The Mdl-principle in Testing Style Homogeneity between Literary Texts

We study a new context-free computationally simple stylometrybased attributor: the sliced conditional compression complexity (sCCC or simply CCC) of literary texts introduced in [8] and inspired by the incomputable Kolmogorov conditional complexity (KC). Other stylometry tools can occasionally almost coincide for different authors. Our CCCattributor is asymptotically strictly minimal for the true author, if the query texts are sufficiently large but much less than the training texts, universal compressor is good and sampling bias is avoided. This classifier simplifies the [13] homogeneity test (partly based on compression) under insignificant difference of unconditional complexities of training and query texts verifiable via its asymptotic normality [15] for IID and Markov sources and normal plots for real literary texts. It is consistent under large text approximation as a stationary ergodic sequence due to the lower bound for the minimax compression redundancy of piecewise stationary strings [11] (see also our elementary combinatorial arguments and simulation in [9] for IID sources). The CCC is based on the t-ratio measuring how many standard deviations are in the mean difference of slices’ CCC which enables evaluation of the corresponding P-value of statistical significance based on slices’ CCC asymptotic normality empirically verified by their normal plots in all cases studied and expected to be proved soon for simplified statistical models of literary texts. The asymptotic CCC study is complemented by many literary case studies (see [8,9, 10]).