The average error of quadrature formulas for functions of bounded variation (Q2640814): Difference between revisions

Using a natural measure of Dubins-Freedman and Mauldin-Ulam on the space H of non-decreasing continuous maps of [0,1] onto itself a probability measure Q is introduced on the space \(BV=\{f: [0,1]\to {\mathbb{R}}| f\) continuous, \(f(0)=0\), Var(f)\(\leq 1\}\). For \(0\leq a_ 1<...<a_ n\leq 1\) let \(a=(a_ 1,...,a_ n)\) and let \(N_ a: BV\to {\mathbb{R}}^ n\) be defined by \(N_ a(f)=(f(a_ 1),...,f(a_ n))\). Let \(A_ n=\{\Phi \circ N_ a|\) \(a_ 1,...,a_ n\in [0,1]\), \(\Phi\) : \({\mathbb{R}}^ n\to {\mathbb{R}}\}\). For \(M\in A_ n\) define the average error by \[ \Delta_ Q(M)=\int^{*}(\int^{1}_{0}f(t)dt-M(f))^ 2dQ(f))^{1/2}, \] where \(\int^{*}\) denotes the upper integral and set \(e^ Q_ n=\inf \{\Delta_ Q(M)|\) \(M\in A_ n\}\). The main result of the paper states that there are constants \(c,C>0\) such that \[ c\cdot n^{-(\log 6/\log 4)}\leq e^ Q_ n\leq Cn^{-(\log 6/\log 4)} \] (where log 6/log 4\(=1,29248...)\). This contrasts the worst case error \[ e_ n:=\inf \{\sup_{f\in BV}| \quad \int^{1}_{0}f(t)dt- M8f)|:\;M\in A_ n\}=1/2n+1. \]