Motivation
Hausdorff separation supplies disjoint neighborhoods of two points. Constructing a continuous real function requires more coordinated information: near a point where the value is small, nearby values must also remain small, and similarly for large values. Normality supplies the needed space between closed sets. A repeated shrinking construction organizes that space into dyadic levels, and taking an infimum turns the levels into a continuous function.
The closure condition is the important extra ingredient. Nested open sets alone can have boundaries that jump, producing a discontinuous infimum. We will prove the neighborhood-shrinking lemma first, use it at every dyadic insertion, and check the two kinds of inverse images separately. No sequence criterion for continuity in an arbitrary space will be assumed.
Definitions
Definition 8.1.1
(, Hausdorff and regular spaces)
A space is if every singleton is closed. It is Hausdorff if distinct points have disjoint open neighborhoods. We say regular to mean together with separation of any point and any disjoint closed set by disjoint open neighborhoods.
Definition 8.1.2
(Normal separation property and normal spaces)
Call the property that any two disjoint closed sets have disjoint open neighborhoods the normal separation property. In this note a normal space means a space with this property. Some books omit from the word normal and use for their conjunction. Thus compare definitions before transferring a statement. The dyadic function construction itself needs the normal separation property, not ; the latter is needed to apply it to arbitrary singleton sets.
Definition 8.1.3
(Support of a function)
For a real-valued continuous function , its support is . Being zero off an open set does not by itself place the support inside : the closure can reach the boundary.
Definition 8.1.4
(Dyadic rationals in the unit interval)
Write and for the dyadics in . Between any two distinct real numbers in this interval there is a dyadic: choose a mesh smaller than their difference. We use the usual classical set-theoretic background for the successive choices of neighborhoods.
Theorem/proposition
Theorem 8.1.5
(Shrinking an open neighborhood of a closed set)
A space has the normal separation property if and only if, whenever is closed and with open, there is an open such that
All inclusions here allow equality.
Theorem 8.1.6
(Compact Hausdorff spaces and closed subspaces)
Every compact Hausdorff space is normal under the stated convention. Every closed subspace of a normal space is normal. Moreover normal spaces are regular and Hausdorff under these conventions.
Theorem 8.1.7
(Urysohn's lemma and its converse)
If are disjoint closed subsets of a space with the normal separation property, there is a continuous with and . Conversely, the existence of such a function for every disjoint closed pair implies the normal separation property. For a space it is therefore equivalent to normality.
Proof sketch or proof idea
Proof
(Shrinking is equivalent to normal separation)
For shrinking, separate the closed sets and by disjoint open sets . Then , since the closed set contains . Conversely, given disjoint closed , shrink inside . The sets and are disjoint open neighborhoods of and .
Proof
(Compact Hausdorff spaces and closed subspaces are normal)
A Hausdorff space is : for a point , every other point has an open neighborhood excluding , and their union is . In a compact Hausdorff space disjoint closed are compact. For each , Hausdorff separation from each point of and a finite subcover of give disjoint opens and : take a finite intersection on the side and a finite union on the side. Now finitely many cover . Their union and are the required disjoint open neighborhoods. Empty sets cause no difficulty. For a closed subspace , sets closed in are also closed in , so restrict the separating open sets to ; is inherited. Normality separates a closed singleton from any disjoint closed set, proving regularity, which separates two closed singletons and proves Hausdorffness.
Counterexample mode
Compact and T1 is not enough
Claim: the compactness argument needs only closed points, so compact spaces are normal. Take an infinite set with the cofinite topology. It is , and it is compact: any nonempty member of an open cover misses only finitely many points, each covered by one further member. Yet for distinct , any open and have finite complements, so is cofinite in an infinite set and nonempty. The closed sets cannot be separated. Repair: the proof separates a point from a compact set using Hausdorffness, which cannot be weakened to .
Proof
(Dyadic levels and the infimum function)
For Urysohn's lemma set . Shrinking gives an open with . Suppose the opens for one finite dyadic level have been chosen, with the closure of each lying inside the next. Between consecutive old levels , insert by shrinking the closed set inside . Thus
Keep the old choices unchanged. Finite induction supplies each level. Any in eventually appear in one common finite level, so its chain gives .
Define
The added makes the infimum defined even when lies in no level set. We have . Points of lie in , so have value zero. Points of lie in none of the , since all are contained in , so have value one.
For ,
Indeed an actual membership level below makes the infimum smaller than , and an infimum below supplies such a level. The artificially added cannot be that witness. Thus these inverse images are open.
The other direction uses closures:
If , it lies in no with , so . Conversely, if , choose dyadics . If , then , contradicting . Hence . These sets too are open. The two families are the inverse images of the basic open rays in and generate its topology. At the endpoints the same formulas handle neighborhoods and , so no endpoint continuity is missing.
For the converse, disjoint opens and contain and . These intervals are open in the subspace . This proves precisely the normal separation property; it does not manufacture without a separate hypothesis.
Proof X-ray
Why the second inverse-image calculation needs a buffer
A point in gives an upper bound for its infimum, so ordinary nesting already controls small values. To control large values on an open neighborhood, we need a point outside a closed set. The relation makes that exterior detectable from . This is the exact job of the inserted closure buffers, rather than a decorative strengthening of nesting.
Worked examples
Worked example 8.1.8
(Metric separation does not need a uniform gap)
For nonempty disjoint closed subsets of a metric space, put
The triangle inequality gives , so each distance function is continuous. Closedness gives exactly on . The denominator is therefore positive at every point, proving continuity and the required values. Pulling back the two separated intervals proves metric spaces normal, without using Urysohn's lemma circularly.
There need not be one positive distance between the sets. In , take the horizontal axis and . Both are closed and disjoint, but and are distance apart. The pointwise denominator stays positive although the global infimum distance is zero.
Worked example 8.1.9
(Two conventions that must not be conflated)
An infinite cofinite space is because finite sets are closed. It is not Hausdorff: two nonempty cofinite opens always intersect. In particular two different closed singletons cannot have disjoint open neighborhoods, so it is not normal.
A two-point indiscrete space has the opposite lesson. Its only closed sets are the empty set and the whole space, so it satisfies the normal separation property. It is not , hence is not normal by our convention. Urysohn's conclusion for its closed pairs is valid but vacuous as a point-separation assertion. In fact every continuous function from this space to is constant: two different values would have disjoint nonempty open inverse images.
Worked example 8.1.10
(Hausdorff need not mean regular)
On , let and take as a basis all ordinary open intervals together with the sets . Intersections have one of these forms, so this is a basis. The resulting topology is finer than the ordinary one, hence Hausdorff. Also is closed, since its complement is open, and .
Every neighborhood of zero contains for some . Suppose an open were disjoint from . Choose . A basic neighborhood of this point cannot be a punctured interval, since those exclude , so contains an ordinary interval around . It contains points not in and still inside , contradicting disjointness from . Thus the point and closed set cannot be separated: the space is not regular and therefore not normal. Hausdorffness alone supplies no Urysohn construction.
Worked example 8.1.11
(A cutoff with support really inside the open set)
For closed with open in a normal space, first choose . Apply Urysohn to and , and subtract its function from one. The resulting is one on , zero outside , and has support contained in .
For , , an explicit choice is one for , the linear value for , and zero for . Its support is . By contrast vanishes off but its support includes the endpoints. Pointwise vanishing and support containment are different claims.
Counterexample mode
Nested open sets alone can give a jump
Let . Take for dyadics , for , and . These open sets are nested. Their infimum function is on negative numbers and one on nonnegative numbers, so is discontinuous at zero. The missing relation is closure containment between positive levels below one: does not lie in . Urysohn's construction repairs exactly this defect.
Common mistakes
Common mistake
(Leaving the normality convention implicit)
Normality is not just point separation, and its convention may include or leave it separate. Declare the convention before deducing Hausdorffness. The compactness proof uses closed subsets being compact, then finite intersections of neighborhoods; an arbitrary infinite intersection would not be known open.
Common mistake
(Treating prescribed values as exact level sets)
The Urysohn function need not take value zero only on , nor one only on . The conclusion prescribes values there without prescribing exact level sets. Its continuity proof uses open inverse images, so it remains valid in spaces whose continuity cannot be tested by sequences. The closed-subspace theorem also uses closedness of that subspace; the argument asserts no general hereditary-normality theorem.
Summary
Normal separation is equivalent to a closure-buffered shrinking property and to continuous separation of disjoint closed sets. The dyadic proof converts those buffers into two open-ray inverse-image formulas. Compact Hausdorff and metric spaces provide important sources of normality, while the examples separate the roles of , Hausdorffness, regularity and support control.
Exercises
Quick check
(If the closedness assumption is dropped from the metric example, can its denominator vanish even for disjoint sets? Give an example and locate the failed step.)
Solution · Answer
Take and in the usual line. They are disjoint but , so the formula is undefined there. Distance zero detects membership in the closure, which coincides with the set only under closedness.
Quick check
(In a normal space, interpolate prescribed values at finitely many distinct points by a continuous function to .)
Solution · Answer
Each singleton and each finite union of the other singletons is closed. Choose equal to one at and zero at the other points by Urysohn. The continuous sum has the prescribed point values and is nonnegative. Set , a continuous function into with the same values, since each . No assertion of disjoint supports was needed.
Quick check
(For a Urysohn family prove for each dyadic . Must either inclusion be equality?)
Solution
If , some level contains , and nesting gives . Membership in implies . Equalities need not hold. On a one-point space, choose empty for and the whole space for . Closure nesting holds, but and is smaller than . Choosing the whole space already at instead makes larger than .
Quick check
(Prove that a closed set in a normal space is an intersection of zero sets of continuous functions, without claiming that one zero set always suffices.)
Solution · Answer
Intersect over all continuous vanishing on . Every such set contains . If , Urysohn applied to the closed pair gives one such with , so the intersection excludes . This family is a set of functions, and does not require choosing one function simultaneously for every outside point. The argument proves an intersection representation, not that itself is a single zero set.
Quick check
(Let be all continuous functions on a normal space. Prove the evaluation map , , is a topological embedding.)
Solution · Answer
Each coordinate of is continuous, so the product universal property makes continuous. Distinct points are distinguished by a Urysohn function on their closed singleton pair, making it injective. For open and , choose with and on . The product-open set pulls back to a neighborhood of contained in . Taking the union of these image-relative neighborhoods shows is open in . Hence the inverse to the continuous bijection is continuous. No countability or metrizability assertion follows from the possibly large coordinate family.