1: \begin{abstract}

2: Independent component analysis (ICA) aims at decomposing an observed

3: random vector into statistically independent

4: variables. Deflation-based implementations, such as the popular

5: one-unit FastICA algorithm and its variants, extract the independent components one after

6: another. A novel method for deflationary ICA, referred to as

7: RobustICA, is put forward in this paper. This simple technique consists of

8: performing exact line search optimization of the kurtosis

9: contrast function. The step size leading to the global maximum of the contrast

10: along the search direction is found among the roots of a fourth-degree

11: polynomial. This polynomial rooting can be performed algebraically, and thus at low cost, at each iteration. Among other practical benefits, RobustICA can avoid

12: prewhitening and deals with real- and complex-valued mixtures of possibly

13: non-circular sources alike. The absence of prewhitening improves

14: asymptotic performance. The algorithm is robust to local

15: extrema and shows a very high convergence speed in

16: terms of the computational cost required to reach a given source extraction

17: quality, particularly for short data

18: records. These features are demonstrated by a comparative numerical

19: analysis on synthetic data. RobustICA's capabilities in

20: processing real-world data involving non-circular complex strongly

21: super-Gaussian sources are

22: illustrated by the biomedical problem of atrial activity (AA) extraction

23: in atrial fibrillation (AF) electrocardiograms (ECGs), where it

24: outperforms an alternative ICA-based technique.

25: \end{abstract}

26: