Let $\{g(\cdot-k),k\in\mathbb Z\}$ be a Riesz basis, and let $\varphi\in L^2(\mathbb R)$ be a function defined by its Fourier transform $$\hat{\varphi}(\xi)=\frac{\hat{g}(\xi)}{\Gamma(\xi)},$$ where $$\Gamma(\xi)=\left(\sum_k |\hat{g}(\xi+2k\pi)|^2\right)^{1/2}$$ Then $\{\varphi(\cdot-k),\ k\in\mathbb Z\}$ is an orthonormal system.

The proof of this result uses Parseval's identity and the fact that the Fourier transform of $\varphi(x-k)$ is $e^{-ik\xi}\hat{\varphi}(\xi)$. Hence, after some steps, one arrives to: $$\frac{1}{2\pi}\int_0^{2\pi}e^{-i(k-l)\xi}d\xi=\delta_{kl}$$ A complete proof is located - for instance - in:

Härdle, Wolfgang, et al. Wavelets, approximation, and statistical applications. Vol. 129. Springer Science & Business Media, 2012.

Let us consider a "perturbed system" $\{g(\cdot-\lambda_k),\ \lambda_k\in\mathbb R, \ k\in\mathbb Z\}$ that is a Riesz basis for $|\lambda_k-k|\leq L<1$. Is it possible to derive an orthonormal system by this Riesz basis?

References and answers are welcome.

  • $\begingroup$ Mark, What do you mean by "derive"? Do you want a function $f$ such that $f(\cdot-\lambda_k)$ form an orthonormal basis with the same span as $(g(\cdot-\lambda_k))$? $\endgroup$ – user71040 Oct 1 '15 at 11:30
  • $\begingroup$ @user71040, Yes. Or a function f such that $f(\cdot−k)$ form an orthonormal basis with the same span as $(g(\cdot−\lambda_k))$. $\endgroup$ – Mark Oct 1 '15 at 19:58

This is an answer to the second version: a function $f$ such that $f(⋅−k)$ form an orthonormal basis with the same span as $(g(⋅−λ_k))$ exists only in very special cases. Assume it exists, then on the Fourier transform side span $(g-\lambda_k)$ is the space of functions of the form $\phi\cdot \hat f$ where $\phi$ is $2\pi$-periodic, so in particular we have $e^{i\lambda_k t}\hat g(t)=\phi_k(t)\hat f(t)$. Let $S$ be the support of $\hat g$. Clearly $\hat f$ is non zero on $S$, so on $S$ we have $e^{i\lambda_k t}=\phi_k(t)[\hat f(t)/\hat g(t)]$. Writing this for integers $k\neq s$ and dividing we see that $e^{i(\lambda_k-\lambda_s)t}$ is $2\pi$ periodic on $S$. If the measure of $S$ is $>2\pi$ there are two sets of positive measure $A$ and $B$ such that $A=2\pi \mu +B$ for some $\mu\in Z$. Thus for $t\in B$ we have $e^{i(\lambda_k-\lambda_s)(t+2\pi \mu)}=e^{i(\lambda_k-\lambda_s)t}$ which gives $e^{i(\lambda_k-\lambda_s)2\pi \mu}=1$ so $\lambda_k-\lambda_s$ is an integer for all $k,s\in Z$ which implies $\lambda_k=k+\delta$. If measure of $S$ is $\leq2\pi$ then such $f$ can exist only if $\sum_{s\in Z}I(t+2\pi s t)=1$ for $I$ an indicator function of $S$.

For the first version here is a counterexample. Let $g$ be the indicator function of the interval $[-1/2,1/2]$. Then $(g(x-k))$ is an orthonormal basis so a Riesz basis. Let $\lambda_k=k+\delta_k$ with $0<\delta_k<1/10$ and and linearly independent over rational numbers for $k$ odd and $\delta_{k}=0$ for $k$ even. It is a Riesz basis. We have $<g_j,g_s>$ equals $1$ if $j=s$ and equals $0$ if $|j-s|>1$. Otherwise it is small, so the Gram matrix is diagonally dominated, so invertible. Let $G$ denote span $(g(x-\lambda_j))$. If $f(x-\lambda_k))$ is orthogonal in $G$ all those translations are in the span so in particular $f\in G$. If $h\in G$ then $h=\sum_{j\in Z} x_j g_j$ with $(x_j)$ square summable. We calculate that $h$ is a piecewise constant function with jumps at points $k-\frac12+\delta_k$ or $k+\frac12+\delta_k$ for $k\in Z$. Thus if $x_0=s\pm\frac12+\delta_s$ with $s\neq 0$ is a jump of $f$ we see that $x_0+\lambda_\mu$ is in this set of jumps for each $\mu$. We check that this is impossible.

Thus there is no non-zero function $f\neq g$ in $G$ such that the whole sequence $(f(x-\lambda_j))$ is $G$.

  • $\begingroup$ In the first answer, can you discard the case where for exemple $S=(0,2\pi)\cup(8\pi,10\pi)$ and $\lambda_k=k,$ for k even, $\lambda_k=k+1/2$ for k odd? $\endgroup$ – user75485 Oct 5 '15 at 16:36
  • $\begingroup$ @Joseph. Thanks for the question. It points out that I was to optimistic. Clearly $e^{i(\lambda_k-\lambda_s)2\pi \mu}=1$ gives $\lambda_k-\lambda_s$ is an integer divided by $\mu$ which still gives a lot of space for counterexamples. $\endgroup$ – user71040 Oct 5 '15 at 19:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.