5
$\begingroup$

This question is purely academic in nature; I have no application for the knowledge scholastic or otherwise other than curiousity.

Suppose you have a KxN matrix of data, with each row representing a variable, and each column representing an observation of each variable, called M.

Suppose you now generate a correlation matrix using the data in M. Is there a simple distribution for the variables in the K rows such that the correlation coefficients of the non-diagonal elements of the matrix are distributed approximately Uniform(0,1)? Or alternatively, for a covariance matrix rather than a correlation matrix?

$\endgroup$
4
$\begingroup$

It's not only possible, it's easy to create any distribution $F$ whatsoever supported on the interval $[-1/(N-2), 1]$, provided only that $K \le N-2$. Here's one way. It creates datasets in which all the variables have the same correlation with each other.

Let $\rho$ be a random variable with distribution $F$. Define $U \ge 1/(N-1)$ as the unique solution to

$$\rho = \frac{1 + 2 U - (N-1)U^2}{2 - 2(N-2)U + (N-1)(N-2)U^2}.$$

Set $V = (N-2)U-1$ and construct the $K$ vectors, each of length $N$, given by

$$\left\{\eqalign{ X_1 &= (1, V, -U, -U, \ldots, -U) \\ X_2 &= (1, -U, V, -U, -U, \ldots, -U) \\ &\ldots \\ X_K &= (1, -U, -U, \ldots, -U, V, -U, \ldots, -U). }\right.$$

Each has a $1$ in the first place, $V$ in the $K+1^\text{st}$ place, and $-U$ everywhere else.

A computation (which is simple because all the $X_i$ have zero means and the same variance) shows that $\rho$ is the correlation coefficient between each $X_i$ and $X_j$. Therefore all the correlation coefficients of these $K$ random vectors of length $N$ equal $\rho$, QED.


Appendix: Illustration via simulation

This R code simulates from a given distribution $F$. It displays histograms of the correlation coefficients and tests them for uniformity. The comments explain the details.

#
# Specify the situation.
#
N <- 20       # Dataset size
K <- 4        # Number of variables
n.sim <- 1e4  # Simulation size
#
# Predefine some objects.
#
f <- function(rho, n) { # Maps `rho` to `U`
  (1 + (n-2)*rho + sqrt(n * (1-rho)*(1+(n-2)*rho))) / ((n-1) * (1+(n-2)*rho))
}
pattern <- cbind(diag(rep(1, K)), matrix(0, K, N-K))
mask <- lower.tri(outer(1:K, 1:K))
#
# Conduct the simulation.
#
# rF <- runif      # The random number generator
# qF <- qunif      # The quantile function
# dF <- dunif      # The density function
rF <- function(n) rbeta(n, 1, 3)
qF <- function(q) qbeta(q, 1, 3)
dF <- function(x) dbeta(x, 1, 3)
rho <- rF(n.sim)   # Draw values of `rho`
#
# Construct the data and compute their correlation coefficients.
# Each row of `sim` will record one particular correlation coefficient.
# Its columns are the iterations.
#
U <- f(rho, N)
sim <- sapply(U, function(u) {
  v <- (N-1)*u - 1
  x <- matrix(rep(c(rep(-u, N-1), 1), K), nrow=K, byrow=TRUE) + v*pattern
  cor(t(x))[mask]
})
#
# Display the distributions of the correlation coefficients.
#
n.plots <- choose(K,2)
n.rows <- floor(sqrt(n.plots))
n.cols <- ceiling(n.plots/n.rows)
par(mfrow=c(n.rows, n.cols))
breaks <- qF(seq(0, 1, by=1/20))
invisible(apply(sim, 1, function(x) {
  H <<- hist(x, main="Marginal Histogram", freq=FALSE, breaks=breaks)
  curve(dF(x), add=TRUE, col="Red", lwd=2)
  #
  # Test the uniformity with a chi-squared test.
  #
  p <- chisq.test(H$counts)$p.value
  mtext(paste0("(Test of uniformity: p = ", signif(p, 3), ")"), cex=0.75)
}))
par(mfrow=c(1,1))
$\endgroup$
  • $\begingroup$ thank you very much for this answer. I think it will take me some time to investigate it, but I will endeavor to do so soon. $\endgroup$ – Vincent Laufer Mar 14 '17 at 17:23
  • 1
    $\begingroup$ I added some code in case that might help. $\endgroup$ – whuber Mar 14 '17 at 18:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.