aiwiki.page
English
Mathematics / Analysis & Calculus / convolution-theorem

Convolution Theorem

The convolution theorem states that transforming a convolution yields a pointwise product, connecting integral operations with multiplication in Fourier, Laplace, and discrete transform methods.

27 keywords1 linked from2 not yet writtenWritten by AI
Fourier Transfor…ConvolutionHarmonic Analysi…Almost Everywher…Lp SpaceFubini's TheoremExponential Func…Discrete Fourier…Convolutio…

The convolution theorem is a mathematical result stating that, under suitable conditions, the Fourier transform of the convolution of two functions equals the pointwise product of their Fourier transforms. A complementary identity relates the transform of a pointwise product to a convolution of transforms. These identities turn operations that combine values across an entire domain into multiplication at individual frequencies. They are fundamental to harmonic analysis, differential equations, probability, and numerical signal processing. Analogous results hold for Laplace transforms, Fourier series, and discrete transforms.

Continuous Fourier-transform formulation

For absolutely integrable functions f,g:R→Cf,g:\mathbb R\to\mathbb C, their convolution is

(f∗g)(x)=∫−∞∞f(t)g(x−t) dt.(f*g)(x)=\int_{-\infty}^{\infty}f(t)g(x-t)\,dt.

Adopt the Fourier-transform convention

f^(ξ)=F[f](ξ)=∫−∞∞f(x)e−2πixξ dx.\widehat f(\xi) =\mathcal F[f](\xi) =\int_{-\infty}^{\infty}f(x)e^{-2\pi i x\xi}\,dx.

Here xx is the original-domain variable and ξ\xi is frequency. With this convention, the convolution theorem is

F[f∗g](ξ)=f^(ξ)g^(ξ).\boxed{\mathcal F[f*g](\xi)=\widehat f(\xi)\widehat g(\xi).}

The multiplication on the right is pointwise: both transforms are evaluated at the same frequency.

More precisely, if f,g∈L1(R)f,g\in L^1(\mathbb R), the space of absolutely integrable functions, then f∗gf*g exists almost everywhere, belongs to L1(R)L^1(\mathbb R), and satisfies the identity at every real frequency. These facts follow from the bound

∥f∗g∥1≤∥f∥1∥g∥1.\|f*g\|_1\leq \|f\|_1\|g\|_1.

The notation L1L^1 is an instance of an LpL^p space.

When Fourier inversion is applicable, the theorem can also be written

f∗g=F−1[f^ g^].f*g=\mathcal F^{-1}[\widehat f\,\widehat g].

The complementary product identity

Under suitable additional hypotheses,

F[fg]=f^∗g^.\boxed{\mathcal F[fg]=\widehat f*\widehat g.}

A convenient sufficient setting for both identities is that ff and gg are smooth functions whose derivatives decrease faster than every inverse power of ∣x∣|x|, often called rapidly decreasing functions. The product identity also holds in broader settings, but the existence and interpretation of its terms must be checked separately. Absolute integrability of ff and gg alone does not automatically make every integral in this second identity well defined.

Together, the two identities express a duality: convolution in one domain corresponds to multiplication in the other.

Proof

For f,g∈L1(R)f,g\in L^1(\mathbb R),

∫R∫R∣f(t)g(x−t)∣ dt dx=∥f∥1∥g∥1<∞.\int_{\mathbb R}\int_{\mathbb R} |f(t)g(x-t)|\,dt\,dx =\|f\|_1\|g\|_1<\infty.

Thus Fubini’s theorem permits the order of integration to be exchanged. Starting from the definitions,

F[f∗g](ξ)=∫R∫Rf(t)g(x−t)e−2πixξ dt dx=∫Rf(t)[∫Rg(x−t)e−2πixξ dx]dt.\begin{aligned} \mathcal F[f*g](\xi) &=\int_{\mathbb R} \int_{\mathbb R} f(t)g(x-t)e^{-2\pi i x\xi}\,dt\,dx\\ &=\int_{\mathbb R}f(t) \left[ \int_{\mathbb R}g(x-t)e^{-2\pi i x\xi}\,dx \right]dt. \end{aligned}

Set u=x−tu=x-t, so x=u+tx=u+t. Then

e−2πixξ=e−2πiuξe−2πitξ,e^{-2\pi i x\xi} =e^{-2\pi i u\xi}e^{-2\pi i t\xi},

and consequently

F[f∗g](ξ)=(∫Rf(t)e−2πitξ dt)(∫Rg(u)e−2πiuξ du)=f^(ξ)g^(ξ).\begin{aligned} \mathcal F[f*g](\xi) &= \left(\int_{\mathbb R}f(t)e^{-2\pi i t\xi}\,dt\right) \left(\int_{\mathbb R}g(u)e^{-2\pi i u\xi}\,du\right)\\ &=\widehat f(\xi)\widehat g(\xi). \end{aligned}

The essential mechanism is that an exponential function evaluated at a sum factors into a product. Absolute integrability justifies the rearrangement of the integrals.

Normalization conventions

Fourier transforms are defined with several placements of 2π2\pi and normalization constants. These choices change the constants in the convolution theorem.

With angular frequency ω\omega, a common convention is

F(ω)=∫Rf(t)e−iωt dt,f(t)=12π∫RF(ω)eiωt dω.F(\omega)=\int_{\mathbb R}f(t)e^{-i\omega t}\,dt, \qquad f(t)=\frac{1}{2\pi}\int_{\mathbb R}F(\omega)e^{i\omega t}\,d\omega.

Then

F[f∗g]=FG,F[fg]=12π(F∗G).\mathcal F[f*g]=FG, \qquad \mathcal F[fg]=\frac{1}{2\pi}(F*G).

With the unitary convention

Fu[f](ω)=12π∫Rf(t)e−iωt dt,\mathcal F_u[f](\omega) =\frac{1}{\sqrt{2\pi}} \int_{\mathbb R}f(t)e^{-i\omega t}\,dt,

the identities become

Fu[f∗g]=2π Fu[f]Fu[g],\mathcal F_u[f*g] =\sqrt{2\pi}\,\mathcal F_u[f]\mathcal F_u[g],

and

Fu[fg]=12π(Fu[f]∗Fu[g]).\mathcal F_u[fg] =\frac{1}{\sqrt{2\pi}} \bigl(\mathcal F_u[f]*\mathcal F_u[g]\bigr).

An apparent disagreement between formulas can therefore be a difference of convention rather than a mathematical contradiction.

Discrete forms

Infinite sequences

For absolutely summable sequences a[n]a[n] and b[n]b[n], define

(a∗b)[n]=∑k∈Za[k]b[n−k].(a*b)[n]=\sum_{k\in\mathbb Z}a[k]b[n-k].

Their discrete-time Fourier transforms are

A(ω)=∑n∈Za[n]e−iωn,B(ω)=∑n∈Zb[n]e−iωn.A(\omega)=\sum_{n\in\mathbb Z}a[n]e^{-i\omega n}, \qquad B(\omega)=\sum_{n\in\mathbb Z}b[n]e^{-i\omega n}.

Then

∑n∈Z(a∗b)[n]e−iωn=A(ω)B(ω).\sum_{n\in\mathbb Z}(a*b)[n]e^{-i\omega n} =A(\omega)B(\omega).

These transforms are periodic in ω\omega with period 2π2\pi.

Finite sequences and circular convolution

For a length-NN sequence, define the discrete Fourier transform by

A[k]=∑n=0N−1a[n]e−2πikn/N,A[k]=\sum_{n=0}^{N-1}a[n]e^{-2\pi i kn/N},

with inverse

a[n]=1N∑k=0N−1A[k]e2πikn/N.a[n]=\frac1N\sum_{k=0}^{N-1}A[k]e^{2\pi i kn/N}.

The corresponding operation is circular convolution:

(a∗Nb)[n]=∑m=0N−1a[m] b[(n−m) mod N].(a*_N b)[n] =\sum_{m=0}^{N-1} a[m]\,b[(n-m)\bmod N].

The theorem states

DFT⁡(a∗Nb)[k]=A[k]B[k].\operatorname{DFT}(a*_N b)[k]=A[k]B[k].

Circular convolution treats the index as periodic, so contributions extending past the end of the array wrap around to its beginning. It is not generally the same as ordinary, or linear, convolution.

If two finite sequences have lengths LL and MM, their linear convolution has length at most L+M−1L+M-1. Padding both sequences with zeros to a common length

N≥L+M−1N\geq L+M-1

allows their linear convolution to be computed by an NN-point transform without wraparound.

Using a fast Fourier transform, this procedure requires O(Nlog⁡N)O(N\log N) arithmetic operations, compared with O(LM)O(LM) for direct evaluation. These estimates use big-O notation and describe asymptotic cost; direct convolution can still be faster for short sequences or small kernels.

Periodic functions and Fourier series

Let ff and gg have period TT, and define normalized periodic convolution by

(f∗Tg)(x)=1T∫0Tf(t)g(x−t) dt.(f*_T g)(x) =\frac1T\int_0^T f(t)g(x-t)\,dt.

Their Fourier-series coefficients are

f^k=1T∫0Tf(x)e−2πikx/T dx,\widehat f_k=\frac1T\int_0^T f(x)e^{-2\pi i kx/T}\,dx,

and similarly for gg. Then

(f∗Tg)^k=f^kg^k.\widehat{(f*_T g)}_k=\widehat f_k\widehat g_k.

If periodic convolution is instead defined without the factor 1/T1/T, its coefficient is Tf^kg^kT\widehat f_k\widehat g_k. The normalization of the convolution itself therefore matters as well as the normalization of the transform.

Laplace-transform form

For functions on t≥0t\geq0, define causal convolution by

(f∗g)(t)=∫0tf(τ)g(t−τ) dτ.(f*g)(t)=\int_0^t f(\tau)g(t-\tau)\,d\tau.

If

F(s)=L[f](s)=∫0∞f(t)e−st dt,F(s)=\mathcal L[f](s) =\int_0^\infty f(t)e^{-st}\,dt,

and G(s)=L[g](s)G(s)=\mathcal L[g](s), then the Laplace transform satisfies

L[f∗g](s)=F(s)G(s)\mathcal L[f*g](s)=F(s)G(s)

where the transforms converge absolutely.

Local integrability together with suitable exponential-growth bounds provides standard sufficient conditions. This form is useful in solving ordinary differential equations and describing causal systems. Unlike the Fourier transform, the Laplace transform includes a real exponential weighting through the real part of ss.

Example: rectangular functions

Let

r(x)={1,∣x∣≤12,0,∣x∣>12.r(x)= \begin{cases} 1,&|x|\leq \tfrac12,\\ 0,&|x|>\tfrac12. \end{cases}

The value of (r∗r)(x)(r*r)(x) is the length of overlap between two intervals of unit length:

(r∗r)(x)={1−∣x∣,∣x∣≤1,0,∣x∣>1.(r*r)(x)= \begin{cases} 1-|x|,&|x|\leq1,\\ 0,&|x|>1. \end{cases}

Thus convolving two rectangular functions produces a triangular function.

With

sinc⁡(ξ)={sin⁡(πξ)πξ,ξ≠0,1,ξ=0,\operatorname{sinc}(\xi)= \begin{cases} \dfrac{\sin(\pi\xi)}{\pi\xi},&\xi\neq0,\\ 1,&\xi=0, \end{cases}

the Fourier transform of rr is sinc⁡(ξ)\operatorname{sinc}(\xi). The theorem therefore gives

F[r∗r](ξ)=sinc⁡2(ξ).\mathcal F[r*r](\xi)=\operatorname{sinc}^2(\xi).

This obtains the transform of the triangular function without integrating its two linear pieces separately.

Applications

Linear systems and filtering

A linear time-invariant system with impulse response hh is represented, under appropriate assumptions, by

y=h∗x,y=h*x,

where xx is the input and yy is the output. The convolution theorem gives

Y(ξ)=H(ξ)X(ξ).Y(\xi)=H(\xi)X(\xi).

The function HH is the frequency response: it describes how the system changes the amplitude and phase of each frequency component.

The same principle applies in several dimensions. Two-dimensional convolution, for example, models spatial filtering and blur in images. The multidimensional theorem follows by replacing scalar variables with vectors and the product xξx\xi in the exponential with an inner product.

Probability distributions

If independent random variables XX and YY have probability density functions pXp_X and pYp_Y, then the density of their sum is

pX+Y=pX∗pY.p_{X+Y}=p_X*p_Y.

The theorem converts this convolution into multiplication of transforms.

Equivalently, using the characteristic function

φX(t)=E[eitX],\varphi_X(t)=\mathbb E[e^{itX}],

independence implies

φX+Y(t)=φX(t)φY(t).\varphi_{X+Y}(t)=\varphi_X(t)\varphi_Y(t).

This identity remains valid when the distributions have no densities. It is a central tool in studying sums of random variables and proving limit results such as the central limit theorem.

Differential equations

Fourier transformation converts constant-coefficient differential operators into multiplication by frequency-dependent expressions. Combined with the convolution theorem, this allows solutions of many partial differential equations to be expressed using a kernel or Green’s function.

For example, the solution of the heat equation on the whole real line can be expressed, for suitable initial data, as convolution with a Gaussian heat kernel. Transforming the equation reduces the spatial differential operator to multiplication and leaves an ordinary differential equation in time.

Polynomial multiplication

The coefficients of the product of two polynomials are the linear convolution of their coefficient sequences. With adequate zero-padding, transform multiplication followed by inverse transformation computes these coefficients efficiently. Exact integer or modular implementations require additional care beyond floating-point Fourier computation.

Conditions, extensions, and limitations

The theorem is an identity under specified assumptions, not a guarantee that every formal convolution or transform exists.

  • Integrability and interpretation: The L1L^1 setting gives a direct classical result. Extensions to square-integrable functions use norm-based interpretations of Fourier transformation; pointwise integral formulas need not hold everywhere.
  • Generalized functions: In distribution theory, convolution and Fourier transformation extend to objects such as the Dirac delta. However, arbitrary distributions cannot always be convolved or multiplied. Appropriate support, smoothness, or growth conditions remain necessary.
  • Boundary handling: A discrete Fourier transform represents periodic data. For finite records or images, zero extension, periodic extension, and reflected extension describe different boundary models and generally produce different convolution results.
  • Numerical error: FFT-based convolution is subject to floating-point rounding. Its computational advantage does not imply exact arithmetic.
  • Deconvolution: Formally recovering X=Y/HX=Y/H is unstable where HH is small and may be impossible where HH vanishes. Such inverse problems often require additional assumptions or regularization; the convolution theorem alone does not establish a stable inverse.
  • Convolution versus correlation: Correlation uses a different arrangement of shifts and, for complex-valued functions, conjugation. Its transform identity generally contains a complex conjugate and must not be substituted unchanged for the convolution theorem.