Motivation
A periodic function has discrete frequencies indexed by integers. A function on the whole line has no distinguished period, and its frequency variable is continuous. The Fourier transform replaces the sequence of Fourier coefficients with a function of frequency. It turns translations into phase factors, convolution into multiplication, and differentiation into multiplication by frequency. These rules explain why the transform is useful, but each rule must be tied to a class of functions on which its integrals and limits have meaning.
There are several common normalizations. We use the symmetric convention throughout, putting the factor in both the transform and its inverse. The frequency variable is angular frequency. Changing to cycles per unit length changes both the exponential and the constants.
Definitions
Definition 2.1.1
(Transform and convolution)
For , its Fourier transform is the continuous bounded function
For , define their convolution almost everywhere by
Fubini's theorem gives and . Absolute integrability permits the integral exchanges in the convolution proof below.
The bound follows immediately. Continuity follows by dominated convergence when varies, since the integrand is dominated by . These facts alone do not imply that is integrable, or that its inverse integral converges absolutely.
Definition 2.1.2
(A class on which the formulas are classical)
The Schwartz space consists of smooth functions for which every is bounded, for all nonnegative integers . Such functions and all their derivatives decay faster than every inverse power; in particular they are integrable and square integrable. The Fourier transform maps into itself. This follows by differentiating under the integral and integrating by parts repeatedly: derivatives in introduce powers of , while powers of can be moved onto derivatives of .
Theorem/Proposition
Theorem 2.1.3
(Convolution, translation and differentiation)
For ,
If , then . If , then
Theorem 2.1.4
(Inversion and Plancherel)
For , both integrals converge absolutely and
The transform extends uniquely by continuity to a surjective linear isometry on . For a general function, the transform is an equivalence class obtained as a limit; the defining integral need not exist pointwise. The inverse is likewise an limit. On , the extension agrees almost everywhere with the integral-defined transform.
Theorem 2.1.5
(Inversion for integrable transforms)
If and , the inverse integral is continuous and equals almost everywhere. It therefore equals at every point where the chosen representative is continuous.
The proof below uses the same Gaussian approximation. This statement is not available merely from .
Proof sketch or proof idea
Proof
(Why the rules and constants work)
For convolution, Fubini applies because
With , the exponential splits as . Each resulting integral is times its transform, leaving the factor after the single prefactor is applied. Translation follows from the substitution . For differentiation, integration by parts has zero boundary term for Schwartz functions, giving ; differentiating under the integral gives the multiplication-by- rule.
For inversion, let for . The Gaussian identity gives
To justify the identity, differentiate its left side with respect to and integrate by parts in : the result solves . Its value at zero follows from the real Gaussian integral, giving the displayed formula. Substituting the definition of and using Fubini yields
The nonnegative kernels have integral one and their mass outside any fixed neighborhood of zero tends to zero. For continuous bounded , the convolution tends to at each continuity point: split the integral into small and large displacements, using continuity on the first part and the vanishing tail mass on the second. For , , so dominated convergence removes on the left and proves inversion. For general , the Gaussian approximate-identity theorem states that at almost every (indeed at Lebesgue points). At a continuity point this also follows directly: choose so that is small for ; on , the contribution of is bounded by as , while that of is controlled by the tail mass. If , dominated convergence makes the inverse integral a continuous function and proves it equals almost everywhere and at each continuity point.
For , insert the inverse formula for into and exchange the absolutely convergent integrals. The result is the inner-product identity
Set for the norm identity. Schwartz functions are dense in : approximate an function by compactly supported continuous functions, then smooth them by convolution with a smooth approximate identity. The norm identity makes the transforms of any approximating sequence Cauchy, and the limit is independent of the sequence. This defines the unique isometric extension. Inversion shows that the transform maps Schwartz space onto itself: for , the inverse of is the transform of , which is again Schwartz. Thus the extended transform has dense range. Its range is closed because it is an isometry from a complete space, so it is all of . Finally, if , truncation and smoothing give Schwartz functions in both norms. Their transforms converge uniformly to the integral-defined by the bound, and converge in to the extended transform. A subsequence of the latter converges almost everywhere, so both definitions agree almost everywhere.
Proof X-ray
Two devices: damping and extension by density
Two analytic devices carry the whole argument. The first is damping: the factor makes the double integral absolutely convergent, so Fubini can exchange the order and turn the damped inverse integral into . Only afterwards is taken, on the space side by the approximate-identity property and on the frequency side by dominated convergence, which is exactly where is used. Without that hypothesis the frequency side has no undamped limit to take, while the space side still converges; this is the source of the uniqueness theorem below. The second device is density plus isometry. The inner-product identity is proved only on , where every exchange is absolute, and the extension to uses nothing but completeness. Surjectivity needs one further fact, that inversion maps onto : an isometry alone has closed range but could miss part of .
The real Gaussian integral behind the kernel
The inversion proof used a real integral and then a differential equation in the frequency parameter. Here is where the normalization comes from. For let . This integral is finite and positive: outside a bounded interval the Gaussian is bounded by a decaying exponential. Tonelli's theorem and polar coordinates give
The positive square root yields . No square root with an arbitrary sign is involved because the original integrand is positive.
Now define . Differentiation under the integral is justified by the integrable bound . Integrating by parts in , with vanishing Gaussian boundary terms, gives
Solving this ordinary differential equation gives , exactly the identity used above. The parameter restriction makes both the real integral and these boundary arguments valid.
Quick check
(Use the real Gaussian integral to verify .)
Solution · Answer
With , the integral is . This normalization is what lets convolution with average rather than rescale a constant.
Proof
(One approximation, two Fourier transforms)
The overlap is where an integral calculation must be connected to the extension. Write for the extension and for the integral transform. We spell out how to choose the same smooth approximation for both norms.
For , first truncate its domain and magnitude:
Each is bounded and compactly supported. For , dominated convergence applied to gives . This step has not made smooth.
Choose a nonnegative smooth function supported in with integral one, and let . For a fixed , Minkowski's integral inequality gives
Here we use the standard continuity of translations in and dominated convergence: the integrand is bounded by . For each , choose a single small enough that both the and the smoothing errors for are below . Then is smooth and compactly supported, hence Schwartz, and in both norms.
The two convergences play different roles. For every , the estimate gives
The first convergence is uniform; the second is in . To connect their limits, choose a subsequence whose squared errors sum to a finite number. Tonelli's theorem shows that the sum of its squared pointwise errors is finite almost everywhere, so the subsequence converges there to . Its uniform limit is , hence almost everywhere. In particular, the continuous integral transform represents the extended class and satisfies Plancherel. This does not imply or absolute convergence of its inverse integral.
Concept lensStructural
One function, two notions of transform
For , the integral gives a continuous representative, while Plancherel constructs an equivalence class. Simultaneous approximation in both norms proves they agree almost everywhere. This agreement permits an explicit integral calculation to be used in an energy identity.
Uniqueness without an integrable transform
The inversion theorem above requires both and to be integrable. Is that extra hypothesis needed merely to know that the transform determines ? No. Gaussian smoothing gives uniqueness for every function, even when the ordinary inverse integral is unavailable.
Theorem 2.1.6
(Uniqueness for integrable functions)
If and almost everywhere, then almost everywhere.
Proof
(Damp, invert, and let the damping vanish)
Suppose and almost everywhere in frequency. Both transforms are continuous, so their difference, being continuous and zero almost everywhere, is zero everywhere: a nonzero value would persist on an interval of positive length. Put . The Gaussian identity above then gives, for every and ,
Here the damped integral is absolutely convergent because is bounded. At every Lebesgue point of , the Gaussian approximate-identity theorem already used above gives as . Almost every point is such a point; hence almost everywhere. We have not assumed or used an undamped inverse integral. Equality almost everywhere is the strongest conclusion for arbitrary representatives: changing a function at one point changes neither its integral transform nor its class.
Worked example 2.1.7
(A transform outside the absolute inversion theorem)
The rectangle example below computes the transform of as , with its continuous value at zero, and explains why this transform is not in . If has the same transform almost everywhere, the uniqueness proof nevertheless gives almost everywhere. Thus uniqueness applies although the absolute inversion theorem's second integrability hypothesis fails. The endpoint values of a representative cannot be recovered from the transform.
Quick check
(If an function has transform zero for all , must it vanish almost everywhere?)
Solution · Answer
Yes. Continuity forces its transform also to vanish at zero. Apply the uniqueness argument with . This does not force a chosen representative to be zero at every individual point.
Worked examples
Worked example 2.1.8
(Exponential decay and a rational transform)
For , put . It belongs to , and splitting the integral at zero gives
The transform is integrable, so inversion recovers everywhere because is continuous. Plancherel also gives
Worked example 2.1.9
(A rectangle and a sinc profile)
For , let . Direct integration gives
The value at zero is the continuous limit. Plancherel, applied to , gives
Although this transform is square integrable, its absolute value is not integrable: on a fixed proportion of each successive half-period, is bounded below by a positive constant, and the resulting harmonic series diverges. Thus the hypothesis for absolute inversion does not apply. Inversion still holds in : the inverse integrals over converge in norm as . Changing at either endpoint also leaves its transform unchanged.
A cross-inner-product calculation: two shifted rectangles
The energy identity compares a function with itself; the full Plancherel identity also compares different functions. It preserves their inner product, including relative phase. The following example turns a frequency integral into a length, and shows why the phase cannot be discarded.
Worked example 2.1.10
(Overlap of two intervals and a sinc integral)
Let , , and . Both belong to , and the integral-defined transforms represent their transforms. With the convention above, translation gives
with each quotient continuously extended at zero. Our inner product is linear in its first argument: . In the original variable the product is the indicator of the intersection, so the inner product is its length. Put and . This formula says that the overlap is the shorter interval's length when one interval is contained in the other, decreases linearly during partial overlap, and is zero once they are disjoint.
Plancherel gives
This is an ordinary absolutely convergent integral: the product of the two sinc factors is bounded near zero and is at infinity. Its imaginary part is an odd integrable function, so it integrates to zero. Consequently
For instance, , , gives overlap and integral . At the overlap is and the integral is . Setting and recovers the sinc-square energy calculation above. The complex exponential arose from conjugating the transform of the shifted function; omitting it would replace by and lose the information about relative position.
Quick check
(For and , evaluate the boxed integral when and when .)
Solution · Answer
At , the smaller interval is contained in , so and the integral is . At , the intervals are disjoint, so and the integral is . The latter cancellation does not mean that the integrand vanishes pointwise.
Worked example 2.1.11
(Convolution of two boxes)
Let . The integral is the length of the overlap of and . Hence
The convolution theorem gives , with the continuous value at zero. The product is integrable even though the transform of one box is not absolutely integrable; convolution has improved the frequency decay in this example.
Worked example 2.1.12
(Unequal boxes: overlap, phase and energy)
Let , , and . Put . Writing , the convolution is the length of the intersection of and . When the shorter interval lies inside the longer one; as it moves out, the overlap decreases linearly. Thus
The plateau has width and each sloping shoulder has width . When the plateau contracts to a point and the triangle above returns. The shift changes the location but not the shape or total area, which is .
Direct integration of the boxes, followed by the convolution theorem, gives
The value at zero is the continuous limit and also equals . The factor records the shift: dropping it gives the transform of the centered trapezoid, not generally that of . Its modulus is one, so a shift does not change the frequency energy. Indeed, integrating the plateau and two shoulders gives
Plancherel gives the same value for . It applies because this bounded, compactly supported lies in ; it does not require us to integrate its oscillatory transform by hand.
Quick check
(With , , , find and . How do these values change if the phase factor is omitted?)
Solution · Answer
The plateau gives , and gives . Omitting the phase centers the trapezoid at zero; its values at and are both zero. The phase therefore carries information that the energy alone loses.
Why the convolution is defined almost everywhere
The box integrals are finite at every , but this is not automatic for arbitrary functions.
Counterexample mode
An integrable convolution can diverge at a point
A tempting claim is that for the integral defining converges at every . For a counterexample, let for and elsewhere, including at zero. Then , yet
This is the integral defining . It does not contradict the convolution theorem. For arbitrary measurable , Tonelli gives
Consequently the inner integral is finite for almost every ; assigning any finite value on the exceptional null set produces the same class. Such a change cannot alter the Fourier transform. One must not infer finiteness at every point from the hypotheses.
Common mistakes
Common mistake
(Treating all inversions as ordinary integrals)
An transform is generally defined as a limit in norm, not by substituting every into an absolutely convergent integral. Even when , its transform need not lie in . Check which inversion theorem applies and whether the conclusion is pointwise or almost everywhere. A value assigned at a single point cannot be recovered from an integral transform.
Common mistake
(Dropping the convolution constant)
Under the symmetric convention there is a factor in the convolution theorem. Writing a formula from another convention without converting its definition changes the answer. Check the formula at against .
Summary
The transform gives a bounded continuous frequency function, and convolution becomes a scaled product. Gaussian approximation proves inversion on Schwartz functions; the inner-product identity then extends the transform to a unitary map on . Inversion and energy identities must be interpreted using the hypotheses and mode of convergence stated.
Exercises
Quick check
(Let and . Find and .)
Solution · Answer
Translation gives . The substitution gives .
Quick check
(For , evaluate in two ways: from the convolution theorem and by integrating the triangle.)
Solution · Answer
Since , the convolution theorem gives . The triangle has area four, giving the same answer after multiplying by .
Quick check
(If and , find and verify that its norm agrees with the spatial norm.)
Solution · Answer
The substitution gives . Therefore .