The Frenet-Serret Frame

Take a strand of wire bent freely in three-dimensional space. Ask how much information you need to describe its shape. This means ignoring where it sits in space. A curve in the plane has one shape-invariant, its curvature. But a curve in space twists as well as bends. Capturing that twisting requires a second invariant. It also requires a cleverer bookkeeping device. The Frenet–Serret frame is exactly this device. At every point of a sufficiently smooth space curve, it attaches a triple of mutually perpendicular unit vectors. These vectors ride along with the curve. They rotate as the curve turns and twists. This construction matters because it converts a geometric object into a dynamical system. The object is a curve sitting inside R3\mathbb{R}^3. The dynamical system is a set of vectors evolving according to linear differential equations. These are the Frenet–Serret formulas. The coefficients in these equations are exactly two scalar functions: curvature κ\kappa and torsion τ\tau. Remarkably, these two functions determine the curve completely. This holds up to a rigid motion of space. This is the seed of an idea ubiquitous in geometry: local invariants control global shape. The big picture is that the Frenet frame is the first and simplest example of a "moving frame." This technique was later vastly generalized by Cartan to arbitrary manifolds and bundles. It is the direct three-dimensional ancestor of connection forms and curvature tensors in Riemannian geometry. Even in this modest setting, it already displays the essential moves. Differentiate a frame. Express the derivatives back in terms of the frame itself. This is what makes the system "closed." Read off the coefficients as geometric invariants. Then prove a uniqueness/existence theorem from the resulting ODE system. This article builds the frame from the ground up. It gives precise definitions of tangent, normal, and binormal vectors. It offers a full proof of the Frenet–Serret differential equations. It includes two worked examples with genuinely asymmetric data, not a circle or a symmetric helix. It presents a counterexample showing why regularity and nonvanishing curvature cannot be dropped. Finally, it looks ahead to the Fundamental Theorem of Space Curves. That theorem upgrades this machinery into a full classification theorem.

Key takeaways
  • A regular, twice-differentiable space curve carries three orthonormal vectors at each point: tangent T\mathbf{T}, normal N\mathbf{N}, binormal B\mathbf{B}.
  • Curvature κ(s)≥0\kappa(s) \ge 0 measures bending; torsion τ(s)\tau(s) measures twisting out of the osculating plane.
  • The Frenet–Serret formulas express T′,N′,B′\mathbf{T}', \mathbf{N}', \mathbf{B}' (with respect to arc length) as a skew-symmetric linear combination of T,N,B\mathbf{T}, \mathbf{N}, \mathbf{B}.
  • The construction requires the curve to be regular and to have strictly positive curvature everywhere the normal is defined.
  • κ\kappa and τ\tau are invariant under rigid motions of R3\mathbb{R}^3, and together they determine the curve up to such a motion.
  • Key idea: encode a curve's shape as a rotating orthonormal frame, and let its rate of rotation — curvature and torsion — be the whole geometric story.

Where this lives today

In classical mechanics and robotics, the Frenet frame supplies a natural coordinate system for motion along a trajectory — used, for instance, in path-following control for autonomous vehicles and robotic arms, where "cross-track error" is measured relative to the moving N,B\mathbf N,\mathbf B plane rather than fixed world axes. In computer graphics and computer-aided design, Frenet frames (suitably regularized to avoid the flips seen in the counterexample, via "rotation-minimizing frames") are used to sweep a cross-section along a spline curve to build tubes, pipes, and camera paths, since a naive Frenet frame twists unpredictably wherever curvature dips toward zero. In physics, the same rotating-frame idea appears in the theory of a charged particle's trajectory in electromagnetic fields and in the kinematics of rigid-body motion, where angular velocity plays the role of the skew-symmetric matrix from the Corollary: Skew-symmetry is automatic. In differential geometry itself, the frame is the direct blueprint for the theory of connections on principal bundles, where the skew-symmetric coefficient matrix generalizes to a connection 1-form valued in a Lie algebra — the entire subject of curvature in modern geometry traces its ancestry to exactly this 3×33\times 3 skew matrix of κ\kappa and τ\tau. Finally, the two invariants κ,τ\kappa,\tau are the essential input to the Fundamental Theorem of Space Curves, which shows that any two smooth functions κ>0,τ\kappa>0,\tau on an interval are realized by an essentially unique curve, so that this ODE system is not just descriptive but fully generative of the classification of space curves.

Ask the tutor: How does the proof of the Fundamental Theorem of Space Curves use the Frenet–Serret equations as a linear ODE system to reconstruct a curve from prescribed curvature and torsion?

DefinitionFrenet–Serret frame, curvature, and torsion

Let γ:I→R3\gamma : I \to \mathbb{R}^3 be a curve of class C3C^3 on an open interval II, parametrized by arc length ss, so that

∣γ′(s)∣=1for all s∈I.|\gamma'(s)| = 1 \quad \text{for all } s \in I.

Assume further that γ\gamma has nonvanishing curvature: κ(s):=∣γ′′(s)∣>0\kappa(s) := |\gamma''(s)| > 0 for all s∈Is \in I. Define:

  • the unit tangent T(s):=γ′(s)\mathbf{T}(s) := \gamma'(s);
  • the curvature κ(s):=∣T′(s)∣=∣γ′′(s)∣\kappa(s) := |\mathbf{T}'(s)| = |\gamma''(s)|;
  • the principal normal N(s):=T′(s)/κ(s)\mathbf{N}(s) := \mathbf{T}'(s)/\kappa(s) (well-defined since κ(s)≠0\kappa(s) \neq 0);
  • the binormal B(s):=T(s)×N(s)\mathbf{B}(s) := \mathbf{T}(s) \times \mathbf{N}(s);
  • the torsion τ(s)\tau(s), defined implicitly by B′(s)=−τ(s) N(s)\mathbf{B}'(s) = -\tau(s)\,\mathbf{N}(s) (this is justified in the theorem below).

The ordered triple {T(s),N(s),B(s)}\{\mathbf{T}(s), \mathbf{N}(s), \mathbf{B}(s)\} is called the Frenet–Serret frame (or moving trihedron) of γ\gamma at parameter ss. By construction it is a positively oriented orthonormal basis of R3\mathbb{R}^3 for every ss.

Intuition

Picture driving a car along the curve at unit speed. At each instant your steering wheel and the road together determine three natural directions: straight ahead (T\mathbf{T}, the direction of travel), the direction your front wheels are turning toward (N\mathbf{N}, pointing into the curve you're carving), and a third direction perpendicular to both (B\mathbf{B}), which you can think of as "up" relative to the plane of your current turn. As you drive, this little rigid frame of three perpendicular directions rotates continuously. Curvature κ\kappa measures how fast you're turning the steering wheel in the plane you're currently in — the osculating plane spanned by T\mathbf{T} and N\mathbf{N}. Torsion τ\tau measures something new: how fast that osculating plane itself is tilting, i.e. how much the road is corkscrewing out of the plane you thought you were turning in. If τ≡0\tau \equiv 0, the osculating plane never tilts, and the whole curve lies flat in a single plane — space curves with zero torsion are exactly planar curves. A helix is the cleanest example of nonzero constant torsion: as you climb it, you're always turning at a constant rate and always corkscrewing upward at a constant rate, so both κ\kappa and τ\tau are constant. The Frenet–Serret equations formalize the "rotating frame" picture: they say that the derivatives T′,N′,B′\mathbf{T}', \mathbf{N}', \mathbf{B}' are themselves expressible purely in terms of the current frame T,N,B\mathbf{T}, \mathbf{N}, \mathbf{B}, with coefficients κ\kappa and τ\tau — the frame's own rotation is self-contained data.

Figure
The Frenet frame {T,N,B}\{\mathbf T,\mathbf N,\mathbf B\} riding along a curving, twisting space curve
T N B T N B As s increases, T, N, B rotate: rates set by κ(s) (bending) and τ(s) (twisting).
TheoremFrenet–Serret formulas

Let γ:I→R3\gamma : I \to \mathbb{R}^3 be a C3C^3 curve parametrized by arc length, with curvature κ(s)>0\kappa(s) > 0 for all s∈Is \in I, and let T,N,B\mathbf{T}, \mathbf{N}, \mathbf{B} be the Frenet–Serret frame defined above. Then there is a well-defined smooth function τ:I→R\tau : I \to \mathbb{R}, the torsion, such that for all s∈Is \in I:

T′(s)=κ(s) N(s),\mathbf{T}'(s) = \kappa(s)\,\mathbf{N}(s),
N′(s)=−κ(s) T(s)+τ(s) B(s),\mathbf{N}'(s) = -\kappa(s)\,\mathbf{T}(s) + \tau(s)\,\mathbf{B}(s),
B′(s)=−τ(s) N(s).\mathbf{B}'(s) = -\tau(s)\,\mathbf{N}(s).

Equivalently, in matrix form with F(s)=(T(s) N(s) B(s))F(s) = \big(\mathbf T(s)\ \mathbf N(s)\ \mathbf B(s)\big) as columns,

F′(s)=F(s)(0−κ(s)0κ(s)0−τ(s)0τ(s)0).F'(s) = F(s)\begin{pmatrix} 0 & -\kappa(s) & 0 \\ \kappa(s) & 0 & -\tau(s) \\ 0 & \tau(s) & 0 \end{pmatrix}.
Proof

Technique: differentiate the orthonormality relations ⟨ei,ej⟩=δij\langle e_i, e_j\rangle = \delta_{ij} and use the resulting skew-symmetry, combined with the direct definitions of T,N,B\mathbf T,\mathbf N,\mathbf B. Step 1 (the T′\mathbf{T}' equation). By definition, T(s)=γ′(s)\mathbf{T}(s) = \gamma'(s), so T′(s)=γ′′(s)\mathbf{T}'(s) = \gamma''(s). By definition κ(s)=∣γ′′(s)∣=∣T′(s)∣\kappa(s) = |\gamma''(s)| = |\mathbf T'(s)| and, when κ(s)≠0\kappa(s) \ne 0, N(s):=T′(s)/κ(s)\mathbf{N}(s) := \mathbf{T}'(s)/\kappa(s). Rearranging gives immediately

T′(s)=κ(s) N(s).\mathbf{T}'(s) = \kappa(s)\,\mathbf N(s).

This holds for every s∈Is \in I since κ(s)>0\kappa(s) > 0 throughout by hypothesis. Step 2 (orthonormality is preserved, and general skew-symmetry). Since {T,N,B}\{\mathbf T,\mathbf N,\mathbf B\} is orthonormal for every ss, we have the three scalar identities, valid on all of II:

⟨T,T⟩≡1,⟨N,N⟩≡1,⟨B,B⟩≡1,\langle \mathbf T,\mathbf T\rangle \equiv 1,\quad \langle \mathbf N,\mathbf N\rangle \equiv 1,\quad \langle \mathbf B,\mathbf B\rangle \equiv 1,
⟨T,N⟩≡0,⟨T,B⟩≡0,⟨N,B⟩≡0.\langle \mathbf T,\mathbf N\rangle \equiv 0,\quad \langle \mathbf T,\mathbf B\rangle \equiv 0,\quad \langle \mathbf N,\mathbf B\rangle \equiv 0.

Differentiating ⟨T,T⟩≡1\langle \mathbf T,\mathbf T\rangle \equiv 1 gives 2⟨T,T′⟩=02\langle\mathbf T,\mathbf T'\rangle = 0, i.e. T′\mathbf T' is orthogonal to T\mathbf T; likewise N′⊥N\mathbf N'\perp \mathbf N and B′⊥B\mathbf B'\perp \mathbf B. Since {T,N,B}\{\mathbf T,\mathbf N,\mathbf B\} is an orthonormal basis at each ss, any vector can be written in this basis; combined with the orthogonality just derived,

T′=a N+b B,N′=c T+d B,B′=e T+f N,\mathbf T' = a\,\mathbf N + b\,\mathbf B, \qquad \mathbf N' = c\,\mathbf T + d\,\mathbf B, \qquad \mathbf B' = e\,\mathbf T + f\,\mathbf N,

for scalar functions a,b,c,d,e,fa,b,c,d,e,f of ss. From Step 1 we already know a=κa = \kappa and b=0b=0 (since T′=κN\mathbf T' = \kappa \mathbf N has no B\mathbf B-component), so

T′=κ N.\mathbf T' = \kappa\,\mathbf N.

Now differentiate ⟨T,N⟩≡0\langle \mathbf T,\mathbf N\rangle \equiv 0:

⟨T′,N⟩+⟨T,N′⟩=0.\langle \mathbf T',\mathbf N\rangle + \langle \mathbf T,\mathbf N'\rangle = 0.

Since T′=κN\mathbf T' = \kappa \mathbf N, we get ⟨T′,N⟩=κ\langle \mathbf T',\mathbf N\rangle = \kappa, so ⟨T,N′⟩=−κ\langle \mathbf T,\mathbf N'\rangle = -\kappa. But N′=c T+d B\mathbf N' = c\,\mathbf T + d\,\mathbf B, and taking the inner product with T\mathbf T (using ⟨T,T⟩=1, ⟨T,B⟩=0\langle \mathbf T,\mathbf T\rangle=1,\ \langle \mathbf T,\mathbf B\rangle = 0) gives ⟨T,N′⟩=c\langle \mathbf T, \mathbf N'\rangle = c. Hence c=−κc = -\kappa. Differentiate ⟨T,B⟩≡0\langle \mathbf T,\mathbf B\rangle \equiv 0:

⟨T′,B⟩+⟨T,B′⟩=0.\langle \mathbf T',\mathbf B\rangle + \langle \mathbf T,\mathbf B'\rangle = 0.

Since T′=κN\mathbf T' = \kappa\mathbf N and ⟨N,B⟩=0\langle \mathbf N,\mathbf B\rangle = 0, the first term vanishes, so ⟨T,B′⟩=0\langle \mathbf T,\mathbf B'\rangle = 0. But B′=e T+f N\mathbf B' = e\,\mathbf T + f\,\mathbf N, so ⟨T,B′⟩=e\langle \mathbf T,\mathbf B'\rangle = e. Hence e=0e = 0. Differentiate ⟨N,B⟩≡0\langle \mathbf N,\mathbf B\rangle \equiv 0:

⟨N′,B⟩+⟨N,B′⟩=0.\langle \mathbf N',\mathbf B\rangle + \langle \mathbf N,\mathbf B'\rangle = 0.

Now N′=c T+d B=−κ T+d B\mathbf N' = c\,\mathbf T + d\,\mathbf B = -\kappa\,\mathbf T + d\,\mathbf B, so ⟨N′,B⟩=d\langle \mathbf N',\mathbf B\rangle = d. And B′=f N\mathbf B' = f\,\mathbf N (using e=0e=0), so ⟨N,B′⟩=f\langle \mathbf N,\mathbf B'\rangle = f. Thus d+f=0d + f = 0, i.e. f=−df = -d. Step 3 (identify dd as τ\tau, and confirm the third equation). Define

τ(s):=d(s)=⟨N′(s),B(s)⟩.\tau(s) := d(s) = \langle \mathbf N'(s), \mathbf B(s)\rangle.

Since N=T′/κ\mathbf N = \mathbf T'/\kappa is C1C^1 (as γ∈C3\gamma \in C^3 makes T′=γ′′\mathbf T' = \gamma'' differentiable, and κ\kappa is C1C^1 where nonvanishing, by the quotient/product rule), N′\mathbf N' is continuous, and B\mathbf B is continuous, so τ\tau is a well-defined continuous — in fact, by the same regularity chain, C1C^1 — function on II. From the relations found above:

N′=−κ T+τ B,B′=f N=−τ N.\mathbf N' = -\kappa\,\mathbf T + \tau\,\mathbf B, \qquad \mathbf B' = f\,\mathbf N = -\tau\,\mathbf N.

These are exactly the second and third Frenet–Serret formulas. Step 4 (matrix form). Collecting the three vector identities

T′=0⋅T+κ N+0⋅B,\mathbf T' = 0\cdot\mathbf T + \kappa\,\mathbf N + 0\cdot \mathbf B,
N′=−κ T+0⋅N+τ B,\mathbf N' = -\kappa\,\mathbf T + 0\cdot\mathbf N + \tau\,\mathbf B,
B′=0⋅T−τ N+0⋅B,\mathbf B' = 0\cdot\mathbf T - \tau\,\mathbf N + 0\cdot \mathbf B,

and writing them as columns of the derivative of F(s)=(T N B)F(s) = (\mathbf T\ \mathbf N\ \mathbf B) (each column differentiated, expressed in the T,N,B\mathbf T,\mathbf N,\mathbf B basis with the rows of the coefficient matrix giving the T,N,B\mathbf T,\mathbf N,\mathbf B-components) gives precisely

F′(s)=F(s)(0−κ0κ0−τ0τ0).F'(s) = F(s)\begin{pmatrix} 0 & -\kappa & 0\\ \kappa & 0 & -\tau \\ 0 & \tau & 0\end{pmatrix}.

The coefficient matrix is skew-symmetric, consistent with F(s)F(s) remaining orthogonal for all ss (this skew-symmetry is not an extra assumption — it is forced by orthonormality, exactly as derived in Steps 2–3). ■\blacksquare

∎
CorollarySkew-symmetry is automatic

For any C1C^1 family of orthonormal frames F(s)∈O(3)F(s) \in O(3), the matrix A(s)=F(s)−1F′(s)=F(s)TF′(s)A(s) = F(s)^{-1}F'(s) = F(s)^{\mathsf T}F'(s) is skew-symmetric. This follows by differentiating FTF=IF^{\mathsf T}F = I: (F′)TF+FTF′=0(F')^{\mathsf T}F + F^{\mathsf T}F' = 0, i.e. AT=−AA^{\mathsf T} = -A. The Frenet–Serret matrix is the special case where, additionally, the (1,3)(1,3) and (3,1)(3,1) entries vanish because T′\mathbf T' has no B\mathbf B-component by definition of N\mathbf N.

Worked examples

Example 1: a non-symmetric, non-circular, non-helical curve. Consider γ(t)=(t,t2,t3)\gamma(t) = (t, t^2, t^3) for tt near 00 — deliberately asymmetric so no coincidental cancellations hide an error. This is not arc-length parametrized, so we must use the general (non-unit-speed) formulas, which follow from the arc-length ones by the chain rule:

κ=∣γ′×γ′′∣∣γ′∣3,τ=(γ′×γ′′)⋅γ′′′∣γ′×γ′′∣2.\kappa = \frac{|\gamma'\times\gamma''|}{|\gamma'|^3}, \qquad \tau = \frac{(\gamma'\times\gamma'')\cdot \gamma'''}{|\gamma'\times\gamma''|^2}.

Compute derivatives at general tt:

γ′(t)=(1,2t,3t2),γ′′(t)=(0,2,6t),γ′′′(t)=(0,0,6).\gamma'(t) = (1, 2t, 3t^2), \qquad \gamma''(t) = (0,2,6t), \qquad \gamma'''(t) = (0,0,6).

Cross product:

γ′×γ′′=(2t⋅6t−3t2⋅2, 3t2⋅0−1⋅6t, 1⋅2−2t⋅0)=(6t2,−6t,2).\gamma'\times\gamma'' = \big(2t\cdot 6t - 3t^2\cdot 2,\ 3t^2\cdot 0 - 1\cdot 6t,\ 1\cdot 2 - 2t\cdot 0\big) = (6t^2, -6t, 2).

Evaluate everything at t=1t=1 (an asymmetric, non-special point):

γ′(1)=(1,2,3),γ′′(1)=(0,2,6),γ′′′(1)=(0,0,6),\gamma'(1) = (1,2,3), \qquad \gamma''(1) = (0,2,6), \qquad \gamma'''(1) = (0,0,6),
γ′×γ′′∣t=1=(6,−6,2).\gamma'\times\gamma''\big|_{t=1} = (6, -6, 2).

Magnitudes:

∣γ′(1)∣=1+4+9=14,∣γ′×γ′′∣t=1=36+36+4=76=219.|\gamma'(1)| = \sqrt{1+4+9} = \sqrt{14}, \qquad |\gamma'\times\gamma''|_{t=1} = \sqrt{36+36+4} = \sqrt{76} = 2\sqrt{19}.

Curvature at t=1t=1:

κ(1)=219(14)3=2191414=19714=19⋅147⋅14=26698.\kappa(1) = \frac{2\sqrt{19}}{(\sqrt{14})^3} = \frac{2\sqrt{19}}{14\sqrt{14}} = \frac{\sqrt{19}}{7\sqrt{14}} = \frac{\sqrt{19}\cdot\sqrt{14}}{7\cdot 14} = \frac{\sqrt{266}}{98}.

Numerically, 266≈16.31\sqrt{266}\approx 16.31, so κ(1)≈0.1665\kappa(1) \approx 0.1665. Torsion: numerator is

(γ′×γ′′)⋅γ′′′=(6,−6,2)⋅(0,0,6)=12.(\gamma'\times\gamma'')\cdot \gamma''' = (6,-6,2)\cdot(0,0,6) = 12.

Denominator is ∣γ′×γ′′∣2=76|\gamma'\times\gamma''|^2 = 76. So

τ(1)=1276=319≈0.1579.\tau(1) = \frac{12}{76} = \frac{3}{19} \approx 0.1579.

So at t=1t=1 this twisted cubic has κ(1)=266/98≈0.1665\kappa(1) = \sqrt{266}/98 \approx 0.1665 and τ(1)=3/19≈0.1579\tau(1) = 3/19 \approx 0.1579 — both strictly positive and unequal, confirming the curve genuinely bends and twists with no special symmetry forcing them to coincide (which they don't). One can also verify directly that T(1)=γ′(1)/∣γ′(1)∣=(1,2,3)/14\mathbf T(1) = \gamma'(1)/|\gamma'(1)| = (1,2,3)/\sqrt{14}, and that B(1)\mathbf B(1), being parallel to γ′×γ′′=(6,−6,2)\gamma'\times\gamma''=(6,-6,2) (parallel to (3,−3,1)(3,-3,1) after scaling), is B(1)=(3,−3,1)/19\mathbf B(1) = (3,-3,1)/\sqrt{19}, consistent with ∣B(1)∣=1|\mathbf B(1)|=1 since 9+9+1=199+9+1=19.

Example 2: an elliptical helix (asymmetric radii) parametrized by arc length is impractical, so we verify the Frenet equations directly on a non-unit-speed but exactly solvable curve. Let γ(t)=(3cos⁡t,4sin⁡t,0)\gamma(t) = (3\cos t, 4\sin t, 0), an ellipse in the xyxy-plane with semi-axes 33 and 44 — chosen asymmetric (a≠ba\ne b) so the curvature is genuinely non-constant, unlike a circle. Since it is planar, we predict τ≡0\tau \equiv 0 everywhere, and we verify this while computing a nontrivial, varying κ(t)\kappa(t). Derivatives:

γ′(t)=(−3sin⁡t,4cos⁡t,0),γ′′(t)=(−3cos⁡t,−4sin⁡t,0),γ′′′(t)=(3sin⁡t,−4cos⁡t,0).\gamma'(t) = (-3\sin t, 4\cos t, 0), \qquad \gamma''(t) = (-3\cos t, -4\sin t, 0), \qquad \gamma'''(t) = (3\sin t, -4\cos t, 0).

Cross product γ′×γ′′\gamma'\times\gamma'': since both lie in the xyxy-plane, the result is purely in z^\hat z:

γ′×γ′′=(0, 0, (−3sin⁡t)(−4sin⁡t)−(4cos⁡t)(−3cos⁡t))=(0,0, 12sin⁡2t+12cos⁡2t)=(0,0,12).\gamma'\times\gamma'' = \big(0,\ 0,\ (-3\sin t)(-4\sin t) - (4\cos t)(-3\cos t)\big) = (0,0,\ 12\sin^2 t + 12\cos^2 t) = (0,0,12).

This is constant — a clean check, since for a planar curve the "binormal direction" γ′×γ′′\gamma'\times\gamma'' should stay perpendicular to the plane, here the constant vector (0,0,12)(0,0,12). Torsion numerator: (γ′×γ′′)⋅γ′′′=(0,0,12)⋅(3sin⁡t,−4cos⁡t,0)=0(\gamma'\times\gamma'')\cdot\gamma''' = (0,0,12)\cdot(3\sin t,-4\cos t,0) = 0 for every tt, since the third coordinate of γ′′′\gamma''' is 00. Hence

τ(t)≡0,\tau(t) \equiv 0,

confirming the general fact that planar curves have identically zero torsion. Now the curvature, which is genuinely non-constant here (unlike a circle):

∣γ′(t)∣2=9sin⁡2t+16cos⁡2t=9+7cos⁡2t,|\gamma'(t)|^2 = 9\sin^2 t + 16\cos^2 t = 9 + 7\cos^2 t,
∣γ′×γ′′∣=12,|\gamma'\times\gamma''| = 12,
κ(t)=12(9+7cos⁡2t)3/2.\kappa(t) = \frac{12}{(9+7\cos^2 t)^{3/2}}.

At t=0t = 0: κ(0)=12/(16)3/2=12/64=3/16=0.1875\kappa(0) = 12/(16)^{3/2} = 12/64 = 3/16 = 0.1875. At t=π/2t=\pi/2: κ(π/2)=12/(9)3/2=12/27=4/9≈0.4444\kappa(\pi/2) = 12/(9)^{3/2} = 12/27 = 4/9 \approx 0.4444. These differ (as they must, since 3≠43 \ne 4), correctly reflecting that curvature is largest at the ends of the short semi-axis (yy-axis, radius of curvature small) and smallest at the ends of the long semi-axis — a standard but non-trivial sanity check that our formula tracks the geometry correctly.

Counterexampledropping "nonvanishing curvature" breaks the frame

The hypothesis κ(s)>0\kappa(s) > 0 for all ss is essential to Step 1 of the proof, where N(s):=T′(s)/κ(s)\mathbf N(s) := \mathbf T'(s)/\kappa(s) requires dividing by κ(s)\kappa(s). Consider

γ(t)=(t,t3,0),t∈R.\gamma(t) = (t, t^3, 0), \qquad t \in \mathbb{R}.

This is a smooth, regular (γ′(t)=(1,3t2,0)≠0\gamma'(t) = (1,3t^2,0) \ne 0 for all tt) planar curve. Its curvature is

κ(t)=∣γ′×γ′′∣∣γ′∣3=∣(1,3t2,0)×(0,6t,0)∣(1+9t4)3/2=∣(0,0,6t)∣(1+9t4)3/2=6∣t∣(1+9t4)3/2.\kappa(t) = \frac{|\gamma'\times\gamma''|}{|\gamma'|^3} = \frac{|(1,3t^2,0)\times(0,6t,0)|}{(1+9t^4)^{3/2}} = \frac{|(0,0,6t)|}{(1+9t^4)^{3/2}} = \frac{6|t|}{(1+9t^4)^{3/2}}.

At t=0t=0, κ(0)=0\kappa(0) = 0: the curve has an inflection point there (it changes concavity, since γ(t)=(t,t3,0)\gamma(t)=(t,t^3,0) looks like a cubic curve through the origin, convex for t>0t>0 and concave for t<0t<0 in the yy-coordinate relative to xx). At t=0t=0, T′(0)=γ′′(0)/∣γ′(0)∣⋅(reparametrization factor)\mathbf T'(0) = \gamma''(0)/|\gamma'(0)| \cdot(\text{reparametrization factor}); more directly, in arc length terms T′(s0)=0\mathbf T'(s_0) = \mathbf 0 at the point corresponding to t=0t=0 (since κ=0\kappa=0 there), so the formula N:=T′/κ\mathbf N := \mathbf T'/\kappa becomes 0/0\mathbf 0/0: undefined. There is no way to choose N(0)\mathbf N(0) continuously that is forced by the construction — indeed, as t→0−t\to 0^-, the normal points one way, and as t→0+t \to 0^+ it points the opposite way (the concavity flips), so N(t)\mathbf N(t) has a jump discontinuity as tt crosses 00, however you try to patch it. Concretely, for t>0t>0 small, γ′′(t)=(0,6t,0)\gamma''(t) = (0,6t,0) points in +y^+\hat y after normalizing perpendicular to T\mathbf T, while for t<0t<0 small it points in −y^-\hat y direction relative to the frame — the sign of the yy-component of N\mathbf N flips as tt passes through 00. Consequently the binormal B=T×N\mathbf B = \mathbf T \times \mathbf N and the torsion τ\tau (defined via B′=−τN\mathbf B' = -\tau \mathbf N) are likewise undefined at t=0t = 0, and the entire Frenet apparatus — theorem included — simply fails to apply at inflection points. This is exactly why the theorem's hypothesis κ(s)>0\kappa(s) > 0 for all s∈Is \in I cannot be weakened to "κ(s)≥0\kappa(s) \ge 0" or "κ\kappa vanishes only at isolated points": the frame is not merely harder to compute there, it is genuinely not well-defined, and no amount of extra smoothness of γ\gamma repairs this.

Scope, edge cases & common traps

SituationWhat happens
κ(s)>0\kappa(s) > 0 everywhere on II, γ∈C3\gamma \in C^3Frenet frame and τ\tau well-defined and smooth; theorem applies fully.
κ(s0)=0\kappa(s_0) = 0 at an isolated pointN\mathbf N, B\mathbf B, τ\tau undefined at s0s_0; frame may have a genuine discontinuity there (see counterexample).
γ\gamma only C2C^2T′\mathbf T' exists and is continuous, but N′\mathbf N' (needed for τ\tau) may fail to exist; need C3C^3 for τ\tau to be defined via a derivative of N\mathbf N.
γ\gamma planar (lies in a fixed plane)τ≡0\tau \equiv 0; conversely τ≡0\tau \equiv 0 with κ>0\kappa>0 forces γ\gamma planar (a short additional argument via the Fundamental Theorem of Space Curves).
γ\gamma a straight lineκ≡0\kappa \equiv 0 everywhere, so the frame is undefined globally, not just at isolated points — the whole apparatus is vacuous for lines.
Not arc-length parametrizedMust convert with chain-rule formulas for κ,τ\kappa,\tau in terms of γ′,γ′′,γ′′′\gamma',\gamma'',\gamma'''; using the arc-length formulas directly on tt gives wrong numerical values.

The mathematician behind it

Origins · 1 minJean Frédéric Frenet and how this was discovered
JF
Jean Frédéric Frenet
1816–1900 · French mathematician and astronomer

Frenet was a student at the École Normale Supérieure and later a professor at Lyon, where he also served as an astronomer and later as rector of the academy at various French cities. His 1847 doctoral thesis introduced what are now called the Frenet formulas for space curves, though he published only a partial announcement at the time; the full, independent development is usually credited jointly to Joseph Serret, who published a more complete set of the formulas in 1851. The two names have been attached together ever since, an example of simultaneous or near-simultaneous discovery common in nineteenth-century differential geometry, when the systematic study of curves and surfaces using calculus was rapidly maturing following the foundational work of Gauss on surfaces.

History
- 1847 — Frenet's doctoral thesis at Toulouse presents the formulas for a moving frame along a space curve, in a form restricted to the case τ≠0\tau \ne 0.
- 1851 — Joseph Alfred Serret publishes an independent and more general derivation of the same system of equations, extending the formulas.
- 1850s–1860s — the formulas become standard tools in the emerging differential geometry of curves, alongside Gauss's earlier (1827) theory of curved surfaces.
- Early 1900s — the moving-frame idea is generalized by Élie Cartan into the powerful method of "repères mobiles" (moving frames) applicable to general manifolds and Lie groups.
- Mid-20th century onward — the Frenet frame becomes a standard topic in differential geometry courses, and its ODE-system viewpoint becomes the model for connection and curvature in modern Riemannian and gauge-theoretic geometry.
Why do we insist on arc-length parametrization in the definition?

Arc-length parametrization makes ∣γ′(s)∣=1|\gamma'(s)| = 1 identically, which is what lets us call T=γ′\mathbf T = \gamma' a unit tangent without further normalization, and crucially makes T′\mathbf T' automatically perpendicular to T\mathbf T (differentiate ∣T∣2=1|\mathbf T|^2=1). This perpendicularity is exactly what lets us define N\mathbf N as the normalized T′\mathbf T' and builds the whole orthonormal frame cleanly. For a curve parametrized by an arbitrary parameter tt, one can still define κ\kappa and τ\tau via the chain-rule formulas involving γ′,γ′′,γ′′′\gamma',\gamma'',\gamma''', and these agree with the arc-length definitions after reparametrizing, but the derivation of the Frenet–Serret equations is cleanest and most transparent in arc length, which is why the theorem is always stated that way first.

Is torsion ever allowed to be negative?

Yes — unlike curvature, which is defined as a norm and hence is always ≥0\ge 0 (and here strictly positive by hypothesis), torsion is a signed quantity, defined via the coefficient in B′=−τN\mathbf B' = -\tau \mathbf N. A right-handed helix has positive torsion and a left-handed helix (its mirror image) has negative torsion of the same magnitude, while curvature is identical for both since curvature cannot detect handedness. This sign is exactly what allows torsion, together with curvature, to distinguish a curve from its mirror image, a distinction curvature alone can never make.

Does the Frenet frame depend on the choice of arc-length starting point or direction?

The frame is unaffected by where you set s=0s=0, since it depends only on derivatives at each point. Reversing the direction of traversal (s↦−ss \mapsto -s) flips T↦−T\mathbf T \mapsto -\mathbf T; since N\mathbf N is built from T′/κ\mathbf T'/\kappa and both the numerator's sign flip (from the extra chain-rule minus sign) and κ≥0\kappa\ge0 combine so that N\mathbf N is actually left unchanged; and then B=T×N\mathbf B = \mathbf T\times \mathbf N flips sign because T\mathbf T does. Correspondingly τ\tau, computed from B′\mathbf B' and N\mathbf N, is left unchanged under this reversal, while κ\kappa (a norm) is trivially unchanged too. So torsion and curvature are honest geometric invariants of the unoriented curve, while T,B\mathbf T,\mathbf B individually depend on a chosen orientation.

How does this relate to the Fundamental Theorem of Space Curves?

The Frenet–Serret formulas show that κ\kappa and τ\tau are exactly the coefficients controlling how the frame rotates, i.e. they are the "differential data" of the curve's shape. The Fundamental Theorem of Space Curves goes one step further: it proves that given any smooth functions κ(s)>0\kappa(s) > 0 and τ(s)\tau(s) on an interval, there exists a space curve realizing them as its curvature and torsion, and that this curve is unique up to a rigid motion (rotation and translation) of R3\mathbb{R}^3. The proof uses exactly the Frenet–Serret system as a linear ODE for the frame F(s)F(s), invokes an existence-and-uniqueness theorem for linear systems of ODEs to solve for FF, and then integrates T(s)=γ′(s)\mathbf T(s)=\gamma'(s) to recover γ\gamma itself.

What goes wrong with the frame at inflection points in practical software (e.g. computer graphics)?

Exactly the phenomenon in the counterexample: wherever curvature dips to zero or near zero along a spline, the classical Frenet frame's normal direction becomes ill-defined or numerically unstable, causing visible "flipping" or twisting artifacts when the frame is used to sweep geometry (like a tube) along the curve. Practitioners instead use "rotation-minimizing frames," which are also orthonormal moving frames with tangent vector T\mathbf T, but are chosen to rotate as little as possible around T\mathbf T rather than being pinned to the curve's exact osculating plane; these coincide with the Frenet frame away from inflection points but remain smooth and well-defined through them.

Can curvature and torsion be recovered directly from a non-arc-length parametrization without reparametrizing?

Yes: for γ(t)\gamma(t) with γ′(t)≠0\gamma'(t) \ne 0 but tt not necessarily arc length, the curvature is κ=∣γ′×γ′′∣/∣γ′∣3\kappa = |\gamma'\times\gamma''|/|\gamma'|^3 and the torsion is τ=(γ′×γ′′)⋅γ′′′/∣γ′×γ′′∣2\tau = (\gamma'\times\gamma'')\cdot\gamma'''/|\gamma'\times\gamma''|^2, both derived by the chain rule from the arc-length definitions (s=s(t)s=s(t) with ds/dt=∣γ′(t)∣ds/dt = |\gamma'(t)|). These formulas are what was used in the first worked example above, since γ(t)=(t,t2,t3)\gamma(t)=(t,t^2,t^3) is not unit speed; they always agree with direct arc-length computation after the substitution s(t)=∫0t∣γ′(u)∣ dus(t) = \int_0^t |\gamma'(u)|\,du, but save the (often impossible in closed form) step of actually inverting s(t)s(t).

Why is the binormal defined via a cross product rather than another derivative?

Once T\mathbf T and N\mathbf N are fixed as two orthonormal vectors, there are exactly two unit vectors orthogonal to both, differing by a sign; the cross product B:=T×N\mathbf B := \mathbf T\times\mathbf N is simply the canonical choice that makes {T,N,B}\{\mathbf T,\mathbf N,\mathbf B\} a right-handed (positively oriented) orthonormal basis, matching the standard orientation of R3\mathbb{R}^3. One could equally well define B\mathbf B some other way and then derive its derivative, but using the cross product keeps the frame's orientation consistent and matches the convention used in virtually all textbooks, which matters when comparing formulas or signs of torsion across sources.

Before this

Where this leads