Differential Equations › Section D · Laplace transforms › Topic D.9
Convolution integrals
The transform of a product is not the product of the transforms. But there is an operation in t whose transform is a product: convolution. It inverts awkward transforms, and it solves equations whose forcing is not even known yet.
Products of transforms
Solving equations with Laplace transforms keeps producing products such as \(H(s)G(s)\): a transfer function times the transform of an input. When both factors are known rational functions, partial fractions usually finish the job. But sometimes the algebra is heavy, and sometimes \(G\) is not known at all, because the input is left general or comes from measured data. The question becomes: if \(F(s)\) and \(G(s)\) are transforms of \(f\) and \(g\), what function has transform \(F(s)G(s)\)? It is not \(f(t)g(t)\). The answer is an integral built from \(f\) and \(g\).
The convolution integral
The convolution of \(f\) and \(g\) is the function
The integration variable is \(\tau\), and \(t\) stays fixed during the integration, so the result is a function of \(t\). At each moment \(t\), one function is flipped and slid along the other, and their overlap from 0 to \(t\) is added up. Convolution is not ordinary multiplication. For example, \(1 * 1 = \int_0^t 1\,d\tau = t\), not 1. Convolution also describes blurring in images, where each pixel becomes a weighted sum of its neighbours, and echo in audio, where each output value is a weighted sum of past inputs.
Properties of convolution
Substituting \(u = t - \tau\) in the integral swaps the roles of the two functions, so convolution is commutative: \(f * g = g * f\). Choose whichever order makes the integral easier. Convolution is also associative and distributes over addition: \(f * (g + k) = f * g + f * k\), and constants pull out. Convolving with zero gives zero. But there is no ordinary function that acts as 1 does in multiplication: \(f * 1 = \int_0^t f(\tau)\,d\tau\), the running integral of \(f\), which is not \(f\) itself. The role of 1 is played by the delta, since \(f * \delta = f\).
The convolution theorem
If \(f\) and \(g\) are piecewise continuous and of exponential order, with transforms \(F\) and \(G\), then
Equivalently, \(\mathcal{L}^{-1}\{F(s)G(s)\} = (f * g)(t)\). The proof writes \(F(s)G(s)\) as a double integral over the quarter-plane, then changes variables and reverses the order of integration so that the inner integral becomes the convolution. The theorem turns a hard operation in \(t\), convolution, into an easy one in \(s\), multiplication. It works in both directions, which is what makes it useful.
Worked example: two routes to one answer
Invert \(\frac{1}{s^2(s + 1)}\) with the convolution theorem, then check with partial fractions. 1. Split the product: \(\frac{1}{s^2}\) inverts to \(t\), and \(\frac{1}{s + 1}\) to \(e^{-t}\). 2. So the inverse is
3. By parts, \(\int_0^t \tau e^{\tau}\,d\tau = (t - 1)e^t + 1\). 4. Multiply by \(e^{-t}\): the answer is \(t - 1 + e^{-t}\). Check: partial fractions give \(-\frac{1}{s} + \frac{1}{s^2} + \frac{1}{s + 1}\), which inverts to \(-1 + t + e^{-t}\), the same function.
Dividing by s means integrating
A special case is worth remembering. Since \(\frac{1}{s}\) is the transform of 1,
Dividing a transform by \(s\) integrates the original function from 0, just as multiplying by \(s\) differentiates it when \(f(0) = 0\). For example, \(\frac{1}{s(s^2 + 4)}\) is \(\frac{1}{s}\) times the transform of \(\frac{1}{2}\sin 2t\), so its inverse is \(\int_0^t \frac{1}{2}\sin 2\tau\,d\tau = \frac{1 - \cos 2t}{4}\). Check: partial fractions give \(\frac{1}{4}\big(\frac{1}{s} - \frac{s}{s^2 + 4}\big)\), the same answer.
Worked example: an unknown input
Solve \(y'' + 2y' + y = g(t)\) with \(y(0) = 1\), \(y'(0) = 0\), for a general \(g\). 1. Transform: \((s^2 + 2s + 1)Y - s - 2 = G(s)\). 2. So \(Y = \frac{s + 2}{(s + 1)^2} + \frac{G(s)}{(s + 1)^2}\). 3. The first term splits as \(\frac{1}{s + 1} + \frac{1}{(s + 1)^2}\), with inverse \(e^{-t} + te^{-t}\). 4. For the second, \(\frac{1}{(s + 1)^2}\) inverts to \(te^{-t}\), so by the convolution theorem
Check: with \(g = 0\) only the first two terms remain, and they satisfy both initial conditions.
Response as weighted memory
The last example shows a general pattern. For a system at rest, \(Y = H(s)G(s)\), so the response is \(y = h * g\), where \(h\) is the impulse response from The Dirac delta function. Read the integral \(\int_0^t h(t - \tau)g(\tau)\,d\tau\) as follows: the input at each earlier time \(\tau\) is a small impulse, and its effect now is weighted by \(h(t - \tau)\), the response that has had \(t - \tau\) time to develop. Adding up all the past inputs gives the present output. With nonzero initial values, the free response from those values is simply added on.
Integral equations
Convolution also solves equations in which the unknown sits inside an integral. Take
The integral is \(\sin t * y\), so transforming gives \(Y = \frac{1}{s^2} + \frac{Y}{s^2 + 1}\). Collect the \(Y\) terms: \(Y \cdot \frac{s^2}{s^2 + 1} = \frac{1}{s^2}\), so \(Y = \frac{s^2 + 1}{s^4} = \frac{1}{s^2} + \frac{1}{s^4}\). Inverting gives \(y = t + \frac{t^3}{6}\). Equations of this type, called Volterra integral equations, arise in population models and in materials with memory.
Common mistakes to avoid
Writing \(\mathcal{L}\{f(t)g(t)\} = F(s)G(s)\): the product rule for transforms belongs to convolution, not to ordinary products. Treating \(f * 1\) as \(f\): it is the running integral. Putting \(t\) where \(\tau\) belongs, such as \(f(t)g(t - \tau)\), or integrating from 0 to \(\infty\) instead of 0 to \(t\). Forgetting the constant needed to match a table entry before convolving, such as the \(\frac{1}{2}\) in \(\frac{1}{s^2 + 4}\). Reaching for convolution when a short partial fraction split is quicker: distinct simple factors rarely need it, while a repeated quadratic such as \((s^2 + 1)^{-2}\), whose inverse is \(\sin t * \sin t\), is a natural job for it.
Key takeaways
The convolution
is commutative, and its transform is \(F(s)G(s)\). Use it to invert products of known transforms, to write \(\frac{F(s)}{s}\) as a running integral, and to solve problems with an unspecified input: from rest, the response is \(h * g\), the impulse response convolved with the input. Integral equations of convolution type become algebra after transforming. Convolution is not multiplication: \(1 * 1 = t\).
Flashcards
Flashcard
(f * g)(t) = ∫₀ᵗ f(τ) g(t - τ) dτ
Test your understanding
Question 1 of 5
What is the convolution 1 * 1?
Question 2 of 5
What is the Laplace transform of t * e^(2t)?
Question 3 of 5
Writing 1/(s(s + 3)) as (1/s) · 1/(s + 3), what does the convolution theorem give as its inverse?
Question 4 of 5
Starting from rest, which formula gives the solution of y″ + 9y = g(t) for a general input g?
Question 5 of 5
Solve the integral equation y(t) = 1 + ∫ from 0 to t of y(τ) dτ.
