Solution (source code)

= Solution

For a positive <Borel measure> $\nu$, the centered <Hardy-Littlewood maximal function> and the <uncentered maximal function of a finite measure> are
$$
m\nu(x)=\sup_{r>0}\frac{\nu(B(x,r))}{|B(x,r)|},\qquad
m_u\nu(x)=\sup_{x\in B(c,r)}\frac{\nu(B(c,r))}{|B(c,r)|}.
$$
We use open balls and allow infinite values. For signed or complex measures the definition uses their <total variation measure>. Since a centered ball is also an admissible uncentered ball, and every ball $B(c,r)$ containing $x$ lies in $B(x,2r)$,
$$
\boxed{m\nu(x)\leq m_u\nu(x)\leq2^d m\nu(x).}
$$
For functions, $m(f)$ means the centered <Hardy-Littlewood maximal function> of $|f|\,dx$.

The set $E_\alpha$ is the union of the open balls $B$ with $\nu(B)>\alpha|B|$. In particular, each witnessing ball lies entirely inside $E_\alpha$. If $K\subset E_\alpha$ is compact, take a finite witnessing cover of $K$. Choose its largest-radius ball, discard all balls meeting it, and repeat. The selected balls $B_j$ are disjoint. Every discarded ball has no larger radius than the selected ball that discarded it and is contained in its threefold dilation. Hence
$$
|K|\leq3^d\sum_j|B_j|\leq\frac{3^d}{\alpha}\sum_j\nu(B_j)
\leq\frac{3^d}{\alpha}\nu(E_\alpha).
$$
Taking the supremum over compact $K$ by <inner regularity of Lebesgue measure> proves the stronger, localized estimate
$$
\boxed{|E_\alpha|\leq\frac{3^d}{\alpha}\nu(E_\alpha).}
$$
If the right side is infinite there is nothing to prove. Positivity of the measure is important: a signed version uses $|\nu|(E_\alpha)$, not $\nu(E_\alpha)$.

For the radial-kernel inequality, integrability implies $\phi(r)\to0$. For $0<s<\phi(0)$, let $r(s)$ be the unique radius with $\phi(r(s))=s$. The <layer cake representation> and <Tonelli theorem> give
$$
\begin{aligned}
t^{-d}\int\phi(|x-y|/t)|f(y)|\,dy
&=\int_0^{\phi(0)}t^{-d}\int_{B(x,tr(s))}|f(y)|\,dy\,ds\\
&\leq m(f)(x)\int_0^{\phi(0)}v_d r(s)^d\,ds
=m(f)(x).
\end{aligned}
$$
The last integral equals the total mass of $\phi(|\cdot|)$, namely one. Taking the absolute value of the original <convolution> therefore proves the claim with constant exactly one. This is <radial decreasing kernel domination by the maximal function>.

An application is boundary convergence of the <Poisson integral>. Its normalized radial profile is $c_d(1+r^2)^{-(d+1)/2}$, so $\sup_{t>0}|P_t*f|\leq m(f)$. For a continuous compactly supported $g$, the <approximate identity> property gives $P_t*g\to g$ pointwise. For arbitrary $f\in L^1$, put $h=f-g$; then
$$
\limsup_{t\downarrow0}|P_t*f(x)-f(x)|\leq m(h)(x)+|h(x)|.
$$
The preceding <weak type (1,1)> estimate and the elementary integral bound for $|h|$ show that the set where this exceeds $2a$ has measure at most $(3^d+1)\|h\|_1/a$. Continuous compactly supported functions are dense in $L^1$, so this measure is zero. Thus \b[the Poisson integrals of an integrable function converge to it almost everywhere]; the same <approximate identity> also gives convergence in $L^1$.