Fundamental Theorem of Space Curves

Suppose two ants trace out paths in space. At every instant they experience the exact same bending and twisting. That means the same curvature and the same torsion. But they start at different points facing different directions. Must their paths, as pure shapes, be identical? This is the central question answered by the Fundamental Theorem of Space Curves. The answer is a resounding yes. Curvature and torsion, read off as functions of arc length, completely determine a space curve. This holds up to a rigid motion of Euclidean space. The result matters because it converts a geometric object into two scalar functions. That object is a curve — an infinite set of points with a shape. The functions are of a single real variable. This is the same philosophy that pervades all of differential geometry. Replace an intractable global object with local invariants. These include curvature, torsion, and later curvature tensors and second fundamental forms. They are computable and secretly carry all the shape information. The theorem is the prototype for later, harder existence-and-uniqueness statements. One example is the theorem that a surface is determined by its first and second fundamental forms. This holds subject to the Gauss and Codazzi–Mainardi equations. The big picture is an interplay between two classical branches of mathematics. These are the theory of ordinary differential equations, and Euclidean geometry. The Frenet-Serret Frame gives a linear system of first-order ODEs. These are the Frenet–Serret equations, satisfied by the moving frame (T,N,B)(T,N,B) attached to a curve. The existence half of the theorem is exactly the Picard–Lindelöf existence-and-uniqueness theorem. This is for linear ODE systems, transplanted into geometry. The uniqueness half is a rigidity statement. Any two solutions with the same curvature and torsion functions differ only by a constant rotation and translation. That is, they differ by an element of the group SE(3)SE(3) of orientation-preserving rigid motions. This article states the Frenet–Serret equations. It proves existence and uniqueness of the frame given arbitrary curvature and torsion functions. It also proves the rigidity theorem in full. Along the way we build the intuition for why curvature and torsion are the complete invariants of a curve's shape. We work two asymmetric examples: a curve with non-constant curvature and torsion, and the classical helix. We show exactly what breaks if either hypothesis is dropped. Those hypotheses are positivity of curvature, or matching endpoint frames. This sets the stage for the analogous but harder rigidity theorem. That theorem concerns Regular Surfaces and Parametrizations.

Key takeaways
  • Two space curves with identical curvature κ(s)>0\kappa(s) > 0 and torsion τ(s)\tau(s) functions (same arc length parametrization) are related by a unique rigid motion of R3\mathbb{R}^3.
  • The Frenet–Serret equations form a linear ODE system for the moving frame (T,N,B)(T,N,B); existence and uniqueness of solutions is exactly Picard–Lindelöf.
  • Reconstructing the curve is a second integration: α(s)=α(0)+∫0sT(u) du\alpha(s) = \alpha(0) + \int_0^s T(u)\,du.
  • Rigid motions are orthogonal linear maps (rotations, det =+1=+1) composed with translations; they preserve arc length, curvature, and torsion.
  • The requirement κ(s)>0\kappa(s) > 0 everywhere is essential — it is what makes the principal normal NN, and hence τ\tau, well defined.
  • Key idea: curvature and torsion are a complete set of local invariants for space curves — knowing them as functions of arc length is exactly as much data as knowing the curve itself, up to rigid motion.

Where this lives today

The Frenet–Serret rigidity theorem is the prototype for a family of "moving frame" rigidity results across geometry: the analogous statement for surfaces (that the first and second fundamental forms determine a surface up to rigid motion, subject to the Gauss–Codazzi compatibility equations) is a direct generalization, central to Regular Surfaces and Parametrizations and classical surface theory. In computer-aided geometric design and computer graphics, curvature and torsion profiles are used to specify smooth space curves (e.g. camera paths, robot end-effector trajectories) with prescribed bending and twisting behavior, essentially inverting the existence half of the theorem numerically. In robotics and motion planning, the Frenet frame gives the natural "road-following" coordinate frame for a vehicle or manipulator tracing a reference path, and torsion quantifies how much the frame must roll to stay aligned with the curve — relevant to trajectory-tracking control design. In molecular biology, the backbone of a DNA double helix or a protein alpha-helix is modeled as a space curve, and curvature/torsion profiles are used to classify and compare secondary structure shapes independent of position and orientation in the lab frame. In physics, the Frenet–Serret apparatus underlies the kinematics of a particle constrained to a wire or track, and the same linear-ODE-plus-rigidity strategy reappears in the theory of moving frames and Cartan's method of equivalence, which underlies modern approaches to curvature in General Relativity-adjacent geometric contexts and in the classification of homogeneous spaces.

Ask the tutor: Can you show me step-by-step how to numerically integrate the Frenet equations for a curve with κ(s) = 1 + sin(s) and τ(s) = cos(s), and plot the resulting shape?

DefinitionRigid motion, Frenet frame, and the data of a curve

A rigid motion of R3\mathbb{R}^3 is a map F:R3→R3F : \mathbb{R}^3 \to \mathbb{R}^3 of the form

F(x)=Ax+b,A∈SO(3), b∈R3,F(x) = Ax + b, \qquad A \in SO(3),\ b \in \mathbb{R}^3,

where SO(3)SO(3) is the group of 3×33\times 3 real orthogonal matrices with det⁡A=1\det A = 1. Rigid motions are exactly the orientation-preserving isometries of R3\mathbb{R}^3. Let α:[0,L]→R3\alpha : [0, L] \to \mathbb{R}^3 be a curve parametrized by arc length (so ∣α′(s)∣=1|\alpha'(s)| = 1 for all ss), of class C3C^3, and suppose the curvature

κ(s)=∣α′′(s)∣\kappa(s) = |\alpha''(s)|

is strictly positive for every s∈[0,L]s \in [0,L]. Define the Frenet frame (T,N,B)(T,N,B) by

T(s)=α′(s),N(s)=α′′(s)κ(s),B(s)=T(s)×N(s).T(s) = \alpha'(s), \qquad N(s) = \frac{\alpha''(s)}{\kappa(s)}, \qquad B(s) = T(s) \times N(s).

This is an orthonormal, positively oriented frame at each ss. The torsion τ(s)\tau(s) is the scalar function defined by

B′(s)=−τ(s) N(s).B'(s) = -\tau(s)\, N(s).

Together (T,N,B)(T, N, B) satisfy the Frenet–Serret equations:

T′=κN,N′=−κT+τB,B′=−τN.T' = \kappa N, \qquad N' = -\kappa T + \tau B, \qquad B' = -\tau N.

The functions κ:[0,L]→(0,∞)\kappa : [0,L] \to (0,\infty) and τ:[0,L]→R\tau : [0,L] \to \mathbb{R} are called the curvature and torsion of α\alpha, and the pair (κ,τ)(\kappa,\tau) is called the natural equations or intrinsic equations of the curve.

Intuition

Think of driving a car along the curve at unit speed, arc length playing the role of time. At each moment you have a steering wheel (curvature κ\kappa, how sharply you turn left or right within your instantaneous plane of motion) and a banking control (torsion τ\tau, how much your plane of turning itself twists out of a plane, like a road with camber). If someone hands you the full log of steering-wheel angle and banking angle for every instant of the trip, you know exactly the shape of the road you drove — you just don't know where on the globe you started or which compass direction you were originally facing. That "starting position and orientation" ambiguity is precisely the ambiguity of a rigid motion: a translation (where you started) composed with a rotation (which way you were facing). This is why the theorem has two halves. The existence half says: given any reasonable functions κ(s)>0\kappa(s) > 0 and τ(s)\tau(s), some curve realizes them — you can always build a road matching any steering/banking log. The uniqueness half says the road built this way is the only one (up to how you place it in space) — no two genuinely different-shaped curves can produce the same log. The proof of existence is an ODE argument: the Frenet equations are a linear system for (T,N,B)(T,N,B), solvable by Picard–Lindelöf, and then α\alpha is recovered by integrating TT. The proof of uniqueness is a rigidity argument: if two curves share (κ,τ)(\kappa,\tau), first align their frames at one point using a rotation, then show the aligned frames coincide for all ss by a differential inequality (Grönwall-type) argument, and finally integrate to see the curves themselves coincide up to translation.

Figure
The Frenet frame data: curvature bends within the osculating plane, torsion twists it out
α(s) T(s) N(s) B(s) κ(s): bending rate of T within the T–N plane τ(s): rate the plane twists out, tracked by B
TheoremFundamental Theorem of Space Curves

Existence. Let κ:[0,L]→(0,∞)\kappa : [0,L] \to (0,\infty) and τ:[0,L]→R\tau : [0,L] \to \mathbb{R} be continuous functions. Then there exists a C3C^3 curve α:[0,L]→R3\alpha : [0,L] \to \mathbb{R}^3, parametrized by arc length, whose curvature and torsion functions are exactly κ\kappa and τ\tau.

Uniqueness. Let α,α~:[0,L]→R3\alpha, \tilde\alpha : [0,L] \to \mathbb{R}^3 both be C3C^3, parametrized by arc length, with curvature functions κ,κ~\kappa, \tilde\kappa strictly positive on [0,L][0,L] and torsion functions τ,τ~\tau, \tilde\tau. If

κ(s)=κ~(s)andτ(s)=τ~(s)for all s∈[0,L],\kappa(s) = \tilde\kappa(s) \quad \text{and} \quad \tau(s) = \tilde\tau(s) \qquad \text{for all } s \in [0,L],

then there exists a unique rigid motion F(x)=Ax+bF(x) = Ax + b, A∈SO(3)A \in SO(3), b∈R3b\in\mathbb{R}^3, such that

α~(s)=F(α(s))=A α(s)+bfor all s∈[0,L].\tilde\alpha(s) = F(\alpha(s)) = A\,\alpha(s) + b \qquad \text{for all } s \in [0,L].

Conversely, if α~=F∘α\tilde\alpha = F \circ \alpha for a rigid motion FF, then α~\tilde\alpha has the same curvature and torsion as α\alpha.

Proof

We prove the four claims in order: the converse (rigid motions preserve κ,τ\kappa,\tau), existence, uniqueness, and we identify the proof techniques used at each stage.

Step 0 (Converse: rigid motions preserve curvature and torsion). Let F(x)=Ax+bF(x) = Ax+b with A∈SO(3)A \in SO(3), and set α~=F∘α\tilde\alpha = F\circ\alpha, i.e. α~(s)=Aα(s)+b\tilde\alpha(s) = A\alpha(s) + b. Since AA is linear, α~′(s)=Aα′(s)\tilde\alpha'(s) = A\alpha'(s), and since AA is orthogonal, ∣α~′(s)∣=∣α′(s)∣=1|\tilde\alpha'(s)| = |\alpha'(s)| = 1; so α~\tilde\alpha is still arc-length parametrized. Then α~′′(s)=Aα′′(s)\tilde\alpha''(s) = A\alpha''(s), so

κ~(s)=∣α~′′(s)∣=∣Aα′′(s)∣=∣α′′(s)∣=κ(s),\tilde\kappa(s) = |\tilde\alpha''(s)| = |A\alpha''(s)| = |\alpha''(s)| = \kappa(s),

using orthogonality of AA again. Consequently N~(s)=Aα′′(s)/κ(s)=AN(s)\tilde N(s) = A\alpha''(s)/\kappa(s) = AN(s), and T~(s)=AT(s)\tilde T(s) = A T(s). For the binormal, we use that A∈SO(3)A \in SO(3) satisfies A(u×v)=(Au)×(Av)A(u\times v) = (Au)\times(Av) for all u,v∈R3u,v \in \mathbb{R}^3 (a standard fact: orthogonal matrices with determinant +1+1 commute with the cross product, since det⁡A=1\det A = 1 is exactly the condition making AA preserve orientation and hence the cross-product structure). Hence

B~(s)=T~(s)×N~(s)=(AT(s))×(AN(s))=A(T(s)×N(s))=AB(s).\tilde B(s) = \tilde T(s) \times \tilde N(s) = (AT(s)) \times (AN(s)) = A(T(s)\times N(s)) = A B(s).

Differentiating, B~′(s)=AB′(s)=A(−τ(s)N(s))=−τ(s)AN(s)=−τ(s)N~(s)\tilde B'(s) = A B'(s) = A(-\tau(s) N(s)) = -\tau(s) AN(s) = -\tau(s)\tilde N(s), so by definition of torsion, τ~(s)=τ(s)\tilde\tau(s) = \tau(s). This proves the converse.

Step 1 (Existence, via Picard–Lindelöf for linear systems). Fix any point p0∈R3p_0 \in \mathbb{R}^3 and any positively oriented orthonormal frame (T0,N0,B0)(T_0, N_0, B_0) at p0p_0 (e.g. the standard basis e1,e2,e3e_1,e_2,e_3). Consider the linear, time-varying (here "time" is arc length ss) system of ODEs for a triple of vector-valued functions T(s),N(s),B(s)∈R3T(s), N(s), B(s) \in \mathbb{R}^3:

T′=κ(s)N,N′=−κ(s)T+τ(s)B,B′=−τ(s)N,T' = \kappa(s) N, \qquad N' = -\kappa(s) T + \tau(s) B, \qquad B' = -\tau(s) N,

with initial condition T(0)=T0,N(0)=N0,B(0)=B0T(0)=T_0, N(0)=N_0, B(0)=B_0. Writing the 99 scalar components as a vector v(s)∈R9v(s) \in \mathbb{R}^9, this system has the form v′(s)=M(s)v(s)v'(s) = M(s) v(s) where M(s)M(s) is a 9×99\times 9 matrix whose entries are continuous functions of κ(s),τ(s)\kappa(s),\tau(s) (hence continuous on [0,L][0,L] since κ,τ\kappa,\tau are). The map (s,v)↦M(s)v(s,v) \mapsto M(s)v is continuous in ss and globally Lipschitz in vv (linear maps are Lipschitz with constant sup⁡[0,L]∥M(s)∥\sup_{[0,L]}\|M(s)\|, finite by continuity on a compact interval). By the Picard–Lindelöf theorem (existence-uniqueness for Lipschitz ODEs) applied on each small subinterval and continued by the standard continuation argument (no blow-up occurs because the system is linear, so solutions exist on the whole interval [0,L][0,L]), there is a unique C1C^1 solution T,N,B:[0,L]→R3T,N,B : [0,L]\to\mathbb{R}^3. We must check (T,N,B)(T,N,B) remains an orthonormal, positively oriented frame for all ss, not merely at s=0s=0. Consider the 66 scalar functions

f1=T⋅T−1, f2=N⋅N−1, f3=B⋅B−1, f4=T⋅N, f5=T⋅B, f6=N⋅B.f_1 = T\cdot T - 1,\ f_2 = N\cdot N - 1,\ f_3 = B\cdot B - 1,\ f_4 = T\cdot N,\ f_5 = T\cdot B,\ f_6 = N\cdot B.

Differentiate using the ODEs:

f1′=2T⋅T′=2κ T⋅N=2κf4,f_1' = 2T\cdot T' = 2\kappa\, T\cdot N = 2\kappa f_4,
f4′=T′⋅N+T⋅N′=κN⋅N+T⋅(−κT+τB)=κ(f2+1)−κ(f1+1)+τf5=κf2−κf1+τf5,f_4' = T'\cdot N + T\cdot N' = \kappa N\cdot N + T\cdot(-\kappa T + \tau B) = \kappa(f_2+1) - \kappa(f_1+1) + \tau f_5 = \kappa f_2 - \kappa f_1 + \tau f_5,

and similarly for the others; one finds each fi′f_i' is a linear combination of f1,…,f6f_1,\dots,f_6 with coefficients depending continuously on κ,τ\kappa,\tau (this is a direct but slightly lengthy computation; the key structural point is that the Frenet matrix is skew-symmetric with respect to the standard inner product, which forces dds(v⋅w)\frac{d}{ds}(v\cdot w) to be expressible purely in terms of inner products among T,N,BT,N,B). Thus (f1,…,f6)(f_1,\dots,f_6) solves a linear homogeneous ODE system in R6\mathbb{R}^6. Since fi(0)=0f_i(0) = 0 for all ii (because (T0,N0,B0)(T_0,N_0,B_0) is orthonormal), and the zero function solves this linear homogeneous system with the same initial condition, uniqueness of solutions to linear ODEs (again Picard–Lindelöf) forces fi(s)≡0f_i(s) \equiv 0 for all s∈[0,L]s \in [0,L]. Hence (T(s),N(s),B(s))(T(s),N(s),B(s)) is orthonormal for every ss. Positivity of orientation (det⁡[T N B]=1\det[T\,N\,B] = 1, not −1-1) holds at s=0s=0 and varies continuously, and since det⁡∈{−1,+1}\det \in \{-1,+1\} always for an orthonormal triple, continuity forces det⁡≡1\det \equiv 1 throughout. Now define

α(s)=p0+∫0sT(u) du.\alpha(s) = p_0 + \int_0^s T(u)\, du.

Then α\alpha is C2C^2 (since TT is C1C^1) with α′(s)=T(s)\alpha'(s) = T(s), so ∣α′(s)∣=1|\alpha'(s)|=1: α\alpha is arc-length parametrized. Also α′′(s)=T′(s)=κ(s)N(s)\alpha''(s) = T'(s) = \kappa(s) N(s), so ∣α′′(s)∣=κ(s)∣N(s)∣=κ(s)|\alpha''(s)| = \kappa(s)|N(s)| = \kappa(s) (as κ(s)>0\kappa(s)>0 and NN is a unit vector), confirming the curvature of α\alpha is exactly κ\kappa, and the principal normal of α\alpha is exactly NN. The binormal of α\alpha is T×N=BT \times N = B by construction, and B′=−τNB' = -\tau N by the defining ODE, so the torsion of α\alpha is exactly τ\tau. Since κ,τ\kappa,\tau are continuous, T′=κNT' = \kappa N is continuous, so α′′\alpha'' is C1C^1, i.e. α\alpha is C3C^3. This proves existence.

Step 2 (Uniqueness, via rigid alignment + Grönwall-type rigidity). Let α,α~\alpha,\tilde\alpha be as in the statement, sharing κ=κ~>0\kappa=\tilde\kappa>0 and τ=τ~\tau=\tilde\tau, with Frenet frames (T,N,B)(T,N,B) and (T~,N~,B~)(\tilde T,\tilde N,\tilde B) respectively. Aligning at a point. Since (T(0),N(0),B(0))(T(0),N(0),B(0)) and (T~(0),N~(0),B~(0))(\tilde T(0),\tilde N(0),\tilde B(0)) are both positively oriented orthonormal bases of R3\mathbb{R}^3, there is a unique A∈SO(3)A \in SO(3) with

AT(0)=T~(0),AN(0)=N~(0),AB(0)=B~(0).A T(0) = \tilde T(0), \qquad A N(0) = \tilde N(0), \qquad A B(0) = \tilde B(0).

(Concretely, if P=[T(0) N(0) B(0)]P = [T(0)\,N(0)\,B(0)] and P~=[T~(0) N~(0) B~(0)]\tilde P = [\tilde T(0)\,\tilde N(0)\,\tilde B(0)] are the orthogonal matrices with these columns, then A=P~P−1=P~PTA = \tilde P P^{-1} = \tilde P P^{T} is orthogonal with det⁡A=det⁡P~det⁡PT=1⋅1=1\det A = \det\tilde P \det P^{T} = 1\cdot 1 = 1.) Comparing frames for all ss. Define U(s)=AT(s)U(s) = AT(s), V(s)=AN(s)V(s) = AN(s), W(s)=AB(s)W(s) = AB(s). Since AA is a constant linear map, differentiating the Frenet equations for (T,N,B)(T,N,B) and applying AA shows (U,V,W)(U,V,W) solves the same Frenet ODE system

U′=κV,V′=−κU+τW,W′=−τVU' = \kappa V, \quad V' = -\kappa U + \tau W, \quad W' = -\tau V

(using τ=τ~,κ=κ~\tau=\tilde\tau, \kappa=\tilde\kappa), with initial condition U(0)=AT(0)=T~(0)U(0) = AT(0) = \tilde T(0), and likewise V(0)=N~(0)V(0) = \tilde N(0), W(0)=B~(0)W(0)=\tilde B(0). But (T~,N~,B~)(\tilde T,\tilde N,\tilde B) also solves this exact system with the exact same initial condition. By the uniqueness half of Picard–Lindelöf (the system is linear, hence globally Lipschitz, so solutions to the IVP are unique on all of [0,L][0,L]), we conclude

U(s)=T~(s),V(s)=N~(s),W(s)=B~(s)for all s∈[0,L].U(s) = \tilde T(s), \quad V(s) = \tilde N(s), \quad W(s) = \tilde B(s) \qquad \text{for all } s \in [0,L].

In particular T~(s)=A T(s)\tilde T(s) = A\,T(s) for every ss. Integrating to compare the curves. Set b=α~(0)−A α(0)b = \tilde\alpha(0) - A\,\alpha(0). Define α^(s)=A α(s)+b\hat\alpha(s) = A\,\alpha(s) + b. Then α^(0)=Aα(0)+b=α~(0)\hat\alpha(0) = A\alpha(0) + b = \tilde\alpha(0), and

α^′(s)=A α′(s)=A T(s)=T~(s)=α~′(s)for all s.\hat\alpha'(s) = A\,\alpha'(s) = A\,T(s) = \tilde T(s) = \tilde\alpha'(s) \qquad \text{for all } s.

So α^\hat\alpha and α~\tilde\alpha have the same derivative everywhere on [0,L][0,L] and agree at s=0s=0; by the fundamental theorem of calculus applied to each component of α^−α~\hat\alpha - \tilde\alpha, the difference α^(s)−α~(s)\hat\alpha(s) - \tilde\alpha(s) is constant, and equals 00 at s=0s=0, hence is 00 for all ss. Thus

α~(s)=α^(s)=A α(s)+bfor all s∈[0,L],\tilde\alpha(s) = \hat\alpha(s) = A\,\alpha(s) + b \qquad \text{for all } s \in [0,L],

which is exactly α~=F∘α\tilde\alpha = F\circ\alpha with F(x)=Ax+b∈SE(3)F(x) = Ax+b \in SE(3). Uniqueness of FF. Suppose F1(x)=A1x+b1F_1(x) = A_1 x + b_1 and F2(x)=A2x+b2F_2(x) = A_2x+b_2 both satisfy α~=Fi∘α\tilde\alpha = F_i \circ \alpha for i=1,2i=1,2. Evaluating at s=0s=0: A1α(0)+b1=A2α(0)+b2=α~(0)A_1\alpha(0)+b_1 = A_2\alpha(0)+b_2 = \tilde\alpha(0). Differentiating α~=Fi∘α\tilde\alpha = F_i\circ\alpha and α~′′=Fi∘α′′\tilde\alpha'' = F_i\circ\alpha'' appropriately (using linearity of AiA_i) at s=0s=0 gives A1T(0)=A2T(0)=T~(0)A_1T(0)=A_2T(0)=\tilde T(0) and A1N(0)=A2N(0)=N~(0)A_1N(0) = A_2N(0) = \tilde N(0), and since AiB(0)=Ai(T(0)×N(0))=AiT(0)×AiN(0)A_i B(0) = A_i(T(0)\times N(0)) = A_iT(0)\times A_i N(0) (as Ai∈SO(3)A_i \in SO(3)) we get A1B(0)=A2B(0)=B~(0)A_1B(0)=A_2B(0)=\tilde B(0) too. So A1A_1 and A2A_2 agree on the basis {T(0),N(0),B(0)}\{T(0),N(0),B(0)\} of R3\mathbb{R}^3, forcing A1=A2A_1=A_2; then b1=b2b_1=b_2 follows from the point equation above. This proves uniqueness of FF, completing the proof. ■\blacksquare

∎

Worked examples

Example 1: reconstructing a curve from a non-constant, asymmetric (κ,τ)(\kappa,\tau) pair. Suppose we are told a curve has constant curvature κ(s)=2\kappa(s) = 2 and linearly increasing torsion τ(s)=s\tau(s) = s, for s≥0s \geq 0; this is deliberately not a symmetric or constant-torsion case. We do not solve this fully in closed form (the Frenet system with τ(s)=s\tau(s)=s does not admit elementary closed-form solutions — this is itself instructive: the theorem guarantees existence even when no elementary formula exists), but we illustrate the mechanics precisely, including a consistency check. Initial frame: T(0)=(1,0,0)T(0) = (1,0,0), N(0)=(0,1,0)N(0) = (0,1,0), B(0)=(0,0,1)B(0)=(0,0,1), and base point α(0)=(0,0,0)\alpha(0) = (0,0,0). The Frenet system is

T′=2N,N′=−2T+sB,B′=−sN.T' = 2N, \qquad N' = -2T + sB, \qquad B' = -sN.

At s=0s=0: T′(0)=2N(0)=(0,2,0)T'(0) = 2N(0) = (0,2,0), N′(0)=−2T(0)+0⋅B(0)=(−2,0,0)N'(0) = -2T(0) + 0\cdot B(0) = (-2,0,0), B′(0)=−0⋅N(0)=(0,0,0)B'(0) = -0\cdot N(0) = (0,0,0). A second-order Taylor step (using T′′(0)=2N′(0)=(−4,0,0)T''(0) = 2N'(0) = (-4,0,0)) gives, near s=0s=0,

T(s)≈(1,0,0)+(0,2,0)s+(−4,0,0)s22+O(s3)=(1−2s2, 2s, 0)+O(s3).T(s) \approx (1,0,0) + (0,2,0)s + (-4,0,0)\tfrac{s^2}{2} + O(s^3) = (1-2s^2,\ 2s,\ 0) + O(s^3).

Check unit length to this order: ∣T(s)∣2≈(1−2s2)2+(2s)2=1−4s2+4s4+4s2=1+4s4=1+O(s4)|T(s)|^2 \approx (1-2s^2)^2 + (2s)^2 = 1 - 4s^2 + 4s^4 + 4s^2 = 1 + 4s^4 = 1+O(s^4), consistent with ∣T∣≡1|T|\equiv 1 to the order computed (the theorem's proof guarantees this holds exactly to all orders, not merely as a coincidence of the Taylor truncation — the discrepancy at O(s4)O(s^4) is an artifact of truncating the series, not a violation). Then α(s)=∫0sT(u) du≈(s−2s33, s2, 0)+O(s4)\alpha(s) = \int_0^s T(u)\,du \approx \left(s - \tfrac{2s^3}{3},\ s^2,\ 0\right) + O(s^4), a curve leaving the origin tangent to the xx-axis, curving up into the xyxy-plane initially (torsion effects appear only at higher order since τ(0)=0\tau(0)=0). This example shows concretely how the existence proof manufactures a curve, term by term if necessary, from prescribed (κ,τ)(\kappa,\tau) data — and that even "unsolvable in elementary functions" data still determines a curve, uniquely up to rigid motion, by the theorem, without needing an explicit formula.

Example 2: verifying the theorem on the circular helix, with mismatched radius and pitch. Let a=3, b=4a=3,\ b=4 (asymmetric, chosen so a2+b2=25a^2+b^2=25 is a clean number, but a≠ba\ne b), and set c=a2+b2=5c = \sqrt{a^2+b^2} = 5. Define

α(s)=(acos⁡sc, asin⁡sc, bsc)=(3cos⁡s5, 3sin⁡s5, 4s5).\alpha(s) = \left(a\cos\frac{s}{c},\ a\sin\frac{s}{c},\ \frac{bs}{c}\right) = \left(3\cos\frac{s}{5},\ 3\sin\frac{s}{5},\ \frac{4s}{5}\right).

Check arc length: α′(s)=(−acsin⁡sc, accos⁡sc, bc)\alpha'(s) = \left(-\frac{a}{c}\sin\frac{s}{c},\ \frac{a}{c}\cos\frac{s}{c},\ \frac{b}{c}\right), so

∣α′(s)∣2=a2c2sin⁡2sc+a2c2cos⁡2sc+b2c2=a2+b2c2=1,|\alpha'(s)|^2 = \frac{a^2}{c^2}\sin^2\frac{s}{c} + \frac{a^2}{c^2}\cos^2\frac{s}{c} + \frac{b^2}{c^2} = \frac{a^2+b^2}{c^2} = 1,

confirming unit speed. Then

T(s)=(−35sin⁡s5, 35cos⁡s5, 45),T′(s)=(−325cos⁡s5, −325sin⁡s5, 0).T(s) = \left(-\frac{3}{5}\sin\frac{s}{5},\ \frac{3}{5}\cos\frac{s}{5},\ \frac{4}{5}\right), \qquad T'(s) = \left(-\frac{3}{25}\cos\frac{s}{5},\ -\frac{3}{25}\sin\frac{s}{5},\ 0\right).

So

κ(s)=∣T′(s)∣=325=ac2=325,\kappa(s) = |T'(s)| = \frac{3}{25} = \frac{a}{c^2} = \frac{3}{25},

a constant, matching the general helix formula κ=a/c2\kappa = a/c^2. The principal normal is N(s)=T′(s)/κ(s)=(−cos⁡s5,−sin⁡s5,0)N(s) = T'(s)/\kappa(s) = \left(-\cos\frac{s}{5}, -\sin\frac{s}{5}, 0\right), and

B(s)=T(s)×N(s)=(45sin⁡s5, −45cos⁡s5, 35)B(s) = T(s)\times N(s) = \left(\frac{4}{5}\sin\frac{s}{5},\ -\frac{4}{5}\cos\frac{s}{5},\ \frac{3}{5}\right)

(direct computation of the cross product using the components above). Then

B′(s)=(425cos⁡s5, 425sin⁡s5, 0)=−425(−cos⁡s5,−sin⁡s5,0)=−425N(s),B'(s) = \left(\frac{4}{25}\cos\frac{s}{5},\ \frac{4}{25}\sin\frac{s}{5},\ 0\right) = -\frac{4}{25}\left(-\cos\frac{s}{5},-\sin\frac{s}{5},0\right) = -\frac{4}{25}N(s),

so torsion τ(s)=4/25=b/c2\tau(s) = 4/25 = b/c^2, also constant. So this specific helix has (κ,τ)=(3/25,4/25)(\kappa,\tau) = (3/25, 4/25), constants but unequal — an asymmetric, non-degenerate check. By the theorem, any other unit-speed curve with these exact constant values of κ\kappa and τ\tau everywhere must be congruent to this helix by a rigid motion: it can be a helix wound the "same way" starting anywhere, facing any direction, but it cannot be, say, a circle (that would require τ≡0\tau \equiv 0) or a helix of different pitch-to-radius ratio (different κ/τ\kappa/\tau ratio). This shows concretely how (κ,τ)(\kappa,\tau) as constant functions single out the entire congruence class of helices with a fixed shape parameter b/a=τ/κ⋅(appropriate scaling)b/a = \tau/\kappa \cdot (\text{appropriate scaling}), here 4/34/3.

CounterexampleWhy κ(s)>0\kappa(s) > 0 everywhere is needed

Drop the hypothesis κ(s)>0\kappa(s) > 0 for all ss, allowing κ\kappa to vanish at isolated points, and consider two curves that agree in curvature (both zero at one point) but genuinely differ in shape in a way not detected by "torsion," because torsion is undefined where κ=0\kappa = 0. Let

α(s)=(s, 16s3cos⁡θ1, 16s3sin⁡θ1),α~(s)=(s, 16s3cos⁡θ2, 16s3sin⁡θ2),\alpha(s) = \left(s, \ \tfrac{1}{6}s^3\cos\theta_1,\ \tfrac{1}{6}s^3\sin\theta_1\right), \qquad \tilde\alpha(s) = \left(s,\ \tfrac{1}{6}s^3\cos\theta_2,\ \tfrac16 s^3\sin\theta_2\right),

for ss near 00, where θ1≠θ2\theta_1 \neq \theta_2 are fixed angles (say θ1=0\theta_1 = 0, θ2=π/2\theta_2 = \pi/2), so that α\alpha bends off the xx-axis into the direction (cos⁡θ1,sin⁡θ1)(\cos\theta_1,\sin\theta_1) in the yzyz-plane as ss increases past 00, while α~\tilde\alpha bends off into the different direction (cos⁡θ2,sin⁡θ2)(\cos\theta_2,\sin\theta_2). Both curves have α′(0)=α~′(0)=(1,0,0)\alpha'(0) = \tilde\alpha'(0) = (1,0,0) and α′′(0)=α~′′(0)=(0,0,0)\alpha''(0) = \tilde\alpha''(0) = (0,0,0), so both have curvature κ(0)=0\kappa(0) = 0. For ss near 00 but s≠0s\neq 0, a direct computation gives α′′(s)=(0,scos⁡θ1,ssin⁡θ1)\alpha''(s) = (0, s\cos\theta_1, s\sin\theta_1), so κ(s)=∣s∣\kappa(s) = |s| for both curves regardless of θ1,θ2\theta_1,\theta_2 — the curvature functions s↦∣s∣s \mapsto |s| are identical for α\alpha and α~\tilde\alpha! Yet for s>0s>0, the principal normal of α\alpha points in direction (0,cos⁡θ1,sin⁡θ1)(0,\cos\theta_1,\sin\theta_1) while that of α~\tilde\alpha points in (0,cos⁡θ2,sin⁡θ2)(0,\cos\theta_2,\sin\theta_2) — for s<0s<0 each normal flips sign (since κ(s)=∣s∣\kappa(s) = |s| forces N(s)=α′′(s)/∣s∣N(s) = \alpha''(s)/|s| to reverse as ss crosses 00 from the scos⁡θs\cos\theta term changing sign), but crucially: as s→0+s\to 0^+ versus s→0−s \to 0^-, the plane spanned by N,BN,B can rotate discontinuously through the vanishing-curvature point, and no torsion function is continuously defined at s=0s=0 to record this rotation. Concretely, for θ1=0≠θ2=π/2\theta_1 = 0 \ne \theta_2 = \pi/2, the curves α\alpha and α~\tilde\alpha osculate the xx-axis to second order identically, agree in the value of κ(s)=∣s∣\kappa(s) = |s| for all ss, yet are not related by any rigid motion near s=0s=0: no rotation A∈SO(3)A \in SO(3) fixing the shared tangent direction (1,0,0)(1,0,0) and the shared vanishing second derivative can simultaneously rotate the direction (0,1,0)(0,1,0) (the "bend direction" of α\alpha just after 00) to (0,0,1)(0,0,1) (that of α~\tilde\alpha) while also fixing the pre-00 bend directions correctly, because the two curves genuinely twist into different half-planes independently on the two sides of s=0s=0. This shows the theorem's uniqueness conclusion can fail once curvature is allowed to vanish: the torsion function, and hence the entire rigidity argument (which required inverting N=α′′/κN = \alpha''/\kappa and used κ>0\kappa > 0 to keep (T,N,B)(T,N,B) a continuously-varying, well-defined orthonormal frame satisfying a bona fide linear ODE system throughout), simply loses meaning at points where κ=0\kappa = 0, and the Picard–Lindelöf uniqueness argument used in the theorem's proof no longer applies across such points.

Scope, edge cases & common traps

SituationWhat happens
κ(s)>0\kappa(s) > 0 on all of [0,L][0,L]Theorem applies exactly as stated; (T,N,B)(T,N,B) well-defined and smooth.
κ(s0)=0\kappa(s_0) = 0 at an isolated pointN,τN,\tau undefined at s0s_0; uniqueness can fail (see counterexample).
Curve not parametrized by arc lengthκ,τ\kappa,\tau must first be recomputed intrinsically (via reparametrization) — the theorem is stated for arc-length parametrizations, and applying it directly to a non-unit-speed parametrization is a common error.
Orientation-reversing map (AA with det⁡A=−1\det A = -1)Not a rigid motion in this theorem's sense; reflections flip the sign of torsion, so a mirror-image curve has torsion −τ-\tau, not τ\tau.
Curve only C2C^2, not C3C^3τ\tau may fail to be well-defined (needs B′B', hence needs α′′′\alpha''' to exist); the theorem's hypotheses require enough smoothness for τ\tau to make sense.
Domain not an interval (e.g. curve on a circle, periodic)Existence/uniqueness argument via ODEs still works locally: extra care is needed for global closure conditions (periodicity of κ,τ\kappa,\tau alone does not guarantee a closed curve).

The mathematician behind it

Origins · 1 minJean Frédéric Frenet and how this was discovered
JF
Jean Frédéric Frenet
1816–1900 · French mathematician, later mayor and academic administrator in Lyon

Frenet introduced the moving trihedron and the equations bearing his name in his 1847 doctoral thesis at the University of Toulouse, studying curves in space via a frame that automatically adapts to the curve's own bending and twisting. Joseph Alfred Serret independently derived the complete system a few years later in a more general and rigorous form, and the two names are now permanently attached to the frame and its governing equations. The fundamental theorem of space curves, showing that this local data is also complete, was assembled from this apparatus over the following decades as part of the broader 19th-century program (Gauss, Riemann, Darboux) of understanding geometric objects through differential invariants.

History
- 1847 — Jean Frédéric Frenet states the moving-frame equations for space curves in his doctoral thesis.
- 1851 — Joseph Alfred Serret publishes an independent, more complete derivation of the same frame equations.
- 1874 — Darboux and contemporaries extend the moving-frame idea to surfaces, prefiguring the surface analogue of this rigidity theorem.
- 1894 — Gaston Darboux's treatise systematizes curve and surface theory using moving frames, cementing the modern proof strategy via ODE existence-uniqueness.
- 1909 — Élie Cartan generalizes the moving-frame method into the broader theory of Cartan connections, of which the Frenet frame is now understood as the simplest example.
Do curvature and torsion have to be constant for the theorem to apply?

No. The theorem holds for any continuous functions κ(s)>0\kappa(s) > 0 and τ(s)\tau(s) on the interval, not just constants. Constant (κ,τ)(\kappa,\tau) happens to produce helices (with a circle as the degenerate case τ=0\tau=0), which is why those examples are so common in textbooks, but the existence and uniqueness proofs use only continuity of κ,τ\kappa,\tau and positivity of κ\kappa. In fact the Frenet ODE system can be solved (numerically, or via power series as in the first worked example) for essentially arbitrary continuous κ,τ\kappa,\tau, and the resulting curve is still unique up to rigid motion. Constant curvature and torsion are simply the case where the resulting ODE system happens to have a closed-form elementary solution.

Is arc-length parametrization essential, or can any parametrization be used?

Arc-length parametrization is essential to the statement as given, because κ\kappa and τ\tau as functions of a parameter depend on the parametrization unless that parameter is arc length. Curvature and torsion are geometric quantities meant to depend only on the shape of the curve, and this is only literally true when read off with respect to arc length; a curve traced twice as fast has different-looking κ,τ\kappa,\tau formulas as functions of the "wrong" parameter even though the geometric bending is the same. In practice, one reparametrizes any regular curve by arc length first (always possible for a regular CkC^k curve, k≥1k\geq 1) and then applies the theorem.

What exactly does "up to rigid motion" rule out?

It rules out reflections and any change of shape, but explicitly allows translating the curve anywhere in space and rotating it to face any direction. Formally, rigid motions are the group SE(3)=R3⋊SO(3)SE(3) = \mathbb{R}^3 \rtimes SO(3) of maps x↦Ax+bx \mapsto Ax+b with AA a rotation (det⁡A=+1\det A = +1). This excludes reflections (det =−1=-1), scalings, shears, and any non-isometric deformation. Two curves related by a rigid motion are called congruent; the theorem says matching (κ,τ)(\kappa,\tau) functions is exactly the condition for congruence.

How is this different from the analogous theorem for plane curves?

Plane curves need only a single function, the signed curvature κ(s)\kappa(s) (which can be positive or negative, unlike the always-nonnegative κ\kappa used for space curves), and are determined up to rigid motion in the plane by κ(s)\kappa(s) alone — torsion is not needed because a plane curve never leaves its plane. The space curve theorem needs two functions because a space curve has an extra degree of freedom: how its osculating plane twists as it moves, captured by τ\tau. In fact, a space curve has τ≡0\tau \equiv 0 everywhere if and only if it lies entirely in a single plane, which is a nice corollary connecting the two theorems.

Can the theorem fail if the curve is only piecewise smooth?

Yes, in the sense that κ\kappa and τ\tau may not be defined, or may be discontinuous, at the breakpoints, and the ODE argument (which relies on continuity of κ,τ\kappa,\tau to invoke Picard–Lindelöf) needs modification. One can still often apply the theorem separately on each smooth piece and then check the pieces glue together consistently (matching position and frame at the junctions), but the single global statement as proved here assumes α\alpha is C3C^3 throughout the interval.

What is the surface-theory analogue this leads to?

For Regular Surfaces and Parametrizations, the analogous statement is the fundamental theorem of surface theory (Bonnet's theorem): the first fundamental form (metric) and second fundamental form (shape/curvature data) together determine a surface up to a rigid motion, provided they satisfy the Gauss equation and the Codazzi–Mainardi equations, which are compatibility conditions with no counterpart in the curve case (curves have no compatibility constraint on κ,τ\kappa,\tau beyond continuity and κ>0\kappa>0). The proof strategy is the same two-step template: an ODE/PDE existence argument (now a system of PDEs rather than ODEs) followed by a rigidity argument using uniqueness of solutions with matching initial data.

Does the theorem say anything about closed curves, like whether a given (κ,τ) traces a loop?

Not directly — the theorem as stated is about curves on an interval [0,L][0,L] and says nothing about when the resulting curve closes up into a loop (α(0)=α(L)\alpha(0)=\alpha(L) with matching frame). Closure is an extra, generally hard, global condition: for a curve built from periodic κ,τ\kappa,\tau (period LL) to close up, one needs the holonomy of the Frenet frame around one period to be trivial, i.e. the frame must return exactly to its starting orientation, which is an additional constraint on κ,τ\kappa,\tau beyond mere periodicity of the functions themselves.

Before this

Where this leads