Basic Fourier Transform

Continuous Fourier transform mapping signals between time/space domains and spectral frequency representations.

What it is

The Fourier transform is a linear integral operator that maps a continuous, non-periodic time or space domain function $f(t)$ into a continuous spectral frequency domain representation $\hat{f}(\omega)$: $\hat{f}(\omega) = \mathcal{F}\{f(t)\} = \int_{-\infty}^\infty f(t) e^{-i \omega t} dt$ The original continuous function is uniquely recovered via the inverse Fourier transform: $f(t) = \mathcal{F}^{-1}\{\hat{f}(\omega)\} = \frac{1}{2\pi} \int_{-\infty}^\infty \hat{f}(\omega) e^{i \omega t} d\omega$ It emerges naturally as the continuous continuum limit of Fourier series as the fundamental period approaches infinity ($T \to \infty$). In this limit, discrete harmonic spacing $\Delta\omega = 2\pi/T$ becomes the infinitesimal differential $d\omega$, and discrete line coefficients merge into a continuous spectral amplitude density $\hat{f}(\omega)$.

Why it appears in physics

It appears in physics because real physical events do not repeat periodically into infinite past and future: an ultrashort laser pulse, a seismic shockwave, an acoustic collision, or a particle traversing a particle accelerator are non-repeating transients with distinct beginnings and endings.

The Fourier transform decomposes arbitrary localized wave packets into a continuous superposition of harmonic plane waves. Furthermore, it converts linear differential operators into simple algebraic multiplication: temporal differentiation $\frac{d}{dt}$ maps to multiplication by $i\omega$, transforming partial differential equations (heat diffusion, wave propagation, free-particle Schrödinger equation) into elementary algebraic problems.

Visual or geometric idea

Imagine a localized time signal $f(t)$ as an intricate musical chord comprising an unbroken continuum of frequencies. The Fourier transform acts like a prism, decomposing white light into its continuous rainbow band of constituent colors.

The magnitude $|\hat{f}(\omega)|$ quantifies the spectral distribution (which frequencies are present and how strongly they contribute), while the argument $\arg(\hat{f}(\omega))$ tracks the phase alignment of each spectral component. An extremely short pulse in time demands an exceptionally broad frequency bandwidth to construct; conversely, a near-monochromatic wave must persist over an extensive duration. This reciprocity reflects the Gabor limit in communications and Heisenberg's uncertainty principle in quantum physics.

Minimal symbolic language

• Forward transform: $\hat{f}(\omega) = \int_{-\infty}^\infty f(t) e^{-i \omega t} dt$.
• Inverse transform: $f(t) = \frac{1}{2\pi} \int_{-\infty}^\infty \hat{f}(\omega) e^{i \omega t} d\omega$.
• Derivative operational property: $\mathcal{F}\left\{\frac{d^n f}{dt^n}\right\} = (i\omega)^n \hat{f}(\omega)$.
• Time translation property: $\mathcal{F}\{f(t - t_0)\} = e^{-i \omega t_0} \hat{f}(\omega)$.
• Frequency modulation property: $\mathcal{F}\{f(t) e^{i \omega_0 t}\} = \hat{f}(\omega - \omega_0)$.
• Convolution theorem: $\mathcal{F}\{(f * g)(t)\} = \hat{f}(\omega) \cdot \hat{g}(\omega)$, where $(f * g)(t) = \int_{-\infty}^\infty f(\tau) g(t - \tau) d\tau$.
• Plancherel / Parseval identity: $\int_{-\infty}^\infty |f(t)|^2 dt = \frac{1}{2\pi} \int_{-\infty}^\infty |\hat{f}(\omega)|^2 d\omega$.
• Bandwidth uncertainty relation: $\Delta t \cdot \Delta \omega \ge \frac{1}{2}$.

What the result means

The complex function $\hat{f}(\omega)$ represents the spectral amplitude density. Its squared modulus, $|\hat{f}(\omega)|^2$, represents the energy spectral density (power spectrum), measuring how many joules of signal energy reside per unit frequency interval (Hz or rad/s).

The convolution theorem is of monumental physical utility: the output of any linear time-invariant (LTI) physical system—such as an electronic filter, an optical imaging lens, or a seismometer—subjected to an input $x(t)$ is the convolution of $x(t)$ with the system impulse response $h(t)$: $y(t) = (x * h)(t)$. In the Fourier domain, this simplifies to pure algebraic multiplication: $\hat{y}(\omega) = \hat{h}(\omega) \hat{x}(\omega)$, where $\hat{h}(\omega)$ is the system transfer function.

When not to use it

Do not attempt standard Fourier transforms on functions that blow up at infinity (such as growing exponentials $e^{a t}$ with $a > 0$ or unbounded polynomials), as the integral fails to converge in the classical $L^2(\mathbb{R})$ sense. Such unstable systems require the Laplace transform.

Additionally, remain vigilant regarding disciplinary normalization conventions: physics often uses symmetric $\frac{1}{\sqrt{2\pi}}$ factors, engineering often uses ordinary cyclic frequency $f$ (eliminating factors of $2\pi$), and quantum mechanics reverses the sign of $i$ in exponential definitions. Inconsistent conventions corrupt amplitude constants and phase angles.

Operational summary

Confirm that the signal is absolutely integrable ($\int |f(t)| dt < \infty$). Compute the forward integral $\int f(t) e^{-i\omega t} dt$ directly or invoke standard transform pairs (Gaussian pulses, rect pulses, decaying exponentials). To solve linear differential equations, substitute $\frac{d}{dt} \to i\omega$, algebraically isolate the output spectrum $\hat{y}(\omega)$, and invert back to the time domain using the inverse transform or the convolution theorem.

Procedure

Before calculating

Confirm that $f(t)$ satisfies absolute integrability over $(-\infty, \infty)$ or determine whether generalized distribution techniques are required (such as for infinite sinusoids or Heaviside steps).

Verify the adopted convention for angular frequency $\omega$ and normalization factors (typically unity on forward, $\frac{1}{2\pi}$ on inverse).

Step 1: identify the mathematical object

Inspect the analytical profile of $f(t)$ and its symmetries:

• If $f(t)$ is real and even, $\hat{f}(\omega)$ is purely real and even.
• If $f(t)$ is real and odd, $\hat{f}(\omega)$ is purely imaginary and odd.
• Rectangular pulses of duration $\tau$ yield sinc cardinal spectral profiles $\frac{\sin(x)}{x}$.
• Gaussian temporal pulses yield Gaussian frequency spectra.

Step 2: choose the operation

For direct calculation of elementary wave profiles, set up the forward integral: $\hat{f}(\omega) = \int_{-\infty}^\infty f(t) e^{-i\omega t} dt$.

For differential equations or filtering, use transform operational theorems: $\frac{d f}{dt} \to i\omega \hat{f}(\omega)$, $f(t - t_0) \to e^{-i\omega t_0} \hat{f}(\omega)$, $(f * g)(t) \to \hat{f}(\omega) \hat{g}(\omega)$.

Step 3: preserve units and signs

Check dimensional consistency: if $f(t)$ has units of $[U]$ and $t$ is in seconds, then $\hat{f}(\omega)$ must carry dimensions of $[U] \cdot \text{s}$ (or $[U]/\text{Hz}$).

Preserve the negative sign in the forward transform exponent: $e^{-i\omega t}$.

Step 4: calculate

Evaluate the integral across the non-zero support domain of the function.

When dealing with oscillatory tails, utilize Euler expansions and symmetric limits.

Step 5: interpret the result

Plot or examine the magnitude spectrum $|\hat{f}(\omega)|$, noting spectral bandwidth at full-width at half-maximum (FWHM).

Validate the reciprocal uncertainty scaling: verify that halving the pulse duration $\Delta t$ precisely doubles the spectral bandwidth $\Delta\omega$.

Self-check checklist

• Are the physical units of $\hat{f}(\omega)$ equal to $[f(t)] \times \text{time}$?
• Did you exploit parity to convert complex exponentials into real cosines or sines?
• Did you apply convolution for system outputs instead of multiplying time signals directly?
• Does the spectrum of a real physical signal satisfy Hermitian symmetry: $\hat{f}(-\omega) = \hat{f}^*(\omega)$?

Guided example

Physical situation

A rectangular optical pulse from a laser passes through a fast electro-optic shutter that stays open for duration $\tau = 2\text{ ps}$ centered at $t=0$. The electric field is constant at $E_0 = 1000\text{ V/m}$ for $-\tau/2 < t < \tau/2$ and zero elsewhere. We wish to calculate its Fourier transform and identify its spectral bandwidth.

Mathematical translation

Model the signal as a symmetric rectangular pulse $f(t) = E_0 \ \text{rect}(t/\tau)$. Because $f(t)$ is real and even, its Fourier transform $\hat{f}(\omega)$ is purely real and even: $\hat{f}(\omega) = \int_{-\tau/2}^{\tau/2} E_0 e^{-i \omega t} dt$.

Step-by-step calculation

1. Set up the definite integral: $\hat{f}(\omega) = E_0 \int_{-\tau/2}^{\tau/2} e^{-i \omega t} dt$.
2. Integrate the complex exponential: $\hat{f}(\omega) = E_0 \left[ \frac{e^{-i \omega t}}{-i \omega} \right]_{-\tau/2}^{\tau/2} = E_0 \frac{e^{-i \omega \tau/2} - e^{i \omega \tau/2}}{-i \omega}$.
3. Apply Euler's identity $\sin(x) = \frac{e^{ix} - e^{-ix}}{2i}$: $\hat{f}(\omega) = E_0 \frac{2i \sin(\omega \tau / 2)}{i \omega} = E_0 \tau \frac{\sin(\omega \tau / 2)}{\omega \tau / 2}$.
4. Express via the normalized cardinal sinc function: $\hat{f}(\omega) = E_0 \tau \ \text{sinc}\left(\frac{\omega \tau}{2\pi}\right)$ where $\text{sinc}(u) = \frac{\sin(\pi u)}{\pi u}$.
5. Find the first spectral nulls: The numerator vanishes when $\frac{\omega \tau}{2} = \pm \pi \implies \omega_{\text{null}} = \pm \frac{2\pi}{\tau}$. In cyclic frequency $f = \frac{\omega}{2\pi}$: $f_{\text{null}} = \pm \frac{1}{\tau} = \pm \frac{1}{2 \times 10^{-12}\text{ s}} = \pm 0.5\text{ THz}$.

Reading the result

The rectangular pulse spectrum features a central lobe centered at $\omega = 0$ with null-to-null bandwidth $\Delta f = \frac{2}{\tau} = 1.0\text{ THz}$. The shorter the pulse duration (smaller $\tau$), the broader its spectral bandwidth, illustrating the fundamental time-frequency reciprocity.

Common error avoided

We avoided assuming that a constant temporal amplitude only contains DC frequency, recognizing that sharp transient edges inject an infinite continuum of high-frequency sinc sidelobes.

Inverse example

Given result

The emission spectrum of an electromagnetic wave packet is a Gaussian profile centered at $\omega_0$: $\hat{E}(\omega) = E_0 \sigma_t \sqrt{2\pi} e^{-\frac{\sigma_t^2 (\omega - \omega_0)^2}{2}}$.

What physical question it can answer

Applying the inverse Fourier transform reconstructs the exact temporal wave packet: $E(t) = E_0 e^{-\frac{t^2}{2\sigma_t^2}} e^{i\omega_0 t}$. This proves that a Gaussian spectrum produces a Gaussian temporal envelope oscillating at carrier frequency $\omega_0$, demonstrating that the Gaussian wave packet is the unique waveform that achieves the theoretical minimum uncertainty product $\Delta t \cdot \Delta \omega = \frac{1}{2}$.

Common errors

Error 1: conflating the continuous Fourier transform with discrete Fourier series

It seems reasonable to integrate over $[0, T]$ when attempting to find the Fourier transform of a non-periodic pulse. It fails because Fourier series apply strictly to infinitely repeating periodic signals yielding discrete line coefficients $c_n$, whereas the Fourier transform applies to finite-energy non-periodic transients integrating across $(-\infty, \infty)$ to produce a continuous spectral density $\hat{f}(\omega)$. It is detected when the calculated output yields discrete values instead of a continuous function of $\omega$. It is corrected by reserving series expansions for periodic signals and continuous transforms for isolated pulses. Mini-example: an isolated pulse does not have discrete harmonic lines.

Error 2: omitting the 1/(2pi) factor in the inverse Fourier transform

It seems reasonable to define the inverse transform identically to the forward transform except for reversing the exponential sign: $\int \hat{f}(\omega) e^{i\omega t} d\omega$. It fails because evaluating the nested identity $\int e^{i(\omega - \omega')t} dt = 2\pi \delta(\omega - \omega')$ proves that omitting $\frac{1}{2\pi}$ inflates the reconstructed signal by a factor of $2\pi \approx 6.28$. It is detected when forward followed by inverse transformation scales amplitudes by $2\pi$. It is corrected by strictly incorporating $\frac{1}{2\pi}$ alongside the $d\omega$ integration. Mini-example: $\mathcal{F}^{-1}\{\mathcal{F}\{f(t)\}\} = f(t)$ holds only when the inverse operator includes $\frac{1}{2\pi}$.

Error 3: multiplying signals in the time domain expecting their spectra to multiply

It seems reasonable to assume that if $y(t) = x(t) \cdot h(t)$, then its frequency spectrum must be $\hat{y}(\omega) = \hat{x}(\omega) \cdot \hat{h}(\omega)$. It fails because the convolution theorem establishes that point-by-point multiplication in time maps to convolution in frequency: $\mathcal{F}\{x(t) h(t)\} = \frac{1}{2\pi}(\hat{x} * \hat{h})(\omega)$, and vice versa. It is detected when modulating an RF carrier fails to shift signal spectra. It is corrected by remembering: multiplication in one domain corresponds to convolution in the other. Mini-example: multiplying a carrier by an audio pulse broadens the spectrum via spectral convolution, not algebraic multiplication.

Error 4: overlooking Hermitian symmetry for real physical observables

It seems reasonable to assume that the real and imaginary parts of $\hat{f}(\omega)$ are mutually independent functions for experimental signals. It fails because any strictly real time-domain observable $f(t) \in \mathbb{R}$ forces its Fourier transform to satisfy Hermitian symmetry $\hat{f}(-\omega) = \hat{f}^*(\omega)$, meaning its real part is even $\text{Re}[\hat{f}(-\omega)] = \text{Re}[\hat{f}(\omega)]$ and its imaginary part is odd $\text{Im}[\hat{f}(-\omega)] = -\text{Im}[\hat{f}(\omega)]$. It is detected if the magnitude spectrum fails to be symmetric: $|\hat{f}(-\omega)| \neq |\hat{f}(\omega)|$. It is corrected by enforcing symmetric magnitude bounds around $\omega = 0$. Mini-example: for negative frequencies, the amplitude spectrum of a real signal mirrors positive frequencies identically.

Error 5: treating negative frequencies as unphysical artifacts

It seems reasonable to discard $\omega < 0$ as meaningless because physical frequency meters report positive numbers. It fails because in complex exponential notation, real sinusoids $\cos(\omega_0 t) = \frac{e^{i\omega_0 t} + e^{-i\omega_0 t}}{2}$ require two counter-rotating phasors at $+\omega_0$ and $-\omega_0$. It is detected when Parseval energy integrals or convolutions omit negative frequency ranges, halving computed energy. It is corrected by integrating across the complete domain $(-\infty, \infty)$ so imaginary components cancel out. Mini-example: the transform of $\cos(\omega_0 t)$ contains two distinct delta peaks: $\pi[\delta(\omega - \omega_0) + \delta(\omega + \omega_0)]$.

Quick detection rule

If your inverse transform inflates signal amplitude by $2\pi$, if you multiply time signals instead of convolving their spectra, or if your real signal transform violates $|\hat{f}(\omega)| = |\hat{f}(-\omega)|$, your Fourier transform operations contain errors.

Physical applications

1. Dispersion and pulse broadening in ultrashort laser physics

It appears in femtosecond laser physics and ultrafast nonlinear optics when propagating ultrashort optical pulses through prisms or transmission fibers. It is calculated by mapping the pulse to the frequency domain $\hat{E}(\omega)$, imparting the medium's dispersive phase factor $e^{i \beta(\omega) z}$, and applying the inverse Fourier transform. The result accurately predicts temporal pulse broadening and chirp development (frequency modulation where different spectral colors travel at different speeds). The dangerous error is assuming that pulse temporal duration remains invariant by neglecting group velocity dispersion (GVD). As a final check, the student must explain how a diffraction grating pair introduces negative anomalous dispersion to recompress the pulse back to its transform-limited duration (the Nobel-winning chirped pulse amplification technique).

2. Fourier-Transform Infrared (FTIR) spectroscopy

It appears in physical chemistry, astrophysics, and materials science to characterize molecular vibrational bonds and chemical compositions. It is calculated by measuring the optical interferogram generated by a Michelson interferometer as a mirror translates, followed by digital Fast Fourier Transformation (FFT) of the recorded interference data. The result transforms raw spatial interferograms into continuous high-resolution infrared absorption spectra with superior signal-to-noise ratio (Fellgett's multiplex advantage). The dangerous error is failing to apply apodization windows to finite mirror scan lengths, which creates artificial sinc sidelobes flanking true absorption peaks. As a final check, the student must relate the physical displacement range of the translation stage to the achievable wavenumber resolution ($\text{cm}^{-1}$).

3. Heisenberg's uncertainty principle in quantum mechanics

It appears when representing the state of a free quantum particle in position space $\psi(x)$ versus momentum space $\phi(p)$. It is calculated by recognizing that $\phi(p)$ is precisely the spatial Fourier transform of $\psi(x)$: $\phi(p) = \frac{1}{\sqrt{2\pi\hbar}} \int \psi(x) e^{-i p x/\hbar} dx$. The result shows that the mathematical Fourier bandwidth inequality $\Delta x \cdot \Delta k \ge 1/2$ transforms, via De Broglie's relation $p = \hbar k$, into Heisenberg's fundamental uncertainty relation $\Delta x \cdot \Delta p \ge \frac{\hbar}{2}$. The dangerous error is interpreting the uncertainty principle as an experimental measurement artifact rather than an intrinsic mathematical property of Fourier wave packet pairs. As a final check, the student must prove that Gaussian wavefunctions uniquely minimize the uncertainty product.

4. Fraunhofer optical diffraction and X-ray crystallography

It appears when analyzing far-field diffraction patterns produced by two-dimensional apertures or atomic crystal lattices illuminated by X-rays. It is calculated by proving that the distant diffracted electric field $E(u, v)$ is proportional to the two-dimensional spatial Fourier transform of the aperture transmission function $t(x, y)$. The result enables structural determination of complex biomolecules and crystalline lattices (such as the historic discovery of DNA double helix structure) from observed diffraction spots. The dangerous error is overlooking the "phase problem": intensity sensors capture only irradiance $|\hat{t}(u,v)|^2$, losing the phase required to invert the transform back to real-space electron density directly. As a final check, the student must derive the Airy disk diffraction pattern $\frac{J_1(k R \theta)}{k R \theta}$ of a circular aperture of radius $R$.

5. Seismic signal processing and geophysical hydrocarbon exploration

It appears in exploration geophysics when imaging subsurface geological formations and identifying mineral or water reservoirs via reflected elastic waves. It is calculated by firing an acoustic pulse or vibroseis sweep, recording seismic reflection traces across geophone arrays, and filtering data in the Fourier frequency-wavenumber ($f-k$) domain to filter out Rayleigh surface waves. The result yields migrated seismic sections mapping rock strata to their true spatial depths. The dangerous error is temporal trace sampling with inadequate time intervals, causing spectral aliasing that folds high-frequency acoustic noise onto subterranean reflections. As a final check, the student must invoke the Nyquist-Shannon sampling theorem to compute the minimum required temporal sampling rate to eliminate aliasing artifacts.

Practice

Level 1: recognition

Using the time-frequency reciprocity relation $\Delta t \cdot \Delta \omega \ge 1/2$, identify which of the following signals requires the largest bandwidth in a communications channel: a) A rectangular pulse of duration $1\text{ ms}$. b) A rectangular pulse of duration $1\text{ ns}$. c) A femtosecond laser pulse of duration $10\text{ fs}$.

Exercise: Order the three pulses from narrowest to broadest spectral bandwidth and justify your ordering.

Level 2: direct application

Compute the Fourier transform of the single-sided decaying exponential signal $f(t) = e^{-a t} u(t)$, where $a > 0$ and $u(t)$ is the Heaviside step function.

Exercise: Evaluate the integral $\int_0^\infty e^{-(a + i\omega)t} dt$, obtain the complex result $\hat{f}(\omega) = \frac{1}{a + i\omega}$, and calculate its magnitude spectrum $|\hat{f}(\omega)|$ and energy power spectrum $|\hat{f}(\omega)|^2$.

Level 3: choosing the tool

A first-order low-pass filter with impulse response $h(t) = \frac{1}{R C} e^{-t/(R C)} u(t)$ is driven by an input signal $x(t) = e^{-t / \tau} u(t)$.

Exercise: Explain why evaluating the output $y(t)$ via Fourier transform using the convolution theorem $\hat{y}(\omega) = \hat{h}(\omega) \hat{x}(\omega)$ followed by partial fractions is significantly more direct than evaluating the time-domain convolution integral $(h * x)(t)$.

Level 4: physical interpretation

Apply Parseval's identity $\int_{-\infty}^\infty |f(t)|^2 dt = \frac{1}{2\pi} \int_{-\infty}^\infty |\hat{f}(\omega)|^2 d\omega$ to the decaying exponential $f(t) = e^{-a t} u(t)$ whose transform is $\hat{f}(\omega) = \frac{1}{a + i\omega}$.

Exercise: Compute the time-domain integral $\int_0^\infty e^{-2at} dt = \frac{1}{2a}$ and prove that the spectral integral $\frac{1}{2\pi} \int_{-\infty}^\infty \frac{d\omega}{a^2 + \omega^2}$ produces the exact same result $\frac{1}{2a}$, confirming strict energy conservation across domains.

Level 5: intentional error

A student attempts to compute the Fourier transform of the constant signal $f(t) = 1$ for all $t$ by evaluating: $\hat{f}(\omega) = \int_{-\infty}^\infty 1 \cdot e^{-i\omega t} dt = \left[ \frac{e^{-i\omega t}}{-i\omega} \right]_{-\infty}^\infty$. The student concludes that the transform does not exist because the exponential oscillates infinitely without converging.

Exercise: Explain why the constant function is not integrable in the classical Riemann sense but admits a rigorous Fourier transform in the space of tempered distributions, and prove that its transform is strictly $2\pi \delta(\omega)$, where $\delta$ is the Dirac delta distribution.