Curves as points in an infinite-dimensional space
The previous rung handed you the geodesic equation nabla_(gamma') gamma' = 0 — a curve whose velocity is parallel-transported along itself, the straightest path a metric allows. We derived it as 'zero acceleration.' This guide reveals where that equation truly comes from, and the answer reorganizes everything: a geodesic is a critical point of a functional on a space of curves. The shift in viewpoint is the whole subject of geometric analysis in miniature — stop looking at one curve, and start looking at the entire space of curves between two points as if it were a single manifold, with length and energy as functions on it.
Fix two points p, q in a Riemannian manifold M and consider all smooth curves gamma: [0,1] -> M with gamma(0) = p, gamma(1) = q. This path space is an infinite-dimensional analogue of a manifold. What is a tangent vector to it at a particular curve gamma? It is a variation field: a vector field V(t) along gamma telling you how to nudge each point of the curve. To wiggle the curve, build a variation — a smooth family gamma_s(t) of curves with gamma_0 = gamma — and its infinitesimal velocity at s = 0 is exactly V(t) = (d/ds) gamma_s(t) at s = 0. Holding the endpoints fixed means V(0) = V(1) = 0. Differentiating a functional in the direction V is then just ordinary calculus, one variable s at a time.
First variation: differentiating energy once
Now compute dE/ds at s = 0. The calculation is a small masterpiece of the moving-frame technology you built last rung: differentiate |gamma_s'|^2 = g(gamma_s', gamma_s') in s, use metric-compatibility of the Levi-Civita connection to move nabla across g, then swap the two derivatives d/ds and d/dt — legal because the connection is torsion-free, so the mixed covariant derivatives of the map (s,t) -> gamma_s(t) commute. One integration by parts in t peels the derivative off V and lands it on gamma'. The boundary term dies because V vanishes at the endpoints, and out drops the first variation formula.
First variation of energy (endpoints fixed, V(0)=V(1)=0): d/ds E(gamma_s) = - integral_0^1 g( V(t) , nabla_(gamma') gamma'(t) ) dt at s=0 Vanishes for ALL admissible V <=> nabla_(gamma') gamma' = 0 (geodesic equation) With FREE endpoints, the surviving boundary term is [ g( V , gamma' ) ]_0^1 (this forces the transversality / Neumann condition)
Read the formula like a gradient. The integrand pairs the direction V you push in against the vector nabla_(gamma') gamma', the curve's acceleration. So the acceleration field IS the gradient of energy on path space — energy decreases fastest when you push the curve in the direction opposite its acceleration, straightening the bends. A curve is critical (dE = 0 for every V) if and only if its acceleration vanishes identically, which is the geodesic equation again, now born as an Euler-Lagrange equation rather than imposed by fiat. This is the cleanest possible answer to 'why is a geodesic straight?': it is the curve nothing can be saved by bending.
Second variation: where curvature enters
A critical point can be a minimum, a saddle, or worse — to tell them apart you need the second derivative, the Hessian of energy. Differentiate twice and evaluate AT a geodesic gamma (so the first-order acceleration term vanishes). The same commuting-derivatives bookkeeping runs again, but now d/ds and d/dt cross over a SECOND time, and swapping a pair of covariant derivatives is no longer free: their commutator is the Riemann curvature tensor itself, R(V, gamma')gamma'. This is the decisive moment of the whole rung — curvature is precisely the obstruction to second derivatives commuting, so it is forced to appear the instant you differentiate energy twice. The second variation formula is the result.
The resulting quadratic form is called the index form I(V, V) = integral_0^1 [ |nabla_(gamma') V|^2 - g(R(V, gamma')gamma', V) ] dt, and reading its sign carries the geometry. The first term |nabla_(gamma') V|^2 is always nonnegative — it is the cost of making the variation field wiggle. The curvature term is SUBTRACTED, and for V orthogonal to gamma' it equals the sectional curvature K times |V|^2. So positive curvature works against you: it lowers the second variation, and if you can find a V making I(V,V) < 0, the geodesic is not a local minimum — you can shorten it by bending.
This is the precise mechanism behind a fact you have always known intuitively: on a sphere, a great-circle arc longer than half the circle is a geodesic but NOT shortest. Two antipodal points are joined by infinitely many shortest arcs, and just past the antipode, positive curvature has destabilized the long way round. The same index form, compared against a constant-curvature model, is exactly what will power the Rauch comparison theorem later in this rung — so this little quadratic functional is the seed of the whole comparison toolkit.
Jacobi fields and conjugate points
The index form I(V,V) is a quadratic functional, and just as in finite dimensions the search for its critical directions yields an Euler-Lagrange equation. Setting the first variation of I to zero gives a linear, second-order ODE along the geodesic: nabla_(gamma') nabla_(gamma') J + R(J, gamma')gamma' = 0. Its solutions are Jacobi fields, and they have a beautiful concrete meaning — a Jacobi field is exactly the velocity field of a variation of gamma THROUGH geodesics. Equivalently, it is the infinitesimal spreading or focusing of a family of geodesics fired from a point in slightly different directions. The Jacobi equation is the linearization of the geodesic equation, with curvature as the restoring (or anti-restoring) force.
The picture is vivid. In flat space, R = 0, the Jacobi equation is J'' = 0, and geodesics fired from a point spread apart linearly forever — parallel lines stay parallel. With positive curvature, the equation becomes J'' = -K J, the harmonic-oscillator equation: nearby geodesics oscillate back TOWARD each other and refocus. On the unit sphere all geodesics from the north pole reconverge exactly at the south pole. That refocusing point is a conjugate point: q is conjugate to p along gamma when a nonzero Jacobi field vanishes at both ends. A conjugate point is where an infinitesimally-close family of geodesics from p crosses gamma again — the place where geodesics, set out in slightly different directions, come back to meet.
The Morse Index Theorem
Now everything converges into one clean accounting statement. The index of a geodesic gamma is the dimension of a maximal subspace of variation fields on which the index form I(V,V) is negative-definite — informally, the number of independent ways you can bend gamma to make it shorter, the count of 'downhill directions' of energy at this critical point. The Morse Index Theorem computes this otherwise-mysterious analytic number purely geometrically: the index of gamma equals the number of points conjugate to the starting point along gamma, each counted with its multiplicity (the dimension of the space of Jacobi fields vanishing there).
- Start with the analytic object: the index of gamma, defined as the maximal number of independent variation fields V (with fixed endpoints) on which the second-variation index form I(V,V) is negative — the count of independent shortening directions.
- Identify the geometric object: walk along gamma from the start and mark every conjugate point, each with its multiplicity (how many independent Jacobi fields vanish at both that point and the start).
- The theorem equates them: index(gamma) = (number of interior conjugate points, counted with multiplicity). A finite, integer-valued analytic quantity is computed by counting geometric crossings.
- Sanity check on the sphere S^n: the south pole is conjugate to the north pole with multiplicity n-1, so once a great-circle geodesic overshoots the south pole its index jumps by n-1 — matching the n-1 independent ways to shorten an over-long great-circle arc.
Why does this matter beyond a slick count? Three reasons frame the rest of geometric analysis. First, it is a finite-dimensional miracle: the index, defined over an infinite-dimensional path space, is always finite and computable — the seed of Morse theory, which reads the topology of the whole path space off the indices of its geodesics. Second, it converts curvature hypotheses into topology: if Ric > 0 with a uniform lower bound, conjugate points must appear within a bounded distance, forcing every long geodesic to have positive index and hence not minimize — that is the engine inside the Bonnet-Myers theorem, which concludes such a manifold is compact with finite fundamental group. Third, the same index form, compared against a constant-curvature model, is exactly what powers the Rauch comparison theorem in the next guides. Stated honestly: the theorem here is genuinely proved from the second variation formula, but its deployment in Bonnet-Myers, Synge, and the sphere theorem is a sequence of further arguments, each a real proof in its own right — we are surveying the road, not collapsing it.