Mathematics · Signal Processing
June 2026 · 14 min read · Spectral Analysis · DFT · FFT · Applications

The Fourier Transform — From Signals to Spectra

In 1822 Joseph Fourier showed that any periodic function can be expressed as an infinite sum of sines and cosines. Extended to non-periodic functions, this insight became the Fourier transform — perhaps the single most versatile mathematical tool in applied science. It decomposes any signal into its frequency components, enabling engineers and scientists to filter noise, compress images, scan human brains, detect gravitational waves, and build the modern internet. This article develops the theory from first principles and traces the transform through its most important applications.

1. Fourier Series: Decomposing Periodic Functions

For a function f(t) that is periodic with period T (so f(t+T) = f(t) for all t), the Fourier series representation is:

f(t) = a₀/2 + Σ[n=1 to ∞] ( aₙcos(2πnt/T) + bₙsin(2πnt/T) ) Coefficients (orthogonality of sin/cos over one period): a₀ = (2/T) · ∫[-T/2 to T/2] f(t) dt aₙ = (2/T) · ∫[-T/2 to T/2] f(t)cos(2πnt/T) dt bₙ = (2/T) · ∫[-T/2 to T/2] f(t)sin(2πnt/T) dt

In complex exponential form (more compact and general):

f(t) = Σ[n=−∞ to ∞] cₙ · e^(i2πnt/T) cₙ = (1/T) · ∫[-T/2 to T/2] f(t) · e^(−i2πnt/T) dt where: cₙ = (aₙ − ibₙ)/2 for n > 0 c₀ = a₀/2 c₋ₙ = cₙ* (complex conjugate, for real f)

The frequencies present are discrete multiples of the fundamental frequency f₀ = 1/T: {0, f₀, 2f₀, 3f₀, ...}. A square wave of period T and amplitude A has only odd harmonics:

Square wave: f(t) = (4A/π) · Σ[n=1,3,5,...] sin(2πnt/T) / n = (4A/π) · [sin(2πt/T) + sin(6πt/T)/3 + sin(10πt/T)/5 + ...]

This decomposition is why a square wave from a synthesiser sounds harsher than a sine wave — it contains energy at many harmonics that stimulate multiple resonances in the human ear.

2. The Continuous Fourier Transform

The Fourier series works for periodic functions. For a non-periodic signal f(t) that exists for all time, we take the limit T → ∞. The discrete frequencies merge into a continuum, and the Fourier series sum becomes an integral — the Fourier transform:

Forward transform: F(ω) = ∫[-∞ to ∞] f(t) · e^(−iωt) dt Inverse transform: f(t) = (1/2π) · ∫[-∞ to ∞] F(ω) · e^(iωt) dω where ω = 2πf is angular frequency (radians per second). Alternative convention using ordinary frequency ν (Hz): F(ν) = ∫[-∞ to ∞] f(t) · e^(−i2πνt) dt f(t) = ∫[-∞ to ∞] F(ν) · e^(i2πνt) dν (symmetric — no 1/2π factor)

F(ω) is generally a complex-valued function of frequency. Its magnitude |F(ω)| is the amplitude spectrum — how much of each frequency is present. The argument arg(F(ω)) is the phase spectrum — the relative timing of each frequency component.

Common Transform Pairs

f(t) = e^(−a|t|) ↔ F(ω) = 2a/(a² + ω²) (Lorentzian spectrum) f(t) = rect(t/T) ↔ F(ω) = T · sinc(ωT/2π) (sinc spectrum, where sinc(x)=sin(πx)/(πx)) f(t) = e^(−πt²) ↔ F(ν) = e^(−πν²) (Gaussian is its own transform!) f(t) = δ(t) ↔ F(ω) = 1 (impulse → flat spectrum) f(t) = 1 ↔ F(ω) = 2πδ(ω) (DC signal → single spike at ω=0) f(t) = cos(ω₀t) ↔ F(ω) = π[δ(ω−ω₀) + δ(ω+ω₀)]

3. Key Properties: Linearity, Shift, Scale

The Fourier transform satisfies several fundamental properties that make it powerful:

Linearity: ℱ{af(t) + bg(t)} = aF(ω) + bG(ω) Time shift: ℱ{f(t − t₀)} = e^(−iωt₀) · F(ω) (shift in time → phase rotation in frequency; magnitudes unchanged) Frequency shift: ℱ{e^(iω₀t) f(t)} = F(ω − ω₀) (multiply by complex exponential → shift frequency spectrum) This is the basis of radio modulation (AM, FM, SSB). Time scaling: ℱ{f(at)} = (1/|a|) · F(ω/a) Compress in time (a > 1) → stretch in frequency, lower amplitude. This is the time-bandwidth uncertainty principle in action. Duality: if f(t) ↔ F(ω), then F(t) ↔ 2π f(−ω) Differentiation: ℱ{f'(t)} = iω · F(ω) Derivatives in time → multiplication by iω in frequency. This converts ODEs to algebraic equations — the basis of Laplace/s-domain.

Uncertainty Principle

A signal cannot be arbitrarily localised in both time and frequency simultaneously. The Heisenberg-Gabor uncertainty principle for signals states:

Δt · Δω ≥ 1/2 where Δt = RMS time duration, Δω = RMS bandwidth. Equality holds for a Gaussian signal (minimum uncertainty packet). This is mathematically identical to the quantum mechanical uncertainty principle Δx · Δp ≥ ℏ/2, since the wavefunction is related to position by a Fourier transform.

4. Convolution Theorem

The convolution of two functions f and g is defined as:

(f * g)(t) = ∫[-∞ to ∞] f(τ) · g(t − τ) dτ

Convolution appears everywhere: the output of a linear time-invariant (LTI) system to input f(t) is the convolution of f with the system's impulse response h(t). A room's acoustic echo is the convolution of the dry signal with the room impulse response. Image blurring is 2D convolution with a blur kernel.

The convolution theorem is one of the most useful results in all of signal processing:

ℱ{f * g} = F(ω) · G(ω) (convolution in time ↔ multiplication in frequency) ℱ{f · g} = (1/2π) F * G (multiplication in time ↔ convolution in frequency)

This is transformative: instead of computing an O(N²) convolution directly, you can FFT both signals (O(N log N)), multiply them pointwise (O(N)), and inverse FFT (O(N log N)). Total: O(N log N) — a dramatic speedup for large signals.

Filtering in the Frequency Domain

A low-pass filter H(ω) = 1 for |ω| ≤ ω_c, 0 otherwise (ideal brick-wall filter) removes all frequency content above the cut-off frequency ω_c. Its time-domain impulse response is a sinc function:

h(t) = ℱ⁻¹{H(ω)} = (ω_c/π) · sinc(ω_c t/π)

Real filters use windowed sincs (Hann, Hamming, Kaiser windows) to avoid the Gibbs phenomenon — the ringing artifacts caused by the sharp cut-off in the ideal filter.

5. Parseval's Theorem and Energy Spectra

Parseval's theorem states that the total energy of a signal is the same whether computed in the time domain or the frequency domain:

∫[-∞ to ∞] |f(t)|² dt = (1/2π) · ∫[-∞ to ∞] |F(ω)|² dω In the symmetric convention (frequency in Hz): ∫[-∞ to ∞] |f(t)|² dt = ∫[-∞ to ∞] |F(ν)|² dν

The function |F(ω)|² is the power spectral density (PSD) — it shows how signal power is distributed across frequencies. Parseval's theorem is physically the statement that the Fourier transform is a unitary operator — it preserves inner products (and hence norms, hence energy).

Power Spectral Density in Practice

For a random stationary process x(t), the PSD S_xx(f) is defined as the Fourier transform of the autocorrelation function R_xx(τ) — the Wiener-Khinchin theorem:

S_xx(f) = ℱ{R_xx(τ)} where R_xx(τ) = E[x(t)·x(t+τ)] (expected value of the correlation) White noise: R_xx(τ) = σ²δ(τ) → S_xx(f) = σ² (flat spectrum) 1/f noise (pink noise): S_xx(f) ∝ 1/f (found in electronics, finance, music)

6. Discrete Fourier Transform and the FFT Algorithm

Digital computers work with discrete, finite sequences. The Discrete Fourier Transform (DFT) operates on N samples x[0], x[1], ..., x[N−1]:

Forward DFT: X[k] = Σ[n=0 to N−1] x[n] · e^(−i2πkn/N) k = 0, 1, ..., N−1 Inverse DFT: x[n] = (1/N) · Σ[k=0 to N−1] X[k] · e^(i2πkn/N) X[k] represents the amplitude and phase of frequency f_k = k·f_s/N where f_s = sampling frequency. Nyquist limit: only frequencies 0 ≤ f < f_s/2 are unambiguously represented. (Aliasing occurs if signal contains f > f_s/2 — Shannon sampling theorem.)

The Fast Fourier Transform (FFT)

Naive DFT computation requires O(N²) complex multiplications. For N = 10⁶ samples, this is 10¹² operations — intractable. The Cooley-Tukey FFT algorithm (1965) reduces this to O(N log₂ N) by recursively splitting the DFT:

Cooley-Tukey decimation-in-time (radix-2, N = 2^m): X[k] = Σ[n even] x[n]·e^(−i2πkn/N) + Σ[n odd] x[n]·e^(−i2πkn/N) = E[k] + e^(−i2πk/N) · O[k] where E[k] = DFT of even-indexed samples (N/2 point DFT) O[k] = DFT of odd-indexed samples (N/2 point DFT) Recursive splitting reduces operations from O(N²) to O(N log₂ N). For N = 10⁶: 10¹² → ~2×10⁷ operations — 50,000× speedup!

The FFT is routinely called one of the most important algorithms of the 20th century. It enabled real-time digital signal processing, digital audio workstations, software radio, and modern telecommunications.

Windowing: Direct DFT of a finite sample assumes the signal repeats periodically outside the window, causing spectral leakage for frequencies not exactly on the DFT grid. Applying a window function (Hann, Hamming, Blackman, Kaiser) tapers the signal smoothly to zero at the edges, greatly reducing leakage at the cost of slightly reduced frequency resolution.

7. Real-World Applications

Audio Equalisation and Music Production

A parametric equaliser decomposes the audio signal into frequency bands (using an FFT or filterbank), adjusts the gain at each frequency, and reconstructs the signal. The convolution theorem enables modern convolution reverb: record the impulse response of a concert hall (firing a starter pistol), then convolve any dry recording with it — placing that recording acoustically inside the concert hall. This requires FFT-based convolution to run in real time on a laptop.

MRI — Magnetic Resonance Imaging

In MRI, the radiofrequency signals detected by the receiver coil are actually the Fourier transform of the spatial spin density of the tissue. The scanner systematically samples this Fourier space (called k-space) and then performs a 2D inverse FFT to reconstruct the spatial image:

k-space: S(k_x, k_y) = ∫∫ ρ(x,y) · e^(−i2π(k_x·x + k_y·y)) dx dy Image reconstruction: ρ(x,y) = ∫∫ S(k_x,k_y) · e^(i2π(k_x·x + k_y·y)) dk_x dk_y

Incomplete k-space sampling (compressed sensing MRI) uses the sparsity of images in the wavelet domain to reconstruct full-resolution images from far fewer measurements, dramatically reducing scan time.

JPEG Image Compression

JPEG divides an image into 8×8 pixel blocks and applies the Discrete Cosine Transform (DCT) — a variant of the Fourier transform using only cosines:

DCT-II (standard JPEG): X[k] = Σ[n=0 to N−1] x[n] · cos( π(2n+1)k / 2N ) k = 0, ..., N−1 Properties: - Real-valued output (no complex numbers) - Energy compaction: natural images concentrate most energy in low-frequency DCT coefficients - Quantisation: high-frequency coefficients divided by larger quantisation steps (lossy step) - Run-length + Huffman encoding of quantised coefficients

The key insight: the human visual system is much more sensitive to low-frequency information (overall brightness and colour gradients) than high-frequency detail. JPEG exploits this by aggressively quantising the high-frequency DCT coefficients, achieving compression ratios of 10:1 or more with minimal perceptible quality loss.

Radio: Modulation and Software-Defined Radio

Every radio transmission involves the Fourier transform. Amplitude Modulation (AM) multiplies the baseband audio signal m(t) by a carrier cos(ω_c t) — by the frequency- shift property, this translates the audio spectrum to centre around ω_c. A software-defined radio receiver digitises a wide swath of RF spectrum, then uses FFT-based filtering to select any channel digitally — replacing analogue hardware with mathematics.

OFDM (Orthogonal Frequency Division Multiplexing), the modulation scheme behind WiFi, 4G/5G, and digital TV, transmits data on hundreds of orthogonal subcarriers simultaneously. The modulator is a single IFFT; the demodulator is a single FFT. The entire modern wireless internet runs on the Fourier transform.

Gravitational Wave Detection (LIGO)

When LIGO detected the first gravitational wave signal GW150914 in 2015 — a 0.2-second chirp from two merging black holes — the signal had an amplitude of 10⁻²¹ strain (the detector changed length by 1/1000th of a proton diameter over 4 km). Extracting this from detector noise required:

Signal-to-noise ratio for matched filter: SNR² = 4 · Re ∫[0 to ∞] |h̃(f)|² / S_n(f) df where h̃(f) = FFT of template waveform S_n(f) = one-sided noise PSD of detector This is the optimal linear filter for detecting a known signal in Gaussian noise (Wiener filter), and its implementation is entirely FFT-based.
〰️
Fourier Series Simulator
Build waveforms from harmonics and watch the spectrum evolve interactively
 🖼️
Image FFT Explorer
Visualise the 2D Fourier transform of images and explore filtering in k-space

Related Articles

The Fourier Transform — How It Works and Where It Appears The Fourier Transform — Decomposing Everything Into Waves Fourier Transform and Sound — DFT, FFT \u0026amp; Spectrograms Fractal Geometry — Self-Similarity and the Geometry of Nature Lotka–Volterra Equations — Predator–Prey Population Cycles