Read the geodesic initial-value chapter first. We retain embedded surfaces and their induced metric. The inverse function theorem from multivariable calculus is a named input; differentiable ODE dependence, Gauss's lemma and local minimization are proved below. No Hopf–Rinow or general global minimizer theorem is used.
Motivation
Initial-value existence tells us how to start a geodesic. A shortest-path theorem must compare it with other paths, which the differential equation alone cannot do. The exponential map organizes the geodesics starting at one point by their initial velocities. Near zero it becomes a coordinate system; Gauss's lemma then shows that radial and angular motions are perpendicular. Angular detours cannot reduce the radial distance that a path must cover.
The comparison must also handle a competitor that leaves the coordinate neighborhood. Merely proving an inequality for paths staying inside it leaves a gap. We will use an intermediate coordinate sphere: a path escaping that smaller ball already spends more length than the short radial geodesic under consideration.
Definitions
Definition 4.4.1
(Exponential map)
Fix . Give the Euclidean inner product inherited from . Let be the affine geodesic with , . Where it is defined up to time one, put
Uniqueness and affine time scaling give whenever both expressions are defined. The vector length is the speed, so the radial segment , , has length .
Definition 4.4.2
(Normal ball and normal coordinates)
A normal ball about is a set , where and is a diffeomorphism on an open set containing the closed tangent ball . The normal coordinates of are .
Shrinking the radius after applying the inverse function theorem ensures this slightly stronger convenient condition. Normal coordinates are centered at one fixed point; the definition does not assert convexity for geodesics between every pair of points in .
For a piecewise path , length is . Pieces are finite in number. Our minimization statements compare all such paths with the prescribed endpoints, including those leaving and those with pauses.
Theorem/proposition
Theorem 4.4.3
(The exponential map gives local coordinates)
Near , is and is the identity under the natural identification of its tangent spaces with . Therefore a normal ball exists at every point.
Theorem 4.4.4
(Gauss's lemma)
For in a normal coordinate domain and ,
Consequently radial unit vectors retain length one, and vectors tangent to a tangent-space sphere map perpendicularly to the radial direction.
Concept lensGeometric
Geodesic polar coordinates
Choose an orthonormal basis of and write a nonzero normal coordinate vector as . On the punctured normal ball, are coordinates, with defined locally. Gauss's lemma with says that has unit length, and with that is perpendicular to . Since is two-dimensional, these two cases give the whole lemma by bilinearity. The lemma is therefore equivalent to the metric having the form
with no term. In the plane ; on the radius- sphere . The length inequality of the proof becomes visible: .
Theorem 4.4.5
(Short radial geodesics minimize length)
For every in a normal ball, the radial geodesic from to has length and minimizes length among all piecewise paths in joining those points. In particular sufficiently short geodesics starting at any point are shortest paths between their endpoints. This is a local endpoint conclusion, not existence for arbitrary pairs on the whole surface.
Proof sketch or proof idea
Proof
(Differentiable dependence and normal coordinates)
We first supply the differentiability omitted from the initial-value construction. On a common short interval , let solve the phase field with initial phase , all in one fixed closed convex ball. Write for a bound on . We use the elementary integral estimate
for nonnegative continuous . To prove it, repeatedly substitute the inequality into its integral. The first terms give , with remainder bounded by , where bounds on the compact interval. The remainder tends to zero. Applied to two solutions, this yields
Define the matrix by
On an interval with , this integral operator is a contraction on the complete space of continuous matrix functions, so the preceding Picard argument applies without a range-ball restriction. The estimate gives . Uniform continuity of on the fixed closed ball and the preceding solution estimate imply that depends continuously on : subtract the two integral equations, bound the coefficient difference uniformly, and apply the same integral estimate.
Let and . The fundamental theorem of calculus along segments in the ball gives a uniform modulus such that
Subtracting the integral equations and using , , gives
Thus , continuously. Together with , this proves joint dependence. Negative times follow by replacing by .
There is a small time-domain issue in defining . Uniform local existence near first gives time , not necessarily one. For sufficiently small, the solution with initial velocity exists to . The time-rescaled curve then has initial velocity and exists to time one. This also makes a function of .
At zero velocity the phase solution is constant. Linearizing , there gives , . An initial velocity variation therefore produces position variation . At time one this is , so is the identity. Apply the inverse function theorem in a surface patch, then choose a smaller closed tangent ball inside the resulting diffeomorphism domain. This proves that the exponential map gives local coordinates.
Proof
(Gauss's lemma by a geodesic variation)
For Gauss's lemma, set , and write , . Covariant differentiation in a patch means adding the Christoffel term to the derivative of vector components. The metric identity from the preceding chapter proves the product rule for inner products. Symmetry gives . The mixed derivative exists here because the phase flow, including its velocity, is in initial data; differentiating the integral formula for position interchanges the two derivatives used.
Every fixed- curve is a geodesic of constant speed . Consequently
At , because every curve starts at . Integrate to and put . There and , yielding the lemma. It is an identity of inner products, not a claim that the whole differential preserves lengths.
Proof
(Radial comparison and the escape barrier)
Now take a competing path inside , from to , and set . Where , write , . Then , and Gauss's lemma makes the radial and angular contributions perpendicular:
At the center, need not be differentiable, so do not integrate this formula blindly through every zero. If , choose and let be the last time that . It exists by continuity and compactness of the parameter interval. Afterwards , so ordinary piecewise differentiation and integration give
Let . If , nonnegativity of length suffices. The radial path attains .
Finally let a competitor move anywhere in . Choose . If it stays in , the previous argument applies. Otherwise consider its first exit from that smaller open ball. The image of is compact, hence closed in the Hausdorff surface, and lies inside . Continuity puts the exit point on the image of the radius- sphere: an interior coordinate point could not be a first exit. Up to that time the path lies in , so the same estimate gives length at least . A path leaving the neighborhood cannot improve on the radial geodesic. This proves the full comparison.
Proof X-ray
Where each ingredient of the minimization proof is used
The minimization argument rests on three separately proved inputs. Differentiable dependence is needed only to apply the inverse function theorem at zero, where the identity derivative is the one computation available; it yields a small domain and nothing more. Gauss's lemma is where the geodesic equations enter: each curve has constant speed, and symmetry of the Christoffel symbols gives . Together they produce the pointwise bound . Two limiting devices then close the gaps. The last crossing time avoids the point where is not differentiable. The intermediate radius , chosen with , handles competitors that leave: the image of the closed -ball is compact, so a first exit lies on its boundary sphere, and before reaching it the path has already spent length at least . The strict inequality makes escaping strictly worse. No global completeness theorem enters this local barrier argument.
Worked examples
Worked example 4.4.6
(The plane recovers the straight-line inequality)
For the Euclidean plane, and its differential is the identity everywhere. The radial-angular decomposition gives the usual fact that every path from to has length at least . Here normal coordinates can be taken on arbitrarily large balls; the global conclusion follows from this special formula, not from a general claim that every surface has a global exponential coordinate chart.
Worked example 4.4.7
(The sphere: an angular collapse at the antipode)
On the radius- sphere, for ,
A unit variation perpendicular to is multiplied in length by for , whereas radial length is unchanged. Thus the differential is invertible before the antipodal radius. The formula is one-to-one on and maps onto the sphere without . Applying the argument on smaller closed balls proves that arcs of length less than are minimizing. At radius , all directions meet at and the angular differential vanishes. A great-circle arc of length with is longer than its complementary arc of length .
Worked example 4.4.8
(Local minima do not supply a missing global minimizer)
In the punctured plane, take and . A path going along the axis to , around the upper semicircle of radius , and then along the axis to has length . Thus the infimum of path lengths is two, using also the Euclidean lower bound. It is not attained. Equality in would force on every smooth piece, hence throughout; continuity would then force passage through the missing origin. Every point still has normal neighborhoods with short minimizing radial paths. The failure concerns far endpoints on an incomplete surface.
A global hypothesis, metric completeness, turns local minimization into a minimizing geodesic between any two points.
A path to generalization
From normal balls to arbitrary endpoints
A global hypothesis, metric completeness, turns local minimization into a minimizing geodesic between any two points.
On a connected surface let be the intrinsic distance, the infimum of lengths of joining paths. The Hopf–Rinow theorem, treated in the note on intrinsic distance, says that if is complete, then every is defined on all of and any two points are joined by a geodesic of length . The punctured plane is not complete: points approaching the missing origin form a Cauchy sequence without a limit, and the minimizer above fails to exist.
Common mistakes
Every result above carries a hypothesis that is easy to drop. The counterexample below shows what happens when the endpoint is not in a normal ball.
Counterexample mode
A radial geodesic that has passed the antipode
Claim: whenever is defined, the radial geodesic , , is a shortest path from to . On the radius- sphere take . The sphere formula gives , a point on the equator relative to the pole . The radial geodesic has length , but the quarter great circle from in the direction reaches the same point with length . The claim fails. The theorem requires to lie in a normal ball; on the sphere that means .
Common mistake
(Gauss's lemma is not a local isometry)
Gauss's lemma preserves radial inner products, not every inner product. On a sphere the angular factor explicitly differs from one. The exponential map is generally not a local isometry from the flat tangent plane.
Common mistake
(Using the inverse function theorem beyond its domain)
The inverse function theorem applies at zero because its derivative was computed, and it only gives a small domain. One cannot assume injectivity at all velocities. Nor can an inequality proved for paths remaining in that domain silently cover paths that leave it; the exit-sphere argument is essential.
Common mistake
(A normal ball is centered at one point)
A normal ball is not automatically a convex normal neighborhood for every pair of its points. Likewise, shortest geometric paths may have many parametrizations; a pause changes the parameter without changing length. Affine parameters are needed when stating initial-value uniqueness.
Summary
Differentiating the phase integral equation supplies a exponential map with identity derivative at zero. The inverse function theorem gives normal coordinates. A geodesic variation proves radial-angular orthogonality, which bounds every path's length by its radial progress. An intermediate sphere handles escaping paths, establishing local length minimization without assuming global completeness.
Exercises
Quick check
(Why is the remainder estimate stronger than continuous dependence of solutions? State the derivative it identifies.)
Compare what is subtracted before dividing by .
Solution
Continuous dependence only gives . The proved statement subtracts the linear part and shows the residual divided by tends uniformly to zero. It identifies the Fréchet derivative of the path-valued solution map as . Continuity of then supplies dependence, the hypothesis needed for the inverse function theorem after evaluation in time.
Quick check
(On the sphere, differentiate the exponential formula in a unit direction perpendicular to . Compute the norm of its image and explain the failure at .)
Write and vary only the direction.
Solution
Write and choose unit . The derivative of at zero is zero, while the derivative of its unit direction is . Therefore , with norm . At this is zero, so the differential cannot be invertible. It does not contradict radial length preservation, because the killed direction is angular.
Quick check
(Prove that a path starting at and reaching the coordinate sphere of radius has length at least , even if it passes through several times before reaching that sphere.)
Use a last crossing time.
Solution
Until its first sphere hit, the path remains in the closed radius- normal ball, inside . For each , take its last radius- crossing before that hit. On the remaining segment the radius is positive, and the Gauss estimate integrates to . The total length is at least this value for every , hence at least . Earlier returns to the center do not matter because the last crossing was chosen.
Quick check
(On a fixed interval , show that the short radial path traversed at constant speed minimizes energy as well as length, and compute that energy.)
Combine the length comparison with the energy–length inequality.
Solution
Put . Every competitor has , and Cauchy–Schwarz gives . The path has constant speed , so its energy is exactly . This uses both the length comparison and constant-speed equality; an arbitrary stationary path need not satisfy either minimizing conclusion.
Quick check
(Let be the north pole of a sphere and . Choose two points at colatitude , with opposite longitudes and . Explain why this normal coordinate domain need not contain a shortest geodesic between those two points.)
Locate the short great-circle arc joining the two points.
Solution
The points are separated along their meridian great circle by a short arc of length through the south pole. The sphere formula, now centered at either endpoint, proves this arc minimizes because its length is less than . Any geodesic joining non-antipodal endpoints lies on their uniquely determined great-circle plane, by the explicit initial-value formula; all other joining arcs have length at least . Thus no shortest geodesic between these points stays in , although is the exponential image of the open radius- tangent ball at the north pole. Applying the chapter's convention literally uses smaller closed balls; this larger open domain still illustrates why coordinates centered at one point do not imply pairwise geodesic convexity.