1: \begin{abstract}

2: PageRank is a ranking of the webpages used by the search engine Google to determine in which order to display webpages.  It measures how often a given webpage would be visited by a random surfer on the webgraph.  It seems realistic that this ranking may influence the reputation of the webpages, and therefore the behavior of websurfers.  We propose a simple model taking in account the mutual influence between webranking and websurfing.

3: 

4: We study this model by considering the following iteration on the set of stochastic row vectors: $\vec r(s+1)=\vec u_T(\vec r(s))$, where for all vector $\vec x$, $\vec u_T(\vec x)$ is the unique invariant measure of a primitive stochastic matrix $M_T(\vec x)$ constructed from the adjacency matrix of the webgraph.  The parameter $T>0$ is fixed and measures the confidence of the surfer in the ranking.

5: We call $T$-PageRank the limit, if it exists, of these iterates when $s$ tends to infinity.

6: We also consider the simple iteration defined by $ \vec{\tilde r}(s+1)= \vec{\tilde r}(s)M_T(\vec{\tilde r}(s))$.

7: 

8: We prove that, when $T$ is large enough, the fixed point of these iterations is unique and the convergence is global on the domain. But for small values of $T$, at least when the matrix $C$ is positive, there are always several fixed points.

9: 

10: Our analysis uses results of nonlinear Perron--Frobenius theory, Hilbert projective metric and Birkhoff's coefficient of ergodicity.

11: \end{abstract}

12: