1: \begin{abstract}

2: We consider robust covariance estimation with group symmetry constraints. Non-Gaussian

3: covariance estimation, e.g.,

4: Tyler scatter estimator and Multivariate Generalized Gaussian distribution methods,

5: usually involve non-convex

6: minimization problems. Recently, it was shown that the underlying principle behind their

7: success is an extended form

8: of convexity over the geodesics in the manifold of positive definite matrices.

9: A modern approach to improve estimation accuracy is to exploit prior knowledge via

10: additional constraints, e.g., restricting the attention to specific classes

11: of covariances which adhere to prior symmetry structures. In this paper, we prove that

12: such group symmetry constraints

13: are also geodesically convex and can therefore be incorporated into various non-Gaussian

14: covariance estimators.

15: Practical examples of such sets include: circulant, persymmetric and complex/quaternion

16: proper structures.

17: We provide a simple numerical technique for finding maximum likelihood estimates under such

18: constraints,

19: and demonstrate their performance advantage using synthetic experiments.

20: \end{abstract}

21: