Strange behavior of correlation estimation

The Gaussian vector is extremely interesting since it remains Gaussian when conditioning. More precisely, if is a Gaussian random vector, then the conditional distribution of is also Gaussian. Further, it is possible to derive explicitly the covariance matrix. And interesting result, that one will not depend on the conditioning. Here, it will we

where the covariance matrix of the random vector was

i.e. it is the Schur complement of bloc in . For instance consider an Gaussian vector in dimension 3,

It is possible to derive explicitly the correlation of

> library(mnormt)
> SIGMA=matrix(c(1,.6,.7,.6,1,.8,.7,.8,1),3,3)
> SIGMA11=SIGMA[1:2,1:2]
> SIGMA21=SIGMA[3,1:2]
> SIGMA12=SIGMA[1:2,3]
> SIGMA22=SIGMA[3,3]
> S=SIGMA11-SIGMA12%*%solve(SIGMA22)%*%SIGMA21
> S[1,2]/sqrt(S[1,1]*S[2,2])
[1] 0.09335201

This concept is discussed in a paper on Conditional Correlation by Kunihiro Baba, Ritei Shibata and Masaaki Sibuya. But what about statistical inference ? In order to see what happens, let us run some simulations, and condition with in some deciles,

> k=10
> correlV=rep(0,k)
> X=rmnorm(n=10000000,varcov=SIGMA)
> QZ=qnorm((0:k)/k)
> ZQ=cut(X[,3],QZ)
> LZ=levels(ZQ)
> correl=rep(NA,k)
> for(i in 1:k){
+ 	correl[i]=cor(X[ZQ==LZ[i],1:2])[1,2]
+ }
> barplot(correl,names.arg=LZ,col="light blue",ylim=range(correl))

or some percentiles,

> k=100

In the middle it looks fine (we can look at it more carefully and fit a spline curve), but for very large values of , then we observe strange patterns (the graph below is based on 100 million simulations),

Note that for all estimates (each bar), there are exactly the same number of observations since when conditioning, we keep only observations such that is between two consecutive percentile (hence, we always keep 1% of the data). If anyone has any idea about what happens, I’d be glad to get explanations… At first, I thought it would be a speed of convergence issue… But if we look at the impact of the number of simulations used in the fifth decile,

(up to 1 million simulations, i.e. 100,000 conditional observations used to estimate the correlation). If we look at the third decile, it converges,

for the second one, it might be a bit slower,

but for the first decile…  either it is extremely slow, or it has a strong bias. But I don’t get why…