At each fixed smooth point, Fourier sums of a piecewise smooth function converge to the function. At a jump, they converge to the midpoint of its two one-sided values. Neither statement describes what happens at points that move toward the jump as the number of frequencies grows. Those moving points reveal a transition layer: its width decreases like the reciprocal of the cutoff, while its overshoot does not disappear.
We will prove that profile, rather than infer it from a plot. Its shape depends on the jump height and not on the smooth background. We will also see why averaging changes the shape. Ordinary sums use a sign-changing kernel; Fejér means use a nonnegative kernel. That difference has an exact consequence for the range of the approximations.
Hypotheses and the unit jump
All functions are real and -periodic. Fourier coefficients and ordinary sums use
Here is a nonnegative integer; scaled statements use . Unless stated otherwise, is a fixed piecewise function: a finite partition makes it on each open piece, with finite one-sided limits of both and at the endpoints. Values assigned at the partition points can be arbitrary. This hypothesis gives bounded derivatives on the pieces and finitely many possible jumps. The word fixed matters: the theorem will not automatically be uniform over functions whose jumps approach each other as increases.
We use the already proved Dirichlet-kernel formula and Riemann–Lebesgue lemma. The latter says that the oscillatory integral of a fixed function tends to zero as its real frequency tends to infinity. We also use the earlier fixed-point convergence theorem for piecewise functions. None of these inputs asserts the moving-point limit that we now seek.
Define the periodic unit-jump function by
Its left and right limits at zero are and , so its jump there is one. Its mean is zero, and integration by parts gives
This function has only one jump on the circle. Using it as a building block avoids creating an unwanted extra jump elsewhere.
Uniform convergence of the continuous remainder
A pointwise theorem alone is insufficient here: our observation point depends on . We first establish the uniform statement needed for the regular part.
Theorem 4.3.1
(A uniform Dini criterion)
Let be a continuous periodic function, and let
where is circular distance. If , then uniformly. In particular this holds for every globally Hölder function with exponent , including every periodic Lipschitz function.
Proof
(Uniform Riemann–Lebesgue over a compact family)
Set . Subtracting the constant contribution from the Dirichlet-kernel formula gives
Since , we have , an integrable bound independent of . The map is continuous into : first make the integral of this bound over small, then use uniform continuity of and the denominator's positive lower bound on . The two regions together prove the claim.
The circle is compact, so the family has a finite net of any prescribed accuracy. Choose finitely many members within of every . Riemann–Lebesgue makes all their oscillatory integrals small simultaneously for large . Replacing a net member by changes its integral by at most , since . Thus the displayed error tends to zero uniformly in . For Hölder functions, the integral condition follows from .
Let the distinct jump locations of be , with jumps . Subtract all of them:
At each , the jump cancels. Assigning the matching continuous value there gives a continuous periodic with bounded derivative on each of finitely many pieces. Integration along the pieces of a shortest circular arc proves a global Lipschitz bound. Therefore uniformly. The decomposition holds almost everywhere, which is exactly what Fourier coefficients require.
We will also need uniform convergence of away from its jump. For circular distance , a finite geometric sum gives
Summation by parts against the decreasing weights yields
The fixed-point theorem identifies the infinite series as away from zero. Hence uniformly on every closed arc separated from the jump. These are the two uniform inputs that allow the other parts of to remain harmless while we zoom in on one jump.
Proof X-ray
Local universality needs a global remainder argument
Removing the chosen jump alone would leave other jumps in the remainder. A global uniform-convergence assertion for that remainder would then be false. We subtract every jump, obtain one genuinely continuous Lipschitz remainder, and treat the other unit jumps on arcs where their own jumps are absent. Each step controls a moving observation point, which a fixed-point convergence statement could not do by itself.
The ordinary-sum transition profile
Define the sine integral by , interpreting the integrand at zero as one. It is an odd continuous function.
Theorem 4.3.2
(The universal jump profile)
Let be a jump of the fixed piecewise function , and put
For every finite ,
The points are interpreted modulo the period, and .
Proof
(Riemann sums for the unit jump, then superposition)
First consider itself. The function
is continuous on . Consequently its Riemann sums converge uniformly in :
For completeness, uniform continuity of bounds the difference between each rectangle and its integral by the same modulus at mesh size ; summing the rectangle lengths preserves that bound. This proves uniformity without exchanging an uncontrolled pointwise limit and a supremum.
Now use the jump decomposition. The continuous remainder contributes in the limit, uniformly for bounded , by its uniform Fourier convergence and its continuity. For every , the point stays in a closed arc away from zero for all sufficiently large . The preceding tail bound and continuity there show that its contribution tends uniformly to . The selected jump contributes . Finally,
because the selected unit jump has midpoint zero and the other terms are continuous at . This proves the profile.
Proof X-ray
Where the scale one over N comes from
The decisive move is to read as a Riemann sum: frequency meets the phase , so the terms sample with mesh on . That is why the layer has width of order and why its shape is an integral, . Uniformity in needs continuity of up to , which is where the value enters. The fixed-function hypothesis is used only in the superposition: every other jump stays a fixed distance from the moving point, so the remainder and the other jumps contribute one constant, turning the unit jump's midpoint into .
At we recover the midpoint theorem. At nonzero fixed , the observation point moves with ; there is no reason its limiting value must equal the one-sided plateau. The next calculation determines the excess precisely.
Why the overshoot survives
Theorem 4.3.3
(The Gibbs excess)
The sine integral satisfies
Set
For an upward jump ,
Thus the excess persists as a fixed fraction of the jump height, although its location approaches the jump. These sample locations need not be exact extrema of the finite sums.
Proof
(The Dirichlet integral and the first lobe)
Here is a proof of the integral limit. For define . Differentiation under the integral is justified on every interval bounded away from , and gives
The last integral follows by integrating the exponential, or by two integrations by parts. Also , so as . Hence . To remove the damping, integration by parts against gives, uniformly for and ,
Indeed the amplitude decreases to zero, and its derivative has total absolute integral equal to its value at . At this argument first proves convergence of the improper tail. On a finite interval, the damped integrals tend to the undamped one; the uniform tail bound then proves that the full integral equals .
To locate the maximum, let
These numbers are positive, strictly decreasing and tend to zero. The derivative alternates sign between consecutive multiples of . Its successive positive local maxima are , , and so on; each is strictly smaller than the preceding one because . Thus the first maximum, at , is the largest. The alternating tail after it is negative, so . The profile theorem now gives the displayed excesses. The decimal is only a convenient description of the exact integral constant, not a new normalization.
Counterexample mode
Fixed points do not control a moving layer
For the unit jump, for every , and at every fixed . Nevertheless . Since , their difference tends to . Thus no uniform convergence holds on a neighborhood reaching the jump from the right, even if the jump point itself is removed. Redefining cannot repair a failure occurring at infinitely many nonzero points.
What Fejér averaging changes
Let , for . Its kernel is nonnegative and has normalized integral one. Convolution uses .
Theorem 4.3.4
(Range preservation and the averaged profile)
If real constants satisfy almost everywhere, then for every and every . For the fixed piecewise function at the jump above,
uniformly on each bounded interval, where
The function is nondecreasing, with limits and at the two ends of the real line. The limiting transition stays between the two one-sided levels.
Proof
(Positivity, then weighted Riemann sums)
The range statement follows immediately by integrating the bounds against the probability kernel . It uses almost-everywhere bounds, so arbitrary assigned values at finitely many jumps do not interfere.
For , the coefficient weights give
Replacing the weight by changes this by a quantity tending uniformly to zero for bounded : the weight error is at most and . Uniform Riemann sums therefore give
At zero the integral is zero. Direct differentiation away from zero gives
The derivative extends at zero with value , and the integral limit for gives the two limiting levels. Finally, the same decomposition as before handles and all other jumps: their ordinary sums converge uniformly on the relevant sets, so their Cesàro averages do too. This proves the whole profile statement.
Global range preservation is stronger than a picture of a single jump, but it is not a finite- local range theorem. If has much larger values elsewhere, those values can contribute through the kernel. What the profile proves locally is the absence of a persistent Gibbs excess in the scaled Fejér limit. For a square wave, its two local levels are also its global bounds, so range preservation excludes excess at every finite .
Common mistake
(No overshoot does not mean uniform convergence)
For the square wave , equal to on and on , the Fejér means stay in and have a monotone limiting profile, yet they do not converge uniformly to on . Each is odd and continuous, so and as ; hence for every . A continuous approximation of a jump must cross the transition layer. Averaging removes the excess, not the layer.
Concept lensAnalytic
A smaller width does not imply a smaller height
The variable keeps the transition visible while its physical width shrinks. Ordinary sums have limiting shape , which passes beyond the one-sided levels. Fejér means have shape , which rises monotonically between them for an upward jump. Both recover the midpoint at the jump and the one-sided levels far out in the scaled variable. The summation method changes the transition, not the jump height of the original data.
Worked examples
Worked example 4.3.5
(A square wave: percentage of which height?)
Let for and for , extended periodically. Its jump at zero has and , and
The profile theorem gives . At , the limiting value is . The excess is about of the upper plateau value one, but about of the jump height two. The latter is the universal fraction. Meanwhile for every , and its scaled profile is . Changing the assigned value to a large number changes none of these statements, because the a.e. bounds are still and .
Worked example 4.3.6
(The exact first peak for the unit jump)
For , differentiating the finite polynomial gives
It is positive for and negative immediately after that point. Hence the first positive local maximum occurs exactly at . Since , uniform convergence of the scaled profile on a compact interval gives
The sample used in the general theorem has the same limiting height, although it is not this exact finite- critical point. This separates an asymptotic scale from an exact location.
Worked example 4.3.7
(A continuous corner has no fixed jump excess)
For on , extended periodically, direct cosine integration gives
The coefficient tail is absolutely bounded by for , so the series converges uniformly; its sum is by the fixed-point theorem. At zero the error is the positive tail , which tends to zero. A derivative jump can produce visible finite-cutoff behavior, but it does not create the nonvanishing height caused here by a jump in the function itself.
Practice
Quick check
(Let , with real constants and . Find its coefficients, its one-sided levels at , and its ordinary and Fejér scaled profiles.)
Solution · 1. Translation, scaling and the jump height
The mean is , and for . The left and right limits are and , so the signed jump is . The profiles are and . If , the Fejér transition is decreasing. Writing and , the ordinary right sample tends to , below the right plateau, while the left sample tends to , above the left plateau. Both excess magnitudes are .
Quick check
(Fix and let . Prove that tends uniformly on bounded intervals to . Why does this not contradict the single-jump profile theorem?)
Solution · 2. Two jumps within one transition scale
Linearity and translation give
The unit-jump profile is uniform on a compact interval containing both and , so both terms pass uniformly to their sine-integral limits. The function being approximated now changes with , and its two jumps have separation . There is no fixed arc around the first jump that stays away from the second one. Thus the uniform-away-from-other-jumps step in the fixed-function theorem is unavailable. The term is nonconstant (its derivative at is ), so the limiting shape is not the isolated unit-jump profile plus a constant. This explicitly exhibits interaction and does not contradict the fixed-function theorem.
Quick check
(The kernel has normalized integral one. Let . Compute and show why unit integral without nonnegativity cannot guarantee range preservation.)
Solution · 3. Mass one is not positivity
Normalized convolution gives
Although , the output lies below zero. The kernel takes negative values, so multiplying an inequality by it cannot preserve the inequality under integration. Its unit integral preserves constants, which is a different property from preserving order and range.
Quick check
(Prove uniform Fourier convergence for a continuous periodic function satisfying for all sufficiently small , where . Explain what the same sufficient test says when the exponent is exactly one.)
Solution · 4. A logarithmic Dini modulus
Near zero the Dini integral is bounded by . The substitution turns this into , which is finite. Away from zero the continuous function is bounded and is integrable on . The uniform Dini theorem applies. At exponent one the upper bound has a divergent integral; that fact does not prove the actual Dini integral diverges, and even failure of the Dini condition would not by itself prove divergence of the Fourier sums. The sufficient estimate is simply inconclusive at that threshold.
Quick check
(At an upward jump , compare and . Find both limits and explain why approaching the jump faster does not improve pointwise accuracy along these moving points.)
Solution · 5. Overshoot and the middle of the layer
At the first sequence of points, the Fourier profile tends to while . Their difference tends to . At the second sequence, the scaled coordinate is . Uniformity of the profile for, say, gives . The actual function values again tend to . Thus the second error tends to . Very close to the jump, the approximation remains in the middle of its transition layer; the fixed-point theorem does not apply to this changing observation point. These errors occur away from itself, so assigning a different value at the jump cannot remove them.
Scope still left open
The result treats a fixed function with finitely many piecewise pieces. It does not prove the same profile for every bounded-variation or merely integrable function, estimates uniform over nearly colliding jumps, convergence rates for the whole scaled remainder, or corresponding profiles for arbitrary filters. Pointwise, uniform and mean-square convergence remain distinct. The proof does establish the universal ordinary and Fejér profiles under the exact stated hypotheses, together with the persistent excess and its normalization by jump height.