The Convolution Theorem

Convolution is the operation $(f*g)(x)=\int f(y)\,g(x-y)\,\mathrm{d}y$ that blends two functions by sliding one across the other and accumulating their overlap. It is the mathematical signature of a system that responds linearly and translation-invariantly to an input: whatever the machine — a lens, a filter, a differential operator, a channel that adds two random variables — the output is the input convolved with the machine's response to a single spike. That spike is the Dirac delta, and its response is the impulse response, or, for differential equations, the Green's function. Convolution is therefore the universal grammar of linear time-invariant physics. The trouble is that convolution is nonlocal: to know the output at one point you must integrate over all of the input, and doing that directly for a signal of $N$ samples costs $N^2$ operations. The convolution theorem is the escape. It says the Fourier transform, which resolves a function into pure frequencies, turns this awkward smearing integral into ordinary pointwise multiplication: $\widehat{f*g}=\sqrt{2\pi}\,\hat f\,\hat g$ . Transform, multiply frequency-by-frequency, transform back — a global integral collapses into a local product. The reason is structural: the Fourier transform diagonalizes translations — every mode $e^{ikx}$ is an eigenfunction of every shift — and convolution is built entirely from shifts, so in the basis that diagonalizes shifts it must act as multiplication by a number, the transform of the kernel. This one identity is the engine behind transform methods for differential equations, Green's functions, digital filtering, the characteristic-function proof of the central limit theorem, optical image formation, and the matched-filter searches that pulled gravitational waves out of the noise at LIGO.

Key takeaways

The convolution theorem states that the Fourier transform turns convolution into pointwise multiplication: $\widehat{f*g}=\sqrt{2\pi}\,\hat f\,\hat g$ in the symmetric convention.
Dually, the transform of a product is the convolution of the transforms, $\widehat{fg}=\tfrac{1}{\sqrt{2\pi}}\,\hat f*\hat g$ : multiplication and convolution swap roles under $\mathcal{F}$ .
It converts integro-differential equations into algebraic ones, so a forced linear ODE becomes $\hat y=\hat G\hat f$ and its solution is $y=G*f$ , a convolution with the Green's function.
The exact $2\pi$ factors depend entirely on the chosen Fourier convention; the structural statement — convolution becomes multiplication — is convention-independent.
Concrete payoffs: two Gaussians convolve to a Gaussian whose widths add in quadrature ( $\sigma=\sqrt{3^2+4^2}=5$ ), and the FFT computes convolutions in $O(N\log N)$ instead of $O(N^2)$ .
Rigorous statements hold for $L^1$ , extend to $L^2$ via Plancherel and density, and reach $\delta$ and other objects through the tempered-distribution framework.
Key idea: The Fourier transform converts the difficult, nonlocal operation of convolution into the trivial, local operation of pointwise multiplication, because it diagonalizes translations — and convolution is nothing but a continuous superposition of translations.

Where this lives today

The convolution theorem is quietly one of the most-executed identities in scientific computing. Gravitational-wave astronomy is a spectacular case: LIGO and Virgo detect black-hole mergers by matched filtering — cross-correlating the strain data against a bank of hundreds of thousands of waveform templates — and because correlation is convolution with a reflected kernel, each match is computed as a frequency-domain multiplication via the FFT, weighted by the detector noise spectrum. Without the theorem the searches would be computationally impossible. In imaging, every recorded picture is the true scene convolved with the instrument's point-spread function; deconvolution — dividing by the optical transfer function in frequency space, regularized against noise and its zeros — is how Hubble's early spherical aberration was fixed in software and how radio interferometry (where the "dirty image" is the sky convolved with the array beam) recovers sharp structure; the convolutional neural networks behind image recognition are named for the same operation, applying learned kernels in the Fourier domain for speed. The open frontier of blind deconvolution — untangling an unknown signal from an unknown blur at once — is an active inverse problem the theorem frames but does not by itself solve.

The physical picture

Think of $g$ as a stencil and $f$ as a sharp picture. Convolution drags the stencil across the picture and, at each position, records the overlap — replacing every point of $f$ by a weighted average of its neighbours dictated by $g$ . Where $g$ is a narrow spike almost nothing changes; where $g$ is broad the picture is blurred. The cleanest case to hold in mind is a box smearing a box: slide one unit-width rectangle across another and the overlap grows linearly, peaks when they align, and shrinks again — the output is a triangle, the fingerprint of the operation. Every convolution is a smoother version of the same sliding-overlap story.

Figure

Convolution as sliding overlap: dragging one unit-width box g across another box f and recording the overlap area at each shift produces a triangle of base 2 and peak height 1 — the peak occurs where the boxes coincide and the overlap equals ∫1·1 over width 1 = 1

Now watch the same operation in frequency space. Translating a function by $a$ multiplies each Fourier mode by the phase $e^{-ika}$ , so a mode's amplitude is untouched in size — only rotated. Convolution is a continuous sum of translations weighted by $g$ , so mode $k$ of $f$ is multiplied by $\int g(a)e^{-ika}\,\mathrm{d}a=\hat g(k)$ : a single number per frequency. Smearing in $x$ is nothing but attenuating high frequencies in $k$ . The most useful concrete instance is the Gaussian, whose transform is again a Gaussian: convolving two Gaussians multiplies two Gaussians in frequency, and the variances simply add — the reason independent errors combine in quadrature.

Figure

Domain swap made numerical: convolving a Gaussian of width σ=3 with one of width σ=4 (left, x-space) corresponds to multiplying their transforms (right, k-space), where the exponents add — e^(−9k²/2)·e^(−16k²/2)=e^(−25k²/2) — so the result is a Gaussian of width σ=√(3²+4²)=5, wider in x and correspondingly narrower in k

Read backwards, the theorem delivers a second lesson. A product in $x$ -space scrambles frequencies together: $\widehat{fg}$ blends every pair with $k=k_1+k_2$ , which is the convolution $\hat f*\hat g$ in frequency. Multiplying a signal by a carrier — modulation — therefore shifts and spreads its spectrum, the "frequency mixing" behind every radio.

The core equations

Fix a convention first, because every numerical factor below depends on it. We use the symmetric transform, with the same $1/\sqrt{2\pi}$ on the forward and inverse. Convolution itself carries no normalizing constant.

(f*g)(x)=\int_{-\infty}^{\infty} f(y)\,g(x-y)\,\mathrm{d}y \qquad (1)

where: $f,g$ — functions on $\mathbb{R}$ , typically in $L^1$ or $L^2$ ; $y$ — the integration (shift) variable; $f*g$ — the convolution, itself a function of $x$ . To get the output at $x$ : flip $g$ , slide it to $x$ , integrate its product with $f$ .

\hat f(k)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty} f(x)\,e^{-ikx}\,\mathrm{d}x,\qquad f(x)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty} \hat f(k)\,e^{ikx}\,\mathrm{d}k \qquad (2)

where: $\hat f=\mathcal{F}[f]$ — the Fourier transform of $f$ ; $k$ — angular frequency (wavenumber). The pair differs only by the sign in the exponent, which is what makes the forward and dual theorems mirror images. The central identity: the transform of a convolution factorizes into the product of transforms, times the one surviving convention constant.

\widehat{f*g}(k)=\sqrt{2\pi}\,\hat f(k)\,\hat g(k) \qquad (3)

where: $\sqrt{2\pi}$ — the convention-dependent constant ( $1$ in the ordinary-frequency convention). Reading it: every output frequency is the same frequency of $f$ scaled by the number $\hat g(k)$ — convolution acts diagonally in the Fourier basis. Applying the transform twice, or equally applying it to a product, gives the dual statement in which the roles of multiplication and convolution are exchanged.

\widehat{f\,g}(k)=\frac{1}{\sqrt{2\pi}}\,(\hat f*\hat g)(k) \qquad (4)

where: $fg$ — the pointwise product in $x$ -space; $\hat f*\hat g$ — convolution in $k$ -space. Multiplying in one domain is convolving in the other, with the constant inverted to $1/\sqrt{2\pi}$ . The Dirac delta is the identity element, which pins the convention and makes convolution an algebra with a unit.

f*\delta=f,\qquad \widehat{\,\delta\,}=\frac{1}{\sqrt{2\pi}},\qquad \frac{\mathrm{d}}{\mathrm{d}x}(f*g)=f'*g=f*g' \qquad (5)

where: $\delta$ — the Dirac delta; $f'$ — a derivative. Reading it: convolving with $\delta$ does nothing, its flat transform makes equation (3) return $\hat f$ , and a derivative may sit on either factor — so a smooth kernel transfers its smoothness to the result. Finally, the correlation cousin of the theorem yields Parseval–Plancherel, the statement that $\mathcal{F}$ preserves inner products and energy.

\int_{-\infty}^{\infty}\overline{f(x)}\,g(x)\,\mathrm{d}x=\int_{-\infty}^{\infty}\overline{\hat f(k)}\,\hat g(k)\,\mathrm{d}k,\qquad \int|f|^2\,\mathrm{d}x=\int|\hat f|^2\,\mathrm{d}k \qquad (6)

where: $\overline{\,\cdot\,}$ — complex conjugate. Setting $g=f$ gives energy conservation under $\mathcal{F}$ .

Derivation AProof of the convolution theorem for integrable functions

The goal is to prove $\widehat{f*g}=\sqrt{2\pi}\,\hat f\,\hat g$ rigorously for $f,g\in L^1(\mathbb{R})$ , not just to shuffle symbols: we must first show that $f*g$ is itself integrable so that its transform even exists, then justify exchanging the order of a double integral, and only then read off the factorization. The one nontrivial analytic step is Fubini's theorem, licensed precisely by the integrability bound we establish first — so the proof is really "bound, then swap, then shift."

Step 1. Show the convolution lives in $L^1$ with $\|f*g\|_1\le\|f\|_1\,\|g\|_1$ . By Tonelli's theorem for the nonnegative integrand $|f(y)||g(x-y)|$ , $\int|f*g|\,\mathrm{d}x\le\int\!\!\int|f(y)|\,|g(x-y)|\,\mathrm{d}y\,\mathrm{d}x=\|f\|_1\,\|g\|_1<\infty$ . Why: this both guarantees $f*g$ is a bona fide integrable function and supplies the absolute convergence that will license Fubini in the next step.

Step 2. Insert the definition into the transform, $\widehat{f*g}(k)=\frac{1}{\sqrt{2\pi}}\int e^{-ikx}\big(\int f(y)\,g(x-y)\,\mathrm{d}y\big)\mathrm{d}x$ . Why: this is a double integral over $(x,y)$ with integrand of absolute value $|f(y)||g(x-y)|$ , whose total mass is finite by Step 1, so it satisfies Fubini's hypothesis exactly.

Step 3. Exchange the order and factor the phase: $\widehat{f*g}(k)=\frac{1}{\sqrt{2\pi}}\int f(y)\big(\int g(x-y)\,e^{-ikx}\,\mathrm{d}x\big)\mathrm{d}y$ . Why: with absolute convergence secured, Fubini permits integrating over $x$ first while $y$ is held fixed — the step that separates the two functions.

Step 4. Evaluate the inner integral by the shift substitution $u=x-y$ : $\int g(x-y)e^{-ikx}\,\mathrm{d}x=e^{-iky}\int g(u)e^{-iku}\,\mathrm{d}u=\sqrt{2\pi}\,e^{-iky}\hat g(k)$ . Why: the translation pulls out a pure phase $e^{-iky}$ and leaves the transform of $g$ intact — this is where "convolution is built from shifts" becomes a multiplicative phase.

Step 5. Collect: $\widehat{f*g}(k)=\sqrt{2\pi}\,\hat g(k)\cdot\frac{1}{\sqrt{2\pi}}\int f(y)e^{-iky}\,\mathrm{d}y=\sqrt{2\pi}\,\hat f(k)\,\hat g(k)$ . Why: the leftover $y$ -integral is just $\sqrt{2\pi}\,\hat f(k)$ , so the two transforms multiply and exactly one factor of $\sqrt{2\pi}$ survives — the convention constant of equation (3). *As a self-consistency check: equation (3) gives $\widehat{f*\delta}=\sqrt{2\pi}\,\hat f\,\hat\delta=\sqrt{2\pi}\,\hat f\cdot\tfrac{1}{\sqrt{2\pi}}=\hat f$ , so $f*\delta=f$ — and this fixes $\hat\delta=1/\sqrt{2\pi}$ unambiguously.*

Go deeper · Derivation B · 1 minSolving a forced damped oscillator by transform

Derivation BSolving a forced damped oscillator by transform

The aim is to show the theorem's headline payoff: a constant-coefficient linear ODE with an arbitrary driving force is converted into an algebraic equation by $\mathcal{F}$ , solved by division, and inverted — by the theorem read backwards — into a convolution of the force with a Green's function. We take the damped oscillator $y''+2\gamma y'+\omega_0^2 y=f(t)$ on the whole line, transform it, solve the algebra, and identify the impulse response, checking causal and undamped limits at the end.

Step 1. Transform the equation, using that each $\mathrm{d}/\mathrm{d}t$ becomes multiplication by $ik$ : $(-k^2+2i\gamma k+\omega_0^2)\,\hat y(k)=\hat f(k)$ . Why: differentiation is diagonal in the Fourier basis, so a differential equation in $t$ becomes an ordinary algebraic equation in $k$ — no calculus left to do.

Step 2. Divide to isolate $\hat y$ , defining the transfer function $\hat G$ : $\hat y(k)=\hat G(k)\,\hat f(k)$ with $\hat G(k)=\dfrac{1}{\sqrt{2\pi}}\,\dfrac{1}{\omega_0^2-k^2+2i\gamma k}$ . Why: the $1/\sqrt{2\pi}$ is inserted deliberately so that $\hat y=\sqrt{2\pi}\,\hat G\,\hat f$ has exactly the shape of the right-hand side of equation (3), ready to invert as a convolution.

Step 3. Invert with the convolution theorem: $y=\mathcal{F}^{-1}[\sqrt{2\pi}\,\hat G\,\hat f]=G*f$ . Why: a product of transforms inverts to a convolution, so the solution is the source $f$ smeared by the impulse response $G$ — every forced linear system answers this way.

Step 4. Identify $G$ by inverting $\hat G$ with a contour closed to enforce causality: $G(t)=\theta(t)\,e^{-\gamma t}\,\dfrac{\sin(\omega t)}{\omega},\ \ \omega=\sqrt{\omega_0^2-\gamma^2}$ . Why: the poles sit at $k=i\gamma\pm\omega$ ; closing in the half-plane that vanishes for $t>0$ selects the retarded (causal) response, and $\theta(t)$ enforces "no output before the kick." *Two limiting checks confirm the result. With $f=\delta$ , $y=G*\delta=G$ — the response to a unit impulse is $G$ itself, as it must be. As $\gamma\to0$ , $G\to\theta(t)\sin(\omega_0 t)/\omega_0$ , recovering the undamped Green's function; and for $\omega_0=5,\ \gamma=3$ the frequency is $\omega=\sqrt{25-9}=4$ , an underdamped ring.*

Go deeper · Derivation C · 1 minThe dual product theorem from inversion symmetry

Derivation CThe dual product theorem from inversion symmetry

The goal is to derive the mirror statement $\widehat{fg}=\tfrac{1}{\sqrt{2\pi}}\,\hat f*\hat g$ — that multiplying in $x$ -space is convolving in $k$ -space — directly, using only the inversion formula and the plane-wave orthogonality relation $\int e^{iqx}\mathrm{d}x=2\pi\,\delta(q)$ . This exposes the forward and dual theorems as one fact seen through the near-symmetry of $\mathcal{F}$ and $\mathcal{F}^{-1}$ , and explains why modulation spreads spectra.

Step 1. Write both factors by inversion, $f(x)=\frac{1}{\sqrt{2\pi}}\int \hat f(k_1)e^{ik_1x}\mathrm{d}k_1$ and $g(x)=\frac{1}{\sqrt{2\pi}}\int \hat g(k_2)e^{ik_2x}\mathrm{d}k_2$ . Why: expressing each function as a superposition of modes turns the product $fg$ into a double sum over mode pairs, which is where the convolution in $k$ will emerge.

Step 2. Transform the product, collecting all three exponentials: $\widehat{fg}(k)=\frac{1}{(2\pi)^{3/2}}\int\!\!\int\!\!\int \hat f(k_1)\hat g(k_2)\,e^{i(k_1+k_2-k)x}\,\mathrm{d}k_1\mathrm{d}k_2\mathrm{d}x$ . Why: the only $x$ -dependence now lives in a single plane-wave phase, so the $x$ -integral can be done first and universally.

Step 3. Do the $x$ -integral to get a delta, $\int e^{i(k_1+k_2-k)x}\mathrm{d}x=2\pi\,\delta(k_1+k_2-k)$ . Why: plane waves are orthogonal — a phase integrates to zero unless it is stationary, i.e. unless the frequencies add up to $k$ — enforcing momentum-like conservation $k=k_1+k_2$ .

Step 4. Collapse the delta over $k_2$ : $\widehat{fg}(k)=\frac{1}{\sqrt{2\pi}}\int \hat f(k_1)\hat g(k-k_1)\,\mathrm{d}k_1=\frac{1}{\sqrt{2\pi}}(\hat f*\hat g)(k)$ . Why: the surviving integral is precisely a convolution in frequency — every output frequency is built from all pairs of input frequencies that sum to it, which is exactly why modulating a signal shifts and broadens its spectrum. *Applying the forward theorem (3) and this dual (4) in succession multiplies $\sqrt{2\pi}$ by $1/\sqrt{2\pi}$ to give $1$ , returning the original function — the two constants are genuine inverses, confirming the convention is self-consistent.*

Limiting cases

How the theorem specializes across conventions and operands — the constant moves, the structure never does:

Case	Result	Meaning
Symmetric convention	$\widehat{f*g}=\sqrt{2\pi}\,\hat f\,\hat g$	With $\hat f(k)=\frac{1}{\sqrt{2\pi}}\int f e^{-ikx}\mathrm{d}x$ an explicit $\sqrt{2\pi}$ appears.
Ordinary-frequency convention	$\widehat{f*g}=\hat f\,\hat g$	With $\hat f(\nu)=\int f e^{-2\pi i\nu x}\mathrm{d}x$ the theorem is cleanest — no extra constant at all.
Convolution with $\delta$	$f*\delta=f,\ \ \hat\delta=\tfrac{1}{\sqrt{2\pi}}$	$\delta$ is the identity; consistency check $\widehat{f*\delta}=\sqrt{2\pi}\,\hat f\,\hat\delta=\hat f$ .
Two Gaussians	$\sigma^2=\sigma_1^2+\sigma_2^2$	Widths add in quadrature — the transform of a Gaussian is a Gaussian, and exponents add.
Correlation / Parseval limit	$\int\|f\|^2\mathrm{d}x=\int\|\hat f\|^2\mathrm{d}k$	Setting $g(x)=\overline{f(-x)}$ and evaluating at $0$ reproduces Plancherel.
Merely bounded, not integrable	needs $L^2$ or distributions	The defining integral may diverge; extend via Plancherel/density or the tempered framework.

Common trapThe 2π factors are physics — they are not.

The single most common error is treating the $2\pi$ and $\sqrt{2\pi}$ constants as though they carried physical content, then mixing conventions mid-calculation. They do not: whether the constant sits in the forward transform, the inverse, split symmetrically as $1/\sqrt{2\pi}$ each, or is banished entirely by using ordinary frequency $\nu=k/2\pi$ , is a bookkeeping choice with no observable consequence. What is invariant is the structure — convolution becomes multiplication. The discipline is non-negotiable: pick one convention at the top, write down where its constant lives, and never switch. A stray factor of $2\pi$ in a Green's function or filter gain almost always traces to a silent convention change, not to physics.

Worked examples

1 · Two Gaussians convolve to a Gaussian — widths add in quadrature (fully worked)

Let $f$ and $g$ be normalized Gaussian densities of standard deviations $\sigma_1=3$ and $\sigma_2=4$ . Find $f*g$ using the convolution theorem, and give the width, peak height, and full width at half maximum of the result.

Step 1. Transform each factor. The Fourier transform of a normalized Gaussian $\frac{1}{\sqrt{2\pi}\sigma}e^{-x^2/2\sigma^2}$ in the symmetric convention is $\hat f(k)=\frac{1}{\sqrt{2\pi}}e^{-\sigma_1^2k^2/2}$ , and likewise $\hat g(k)=\frac{1}{\sqrt{2\pi}}e^{-\sigma_2^2k^2/2}$ . A Gaussian transforms to a Gaussian, with width and reciprocal-width swapping roles.

Step 2. Multiply and apply the theorem. Equation (3) gives $\widehat{f*g}=\sqrt{2\pi}\,\hat f\,\hat g=\sqrt{2\pi}\cdot\frac{1}{2\pi}\,e^{-(\sigma_1^2+\sigma_2^2)k^2/2}=\frac{1}{\sqrt{2\pi}}\,e^{-\sigma^2k^2/2}$ with $\sigma^2=\sigma_1^2+\sigma_2^2$ . The exponents simply add — this is the whole mechanism.

Step 3. Invert. $\frac{1}{\sqrt{2\pi}}e^{-\sigma^2k^2/2}$ is the transform of the normalized Gaussian of variance $\sigma^2$ , so $f*g$ is that Gaussian. The width follows from $\sigma=\sqrt{\sigma_1^2+\sigma_2^2}=\sqrt{3^2+4^2}=\sqrt{25}=5$ — the variances $9$ and $16$ add to $25$ , never the widths $3$ and $4$ .

Step 4. Read off the shape. Peak height $=\frac{1}{\sqrt{2\pi}\,\sigma}=\frac{1}{5\sqrt{2\pi}}=\mathbf{0.0798}$ ; full width at half maximum $=2\sqrt{2\ln 2}\,\sigma=2.3548\times5=\mathbf{11.77}$ ; a direct numerical convolution confirms $(f*g)(0)=0.07979$ . The convolution of two Gaussians is a Gaussian with $\sigma=5$ — this quadrature rule is exactly why independent Gaussian errors combine as $\sqrt{\sigma_1^2+\sigma_2^2}$ .

2 · The sum of two independent uniforms is triangular (fully worked)

Let $X,Y$ be independent and uniform on $[0,1]$ , with box densities $f=g=\mathbf{1}_{[0,1]}$ . Find the density of $X+Y$ and its variance.

Step 1. The density of a sum of independent variables is the convolution of their densities: $p_{X+Y}=f*g$ . This is the probabilistic face of the theorem — in Fourier space it says characteristic functions multiply, $\varphi_{X+Y}=\varphi_X\,\varphi_Y$ .

Step 2. Convolve the two boxes directly: $(f*g)(z)=\int_0^1 \mathbf{1}_{[0,1]}(z-y)\,\mathrm{d}y$ is the length of the overlap of $[0,1]$ with $[z-1,z]$ . This gives $z$ for $0\le z\le1$ and $2-z$ for $1\le z\le2$ , zero otherwise — a triangle on $[0,2]$ with peak $1$ at $z=1$ .

Step 3. Check normalization and moments. Area $=\frac{1}{2}\cdot 2\cdot 1=1$ (a probability density). Mean $=1$ by symmetry, and variance adds: $\mathrm{Var}(X+Y)=\tfrac{1}{12}+\tfrac{1}{12}=\mathbf{\tfrac{1}{6}}$ . Convolving two flat distributions already produces the peaked triangular shape — one more convolution smooths it toward a Gaussian, which is the central limit theorem in action.

3 · Fast convolution: why the FFT changed computing

A digital filter must convolve two signals of $N=10^6$ samples each. Compare the cost of the direct sum with the transform-multiply-invert route.

Step 1. Direct convolution evaluates $\sum_y f(y)g(x-y)$ for every output $x$ : about $N$ multiply-adds per output, times $N$ outputs, so $\sim N^2=10^{12}$ operations — minutes to hours even on fast hardware.

Step 2. The theorem replaces this with two forward FFTs, a pointwise multiply, and one inverse FFT. Each FFT costs $\sim 5N\log_2 N$ , the multiply $\sim N$ , so for $N=10^6$ the total is $\approx 10^{8}$ operations versus $10^{12}$ .

Step 3. The speedup is $\dfrac{N^2}{5N\log_2 N}=\dfrac{N}{5\log_2 N}\approx\mathbf{10^{4}}$ at $N=10^6$ : a four-order-of-magnitude acceleration, growing without bound as $N$ increases. This is why essentially every large convolution in the world — audio, imaging, radio astronomy, machine learning — is computed in the frequency domain via the FFT rather than directly.

Physics of the portrait

The convolution theorem sits at a crossroads of three centuries: an integral written down by eighteenth-century analysts, a transform founded by Fourier to solve heat flow, and an algebra of operations that only the twentieth century recognized as one structure.

Origins · 1 minconvolution theorem and how this was discovered

convolution theorem

the folding integral (f∗g)(x)=∫f(y)g(x−y)dy and its Fourier diagonalization · linear systems and integral transforms

The convolution integral is older than the theorem that tames it. It surfaces in d'Alembert's work on vibrating strings in 1754 and, more consequentially, in Laplace's probability calculus of the 1780s, where the density of a sum of independent quantities is a "composition" of their densities — convolution in all but name. The transform that would diagonalize it arrived with Joseph Fourier's 1822 Théorie analytique de la chaleur, whose decomposition of heat flow into sinusoids is precisely the basis in which every translation, and hence every convolution, becomes a multiplication. For a century the pieces sat apart: German analysts spoke of the Faltung — the "folding" of one function onto another — and Vito Volterra, in his theory of integral equations around 1910, treated convolution as the product of a commutative algebra. Harmonic analysis fused them: Norbert Wiener's Tauberian theorems of the early 1930s placed convolution at the centre of the subject, and the rigorous $L^1$ , $L^2$ , and distributional statements were settled by Plancherel, Bochner, and finally Laurent Schwartz's tempered distributions, which gave $f*\delta=f$ and $\hat\delta=\mathrm{const}$ a precise meaning. The decisive practical turn came in 1965, when Cooley and Tukey rediscovered the fast Fourier transform: it made "transform, multiply, invert" cheaper than direct convolution and turned a theorem of pure analysis into the workhorse of signal processing, imaging, and machine learning.

History

1754

— d'Alembert writes an early convolution-type integral in work on vibrating strings

1782

— Laplace composes probability densities of sums, convolution in all but name

1822

— Fourier's heat treatise founds the transform that diagonalizes convolution

1910

— Volterra casts convolution as the product in his algebra of integral equations

1933

— Wiener's Tauberian theorems place convolution at the centre of harmonic analysis

1965

— Cooley and Tukey's FFT makes transform-domain convolution the everyday method

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. on Fourier methods (the symmetric convention used here) · Arfken & Weber, Mathematical Methods for Physicists, Fourier-transform chapter (Green's functions and contour inversion) · Hassani, Mathematical Physics, integral-transforms chapters · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, §13 (integral transforms) · Reed & Simon, Methods of Modern Mathematical Physics II (the $L^1$ / $L^2$ /distribution statements) · Bracewell, The Fourier Transform and Its Applications (the engineering and imaging face). Key numerical anchors used throughout: the norm bound $\|f*g\|_1\le\|f\|_1\|g\|_1$ ; the underdamped Green's function $G(t)=\theta(t)e^{-\gamma t}\sin(\omega t)/\omega$ with $\omega=\sqrt{\omega_0^2-\gamma^2}$ , giving $\omega=4$ for $\omega_0=5,\gamma=3$ ; the Gaussian quadrature $\sigma=\sqrt{3^2+4^2}=5$ and peak $1/(5\sqrt{2\pi})=0.0798$ ; the uniform-uniform triangular density with variance $1/6$ ; and the FFT cost $\sim 5N\log_2 N$ versus $N^2$ , speedup $\approx10^4$ at $N=10^6$ .

Prerequisites. The Fourier Transform · Properties of the Fourier Transform · The Dirac Delta Function. Leads to. Solving Differential Equations by Transform Methods · Green's Functions via Integral Transforms · Tempered Distributions and Fourier Transforms · Parseval's and Plancherel's Theorems.

Ask the tutor: Ask me to walk through inverting a specific Green's function by contour integration, or to show how the characteristic-function version of this theorem drives the central limit theorem.

Does the convolution theorem depend on which Fourier convention I use?

The structure does not; only the numerical constant does. In the symmetric convention the theorem reads $\widehat{f*g}=\sqrt{2\pi}\,\hat f\,\hat g$ , while in the ordinary-frequency convention $\hat f(\nu)=\int f\,e^{-2\pi i\nu x}\mathrm{d}x$ it is simply $\widehat{f*g}=\hat f\,\hat g$ with no constant at all. Physicists often use the "angular frequency, constant on the inverse" convention, which puts a bare $\hat f\hat g$ on one side too. The safe practice is to fix one convention at the start, note where its $2\pi$ lives, and never mix — a stray $\sqrt{2\pi}$ is almost always a silent convention switch, not real physics.

Why does convolution become multiplication, intuitively?

Because the Fourier transform is the basis that diagonalizes translation. Shifting a function by $a$ multiplies its mode $e^{ikx}$ by the number $e^{-ika}$ — the modes are eigenvectors of every shift operator. Convolution with $g$ is nothing but a weighted continuous sum of shifts, the weights being $g$ itself, so in the shift-diagonalizing basis it can only multiply each mode by a number, namely $\hat g(k)$ . It is the exact analogue of how a matrix that commutes with a diagonalizable operator is itself diagonal in that operator's eigenbasis: convolutions all commute with translations, so they are all "diagonal" in the Fourier basis.

What is the Green's function, and how does the theorem produce it?

The Green's function $G$ is the response of a linear differential operator to a unit impulse $\delta$ — its impulse response. For a constant-coefficient operator $L$ , transforming $Ly=f$ gives $P(ik)\,\hat y=\hat f$ , where $P$ is the operator's symbol; solving gives $\hat y=\hat f/P(ik)$ , a product of transforms. Reading the convolution theorem backwards, a product of transforms inverts to a convolution, so $y=G*f$ with $\hat G\propto 1/P(ik)$ . The physics is that any linear system's output is its impulse response smeared over the input — which is why $G$ , computed once, solves the equation for every forcing $f$ .

Can I convolve functions that are not integrable, like sin or a constant?

Not within the elementary $L^1$ theory, where the defining integral can diverge and the transforms may not exist as functions. Two extensions rescue most cases. First, $L^2$ : for square-integrable functions the transform is defined via Plancherel and a density argument, and the theorem persists. Second, tempered distributions: objects like $\delta$ , constants, and $e^{ik_0x}$ have Fourier transforms as distributions, and convolution of a tempered distribution with a Schwartz function is well defined, with the theorem intact. What you cannot do is blindly convolve two arbitrary distributions — $\delta'*\delta'$ and similar products need care, because multiplication of distributions (the dual side) is not always defined.

How is the FFT related to the convolution theorem?

The FFT is the algorithm that makes the theorem cheap. Direct convolution of two length- $N$ signals costs $O(N^2)$ ; the theorem says instead to Fourier-transform both, multiply pointwise, and inverse-transform. The discrete Fourier transform naively also costs $O(N^2)$ — but the fast Fourier transform of Cooley and Tukey computes it in $O(N\log N)$ by recursively splitting the sum, giving $O(N\log N)$ convolution overall (about $10^4\times$ faster at $N=10^6$ ). One caveat: the DFT convolution is circular, so real applications zero-pad to avoid wrap-around before transforming.

What is the difference between convolution and correlation?

They differ only by a reflection of one argument. Convolution flips $g$ before sliding, $(f*g)(x)=\int f(y)g(x-y)\mathrm{d}y$ ; cross-correlation does not, $(f\star g)(x)=\int \overline{f(y)}\,g(x+y)\mathrm{d}y$ . In Fourier space convolution multiplies the transforms, while correlation multiplies with a conjugate, $\widehat{f\star g}\propto\overline{\hat f}\,\hat g$ — which is what makes correlation the natural tool for matching a template to data (as in LIGO's searches) rather than blurring. Convolution is commutative and associative; correlation is neither, since the reflection breaks the symmetry.