The Gauss Map

Take a smooth surface sitting in three-dimensional space. At every point, erect the unit vector perpendicular to it. As the point moves across the surface, this arrow sweeps out a path. It traces a path on the unit sphere. This assignment — point on the surface to point on the sphere — is the Gauss map. It is one of the most consequential ideas in the history of geometry. The reason it matters is that the sphere records direction. Direction is exactly what curvature is about. A flat plane has a constant normal. So its Gauss map collapses to a single point: no curvature, no spread. A sphere of radius rr has a Gauss map that is, up to scaling and orientation, the identity. The normal direction changes as fast as possible. The surface curves as tightly as possible. Between these extremes, the rate at which the Gauss map distorts area and direction matters. It is a precise numerical measure of how curved the surface is at each point. This is the seed from which Gauss's Theorema Egregium grows. The whole theory of intrinsic curvature grows from it too. The big picture is that the Gauss map converts a geometric question into an analytic one. The geometric question is "how curved is this surface?" The analytic one is "what is the differential of this particular map between surfaces?" Both the surface and the sphere are two-dimensional. So the differential of the Gauss map at a point is a linear operator. This operator acts on a two-dimensional tangent space. Linear operators on a plane are things we know how to dissect completely. We do this via eigenvalues, determinants, and traces. Gauss's genius was to recognize something important. These algebraic invariants of one linear map encode all the local curvature information of the surface. This article builds the Gauss map rigorously. We define it as a genuine smooth map between regular surfaces. We examine its differential as a self-adjoint linear operator, called the shape operator. We prove the key structural facts that make it worth studying. This sets the stage precisely for the second fundamental form and the curvatures it defines.

Key takeaways
  • The Gauss map NN sends each point of an oriented surface to the tip of its unit normal vector on the unit sphere S2S^2.
  • Both the surface and S2S^2 are 2-dimensional, so the differential dNpdN_p is a linear map from a tangent plane to a (parallel) tangent plane.
  • A key algebraic fact: dNpdN_p is self-adjoint with respect to the first fundamental form, which forces it to have real eigenvalues and orthogonal eigenvectors.
  • These eigenvalues are the principal curvatures; their product is the Gaussian curvature and their average is the mean curvature.
  • The Gauss map is only defined on orientable patches (or orientable surfaces), since a consistent choice of unit normal is required.
  • Key idea: the differential of the Gauss map is exactly the linear operator whose invariants (determinant, trace, eigenvalues) are the curvatures of the surface.

Where this lives today

The Gauss map's differential is the direct ancestor of the second fundamental form and the principal, mean, and Gaussian curvatures used throughout classical and modern Differential Geometry. In general relativity, the analogous construction on spacelike hypersurfaces (the extrinsic curvature tensor) governs how a slice of spacetime bends within the ambient four-dimensional geometry, entering the Gauss–Codazzi equations used in numerical relativity. In computer graphics and geometric modeling, discretized Gauss maps and curvature estimates from triangulated meshes drive shading, remeshing, and feature-detection algorithms. In computer vision, the Gauss map underlies shape-from-shading and surface reconstruction, since apparent shading of an object is a function of the surface normal direction relative to a light source. In materials science and elasticity, Gaussian curvature computed via the Gauss map explains why doubly-curved shells (domes, saddle roofs) resist bending in ways flat or singly-curved sheets do not — a fact traceable to the regularity of the shape operator's eigenvalues. The topological degree of the Gauss map on a closed orientable surface, via the Gauss–Bonnet theorem, links the total curvature to the surface's Euler characteristic, forming a bridge to algebraic topology.

Ask the tutor: How does the self-adjointness of dNpdN_p shown here directly produce the symmetric matrix of the second fundamental form?

DefinitionThe Gauss Map

Let S⊂R3S \subset \mathbb{R}^3 be a regular surface that is orientable, meaning there exists a smooth choice of unit normal vector field

N:S→R3,N(p)⊥TpS,∥N(p)∥=1for all p∈S,N : S \to \mathbb{R}^3, \qquad N(p) \perp T_pS, \qquad \lVert N(p) \rVert = 1 \quad \text{for all } p \in S,

varying smoothly with pp (i.e., in any local parametrization x(u,v)\mathbf{x}(u,v) with coordinate vectors xu,xv\mathbf{x}_u, \mathbf{x}_v, one may take

N=xu×xv∥xu×xv∥N = \frac{\mathbf{x}_u \times \mathbf{x}_v}{\lVert \mathbf{x}_u \times \mathbf{x}_v \rVert}

and this expression is smooth in (u,v)(u,v)). Since N(p)N(p) is a unit vector, we may regard it as a point of the unit sphere S2={w∈R3:∥w∥=1}S^2 = \{ w \in \mathbb{R}^3 : \lVert w \rVert = 1\}. The map

N:S→S2N : S \to S^2

is called the Gauss map of the oriented surface SS. Because TpST_pS and TN(p)S2T_{N(p)}S^2 are both, as vector subspaces of R3\mathbb{R}^3, equal to the plane orthogonal to N(p)N(p) (this is immediate: TN(p)S2=N(p)⊥T_{N(p)}S^2 = N(p)^\perp by definition of the sphere's tangent plane, and TpS=N(p)⊥T_pS = N(p)^\perp by definition of NN), we may canonically identify

TpS=TN(p)S2.T_pS = T_{N(p)}S^2.

Under this identification, the differential

dNp:TpS→TpSdN_p : T_pS \to T_pS

is a well-defined linear operator on the single vector space TpST_pS. This operator is called the shape operator (or Weingarten map) at pp, often written Sp:=−dNpS_p := -dN_p (the sign convention varies by author; we follow the convention dNpdN_p directly for now and note the sign issue explicitly later).

Intuition

Picture the surface as a landscape and imagine standing an arrow straight up out of the ground at every point — perpendicular to the local ground plane, always pointing to the same "outside." As you slide from point to point on the landscape, the arrow tips trace a path on a unit sphere hovering nearby. On a perfectly flat table, the arrow never turns: all the tips land on one point of the sphere, and the Gauss map is constant. On a sphere of radius rr, if you walk a distance r θr\,\theta along a great circle, the arrow (which always points radially outward) rotates by exactly θ\theta — the map from the surface to S2S^2 is a rescaled copy of the identity. The interesting case is a saddle, like the middle of a Pringle chip or the surface z=xyz = xy near the origin. Moving in one direction, the normal tips lean one way; moving in the perpendicular direction, they lean the opposite way. This reversal of orientation under the Gauss map is the geometric signature of negative curvature, and it is invisible if you only look at the map pointwise — you need to track how it distorts a small patch of tangent directions, which is precisely what the differential dNpdN_p captures. The key structural fact, proved below, is that dNpdN_p is a self-adjoint linear operator on the 2-dimensional space TpST_pS. Self-adjoint operators on an inner product space always diagonalize with real eigenvalues and orthogonal eigenvectors — this is pure linear algebra, but it lands squarely on the surface: the eigenvalues are the principal curvatures k1,k2k_1, k_2, and the eigenvectors are the principal directions along which the surface bends the least and the most.

Figure
The Gauss map sends unit normals at three points of a curved surface to three points on the unit sphere.
surface S N(p₁) N(p₂) N(p₃) unit sphere S² N(p₁), N(p₂), N(p₃) Gauss map N
TheoremSelf-adjointness of the differential of the Gauss map

Let S⊂R3S \subset \mathbb{R}^3 be a regular, orientable surface with a chosen smooth unit normal field N:S→S2N: S \to S^2, and let p∈Sp \in S. Equip TpST_pS with the inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle inherited from R3\mathbb{R}^3 (the first fundamental form). Then the differential

dNp:TpS→TpSdN_p : T_pS \to T_pS

is a self-adjoint linear operator, i.e.

⟨dNp(w1),w2⟩=⟨w1,dNp(w2)⟩for all w1,w2∈TpS.\langle dN_p(w_1), w_2 \rangle = \langle w_1, dN_p(w_2) \rangle \qquad \text{for all } w_1, w_2 \in T_pS.

Consequently dNpdN_p has two real eigenvalues k1≤k2k_1 \le k_2 (the principal curvatures at pp, possibly equal), and if k1≠k2k_1 \ne k_2 the corresponding eigenvectors are orthogonal; if k1=k2k_1 = k_2 then dNpdN_p is a scalar multiple of the identity and every direction is an eigenvector.

Proof

Setup. Let x:U→S\mathbf{x} : U \to S be a local parametrization around pp, with U⊂R2U \subset \mathbb{R}^2 open and x(u0,v0)=p\mathbf{x}(u_0,v_0) = p. Write xu,xv\mathbf{x}_u, \mathbf{x}_v for the partial derivative vectors, which by regularity of the parametrization form a basis of TpST_pS at each point of the patch. It suffices to verify self-adjointness on this basis, since a bilinear form that is symmetric on a spanning set is symmetric everywhere (bilinearity distributes the check to all pairs of basis vectors, and symmetry on each pair extends by linearity to all linear combinations).

Step 1: Express dNpdN_p on the basis {xu,xv}\{\mathbf{x}_u, \mathbf{x}_v\}. Since N∘x:U→S2⊂R3N \circ \mathbf{x} : U \to S^2 \subset \mathbb{R}^3 is a smooth map, by the chain rule (definition of the differential of a map between surfaces, as established for The Tangent Plane and Differential of a Map) we have

dNp(xu)=Nu,dNp(xv)=Nv,dN_p(\mathbf{x}_u) = N_u, \qquad dN_p(\mathbf{x}_v) = N_v,

where Nu=∂(N∘x)/∂uN_u = \partial (N\circ\mathbf{x})/\partial u and Nv=∂(N∘x)/∂vN_v = \partial(N \circ \mathbf{x})/\partial v, evaluated at (u0,v0)(u_0,v_0). This is simply the statement that the differential of NN applied to a tangent vector realized as a velocity of a coordinate curve equals the velocity of the image curve.

Step 2: Self-adjointness reduces to a single scalar identity. Self-adjointness of a linear operator on a 2-dimensional inner product space, verified on a basis {xu,xv}\{\mathbf{x}_u, \mathbf{x}_v\}, requires checking

⟨dNp(xu),xv⟩=⟨xu,dNp(xv)⟩,\langle dN_p(\mathbf{x}_u), \mathbf{x}_v \rangle = \langle \mathbf{x}_u, dN_p(\mathbf{x}_v) \rangle,

i.e.

⟨Nu,xv⟩=⟨xu,Nv⟩.\langle N_u, \mathbf{x}_v \rangle = \langle \mathbf{x}_u, N_v \rangle.

(The two diagonal identities ⟨dNpxu,xu⟩=⟨xu,dNpxu⟩\langle dN_p \mathbf{x}_u, \mathbf{x}_u\rangle = \langle \mathbf{x}_u, dN_p\mathbf{x}_u\rangle and similarly for xv\mathbf{x}_v are trivially true for any bilinear form, being of the form a=aa=a; only the cross term is a genuine constraint.)

Step 3: Prove the cross-term identity using ⟨N,xu⟩≡0\langle N, \mathbf{x}_u\rangle \equiv 0 and ⟨N,xv⟩≡0\langle N, \mathbf{x}_v \rangle \equiv 0. Since NN is by definition orthogonal to the tangent plane at every point of the patch, the two smooth scalar functions on UU,

⟨N,xu⟩and⟨N,xv⟩,\langle N, \mathbf{x}_u \rangle \quad \text{and} \quad \langle N, \mathbf{x}_v \rangle,

are identically zero on UU. Differentiate the first identity with respect to vv, and the second with respect to uu, using the product rule for the inner product (valid since ⟨⋅,⋅⟩\langle \cdot,\cdot\rangle is bilinear and N,xu,xvN, \mathbf{x}_u, \mathbf{x}_v are smooth vector-valued functions of (u,v)(u,v)):

∂∂v⟨N,xu⟩=⟨Nv,xu⟩+⟨N,xuv⟩=0,\frac{\partial}{\partial v}\langle N, \mathbf{x}_u \rangle = \langle N_v, \mathbf{x}_u \rangle + \langle N, \mathbf{x}_{uv} \rangle = 0,
∂∂u⟨N,xv⟩=⟨Nu,xv⟩+⟨N,xvu⟩=0.\frac{\partial}{\partial u}\langle N, \mathbf{x}_v \rangle = \langle N_u, \mathbf{x}_v \rangle + \langle N, \mathbf{x}_{vu} \rangle = 0.

Because x\mathbf{x} is a smooth (in particular C2C^2) parametrization, mixed partials commute: xuv=xvu\mathbf{x}_{uv} = \mathbf{x}_{vu}. Subtracting the two equations above (or simply comparing them, since both contain the identical term ⟨N,xuv⟩\langle N, \mathbf{x}_{uv}\rangle) gives

⟨Nv,xu⟩=−⟨N,xuv⟩=⟨Nu,xv⟩.\langle N_v, \mathbf{x}_u \rangle = -\langle N, \mathbf{x}_{uv} \rangle = \langle N_u, \mathbf{x}_v \rangle.

That is,

⟨Nu,xv⟩=⟨xu,Nv⟩,\langle N_u, \mathbf{x}_v \rangle = \langle \mathbf{x}_u, N_v \rangle,

which is exactly the cross-term identity required in Step 2. This closes the verification that dNpdN_p is self-adjoint on the basis {xu,xv}\{\mathbf{x}_u,\mathbf{x}_v\}, and hence (by the basis-extension remark in the setup) self-adjoint on all of TpST_pS.

Step 4: Spectral consequences. A self-adjoint linear operator AA on a finite-dimensional real inner product space is diagonalizable with real eigenvalues, and eigenvectors for distinct eigenvalues are orthogonal — this is the real spectral theorem, a standard fact of linear algebra (proved, e.g., by noting the associated bilinear form B(w1,w2)=⟨Aw1,w2⟩B(w_1,w_2) = \langle A w_1, w_2\rangle is symmetric, hence represented in an orthonormal basis by a symmetric matrix, which is orthogonally diagonalizable by the standard spectral theorem for symmetric matrices). Applying this to A=dNpA = dN_p on the 2-dimensional space TpST_pS yields exactly the stated conclusion: two real eigenvalues k1≤k2k_1 \le k_2, with orthogonal eigenvectors if k1<k2k_1 < k_2, and dNp=k1 IddN_p = k_1 \,\mathrm{Id} if k1=k2k_1 = k_2. ■\blacksquare

∎
CorollaryCoordinate formula for dNpdN_p

In the parametrization x\mathbf{x} above, write dNp(xu)=a11xu+a21xvdN_p(\mathbf{x}_u) = a_{11}\mathbf{x}_u + a_{21}\mathbf{x}_v and dNp(xv)=a12xu+a22xvdN_p(\mathbf{x}_v) = a_{12}\mathbf{x}_u + a_{22}\mathbf{x}_v for scalars aija_{ij} (possible since {xu,xv}\{\mathbf{x}_u,\mathbf{x}_v\} is a basis of TpST_pS and dNpdN_p maps TpST_pS to itself). Then the matrix (aij)(a_{ij}), though generally not symmetric as a literal matrix (because the basis {xu,xv}\{\mathbf{x}_u,\mathbf{x}_v\} need not be orthonormal), represents a self-adjoint operator; equivalently (aij)=−I−1II(a_{ij}) = -I^{-1}II where II is the matrix of the first fundamental form and IIII the matrix of the second fundamental form — a relation made precise once the second fundamental form is introduced.

Worked examples

Example 1: The Gauss map of a circular cylinder. Let SS be the circular cylinder of radius 22,

x(u,v)=(2cos⁡u, 2sin⁡u, v),(u,v)∈(0,2π)×R.\mathbf{x}(u,v) = (2\cos u,\ 2\sin u,\ v), \qquad (u,v) \in (0, 2\pi) \times \mathbb{R}.

Compute the coordinate vectors:

xu=(−2sin⁡u, 2cos⁡u, 0),xv=(0,0,1).\mathbf{x}_u = (-2\sin u,\ 2\cos u,\ 0), \qquad \mathbf{x}_v = (0,0,1).

These are orthogonal (⟨xu,xv⟩=0\langle \mathbf{x}_u,\mathbf{x}_v\rangle = 0), with ∥xu∥=2\lVert \mathbf{x}_u \rVert = 2 and ∥xv∥=1\lVert \mathbf{x}_v\rVert = 1. The cross product is

xu×xv=(2cos⁡u, 2sin⁡u, 0),∥xu×xv∥=2,\mathbf{x}_u \times \mathbf{x}_v = (2\cos u,\ 2\sin u,\ 0), \qquad \lVert \mathbf{x}_u \times \mathbf{x}_v\rVert = 2,

so the unit normal is

N(u,v)=(cos⁡u, sin⁡u, 0).N(u,v) = (\cos u,\ \sin u,\ 0).

Notice NN does not depend on vv at all — geometrically, sliding along the axis of the cylinder does not tilt the normal. As a map to S2S^2, the image of the Gauss map is exactly the equator {(cos⁡u,sin⁡u,0)}⊂S2\{(\cos u,\sin u,0)\} \subset S^2: the cylinder's Gauss map is degenerate, collapsing all of R\mathbb{R}-worth of the vv-direction onto a single circle. Now differentiate:

Nu=(−sin⁡u, cos⁡u, 0),Nv=(0,0,0).N_u = (-\sin u,\ \cos u,\ 0), \qquad N_v = (0,0,0).

So in the basis {xu,xv}\{\mathbf{x}_u,\mathbf{x}_v\}:

dNp(xv)=Nv=0=0⋅xv,dN_p(\mathbf{x}_v) = N_v = 0 = 0\cdot\mathbf{x}_v,

and since Nu=12xuN_u = \tfrac12 \mathbf{x}_u (check: 12(−2sin⁡u,2cos⁡u,0)=(−sin⁡u,cos⁡u,0)\tfrac12(-2\sin u, 2\cos u, 0) = (-\sin u,\cos u,0), matching NuN_u exactly),

dNp(xu)=12xu.dN_p(\mathbf{x}_u) = \tfrac12 \mathbf{x}_u.

Thus {xu,xv}\{\mathbf{x}_u, \mathbf{x}_v\} is already an eigenbasis (orthogonal, as required by the theorem, and indeed here orthogonal as raw coordinate vectors too), with eigenvalues

k1=0(eigenvector xv, the axial direction),k2=12(eigenvector xu, the circular direction).k_1 = 0 \quad(\text{eigenvector } \mathbf{x}_v, \text{ the axial direction}), \qquad k_2 = \tfrac12 \quad(\text{eigenvector } \mathbf{x}_u, \text{ the circular direction}).

This matches intuition: moving along the axis is "straight" (curvature 00), while moving around the circular cross-section of radius 22 bends with curvature 1/21/2 (reciprocal of the radius) — exactly the principal curvatures of a cylinder.

Example 2: The Gauss map of an elliptic paraboloid at an off-center point. Let SS be the graph z=x2+2y2z = x^2 + 2y^2, parametrized by

x(u,v)=(u, v, u2+2v2).\mathbf{x}(u,v) = (u,\ v,\ u^2 + 2v^2).

We evaluate everything at the asymmetric point (u0,v0)=(1,1)(u_0,v_0) = (1,1), i.e. the surface point p=(1,1,3)p = (1,1,3). Compute:

xu=(1,0,2u),xv=(0,1,4v).\mathbf{x}_u = (1,0,2u), \qquad \mathbf{x}_v = (0,1,4v).

At (1,1)(1,1): xu=(1,0,2)\mathbf{x}_u = (1,0,2), xv=(0,1,4)\mathbf{x}_v = (0,1,4). Cross product:

xu×xv=(0⋅4−2⋅1, 2⋅0−1⋅4, 1⋅1−0⋅0)=(−2, −4, 1).\mathbf{x}_u \times \mathbf{x}_v = (0\cdot4 - 2\cdot1,\ 2\cdot0-1\cdot4,\ 1\cdot1-0\cdot0) = (-2,\,-4,\,1).

Its norm is 4+16+1=21\sqrt{4+16+1} = \sqrt{21}, so the unit normal at pp is

N(p)=121(−2,−4,1).N(p) = \frac{1}{\sqrt{21}}(-2,-4,1).

For general (u,v)(u,v), xu×xv=(−2u,−4v,1)\mathbf{x}_u \times \mathbf{x}_v = (-2u,-4v,1) with norm W(u,v)=4u2+16v2+1W(u,v) = \sqrt{4u^2+16v^2+1}, so

N(u,v)=1W(u,v)(−2u,−4v,1).N(u,v) = \frac{1}{W(u,v)}(-2u,-4v,1).

Differentiate with the quotient rule. Let W=(4u2+16v2+1)1/2W = (4u^2+16v^2+1)^{1/2}, so Wu=4uWW_u = \dfrac{4u}{W}, Wv=16vWW_v = \dfrac{16v}{W}. Then

Nu=1W(−2,0,0)−WuW2(−2u,−4v,1),N_u = \frac{1}{W}(-2,0,0) - \frac{W_u}{W^2}(-2u,-4v,1),
Nv=1W(0,−4,0)−WvW2(−2u,−4v,1).N_v = \frac{1}{W}(0,-4,0) - \frac{W_v}{W^2}(-2u,-4v,1).

At (u,v)=(1,1)(u,v) = (1,1): W=21W = \sqrt{21}, Wu=4/21W_u = 4/\sqrt{21}, Wv=16/21W_v = 16/\sqrt{21}. So

Nu=121(−2,0,0)−4/2121(−2,−4,1)=121(−2,0,0)−42121(−2,−4,1).N_u = \frac{1}{\sqrt{21}}(-2,0,0) - \frac{4/\sqrt{21}}{21}(-2,-4,1) = \frac{1}{\sqrt{21}}(-2,0,0) - \frac{4}{21\sqrt{21}}(-2,-4,1).

Multiply through by 212121\sqrt{21} to clear denominators and read off components (we keep it symbolic to avoid rounding error): writing Nu=12121(21(−2,0,0)−4(−2,−4,1))N_u = \frac{1}{21\sqrt{21}}\big(21(-2,0,0) - 4(-2,-4,1)\big),

21(−2,0,0)−4(−2,−4,1)=(−42,0,0)−(−8,−16,4)=(−34,16,−4).21(-2,0,0) - 4(-2,-4,1) = (-42,0,0) - (-8,-16,4) = (-34,16,-4).

So

Nu=12121(−34, 16, −4).N_u = \frac{1}{21\sqrt{21}}(-34,\,16,\,-4).

Similarly Nv=1W(0,−4,0)−WvW2(−2u,−4v,1)N_v = \frac{1}{W}(0,-4,0) - \frac{W_v}{W^2}(-2u,-4v,1) with Wv=16/21W_v = 16/\sqrt{21}:

Nv=12121(21(0,−4,0)−16(−2,−4,1))=12121((0,−84,0)−(−32,−64,16))=12121(32, −20, −16).N_v = \frac{1}{21\sqrt{21}}\Big(21(0,-4,0) - 16(-2,-4,1)\Big) = \frac{1}{21\sqrt{21}}\big((0,-84,0) - (-32,-64,16)\big) = \frac{1}{21\sqrt{21}}(32,\,-20,\,-16).

Now express Nu,NvN_u, N_v in the basis {xu,xv}={(1,0,2),(0,1,4)}\{\mathbf{x}_u,\mathbf{x}_v\} = \{(1,0,2),(0,1,4)\}. We seek a11,a21a_{11},a_{21} with

Nu=a11xu+a21xv=(a11, a21, 2a11+4a21).N_u = a_{11}\mathbf{x}_u + a_{21}\mathbf{x}_v = (a_{11},\,a_{21},\,2a_{11}+4a_{21}).

Matching the first two components of Nu=12121(−34,16,−4)N_u = \frac{1}{21\sqrt{21}}(-34,16,-4) directly gives

a11=−342121,a21=162121.a_{11} = \frac{-34}{21\sqrt{21}}, \qquad a_{21} = \frac{16}{21\sqrt{21}}.

(As a consistency check, the third component must satisfy 2a11+4a21=−68+642121=−421212a_{11} + 4a_{21} = \frac{-68 + 64}{21\sqrt{21}} = \frac{-4}{21\sqrt{21}}, matching the given third component of NuN_u exactly — confirming the computation.) Similarly for Nv=a12xu+a22xvN_v = a_{12}\mathbf{x}_u + a_{22}\mathbf{x}_v with Nv=12121(32,−20,−16)N_v = \frac{1}{21\sqrt{21}}(32,-20,-16):

a12=322121,a22=−202121,a_{12} = \frac{32}{21\sqrt{21}}, \qquad a_{22} = \frac{-20}{21\sqrt{21}},

and the consistency check: 2a12+4a22=64−802121=−1621212a_{12}+4a_{22} = \frac{64 - 80}{21\sqrt{21}} = \frac{-16}{21\sqrt{21}}, matching the third component of NvN_v. Good. So in this (non-orthonormal) basis, the matrix of dNpdN_p is

(aij)=12121(−343216−20).(a_{ij}) = \frac{1}{21\sqrt{21}}\begin{pmatrix} -34 & 32 \\ 16 & -20 \end{pmatrix}.

This matrix is not symmetric as written — exactly as the corollary warned, since xu,xv\mathbf{x}_u,\mathbf{x}_v are not orthonormal (⟨xu,xv⟩=8≠0\langle \mathbf{x}_u,\mathbf{x}_v\rangle = 8 \ne 0 here) — yet dNpdN_p is still self-adjoint with respect to the inner product of TpST_pS, and its eigenvalues are basis-independent real numbers (the principal curvatures at pp), computable from the trace and determinant of this matrix without ever orthonormalizing the basis:

det⁡(aij)=(−34)(−20)−(32)(16)(2121)2=680−512441⋅21=1689261=8441,\det(a_{ij}) = \frac{(-34)(-20) - (32)(16)}{(21\sqrt{21})^2} = \frac{680 - 512}{441\cdot 21} = \frac{168}{9261} = \frac{8}{441},

which will reappear (once the second fundamental form is defined) as the Gaussian curvature K=k1k2K = k_1 k_2 of the paraboloid at p=(1,1,3)p=(1,1,3).

Counterexample

Counterexampledropping orientability breaks the definition of NN, not merely its niceness

The theorem requires SS to admit a globally consistent smooth unit normal field — orientability. Without it, the Gauss map cannot even be well-defined, let alone self-adjoint. The standard example is the Möbius band. Take the strip (−1,1)×[0,2π](-1,1)\times[0, 2\pi] and glue ( t, 0 )(\,t,\,0\,) to (−t, 2π)(-t,\,2\pi) to form a Möbius band MM, realized as a regular surface in R3\mathbb{R}^3 by the parametrization

x(t,θ)=((2+tcos⁡θ2)cos⁡θ, (2+tcos⁡θ2)sin⁡θ, tsin⁡θ2),t∈(−1,1), θ∈[0,2π].\mathbf{x}(t,\theta) = \Big(\big(2+t\cos\tfrac{\theta}{2}\big)\cos\theta,\ \big(2+t\cos\tfrac\theta2\big)\sin\theta,\ t\sin\tfrac\theta2\Big), \quad t\in(-1,1),\ \theta\in[0,2\pi].

Locally, near any point, one can compute xt×xθ\mathbf{x}_t \times \mathbf{x}_\theta and normalize it to get a smooth local unit normal NN, exactly as in a regular parametrization. This local NN is perfectly smooth on θ∈(0,2π)\theta \in (0, 2\pi). The problem appears when we try to extend this local choice all the way around the band and back to the start. Tracking the normal continuously as θ\theta runs from 00 to 2π2\pi, one finds

N(t,θ=2π−)=−N(−t,θ=0+),N(t,\theta=2\pi^-) = -N(-t, \theta = 0^+),

i.e. after a full traversal the normal has flipped to point the opposite way relative to the identified point. (This is the defining topological feature of the Möbius band: parallel transport of a normal frame around the core circle reverses orientation.) Consequently there is no way to choose NN consistently and smoothly on all of MM: any attempted global unit normal field must be discontinuous somewhere, because continuously following it around the band forces N→−NN \to -N upon return. Hence the Möbius band, though a perfectly good regular surface satisfying every hypothesis of surface regularity, admits no Gauss map N:M→S2N: M \to S^2 at all — the very object the theorem is about fails to exist. This is exactly why orientability is listed as a standing hypothesis rather than a convenience: remove it, and there is nothing left to differentiate, let alone show is self-adjoint. (One can still speak of the shape operator locally, on an orientable coordinate patch of MM minus a slit, but it does not glue to a global operator on MM.)

Scope, edge cases & common traps

SituationWhat happens
Surface is compact and orientable (e.g. sphere, torus, ellipsoid)Gauss map is globally defined; its degree relates to Euler characteristic (Gauss–Bonnet)
Surface is non-orientable (Möbius band, Klein bottle immersed in R3\mathbb{R}^3)No global NN exists; only locally-defined Gauss maps on orientable subpatches
Surface is a planeNN is constant; dNp≡0dN_p \equiv 0; both principal curvatures are 00
Ruled surface, e.g. a cylinderdNpdN_p has a zero eigenvalue along the ruling direction (Example 1)
Non-orthonormal parametrization basisThe coordinate matrix (aij)(a_{ij}) of dNpdN_p need not look symmetric; self-adjointness is with respect to the first fundamental form, not the raw dot product of coefficients
Sign conventionSome texts define the shape operator as Sp=−dNpS_p = -dN_p so that principal curvatures of a sphere with outward normal come out positive; always check the author's convention

The mathematician behind it

Origins · 1 minCarl Friedrich Gauss and how this was discovered
CF
Carl Friedrich Gauss
1777–1855 · Göttingen, Kingdom of Hanover

Gauss was a mathematician, astronomer, and physicist whose work reshaped number theory, statistics, geodesy, and geometry. Commissioned by the Hanoverian government to conduct a geodetic survey of the kingdom in the 1820s, he became deeply engaged with the practical problem of measuring curvature on the literal surface of the Earth using only surface measurements — no access to the third dimension. This drove him to the map now bearing his name, and to the astonishing discovery that the product of principal curvatures, K=k1k2K = k_1 k_2, could be computed purely from lengths and angles within the surface, without reference to how the surface sits in space. He announced this in Disquisitiones generales circa superficies curvas (1827), calling it his Theorema Egregium — the "remarkable theorem."

History
- 1777: Gauss born in Braunschweig.
- 1818–1825: Gauss directs the geodetic survey of the Kingdom of Hanover, motivating detailed study of curved surfaces.
- 1827: Publication of Disquisitiones generales circa superficies curvas, introducing the Gauss map and the Theorema Egregium.
- 1854: Riemann's habilitation lecture extends curvature to abstract manifolds without any ambient space, building on Gauss's intrinsic viewpoint.
- 1855: Gauss dies in Göttingen.
Is the Gauss map always injective?

No. Injectivity would require distinct points of SS to always have distinct unit normals, which fails constantly. On a sphere, the Gauss map (suitably normalized) is a bijection, but on a cylinder every point along a given ruling line shares the same normal direction, so the Gauss map collapses an entire line to a single point of S2S^2 — hugely non-injective. On a torus, opposite "sides" of the tube can share parallel normals too. Injectivity of the Gauss map is a special, geometrically strong condition (related to convexity), not a general feature; what is always true, under the stated hypotheses, is that its differential is a well-defined self-adjoint linear operator at each point, regardless of global injectivity.

Does the Gauss map depend on the choice of unit normal, NN versus −N-N?

Yes, and this dependence is precisely tracked by the surface's orientation. Flipping N→−NN \to -N sends the Gauss map to the antipodal composition, and it negates the shape operator: d(−N)p=−dNpd(-N)_p = -dN_p. This flips the sign of both principal curvatures and hence the sign of the mean curvature, but leaves the Gaussian curvature K=k1k2K = k_1k_2 unchanged, since it is a product of two negated numbers. This is why Gaussian curvature is often regarded as more geometrically fundamental than mean curvature — it doesn't care which way you point the normal, while mean curvature (and the shape operator itself) does.

Why is the codomain of the Gauss map the unit sphere rather than some other sphere?

Because N(p)N(p) is constructed to have length exactly 1 by definition — it is the unit normal vector. This normalization is a deliberate choice: it makes the codomain a fixed, universal object (independent of the particular surface SS) so that Gauss maps of different surfaces can be compared on the same target space S2S^2. It also ensures TN(p)S2T_{N(p)}S^2 is literally the orthogonal complement of N(p)N(p), matching TpST_pS exactly and making the identification TpS=TN(p)S2T_pS = T_{N(p)}S^2 possible without any extra rescaling.

What replaces the Gauss map when the ambient space is not R3\mathbb{R}^3?

For a hypersurface in a general Riemannian manifold, there is no fixed external "sphere" to map into in the same literal sense, so the Gauss map is generalized to the shape operator (or Weingarten map) directly: a self-adjoint operator on each tangent space defined via the ambient connection's derivative of the unit normal field, without reference to any target sphere. All the linear-algebraic content — eigenvalues as principal curvatures, determinant as Gaussian-type curvature — survives verbatim; what is lost is the literal geometric picture of "a map to a sphere," which is special to hypersurfaces of flat Euclidean space.

Is self-adjointness of dNpdN_p special to surfaces in R3\mathbb{R}^3, or does it hold more generally?

The proof given here uses only that NN is a unit normal field to a hypersurface in a flat ambient space with a symmetric (torsion-free) derivative, via the symmetry of mixed partial derivatives xuv=xvu\mathbf{x}_{uv} = \mathbf{x}_{vu}. The same argument goes through verbatim for hypersurfaces in any flat space Rn\mathbb{R}^n, and with the ambient derivative replaced by a torsion-free covariant derivative, it extends to hypersurfaces in any Riemannian manifold. Self-adjointness of the shape operator is thus not a quirk of three-dimensional Euclidean space; it is a structural consequence of working with torsion-free connections, which is the standard setting throughout Riemannian geometry.

How does the Gauss map relate to the second fundamental form?

The second fundamental form is essentially a repackaging of dNpdN_p as a bilinear form rather than a linear operator: one defines IIp(w1,w2)=⟨dNp(w1),w2⟩\mathrm{II}_p(w_1,w_2) = \langle dN_p(w_1), w_2\rangle (up to a sign convention). Self-adjointness of dNpdN_p, proved above, is exactly what guarantees this bilinear form is symmetric, i.e. that it deserves to be called a "form" at all in the usual sense of quadratic forms. The Gauss map is thus the geometric object; the second fundamental form is its algebraic shadow as a symmetric bilinear pairing on TpST_pS, and the two carry identical information.

Can the Gauss map have a differential that is not diagonalizable?

No — this is precisely the force of the self-adjointness theorem. Over the real numbers, an operator that is merely linear can fail to be diagonalizable (a shear map, for instance, has repeated eigenvalues but only a one-dimensional eigenspace). Self-adjoint operators are never like this: the real spectral theorem guarantees a full orthogonal eigenbasis always exists, with no exceptions, once self-adjointness relative to the given inner product is established. So dNpdN_p is always diagonalizable, in every case satisfying the theorem's hypotheses — this is not an additional assumption but a guaranteed consequence.

Before this

Where this leads