Spring 2015 - Problem 4

Fourier analysis, Hahn-Banach

Let fLloc1(R)f \in L^1_{\mathrm{loc}}\p{\R} be 2π2\pi-periodic. Show that linear combinations of the translates f(xa)f\p{x - a}, aRa \in \R, are dense in L1((0,2π))L^1\p{\p{0, 2\pi}} if and only if each Fourier coefficient of ff is 0\neq 0.

Solution.

"    \implies"

Suppose f^(n)=0\hat{f}\p{n} = 0 for some nn. Observe that for any linear combination h(x)=n=1Ncjf(xaj)h\p{x} = \sum_{n=1}^N c_j f\p{x - a_j},

h^(n)=n=1Ncjeinajf^(n)=0.\hat{h}\p{n} = \sum_{n=1}^N c_j e^{-ina_j} \hat{f}\p{n} = 0.

Hence, linear combinations of translates of ff cannot possibly be dense. Indeed, consider g(x)=einxg\p{x} = e^{inx} has g^(n)=1\hat{g}\p{n} = 1. Because the Fourier transform is a continuous function from L1C0(R)L^1 \to C_0\p{\R}, if {fk}k\set{f_k}_k is a sequence of linear combinations of translates of ff such that fkgf_k \to g in L1L^1, then we would havefk^g^\hat{f_k} \to \hat{g} uniformly. However, this is impossible, since fk^(n)g^(n)=1\abs{\hat{f_k}\p{n} - \hat{g}\p{n}} = 1 for all k1k \geq 1.

"    \impliedby"

Let M=span{f(xa)}aRM = \cl{\span\,\set{f\p{x - a}}_{a \in \R}}. Suppose that ML1((0,2π))M \neq L^1\p{\p{0, 2\pi}} and so there exists x0L1((0,2π))Mx_0 \in L^1\p{\p{0, 2\pi}} \setminus M. By Hahn-Banach, there exists a linear functional Lg(L1((0,2π)))L((0,2π))L_g \in \p{L^1\p{\p{0, 2\pi}}}^* \simeq L^\infty\p{\p{0, 2\pi}}, i.e., integration against gg, such that Lg(x0)0L_g\p{x_0} \neq 0 but LgM=0\res{L_g}{M} = 0. Hence,

h(a)=02πf(xa)g(x)dx=Lg(f(xa))=0,h\p{a} = \int_0^{2\pi} f\p{x - a}g\p{x} \,\diff{x} = L_g\p{f\p{x - a}} = 0,

for all aRa \in \R by construction. Notice that by Hölder's inequality, eiξxfgL1e^{-i\xi x}fg \in L^1. Hence, we may apply Fubini's theorem in the following:

h^(ξ)=02πeiξa02πf(ya)g(y)dyda=02π02πeiξaf(ya)g(y)dady=02πg(y)02πeiξ(ya)f(a)dady(aya, f is periodic)=02πeiξyg(y)02πei(ξ)af(a)dady=g^(ξ)f^(ξ).\begin{aligned} \hat{h}\p{\xi} &= \int_0^{2\pi} e^{-i\xi a} \int_0^{2\pi} f\p{y - a}g\p{y} \,\diff{y} \,\diff{a} \\ &= \int_0^{2\pi} \int_0^{2\pi} e^{-i\xi a} f\p{y - a}g\p{y} \,\diff{a} \,\diff{y} \\ &= -\int_0^{2\pi} g\p{y} \int_0^{2\pi} e^{-i\xi \p{y-a}} f\p{a} \,\diff{a} \,\diff{y} && \p{a \mapsto y - a,\ f \text{ is periodic}} \\ &= -\int_0^{2\pi} e^{-i\xi y} g\p{y} \int_0^{2\pi} e^{-i\p{-\xi} a} f\p{a} \,\diff{a} \,\diff{y} \\ &= -\hat{g}\p{\xi} \hat{f}\p{-\xi}. \end{aligned}

But h=0h = 0 almost everywhere, so h^=0\hat{h} = 0 everywhere. Thus,

g^(n)f^(n)=h^(n)=0    g^(n)=0-\hat{g}\p{n}\hat{f}\p{-n} = \hat{h}\p{n} = 0 \implies \hat{g}\p{n} = 0

for all nn, since f^(n)0\hat{f}\p{n} \neq 0. Since L((0,2π))L2((0,2π))L^\infty\p{\p{0, 2\pi}} \subseteq L^2\p{\p{0, 2\pi}}, we see that

g(x)=nZg^(n)einx=0,g\p{x} = \sum_{n \in \Z} \hat{g}\p{n} e^{inx} = 0,

which is impossible as gg was non-zero. Hence, M=L1((0,2π))M = L^1\p{\p{0, 2\pi}} to begin with.