Numerical methods for a Volterra integral equation with nonsmooth solutions (Q818196)

The authors consider the numerical treatment of a singular Volterra integral equation of the form \(u(t)=\int_0^t \frac{s^{\mu -1}}{t^\mu}u(s)\,ds+g(t)\). This equation has an infinite set of solutions, one of which is smooth and all others have infinite gradient at the origin. Building on their earlier work, in which an approximation to the smooth solution was obtained, they obtain approximations to any of the infinite class of solutions. The method adopted involves choosing a fixed real number \(\varepsilon\) \((> 0)\) and splitting the integral into two terms (\(\int_0^t=\int_0^\varepsilon+\int_\varepsilon ^t\)), leading to an equation with a kernel which is regular in \(\{(t,s): 0\leq s \leq t \leq T-\varepsilon \}\). Any standard numerical scheme for regular second kind integral equations can then be applied. Initially they consider the case when \(I_\varepsilon= \int_0^\varepsilon s^{\mu-1}u(s)\,ds\) is known exactly. Euler's method and the trapezoidal method are then applied, along with suitable approximations for \(u(t)\). The authors state that for the case \(\varepsilon=0\) a convergence proof is not yet available but they make a conjecture (based on the results of numerical experiments) about the convergence order of the trapezoidal method under stated conditions on \(g(t)\). Theorems concerning the convergence of the two methods for the case \(\epsilon \neq 0\) are then presented and proved. Varying \(\varepsilon\) is then considered and a theorem giving an upper bound for the error in the approximate solution obtained through use of the product Euler method is presented and proved. For the case when \(I_\varepsilon\) is only known approximately the authors estimate how this will affect the error in the solution. The paper concludes with illustrative numerical examples.