Common Errors in Covering and Differentiation Arguments

What Covering and Differentiation Arguments Actually Say

Covering arguments often replace a complicated family of balls by a pairwise disjoint subfamily. The replacement is useful because the measures of disjoint balls add, but it does not usually preserve the original union: the selected balls may leave points uncovered. The missing points are handled by enlarging selected balls, and the enlargement has a measurable cost. Differentiation arguments have a different source of error: their conclusions hold almost everywhere, and pointwise values can change on a null set without changing any ball averages.

Throughout, balls are open Euclidean balls $B(x,r)=\{y\in\mathbb{R}^n:|y-x|<r\}$ with $r>0$, and $m$ is Lebesgue measure on $\mathbb{R}^n$. Lebesgue measure is complete and sigma-finite. For a measurable set $E$, $m(E)$ may be infinite. For a real-valued measurable function $f$, local integrability means $\int_K |f|\,dm<\infty$ for every compact set $K\subset\mathbb{R}^n$. A locally integrable function has a finite integral of its absolute value over every ball.

Definition: For a real-valued $f\in L^1_{\mathrm{loc}}(m)$, its centered Hardy–Littlewood maximal function is $$ Mf(x)=\sup_{r>0}\frac{1}{m(B(x,r))}\int_{B(x,r)}|f(y)|\,dm(y),\qquad x\in\mathbb{R}^n. $$ The supremum may be $+\infty$. It uses averages of $|f|$, not signed averages of $f$.

In the differentiation theorem, the relevant pointwise conclusion for $f\in L^1_{\mathrm{loc}}(m)$ is that $\lim_{r\downarrow0}m(B(x,r))^{-1}\int_{B(x,r)}|f(y)-f(x)|\,dm(y)=0$ for $m$-almost every $x$. This is the Lebesgue Differentiation Theorem on $\mathbb{R}^n$, established in the earlier material on differentiation. “Almost every” cannot be replaced by “every,” even though changing a function on a null set leaves all its ball integrals unchanged.

A Disjoint Selection Is Not a Cover

Worked Example: Two Overlapping Intervals

Work in $\mathbb{R}$ with Lebesgue measure $m$. Consider $B_1=B(-\tfrac12,1)=(-\tfrac32,\tfrac12)$ and $B_2=B(\tfrac12,1)=(-\tfrac12,\tfrac32)$. The intervals intersect, so a pairwise disjoint subcollection can contain at most one of them. If the selection keeps $B_1$, then the point $1$ belongs to $B_2$ but not to $B_1$: indeed $|1-\tfrac12|=\tfrac12<1$, whereas $|1-(-\tfrac12)|=\tfrac32>1$. Thus the selected interval does not cover $B_1\cup B_2$.

The correction is not to claim that disjoint intervals cover the union. Instead, enlarge the selected interval. Its threefold dilation about its center is $3B_1=B(-\tfrac12,3)=(-\tfrac72,\tfrac52)$, which contains both original intervals. Their union has measure $3$, while $m(3B_1)=6$. This example exhibits the role of enlargement: the covering conclusion concerns dilated selected balls, and the estimate must account for their larger measure. Check: $m(B_1)=2$, $m(B_2)=2$, and their intersection has length $1$, so $m(B_1\cup B_2)=2+2-1=3$, as stated.

Here is the finite selection argument that justifies this type of correction. It is a special finite-family argument; it does not assert that an arbitrary infinite family can be handled by the same greedy procedure. For a finite family of balls, choose a ball of largest radius, retain it, discard every ball intersecting it, and repeat on the balls that remain. The retained balls are pairwise disjoint. If a discarded ball $B(c,r)$ intersects a retained ball $B(d,R)$ chosen earlier, then $r\le R$ and $|c-d|<r+R$. For $z\in B(c,r)$,

|z-d|\le |z-c|+|c-d|<r+(r+R)=2r+R\le3R.

Consequently $B(c,r)\subset 3B(d,R)$. Every ball in the original finite family is either retained or discarded, so the union of the original family is contained in the union of the threefold dilations of the retained balls. This proves the claimed finite covering statement. The selection does not say that the retained balls themselves cover the original family.

Takeaway: In a finite family, a greedy largest-radius selection produces pairwise disjoint balls whose threefold dilations cover the original union. In the full Vitali covering theorem and the fivefold covering lemma, the precise hypotheses and conclusion matter; these are established results from the earlier material on covering lemmas. Neither result licenses the claim that a disjoint selection alone covers the original set.

Dilation Changes the Measure

For $a>0$, write $aB(x,r)=B(x,ar)$. The scaling rule for Lebesgue measure gives $m(aB)=a^n m(B)$. Therefore a covering by dilated balls produces a factor in a measure estimate; dilation is not a measure-preserving operation.

Worked Example: The Cost of a Fivefold Dilation

In $\mathbb{R}$, take $B=B(0,2)=(-2,2)$. Its measure is $m(B)=4$. The fivefold dilation is $5B=B(0,10)=(-10,10)$, and $m(5B)=20$. Thus $m(5B)=5m(B)$, not $m(B)$. In $\mathbb{R}^n$, a ball of radius $2$ has measure $v_n2^n$, where $v_n=m(B(0,1))$; its fivefold dilation has measure $v_n10^n=5^n v_n2^n$. Check: the ratio is $(v_n10^n)/(v_n2^n)=5^n$, and for $n=1$ this is $5$, matching the interval calculation.

This is why covering estimates retain a dimension-dependent constant. For example, if a set $A$ is covered by dilated selected balls $5B_j$, countable subadditivity gives $m(A)\le\sum_jm(5B_j)=5^n\sum_jm(B_j)$. If the $B_j$ are pairwise disjoint and all lie in a measurable set $U$, then countable additivity gives $\sum_jm(B_j)=m(\bigcup_jB_j)\le m(U)$, hence $m(A)\le5^n m(U)$. The containment in $U$ is needed for the last inequality; disjointness by itself does not bound the sum by $m(U)$.

Common mistake. A tempting but incorrect line is $m(5B_j)=m(B_j)$, followed by an estimate that omits the dilation factor. For the interval $B=(-2,2)$, this would assert $20=4$. The correction is to use the scaling identity $m(5B)=5^n m(B)$, and then carry the factor $5^n$ through the sum. Similarly, one must not infer that disjoint selected balls cover a level set unless the relevant covering theorem actually gives that conclusion, possibly after dilation.

Pointwise Representatives and Maximal Functions

Two functions equal $m$-almost everywhere have the same integrals of their absolute values over every ball, even if they take different values at individual points. To make a statement about signed integrals, however, one must also ensure that those integrals exist. Local integrability supplies that condition.

Proposition: If $f,g\in L^1_{\mathrm{loc}}(m)$ are real-valued measurable functions on $\mathbb{R}^n$ and $f=g$ $m$-almost everywhere, then their signed integrals over every ball agree and $Mf(x)=Mg(x)$ for every $x\in\mathbb{R}^n$.

Proof. Fix a ball $B$. Since $f=g$ outside an $m$-null set, $|f|=|g|$ there as well, and $f-g=0$ almost everywhere on $B$. Local integrability gives $\int_B|f-g|\,dm\le\int_B(|f|+|g|)\,dm<\infty$. Therefore $\int_B(f-g)\,dm=0$, so $\int_Bf\,dm=\int_Bg\,dm$; both signed integrals exist and are finite. Also $\int_B|f|\,dm=\int_B|g|\,dm$. For each fixed $x$, this equality holds for every ball $B(x,r)$, so the averages agree for every $r>0$. Taking their suprema gives $Mf(x)=Mg(x)$. Since $x$ was arbitrary, the equality holds everywhere. $\square$

Worked Example: A Null-Set Modification Can Break Differentiation at a Point

On $\mathbb{R}$ with Lebesgue measure, let $f=\chi_{\mathbb{Q}}$, where $\mathbb{Q}$ is the set of rational numbers. The rationals are countable and therefore have measure zero. For every $x\in\mathbb{R}$ and $r>0$, the integral $\int_{B(x,r)}|f|\,dm=0$, so $Mf(x)=0$ for every $x$.

If $x\in\mathbb{Q}$, then $f(x)=1$, while $f(y)=0$ for almost every $y$ in $B(x,r)$. The interval has length $2r$, and hence $$ \frac{1}{m(B(x,r))}\int_{B(x,r)}|f(y)-f(x)|\,dm(y) =\frac{1}{2r}\int_{B(x,r)}1\,dm(y)=\frac{2r}{2r}=1. $$ This average does not tend to zero as $r\downarrow0$. If $x\notin\mathbb{Q}$, then $f(x)=0$, and the same average is $0$, because $f=0$ almost everywhere on the ball. Thus differentiation fails at each rational point and holds at each irrational point. Check: the exceptional set is $\mathbb{Q}$, which is countable and has measure zero; the example contradicts “at every point,” not the almost-everywhere differentiation theorem.

The maximal function and the differentiation conclusion answer different questions. The maximal function uses averages of $|f|$, so it is unchanged everywhere by a null-set modification when the functions are locally integrable. The differentiation expression also compares those averages with the specific point value $f(x)$; that value can change on the null set. Therefore an almost-everywhere differentiation conclusion cannot be promoted to a claim about every point of a chosen representative.

When Signed Integrals Are Not Defined

A further error is to infer that signed integrals agree merely because two functions agree almost everywhere. The inference is valid when both signed integrals exist, as in the proposition, but not when their positive and negative parts both have infinite integral. In that case the Lebesgue integral is undefined, rather than a number that can be compared.

Worked Example: Equality Almost Everywhere Does Not Ensure Integrability

On $\mathbb{R}$, define $f(x)=g(x)=1/x$ for $x\ne0$, and set $f(0)=g(0)=0$. The functions agree everywhere, but neither is locally integrable near $0$. On $(-1,1)$, the positive part has integral $\int_0^1(1/x)\,dx=+\infty$, and the negative part also has integral $\int_{-1}^0(1/|x|)\,dx=+\infty$. Thus the signed Lebesgue integral over $(-1,1)$ is undefined: it would require subtracting $+\infty$ from $+\infty$. The absolute-value integral is $+\infty$ as well. Correction: to assert equality of signed integrals over every ball, assume local integrability. To define the maximal function using averages of absolute values, local integrability is also a natural standing assumption; for this example the absolute-value averages are infinite for balls containing $0$, so it is not in the stated $L^1_{\mathrm{loc}}$ domain.

Use the Hypotheses, Not a Stronger Conclusion

Covering and differentiation arguments are reliable when their conclusions are stated at the strength actually proved. A covering lemma gives a specific relation between a family, a selected subfamily, and any dilations; its measure estimate also uses countable subadditivity, scaling, and any required containment in a finite-measure set. The differentiation theorem gives an almost-everywhere conclusion for a locally integrable function, not a pointwise conclusion for every representative.

Before using disjointness, identify what the selected balls cover. If coverage is only after dilation, retain the dilation.
Before estimating measure, compute how the dilation changes each ball's measure.
Before comparing signed integrals, check that they exist; local integrability ensures finite integrals on balls.
When changing a function on a null set, distinguish unchanged ball averages from potentially changed point values.
State exceptional-set conclusions as almost everywhere, and do not silently replace them by everywhere.

Check Your Understanding

For each question, identify both the claim justified by the hypotheses and any stronger claim that is not justified.

Give two overlapping intervals for which a pairwise disjoint subcollection fails to cover their union. What role does dilation play in the finite selection argument?
In $\mathbb{R}^n$, by what factor does Lebesgue measure change when a ball is dilated by $a>0$? Compute the factor when $a=3$ and $n=2$.
Why does $f=g$ almost everywhere imply equality of signed integrals over balls when $f,g\in L^1_{\mathrm{loc}}(m)$, and what can go wrong without that assumption?
For $f=\chi_{\mathbb{Q}}$ on $\mathbb{R}$, compute the differentiation average at rational and irrational points and identify the exceptional set.
Explain why equality of centered maximal functions everywhere does not imply that the differentiation expression agrees at every point for two representatives equal almost everywhere.