps all finite distance tables in one bounded range. Together they give a diagonal process over a sequence of scales, where each space is replaced by a finite metric table. The next theorem is needed because it says that no further hidden regularity is required: uniform covering data at all scales, together with large-scale boundedness, is exactly the precompactness condition. [quotetheorem:5389] [citeproof:5389] The covering bound is not a technical convenience; it is the condition that rules out new independent points appearing at a fixed scale. If it is removed, bounded diameter alone does not prevent a sequence from having larger and larger separated subsets, so no finite limiting compact metric space can contain all the required mutual distances. The following example isolates this failure at scale $1/3$: every pair of points remains distance $1$ apart, and the number of such points tends to infinity. [example: Why Diameter Alone Is Not Enough] Let $X_m=\{x_1,\dots,x_m\}$, equipped with the metric \begin{align*} d(x_i,x_j)= \begin{cases} 0, & i=j,\\ 1, & i\ne j. \end{cases} \end{align*} Since $m\ge 2$ gives pairs of distinct points, its diameter is \begin{align*} \operatorname{diam}(X_m) &=\sup_{1\le i,j\le m} d(x_i,x_j) \\ &=1. \end{align*} If $A\subset X_m$ and $a\in A$, then the open ball of radius $1/3$ around $a$ is \begin{align*} B(a,1/3) &=\{x\in X_m:d(x,a)<1/3\}\\ &=\{a\}, \end{align*} because every $x\ne a$ satisfies $d(x,a)=1>1/3$. Hence \begin{align*} \bigcup_{a\in A}B(a,1/3)=A. \end{align*} For $A$ to be a $1/3$-net, this union must equal $X_m$, so $A=X_m$ and therefore \begin{align*} |A|=m. \end{align*} Thus, at the single fixed scale $1/3$, the minimum number of centres needed to cover $X_m$ is $m$. If a uniform covering bound existed, there would be an integer $N(1/3)$ such that every $X_m$ had a $1/3$-net with at most $N(1/3)$ points; choosing $m>N(1/3)$ contradicts the calculation above. The diameters are uniformly bounded by $1$, but the covering numbers are not uniformly bounded, so the class is not precompact in the Gromov-Hausdorff topology by the compactness criterion stated above. [/example] For Riemannian manifolds, comparison geometry enters by proving covering bounds from curvature and volume information. The next section translates Bishop-Gromov type estimates into the hypotheses of the compactness theorem. ## Curvature Bounds And Precompactness A lower Ricci or sectional curvature bound does not directly say that a manifold has finitely many shapes. It gives comparison estimates for volumes of balls, and these estimates become covering estimates when paired with diameter and noncollapse assumptions. [definition: Noncollapsed Riemannian Class] Let $n\in\mathbb N$, $D>0$, $v>0$, and $K\in\mathbb R$. A class of closed $n$-dimensional Riemannian manifolds $(M,g)$ is noncollapsed with Ricci lower bound, diameter bound, and volume lower bound if \begin{align*} \operatorname{Ric}_g &\ge (n-1)K g,\\ \operatorname{diam}(M,g) &\le D,\\ \operatorname{Vol}_g(M) &\ge v. \end{align*} [/definition] The volume lower bound prevents the entire manifold from shrinking to a lower-dimensional object. For plain Gromov-Hausdorff precompactness, what is needed is a uniform covering bound at every scale; the Ricci lower bound and diameter bound supply the comparison framework, while the volume lower bound gives enough noncollapse to turn Bishop-Gromov estimates into those covering numbers with constants depending only on $n,K,D,v$. The next theorem is therefore a metric precompactness theorem for the noncollapsed class, not a smooth compactness theorem and not a claim that every Ricci-and-diameter-bounded family is noncollapsed. [quotetheorem:5390] [citeproof:5390] The conclusion is compactness of metric spaces, not smooth compactness of manifolds. Each assumption prevents a different escape. Without a diameter bound, round spheres or hyperbolic balls with radii tending to infinity leave every compact Gromov-Hausdorff class. Without a Ricci lower bound, sequences with more and more thin negatively curved necks inside a fixed diameter and fixed total volume range can have no uniform covering bound at the neck scale. Without volume noncollapse, the flat tori $S^1\times S^1_{\varepsilon_j}$ have Ricci curvature $0$ and bounded diameter but converge to a circle, so dimension and Riemannian volume information are lost. Even with all three hypotheses, limits are compact length spaces and may have singular sets rather than smooth manifold structure. [example: Pointed Limits Of Hyperbolic Balls] Let $\mathbb H^n$ have sectional curvature $-1$, fix $o\in \mathbb H^n$, and set \begin{align*} X_j=B_{\mathbb H^n}(o,R_j), \end{align*} with the restricted hyperbolic distance and $R_j\to\infty$. For any $x,y\in X_j$, the triangle inequality gives \begin{align*} d_{\mathbb H^n}(x,y) &\le d_{\mathbb H^n}(x,o)+d_{\mathbb H^n}(o,y)\\ &\le R_j+R_j\\ &=2R_j. \end{align*} Conversely, choose a unit-speed complete geodesic $\gamma:\mathbb R\to \mathbb H^n$ with $\gamma(0)=o$. Then $\gamma(R_j),\gamma(-R_j)\in X_j$, and since $\gamma$ is distance-minimizing on each interval, \begin{align*} d_{\mathbb H^n}(\gamma(R_j),\gamma(-R_j)) &=|R_j-(-R_j)|\\ &=2R_j. \end{align*} Therefore \begin{align*} \operatorname{diam}(X_j)=2R_j\to\infty. \end{align*} So the compact spaces $X_j$ cannot form a compact Gromov-Hausdorff convergent sequence: if $X_j\to Y$ for a compact metric space $Y$, then for all large $j$ there would be $\varepsilon_j$-approximations $X_j\to Y$ with $\varepsilon_j\to 0$, and hence \begin{align*} \operatorname{diam}(X_j) &=\sup_{x,x'\in X_j} d_{X_j}(x,x')\\ &\le \sup_{x,x'\in X_j}\bigl(d_Y(f_j(x),f_j(x'))+\varepsilon_j\bigr)\\ &\le \operatorname{diam}(Y)+\varepsilon_j, \end{align*} contradicting $\operatorname{diam}(X_j)\to\infty$. Pointed convergence only asks for convergence on each fixed radius around the basepoint. Fix $R>0$. Since $R_j\to\infty$, there is $j_0$ such that $R_j\ge R$ for every $j\ge j_0$. For those $j$, \begin{align*} B_{X_j}(o,R) &=\{x\in X_j:d_{X_j}(o,x)\le R\}\\ &=\{x\in \mathbb H^n:d_{\mathbb H^n}(o,x)\le R\}\\ &=B_{\mathbb H^n}(o,R). \end{align*} The identity map on this ball preserves every distance exactly, so the Gromov-Hausdorff distance between the radius-$R$ pointed balls is $0$ for all $j\ge j_0$. Thus $(B(o,R_j),o)$ converges in the pointed Gromov-Hausdorff sense to $(\mathbb H^n,o)$, even though the unpointed compact spaces have no compact Gromov-Hausdorff limit. [/example] The next theorem is needed because compactness from sectional curvature uses a different mechanism from the Ricci noncollapse argument. The pointed example separates compactness from large-scale convergence and returns us to the closed diameter-bounded setting. Sectional comparison controls metric packing even when total volume is allowed to collapse, so it gives precompactness without adding a volume lower bound. [remark: Quoted result: Gromov Precompactness Theorem for Manifolds with Sectional Curvature Bounded Below] Let $n \in \mathbb{N}$, let $K \in \mathbb{R}$, and let $D > 0$. Let $\mathcal{M}(n,K,D)$ be the class of isometry classes of closed $n$-dimensional Riemannian manifolds $(M,g)$ such that \begin{align*} \sec_g(\sigma) \geq K \end{align*} for every $p \in M$ and every two-dimensional subspace $\sigma \subset T_pM$, and \begin{align*} \operatorname{diam}(M,d_g) \leq D, \end{align*} where $d_g$ denotes the Riemannian distance induced by $g$. Then $\mathcal{M}(n,K,D)$ is precompact in the Gromov-Hausdorff topology on compact metric spaces. Equivalently, every sequence $((M_j,g_j))_{j=1}^{\infty}$ in $\mathcal{M}(n,K,D)$ has a subsequence whose associated compact metric spaces $(M_{j_\ell},d_{g_{j_\ell}})$ converge in the Gromov-Hausdorff topology. [/remark] The sectional version does not require a separate noncollapse hypothesis for precompactness, although collapse can still occur. Its remaining hypotheses are still essential. Fixed dimension is needed because the unit spheres $S^n$ have diameter $\pi$ and sectional curvature $1$, but their covering numbers grow with $n$ at any small fixed scale. The diameter bound is needed because rescaled round circles of length tending to infinity have sectional curvature $0$ and no compact limit. A uniform lower sectional curvature bound is needed because increasingly branched or necked surfaces of bounded diameter can create arbitrarily large packing numbers at a fixed scale. What changes under allowed collapse is the dimension and regularity of the limit, not the existence of a subsequential metric limit. ## Collapsing And Noncollapsing Limits Compactness theorems guarantee subsequences, but they do not decide the dimension or smoothness of the limit. The next question is what geometric information survives when volume degenerates, and which comparison estimates are stable under the limiting process. [definition: Collapsing Sequence] Let $(M_j,g_j)$ be a sequence of closed $n$-dimensional Riemannian manifolds with uniformly bounded diameter. In the Ricci setting, assume $\operatorname{Ric}_{g_j}\ge (n-1)K g_j$ for a fixed $K\in\mathbb R$; in the sectional or Alexandrov setting, assume $\sec_{g_j}\ge K$ for a fixed $K\in\mathbb R$. The sequence collapses if \begin{align*} \operatorname{Vol}_{g_j}(M_j)\to 0. \end{align*} It is noncollapsing if there exists $v>0$ such that $\operatorname{Vol}_{g_j}(M_j)\ge v$ for all $j$. [/definition] Collapse means that the metric spaces may converge to an object of smaller Hausdorff dimension. The basic model is a product where one factor is being shrunk while curvature remains controlled. [illustration:geometric-analysis-i-flat-tori-collapse] [example: Collapsing Flat Tori] Let $M_j=S^1\times S^1$ with angular coordinates $\theta,\phi\in \mathbb R/2\pi\mathbb Z$ and flat product metric \begin{align*} g_j=d\theta^2+\varepsilon_j^2 d\phi^2, \end{align*} where $\varepsilon_j>0$ and $\varepsilon_j\to 0$. Since this is the product of two flat one-dimensional metrics with constant scaling in the second factor, every coordinate vector field is parallel in these coordinates, so the curvature tensor vanishes and every sectional curvature is $0$. The Riemannian area element is \begin{align*} dA_j &=\sqrt{\det\begin{pmatrix}1&0\\0&\varepsilon_j^2\end{pmatrix}}\,d\theta\,d\phi\\ &=\sqrt{\varepsilon_j^2}\,d\theta\,d\phi\\ &=\varepsilon_j\,d\theta\,d\phi. \end{align*} Therefore \begin{align*} \operatorname{Area}(M_j) &=\int_0^{2\pi}\int_0^{2\pi}\varepsilon_j\,d\phi\,d\theta\\ &=\int_0^{2\pi} 2\pi\varepsilon_j\,d\theta\\ &=4\pi^2\varepsilon_j\to 0. \end{align*} If $d_{S^1}$ denotes intrinsic distance on the unit circle, then the product distance is \begin{align*} d_j\bigl((\theta,\phi),(\theta',\phi')\bigr) =\sqrt{d_{S^1}(\theta,\theta')^2+\varepsilon_j^2 d_{S^1}(\phi,\phi')^2}. \end{align*} Since $d_{S^1}(\theta,\theta')\le \pi$ and $d_{S^1}(\phi,\phi')\le \pi$, \begin{align*} d_j\bigl((\theta,\phi),(\theta',\phi')\bigr) &\le \sqrt{\pi^2+\varepsilon_j^2\pi^2}\\ &=\pi\sqrt{1+\varepsilon_j^2}. \end{align*} The sequence $(\varepsilon_j)$ is bounded because it converges, so the diameters are uniformly bounded. Let \begin{align*} \pi_j:M_j\to S^1,\qquad \pi_j(\theta,\phi)=\theta. \end{align*} For $p=(\theta,\phi)$ and $q=(\theta',\phi')$, set \begin{align*} a=d_{S^1}(\theta,\theta'),\qquad b=d_{S^1}(\phi,\phi'). \end{align*} Then $0\le b\le \pi$ and \begin{align*} d_j(p,q) &=\sqrt{a^2+\varepsilon_j^2b^2}\\ &\ge a\\ &=d_{S^1}(\pi_j(p),\pi_j(q)). \end{align*} Also, since $\sqrt{u^2+v^2}\le u+v$ for $u,v\ge 0$, \begin{align*} d_j(p,q) &=\sqrt{a^2+(\varepsilon_j b)^2}\\ &\le a+\varepsilon_j b\\ &\le a+\pi\varepsilon_j. \end{align*} Hence \begin{align*} \left|d_{S^1}(\pi_j(p),\pi_j(q))-d_j(p,q)\right| \le \pi\varepsilon_j <2\pi\varepsilon_j. \end{align*} The map $\pi_j$ is onto, because every $\theta\in S^1$ is $\pi_j(\theta,0)$. Thus its image is $2\pi\varepsilon_j$-dense in $S^1$, and $\pi_j$ is a $2\pi\varepsilon_j$-approximation. Since $2\pi\varepsilon_j\to 0$, the *Epsilon Approximation Criterion* gives \begin{align*} d_{GH}(M_j,S^1)\to 0. \end{align*} So the flat tori collapse with bounded curvature and bounded diameter to a one-dimensional circle, while their two-dimensional areas vanish. [/example] This example shows why Gromov-Hausdorff limits need not preserve dimension. Nevertheless, many comparison inequalities are formulated purely in terms of distances and measures, so they have limiting forms. [quotetheorem:5391] [citeproof:5391] The hypotheses describe exactly which information can be passed through the limit. Geodesicity is needed for triangle comparison statements because comparison angles and sides are defined using geodesic triangles. For instance, $\mathbb Q\cap[0,1]$ with the Euclidean distance inherits all finite distance equalities from the interval, but it is not a geodesic space because most pairs of rational endpoints are not joined by an isometric copy of the whole interval inside $\mathbb Q\cap[0,1]$. Compactness is what lets all finitely many chosen points be approximated at once and supplies convergent midpoint or almost-geodesic choices; without it, only pointed or local statements are available. The closed-inequality condition is also necessary: a sequence of strict inequalities such as $a_j<b_j$ can converge to equality, so open comparison conditions need not survive. Lower sectional curvature bounds pass naturally to Alexandrov limits for this reason. Ricci lower bounds are subtler because Ricci curvature is not determined by individual triangles; ordinary Gromov-Hausdorff convergence can collapse flat tori to a circle, where the metric limit alone no longer remembers the original two-dimensional volume comparison. The stable replacement involves volume comparison, optimal transport, or measured Gromov-Hausdorff convergence. The next theorem records the noncollapsed form closest to the Bishop-Gromov estimates proved earlier in the course. [quotetheorem:5392] [citeproof:5392] The theorem explains the slogan that Ricci lower bounds survive in measured form. Without noncollapse, the limiting measure may live on a lower-dimensional space or depend on normalization, so the statement has to remember both the metric and the chosen measures. This measured viewpoint is also the point at which comparison geometry connects to analysis on metric spaces: doubling, Poincare-type inequalities, heat kernels, and spectral convergence require measures as well as distances. In moduli problems, Gromov-Hausdorff compactness plays the same organizing role as compactness theorems in topology and algebraic geometry, separating the existence of a limiting object from the extra work needed to identify its regularity. [remark: What Compactness Does Not Give] Gromov-Hausdorff convergence alone does not provide differentiable coordinate charts, curvature tensors on the limit, or convergence of eigenfunctions and heat kernels. Such conclusions require additional regularity theory, for example two-sided Ricci bounds, injectivity radius lower bounds, harmonic-coordinate estimates, or synthetic analysis on metric measure spaces. [/remark] The compactness viewpoint closes the comparison part of the course. Curvature bounds first produced inequalities for geodesics, triangles, and volumes; Gromov-Hausdorff convergence packages those inequalities into a language stable under limits, while the distinction between collapsing and noncollapsing records exactly where smooth geometry can be lost. Compactness and Gromov-Hausdorff convergence package the comparison results into a limit theory for manifolds with curvature bounds. The final chapter uses that framework to organize the equality cases, stability statements, and core tools that make comparison geometry effective across the course. # 12. Rigidity, Stability, and the Comparison Geometry Toolkit ## Equality Cases in the Main Comparison Theorems This final chapter explains how the comparison theorems proved earlier become rigidity and compactness tools. The prerequisites are Rauch comparison from Chapter 3, Laplacian comparison from Chapter 4, Bonnet--Myers from Chapter 5, Toponogov from Chapter 6, Bishop--Gromov from Chapter 7, and Gromov--Hausdorff convergence from Chapter 11, together with the basic language of geodesics, Jacobi fields, and Ricci curvature. The main goal is to understand three related questions: when equality forces a model geometry, when almost equality forces closeness to a model, and how these estimates supply compactness for later geometric analysis. The recurring question is how much geometry is hidden in equality. Comparison theorems are inequalities against model spaces, but equality often says that the manifold has no freedom left: the relevant Jacobi fields, triangles, or volume annuli must be the model ones. [quotetheorem:5393] [citeproof:5393] This is the local prototype for all later rigidity: equality in a global statement is traced back to equality in a second variation inequality. Each hypothesis is doing visible work. The curvature upper bound is the input that makes the model Jacobi field extremal; if radial curvatures are allowed to exceed $k$, the length inequality can reverse. The no-conjugate-point condition prevents endpoint data from becoming nonunique, since conjugate points allow nonzero Jacobi fields vanishing at both ends. Completeness is included to place the statement in the global comparison setting where the relevant geodesics exist on the prescribed interval; without a global existence hypothesis, an open round hemisphere has the same local curvature as the sphere but radial geodesics can run into the missing boundary before a comparison interval reaches the model endpoint. The theorem also does not say the whole manifold has constant curvature $k$; it only identifies the radial two-planes detected by this particular Jacobi field. The next example shows how a single equality can diagnose a space form direction. [illustration:geometric-analysis-i-model-jacobi-growth] [example: Radial Jacobi Field With Model Growth] Fix $t_0\in(0,r]$ with $J(t_0)\ne 0$. Then $\operatorname{sn}_k(t_0)=|J(t_0)|>0$, so the scalar comparison is being used on an interval where $\operatorname{sn}_k>0$. Since the assumed equality is $|J(t)|=\operatorname{sn}_k(t)$, for every $t$ in this interval we have \begin{align*} \frac{d}{dt}\log |J(t)| &=\frac{d}{dt}\log \operatorname{sn}_k(t)\\ &=\frac{\operatorname{sn}_k'(t)}{\operatorname{sn}_k(t)}\\ &=\operatorname{ct}_k(t). \end{align*} Thus the logarithmic growth of $J$ agrees with the model logarithmic growth at every regular time before $t_0$. Applying the equality case of *Rauch comparison* on $[0,t_0]$ gives \begin{align*} J(t)=\operatorname{sn}_k(t)E(t) \end{align*} for a parallel unit field $E(t)\perp \dot{\gamma}(t)$. Differentiating this identity and using $D_tE=0$ gives \begin{align*} D_tJ(t)&=\operatorname{sn}_k'(t)E(t),\\ D_t^2J(t)&=\operatorname{sn}_k''(t)E(t). \end{align*} By the defining equation $\operatorname{sn}_k''+k\operatorname{sn}_k=0$, this becomes \begin{align*} D_t^2J(t)=-k\operatorname{sn}_k(t)E(t)=-kJ(t). \end{align*} The Jacobi equation is \begin{align*} D_t^2J(t)+R(J(t),\dot{\gamma}(t))\dot{\gamma}(t)=0, \end{align*} so substitution gives \begin{align*} -kJ(t)+R(J(t),\dot{\gamma}(t))\dot{\gamma}(t)=0, \end{align*} hence \begin{align*} R(J(t),\dot{\gamma}(t))\dot{\gamma}(t)=kJ(t). \end{align*} Where $J(t)\ne 0$, write $J(t)=|J(t)|E(t)$. Taking the inner product with $E(t)$ yields \begin{align*} \operatorname{sec}(\dot{\gamma}(t),E(t)) &=\langle R(E(t),\dot{\gamma}(t))\dot{\gamma}(t),E(t)\rangle\\ &=k\langle E(t),E(t)\rangle\\ &=k. \end{align*} Thus equality of the Jacobi-field length with the model function detects exactly the model radial sectional curvature in the plane spanned by $\dot{\gamma}(t)$ and $J(t)$. [/example] The Jacobi-field example is pointwise along one geodesic, so it does not yet explain how equality in a coarse quantity such as volume can determine geometry. This motivates the Bishop-Gromov equality statement, which converts equality of integrated volume ratios into equality of radial Jacobians. [quotetheorem:5394] [citeproof:5394] The volume statement is especially useful because its hypotheses are often measurable or coarse. Completeness ensures that metric balls and minimizing radial geodesics from $p$ are globally available, so Bishop-Gromov monotonicity applies up to the stated radii; a punctured or truncated model ball can have model local geometry while losing the global ball structure needed for the comparison. The Ricci lower bound is essential: without it, volume ratios can match at two radii for reasons unrelated to model geometry. For instance, on a rotationally symmetric warped product with metric $dr^2+f(r)^2g_{S^{n-1}}$, the annular volume is controlled by \begin{align*} \int f(r)^{n-1}\,dr. \end{align*} By changing $f$ above and below the model profile on different subannuli, the same endpoint volume ratio can hold while the radial sectional curvature $-f''/f$ is not the model curvature. The two-radius condition is also doing separate work: equality at one radius is only a normalization, while equality at two radii makes the monotone Bishop-Gromov ratio constant throughout the intervening interval. The conclusion is radial and annular rather than global; cut-locus behaviour and topology outside the chosen annulus are not determined by a single equality of two ratios. When equality holds for every radius from a point, the manifold is forced to look like the model ball from that point until the cut locus intervenes. [example: Diagnosing What Equality In Volume Ratios Means] Assume $\operatorname{Ric}\ge 0$ and \begin{align*} \operatorname{Vol}(B(p,r))=\omega_n r^n \end{align*} for every $0<r<R$. Since the Euclidean model volume is $V_0(r)=\omega_n r^n$, the Bishop--Gromov ratio \begin{align*} \frac{\operatorname{Vol}(B(p,r))}{V_0(r)} = \frac{\omega_n r^n}{\omega_n r^n} = 1 \end{align*} is constant on $(0,R)$. The equality case of *Bishop--Gromov comparison* says that this is not merely a numerical coincidence: the radial volume density has the same scale dependence as in Euclidean space on the regular part of the distance function. The point of the example is diagnostic. Equality in the integrated volume ratio points back to equality in the comparison inputs that produced the monotonicity theorem. In a full rigidity argument, one then analyzes the regular radial geometry, the cut locus, and the angular metric data to decide whether the ball is actually Euclidean. Those extra steps are theorem-level content; for the present course page, the lesson is simply that equality of all Euclidean ball volumes is a much stronger signal than a single sharp volume bound at one radius. [/example] The volume example detects radial rigidity, but many comparison arguments use two geodesic sides and an included angle rather than all geodesics from one center. This motivates the Toponogov equality principle, where the rigid object is a comparison triangle. [remark: Quoted result: Toponogov Equality Principle] Let $(M,g)$ be complete with $\operatorname{sec}\ge k$. Let $\alpha:[0,a]\to M$ and $\beta:[0,b]\to M$ be unit-speed minimizing geodesics with common initial point $p$, initial angle $\theta\in(0,\pi)$, and endpoints $x=\alpha(a)$, $y=\beta(b)$. Assume the geodesic segment $\sigma$ from $x$ to $y$ is minimizing with length $c=d(x,y)$, and let $\bar c$ be the opposite side length of the comparison hinge in the simply connected space form $\bar M_k$. Toponogov gives $c\le \bar c$. If $c=\bar c$ and the geodesic triangle admits a smooth geodesic filling by minimizing segments from $p$ to points of $\sigma$ whose interior avoids the cut locus of $p$, then that filled geodesic triangle is isometric to the corresponding model triangle in $\bar M_k$. In particular, on the regular part of the filled triangle the tangent two-plane has sectional curvature $k$ and the second fundamental form in $M$ vanishes. [/remark] Toponogov turns metric equality into two-dimensional model geometry, but it is sensitive to each hypothesis. If the sides are not minimizing, the side lengths no longer encode the hinge geometry; if the curvature lower bound is removed, thin triangles can imitate the model endpoint equality without having model curvature; if the interior of the variation crosses the cut locus, equality of the three side lengths need not give a smooth ruled surface. The theorem therefore gives a local sector rigidity statement, not a classification of the whole manifold. The next rigidity theorem uses the same philosophy at the largest possible scale: if the Bonnet-Myers diameter bound is sharp, then every radial comparison from a pair of antipodal points must be sharp. [illustration:geometric-analysis-i-hinge-equality] Bonnet-Myers says that a positive Ricci lower bound makes the whole manifold fit inside a sphere-sized diameter bound. The rigidity question asks what remains possible when two points achieve that largest allowed separation. The analytic input is Laplacian comparison for distance functions, because it lets the maximum principle detect when all radial mean-curvature inequalities have become equalities. [quotetheorem:5395] [citeproof:5395] This equality statement is the bridge from scalar maximum-principle information to metric rigidity, but it is deliberately local. The connected set $U$ must lie in the regular part of the distance function because $r=d(p,\cdot)$ is not smooth on the cut locus or at $p$, and the displayed Riccati equation is a smooth radial equation only before those singularities. In applications the comparison inequality is often interpreted in the barrier or weak sense across the cut locus, but equality in that weak sense needs extra work before it can be converted into a smooth shape-operator conclusion. The Ricci lower bound is also indispensable. On a warped product end $dr^2+f(r)^2h$, the quantity $\Delta r$ is governed by $(n-1)f'(r)/f(r)$ on the regular radial region; by arranging this trace to match the model at selected radii while allowing the individual radial curvatures to vary, one can imitate the scalar mean-curvature equality without having the model radial geometry. Even under the stated hypotheses, equality on one regular open set does not classify the whole manifold, does not control topology outside the radial sector, and does not say anything across cut-locus branches. The answer to the maximal-diameter question is much stronger than radial equality from one point: maximal separation forces all comparison data to fit together as the round model. [quotetheorem:5396] [citeproof:5396] The maximal-diameter theorem is sharp in three separate ways. Completeness is required because Bonnet-Myers and the global maximum-principle argument need geodesics and distance barriers to exist at the relevant endpoints; an open hemisphere has the same local round curvature as the sphere but is not a complete maximal-diameter model. The normalization $\operatorname{Ric}\ge (n-1)k>0$ fixes the comparison radius: rescaling the metric changes both the Ricci lower bound and the diameter, so the statement is not a scale-free assertion about having large diameter. The equality $\operatorname{diam}(M)=\pi/\sqrt{k}$ is also exact. Spherical space forms and round real projective space satisfy the same Ricci normalization but have smaller diameter, and nearly maximal diameter gives stability rather than an isometric classification. The theorem also does not say that any manifold with two far-apart points is round, nor does it classify manifolds whose diameter is merely close to the bound. The proof uses equality in the barrier Laplacian comparison for the two antipodal distance functions; if the diameter is below the bound, even by a small amount, those equalities become estimates and singular or nonsmooth limits can appear. The splitting theorem is the rigidity statement for nonpositive global behaviour under nonnegative Ricci curvature. The equality object is a line, meaning a geodesic that minimizes between any two of its points. [quotetheorem:2767] [citeproof:2767] The existence of a line is the exact equality hypothesis: a ray alone is not enough, since Euclidean cones and many noncompact manifolds contain minimizing rays without splitting off an $\mathbb R$ factor. Completeness is needed so that Busemann functions and their gradient flow exist globally. A concrete obstruction is the Euclidean strip $(0,1)\times\mathbb R$ with the flat metric. It has $\operatorname{Ric}=0$ and contains the line $t\mapsto (1/2,t)$, but it is not covered by the theorem because it is incomplete; the only visible splitting would use the incomplete factor $(0,1)$. The Ricci lower bound is the analytic input for the Laplacian comparison and maximum principle; with negative Ricci curvature, hyperbolic space contains complete geodesics but has no product splitting. The theorem also does not assert that $N$ is compact or flat. It extracts only the direction certified by the line, and this is the model for the quantitative almost-line theorem below. The equality cases above share a common pattern: a comparison inequality has a nonnegative deficit, and rigidity follows when the deficit vanishes. This viewpoint prepares the almost-rigidity statements, where the deficit is small rather than zero. ## Almost Rigidity and Stability Exact equality is rare in geometric analysis, so the practical question is what a small comparison deficit controls. The general philosophy is that small deficits should force the space to be close to a model space in a topology appropriate to the estimate, often Gromov-Hausdorff distance or measured Gromov-Hausdorff distance. [definition: Comparison Deficit] A comparison deficit is a nonnegative quantity measuring the failure of a geometric inequality to be an equality, usually normalized so that it is invariant under the natural scaling of the problem. [/definition] For example, in Bishop-Gromov comparison under $\operatorname{Ric}\ge 0$, the quantity \begin{align*} 1-\frac{\operatorname{Vol}(B(p,r))}{\omega_n r^n} \end{align*} is a normalized volume deficit for a ball centered at $p$. To make this philosophy precise in a familiar setting, we record the standard almost-maximal-diameter principle in the form used later in geometric analysis. [remark: Quoted result: Almost Maximal Diameter Principle for Ricci Limit Spaces] Let $n \geq 2$, and let $(M_i^n,g_i)$ be complete connected $n$-dimensional Riemannian manifolds with \begin{align*} \operatorname{Ric}_{g_i} \geq (n-1)g_i \end{align*} and \begin{align*} \operatorname{diam}(M_i,d_{g_i})\to \pi. \end{align*} After passing to a Gromov-Hausdorff convergent subsequence, the limit is a compact Ricci-limit space of diameter $\pi$ and is a spherical suspension \begin{align*} [0,\pi]\times_{\sin}Y \end{align*} over a compact length space $Y$. Additional noncollapsing or regularity hypotheses are needed to identify the limit with the round sphere $S^n$. [/remark] Almost maximal diameter is a stability version of Bonnet-Myers, and each normalization is part of the statement. The fixed dimension prevents escape through a sequence of unrelated model dimensions: the round spheres $S^{n_i}$ with $n_i\to\infty$ satisfy their own normalized Ricci bounds but have no fixed finite-dimensional spherical target. The lower bound $\operatorname{Ric}\ge (n-1)g$ fixes the radius of the model sphere; a round sphere of radius $a$ has a different Ricci normalization and diameter $\pi a$. Completeness is needed for Bonnet-Myers and compactness, so a punctured round sphere is not covered even though it has the same local curvature. The diameter hypothesis is also essential: round real projective space $\mathbb{RP}^n$ has $\operatorname{Ric}=(n-1)g$ and diameter $\pi/2$, but it is not Gromov-Hausdorff close to $S^n$. In applications, the conclusion is less about identifying a manifold exactly and more about ruling out uncontrolled degeneration. [example: Near Spherical Geometry From Almost Maximal Diameter] Let $L_i=d_{g_i}(p_i,q_i)$ and define the excess function \begin{align*} e_i(x)=d_{g_i}(p_i,x)+d_{g_i}(x,q_i)-L_i. \end{align*} The triangle inequality gives $e_i(x)\ge 0$ for every $x\in M_i$. Since $\operatorname{Ric}_{g_i}\ge (n-1)g_i$, Bonnet--Myers gives $\operatorname{diam}(M_i)\le \pi$, so $L_i\le \pi$; the hypothesis $L_i\to \pi$ says that the diameter deficit \begin{align*} \delta_i=\pi-L_i \end{align*} satisfies $\delta_i\to 0$. For any point $x$ with $e_i(x)\le \eta_i$, we have \begin{align*} d_{g_i}(p_i,x)+d_{g_i}(x,q_i) &=L_i+e_i(x)\\ &=\pi-\delta_i+e_i(x). \end{align*} Hence \begin{align*} -\delta_i \le d_{g_i}(p_i,x)+d_{g_i}(x,q_i)-\pi \le \eta_i, \end{align*} so if both $\delta_i$ and $\eta_i$ are small, the two distance functions nearly satisfy the spherical suspension identity \begin{align*} d(p,x)+d(x,q)=\pi. \end{align*} The geometric meaning of small excess is that the broken path from $p_i$ to $q_i$ through $x$ has almost the same length as a minimizing geodesic from $p_i$ to $q_i$. Indeed, if $\sigma_i:[0,L_i]\to M_i$ is a minimizing geodesic from $p_i$ to $q_i$ and $s=d_{g_i}(p_i,x)$, then the exact equality case $e_i(x)=0$ would force $x=\sigma_i(s)$ along the geodesic segment. The almost maximal diameter argument uses the same quantity in the limit: as $\delta_i\to 0$ and the relevant excess becomes small on large subsets, the manifolds become Gromov-Hausdorff close to a spherical suspension, with $p_i$ and $q_i$ playing the roles of the two suspension tips. [/example] The near-spherical example uses almost antipodal points; product stability uses almost collinear points instead. This motivates the almost splitting theorem, where a small excess function replaces the exact line in Cheeger-Gromoll splitting. [remark: Quoted result: Almost Splitting Theorem] For every $n\in\mathbb N$, $R>1$, and $\varepsilon>0$, there exists $\delta=\delta(n,R,\varepsilon)>0$ with the following property. Let $(M^n,g)$ be complete, let $x\in M$, and suppose \begin{align*} \operatorname{Ric}_g \ge -(n-1)\delta \end{align*} on $B(x,R)$. Suppose there are points $p,q\in M$ with $d(p,x),d(q,x)\ge R$ such that the excess at $x$ satisfies \begin{align*} e_{p,q}(x)=d(p,x)+d(x,q)-d(p,q)\le \delta. \end{align*} Then there is a pointed metric space $(Y,y)$ such that \begin{align*} d_{GH}\bigl(B(x,1), B((y,0),1)\subset Y\times \mathbb R\bigr)<\varepsilon, \end{align*} where $Y\times \mathbb R$ carries the product metric. After rescaling, the same statement holds on any fixed smaller ball once the Ricci lower bound, distances to $p,q$, and excess are scaled accordingly. [/remark] The course uses this theorem as the quantitative analogue of Cheeger-Gromoll splitting. The exact Busemann functions from the rigidity proof are replaced by harmonic replacements with controlled Hessian, and the small excess assumption supplies the approximate affine coordinate in the $\mathbb R$ direction. The hypotheses rule out specific ways product structure can fail. Without a Ricci lower bound, negatively curved necks can make a long geodesic pass through $x$ while nearby transverse balls expand or pinch in a non-product way. If $p$ and $q$ are not far from $x$ compared with the ball being studied, a short geodesic segment gives no stable direction across $B(x,1)$. If the excess is not small, the vertex of a metric cone is the model obstruction: two endpoints on different rays can be far away, but the broken path through the vertex has a visible angle defect and no product neighbourhood. The statement does not assert smooth product structure or identify the transverse factor as a manifold; under only Ricci lower bounds the natural conclusion is metric closeness. [example: Near Product Behaviour From Almost Lines] Consider a sequence $(M_i^n,g_i,p_i)$ with \begin{align*} \operatorname{Ric}_{g_i}\ge -\varepsilon_i g_i, \qquad \varepsilon_i\to 0, \end{align*} and suppose there are points $a_i,b_i$ with $d_{g_i}(a_i,p_i)$ and $d_{g_i}(b_i,p_i)$ tending to infinity, while the excess at $p_i$ satisfies \begin{align*} e_i(p_i) &=d_{g_i}(a_i,p_i)+d_{g_i}(p_i,b_i)-d_{g_i}(a_i,b_i)\\ &\to 0. \end{align*} The triangle inequality gives \begin{align*} d_{g_i}(a_i,b_i) \le d_{g_i}(a_i,p_i)+d_{g_i}(p_i,b_i), \end{align*} so $e_i(p_i)\ge 0$. Thus the assumption says that the broken path from $a_i$ to $b_i$ through $p_i$ has length only $e_i(p_i)$ more than the minimizing distance between $a_i$ and $b_i$. Fix a radius $R_0>0$. Since $d_{g_i}(a_i,p_i)\to\infty$ and $d_{g_i}(b_i,p_i)\to\infty$, for all sufficiently large $i$ the two endpoints lie outside $B(p_i,R)$ for any fixed scale $R>R_0$. Also, after replacing $\varepsilon_i$ by a scale-dependent bound, the Ricci inequality has the form required by the *Almost Splitting Theorem* on $B(p_i,R)$. Since the excess tends to $0$, the theorem applies with parameters tending to the equality case and gives pointed metric spaces $(Y_i,y_i)$ such that \begin{align*} d_{GH}\bigl(B(p_i,R_0),B((y_i,0),R_0)\subset Y_i\times\mathbb R\bigr)\to 0. \end{align*} In particular, for every fixed observation radius, the geometry near $p_i$ becomes arbitrarily close to a product ball. The excess condition is the quantitative replacement for lying on a line: if $e_i(p_i)=0$, then \begin{align*} d_{g_i}(a_i,b_i) = d_{g_i}(a_i,p_i)+d_{g_i}(p_i,b_i), \end{align*} so $p_i$ lies on a minimizing geodesic from $a_i$ to $b_i$. When the same equality persists at larger and larger scales, the limiting geodesic is a line; when the excess merely tends to $0$, the almost splitting theorem produces an approximate $\mathbb R$-factor instead. Thus the long almost minimizing segment from $a_i$ to $b_i$ supplies the product direction, and the spaces $Y_i$ record the transverse geometry. [/example] The almost-line example explains product formation, while volume rigidity explains cone formation. This motivates the next bridge theorem: equality in volume ratios on a Ricci limit space should determine the metric cone structure, not only the measure growth. [remark: Quoted result: Volume Cone Implies Metric Cone] Let $(X,d,m,p)$ be the pointed measured Gromov-Hausdorff limit of a noncollapsed sequence of pointed complete $n$-dimensional Riemannian manifolds $(M_i,g_i,p_i)$ with \begin{align*} \operatorname{Ric}_{g_i}\ge -(n-1)\varepsilon_i g_i, \qquad \varepsilon_i\to 0, \end{align*} where noncollapsed means $\operatorname{Vol}_{g_i}(B(p_i,1))\ge v>0$ for some fixed $v$, and where $m$ is the renormalized limit measure obtained from the Riemannian measures, for instance after normalizing $m(B(p,1))=1$ when $B(p,1)$ is in the pointed region under consideration. Assume $0<r<R$ and the closed annulus $\overline{A}(p;r,R)$ is compact in $X$. If the Euclidean normalized volume ratio satisfies \begin{align*} \frac{m(B(p,R))}{R^n}=\frac{m(B(p,r))}{r^n}, \end{align*} where $B(p,s)$ denotes the metric ball in $X$ and the equality is interpreted for the Bishop-Gromov limit measure $m$, then the equality case of Bishop-Gromov holds for every scale $s\in[r,R]$. Consequently there is a compact length space $Z$ such that the open annulus \begin{align*} A(p;r,R)=\{x\in X:r<d(p,x)<R\} \end{align*} is isometric, as an intrinsic length space with the restricted distance induced by $X$, to the cone annulus $(r,R)\times Z$ with metric $ds^2+s^2d_Z^2$. Equivalently, every compact subannulus $A(p;s_1,s_2)$ with $r<s_1<s_2<R$ has radial distances, annular cross-sections, and homothetic scaling as in the corresponding subannulus of the metric cone $C(Z)$. The theorem does not assert a smooth polar-coordinate description, and it does not identify behaviour at the endpoints $r$ and $R$ beyond the intrinsic annular conclusion. [/remark] This theorem is the metric-measure version of Bishop-Gromov equality. The noncollapsed hypothesis prevents lower-dimensional measured cones from entering the conclusion: collapsing flat tori, for instance, can have controlled Ricci curvature while their limits lose dimension, so Euclidean volume normalization no longer detects an $n$-dimensional cone. The Ricci lower bound is also essential because it supplies Bishop-Gromov monotonicity; without it, a warped product annulus can be arranged to have matching endpoint volume ratios while its radial metric is not conical between those endpoints. Equality between the two endpoint radii is enough here because monotonicity fills the whole interval $[r,R]$; equality at a single radius, with no second scale for comparison, would not determine an annulus. The conclusion is local to the annulus: it identifies the conical metric structure between radii $r$ and $R$, but it does not determine the geometry beyond those radii or require $X$ to be a smooth manifold. [illustration:geometric-analysis-i-metric-cone-annuli] [example: Detecting a Cone From Volume Ratios] Suppose $(M_i^n,g_i,p_i)$ converges noncollapsingly to $(X,d,m,p)$, and assume that for some $0<r<R$ the Euclidean normalized Bishop--Gromov ratio satisfies \begin{align*} \frac{m(B(p,R))}{R^n} = \frac{m(B(p,r))}{r^n}. \end{align*} Write \begin{align*} \Phi(s)=\frac{m(B(p,s))}{s^n}. \end{align*} Bishop--Gromov monotonicity on the Ricci-limit space says that $\Phi$ is nonincreasing in $s$. Hence for every $s$ with $r<s<R$, \begin{align*} \Phi(R)\le \Phi(s)\le \Phi(r). \end{align*} The endpoint equality gives $\Phi(R)=\Phi(r)$, so the preceding inequalities become \begin{align*} \Phi(r)\le \Phi(s)\le \Phi(r), \end{align*} and therefore \begin{align*} \Phi(s)=\Phi(r)=\Phi(R) \end{align*} for every $s\in[r,R]$. Thus the volume ratio is constant at every intermediate radius, not only at the two endpoints. By the equality case in *Volume Cone Implies Metric Cone*, this constant volume ratio forces the open annulus \begin{align*} A(p;r,R)=\{x\in X:r<d(p,x)<R\} \end{align*} to be isometric, as an intrinsic length space, to a cone annulus \begin{align*} (r,R)\times Z \end{align*} with metric \begin{align*} ds^2+s^2d_Z^2 \end{align*} for some compact length space $Z$. Under this identification, the distance from the cone vertex is exactly the coordinate $s$, the cross-section at radius $s$ is $\{s\}\times Z$, and changing from radius $s_1$ to radius $s_2$ rescales transverse distances by \begin{align*} d_{\{s_2\}\times Z}((s_2,z_1),(s_2,z_2)) = \frac{s_2}{s_1} d_{\{s_1\}\times Z}((s_1,z_1),(s_1,z_2)). \end{align*} Thus equality of the normalized volume ratios detects an actual metric cone structure on the annulus: radial geodesics, cross-sections, and homothetic scaling all behave as they do in a cone. [/example] Almost-rigidity statements should be read as compactness arguments with a contradiction built in. If a desired stability conclusion failed, a sequence with deficits tending to zero would converge to a limit with equality, and the rigidity theorem for the limit would contradict the assumed failure. ## Compactness and A Priori Estimates for Geometric Analysis The last question is why comparison geometry is a toolkit rather than a collection of isolated theorems. Later geometric analysis often studies sequences of manifolds, maps, submanifolds, or metrics; comparison estimates provide the uniform bounds needed to extract subsequences and control limiting behaviour. [definition: Noncollapsing] A class of pointed Riemannian manifolds $(M_i,g_i,p_i)$ is noncollapsed at scale $r_0>0$ if there is $v>0$ such that \begin{align*} \operatorname{Vol}_{g_i}(B(p_i,r_0))\ge v \end{align*} for every $i$. [/definition] Noncollapsing prevents lower-dimensional limits from appearing under Gromov-Hausdorff convergence. Without it, flat tori with one circle factor shrinking to length $0$ converge to a lower-dimensional flat torus, even though curvature and diameter remain controlled. Metric precompactness, however, is a separate statement: it needs uniform covering numbers, not a lower bound on the limiting dimension. The compactness material from Chapter 11 should be read in this division. The compact metric-space criterion says that uniform covering numbers plus a diameter bound give Gromov-Hausdorff precompactness. The noncollapsed Ricci theorem supplies those covering numbers when a Ricci lower bound, diameter bound, and volume lower bound are available. Sectional curvature lower bounds give a related covering theory in Alexandrov-style compactness. Thus curvature and diameter estimates become covering estimates, and covering estimates become subsequential metric limits. A separate noncollapsing assumption is required when the intended conclusion remembers the original $n$-dimensional volume measure rather than only the underlying compact metric space. None of these compactness statements gives smooth convergence by itself, since curvature may concentrate or injectivity radius may collapse without stronger hypotheses. [example: Compactness With Uniform Curvature and Diameter Bounds] Let $(M_i^n,g_i)$ be compact and suppose \begin{align*} |\operatorname{sec}_{g_i}|\le K,\qquad \operatorname{diam}(M_i)\le D,\qquad \operatorname{Vol}_{g_i}(M_i)\ge v>0. \end{align*} For a unit vector $u\in T_xM_i$ and an orthonormal basis $u,e_2,\dots,e_n$, the Ricci curvature in the $u$-direction is \begin{align*} \operatorname{Ric}_{g_i}(u,u) &=\sum_{a=2}^n \langle R(e_a,u)u,e_a\rangle\\ &=\sum_{a=2}^n \operatorname{sec}_{g_i}(u,e_a). \end{align*} Since $|\operatorname{sec}_{g_i}|\le K$, each summand satisfies $\operatorname{sec}_{g_i}(u,e_a)\ge -K$, so \begin{align*} \operatorname{Ric}_{g_i}(u,u) \ge \sum_{a=2}^n (-K) = -(n-1)K. \end{align*} Thus every $M_i$ satisfies the Ricci lower bound \begin{align*} \operatorname{Ric}_{g_i}\ge (n-1)(-K)g_i. \end{align*} Applying the compactness criterion from Chapter 11 with the Ricci lower bound $\operatorname{Ric}_{g_i}\ge -(n-1)K g_i$ and the common diameter bound $D$, for each $\varepsilon>0$ there is a number $N(n,K,D,\varepsilon)$ such that every $M_i$ is covered by at most $N(n,K,D,\varepsilon)$ balls of radius $\varepsilon$. These uniform covering numbers are exactly the hypothesis in Gromov's compactness criterion for compact metric spaces, so a subsequence of $(M_i,d_{g_i})$ converges in the Gromov-Hausdorff topology. The lower volume bound is the noncollapsing input: \begin{align*} \operatorname{Vol}_{g_i}(M_i)\ge v>0. \end{align*} Together with the two-sided sectional curvature bound, it rules out the standard collapse mechanism in which a direction shrinks to zero size while curvature and diameter remain bounded. If one also has a uniform injectivity-radius bound \begin{align*} \operatorname{inj}(M_i,g_i)\ge i_0>0, \end{align*} then *Cheeger compactness* upgrades the metric subsequential convergence to smooth subsequential convergence after passing to diffeomorphisms on a fixed smooth manifold. Thus the comparison bounds alone give Gromov-Hausdorff compactness, while the extra injectivity-radius control is what turns the limit from a metric limit into a smooth Riemannian limit. [/example] Comparison estimates also give bounds for functions naturally associated to geometry. Distance functions, heat kernels, harmonic replacements, and eigenfunctions all use volume doubling, Poincare-type inequalities, and Laplacian comparison as input estimates. [remark: Role of Laplacian Comparison] Laplacian comparison controls $\Delta r$ for the distance function $r=d(p,\cdot)$ away from the cut locus, usually in a weak or barrier sense across the cut locus. This estimate is the analytic form of radial volume comparison: it bounds the mean curvature of geodesic spheres and enters maximum-principle arguments for Busemann functions, excess functions, and harmonic replacements. [/remark] The following summary records the main toolkit that later courses use without reproving the full comparison theory. [explanation: Comparison Geometry Toolkit] Rauch comparison turns sectional curvature bounds into control of the differential of the exponential map. It is the local mechanism behind conjugate-point estimates, injectivity-radius arguments, and the rigidity of radial coordinates. Toponogov comparison turns sectional curvature lower bounds into metric triangle control. It is the source of convexity, diameter rigidity, and many arguments where geodesic triangles replace coordinate computations. Bishop-Gromov comparison turns Ricci lower bounds into monotonicity of volume ratios. It supplies volume doubling, noncollapsing propagation, compactness criteria, and the volume-rigidity input for cone theorems. Splitting and almost splitting convert the existence of lines or almost lines into product structure. This is the bridge from smooth comparison geometry to the structure theory of Ricci limit spaces. [/explanation] The chapter closes the course by reframing the earlier theorems as reusable estimates. Exact rigidity identifies the model spaces; almost rigidity says that small errors force closeness to those models; compactness packages these facts into subsequential convergence and a priori control. This is why comparison geometry is a starting point for geometric analysis: it supplies geometric bounds strong enough to survive limits, weak formulations, and variational arguments. ## Connections and Further Directions Comparison geometry feeds directly into Ricci flow, metric-measure geometry, and the analysis of spaces with synthetic curvature bounds. Ricci flow uses the same distance, volume, and splitting estimates to control singularity models and compactness of evolving metrics. Metric-measure geometry keeps Bishop-Gromov type monotonicity and Poincare inequalities while weakening smoothness, so the comparison theorems become stability tools for limits rather than only pointwise smooth statements. The subject also connects to minimal submanifold theory and nonlinear elliptic PDE. Laplacian comparison and volume doubling support gradient estimates, heat-kernel bounds, and Sobolev inequalities, while convexity and splitting arguments help identify equality cases. In global Riemannian geometry, the same toolkit underlies finiteness theorems, rigidity theorems, and the study of manifolds whose curvature is almost, but not exactly, modeled on a space form. On Androma, the most useful companion results are the core geodesic and comparison theorems: Hopf-Rinow for completeness, second variation and Sturm comparison for Jacobi-field estimates, Bonnet-Myers for positive Ricci curvature, and Cheeger-Gromoll splitting for the product structure forced by lines. The course's next natural direction is to combine these comparison estimates with analytic compactness methods, especially in Ricci limit spaces and geometric flows. ## References - Androma, [Hopf-Rinow Theorem](/theorems/2726). - Androma, [Second Variation Formula](/theorems/2729). - Androma, [Sturm Comparison Theorem](/theorems/3510). - Androma, [Bonnet-Myers Diameter Theorem](/theorems/2734). - Androma, [Cheeger-Gromoll Line-Splitting Theorem](/theorems/2767). - Jeff Cheeger and David Ebin, *Comparison Theorems in Riemannian Geometry*, AMS Chelsea. - Peter Petersen, *Riemannian Geometry*, Springer. - John M. Lee, *Riemannian Manifolds: An Introduction to Curvature*, Springer.