A measurable function can have sharp jumps and still be integrable. Smooth functions are easier to differentiate, but replacing a rough function by a smooth one is useful only when the replacement converges in a specified sense. Averaging nearby translates gives a systematic construction. Its success in for finite rests on translation continuity; a jump shows why the same argument cannot give uniform approximation in general.
Throughout, measure is Lebesgue measure on , , and . Real or complex functions equal almost everywhere represent the same element. We use Hölder, Tonelli, dominated convergence, translation invariance of Lebesgue measure and its inner/outer regularity. For each class choose a finite-valued Borel representative, available after modifying a Lebesgue-measurable representative on a null set; this also makes jointly measurable. Define .
Why translates converge in mean
Lemma 4.4.1
(Density of continuous compactly supported functions)
is dense in .
Proof
(Simple functions, then Urysohn-type cutoffs)
Simple functions supported on finite-measure sets are dense in : truncate both the size and support of and approximate the resulting bounded measurable function by simple functions. For a measurable of finite measure, regularity gives compact with open and . If is empty, use the zero function. Otherwise choose with the closed -neighborhood of contained in , and put . Then , , on , and outside . Hence . Approximate each term of a finite simple sum in this manner and use the triangle inequality.
Theorem 4.4.2
(Continuity of translation in Lp)
For every , as .
Proof
(Prove it for continuous functions, then approximate)
For , uniform continuity and compact support imply pointwise convergence dominated, for , by a constant times the indicator of a fixed compact set. Dominated convergence gives . Given , choose with . Translation invariance gives , so
First take , then .
Worked example 4.4.3
(Translating an indicator)
For on and , the function is the indicator of the symmetric difference of and , which is of measure . So . In contrast, for every such .
The last computation is not an accident of the example; it marks exactly where the theorem stops.
Counterexample mode
Translation is not continuous in $L^\infty$
Claim: the translation theorem also holds for . Take . For every , on the set of positive measure, so , which does not tend to . The proof breaks at the density step: is not dense in , because a uniform limit of continuous functions is continuous, and no continuous satisfies (if the error were at most some , letting from both sides would give and , so ). The hypothesis is essential.
Averaging translates
Definition 4.4.4
(Mollifier and mollification)
A mollifier is a nonnegative supported in the closed unit ball with . For put . The mollification of is
One choice of is a normalized multiple of for , extended by zero elsewhere; it is even. Each is nonnegative, has mass one, and is supported in the closed ball of radius . The integral (1) is finite for every for a locally integrable representative: by Hölder and the kernel is bounded and compactly supported. Changing the representative on a null set leaves (1) unchanged.
Theorem 4.4.5
(Approximation by mollification)
For , each is smooth and as . If is supported almost everywhere in a compact set , then is supported in .
Proof
(Jensen, Tonelli and translation continuity)
Regard as a probability measure. Jensen's inequality applied to gives
for almost every . Jensen and Tonelli also give , so . Integrating the displayed error bound and using Tonelli,
The last step is precisely translation continuity at zero; no uniform pointwise continuity of is assumed.
For smoothness, change variables in (1) to write . Differentiating the kernel under the integral is justified locally by dominated convergence: near a fixed , all relevant lie in one compact set, is integrable there, and each kernel derivative is bounded. Repeating gives
for every multi-index ; continuity follows by the same domination. If almost everywhere off compact , the integrand vanishes whenever , proving the support assertion.
Proof X-ray
Which property of the kernel does which job
The proof separates three roles. Mass one and nonnegativity make a probability measure, which is what licenses Jensen's inequality and turns the error into an average of . Support in the ball of radius restricts that average to , so translation continuity at zero closes the convergence argument; this is the only place where enters. Smoothness and compact support of are used only for differentiating under the integral, and say nothing about convergence. The argument controls the integral of the error, not its maximum, so it gives no uniform statement at a jump.
Common mistake
(Expecting mollification to supply compact support)
Mollifying an arbitrary gives a smooth function, but its support is as large as that of , enlarged by . To obtain smooth compactly supported approximants, first choose with small and then mollify ; the support conclusion applies to .
Worked example 4.4.6
(Mollifying an indicator)
On let , take an even mollifier and . Then . Since vanishes off , on and off ; by evenness . So outside , a set of measure , and gives . Hence . Just to the right of , however, while is close to , so the uniform error does not tend to .
Worked example 4.4.7
(Mollification preserves mass)
If , Fubini and translation invariance give
Absolute integrability of the product integrand follows from . For the indicator above, for every , even though is spread over .
What changes at the endpoint
On , let . No continuous can satisfy . Indeed, if the essential supremum were below , continuity would force for every and for every after choosing a slightly larger error bound still below . Letting from both sides gives incompatible bounds on . Thus even smooth approximants need not converge in ; point values at the jump are irrelevant to this essential-supremum argument.
Practice
Quick check
(For and , compute and .)
Solution · Answer
The symmetric difference has measure ; the answers are and .
Quick check
(If is supported in , where can be nonzero?)
Solution · Answer
The support is contained in .
Quick check
(If is nonnegative, what are the sign and integral of ?)
Solution · Answer
The integrand in (1) is nonnegative almost everywhere, so , and its integral equals by mass preservation.
Quick check
(Can continuous functions converge to in ? Explain.)
Solution · Answer
No: the endpoint argument shows every continuous function has essential-supremum error at least .