This note uses fundamental forms, normal curvature and the metric Gauss equation. Let be a nonempty compact surface without boundary and a immersion. Self-intersections are allowed. A global orientation is not required: normals below are chosen locally. Compactness and absence of boundary are hypotheses that will be used, not inferred from a drawing.
Motivation
Local metric data need not admit a requested global realization in three-dimensional Euclidean space. A flat square torus is locally Euclidean, but the familiar doughnut surface has both positive and negative curvature. This difference can be proved without a classification of all immersed surfaces.
Choose any center in space and enclose the entire compact image by the smallest concentric sphere. At a farthest point, the image is tangent to that sphere and remains on its inner side. An exact second-derivative inequality forces both principal curvatures there to bend at least as strongly as the enclosing sphere. In particular Gaussian curvature is positive somewhere. This elementary support argument yields global consequences that no single arbitrary coordinate patch could establish.
Definitions
Definition 7.1.1
(Fundamental forms and principal curvatures)
In a local chart write and choose a unit normal . The second fundamental form is . Its principal curvatures are the eigenvalues of the -self-adjoint shape operator determined by . Put and .
Reversing reverses both and but preserves and . These two quantities are therefore defined at every point even when no global normal exists.
Definition 7.1.2
(Elliptic points, minimal immersions and flat metrics)
An elliptic point is a point with . A minimal immersion is one with in every local choice of normal. A flat metric is a Riemannian metric whose intrinsic Gaussian curvature is identically zero.
For the metric-isometric corollary we assume , matching the metric Gauss equation: a isometric immersion into must have the same as its prescribed metric. No extension of that theorem is used.
For a fixed center put
A farthest point realizes the maximum by compactness. This number is positive: if it were zero, would be constant and could not be an immersion. Tangent vectors below belong to , and for nonzero , even if another domain point has the same image.
Theorem/proposition
Theorem 7.1.3
(A supporting sphere forces an elliptic point)
At a point maximizing , choose . Then, as quadratic forms on ,
Consequently both principal curvatures are at most , and
The normal choice is local; the assertion applies to nonorientable immersions as well.
The first theorem is local at one point, but the point exists only because of compactness. Two global consequences follow at once.
Corollary 7.1.4
(Global obstructions from one maximum)
No nonempty compact boundaryless surface has a immersion into with everywhere, or with everywhere. In particular a compact flat Riemannian surface has no isometric immersion into . These statements allow self-intersections and therefore exclude more than embedded realizations.
Theorem 7.1.5
(Optimal enclosing-radius bounds)
The function has a minimum over all centers . Every such immersion satisfies
A round sphere attains both bounds. These inequalities do not claim that attaining just one bound characterizes every possible surface.
Proof sketch or proof idea
Proof
(The Hessian of squared distance at a farthest point)
Let . At its maximum , its differential vanishes. Thus is perpendicular to ; dividing by its positive length gives the stated unit normal. Extend that normal to a local continuous unit normal field with the chosen value; a immersion supplies a normal through the normalized coordinate cross product.
Differentiate in coordinates:
At this is . At a critical point the Hessian quadratic form is coordinate independent: the extra coordinate-change terms multiply first derivatives of , which are zero. For any choose a chart curve through with that tangent. Since has no boundary, it exists for both positive and negative small parameters. The real second-derivative test gives
This is exactly the quadratic-form inequality. Choose a -orthonormal principal basis, using the spectral theorem for a self-adjoint operator on a two-dimensional inner-product space. Applying the inequality to each unit principal vector gives , hence . Multiplying two negative numbers bounded away from zero gives , and averaging gives in this normal choice. With the opposite normal has the opposite sign; gives the orientation-independent result.
Proof X-ray
Where the global hypotheses enter
Compactness is used once, to produce a farthest point; nothing else in the argument is global. The immersion hypothesis is used twice: it makes positive definite, and it rules out , so the normal is defined. Absence of boundary is what makes the second-derivative test two-sided in every tangent direction. At a boundary point only inward directions are available, and a maximum there gives no sign for the Hessian in the outward direction.
The key move is to compute the Hessian of instead of appealing to a tangency picture. At a critical point that Hessian is a well-defined quadratic form, and in coordinates it is exactly . The proof then closes by linear algebra: is positive definite, so both principal curvatures have the same sign and are bounded away from zero. Injectivity and orientability never enter, because every step takes place at one domain point with a locally chosen normal.
The positive product also persists on a neighborhood of by continuity of . Thus the obstruction is not an isolated numerical accident: a nonempty open set of positive Gaussian curvature is forced.
Proof
(From one elliptic point to the global obstructions)
If everywhere, the positive lower bound at contradicts it. If everywhere, the lower bound for contradicts it instead. A isometric immersion of a compact flat metric would have everywhere by the metric Gauss equation, so it is excluded by the first obstruction. One cannot evade this by allowing the image to cross itself, since injectivity was never used.
Proof
(Existence of an optimal center)
The reverse triangle inequality gives
Also, fixing any gives . Thus is continuous and tends to infinity as . Choose a closed ball outside which exceeds ; inside that ball compactness supplies a minimum, which is global. It is positive by the immersion argument already given. Apply the first theorem with a minimizing center. The continuous functions and attain maxima on compact , and these maxima dominate the bounds at the selected farthest point. No assertion about uniqueness of the minimizing center is needed.
The inverse function theorem adds one local consequence. The differential of a chosen Gauss map is minus the shape operator, whose determinant is . At the elliptic point this is invertible, so the Gauss map sends a small surface neighborhood diffeomorphically onto a neighborhood of its unit normal. This does not by itself prove a classification or a global degree formula for the Gauss map.
Worked examples
Worked example 7.1.6
(A round sphere realizes the bounds)
For , choose the outward normal . Its principal curvatures in our convention are both , so and . Its optimal enclosing radius is : any ball containing two opposite sphere points must have radius at least half their distance, namely , and the center achieves it. Both inequalities are equalities, with the sign of reversed under the inward orientation.
Worked example 7.1.7
(A ring torus bends in two different ways)
For constants parametrize
The outward normal is . Direct differentiation gives , and , with second-form coefficients , and . Thus
The outer equator is elliptic, the inner equator has negative and the two intermediate circles have . From the center the farthest radius is , since . At the curvatures are and , satisfying the support inequality, the first with equality. The area element is , so the total curvature is . Positive curvature somewhere is compatible with zero total curvature; no Gauss–Bonnet input was needed for this computation.
Worked example 7.1.8
(The flat torus is a metric, not this doughnut)
Identify points of differing by integer translations. Small disks of radius less than project injectively, and translation transition maps preserve the Euclidean metric. The quotient therefore has a well-defined locally Euclidean metric with . It is compact because the closed unit square surjects onto it, and has no boundary because those small charts are open disks.
If that particular metric admitted a isometric immersion into , the obstruction theorem would give both everywhere and somewhere. It cannot. A ring torus is a smooth embedded realization of the same topological torus, but the computed curvature shows that it realizes a different metric. This proves a geometric obstruction without claiming that the abstract torus cannot be embedded at all, or making any assertion about lower-regularity isometries.
Yes, in four-dimensional space; in three-dimensional space only with less regularity than the curvature argument needs.
A natural next question
Can the flat torus be realized isometrically elsewhere?
Yes, in four-dimensional space; in three-dimensional space only with less regularity than the curvature argument needs.
The map
is invariant under integer translations and injective on the quotient. Its two partial derivatives are orthonormal, so it is an isometric embedding of the flat torus into . The obstruction above is specific to and to second derivatives: the Nash–Kuiper theorem gives isometric embeddings of this flat torus into , for which no second fundamental form is defined.
Common mistakes
Common mistake
(Switching sign conventions halfway)
At the farthest point the normal pointing away from the center gives negative principal curvatures in this convention. The useful statement is that their product is positive and their absolute mean is bounded below. Silently switching to the inward normal halfway through the argument would reverse the inequality while keeping the old signs of , which is inconsistent.
Common mistake
(Reading “somewhere” as “everywhere”)
The theorem gives somewhere, not everywhere: the ring torus has on its inner equator. Nor does an obstruction to an isometric immersion exclude a topological embedding with another metric. The ring torus embeds the torus, but with a metric whose curvature changes sign.
The proof does not require to be embedded, orientable or simply connected; adding those assumptions would weaken the statement. Compactness and absence of boundary, by contrast, cannot be dropped.
Counterexample mode
Boundary or noncompactness removes the elliptic point
The plausible claim is that every surface in has an elliptic point. A closed flat planar disk is compact and has everywhere. Its farthest points from the origin lie on the boundary circle, where the outward radial direction leaves the parameter domain, so the two-sided second-derivative test is unavailable.
The whole plane is boundaryless, minimal and flat, but noncompact, and has no farthest point from a fixed center. For a curved noncompact example, the graph has everywhere. The correct statement needs both compactness and empty boundary.
Summary
At a farthest point, differentiating squared distance yields . Positive definiteness forces two same-sign nonzero principal curvatures and hence positive , as well as a mean-curvature lower bound. This excludes compact boundaryless minimal and flat immersions, supplies optimal enclosing-radius estimates, and explains why the locally flat torus metric cannot be the metric of an ordinary smooth doughnut in .
Exercises
Quick check
(If the image is contained in a ball of radius , prove that and , even if the chosen ball is not optimal.)
Compare with the actual farthest distance from the ball's center.
Solution
For the given center the actual farthest radius satisfies . At its farthest point and . Taking global maxima retains the inequalities. No contact with the boundary of the originally chosen ball is necessary.
Quick check
(Suppose is a positive constant . What lower bound follows for every enclosing radius? Show that the sphere is sharp.)
Apply the support theorem at a farthest point.
Solution
At a farthest point , so . Every enclosing ball has radius at least its actual farthest distance, hence radius at least . A sphere of radius has constant and is enclosed by its own ball of radius , proving sharpness without a classification of constant-mean-curvature surfaces.
Quick check
(In the ring torus example, identify the parabolic circles. Why do they not contradict the global obstruction?)
Find where and compare with what the theorem excludes.
Solution
Since , exactly when , namely or modulo . The theorem only excludes at every point. The outer equator has positive and supplies the required elliptic point; some flat points are perfectly compatible with it.
Quick check
(Rewrite the support inequality using the opposite normal. Which curvature conclusions survive unchanged?)
Reversing changes the sign of .
Solution
Reversing changes to , so the inequality becomes . Both new principal curvatures are at least . and keep their original bounds, while signed changes from at most to at least . This also explains why a global orientation was not needed.
Quick check
(For any unit vector , prove that the orthogonal projection of a compact immersed boundaryless surface onto the plane cannot itself be an immersion everywhere.)
Use the inverse function theorem and compactness of the image.
Solution
If the projected map were an immersion, it would have invertible differential between two-dimensional spaces everywhere. The inverse function theorem would make it a local diffeomorphism and hence an open map. Its image is also a nonempty compact subset of the plane, hence closed. A nonempty subset both open and closed in the connected plane must be the whole plane, which is not compact. This contradiction needs no global injectivity, and it shows that every projection direction has a singular point.