On stable parameter estimation and forecasting in epidemiology by the Levenberg-Marquardt algorithm with Broyden's rank-one updates for the Jacobian operator (Q2008270)

This paper studies an inverse problem in the form of nonlinear least squares between Hilbert spaces \(X\) and \(Y\) as \(\min\|F(x)\|^2\) with \(F:D_F\subset X\rightarrow Y\), where \(F\) is Fréchet differentiable. A regularized Levenberg-Marquardt procedure with rank-one Broyden's updates is proposed for the Fréchet derivative operator. Namely, \(x_{n+1}=P_n\{x_n-[J_n^*J_n+\tau I]^{-1}J_n^*F(x_n)\}\) and \(J_{n+1}=J_n+\frac{(s_n,\cdot)}{\|s_n\|^2}(y_n-J_ns_n)\), where \(s_n=x_{n+1}-x_n\), \(y_n=F(x_{n+1})-F(x_n)\), \(\tau\ge0\), and \(P_n\) is the orthogonal projection operator onto the subspace spanned by the first some eigenfunctions of \(J_n^*J_n\). Assume that \(\|(I-P_n)\hat{x}\|\le\delta\) for some \(0\le\delta\ll R\), where \(\hat{x}\) satisfies \(\|F(\hat{x})\|=0\). Under some regularity conditions, if the iterations are terminated at \(n=N(\delta)\), the first time the stopping criterion \(q^nR\le\frac{\delta}{1-q}\) is fulfilled with \(N(\delta)=[\ln(\delta/R(1-q))/\ln q]\), then it is shown that \(\|x_{N(\delta)}-\hat{x}\|\le 2\delta/(1-q)\). Some simulations are presented to illustrate the results.