Accuracy and stability of computing high-order derivatives of analytic functions by Cauchy integrals (Q626444)

The magic of complex analysis makes it possible to replace differentiation by integration: just take a circle of radius \(r\) around the point of interest and use the Cauchy formula. In theory, any \(r>0\) will do, as long as you stay within the domain of analyticity of the function. So, computing the integral by a quadrature formula (say, the trapezoidal rule) gives us a numerical algorithm for differentiation of an analytic function. But once we leave the analytic function theory and enter the realm of the machine arithmetics, everything changes. The apparently benign step of selecting \(r\) (``choose any'') becomes a crucial problem for accuracy and stability. For small \(r\) there is an important loss of accuracy due to cancellation errors in the Cauchy integral; large \(r\), close to the radius of convergence of the Taylor series, also yield some numerical stability problems. This problem becomes even more apparent for higher order derivatives. So, the right choice for \(r\) is somewhere in between, but what is the optimal? And how does this optimal choice depend on the order of the derivative and the features of the function? These are precisely the questions addressed in this nice paper. Among other results the author shows that there is indeed the best choice: a unique \(r\) that minimizes the loss of accuracy caused by round-off errors. For many classes of functions, but nor for all, this optimal radius allows to attain full accuracy (machine precision). The loss of accuracy of the problem of computing \(f^{(n)}(0)\) for a given a function \(f\) and natural \(n\) is modeled by the condition number \[ \kappa(n,r)=\frac{\int_0^{2\pi} |f(r e^{i\theta})| \, d\theta}{|\int_0^{2\pi} e^{-in\theta} f(r e^{i\theta}) \, d\theta|}\geq 1. \] Its behavior is connected with several fine results from complex analysis, discussed in the paper. Especially appealing are the examples presented in the text in a large amount, ranging from the exponential and power functions to a Fredholm determinant for the limiting gap probability of a typical random matrix ensemble (where the author has made significant contributions, see \textit{F.~Bornemann} [Math. Comput. 79, 871--915 (2010; Zbl 1208.65182)]). This is a recommended paper for those interested in a nice interplay of the complex analysis with scientific computation.