9.8 Fourier analysis

Foundations of Fourier analysis

Fourier analysis is a method for breaking down complex functions into sums of simple sine and cosine waves. The core idea: any sufficiently well-behaved periodic function can be expressed as a combination of sinusoidal components, each with a specific frequency and amplitude. This turns complicated problems into manageable ones, because sine and cosine functions are extremely well understood.

This technique shows up everywhere, from compressing audio files to solving differential equations to processing medical images. Understanding Fourier analysis means understanding how to move between two ways of looking at the same function: what it does over time, and what frequencies it contains.

Periodic functions

A periodic function repeats its values at regular intervals. The smallest such interval is called the period, denoted $T$. Formally:

$f(x + T) = f(x)$

Sine and cosine are the most basic periodic functions, with period $2\pi$. But periodic functions also model real-world cycles: sound waves, seasonal temperatures, alternating current, and orbital motion. Fourier analysis takes advantage of the fact that these repeating patterns can always be built from sinusoidal pieces.

Trigonometric series

A trigonometric series is an infinite sum of sine and cosine terms used to represent a periodic function:

$f(x) = a_0 + \sum_{n=1}^{\infty} (a_n \cos(nx) + b_n \sin(nx))$

• The constant $a_0$ captures the average value of the function.
• Each pair $a_n \cos(nx) + b_n \sin(nx)$ represents the contribution of the $n$th harmonic, a sinusoidal component oscillating $n$ times faster than the fundamental frequency.
• The coefficients $a_n$ and $b_n$ control how much each harmonic contributes.

This series provides the link between a periodic function and its frequency content: you go from one complicated shape to a recipe of simple waves.

Fourier series representation

The Fourier series expresses a periodic function as a sum of sinusoidal components. The lowest-frequency term ($n = 1$) oscillates at the fundamental frequency, which matches the period of the original function. Higher harmonics ($n = 2, 3, \ldots$) oscillate at integer multiples of that fundamental frequency.

To find the coefficients, you use inner products with the trigonometric basis functions (the integral formulas below). The result is a decomposition that captures both the coarse shape and fine detail of the original waveform, with each term adding more precision.

Key concepts in Fourier analysis

Frequency domain vs. time domain

These are two different ways of representing the same signal:

• The time domain shows how a signal changes over time: amplitude as a function of $t$.
• The frequency domain shows which frequencies are present and how strong each one is.

The Fourier transform converts between these two representations. A signal that looks complicated in the time domain might reveal a simple structure in the frequency domain. For example, a musical chord that appears as a messy waveform in time becomes a few distinct spikes at specific frequencies.

Fourier coefficients

The Fourier coefficients tell you exactly how much each frequency contributes to the overall function. For a function with period $2\pi$:

$a_n = \frac{1}{\pi} \int_{-\pi}^{\pi} f(x) \cos(nx) \, dx$

$b_n = \frac{1}{\pi} \int_{-\pi}^{\pi} f(x) \sin(nx) \, dx$

The magnitude of a coefficient tells you the strength of that frequency component. Larger coefficients mean that frequency plays a bigger role in shaping the function. When working with complex exponential form, the phase of a coefficient encodes the timing offset of that component.

Convergence of Fourier series

A natural question: does the Fourier series actually converge to the original function? The answer depends on the function's properties.

• Pointwise convergence: the series converges to $f(x)$ at each individual point. This holds at points where $f$ is continuous (under mild conditions).
• Uniform convergence: the series converges to $f(x)$ at the same rate everywhere. This requires stronger smoothness conditions.
• Smoother functions generally have faster-converging Fourier series, because their coefficients decay more rapidly.
• Near discontinuities, the series exhibits the Gibbs phenomenon (covered below), where the partial sums overshoot the true value.

Applications of Fourier analysis

Signal processing

Fourier analysis is the backbone of modern signal processing. By converting a signal to the frequency domain, you can:

• Filter out unwanted noise by zeroing out certain frequency components
• Compress audio and image data (MP3 and JPEG both rely on frequency-domain representations)
• Analyze speech patterns, musical compositions, and acoustic environments
• Design communication systems that modulate and demodulate signals at specific carrier frequencies

Radar and sonar technologies also depend on Fourier methods to extract useful information from reflected signals.

Data compression

Compression works by representing data in the frequency domain and then discarding components that contribute the least. A natural image, for instance, has most of its energy concentrated in low-frequency components. You can throw away the high-frequency detail with minimal perceptible loss.

• Lossy compression (JPEG, MP3) deliberately removes less significant frequency components to reduce file size.
• The tradeoff is always between file size and quality.
• This approach exploits the fact that real-world signals tend to have redundancy that becomes visible in the frequency domain.

Partial differential equations

Fourier methods transform certain PDEs into simpler algebraic equations. The general strategy:

1. Express the unknown function as a Fourier series or apply a Fourier transform.
2. The differential equation becomes an algebraic equation in the frequency domain (derivatives turn into multiplications).
3. Solve the simpler equation for the Fourier coefficients or transform.
4. Transform back to get the solution in the original domain.

This technique is standard for problems involving heat conduction, wave propagation, and fluid dynamics. It also appears in quantum mechanics, where Fourier transforms help solve the Schrödinger equation.

Fourier transforms

Fourier series work for periodic functions. The Fourier transform generalizes the idea to functions that are not periodic, defined over the entire real line.

Continuous Fourier transform

The continuous Fourier transform converts a time-domain function $f(t)$ into a frequency-domain function $F(\omega)$:

$F(\omega) = \int_{-\infty}^{\infty} f(t) \, e^{-i\omega t} \, dt$

The inverse transform recovers the original function:

$f(t) = \frac{1}{2\pi} \int_{-\infty}^{\infty} F(\omega) \, e^{i\omega t} \, d\omega$

Think of this as decomposing $f(t)$ into a continuous superposition of complex exponentials $e^{i\omega t}$, each weighted by $F(\omega)$. Where Fourier series give you discrete frequency components, the Fourier transform gives you a continuous spectrum.

Discrete Fourier transform

In practice, you work with sampled data: a finite sequence of $N$ values rather than a continuous function. The discrete Fourier transform (DFT) handles this case:

$X_k = \sum_{n=0}^{N-1} x_n \, e^{-i2\pi kn/N}$

The inverse:

$x_n = \frac{1}{N} \sum_{k=0}^{N-1} X_k \, e^{i2\pi kn/N}$

The DFT produces $N$ frequency components from $N$ data points. It's the foundation of digital signal processing, spectral analysis, and digital filtering.

Fast Fourier transform (FFT)

Computing the DFT directly requires $O(N^2)$ operations, which becomes impractical for large datasets. The Fast Fourier Transform is an algorithm that computes the same result in $O(N \log N)$ operations.

The most common implementation is the Cooley-Tukey algorithm, which recursively splits the DFT into smaller sub-problems. The FFT is what makes real-time audio processing, medical imaging, and large-scale spectral analysis computationally feasible.

Properties of Fourier transforms

These properties make Fourier analysis a practical tool, not just a theoretical one. They let you predict how operations in one domain affect the other domain, often simplifying calculations dramatically.

Linearity

$F\{af(t) + bg(t)\} = aF\{f(t)\} + bF\{g(t)\}$

The transform of a sum is the sum of the transforms. This means you can analyze complex signals by breaking them into simpler parts, transforming each part separately, and adding the results. It's the mathematical basis for the superposition principle in linear systems.

Time-shifting

$F\{f(t - t_0)\} = e^{-i\omega t_0} F(\omega)$

Delaying a signal by $t_0$ in the time domain multiplies its frequency-domain representation by a complex exponential. The magnitude spectrum stays the same; only the phase changes. This is why a delayed signal sounds the same but arrives later.

Frequency-shifting

$F\{e^{i\omega_0 t} f(t)\} = F(\omega - \omega_0)$

Multiplying a signal by a complex exponential in the time domain shifts its entire spectrum by $\omega_0$. This is the principle behind frequency modulation: you move a signal's frequency content to a different part of the spectrum. Radio broadcasting relies on this to assign different stations to different carrier frequencies.

Convolution theorem

$F\{f(t) * g(t)\} = F(\omega) \cdot G(\omega)$

Convolution in the time domain becomes simple multiplication in the frequency domain. This is enormously useful because convolution (which describes how linear systems respond to inputs) is computationally expensive, while multiplication is cheap. In practice, you can compute convolutions efficiently by:

1. Transforming both signals to the frequency domain (using FFT).
2. Multiplying the transforms pointwise.
3. Inverse-transforming the result.

Fourier analysis in higher dimensions

Multidimensional Fourier transforms

Fourier analysis extends naturally to functions of multiple variables. The 2D Fourier transform, for example:

$F(u,v) = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} f(x,y) \, e^{-i2\pi(ux + vy)} \, dx \, dy$

This decomposes a 2D function (like an image) into spatial frequency components. Low spatial frequencies correspond to gradual changes across the image; high spatial frequencies correspond to sharp edges and fine detail. Applications include:

• Medical imaging: MRI machines actually acquire data in the frequency domain and use inverse Fourier transforms to reconstruct images.
• Computer vision: frequency-domain filtering for edge detection and texture analysis.
• Image processing: noise removal, sharpening, and compression.

Fourier analysis on groups

At a more abstract level, Fourier analysis can be extended to algebraic structures called groups. Classical Fourier series correspond to analysis on the circle group, and the Fourier transform on $\mathbb{R}$ is a special case of analysis on locally compact abelian groups. This generalization, called harmonic analysis, connects to representation theory, number theory (analytic number theory, zeta functions), and quantum mechanics.

Limitations and extensions

Gibbs phenomenon

When a Fourier series approximates a function with a jump discontinuity, the partial sums overshoot near the discontinuity by about 9% of the jump size. Adding more terms doesn't reduce this overshoot; it just makes the overshoot region narrower.

This is the Gibbs phenomenon, and it's an inherent limitation of representing discontinuous functions with smooth sinusoidal components. Techniques like sigma approximation (applying a smoothing factor to the coefficients) can reduce the effect.

Wavelet analysis

Wavelets address a key limitation of Fourier analysis: standard Fourier methods tell you which frequencies are present but not when they occur. Wavelets are localized basis functions that provide simultaneous time and frequency information.

• Wavelets are better suited for analyzing transient signals and signals with discontinuities.
• Multi-resolution analysis lets you examine a signal at different scales.
• Applications include signal denoising, data compression, and feature extraction.

Short-time Fourier transform

The short-time Fourier transform (STFT) is a compromise between the time and frequency domains. It works by:

1. Sliding a window function $w(t - \tau)$ along the signal.

2. Computing the Fourier transform of each windowed segment.

$STFT\{x(t)\}(\tau, \omega) = \int_{-\infty}^{\infty} x(t) \, w(t - \tau) \, e^{-i\omega t} \, dt$

The result is a time-frequency representation. There's a fundamental tradeoff: a narrow window gives good time resolution but poor frequency resolution, and vice versa. This tradeoff is related to the uncertainty principle (yes, the same concept from quantum mechanics). The STFT is widely used in speech processing, music analysis, and radar.

Computational aspects

Numerical methods for Fourier analysis

When implementing Fourier analysis computationally, several practical issues arise:

• Spectral leakage: analyzing a finite segment of a signal can introduce artifacts. Windowing functions (Hamming, Hanning, Blackman) taper the signal at the edges to reduce this effect.
• Zero-padding: adding zeros to a data sequence before computing the DFT increases the number of frequency bins, giving finer frequency resolution in the output (though it doesn't add new information).
• Aliasing: if you sample a signal too slowly, high frequencies masquerade as low frequencies. The Nyquist theorem states you must sample at least twice the highest frequency present.
• Numerical stability and error accumulation matter in large-scale computations.

Software tools for Fourier analysis

• Python: NumPy's numpy.fft module and SciPy's scipy.fft provide standard FFT implementations.
• MATLAB: the Signal Processing Toolbox offers comprehensive Fourier analysis functions.
• FFTW ("Fastest Fourier Transform in the West"): a high-performance C library for FFT computations.
• Open-source alternatives like GNU Octave and Scilab offer similar functionality for academic use.

Fourier analysis in mathematics

Functional analysis connections

Fourier analysis connects deeply to functional analysis and the theory of Hilbert spaces. The set of functions $\{1, \cos(nx), \sin(nx)\}$ forms an orthonormal basis for the space of square-integrable periodic functions. This means:

• Fourier series are just expansions in this orthonormal basis, analogous to expressing a vector in terms of basis vectors.
• Parseval's theorem states that the total energy of a function equals the sum of the squared magnitudes of its Fourier coefficients: $\frac{1}{\pi}\int_{-\pi}^{\pi} |f(x)|^2 dx = 2a_0^2 + \sum_{n=1}^{\infty}(a_n^2 + b_n^2)$. Energy is preserved when you move between domains.
• Spectral theory uses Fourier techniques to analyze linear operators, with applications in quantum mechanics and operator algebras.

Harmonic analysis foundations

Harmonic analysis generalizes Fourier methods to more abstract mathematical settings, studying functions on topological groups and homogeneous spaces. It connects to the representation theory of Lie groups and has deep applications in number theory (particularly through zeta functions and L-functions). At its core, harmonic analysis provides a framework for understanding symmetries in mathematical structures, extending the same decomposition philosophy that makes classical Fourier analysis so effective.