The Riemann-Lebesgue lemma states that if f∈L1(T) then ∣f^(n)∣→0 when ∣n∣→∞. In particular, this implies that for an integrable function on the circle, ∫02πf(x)sin(nx)dx→0 and ∫02πf(x)cos(nx)→0 when ∣n∣→∞.
Proof.
If f is continuous, then the Parsval identity implies the positive infinite series ∑N=−∞+∞∣f^(n)∣=∥f∥L2≤+∞ so this implies ∣f^(n)∣→0 when ∣n∣→∞.
In general, let g be a continuous function such that ∥f−g∥L1<ϵ. Then we have ∣f^(n)−g^(n)∣=∣∫01(f(x)−g(x))e−2πinxdx∣≤∫01∣f(x)−g(x)∣dx=∥f−g∥L1<ϵ. So ∣g^(n)∣→∞ implies ∣f^(n)∣→∞.
Let f be a 2π-periodic function with mean value 0 (i.e. f^(0)=∫02πf(x)2πdx = 0 ). Then ∥f′∥L2≥∥f∥L2, i.e. ∫02π∣f′(x)∣22πdx≥∫02π∣f(x)∣22πdx.
Proof. By the derivative theorem f′^(n)=inf^(n). Then the Parsval identity implies ∥f′∥L22=∑n∈Zn2∣f^(n)∣2.
We apply the simple fact that multiplying a strictly positive number by n2 will make it bigger for ∣n∣≥2 and equal when ∣n∣=1, so when f^(0)=0, ∥SN(f)∥L22=∑∣n∣≤N,n=0∣f^(n)∣2≤∑n∈Z,n=0n2∣f^(n)∣2=∥SN(f′)∥L22. If f^(n)=0 for all ∣n∣≥2, the above inequality will be equality (converse also true).
Then let N→∞, the left hand side goes to ∥f∥L22, the right hand side goes to ∥f′∥L22.
As remarked, the inequality holds if and only if f^(n)=0 for all n except ∣n∣=1, so f(x)=f^(1)e2πix+f(−1)e−2πix. If f is real-valued, then f(x)=asin(x)+bcos(x).
A regular curve is a C1 mapping γ:[a,b]→R2, t↦(x(t),y(t)):=γ(t), such that γ′(t)=(x′(t),y′(t))=0 for every t∈[a,b]. Here the C1 means both x(t) and y(t) has continuous first derivatives.
The length of the curve is given by the line integral ∫ab∣γ′(t)∣dt=∫ab(x′(t)2+y′(t)2)dt.
A reparametrization of the curve is determined by composition γ(t(s)):[c,d]↦R2, where t(s) is a C1 function s∈[c,d]↦t(s)∈[a,b], such that t(c)=a,t(b)=d and t′(s)=0 for every s. In particular, t(s) is invertible, and either strictly increasing or strictly decreasing. If t′(s)>0, the reparametrization keeps orientation, and reverses the orientation if t′(s)<0.
The length l of a curve is invariant under reparametrization.
An arc-length parametrization of the curve C is γ(s):[0,l]→C such that ∣γ′(s)∣≡1. Any regular curve can be arc-length parametrized. To see this, let γ(t):[a,b]→C be any parametrization of the curve. Let s(t)=∫at∣γ′(x)∣dx be the length up to γ(t). For every t0∈(a,b), s′(t0)=∣γ′(t0)∣=0 because the curve is regular, then by the inverse function theorem, the function s(t) has inverse, denoted by t(s) such that s(t(s))=s, on a neighbourhood of s(t0). Repeating this step we get a reparemetrization γ(t(s)), which is piecewisely defined on [0,l], such that ∣dsdγ(t(s))∣=∣γ′(t(s))∣∣t′(s)∣=∣γ′(t(s))∣∣γ′(t(s))∣1=1. This is the arc-length parametrization of C.
A curve γ(t) is closed provided that γ(a)=γ(b). A simple closed curve is a closed curve without self intersection. Let C be a regular simple closed curve, then C surronds a domain D with ∂D=C. By the Green formula, the area of D can be calculated by the line integral 21∣∫Cxdy−ydx∣. One can also verify this by Stokes theorem.
Let C be a simple closed curve on R2 with length l. Let A be the area of the region enclosed by C. Then l2≥4πA, and the equality holds if and only if C is a circle.
We prove the basic case when l=2π, so the inequality reduces to A≤π. The general case follows directly from the basic case by scaling.
If C is the unit circle, then l=2π, A=π so 4πA=4π2=l2.
Let C be a C1 curve.
Let γ(s):[0,2π] be the arc-length parametrization of C. Write γ(s)=(x(s),y(s)), then x(s),y(s)∈C1(S1). Since γ(s) is arc-length parametrization, we have ∣γ′(s)∣≡1 i.e. x′(s)2+y′(s)2≡1, so 2π1∫02πx′(s)2+y′(s)2dx=1, this implies
where (⋅) is the inner product of functions on the circle. We want to show that A≤2π, so it reduces to show that ∣(x,y′)−(y,x′)∣≤1 and find condition of equality.
Since x(s),y(s) are C1 functions, x(s)=∑n=−∞+∞aneins and y(s)=∑n=−∞+∞bneins. By the derivative theorem, x′(s)=∑n=−∞+∞anineins and y′(s)=∑n=−∞+∞bnineins. Note that here writing “=” is an abuse of notation because the Fourier series of x′ and y′ may not converge pointwise but here we just use it to calculate integrals, which will not cause problem.
Then the Parsval identity and (2) implies ∑n=−∞+∞n2(∣an∣2+∣bn∣2)=1. Moreover, (x,y′)=∑n=−∞+∞aninbn, (y,x′)=∑n=−∞+∞bninan (this is a generalization of the finite dimensional identity, but can be proved using Paraval and the polarization identity), see Page 81, lemma 1.5 of Stein-Shakarchi).
Now we analyze the condition when the equality holds. First, if C is a circle, we know the equality holds. Conversely, suppose the equality hold, necessarily the (a)(b)(c) part should take equal.
When (c) takes “=”, the summation should not have any term with ∣n∣≥2. So x(s)=a0+a1eis+a−1e−is, y(s)=b0+b1eis+b−1e−is. Note that x,y are real functions, so a−1=a1, b−1=b1, a0 and b0 are real. Let a1=r1eiα, b1=r2e−iβ. Then x(s)=a0+r1cos(α+s), y=b0+r2cos(β+s). Now to show that γ(s)=(x(s),y(s)) is a circle, it reduces to show that r1=r2 and α−β=2kπ. These shall follow from equality (b) and (a).
When (b) takes “=”, we conclude that ∣a1∣=∣b1∣, this implies r1=r2=21.
When (a) takes “=”, by writing the equation explicitly we get ∣a1b1−b1a1∣=2∣a1∣∣b1∣=2⋅21⋅21=21. Then plug in a1=21eiα and b1=21eiβ to the left hand side we get ∣41⋅2isin(α−β)∣=21, so ∣sin(α−β)∣=1. This implies α−β=2kπ. The proof is complete.
Proof of isoperimetric inequality using Wirtinger’s inequality¶
Still, we consider the basic case l=2π and take arc-length parametrization (x(t),y(t)) of C. The idea is to apply Wirtinger’s inequality to x(t), so we need its mean value to be zero, this can be done by replacing x(t) by x(t)−∫02πx(t)dt.
Also, the area A=21∫Cxdy−ydx=∫02πxdy. To see this, integration by parts implies ∫Cxdy=∫02πx(t)dy(t)=x(t)y(t)∣02π−∫02πy(t)dx(t)=0−∫Cydx.
The first term is ≥0 by Wirtinger inequality, the second is automatically ≥0. If the curve C acheives equality l2=4πA, then both terms should be zero. The first term is zero when x(t)=asint+bcost; The second term is zero when y′=x, i.e. y=−acost+bsint. Then x′2+y′2=1 implies a2+b2=1, so a=sinα,b=cosα for some α. So x(t)=cos(t−α),y(t)=sin(t−α).