9.3 Separation Axioms
How finely can open sets tell points and closed sets apart? The ladder $T_4\Rightarrow T_3\Rightarrow T_2\Rightarrow T_1\Rightarrow T_0$, its standard non-examples (cofinite line, line with two origins, $K$-topology, Sorgenfrey plane), metric spaces as the model of normality, and Urysohn's lemma.
In a metric space two different points are always a positive distance apart, so small balls around them are disjoint; a point outside a closed set is a positive distance from it; and two disjoint closed sets can be pulled apart by the continuous function $d(x,A)/\big(d(x,A)+d(x,B)\big)$. A general topology has no distance, and each of these facts can fail. In the cofinite topology on $\mathbb{R}$ any two nonempty open sets meet, so a sequence of distinct points converges to every point at once. On the "line with two origins" a sequence has two limits although every point has a neighbourhood that looks exactly like an open interval. The separation axioms measure how far a space is from these pathologies: $T_0$ and $T_1$ ask that open sets distinguish points, $T_2$ (Hausdorff) that disjoint open sets surround distinct points, and regularity and normality that points and closed sets, or pairs of closed sets, can be surrounded by disjoint open sets.
Separation is what makes the rest of topology work. Limits are unique in Hausdorff spaces; compact subsets of Hausdorff spaces are closed (§9.4); normal spaces carry enough continuous real functions to separate closed sets (Urysohn's lemma) and to extend functions from closed subsets (Tietze); and a regular space with a countable basis is metrizable (Urysohn's metrization theorem). In CSIR NET this section supplies a steady stream of Part B and Part C questions of the type "which of the following spaces is Hausdorff / regular / normal / metrizable?", and they are always settled by the same handful of examples: the cofinite and cocountable topologies, the $K$-topology, the Sorgenfrey line $\mathbb{R}_\ell$ and its square. We meet all of them here, with proofs.
Throughout, $(X,\mathcal{T})$ is a topological space, "open" means "in $\mathcal{T}$", $\overline{A}$ is the closure of $A$, and $\mathbb{R}_\ell$ denotes $\mathbb{R}$ with the lower limit topology (basis $[a,b)$) from §9.1. Subspaces, products and quotients carry the topologies of §9.2.
Distinguishing Points: $T_0$, $T_1$ and Hausdorff Spaces
A topological space $X$ is
- $T_0$ (a Kolmogorov space) if for any two distinct points some open set contains one of them but not the other;
- $T_1$ (a Fréchet or accessible space) if for any two distinct points $x,y$ there is an open set containing $x$ but not $y$ (and, by symmetry, one containing $y$ but not $x$);
- $T_2$, a Hausdorff space, if for any two distinct points $x,y$ there are open sets $U\ni x$ and $V\ni y$ with $U\cap V=\varnothing$.
Directly from the definitions, $T_2\Rightarrow T_1\Rightarrow T_0$: disjoint $U\ni x$, $V\ni y$ give $y\notin U$ and $x\notin V$.
Four small examples fix the ideas. The indiscrete topology $\{\varnothing,X\}$ on a set with at least two points is not $T_0$: the only nonempty open set contains everything. The Sierpiński space $S=\{0,1\}$ with open sets $\varnothing,\{1\},S$ is $T_0$ (the set $\{1\}$ contains $1$ but not $0$) but not $T_1$ (no open set contains $0$ but not $1$). The cofinite topology on an infinite set $X$ is $T_1$ — the set $X\setminus\{y\}$ is open, contains $x$ and misses $y$ — but not $T_2$, since two nonempty open sets $X\setminus F$, $X\setminus G$ ($F,G$ finite) meet in $X\setminus(F\cup G)\ne\varnothing$. Finally every metric space is Hausdorff: if $x\ne y$ and $r=d(x,y)/2$, the balls $B(x,r)$ and $B(y,r)$ are disjoint, for a point $z$ in both would give $d(x,y)\le d(x,z)+d(z,y)\lt2r=d(x,y)$.
For a topological space $X$ the following are equivalent:
- $X$ is $T_1$;
- every singleton $\{x\}$ is closed;
- every finite subset of $X$ is closed;
- $\mathcal{T}$ is finer than the cofinite topology on $X$;
- for each $x\in X$, the intersection of all open sets containing $x$ is $\{x\}$.
In particular a finite $T_1$ space is discrete.
Proof. (1)$\Rightarrow$(2): given $x$, every $y\ne x$ has an open $U_y\ni y$ with $x\notin U_y$, so $X\setminus\{x\}=\bigcup_{y\ne x}U_y$ is open. (2)$\Rightarrow$(3): a finite union of closed sets is closed. (3)$\Rightarrow$(4): a nonempty cofinite-open set is $X\setminus F$ with $F$ finite, and $X\setminus F$ is open because $F$ is closed. (4)$\Rightarrow$(5): $X\setminus\{y\}$ is cofinite-open, hence open, for each $y\ne x$; so the intersection of the open sets containing $x$ excludes every $y\ne x$. (5)$\Rightarrow$(1): if $y\ne x$, then $y$ is not in the intersection of the open sets containing $x$, so some open $U\ni x$ misses $y$. Finally, in a finite $T_1$ space every subset is a finite union of closed singletons, hence closed; so every subset is also open (its complement is closed). ∎
Proof. "If" is immediate: infinitely many points of $A$ include one different from $x$. Conversely, suppose some open $U\ni x$ meets $A$ in only finitely many points, and let $F=(U\cap A)\setminus\{x\}$, a finite set. By Theorem 9.33, $F$ is closed, so $V=U\setminus F$ is open; it contains $x$ and $V\cap A\subseteq\{x\}$. Hence $x$ is not a limit point of $A$. ∎
The theorem fails without $T_1$: in the Sierpiński space, $0$ is a limit point of the one-point set $\{1\}$, because the only open set containing $0$ is $S$ itself.
To talk about limits we recall the general definition: a sequence $(x_n)$ in $X$ converges to $x$, written $x_n\to x$, if every open set containing $x$ contains $x_n$ for all but finitely many $n$. Nothing in this definition forces a limit to be unique.
(a) In a Hausdorff space a sequence has at most one limit.
(b) Conversely, if $X$ is first countable and every convergent sequence in $X$ has exactly one limit, then $X$ is Hausdorff. (Without first countability the converse fails: Example 2(b).)
Proof. (a) Suppose $x_n\to x$, $x_n\to y$ and $x\ne y$. Choose disjoint open $U\ni x$, $V\ni y$. There are $N_1,N_2$ with $x_n\in U$ for $n\ge N_1$ and $x_n\in V$ for $n\ge N_2$; then $x_n\in U\cap V=\varnothing$ for $n\ge\max(N_1,N_2)$, which is absurd.
(b) Suppose $X$ is not Hausdorff: there are $x\ne y$ such that every open set containing $x$ meets every open set containing $y$. Let $\{B_n\}$ and $\{C_n\}$ be countable neighbourhood bases at $x$ and $y$; replacing $B_n$ by $B_1\cap\cdots\cap B_n$ (and similarly $C_n$) we may assume $B_1\supseteq B_2\supseteq\cdots$ and $C_1\supseteq C_2\supseteq\cdots$. Choose $z_n\in B_n\cap C_n$. Any open $U\ni x$ contains some $B_N$, hence contains $z_n$ for all $n\ge N$; so $z_n\to x$, and in the same way $z_n\to y$. The sequence $(z_n)$ has two limits. ∎
The next criterion turns the Hausdorff property into a statement about a single closed set, and it is the source of most of the "nice" consequences of $T_2$.
Proof. A basic open set $U\times V$ of $X\times X$ misses $\Delta$ exactly when no $x$ lies in both $U$ and $V$, i.e. when $U\cap V=\varnothing$. Now $\Delta$ is closed $\iff$ $(X\times X)\setminus\Delta$ is open $\iff$ every $(x,y)$ with $x\ne y$ lies in a basic open $U\times V$ missing $\Delta$ $\iff$ every pair $x\ne y$ has disjoint open neighbourhoods $U\ni x$, $V\ni y$. ∎
Let $f,g:X\to Y$ be continuous and let $Y$ be Hausdorff. Then:
- the equaliser $E=\{x\in X: f(x)=g(x)\}$ is closed in $X$;
- if $f=g$ on a dense subset of $X$, then $f=g$ everywhere;
- the graph $\Gamma_f=\{(x,f(x)):x\in X\}$ is closed in $X\times Y$;
- if $Y=X$, the fixed-point set $\{x: f(x)=x\}$ is closed.
Proof. (1) The map $h:X\to Y\times Y$, $h(x)=(f(x),g(x))$, is continuous, because its two coordinate functions are (§9.2), and $E=h^{-1}(\Delta_Y)$ is the preimage of a closed set (Theorem 9.36). (2) $E$ is closed and contains a dense set $D$, so $E\supseteq\overline{D}=X$. (3) $\Gamma_f=k^{-1}(\Delta_Y)$ for the continuous map $k:X\times Y\to Y\times Y$, $k(x,y)=(f(x),y)$. (4) Apply (1) with $g=\mathrm{id}_X$. ∎
Part (2) is why a continuous function on $\mathbb{R}$ is determined by its values on $\mathbb{Q}$ — and it fails for maps into non-Hausdorff spaces: every function into an indiscrete space is continuous, so the constant map $0$ and the indicator of $\mathbb{R}\setminus\mathbb{Q}$, both maps $\mathbb{R}\to\{0,1\}_{\text{indiscrete}}$, agree on the dense set $\mathbb{Q}$ without being equal.
Proof. We give the Hausdorff case; $T_0$ and $T_1$ are identical with "disjoint" replaced by "not containing". Subspaces: if $A\subseteq X$ and $x\ne y$ in $A$, take disjoint open $U\ni x$, $V\ni y$ in $X$; then $U\cap A$ and $V\cap A$ are disjoint and open in $A$. Products: if $x=(x_\alpha)\ne y=(y_\alpha)$, then $x_\beta\ne y_\beta$ for some index $\beta$. Separate them by disjoint open $U,V\subseteq X_\beta$; then $\pi_\beta^{-1}(U)\ni x$ and $\pi_\beta^{-1}(V)\ni y$ are open and disjoint. Converse: fix a point $c=(c_\alpha)$ of the product. For each $\beta$, the map $X_\beta\to\prod_\alpha X_\alpha$ sending $t$ to the point with $\beta$-coordinate $t$ and all other coordinates $c_\alpha$ is an embedding (§9.2), so $X_\beta$ is homeomorphic to a subspace of a $T_i$ space. ∎
Quotients are the exception: identifying points can destroy every separation property. The quotient of $\mathbb{R}$ by $x\sim y\iff x-y\in\mathbb{Q}$ is indiscrete (Exercise III.7), and the next example is a quotient of two copies of $\mathbb{R}$ that is $T_1$ but not Hausdorff.
Regular and Normal Spaces
Hausdorff spaces separate points from points. The next two axioms separate points from closed sets and closed sets from closed sets. Since a closed set may be "infinitely long" (think of $\mathbb{Z}$ in $\mathbb{R}$, or the graph of $1/x$ in the plane), these are genuinely stronger requirements.
Let $X$ be a $T_1$ space.
- $X$ is a regular space ($T_3$) if for every closed set $C$ and every point $x\notin C$ there are disjoint open sets $U\ni x$ and $V\supseteq C$.
- $X$ is a normal space ($T_4$) if for every pair of disjoint closed sets $A,B$ there are disjoint open sets $U\supseteq A$ and $V\supseteq B$.
We say that $U$ and $V$ separate the two sets. Convention warning: many texts (and some exam papers) define "regular" and "normal" by the separation condition alone, without $T_1$, and then write $T_3=$ regular $+\,T_1$, $T_4=$ normal $+\,T_1$; a few even swap the names. Without $T_1$ the indiscrete space is "normal" and "regular" vacuously. In this course regular and normal spaces are always $T_1$; when we need the bare separation condition we say "satisfies the normality condition".
Proof. In a $T_1$ space singletons are closed (Theorem 9.33). If $X$ is normal and $x\notin C$, then $\{x\}$ and $C$ are disjoint closed sets, so normality separates them: $X$ is regular. If $X$ is regular and $x\ne y$, then $\{y\}$ is a closed set not containing $x$, and regularity separates $x$ from $\{y\}$: $X$ is Hausdorff. The last two implications were noted after Definition 9.20. ∎
Let $X$ be $T_1$.
(a) $X$ is regular if and only if for every point $x$ and every open $U\ni x$ there is an open $V$ with $x\in V\subseteq\overline{V}\subseteq U$.
(b) $X$ is normal if and only if for every closed $A$ and every open $U\supseteq A$ there is an open $V$ with $A\subseteq V\subseteq\overline{V}\subseteq U$.
Proof. (a) Suppose $X$ is regular, $x\in U$, $U$ open. Then $C=X\setminus U$ is closed and $x\notin C$, so there are disjoint open $V\ni x$, $W\supseteq C$. Since $V\subseteq X\setminus W$ and $X\setminus W$ is closed, $\overline{V}\subseteq X\setminus W\subseteq X\setminus C=U$. Conversely, given $x\notin C$ with $C$ closed, apply the condition to $U=X\setminus C$ to get $V$ with $x\in V\subseteq\overline{V}\subseteq U$; then $V$ and $X\setminus\overline{V}$ are disjoint open sets containing $x$ and $C$ respectively. (b) is the same argument with the point $x$ replaced by the closed set $A$. ∎
In words: regularity says that every point has a neighbourhood base of closed neighbourhoods. This is the form in which regularity is used in proofs.
Proof. For any $x,y$ and $a\in A$, $d(x,a)\le d(x,y)+d(y,a)$; taking the infimum over $a$ gives $d(x,A)\le d(x,y)+d(y,A)$, and by symmetry $|d(x,A)-d(y,A)|\le d(x,y)$. So $x\mapsto d(x,A)$ is continuous (1-Lipschitz). Moreover $d(x,A)=0$ iff every ball around $x$ meets $A$ iff $x\in\overline{A}=A$. The function $h(x)=d(x,A)-d(x,B)$ is continuous, so $U=h^{-1}(-\infty,0)$ and $V=h^{-1}(0,\infty)$ are open, and they are disjoint because no point satisfies both strict inequalities. If $a\in A$, then $d(a,A)=0$ while $d(a,B)\gt0$ because $a\notin B$ and $B$ is closed; so $a\in U$. Similarly $B\subseteq V$. Metric spaces are $T_1$ (they are Hausdorff), so $X$ is normal. ∎
Notice that the proof never needed $d(A,B)=\inf\{d(a,b)\}$ to be positive — and it need not be: the closed sets $\{y=0\}$ and $\{xy=1\}$ in $\mathbb{R}^2$ are at distance $0$. Example 5 makes this concrete. (If $A$ is compact, then $d(A,B)\gt0$; see §9.4.)
Proof. Every standard basic set $(a,b)$ is a $K$-basic set, so the $K$-topology is finer; it is strictly finer because $(-1,1)\setminus K$ is $K$-open and contains $0$, but contains no standard open interval around $0$ (every such interval contains points $1/n$). A finer topology than a Hausdorff one is Hausdorff (the same disjoint open sets work), so $\mathbb{R}_K$ is Hausdorff. The complement $\mathbb{R}\setminus K=\bigcup_{n\ge1}\big((-n,n)\setminus K\big)$ is a union of basic sets, so $K$ is closed, and $0\notin K$.
Suppose $U\ni0$ and $V\supseteq K$ were disjoint open sets. $U$ contains a basic set around $0$; a standard interval around $0$ contains points of $K\subseteq V$, so this basic set must be of the form $(a,b)\setminus K$ with $a\lt0\lt b$. Choose $n$ with $1/n\lt b$. Since $1/n\in K\subseteq V$ and $V$ is open, $V$ contains a basic set around $1/n$; this cannot be of the form $(c,d)\setminus K$ (it must contain $1/n\in K$), so it is a standard interval $(c,d)\ni1/n$. Pick $z$ with $\max\!\big(c,\tfrac1{n+1}\big)\lt z\lt\tfrac1n$. Then $z\notin K$, $z\in(c,d)\subseteq V$, and $0\lt z\lt1/n\lt b$ so $z\in(a,b)\setminus K\subseteq U$. Thus $z\in U\cap V$, a contradiction. ∎
The lesson: making a topology finer keeps $T_0,T_1,T_2$ (more open sets can only help to separate points) but can destroy regularity and normality, because a finer topology also has more closed sets that need to be separated.
Proof. (a) Let $Y\subseteq X$, $x\in Y$, and let $C$ be closed in $Y$ with $x\notin C$. Then $C=Y\cap\overline{C}$, where $\overline{C}$ is the closure in $X$ (closure in a subspace is the trace of the closure, §9.2), so $x\notin\overline{C}$. Regularity of $X$ separates $x$ from $\overline{C}$ by disjoint open $U,V$; then $U\cap Y$ and $V\cap Y$ separate $x$ from $C$ in $Y$. $Y$ is $T_1$ by Theorem 9.38.
(b) Let $X=\prod_\alpha X_\alpha$, $x\in X$, and $U\ni x$ open. Some basic open set $\prod_\alpha U_\alpha$ satisfies $x\in\prod U_\alpha\subseteq U$, where $U_\alpha=X_\alpha$ except for finitely many $\alpha$. For those finitely many $\alpha$, Theorem 9.40(a) gives open $V_\alpha$ with $x_\alpha\in V_\alpha\subseteq\overline{V_\alpha}\subseteq U_\alpha$; for the others put $V_\alpha=X_\alpha$. Then $V=\prod V_\alpha$ is open and contains $x$. The set $\prod\overline{V_\alpha}=\bigcap_\alpha\pi_\alpha^{-1}(\overline{V_\alpha})$ is closed and contains $V$, so $\overline{V}\subseteq\prod\overline{V_\alpha}\subseteq\prod U_\alpha\subseteq U$. By Theorem 9.40(a), $X$ is regular ($T_1$ by Theorem 9.38).
(c) If $Y$ is closed in $X$, a set closed in $Y$ is closed in $X$. So two disjoint closed subsets of $Y$ are disjoint closed subsets of $X$; separate them in $X$ and intersect the open sets with $Y$. ∎
Proof. $\mathbb{R}_\ell$ is finer than the standard topology, so singletons are closed and it is $T_1$. Let $A,B$ be disjoint closed sets. For each $a\in A$, the open set $\mathbb{R}\setminus B$ contains $a$, hence contains a basic set $[a,x_a)$. Likewise each $b\in B$ has $[b,y_b)$ disjoint from $A$. Put $U=\bigcup_{a\in A}[a,x_a)\supseteq A$ and $V=\bigcup_{b\in B}[b,y_b)\supseteq B$, both open. If some $z\in[a,x_a)\cap[b,y_b)$, then $a\ne b$; if $a\lt b$, then $a\lt b\le z\lt x_a$, so $b\in[a,x_a)$, contradicting $[a,x_a)\cap B=\varnothing$; if $b\lt a$, symmetrically $a\in[b,y_b)$, contradicting $[b,y_b)\cap A=\varnothing$. So $U\cap V=\varnothing$. ∎
Proof. Let $D$ be a countable dense subset. Because $S$ is discrete, every subset $T\subseteq S$ is closed in $S$, hence closed in $X$ ($S$ is closed). For each $T\subseteq S$ the sets $T$ and $S\setminus T$ are disjoint and closed, so we may choose disjoint open sets $U_T\supseteq T$ and $V_T\supseteq S\setminus T$. Define $\Phi(T)=U_T\cap D\in\mathcal{P}(D)$. We claim $\Phi$ is injective. If $T\ne T'$, say $x\in T\setminus T'$, then $x\in U_T$ and $x\in S\setminus T'\subseteq V_{T'}$. The open set $U_T\cap V_{T'}$ is nonempty, so it contains a point $d\in D$ ($D$ is dense). Then $d\in U_T\cap D=\Phi(T)$, while $d\in V_{T'}$ forces $d\notin U_{T'}$, so $d\notin\Phi(T')$. Hence $|\mathcal{P}(S)|\le|\mathcal{P}(D)|=2^{\aleph_0}=\mathfrak{c}$. If $|S|=\mathfrak{c}$, this would give $2^{\mathfrak{c}}\le\mathfrak{c}$, contradicting Cantor's theorem (Unit 1, §1.1). ∎
Proof. Regular: Theorems 9.44, 9.39 and 9.43(b). Separable: each basic set $[a,b)\times[c,d)$ contains a point of $\mathbb{Q}^2$. Let $L=\{(x,-x):x\in\mathbb{R}\}$. $L$ is closed in the usual topology of $\mathbb{R}^2$, and the Sorgenfrey plane is finer, so $L$ is closed. $L$ is discrete: the open set $[x,x+1)\times[-x,-x+1)$ contains $(t,-t)$ only if $x\le t$ and $-x\le-t$, i.e. only if $t=x$. As $|L|=\mathfrak{c}$, Jones's lemma shows that $\mathbb{R}_\ell^2$ does not satisfy the normality condition. Concretely, the disjoint closed sets $\{(x,-x):x\in\mathbb{Q}\}$ and $\{(x,-x):x\notin\mathbb{Q}\}$ cannot be separated. ∎
A second route to normality, which covers $\mathbb{R}_\ell$ again and every second countable regular space, uses the Lindelöf property: every open cover has a countable subcover. Second countable spaces are Lindelöf (from any cover, keep for each basic set that lies inside some member of the cover one such member), and so is $\mathbb{R}_\ell$ (a classical fact we only quote).
Proof. Let $A,B$ be disjoint closed sets. For each $a\in A$, regularity (Theorem 9.40(a) applied to $X\setminus B$) gives an open $W\ni a$ with $\overline{W}\cap B=\varnothing$. The closed set $A$ is Lindelöf too (add the open set $X\setminus A$ to a cover of $A$), so countably many such sets $W_1,W_2,\dots$ cover $A$, each with $\overline{W_n}\cap B=\varnothing$. Similarly, open $Z_1,Z_2,\dots$ cover $B$ with $\overline{Z_n}\cap A=\varnothing$. Put
$$W_n'=W_n\setminus\textstyle\bigcup_{k\le n}\overline{Z_k},$$ $$Z_n'=Z_n\setminus\textstyle\bigcup_{k\le n}\overline{W_k}.$$These are open (an open set minus a closed set). $U=\bigcup W_n'$ contains $A$: a point $a\in A$ lies in some $W_n$ and in no $\overline{Z_k}$. Likewise $V=\bigcup Z_n'\supseteq B$. If $z\in W_j'\cap Z_k'$ with $j\le k$, then $z\in W_j\subseteq\overline{W_j}$, but $Z_k'$ excludes $\overline{W_j}$; if $j\gt k$, then $z\in Z_k$, which $W_j'$ excludes. So $U\cap V=\varnothing$. ∎
Urysohn's Lemma and Its Consequences
In a metric space the function $f(x)=\dfrac{d(x,A)}{d(x,A)+d(x,B)}$ is continuous (the denominator never vanishes because $A\cap B=\varnothing$ and both sets are closed), equals $0$ exactly on $A$ and $1$ exactly on $B$. Urysohn's lemma says that the much weaker hypothesis of normality — disjoint closed sets have disjoint neighbourhoods — already produces such a function. It is the bridge from "open sets" to "real-valued continuous functions", and from there to metrizability.
Proof (careful sketch). Let $P$ be the set of dyadic rationals $k/2^m$ in $[0,1]$. Step 1: a nested family of open sets. We build open sets $U_p$, $p\in P$, with $A\subseteq U_0$, $U_1=X\setminus B$, and
$$p\lt q\ \Longrightarrow\ \overline{U_p}\subseteq U_q.$$Since $A\subseteq X\setminus B$, Theorem 9.40(b) gives an open $U_0$ with $A\subseteq U_0\subseteq\overline{U_0}\subseteq U_1$. Suppose $U_p$ has been defined for all $p$ with denominator $2^m$, satisfying the condition. A new point $r=(2k+1)/2^{m+1}$ lies between consecutive old points $p=k/2^m$ and $q=(k+1)/2^m$, and $\overline{U_p}\subseteq U_q$. Applying Theorem 9.40(b) to the closed set $\overline{U_p}$ inside the open set $U_q$ gives $U_r$ with $\overline{U_p}\subseteq U_r\subseteq\overline{U_r}\subseteq U_q$. The nesting condition now holds for all points with denominator $2^{m+1}$ (compare any two through the old points between them). Also put $U_p=X$ for dyadic $p\gt1$.
Step 2: the function. Define $f(x)=\inf\{p: x\in U_p\}$ (dyadic $p\ge0$). Then $0\le f\le1$, $f=0$ on $A\subseteq U_0$, and $f=1$ on $B$ because a point of $B$ lies in no $U_p$ with $p\le1$ (all are inside $U_1=X\setminus B$). Two facts follow from the nesting: (i) if $x\in\overline{U_r}$ then $x\in U_s$ for every $s\gt r$, so $f(x)\le r$; (ii) if $x\notin U_r$ then $x\notin U_s$ for every $s\lt r$, so $f(x)\ge r$.
Step 3: continuity. It suffices that $f^{-1}(-\infty,\alpha)$ and $f^{-1}(\alpha,\infty)$ are open for every real $\alpha$, because such rays form a subbasis of $\mathbb{R}$. First, $f(x)\lt\alpha$ iff $x\in U_r$ for some dyadic $r\lt\alpha$ (by definition of the infimum, and by (i)); so $f^{-1}(-\infty,\alpha)=\bigcup_{r\lt\alpha}U_r$ is open. Second, $f(x)\gt\alpha$ iff $x\notin\overline{U_s}$ for some dyadic $s\gt\alpha$: if $f(x)\gt\alpha$, choose dyadic $\alpha\lt s\lt r\lt f(x)$; then $x\notin U_r$ (else $f(x)\le r$), and $\overline{U_s}\subseteq U_r$, so $x\notin\overline{U_s}$; conversely, $x\notin\overline{U_s}$ gives $x\notin U_s$ and $f(x)\ge s\gt\alpha$ by (ii). So $f^{-1}(\alpha,\infty)=\bigcup_{s\gt\alpha}\big(X\setminus\overline{U_s}\big)$ is open. ∎
Proof. $U=f^{-1}\big([0,\tfrac12)\big)$ and $V=f^{-1}\big((\tfrac12,1]\big)$ are open in $X$ (preimages of open subsets of $[0,1]$), disjoint, and contain $A$ and $B$ respectively. ∎
By Urysohn's lemma (with $A=\{x\}$) normal spaces are completely regular, and completely regular spaces are regular ($f^{-1}[0,\tfrac12)$ and $f^{-1}(\tfrac12,1]$ separate $x$ from $C$). So the ladder refines to $T_4\Rightarrow T_{3\frac12}\Rightarrow T_3$. Subspaces and products of completely regular spaces are completely regular (we omit the proof); hence the Sorgenfrey plane is completely regular but not normal. Every completely regular space embeds in a product of copies of $[0,1]$ — a fact we only quote.
Idea of proof. For $f:A\to[-1,1]$, apply Urysohn to $f^{-1}[-1,-\tfrac13]$ and $f^{-1}[\tfrac13,1]$ to get $g_1:X\to[-\tfrac13,\tfrac13]$ with $|f-g_1|\le\tfrac23$ on $A$; repeat with $f-g_1$, and so on. The series $\sum g_n$ converges uniformly (its terms are bounded by $\tfrac13(\tfrac23)^{n-1}$) to the extension. Conversely, Tietze implies Urysohn (extend the function equal to $0$ on $A$ and $1$ on $B$ from the closed set $A\cup B$), so for $T_1$ spaces "normal", "Urysohn functions exist" and "Tietze extensions exist" are equivalent.
(a) Every regular second countable space is metrizable.
(b) Every separable metric space is second countable. Hence a separable space is metrizable if and only if it is regular and second countable.
Proof. (a) Idea: by Theorem 9.47 the space is normal; for each of the countably many pairs of basic sets $B_m,B_n$ with $\overline{B_m}\subseteq B_n$, Urysohn's lemma gives a function $f_{mn}$ equal to $1$ on $\overline{B_m}$ and $0$ off $B_n$, and the countably many $f_{mn}$ define an embedding of $X$ into the metric space $[0,1]^{\mathbb{N}}$ (the Hilbert cube). (b) Let $D$ be a countable dense set. The countable family $\mathcal{B}=\{B(d,1/n):d\in D,n\in\mathbb{N}\}$ is a basis: if $x\in W$ with $W$ open, choose $n$ with $B(x,2/n)\subseteq W$ and $d\in D\cap B(x,1/n)$; then $x\in B(d,1/n)\subseteq B(x,2/n)\subseteq W$ by the triangle inequality. The last statement combines (a), (b) and Theorem 9.41. ∎
- Metric in disguise? If the topology comes from a metric (or is homeomorphic to a metric space), it is $T_4$ and you are done (Theorem 9.41). Discrete spaces are metric.
- Test $T_1$ first: are singletons closed? A finite space is $T_1$ only if discrete. If $T_1$ fails, all of $T_1$–$T_4$ fail (with this course's convention), though the bare normality condition may still hold vacuously.
- Test $T_2$: do any two nonempty open sets meet? (Cofinite and cocountable topologies on large sets: yes, so not Hausdorff.) Is there a sequence with two limits? (Then not Hausdorff.) Was the space made by gluing (a quotient)? Look for doubled points.
- Test $T_3$: look for a closed set that "accumulates" at a point only in a coarser topology — the $K$-topology pattern. Use the shrinking form, Theorem 9.40(a).
- Test $T_4$: separable plus a closed discrete set of size $\mathfrak{c}$ means not normal (Jones's lemma). Regular plus Lindelöf or second countable means normal (Theorem 9.47). Compact Hausdorff means normal (§9.4).
- Use permanence: subspaces and products keep $T_0$–$T_3$; closed subspaces keep $T_4$; products and open subspaces may lose $T_4$; quotients may lose everything; finer topologies keep $T_0$–$T_2$ only.
The zoo of standard examples, which every CSIR Part C question about separation draws on:
| space | $T_0$ | $T_1$ | $T_2$ | $T_3$ | $T_4$ | metrizable |
|---|---|---|---|---|---|---|
| indiscrete, $\ge2$ points | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Sierpiński space | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
| cofinite, infinite set | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ |
| cocountable, uncountable set | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ |
| line with two origins | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ |
| $\mathbb{R}_K$ | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ |
| Sorgenfrey plane $\mathbb{R}_\ell^2$ | ✓ | ✓ | ✓ | ✓ | ✗ | ✗ |
| Sorgenfrey line $\mathbb{R}_\ell$ | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ |
| metric spaces, discrete spaces | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
(The indiscrete and Sierpiński spaces satisfy the bare normality condition vacuously — there are no two disjoint nonempty closed sets — which is exactly why the $T_1$ clause is built into our definitions.)
- Check the convention. "Regular" and "normal" with or without $T_1$; "$T_3$" and "$T_4$" with or without it. Without $T_1$, an indiscrete space is regular and normal, and normal does not imply regular or Hausdorff (Sierpiński space).
- Unique limits do not imply Hausdorff. In the cocountable topology on $\mathbb{R}$ only eventually constant sequences converge, so limits are unique, yet the space is not $T_2$. The converse of Theorem 9.35(a) needs first countability.
- $T_1$ is not $T_2$. The cofinite topology on an infinite set: singletons closed, but any two nonempty open sets meet, and a sequence of distinct points converges to every point.
- Locally Euclidean is not Hausdorff. The line with two origins is locally homeomorphic to $\mathbb{R}$ and $T_1$, but $1/n$ converges to both origins.
- Finer topologies can lose regularity. $\mathbb{R}_K$ is finer than $\mathbb{R}$ and Hausdorff but not regular: $0$ and $K=\{1/n\}$ cannot be separated.
- Products of normal spaces. $\mathbb{R}_\ell$ is normal, $\mathbb{R}_\ell\times\mathbb{R}_\ell$ is not (antidiagonal + Jones's lemma). Products of regular, Hausdorff, $T_1$, $T_0$ spaces are fine.
- Subspaces of normal spaces. Closed subspaces are normal; arbitrary ones need not be: the Sorgenfrey plane, being completely regular, embeds in a product of intervals $[0,1]^J$, which is compact Hausdorff and hence normal (§9.4).
- Continuous images and quotients. A continuous image of a Hausdorff space need not be Hausdorff (discrete $\to$ indiscrete), and the quotient $\mathbb{R}/\mathbb{Q}$ is indiscrete. Only homeomorphisms preserve separation axioms in general.
- Finite spaces. A finite $T_1$ (even more so, Hausdorff) space is discrete. Questions asking for "the number of Hausdorff topologies on an $n$-point set" have answer $1$.
- Distance zero does not prevent separation. In a metric space disjoint closed sets can be at distance $0$ ($\mathbb{N}$ and $\{n+\tfrac1n\}$; the axis and the hyperbola $xy=1$), yet they are still separated by open sets. Distance is positive when one of them is compact.
- Normal is not metrizable. $\mathbb{R}_\ell$ is $T_4$, separable and first countable but not second countable, hence not metrizable. Separable metric spaces are always second countable.
- Agreement on a dense set. Two continuous maps agreeing on a dense set are equal only when the target is Hausdorff.
Exercises
- 1 List all four topologies on $X=\{a,b\}$ and decide which are $T_0$, $T_1$, $T_2$.
- 2 On $X=\{a,b,c\}$ let $\mathcal{T}=\{\varnothing,\{a\},\{b\},\{a,b\},\{a,c\},X\}$. Is $\mathcal{T}$ a topology? Is it $T_0$? $T_1$?
- 3 Prove directly that a finite Hausdorff space is discrete, without quoting Theorem 9.33.
- 4 Let $X$ be an infinite set with the cofinite topology. Show that every infinite subset of $X$ is dense, and that every nonempty open set is dense.
- 5 Show that the cocountable topology on $\mathbb{R}$ is strictly finer than the cofinite topology, and that both are $T_1$ but not $T_2$.
- 6 Let $f,g:X\to Y$ be continuous with $Y$ Hausdorff, and suppose $f=g$ on a dense set. Show $f=g$. Then show by example that the conclusion fails when $Y$ is the Sierpiński space.
- 7 On $\mathbb{R}$ let $\mathcal{T}_{\gt}=\{\varnothing,\mathbb{R}\}\cup\{(a,\infty):a\in\mathbb{R}\}$. Show that $\mathcal{T}_{\gt}$ is a topology which is $T_0$ but not $T_1$, and find all limits of the sequence $x_n=n$ and of the constant sequence $x_n=0$.
- 8 Show that a space $X$ is $T_0$ if and only if distinct points have distinct closures, i.e. $x\ne y\Rightarrow\overline{\{x\}}\ne\overline{\{y\}}$.
- 1 On $X=\{a,b,c\}$ let $\mathcal{T}=\{\varnothing,\{a\},\{a,b\},\{a,c\},X\}$. Find two disjoint closed sets that cannot be separated, so $X$ fails the normality condition. Does it satisfy the regularity condition? Is it $T_0$?
- 2 Show that every discrete space is normal, and that a space with at least two points with the indiscrete topology satisfies the normality condition but is not normal.
- 3 Prove directly that $\mathbb{R}_\ell$ is regular: given $x\notin C$, $C$ closed, find a clopen set containing $x$ and missing $C$.
- 4 Let $X$ be normal and $A,B$ disjoint closed sets. Show that there are open $U\supseteq A$, $V\supseteq B$ with $\overline{U}\cap\overline{V}=\varnothing$.
- 5 In $\mathbb{R}^2$ let $A=\{(x,0):x\in\mathbb{R}\}$ and $B=\{(x,1/x):x\gt0\}$. Show that $A,B$ are disjoint, closed and at distance $0$, and that $U=\{(x,y):2xy\lt1\}$ and $V=\{(x,y):x\gt0,\ 4xy\gt3\}$ separate them.
- 6 In $\mathbb{R}$ let $A=(-\infty,0]$ and $B=[1,\infty)$. Compute the metric Urysohn function $f=d(\cdot,A)/(d(\cdot,A)+d(\cdot,B))$.
- 7 With $A$ the closed unit disc and $B=\{x\ge3\}$ in $\mathbb{R}^2$ (Example 6), compute $f(1.5,1.5)$ to three decimal places and decide whether $(1.5,1.5)\in U$.
- 8 Prove that every regular space $X$ satisfies: for every $x$ and closed $C\not\ni x$ there are open $U\ni x$, $V\supseteq C$ with $\overline{U}\cap\overline{V}=\varnothing$.
- 1 Prove that $X\times Y$ is Hausdorff if $X$ and $Y$ are, and that if $X\times Y$ is Hausdorff and $Y\ne\varnothing$ then $X$ is Hausdorff.
- 2 Let $A\subseteq X$ be a retract of a Hausdorff space $X$: there is a continuous $r:X\to A$ with $r(a)=a$ for $a\in A$. Prove that $A$ is closed in $X$.
- 3 In the Sorgenfrey plane, show that $L_{\mathbb{Q}}=\{(x,-x):x\in\mathbb{Q}\}$ and $L\setminus L_{\mathbb{Q}}$ are disjoint closed sets.
- 4 Show that every subspace $W\subseteq\mathbb{R}\setminus\{0\}$ of $\mathbb{R}_K$ carries the standard subspace topology, so it is metrizable and regular — while $\mathbb{R}_K$ itself is not regular.
- 5 Let $X$ be Hausdorff and $f:X\to X$ continuous with $f\circ f=f$. Show that $f(X)$ is closed.
- 6 Decide (with reason) which of $T_0,\dots,T_4$ hold for $\mathbb{R}_\ell\times\mathbb{R}$ (Sorgenfrey line times the usual line).
- 7 Let $\mathbb{R}/\mathbb{Q}$ be the quotient of $\mathbb{R}$ by $x\sim y\iff x-y\in\mathbb{Q}$. Show that its quotient topology is indiscrete.
- 1 Use Theorem 9.49 to show that if $X$ is $T_1$ and admits, for every pair of disjoint closed sets, a continuous $f:X\to\mathbb{R}$ with $f\le0$ on $A$ and $f\ge1$ on $B$, then $X$ is normal.
- 2 Show that $\mathbb{R}_K$ is not metrizable, and that $\mathbb{R}_K$ is second countable.
- 3 Show that the Sorgenfrey plane is not metrizable, and that it is separable but not second countable.
- 4 Prove that a closed subspace of a normal space is normal by using Urysohn functions instead of open sets (Tietze allowed).
- 5 (Medical imaging.) In a 2-D scan a tumour is modelled as the closed disc $A=\{|p|\le1\}$ (in cm) and the organ boundary as the circle $B=\{|p|=4\}$. A "soft segmentation" map is a continuous $f$ with $f=0$ on $A$, $f=1$ on $B$. For the metric Urysohn function, find $f$ on $1\le|p|\le4$, the radius of the level set $f=\tfrac12$, and $f$ at $|p|=2$ cm.
- 6 (Challenge.) Prove that every infinite Hausdorff space contains an infinite sequence $W_1,W_2,\dots$ of pairwise disjoint nonempty open sets.
E1. CSIR Part B GATE MA Let $X$ be an infinite set with the cofinite topology. Which of the following is true? (A) $X$ is Hausdorff; (B) every sequence of pairwise distinct points of $X$ converges to every point of $X$; (C) some singleton in $X$ is not closed; (D) $X$ is normal.
E2. CSIR Part B Which of the following spaces is Hausdorff but not regular? (A) the Sorgenfrey line $\mathbb{R}_\ell$; (B) $\mathbb{R}$ with the $K$-topology; (C) $\mathbb{R}$ with the cocountable topology; (D) the Sorgenfrey plane $\mathbb{R}_\ell\times\mathbb{R}_\ell$.
E3. CSIR Part B Let $X$ be a set with exactly $4$ elements. The number of topologies on $X$ in which every singleton is closed is (A) $0$; (B) $1$; (C) $15$; (D) $219$.
E4. CSIR Part C Let $\mathbb{R}_\ell$ be $\mathbb{R}$ with the lower limit topology. Which of the following statements are true? (One or more options may be correct.) (A) $\mathbb{R}_\ell$ is normal; (B) $\mathbb{R}_\ell\times\mathbb{R}_\ell$ is normal; (C) $\mathbb{R}_\ell$ is metrizable; (D) $\mathbb{R}_\ell\times\mathbb{R}_\ell$ is regular.
E5. CSIR Part C GATE MA Let $f,g:X\to Y$ be continuous maps between topological spaces. Which of the following are necessarily true? (A) If $Y$ is Hausdorff, $\{x:f(x)=g(x)\}$ is closed in $X$. (B) If $X$ is Hausdorff, $\{x:f(x)=g(x)\}$ is closed in $X$. (C) If $Y$ is Hausdorff, the graph of $f$ is closed in $X\times Y$. (D) If $Y$ is $T_1$, $\{x:f(x)=g(x)\}$ is closed in $X$.
E6. CSIR Part C Let $(X,d)$ be a metric space and $A,B$ disjoint nonempty closed subsets. Which of the following are true? (A) $\inf\{d(a,b):a\in A,b\in B\}\gt0$. (B) There are disjoint open sets $U\supseteq A$, $V\supseteq B$. (C) There is a continuous $f:X\to[0,1]$ with $f^{-1}(0)=A$ and $f^{-1}(1)=B$. (D) If $A$ is compact then $\inf\{d(a,b):a\in A,b\in B\}\gt0$.
Suggested plan — 2 lectures × 60 min.
- $T_0$, $T_1$, $T_2$. Start with Interactive 1: standard line, then the cofinite line and the line with two origins — let the class discover that "shrinking" does not help. Definition 9.20, Theorem 9.33 (finite $T_1$ is discrete; Example 1 on the board), uniqueness of limits and Example 2 (the cocountable topology as the counterexample to the converse), the diagonal criterion and Theorem 9.37. End with Example 3 drawn as two lines glued except at the origin.
- Regular, normal, Urysohn. Definition 9.21 with the convention warning, Theorem 9.40, metric spaces are normal (Theorem 9.41) and Interactive 2 (config "axis and hyperbola" to show distance $0$ does not matter), the $K$-topology (Theorem 9.42 — draw the picture of $(a,b)\setminus K$), the Sorgenfrey line and plane (Theorems 9.44–9.46), Urysohn's lemma (state, explain the dyadic nesting with a picture of nested ovals; leave Step 3 for reading), metrizability and Example 7. Finish with the zoo table and Exam corner E4.
Discussion questions. (1) Why does making a topology finer help $T_2$ but hurt $T_3$? Find the exact step in the proof of Theorem 9.42 where the extra open sets become extra closed sets. (2) In Jones's lemma we never wrote down two closed sets that cannot be separated. Is this an existence proof like Cantor's proof that transcendental numbers exist? (3) Urysohn's lemma produces continuous functions from open sets alone. Where in the proof is normality used, and how many times?
Misconceptions to watch for. "Unique limits $\Rightarrow$ Hausdorff"; "locally like $\mathbb{R}$ $\Rightarrow$ Hausdorff"; "normal $\Rightarrow$ metrizable"; "the product of normal spaces is normal"; "disjoint closed sets are at positive distance"; forgetting the $T_1$ clause (or forgetting that some exam papers omit it); thinking a continuous image of a Hausdorff space is Hausdorff.
Suggested homework. I(1, 2, 4, 7), II(1, 3, 5, 6), III(2, 3, 7), IV(2, 3, 5); Exam corner E2, E5, E6.
Research question. Finite topological spaces can never be Hausdorff unless discrete — yet image-processing software uses finite non-Hausdorff topologies on pixel grids every day. How many $T_0$ topologies are there on an $n$-point set, and why are they the same thing as partial orders?
Why it matters. In digital topology (developed from the 1970s for computer vision and medical imaging) the Khalimsky line — $\mathbb{Z}$ with $\{2k+1\}$ open and $\{2k-1,2k,2k+1\}$ the smallest open set around an even $2k$ — and its square model pixel grids on which a digital Jordan curve theorem holds, so that a closed digital curve really separates the image into an inside and an outside. Such spaces are $T_0$ but not $T_1$, and their topology is completely encoded by a partial order.
What to do.
- For a finite space define $x\preceq y\iff x\in\overline{\{y\}}$. Prove that $\preceq$ is reflexive and transitive, and that it is antisymmetric exactly when the space is $T_0$ (use Exercise I.8).
- Conversely, from a partial order build the topology whose open sets are the up-sets ($y\in U$, $y\preceq z\Rightarrow z\in U$), and check the two constructions are inverse.
- Write a program counting all topologies and the $T_0$ ones on $n=1,2,3,4$ points by brute force; compare with a count of partial orders on $n=5$ points (enumerate the $3^{10}$ choices of comparability for each pair and test transitivity).
- Draw the Khalimsky topology on $\{0,1,2,3,4\}$, list its open sets, and check that it is connected (§9.4) but not $T_1$.
Tools: Python (itertools) or any language with bit operations; a sheet of paper for Hasse diagrams.
Hand in: A one-page note with your table of counts, a proof sketch of the order–topology correspondence, and the open sets of the Khalimsky segment.
Going further. Read about McCord's theorem (1966): every finite $T_0$ space is weakly homotopy equivalent to a simplicial complex, so finite spaces can model the circle and the sphere — a striking use of spaces that fail every separation axiom beyond $T_0$.
🧪 Self-check quiz
Q1. A topological space is $T_1$ if and only if
Q2. In $\mathbb{R}$ with the cocountable topology,
Q3. The $K$-topology on $\mathbb{R}$ (basis: $(a,b)$ and $(a,b)\setminus\{1/n\}$) is
Q4. Which statement is true?
Q5. Urysohn's lemma asserts that in a normal space $X$, for disjoint closed $A,B$,