Fejér means converge uniformly for continuous periodic data and in for integrable data. A stronger local question remains: do they recover the value at each Lebesgue point? The Fejér kernel is nonnegative but oscillates, so a proof that assumes the kernel itself decreases away from zero would miss a real issue. A simple decreasing majorant supplies the needed control.
Definition 1.5.1
(Fejér kernel and Fejér mean)
On use , , and normalized circular convolution. For set and
Finite expansion shows , , and
Each is a trigonometric polynomial, even when has jumps.
Theorem 1.5.2
(A decreasing envelope for the Fejér kernel)
For ,
The envelope is even and nonincreasing in . Its constant is convenient, not claimed optimal.
Proof
(Combine a peak bound with a tail bound)
The finite-sum expression gives . For , , so (1) gives . Put . If , then . If , then . This proves the majorization. Finally,
Worked example 1.5.3
(The Fejér kernel is not monotone)
Take and compare two points of . At , formula (1) gives , since . At , , and gives
So increases from to : it is not nonincreasing in , and its superlevel sets are not centered intervals. The envelope is consistent with both values. The tail bound gives , while
because . The envelope stays above the oscillation while decreasing in .
Recovery at every Lebesgue point
Definition 1.5.4
(Lebesgue point)
A point is a Lebesgue point of with selected value if
By the Lebesgue differentiation theorem this holds almost everywhere, for any chosen representative.
Common mistake
(Testing the Lebesgue condition with the wrong value)
Condition (3) is about a selected value , and changing a representative on a null set changes the meaning of “the value at ”. For redefined so that , the point is not a Lebesgue point with value : the averages in (3) tend to . It is a Lebesgue point with value , and , because the convolutions ignore the assigned value. The theorem recovers the Lebesgue value, not an arbitrary assigned one.
Theorem 1.5.5
(Fejér recovery)
If , then at every Lebesgue point , and hence for almost every . If instead the finite essential one-sided limits and exist—meaning, for example, —then
At a nontrivial jump, (4) is a separate statement: that point is not a Lebesgue point for any assigned value.
Proof
(Use the envelope locally and kernel concentration far away)
Fix a Lebesgue point and put . Given , take so that the average in (3) is at most whenever . The function is even and decreasing in . Its superlevel sets are centered intervals of radius at most , apart from endpoints. The layer-cake formula gives
For the complement, and . Thus its contribution is at most
Since and , (5)–(6) control . First send , then .
For (4), use evenness to pair with on . Their sum approaches in the essential one-sided sense. The local difference is essentially uniformly small; the far contribution vanishes by the same tail bounds. Each full half of the kernel has mass , and the mass outside tends to zero, so each local half has mass tending to . This gives the midpoint. At a genuine jump, the average absolute deviation from any value tends to , so (3) cannot hold.
Proof X-ray
Where the envelope and the unit mass are each used
The Lebesgue-point hypothesis enters only in (5), and only through averages over centered intervals of radius at most . The key move is to replace , which oscillates and whose superlevel sets are unions of several intervals, by the even decreasing envelope , whose superlevel sets are centered intervals. The layer-cake formula then writes the local integral as a superposition of exactly the averages that (3) controls. The price is the mass in place of one, which is harmless because it only multiplies .
The unit mass of is used at a different place: in writing . That identity fixes the limit as ; the envelope, with the wrong mass, could never do so. The far part uses only and the decay of like off the window. The order of limits closes the argument: is fixed by , then , then . At a jump (3) fails for every , so the midpoint statement is proved by pairing with instead.
Common mistake
(Letting the window shrink with )
The window must be fixed before . With , for instance, the tail bound (6) becomes , which does not tend to zero. The local estimate (5) needs small, the tail estimate needs , and a fixed achieves both.
A maximal inequality for all truncation levels
Definition 1.5.6
(Centered circular maximal function)
Define the centered circular maximal function
The same layer-cake argument used in (5), without truncation, yields
The supremum on the left is measurable because it is a countable supremum of continuous polynomials.
Theorem 1.5.7
(Weak Fejér maximal bound)
For every and ,
For the centered maximal inequality on the circle, ; (7) then gives (8). To justify that inequality, use the usual finite-cover proof on circular arcs: from a finite family of witnessing arcs choose a longest one, discard intersecting arcs and repeat. The selected arcs are disjoint, while their concentric triples cover the discarded arcs. Each triple has measure at most three times that of its selected arc. Sum the witnessing inequalities, then use compact exhaustion of the open superlevel set ; it is open because each fixed-radius average is continuous in its center. This is a weak-type estimate, not an assertion that the maximal function is integrable.
Worked example 1.5.8
(A trigonometric input is averaged by explicit weights)
For , the finite expansion of gives
The weight of frequency is for and zero outside. This finite example shows how the average damps frequencies before their weights tend to one.
Worked example 1.5.9
(The weak bound for a short indicator)
Let with , so . Since , and , every Fejér mean satisfies . For the set in (8) is therefore empty, and the bound has content only for . In that range, (8) gives
From the other side, every is a point of continuity, so , and the set has measure at least . Both bounds are proportional to : the weak estimate captures the correct dependence on the size of , up to the factor . It is informative, that is below the total measure one, only when .
Counterexample mode
A jump prevents $L^\infty$ recovery
Let periodically. Evenness and unit mass give for every . Each is continuous; for every there is a positive-measure interval immediately to the right of zero where . Hence for every . The a.e. and convergence theorems remain true, and the assigned value at zero has no effect on the convolutions.
Practice with solutions
Quick check
(Compute and its mean.)
Solution · Answer
. Its normalized mean is one.
Quick check
(For periodic , find the Fejér limits at , and .)
Solution · Answer
The two jumps have one-sided values zero and one, so their limits are . At , is continuous and the limit is .
Quick check
(Compute for integer .)
Solution · Answer
It is when , and zero when . For each fixed the weight tends to one.
Quick check
(Find and explain why its growth does not contradict .)
Solution · Answer
. The kernels become concentrated in a narrower region; a growing peak can coexist with fixed integral.
Quick check
(If , give an upper bound for the measure on which some Fejér mean exceeds in modulus.)
Solution · Answer
Equation (8) gives at most . This is a bound on a union over all , not a separate bound requiring a union sum.