Convergence of Fourier Series

Introduction

When we represent a periodic signal as a Fourier series — an infinite sum of sinusoids — a natural and important question arises: does this infinite sum actually converge to the original signal? And if so, in what sense? These are not merely mathematical curiosities. They have practical implications for how accurately a truncated Fourier series approximates a signal, why ringing appears near discontinuities in digital filters, and what types of signals can be analyzed using Fourier methods at all.

For B.Tech students, understanding convergence helps you know when Fourier analysis will work well and when you should expect problems. The Gibbs phenomenon, in particular, directly affects filter design, signal reconstruction, and spectral analysis.

Dirichlet Conditions for Convergence

A periodic signal $x(t)$ with period $T$ has a convergent Fourier series if it satisfies the Dirichlet conditions:

1. Absolutely integrable over one period: $\int_0^T |x(t)| dt < \infty$
2. Finite number of maxima and minima in one period (bounded variation)
3. Finite number of discontinuities in one period, each of finite size (jump discontinuities only)

If these conditions are met, the Fourier series converges to:

• $x(t)$ at every point where $x(t)$ is continuous
• The average of left and right limits at discontinuities: $\frac{x(t^+) + x(t^-)}{2}$

Signals That Satisfy Dirichlet Conditions

Most signals encountered in engineering satisfy these conditions: square waves, sawtooth waves, triangular waves, pulse trains, and any signal composed of smooth segments joined at finite discontinuities.

A Pathological Example

The function $x(t) = \sin(1/t)$ near $t = 0$ oscillates infinitely many times in any interval containing the origin — violating condition 2. However, such signals rarely arise in physical systems.

Types of Convergence

Pointwise Convergence

The partial sum $S_N(t) = \sum_{n=-N}^{N} c_n e^{jn\omega_0 t}$ converges pointwise to $x(t)$ if:

$$\lim_{N \to \infty} S_N(t_0) = x(t_0)$$

for each specific point $t_0$. Under Dirichlet conditions, pointwise convergence holds at every point of continuity. At discontinuities, it converges to the midpoint value.

Uniform Convergence

Uniform convergence means the partial sum approximation is equally good everywhere simultaneously:

$$\lim_{N \to \infty} \sup_t |S_N(t) - x(t)| = 0$$

This is stronger than pointwise convergence. Uniform convergence holds only for continuous periodic signals. At discontinuities, uniform convergence fails due to the Gibbs phenomenon — the overshoot near the discontinuity never goes away no matter how many terms you include.

Mean-Square (L²) Convergence

The Fourier series minimizes the mean-square error:

$$\lim_{N \to \infty} \frac{1}{T}\int_0^T |x(t) - S_N(t)|^2 dt = 0$$

This is the sense in which the Fourier series is the "best" approximation with $N$ harmonics. Mean-square convergence holds for any signal with finite power (energy per period), even at discontinuities. The integrated error goes to zero, even though the pointwise error at discontinuities doesn't.

Parseval's Theorem (Energy Conservation)

Mean-square convergence implies Parseval's theorem:

$$\frac{1}{T}\int_0^T |x(t)|^2 dt = \sum_{n=-\infty}^{\infty} |c_n|^2$$

The total power in the time domain equals the sum of powers of all harmonic components. This is energy conservation in the frequency domain — no energy is created or destroyed by the Fourier representation.

The Gibbs Phenomenon

What It Is

When a Fourier series is truncated to $N$ terms for a signal with discontinuities, the partial sum exhibits oscillatory overshoot near each discontinuity. As $N$ increases:

• The oscillations become narrower (compressed toward the discontinuity)
• But the peak overshoot does NOT decrease — it remains at approximately 8.95% of the jump size

For a unit step discontinuity (jump of size 1), the Fourier partial sum overshoots to approximately $1 + 0.0895 = 1.0895$ just beyond the discontinuity.

Mathematical Explanation

The peak overshoot is related to the integral:

$$\frac{1}{\pi}\int_0^{\pi} \frac{\sin t}{t} dt \approx 1.0895$$

The $\frac{\sin t}{t}$ integral (sine integral) evaluates to a value greater than 1, causing the persistent overshoot. This is a fundamental property of the sinc kernel used in Fourier analysis.

Why It Matters in Practice

• Digital filter design: Sharp cutoff filters exhibit Gibbs ringing at transitions, motivating the use of window functions
• Signal reconstruction: Reconstructing a signal from bandlimited samples shows ringing near discontinuities
• Image processing: Sharp edges in images show ringing when processed with frequency-domain filters
• Spectral analysis: Truncating a time-domain signal (multiplying by a rectangular window) introduces Gibbs-like sidelobes in the spectrum

Mitigating Gibbs Phenomenon

Windowing: Instead of abruptly truncating the Fourier series, apply a smooth window to the coefficients:

• Fejér summation: Average partial sums → eliminates overshoot entirely (converges uniformly for continuous signals)
• Lanczos sigma factor: Multiply $c_n$ by $\text{sinc}(n/N)$ to reduce overshoot
• Smooth windows (Hamming, Blackman): Taper coefficients to reduce discontinuity effects

Fejér kernel (Cesàro summation):

$$\sigma_N(t) = \frac{1}{N+1}\sum_{n=0}^{N} S_n(t)$$

This arithmetic average of partial sums always converges uniformly to $x(t)$ at points of continuity and eliminates the Gibbs overshoot completely.

Rate of Convergence

The rate at which Fourier coefficients $c_n$ decay determines how quickly the series converges:

• Discontinuous signals (square wave): $|c_n| \sim 1/n$ — slow decay, need many terms
• Continuous signals (triangle wave): $|c_n| \sim 1/n^2$ — moderate decay
• Smooth signals (continuously differentiable): $|c_n| \sim 1/n^k$ for $k-1$ continuous derivatives
• Infinitely smooth signals (Gaussian): $|c_n|$ decays faster than any power of $1/n$ (exponential decay)

Rule of thumb: Each additional order of continuity in $x(t)$ adds one power of $1/n$ to the coefficient decay rate. Smoother signals are better represented by fewer Fourier terms.

Worked Examples

Example 1: The Fourier series of a square wave (period $2\pi$, amplitude 1) is:

$$x(t) = \frac{4}{\pi}\sum_{k=0}^{\infty}\frac{\sin((2k+1)t)}{2k+1}$$

At $t = 0$ (discontinuity): Series gives $\frac{x(0^+)+x(0^-)}{2} = \frac{1+(-1)}{2} = 0$ ✓

The $N$-term partial sum overshoots to $\approx 1.09$ near $t = 0^+$.

Example 2: The triangular wave has coefficients decaying as $1/n^2$. Since it is continuous, there is no Gibbs phenomenon — the Fourier series converges uniformly.

Key Takeaways

• Dirichlet conditions guarantee Fourier series convergence for most engineering signals
• At discontinuities, the series converges to the midpoint value $(x(t^+) + x(t^-))/2$
• The Gibbs phenomenon causes ~9% overshoot at discontinuities regardless of how many terms are used
• Mean-square convergence always holds for finite-power signals — the integrated error goes to zero
• Coefficient decay rate: $1/n^k$ where $k$ relates to signal smoothness — smoother signals converge faster
• Windowing (Fejér, Hamming, etc.) mitigates Gibbs phenomenon in practical applications
• Parseval's theorem confirms energy conservation between time and frequency representations