Lower estimation of the difference between quasi-arithmetic means (Q1692279)

Let \(U\) be an arbitrary interval and \(f:U \rightarrow \mathbb{R}\) a continuous, strictly monotone function. If \(a=(a_1, \ldots ,a_n)\) is an arbitrary sequence of points in \(U\) and \(w=(w_1, \ldots ,w_n)\) a sequence of corresponding weights (\(w_i >0\), \(\sum w_i=1\)), then the quasi-arithmetic mean \(A^{[f]}(a,w)\) is defined by \[ A^{[f]}(a,w) := f^{-1} \left( \sum_{i=1}^n w_i f(a_i) \right). \] In the 1960s, a metric among quasi-arithmetic means was introduced. Namely, if \(f\) and \(g\) are both continuous, strictly monotone and have the same domain, then one can define the distance \[ \rho(A^{[f]}, A^{[g]}) := \sup \{ |A^{[f]}(a,w)-A^{[g]}(a,w)| : a \text{ and } w \text{ admissible}\}. \] For a twice differentiable function \(f:U \rightarrow \mathbb{R}\) having a nowhere vanishing first derivative (we denote such a family of functions by \(S(U)\)) let us define an operator \(P_f:= \frac{f''}{f'}\). Also, for \(K>0\) we denote \(S_K(U):= \{f\in S(U) : \|P_f\|_{\infty} \leq K \} \). The main theorem is the following result. Theorem 1. Let \(f,g \in S_K(U)\). If \(\varepsilon := \|P_f -P_g\|_*\), then \[ \rho(A^{[f]}, A^{[g]}) \geq \frac{\varepsilon (e^{\varepsilon/4}-1)(e^{\varepsilon/6}-1)}{16K \exp (\max(\|P_f\|_*, \|P_g\|_*))(e^{K|U|}-1)}, \] where \(\|f\|_*:= \sup_{x,y \in dom f} \left| \int_x^y f(t) dt \right|\). In fact, the main theorem states that \(\rho(A^{[f]}, A^{[g]})\) can be estimated from below by some term involving \(\|P_f -P_g\|_*\), the lenght of the interval and the number \(K\). In the second theorem a \((\phi,K,\delta)\)-separation occurs. Namely, we say that \(f,g\in S(U)\) are \((\phi,K,\delta) \)-separated if there exists a closed interval \(V\subset U\), \(|V|=\phi\) such that for all \(x\in V\) the following inequalities are satisfied: \(|P_f(x)| \leq K\), \(|P_g(x)| \leq K\), and \(|P_f(x)-P_g(x)| \geq \delta\). Theorem 2. If \(f,g\in S(U)\) are \((\phi,K,\delta) \)-separated, then \[ \rho(A^{[f]}, A^{[g]}) \geq \frac 1K \ln (1+K\alpha), \] where \[ \alpha:= \frac{e^{-K\phi /2}-1}{K} -\frac{e^{(\delta-K)\phi /2}-1}{K-\delta}. \] The following corollary is also given. Corollary 3. If \(f,g\in S(U)\) are \((\phi,K,\delta)\)-separated, then \[ \rho(A^{[f]}, A^{[g]}) \geq \frac 1K \ln \left( 1+\frac{\delta}{K}\cdot \Theta(\frac{K\phi}{2})\right), \] where \(\Theta(x):= 1-e^{-x}-xe^{-x}\).