You are on page 1of 2

LECTURE 9

We can particularise the previous formula to the case when the curve is in fact included
in a lower dimensional subspace:
Definition 79. We say that α : I → R3 is a plane, or planar, curve if there is a plane
P ⊂ R3 such that α(s) ∈ P for all s ∈ I.
Proposition 80. Given a regular curve α : I → R3 parametrised by arc-length so that
κ(s) > 0 for all s ∈ I, then α is planar if and only if its torsion τ (s) = 0 for all s ∈ I.
Proof. If ḃ = 0 then b = t × n is constant and α̇ and α̈ belong to the plane orthogonal to
this constant vector, so that α remains in the affine plane determined by its initial condition
and this vector plane. Conversely α ∈ P implies α̇ ∈ P and α̈ ∈ P so b constant.

So we have seen that given a regular curve α, we get an orthonormal basis of R3 as well
as notions of curvature and torsion via the Frenet formulas. The remarkable thing is that the
Frenet frame satisfy an ODE system which just depends on the curvature and the torsion. So
this suggests that if we specify the curvature and torsion of the curve, then this completely
determines the curve up to initial data. This is exactly:
Theorem 81 (Fundamental Theoreom of Curves in R3 ). Let κ(s) and τ (s) be smooth func-
tions from an interval I to R3 so that κ(s) > 0 for all s ∈ I. Then there is a regular curve
α : I → R3 parametrised by arc-length so that κ is the curvature of α and τ is the torsion of
α. Moreover if α and β are two regular curves parametrised by arc-length with curvatures and
torsions given by κ and τ then there R ∈ SO(3) (a rotation) and b ∈ R3 so that β = R ◦ α + b,
i.e. β is just a rotated and translated version of α.
Proof. The existence part of the theorem follows from the Picard-Lindelöf theorem, given that
the equations on (t, n, b) are a system of linear ordinary differential equations, provided the
functions κ and τ are given. Such system has unique solutions that can be continued globally
on I because of the linearity. The resulting curve Frenet frame is unique up to the choice
of initial conditions at some s0 ∈ I, but there is always a rotation R ∈ SO(3) that maps
an orthonormal basis (t, n, b) to an other orthonormal basis (t̃, ñ, b̃) at this point s0 . Once
this choice is made, it determines α̇, and it remains to choose the initial condition α(s0 ), but
there is always a translation by some b ∈ R3 that maps initial condition α(s0 ) to another
initial condition β(s0 ). Note that is it easy to check that if α is a regular curve parametrised
by arc-length with curvature κ and torsion τ , then β = R ◦ α + b is again a regular curve
parametrised by arc-length with curvature κ and torsion τ (exercise).

3.2 The isoperimetric inequality in the plane


We now focus on curves in the plane, and use the tools we have introduced to address
rigorously one of the oldest problems of mathematics (search “Dido’s problem” online): an
inequality between perimeter and area, i.e. a global ridigity property.
Theorem 82. Given Ω ⊂ R2 a connected open set with compact closure and so that ∂Ω is a
connected 1-manifold (say smooth, but of class C 1 would be suffice), then

4πA(Ω) ≤ ℓ(∂Ω)2

20
where ℓ(∂Ω) is the length of the curve and A(Ω) is the area enclosed.
Remark 83. There are many proofs of this result. We present the proof of Hurwitz (1901).
Its idea is to reduce the inequality to the estimation of the first Laplacian eigenvalue on the
circle, captured by the so-called Wirtinger inequality and proved by basic Fourier calculus.

Proof. Since we want to relate perimeter and area it is natural to search for a vector field
whose divergence is constant. The simplest one is F(x, v) := (x + a, y + b) for (a, b) ∈ R2 to
be determined. We choose (a, b) so that
Z
F(x, v) dS = 0 (3.1)
∂Ω

where dS is the element of length on the curve. Then we apply the divergence theorem:
Z Z Z Z 1
1 2
2
2A(Ω) = divF = (F · n) dS ≤ |F| dS ≤ ℓ(Ω) 2 |F| dS (3.2)
Ω ∂Ω ∂Ω ∂Ω

where n is the outgoing normal unit vector. If we denote (x(s), y(s)) the closed curve describ-
ing ∂Ω in arc-length parametrisation (smooth by assumption) we have G(s) = F(x(s), y(s)) =
(x(s) − a, y(s) − b) is a smooth periodic function on [0, ℓ(Ω)] with zero average on this interval
thanks to (3.1). We can therefore use the Wirtinger inequality (aka Poincaré inequality
with periodic boundary condition):
Z ℓ(Ω) Z
2 ℓ(Ω)2 ℓ(Ω) ′ 2
|G(s)| ds ≤ 2
G (s) ds. (3.3)
0 4π 0

Let us first apply this inequality to conclude, and prove it afterwards. Since (x(s), y(s)) is
parametrised by arc-length, |G′ (s)| = |(x′ (s), y ′ (s))| = 1 and (3.2) implies
Z 1 Z !1
1 2 1
ℓ(Ω) 2

2A(Ω) ≤ ℓ(Ω) 2 |F|2 ds = ℓ(Ω) 2 |G(s)|2 ds


∂Ω 0
3 Z !1
ℓ(Ω)
ℓ(Ω) 2 2
2
ℓ(Ω)2
≤ G′ (s) ds ≤
2π 0 2π

which concludes the proof.


It remains to prove the inequality (3.3). Denoting L = ℓ(Ω) and working on each com-
ponent it is enough to prove that for f : R → R smooth L-periodic and with average zero
Z L Z L
2 L2 2
|f | ds ≤ 2 f ′ ds.
0 4π 0
P
One decomposes in Fourier series f = k∈Z ak e2iπkt/L . The zero average condition imposes
P
a0 = 0. Moreover f ′ = k∈Z 2iπkaL e
k 2iπkt/L
and by orthogonality
Z L ! Z L
2
X L2 X 2iπkak 2 L2 2
|f | ds = 2
|ak | L ≤ 2 L = 2 f ′ ds
0 ∗
4π ∗
L 4π 0
k∈Z k∈Z

which concludes the proof of (3.2).

21

You might also like