# Nonmetrizable uniformities with metrizable topologies

I'm looking for such pathological examples of uniform spaces which are not metrizable, but whose underlying topology is metrizable. Willard in his General Topology text constructs such a uniformity using ordinals. I am asking for examples which do not rely on ordinals.

EDIT: Below is an example by Daniel Tausk using families of pseudometrics. I forgot to mention that, if possible, I would like an example that uses the definition through a diagonal uniformity, not the definition through pseudometrics nor the definition through uniform covers. See Cryptomorphisms for some elaboration on this phenomenon.

-

The family of all neighbourhoods of the diagonal of $\mathbb{R}$ (with its normal topology) is a uniformity without a countable basis but it generates the normal topology. The normal topology of $\mathbb{R}$ is the topology generated by the usual distance. To see that the open sets in the plane that contain the diagonal generate a uniformity let $U$ be such an open set. For each $x$ there is $r_x$ such that $B(x,r_x)^2\subseteq U$. Now let $V$ be the union of the squares $B(x,\frac13r_x)^2$; this is again an open set and it is not hard to show that $V\circ V\subseteq U$. If $U$ is an open set in the plane that contains the diagonal then the vertical section $U[x]$ is open in $\mathbb{R}$, so this uniformity generates at best a subtopology of the normal topology and hence certainly not the discrete topology. To see that it generates the normal topology consider, given $x$ and $r$, the open set $U_{x,r}=B(x,r)^2\cup(\mathbb{R}\setminus\lbrace x\rbrace)^2$; then $U_{x,r}[x]=B(x,r)$. The uniformity does not have a countable base: if $\langle U_n:n\in\mathbb{N}\rangle$ is a sequence of neighbourhoods of the diagonal then you can use diagonalisation to produce a neighbourhood such that $U_n\not\subseteq U$ for all $n$: make sure that the square $[n,n+1]^2$ contains a point in $U_n$ but not in $U$.

-
What do you mean by "normal topology"? Do you mean the usual topology? If so, then why would the family of all neighborhoods of the diagonal (which is the discrete uniformity) generate the usual topology? (it generates the discrete topology, no?) –  lentic catachresis Nov 7 '10 at 15:29
So, when you are saying "The family of all neighbourhoods of the diagonal of $\mathbb{R}$", what you actually mean is "the family of all \textbf{open} neighborhoods $U\subset \mathbb{R}^2$ of the diagonal, where "open" means open with the usual topology of $\mathbb{R}^2$", right? It is not the same: this way, it does not generate the discrete uniformity (if it did, it really would generate the discrete topology -- see en.wikipedia.org/wiki/Discrete_space ). –  lentic catachresis Nov 11 '10 at 20:00
You asked for a metrizable topology, not a discrete one. Though Isbell wrote that all countereamples are discrete'' I thought it'd be interesting to see that the usual topology of $\mathbb{R}$ can be generated by a nonmetrizable uniformity. I use `neighbourhood' in the standard sense $U$ is a neighbourhood of $A$ if there is an open set $O$ such that $A\subseteq O\subseteq U$. –  KP Hart Nov 21 '10 at 12:09

Consider the interval $[0,1]$ and for each $x\in[0,1]$ let $d_x$ be the metric in $[0,1]$ defined by $d_x(y,z)=|y-z|$ if $y\ne x$ and $z\ne x$, $d_x(x,y)=d_x(y,x)=1$ if $y\ne x$ and $d_x(x,x)=0$.

Let $\mathcal U$ be the uniform structure defined by the family of metrics $d_x$, $x\in[0,1]$. For each $x\in[0,1]$, the open ball $B_{d_x}(x,1)$ equals $\{x\}$ and thus the topology defined by $\mathcal U$ is discrete (hence metrizable). Assume by contradiction that there exists a metric $d$ in $[0,1]$ defining the uniform structure $\mathcal U$. In particular, $d$ must define the same topology as $\mathcal U$, i.e., $d$ must be discrete. Thus, if $A_n$ denotes the set of those $x\in[0,1]$ such that the open ball $B_d(x,1/n)$ equals $\{x\}$ then $\bigcup_{n=1}^\infty A_n=[0,1]$ and therefore there exists $n\ge1$ such that $A_n$ is infinite. Since the identity map:

$\mathrm{Id}:([0,1],\mathcal U)\to([0,1],d)$

is uniformly continuous, there exists a finite set $F\subset[0,1]$ and $\delta>0$ such that, for all $y,z\in[0,1]$:

$d_x(y,z)<\delta$ for all $x\in F$ implies $d(y,z)<1/n$.

Choose $y\in A_n\setminus F$ and let $z\in[0,1]$ be such that $|z-y|<\delta$, $z\not\in F$ and $z\ne y$. Then $d_x(y,z)=|y-z|<\delta$ for all $x\in F$ and, since $y\in A_n$ and $z\ne y$, we have also $d(y,z)\ge1/n$, a contradiction.

-

Let $X$ be an infinite set. Take the uniformity $\mathcal U$ generated by those equivalence relations on $X$ that have only finitely many equivalence classes. The induced topology on $X$ is discrete and thus metrizable. But the uniformity $\mathcal U$ is not metrizable, because, given any countably many entourages (without loss of generality, equivalence relations with finitely many classes each), it is easy to construct another equivalence relation, with only two classes, that does not include any of the countably many given ones. (Alternative proof: The completion of this uniform space is the Stone-Cech compactification of the discrete space $X$, which is not metrizable as a topological space.)

-