Oscillations cause cancellation. But how much cancellation follows from integrability alone? The Riemann–Lebesgue lemma guarantees that an Fourier transform tends to zero at large frequencies. No prescribed rate, however slow, is guaranteed for every function, even if its constant may depend on that function. We will prove both statements and distinguish decay from differentiability of the transform.
Use the unitary angular-frequency convention
All integrals use Lebesgue measure and functions may be complex-valued. Translation preserves the norm. We first justify the small-translation fact that drives the cancellation estimate.
Theorem 2.6.1
(Translation is continuous in )
For every ,
Proof
(Approximate by finite interval steps)
For an interval , the symmetric difference with its translate has length at most , so . The same limit holds for finite linear combinations of bounded-interval indicators.
Such steps are dense in : truncate a function in size and support, approximate its bounded range by finitely many values, and use Lebesgue regularity to approximate each finite-measure level set in measure by a finite union of bounded intervals. Given , choose a step function with . Translation invariance and the triangle inequality give
First let , then . This proves (1). Lebesgue regularity is the measure-theoretic input; pointwise continuity of is not assumed.
Theorem 2.6.2
(Riemann–Lebesgue on the real line)
If , then is bounded and uniformly continuous, and
More precisely, for ,
Proof
(A half-period shift reverses the phase)
The integral gives . For an increment ,
The right side is independent of and tends to zero by dominated convergence, giving uniform continuity.
A change of variables gives . Choose , so the phase is . Taking the transform of and applying the bound gives (3). Since as , (1) proves (2).
Proof X-ray
The cancellation is an $L^1$ translation estimate
Three inputs are used, each at one point. Integrability of alone gives the bound and, through dominated convergence with the frequency-independent majorant , uniform continuity. The key move is the choice : translation by half a period multiplies the transform by , so equals half the transform of , and the crude bound is applied to a difference rather than to itself.
The argument closes because as , which is exactly the regime controlled by translation continuity (1); that continuity is where Lebesgue regularity and the density of step functions enter. Estimate (3) transfers only what (1) knows. Since (1) carries no rate for a general , neither does (2), and the slow-decay theorem below shows that this loss is genuine.
Moments and derivatives answer different questions
Theorem 2.6.3
(Moments give frequency derivatives)
Let be a nonnegative integer. If and , then and for ,
Every derivative in (4) is bounded, uniformly continuous and tends to zero at infinity.
For , , so . The difference quotient of is bounded in absolute value by ; at the step from derivative to (where ), the dominating function is . Dominated convergence applied successively gives (4). The preceding theorem applies to each . A moment condition controls differentiation with respect to frequency. It is different from smoothness of with respect to its original variable.
If instead with , integration by parts times gives
The boundary terms vanish because the derivatives have compact support. Since , the Riemann–Lebesgue lemma yields as . Thus spatial smoothness can provide a quantitative power of decay. These hypotheses are sufficient, not necessary.
Decay from an integrable derivative
Compact support is convenient, but it is not needed for integration by parts. What matters is both the behavior at infinity and the regularity between points. Suppose is absolutely continuous on every bounded interval and , where is its almost-everywhere derivative. Then
To see why the boundary terms vanish, the fundamental theorem for absolutely continuous functions gives . The integrable tails of make approach finite limits at both ends of the line. Each limit must be zero: a nonzero limit would contradict . Integrating by parts on and then letting proves the identity in (5); the Riemann–Lebesgue lemma applied to proves the limit. In particular, for ,
The limit in (5) is stronger than this displayed bound. The same argument can be repeated: if is locally absolutely continuous for and for , where each derivative is understood almost everywhere and each locally absolutely continuous derivative denotes such a representative, then and . The hypotheses are sufficient; the formula should not be used just because an ordinary derivative exists almost everywhere.
Worked example 2.6.4
(A corner is allowed)
Let with . It is absolutely continuous on each bounded interval even though it is not differentiable at zero. Its almost-everywhere derivative is for and for ; both and are integrable. Splitting the original transform integral at zero gives
The formula confirms . It even decays as , but tends to the nonzero constant . There is no contradiction with the repeated-derivative result: has a jump at zero and has no locally absolutely continuous representative agreeing with it almost everywhere. Applying integration by parts twice while ignoring that jump loses a boundary contribution.
Worked example 2.6.5
(A triangle has one integrable derivative but not two)
Let , equal to on and zero elsewhere. It is locally absolutely continuous, with integrable almost-everywhere derivative, so (5) applies. Splitting at zero and integrating gives
The value at zero is the integral of divided by . Along , , so a second little- conclusion is false. Here too the derivative has jumps.
Common mistake
(Integrating by parts with an almost-everywhere derivative)
The box has ordinary derivative zero almost everywhere, but its transform is not zero: its jumps prevent local absolute continuity. An almost-everywhere derivative alone does not recover the changes of the function, and cannot justify the integration-by-parts identity.
Worked example 2.6.6
(A jump leaves a first-order oscillatory tail)
For , ,
with its continuous value at zero. It tends to zero and is , but it is not : along , the modulus of is . Despite the jumps of , all moments are integrable, so has derivatives of every order. Smoothness of the transform does not imply comparable smoothness of the original function.
Common mistake
(Vanishing at infinity is not integrability)
The lemma places among continuous functions vanishing at infinity; it does not place in . For the transform is . On each interval with , the integral of equals and , so that interval contributes at least to . These lower bounds form a divergent harmonic series. An inversion integral for such an therefore needs an interpretation, such as a symmetric limit or an limit, rather than absolute convergence.
Integrability gives no prescribed rate
Theorem 2.6.7
(A single function can decay slower than any proposed rate)
Let satisfy as . There exist a nonnegative and a sequence such that
Thus the lemma cannot be strengthened to for every function.
Proof
(Place small positive Gaussian peaks far out)
Choose increasing with . Put and
The cosine series converges uniformly and has absolute value at most one, so and . The Gaussian transform identity under our convention is . Dominated convergence allows termwise transformation; modulation gives
Every summand is nonnegative. At , the th shifted Gaussian contributes . Therefore
The peaks are chosen far enough out to defeat the proposed rate; no cancellation estimate between them is needed. This does not contradict (2): their heights still tend to zero.
Counterexample mode
A tempting bound cannot be uniform over the $L^1$ unit ball
Fix and let . Then but . Thus no function can satisfy for all and . Here the input changes with the frequency. The preceding theorem is stronger in a different direction: it constructs one fixed input that defeats a given rate along a sequence.
Practice with solutions
Quick check
(Compute the transform of , , and verify the lemma.)
Solution · Answer
For , ; at zero it is . The numerator is bounded, so the transform tends to zero.
Quick check
(If as well as , compute .)
Solution · Answer
Equation (4) gives . The moment hypothesis justifies differentiating under the integral.
Quick check
(If , what can be concluded about ?)
Solution · Answer
Integration by parts gives . Since , the right side tends to zero. Thus , including the sign change from .
Quick check
(If almost everywhere and , can a different assigned value of change the transform or its decay?)
Solution · Answer
No. The integrals defining the two transforms agree for every frequency. Point values are invisible to these integrals.
Quick check
(Does (2) imply that for every there is a constant with ?)
Solution · Answer
No. Apply the slow-decay theorem with . The resulting single has an unbounded ratio , so no finite works. The lemma only asserts convergence to zero.