From Continuous Approximation to Pointwise Oscillation
Ball averages compare a function near a point with its value at that point. For a continuous function, the difference becomes small on sufficiently small balls. For a merely locally integrable function, there may be no pointwise continuity at all. The useful strategy is to compare the function with a continuous one and estimate the remaining error by a maximal function.
Throughout, the measure space is \((\mathbb{R}^n,\mathcal{L},m)\), where \(\mathcal{L}\) is the Lebesgue sigma-algebra and \(m\) is Lebesgue measure. This measure space is complete and sigma-finite, but not finite. Let \(f:\mathbb{R}^n\to\mathbb{R}\) be Lebesgue measurable and locally integrable, meaning \(\int_K |f|\,dm<\infty\) for every compact set \(K\). Let \(g:\mathbb{R}^n\to\mathbb{R}\) be continuous. Values of \(f\) at individual points are part of the chosen representative; they are not determined by its integrals.
The oscillation here is measured relative to the specific value \(f(x)\), not relative to the average of \(f\) on the ball. This distinction matters: changing \(f\) at just one point leaves all the ball integrals unchanged, but can change \(\omega_f(x)\) at that point.
The Oscillation Bound
Proof. Fix \(x\) and \(\varepsilon>0\). Continuity at \(x\) gives \(\rho>0\) such that \(|g(y)-g(x)|<\varepsilon\) whenever \(|y-x|<\rho\). Thus for every \(0<r<\rho\), the integrand is bounded by \(\varepsilon\) throughout \(B(x,r)\). Consequently its average is at most \(\varepsilon\). The averages are nonnegative, and this bound holds for every sufficiently small \(r\); since \(\varepsilon>0\) was arbitrary, their limit is zero. \(\square\)
Proof. Fix such an \(x\). For every \(y\), the triangle inequality gives $$ |f(y)-f(x)|\le |f(y)-g(y)|+|g(y)-g(x)|+|g(x)-f(x)|. $$ Integrate this inequality over \(B(x,r)\) and divide by its positive, finite measure. The constant term averages to \(|g(x)-f(x)|\), so $$ \frac{1}{|B(x,r)|}\int_{B(x,r)}|f(y)-f(x)|\,dm(y) \le \frac{1}{|B(x,r)|}\int_{B(x,r)}|f(y)-g(y)|\,dm(y) +\frac{1}{|B(x,r)|}\int_{B(x,r)}|g(y)-g(x)|\,dm(y) +|g(x)-f(x)|. $$ The first average is at most \(M(f-g)(x)\) for every \(r>0\), by the definition of the centered maximal function. The second average tends to zero by the lemma. Taking the upper limit as \(r\downarrow0\) proves the stated inequality. If \(M(f-g)(x)=+\infty\), the inequality holds in the extended order; otherwise the same estimate and limit give the finite bound. \(\square\)
The two error terms have different roles. The maximal function controls the approximation error in neighborhoods of \(x\), while \(|f(x)-g(x)|\) records the error at the center itself. Continuous approximation controls the oscillation only when both contributions are small.
Proof. If both quantities on the right were at most \(\alpha/2\), the theorem would give \(\omega_f(x)\le\alpha\), contrary to \(\omega_f(x)>\alpha\). Hence at least one is strictly greater than \(\alpha/2\). \(\square\)
Worked Examples
Worked Example: A Continuous Tent at Its Peak
On \(\mathbb{R}\) with Lebesgue measure, let \(f(x)=\max(1-|x|,0)\), take \(g=f\), and evaluate the oscillation and the bound at \(x=0\).
For \(0<r\le1\), \(f(0)=1\) and \(|f(y)-f(0)|=|y|\) throughout \((-r,r)\). Therefore $$ \frac{1}{2r}\int_{-r}^{r}|f(y)-f(0)|\,dy =\frac{1}{2r}\left(2\int_0^r y\,dy\right) =\frac{r^2}{2r} =\frac r2. $$ As \(r\downarrow0\), this tends to \(0\), so \(\omega_f(0)=0\). Since \(f-g=0\) everywhere, \(M(f-g)(0)=0\), and \(|f(0)-g(0)|=0\). The theorem gives \(0\le0\), with equality. Check: the direct average \(r/2\) also tends to zero, as required for a continuous function.
Worked Example: A Jump at the Center
Let \(f=\chi_{(0,1)}\) on \(\mathbb{R}\), so in particular \(f(0)=0\), and choose the continuous approximation \(g=0\). Compute both sides of the oscillation bound at \(x=0\).
For \(0<r\le1\), the set where \(|f(y)-f(0)|=1\) inside \((-r,r)\) is exactly \((0,r)\), whose measure is \(r\). Thus the average is \(r/(2r)=1/2\), and \(\omega_f(0)=1/2\). For the maximal function, if \(0<r\le1\), the average of \(|f|\) on \((-r,r)\) is \(1/2\). If \(r>1\), the intersection \((-r,r)\cap(0,1)\) has measure \(1\), so the average is \(1/(2r)<1/2\). Therefore \(M(f-g)(0)=Mf(0)=1/2\), while \(|f(0)-g(0)|=0\). The bound is exact: \(1/2\le1/2\). Check: the strict inequality for \(r>1\) follows from \(2r>2\), and no ball of radius \(r>1\) raises the supremum.
Worked Example: A Ramp Approximation with Sharp Maximal Error
Keep \(f=\chi_{(0,1)}\), and define the continuous function $$ g(y)= \begin{cases} 0,&y\le0,\\ y,&0<y<1,\\ 2-y,&1\le y<2,\\ 0,&y\ge2. \end{cases} $$ The formulas agree at the joining points, so \(g\) is continuous. At \(x=0\), \(f(0)=g(0)=0\), and the direct calculation in the preceding example gives \(\omega_f(0)=1/2\).
Set \(h=f-g\). Except at endpoints, which have measure zero, \(h(y)=1-y\) on \((0,1)\), \(h(y)=-(2-y)\) on \((1,2)\), and \(h(y)=0\) elsewhere. For \(0<r\le1\), $$ \frac{1}{2r}\int_{-r}^r|h(y)|\,dy =\frac{1}{2r}\int_0^r(1-y)\,dy =\frac{r-r^2/2}{2r} =\frac12-\frac r4. $$ For \(1\le r\le2\), the integral of \(|h|\) over \((-r,r)\) is $$ \int_0^1(1-y)\,dy+\int_1^r(2-y)\,dy =\frac12+2r-\frac{r^2}{2}-\frac32 =2r-\frac{r^2}{2}-1. $$ This is at most \(r\), since \(r-(r^2/2)-1=-((r-1)^2+1)/2<0\). Hence the average is at most \(1/2\). For \(r\ge2\), the full integral of \(|h|\) is \(1/2+1/2=1\), giving average \(1/(2r)\le1/4\). The averages for \(0<r\le1\) have supremum \(1/2\), approached as \(r\downarrow0\). Thus \(M(f-g)(0)=1/2\). The theorem again gives the sharp bound \(1/2\le1/2+0\). Check: the two triangular error pieces have integrals \(1/2\) each, so the total error integral used for \(r\ge2\) is \(1\).
A Representative-Sensitive Pitfall
It is incorrect to omit the pointwise term \(|f(x)-g(x)|\), even if \(f-g=0\) almost everywhere. For a concrete example, let \(f(0)=1\), let \(f(y)=0\) for \(y\ne0\), and let \(g=0\). This \(f\) is measurable and locally integrable, and \(f-g=0\) almost everywhere. Every ball integral of \(|f-g|\) is zero, so \(M(f-g)(0)=0\). But \(f(y)=0\) for almost every \(y\) in every ball centered at zero, while \(f(0)=1\). Consequently $$ \frac{1}{|B(0,r)|}\int_{B(0,r)}|f(y)-f(0)|\,dm(y)=1 $$ for every \(r>0\), and \(\omega_f(0)=1\). The full inequality gives \(1\le0+1\), as it should. The error at the center cannot be inferred from averages, because changing a function on a null set does not change those averages.
The estimate is useful because a continuous approximation has no small-scale oscillation of its own: all remaining oscillation is charged to the approximation error, both near the point and at the point. In applications, maximal-function estimates control where the neighborhood error is large, while approximation in \(L^1\) makes such errors manageable. The inequality by itself does not assert that \(\omega_f(x)=0\) almost everywhere; that conclusion requires the further differentiation argument.
Check Your Understanding
Use the definitions and the pointwise inequality to answer the following.
- For \(f(x)=\max(1-|x|,0)\) on \(\mathbb{R}\), compute the average oscillation at \(x=0\) for \(0<r\le1\).
- For \(f=\chi_{(0,1)}\) and \(x=0\), why is the centered maximal average \(1/2\) when \(0<r\le1\), but strictly smaller when \(r>1\)?
- In the oscillation bound, which term accounts for changing the value of \(f\) at the single point \(x\)?
- For the ramp approximation in the third example, compute \(\int_0^1(1-y)\,dy\) and \(\int_1^2(2-y)\,dy\).
- State the triangle-inequality decomposition that leads from \(f\) and \(g\) to the maximal-function bound.