Averaging Fourier partial sums gives Fejér's theorem for continuous periodic functions, but ordinary partial sums use a sign-changing kernel. To prove convergence at a particular point we need control of how the function behaves near that point, together with cancellation of oscillations away from it. The Dini criterion makes those two roles precise. It is sufficient, not necessary: failure of its integral test is not by itself a proof of divergence.
Definition 4.2.1
(Fourier coefficients and partial sums)
Let be a real or complex -periodic function, with coefficients
The integrals are Lebesgue integrals. Changing on a null set does not change its coefficients; a claimed limit at a point must therefore be interpreted with care if the representative is changed there.
Oscillation and the Riemann–Lebesgue lemma
Lemma 4.2.2
(Riemann–Lebesgue lemma)
If , then
The frequency is real and need not be an integer.
Proof
(Approximate by step functions)
For an interval indicator , direct integration bounds the absolute value of its oscillatory integral by . The same limit holds for finite linear combinations of such indicators. These step functions are dense in : approximate by a simple function and approximate its finitely many measurable level sets in measure by finite unions of intervals, using Lebesgue regularity and truncating the countable disjoint interval representation of an open set. Given , choose a step function with . Then
The second term tends to zero and the first can be arbitrarily small. For complex , apply the result at both frequencies and and use and ; this gives the sine and cosine conclusions without assuming is real.
Common mistake
(Absolute values destroy the cancellation)
The lemma controls integrals of a fixed function against increasingly rapid oscillations. It does not say that tends to zero: since , that integral equals for every . The limit comes from cancellation, which the absolute value removes.
The Dirichlet kernel and symmetric increments
Finite summation gives
At multiples of , the quotient is interpreted by its continuous limit. The kernel is even and has normalized integral one, since its nonconstant exponential terms integrate to zero. For a chosen real or complex number , put
Pairing positive and negative and subtracting the integral of the constant gives
Changing on a null set changes the translated functions only on null sets, so none of these integrals or the Dini condition changes. The particular value does not enter the formula: is specified independently. The symmetric average is the local quantity detected by this even kernel; one-sided behavior is combined before the limit is taken.
Dini's pointwise criterion
Theorem 4.2.3
(Dini's pointwise criterion)
Suppose that for a point and a number ,
Then as .
Proof
(Make the integrand a fixed function)
On , . Hence belongs to under the displayed assumption. The identity above expresses as times the integral of . The Riemann–Lebesgue lemma applies to the real frequencies , giving the limit zero.
Proof X-ray
Two distinct parts of the proof
The Dini integral controls the apparent singularity of near zero; it converts the symmetric increment into one fixed function. Only then does the Riemann–Lebesgue lemma use oscillation to make its integral vanish. Positivity of the kernel is not used, and ordinary partial sums are not Fejér means.
A useful sufficient hypothesis is local Hölder control: if for sufficiently small , with , then near zero, which is integrable. Away from zero the denominator is bounded below and the periodic assumption handles the rest. For continuous with , this includes local Lipschitz continuity, but permits weaker powers .
Values at jumps and examples
Suppose has finite one-sided limits and is one-sided Hölder at : for small , both and are at most . Taking
gives the Dini condition, so ordinary symmetric partial sums tend to the midpoint. The assigned value can be different without changing the coefficients or the conclusion.
Worked example 4.2.4
(Square wave)
Define a -periodic square wave by for and for , with arbitrary values at the jump points. Direct integration gives only odd sine coefficients, and
The wave is locally constant away from its jumps, so the criterion gives there. At a jump its two one-sided limits are and , so the limit is zero. These continuous partial sums cannot converge uniformly to the resulting discontinuous limit on the circle: a uniform limit of continuous functions is continuous.
Common mistake
(At a jump the limit is the midpoint, not the assigned value)
For the square wave every term of contains , so for every , whatever value has been given; suppose it is . Using in the Dini integral fails, because then for and is not integrable near zero. The number must be chosen from the one-sided limits.
Worked example 4.2.5
(A continuous, non-Lipschitz example)
For , take . At , for small positive , so . The function is not Lipschitz at zero because . The criterion therefore applies beyond the local Lipschitz case.
Worked example 4.2.6
(A sawtooth at an interior point and at a jump)
Let be -periodic with for . Integration by parts gives and for . Pairing with ,
At the function is linear nearby, hence Lipschitz, so the criterion with gives . Only odd contribute, with and . Hence , which is the series .
At the jump the one-sided limits are from the left and from the right, so the criterion predicts the midpoint . This edge case is exact: every term contains , so for all .
The local hypothesis cannot be weakened to continuity alone.
Counterexample mode
Continuity does not give the Dini condition
Claim: if is continuous at , the Dini condition holds there with . Near zero set for and , and extend it evenly and continuously to a periodic function. It is continuous at zero, and with we have for small . Substituting ,
Any other is worse, since then and is not integrable near zero. So the test does not apply at zero. The repair is a quantitative modulus, such as the Hölder bound above. The calculation says nothing about whether converges: the criterion is only sufficient.
Practice with solutions
Quick check
(Use the oscillatory integral of to prove the Riemann–Lebesgue lemma for a finite step function.)
Solution · Answer
For , has absolute value at most . A finite linear combination therefore has oscillatory integral at most by the triangle inequality, which tends to zero.
Quick check
(Suppose and for , with . Prove the Dini condition with .)
Solution · Answer
The symmetric increment satisfies for , giving the integrable bound there. On , and is integrable because is periodic, so the whole Dini integral is finite.
Quick check
(For the square wave above, find the limit of and explain why changing the assigned value at cannot change it.)
Solution · Answer
The one-sided limits are and , so the limit is the midpoint . A single point is a null set, so changing the assigned value there changes no Fourier coefficient and hence no partial sum.
Quick check
(If a Dini integral diverges at a point, does that prove the Fourier sums diverge there?)
Solution · Answer
No. The condition is sufficient for convergence; when it fails, the theorem gives no conclusion either way.