The Walsh-Fourier transform on the real line (Q2676602)

The Walsh functions map the unit interval to the set \(\{-1,1\}\). They can be extended to the set of nonnegative reals, but not to the whole real line. One of the goals of this article is to give a Walsh-like orthonormal and complete function system that can be extended on the real line and therefore be able to extend the Walsh-Fourier transform and the Walsh-Fourier series. Let \(\mathcal{F}\) be the set of double infinite sequences \[ x=x_n\ (n \in \mathbb{Z}), \] where \(x_n=0,1\) or \(-1\) and \(x_n \rightarrow 0\) as \(n \rightarrow-\infty\). Thus, to each \(x \in \mathcal{F}\) with \(x \neq 0\) there corresponds an integer \(s(x) \in \mathbb{Z}\) such that \[ x_{s(x)} \neq 0 \text { but } x_n=0 \text { for } n<s(x). \] Let \(x=x_n\) \((n \in \mathbb{Z})\) and \(y=y_n\) \((n \in \mathbb{Z})\) be elements of \(\mathcal{F}\). Define the sum of \(x\) and \(y\) by \[ x+y=x_n+y_n \bmod 3\ (n \in \mathbb{Z}). \] Notice that \((\mathcal{F},+)\) is an abelian group. Moreover, \(\mathcal{F}\) is a normed algebra, indeed, defining the product of \(x\) and \(y\) by \[ x\cdot y=x_n\cdot y_n\ (n\in \mathbb{Z}) \] and for \(x=x_n\), \(n \in \mathbb{Z} \in \mathcal{F}\) define the norm \[ |x|:=\sum_{n \in \mathbb{Z}} \frac{\left|x_n\right|}{3^{n+1}}. \] A character on \(\mathcal{F}\) is a continuous complex-valued map which satisfies \((x, y \in \mathcal{F})\) \[ \varphi(x+y)=\varphi(x) \varphi(y)\text { and }|\varphi(x)|=1. \] It is evident that \(\varphi(0)=1\). Let \(e_j\) denote the element \(x=x_n\), \(n \in \mathbb{Z}\) which satisfies \(x_j=1\) for some \(j \in \mathbb{Z}\) and \(x_n=0\) for \(n \neq j\), \(n \in \mathbb{Z}\). Since \(\varphi\) is continuous on \(\mathcal{F}\) and \(e_j \rightarrow 0\) in \(\mathcal{F}\) as \(j \rightarrow \infty\), we have \[ \varphi\left(e_j\right) \rightarrow \varphi(0)=1 \] as \(j \rightarrow \infty\). On the other hand, \(e_j+e_j+e_j=0\). Hence, \[ 1=\varphi(0)=\varphi\left(e_j+e_j+e_j\right)=\varphi^3\left(e_j\right), \] \[ \varphi\left(e_j\right)=\sqrt[3]{1}=e^{\frac{2 \pi i k}{3}},\quad k=-1,0,1. \] Consequently, there is a sequence \(y=\left\{y_j: j \in \mathbb{Z}, y_j=-1,0,1\right\}\) such that for every \(-j-1>s(y)\) and \[ \varphi_y\left(e_j\right)=e^{\frac{2}{3} \pi i y_{-j-1}}. \] Moreover, \[ \varphi\left(x_j \cdot e_j\right)=\left(\varphi\left(e_j\right)\right)^{x_j}. \] Then, from the continuity of \(\varphi\), we can write \[ \varphi_y(x)=\prod_{j \in \mathbb{Z}}\left(\varphi\left(e_j\right)\right)^{x_j}=e^{\frac{2}{3} \pi i \sum_{j \in \mathbb{Z}} y_{-j-1 x_j}} \quad(x, y \in \mathcal{F}). \] All characters of the additive group \((\mathcal{F},+)\) can be written as \(\varphi_y\) \(y \in \mathcal{F}\). For a given \(f \in L_1(\mathcal{F})\) the Walsh-Fourier transform of \(f\) is the function on \(\mathcal{F}\) will be defined by \[ \widehat{f}(y):=\frac{1}{9} \int_{\mathcal{F}} f(x) \bar{\varphi}_y(x) d \mu(x) \quad(y \in \mathcal{F}). \] It is quite well-known that a classical Walsh function maps from the unit interval to the set \(\{-1,1\}\) and also that it can be extended to the set of nonnegative real numbers, but cannot be extended to the whole set of reals. Therefore, to get these notions to the real line, something different must be done. The authors introduce Walsh-like functions on the real line as follows. For every real number \(y \in \mathbb{R}\) can be decomposed in its triadic decomposition, i.e. it can be expressed as \[ y=\sum_{k=-\infty}^{+\infty} \frac{y_k}{3^k}, \] where \(y_k \in\{-1,0,1\}\) for all \(k \in \mathbb{N}\). The digits -1, 0, 1 mean that we are going to the left or getting nowhere, or we are going to the right by \(1 / 3^k\). There is no convergence problems, since \(y\) is a finite real, and, consequently, \(y_k=0\) for \(k\)'s small enough \(\left(\lim _{k \rightarrow-\infty} y_k=0\right.\)). We can identify \(y\) as the sequence \(y_k\), \(k \in \mathbb{Z}\). Unfortunately, this identification is not always a bijection, since for all \(j \in \mathbb{N}\) the numbers \( \frac{6 j+1}{2 \times 3^{n+1}}\) have two corresponding -1, 0, 1 sequences. These numbers are called triadic rationals. In this situation, one can choose the one that terminates in -1's, for instance, since the set of these numbers is countable \((j, n \in \mathbb{N})\) this is of minor importance. Define the addition \(\oplus:\{-1,0,1\}^2 \rightarrow\{-1,0,1\}\) as the mod 3 addition. Then, define the addition \(\oplus\) on \(\mathbb{R}\) as \(x \oplus y:=x_k \oplus y_k\), \(k \in \mathbb{Z}\). The inverse operation is denoted by \(\ominus\). The set of triadic intervals on \(\mathbb{R}\) is defined as follows. Set \(n \in \mathbb{Z}\), and \(t \in \mathbb{R}\), then the set \[ I_n(t):=\left\{y \in \mathbb{R}: y_i=t_i \text { for } i \leqslant n\right\} \] is called a triadic interval. We also use the notation \(I_{-\infty}(t)=\mathbb{R}\). Given, \[ t_{(n)}=\sum_{l=-\infty}^n t_l / 3^l\text{ and }k=\sum_{l=-\infty}^n t_l 3^{n-l} \] then \[ I_n(t)=\left[t_{(n)}-\frac{1}{2 \times 3^n}, t_{(n)}+\frac{1}{2 \times 3^n}\right):=\left[\frac{2 k-1}{2 \times 3^n}, \frac{2 k+1}{2 \times 3^n}\right). \] Observe that the Lebesgue measure of \(I_n(t)\) is \(3^{-n}\) for all \(n \in \mathbb{Z}\) and \(t \in \mathbb{R}\). Then, the authors introduce the generalized Walsh functions. Let \(x, y \in \mathbb{R}\) and \[ \omega(x, y):=r^{\sum_{k=-\infty}^{\infty} x_k y_{-k-1}},\quad r=\exp \left(2 \pi \imath \frac{1}{3}\right), \quad \imath=\sqrt{-1} \] and it is easy to see that the Walsh functions satisfy the following properties: \[ |\omega|=1, \quad \omega(x, y)=\omega(y, x), \quad \omega(x \oplus z, y)=\omega(x, y) \omega(z, y) \] for \(x, y, z \in \mathbb{R}\), and \(x \oplus z\) a triadic irrational. The Walsh-Fourier transform of an \(f \in L_1(\mathbb{R})\) is defined by \[ \widehat{f}(y):=\frac{1}{9} \int_{\mathbb{R}} f(x) \bar{\omega}(x, y) d x \quad(y \in \mathbb{R}) . \] For each \(f \in L_1(\mathbb{R})\) and \(t>0\) the Walsh-Dirichlet integral is defined by \[ S_t(x ; f):=\int_{-t}^t \widehat{f}(y) \omega(x, y) d y . \] By Fubin's theorem, we have \[ S_t(x ; f)=\int_{-t}^t \frac{1}{9}\left(\int_{\mathbb{R}} f(u) \bar{\omega}(u, y) d u\right) \omega(x, y) d y = \frac{1}{9} \int_{\mathbb{R}} f(u) D_t(u \ominus x) d u, \] where \[ D_t(x)=\int_{-t}^t \omega(x, y) d y \] for \(t \in \mathbb{R}_{+}\) and \(f \in L_1(\mathbb{R})\). Then, they prove the following representation for the Walsh-Dirichlet integral. Theorem 3.1. Let \(N \in \mathbb{Z}\). Then \[ D_{\frac{3^N}{2}}(x)=3^N \mathbb{I}_{I_{N-2}(0)}(x), \] where \(\mathbb{I}_E(x)\) is the characteristic function of the set \(E\). Now, they prove an inversion result for the Walsh-Fourier transform. Theorem 3.2. Let \(f \in L_1(\mathbb{R})\) be \(W\)-continuous on \(\mathbb{R}\). If \(\widehat{f} \in L_1(\mathbb{R})\), then \[ f(y)=\int_{\mathbb{R}} \widehat{f}(x) \omega(x, y) d x . \] On the other hand, with the generalized Walsh functions we can obtain an orthonormal and complete system in \(L_2\left(I_n(0)\right)\). Theorem 4.1. Let \(n \in \mathbb{Z}\). Then the system \[ \left\{3^{n / 2} \omega\left(x, 3^{n+2} j\right), j \in \mathbb{Z}\right\} \] is orthonormal and complete in \(L_2\left(I_n(0)\right)\). The Fourier coefficients of the integrable function \(f: I_n(0) \rightarrow \mathbb{C}\) are defined as usual, as \[ \hat{f}(j):=\int_{I_n(0)} f(x) 3^{n / 2} \bar{\omega}\left(x, 3^{n+2} j\right) d x, \] where \(j \in \mathbb{Z}\) and then the Walsh-Fourier series of \(f\) is given by \[ \sum_{j \in \mathbb{Z}} \hat{f}(j) 2^{n / 2} \omega\left(y, 3^{n+2} j\right),\] and its \(N\)-partial sums is defined as \[ S_N f(y) :=\sum_{\{j \in \mathbb{Z}:|j|<N\}} \hat{f}(j) 2^{n / 2} \omega\left(y, 3^{n+2} j\right). \] Consequently, \[ S_{\frac{3^N+1}{2}} f(y)=3^{n+N} \int_{I_{n+N}(y)} f(x) d x. \] The Dirichlet kernels with respect to the system \(\left\{3^{n / 2} \omega\left(x, 3^{n+1} j\right), j \in \mathbb{Z}\right\}\) are defined as: \[ \mathcal{D}_N(x):=\sum_{\{j \in \mathbb{Z}:|j|<N\}} 3^n \omega\left(x, 3^{n+2} j\right) \quad N \in \mathbb{N},\ n \in \mathbb{Z},\ x \in \mathbb{R}. \] The formula for the Dirichlet kernels \(\mathcal{D}_{\frac{3^N+1}{2}}\) are then given. Theorem 4.2. Let \(x \in I_n(0)\) and \(N \in \mathbb{N}\). Then \[ \mathcal{D}_{\frac{3 N+1}{2}}(x)=\left\{\begin{array}{l} 3^{n+N}, \text { if } x \in I_{n+N}(0) \\ 0, \text { if } x \notin I_{n+N}(0) \end{array}. \right. \] As usual, the integral representation of the partial sums is given in terms of \( \mathcal{D}_N\) \[ S_N f(y) =\int_{I_n(0)} f(x) \mathcal{D}_N(y \ominus x) d x .\] Then they discuss the uniform convergence of these partial sums of continuous functions. Denote the (triadic) modulus of continuity of the function \(f: I_n(0) \rightarrow \mathbb{C}\) by \[ w(N, f)= w\left(I_n(0), N, f\right):=\sup _{h \in I_{n+N}, x \in I_n(0)}|f(x)-f(x \oplus h)| \quad(N \in \mathbb{N}). \] It is not difficult to prove that a function \(f\) on \(I_n(0)\) is continuous if and only if its modulus of continuity converges to zero and moreover, \[\left|S_{\frac{3^N+1}{2}} f(y)-f(y)\right| \leqslant\left|3^{n+N} \int_{I_{n+N}(0)} f(y \oplus h)-f(y) d h\right| \leqslant w(n+N, f),\] thus, uniform convergence follows. Finally, they consider the Walsh-Fourier transform in \(L_p\) \((1<p \leqslant 2)\). They prove that in cases when \(f \in L_p(\mathbb{R})\) \((1<p \leqslant 2)\) the Walsh-Fourier transform is the limit of truncated Walsh-Fourier transforms. Theorem 5.1. Let \(f \in L_p(\mathbb{R})\) \((1<p \leqslant 2)\) and \(\frac{1}{p}+\frac{1}{q}=1\). Then \[ \widehat{f}(x, a):=\int_{-a}^a f(y) \omega(x, y) d y \] converges in \(L_p\) norm. Moreover, \[ \int_{\mathbb{R}}\left|\int_{\mathbb{R}} f(y) \omega(x, y) d y\right|^q d x \leqslant 9\left(\int_{\mathbb{R}}|f(x)|^p d x\right)^{\frac{1}{p-1}}. \] As open questions, the authors list the norm and pointwise convergence of the partial sums \(S_N\), and the Fejér kernels and Césaro means.\\