Motivation
Fourier partial sums are continuous trigonometric polynomials, yet for the wave below they converge to the function at every fixed continuity point and to the midpoint of the one-sided limits at a jump. What happens if the observation point itself moves toward the jump as more frequencies are added? This question has a different answer from pointwise convergence. A narrow overshoot can approach the jump while retaining a nonzero height.
We will calculate that behavior for one square wave, from its coefficients through a Riemann-sum limit. The calculation also identifies an actual local maximum of every partial sum in the chosen sequence. It does not prove a universal Gibbs theorem for arbitrary discontinuous functions. Generalizing the same limiting constant to other functions requires additional hypotheses and a separate argument.
Definitions
Definition 1.3.1
(The wave and its Fourier partial sums)
Let be the -periodic function with
The values at the identified endpoints agree, so this defines a periodic representative. At the left and right limits are and ; the jump magnitude there is . There is another, oppositely oriented jump at the identified endpoints . The local analysis below concerns the jump at zero. The symmetric partial sum includes all complex frequencies , equivalently the sine and cosine terms up to frequency .
Oddness gives zero cosine coefficients. For , the sine coefficient is
Only odd frequencies remain. Choose the sequence , , and put . Then
The next even partial sum is the same, because its new coefficient is zero. Thus the selected sequence captures all changes in the ordinary symmetric partial sums. The value at the jump is , the midpoint of the one-sided limits.
Definition 1.3.2
(The sine integral and the moving observation point)
For real , define
The integrand has a removable singularity at zero, so the integral is ordinary and finite for every finite . To examine a shrinking neighborhood of the jump, set and keep fixed while . A fixed then describes a point whose distance from the jump is proportional to .
Two limits near the jump
Theorem 1.3.3
(The local limiting profile)
For every , as ,
Thus the limiting profile is . The convergence is uniform on each bounded interval of the rescaled variable , but not on the whole real line: taking gives while .
Theorem 1.3.4
(The first peak and the Gibbs constant)
For each , the first positive local maximum of occurs at
Its location tends to zero while
The limiting excess above the right-hand value is , or a fraction
of the jump . This persistent overshoot is the Gibbs phenomenon for this wave.
The two statements answer different questions. At the fixed point , every partial sum equals zero. At the moving point , the sum approaches a value greater than one. The points themselves converge to zero; a pointwise theorem does not control a sequence of inputs that changes with .
Concept lensAnalytic
The profile is a band-limited sign function
On the line, let ; its Fourier transform is the indicator of the band . Then, as an improper integral,
So is the sign function with every frequency above removed. On the scale the wave near zero looks like , and the retained frequencies become . The overshoot is the price of a sharp frequency cutoff, not of the particular wave.
Why the first peak survives
Proof
(A midpoint Riemann sum on the jump scale)
Set . These are midpoints of the equal subintervals of . Substituting gives
For fixed , define when , and . It is continuous on , so the displayed sum converges to
The substitution also holds for negative , with the usual orientation of the integral. To see uniformity for , use . Its derivative has magnitude at most . On a mesh interval of length , the integral of distance from its midpoint is . Therefore the error in the midpoint sum, uniformly over , is at most before multiplying by . This proves the stated compact-interval convergence and displays a finite- bound.
Proof
(Locate the actual peak)
Differentiate the finite sum and use the finite cosine identity:
The denominator is positive. The numerator is positive for and negative immediately to the right of , so is the first positive local maximum. In the rescaled variable it is exactly , and the Riemann-sum theorem yields its limiting height. The power series of sine can be integrated term by term on to give . Its term magnitudes decrease, so the alternating-series error is at most the first omitted term. In particular , and further terms give .
The limit profile itself increases on and decreases on , since for . Its later oscillations do not exceed the first peak: the areas decrease strictly as increases, because the numerator repeats after a shift by while the denominator grows. Pairing each later positive lobe with the larger preceding negative lobe shows that every later positive-side local maximum is lower than the one at . This compares extrema of the limiting profile; we have proved a first local maximum for each finite sum, not a global maximum for that finite sum.
Proof X-ray
Exact location, limiting height
The argument has two independent parts, and the key move is to keep them separate. Locating the peak uses only the closed form of the derivative, valid at every finite . Its first positive zero is , so in the rescaled variable the peak sits exactly at for every , not merely in the limit. That exactness is what allows the profile theorem, which is a statement for fixed , to be evaluated at the peak.
The height comes from the Riemann-sum identity. The substitution turns the odd frequencies into the midpoints of a uniform mesh on ; only odd frequencies occur, which is why the points are midpoints. Convergence of the sum needs continuity of on the closed interval, which is where the removable singularity at is used. Uniformity in is not needed for the single value ; it is what makes the profile a statement about shapes, and it fails on the whole line, so fixed- convergence needs its own argument. Finally, the peak exceeds the plateau because : after the lobes of alternate in sign with strictly decreasing areas, so their total contribution is negative.
The first peak is in fact above its own limit at every finite , not merely in the numerical examples. Integrating the derivative from zero to and setting gives
For , and . The integrand is therefore strictly larger than on that interval, proving for every . The finite peak approaches its limit from above, although this comparison alone does not prove that the sequence of peak heights is monotone.
The plateau in the peak comparison is not guessed from the graph. The Dirichlet integral gives , so as . One way to justify the integral is to introduce for . Differentiation gives and as , hence . For , integration by parts bounds the tail beyond by : is nonnegative and decreasing for all . First pass on , then let ; the undamped integral is .
For a fixed , the same finite derivative gives a direct pointwise check:
Write , with . The function is continuous on , so the oscillatory integral of tends to zero. The remaining integral becomes . Thus . Oddness gives the limit for fixed . There is no conflict with the moving peak: its input is not fixed.
Worked examples
Worked example 1.3.5
(A finite partial sum already overshoots)
For , and
The first positive local maximum is at , where
This exceeds the right-hand level one. The limit of the peak heights is approximately , not the finite- value above. A graph at a single truncation can illustrate the effect but cannot by itself establish the limiting height.
Worked example 1.3.6
(A jump of another size)
Let and set . It has left limit , right limit , and jump . Linearity of Fourier coefficients gives
At , its limiting height is , where . For and , the predicted limiting peak is . The overshoot fraction refers to the jump size three, not to the absolute height five.
Worked example 1.3.7
(A bounded averaging kernel behaves differently)
The Fejér mean is the convolution of with a nonnegative kernel for which . Since almost everywhere, the same bounds hold for every : the integral is a weighted average. Hence Fejér means cannot overshoot above one for this wave. Their averaging changes the summation method; it does not change the ordinary partial sums whose peak we calculated.
Counterexample mode
Changing a point is invisible to the coefficients
Change only the value at zero and its periodic copies, assigning for every integer . The integrals defining all Fourier coefficients are unchanged, and every partial sum is unchanged. In particular its value at zero remains zero, not ten. This example shows why a Fourier coefficient calculation cannot recover arbitrary assigned values at individual points. The midpoint conclusion refers to the two one-sided limits, not to a chosen jump value.
Common mistakes
Common mistake
(Using the profile limit at a fixed physical point)
The limit holds for fixed , uniformly only on bounded -intervals. If is fixed while grows, then is unbounded, so this compact-uniform statement alone cannot be applied. The separate oscillatory-integral argument above establishes convergence at fixed .
Common mistake
(Confusing the overshoot height with its fraction of the jump)
Here the peak rises about above the level one, but the jump has magnitude two. Dividing gives . For another jump magnitude, first state the left and right limits and the normalization; then multiply by the jump magnitude in this affine example.
Summary
For the specified square wave, a midpoint Riemann sum gives the profile on the scale . The first positive local maximum occurs at , approaches the jump, and retains an excess equal to approximately of the jump size. Fixed-point convergence, a moving peak, and Fejér averaging are distinct statements.
Exercises
Quick check
(Compute and its first positive local maximum. Compare the peak with the limiting value .)
Solution · Answer
. Its first positive local maximum is at . The limiting peak is about ; they need not agree at finite .
Quick check
(Show directly that the even partial sum equals , and explain why this lets the results cover the full sequence of symmetric partial sums.)
Solution · Answer
The coefficient of is zero, as are all cosine coefficients, so adding the even frequency changes nothing. Every is either odd or immediately follows an odd index, apart from . Therefore the odd-index convergence and peak heights determine the asymptotics of the full sequence, with the same spatial scaling : it equals at odd and at the following even index.
Quick check
(A wave equals on and on , with midpoint values at its jumps. Find the limiting first peak above the right-hand level and its numerical excess.)
Solution · Answer
The wave is , with jump . Its limiting peak is ; the excess over the right-hand level is .
Quick check
(Prove from positivity and normalization of the Fejér kernel that if almost everywhere, then for every .)
Solution · Answer
Write , with and normalized integral one. Multiplying by and integrating gives the two inequalities. A change on a null set does not affect the integral.