A vertical Poisson limit fixes the horizontal coordinate while the height tends to zero. A stronger question allows the point to move horizontally as well, provided its horizontal displacement is at most a fixed multiple of its height. The shape of that approach region matters: an arbitrarily thin excursion can find a nearby spike even when the boundary point itself is a Lebesgue point.
The cone maximal function
For put
For this integral is absolutely defined at every point of the upper half-plane. Let denote the centered Hardy–Littlewood maximal function.
Definition 2.7.1
(Cone maximal function)
For , the cone (nontangential) maximal function of is
It is the supremum of over the cone of aperture with vertex .
The chapter on maximal averages proves the weak estimate with constant 3, the strong estimate for , and domination of centered Poisson averages by ; we use those results here. For , is jointly continuous in the upper half-plane, so , a supremum of continuous functions of , is lower semicontinuous and measurable.
Theorem 2.7.2
(The cone maximal function is controlled by centered averages)
For every and ,
Consequently, for ,
For the same pointwise argument gives for , where is a strong maximal-function constant.
Proof
(Compare a shifted kernel with a centered one)
For and real ,
Thus . Substitute in the convolution and take absolute values:
The last inequality is the layer-cake domination of an even decreasing probability kernel by centered interval averages. Taking the supremum proves (1), and the weak and strong estimates follow from the maximal inequalities. For , , the convolution is well-defined by Hölder because ; the same nonnegative layer-cake argument applies. Continuity of translated Poisson kernels in gives continuity of and hence measurability of in this case as well.
Nontangential limits at Lebesgue points
Theorem 2.7.3
(Every Lebesgue point has a nontangential limit)
Let and let be a Lebesgue point with value . For every fixed ,
Consequently, at almost every this limit holds simultaneously for every finite cone aperture: the same set of Lebesgue points works for all . The function is harmonic on the upper half-plane.
Proof
(Separate the local oscillation from the distant tail)
Write . Given , the Lebesgue-point condition supplies such that
Set . For its centered average at is at most . For , its total integral is at most , so that average is still at most . Thus , and (1) bounds the contribution of inside the cone by .
For the complementary region , take small enough that . There , so the part involving is at most . The constant part is bounded by
Because , these estimates apply to . First let tend to zero uniformly over the cone, then let , proving (3). Lebesgue differentiation supplies almost every boundary point.
For harmonicity, on any neighborhood whose height stays above a positive number, derivatives of the Poisson kernel through order two are bounded by constants depending on that neighborhood. Differentiate under the integral against ; the identity gives .
Proof X-ray
A cone, rather than a single path
The proof splits into a local part and a tail, and the cone enters the two parts differently. The local part is handled at the vertex: (1) bounds its Poisson averages over the whole cone by , and the Lebesgue-point hypothesis is used exactly once, to make . The tail needs the moving point to stay away from the distant mass: requiring keeps within of , so the kernel there is at most . The order of choices is , then , then . A tangential path breaks the local step: the comparison constant grows without bound with , so a small no longer controls the averages of along the path, and the counterexample below shows that this loss is real. All constants depend on the fixed aperture and say nothing uniform as .
Examples
Worked example 2.7.4
(How much the constant in the kernel comparison loses)
The proof of (1) used for . Scale and with ; the case is symmetric. The best constant is then the maximum over of
A value is attained exactly when has a real root, and the maximum is the value at which this quadratic in has a double root: . Its larger solution is
which increases with , so the sharp constant for the cone is . For it is , attained at , whereas . In the edge case no shift occurs and , while . Since , we have , so (1) holds with in place of and the displayed constant loses at most a factor .
Worked example 2.7.5
(A Cauchy profile: explicit limits and a non-integrable maximal function)
Let , so . The Poisson kernels satisfy : in the symmetric convention both sides have Fourier transform . Hence
For , as , as (3) predicts at this point of continuity. For the vertical maximal function put . The function increases for and decreases for , so
The tail is not integrable, so although . The weak-type bound (2) is consistent with this: for the set is the interval of measure , while (2) with allows .
Worked example 2.7.6
(Different slopes at a jump)
Take and approach zero along , for a fixed real . Direct integration gives
The limit changes with . Zero is not a Lebesgue point for any assigned value of there, so this does not contradict (3). Even within one nonzero-aperture cone, different slopes can have different limits at a jump.
Counterexample mode
A tangential path can find vanishingly narrow spikes
For set , , and let be the indicator of the union of the intervals , with . The intervals are disjoint and have summable lengths, so . The origin is a Lebesgue point with value zero: if an interval meets , then and , hence . Summing its lengths and those of all smaller intervals gives for small , so the average of is .
Nevertheless
The points approach outside every fixed cone. Thus arbitrary tangential approach need not recover even a Lebesgue-point value; (3) remains valid for each fixed aperture.
Common mistakes
Common mistake
(Evaluating the maximal function at the moving point)
Domination of centered averages gives . This is true but does not bound : the right side changes with , so the supremum over the cone would involve on every interval , and for these cover the whole line. The kernel comparison moves every estimate back to the vertex, at the price of the factor .
Common mistake
(Expecting a strong bound)
For the estimate (2) is only of weak type. The Cauchy profile above has for , which is not integrable; since , no bound can hold. The strong estimates start at .
Practice
Quick check
(What does (3) say when ?)
Solution · Answer
The only permitted displacement is , so vertically at every Lebesgue point.
Quick check
(Give the weak-type bound for using the displayed constant.)
Solution · Answer
, so . This is a valid bound, not an assertion that 12 is optimal.
Quick check
(For the jump example, compute the limits along and .)
Solution · Answer
The limits are and , respectively.
Quick check
(Why does the spike path eventually leave every fixed cone?)
Solution · Answer
Its displacement-to-height ratio is . For any fixed , this eventually exceeds , so .
Quick check
(Does a nontangential limit at a point follow from assigning a value there, without a Lebesgue-point hypothesis?)
Solution · Answer
No. Changing at a single point changes none of the Poisson convolutions. At the jump above, even two slopes in a cone can have different limits, regardless of the chosen point value.