0
$\begingroup$

I'm looking for a simple, symmetric multivariate distribution for $N$ Bernoulli variables with the following properties:

  • Each individual variable takes on values 1 or 0

  • Fix a subset of $M$ variables. Let $P_K$ be the probability that conditional on $K$ of these $M$ variables taking on value 1, the $N$th variable is 1. Want: $P_K$ increasing in $K$.

  • A parameterized functional form with a single parameter, so that at the extremes, we arrive at the case of perfect correlation and at independence.

Hopefully this is not too simple a question to ask on mathoverflow, but I've had a tough time coming up with a simple distribution that does this. Many thanks.

$\endgroup$
1
  • $\begingroup$ Take a multivariate Gaussian distribution on $(y_1,..,y_N)$ with a constant pairwise correlation $\rho$. Pick some threshold $T$ and set $x_i = 1$ if and only if $Y_i \geq T$ (otherwise zero). The $x_i$'s are correlated Bernoulli r.v.s. with a distribution parametrized by $\rho$ which satisfies your conditions. $\endgroup$ – Or Zuk Apr 27 '11 at 21:50
1
$\begingroup$

Gibbs measures provide a solution. Introduce the energy $H(x)$ of a configuration $x=(x_k)_{1\le k\le N}$ in $S=\{0,1\}^N$ as $$ H(x)=\sum_kx_k+\sum_{k\ne\ell}x_kx_\ell, $$ and, for every parameter $a>0$, define a probability $P_a$ on $S$ by $$ P_a(x)=a^{H(x)}/Z_a,\qquad Z_a=\sum_{x\in S}a^{H(x)}. $$ Special values are $P_1$ uniform on $S$ and $P_\infty$ the Dirac mass at the configuration all ones. For every $a>1$, the random variables $x_k$ are positively correlated due to the $x_kx_\ell$ terms in the energy $H(x)$ hence the properties you ask are satisfied.

$\endgroup$

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.