Biproportional scaling of matrices and the iterative proportional fitting procedure (Q744690): Difference between revisions

Given are \(A=[a_{ij}]\in\mathbb{R}^{k\times l}\), \(a_{ij}\geq0\), \(r\in\mathbb{R}^k\), \(r_i>0\), and \(s\in\mathbb{R}^l\), \(s_j>0\). When an subscript letter is replaced by \(*\), it means that we sum over that index. Then the iterative proportional fitting (IPF) procedure generates a sequence of scaled matrices \(A(t)\), \(t=0,1,\dots\), by alternatingly rescaling even rows to satisfy \(a_{i*}(t)=r_i\) and odd columns so that \(a_{j*}(t)=s_j\). The idea is to minimize the \(L_1\)-norm \(\sum_i|a_{i*}(t)-r_i|+\sum_j|a_{*j}(t)-s_j|\). If a solution exists that fits all row and column sums (error is zero), the IPF procedure will converge. A short proof is given of this result, relying on properties for connected matrices \(A\) for which all solutions with zero error maintain the zero structure of \(A\). This paper appeared in a special issue of the journal and another paper in that issue discusses a similar problem (cf. [\textit{E. Aas}, ibid. 215, 15--23 (2014; Zbl 1302.65112)]).