Typical rank and indscal dimensionality for symmetric three-way arrays of order \(I\times 2\times 2\) or \(I\times 3\times 3\) (Q1881083): Difference between revisions

The rank of a three-way array (over the real field) is defined as the smallest number of (real valued) rank-one three-way arrays (outer products of three vectors) that generate the array as their sum. The rank of a three-way array is also the smallest number of components that give a perfect fit in CANDECOMP/PARAFAC (CP) [cf. \textit{J. D. Carroll} and \textit{J. J. Chang}, Psychometrika 35, 283--319 (1970; Zbl 0202.19101)]. If, e. g., the three-way array \(\underline{\mathbf X}\) of order \(I \times J \times K\) is composed of \(I\) slices \({\mathbf X}_1, \dots,{\mathbf X}_I\) of order \(J \times K\) then a perfect fit in CP implies that there exist matrices \({\mathbf A}\) \((J \times R)\), \({\mathbf B}\) \((K \times R)\) and diagonal matrices \({\mathbf C}_1, \dots, {\mathbf C}_I\) of order \(R \times R\) such that, for \(i=1, \dots, I\), \({\mathbf X}_i=\)\textbf{AC}\(_i{\mathbf B}'\). The smallest value of \(R\) for which this equation can be solved is the three-way rank of the array \(\underline{\mathbf X}\). A peculiar feature of the rank of a three-way array is the distinction between the maximum rank an array may have, and its typical rank, i. e., the rank it has with positive probability. Recently, some advances have been made in the study of the typical rank of three-way arrays over the real field, but these results rely on the assumption of random sampling of the entire array from a continuous distribution. Symmetric arrays cannot be dealt with in this way. This paper offers a first examination of the typical rank of symmetric arrays. The paper is not confined only to the rank of symmetric arrays, but also deals with the smallest number of dimensions that suffices for a perfect fit in INDSCAL. INDSCAL is a program designed for the analysis of individual differences for two or more subjects. Firstly, the typical rank of symmetric arrays, consisting of \(I\) symmetric \(J \times J\) matrices \({\mathbf S}_1, \dots, {\mathbf S}_I\), where \(J=2\) or \(3\), is analyzed. Then, to relate the rank values that arise with positive probability to the smallest number of dimensions for a perfect fit in INDSCAL for the same values, the existence of CP solutions with \({\mathbf A}\) and \({\mathbf B}\) equal is examined. In all considered cases the existence of such solution is established which implies that the smallest number of dimensions equals to the rank for these cases. Next, typical values of the smallest number of dimensions for symmetric arrays are compared to typical rank values of asymmetric arrays of the same size. Also, CP solutions which inevitably have \({\mathbf A}\) and \({\mathbf B}\) equal are determined. Finally, it is shown that double centering has a strong rank-reducing effect on the symmetric array.