A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Convolution in time is equivalent to multiplication in the frequency domain: y(t) = x(t) * h(t) ⟺ Y(f) = X(f)H(f). This theorem greatly simplifies system analysis and is the basis for fast filtering algorithms and spectral methods.
You already know the Fourier transform: a signal x(t) maps to its spectrum X(f), and the transform pair encodes how much energy exists at each frequency. You also know the key Fourier properties — linearity, time shift as a phase rotation, and so on. The convolution theorem is the deepest of these properties, and it connects two operations that seem unrelated: the integral of one signal "sliding over" another in the time domain, and simple pointwise multiplication in the frequency domain.
Recall what convolution computes. If h(t) is a system's impulse response — the output when the input is a brief spike — then the output for any arbitrary input x(t) is the convolution y(t) = ∫ x(τ) h(t−τ) dτ. Intuitively, you're decomposing x into a sum of scaled, shifted spikes, computing the system's response to each, and superimposing. This is a complete and correct description of any linear time-invariant (LTI) system. The problem is computational cost: if x and h each have N samples, computing this sum directly requires N² multiplications. For audio at 44,100 samples per second convolved with a room impulse response of 1 second, that's about 2 billion multiplications per second — far too slow.
The convolution theorem cuts through this. Take the Fourier transform of both x and h, multiply their spectra pointwise (Y(f) = X(f)H(f)), and take the inverse Fourier transform. The result is identical to direct convolution, but the frequency-domain multiplication is O(N) once you have the spectra. Since computing the Fourier transform can be done in O(N log N) operations using the Fast Fourier Transform (FFT), the entire convolution reduces from O(N²) to O(N log N) — a massive speedup that makes real-time audio reverb, image blurring, and radar signal processing computationally practical.
The physical interpretation is equally important. H(f) = ℱ{h(t)} is the system's frequency response: for each frequency f, H(f) tells you how much the system scales and phase-shifts a pure sinusoid of that frequency. A low-pass filter has |H(f)| ≈ 1 for low frequencies and |H(f)| ≈ 0 for high frequencies. When you compute Y(f) = X(f)H(f), you are literally multiplying each frequency component of the input by the filter's gain at that frequency. This is why designing filters in the frequency domain is so natural: you specify the desired H(f) directly (flat passband, sharp rolloff, etc.), and the convolution theorem guarantees that applying this filter to any input in the time domain produces exactly the effect you specified. Every digital filter, image blur, EQ effect, and spectral analysis tool rests on this equivalence.