On the characterization of totally nonpositive matrices (Q2361627): Difference between revisions

The paper deals with bidiagonal factorizations of a nonsingular totally nonpositive or totally negative matrix. A nonpositive real matrix \(A\) is said to be totally nonpositive (negative) if all its minors are nonpositive (negative) and it is abbreviated as t.n.p. (t.n.). In this manuscript, a bidiagonal factorization of a nonsingular totally nonpositive matrix \(A\), of size \(n \times n\), is computed and it is stored in a matrix represented by \(\mathcal{BD}_{(\mathrm{t.n.p.})}(A)\) when \(a_{11}<0\) and \(a_{nn} \leq 0\) (or \(\mathcal{BD}_{(\mathrm{zero})}(A)\) when \(a_{11}=0\)), \[ \mathcal{BD}_{(\mathrm{t.n.p.})}(A)= \begin{pmatrix} -d_1 & \alpha_{1,2} & \alpha_{1,3} & \dots & \alpha_{1,n-1} & \alpha_{1,n} \\ \beta_{1,2} & d_2 & \alpha_{2,3} & \ldots & \alpha_{2,n-1} & \alpha_{2,n} \\ \beta_{1,3} & \beta_{2,3} & d_3 & \dots & \alpha_{3,n-1} & \alpha_{3,n} \\ \vdots & \vdots & \vdots & & \vdots & \vdots \\ \beta_{1,n-1} & \beta_{2,n-1} & \beta_{3,n-1} & \dots & d_{n-1} & \alpha_{n-1,n} \\ \beta_{1,n} & \beta_{2,n} & \beta_{3,n} & \dots & \beta_{n-1,n} & d_n \end{pmatrix}, \] where \[ \begin{aligned} d_i > 0, \;\beta_{i,j} \geq 0, \;\alpha_{i,j} \geq 0, \;\text{for} \;i,j=1,2,\dots,n, \\ \beta_{1,j} > 0, \;\alpha_{1,j} > 0, \;\text{for } \;j=2,3,\dots,n, \\ \text{if} \;\beta{i,j}=0, \;(\alpha_{i,j}=0) \;\Rightarrow \;\beta{i,h}=0, \;(\alpha_{i,h}=0) \;\text{for all} \;h>j. \end{aligned} \] As a converse result, the authors present an efficient algorithm to know if a matrix \(\mathcal{BD}_{(\mathrm{t.n.p.})}(A)\) (\(\mathcal{BD}_{(\mathrm{zero})}(A)\)) is the bidiagonal factorization of a t.n.p. matrix with \(a_{11}<0\) (\(a_{11}=0\)). Similar results are obtained for t.n. matrices with \(a_{11}=0\), using the matrix \(\mathcal{BD}_{(\mathrm{t.n.} )}(A)\). These characterizations are extended to rectangular t.n.p. (t.n.) matrices. Finally, the bidiagonal factorization of the inverse of a nonsingular t.n.p. (t.n.) matrix \(A\) is directly obtained from \(\mathcal{BD}_{(\mathrm{t.n.p.})}(A)\) or \(\mathcal{BD}_{(\mathrm{zero})}(A)\) (\(\mathcal{BD}_{(\mathrm{t.n.})}(A)\)).