The resolvent average for positive semidefinite matrices (Q2267407): Difference between revisions

Given positive semidefinite matrices \(A_1,\dots,A_n\) of the same size, and \(\lambda_1,\dots,\lambda_n\in (0,\infty)\) with \(\sum_{i=1}^n \lambda_i=1\), the authors define the resolvent average to be \[ R_\mu(A,\lambda):= \big( \lambda_1 (A_1+\mu^{-1})^{-1} + \cdots + \lambda_n (A_n+\mu^{-1})^{-1} \big)^{-1} - \mu^{-1}. \] Note \(R_1(A,\lambda)\) is a weighted average of resolvents of the \(A_i\). The authors study this notion and compare it to the well-known harmonic and arithmetic average defined as \( H(A,\lambda)= \big( \lambda_1 A_1^{-1} + \cdots + \lambda_n A_n^{-1} \big)^{-1}\) and \( A(A,\lambda)= \lambda_1 A_1 + \cdots + \lambda_n A_n\), respectively. As a sample of the results obtained, we mention \[ H(A,\lambda) \preceq R_\mu(A,\lambda) \preceq A(A,\lambda), \] where \(\preceq\) stands for the Löwner partial ordering, and \[ \lim_{\mu\searrow 0} R_\mu(A,\lambda) = A(A,\lambda), \qquad \lim_{\mu\to\infty} R_\mu(A,\lambda) = H(A,\lambda). \] The paper is well-written, and the proofs are based on convexity arguments.