[Beginner] — what a signal is, what a system does to it, and the one class of system that makes the whole field tractable.

The kid version first

A signal is just a measurement that changes over time: the air pressure hitting a microphone, the voltage on a wire, the brightness along a row of pixels, your heart rate. A system is anything that takes a signal in and puts a (changed) signal out — a volume knob, an equaliser, a noise filter, an echo.

Digital signal processing is doing this with numbers, on a computer, instead of with analog circuits — and it turns out that almost everything interesting (removing noise, finding a heartbeat, decoding Wi-Fi, recognising speech) is one system acting on one signal.

What a signal is, precisely

A signal is a function mapping an independent variable (usually time) to a value:

  • Continuous-time — defined at every instant. Real-world, analog: sound, voltage, temperature. Written x(t)
  • Discrete-time — defined only at sample points. What a computer holds: a sequence of numbers. Written x[n], where n is an integer index

The whole discipline lives at the boundary between these two — the real world is continuous, computers are discrete, and sampling is the bridge.

Signals can be 1-D (audio, x[n]), 2-D (an image, x[m,n]), or higher (video). The maths is the same in every dimension, which is why the ideas here transfer straight to image processing and to convolutional networks.

The building-block signals

A few idealised signals recur everywhere, because complex signals are built from them:

  • The sinusoid A·cos(2πft + φ) — a pure tone at frequency f. The single most important signal in DSP, because every signal is a sum of these
  • The impulse δ[n] — 1 at n=0, zero everywhere else. A single spike. It’s how you probe a system → the impulse response
  • The step — 0 then 1. A switch turning on
  • The exponential — decay or growth, and the complex exponential e^{jωn}, which is a rotating sinusoid and the mathematical heart of the frequency domain

What a system does

A system transforms an input signal into an output signal. We classify systems by their properties, because the properties determine what maths applies:

  • Linear — scaling the input scales the output, and inputs add: .
  • Superposition. This is the property that lets you decompose a hard signal into simple parts, process each, and add the results
  • Time-invariant — the system behaves the same today as tomorrow; delaying the input just delays the output. The rules don’t change over time
  • Causal — the output depends only on present and past inputs, not the future. Required for real-time processing (you can’t use samples you haven’t received yet)
  • Stable — a bounded input produces a bounded output; it doesn’t blow up → filter stability

LTI — the abstraction that makes DSP tractable

Linear Time-Invariant systems are the ones that are both linear and time-invariant, and almost all of classical DSP is the study of LTI systems — because they have an almost magical property:

An LTI system is completely described by its response to a single impulse.

Poke an LTI system with one spike, record what comes out (the impulse response), and you now know exactly what it will do to any input — via convolution. You never have to test it with anything else.

And there’s a second gift: feed an LTI system a pure sinusoid and you get back the same-frequency sinusoid, only scaled and phase-shifted — never a new frequency. This is why the frequency domain is so powerful: an LTI system just applies a “gain and delay per frequency,” which turns hard time-domain problems into simple frequency-domain multiplication.

Why this matters: most useful systems either are LTI (filters, echoes, most audio and comms processing) or are usefully approximated as LTI over a working range. The moment a system is LTI, the entire toolkit — impulse response, convolution, Fourier, transfer functions — applies.

Why DSP is everywhere

Once signals are numbers, you can do anything computation can do, precisely and reproducibly:

  • Audio — noise removal, EQ, compression, effects, speech recognition
  • Communications — every phone, Wi-Fi and satellite link modulates and demodulates digitally → modulation
  • Images and video — filtering, compression (JPEG is DSP), enhancement → graphics
  • Sensors and control — filtering noisy readings, extracting signals → control theory, robotics
  • Biomedical — ECG, EEG, ultrasound, MRI reconstruction
  • Machine learning — a CNN’s core operation is convolution; audio and speech models run on spectrograms → ai-ml

Analog vs digital, briefly: analog processing (op-amps, RC circuits) is instantaneous and continuous but drifts, is hard to change, and can’t do anything complex. Digital is a sampled approximation but is exact, reproducible, reprogrammable, and can run arbitrarily sophisticated algorithms. The world went digital because “reprogrammable and exact” beats “instant and continuous” for almost everything → digital and analog.

Key insight

A signal is a measurement over time, a system transforms it, and the entire classical toolkit exists because Linear Time-Invariant systems are fully described by their response to a single impulse. That one fact — poke it once and you know everything it does — is what makes convolution, the frequency domain, and filter design all work. Learn to recognise when a system is LTI, and the rest of DSP is the machinery for exploiting it.

Source: [reference] — Aug 2026.

↑ Back to Digital Signal Processing