Why a second derivative: from minimizing to the second variation
Guide 1 left you with a sharp picture: a minimizing geodesic is a critical point of the energy functional E(c) = (1/2) integral of g(c', c') dt, and the first variation formula says the energy stops changing to first order exactly when the acceleration nabla_{c'} c' vanishes — that is the geodesic equation. But a critical point of a function need not be a minimum: it could be a saddle or a max. To tell which, you do in infinite dimensions exactly what single-variable calculus does — you look at the second derivative. The second variation of energy is the curved-space analogue of f''(x_0), and its sign decides whether a geodesic genuinely minimizes length nearby or has quietly stopped doing so.
Set up the variation honestly. Fix a geodesic gamma: [0, L] -> M from p to q, and wiggle it through a family of curves gamma_s sharing the same endpoints. The infinitesimal wiggle is a variation field V(t) = d/ds gamma_s(t) at s = 0 — a vector field along gamma that vanishes at both ends, V(0) = V(L) = 0, because the endpoints are pinned. Differentiate the energy twice in s and evaluate at the geodesic. After an integration by parts (using metric compatibility from the previous rung) the messy formula collapses to a clean, symmetric expression in V — and crucially, the curvature tensor appears, uninvited but unavoidable.
The index form: the second variation written down
Here is the object the whole guide turns on. The second variation of energy at a geodesic gamma, for a variation field V vanishing at the endpoints, is the index form I(V, V) = integral from 0 to L of [ g(V', V') - g(R(V, gamma') gamma', V) ] dt, where V' = nabla_{gamma'} V is the covariant derivative of V along gamma. Read the two terms physically. The first, g(V', V'), is a kinetic 'cost of bending' — pushing the curve costs energy, like stretching a spring, and it is always nonnegative. The second is the curvature term, and it carries the Riemann curvature tensor R; when sectional curvature is positive it makes the integral SMALLER, working to destabilize the geodesic, while negative curvature makes it larger and reinforces minimization.
Jacobi fields: the shadow of a nearby geodesic
Now ask the natural minimization question: among all variation fields V, which one makes the index form stationary? Setting the first variation of I to zero (another integration by parts) produces the Euler-Lagrange equation of the index form, the Jacobi equation: V'' + R(V, gamma') gamma' = 0, where V'' = nabla_{gamma'} nabla_{gamma'} V. A solution is a Jacobi field. Conceptually it is not abstract at all: a Jacobi field is exactly the variation field of a family of geodesics — the infinitesimal displacement to a neighboring geodesic that also starts at p. It is the shadow that one geodesic casts on its neighbors, and the Jacobi equation says how curvature steers that shadow.
The constant-curvature cases make this vivid and you should hold them as touchstones. Take a Jacobi field V with V(0) = 0 and write J(t) = |V(t)| for its growing length. On flat R^n (K = 0) the Jacobi equation is V'' = 0, so J(t) = t grows linearly — neighboring straight lines through a point separate at constant rate. On the sphere of curvature K = 1 it becomes J'' + J = 0, so J(t) = sin(t): the shadow grows, peaks, and returns to zero at t = pi, because great circles through the north pole all reconverge at the south pole. On hyperbolic space of curvature K = -1 it is J'' - J = 0, so J(t) = sinh(t) blows up exponentially — geodesics in negative curvature flee each other. Positive curvature focuses, negative curvature defocuses; that single sentence is most of comparison geometry.
Jacobi equation: V'' + R(V, gamma') gamma' = 0 ( V'' = nabla_{gamma'} nabla_{gamma'} V )
constant curvature K, with J(t) = |V(t)|, V(0)=0:
K = 0 (flat R^n): J'' = 0 -> J(t) = t (linear, never returns)
K = 1 (sphere S^n): J'' + J = 0 -> J(t) = sin(t) (zero again at t = pi)
K = -1 (hyperbolic): J'' - J = 0 -> J(t) = sinh(t) (exponential blow-up)Conjugate points: where a geodesic stops being shortest
The sphere example flagged something dramatic: a Jacobi field that starts at zero can return to zero. When a nontrivial Jacobi field V along gamma satisfies V(0) = 0 AND V(t_0) = 0 for some t_0 > 0, we say gamma(t_0) is a conjugate point to p along gamma. Geometrically it means infinitely many geodesics leaving p in nearly the same direction refocus, to first order, at gamma(t_0) — the north and south poles of a sphere are the cleanest instance, where every meridian reconverges. The exponential map exp_p tells the same story analytically: gamma(t_0) is conjugate to p exactly when the differential d(exp_p) is singular there, so exp_p folds tangent vectors together and stops being a local diffeomorphism.
The payoff is a precise minimization rule, and it deserves stating carefully because the slogan is easy to over-claim. A geodesic minimizes length up to the first conjugate point; past the first conjugate point it can no longer be even a local minimum — there is always a nearby curve that is strictly shorter. On the sphere, the arc of a great circle from the north pole stays shortest until you reach the south pole (the conjugate point at distance pi); push beyond and the 'short way round' the other side beats it. Be careful though: a geodesic can fail to minimize BEFORE reaching a conjugate point, if another geodesic of equal length reaches the same point — that is the difference between the conjugate locus and the cut locus, which is where global minimization actually ends.
The Morse index theorem: counting the directions that fail
We can now make 'how badly does a geodesic fail to minimize' into a number. The index of a geodesic gamma is the index of the index form I, treated as a symmetric bilinear form on the infinite-dimensional space of variation fields vanishing at the endpoints — that is, the maximal dimension of a subspace on which I(V, V) < 0. It counts independent directions in which you can push the geodesic to strictly decrease its energy. Index zero means I is positive (a genuine local minimum); positive index means the geodesic is a saddle of the energy, beaten in that many independent directions. This is Morse theory of the path space, the same index that organizes critical points of any function, now living on the space of curves.
The bridge between the analysis (the index) and the geometry (conjugate points) is the Morse index theorem: the index of a geodesic gamma equals the number of conjugate points to its starting point in the open interval (0, L), each counted with multiplicity (the dimension of the space of Jacobi fields vanishing at both ends). The index is finite, so only finitely many conjugate points occur on a bounded geodesic. This is the clean accounting identity that makes everything earlier rigorous: 'minimizes up to the first conjugate point' is now literally 'index is zero before the first conjugate point, and jumps up by the multiplicity as you cross each one'. A worked check: on S^2 a great-circle arc of length slightly past pi has exactly one interior conjugate point, hence index 1 — one independent way to shorten it, namely sliding it off the antipode.
What you have built, and what runs on it next
Step back and see the assembled machine. From the second variation you extracted one symmetric form, the index form I; its critical fields are Jacobi fields, governed by the curvature-driven Jacobi equation; their reconvergence marks conjugate points, where minimization locally ends; and the Morse index theorem ties the count of those points to the index of I. The whole apparatus exists for one purpose: to convert a bound on curvature into a bound on the behavior of geodesics, and thence into topology. The constant-curvature trichotomy — sin(t), t, sinh(t) — is the template every comparison theorem measures the real manifold against.
- Guide 3 (positive curvature) feeds a positive lower Ricci bound into the index form to force a conjugate point by a fixed distance, proving the Bonnet-Myers theorem (the manifold is compact with finite fundamental group) and Synge's theorem.
- Guide 4 (nonpositive curvature) runs the same form with the opposite sign: no conjugate points ever, so exp_p is a covering map, giving the Cartan-Hadamard theorem and contractible universal covers.
- Guide 5 generalizes the Jacobi comparison to manifolds whose curvature merely lies between two bounds — the Rauch comparison theorem compares Jacobi field growth, and Toponogov compares whole triangles, leading to the sphere and splitting theorems.