A Fourier series sees the values of a function on intervals, not the value assigned at one isolated point. At a jump, its natural limit is the midpoint of the two one-sided limits. Piecewise differentiability is enough for this behavior, but bounded variation permits much less regular functions, including continuous functions that fail the Dini condition. We prove the midpoint result directly from the Dirichlet kernel.
Bounded variation and the midpoint theorem
Definition 1.6.1
(Total variation and bounded variation)
For a real function on an interval , its total variation is
The function has bounded variation on if this supremum is finite. A complex function has bounded variation if its real and imaginary parts do. A periodic function has bounded variation on the circle if it has finite variation on a full period, including the jump across the identified endpoint.
Such a function has finite left and right limits at every point. Changing isolated point values may change its variation but does not change its Fourier coefficients.
Theorem 1.6.2
(Dirichlet–Jordan midpoint convergence)
Let be a -periodic integrable function of bounded variation on the circle. If is its symmetric th Fourier partial sum, then at every ,
In particular the limit equals at a continuity point, but an arbitrary assigned value at a jump need not equal the limit.
Proof of the midpoint theorem
The proof uses one cancellation lemma. It is stronger than an absolute-integrability estimate: the expression below may be meaningful even if is not integrable near zero.
Lemma 1.6.3
(A bounded-variation sine-integral lemma)
Suppose has bounded variation on and as . Then
Proof
(Bounded sine integrals and a monotone decomposition)
The integral is ordinary for each fixed , since near zero.
First the partial sine integrals are uniformly bounded in : bound the part below one by and integrate by parts above one: the boundary terms have total magnitude at most and the remaining integral is at most . Thus one may take .
Replace the isolated value by zero; the change alters the variation by at most the finite amount , so bounded variation is preserved. A real bounded-variation function is a difference of two nondecreasing bounded functions. For example, if , take and . Their right limits at zero exist and are equal because . Subtract this common limit from each for and set both values at to zero; the resulting nondecreasing functions are nonnegative and tend to zero at zero. Treat real and imaginary parts separately for complex .
For one such nondecreasing function , the layer-cake identity and Fubini give
For each the inner domain is an interval up to endpoints, since is nondecreasing. Its integral is a difference of two bounded partial sine integrals, so its absolute value is at most . Fubini is legitimate for fixed because the integrand is absolutely integrable on the finite interval. Consequently
which can be made arbitrarily small uniformly in by choosing small. On , , so the Riemann–Lebesgue lemma makes the remaining integral tend to zero. This proves (2).
Proof
(Apply the lemma to the Dirichlet kernel)
Put and, for ,
The periodic bounded-variation hypothesis makes a bounded-variation function on , and its right limit at zero is zero. With , the kernel formula and give
Write , where near zero and is integrable on . The contribution of tends to zero by (2); the contribution of tends to zero by the Riemann–Lebesgue lemma, since is bounded. This proves (1).
Proof X-ray
Two different uses of cancellation
Bounded variation enters once, to write as a difference of monotone functions. Monotonicity is what makes every superlevel set of inside an interval, so the layer-cake integral meets only partial sine integrals over intervals, each bounded by whatever the interval and whatever . This is the key move: near zero the oscillation of is never estimated with absolute values, which is why the lemma survives when is not absolutely integrable.
The hypothesis enters only at the end of that estimate, by making small. Away from zero, ordinary oscillatory cancellation, the Riemann–Lebesgue lemma, suffices. In the application, subtracting is exactly what produces , and replacing by costs only a bounded error kernel, handled again by Riemann–Lebesgue. Neither step assigns a Fourier series a value by substituting into a divergent integral.
The two proofs together establish (1) at every point of the circle, with no exceptional set.
A jump and an assigned endpoint value
Worked example 1.6.4
(The periodic sawtooth)
Let for , extend it periodically, and assign the common endpoint value . Its Fourier series is
For , (1) gives convergence to . At , the one-sided limits are and , so the midpoint is zero. In fact every finite sine sum already equals zero there. The assigned value does not alter any coefficient and is not the Fourier limit.
Worked example 1.6.5
(The square wave and the series for )
Let , with value zero at multiples of . On a period it moves from to at and back at , so its variation on the circle is . Its sine coefficients are
which equals for odd and for even . At the jump every partial sum vanishes, in agreement with the midpoint . The point is a continuity point, and , so (1) gives
The theorem identifies the value of this convergent series; it gives no rate, and the partial sums of this alternating series indeed approach only at rate of order .
Worked example 1.6.6
(A removable discontinuity)
Let except at multiples of , where . On the function rises from to the value , and on it falls back to , so its variation on the circle is , finite. The coefficients are those of , because one point has measure zero. Hence for every , and .
The theorem predicts the same value: , so the midpoint is , not the assigned value . In the proof, for with ; the definition of at zero already discards the assigned value. This edge case is neither a continuity point of the chosen representative nor a jump.
Counterexample mode
Dini's absolute condition is sufficient, not necessary
Define a periodic even function by
It is continuous, nondecreasing on and nonincreasing on , so it has bounded variation on the circle. At zero,
The usual absolute Dini condition fails there, yet (1) gives . Its failure does not prove divergence; bounded variation supplies a different sufficient hypothesis.
The theorem concerns pointwise limits. It does not assert a universal convergence rate, and it does not say that a continuous function of arbitrary variation has pointwise-convergent Fourier partial sums. The earlier localization and Lebesgue-constants chapter constructs a continuous function whose partial sums diverge at a point.
Common mistakes
Common mistake
(Applying the lemma before subtracting the midpoint)
The lemma needs , and the hypothesis cannot be dropped. For a constant ,
which is not zero when . Applying the lemma directly to at a jump is therefore wrong. Subtracting removes exactly this constant, whose contribution is computed separately from .
Common mistake
(Assuming continuity gives bounded variation)
Let for and . It is continuous, but at it takes the values , so consecutive differences are and
Its variation is infinite, and the Dirichlet–Jordan theorem says nothing about such a function. Bounded variation is a genuine restriction, not a consequence of continuity.
Practice
Quick check
(What limit does the midpoint formula give at a continuity point ?)
Solution · Answer
Both one-sided limits equal , so the limit is .
Quick check
(If and , what is the limit of ?)
Solution · Answer
The limit is , regardless of the value assigned to at .
Quick check
(What is the Fourier limit of the sawtooth at its identified endpoint ?)
Solution · Answer
It is zero, the midpoint of and .
Quick check
(Why can changing the value of at a single point not change any Fourier coefficient?)
Solution · Answer
A single point has Lebesgue measure zero, so every coefficient integral is unchanged.
Quick check
(Does failure of the absolute Dini condition at a point imply that the Fourier series diverges there?)
Solution · Answer
No. The continuous bounded-variation example above fails that absolute condition at zero but its Fourier sums converge there by the Dirichlet–Jordan theorem.