> ## Documentation Index > Fetch the complete documentation index at: https://agentconcepts.io/llms.txt > Use this file to discover all available pages before exploring further. # signal-to-noise > The ratio of meaningful signal to background noise, which bounds how reliably the signal can be discriminated or extracted. You improve it only by changing the ratio — amplifying the signal or suppressing the noise — never by raising the overall level; louder is not clearer. engineering-and-technology psychology statistics # Signal-to-noise ## Description Signal-to-noise is the ratio between the meaningful component of an observation — the *signal* — and the random or irrelevant variation competing with it — the *noise*. The ratio, not the absolute size of either part, is what bounds how reliably the signal can be discriminated, extracted, or acted on. A faint signal in near-silence is easy to read; a strong signal buried in stronger noise is not. The load-bearing structural claim is that the situation improves only by *changing the ratio* — amplifying the signal relative to the noise, or suppressing the noise relative to the signal — and never by raising the overall level. Turning up the gain scales both parts together and leaves discriminability exactly where it was: *louder is not clearer.* In its home domain of communications and signal processing the ratio is usually quoted in decibels, SNR(dB) = 10·log₁₀(P\_signal / P\_noise), and it is the quantity that, through the Shannon-Hartley relation C = B·log₂(1 + S/N), sets the ceiling on reliable transmission rate — capacity rises with the ratio. But the shape is not specific to engineering. It recurs wherever a meaningful pattern must be pulled out of a competing background: a true effect against sampling variance in statistics, a target voice against a room of talkers in psychoacoustics, a genuine trend against churn in a market, a diagnostic finding against measurement artifact in medicine. The diagnostic question — *"is the meaningful part large relative to the background, and if not, can I raise the signal or lower the noise?"* — separates a signal-to-noise problem from an effort or level problem. When results are unreliable, the move is not "measure harder" or "amplify everything," but "improve the ratio": average more samples (noise falls as √n while the signal holds), filter to the band the signal occupies, remove a source of interference, or strengthen the signal at its source. ## Aliases "SNR" and "signal-to-noise ratio" are the same quantity. The decibel convention — a logarithmic ratio in which +3 dB is roughly a doubling of the power ratio — comes from the telephony origins of the term at Bell Labs, the same lineage as [channel-capacity](/concepts/channel-capacity). The metaphor has since spread far past engineering: "signal vs noise" is now everyday vocabulary for the informative part versus the distracting background in writing, hiring, research triage, and social feeds. That spread is itself evidence of a portable structural shape rather than a domain-bound piece of jargon. ## Triggers **User-initiated:** User describes a meaningful pattern being drowned out, buried, or hard to pick out of a noisy background, or reaches for "amplify everything" / "look harder." Vocabulary cues: "signal-to-noise," "SNR," "buried in noise," "drowned out," "too noisy," "can't separate signal from noise," "false positives." **Agent-initiated:** Agent notices the user trying to raise the overall level — more data indiscriminately, louder output, higher gain — when the problem is a *ratio*, and the noise is scaling right alongside the signal. Candidate inference: "this is a signal-to-noise problem, not a level problem; the lever is raising signal *relative to* noise or suppressing noise *relative to* signal." **Situation-shape signals:** Unreliable discrimination that does not improve with more effort or amplification; a real pattern competing with a background of irrelevant variation; interventions that scale signal and noise together and so do not help; debates about extraction quality with no diagnosis of the underlying ratio. ## Exclusions * **The limit is throughput, not discriminability** — signal-to-noise is the upstream ratio that *determines* the maximum reliable rate; [channel-capacity](/concepts/channel-capacity) is that rate itself. Diagnosing a capacity ceiling as an SNR problem (or the reverse) confuses how *cleanly* the signal reads with how *fast* it can be sent. * **The signal is already corrupted and must be reconstructed** — signal-to-noise is raw discriminability *before* any repair; [error-correction](/concepts/error-correction) uses redundancy to rebuild content that noise has already damaged. Raising SNR (a cleaner input) and correcting errors (a repaired output) are different levers at different stages. * **Level, not ratio** — turning up the gain amplifies signal and noise together and leaves the ratio, and therefore discriminability, exactly where it was. An intervention that raises the overall level without changing the ratio is not an SNR move. * **No definable signal to separate** — SNR presupposes a canonical signal to extract from the noise. When what is being called "noise" is really structure not yet understood, or there is no ground-truth signal at all, the ratio is undefined and the framing misleads. ## Structure Internal structure of signal-to-noise: a table of its component slots and the concepts that fill them. ## Relationships Relationship neighborhood of signal-to-noise: a graph of the concepts it connects to and the concepts it is a part of. * [channel-capacity](/concepts/channel-capacity) — SNR is the input; capacity is the output. Shannon-Hartley (C = B·log₂(1 + S/N)) makes the reliable-rate ceiling an increasing function of the ratio. * [error-correction](/concepts/error-correction) — the downstream repair that buys reliability when the ratio is low; complementary to raising the ratio, not a substitute for it. * [figure-ground](/concepts/figure-ground) — the perceptual cousin: separating attended foreground from receding background is the same shape as separating signal from noise, one domain over. ## Examples In communications the signal-to-noise ratio, quoted in decibels, is the quantity that sets a channel's capacity: the Shannon-Hartley theorem gives C = B·log₂(1 + S/N), so the maximum reliable bit-rate rises with the ratio rather than with raw signal power alone. Doubling both signal and noise together — turning up the gain — leaves S/N, and therefore capacity, unchanged; only improving the ratio helps. **Inference**: When a link is unreliable, the lever is the ratio, not the power. Raise the signal at its source, or cut a noise contributor (shielding, filtering, a quieter band) — amplifying everything downstream buys nothing. Cherry's cocktail-party problem is signal-to-noise in perception: a listener extracts one target voice (the signal) from a background of competing talkers (the noise). The brain raises the *effective* ratio using cues the raw acoustics do not carry on their own — the spatial separation between the two ears, a familiar voice's pitch, semantic prediction of what comes next — which is why two ears beat one, and why one's own name spoken across a room can cut through. Selective attention is a noise-suppression mechanism operating on an otherwise fixed physical ratio. A standardized effect size such as Cohen's d is a signal-to-noise ratio — the difference between two group means (the signal) divided by the pooled standard deviation (the noise). A real effect can be undetectable not because it is small but because the noise is large relative to it, and raising the ratio is what makes it detectable: a bigger effect, or less variance through better measurement or more averaging. Statistical power is downstream of this ratio, which is exactly why "collect a larger sample" works — averaging shrinks the noise term as √n while leaving the signal intact.