Geometrical and statistical properties of M-estimates of scatter on Grassmann manifolds

Publication date: 30 December 2018

Abstract: We consider data from the Grassmann manifold

G (m, r)

of all vector subspaces of dimension

r

of

m a t h b b R^{m}

, and focus on the Grassmannian statistical model which is of common use in signal processing and statistics. Canonical Grassmannian distributions

m a t h b b G_{S i g m a}

on

G (m, r)

are indexed by parameters

S i g m a

from the manifold

m a t h c a l M = P o s_{s y m}^{1} (m)

of positive definite symmetric matrices of determinant

1

. Robust M-estimates of scatter (GE) for general probability measures

m a t h c a l P

on

G (m, r)

are studied. Such estimators are defined to be the maximizers of the Grassmannian log-likelihood

- e l l_{m a t h c a l P} (S i g m a)

as function of

S i g m a

. One of the novel features of this work is a strong use of the fact that

m a t h c a l M

is a CAT(0) space with known visual boundary at infinity

p a r t i a l m a t h c a l M

. We also recall that the sample space

G (m, r)

is a part of

p a r t i a l m a t h c a l M

, show the distributions

m a t h b b G_{S i g m a}

are

S L (m, m a t h b b R)

--quasi-invariant, and that

e l l_{m a t h c a l P} (S i g m a)

is a weighted Busemann function. Let

m a t h c a l P_{n} = (d e l t a_{U_{1}} + c d o t s + d e l t a_{U_{n}}) / n

be the empirical probability measure for

n

-samples of random i.i.d. subspaces

U_{i} i n G (m, r)

of common distribution

m a t h c a l P

, whose support spans

m a t h b b R^{m}

. For

S i g m a_{n}

and

S i g m a_{m a t h c a l P}

the GEs of

m a t h c a l P_{n}

and

m a t h c a l P

, we show the almost sure convergence of

S i g m a_{n}

towards

S i g m a

as

n o i n f t y

using methods from geometry, and provide a central limit theorem for the rescaled process

C_{n} = f r a c m t r (S i g m a_{m a t h c a l P}^{- 1} S i g m a_{n}) g^{- 1} S i g m a_{n} g^{- 1}

, where

S i g m a = g g

with

g i n S L (m, m a t h b b R)

the unique symmetric positive-definite square root of

S i g m a

.

This page was built for publication: Geometrical and statistical properties of M-estimates of scatter on Grassmann manifolds