Writing a series is not the same as trusting it
In the previous guides you built a trigonometric Fourier series: you took a periodic signal, computed its coefficients by the Euler-Fourier integrals, and wrote f(x) = a_0/2 + sum (a_n cos(n x) + b_n sin(n x)). That equals sign deserves a hard stare. Computing coefficients only manufactures a candidate series; whether that infinite sum actually adds back up to the original f — and at which points — is a separate question entirely. This is the same distinction you met in Volume I between an infinite series you can write down and one that genuinely converges.
What makes this urgent is that Fourier's whole selling point is jagged signals — square pulses, sawtooth ramps, the on/off switch of a circuit. Those functions have jumps, and a jump is exactly where you should worry: how can a sum of perfectly smooth sines and cosines, each continuous everywhere, possibly reproduce a cliff edge? It cannot do so pointwise in the naive way, and the honest accounting of what it does instead is the subject of this guide. We need a working test for convergence, a clear verdict at jumps, and an unflinching look at the one place the series misbehaves.
The Dirichlet conditions: a test you can run by eye
The clean, practical answer comes from the Dirichlet conditions. They are a friendly set of hypotheses — checkable at a glance, no heavy machinery — that guarantee the series converges and tell you precisely what value it converges to at every single point. Stated for a periodic f: if, over one period, f is absolutely integrable (the definite integral of |f| is finite), has only finitely many maxima and minima, and only finitely many jump discontinuities each of finite size, then its Fourier series converges everywhere. That is it — three things you confirm by looking at the graph.
Now the verdict, point by point. Where f is continuous, the series converges to f(x) itself — no surprise there. Where f has a jump, the series converges to the midpoint of the jump: the average (f(x+) + f(x-))/2 of the right-hand and left-hand limits. That single clause is the practically important one. The series refuses to take a side; it splits the difference. A square wave that leaps from -1 to +1 at the origin has a Fourier series that, exactly at the origin, converges to 0 — and indeed every sine term sin(n·0) is zero there, so the arithmetic agrees on the nose. This is convergence at a jump, and it is not a defect to be apologized for; it is the precise, correct answer.
How to use the test in practice
The payoff of having a checkable test is that justifying a Fourier computation becomes a quick inspection rather than a research project. You almost never need the heavy L^2 theory; you confirm your signal is piecewise smooth with finite jumps, and you immediately know the series is legitimate and where it lands at each point. Here is the whole drill, run on the sawtooth f(x) = x extended periodically from (-pi, pi).
- Check absolute integrability: integral of |x| over (-pi, pi) is pi^2, finite. Pass.
- Check finitely many extrema: x is monotone increasing on the open interval — zero interior turning points. Pass.
- Check finitely many finite jumps: the periodic extension leaps from +pi down to -pi at x = pi (and its copies). One jump per period, of finite size 2pi. Pass.
- Read off the verdict: convergent everywhere; equal to x at every interior point; and at the jump x = pi the series converges to the midpoint (pi + (-pi))/2 = 0 — NOT to the function value pi.
Two cautions worth internalizing. First, the value the series returns at x = pi (namely 0) simply is not the value pi you might have written for f there — and that is fine, because a Fourier series only ever cares about f up to its values at isolated jump points. If it bothers you, redefine f at the jump to be its own midpoint and the series and function agree exactly. Second, this same midpoint rule is what bites in half-range expansions: taking the odd extension of a function that is nonzero at an endpoint manufactures a brand-new jump there, so its sine series collapses to 0 at the ends even though your original f was perfectly continuous on the interval. The jump was created by the extension, not present in the data — a classic source of confusion when you only stare at the formula and forget the picture.
The Gibbs phenomenon: an overshoot that will not die
So the series converges everywhere — at jumps, to the midpoint. You might now expect that a partial sum with enough terms just looks like a slightly rounded version of f, hugging it more and more tightly as you add terms. Watch what actually happens. Plot the partial sum of the square wave (jumping from -1 to +1) with, say, ten terms: it does follow the flat plateaus nicely, but right beside the jump it shoots up to about 1.179 before turning back — a bump rising roughly 9% of the jump height above the true value of 1. Add ninety more terms. The bump gets narrower and crowds closer to the jump, but its height does not shrink: it still overshoots to about 1.179. Sum a million terms; same overshoot. This stubborn, undying spike is the Gibbs phenomenon.
Sit with how strange this is. We just proved, via the Dirichlet conditions, that the series converges at every point. The Gibbs overshoot does not contradict that for a beautiful and subtle reason: convergence is pointwise, and the overshoot survives because its location keeps moving. Fix any single point x_0 to the right of the jump; as you add terms, the bump eventually slides past x_0 and the partial sums there do settle down to the true value. No fixed point is ever permanently overshot — so the series converges at each one. But there is always SOME point, ever closer to the jump, where the partial sum is still overshooting by that fixed ~9%. The defect never lands on a stationary point long enough for pointwise convergence to erase it.
Where the 9% comes from
The overshoot is not a quirk of the square wave; it is a universal constant of jumps, and it traces back to one specific integral. When you sum the first N terms of a Fourier series, you can rewrite the partial sum as f convolved with the Dirichlet kernel, and right at a jump that convolution reduces to a partial integral of (sin t)/t — the integrand behind the sine integral Si(x) = integral from 0 to x of (sin t)/t dt. The height of the overshoot is governed entirely by how Si behaves before it settles, and Si overshoots its own limit. As x increases to infinity, Si(x) climbs to pi/2 approximately 1.571 — but it does not climb monotonically; its first peak, at x = pi, reaches Si(pi) approximately 1.852, sailing above the final value.
Final limit: Si(infinity) = pi/2 ~ 1.5708 (the Dirichlet integral)
First peak: Si(pi) ~ 1.8519 (where the partial sum tops out)
Fractional overshoot of a unit step:
Si(pi)/(pi/2) - 1 = 1.8519/1.5708 - 1 ~ 0.1789
so the partial sum reaches ~ 1.0895 on a step of height 1
==> excess ~ 0.0895 = about 8.95% of the JUMP height
Square wave -1 -> +1 (jump = 2):
peak ~ 1 + 0.0895 * 2 ~ 1.179 (vs. true value 1)Read off the consequence: the partial sum tops out near 1 + 0.0895 times the jump, so for a square wave of jump 2 it reaches about 1.179, the number we saw. Two clarifications keep this honest. The PERCENTAGE of overshoot is the universal constant — about 8.95% of the jump, the same for any jump in any Dirichlet-class function — but the absolute height of the bump of course scales with how big the jump is. And what DOES improve with more terms is the WIDTH of the ringing: the overshoot squeezes into an ever-thinner sliver hugging the jump, even as its peak height refuses to drop. More terms localize the damage; they never cure it.
Living with it: smoothness, energy, and a fix
Why does the square wave suffer this while a smooth signal sails through? It comes down to how fast the coefficients decay, which is set by the smoothness of f — a theme you can feel echoing the way a Taylor series of a smooth function converges so willingly. A function with a jump has Fourier coefficients dying off only like 1/n, far too slowly for the tail to vanish uniformly, and that sluggish tail is precisely the ringing. Smooth out the corner — make f continuous, or differentiable — and the coefficients fall off like 1/n^2 or faster, the tail becomes negligible, and the convergence turns uniform with no overshoot at all. Gibbs is the signature of a genuine discontinuity, nothing else.
There is a comforting bookkeeping note here. The Gibbs spike is tall but vanishingly thin, so it carries essentially no energy: its contribution to the total integral of f^2 shrinks to nothing as N grows. That is why Parseval's theorem — the energy ledger equating the integral of f^2 to the sum of squared coefficients — stays exactly true despite the overshoot, and why the series still converges in the mean-square sense even where it fails to converge uniformly. The mismatch is real to the eye but invisible to the energy. Two different questions, two different answers; the honesty is in keeping them apart.
And if the ringing actually hurts your application — visible halos beside sharp edges in a reconstructed image, spurious oscillation in a numerical solution — there is a practical remedy. Instead of chopping the series off abruptly at term N (a hard rectangular cut-off, which is what causes the ringing), you taper the high-frequency coefficients down gently with a smooth weighting before truncating. Methods like Fejer's averaging or Lanczos sigma factors do exactly this, trading a little blur at the edge for the elimination of the overshoot. The lesson worth carrying forward: Gibbs is not a bug in Fourier's idea, it is the honest price of representing an instantaneous jump with smooth waves — and you manage it by softening either the function or the way you cut the sum, never by simply adding more terms.