Consider a sequence of expander graphs ($G_n$); say $G_n$ has $n$ vertices. Remove $o(n)$ vertices (and the edges emanating from these vertices) and cut $o(n)$ edges. Call $G'_n$ the largest connected component of the resulting graph. Are the ($G'_n$) still expanders ?
-
5$\begingroup$ What are you taking to be the definition of an expander? $\endgroup$– Ben BarberCommented Jul 7, 2016 at 10:15
-
1$\begingroup$ The definition is that the Cheeger constant is bounded below by a positive quantity, independent of $n$. The Cheeger constant (isoperimetric constant) is $$h(G_n)=\inf_{S\subset G_n}\frac{\mbox{number of edges between } S \mbox{ and } S^c}{\min(\sharp S, \sharp S^c)}.$$ $\endgroup$– NaliniCommented Jul 7, 2016 at 11:41
1 Answer
No. Assume that $G_n$ has bounded degree (this is probably an assumption of yours).
By removing $0$ vertices and $O(\log n)$ edges, you can make sure that $G'_n$ has $\geq n/2$ vertives and contains a "segment of length $\geq c\log n$" (that is a sequence $x_1,\dots,x_{k}$ of vertices in $G'_n$ such that $k \geq c \log n$ and for all $i <k $, the set of neighbours of $x_i$ is $\{x_{i-1},x_{i+1}\}$ (without $x_{0}$ if $i=1$). In particular the boundary of $\{x_1,\dots,x_{k-1}\}$ reduces to $\{x_k\}$, so $G'_n$ is not an expander sequence.
There are probably many ways to justify the preceding ; here is one. Pick a spanning tree $T$ in $G_n$. By the bounded degree assumption, $T$ has diameter $\geq c' \log n$, so there is a path $x_1,\dots,x_{2k}$ in $T$ of length $2k = c \log n$. By reversing the order of this path you can assume that ($*$) by removing the edge between $x_k$ and $x_{k+1}$ in the tree, the subtree containing $x_1$ contains less than half of the vertices. Remove from $G_n$ all the $O(\log(n))$ edges adjacent to $x_1,\dots,x_{k-1}$ except those that appear in $T$. Then the connected component of $x_1$ in the remaining graph contains more that half of the vertices of $G_n$ by ($*$). So $x_1$ belongs to $G'_n$, and by construction $x_1,\dots,x_k$ is a segment in $G'_n$.