Read the normal-space and Urysohn chapter first.
Motivation
Urysohn's lemma prescribes two values on two disjoint closed sets. Extension asks for much more: preserve every value of a continuous function already defined on a closed subset. A single separating function will not usually fit all the data, but it can make a bounded correction that reduces the error by a fixed factor. Repeating that correction produces a uniformly convergent series.
The proof has two separate convergence tasks. On the closed subset the residual must approach zero, giving the right values. On the whole space the correction series must converge uniformly, giving continuity. Controlling only the first task would not justify a continuous extension. The geometric error bounds make both tasks work at once.
We retain the preceding convention: normal means and separation of disjoint closed sets by disjoint open neighborhoods. The proofs below use that closed-set separation property through Urysohn's lemma. They require no metric or countability assumption on the domain.
Definitions
Definition 8.2.1
(Extension of a continuous function)
Let . An extension of is a continuous with . Continuity of always means continuity in the subspace topology on ; an arbitrary assignment of values is not enough.
For real functions on any set, a bound with makes converge uniformly: each tail is bounded independently of by the corresponding numerical tail. This definition uses the absolute-value distance in the target, not a metric on .
Definition 8.2.2
(Partial sums and residuals)
Write for a partial sum of corrections and for the residual. The index counts completed corrections. Empty sums are zero.
Definition 8.2.3
(Interior and boundary of a closed set)
The interior is the union of all open subsets of contained in . For closed , its boundary is . These will locate the compatibility condition in the zero-extension exercise.
Theorem/proposition
Theorem 8.2.4
(One bounded correction reduces the error)
Let be closed in a normal space , let , and let be continuous. There is a continuous such that
This is an approximation on , not yet an exact extension of .
Theorem 8.2.5
(Tietze's extension theorem, bounded and unbounded)
If is closed in a normal space , every continuous has a continuous real-valued extension to . If for real , an extension can be chosen with its whole image in the same interval.
Theorem 8.2.6
(Extension characterizes normality among spaces)
If a space extends every continuous bounded real function on every closed subset to a continuous real function on the whole space, it is normal. Hence the closed-subset scalar extension property and normality are equivalent under the stated convention.
Proof sketch or proof idea
Proof
(Thresholds and one Urysohn function)
For the correction lemma, the two sets
are disjoint and closed in . Closedness of makes them closed in . Urysohn gives equal to zero on and one on . Put . On , , so lies between and zero. On , it lies between zero and . At the remaining points, both and lie between and , giving the same error bound. The case of an identically zero bound, when needed, is handled by the zero extension rather than by disjoint sets with overlapping thresholds.
Proof
(Geometric corrections for bounded data)
First extend . Set . Begin with . Given the continuous residual , apply the lemma to obtain a continuous correction on with
The next residual is . Induction proves every residual bound. On all of ,
Consequently converges uniformly and has values in . On , , so .
Here is the continuity check for an arbitrary domain. Given and , choose so the uniform tail is less than . The finite sum is continuous, so some neighborhood of has . For , the two tails plus this difference bound by . Thus uniform convergence preserves continuity without any sequence criterion in . Affine rescaling proves the interval statement when ; when a constant extension suffices.
Proof
(Compression and endpoint removal for unbounded data)
For an unbounded , the homeomorphism
compresses its values. The bounded result extends to . We cannot immediately apply : may reach an endpoint outside .
Let . It is closed and disjoint from . Urysohn gives with on and on . Set . On , ; off , and , so . On , . Therefore is a well-defined continuous extension on all of . No uniform gap from the endpoints is needed; strict inequality holds at each point.
Proof X-ray
Two error estimates have different jobs
The residual estimate holds only on the original closed subset and forces agreement there. The absolute correction estimate holds throughout and makes the series uniformly convergent. Neither estimate alone proves the theorem. The last step for unbounded functions is separate again: it removes possible endpoint values before applying the inverse compression map.
Proof
(Extension forces normal separation)
For the converse, take disjoint closed and let . The function on equal to zero on and one on is continuous, since the two sets are both closed and open in . An assumed continuous extension has disjoint open inverse images of and containing . This is the normal separation property, and the stipulated condition completes normality.
Worked examples
Worked example 8.2.7
(A first correction is deliberately smaller than the data)
For and , choose . It is continuous, globally bounded by , and differs from on by at most . Near zero it already agrees exactly; near the endpoints a residual remains. The construction is not trying to solve the whole problem in one step. Its globally small size leaves a summable budget for later corrections.
Worked example 8.2.8
(Unbounded data on a closed discrete set)
The positive integers form a closed discrete subset of , so is continuous there. An explicit extension joins successive data points linearly:
and takes the constant value for . Adjacent formulas agree at each integer. Each real point has a neighborhood meeting at most two relevant pieces, proving local continuity; no unjustified pasting over an accumulating infinite family is involved. The extension is unbounded, as the prescribed data require.
Worked example 8.2.9
(Extending two coordinates does not extend a circle-valued map)
Each coordinate of the identity map on extends to the disk as the corresponding Cartesian coordinate. Thus maps into extend coordinate by coordinate. But the identity has no continuous extension : it would be a retraction, excluded by the fundamental-group calculation. Normality of the metric disk does not repair a target-space obstruction. Normalizing the coordinate extension fails at the origin, where both coordinates vanish.
Worked example 8.2.10
(An extension can be localized with a cutoff)
Suppose is continuous, is closed, and with open. Extend boundedly to , and use the previous chapter's cutoff with on and . Then is still an extension, still bounded by one, and has support in : its nonzero set is contained in that of . This combines two different constructions instead of assuming the original extension automatically vanishes outside a desired region.
Worked example 8.2.11
(Finite partitions and supported local patching)
Let be a finite open cover of the normal space. Replace the sets one at a time by smaller opens while retaining a cover. At step , the closed set
lies in , because the current family still covers . Shrink it to . The updated family still covers, and at the end the closed sets cover inside their respective .
Choose cutoffs equal to one on , with support in . Then , so are continuous nonnegative functions with sum one and . This is a finite partition of unity subordinate to the cover. The sum in the denominator never vanishes because the closed shrunk sets cover; merely choosing one arbitrary bump in each original open set would not ensure that.
If are continuous local data, extend by zero off . This is continuous even when is unbounded: every point outside has a neighborhood on which is identically zero, by its support condition. The finite sum of these extensions is a continuous weighted blend. If the local data agree on overlaps, the sum equals their common value wherever defined, since the active weights add to one. Without agreement it still provides a continuous blend, not an exact simultaneous extension of incompatible data. Only a finite cover was used; no assertion about partitions for arbitrary covers or paracompactness has been proved.
Counterexample mode
Zero outside the closed set is not a general extension rule
On , take . Assigning zero off gives jumps at both endpoints. Tietze gives a continuous extension, such as the constant one function, but it does not promise that arbitrary outside values can be prescribed simultaneously. A tapered cutoff is possible if a larger open transition region is supplied. The failure is compatibility at the boundary, not a lack of normality of the line.
Common mistakes
Common mistake
(Dropping closedness or continuity of the data)
Closedness of is used when level sets closed in must be closed in . Continuity of the given data is a separate hypothesis. Boundedness alone does not make arbitrary data extendible. Conversely, unboundedness is allowed in the full real-valued theorem and must not be ruled out by the bounded proof's first step.
Common mistake
(Pointwise convergence of corrections is not enough)
A uniformly convergent series gives continuity because its terms are continuous and its tails are uniformly small on the whole domain. Pointwise convergence alone would not justify this step. The theorem also gives existence, not a unique extension or an extension into every possible target. No smoothness or linearity of the extension procedure is asserted.
Summary
Urysohn separation produces a correction with a fixed error reduction. A geometric series controls both residuals on the closed subset and tails on the whole space. Bounded extension follows, while compression plus endpoint removal handles arbitrary real values. The examples identify the independent roles of closedness, continuous data, support control and the target space.
Exercises
Quick check
(How many terms of the normalized correction series suffice to make the uniform tail bound smaller than ? State a general bound for .)
Solution · Answer
After terms the bound is . It is below at , whereas at it is still above . In general any integer suffices. The indexing means terms have been included.
Quick check
(Extend a continuous to a strictly positive function on . Then explain how to extend data taking values in without reaching one outside .)
Solution · Answer
Extend the real function by Tietze and exponentiate; this is continuous, positive everywhere and agrees with . For , first extend the nonnegative real function to a real function . Replace it by , preserving its values on , then use , which stays below one everywhere and recovers .
Quick check
(For closed and continuous , prove that the function equal to on and zero off is continuous if and only if vanishes on .)
Solution · Answer
If the zero extension is continuous and is a boundary point of with , an interval about excluding zero pulls back to a neighborhood of inside , contradicting that it is a boundary point. Conversely, the closed sets and cover . On their intersection the functions and zero agree. The finite closed-set pasting lemma therefore gives continuity.
Quick check
(Give a bounded continuous function on that has no continuous extension to zero, and prove the obstruction.)
Solution · Answer
Use . The sequences and approach zero through the domain, with values one and minus one. A continuous extension at zero would have to have both limiting values, impossible. Here the data are continuous and bounded, but the domain subset is not closed.
Quick check
(On the closed set in , compare the data with the alternating assignment . Can either be extended continuously?)
Solution · Answer
The first data are the restriction of the continuous function , hence extend. The alternating data have no limit along , regardless of the value assigned at zero: even and odd subsequences give different limits. Thus they are not continuous on and cannot be the restriction of a continuous extension. Closedness of does not compensate for discontinuous input data.