1: \begin{abstract}

2: Given a large square matrix $A$ and a sufficiently regular function $f$ so that $f(A)$ is

3: well defined,

4: we are interested in the approximation of the leading singular values and corresponding

5: singular vectors of $f(A)$, and in particular of $\|f(A)\|$, where $\|\cdot \|$ is the matrix norm

6: induced by the Euclidean vector norm. Since neither $f(A)$ nor $f(A)v$ can

7: be computed exactly, we introduce and analyze

8: an {\it inexact} Golub-Kahan-Lanczos bidiagonalization procedure, where

9: the inexactness is related to the inaccuracy of the operations $f(A)v$, $f(A)^*v$.

10: Particular outer and

11: inner stopping criteria are devised so as to cope with the lack of a true residual.

12: Numerical experiments with the new algorithm on typical application problems are reported.

13: %

14: %, assuming that both the computation of $f(A)$ itself and the multiplication of $f(A)$ with a vector are not feasible.  Therefore, we provide an inexact bidiagonalization procedure to approximate $\vectornorm{f(A)}_2$. The method is inexact since in the inner iterations it approximates the multiplication of $f(A)$ with a vector to a certain accuracy. We will first show that to obtain convergence, a specific tolerance for the inner iteration is needed. But it turns out that this accuracy can be relaxed as the number of iterations increases, leading to a variable inner accuracy, and that we still obtain convergence.

15: \end{abstract}