# Labelings of infinite directed acyclic graphs

Let $G=(V,E)$ be a countably infinite directed acyclic graph and $L$ be a finite set of vertex labels. The number $\left|V\right|$ of vertices is countable infinity and some vertices may have an infinite number of ingoing edges. I am interested in the existence of labelings $f: V \rightarrow L$ where the label $f(v)$ is determined by the labels of the predecessors of $v$.

More formally, let there be a function $g_v:L^{\mathrm{indeg}(v)} \rightarrow L$ for each vertex $v \in V$. Does there exist a labeling $f : V \rightarrow L$ such that for every vertex $v \in V$ with predecessors $p_0, p_1,\ldots$ the following holds? $$f(v) = g_v(f(p_0), f(p_1), \ldots)$$

Intuitively, it appears very clear to me that the acyclicity of $G$ guarantees the existence of such labelings $f$. But how do I prove that?

For a trivial example, consider the infinite path $G = (\mathbb{Z}, \{(v,v+1):v\in\mathbb{Z}\})$, $L = \{0, 1\}$ and $g_v:L\rightarrow L:x\mapsto 1 - x$. In this case there are exactly two labelings, one that assigns 0 and 1 to every even and odd vertex, respectively, and the other with 0 and 1 switched.

If we drop the requirement of acyclicity, a counterexample would be an odd directed cycle graph and $L$ and $g$ as in the example above. Here no labeling exists.

In the case of finite DAGs, I would prove the existence inductively over an topological ordering of $G$, starting by an arbitrary label-assignment for vertices without ingoing edges and then iteratively constructing the labeling for vertices whose predecessors have already been labeled. However, I fail to generalize this technique to infinite graphs in the absence of an induction basis.

## 2 Answers

A clear and good question.

In general there is a simple counterexample. Let $V=\omega=\{0,1,2,\dots\}$, $E=\{(m,n)\in\omega^2: m>n\}$, $L=\{0,1\}$. Let $\mathcal F$ be a free ultrafilter on the set $\omega$. Define a function $g:L^\omega\to L$ by putting for each labelling $f:V\to L$ $g(f)=1$, if $\{n\in\omega: f(n)=0\}\in\mathcal F$ and $g(f)=0$, otherwise. For each vertex $n\in\omega$ and each labelling $f’:\omega\setminus\{n+1\}\to L$ (that is, $f’\in \{0,1\}^{\{n+1,n+2,\dots\}}$) put $g_n(f’)=g(f)$, where $f:V\to L$ is an extension of the map $f'$ (that is, $f|_{\omega\setminus\{n+1\}}=f’$) (since the ultrafilter $\mathcal F$ is free, the map $g_n$ does not depend on an extension $f$). Assume that there exists a labelling $f:V\to\omega$ such that $f(n)=g_n(f|_{\omega\setminus\{n+1\}})$ for each $n\in\omega$. Let $F=f^{-1}(0)$. Then $f\equiv 1$ provided $F\in\mathcal F$ and $f\equiv 0$ provided $F\not\in\mathcal F$, a contradiction.

It seems the following.

If for each vertex $v\in V$ the set $p(v)=\{w\in V:(w,v)\in E\}$ of all its predecessors is finite, then the answer is positive. To show this we shall procced as follows. Endow the set $L^V$ of all labellings on the graph $G$ by a topology of a Tychonoff power of a finite discrete space $V$. By Tychonoff Theorem, the space $L^V$ is compact. Let $F$ be a finite subset of $V$. Put $C_F=\{f\in L^V: f(v)=g_v(f|_{p(v)})$ for all $v\in F\}$. We claim that the set $C_F$ is closed. Indeed, assume that $f\in L^V\setminus C_F$ be an arbitrary labelling. Since $f\not\in L^V$, there exists a vertex $v\in F$ such that $f(v)\ne g_v(f|_{p(v)})$. Then $O_f=\{h\in L^V: h(w)=f(w)$ for each $w\in \{v\}\cup p(v)\}$ is an open neighborhood of the labelling $f$ and $O_f\cap C_F=\varnothing$. Next we claim that the set $C_F$ is non-empty. To show this we construct a labelling $f\in C_F$ as follows. For each vertex $v\in V\setminus F$ let $f(v)$ be an arbitrary element of the set $L$. Since the graph $G$ has no directed cycles, we can enumerate the set $F$ (that is, to define a bijective map $i:F\to \{0,1,\dots n\}$) such that $i(v)>i(w)$ for each vertices $v\in F$ and $w\in F\cap p(v)$. Then we can inductive define $f|F$ by putting $f(v)=g(f|_{p(v)})$ for each vertex $v\in F$. Put $\mathcal C=\{C_F: F$ is a finite subset of $V\}$. Since $\mathcal C$ is a centered family of closed subsets of a compact space, there is a map $f\in\bigcap\mathcal C$. By construction, $f$ is a required labelling.

It rests for me to consider a case, when the graph $G$ has no undirected cycles.

Counterexample. As in Alex Ravsky's example, let $V=\omega$, $E=\{(m,n):m\gt n\}$, $L=\{0,1\}.$ Now simply define $f(p_0,p_1,\dots)$ to take the value $1$ if all of the inputs are $0$, and to take the value $0$ otherwise.

• Nice. :-) I work too much in set-theoretic topology. :-) – Alex Ravsky May 8 '15 at 10:10