The question we quietly skipped
In the last guide you used orthogonality to pluck out each Fourier coefficient b_n, one clean integral at a time. That gave you a recipe: take any reasonable starting profile f(x) and write down the numbers b_n that ought to rebuild it as a Fourier series, f(x) = sum of b_n sin(n pi x / L). But notice the sleight of hand. Computing the coefficients of an infinite series is not the same as proving the series adds back up to f. We assumed the sum exists, assumed it equals f, and only then went hunting for the weights.
That gap matters, because nature hands us profiles that are not smooth at all. A plucked string starts as a sharp triangular kink. A bar pressed against ice on one side has a temperature that jumps. The honest question for this rung is: when we feed such a ragged f(x) into our separation of variables machine and superpose the modes, does the result really converge to the data we asked for — and converge in what sense? It turns out there are three different honest answers, and they genuinely disagree at the worst points.
Three honest senses of convergence
The gentlest and most natural sense is mean-square convergence: the area of the squared error between S_N and f shrinks to zero, that is the integral of (f - S_N)^2 over the interval goes to 0 as N grows. This is the sense the orthogonality machinery was built for, and it almost never fails. As long as f has finite energy — finite integral of f^2 — the Fourier series converges to it in mean square, full stop. No smoothness required, no continuity required. This is also exactly the statement that the sine modes form a complete set: nothing with finite energy escapes them.
The second sense is pointwise convergence: fix a single location x and ask whether the number S_N(x) settles down to f(x). This is fussier. The classic guarantee is that if f is piecewise smooth, then at every point where f is continuous the series converges to f(x) exactly — and at a jump it converges to the midpoint of the jump, the average of the left and right limits. That midpoint rule is wonderfully fair: the series cannot pick a side at a discontinuity, so it splits the difference. Pointwise convergence is stronger than mean-square in spirit, but it can still fail at isolated nasty points even for some continuous functions.
The third and strongest sense is uniform convergence: S_N approaches f evenly everywhere at once, so the worst error anywhere on the interval shrinks to zero. Uniform convergence is the gold standard — it lets you swap limits, integrate term by term, and trust the picture. But it carries a steep price of admission: a uniformly convergent series of continuous sines must add up to a continuous function. So the instant f has a genuine jump, uniform convergence is impossible — and that impossibility is not a technicality. It is the doorway to the Gibbs phenomenon.
f piecewise smooth, finite energy:
mean-square integral of (f - S_N)^2 -> 0 (always holds)
pointwise S_N(x) -> f(x) where f continuous
S_N(x) -> (f(x-) + f(x+))/2 at a jump (midpoint rule)
uniform max error over interval -> 0 ONLY if f is continuous
at a jump: uniform convergence is IMPOSSIBLE -> Gibbs overshootThe overshoot that will not go away
Here is the picture that surprised everyone, including the physicists who first plotted it. Take the simplest possible jump: a square wave that sits at +1 on one half of the interval and -1 on the other. Build its Fourier sine series and plot the partial sum S_N. Across the flat stretches, S_N tracks the function beautifully. But right at the jump it does something stubborn: it shoots past the true value, overshooting by a fixed fraction — about 9 percent of the jump height — and it ripples for a moment before settling. This is the Gibbs phenomenon.
Now add more terms — go from N = 10 to N = 100 to N = 1000 — and watch what happens. The good news: the overshooting spike gets narrower, squeezing closer and closer to the jump. The unsettling news: the height of the spike does not shrink. It stalls at that same roughly 9 percent overshoot no matter how many terms you take. The bump never flattens; it only slides toward the cliff edge and gets thinner. The maximum error near the jump refuses to go to zero — which is precisely why uniform convergence is forbidden here, just as the continuity argument warned.
So is the series wrong? No — and reconciling this is the heart of the matter. The overshoot lives in an ever-thinner sliver of x. At any fixed point away from the jump, the spike has long since slid past you and S_N converges to f exactly: pointwise convergence holds. And because the bad region shrinks toward zero width, the area of the squared error vanishes too: mean-square convergence holds. The Gibbs overshoot is invisible to both of those gentler senses. It is only the strictest, uniform sense — which insists on controlling the worst error everywhere at once — that the overshoot defeats.
Why the overshoot is built in, not a bug
It is tempting to blame the overshoot on rounding, on too few terms, on a coding mistake. It is none of those. The Gibbs overshoot is a structural consequence of trying to assemble a discontinuous shape out of perfectly smooth, gently rolling sine waves. Each sine is infinitely smooth; a finite sum of them is infinitely smooth too. To mimic a vertical cliff with smooth curves, the partial sum has to rush upward very steeply right at the jump — and anything that rushes that hard overshoots before it can turn around, like a car braking too late at a wall. The ringing on either side is the smooth curve oscillating as it recovers.
There is a deeper reason hiding in the rates. The coefficients of a smooth function shrink very fast — the smoother f is, the faster its b_n decay, which is the flip side of the decay of Fourier modes you met in the heat equation. But a function with a jump has coefficients that fall off only like 1/n, agonizingly slowly. That sluggish 1/n tail is the mathematical fingerprint of a discontinuity, and it is exactly what leaves enough high-frequency energy lingering to build the persistent overshoot. Smoothness and fast convergence are two faces of one coin; jumps and Gibbs are the other.
What this means when you actually solve a PDE
Step back to the physics, because convergence is not an abstract worry here — it changes how you trust your answer. Suppose you solve the heat equation u_t = k u_xx with that discontinuous initial profile. The full solution carries each mode with its own e^(-k lambda_n t) damping. At t = 0 your series is exactly the Gibbs-ridden reconstruction of the jumpy data, complete with overshoot. But the instant time moves forward, those exponentials crush the high modes — the very modes responsible for the ringing — and the smoothing effect of diffusion erases the overshoot almost immediately. Gibbs is a t = 0 artifact for the heat equation; the equation itself heals it.
The wave equation u_tt = c^2 u_xx tells the opposite story, and the contrast is exactly the well-posedness lesson from the classification rung. The wave equation does not smooth: it has finite speed and faithfully carries corners along. A plucked string starts as a genuine kink, and that kink propagates as a real feature of the motion, not an error to be damped. So a truncated Fourier solution of a vibrating string keeps ringing near the corner for all time — the Gibbs overshoot is not an artifact you can wait out; it is your finite sum honestly struggling to represent a corner that the true solution really does have.
So the practical reading is this. Mean-square convergence is what guarantees your separated solution is the solution — it captures all the energy, and that is the sense in which the eigenfunction expansion is complete and correct. Pointwise convergence tells you the value at honest interior points, with the midpoint rule keeping you sane at jumps. And the missing uniform convergence, announced by Gibbs ringing, is your standing reminder that smooth pieces can only approximate a sharp edge, never reproduce it perfectly with finitely many terms. Knowing which sense applies is what separates a number you can trust from a wiggle you should expect.