[Intermediate] — the central idea of DSP: any signal is a sum of sinusoids, and switching to that view makes hard problems easy.
The kid version first
A musical chord looks like a single messy wiggle on an oscilloscope. But your ear hears three distinct notes — it takes that one wiggle apart into its constituent pure tones.
The frequency domain is doing exactly that, mathematically: taking any signal and finding which pure sinusoids, at which strengths, add up to make it. The wiggle over time (the time domain) and the list of frequencies it contains (the frequency domain) are two views of the same signal — and some problems are trivial in one view and impossible in the other.
The big idea
Any signal can be written as a sum of sinusoids of different frequencies, amplitudes and phases.
This is Fourier’s theorem, and it’s one of the most consequential ideas in all of applied mathematics. A square wave, a spoken word, an image, a stock chart — each is a combination of pure sine waves, and Fourier analysis finds the recipe.
TIME DOMAIN FREQUENCY DOMAIN
"how the signal "which frequencies it
changes over time" is made of, and how much"
a messy wiggle ◄──► a few spikes: 100Hz strong,
250Hz medium, 3kHz faint
Neither view is more “real.” They contain identical information — the transform is reversible, losing nothing. You switch to whichever makes your problem easier, which is the whole trick.
Why the frequency view is so powerful
Three things become obvious in the frequency domain that are hard in the time domain:
1. Filtering becomes selection. “Remove the hum” is hopeless to express as time-domain arithmetic, but in the frequency domain it’s “delete the spike at 50 Hz.” Noise, hum, and unwanted bands are visible and separable → filters.
2. LTI systems become multiplication. An LTI system applies a gain and phase shift per frequency — so its effect is just multiplying the input’s spectrum by the system’s frequency response. The messy time-domain operation (convolution) becomes simple multiplication. This is the single biggest reason the frequency domain matters.
3. Structure appears. A heartbeat’s rhythm, a rotating machine’s vibration signature, the pitch of a voice — periodic structure that’s buried in the time wiggle shows up as clear peaks in the spectrum.
The Fourier family
There are four Fourier transforms, and the differences are just “is the signal periodic/aperiodic, continuous/discrete.” You mostly use the last one, but the family clarifies which is which:
| Time signal | Result | |
|---|---|---|
| Fourier Series | Continuous, periodic | Discrete frequencies (harmonics) |
| Fourier Transform (FT) | Continuous, aperiodic | Continuous spectrum |
| Discrete-Time FT (DTFT) | Discrete, aperiodic | Continuous, periodic spectrum |
| Discrete Fourier Transform (DFT) | Discrete, finite | Discrete, finite — the one computers use |
The DFT is the practical one — it takes N samples in and gives N frequency values out, all discrete and finite, so a computer can actually compute it. Everything a program does is a DFT (usually via its fast algorithm, the FFT).
Reading a DFT output
The DFT of N samples gives N complex numbers, one per frequency bin. Each complex number encodes two things:
- Magnitude — how much of that frequency is present (the amplitude). This is what a spectrum plot shows
- Phase — the timing/alignment of that frequency component. Often ignored in visualisation, but essential — throw away phase and you can’t reconstruct the signal, and it’s why two signals with identical spectra can sound or look different
The bins span from 0 (DC, the average) up to the Nyquist frequency fs/2. For real signals the upper half mirrors the lower (conjugate symmetry), so you usually plot only the first N/2 bins.
Frequency resolution = fs / N: more samples → finer bins → you can distinguish closer frequencies. This is a fundamental trade — capturing more time gives better frequency resolution, which is the seed of the time-frequency uncertainty.
Why complex numbers?
DSP is full of e^{jω} and complex spectra, which can feel like unnecessary abstraction. The reason is clean: a complex exponential e^{jωt} is a rotating point — it encodes both a cosine (real part) and a sine (imaginary part) at once, and rotation is the natural language of oscillation → the maths.
Practically, the two numbers per frequency (real/imaginary, or magnitude/phase) are amplitude-and-timing. Complex numbers aren’t a complication bolted on; they’re the compact, correct way to carry “how much and when” for each frequency. Euler’s formula e^{jθ} = cos θ + j·sin θ is the bridge.
Key insight
The frequency domain is the same signal seen as “which sinusoids it’s made of,” and switching to that view is powerful because filtering becomes selection and LTI processing becomes multiplication. Fourier’s theorem — any signal is a sum of sinusoids — is what lets you move between the two views losslessly, and the DFT is the finite, discrete version a computer can actually run. Once you’re comfortable that a signal is its spectrum, half of DSP becomes “do the easy thing in whichever domain it’s easy in.”
Related
- the FFT — the algorithm that computes the DFT fast enough to matter
- convolution — why “convolution in time = multiplication in frequency”
- spectral analysis — the DFT applied to real, noisy, finite data
- approximation — Fourier as a basis expansion
Source: [reference] — Aug 2026.
Software engineering
Infrastructure & security
Computer science
Hardware & engineering