Limiting distribution of the sample canonical correlation coefficients of high-dimensional random vectors

Abstract: Consider two high-dimensional random vectors

w i d e t i l d e m a t h b f x i n m a t h b b R^{p}

and

w i d e t i l d e m a t h b f y i n m a t h b b R^{q}

with finite rank correlations. More precisely, suppose that

w i d e t i l d e m a t h b f x = m a t h b f x + A m a t h b f z

and

w i d e t i l d e m a t h b f y = m a t h b f y + B m a t h b f z

, for independent random vectors

m a t h b f x i n m a t h b b R^{p}

,

m a t h b f y i n m a t h b b R^{q}

and

m a t h b f z i n m a t h b b R^{r}

with iid entries of mean 0 and variance 1, and two deterministic matrices

A i n m a t h b b R^{p i m e s r}

and

B i n m a t h b b R^{q i m e s r}

. With

n

iid observations of

(w i d e t i l d e m a t h b f x, w i d e t i l d e m a t h b f y)

, we study the sample canonical correlations between them. In this paper, we focus on the high-dimensional setting with a rank-

r

correlation. Let

t_{1} g e c d o t s g e t_{r}

be the squares of the population canonical correlation coefficients (CCC) between

w i d e t i l d e m a t h b f x

and

w i d e t i l d e m a t h b f y

, and

w i d e t i l d e l a m b d a_{1} g e c d o t s g e w i d e t i l d e l a m b d a_{r}

be the squares of the largest

r

sample CCC. Under certain moment assumptions on the entries of

m a t h b f x

,

m a t h b f y

and

m a t h b f z

, we show that there exists a threshold

t_{c} i n (0, 1)

such that if

t_{i} > t_{c}

, then

s q r t n (w i d e t i l d e l a m b d a_{i} - h e t a_{i})

converges in law to a centered normal distribution, where

h e t a_{i} > l a m b d a_{+}

is a fixed outlier location determined by

t_{i}

. Our results extend the ones in [4] for Gaussian vectors. Moreover, we find that the variance of the limiting distribution of

s q r t n (w i d e t i l d e l a m b d a_{i} - h e t a_{i})

also depends on the fourth cumulants of the entries of

m a t h b f x

,

m a t h b f y

and

m a t h b f z

, a phenomenon that cannot be observed in the Gaussian case.

Recommendations

Cites work

Cited in

(10)

This page was built for publication: Limiting distribution of the sample canonical correlation coefficients of high-dimensional random vectors