On paper it works. On the bench it hums.
Every guide so far quietly assumed a clean world: the op-amp's golden rules held, the divider divided, and the signal you fed in was the only thing on the wire. Reality is messier. The instant you build a real circuit, a faint, restless fuzz rides on every node — a wobble that was never part of your design and never sits still. Engineers call anything that is not your wanted signal noise. This whole rung is about where it comes from and how the pros tame it, and this first guide draws the map.
Noise splits cleanly into two families, and telling them apart is half the battle. Intrinsic noise is born inside the components themselves — the ceaseless jiggle of electrons, the graininess of current. It is set by physics and temperature; you can reduce it, but never drive it to zero. Interference (extrinsic noise) is picked up from the outside world — mains hum, a nearby motor, a digital clock, a radio station. In principle it can be eliminated entirely, given good layout, grounding, and shielding. One is a floor you can only lower; the other is a mess you can, with care, sweep away completely.
The hiss that never sleeps: thermal noise
The most fundamental noise needs no signal, no current, nothing but warmth. Inside any resistor the electrons are in perpetual thermal motion, jostling like dust motes in a sunbeam — and that random jiggle is itself a tiny, fluctuating voltage appearing across the resistor's ends. This is thermal noise (also called Johnson-Nyquist noise), and every resistor on earth makes it, all the time, for the single reason that it is above absolute zero. Cool it toward absolute zero and the jiggle fades — which is exactly why radio telescopes chill their front-end amplifiers in liquid nitrogen.
The amount is wonderfully predictable. The RMS noise voltage is Vn = sqrt(4 times k times T times R times B), where k is Boltzmann's constant (1.38 x 10^-23 J/K), T the absolute temperature in kelvin, R the resistance, and B the bandwidth in hertz over which you are listening. Three levers jump out: hotter, larger resistance, or wider bandwidth each means more noise. One handy number is worth memorizing — a 1 kΩ resistor at room temperature (about 300 K) makes roughly 4 nV for every square-root-hertz of bandwidth, written 4 nV/sqrt(Hz).
Put that to work. Listen to that 1 kΩ resistor across a 1 MHz bandwidth and the noise is 4 nV/sqrt(Hz) times sqrt(1,000,000) = 4 nV times 1000 = 4 uV RMS. Two design rules drop straight out of those square roots: quadruple the resistance and the noise only doubles (it grows as sqrt(R), not R); halve the bandwidth and the noise falls by about 30 percent (a factor of sqrt(2), or 3 dB). That last lever is the quiet hero of low-noise design — never let a circuit listen to more bandwidth than the signal genuinely needs.
Thermal noise is white: its power is spread evenly across all frequencies, just as white light contains every colour, so it sounds like the steady hiss between radio stations. White also means you cannot dodge it by tuning — it is present at every frequency you keep. It sets the absolute floor of the whole craft: no resistor, amplifier, or sensor can ever be quieter than its own thermal noise.
Two more from within: shot noise and flicker noise
Thermal noise is about voltage jiggling at rest; the next kind shows up only when current flows, and it springs from current being grainy. Current is not a smooth fluid but a hail of individual electrons, each carrying one fixed packet of charge. When they cross a barrier such as a junction in a diode or transistor, they arrive at random instants — like rain drumming a tin roof, steady on average but built from countless discrete hits. That randomness is shot noise, In = sqrt(2 times q times I times B), where q is the electron charge (1.6 x 10^-19 C).
The telling feature hides in the formula: shot noise grows as the square root of the current, so relative to the current it gets worse as the current gets smaller. A robust 1 mA carries shot noise of about 18 nA over a 1 MHz bandwidth — utterly negligible, a hundred-thousandth of the signal. But shrink to the nanoamps creeping out of a photodiode in dim light and shot noise can become the dominant limit. It is precisely why a faint-light sensor struggles where a bright-light one sails through untroubled.
The third intrinsic noise is the oddball. Flicker noise, or 1/f noise, is not flat at all: its power rises as you go down in frequency, climbing without bound toward DC. It comes from charge carriers being briefly trapped and released by imperfections in the silicon, a slow restless flickering. Because it piles up at low frequencies it is the curse of slow and DC measurements — weighing on a strain gauge, reading a thermocouple, holding a precise reference steady. It sounds 'pink', weighted toward the low rumble. Above a part's corner frequency (often a few Hz to a few hundred Hz) thermal noise takes over and flicker fades back into the white hiss.
INTRINSIC NOISE spectrum physical source grows with tame it by
---------------- --------------- --------------------------- --------------- --------------------------
thermal (Johnson) white (flat) electrons jiggling w/ heat sqrt(R x T x B) smaller R, colder, less BW
shot white (flat) current = discrete charges sqrt(I x B) (set by I) less bandwidth
flicker (1/f) pink (rises carriers trapped/freed in 1 / sqrt(f) chopper/auto-zero, avoid
toward DC) material defects precision DC work
rule of thumb: thermal floor of a 1 kohm resistor at room temp ~ 4 nV / sqrt(Hz)The other family: interference from the outside
Everything so far is born inside the circuit. The second family barges in from outside, and on a real bench it usually dwarfs the intrinsic floor. Your wires are unintentional antennas, and the world is loud: the 50 or 60 Hz field radiating from every mains cable, the buzz of a switching power supply, the sharp edges of a digital clock, a phone's radio chatter. Three doors let this interference in — an electric field sneaks through stray capacitance (capacitive coupling), a changing magnetic field induces currents in wire loops (inductive coupling), and noise rides straight in along shared wires and a shared ground (conducted coupling). When one trace bleeds into its neighbour, we name it crosstalk.
The classic symptom is mains hum — a steady 50 or 60 Hz buzz, with its harmonics, that capacitively couples from the power wiring into any high-impedance node. Audio folk know it as the curse of a guitar lead; instrument builders fight it on every sensitive front end. Unlike the intrinsic floor, hum is not destiny: it has a source, a path, and a victim, and breaking any one of the three kills it. That is exactly what the rest of this rung does — decoupling to give each chip a local charge reservoir, proper grounding to break loops, and shielding, twisted pairs, and ferrite beads to block the coupling paths.
- Look at the fuzz on a scope. Broadband, patternless hiss that stays the same whether or not the circuit is near other gear is intrinsic noise — it sits at the physics floor, and chasing it means lower R, less bandwidth, or a quieter first stage.
- See a clean 50 or 60 Hz wobble (or its harmonics)? That is mains interference coupling in. Move the cabling and shorten the high-impedance leads, then watch it change — interference responds to where things physically sit; intrinsic noise does not.
- See sharp spikes locked in step with something else switching — a clock, a motor, a relay? That is conducted or coupled digital interference. It appears and vanishes with the aggressor, which is exactly how you fingerprint it.
- Touch the suspect node with a finger and the hum jumps? You have just proved it is high-impedance capacitive pickup — the cure is lower impedance, shielding, or a ground done right, never a quieter resistor.
Measuring the enemy: SNR and the noise floor
A noise figure in volts means nothing on its own; what matters is how it compares to your signal. That ratio is the signal-to-noise ratio (SNR), almost always quoted in decibels: SNR(dB) = 20 times log10(Vsignal / Vnoise), with both taken as RMS values. A 1 V RMS signal riding on 1 mV RMS of noise has a ratio of 1000, which is 20 times log10(1000) = 60 dB — a comfortably clean signal. At 0 dB the signal and the noise are equal in size; below that, the signal is drowning.
The lowest your signal can shrink before it disappears into the fuzz is the noise floor — the total RMS noise at your circuit's output, intrinsic plus whatever interference you failed to keep out. The span from the largest signal the circuit can handle without distortion down to that floor is its dynamic range, and it tells you, in one number, how delicate a measurement the circuit can make. Push the floor down and you can resolve quieter signals — that is the entire point of low-noise design.
Here is the design lesson that governs the whole signal chain, and it is not obvious. Gain amplifies your signal and the stage's own noise by the very same factor, so you cannot 'amplify your way out' of a noisy front end — once noise is mixed into the signal, every later stage boosts both together. The fix is to make the very first stage the quietest and give it most of the gain, lifting the signal well clear of the noise before anything downstream can add more. That is why sensitive systems open with a dedicated low-noise amplifier: the first stage sets the floor for everything that follows it in the signal chain.