Wiener's lemma

In mathematics, Wiener's lemma is a well-known identity which relates the asymptotic behaviour of the Fourier coefficients of a Borel measure on the circle to its discrete part. This result admits an analogous statement for measures on the real line. It was first discovered by Norbert Wiener.

Definition

Consider the space M(𝕋) of all (finite) complex Borel measures on the unit circle 𝕋 and the space C(𝕋) of continuous functions on 𝕋 as its dual space. Then C(𝕋)⊂Lp(𝕋) for all 1≤p<∞ and L1(𝕋)⊂M(𝕋).

Given μ∈M(𝕋), let μpp=∑jcjδzj, be its discrete part (meaning that μ({zj})=cj≠0 and μ({z})=0 for z∉{zj}. Then limN→∞12N+1∑n=−NN|μ^(n)|2=∑j|cj|2, where μ^(n)=∫𝕋z−ndμ(z) is the n-th Fourier-Stieltjes coefficient of μ.

Similarly, on the real line ℝ, the space M(ℝ) is the dual space of C0(ℝ), continuous functions which vanish at infinity, and C0(ℝ)⊂Lp(ℝ) for all 1≤p≤∞.

Given μ∈M(ℝ), let μpp=∑jcjδxj, its discrete part. Then limR→∞12R∫−RR|μ^(ξ)|2dξ=∑j|cj|2, where μ^(ξ)=∫ℝe−2πiξxdμ(x) is the Fourier-Stieltjes transform of μ.

Consequences

If μ∈M(𝕋) is continuous, then limN→∞12N+1∑n=−NN|μ^(n)|2=0. Furthermore, μ̂ tends to zero if μ is absolutely continuous. Equivalently, μ is absolutely continuous if its Fourier-Stieltjes sequence belongs to the sequence space ℓ2. That is, if μ places no mass on the sets of Lebesgue measure zero (i.e. μpp=0), then μ̂→0 as |N|→∞. Conversely, if μ̂→0 as |N|→∞, then μ places no mass on the countable sets.

A probability measure μ on the circle is a Dirac mass if and only if limN→∞12N+1∑n=−NN|μ^(n)|2=1. Here, the nontrivial implication follows from the fact that the weights cj are positive and satisfy 1=∑jcj2≤∑jcj≤1, which forces cj2=cj and thus cj=1, so that there must be a single atom with mass 1.

Proof

12N+1∑n=−NNν^(n)=∫𝕋fN(z)dν(z),

with fN(z)=12N+1∑n=−NNz−n. The function fN is bounded by 1 in absolute value and has fN(1)=1, while fN(z)=zN+1−z−N(2N+1)(z−1) for z∈𝕋∖{1}, which converges to 0 as N→∞. Hence, by the dominated convergence theorem,

limN→∞12N+1∑n=−NNν^(n)=∫𝕋1{1}(z)dν(z)=ν({1}).

We now take μ′ to be the pushforward of μ‾ under the inverse map on 𝕋, namely μ′(B)=μ(B−1)‾ for any Borel set B⊆𝕋. This complex measure has Fourier coefficients μ′^(n)=μ^(n)‾. We are going to apply the above to the convolution between μ and μ′, namely we choose ν=μ∗μ′, meaning that ν is the pushforward of the measure μ×μ′ (on 𝕋×𝕋) under the product map ⋅:𝕋×𝕋→𝕋. By Fubini's theorem

ν^(n)=∫𝕋×𝕋(zw)−nd(μ×μ′)(z,w)=∫𝕋∫𝕋z−nw−ndμ′(w)dμ(z)=μ^(n)μ′^(n)=|μ^(n)|2.

So, by the identity derived earlier, limN→∞12N+1∑n=−NN|μ^(n)|2=ν({1})=∫𝕋×𝕋1{zw=1}d(μ×μ′)(z,w). By Fubini's theorem again, the right-hand side equals

∫𝕋μ′({z−1})dμ(z)=∫𝕋μ({z})‾dμ(z)=∑j|μ({zj})|2=∑j|cj|2.
12R∫−RRν^(ξ)dξ=∫ℝfR(x)dν(x)

(which follows from Fubini's theorem), where fR(x)=12R∫−RRe−2πiξxdξ. We observe that |fR|≤1, fR(0)=1 and fR(x)=e2πiRx−e−2πiRx4πiRx for x≠0, which converges to 0 as R→∞. So, by dominated convergence, we have the analogous identity

limR→∞12R∫−RRν^(ξ)dξ=ν({0}).

See also