Convergence Issues for Fourier Series

Part 13, Chapter 13: Fourier Series and Transforms

Learning objectives

State the Dirichlet conditions and apply them at jumps and points of continuity
Recognize the Gibbs phenomenon visually and predict its size (about 9% of the jump)
State Parseval's theorem and use it to evaluate numerical series
Distinguish pointwise, uniform, and $L^2$ convergence

"Every reasonable function equals its Fourier series" sounds simple, until you ask what "reasonable" and "equals" actually mean. A discontinuous function like the square wave is the limit of smooth partial sums, but at the jump the partial sums overshoot by a stubborn 9% no matter how many terms you add. This is the Gibbs phenomenon, and it is real: it shows up every time a digital filter handles a sharp edge. This section lays out three different notions of convergence (pointwise, uniform, $L^2$ ), states the sharpest classical pointwise result (Dirichlet), and ends with the energy identity (Parseval) that gives surprisingly direct evaluations of $\sum 1/n^2$ and friends.

Dirichlet's pointwise theorem

Suppose $f$ is $2\pi$ -periodic and satisfies the Dirichlet conditions on $[-\pi, \pi]$ : piecewise continuous, with at most finitely many maxima and minima in any period, and at most finitely many jump discontinuities (each of finite size). Then the Fourier series of $f$ converges pointwise to:

$f(x)$ at every point of continuity, and
the average of the one-sided limits $\tfrac{1}{2}[f(x^+) + f(x^-)]$ at every jump.

So at a clean point the series gives you back the function, and at a jump it splits the difference. This is the cleanest classical convergence theorem and covers essentially every Fourier series you will meet in practice.

The Gibbs phenomenon

Even though the partial sums converge at every point, they do not converge uniformly near a jump. Plot the $N$ -th partial sum $S_N(x)$ N(x) of a square wave: just inside the jump, $S_N$ N overshoots the limit by about $8.95\%$ of the jump height. As $N$ grows the overshoot spike becomes narrower (shifting toward the discontinuity), but its height does not shrink. This is the Gibbs phenomenon, and it is a fundamental obstruction: no truncated Fourier series can be uniformly close to a discontinuous function on any neighborhood of the jump.

Plot the partial sums $S_N(x) = \sum_{k=1}^{N} (4/(\pi(2k-1))) \sin((2k-1)x)$ pi(2k−1)))sin((2k−1)x) of the square wave for $N = 4, 16, 64$ . Zoom in near $x = 0$ : you should see the overshoot spike maintain its height while moving closer to the jump as $N$ grows. That stubborn ~9% ridge is the Gibbs phenomenon caught on camera.

$L^2$ convergence and Parseval's theorem

The right notion of convergence for Fourier series is $L^2$ convergence in the inner-product norm: $\|S_N - f\|_{L^2}^2 = \int_{-\pi}^{\pi} |S_N(x) - f(x)|^2 \, dx \to 0$ int−pipi∣SN(x)−f(x)∣2,dxto0 as $N \to \infty$ . In this norm, every $L^2$ function is exactly recovered by its Fourier series. Pointwise convergence may fail on a small bad set; $L^2$ convergence never does.

The energy identity is Parseval's theorem:

$\dfrac{1}{\pi} \displaystyle\int_{-\pi}^{\pi} |f(x)|^2 \, dx = \dfrac{a_0^2}{2} + \sum_{n=1}^{\infty} \bigl(a_n^2 + b_n^2\bigr)$ −pipi∣f(x)∣2,dx=dfraca022+sumn=1inftybigl(an2+bn2bigr)

Read this physically: the total "energy" of the signal (left side) equals the sum of energies in each harmonic (right side). Parseval is the discrete analog of Plancherel for the Fourier transform (section 13.4) and is the workhorse for both energy estimates in signal processing and explicit numerical-series evaluations.

Where this shows up

Digital signal processing: Every digital low-pass filter chops off Fourier coefficients above some cutoff. The Gibbs overshoot appears as "ringing" near sharp transitions in the filtered signal, you hear it as a pre-echo on impulsive sounds, and you see it as halos around edges in compressed images. Engineers tame it with windowing (multiplying coefficients by smooth tapers) at the cost of less sharp frequency cutoff.
JPEG ringing artifacts: When JPEG quantises away high-frequency DCT coefficients, the reconstructed block exhibits Gibbs-style ringing near sharp edges, the visible "mosquito noise" around text and high-contrast borders in low-quality JPEGs.
Basel problem: Applying Parseval to $f(x) = x$ on $[-\pi, \pi]$ gives $\sum_{n=1}^{\infty} 1/n^2 = \pi^2 / 6$ n=1infty1/n2=pi2/6 in two lines. Euler computed this in 1734 by other means; Parseval makes it a routine calculation.
Spectroscopy: The total energy under a measured spectrum (line intensities summed) must equal the integrated power of the source, Parseval is the conservation law that lets spectroscopists cross-check their calibration.

Pause and think: If $f$ is continuous everywhere and piecewise smooth, must the Gibbs phenomenon appear? (Hint: where does the 9% overshoot live?)

Try it

The square wave equals $1$ for $0 < x < \pi$ and $-1$ for $-\pi < x < 0$ . At $x = 0$ the function has a jump. To what value does the Fourier series converge there?
The Fourier series of $f(x) = x$ on $[-\pi, \pi]$ has $b_n = 2(-1)^{n+1}/n$ n=2(−1)n+1/n. Use Parseval to compute $\sum_{n=1}^{\infty} 1/n^2$ n=1infty1/n2.
Predict first: at a jump of height $h$ , the partial-sum overshoot is approximately how big? Then verify with the function grapher using a square wave of height $1$ .
True or false: if the Fourier series of $f$ converges in $L^2$ , it converges pointwise at every point. (Answer: false. Counterexamples exist; $L^2$ convergence is strictly weaker than pointwise.)
A function $f$ has Fourier coefficients $a_n = 0$ n=0, $b_n = 1/n$ n=1/n. Does its Fourier series converge in $L^2$ ? (Use Parseval to check whether $\sum b_n^2 < \infty$ n2<infty.)

A trap to watch for

"Convergence" without an adjective is ambiguous. The Fourier series of a discontinuous function in $L^2[-\pi, \pi]$ converges in $L^2$ and converges pointwise at points of continuity, but it does not converge uniformly anywhere near a jump, the Gibbs spike rules that out forever. Whenever someone writes " $S_N \to f$ Ntof," train yourself to ask: in what norm? At which points? Without that clarification, the statement can be true and false at the same time.

What you now know

You can apply Dirichlet at any specific point of a piecewise smooth function, predict the Gibbs overshoot at a jump, and use Parseval both to certify $L^2$ convergence and to evaluate numerical series. The next section extends all of this from periodic functions on $[-\pi, \pi]$ to integrable functions on the entire real line, replacing the Fourier series with the Fourier transform.

Mark section complete →

References

Garrity, T. (2002). All the Mathematics You Missed: But Need to Know for Graduate School. Cambridge University Press, ch. 13.
Stein, E. M., Shakarchi, R. (2003). Fourier Analysis: An Introduction. Princeton University Press, ch. 2-3.
Folland, G. B. (1992). Fourier Analysis and Its Applications. Wadsworth & Brooks/Cole, ch. 2.
Korner, T. W. (1989). Fourier Analysis. Cambridge University Press, ch. 17 (Gibbs phenomenon), ch. 32 (Parseval).
Zygmund, A. (2002). Trigonometric Series (3rd ed.). Cambridge University Press, ch. 2 (Dirichlet kernel and pointwise convergence).