White noise

The sound of static, rain, and a fan running all night — what do they have in common?

Tune an old analog TV to a dead channel and you get that hissy "shhhhh" sound — that's white noise. Unlike a musical note, which is dominated by one or a few specific frequencies, white noise packs every frequency together at roughly the same intensity. There's no dominant pitch to latch onto, which is exactly why it sounds so featureless and "flat" to the ear.

The name comes from an analogy with light. White light (like sunlight) is a mix of every visible wavelength in roughly equal amounts — that's why a prism can split it into a full rainbow. White noise is the acoustic equivalent: a mix of every audible frequency in roughly equal amounts of power.

Mathematically, a signal is white noise if its power spectral density (PSD) is constant across frequency:

Here S(f) is the power at frequency f, and N₀ is just a constant — the same value no matter which frequency you plug in. In the time domain, this translates into a signal whose samples are statistically independent of one another: knowing the value of the signal right now tells you nothing about what it will be a moment later. That "no memory, no pattern" property is really the defining feature of white noise — the flat spectrum is really a consequence of it.

The plot below shows a short segment of simulated white noise in the time domain (top), together with its estimated power spectral density (bottom). Notice how jagged and unpredictable the waveform looks — there's no repeating pattern to find — while the spectrum stays roughly flat across the whole frequency range, with only the small random fluctuations you'd expect from a finite-length measurement.

Figure 1: Top: a short segment of simulated white noise in the time domain — random, with no visible pattern. Bottom: its estimated power spectral density, which stays roughly flat across frequency (the defining trait of white noise), aside from normal statistical fluctuation.

"White" is only one color in a whole family of noise types, each named after a rough analogy to light and each defined by how its power changes with frequency:

• White noise: Flat power spectral density — equal power at every frequency. Sounds like a hiss or static.
• Pink noise: Power falls off as 1/f — each octave carries equal total power. Sounds deeper and softer than white noise, closer to steady rainfall.
• Brown (red) noise: Power falls off as 1/f², an even steeper roll-off than pink noise. Sounds like a low rumble, similar to a waterfall or distant thunder.
• Blue and violet noise: Power increases with frequency (proportional to f and f², respectively). These sound thin and hissy, emphasizing high frequencies, and are less commonly encountered outside of specific technical contexts like dithering.

Why is white noise so useful?

Precisely because it's "featureless" and unpredictable, white noise turns out to be one of the most useful tools in signal processing, acoustics, and beyond:

• Masking other sounds: Its flat, broadband character makes white noise effective at masking intermittent sounds like traffic or conversation, which is why it's a common feature in sleep machines and office sound-masking systems.
• Testing audio equipment: Because it excites every frequency equally, white noise is a standard test signal for measuring the frequency response of speakers, microphones, and amplifiers — problems at any frequency will show up in the measurement.
• System identification: Engineers feed white noise into an unknown system (a filter, a room's acoustics, a communication channel) and analyze the output, because a signal containing all frequencies equally reveals how the system responds across its entire operating range.
• Statistics and modeling: White noise is the standard idealized model for random measurement error in statistics and time-series analysis. Many models (like ARMA processes) are built by describing how a system transforms white noise into structured output.
• Dithering: In digital audio and image processing, small amounts of white (or blue) noise are deliberately added before quantization to break up perceptible patterns like banding, trading a barely-audible or barely-visible hiss for smoother, more natural-looking results.
• Cryptography and random number generation: Physical white-noise sources (like thermal noise in a resistor) are sometimes used as a seed for generating true random numbers, since the signal is unpredictable by nature.

Practical limits and common pitfalls

White noise is a clean theoretical concept, but a few caveats matter when working with it in the real world:

• Truly infinite bandwidth is impossible: A mathematically perfect white noise signal would need infinite power, since it has equal power at every frequency out to infinity. Real-world "white" noise is always band-limited by the physics of the device generating or capturing it (a speaker, a microphone, a sampling rate) — so it's white only up to some practical cutoff.
• "White" describes the spectrum, not the amplitude distribution: A signal can have a flat spectrum whether its instantaneous values follow a Gaussian (bell-curve) distribution or some other distribution entirely. "White Gaussian noise," used throughout this article and commonly in engineering, specifies both properties together, but they are technically independent choices.
• Estimated spectra are never perfectly flat: Any real measurement is based on a finite chunk of data, so the estimated power spectral density will always show some random ripple around the flat average, as seen in Figure 1. This is expected statistical variation, not evidence that the noise isn't "white."
• Not all masking noise is actually white: Many commercial "white noise" machines and apps actually play pink or brown noise, because the emphasis on lower frequencies tends to sound more pleasant and less harsh to human ears over long periods. It's worth checking the actual spectrum if precision matters for your use case.

White noise also shows up constantly as a baseline in machine learning and signal processing pipelines: it's the default assumption for background noise in many denoising and enhancement algorithms, and it's frequently used to initialize or regularize generative models, precisely because it carries no built-in structure of its own for the model to accidentally rely on.

The takeaway: white noise is defined by having equal power across every frequency — a simple, useful idealization for testing, modeling, and masking — but real-world noise almost always has some color to it, so it's worth checking the actual spectrum rather than assuming "noise" automatically means "white."