1: \begin{abstract}

2: Motivated by the need of

3: observers that are both robust to disturbances

4: and guarantee fast convergence to zero of the estimation error,

5: we propose an observer for linear time-invariant systems

6: with noisy output

7: that consists of the combination of

8: $N$ coupled observers over a connectivity graph.

9: At each

10: node of the graph, the output of these {\em interconnected observers}

11: is defined

12: as the average of the estimates obtained using

13: local information.

14: The convergence rate and the robustness to measurement

15: noise of the proposed observer's output are characterized

16: in terms of

17: $\cal KL$ bounds.

18: Several optimization problems are formulated

19: to design the proposed observer so as to satisfy

20: a given rate of convergence specification while minimizing

21: the $H_{\infty}$ gain from noise to estimates

22: or the size of the connectivity graph.

23: It is shown that that the interconnected observers

24: relax the well-known tradeoff between rate of convergence

25: and noise amplification, which is a property attributed to the

26: proposed innovation term that, over the graph, couples the estimates between

27: the individual observers.

28: Sufficient conditions involving information

29: of the plant only, assuring that

30: the estimate

31: obtained at each node of the graph

32: outperforms the one obtained with a single, standard

33: Luenberger observer are given.

34: The results are illustrated in

35: several examples throughout the paper.

36: \end{abstract}

37: