Long-range correlations between letters and sentences in texts

We mapped three long texts to random walks and calculated several correlation measures as Hölder exponents, higher-order cumulants and power spectra. By means of computer experiments we have found that shuffling on/or below the sentence level generates strings showing no anomalous diffusion, no higher-order cumulants and no power spectra with 1/fδ-shape. In this way we have shown that the long correlations reflected in these measures are not based on correlations inside sentences but reflect the large-scale structure beyond the sentence level.