Suppose we pick $n$ random points in the unit cube in $\mathbb{R}_3$, $p_1=\left(x_1,y_1,z_1\right),$ $p_2=\left(x_2,y_2,z_2\right),$ etc. (So, $x_i,y_i,z_i$ are $3n$ uniformly distributed random variables between $0$ and $1$.) Let $\Gamma$ be a complete graph on these $n$ points, and weight each edge $\{p_i,p_j\}$ by $$w_{ij}=\sqrt{\left(x_i-x_j\right)^2+\left(y_i-y_j\right)^2+\left(z_i-z_j\right)^2}.$$

Question: What is the expected value of the total weight of a minimal spanning tree of $\Gamma$?

(Note: Here total weight means the sum of all edges in the minimal spanning tree.)

A peripheral request: The answer is probably a function of $n$, but I don't have the computing power or a good implementation of Kruskall's algorithm to suggest what this should look like. If someone could run a simulation to generate this average over many $n$, it might help towards a solution to see this data.


According to the Wikipedia article on the Euclidean minimum spanning tree (or EMST), the expected total weight is asymptotic to $$ c(d)n^{\frac{d-1}{d}}\int_{\mathbb{R}^{d}}f(x)^{\frac{d-1}{d}}dx $$ as $n\rightarrow\infty$, where $d$ is the dimensionality (here, $d=3$), $f(x)$ is the probability density of point selection (here, $f(x)=1$ inside the cube), and $c(d)$ depends only on the dimensionality. For uniform selection within the unit cube, the expected size should satisfy $$ E(n) \sim c_{3}n^{2/3} $$ for some constant $c_{3}$.

Numerical experiments (for modest $n$, up to about $8000$) suggest that $c_{3}=0.65 \pm 0.01$.

  • $\begingroup$ If the minimum spanning tree were non-Euclidean, how would expected size formula differ? Thank you. $\endgroup$ – Frank Feb 20 '17 at 0:23
  • $\begingroup$ What about the rectilinear non-Euclidean minimum spanning tree expected size? Is there a back of the envelope calculation for Euclidean minimum spanning tree expected size? Thank you. $\endgroup$ – Frank Feb 20 '17 at 7:05
  • $\begingroup$ By rectilinear I assume you mean Manhattan or taxicab metric? I don't know; you might get a good response if you pose it as a new stand-alone question. If I had to guess, it'd still scale as $n^{2/3}$, with a different prefactor. $\endgroup$ – mjqxxxx Feb 20 '17 at 19:31
  • $\begingroup$ Thank you for your comment. In addition, I would like to find out why expected size E(n)∼ n / log (n) when an independent random number generator is used to derive the edge weights according to a [0,1] uniform probability distribution. $\endgroup$ – Frank Feb 21 '17 at 4:06

If $n = 0$ or $n = 1$ the answer obviously is 0. If $n = 2$ we have $$E\left((x_1 - x_2)^2\right) = E(x_1^2 - 2x_1x_2 + x_2^2) = E(x_1^2) - 2E(x_1)\cdot E(x_2) + E(x_2^2) \\= \frac13 - 2\frac12\cdot\frac12 + \frac13 = \frac16.$$ The same for $y$- and $z$-coordinates. So $E(w_{12}) = \sqrt{\frac16 + \frac 16 + \frac16} = \frac1{\sqrt2}$ and spanning tree contains the edge $\{\,1, 2\,\}$ only.

I see it is possible to consider several cases for $n = 3$, however for arbitrary $n$ I don't expect to get close form of the answer.


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.