On optimal regularization methods for fractional differentiation (Q1301262)

This paper is devoted to Abel type integral equations of the following kind \[ (A_\beta u)(x) = \frac{1}{\Gamma(\beta)} \int_{-\infty}^x \frac{u(t) dt}{(x-t)^{1-\beta}} = f(x), \;x \in\mathbb{R},\tag{1} \] where \(\beta > 0\) is fixed, and \(f:\mathbb{R}\to\mathbb{R}\) belongs to the class of functions such that the corresponding solution \(u:\mathbb{R}\to\mathbb{R}\) of equation (1) exists and satisfies \( \|u \|_p \leq E \). Here, \( E > 0 \) is a real number, and \( \|\cdot \|_p \) denotes the standard Sobolev space of order \( p > 0 \). A basic purpose is the optimal reconstruction (with respect to the \( L^2(\mathbb{R}) \)-norm \( \|\cdot \|\)) of \( u \) from noisy data \( f^\delta \in L^2(\mathbb{R}) \) with \( \|f^\delta - f \|\leq \delta \). More precisely, it is shown in this paper that \[ \omega(\delta) = E^{\beta/(p+\beta)} \delta^{p/(p+\beta)} ( 1 + o(1)) \] holds as \( \delta \to 0 \). Here \( \omega(\delta) := \inf_{R:L^2(\mathbb{R})\to L^2(\mathbb{R})}\Delta(\delta,R) \) is the best possible worst case error, with \[ \Delta (\delta,R) := \sup\{ \|R f^\delta - u \||u:\mathbb{R}\to\mathbb{R}, \|u \|_p \leq E, f^\delta \in L^2(\mathbb{R}) \text{ with } \|f^\delta - A_\beta u \|\leq \delta \} \] denoting the worst case error of an algorithm \(R: L^2(\mathbb{R}) \to L^2(\mathbb{R})\). Moreover, specific a priori parameter choices for Tikhonov's regularization method and two spectral methods are presented that yield optimal methods attaining the optimal value \( \omega(\delta) \), respectively. These results are generalized to the situation where the reconstruction error \( R f^\delta - u \) is measured with respect to the norm \( \|\cdot \|_q \) for \( 0 \leq q \leq p \).