Read the exponential-map and intrinsic-distance chapters first. For this chapter the surface and the variations used in differentiations are smooth; this stronger regularity avoids unnecessary mixed-derivative bookkeeping. The earlier initial-data dependence supplies the linearization. No Morse index theorem or comparison theorem is assumed.
Motivation
Neighboring geodesics may separate or refocus. The derivative of a geodesic family measures this change, and satisfies a linear equation involving curvature. The same curvature appears with a negative sign in the second derivative of energy: positive curvature can make transverse variations less costly and eventually create a shortening direction.
A zero second variation does not by itself produce a shorter path. On a sphere, antipodal endpoints are conjugate and still joined by minimizing semicircles. The obstruction becomes decisive when a conjugate point lies strictly inside the segment: one can splice the vanishing Jacobi field to zero and perturb it to obtain genuinely negative second variation. Even absence of conjugate points does not guarantee global minimization, as the flat cylinder will show.
Definitions
Let be a unit-speed geodesic, , and write . Covariant differentiation along it is denoted . In coordinates, . We use the curvature sign
so for orthonormal , consistently with the earlier Gauss equation.
Definition 4.6.1
(Jacobi field and conjugate point)
A Jacobi field is a vector field along satisfying
The point at time is conjugate to the initial point along this geodesic if some nonzero Jacobi field has .
The definition depends on the geodesic, not merely on two unlabeled points.
Definition 4.6.2
(Endpoint-fixed variation)
An endpoint-fixed variation of is a map with , and . Its variation field is , so . Its energy is .
Piecewise smooth paths and continuous piecewise smooth fields are allowed on a common finite partition; the first two derivatives with respect to agree at the joins.
Theorem/proposition
Theorem 4.6.3
(Linearized geodesics are Jacobi fields)
The field obtained by differentiating a family of geodesics satisfies the Jacobi equation. Choose a parallel unit field perpendicular to . Writing reduces it to
Initial values and determine one Jacobi field.
Theorem 4.6.4
(The second variation and index form)
For an endpoint-fixed variation of the geodesic,
where
Every continuous piecewise smooth field vanishing at the endpoints can be realized by such a variation. If for one such field, the geodesic is not length minimizing.
Theorem 4.6.5
(An interior conjugate point prevents minimization)
If some is conjugate to zero along , then is not length minimizing. The strict inequality is essential; conjugacy at the final endpoint alone does not yield this conclusion.
Proof sketch or proof idea
Proof
(Linearizing a family of geodesics)
For a geodesic variation write and . Symmetry of the Christoffel functions gives . Curvature measures their next commutation:
The first term is zero because each curve is geodesic. Restrict to to obtain the Jacobi equation. The phase dependence established earlier suffices: differentiating its integral equation in initial data gives a continuous derivative of velocity and a twice-time-differentiable position variation satisfying the linearized equation. Thus no unproved higher smoothness of the flow is needed for this conclusion.
A parallel normal field is obtained by solving the linear equation from one unit vector perpendicular to . Local linear ODE existence and the integral estimate bound solutions on compact intervals, allowing finite chart gluing along . The metric product rule preserves and . The surface curvature identity gives , proving the scalar equations. Their second-order linear initial-value problems have existence and uniqueness by the same ODE method.
Every Jacobi field can in fact be obtained by varying initial data. Here is the coordinate construction, which also explains the four scalar degrees of freedom. Suppose its desired initial values are and , and the base geodesic has coordinate initial velocity . Move its initial position along in a chart. At that moving point prescribe velocity components
The connection term makes the covariant derivative of this moving velocity equal to , rather than merely making its coordinate derivative equal to . The common-time local flows, composed along a finite partition if necessary, produce a geodesic variation. Its field has initial values , so Jacobi uniqueness identifies it with the prescribed solution. The varied geodesics are not required to retain unit speed; requiring that would unnecessarily exclude some initial variations.
Proof
(Differentiating energy twice and realizing fields)
For the second variation, differentiation under the energy integral first gives , using endpoints and . For a second derivative, put and commute derivatives before integrating:
The middle term has zero integral: integrate the metric product rule, use the fixed endpoints, which give there, and use . On a surface for , while . This proves the displayed index form without leaving its sign convention ambiguous. In the parallel frame,
which also proves symmetry. For piecewise fields, integrate over each piece. The boundary terms involving cancel at internal joins because these fields agree there. Derivative jumps in are therefore permitted.
To realize , cover the time interval by finitely many chart intervals and split with smooth time cutoffs subordinate to them. Such cutoffs follow by choosing finitely many positive bump functions and normalizing their sum. In each chart perturb coordinates by times the coordinate components of , then compose these finitely many small perturbations in a fixed order. Compactness and small keep every operation in its chart. At their first derivatives add to , and the endpoint shifts vanish. This constructs the needed continuous piecewise smooth variation with the required matching derivatives.
If its index is negative, Taylor's formula gives for sufficiently small nonzero . Cauchy–Schwarz gives , hence a shorter path. The implication uses the fixed interval and the original unit speed; it is not an unsupported identification of energy and length.
Proof
(Truncate the Jacobi field, then perturb the corner)
Now let be a nonzero Jacobi field vanishing at zero and at . Its tangential component is affine and vanishes twice, so is identically zero. Also , since zero position and derivative would force the whole Jacobi solution to vanish. Define a continuous piecewise smooth field to equal on and zero on . Integration by parts and the Jacobi equation give
for every endpoint-zero field . Choose a smooth field supported near with , using the parallel frame and a scalar bump. Then
for sufficiently small positive . Realization and the preceding energy argument give a shorter path. At , endpoint-fixed fields have , so this negative cross term is unavailable; that is where strict interior location enters the proof.
Proof X-ray
Where the strict interior position of the conjugate point is used
The proof has two layers. The analytic layer turns any endpoint-zero field of negative index into a genuinely shorter path: realization gives a variation, Taylor's formula gives , and Cauchy–Schwarz on the fixed interval gives . The geometric layer produces such a field. The truncated Jacobi field has zero index and shortens nothing by itself. The key move is the cross term: integration by parts on leaves only , and by Jacobi uniqueness, which is where is used. With the linear term dominates . The hypothesis enters exactly here: must be nonzero at time while still vanishing at . At this is forbidden, which is why the conjugate semicircle on the sphere can still minimize.
Worked examples
Worked example 4.6.6
(The sphere focuses at the antipode)
For , the normal solution with , is . Its first positive zero is . Thus a unit-speed great circle has an antipodal conjugate point at that time. The preceding distance computation shows the semicircle still minimizes. If , that point is interior and the segment cannot minimize. One can also test directly in the index form to obtain .
Worked example 4.6.7
(A flat cylinder loses global optimality without conjugate points)
The cylinder has , so normal Jacobi fields have affine coefficient . No nonzero one vanishes at zero and a positive time. Nevertheless a unit-speed horizontal circle stops being minimizing beyond length , by the cylinder distance formula: the opposite angular route becomes shorter. Every fixed-endpoint index is positive for nonzero fields, yet long wrapped segments are globally beaten. Local variation stability does not erase topology or competing routes far from the path.
Worked example 4.6.8
(Negative curvature separates nearby directions)
Along a segment where , the normalized normal Jacobi coefficient is , with no positive zero. It grows faster than the flat value . This is a calculation along a specified segment, not a claim that a complete constant-negative-curvature plane embeds in . It describes infinitesimal variation of direction; finite distances between arbitrary geodesics require additional comparison results.
Common mistakes
Common mistake
(A Jacobi field need not be normal)
A Jacobi field describes first-order displacement of geodesics and need not point perpendicular to the base curve. Its tangential component is affine; vanishing at two endpoints forces that component away. Curvature and parameter speed also matter: the scalar equation above assumes unit speed.
Common mistake
(Positive index is not global minimality)
Zero second variation does not decide minimization. Negative index proves shortening, but positivity alone is not a proof of global minimality. The flat cylinder supplies an explicit warning.
Nor does the interior-conjugate theorem say that the first conjugate endpoint must already fail to minimize.
Counterexample mode
A conjugate final endpoint can still be minimizing
Claim: if is conjugate to along , then is not length minimizing. On the radius- sphere take a unit-speed great circle with . The field satisfies , so it is a nonzero Jacobi field, and it vanishes at and . Yet the semicircle has length , which equals the distance between antipodal points, so it is minimizing. The claim fails. The theorem needs the conjugate time strictly inside ; for it is, and the segment does not minimize.
Common mistake
(Reversing the curvature sign convention)
All signs follow the stated curvature convention. Reversing the definition of without changing the Jacobi and index formulas would falsely predict hyperbolic focusing and spherical defocusing.
The scalar equation becomes a linear system, and the Morse index theorem counts conjugate points.
A path to generalization
Jacobi fields in higher dimensions
The scalar equation becomes a linear system, and the Morse index theorem counts conjugate points.
On an -dimensional Riemannian manifold, the normal part of a Jacobi field lies in an -dimensional space, and becomes a linear system governed by ; in a parallel orthonormal frame its diagonal entries are sectional curvatures. A conjugate point can then have multiplicity up to . The Morse index theorem, not proved here, says that the index of on endpoint-zero fields equals the number of conjugate points in , counted with multiplicity.
Summary
Linearizing geodesics yields a curvature-controlled Jacobi equation. Differentiating energy twice produces the symmetric index form. An interior conjugate point creates a negative direction through truncation and a cross-term perturbation, whereas an endpoint conjugate point may still minimize. Curvature controls infinitesimal stability; global shortest paths also depend on the surrounding geometry and topology.
Exercises
Quick check
(Show every Jacobi field has . Why must a field vanishing twice be normal?)
Solution
Since is parallel, the scalar derivative calculation gives . Thus it is affine. Two distinct zeros force , leaving only the parallel normal component. This does not force a general Jacobi field to be normal.
Quick check
(Compute the index of on a sphere. Interpret the positive, zero and negative cases without claiming more than the calculation proves.)
Solution
Here , and the integrals of both squared sine and squared cosine are . Hence . It is positive for , zero at equality and negative above it. Only the last sign alone proves shortening. Minimization at and below the antipodal length follows from the earlier distance proof, not from this one test field.
Quick check
(If along the segment, prove the normal solution , satisfies for . Show the index is strictly positive for every nonzero endpoint-zero field. Does that imply global minimization?)
Solution
Initially . As long as it stays positive, , so and . A first positive zero is therefore impossible. The index is , strictly positive for any nonzero continuous piecewise smooth endpoint-zero field: zero derivative norm would make both coefficients constant and then zero. Global minimization does not follow; the flat cylinder has exactly this curvature and long nonminimizing wrapped segments.
Quick check
(Suppose , , , and on . For smooth with , prove the identity below. Explain the boundary limit at zero and the strict equality case.)
Solution
On , put and expand to get . The boundary term is . It vanishes at since . Near zero, , , and , so the boundary term tends to zero. The nonnegative integral therefore has the stated improper limit. If it vanishes, is constant on ; and force that constant to be zero. Thus only has zero index under these hypotheses.
Quick check
(Prove time is conjugate along a unit-speed geodesic starting with if and only if is singular. Assume the geodesic and nearby initial data are defined through time .)
Solution
Vary the initial unit-direction vector without imposing unit length on the varied data: , . Its Jacobi field has , , and . A nonzero kernel vector therefore gives a nonzero Jacobi field vanishing at both endpoints. Conversely, such a field has nonzero initial derivative by uniqueness. Choose ; linearized initial-value uniqueness identifies it with the displayed variation field, so lies in the kernel. The two-dimensional differential is singular exactly when its kernel is nontrivial.