Little Problem About Small World Graphs

Little Problem About Small World Graphs

Consider a graph G:=(V,E)G := (V, E) with |V|=n\lvert V \rvert = n and uv∈Euv \in E if and only if |u−v|=2k\lvert u - v \rvert = 2^k for some k∈ℕ0k \in \mathbb{N}_0. What is E[d(u,v)∣n]E\left[d\big(u, v) \mid n\right]? I was interested in showing that this is in fact a “small-world” graph, i.e. that L∝log⁡(n)L \propto \log\big(n). This is not so difficult, as the maximum distance between two nodes uu and vv cannot exceed log⁡2(n)\log_2\big(n).

My first attempt to find the average involved bounding the average as follows. In this setting, there’s no trip with greater distance than from a node to the node directly behind it, i.e. dn(0,n−1)≥E[dn(0,v)]d_{n}\big(0, n - 1) \geq E[d_{n}\big(0, v)]. But this is just a minor improvement over the maximum.

∫0nln⁡(u)nln⁡(2)du=log⁡2(n)−1ln⁡(2)\begin{equation} \int_{0}^{n}\frac{\ln\left(u\right)}{n \ln\left(2\right)}\ du = \log_2\big(n) - \frac{1}{\ln\big(2)} \end{equation}

Instead, notice that the distance between nodes indexed uu and vv is equal to the number of bits set in uxorvu \ \text{xor} \ v. This gives us an expected distance of 0.5⋅log⁡2(n)0.5 \cdot \log_2\big(n) and a distribution of distances that’s very roughly N⁡(log⁡2(n)/2,log⁡2(n)/4)\operatorname{N}\big(\log_2\big(n)/2, \log_2\big(n)/4).