What brings you to Androma?

Riemannian geometry studies smooth manifolds equipped with a metric—an inner product structure on each tangent space that measures distances and angles intrinsically. Unlike extrinsic geometry, where a surface is embedded in ambient space, Riemannian geometry examines manifolds from within, developing tools to understand their shape and structure through purely intrinsic data. This course explores how a Riemannian metric gives rise to fundamental objects: the Riemann curvature tensor, which encodes infinitesimal bending; geodesics, which generalize straight lines as locally shortest paths; and the deep interplay between local curvature and global topological properties. The course progresses from foundations to increasingly sophisticated structures, revealing how curvature acts as the bridge between local and global geometry. After establishing basics of Riemannian manifolds, we study Riemann curvature in detail—understanding the Riemann tensor, sectional curvature, and Ricci curvature as different lenses on the same intrinsic information. Geodesics emerge as the curves that respect the metric, enabling us to define distance and explore the geometry at large scales. The course then shifts to analysis on Riemannian manifolds through Hodge theory, which reveals deep connections between differential forms, harmonic analysis, and topology via the Laplacian operator. The final chapters introduce algebraic and structural perspectives. Riemannian holonomy groups—the symmetry groups acting on tangent spaces along parallel transport—provide an algebraic window into the manifold's curvature constraints. The Cheeger–Gromoll splitting theorem then exemplifies global rigidity: under conditions on Ricci curvature, a manifold's topology must split into a product, showing how curvature growth strictly constrains what shapes are geometrically possible. Together, these chapters demonstrate that Riemannian geometry unifies local differential-theoretic analysis with global topological consequences through the universal language of curvature. # 1. Basics of Riemannian manifolds This chapter reviews the foundational material on Riemannian manifolds that underpins the rest of the course. We fix index conventions, recall the definition of a Riemannian metric and its local coordinate expression, establish that every smooth manifold carries a metric via Whitney's embedding theorem, and introduce the two central objects that make a metric geometrically effective: the notion of isometry as the correct notion of equivalence, and the Levi-Civita connection as the canonical way to differentiate tensor fields on a Riemannian manifold. The Christoffel symbols computed here reappear in Chapter 2 to evaluate the curvature of $S^2$, and the Levi-Civita connection is the foundation for geodesic theory (Chapter 3), Hodge theory (Chapter 4), holonomy (Chapter 5), and the splitting theorem (Chapter 6). ## Index Conventions and Einstein Summation Without a consistent bookkeeping convention, tensor expressions become ambiguous: the notation $A_i B_i$ could mean a contraction (applying a covector to a vector, an invariant operation) or a mere coordinate-wise product (which changes under a change of basis). In high-dimensional computations the resulting sign errors and missing factors of the metric are notoriously hard to trace. The up/down index convention fixes this: it makes the geometrically meaningful operation of contraction visually explicit, so that any formula where an index appears in the wrong position is immediately recognisable as erroneous. Given a chart $(U, \varphi)$ with local coordinates $\{x^k\}$, a vector field $X$ on $U$ is written \begin{align*} X &= \sum_k X^k\, \partial_k, \qquad \partial_k := \frac{\partial}{\partial x^k}. \end{align*} Upper indices label **contravariant** components (vector fields, tangent vectors), lower indices label **covariant** components (covector fields, $1$-forms). A general tensor field $T \in \Gamma(\mathcal{T}^{q,p}M)$, where $\mathcal{T}^{q,p}M := (TM)^{\otimes q} \otimes (T^*M)^{\otimes p}$, is written in local coordinates as \begin{align*} T &= T^{k_1 \cdots k_q}_{\ell_1 \cdots \ell_p}\; \partial_{k_1} \otimes \cdots \otimes \partial_{k_q} \otimes dx^{\ell_1} \otimes \cdots \otimes dx^{\ell_p}, \end{align*} where the tensor product symbols are often omitted. Note the deliberate choice of writing basis cotangent vectors as $dx^k$ (with an upper index), so that in an expression such as $A_k\, dx^k$ the dummy index $k$ appears once up and once down. **Einstein summation convention.** Whenever an index appears once as a superscript and once as a subscript in the same expression, a summation over that index is implied. For example, \begin{align*} T^i{}_{jk} S^{jk}{}_i \end{align*} means $\sum_{i,j,k} T^i{}_{jk} S^{jk}{}_i$. We sum an upper index against a lower index because this corresponds to the invariant operation of applying a covector to a vector. We will write out summations explicitly whenever we deviate from this convention. We also adopt the shorthand $\nabla_k := \nabla_{\partial_k}$ for the covariant derivative along the $k$-th coordinate vector field. ## Riemannian Metrics A smooth manifold has no intrinsic notion of distance, angle, or length — these are additional structures that must be imposed. The Riemannian metric supplies all of them at once by equipping each tangent space with an inner product that varies smoothly from point to point. [definition: Riemannian Metric] Let $M$ be a smooth manifold. A **Riemannian metric** $g$ on $M$ is a smooth global section of $T^*M \otimes T^*M$ that is fiberwise symmetric and positive definite. Concretely, for each $p \in M$, $g_p : T_pM \times T_pM \to \mathbb{R}$ is an inner product, and the assignment $p \mapsto g_p$ is smooth. The pair $(M, g)$ is called a **Riemannian manifold**. [/definition] In a chart $(U, \varphi)$ with coordinates $(x_1, \ldots, x_n)$, the metric restricts to \begin{align*} g\big|_U &= g_{ij}(x)\, dx^i\, dx^j, \end{align*} where the coefficient functions $g_{ij} = g(\partial_i, \partial_j)$ are $C^\infty$ on $U$. Symmetry forces $g_{ij} = g_{ji}$, and positive definiteness forces the matrix $(g_{ij}(x))$ to be positive definite at every point $x \in \varphi(U)$. [example: Euclidean Space] The manifold $\mathbb{R}^n$ carries the **Euclidean metric**: in standard coordinates, $g_{ij} = \delta_{ij}$. This means $g(u, v) = \sum_i u_i v_i$ for tangent vectors $u = (u_1, \ldots, u_n)$ and $v = (v_1, \ldots, v_n)$ at any point. The resulting notion of distance is the ordinary Euclidean distance $d(x, y) = |x - y|$. [/example] ### Existence of Riemannian Metrics The Euclidean example raises a natural question: does every smooth manifold admit a Riemannian metric? The answer is yes, and the proof uses a classical embedding theorem. [quotetheorem:1519] Given such an embedding $\iota: M \hookrightarrow \mathbb{R}^k$, each tangent space $T_pM$ injects into $T_p\mathbb{R}^k \cong \mathbb{R}^k$, and we can restrict the standard inner product of $\mathbb{R}^k$ to $T_pM$. This induced metric is called the **pullback metric** $g = \iota^*g_{\mathbb{R}^k}$. A few remarks are in order about the sharpness and scope of the Whitney theorem. The bound $k \leq 2\dim M$ on the ambient dimension is optimal: there exist smooth manifolds of dimension $n$ that cannot be embedded into $\mathbb{R}^{2n-1}$. The properness hypothesis is essential — a non-proper embedding can fail to be a closed submanifold (the image might accumulate on itself), and then the pullback metric may not extend smoothly. The theorem also depends crucially on the manifold being Hausdorff and second-countable: without Hausdorff separation, distinct points need not have disjoint neighbourhoods and the embedding cannot be injective on a global scale; without second-countability, the partition-of-unity argument used in the proof breaks down. For example, the long line is a non-second-countable $1$-manifold that does not embed in any $\mathbb{R}^k$. The same construction applies in a much more general setting: [quotetheorem:2702] The immersion condition — that $dF_p$ is injective at every $p$ — is precisely what ensures $g_p$ is positive definite (not merely positive semidefinite): $g_p(u, u) = h_{F(p)}(dF_p\, u, dF_p\, u) = 0$ implies $dF_p\, u = 0$, hence $u = 0$. It is important to understand what this theorem does *not* say. A smooth immersion $F: M \to N$ pulling back a metric is not in general an isometric embedding — distances in $(M, g)$ and in $(N, h)$ can differ substantially. For instance, the standard inclusion of $S^1$ into $\mathbb{R}^2$ is an isometric embedding, but an immersion that winds $S^1$ around twice in $\mathbb{R}^2$ pulls back the Euclidean metric to a metric on $S^1$ whose total length is $4\pi$ rather than $2\pi$. The pullback metric records only the infinitesimal geometry seen by $dF$. This construction will reappear when we study induced metrics on submanifolds: a submanifold $S \subset M$ always carries the induced (pullback) metric from the inclusion, and the interplay between intrinsic geometry of $S$ and the ambient geometry of $M$ is a central theme of Riemannian geometry. Combining the two results: every smooth manifold admits a Riemannian metric, obtained by composing the Whitney embedding with the pullback construction. For non-compact manifolds, a partition-of-unity argument gives the same conclusion directly. ## Isometries On a smooth manifold without additional structure, the natural notion of equivalence is diffeomorphism. Once a Riemannian metric is present, the metric itself must be preserved. [definition: Isometry] Let $(M, g)$ and $(N, h)$ be Riemannian manifolds. A smooth map $f: M \to N$ is an **isometry** if it is a diffeomorphism and $f^*h = g$. Explicitly, for every $p \in M$ and every $u, v \in T_pM$, \begin{align*} h_{f(p)}\!\left(df_p\, u,\, df_p\, v\right) &= g_p(u, v). \end{align*} [/definition] The condition $f^*h = g$ says that $f$ preserves lengths and angles of tangent vectors, hence distances along curves. In particular, the length of any smooth curve $\gamma: [a,b] \to M$, given by $\int_a^b |\dot\gamma(t)|_g\, dt$, equals the length of its image under $f$. Note that isometry is strictly stronger than a mere bijection that preserves distances in the sense of metric spaces: we additionally require $f$ to be a diffeomorphism, so that $f^{-1}$ is also smooth. When $M = N$, an isometry from $(M, g)$ to itself is called an **isometry of $(M, g)$**; these form the **isometry group** $\mathrm{Isom}(M, g)$. [example: Left-Invariant Metrics on Lie Groups] Let $G$ be a Lie group. For any $x \in G$, the left and right translations $L_x, R_x: G \to G$ defined by $L_x(y) = xy$ and $R_x(y) = yx$ are diffeomorphisms of $G$. Since $G$ admits a Riemannian metric (by Whitney), we can ask for something stronger: does $G$ admit a **left-invariant metric**, i.e., a metric for which every $L_x$ is an isometry? Recall that the **Lie algebra** of $G$ is $\mathfrak{g} = T_eG$, where $e$ is the identity. A vector field $X$ on $G$ is **left-invariant** if $d(L_x)_y X_y = X_{xy}$ for all $x, y \in G$, i.e., if it is invariant under pushforward by all left translations. Left-invariant vector fields are in bijection with elements of $\mathfrak{g}$: given $X_e \in \mathfrak{g}$, we set $X_a = d(L_a)_e\, X_e$. Analogously, a left-invariant Riemannian metric on $G$ is constructed by choosing any inner product $\langle \cdot, \cdot \rangle$ on $T_eG = \mathfrak{g}$, and propagating it via left translation: for $u, v \in T_xG$, set \begin{align*} g_x(u, v) &= \left\langle d(L_{x^{-1}})_x\, u,\; d(L_{x^{-1}})_x\, v \right\rangle. \end{align*} Smoothness of $g$ follows from smoothness of the group multiplication. By construction, each $L_x$ satisfies $L_x^*g = g$, so $L_x$ is an isometry. Everything works symmetrically with $R_x$ in place of $L_x$. A metric that is both left- and right-invariant is called **bi-invariant**. Bi-invariant metrics are harder to construct in general. The key result is that every **compact** Lie group admits a bi-invariant metric: start with any left-invariant metric $g$, and average it over all right translations, \begin{align*} \bar{g}_e(u,v) &= \int_G g_e\!\left(d(R_x)_e\, u,\; d(R_x)_e\, v\right) d\mu(x), \end{align*} where $\mu$ is the Haar measure on $G$. Compactness guarantees the integral is finite. The resulting $\bar{g}$ is both left- and right-invariant. For non-compact groups the construction fails; for instance, $\mathrm{SL}(2, \mathbb{R})$ does not admit a bi-invariant metric. [/example] The left-invariant metric construction shows that on a Lie group the entire Riemannian structure is encoded by a single inner product on the Lie algebra $\mathfrak{g}$, and the symmetry group of the metric is at least as large as $G$ acting on itself by left translation. This pattern — a homogeneous space where geometry is fully determined by algebraic data at the identity — recurs throughout differential geometry and is one reason Lie groups serve as the central source of explicit examples. With existence of metrics and the correct notion of equivalence in hand, we now turn to the question of how to differentiate on a Riemannian manifold. ## The Levi-Civita Connection To differentiate vector fields and tensors on a Riemannian manifold, we need a connection — a rule for differentiating sections of a vector bundle. On a general smooth manifold there are infinitely many connections on $TM$. The Riemannian metric singles out a canonical one. [motivation] ### Why connections are needed On a smooth manifold, there is no coordinate-independent way to "subtract" two vectors at different points: the tangent spaces $T_pM$ and $T_qM$ are distinct vector spaces with no natural identification between them. A connection provides exactly such an identification, by specifying how to transport vectors along curves. Without a connection, the derivative of a vector field along a curve is not well-defined. ### Why a unique connection is singled out On a general manifold, the space of connections on $TM$ is an affine space — the difference of two connections is a tensor, so there is no canonical choice. The metric imposes two natural conditions that, together, uniquely determine a connection: 1. **Metric compatibility** ("the connection is compatible with $g$"): inner products of parallel-transported vectors should be preserved. This rules out connections that distort lengths. 2. **Torsion-freeness** ("the connection is symmetric"): this is a condition that roughly says the connection behaves consistently with the coordinate partial derivatives, i.e., it has no "twisting". It is the Riemannian generalisation of the symmetry $\partial_i \partial_j = \partial_j \partial_i$. The fundamental theorem of Riemannian geometry says these two conditions together uniquely characterise the connection. [/motivation] To see why uniqueness is non-trivial, observe that metric compatibility alone is a large constraint but still leaves a family of connections: any torsion-ful connection obtained by adding a skew-symmetric tensor to a metric-compatible one is again metric-compatible. Conversely, torsion-freeness alone has nothing to do with the metric and admits many solutions. The remarkable fact — and the content of the fundamental theorem — is that demanding both properties simultaneously collapses the space of solutions to a single point. [definition: Levi-Civita Connection] Let $(M, g)$ be a Riemannian manifold. The **Levi-Civita connection** is the unique connection $\nabla: \mathfrak{X}(M) \times \mathfrak{X}(M) \to \mathfrak{X}(M)$, written $(Z, Y) \mapsto \nabla_Z Y$, satisfying: 1. **Metric compatibility**: for all vector fields $X, Y, Z \in \mathfrak{X}(M)$, \begin{align*} Z\, g(X, Y) &= g(\nabla_Z X,\, Y) + g(X,\, \nabla_Z Y); \end{align*} 2. **Torsion-freeness**: for all $X, Y \in \mathfrak{X}(M)$, \begin{align*} \nabla_X Y - \nabla_Y X &= [X, Y]. \end{align*} [/definition] Metric compatibility says that $g$ is parallel with respect to $\nabla$. In the compact tensor notation it reads $\nabla g = 0$. Torsion-freeness says the torsion tensor $T(X,Y) := \nabla_X Y - \nabla_Y X - [X,Y]$ vanishes identically. Dropping either condition leads to qualitatively different geometries. If one keeps torsion-freeness but drops metric compatibility, one obtains a class of connections studied in the context of projective differential geometry; these are used in Weyl geometry and certain formulations of general relativity. If one keeps metric compatibility but drops torsion-freeness, the torsion tensor $T$ is a new geometric datum encoding "twisting" of the parallel transport; such connections appear in the **Weitzenböck connection** used in the proof of the Bochner vanishing theorem, and in **Cartan geometry** where torsion encodes translational curvature. In theoretical physics, connections with non-zero torsion (Einstein–Cartan theory) are used to couple fermions to gravity. The uniqueness of the Levi-Civita connection is therefore not just a mathematical nicety — it reflects a specific geometric choice that prioritises the metric structure. [definition: Christoffel Symbols] In local coordinates $(x_1, \ldots, x_n)$, the **Christoffel symbols** $\Gamma_{jk}^i$ of the Levi-Civita connection are defined by \begin{align*} \nabla_{\partial_j} \partial_k &= \Gamma_{jk}^i\, \partial_i. \end{align*} [/definition] Torsion-freeness translates immediately into the symmetry $\Gamma_{jk}^i = \Gamma_{kj}^i$. Metric compatibility gives an explicit formula for the Christoffel symbols in terms of the metric coefficients: \begin{align*} \Gamma_{jk}^i &= \frac{1}{2}\, g^{i\ell}\!\left(\partial_j g_{k\ell} + \partial_k g_{j\ell} - \partial_\ell g_{jk}\right), \end{align*} where $(g^{i\ell})$ denotes the inverse matrix of $(g_{i\ell})$. This formula — sometimes called the **Koszul formula** in coordinate form — shows that the Levi-Civita connection is entirely determined by the metric. [example: Christoffel Symbols on $S^2$ with Round Metric] We compute the Christoffel symbols for the unit $2$-sphere $S^2$ with the round metric in standard spherical coordinates $(x_1, x_2) = (\theta, \phi)$, where $\theta \in (0, \pi)$ is the polar angle and $\phi \in (0, 2\pi)$ is the azimuthal angle. The metric takes the form \begin{align*} g &= d\theta^2 + \sin^2\theta\, d\phi^2, \end{align*} so the metric coefficients are $g_{11} = 1$, $g_{12} = g_{21} = 0$, $g_{22} = \sin^2\theta$, and the inverse is $g^{11} = 1$, $g^{12} = g^{21} = 0$, $g^{22} = \csc^2\theta$. The Koszul formula gives \begin{align*} \Gamma_{jk}^i &= \frac{1}{2}\, g^{i\ell}\left(\partial_j g_{k\ell} + \partial_k g_{j\ell} - \partial_\ell g_{jk}\right). \end{align*} Since $g^{i\ell} = 0$ for $i \neq \ell$, only the diagonal terms $g^{11}$ and $g^{22}$ contribute. We compute each non-zero symbol. **$\Gamma_{22}^1$:** Take $i = 1$, $j = k = 2$. \begin{align*} \Gamma_{22}^1 &= \frac{1}{2}\, g^{11}\left(\partial_2 g_{21} + \partial_2 g_{21} - \partial_1 g_{22}\right) = \frac{1}{2}\cdot 1 \cdot\left(0 + 0 - \partial_\theta(\sin^2\theta)\right) = -\sin\theta\cos\theta. \end{align*} **$\Gamma_{12}^2 = \Gamma_{21}^2$:** Take $i = 2$, $j = 1$, $k = 2$. \begin{align*} \Gamma_{12}^2 &= \frac{1}{2}\, g^{22}\left(\partial_1 g_{22} + \partial_2 g_{12} - \partial_2 g_{12}\right) = \frac{1}{2}\csc^2\theta \cdot \partial_\theta(\sin^2\theta) = \frac{1}{2}\csc^2\theta \cdot 2\sin\theta\cos\theta = \cot\theta. \end{align*} All other Christoffel symbols vanish. The full non-zero list is \begin{align*} \Gamma_{22}^1 &= -\sin\theta\cos\theta, \qquad \Gamma_{12}^2 = \Gamma_{21}^2 = \cot\theta. \end{align*} These reflect the geometry of the sphere: the $\Gamma_{22}^1$ term accounts for the convergence of meridians toward the poles, while $\Gamma_{12}^2 = \cot\theta$ records how the azimuthal basis vector $\partial_\phi$ must be adjusted as $\theta$ changes to remain parallel. [/example] The sphere computation illustrates how the Koszul formula extracts geometric content from the metric coefficients alone. It is also instructive to see what happens when one of the two defining conditions of the Levi-Civita connection is dropped — both to confirm that the conditions are genuinely independent, and to develop intuition for what each one rules out. The next example exhibits a torsion-free connection on $\mathbb{R}^2$ that fails metric compatibility, showing that torsion-freeness alone does not pin down a unique connection adapted to the metric. [example: A Non-Metric-Compatible Connection] To illustrate why metric compatibility is not automatic, consider $\mathbb{R}^2$ with the standard Euclidean metric $g = dx_1^2 + dx_2^2$ (so $g_{ij} = \delta_{ij}$), but define a connection $\tilde\nabla$ by \begin{align*} \tilde\nabla_{\partial_1}\partial_1 &= \partial_2, \qquad \tilde\nabla_{\partial_i}\partial_j = 0 \text{ otherwise.} \end{align*} This is a torsion-free connection (one checks $\tilde\nabla_{\partial_i}\partial_j = \tilde\nabla_{\partial_j}\partial_i$ for all $i, j$). Metric compatibility requires $\nabla g = 0$, which expands to \begin{align*} (\tilde\nabla_{\partial_k} g)(\partial_i, \partial_j) &= \partial_k g(\partial_i, \partial_j) - g(\tilde\nabla_{\partial_k}\partial_i, \partial_j) - g(\partial_i, \tilde\nabla_{\partial_k}\partial_j). \end{align*} For $k = i = 1$, $j = 2$ this evaluates to $0 - g(\partial_2, \partial_2) - g(\partial_1, 0) = -1 \neq 0$. So $\tilde\nabla$ is torsion-free but not metric-compatible, demonstrating that the two conditions are genuinely independent. [/example]  So far the connection has been described purely as an operation on vector fields. To do meaningful Riemannian calculus, however, we need to differentiate arbitrary tensors — covector fields, the metric itself, curvature tensors yet to come. Fortunately, the connection on $TM$ determines a connection on every tensor bundle by a unique extension that respects duality and tensor products. We describe this extension now, and use it to give the metric-compatibility condition its cleanest global formulation. ### Extending the Connection to All Tensor Bundles The Levi-Civita connection is defined on $TM$, but it extends canonically to all tensor bundles. The extension is built in two steps. **Step 1: Extension to $T^*M$.** The connection $\nabla$ on $TM$ induces a unique connection on $T^*M$, also denoted $\nabla$, determined by the Leibniz rule for the natural pairing: for all $\alpha \in \Omega^1(M)$ and $X, Y \in \mathfrak{X}(M)$, \begin{align*} X\langle \alpha, Y \rangle &= \langle \nabla_X \alpha,\, Y \rangle + \langle \alpha,\, \nabla_X Y \rangle. \end{align*} This defines $\nabla_X \alpha$ uniquely since the pairing is non-degenerate. **Step 2: Extension to tensor products.** For vector bundles $E$ and $F$ with connections $\nabla^E$ and $\nabla^F$, the product rule \begin{align*} \nabla^{E \otimes F}(s_1 \otimes s_2) &= (\nabla^E s_1) \otimes s_2 + s_1 \otimes (\nabla^F s_2) \end{align*} defines a connection on $E \otimes F$. Applying this inductively to copies of $TM$ and $T^*M$ extends $\nabla$ to any tensor bundle $\mathcal{T}^{q,p}M$. The metric itself is a section $g \in \Gamma(T^*M \otimes T^*M)$. Metric compatibility then takes the elegant global form: \begin{align*} \nabla g &= 0. \end{align*} This says that the metric is a parallel section of $T^*M \otimes T^*M$ — it is constant along parallel transport in every direction. This is the sense in which the Levi-Civita connection is "adapted to" the metric. [remark: Uniqueness of the Levi-Civita Connection] The two defining properties — metric compatibility and torsion-freeness — together uniquely determine $\nabla$. Existence and uniqueness are proved by deriving the Koszul formula: applying metric compatibility to each of $\nabla_X g(Y, Z)$, $\nabla_Y g(X, Z)$, $\nabla_Z g(X, Y)$, then combining using torsion-freeness, one isolates $g(\nabla_X Y, Z)$ in terms of metric coefficients alone. This derivation shows that any torsion-free metric-compatible connection must equal the one given by the Koszul formula. [/remark] With the Levi-Civita connection and metric fixed, we turn to curvature—the fundamental measure of how the manifold twists and fails to be flat—captured by the Riemann tensor's role in comparing covariant derivatives along different paths. # 2. Riemann curvature This chapter develops the central object of Riemannian geometry: the Riemann curvature tensor. Building on the Levi-Civita connection from Chapter 1, we ask how much the connection fails to commute when applied in two different directions. This failure is precisely curvature. We then extract from the full curvature tensor several simpler, geometrically meaningful quantities — sectional, Ricci, and scalar curvature — that will play a central role in the rest of the course. ## The Curvature Tensor The key idea is that on a flat space, covariant derivatives commute: differentiating a vector field first in direction $X$ and then in direction $Y$ gives the same result as differentiating first in direction $Y$ and then in direction $X$. On a curved manifold, this commutativity fails, and the failure is measured by a tensor. [definition: Riemann Curvature Tensor] Let $(M, g)$ be a Riemannian manifold with Levi-Civita connection $\nabla$. The **curvature $2$-form** is the section \begin{align*} R = -\nabla \circ \nabla \in \Gamma(\Lambda^2 T^*M \otimes T^*M \otimes TM) \subseteq \Gamma(T^{1,3}M). \end{align*} This is a $(1,3)$-tensor field; we call it the **Riemann curvature tensor** of $(M,g)$. [/definition] It is helpful to think of $R$ as a $2$-form with values in $\operatorname{End}(TM) = T^*M \otimes TM$. For any vector fields $X, Y \in \mathfrak{X}(M)$, the expression $R(X,Y)$ is a section of $\operatorname{End}(TM)$, meaning it acts on vector fields. The following formula makes the curvature explicit and is the one most often used in computations. [quotetheorem:2703] **Convention note.** The sign convention $R = -\nabla \circ \nabla$ used here means $R(X,Y) = \nabla_{[X,Y]} - [\nabla_X, \nabla_Y]$. This is the **negative** of the convention used in many textbooks (e.g. do Carmo, Petersen), which define $R(X,Y) = [\nabla_X, \nabla_Y] - \nabla_{[X,Y]}$. The sign difference propagates into the Jacobi equation in Chapter 3, the Weitzenböck Ricci identity in Chapter 4, and the infinitesimal holonomy expansion in Chapter 5; whenever comparing with other sources, check their sign convention first. This formula reveals precisely why the bracket correction term $\nabla_{[X,Y]}Z$ is necessary. Without it, the expression $[\nabla_X, \nabla_Y]Z = \nabla_X \nabla_Y Z - \nabla_Y \nabla_X Z$ would not be tensorial — it would depend on derivatives of $X$ and $Y$, not just their values at a point. The bracket correction restores tensoriality, leaving a genuine tensor that measures non-commutativity geometrically rather than as an artifact of the choice of vector fields.  **Local coordinates.** In a chart $(U, \varphi)$ with coordinates $(x_1, \ldots, x_n)$, we write \begin{align*} R = \bigl(R^i_{j,k\ell}\, dx^k\, dx^\ell\bigr)_{i,j=1,\ldots,n} \in \Omega^2_M(\operatorname{End}(TM)), \end{align*} so that $R(X,Y)^i_j = R^i_{j,k\ell} X^k Y^\ell$ (summation convention in force). The comma between $j$ and $k\ell$ is a notational convenience to separate the endomorphism index from the $2$-form indices. ### The Fully Covariant Tensor It is often useful to lower all indices using the metric and work with a $(0,4)$-tensor instead: \begin{align*} R(X, Y, Z, T) = g(R(X,Y)Z,\, T), \qquad X,Y,Z,T \in T_pM. \end{align*} In coordinates this reads $R_{ij,k\ell} = g_{iq} R^q_{j,k\ell}$. The $(0,4)$ form is the version in which the symmetries of the curvature tensor are most cleanly stated. ## Symmetries of the Curvature Tensor Having defined the curvature tensor, a natural question is: how many independent components does $R_{ij,k\ell}$ actually have? A priori it is a $(0,4)$-tensor with $n^4$ components. The four symmetry identities below cut this count dramatically — to $\frac{1}{12}n^2(n^2-1)$ independent components in dimension $n$, which equals $1$ for $n=2$, $6$ for $n=3$, and $20$ for $n=4$. The Riemann curvature tensor satisfies several non-obvious symmetry identities. Far from being a coincidence, these reflect deep properties of the Levi-Civita connection — specifically, its compatibility with the metric and its torsion-free character. [quotetheorem:2704] [citeproof:2704] The proof above is more than a sequence of computations: it allocates each symmetry to a specific structural feature of the Levi-Civita connection. The skew-symmetry in the last two indices is built into the definition of $R$ as a $2$-form, the antisymmetry in the first pair traces back to metric compatibility, and the first Bianchi identity emerges from the torsion-free condition. The pair symmetry is not an independent axiom — it is a consequence of the other three identities together. The next remark records this allocation explicitly, since it is useful to know which conclusions survive when one or another axiom is dropped. [remark: Which Properties Use Which Axioms] The skew-symmetry $R_{ij,k\ell} = -R_{ij,\ell k}$ uses no properties of $\nabla$ at all. The antisymmetry in the first pair $R_{ij,k\ell} = -R_{ji,k\ell}$ uses only metric compatibility ($\nabla g = 0$). The first Bianchi identity uses only torsion-freeness ($\Gamma^k_{ij} = \Gamma^k_{ji}$). The pair symmetry (3) uses both. [/remark] These symmetry identities are not mere algebraic curiosities — they encode exactly which data from the ambient geometry the curvature tensor captures. The count of independent components confirms that the curvature tensor, while derived from a connection, carries strictly less redundancy than a generic $(0,4)$-tensor: in dimension $n$, the symmetry-imposed quotient leaves the Riemann number $\frac{1}{12}n^2(n^2-1)$ as the dimension of the space of curvature-like tensors. The pair symmetry (3) has a neat algebraic formulation: $R_{ij,k\ell}$ is a symmetric bilinear form on $\Lambda^2 T_p^*M$. ## Sectional Curvature The Riemann tensor in dimension $n$ has $\frac{1}{12}n^2(n^2-1)$ independent components — for $n = 4$ that is already $20$ numbers at each point. The challenge is to find the minimal data that still recovers the full tensor. The answer is beautiful: a single real-valued function on $2$-planes, the sectional curvature, suffices. The full curvature tensor has $O(n^4)$ components, but all the geometric information is already captured by this much simpler function, which assigns a real number to each $2$-plane in the tangent space. Recall that for tangent vectors $X, Y \in T_pM$, the area of the parallelogram they span is $|X \wedge Y| = \sqrt{g(X,X)g(Y,Y) - g(X,Y)^2}$. [definition: Sectional Curvature] Let $(M,g)$ be a Riemannian manifold and $p \in M$. For a $2$-plane $\sigma = \operatorname{span}(X,Y) \subseteq T_pM$, the **sectional curvature** of $\sigma$ is \begin{align*} K(\sigma) = K(X,Y) = \frac{R(X,Y,X,Y)}{|X \wedge Y|^2}. \end{align*} This defines a smooth function $K: \operatorname{Gr}(2, T_pM) \to \mathbb{R}$. [/definition] The definition is well-posed: $K(X,Y)$ is unchanged when $X$ or $Y$ are scaled by non-zero constants, and is also invariant under the shear $(X,Y) \mapsto (X + \lambda Y, Y)$. These operations generate all isomorphisms of a two-dimensional vector space, so $K(X,Y)$ depends only on the $2$-plane $\sigma$ spanned by $X$ and $Y$, not on the particular basis chosen. The remarkable fact is that the sectional curvature completely determines the full Riemann curvature tensor: [quotetheorem:2705] [citeproof:2705] Note carefully that this result requires $\dim V \geq 2$: in dimension $1$ there are no $2$-planes, so the sectional curvature function is vacuously defined on an empty domain and carries no information whatsoever — the curvature tensor is not determined. The proof also requires all four symmetries: without the pair symmetry $R(X,Y,Z,T) = R(Z,T,X,Y)$, the polarization step would fail to produce the cyclic invariance needed to invoke the first Bianchi identity. [example: Non-Commutativity of Covariant Derivatives on the Sphere vs. the Plane] The defining property of curvature — that covariant derivatives fail to commute — is best understood by contrasting two explicit cases: the flat plane $\mathbb{R}^2$ and the round sphere $S^2$. **Flat plane.** On $\mathbb{R}^2$ with the standard Euclidean metric, the Christoffel symbols all vanish in Cartesian coordinates, so $\nabla_{\partial_i} = \partial_i$. For any smooth vector field $Z = Z^1 \partial_1 + Z^2 \partial_2$, we compute \begin{align*} [\nabla_{\partial_1}, \nabla_{\partial_2}]Z = \partial_1 \partial_2 Z - \partial_2 \partial_1 Z = 0, \end{align*} because partial derivatives commute for smooth functions. Since $[\partial_1, \partial_2] = 0$ as well, the curvature formula gives $R(\partial_1, \partial_2) = \nabla_{[\partial_1,\partial_2]} - [\nabla_{\partial_1}, \nabla_{\partial_2}] = 0$. Covariant derivatives commute — the plane is flat. **Round sphere.** On $S^2$ with the round metric of radius $1$, use the chart $(\theta, \phi)$. The nonzero Christoffel symbols computed in Chapter 1 are $\Gamma^\phi_{\theta\phi} = \Gamma^\phi_{\phi\theta} = \cot\theta$ and $\Gamma^\theta_{\phi\phi} = -\sin\theta\cos\theta$. Computing covariant derivatives of $\partial_\theta$ in the two orders gives a non-zero commutator: one finds $R(\partial_\theta, \partial_\phi)\partial_\theta = -\partial_\phi \neq 0$ at every interior point. Covariant derivatives do not commute on the sphere — it is genuinely curved. [/example] The plane-versus-sphere contrast shows directly that curvature is the obstruction to commutativity of covariant differentiation. The asymmetry between these two cases is sharp: on the plane every commutator vanishes identically, while on the sphere a single commutator computation already produces a non-zero vector field, and the magnitude of that field is detectable from the metric components alone. This is a useful sanity check on the formalism — the commutator interpretation of curvature genuinely captures geometric content rather than reflecting an artefact of coordinates. It also raises a natural follow-up question. The plane is the simplest case of constant sectional curvature ($K \equiv 0$), and the sphere will turn out to be the next simplest ($K \equiv 1$). What does the entire curvature tensor look like when sectional curvature is constant, and how much of the previous structure can be recovered from this single number? [example: Constant Sectional Curvature] Suppose $(M,g)$ has the property that $K_p: \operatorname{Gr}(2, T_pM) \to \mathbb{R}$ is constant (equal to some real number $K_p$) for each $p$. Define \begin{align*} R^0_p(X,Y,Z,T) = g_p(X,Z)\,g_p(Y,T) - g_p(X,T)\,g_p(Y,Z). \end{align*} Then $R^0_p$ satisfies all the curvature symmetries, and a direct computation shows $K^0(X,Y) = 1$ for any orthonormal $X,Y$. Applying the previous theorem with $R' = K_p R^0_p$ and $R'' = R_p$ gives \begin{align*} R_p = K_p R^0_p. \end{align*} In a chart where the metric at $p$ is the identity $\delta_{ij}$, this reads $R_{ij,ij} = -R_{ij,ji} = K_p$ for all $i \neq j$, with all other independent components zero. The function $p \mapsto K_p$ is smooth since both $g$ and $R$ are smooth. Note that when $\dim M > 2$, constant sectional curvature at each point actually forces $K_p$ to be independent of $p$ (i.e., the manifold has globally constant sectional curvature). This is a non-trivial result — **Schur's lemma** — whose proof requires the second Bianchi identity (see below). It is stated here as a fact quoted without proof; the argument can be found in any standard reference (e.g. do Carmo, Riemannian Geometry, Chapter 4). [/example] ## Ricci Curvature and Scalar Curvature The sectional curvature recovers the full Riemann tensor, but it requires knowing a real number for every $2$-plane at every point — a vast amount of data. What can we extract that is both simpler and geometrically meaningful? The answer is to trace: contracting one index of the curvature tensor produces the Ricci curvature, a symmetric $2$-tensor; contracting again gives the scalar curvature, a single function. These reductions lose information but retain the coarse geometric data needed for global questions like comparison theorems and Einstein's field equations. The cost is that two manifolds can have the same Ricci curvature but different Riemann tensors — except in dimension $3$, as we shall see. The sectional curvature contains all the information in the Riemann tensor, but it is a function on $2$-planes rather than on the manifold itself. By taking traces, we obtain tensors that are simpler to work with, at the cost of losing some information. [definition: Ricci Curvature] The **Ricci curvature** of $(M,g)$ at $p \in M$ is the symmetric bilinear form \begin{align*} \operatorname{Ric}_p(X,Y) = \operatorname{tr}\bigl(v \mapsto R_p(X,v)Y\bigr), \qquad X,Y \in T_pM. \end{align*} In local coordinates, \begin{align*} \operatorname{Ric}_{ij} = R^q_{i,jq} = g^{pq} R_{pi,jq}, \end{align*} where $g^{pq}$ is the inverse of the metric matrix. The associated quadratic form is \begin{align*} \operatorname{Ric}(X) = \frac{1}{n-1}\operatorname{Ric}_p(X,X), \end{align*} with the factor $\frac{1}{n-1}$ a normalising convention. [/definition] The definition above is a partial trace, so its content is best understood by asking what averaging operation it performs on the curvature tensor. The next remark unpacks $\operatorname{Ric}(X,X)$ in an orthonormal basis and identifies it as a sum of sectional curvatures of the $2$-planes through $X$. [remark: Ricci as Average Sectional Curvature] If $e_1, \ldots, e_n$ is an orthonormal basis of $T_pM$ and $X = e_i$, then $\operatorname{Ric}(e_i, e_i) = \sum_{j \neq i} K(\operatorname{span}(e_i, e_j))$. So $\operatorname{Ric}(X,X)/(n-1)$ is the average sectional curvature over all $2$-planes containing $X$. [/remark] This averaging interpretation explains what Ricci curvature detects and what it misses: it measures the average of sectional curvatures in directions transverse to $X$, but it discards the variation among those individual sectional curvatures. Two manifolds can agree on all such averages while having very different sectional curvature functions — the difference is captured by the Weyl tensor, which is the traceless part of the curvature tensor. [definition: Scalar Curvature] The **scalar curvature** of $(M,g)$ is the trace of $\operatorname{Ric}$ with respect to $g$: \begin{align*} s = g^{ij}\operatorname{Ric}_{ij} = g^{ij} R^q_{i,jq} = R^{qi}{}_{iq}. \end{align*} [/definition] The scalar curvature is a single real-valued function on $M$. Some sources define it as $s/n(n-1)$ so that a unit sphere has scalar curvature $1$. For a manifold of constant sectional curvature $K_p$, the Ricci and scalar curvatures take simple forms. Since $R_p = K_p R^0_p$, taking a trace gives \begin{align*} \operatorname{Ric}_p = (n-1)K_p\, g_p, \qquad s(p) = n(n-1)K_p. \end{align*} These formulas confirm the intuition that positively curved manifolds (like spheres) have positive Ricci and scalar curvature. [example: Curvature of the Round Sphere $S^2$] We compute sectional curvature, Ricci, and scalar curvature for $S^2$ with the round metric of radius $1$, using the Christoffel symbols from Chapter 1. In the $( \theta, \phi)$-chart, the metric is $g = d\theta^2 + \sin^2\theta\, d\phi^2$, giving $g_{11} = 1$, $g_{12} = 0$, $g_{22} = \sin^2\theta$. The nonzero Christoffel symbols are $\Gamma^\phi_{\theta\phi} = \Gamma^\phi_{\phi\theta} = \cot\theta$ and $\Gamma^\theta_{\phi\phi} = -\sin\theta\cos\theta$. A direct computation using the coordinate formula for the curvature tensor yields the single independent fully-covariant component \begin{align*} R_{12,12} = \sin^2\theta. \end{align*} Since $|\partial_\theta \wedge \partial_\phi|^2 = g_{11}g_{22} - g_{12}^2 = \sin^2\theta$, the sectional curvature is \begin{align*} K = \frac{R_{12,12}}{|\partial_\theta \wedge \partial_\phi|^2} = \frac{\sin^2\theta}{\sin^2\theta} = 1. \end{align*} The sphere has constant sectional curvature $K = 1$. Applying the formulas for constant curvature spaces: \begin{align*} \operatorname{Ric}_p &= (n-1)K\, g_p = 1 \cdot g_p = g_p, \qquad s = n(n-1)K = 2. \end{align*} So on $S^2$ the Ricci tensor equals the metric, and the scalar curvature is $2$. **The Theorema Egregium.** The Gaussian curvature of a surface in $\mathbb{R}^3$ was originally defined extrinsically using the shape operator. For $S^2 \subset \mathbb{R}^3$ of radius $1$, the principal curvatures are both $1$, so the classical Gaussian curvature is $K = 1 \cdot 1 = 1$. Our purely intrinsic computation — using only $g_{ij}$ and its derivatives — gives the same value. This is the Theorema Egregium in action: the Gaussian curvature is intrinsic, computed entirely from the first fundamental form. [/example] The sphere computation makes concrete the theoretical machinery developed so far: the Christoffel symbols give a connection, the connection gives a curvature tensor, and the curvature tensor recovers both classical Gaussian curvature and the modern Ricci/scalar reductions. With this foundation, we now examine how curvature simplifies in low dimensions, where component counts force striking equivalences. ## Low-Dimensional Cases A dimension count shows that in dimensions $2$ and $3$, the curvature tensor is far more constrained than in higher dimensions. In dimension $2$ there is only one $2$-plane, so the curvature tensor reduces to a single number. In dimension $3$, a counting argument shows the curvature and Ricci tensors have the same number of independent components — raising the question of whether they are equivalent. How many independent components does the curvature tensor carry, and when does that count collapse enough for simpler data to determine it? The curvature theory simplifies greatly in low dimensions, because certain tensors that are independent in general dimension become determined by simpler data. ### Surfaces ($n = 2$) On a surface, the Riemannian metric $g$ is traditionally called the **first fundamental form** and is written in local coordinates $(u,v)$ as \begin{align*} g = E\,du^2 + 2F\,du\,dv + G\,dv^2. \end{align*} There is only one $2$-plane in $T_pM$ (the whole tangent space), so the sectional curvature is a single real number at each point — the **Gaussian curvature** $K$. Up to the symmetries, the only independent component of the curvature tensor is $R_{12,12}$, and one verifies that \begin{align*} R_{12,12} = \frac{1}{2}s(EG - F^2). \end{align*} The scalar curvature and the sectional curvature therefore satisfy $s/2 = K$. One further checks using the second fundamental form (with components $L,M,N$) that \begin{align*} K = \frac{LN - M^2}{EG - F^2}, \end{align*} the classical formula for Gaussian curvature. Note that $R_{12,21} = -R_{12,12} = -(LN - M^2)$ by skew-symmetry in the last two indices. [quotetheorem:2706] The Theorema Egregium asserts that $K$ — originally defined extrinsically via the shape operator — is completely determined by the intrinsic geometry. This is a non-trivial fact: it says that a flat map of the Earth must distort distances (since the sphere has $K = 1$ while the plane has $K = 0$, and an isometry would have to preserve $K$). The proof amounts to the explicit Brioschi formula expressing $K$ in terms of $E, F, G$ alone, which follows from the coordinate expression for $R_{12,12}$ computed via Christoffel symbols. ### Three-Manifolds ($n = 3$) [quotetheorem:2707] This formula is verified by checking that both sides have the same symmetries and produce the same Ricci curvature on contraction. The key dimension count: the Riemann tensor in dimension $3$ has $\frac{1}{12} \cdot 9 \cdot 8 = 6$ independent components, matching the $6$ independent components of a symmetric $3 \times 3$ matrix (the Ricci tensor). The map is bijective at this count. **The Second Bianchi Identity.** The curvature tensor satisfies a differential identity: \begin{align*} (\nabla_X R)(Y,Z,W,V) + (\nabla_Y R)(Z,X,W,V) + (\nabla_Z R)(X,Y,W,V) = 0. \end{align*} Contracted, this gives the **contracted Bianchi identity** $\operatorname{div}\operatorname{Ric} = \frac{1}{2}ds$, or in coordinates $\nabla^j \operatorname{Ric}_{ij} = \frac{1}{2}\partial_i s$. This identity plays a central role in general relativity (it ensures the Einstein tensor $G_{ij} = \operatorname{Ric}_{ij} - \frac{s}{2}g_{ij}$ is divergence-free) and in the study of Ricci flow. It is also the key tool in proving Schur's lemma (that pointwise constant sectional curvature implies globally constant curvature when $\dim M > 2$). This chapter has developed the full apparatus of Riemannian curvature: from the tensor $R$ measuring failure of commutativity of covariant derivatives, through the sectional curvature that determines $R$ completely, to the Ricci and scalar curvature that distill the tensor to simpler data. In the next chapter we turn to **geodesics** — the curves of zero acceleration. Curvature will reappear immediately: the sectional curvature controls how nearby geodesics spread apart (the Jacobi equation), and the sign of the Ricci curvature governs global comparison theorems for geodesic lengths. The Riemann curvature tensor encodes subtle symmetries; sectional, Ricci, and scalar curvatures are its key reductions. Geodesics represent the straightest paths in this curved space, and the Jacobi equation will show precisely how curvature forces nearby geodesics to diverge or converge. # 3. Geodesics This chapter studies geodesics — the curves of zero acceleration on a Riemannian manifold — and the global geometric consequences of their behaviour. We begin by defining geodesics through the Levi-Civita connection and developing the basic local theory: the geodesic equation, the exponential map, and short-time existence. Jacobi fields measure how families of geodesics spread apart and connect curvature to global geometry. We then develop minimization properties via Gauss's lemma and normal coordinates, prove the Hopf--Rinow theorem characterising metric completeness, develop the first and second variation formulas, and apply this machinery to obtain global theorems: the Bonnet--Myers diameter bound, Synge's theorem on simply-connectedness, and related comparison results. ## Definitions and basic properties With a Riemannian manifold $(M, g)$ and its Levi-Civita connection $\nabla$ in hand, we can ask a fundamental question: what are the "straightest possible" curves on $M$? On flat Euclidean space the answer is obvious — straight lines — but on a curved manifold like a sphere, the notion of straightness must be defined intrinsically using only the geometry of $M$ itself. The answer is the concept of a **geodesic**, a curve whose velocity vector is parallel transported along itself. Before defining geodesics directly, it is natural to set up the machinery of covariant differentiation along curves in a general vector bundle, since this context makes the structure cleaner and will be reused throughout the chapter. ### Covariant differentiation along curves Let $\pi: E \to M$ be a smooth vector bundle with typical fiber $V$, and let $\gamma: (-\varepsilon, \varepsilon) \to M$ be a smooth curve. [definition: Lift of a Curve] Let $\pi: E \to M$ be a vector bundle. A **lift** of a curve $\gamma: (-\varepsilon, \varepsilon) \to M$ is a smooth map $\gamma^E: (-\varepsilon, \varepsilon) \to E$ satisfying $\pi \circ \gamma^E = \gamma$, i.e., $\gamma^E(t) \in E_{\gamma(t)}$ for all $t$. [/definition] Concretely, a lift assigns an element of the fiber $E_{\gamma(t)}$ to each point along the curve. If $E = TM$, the canonical lift is simply $t \mapsto \dot\gamma(t)$; in a physical interpretation where $M$ is spacetime and the fiber encodes a field, a lift is a field value recorded along a worldline. In a local trivialization over a coordinate neighborhood $U \subseteq M$ on which $E$ is trivial, a lift takes the form $\gamma^E(t) = (\gamma(t), a(t))$ for a smooth function $a: (-\varepsilon, \varepsilon) \to V$. The next step is to differentiate lifts along $\gamma$. [quotetheorem:2708] [citeproof:2708] [definition: Covariant Derivative Along a Curve] The uniquely defined operator $\frac{\nabla}{dt}$ constructed above is called the **covariant derivative along $\gamma$**. [/definition] The covariant derivative along $\gamma$ measures how a lift changes relative to the geometry of $E$, not just as a map into the total space. ### Horizontal lifts and parallel transport A natural special class of lifts are those that are "constant" in the sense of $\nabla$. [definition: Horizontal Lift] A lift $\gamma^E$ of $\gamma$ is called **horizontal** (or **parallel along $\gamma$**) if \begin{align*} \frac{\nabla \gamma^E}{dt} = 0. \end{align*} [/definition] In local coordinates, horizontality is the first-order linear ODE system $\dot a_i + \sum_{j,k}\Gamma_{jk}^i\,a_j\,\dot x_k = 0$. Since this system is linear with smooth coefficients, the theory of ODEs gives global existence and uniqueness for any initial value $a(0) \in E_{\gamma(0)}$. [definition: Parallel Transport] Let $\gamma: [0,1] \to M$ be a smooth curve and $a_0 \in E_{\gamma(0)}$. The unique horizontal lift $\gamma^E$ of $\gamma$ with $\gamma^E(0) = a_0$ is called the **parallel transport** of $a_0$ along $\gamma$. The resulting element $\gamma^E(1) \in E_{\gamma(1)}$ is also called the parallel transport of $a_0$ from $\gamma(0)$ to $\gamma(1)$ along $\gamma$. [/definition] Parallel transport provides an isomorphism $P_\gamma^{0,1}: E_{\gamma(0)} \to E_{\gamma(1)}$ that depends on the path $\gamma$ (not just the endpoints), reflecting the curvature of the connection. ### Geodesics We now specialize to the Riemannian setting: $M$ is a Riemannian manifold with metric $g$, $E = TM$, and $\nabla$ is the Levi-Civita connection. A curve $\gamma$ has a canonical lift to $TM$ given by its velocity $\dot\gamma(t) \in T_{\gamma(t)}M$. [definition: Geodesic] A smooth curve $\gamma: I \to M$ on a Riemannian manifold $(M, g)$ is a **geodesic** if its velocity vector is parallel along itself, i.e., \begin{align*} \frac{\nabla \dot\gamma}{dt} = 0. \end{align*} [/definition]  In local coordinates $(x_1, \ldots, x_n)$, the geodesic equation reads \begin{align*} \ddot x_i + \sum_{j,k} \Gamma_{jk}^i\, \dot x_j\, \dot x_k = 0. \end{align*} This is a second-order nonlinear ODE system. By the Picard–Lindelöf theorem, given any initial point $p \in M$ and initial velocity $a \in T_pM$, there exists a unique maximal geodesic $\gamma_{p,a}: I \to M$ with $\gamma_{p,a}(0) = p$ and $\dot\gamma_{p,a}(0) = a$, defined on some open interval $I$ containing $0$. Unlike linear ODEs, global existence is not guaranteed for all time. We write $\gamma_p(t, a)$ for this geodesic when we want to emphasize the dependence on the initial data. ### First properties of geodesics A key immediate consequence of the geodesic equation is that geodesics travel at constant speed. [quotetheorem:2709] [citeproof:2709] [remark: Speed vs. Parametrization] This constant speed is a property of the parametrization, not merely of the image. Two geodesics can trace the same path but at different (constant) speeds, and a constant-speed reparametrization of a geodesic is again a geodesic. A geodesic with $|\dot\gamma|_g = 1$ is called a **unit-speed geodesic**. [/remark] The next result is a rescaling property that will be crucial when we define the exponential map. [quotetheorem:2710] [citeproof:2710] The rescaling lemma allows us to fix $t = 1$ and vary $a \in T_pM$, packaging all geodesic information into a single smooth map. The lemma does not require $\lambda > 0$: reversing the direction of travel ($\lambda = -1$) reverses the geodesic, reflecting the time-reversal symmetry of the geodesic equation. Note also that the lemma is a statement about parametrized curves — two distinct values of $\lambda$ in general give distinct curves (one travels faster than the other), though they trace the same image in $M$. ### Examples of geodesics [example: Euclidean Space] On $\mathbb{R}^n$ with the Euclidean metric, all Christoffel symbols vanish: $\Gamma_{jk}^i = 0$. The geodesic equation reduces to $\ddot x_i = 0$, whose solutions are $x(t) = p + ta$ for $p, a \in \mathbb{R}^n$. Hence the geodesics are precisely the straight lines, consistent with our Euclidean intuition. [/example] The Euclidean case is degenerate in an instructive way: the vanishing Christoffel symbols mean the geodesic equation decouples into $n$ independent scalar equations $\ddot{x}_i = 0$, and there is no curvature to deflect geodesics from straight lines. On a curved manifold, even the local behavior of geodesics is more subtle — the Christoffel symbols couple the coordinate functions together, and geodesics can curve, cross, and focus depending on the sign and magnitude of the curvature. [example: Geodesics on the Sphere] On the round sphere $S^n$ with the metric induced by the standard embedding $S^n \hookrightarrow \mathbb{R}^{n+1}$, the geodesics are the great circles (intersections of $S^n$ with two-dimensional planes through the origin). To verify this, it suffices to identify the geodesic through $p = e_0$ with initial velocity $a = e_1$ (using the standard basis $\{e_i\}$ of $\mathbb{R}^{n+1}$). Consider the map $\varphi: (x_0, x_1, x_2, \ldots, x_n) \mapsto (x_0, x_1, -x_2, \ldots, -x_n)$. This is an isometry of $S^n$ (it is the restriction of an orthogonal reflection) that fixes both $p$ and $a$. Since isometries preserve geodesics and the geodesic through $(p,a)$ is unique, $\varphi(\gamma) = \gamma$, so $\gamma$ must lie in the fixed-point set of $\varphi$, which is the great circle in the $(e_0, e_1)$-plane. All other geodesics are obtained by applying isometries of $S^n$. [/example] ### The exponential map The geodesic rescaling lemma shows that for a fixed $p \in M$, the geodesic $\gamma_p(t, a)$ depends on the combination $ta$, so we may set $t = 1$ and let $a$ range over a neighborhood of $0 \in T_pM$. [definition: Exponential Map] Let $(M, g)$ be a Riemannian manifold and $p \in M$. The **exponential map** at $p$ is \begin{align*} \exp_p: \mathcal{U}_p \to M, \qquad \exp_p(a) = \gamma_p(1, a), \end{align*} where $\mathcal{U}_p \subseteq T_pM$ is the open set of vectors $a$ for which the geodesic $\gamma_p(\cdot, a)$ is defined at least on $[0,1]$. [/definition] By standard ODE smooth dependence on parameters, $\mathcal{U}_p$ is an open neighborhood of $0 \in T_pM$ and $\exp_p$ is smooth. By construction, $\exp_p(0) = p$ and the geodesic from $p$ in direction $a$ is $t \mapsto \exp_p(ta)$. The following proposition identifies the differential of $\exp_p$ at the origin, which is the key to using $\exp_p$ as a local chart. [quotetheorem:2711] [citeproof:2711] The differential being the identity has a concrete geometric meaning: to first order, the exponential map simply carries a tangent vector $v$ to the point reachable by traveling a unit time in direction $v$. The theorem says nothing about higher-order behavior — the exponential map is generally not an isometry, only a diffeomorphism near $0$, and the deviation from isometry is encoded by the curvature tensor. This first-order agreement between $\exp_p$ and the identity is what makes normal coordinates well-behaved at their center point. Since $(d\exp_p)_0 = \operatorname{id}_{T_pM}$ is a linear isomorphism, the inverse function theorem immediately gives: [quotetheorem:2712] The inverse $\exp_p^{-1}$ provides a local coordinate chart centered at $p$, known as **geodesic local coordinates** (or **normal coordinates**). Choosing an orthonormal basis for $(T_pM, g_p)$ to identify $T_pM \cong \mathbb{R}^n$ isometrically, these coordinates have especially pleasant properties. [quotetheorem:2713] [citeproof:2713] Normal coordinates at $p$ are a powerful computational tool: any coordinate-independent identity can be verified at an arbitrary point, and at that point we may work in normal coordinates where $g_{ij}(p) = \delta_{ij}$ and $\Gamma_{jk}^i(p) = 0$, greatly simplifying calculations. ### Geodesic polar coordinates and Gauss' lemma By identifying $(T_pM, g_p) \cong (\mathbb{R}^n, \mathrm{eucl})$, we can introduce polar-type coordinates on a punctured neighborhood of $p$. For $\delta > 0$ sufficiently small, the map \begin{align*} (r, v) \in (0, \delta) \times S^{n-1} \longmapsto \exp_p(rv) \in M \end{align*} is a smooth embedding. These are the **geodesic polar coordinates** centered at $p$. For each fixed $r$, the image \begin{align*} \Sigma_r = \exp_p(r\, S^{n-1}) \subseteq M \end{align*} is called the **geodesic sphere** of (geodesic) radius $r$ centered at $p$; it is a smooth embedded hypersurface. A natural question: how does the Riemannian metric look in geodesic polar coordinates? The radial directions are unit-speed geodesics, so $g(\partial_r, \partial_r) = 1$ everywhere. The deeper fact — that the radii are also orthogonal to the geodesic spheres — is the content of Gauss' lemma. [quotetheorem:1558] [citeproof:1558] Gauss' lemma has a clean covariant formulation in terms of the exponential map's differential. [quotetheorem:2714] This says the exponential map preserves inner products involving the radial direction, even though it is not in general an isometry. In the next section, we will see how geodesics arise as Jacobi fields and how the curvature of $M$ influences their behavior. ## Jacobi Fields [motivation] ### The space of geodesics Geodesics on a Riemannian manifold form a rich family of curves, and a natural question is: what does it mean to deform a geodesic through other geodesics? Imagine the "manifold of all smooth curves" on $M$. A tangent vector at a curve $\gamma$ in this infinite-dimensional space should assign a tangent vector in $M$ to each point along $\gamma$ — a vector field along $\gamma$ that records how the curve is moving as we vary it. When we restrict to the "submanifold" of geodesics, these tangent vectors are constrained: not every vector field along a geodesic arises as the variation field of a one-parameter family of geodesics. The Jacobi equation characterises exactly which vector fields do. This motivates the central definition of this section: a **Jacobi field** is a vector field along a geodesic satisfying the Jacobi equation, and the main theorem of the section asserts that these are precisely the variation fields of geodesic families. ### Why this matters Jacobi fields encode how neighbouring geodesics spread apart or converge. In a positively curved space, geodesics focus toward each other; in a negatively curved space, they diverge. The rate at which $|J(t)|$ grows or shrinks encodes sectional curvature, and this connection — made precise by the formula $|J(t)| = t - \frac{1}{6} K(\sigma) t^3 + o(t^3)$ — is the bridge between the local invariant $K$ and the global geometry of $M$. [/motivation] ### Deriving the Jacobi equation Fix a Riemannian manifold $(M, g)$. Suppose $f: [0, L] \times (-\varepsilon, \varepsilon) \to M$ is a smooth map such that for each fixed $s$, the curve $t \mapsto f(t, s)$ is a geodesic. We write $\partial_t = \frac{\partial f}{\partial t}$ and $\partial_s = \frac{\partial f}{\partial s}$ for the induced vector fields along $f$.  Since each curve $t \mapsto f(t, s)$ is a geodesic, the geodesic equation gives \begin{align*} \nabla_{\partial_t} \partial_t = 0. \end{align*} We now differentiate this in the $s$-direction. Applying $\nabla_{\partial_s}$ to both sides: \begin{align*} 0 = \nabla_{\partial_s} \nabla_{\partial_t} \partial_t. \end{align*} We rearrange using the curvature tensor. Recall from Chapter 2 that the sign convention adopted in this course is $R = -\nabla \circ \nabla$, giving $R(X,Y) = \nabla_{[X,Y]} - [\nabla_X, \nabla_Y]$, equivalently $R(X,Y)Z = \nabla_{[X,Y]}Z - \nabla_X\nabla_YZ + \nabla_Y\nabla_XZ$. Rearranging: \begin{align*} 0 = \nabla_{\partial_s} \nabla_{\partial_t} \partial_t &= \nabla_{\partial_t} \nabla_{\partial_s} \partial_t - R(\partial_s, \partial_t)\partial_t \\ &= \nabla_{\partial_t} \nabla_{\partial_s} \partial_t + R(\partial_t, \partial_s)\partial_t. \end{align*} A key symmetry of the mixed covariant derivatives is that $\nabla_{\partial_s} \partial_t = \nabla_{\partial_t} \partial_s$ (since $[\partial_t, \partial_s] = 0$ for the coordinate vector fields of $f$, and the Levi-Civita connection is torsion-free). Substituting this and setting $s = 0$, with $J(t) = \frac{\partial f}{\partial s}(t, 0)$ the variation field along $\gamma(t) = f(t,0)$, we obtain: \begin{align*} \frac{\nabla^2 J}{dt^2} + R(\dot\gamma, J)\dot\gamma = 0. \end{align*} This is a **linear** second-order ODE along $\gamma$. [definition: Jacobi Field] Let $\gamma: [0, L] \to M$ be a geodesic in a Riemannian manifold $(M, g)$. A **Jacobi field** along $\gamma$ is a smooth vector field $J$ along $\gamma$ satisfying the **Jacobi equation**: \begin{align*} \frac{\nabla^2 J}{dt^2} + R(\dot\gamma, J)\dot\gamma = 0. \end{align*} [/definition] [remark: Tangential Jacobi fields] The vector fields $\dot\gamma(t)$ and $t\dot\gamma(t)$ both satisfy the Jacobi equation. This can be verified directly: $R(\dot\gamma, \dot\gamma)\dot\gamma = 0$ by skew-symmetry of $R$, so $\frac{\nabla^2 \dot\gamma}{dt^2} = 0$ (since $\dot\gamma$ is parallel along $\gamma$), giving $\dot\gamma$ as a Jacobi field; a short computation confirms $t\dot\gamma$ as well. These are the "longitudinal" Jacobi fields, arising from reparametrising the geodesic. [/remark] ### Existence, uniqueness, and structure Since the Jacobi equation is a linear second-order ODE, the standard theory of ODEs applies. The space of Jacobi fields is a vector space of dimension $2n$ (where $n = \dim M$), determined by initial data $(J(0), \frac{\nabla J}{dt}(0)) \in T_{\gamma(0)}M \times T_{\gamma(0)}M$. [quotetheorem:2715] [citeproof:2715] The theorem gives a clean decomposition: the $2n$-dimensional space of Jacobi fields splits as a $2$-dimensional "longitudinal" subspace (parallel to $\dot\gamma$) and a $2(n-1)$-dimensional "transverse" subspace (orthogonal to $\dot\gamma$). For the geometry of the manifold, the transverse Jacobi fields are the interesting ones. The theorem does not say that Jacobi fields are determined by their values at a single point — it requires two pieces of data, $J(0)$ and $J'(0)$, reflecting the second-order nature of the Jacobi equation. In particular, there is no reason a Jacobi field that vanishes at $\gamma(0)$ must also vanish elsewhere; when it does vanish again, this signals a conjugate point — a concept developed in §5. ### Jacobi fields in constant sectional curvature With the structure theorem in hand, we can compute Jacobi fields explicitly in the most symmetric case. [example: Constant Sectional Curvature] Suppose $(M, g)$ has constant sectional curvature $K \in \mathbb{R}$, with $\dim M \geq 2$. Normalise so that $|\dot\gamma| = 1$, and let $J$ be a Jacobi field orthogonal to $\dot\gamma$. For any vector field $T$ along $\gamma$, the definition of sectional curvature gives \begin{align*} g(R(\dot\gamma, J)\dot\gamma, T) = K\bigl(g(\dot\gamma, \dot\gamma)\,g(J, T) - g(\dot\gamma, J)\,g(\dot\gamma, T)\bigr) = K\,g(J, T), \end{align*} where we used $|\dot\gamma| = 1$ and $g(\dot\gamma, J) = 0$. Since this holds for all $T$, we have \begin{align*} R(\dot\gamma, J)\dot\gamma = K J. \end{align*} The Jacobi equation becomes \begin{align*} \frac{\nabla^2 J}{dt^2} + K J = 0. \end{align*} Using the parallel frame $\{X_i(t)\}_{i=2}^n$ orthogonal to $\dot\gamma$ from the structure theorem proof, the unique Jacobi fields with $J(0) = 0$ and $\frac{\nabla J}{dt}(0) = e_i$ are: \begin{align*} J(t) = \begin{cases} \dfrac{\sin(t\sqrt{K})}{\sqrt{K}}\, X_i(t) & K > 0, \\ t\, X_i(t) & K = 0, \\ \dfrac{\sinh(t\sqrt{-K})}{\sqrt{-K}}\, X_i(t) & K < 0. \end{cases} \end{align*} To verify: in each case, $J(0) = 0$ and $\frac{\nabla J}{dt}(0) = X_i(0) = e_i$ as required. In positive curvature ($K > 0$), the Jacobi field vanishes again at $t = \pi/\sqrt{K}$ — this is the first conjugate point along $\gamma$. In flat space ($K = 0$), the Jacobi field grows linearly. In negative curvature ($K < 0$), it grows exponentially, reflecting the divergence of geodesics. [/example] ### Jacobi fields as geodesic variations We have seen that every family of geodesics produces a Jacobi field. The converse is also true: every Jacobi field is the variation field of some family of geodesics. [quotetheorem:2716] [citeproof:2716] The theorem shows that Jacobi fields are not merely solutions to an ODE imposed along a geodesic, but carry genuine geometric information about the space of geodesics. One consequence is that the dimension count is consistent: the $2n$-dimensional solution space of the Jacobi equation matches the $2n$-dimensional space of initial conditions for the geodesic ODE, and both are parametrized by choosing a base geodesic and a nearby geodesic. The theorem also implies that Jacobi fields are the only obstruction to the exponential map being a local diffeomorphism — when a Jacobi field vanishing at the basepoint also vanishes at a later point, the exponential map fails to be a local diffeomorphism there. ### The exponential map description of Jacobi fields The converse in the theorem above has a particularly clean form when the Jacobi field vanishes at the basepoint. [quotetheorem:2717] [citeproof:2717]  [remark: Geometric meaning] This formula says that $J(t)$ is the image under $d\exp_p$ of the vector $tw$ in $T_p M$, where $w = J'(0)$ is the initial velocity of the Jacobi field. In geodesic normal coordinates (where $\exp_p$ is the identity to first order), $J(t) \approx tw$ for small $t$. The exponential map is a local diffeomorphism near $0$, so for small $t$ the Jacobi field truly grows like $tw$. The failure of the exponential map to be a diffeomorphism at a point $ta$ — detected by $d\exp_p$ being singular there — is precisely when a Jacobi field with $J(0) = 0$ vanishes again: the notion of a **conjugate point**. [/remark] ### Jacobi fields and sectional curvature: the size formula The exponential map formula yields a precise quantitative relationship between the size of a Jacobi field and the sectional curvature of $M$. [quotetheorem:2718] This is the explicit sense in which Jacobi fields measure curvature. In positive sectional curvature, $|J(t)|$ grows slower than $t$ (geodesics bend toward each other); in negative sectional curvature, $|J(t)|$ grows faster than $t$ (geodesics spread apart). The constant sectional curvature example above is consistent: for $K > 0$, $\frac{\sin(t\sqrt{K})}{\sqrt{K}} = t - \frac{K}{6}t^3 + O(t^5)$. ## Gauss's Lemma and Local Length Minimization We now know that geodesics exist locally and are unique given initial data. But do they actually minimize length? A geodesic on a sphere, continued past the antipodal point, eventually becomes longer than competing paths — so local existence alone does not guarantee minimization. The question is: over what range does a geodesic remain the shortest curve between its endpoints, and can we characterize this range intrinsically in terms of the geometry of $M$? The answer requires understanding how the exponential map distorts distances, which is the content of the strengthened Gauss's lemma developed in this section. From it we will deduce that every sufficiently short geodesic segment is the unique length-minimizer between its endpoints. ### A Strengthened Gauss's Lemma Recall that the exponential map $\exp_p: T_pM \to M$ sends a tangent vector $v$ to the point $\gamma_v(1)$, where $\gamma_v$ is the geodesic with $\gamma_v(0) = p$ and $\dot\gamma_v(0) = v$. An earlier version of Gauss's lemma, proved in geodesic normal coordinates, said that the radial direction is orthogonal to geodesic spheres. Here we prove a quantitative version that holds for all tangent vectors, not just unit ones. [quotetheorem:858] The force of this statement is that the inner product of the radial direction $(d\exp_p)_{ta}\, a$ with any direction $(d\exp_p)_{ta}\, w$ is determined entirely by the initial data $g_p(a, w)$. The exponential map preserves this pairing, even though it need not be an isometry. [citeproof:858] ### Local Minimization of Length The Gauss lemma lets us compare the length of an arbitrary curve starting at $p$ against the length of a geodesic from $p$ with the same endpoint in $T_pM$. [quotetheorem:2719] The hypothesis is about curves in $T_pM$ sharing the same endpoint $a$, not merely curves in $M$ sharing the image point $\exp_p(a)$. Two different curves in $T_pM$ can map to the same point in $M$ without passing through the same point in $T_pM$ — the torus $T^n = \mathbb{R}^n/\mathbb{Z}^n$ provides an example where this subtlety matters. [citeproof:2719] ### The Riemannian Distance and Geodesic Balls Local length minimization gives us a clean way to turn a Riemannian manifold into a metric space. [definition: Path Space] Write $\Omega(p, q)$ for the set of all piecewise $C^1$ curves from $p$ to $q$ in $M$. [/definition] [definition: Riemannian Distance] Let $M$ be a connected Riemannian manifold. For $p, q \in M$, define \begin{align*} d(p, q) = \inf_{\xi \in \Omega(p, q)} \operatorname{length}(\xi). \end{align*} [/definition] The infimum in the definition of $d(p, q)$ is taken over a non-empty set whenever $M$ is connected, but it need not be achieved — a sequence of paths whose lengths approach $d(p,q)$ may fail to have a convergent limit in $\Omega(p, q)$. The existence of a length-realizing curve (a minimal geodesic) is a separate question, addressed by the Hopf–Rinow theorem: completeness of $M$ is exactly the condition that guarantees it. Symmetry and the triangle inequality are immediate. That $d(p,q) > 0$ for $p \neq q$ is the content of the following theorem, which also characterizes minimizing curves. [quotetheorem:2720] [citeproof:2720] [remark: Distance Induces the Manifold Topology] The distance $d$ is a genuine metric on $M$: all metric axioms hold, with positivity being the non-trivial part. Moreover, the metric topology coincides with the $C^\infty$ manifold topology on $M$. This is a consequence of the theorem above and the fact that geodesic balls $\exp_p(B(0,\varepsilon))$ form a neighbourhood basis at each point. [/remark] ### Minimal Geodesics [definition: Minimal Geodesic] A curve $\gamma: [0,1] \to M$ is a *minimal geodesic* if \begin{align*} d(\gamma(0), \gamma(1)) = \operatorname{length}(\gamma). \end{align*} [/definition] The name is only justified if minimal geodesics are actually geodesics. This follows from the observation that any sub-arc of a minimizing curve is still minimizing. [quotetheorem:2721] [citeproof:2721] The smoothness conclusion is a genuine regularity result: the hypothesis is only piecewise $C^1$ (a very weak regularity assumption), yet the minimizing property forces the curve to satisfy the smooth geodesic equation everywhere. The argument works because any corner in a constant-speed curve can be smoothed out by the local minimization theorem to produce a strictly shorter curve — so a minimizing curve can have no corners, and once it has no corners the first-variation argument forces it to satisfy the geodesic equation. ### Geodesics Are Locally Minimal A geodesic need not be globally minimizing — on the sphere, for instance, continuing past the antipodal point begins increasing length. But geodesics are always *locally* minimal, in two related senses. The first sense chops up the geodesic into short segments. [quotetheorem:2722] [citeproof:2722] The second sense of local minimality concerns nearby curves in $\Omega(p,q)$, equipped with the topology of uniform convergence. [quotetheorem:2723] The regularity hypothesis is necessary. At conjugate points (where $d(\exp_p)_{ta}$ is singular), nearby geodesics can produce curves of the same or shorter length. The sphere illustrates this: all geodesics from the north pole $p$ to the south pole $q$ have the same length, so the minimizing geodesic is not strictly minimal in any neighbourhood of itself in $\Omega(p,q)$.  [citeproof:2723] [remark: Global vs. Local Minimization] The flat torus $T^n = \mathbb{R}^n/\mathbb{Z}^n$ illustrates why global minimization fails: the exponential map is everywhere regular (its derivative is the identity in the flat chart), so the theorem applies, but there are infinitely many geodesics connecting any two points. Only one of these is the globally shortest path; the others are local but not global minima in $\Omega(p,q)$. This shows the theorem gives the sharpest statement possible: regularity of $\exp_p$ along $\gamma$ is the precise condition for strict local minimality in path space. [/remark] ## Completeness and the Hopf–Rinow Theorem Two natural questions arise once we have the theory of geodesics in hand. First: can every geodesic be extended indefinitely, or might a geodesic "run off the edge" of the manifold in finite time? Second: given any two points, is there always a geodesic realizing the distance between them? Remarkably, these two questions have the same answer — a manifold satisfies one if and only if it satisfies the other. This equivalence, together with two further characterisations via metric-space completeness and the Heine–Borel property, is the content of the Hopf–Rinow theorem. ### Geodesic Completeness [definition: Geodesically Complete] A Riemannian manifold $(M, g)$ is **geodesically complete** if every geodesic can be extended to exist for all time. Equivalently, for every point $p \in M$, the exponential map $\exp_p$ is defined on all of $T_p M$. [/definition] The equivalence between extending geodesics and having a globally defined exponential map follows directly from the fact that $\exp_p(tv) = \gamma_v(t)$ where $\gamma_v$ is the geodesic with initial velocity $v$. [example: Incomplete and Complete Surfaces] The upper half-plane $H^2 = \{(x, y) \in \mathbb{R}^2 : y > 0\}$ equipped with the metric induced from $\mathbb{R}^2$ (i.e., the flat metric restricted to the open half-plane) is **not** geodesically complete. A horizontal geodesic moving downward reaches the boundary $y = 0$ in finite arc length and cannot be extended. In contrast, $\mathbb{R}^2$ itself with the flat metric is geodesically complete — straight lines extend forever. Since $H^2$ and $\mathbb{R}^2$ are diffeomorphic as smooth manifolds, geodesic completeness is genuinely a metric property, not merely a topological one. [/example] ### Existence of Minimal Geodesics Before stating the full Hopf–Rinow theorem, we establish its key geometric content: geodesic completeness at a single point implies that any two points are connected by a distance-realizing geodesic. [quotetheorem:2724] The proof proceeds by a bootstrap argument: start with a small geodesic sphere around $p$, find the optimal point on it, shoot a geodesic through it, and use a closed-set argument to show this geodesic reaches $q$ exactly. The key lemma identifies a "best starting direction." [quotetheorem:2725] [citeproof:2725]  [citeproof:2725] ### The Hopf–Rinow Theorem The theorem collects five conditions — two geodesic, one Heine–Borel, and two metric-space — and asserts they are all equivalent for connected Riemannian manifolds. [quotetheorem:2726] [citeproof:2726] The lemma identifies the optimal point $p_0$ on the geodesic sphere, but the interesting content is not its existence (which follows from compactness of the sphere and continuity of $d(\cdot, q)$) but its role in the bootstrap argument for the main theorem. The triangle equality $d(p, p_0) + d(p_0, q) = d(p,q)$ means the geodesic from $p$ through $p_0$ is already pointing in exactly the right direction to reach $q$ at minimal cost. The subsequent closed-set argument shows this optimal direction is maintained all the way to $q$. [remark: Completeness Conditions Are Not All Equivalent in General] The equivalence of the five conditions is special to Riemannian manifolds. In a general metric space, metric completeness does not imply the Heine–Borel property (e.g., infinite-dimensional Banach spaces are complete but their closed unit ball is not compact). For Riemannian manifolds, the Riemannian structure tightly constrains the geometry and forces all five notions to coincide. [/remark] The five-way equivalence in Hopf–Rinow may seem redundant, but each formulation is useful in a different context. Metric completeness (condition 5) is verified via Cauchy sequences and is natural in analysis. The Heine–Borel property (condition 4) is what compactness arguments require. Geodesic completeness (conditions 1–2) is what geometric constructions — parallel transport, holonomy, comparison theorems — demand. The single-point condition (condition 3) is the most surprising: completeness at one base point forces completeness everywhere, a striking rigidity statement. [explanation: What Hopf–Rinow Does and Does Not Say] The Hopf–Rinow theorem guarantees the **existence** of a minimal geodesic between any two points of a complete Riemannian manifold, but it does not guarantee **uniqueness**. On $S^2$ with the round metric, for example, the exponential map $\exp_p$ is defined on all of $T_p S^2$, confirming completeness, yet any two antipodal points are connected by infinitely many minimal geodesics (the great circle arcs). The theorem also highlights the role of completeness as a hypothesis in global theorems of Riemannian geometry. Results such as the Cartan–Hadamard theorem (on non-positive curvature) and the Bonnet–Myers theorem (on positive Ricci curvature) both require completeness as an input. Hopf–Rinow guarantees that completeness is a robust condition: it can be verified via any of the five equivalent formulations, and the most convenient for a given application can be chosen freely. The implication (3) $\implies$ (1) — completeness at a single point implies global completeness — is particularly striking. It means the global topology and geometry of $M$ are constrained by the behavior of geodesics from just one base point. [/explanation] ## Variations of Arc Length and Energy Jacobi field theory tells us what nearby geodesics look like infinitesimally — but is this picture consistent with the variational principle? If geodesics are critical points of arc length or energy, the Jacobi equation should emerge from the second variation of one of these functionals, and the objects that obstruct local minimization should be precisely the Jacobi fields vanishing at both endpoints. Getting this right requires choosing the correct functional: arc length is reparametrization-invariant, which makes its critical points non-isolated and its second variation degenerate, while the energy functional breaks this symmetry in a controlled way. This section develops both first and second variation formulas, identifies the right functional for variational analysis, and sets up the index form machinery that connects curvature to the global theorems of §6. ### The Energy Functional The arc length $\ell(\gamma) = \int_0^T |\dot{\gamma}(t)|\, dt$ is independent of reparametrization, which means its critical points cannot be isolated: any reparametrization of a critical point is again a critical point with the same length. This technical inconvenience motivates passing to a different functional. [definition: Energy of a Curve] Let $M$ be a Riemannian manifold and $\Omega(p, q)$ the space of piecewise $C^1$ curves from $p$ to $q$. The **energy functional** $E: \Omega(p, q) \to \mathbb{R}$ is defined by \begin{align*} E(\gamma) = \frac{1}{2} \int_0^T |\dot{\gamma}(t)|^2\, dt, \end{align*} where $\gamma: [0, T] \to M$. [/definition] Unlike length, energy does depend on reparametrization — it penalizes a curve for failing to travel at constant speed. As a result, the energy functional has isolated critical points, which is technically far more convenient for variational analysis. The following result clarifies the relationship between energy minimizers and length minimizers via the Cauchy–Schwarz inequality. [quotetheorem:2727] [citeproof:2727] The theorem is sharp in the following sense: minimizing energy among all parametrized curves from $p$ to $q$ defined on $[0,T]$ necessarily produces a constant-speed curve. In contrast, minimizing length among all curves from $p$ to $q$ (with any parametrization) does not single out a unique parametrization — the length functional is blind to speed. This is why the energy functional, rather than arc length, is the natural one for the Euler–Lagrange analysis of geodesics: energy has isolated critical points, each carrying a canonical parametrization. ### Variations and Variational Vector Fields To compute derivatives of $E$ and $\ell$ along a one-parameter family of curves, we introduce the notion of a variation. [definition: Variation of a Curve] A **smooth variation** of a curve $\gamma_0: [0, T] \to M$ is a smooth map $H: [0, T] \times (-\varepsilon, \varepsilon) \to M$ such that $H(t, 0) = \gamma_0(t)$ for all $t$. We write $\gamma_s(t) = H(t, s)$. The variation is **endpoint-fixing** if $H(0, s) = p$ and $H(T, s) = q$ for all $s$. The **variational vector field** of $H$ is the vector field along $\gamma_0$ defined by \begin{align*} Y(t) = \left.\frac{\partial H}{\partial s}\right|_{s = 0} = (dH)_{(t,0)}\frac{\partial}{\partial s}. \end{align*} [/definition] Conversely, every vector field $Y$ along $\gamma_0$ arises as the variational vector field of some variation. One explicit construction uses the exponential map: \begin{align*} H(t, s) = \exp_{\gamma_0(t)}(s Y(t)), \end{align*} which is valid on some neighbourhood of $[0, T] \times \{0\}$. If $Y(0) = 0 = Y(T)$, this variation fixes endpoints.  ### The First Variation Formula [quotetheorem:2728] [citeproof:2728] [remark: Boundary Terms] When the variation is endpoint-fixing, $Y(0) = Y(T) = 0$, so the boundary terms $g(Y(t), \dot{\gamma}(t))\big|_0^T$ in the first variation formula vanish. In that case the formula reduces to $\left.\frac{d}{ds}E\right|_{s=0} = -\int_0^T g(Y, \frac{\nabla}{dt}\dot{\gamma})\, dt$, and the geodesic equation $\frac{\nabla}{dt}\dot{\gamma} = 0$ is the precise condition for this to vanish for all $Y$. [/remark] ### The Second Variation Formula To determine whether a geodesic is a local minimum, maximum, or saddle point of energy or length, we need the second derivative of these functionals. The second variation introduces curvature explicitly. [quotetheorem:2729] [citeproof:2729] [remark: Fixed Endpoints] For endpoint-fixing variations, $Y(0) = Y(T) = 0$, so $\frac{\nabla Y}{ds}(t, 0)\big|_0^T$ vanishes when contracted with $\dot{\gamma}$. The second variation formulas then reduce to purely integral expressions: \begin{align*} \left.\frac{d^2}{ds^2} E(\gamma_s)\right|_{s=0} = \int_0^T \left(|Y'|^2 - R(Y, \dot{\gamma}, Y, \dot{\gamma})\right) dt. \end{align*} The sign of this expression — equivalently the interplay between $|Y'|^2$ (a stabilizing term) and the curvature term $R(Y, \dot{\gamma}, Y, \dot{\gamma})$ (which is positive when sectional curvatures are positive) — determines whether the geodesic is stable. This is precisely the structure that the index form analysis and the Bonnet–Myers theorem exploit in §6. [/remark] The competition between the $|Y'|^2$ and $R(Y, \dot{\gamma}, Y, \dot{\gamma})$ terms in the second variation is the geometric core of geodesic stability theory. When sectional curvatures are positive, the curvature term contributes positively to the second variation for suitable $Y$, making it possible to push the second variation negative — meaning the geodesic can be shortened. When curvatures are non-positive, the curvature term has the wrong sign to compete with $|Y'|^2$, and the geodesic remains locally minimizing. The first conjugate point is the precise moment at which the curvature term wins. [explanation: Role of the Index Form] The integrand $I(Y, Y) = \int_0^T (|Y'|^2 - R(Y, \dot{\gamma}, Y, \dot{\gamma}))\, dt$ appearing in the second variation of energy is called the **index form** of the geodesic $\gamma$. It is a symmetric bilinear form on vector fields along $\gamma$ vanishing at the endpoints. The Jacobi equation $J'' + R(J, \dot{\gamma})\dot{\gamma} = 0$ derived in §2 characterizes exactly the null space of the index form: a vector field $J$ satisfies $I(J, Y) = 0$ for all $Y$ if and only if $J$ is a Jacobi field vanishing at both endpoints, i.e., $\gamma(T)$ is conjugate to $\gamma(0)$ along $\gamma$. This connection explains why conjugate points, introduced in §2, are the critical obstruction to length minimization: a geodesic ceases to minimize length beyond the first conjugate point precisely because the index form acquires a negative direction there. The precise version of this statement, and its application to manifolds of positive curvature, is the subject of §6. [/explanation] ## Applications The variational machinery assembled in the preceding sections — the second variation formula, Jacobi fields, and the index form — is not merely an end in itself. In this section we harvest the rewards: three landmark theorems that connect the infinitesimal data of curvature to the global topological and metric structure of the manifold. Synge's theorem shows that positive sectional curvature forces simple connectivity in even dimensions. The Bonnet–Myers diameter theorem shows that a lower bound on Ricci curvature forces compactness and a finite diameter bound. The Hadamard–Cartan theorem shows that non-positive sectional curvature forces the exponential map to be a covering map, and in the simply-connected case yields a global diffeomorphism to $\mathbb{R}^n$. ### Synge's Theorem We begin with a remarkable result relating curvature to the fundamental group. [quotetheorem:2730] Before proving this, it is worth checking that all the hypotheses are genuinely necessary. [example: Necessity of the Hypotheses in Synge's Theorem] The real projective plane $\mathbb{RP}^2 = S^2/\{\pm 1\}$ with the metric induced from $S^2$ is compact with positive sectional curvature and even-dimensional — but it is not orientable, and it is not simply connected. So orientability cannot be dropped. The real projective space $\mathbb{RP}^3$, also with the round metric, is compact, orientable, and positively curved, but has odd dimension and is not simply connected. So the evenness of $\dim M$ cannot be dropped. The flat torus $\mathbb{T}^n = \mathbb{R}^n / \mathbb{Z}^n$ is compact, orientable, and even-dimensional, but has $K \equiv 0$. It is not simply connected. So the strict inequality $K > 0$ cannot be weakened to $K \geq 0$. [/example] The key ingredient in the proof is the following existence lemma. [quotetheorem:2731] [citeproof:2731] [citeproof:2731] ### Conjugate Points Recall from §2 that Jacobi fields measure the infinitesimal behaviour of nearby geodesics. They now serve as the right tool to characterise the precise moment at which a geodesic ceases to be locally length-minimising. We adopt the shorthand $J' = \nabla_{\dot{\gamma}} J$ for covariant differentiation along a curve $\gamma$. [definition: Conjugate Points] Let $\gamma: [a, b] \to M$ be a geodesic. Points $p = \gamma(\alpha)$ and $q = \gamma(\beta)$ are **conjugate along $\gamma$** if there exists a non-trivial Jacobi field $J$ along $\gamma$ with $J(\alpha) = 0 = J(\beta)$. [/definition] The definition is independent of reparametrisation because Jacobi fields transform covariantly under it. [quotetheorem:2732] [citeproof:2732] [remark: Dimension of Normal Jacobi Fields] The proof shows that for any Jacobi field with $J(0) = 0$, the condition $g(J'(0), \dot{\gamma}(0)) = 0$ is equivalent to $g(J(t), \dot{\gamma}(t))$ being identically zero. The space of normal Jacobi fields along $\gamma$ vanishing at $\gamma(0)$ has dimension $\dim M - 1$. [/remark] The intuition behind conjugate points is vivid on the sphere. [example: Conjugate Points on $S^2$] On $S^2 \subset \mathbb{R}^3$ with the round metric, the north pole $N = (0,0,1)$ and the south pole $S = (0,0,-1)$ are conjugate. Define a family of great circles by \begin{align*} f(t, s) &= \begin{pmatrix} \cos s \sin t \\ \sin s \sin t \\ \cos t \end{pmatrix}. \end{align*} For $s = 0$ this is the geodesic in the $(x,z)$-plane from $N$ to $S$. The variational field \begin{align*} J(t) = \left.\frac{\partial f}{\partial s}\right|_{s=0} = \begin{pmatrix} 0 \\ \sin t \\ 0 \end{pmatrix} \end{align*} is a Jacobi field vanishing at $t = 0$ (the north pole) and $t = \pi$ (the south pole), confirming that $N$ and $S$ are conjugate. Geometrically, all longitudes leave $N$ in different directions but converge again at $S$ — the geodesics focus at the antipodal point. [/example]  The significance of conjugate points for the length-minimising property is the following theorem. [quotetheorem:2733] [citeproof:2733] [remark: Smoothing the Variation] The variation $Y_\alpha$ constructed above is only piecewise smooth (with a corner at $t_0$). This can be repaired by a smooth approximation that perturbs $Y_\alpha$ slightly near $t_0$, which changes the second variation by an arbitrarily small amount and leaves the conclusion intact. [/remark] Together with the result from §5 that geodesics are locally minimising up to their first conjugate point, this gives a complete picture: a geodesic is locally length-minimising if and only if it contains no conjugate points in its interior. ### The Bonnet–Myers Diameter Theorem We now turn to a comparison theorem of a different kind: a lower bound on Ricci curvature forces a global upper bound on the diameter, and in particular forces compactness. [definition: Diameter] The **diameter** of a Riemannian manifold $(M, g)$ is \begin{align*} \operatorname{diam}(M, g) &= \sup_{p, q \in M} d(p, q). \end{align*} [/definition] The diameter may be infinite: any non-compact complete Riemannian manifold, such as $\mathbb{R}^n$ with the flat metric, has infinite diameter. On a compact manifold the diameter is always finite, since the distance function is continuous on the compact set $M \times M$. The Bonnet–Myers theorem gives a converse for a special class of manifolds: a purely infinitesimal condition on the Ricci curvature forces the diameter to be finite and bounded by an explicit constant depending only on the curvature bound and the dimension. [example: Diameter of a Round Sphere] For the sphere $S^{n-1}(r) = \{x \in \mathbb{R}^n : |x| = r\}$ with the round metric induced from $\mathbb{R}^n$, the diameter is $\operatorname{diam}(S^{n-1}(r)) = \pi r$, achieved by any pair of antipodal points. The sectional curvature of $S^{n-1}(r)$ is identically $1/r^2$, and the Ricci curvature satisfies $\operatorname{Ric} = \frac{n-2}{r^2} g$. [/example] We also need the notion of a Riemannian covering map, which will allow us to pass to the universal cover in the argument. [definition: Riemannian Covering Map] Let $(M, g)$ and $(\widetilde{M}, \tilde{g})$ be Riemannian manifolds. A smooth covering map $f: \widetilde{M} \to M$ is a **Riemannian covering map** if it is a local isometry, i.e., $f^* g = \tilde{g}$. We call $\widetilde{M}$ a Riemannian cover of $M$. [/definition] If $f: \widetilde{M} \to M$ is a universal Riemannian cover (so $\widetilde{M}$ is simply connected), then the fundamental group $\pi_1(M)$ is identified with the deck transformation group, and for any $p \in M$ the fibre $f^{-1}(p)$ is a torsor under $\pi_1(M)$. [quotetheorem:2734] The comparison model is the sphere $S^n(r)$, whose Ricci curvature is exactly $\frac{n-1}{r^2} g$ and whose diameter is exactly $\pi r$. The theorem says that any manifold with at least as much Ricci curvature is at least as compact. [citeproof:2734] [remark: Sharpness of the Ricci Hypothesis] The hypothesis cannot be weakened to $\operatorname{Ric} > 0$ (positive definiteness without a uniform lower bound). The paraboloid $\{(x,y,z) : z = x^2 + y^2\} \subset \mathbb{R}^3$ with the induced metric has positive Ricci curvature everywhere but is non-compact with infinite diameter, showing that a quantitative lower bound is essential. [/remark] ### The Hadamard–Cartan Theorem Non-positive sectional curvature has the opposite effect: geodesics spread apart rather than focus, there are no conjugate points, and the exponential map is a global diffeomorphism on the universal cover. The main technical input is a criterion for a local isometry to be a covering map. [quotetheorem:2735] [citeproof:2735] [quotetheorem:2736] [citeproof:2736] We can now state and prove the central result of this subsection. [quotetheorem:2737] The proof rests on the following lemma. [quotetheorem:2738] [citeproof:2738] [citeproof:2738] The Hadamard–Cartan theorem is a striking global rigidity statement: non-positive curvature not only prevents conjugate points, it forces the universal cover to be diffeomorphic to $\mathbb{R}^n$. This completes the picture of how curvature controls global structure: positive curvature compresses (Bonnet–Myers, Synge), while non-positive curvature expands and flattens the topology of the universal cover. From local variational principles and global rigidity theorems (Hopf-Rinow, Bonnet-Myers, Synge), we shift to the analytic perspective: harmonic differential forms and Hodge's decomposition theorem, tied together through curvature via Bochner-Weitzenböck identities. # 4. Hodge theory on Riemannian manifolds This chapter develops Hodge theory on Riemannian manifolds — the analytic framework that turns the topological invariants of de Rham cohomology into concrete geometric objects. We define the Hodge star operator and the codifferential, prove the Hodge decomposition theorem giving every form a unique split into exact, coexact, and harmonic pieces, and use the divergence operator to relate Hodge theory to differential geometry on functions. The chapter culminates in Bochner-Weitzenböck-type formulas, which connect curvature to harmonic forms and yield striking vanishing theorems. ## Hodge Star and Operators [motivation] ### Why Hodge Theory? The exterior derivative $d: \Omega^p(M) \to \Omega^{p+1}(M)$ is a natural operator on differential forms, but it only goes in one direction — it increases degree. On a Riemannian manifold, the metric $g$ provides additional structure that allows us to go in the opposite direction, producing a "formal adjoint" $\delta: \Omega^p(M) \to \Omega^{p-1}(M)$. Combining these gives the Laplace–Beltrami operator $\Delta = d\delta + \delta d$, which generalises the classical Laplacian from $\mathbb{R}^n$ to curved spaces. ### The Bridge Between Topology and Analysis The kernel of $\Delta$ — the harmonic forms — will turn out (in the Hodge Decomposition Theorem of the next section) to give canonical representatives for de Rham cohomology classes. This remarkable fact connects the analytic property of being harmonic to purely topological information. The Hodge star operator is the key algebraic device that makes all of this possible, converting a metric into a way of "turning forms around" in degree. [/motivation] Throughout this chapter, all manifolds are assumed to be oriented. We fix an orientation by a non-vanishing form $\varepsilon \in \Omega^n(M)$, where $n = \dim M$. ### The Volume Form A smooth manifold, on its own, carries no canonical measure: to integrate a function one must choose a volume form, and any two choices differ by a positive scalar function. There is no intrinsic way to single one out. The Riemannian metric resolves this ambiguity. Once we fix an orientation, the metric imposes a rigid length scale on tangent vectors, and we can demand that our volume form assigns unit volume to any orthonormal basis of each tangent space — a condition that uniquely pins down the form. On a Riemannian manifold $(M, g)$, the metric therefore allows us to single out a canonical $n$-form. Working in a coordinate patch $U \subseteq M$, apply Gram–Schmidt to obtain a positively-oriented orthonormal frame $e_1, \ldots, e_n$ for $TU$. The dual coframe $\omega_1, \ldots, \omega_n \in \Omega^1(U)$, defined by $\omega_i(e_j) = \delta_{ij}$, consists of linearly independent $1$-forms, and their wedge product satisfies \begin{align*} \omega_1 \wedge \cdots \wedge \omega_n = a\,\varepsilon \end{align*} for some $a \in C^\infty(U)$ with $a > 0$. This construction is independent of the choice of orthonormal frame: any two positively-oriented orthonormal coframes differ by a matrix $\Phi \in SO(n)$, and $\det(\Phi) = 1$ ensures \begin{align*} \omega_1' \wedge \cdots \wedge \omega_n' = \det(\Phi)\,\omega_1 \wedge \cdots \wedge \omega_n = \omega_1 \wedge \cdots \wedge \omega_n. \end{align*} These local forms therefore agree on overlapping charts and patch together to give a globally defined $n$-form. The volume form is not merely a convenience: it is the unique metric-adapted measure that makes integration of functions on $M$ well-defined and independent of any auxiliary choice. In coordinates it provides the familiar $\sqrt{|g|}\,dx_1\wedge\cdots\wedge dx_n$ factor, which is the Jacobian correcting for the non-Euclidean metric. [definition: Volume Form] Let $(M^n, g)$ be an oriented Riemannian manifold. The **Riemannian volume form** $\omega_g \in \Omega^n(M)$ is the unique $n$-form that, on any coordinate patch with positively-oriented orthonormal coframe $\omega_1, \ldots, \omega_n$, equals \begin{align*} \omega_g = \omega_1 \wedge \cdots \wedge \omega_n. \end{align*} It depends only on the metric $g$ and the chosen orientation. [/definition] In local coordinates $(x_1, \ldots, x_n)$ with metric $g = g_{ij}\,dx_i \otimes dx_j$, the volume form takes the familiar coordinate expression \begin{align*} \omega_g = \sqrt{|g|}\,dx_1 \wedge \cdots \wedge dx_n, \end{align*} where $|g| = \det(g_{ij})$. ### The Inner Product on Forms To define the Hodge star operator, we need to compare $p$-forms of equal degree pointwise. The metric $g$ gives an inner product on each tangent space $T_xM$, but this does not directly produce an inner product on $\Lambda^p T_x^*M$: the exterior power is a higher-order algebraic construction, and it is not obvious how to extend a bilinear pairing from vectors to multi-vectors. The question is how to induce a canonical inner product on $\Lambda^p T^*M$ from $g$. The answer is to use the fact that $g$ identifies an orthonormal basis of $T_xM$ with an orthonormal basis of $T_x^*M$, and then declare the wedge products of these dual basis elements to be an orthonormal basis for $\Lambda^p T_x^*M$. On a fiber at $x \in M$, with orthonormal coframe $\omega_1, \ldots, \omega_n$, the collection \begin{align*} \{\,\omega_{i_1} \wedge \cdots \wedge \omega_{i_p} : 1 \leq i_1 < \cdots < i_p \leq n\,\} \end{align*} forms an orthonormal basis of $\Lambda^p T_x^*M$. This makes the volume form a form of unit length, since $\langle \omega_g, \omega_g \rangle_g = 1$ at every point. ### The Hodge Star Operator With a metric in hand, we can now "swap" a $p$-form for an $(n-p)$-form in a canonical way. [definition: Hodge Star] The **Hodge star operator** on $(M^n, g)$ is the bundle map \begin{align*} \star: \Lambda^p T_x^*M \to \Lambda^{n-p} T_x^*M \end{align*} characterised by the identity: for all $\alpha, \beta \in \Lambda^p T_x^*M$, \begin{align*} \alpha \wedge \star\beta = \langle \alpha, \beta \rangle_g\,\omega_g. \end{align*} [/definition] Since $g$ is non-degenerate, this identity uniquely determines $\star\beta$ for each $\beta$: the right-hand side is a linear functional in $\alpha$ on $\Lambda^p T_x^*M$, and non-degeneracy of $g$ on this space ensures there is a unique element $\star\beta \in \Lambda^{n-p} T_x^*M$ producing that functional via the wedge pairing. To compute $\star$ explicitly, one works in an orthonormal basis. [quotetheorem:2739] [citeproof:2739] Two features of this theorem deserve emphasis. Orientation is essential: the ordering $\omega_{i_1} \wedge \cdots \wedge \omega_{i_p} \wedge \omega_{j_1} \wedge \cdots \wedge \omega_{j_{n-p}} = \omega_g$ in the hypothesis selects a specific sign, and a different orientation choice would negate $\omega_g$ and hence negate $\star$. Orthonormality is equally essential: the computation breaks down for a non-orthonormal coframe because the inner product $\langle \omega_J, \omega_K \rangle_g$ would no longer equal $\delta_{JK}$, producing off-diagonal cross-terms that invalidate the clean formula. If either condition is dropped, one can still define $\star$ via the universal characterisation $\alpha \wedge \star\beta = \langle \alpha, \beta \rangle_g \omega_g$, but explicit computation requires extra work with the Gram matrix. Applying $\star$ twice gives a scalar endomorphism. [quotetheorem:2740] [citeproof:2740] Two special cases are worth noting immediately: \begin{align*} \star 1 = \omega_g, \qquad \star\omega_g = 1. \end{align*}  ### The Co-differential Using $\star$, we can define an operator that decreases the degree of a form, serving as a counterpart to $d$. [definition: Co-differential] The **co-differential** $\delta: \Omega^p(M) \to \Omega^{p-1}(M)$ is defined by \begin{align*} \delta = \begin{cases} (-1)^{n(p+1)+1}\,\star\,d\,\star & p \neq 0 \\ 0 & p = 0. \end{cases} \end{align*} [/definition] The specific sign $(-1)^{n(p+1)+1}$ is chosen so that $\delta$ is the formal $L^2$ adjoint of $d$, as we will see in the proposition below. The formula makes clear that $\delta$ is built from $d$ and $\star$: it converts a $p$-form to an $(n-p)$-form via $\star$, differentiates with $d$, then converts back with $\star$. Since $d^2 = 0$ and $\star$ is an isomorphism on each fiber, it follows immediately that $\delta^2 = 0$. ### The Laplace–Beltrami Operator [definition: Laplace–Beltrami Operator] The **Laplace–Beltrami operator** (or **Hodge Laplacian**) is \begin{align*} \Delta = d\delta + \delta d: \Omega^p(M) \to \Omega^p(M). \end{align*} [/definition] [quotetheorem:2741] [citeproof:2741] The commutativity $\star\Delta = \Delta\star$ is more than an algebraic curiosity. It says that $\star$ maps harmonic $p$-forms to harmonic $(n-p)$-forms, providing the analytic underpinning of Poincaré duality: since $\mathcal{H}^p \cong H^p_{\mathrm{dR}}(M)$ and $\mathcal{H}^{n-p} \cong H^{n-p}_{\mathrm{dR}}(M)$, the isomorphism $\star: \mathcal{H}^p \to \mathcal{H}^{n-p}$ recovers the classical duality $H^p_{\mathrm{dR}}(M) \cong H^{n-p}_{\mathrm{dR}}(M)$. Without orientation there is no $\star$, and without $\star$ this transfer of analytic information between complementary degrees is lost: questions about degree $p$ and degree $n-p$ decouple entirely. **Example: The Laplacian on flat $\mathbb{R}^n$.** Take $(M, g) = (\mathbb{R}^n, g_{\mathrm{eucl}})$ and $p = 0$. For $f \in C^\infty(\mathbb{R}^n)$, one has $\delta f = 0$ (since $\delta = 0$ on $0$-forms), and $\Delta f = \delta df$. A direct computation in coordinates shows \begin{align*} \Delta f = -\frac{\partial^2 f}{\partial x_1^2} - \cdots - \frac{\partial^2 f}{\partial x_n^2}. \end{align*} This is the usual Laplacian up to the overall minus sign. The sign convention here is chosen so that $\Delta$ is a non-negative operator (its $L^2$ eigenvalues are non-negative), which is the geometrically natural normalisation. **Example: The Laplacian in local coordinates.** For a metric $g = g_{ij}\,dx_i \otimes dx_j$ on a coordinate patch, the volume form is $\omega_g = \sqrt{|g|}\,dx_1 \wedge \cdots \wedge dx_n$. For $f \in C^\infty(M)$, \begin{align*} \Delta_g f = -\frac{1}{\sqrt{|g|}}\,\partial_j\!\left(\sqrt{|g|}\,g^{ij}\,\partial_i f\right), \end{align*} where $g^{ij}$ denotes the inverse metric tensor. The leading term is $-g^{ij}\partial_i\partial_j f$, with lower-order correction terms arising from the variation of $\sqrt{|g|}$. ### The Co-differential as Formal Adjoint To make the adjoint relationship precise, we define a global inner product on forms. [definition: L Squared Inner Product on Forms] For $\xi, \eta \in \Omega^p(M)$, the **$L^2$ inner product** is \begin{align*} \langle\langle \xi, \eta \rangle\rangle_g = \int_M \langle \xi, \eta \rangle_g\,\omega_g. \end{align*} This is well-defined when $M$ is compact (or when the forms have compact support). [/definition] [quotetheorem:2742] [citeproof:2742] The specific sign in the definition of $\delta$ is precisely what makes the signs in this proof work out correctly. [quotetheorem:2743] [citeproof:2743] Self-adjointness of $\Delta$ is the gateway to the Hodge decomposition. The proof of the decomposition theorem relies on the orthogonal complement $\Delta\Omega^p(M) = (\mathcal{H}^p)^\perp$, which is a statement in the $L^2$ inner product space $\Omega^p(M)$ — and self-adjointness is precisely what makes this orthogonality hold. The hypotheses of compact support (or compact $M$) are crucial: without them the integration by parts in the adjoint computation for $\delta$ produces boundary terms, self-adjointness fails, and the decomposition breaks down. On a non-compact manifold without boundary conditions one can have $\langle\langle \Delta\alpha, \beta \rangle\rangle \neq \langle\langle \alpha, \Delta\beta \rangle\rangle$ for compactly supported $\alpha, \beta$ supported near infinity. ### Harmonic Forms The Laplacian $\Delta = d\delta + \delta d$ is a second-order differential operator on forms, and its kernel is the space of harmonic forms. But the kernel of a differential operator can easily be trivial or infinite-dimensional, and there is no reason, a priori, for it to carry special topological significance. What makes harmonic forms remarkable is that on a compact manifold they are precisely the forms that are simultaneously closed and co-closed — they minimise the $L^2$ norm within their cohomology class. If one collapses the kernel into a single definition and moves on, this geometric meaning is lost. By isolating harmonic forms as a distinct class, we can later show (Hodge Decomposition Theorem) that they are canonical representatives for de Rham cohomology: every cohomology class contains exactly one harmonic form, and the count of linearly independent harmonic forms is a topological invariant of the manifold. [definition: Harmonic Forms] A **harmonic $p$-form** is a $p$-form $\omega \in \Omega^p(M)$ such that $\Delta\omega = 0$. The space of harmonic $p$-forms is denoted \begin{align*} \mathcal{H}^p = \{\alpha \in \Omega^p(M) : \Delta\alpha = 0\}. \end{align*} [/definition] On a compact manifold, the condition $\Delta\alpha = 0$ splits cleanly. [quotetheorem:2744] [citeproof:2744] [remark: Harmonic Functions on Compact Manifolds] In degree $p = 0$, the co-differential $\delta$ vanishes identically, so co-closedness is automatic. The theorem then says: a function $f \in C^\infty(M)$ is harmonic if and only if $df = 0$, i.e., $f$ is locally constant. Since $M$ is connected and compact, this means $f$ must be a global constant. This recovers the classical principle that harmonic functions on compact manifolds without boundary are constant. By contrast, $\mathbb{R}^n$ admits many non-constant harmonic functions (e.g., any linear function $x \mapsto c_1 x_1 + \cdots + c_n x_n$), showing that compactness is essential. [/remark] ### Simplifications in Even Dimensions When $n = \dim M = 2m$ is even, the double-star formula gives $\star\star = (-1)^p$ on $p$-forms. In this case, the co-differential simplifies to \begin{align*} \delta = -\star d\star \qquad (p \neq 0). \end{align*} This applies in particular to complex manifolds viewed as real manifolds: $\mathbb{C}^m \cong \mathbb{R}^{2m}$ with the Hermitian metric provides the canonical example. The interaction between the Hodge theory developed here and the complex structure will be explored further in the context of Kähler manifolds. ## Hodge Decomposition Theorem The goal of this section is to prove the central result of Hodge theory: every smooth $k$-form on a compact oriented Riemannian manifold decomposes, orthogonally and uniquely, into an exact part, a coexact part, and a harmonic part. This is a deep bridge between analysis and topology — the analytic data of the Laplacian turns out to encode purely topological invariants. ### Statement and First Consequences [quotetheorem:2745] Compactness of $M$ is essential here. On a non-compact manifold, the space of harmonic forms can be infinite-dimensional and the decomposition can fail entirely. By splitting $\Delta = d\delta + \delta d$, we immediately obtain a finer decomposition. [quotetheorem:2746] [citeproof:2746] ### Harmonic Representatives of de Rham Cohomology The refined decomposition has a striking consequence for de Rham cohomology. A closed form $\alpha$ satisfies $d\alpha = 0$; in the decomposition $\alpha = \alpha_1 + d\alpha_2 + \delta\alpha_3$, applying $d$ gives $d\delta\alpha_3 = 0$, which forces $\delta\alpha_3 = 0$ (take the inner product with $\delta\alpha_3$ itself). So every closed form is cohomologous to its harmonic component $\alpha_1 \in \mathcal{H}^p$. [quotetheorem:2747] [citeproof:2747] This result is remarkable because the left-hand side is analytic — $\mathcal{H}^p$ is defined by a PDE, the Laplace equation $\Delta\alpha = 0$ — while the right-hand side $H_{\mathrm{dR}}^p(M)$ is a homotopy invariant of $M$, independent of the metric. By de Rham's theorem, $H_{\mathrm{dR}}^p(M) \cong H^p(M; \mathbb{R})$ (singular cohomology), so the Hodge theorem gives a canonical metric-dependent way to select representatives for purely topological cohomology classes. [remark: Variational Interpretation] There is a clean variational reason to expect this theorem. Fix a cohomology class $a \in H_{\mathrm{dR}}^p(M)$ and consider the problem of minimizing $\|\xi\|_g^2$ over all closed $p$-forms $\xi$ with $[\xi] = a$. Any minimizer must be stationary under perturbations $\xi \mapsto \xi + t\, d\beta$; differentiating gives $\langle\langle \xi, d\beta \rangle\rangle_g = \langle\langle \delta\xi, \beta \rangle\rangle_g = 0$ for all $\beta$, so $\delta\xi = 0$. Combined with $d\xi = 0$, the minimizer is harmonic. The Hodge theorem confirms that such a minimizer exists and is unique. [/remark] ### Proof of the Hodge Decomposition Theorem The proof of the Hodge decomposition theorem requires three analytic ingredients, which we now state and then use to complete the argument. The first ingredient controls sequences of forms with bounded Laplacian. [quotetheorem:2748] The space $\Omega^p(M)$ is not complete, so we cannot conclude the subsequence converges in $\Omega^p(M)$ — only that it is Cauchy. This is enough for our purposes. [quotetheorem:2749] [citeproof:2749] The second ingredient converts weak solutions into genuine smooth solutions. [definition: Weak Solution] A **weak solution** to the equation $\Delta\omega = \alpha$ (for $\alpha \in \Omega^p(M)$ given) is a bounded linear functional $\ell: \Omega^p(M) \to \mathbb{R}$ satisfying \begin{align*} \ell(\Delta\varphi) = \langle\langle \alpha, \varphi \rangle\rangle_g \quad \text{for all } \varphi \in \Omega^p(M). \end{align*} Here "bounded" means there exists $C > 0$ such that $|\ell(\beta)| \leq C\|\beta\|$ for all $\beta \in \Omega^p(M)$. [/definition] The motivation for this definition comes from self-adjointness: if $\omega$ is a genuine solution, then $\langle\langle \omega, \Delta\varphi \rangle\rangle_g = \langle\langle \Delta\omega, \varphi \rangle\rangle_g = \langle\langle \alpha, \varphi \rangle\rangle_g$, so the functional $\ell(\cdot) = \langle\langle \omega, \cdot \rangle\rangle_g$ is a weak solution. The Regularity Theorem reverses this: [quotetheorem:2750] The third ingredient is a classical extension theorem. [quotetheorem:2751] We can now carry out the proof of the main theorem. [citeproof:2751] ### The Flat Torus: A Concrete Verification The abstract proof above appeals to regularity and compactness theorems from PDE theory. On the flat torus, the decomposition can be verified directly using Fourier series, which makes every step fully explicit. Let $M = T^n = \mathbb{R}^n / (2\pi\mathbb{Z})^n$ carry the flat metric. Since the cotangent bundle is trivialised by $\{dx_{i_1} \wedge \cdots \wedge dx_{i_p}\}$, a $p$-form is just a collection of $2\pi$-periodic scalar functions as coefficients. The Laplacian acts coefficient-by-coefficient: \begin{align*} \Delta\bigl(\alpha\, dx_{i_1} \wedge \cdots \wedge dx_{i_p}\bigr) = -\sum_{i=1}^n \frac{\partial^2 \alpha}{\partial x_i^2}\, dx_{i_1} \wedge \cdots \wedge dx_{i_p}. \end{align*} So Hodge decomposition for $p$-forms on $T^n$ reduces to Hodge decomposition for functions, i.e., to the case $p = 0$. For functions, Fourier theory gives a complete picture. Every $\varphi \in C^\infty(T^n)$ expands as \begin{align*} \varphi(x) = \sum_{k \in \mathbb{Z}^n} \hat\varphi_k\, e^{ik \cdot x}, \qquad \hat\varphi_k = \frac{1}{(2\pi)^n}\int_{T^n} \varphi(x)\, e^{-ik \cdot x}\, dx, \end{align*} with uniform convergence of all derivatives. The Laplacian diagonalises in this basis: \begin{align*} \widehat{\Delta\varphi}_k = -|k|^2 \hat\varphi_k. \end{align*} From this, the Hodge decomposition is transparent. The harmonic functions are exactly those with $\hat\varphi_k = 0$ for all $k \neq 0$, i.e., the constants: \begin{align*} \mathcal{H}^0 = \bigl\{\varphi \in C^\infty(T^n) : \hat\varphi_k = 0 \text{ for all } k \neq 0\bigr\} = \{\text{constants}\}. \end{align*} The orthogonal complement consists of zero-mean functions: \begin{align*} (\mathcal{H}^0)^\perp = \bigl\{\varphi \in C^\infty(T^n) : \hat\varphi_0 = 0\bigr\}. \end{align*} For any $\alpha \in (\mathcal{H}^0)^\perp$, the equation $\Delta\omega = \alpha$ is solved explicitly: since $-|k|^2 \neq 0$ for $k \neq 0$, we can define \begin{align*} \hat\omega_k = \frac{\hat\alpha_k}{-|k|^2} \quad (k \neq 0), \qquad \hat\omega_0 = 0. \end{align*} The decay condition $\hat\alpha_k = o(|k|^{-m})$ for all $m$ (guaranteed by smoothness) ensures $\hat\omega_k$ also decays rapidly, so $\omega \in C^\infty(T^n)$. This gives $(\mathcal{H}^0)^\perp = \Delta C^\infty(T^n)$, verifying the Hodge decomposition directly. The mechanism is the same as in the general proof: the Laplacian is a bijection on the complement of its kernel, and the key analytic facts (the Compactness Theorem, the Regularity Theorem) reduce in this flat case to the completeness and invertibility properties of Fourier series. ## Divergence The Hodge codifferential $\delta$ on 1-forms was defined abstractly as $\star\,d\,\star$ up to sign, but to compute it and connect it to classical analysis one needs a more concrete handle. The natural candidate is the divergence operator, but why should divergence appear in a chapter about differential forms? The answer is that on a Riemannian manifold the metric identifies 1-forms with vector fields, and the divergence — defined intrinsically via the trace of the covariant derivative — turns out to equal $-\delta$ under this identification. Establishing this link requires first building the right notion of divergence. On a Riemannian manifold, the Levi-Civita connection $\nabla$ allows us to define an intrinsic version: given $X \in \mathfrak{X}(M)$, the covariant derivative $\nabla X$ is a section of $TM \otimes T^*M = \operatorname{End}(TM)$, and the trace of an endomorphism is basis-independent. [definition: Divergence of a Vector Field] Let $(M, g)$ be an oriented Riemannian manifold and $X \in \mathfrak{X}(M)$. The **divergence** of $X$ is \begin{align*} \operatorname{div} X = \operatorname{tr}(\nabla X). \end{align*} [/definition] In terms of a local frame field $\{e_i\}$, the trace unfolds as \begin{align*} \operatorname{div} X = \sum_{i=1}^n g(\nabla_{e_i} X, e_i). \end{align*} This formula reduces to $\sum_i \partial_{x_i} X_i$ in an orthonormal coordinate frame on $\mathbb{R}^n$, confirming that the Riemannian definition extends the classical one. The divergence satisfies a Leibniz rule with respect to scalar multiplication by smooth functions. [quotetheorem:2752] [citeproof:2752] ### The Interior Product To establish the relationship between divergence and the codifferential, it is convenient to have the interior product, which contracts a form against a vector field. [definition: Interior Product] Let $X \in \mathfrak{X}(M)$. The **interior product** $i(X): \Omega^p(M) \to \Omega^{p-1}(M)$ is defined by \begin{align*} (i(X)\psi)(Y_1, \ldots, Y_{p-1}) = \psi(X, Y_1, \ldots, Y_{p-1}). \end{align*} This is also written $i(X)\psi = X \lrcorner \psi$. [/definition] [remark: Interior Product and Anti-symmetry] When computing $i(X)(\alpha \wedge \beta)$, one must track signs carefully: the result is not simply $\alpha(X)\beta$, because that would break the anti-symmetry of the wedge product. The correct expansion follows from the definition and the alternating nature of forms. [/remark] ### Divergence and the Riemannian Volume Form A key structural lemma connects the divergence to the exterior derivative via the Riemannian volume form $\omega_g$. [quotetheorem:2753] [citeproof:2753] This identity is the geometric heart of the divergence theorem: it says that $(\operatorname{div} X)\,\omega_g$ is an exact $n$-form, so its integral over a closed manifold vanishes by Stokes' theorem. [quotetheorem:2754] [citeproof:2754] The role of closedness here is not incidental: the argument uses Stokes' theorem, which requires $\int_M d\omega = 0$, and this holds only when $M$ has no boundary. On a non-compact manifold the form $i(X)\omega_g$ may have non-trivial support extending to infinity, and the boundary term at infinity need not vanish — the integral $\int_M \operatorname{div}(X)\,\omega_g$ can then be non-zero. The proof that $\delta\theta = -\operatorname{div} X_\theta$ in the next section also depends on integrating $\operatorname{div}(fX_\theta)$ against $\omega_g$ and setting it to zero; this integration by parts is again only valid on closed manifolds (or for compactly supported data). Thus closedness is not merely a technical convenience but the precise analytic hypothesis that makes $\delta = -\operatorname{div}$ a global identity. [remark: Divergence Theorem with Boundary] When $M$ has boundary $\partial M$, the same computation yields the familiar form: $\int_M \operatorname{div}(X)\,\omega_g = \int_{\partial M} i(X)\omega_g$. The right-hand side is the flux of $X$ through the boundary. [/remark] ### Divergence as the Codifferential The divergence is not an independent operator: it is the codifferential $\delta$ in disguise. Recall from §1 that $\delta: \Omega^p(M) \to \Omega^{p-1}(M)$ is the formal $L^2$-adjoint of $d$. [quotetheorem:2755] [citeproof:2755] This theorem shows that $\delta$ on 1-forms is the negative divergence of the metrically dual vector field. It is not merely an abstract identification: it gives us explicit local formulas for $\delta$ that will be indispensable in the next section on Bochner's method. ### Explicit Coordinate Formulas for $\delta$ [quotetheorem:2756] [citeproof:2756] [example: Divergence on $\mathbb{R}^n$] Take $M = \mathbb{R}^n$ with the Euclidean metric and $\theta = \sum_i f_i\,dx_i$ a 1-form with smooth coefficients. The standard basis $\{e_k = \partial_{x_k}\}$ is a global orthonormal frame and the Levi-Civita connection satisfies $\nabla_{e_k}(f_i\,dx_i) = \frac{\partial f_i}{\partial x_k}dx_i$. Therefore \begin{align*} \delta\theta = -\sum_{k=1}^n \langle\nabla_{e_k}\theta, e_k\rangle = -\sum_{k=1}^n \frac{\partial f_k}{\partial x_k} = -\operatorname{div}(f_1, \ldots, f_n). \end{align*} This confirms that $\delta$ on 1-forms is exactly the negative classical divergence when $M$ is flat. [/example] For the Bochner technique in §4, we will also need an analogous formula for 2-forms. The proof of the following proposition is on the example sheet and parallels the argument above. [quotetheorem:2757] The proof for general $p$-forms follows the same pattern: $\delta$ on a $p$-form is the negative divergence in the sense of summing covariant derivatives contracted against a frame. These explicit formulas are the starting point for Bochner's method, where one computes $\Delta = d\delta + \delta d$ in terms of curvature to derive vanishing theorems. ## Introduction to Bochner's Method The Hodge decomposition theorem, developed in the preceding sections, identifies the de Rham cohomology group $H^k_{\mathrm{dR}}(M)$ with the kernel of the Hodge Laplacian $\Delta$. This identification turns topological questions about cohomology into analytic questions about differential operators. Bochner's method is a systematic technique for exploiting this bridge: by decomposing $\Delta$ into a manifestly non-negative piece and a curvature term, one can force harmonic forms to vanish whenever the curvature is sufficiently positive. ### The Strategy: Splitting the Laplacian The motivating question is: when does $H^1_{\mathrm{dR}}(M) = 0$? By Hodge theory, this is equivalent to showing that every harmonic $1$-form is zero. To prove $\Delta \alpha \neq 0$ for all $\alpha \neq 0$, it suffices to show the $L^2$ inner product \begin{align*} \langle\langle \alpha, \Delta \alpha \rangle\rangle > 0 \end{align*} for all non-zero $\alpha$. The key idea is to write \begin{align*} \Delta = T^* T + C \end{align*} for some operator $T$ with formal adjoint $T^*$, chosen so that $C$ has a geometric meaning. Since $\langle\langle T^* T \alpha, \alpha \rangle\rangle = \|T\alpha\|_2^2 \geq 0$, if the operator $C$ is also non-negative then so is $\Delta$, and no non-zero harmonic form can exist. The natural choice of $T$ is the covariant derivative $\nabla$, and the resulting operator $C$ turns out to be the Ricci curvature. ### The Covariant Laplacian The construction works in the following generality. Let $(M, g)$ be a Riemannian manifold, $E \to M$ a vector bundle with inner product $h$, and $\nabla^E : \Gamma(E) \to \Omega^1(E)$ a connection on $E$. The formal adjoint $(\nabla^E)^* : \Omega^1(E) \to \Gamma(E)$ is defined by the $L^2$ adjoint relation \begin{align*} \int_M \langle \nabla^E \alpha, \beta \rangle_{E,g}\, \omega_g = \int_M \langle \alpha, (\nabla^E)^* \beta \rangle_E\, \omega_g \end{align*} for all $\alpha \in \Gamma(E)$ and $\beta \in \Omega^1(E)$. Since $h$ is non-degenerate, this defines $(\nabla^E)^*$ uniquely. [definition: Covariant Laplacian] Let $(M, g)$ be a Riemannian manifold, $E \to M$ a vector bundle with metric $h$, and $\nabla^E$ a connection on $E$. The **covariant Laplacian** (also called the **connection Laplacian** or **rough Laplacian**) is the operator \begin{align*} \nabla^* \nabla : \Gamma(E) \to \Gamma(E) \end{align*} defined by composing $\nabla^E$ with its formal adjoint $(\nabla^E)^*$. [/definition] For the case $E = T^*M$ with the Levi-Civita connection, there is a concrete formula in terms of a local orthonormal frame. [quotetheorem:2758] [citeproof:2758] As an immediate corollary, the covariant Laplacian takes the form \begin{align*} \nabla^* \nabla \alpha = -\sum_{i=1}^n \nabla_{e_i} \nabla_{e_i} \alpha \end{align*} for any local orthonormal frame $\{e_i\}$. This is the rough Laplacian: it looks like the sum of second-order covariant derivatives in each frame direction, with no curvature correction. ### The Ricci Endomorphism The Bochner–Weitzenböck formula will express the Hodge Laplacian $\Delta$ as the sum of the covariant Laplacian $\nabla^*\nabla$ and a curvature term. That curvature term needs to act on $1$-forms as a linear endomorphism, but the Ricci tensor is a $(0,2)$-tensor — a symmetric bilinear form on $TM$. As it stands, $\operatorname{Ric}$ eats two vector fields and returns a scalar; it does not act on $1$-forms. The question is how to canonically convert a $(0,2)$-tensor into an endomorphism of $\Omega^1(M)$. The metric provides the canonical way: it identifies $1$-forms with vector fields via index raising, so one can feed $\operatorname{Ric}$ the metric-dual vector field of a given $1$-form. The result is a $(0,1)$-tensor, i.e., another $1$-form. The Ricci curvature $\operatorname{Ric}$ is a priori a symmetric bilinear form on $TM$. To make it act on $1$-forms, one uses the metric to raise an index. Given $\alpha \in \Omega^1(M)$, let $X_\alpha \in \mathfrak{X}(M)$ be the unique vector field satisfying $\alpha(Z) = g(X_\alpha, Z)$ for all $Z \in \mathfrak{X}(M)$ (the metric dual of $\alpha$). Then define $\operatorname{Ric}(\alpha) \in \Omega^1(M)$ by \begin{align*} \operatorname{Ric}(\alpha)(X) = \operatorname{Ric}(X, X_\alpha). \end{align*} In local coordinates, this corresponds to contracting $\operatorname{Ric}_{ij}$ with $g^{jk} \alpha_k$, i.e., the operation $\alpha_k \mapsto \sum_{i,j} g^{jk} \operatorname{Ric}_{ij} \alpha_k$. With this definition, $\operatorname{Ric}$ becomes a linear endomorphism of $\Omega^1(M)$, and the Bochner–Weitzenböck formula can be stated as a clean operator identity. ### The Bochner–Weitzenböck Formula [quotetheorem:2759] [citeproof:2759] [remark: Pointwise Argument] The proof computes at a single point $p$ using a normal frame, so the result $\nabla e_k|_p = 0$ eliminates many terms. Since $p$ was arbitrary, the identity holds everywhere on $M$. [/remark] ### Vanishing Theorems The Bochner–Weitzenböck formula has powerful topological consequences when the Ricci curvature is sign-definite. Taking the $L^2$ inner product with $\alpha$: \begin{align*} \langle\langle \Delta\alpha, \alpha \rangle\rangle = \|\nabla\alpha\|_2^2 + \int_M \operatorname{Ric}(\alpha, \alpha)\,\omega_g. \end{align*} If $\alpha$ is harmonic, then $\Delta\alpha = 0$, so both $\|\nabla\alpha\|_2^2$ and $\int_M \operatorname{Ric}(\alpha,\alpha)\,\omega_g$ are constrained by their sum being zero. [quotetheorem:2760] [citeproof:2760] [remark: Betti Number Bound] The bound $b^1(M) \leq n$ in Part (2) is sharp: the $n$-torus $T^n = (S^1)^n$ with the flat metric satisfies $\operatorname{Ric} = 0$ and $b^1 = n$. [/remark] ### The Flat Torus Rigidity When $b^1(M) = n$ and $\operatorname{Ric} \geq 0$, the manifold is not merely flat — it is globally isometric to a flat torus. This sharpening requires some topology, and the argument is worth recording. [quotetheorem:2761] [citeproof:2761] ### The General Weitzenböck Formula for $p$-Forms The Bochner–Weitzenböck formula extends to differential forms of arbitrary degree. For $E = \bigwedge^p T^*M$, the covariant derivative $\nabla: \Omega^0(E) \to \Omega^1(E)$ has a formal adjoint $\nabla^*$, and for any $\alpha \in \Omega^p(M)$: \begin{align*} \Delta\alpha = \nabla^*\nabla\alpha + \mathfrak{R}(\alpha), \end{align*} where $\mathfrak{R}: \Omega^p(M) \to \Omega^p(M)$ is a curvature endomorphism whose explicit form involves the full Riemann tensor (not just the Ricci tensor). The formula for $p = 1$ uses only $\operatorname{Ric}$ because the Riemann tensor contracts to Ricci precisely when acting on $1$-forms; for $p \geq 2$, the curvature endomorphism $\mathfrak{R}$ mixes all components of $R_{ijkl}$ and cannot be reduced to the Ricci tensor alone. This is an important limitation: many arguments that work for $1$-forms via Bochner's method do not extend cleanly to higher forms, and the vanishing theorems in those degrees require additional curvature hypotheses (such as positivity of the curvature operator, not just the Ricci curvature). The formula also requires the Levi-Civita connection: for a connection with torsion the torsion terms would contribute additional lower-order corrections to the Bochner identity, and the clean separation $\Delta = \nabla^*\nabla + \mathfrak{R}$ breaks down. By the same proof strategy: - If $\mathfrak{R} > 0$ at all points, then $\mathcal{H}^k(M) = 0$ for $k = 1, \ldots, n-1$, and hence $H^k_{\mathrm{dR}}(M) = 0$ for all middle degrees. - If $\mathfrak{R} \geq 0$ (which holds, for instance, when the metric is flat), then $b^k(M) \leq \binom{n}{k}$, and a $p$-form is harmonic if and only if it is parallel. These bounds are again sharp: on a flat torus $T^n$, harmonic $p$-forms correspond to constant-coefficient forms, giving $b^k(T^n) = \binom{n}{k}$. [example: Positive Ricci and Spheres] On the round sphere $S^n$ with $n \geq 2$, the Ricci curvature satisfies $\operatorname{Ric} = (n-1)g > 0$. The Bochner vanishing theorem then gives $H^1_{\mathrm{dR}}(S^n) = 0$, consistent with the fact that $S^n$ is simply connected. More precisely, $b^1(S^n) = 0$ for $n \geq 2$. The Ricci endomorphism on $1$-forms acts as multiplication by $n-1$, so for any harmonic $\alpha$, the integral Bochner identity gives $\|\nabla\alpha\|_2^2 + (n-1)\|\alpha\|_2^2 = 0$, forcing $\alpha = 0$. [/example] The interplay between curvature and topology demonstrated throughout this section is a hallmark of global Riemannian geometry. The Bochner–Weitzenböck technique uses the sign of curvature as an analytic input to derive topological vanishing theorems — a philosophically elegant synthesis of geometry and topology. The next chapter (Chapter 5) turns to a different but related structural question: how curvature constrains the holonomy of a Riemannian manifold. While Bochner's method exploits the Ricci curvature (introduced in Chapter 2) through its sign on form bundles, holonomy theory studies the global parallel transport group generated by the full curvature tensor $R$. The two perspectives are deeply connected: the holonomy decomposition of Chapter 5 refines the Hodge decomposition proved in this chapter, and restricted holonomy is intimately tied to $R$ via the Ambrose–Singer theorem. Harmonic forms are rigid—preserved by parallel transport along closed loops. Holonomy groups algebraically encode this parallel structure, refining our understanding of which forms can exist globally and sharpening the topological constraints that curvature imposes. # 5. Riemannian holonomy groups The holonomy group of a Riemannian manifold encodes how vectors are affected when they are parallel-transported around loops. While each previous chapter has studied curvature and topology through differential or integral analytic methods, holonomy provides a more global, algebraic lens: instead of measuring how much a vector field fails to commute with differentiation, we track how the tangent space at a point is transformed by every possible loop based there. The connection between holonomy and curvature runs deep — the infinitesimal version of holonomy is precisely the curvature tensor — and this chapter makes that connection precise. The main reward is a decomposition of de Rham cohomology that refines the classical Hodge decomposition whenever the holonomy group is smaller than $\mathrm{SO}(n)$. ## Holonomy Transformations and the Holonomy Group When a vector is parallel-transported around a closed loop, it may return transformed — rotated, reflected, or otherwise changed from its initial value. This transformation is entirely invisible to pointwise curvature but encodes genuinely global information about the metric. The question is: what algebraic structure do all such transformations, taken over every loop through a base point, form? And what do we lose if we fail to track this structure — that is, if we treat parallel transport only along paths between distinct points, never demanding that paths close? The answer is the holonomy group. Without restricting to loops based at a single point $x$, the composition law for parallel transport breaks down: $P(\gamma_2) \circ P(\gamma_1)$ only makes sense when $\gamma_1$ ends where $\gamma_2$ begins, so we cannot form a group. Restricting to loops based at $x$ gives a genuine group under composition, and the study of this group is the subject of the section. Throughout, $(M, g)$ is a connected Riemannian manifold of dimension $n$, with Levi-Civita connection $\nabla$. Fix a point $x \in M$ and a path $\gamma \in \Omega(x, y)$, meaning $\gamma: [0,1] \to M$ with $\gamma(0) = x$ and $\gamma(1) = y$. From the theory of ODEs on vector bundles, for any $X_0 \in T_x M$ there exists a unique vector field $X$ along $\gamma$ satisfying \begin{align*} \frac{\nabla X}{dt} = 0, \quad X(0) = X_0. \end{align*} This is the parallel transport equation, and $X$ is called the parallel transport of $X_0$ along $\gamma$. [definition: Holonomy Transformation] Let $\gamma \in \Omega(x, y)$. The **holonomy transformation** $P(\gamma): T_x M \to T_y M$ is the linear map sending $X_0 \in T_x M$ to $X(1) \in T_y M$, where $X$ is the unique parallel vector field along $\gamma$ with $X(0) = X_0$. [/definition] Since parallel transport is an isomorphism (it preserves the inner product and is invertible), $P(\gamma)$ is an orthogonal isomorphism from $T_x M$ to $T_y M$. When $x = y$, parallel transport around a loop sends $T_x M$ to itself by an orthogonal transformation, so $P(\gamma) \in \mathrm{O}(T_x M) \cong \mathrm{O}(n)$. [definition: Holonomy Group] The **holonomy group** of $(M, g)$ at $x \in M$ is \begin{align*} \mathrm{Hol}_x(M) = \{P(\gamma) : \gamma \in \Omega(x, x)\} \subseteq \mathrm{O}(T_x M). \end{align*} The group operation is composition of linear maps, which corresponds to concatenation of loops. [/definition] This definition raises an immediate question: the holonomy group at $x$ a priori depends on the choice of base point, but the geometric content we care about — how the metric twists tangent spaces under parallel transport — should not. We therefore need to know how the groups at different base points compare. The answer is the cleanest possible: they are all conjugate inside $\mathrm{O}(n)$, so the holonomy group is well-defined as an abstract group and as a subgroup of $\mathrm{O}(n)$ up to conjugacy. [remark: Independence of Base Point] The holonomy group is, up to conjugacy, independent of the base point. Given any $y \in M$ and a path $\beta \in \Omega(x, y)$, there is a group isomorphism \begin{align*} \mathrm{Hol}_x(M) &\longrightarrow \mathrm{Hol}_y(M) \\ P_\gamma &\longmapsto P_\beta \circ P_\gamma \circ P_{\beta^{-1}}. \end{align*} After fixing isomorphisms $\mathrm{O}(T_x M) \cong \mathrm{O}(n)$ and $\mathrm{O}(T_y M) \cong \mathrm{O}(n)$, the groups $\mathrm{Hol}_x(M)$ and $\mathrm{Hol}_y(M)$ appear as conjugate subgroups of $\mathrm{O}(n)$. We therefore write $\mathrm{Hol}(M)$ to denote this conjugacy class, or $\mathrm{Hol}(M, g)$ or $\mathrm{Hol}(g)$ when we wish to emphasise the metric. From the representation-theoretic perspective, $\mathrm{Hol}(M)$ is a representation of $\mathrm{O}(n)$ on $\mathbb{R}^n$, defined up to conjugacy. [/remark] ## The Restricted Holonomy Group It is often useful to restrict attention to loops that are contractible. This separates the topological contribution of $\pi_1(M)$ from the purely geometric contribution of curvature. [definition: Restricted Holonomy Group] The **restricted holonomy group** at $x \in M$ is \begin{align*} \mathrm{Hol}^0_x(M) = \{P(\gamma) : \gamma \in \Omega(x, x),\ \gamma \text{ is nullhomotopic}\}. \end{align*} As with the full holonomy group, this is independent of base point up to conjugacy, and we write $\mathrm{Hol}^0(M)$. [/definition] By definition, $\mathrm{Hol}^0(M) \subseteq \mathrm{Hol}(M)$, with equality when $M$ is simply connected ($\pi_1(M) = 0$). The quotient $\mathrm{Hol}(M)/\mathrm{Hol}^0(M)$ is therefore a measure of the topological contribution of $\pi_1(M)$ to holonomy, separated cleanly from the geometric contribution carried by nullhomotopic loops. The first basic structural fact about $\mathrm{Hol}^0(M)$ is that, when $M$ is simply connected, the two groups coincide and the holonomy group is well-behaved as a topological subgroup of $\mathrm{O}(n)$. [quotetheorem:2762] [citeproof:2762] Simple connectivity is essential here. Without it, non-contractible loops can contribute holonomy transformations that disconnect the holonomy group. The Klein bottle is the canonical counterexample: it carries a flat metric, and the non-contractible loop responsible for the orientation reversal produces a holonomy element of determinant $-1$, so $\mathrm{Hol}(K) \cong \mathbb{Z}_2$ has two path components and is not path connected. More generally, for a manifold with $\pi_1(M) \neq 0$, the full holonomy group $\mathrm{Hol}(M)$ can be strictly larger and disconnected relative to the restricted holonomy group $\mathrm{Hol}^0(M)$: the latter is always path connected by the same argument applied to nullhomotopic loops. [quotetheorem:2763] [citeproof:2763] The restriction to nullhomotopic loops is what makes this work. The full holonomy group $\mathrm{Hol}(M)$ need not lie in $\mathrm{SO}(n)$: on the Klein bottle, the non-contractible loop contributes $P_\gamma$ with $\det P_\gamma = -1$, which sits outside $\mathrm{SO}(2)$ but inside $\mathrm{O}(2)$. Geometrically, this reflects the fact that the Klein bottle is non-orientable — a global orientation is precisely a parallel volume form, and the fundamental principle will show this requires $\mathrm{Hol}(M) \subseteq \mathrm{SO}(n)$. The theorem tells us that at least the curvature-generated part of the holonomy (captured by $\mathrm{Hol}^0$) always preserves orientation. [remark: Structural Facts about Holonomy] Two important structural results hold, both quoted without proof: - $\mathrm{Hol}^0(M)$ is the connected component of the identity in $\mathrm{Hol}(M)$. - $\mathrm{Hol}^0(M)$ is a **Lie subgroup** of $\mathrm{SO}(n)$: it is simultaneously a subgroup and an immersed submanifold. Consequently, its Lie algebra $\mathfrak{hol}(M)$ is a Lie subalgebra of $\mathfrak{so}(n)$, the algebra of skew-symmetric $n \times n$ matrices. The second fact is a consequence of **Yamabe's theorem**: any path-connected subgroup of a Lie group is automatically a Lie subgroup. There is also a natural action of $\mathrm{Hol}_x(M)$ on the tensor algebra at $x$: it acts on $T_x^* M$, on $\bigwedge^p T_x^* M$ for all $p$, and on general tensor products $T_x M^{\otimes q} \otimes T_x^* M^{\otimes r}$. [/remark] The abstract picture above is best anchored by an explicit example in which the full and restricted holonomy groups genuinely differ. Such examples must be non-simply-connected, and they should ideally be flat — so that the only contribution to holonomy comes from the topology, not from curvature. The Klein bottle satisfies both requirements and exhibits the simplest possible behaviour: a finite full holonomy group with an identity restricted holonomy group. [example: Klein Bottle — Holonomy Can Be Non-Connected] The Klein bottle $K$ carries a flat metric. Consider the non-contractible loop $\gamma$ that runs parallel to the identification direction. Parallel transport around $\gamma$ acts on $T_x K \cong \mathbb{R}^2$ by \begin{align*} P_\gamma = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}. \end{align*} A direct computation shows that $\mathrm{Hol}(K) \cong \mathbb{Z}_2$. Since $P_\gamma$ has determinant $-1$, it does not lie in $\mathrm{SO}(2)$, which is consistent with $K$ being non-orientable. The restricted holonomy group $\mathrm{Hol}^0(K) = \{\mathrm{id}\}$, which is also consistent with the flat metric: a manifold is flat if and only if $\mathrm{Hol}^0(M, g) = \{\mathrm{id}\}$. [/example] ## The Holonomy Algebra and Its Relation to Curvature We now know that $\mathrm{Hol}^0(M)$ is a Lie subgroup of $\mathrm{SO}(n)$, so it has a Lie algebra. The question is: what generates it? A Lie algebra is, at its core, determined by the infinitesimal data of the group — and for holonomy, the infinitesimal data is exactly what you get by shrinking a loop to a point. As a loop collapses, the holonomy transformation it produces approaches the identity; the rate at which it departs from the identity is precisely the curvature of the connection. This is not merely an analogy: the curvature tensor $R(X,Y)$, which measures the failure of parallel transport to commute around an infinitesimal parallelogram in the $X$-$Y$ plane, should sit inside the holonomy algebra as a tangent vector at the identity. [definition: Holonomy Algebra] The **holonomy algebra** $\mathfrak{hol}(M)$ is the Lie algebra of $\mathrm{Hol}^0(M)$, viewed as a Lie subalgebra of $\mathfrak{so}(n)$ (up to conjugation). [/definition] The key insight is that the curvature tensor generates the holonomy algebra. To make this precise, let $(U, \varphi)$ be a coordinate chart with coordinates $(x_1, \ldots, x_n)$, and write $\partial_i = \partial/\partial x_i$ and $\nabla_i = \nabla_{\partial_i}$ as before. In coordinates, the curvature satisfies \begin{align*} R(\partial_k, \partial_\ell) = -[\nabla_k, \nabla_\ell]. \end{align*} Consider the family of square loops $\gamma_t$ in the $(x_k, x_\ell)$-plane of side length $\sqrt{t}$. A direct computation using the parallel transport ODE shows that \begin{align*} P(\gamma_t) = I + \lambda t\, R(\partial_k, \partial_\ell) + o(t), \end{align*} where the coefficient of $t$ is exactly $+1$ under the chapter's sign convention $R(\partial_k, \partial_\ell) = -[\nabla_k, \nabla_\ell]$. To see why: traversing the square counterclockwise, the four legs of $\gamma_t$ contribute, to first order in $t$, the commutator $[\nabla_k, \nabla_\ell]$ with a sign that matches the sign of $R$. Since $R(\partial_k,\partial_\ell) = \nabla_k\nabla_\ell - \nabla_\ell\nabla_k = -[\nabla_k, \nabla_\ell]$, the expansion becomes $P(\gamma_t) = I + t\,R(\partial_k, \partial_\ell) + o(t)$. Differentiating with respect to $t$ at $t = 0$ shows that $R(\partial_k, \partial_\ell)|_p$ lies in $\mathfrak{hol}_p(M) \subseteq \mathrm{End}(T_p M)$ for every $p \in U$.  More precisely, at any point $p \in U$, the curvature tensor satisfies \begin{align*} R_p = (R^i_{j,k\ell})_p \in \bigwedge^2 T_p^* M \otimes \mathfrak{hol}_p(M), \end{align*} where we view $\mathfrak{hol}_p(M) \subseteq \mathrm{End}(T_p M)$. Using the metric to identify $T_p M \cong T_p^* M$, the symmetry properties of the curvature tensor (in particular $R_{ij,k\ell} = R_{k\ell,ij}$) give \begin{align*} (R_{ij,k\ell})_p \in S^2 \mathfrak{hol}_p(M) \subseteq \bigwedge^2 T_p^* M \otimes \bigwedge^2 T_p^* M. \end{align*} This is the beginning of the deep relationship between curvature and holonomy: the curvature tensor at every point is an element of the symmetric square of the holonomy algebra. ## The Fundamental Principle of Riemannian Holonomy The most important theorem of the chapter establishes a perfect correspondence between parallel tensor fields on $M$ and holonomy-invariant tensors at a single point. This principle is the bridge between local curvature data and global geometric structures. [quotetheorem:2764] [citeproof:2764] The connectedness hypothesis is essential. The theorem uses $\mathrm{Hol}_x(M)$, the full holonomy group at $x$, not merely $\mathrm{Hol}^0_x(M)$. If one were to use only the restricted holonomy group, the invariance condition would be weaker — it would not detect elements of $\mathrm{Hol}(M)$ coming from non-contractible loops, and the extension of $\alpha_0$ to a global tensor field could fail to be well-defined. The connectedness of $M$ is needed so that any two points can be joined by a path, ensuring the extension via parallel transport produces a single well-defined global section; if $M$ were disconnected, one could prescribe different $\mathrm{Hol}$-invariant tensors on different components with no contradiction. The power of this result lies in what it tells us about special geometries: any reduction of the holonomy group from $\mathrm{O}(n)$ to a proper subgroup $H$ corresponds exactly to the existence of a parallel tensor field whose stabilizer in $\mathrm{O}(n)$ is $H$. [example: Orientation and SO(n)] Suppose $M$ is oriented. Then the volume form $\omega_g \in \Omega^n(M)$ satisfies $\nabla \omega_g = 0$ (compatibility of the Levi-Civita connection with the metric). By the fundamental principle, $(\omega_g)_x$ is invariant under $\mathrm{Hol}_x(M)$. The stabilizer of a nonzero $n$-form in $\mathrm{O}(n)$ is exactly $\mathrm{SO}(n)$, so we deduce $\mathrm{Hol}(M) \subseteq \mathrm{SO}(n)$. Conversely, if $\mathrm{Hol}(M) \subseteq \mathrm{SO}(n)$, we can take any nonzero $n$-form $\alpha_0$ at a point $x$ and use parallel transport to extend it globally; the invariance condition guarantees this extension is well-defined and nonzero everywhere, hence an orientation. So $M$ is orientable if and only if $\mathrm{Hol}(M) \subseteq \mathrm{SO}(n)$. [/example] The orientation example illustrates the simplest case of the fundamental principle: a single parallel form, with stabilizer the maximal connected subgroup of $\mathrm{O}(n)$. Smaller stabilizers correspond to richer parallel structures and stronger geometric constraints. The next natural reduction is from $\mathrm{SO}(2n)$ to $\mathrm{U}(n)$, where the parallel object is no longer a top-degree form but an endomorphism of the tangent bundle squaring to $-\mathrm{id}$. This reduction characterises Kähler geometry. [example: Almost Complex Structures and U(n)] Let $\dim M = 2n$ and suppose \begin{align*} \mathrm{Hol}_x(M) \subseteq \mathrm{U}(n) = \{g \in \mathrm{SO}(2n) : g J_0 g^{-1} = J_0\}, \end{align*} where $J_0 = \begin{pmatrix} 0 & I \\ -I & 0 \end{pmatrix}$ is the standard complex structure on $\mathbb{R}^{2n}$. Taking $\alpha_0 = J_0 \in \mathrm{End}(T_x M)$, the fundamental principle produces a global endomorphism $J \in \Gamma(\mathrm{End}(TM))$ with $\nabla J = 0$ and $J^2 = -\mathrm{id}$. Such a $J$ is an **almost complex structure** on $M$. The condition $\nabla J = 0$ (integrability in the strongest sense) makes $M$ a Kähler manifold, and this is one of the entry points into complex Riemannian geometry. [/example] The two preceding examples have shown that a proper subgroup of $\mathrm{O}(n)$ as the holonomy group corresponds to extra parallel structure. The opposite extreme — the smallest possible holonomy group — should correspond to the strongest possible structural constraint, namely a full parallel framing of the tangent bundle. This is the case of flat metrics, where one must be careful to use the restricted holonomy group: as the Klein bottle showed, the full holonomy can detect topology that the curvature cannot. [example: Flat Manifolds and Identity Restricted Holonomy] A Riemannian manifold $(M, g)$ is flat — meaning $R(g) \equiv 0$ — if and only if around every point there exists a basis of parallel vector fields. By the fundamental principle, this is equivalent to $\mathrm{Hol}^0(M, g) = \{\mathrm{id}\}$. It is essential to use the restricted holonomy group here, not the full $\mathrm{Hol}(M, g)$. The Klein bottle with its flat metric has $\mathrm{Hol}(K) = \mathbb{Z}_2 \neq \{\mathrm{id}\}$, yet the metric is flat because $\mathrm{Hol}^0(K) = \{\mathrm{id}\}$. The non-identity element of $\mathrm{Hol}(K)$ comes from a non-contractible loop, not from curvature. [/example] Taken together, these examples show how reductions of the holonomy group encode geometric structure: $\mathrm{Hol}(M) \subseteq \mathrm{SO}(n)$ encodes orientability, $\mathrm{Hol}(M) \subseteq \mathrm{U}(n)$ encodes a Kähler structure, and $\mathrm{Hol}^0(M) = \{\mathrm{id}\}$ encodes flatness. Each reduction also imposes constraints on what tensors can exist parallel on $M$, and — as the next section will show — on the algebraic structure of de Rham cohomology itself. The bridge between these two consequences is the way the holonomy representation decomposes the bundle of differential forms. ## Holonomy and the Decomposition of de Rham Cohomology The classical Hodge decomposition established in Chapter 4 — $H^k_{\mathrm{dR}}(M) \cong \ker(\Delta|_{\Omega^k})$ for compact oriented Riemannian manifolds — says nothing about the algebraic structure of the holonomy representation. Every harmonic $k$-form is treated equally, regardless of which $\mathrm{Hol}(M)$-representation it belongs to. The question is whether cohomology itself naturally splits when the holonomy group is smaller than $\mathrm{SO}(n)$. If $\bigwedge^k(\mathbb{R}^{n*})$ decomposes into several irreducible $\mathrm{Hol}(M)$-representations, do the corresponding sub-bundles of $k$-forms carry their own separate cohomology, or does the Laplacian mix them? The answer is that they do not mix. Whenever $\mathrm{Hol}(M) \leq G \leq \mathrm{O}(n)$ for some subgroup $G$, the bundle of differential forms decomposes into parallel sub-bundles, and this decomposition is respected by the Hodge Laplacian. The result is a refinement of de Rham cohomology into **refined Betti numbers**. To set this up, consider an arbitrary Lie subgroup $G \subseteq \mathrm{GL}_n(\mathbb{R})$. The standard representation of $\mathrm{GL}_n(\mathbb{R})$ on $\mathbb{R}^n$ restricts to a representation $(\rho, \mathbb{R}^n)$ of $G$, which induces a representation $(\rho^k, \bigwedge^k(\mathbb{R}^{n*}))$ of $G$. This induced representation is generally reducible. Decompose it into irreducible components: \begin{align*} \bigwedge^k (\mathbb{R}^{n*}) = \bigoplus_i W_i^k. \end{align*} Now suppose $M$ carries a $G$-structure — an atlas where all transition maps satisfy $\left(\partial x_\alpha / \partial x_\beta'\right)_p \in G$ at every point. This allows the decomposition of $\bigwedge^k(\mathbb{R}^{n*})$ into irreducibles to be promoted to a direct sum decomposition of the bundle of $k$-forms into well-defined vector sub-bundles: \begin{align*} \bigwedge^k T^* M = \bigoplus_i \Lambda_i^k, \end{align*} where the typical fiber of $\Lambda_i^k$ is $W_i^k$. Moreover, every $G$-equivariant linear map $\varphi: W_i^k \to W_j^\ell$ induces a morphism of vector bundles $\phi: \Lambda_i^k \to \Lambda_j^\ell$. The crucial geometric fact is that when $\mathrm{Hol}(M) \leq G$, parallel transport preserves this decomposition. Therefore $\nabla$ restricts to a well-defined connection on each sub-bundle $\Lambda_i^k$, and if $\xi \in \Gamma(\Lambda_i^k)$ then $\nabla \xi \in \Gamma(T^*M \otimes \Lambda_i^k)$. This is what one needs to show that the Hodge Laplacian $\Delta$ preserves each sub-bundle. For $k = 1$ and $\xi \in \Omega^1(M)$, the Bochner–Weitzenböck formula gives \begin{align*} \Delta \xi = \nabla^* \nabla \xi + \operatorname{Ric}(\xi). \end{align*} Both $\nabla^* \nabla$ and $\operatorname{Ric}$ (which is a self-adjoint endomorphism of $T^*M$, introduced in Chapter 2 and used in the Bochner vanishing theorem of Chapter 4) preserve each $\Lambda_j^1$, so $\Delta: \Gamma(\Lambda_j^1) \to \Gamma(\Lambda_j^1)$ is well-defined. The same argument applies for all $k$ using the analogous Weitzenböck formula for $k$-forms. [quotetheorem:2765] [citeproof:2765] Compactness is essential for two reasons. First, the Hodge theorem — which identifies harmonic forms with cohomology classes — requires compactness (and completeness); on a non-compact manifold, a harmonic $k$-form need not represent a cohomology class, and the identification breaks down. Second, without compactness, the space of harmonic forms can be infinite-dimensional, making the "decomposition" of cohomology ill-defined. Orientability is needed to make sense of the Hodge star operator $*: \Omega^k \to \Omega^{n-k}$, which enters through the formal adjoint $d^* = (-1)^{nk+n+1} * d *$; without a consistent orientation, the sign of $*$ is not globally well-defined and $d^*$ cannot be constructed. On a non-compact manifold, even if harmonic forms exist, they may not lie in $L^2$ and so fall outside the domain where the Weitzenböck identity applies. [explanation: Why the Refined Decomposition Matters] When $\mathrm{Hol}(M) = \mathrm{SO}(n)$ (generic holonomy), the representation $\bigwedge^k(\mathbb{R}^{n*})$ is irreducible, so the decomposition has only one summand and the refined Betti numbers recover the ordinary Betti numbers. The interesting cases arise when holonomy is smaller. For a Kähler manifold, $\mathrm{Hol}(M) \leq \mathrm{U}(n)$, and the decomposition $\bigwedge^k(\mathbb{R}^{2n*}) = \bigoplus_{p+q=k} \bigwedge^{p,q}$ into $(p,q)$-forms is a decomposition into $\mathrm{U}(n)$-irreducibles. The resulting refined Betti numbers are the **Hodge numbers** $h^{p,q}$, and the cohomology decomposition $H_{\mathrm{dR}}^k(M; \mathbb{C}) = \bigoplus_{p+q=k} H^{p,q}(M)$ is the Hodge decomposition of complex geometry. The holonomy framework thus recovers the Lefschetz decomposition as a special case. This illustrates the general philosophy: special holonomy implies the existence of parallel structures (by the fundamental principle), which impose strong constraints on topology (via the cohomology decomposition). The classification of possible holonomy groups — due to Berger — organises the landscape of Riemannian geometry into distinct types, each with its own characteristic parallel structures and topological consequences. [/explanation] Holonomy classifies the parallel structures a manifold admits; Ricci curvature governs their global behavior. The Cheeger–Gromoll splitting theorem shows how positive Ricci curvature forces rigidity, while zero Ricci curvature allows complete manifolds to decompose into product structures. # 6. The Cheeger–Gromoll splitting theorem The preceding chapters developed the three main curvature theories of Riemannian geometry — sectional, Ricci, and scalar — and used them to prove diameter and topology theorems under positive curvature hypotheses. This final chapter turns to the opposite regime: what can be said when Ricci curvature is merely nonnegative? The answer is a remarkable structural theorem due to Cheeger and Gromoll (1971): any complete connected Riemannian manifold with $\operatorname{Ric}(g) \geq 0$ that contains a globally minimizing geodesic line must split isometrically as a Riemannian product with $\mathbb{R}$. The theorem links a global geometric condition (the existence of a line) to a strong topological conclusion (a product structure), and its iterative application yields a near-complete classification of the universal covers of compact manifolds with nonnegative Ricci curvature. ## Rays, Lines, and Ends of a Manifold The splitting theorem to come pivots on a single global geometric datum: the existence of a *line*, a unit-speed geodesic that minimizes distance between every pair of its points across all of $\mathbb{R}$. Why insist on a two-sided minimizing geodesic rather than a one-sided ray? Rays are cheap — every complete unbounded manifold issues a ray from every point, by a Hopf–Rinow compactness argument — but they are too cheap to force a splitting. The Busemann function $b^+$ associated to a forward ray is superharmonic under $\operatorname{Ric} \geq 0$, but a single superharmonic function carries only one-sided information: it pins down a half-space at best. The maximum-principle argument that drives the proof of the splitting theorem requires a *second* superharmonic function whose sum with $b^+$ is forced to be identically zero on $M$, and that second function is the Busemann function $b^-$ of the *backward* ray. Without a backward ray — without a line — there is no $b^-$, and no rigidity. We therefore need both the one-sided notion (rays, the right object for many compactness arguments) and the two-sided notion (lines, the right object for splitting), and we need a topological criterion that detects when lines exist. [definition: Ray] Let $(M, g)$ be a Riemannian manifold. A **ray** is a map $r : [0, \infty) \to M$ that is a unit-speed geodesic minimizing the distance between any two of its points: for all $0 \leq s < t < \infty$, \begin{align*} d(r(s), r(t)) = t - s. \end{align*} [/definition] A ray is a half-infinite geodesic that never stops being a shortest path. If $M$ is complete and unbounded, then from every point of $M$ there issues at least one ray — this follows from the Hopf–Rinow theorem: pick a sequence of points $q_k$ with $d(p, q_k) \to \infty$, take unit-speed minimizing geodesics $\gamma_k$ from $p$ to $q_k$, and extract a limit using compactness of the unit sphere in $T_p M$. [definition: Line] A **line** in $(M, g)$ is a map $\ell : \mathbb{R} \to M$ that is a unit-speed geodesic globally minimizing distance between any two of its points: for all $s, t \in \mathbb{R}$, \begin{align*} d(\ell(s), \ell(t)) = |t - s|. \end{align*} [/definition] A line is simultaneously a ray in the forward direction and a ray in the backward direction. Its existence requires the manifold to be, in some sense, "large in two opposite directions at once." The precise topological condition that forces this is the following. [definition: Connected at Infinity] A complete manifold $M$ is **connected at infinity** if for every compact set $K \subseteq M$ there exists a compact set $C \supseteq K$ such that for every two points $p, q \in M \setminus C$ there is a path $\gamma$ from $p$ to $q$ staying entirely in $M \setminus K$. If $M$ fails this condition, it is **disconnected at infinity**. [/definition] Informally, $M$ is connected at infinity if "going far enough out, the complement of any compact set is connected." The simplest example of a manifold disconnected at infinity is $S^{n-1} \times \mathbb{R}$: removing a slab $S^{n-1} \times [-R, R]$ for large $R$ leaves two disconnected ends. [quotetheorem:2766] [citeproof:2766] The hypothesis of disconnection at infinity, rather than mere noncompactness, is what makes the argument work: noncompactness alone gives only that *some* sequence of points escapes every compact set, but a single ray is enough to witness this. Disconnection forces the escaping sequences to live on opposite sides of a fixed compact obstruction, and any minimizing geodesic between the two sides must thread through that compact set — this is exactly what produces a midpoint in a fixed compact region around which to extract a two-sided limit. This lemma will power the compact-manifold splitting argument later in the chapter: the universal cover of a compact manifold with infinite fundamental group, while not visibly disconnected at infinity, admits a related compactness argument (using a fundamental domain in place of the obstructing compact set) that produces a line by the same limiting procedure. The converse is false — a manifold may contain a line yet be connected at infinity (think of $\mathbb{R}^n$ for $n \geq 2$) — so disconnection at infinity is sufficient but not necessary for the existence of a line. [remark: Lines Require Noncompactness] A compact Riemannian manifold cannot contain a line: any geodesic must eventually self-intersect or fail to minimize, since the manifold has finite diameter. Thus the existence of a line is purely a noncompact phenomenon. [/remark] ### Examples [example: Paraboloid Contains No Line] Consider the elliptic paraboloid $P = \{(x, y, z) \in \mathbb{R}^3 : z = x^2 + y^2\}$ with the induced metric from $\mathbb{R}^3$. The paraboloid is connected at infinity: going far from the origin, the complement of any compact region is connected, because the surface flares out in all directions. Therefore the theorem above does not force the existence of a line, and indeed $P$ contains no line. Any geodesic that is not a meridian (a curve of the form $\{(\cos\theta \cdot r, \sin\theta \cdot r, r^2) : r \geq 0\}$ for fixed $\theta$) must at some point curve back and intersect itself, preventing global minimization. [/example] [example: Cylinder Always Contains a Line] For any complete Riemannian metric $g$ on $S^{n-1} \times \mathbb{R}$, the manifold is disconnected at infinity — removing a sufficiently wide cylinder separates the two ends. The theorem therefore guarantees that $S^{n-1} \times \mathbb{R}$ contains a line for every complete metric, regardless of curvature. [/example] ## The Cheeger–Gromoll Splitting Theorem When can global geometry force a metric splitting? Local curvature conditions — even strong ones — typically say nothing about product structure: a manifold can be locally curved everywhere and yet globally rigid. The question is whether the existence of a globally minimizing line, a single geodesic that minimizes distance between every pair of its points, can force the entire manifold to split. The obstacle is that a line is a global datum, and local conditions like bounded curvature cannot detect it. The precise question: if $(M, g)$ has $\operatorname{Ric}(g) \geq 0$ everywhere and contains a line, must $M$ split isometrically as a product with $\mathbb{R}$? [quotetheorem:2767] The proof uses the theory of Busemann functions, which measure "distance to infinity along a ray," and is a deep application of the Bochner technique and the maximum principle for subharmonic functions. We state the theorem without proof. Several features of the conclusion deserve emphasis. First, completeness is genuinely necessary: on an incomplete manifold with $\operatorname{Ric} \geq 0$ containing a line, splitting can fail — one can remove a point from $\mathbb{R}^2$ to produce an incomplete flat manifold that contains geodesics extending to infinity in both directions but does not globally split. Second, a ray alone is not sufficient: the Busemann function for a single ray $r : [0, \infty) \to M$ is superharmonic under $\operatorname{Ric} \geq 0$, but without a backward ray one cannot produce a second superharmonic function whose sum is forced to be identically zero by the maximum principle; the argument breaks down and splitting need not occur. Third, the theorem asserts an isometric splitting $M \cong N \times \mathbb{R}$, not merely a topological product: the metric must decompose as a product metric, and the factor $N$ is the common level set of the Busemann function $b^+$. What the theorem does not assert is that $N$ is topologically $M / \mathbb{R}$ in any naive sense; the splitting is a geometric isometry, and the direction $\mathbb{R}$ is precisely the gradient flow of $b^+$. As a forward consequence, since $N$ is again complete with $\operatorname{Ric}(g_N) \geq 0$, the splitting theorem can be applied iteratively to $N$. [explanation: Strategy of the Proof] The key analytic tool is the Busemann function associated to the line $\ell$. For the forward ray $\ell|_{[0,\infty)}$, define \begin{align*} b^+(x) = \lim_{t \to \infty} \bigl(t - d(x, \ell(t))\bigr). \end{align*} This limit exists for every $x \in M$ because $t \mapsto t - d(x, \ell(t))$ is nondecreasing and bounded above. For the upper bound: by the triangle inequality, $d(x, \ell(t)) \geq t - d(x, \ell(0))$, so $b^+_t(x) := t - d(x, \ell(t)) \leq d(x, \ell(0))$. The function $t \mapsto b^+_t(x)$ is therefore bounded above by $d(x, \ell(0))$. Monotonicity (which follows from the fact that $d(x, \ell(t))$ grows at most as fast as $t$) together with this upper bound gives existence of the limit. Similarly, from the backward ray, one defines $b^-(x) = \lim_{t \to \infty}(t - d(x, \ell(-t)))$. By the triangle inequality, $b^+ + b^- \leq 0$ everywhere. Since $\ell$ is a line, one also shows $b^+(x) + b^-(x) \geq 0$ along $\ell$ itself (equality holds). The function $b^+ + b^-$ is thus a nonpositive function that equals zero on $\ell$. Under the assumption $\operatorname{Ric}(g) \geq 0$, a Bochner-type estimate shows that $b^+$ is superharmonic (i.e., $\Delta b^+ \leq 0$) and $b^-$ is superharmonic. Their sum $b^+ + b^- \leq 0$ is therefore superharmonic. Since it achieves its maximum value $0$ on the line, the strong maximum principle forces $b^+ + b^- \equiv 0$ on all of $M$, and in fact $b^+$ is harmonic and $|\nabla b^+| \equiv 1$. One then uses the condition $|\nabla b^+| \equiv 1$ and harmonicity to conclude that $b^+$ is a smooth function on $M$ with unit-norm gradient everywhere. The gradient vector field $\nabla b^+$ is a complete unit vector field, and its flow defines a one-parameter group of isometries (this uses harmonicity together with the metric-compatibility of the Levi-Civita connection). Setting $N = (b^+)^{-1}(0)$, the gradient flow provides a global diffeomorphism $N \times \mathbb{R} \to M$, $(x, t) \mapsto \Phi_t(x)$, under which the metric pulls back to $g_N + dt^2$. This gives the isometric splitting $M \cong N \times \mathbb{R}$. [/explanation] [remark: Sharpness] The curvature hypothesis $\operatorname{Ric}(g) \geq 0$ cannot be weakened to scalar curvature $S \geq 0$: there exist complete manifolds with positive scalar curvature that contain lines and do not split as products. The Ricci condition is essential for the Bochner estimate that forces $b^+$ to be harmonic. [/remark] ## Iterated Splitting and the de Rham Decomposition What happens when $M$ contains not just one line, but multiple lines in independent directions? And what can be said when the transverse factor contains no line at all — is that the end of the splitting, or is there further structure? The question is how far the splitting theorem can be iterated, what the irreducible piece looks like once no further lines exist, and whether the decomposition is unique. [quotetheorem:2768] [citeproof:2768] The integer $q$ is the **rank** of the Euclidean factor, measuring how many linearly independent "directions to infinity" the manifold possesses. When $q = n = \dim M$, the manifold is flat. When $q = 0$, no splitting is possible and $M$ has a genuinely curved transverse geometry. Without the hypothesis $\operatorname{Ric}(g) \geq 0$, the decomposition fails: $\mathbb{H}^2 \times \mathbb{R}$ contains lines (the $\mathbb{R}$ factor), but one could deform the metric on $\mathbb{H}^2$ to destroy the product structure while keeping a geodesic line; conversely, the hyperbolic plane $\mathbb{H}^2$ has no lines yet has uncountably many geodesics extending to infinity in both directions that fail to globally minimize. The uniqueness claim — that $X$ and $q$ are determined up to isometry — requires care: in the stated generality, one passes to the universal cover $\tilde{M}$, which is simply connected by definition, applies the de Rham decomposition theorem (which requires simple connectivity), and then descends the splitting back to $M$ by observing that the deck group of $\tilde{M} \to M$ must preserve the de Rham factors. The claim as stated therefore holds, but the uniqueness argument is not immediate from de Rham alone without this descent step. [explanation: Low-Dimensional Consequences] The splitting decomposition $M \cong X \times \mathbb{R}^q$ has particularly clean consequences in low dimensions. If $\dim X = 0$, then $M \cong \mathbb{R}^q$ and $M$ is flat. If $\dim X = 1$, then $X$ is a complete one-dimensional manifold with $\operatorname{Ric} \geq 0$ and no line. The only such manifold is $S^1$ (the circle), since $\mathbb{R}$ contains its own parametrization as a line. So $M \cong S^1 \times \mathbb{R}^q$, which is again flat. Now suppose $\operatorname{Ric}(g) = 0$ exactly (Ricci-flat). One checks from the definition of the Ricci tensor that the product component $X$ satisfies $\operatorname{Ric}(g_X) = 0$ as well. In dimensions $\leq 3$, the Ricci tensor determines the full Riemann curvature tensor — as proved in Chapter 2, where the explicit reconstruction formula $R_{ijkl} = g_{ik}\operatorname{Ric}_{jl} + g_{jl}\operatorname{Ric}_{ik} - g_{il}\operatorname{Ric}_{jk} - g_{jk}\operatorname{Ric}_{il} - \tfrac{s}{2}(g_{ik}g_{jl} - g_{il}g_{jk})$ was established. So if $\dim X \leq 3$, then $\operatorname{Ric}(g_X) = 0$ implies the full Riemann tensor of $X$ vanishes, meaning $X$ is flat. Hence $M$ itself is flat. [/explanation] ## Applications to Compact Manifolds The splitting theorem requires a line, and compact Riemannian manifolds never contain lines — any geodesic in a compact manifold must eventually stop being globally minimizing, since the manifold has finite diameter. So how can the splitting theorem bear on compact manifolds at all? The bridge is the universal cover: a compact manifold $M$ with infinite fundamental group has a noncompact universal cover $\tilde{M}$, and noncompact complete manifolds can and do contain lines. The question then becomes what structural conclusions about $\tilde{M}$ descend to $M$. The key technical lemma underlying everything is the existence of a line in the universal cover of a compact manifold (when the cover is non-compact). It applies to any compact Riemannian manifold, not just those with curvature hypotheses, and serves as the geometric engine for the Ricci-nonnegative result that follows. [quotetheorem:2769] [citeproof:2769] Compactness of $M$ is genuinely essential here: it is what guarantees a *compact* fundamental domain $K$ in which to corral the midpoints of the diverging minimizing geodesics, and hence it is what lets Arzelà–Ascoli extract a limiting line. If $M$ is non-compact, then no compact fundamental domain exists, the midpoints of long minimizing geodesics in $\tilde{M}$ can escape every compact set, and the limiting construction collapses. The geometric content is that an infinite deck group acting cocompactly on a non-compact $\tilde{M}$ produces translates that pile up at infinity; remove cocompactness, and that piling-up no longer happens. Building on this lemma, we can now state the existence-of-line result for the universal cover in the Ricci-nonnegative setting. [quotetheorem:2770] [citeproof:2770] This result is the gateway to applying Cheeger–Gromoll: it translates the topological hypothesis "$\pi_1(M)$ infinite" into the geometric input the splitting theorem needs. Cheeger–Colding extensions weaken $\operatorname{Ric} \geq 0$ to almost-nonnegative Ricci ($\operatorname{Ric} \geq -\varepsilon$) and recover an approximate splitting on the universal cover, with quantitative control of how close the geometry is to a true product. Without compactness of $M$, the entire bridge collapses: the deck group need not act cocompactly, no compact fundamental domain exists in $\tilde{M}$, and the line-extraction argument fails — there exist non-compact manifolds with $\operatorname{Ric} \geq 0$ and infinite $\pi_1$ whose universal covers contain no line at all. [quotetheorem:2771] [citeproof:2771] [remark: Compactness of X] The compactness of $X$ in part (1) can also be seen directly: if $X$ were non-compact, the deck group action on $X \times \mathbb{R}^q$ could not produce a compact quotient, since $X$ itself cannot be quotiented to a compact space by a discrete group of isometries unless $X$ is already compact (by the definition of a covering space). [/remark] ## Obstructions to Ricci-Flat Metrics Ricci-flat metrics are rare and difficult to construct: on most manifolds they simply do not exist, but proving nonexistence requires an argument, not just an absence of examples. Why should a topologically simple manifold like $S^{n-1} \times \mathbb{R}$ fail to carry a complete Ricci-flat metric? The product metric is certainly not Ricci-flat (the $S^{n-1}$ factor has positive Ricci curvature), but that says nothing about other metrics. The question is whether any complete Ricci-flat metric on $S^{n-1} \times \mathbb{R}$ can exist, and for small $n$ the splitting theorem gives a decisive obstruction. [quotetheorem:2772] [citeproof:2772] [example: Lie Groups with Finite Center and Positive Ricci] Let $G$ be a Lie group with a bi-invariant metric $g$. Suppose the center $Z(G)$ is finite. Then the center of the Lie algebra $\mathfrak{g}$ is trivial (it is the Lie algebra of $G / Z(G)$, which has trivial center since $Z(G)$ is finite). An earlier Bochner-type computation shows that for bi-invariant metrics, the Ricci tensor satisfies \begin{align*} \operatorname{Ric}(X, X) = \frac{1}{4} \sum_i |[e_i, X]|^2 \end{align*} for an orthonormal basis $\{e_i\}$. When the center of $\mathfrak{g}$ is trivial, for every nonzero $X$ there exists some $e_i$ with $[e_i, X] \neq 0$, so $\operatorname{Ric}(X, X) > 0$. By the structural theorem for compact manifolds with nonnegative Ricci curvature (part (2)), $\pi_1(G)$ is finite. The converse — if $\pi_1(G)$ is finite then $Z(G)$ is finite — also holds, though its proof is more involved. [/example] [explanation: Homogeneous Manifolds with Nonneg Ricci] A further consequence concerns homogeneous manifolds. Recall that $(M, g)$ is homogeneous if its isometry group $I(M)$ acts transitively. For a complete homogeneous manifold with $\operatorname{Ric}(g) \geq 0$, the transverse factor $X$ from the splitting theorem is also homogeneous (since isometries of $M = X \times \mathbb{R}^q$ decompose as products). Moreover, in this case $X$ must be compact. To see this, suppose $X$ is non-compact. Since $X$ is homogeneous, its isometry group acts transitively, so for any two points there is an isometry taking one to the other. Given any $n \in \mathbb{N}$, pick points $p_n, q_n \in X$ with $d(p_n, q_n) \geq 2n$, and take a minimizing geodesic $\gamma_n$ from $p_n$ to $q_n$. By homogeneity, apply an isometry to place the midpoint of $\gamma_n$ at a fixed base point $x_0 \in X$. The directions $\dot\gamma_n(0)$ lie in the unit sphere of $T_{x_0}X$, which is compact. Passing to a subsequence, $\dot\gamma_n(0) \to v \in T_{x_0}X$, and the geodesic $\ell$ with initial data $(x_0, v)$ is a line in $X$. But $X$ was chosen to contain no lines — contradiction. So $X$ is compact. [/explanation] ## Closing Remarks on the Course This course has traced a path through the central ideas of Riemannian geometry, beginning with the basic definitions of metrics, connections, and curvature, and developing through the major theorems that relate curvature to global topology. We started with geodesics and the Hopf–Rinow completeness theorem, which gives the foundational bridge between the local differential structure and global distance geometry. The Jacobi field theory then quantified how nearby geodesics spread or converge, controlled directly by sectional curvature. Positive curvature bounds led to the Bonnet–Myers diameter theorem and Synge's theorem, demonstrating how curvature constrains the fundamental group. In the opposite direction, Hodge theory and the Bochner–Weitzenböck formula showed that nonnegative curvature forces vanishing of cohomology, connecting analysis and topology through curvature in a different and equally powerful way. The present chapter brought these threads together: the Cheeger–Gromoll theorem shows that nonnegative Ricci curvature, combined with a single global geometric datum (the existence of a line), pins down the entire product structure of the universal cover. Throughout, the recurring theme has been that curvature — a local second-order invariant — exerts profound and sometimes rigid control over the global shape of the manifold. ## References Kovalev, A. G. (2017). *Riemannian Geometry* — lecture notes for Cambridge Part III, Lent term 2017.

Created by admin on 4/29/2026 | Last updated on 4/29/2026

What brings you to Androma?

Start with a route through the knowledge graph.

Cambridge III Riemannian Geometry

Sign in to Androma

Check your inbox

One last step

Cambridge III Riemannian Geometry

Prerequisites (0/74 completed)

Prerequisites Graph

Rate this page