This paper presents a robust information filter which inherits the structural simplicity of the information filter and the robustness property of the H-infinity filter with respect to noise statistics. In this filter, an assurance level on the noise bound is reflected in the newly defined covariance matrix. It provides robustness against uncertainty in the noise model by sacrificing performance in the minimum variance sense. All these are achieved while retaining the simple structure of the standard information filter. Thus, it is able to realize robust decentralized estimation with less communication and computational load while without the need to model the system noise accurately. (c) 2005 Elsevier Ltd. All rights reserved.