noise analysis
Even with no signal applied, every resistor hisses and every transistor crackles, a faint random fuzz set by physics that limits the quietest signal a circuit can handle. Noise analysis predicts that fuzz: it estimates how much random noise appears at your output, and from which components it came, all without you ever wiring up a real low-noise measurement.
It is an AC-style analysis. SPICE linearizes the circuit at its operating point, then adds each device's intrinsic noise sources, a resistor's thermal (Johnson) noise, a transistor's shot and flicker noise, and computes how each contributes at the output across frequency. The command names an output and an input, like '.noise V(out) Vin dec 100 1 100k'. Noise is random, so it does not add as plain voltages but as power: contributions combine as the square root of the sum of squares. The output is usually given as a noise density in nanovolts per root-hertz, which you integrate over your bandwidth to get a total.
The real gift of noise analysis is the breakdown by source: SPICE will tell you that, say, one bias resistor contributes most of the output noise, pointing straight at what to change. But it is only as honest as the device noise models, and it captures none of the noise that dominates real benches, mains hum, digital switching pickup, ground loops. So it sets a best-case floor; the real circuit is usually noisier.
A single resistor R at room temperature generates a thermal noise density of about 0.13 nV per root-hertz times the square root of R in kilohm, so a 10 kilohm resistor hisses at roughly 13 nV per root-hertz, exactly the figure a noise analysis attributes to it.
Predicted output noise versus frequency, broken down by which component contributes most.
Noise analysis only includes the intrinsic device noise in the models. It cannot predict interference picked up from outside, mains hum, crosstalk, or grounding problems, which often dominate a real measurement.