Convolution, Intuitively — What a Cymbal Can Teach You About Signal Processing
Most people encounter convolution in a textbook as an intimidating integral:
And immediately feel lost. What is τ? Why is h flipped? What is this even computing?
Here’s the thing — you already have the intuition. You just need to connect the dots!
Let’s play some music! Hitting a Cymbal Once…
Imagine you strike a cymbal at time t = 0. What happens?
Sound radiates outward from the point of impact. It spreads, bounces, and slowly decays. If you were to measure the amplitude of that vibration over time, you’d see something like this — a sharp peak right at the moment of impact, then a smooth exponential decay:
This is called the impulse response of the system. It answers a simple question: if I poke this system exactly once, how does it respond?
The key insight for this practical scenario: the system is causal. That means it only reacts after you hit it, not before. That’s why h(t) is zero for t < 0.
Now hit the Cymbal to the rhythm…
What if you don’t hit the cymbal once, but three times, at t = 0, t = 1, and t = 3; with the last hit being a softer tap?
Each hit triggers its own independent decaying response. The hit at t = 0 starts its decay immediately. The hit at t = 1 starts a second, identical decay one second later. The soft tap at t = 3 launches a half-amplitude decay two seconds after that.
This is your input signal, x(t):
- An impulse of amplitude 1 at t = 0
- An impulse of amplitude 1 at t = 1
- An impulse of amplitude 0.5 at t = 3
What you hear — the total sound at any given moment — is all three decaying responses playing simultaneously, layered on top of each other.
That layered sum is the convolution!
Get Chirantha Kurukulasuriya’s stories in your inbox
Join Medium for free to get updates from this writer.
Let’s find the output signal mathematically from the convolution of the hitting pattern x(t) and the impulse response h(t) now.
The Formal Picture
The cymbal’s total output y(t) is the convolution of the hitting pattern x(t) and the impulse response h(t):
Let’s break this down visually.
Step 1: Fix x(τ)
Let’s plot the hitting pattern along the τ-axis. Three impulses at τ = 0, 1, 3.
Step 2: Flip h(τ) to get h(−τ)
Take the impulse response and flip it about the vertical axis. The decaying tail that previously ran to the right now runs to the left.
Why the flip? Because we’re asking: “For a given output time t, which past inputs are still contributing?” Flipping reverses the natural flow of time so we can look backward from t.
Step 3: Slide h(t − τ) across x(τ) and build the output
Now shift the flipped curve to the right by an amount t. As t increases from 0, the flipped curve slides across the fixed impulses.
Each time h(t − τ) passes over an impulse in x(τ), that impulse contributes to the output — scaled by its amplitude and weighted by where it falls on the decaying curve.
Notice the jump discontinuities at t = 1 and t = 3 — that’s each new cymbal hit arriving instantly while the previous responses are still ringing.
Why This Matters Beyond Cymbals
The cymbal is just a friendly way to picture something deeply general. The same mathematics describes:
- Audio reverb — a dry vocal convolved with a room’s impulse response produces the sound of singing in that room
- Image blurring — a sharp photo convolved with a Gaussian kernel produces a blurred one
- Edge detection — a Sobel filter is just convolution with a hand-designed kernel
- Probability — the distribution of a sum of two random variables is the convolution of their individual distributions
- Neural networks — convolutional layers are literally doing this, shifted and stacked millions of times
In every case, the story is the same.
One-Sentence Summary
Convolution asks: at each output moment, how much do all past inputs still contribute, weighted by how the system responds to each one?
The flip reverses time so you can look backward. The slide moves you forward through the output. The integral sums it all up.
Once you see the cymbal, you can’t unsee it.
Tags: Signal Processing Mathematics Machine Learning Education Data Science
