1.  Sites, days, temperatures; sd - (# of NAs).

2.  Ridge has a quadratic/linear objective function, hence is
differentiable.  The LASSO, with its absolute value, is not
differentiable.

3.  We now have a square instead of a diamond, with corners at
(-eta,eta), (eta,eta), (eta,-eta) and (-eta,eta).  Almost certainly, the
ellipse that barely touches the square will touch at one of the corners,
all of which are nonsparse.

4.  The intended solution below is not quite correct.  It sets up a
structure in which the variables 2,3,...,p each have a correlation of
rho with variable 1 but only have a correlation of rho^2 with each
other.  (Use our formula Cov(AX) = A Cov(X) A' to see this.)  To fix it
would require not only replacing rho with sqrt(rho) but also adjusting
the new standard deviations.

simProbAccident <- function(p,n,rho,accidentLevel,nSamp)
{
   simout <- replicate(nSamp,
   {
      v <- matrix(rnorm(p*n),ncol=p);
      w <- matrix(nrow=n,ncol=p);
      w[,1] <- v[,1];
      for (i in 2:p) {
         w[,i] <- rho * w[,1] + v[,i]
      }; 
      corMat <- cor(w);   # element i,j is rhoHat[i,j]
      diag(corMat) <- 0; 
      max(corMat);
    })
   mean(simout >= accidentLevel)
}