Noise, Randomness, and the Boundary Between Signal and Systematic Error

The highlighted sentence says: “Noise is random variation that can obscure it.” In context, “it” means the signal: the effect the experimenter is trying to detect. This is a good first definition for a physicist learning how evidence is separated from confusion. But it is also a compressed definition. To use it well, we need to ask what “random” means, whether all noise must be random, and whether quantum randomness is therefore the deepest or “true” model of noise.

The short answer is: in the strict statistical sense, noise is usually modeled as random variation; in broader laboratory language, people sometimes call nonrandom unwanted disturbances “noise” too. But if an unwanted effect is reproducible and biased, the parent document is right to separate it from noise and call it a systematic error. Quantum randomness is one important physical source of noise, especially in quantum measurement, photon counting, electronics, and precision metrology, but it is not the universal correct model for all noise.

Signal, noise, and systematics as a model of measurement

A measurement is rarely just “nature’s answer.” It is usually a mixture of the effect we care about, random fluctuations, and biases introduced by the apparatus or analysis. A simple way to write this is

$$ y_i = s_i(\theta) + \epsilon_i + b_i . $$

Here $y_i$ is the $i$-th measured value, $s_i(\theta)$ is the signal predicted by a model with parameters $\theta$, $\epsilon_i$ is noise, and $b_i$ is a systematic contribution or bias. This equation is not a law of nature; it is a useful bookkeeping structure.

In this structure, noise is the part that varies unpredictably from trial to trial. Often we assume

$$ \mathbb{E}[\epsilon_i] = 0, $$

meaning that if we could repeat the measurement many times under the same conditions, the random fluctuation would average to zero. The noise still matters because any one measurement can be pushed upward or downward by it. Its typical size might be described by a standard deviation $\sigma$, where

$$ \sigma^2 = \mathbb{E}\left[(\epsilon_i - \mathbb{E}[\epsilon_i])^2\right]. $$

This is why noise can obscure a signal. If the signal is smaller than, or comparable to, the typical random fluctuation, then the measurement may not reliably reveal the effect. In simple cases, repeated measurements help: random errors tend to average down roughly like $1/\sqrt{N}$, where $N$ is the number of independent measurements [Taylor 1997; Bevington and Robinson 2003].

A systematic error behaves differently. If $b_i$ is a constant offset, then repeating the experiment does not make it go away. You can average a thousand measurements and still get the wrong answer with great precision. This is why the parent document says systematic error is often more dangerous than noise: it can imitate a real effect.

Is noise always random?

In careful statistical language, noise is random, or at least modeled as random. But this sentence hides an important distinction: randomness can be a property of the physical process, or it can be a property of our model of incomplete knowledge.

For example, thermal noise in a resistor is caused by microscopic charge motion. At the macroscopic level, we cannot track every carrier, so we describe the voltage fluctuations statistically. The classic Johnson-Nyquist result gives the mean-square voltage noise over a bandwidth $\Delta f$ as

$$ \langle V^2 \rangle = 4 k_B T R \Delta f, $$

where $k_B$ is Boltzmann’s constant, $T$ is temperature, $R$ is resistance, and $\Delta f$ is the measurement bandwidth [Johnson 1928; Nyquist 1928]. This is random in the operational sense: the exact voltage fluctuation at a particular instant is not predictable from the coarse variables $T$, $R$, and $\Delta f$.

But not every unwanted disturbance is random in the same way. Suppose a laboratory power supply leaks a 60 Hz or 50 Hz sinusoidal signal into a detector. The unwanted term might look like

$$ d(t) = A \sin(2\pi f t + \phi), $$

where $A$ is amplitude, $f$ is the power-line frequency, and $\phi$ is phase. If $A$, $f$, and $\phi$ are stable, this is not random. It is a deterministic interference. Experimenters may casually say, “There is line noise in the data,” but conceptually it is closer to a reproducible disturbance or systematic contamination than to ideal random noise.

There are intermediate cases. The amplitude or phase of that sinusoidal disturbance may drift unpredictably. A vibration from a pump may be deterministic over short intervals but irregular over long intervals. A chaotic system may be deterministic in its equations but practically unpredictable because small uncertainties in initial conditions grow rapidly. Such effects are often treated statistically, not because they are necessarily metaphysically random, but because the statistical model is the best usable description.

So the parent document’s definition is best read as a foundational distinction, not as a complete taxonomy. It is teaching the reader to separate three roles:

• the effect sought: signal;
• unpredictable fluctuation: noise;
• reproducible bias or distortion: systematic error.

In real experiments, the boundaries can be messy.

Random does not mean simple

Another common misunderstanding is that “random noise” means independent, Gaussian, and featureless. That is not true.

Some noise is approximately white noise, meaning its power is roughly equal over a range of frequencies. Some is colored noise, meaning low or high frequencies dominate. A famous example is $1/f$ noise, whose power spectral density scales roughly like

$$ S(f) \propto \frac{1}{f^\alpha}, $$

with $\alpha$ often near 1. Such noise is random, but it has correlations across time. If a detector’s noise today is statistically related to its noise a minute from now, then repeated samples are not independent. Averaging them does not reduce uncertainty as quickly as the simple $1/\sqrt{N}$ rule would suggest.

This matters for discovery. A small excess in a spectrum, a faint periodicity, or a weak correlation may look like a signal if the noise model is too simple. In precision physics, the question is not merely “Is there noise?” but “What kind of stochastic process describes the noise, over what bandwidth, under what environmental conditions, and with what correlations?” Standard treatments of stochastic processes and data analysis emphasize this distinction [Gardiner 2009; Bevington and Robinson 2003].

Can noise be nonrandom?

If we are being strict: nonrandom noise is a contradiction in the narrow statistical definition. A nonrandom unwanted contribution is better called interference, background, drift, bias, or systematic error.

If we are describing actual laboratory speech: yes, physicists sometimes call nonrandom unwanted effects “noise.” They may speak of “acoustic noise,” “electrical noise,” or “mechanical noise” even when part of the disturbance is deterministic. This is practical language, not a precise classification.

The distinction matters because different remedies apply. Random noise may be reduced by averaging, filtering, increasing signal strength, narrowing bandwidth, cooling an apparatus, or improving detector sensitivity. A systematic effect requires calibration, redesign, independent cross-checks, environmental control, or a better model of the apparatus. Treating a systematic error as random noise can produce false confidence.

This is exactly the danger the parent document is pointing toward. If an effect “averages away,” it behaves like noise. If it remains and shifts the result, it behaves like a systematic error. If it mimics the expected signature of a new phenomenon, it can produce a false discovery.

Does quantum randomness make the “true” model of noise?

Quantum mechanics introduces forms of randomness that are not merely ignorance about complicated classical variables, at least according to the standard interpretation and the empirical constraints from Bell-type experiments. In quantum theory, measurement outcomes are governed by probabilities given by the Born rule. For a state $|\psi\rangle$ and a measurement outcome associated with a projector $P$, the probability is

$$ p = \langle \psi | P | \psi \rangle . $$

This probability is not usually interpreted as ordinary ignorance about a pre-existing classical value. Bell’s theorem and later experiments show that broad classes of local hidden-variable explanations cannot reproduce the predictions of quantum mechanics [Bell 1964; Aspect, Dalibard, and Roger 1982].

Many important noise sources have quantum aspects. Photon counting has shot noise because photons arrive as discrete quanta. Vacuum fluctuations matter in quantum optics. Radiation pressure noise and photon shot noise appear in precision interferometry. At sufficiently low temperatures or high frequencies, even electrical noise requires quantum corrections to the classical Johnson-Nyquist formula [Callen and Welton 1951; Clerk et al. 2010].

But it would be wrong to conclude that quantum randomness is the correct true model for all noise. The right model depends on the physical source and the scale of description. Thermal noise in ordinary electronics may often be modeled classically with excellent accuracy. Wind shaking a telescope, seismic vibration, digitization error, amplifier drift, cosmic-ray hits, biological variability, and turbulent flow may each require different models. Some are fundamentally quantum at the deepest microscopic level, but that does not mean a quantum model is the most useful or correct effective model for the experiment.

Physics often works through effective descriptions. A gas is made of quantum particles, but classical thermodynamics can still correctly describe pressure and temperature in many regimes. Similarly, a detector is made of quantum matter, but its laboratory noise may be best modeled by a classical stochastic process over the relevant bandwidth. A deeper microscopic model is not automatically a better model for a particular inference.

The useful lesson for experimental discovery

The highlighted passage is scientifically sound as an introductory statement: noise is random variation that can hide a signal. But for research-level judgment, it needs two refinements.

First, randomness is often a modeling claim. To call something noise is to say that, for the purpose of the analysis, its detailed fluctuations are not predictable and must be described statistically. That model may be excellent, approximate, or dangerously oversimplified.

Second, not all unwanted variation should be treated as noise. Reproducible offsets, drifting calibrations, environmental couplings, analysis choices, and detector artifacts can imitate signal. These are systematic effects, and they are often harder to defeat than random noise because more data alone may not expose them.

So the mature version of the sentence is: noise is the unpredictable component of a measurement that can obscure the signal, while systematic error is a reproducible or structured distortion that can imitate or bias the signal. Quantum randomness is one profound source of noise, but not the universal explanation of all noisy data.

References

Aspect, A., Dalibard, J., & Roger, G. (1982). Experimental test of Bell’s inequalities using time-varying analyzers. Physical Review Letters, 49(25), 1804–1807. https://doi.org/10.1103/PhysRevLett.49.1804

Bell, J. S. (1964). On the Einstein Podolsky Rosen paradox. Physics, 1, 195–200. https://doi.org/10.1103/PhysicsPhysiqueFizika.1.195

Bevington, P. R., & Robinson, D. K. (2003). Data Reduction and Error Analysis for the Physical Sciences (3rd ed.). McGraw-Hill.

Callen, H. B., & Welton, T. A. (1951). Irreversibility and generalized noise. Physical Review, 83(1), 34–40. https://doi.org/10.1103/PhysRev.83.34

Clerk, A. A., Devoret, M. H., Girvin, S. M., Marquardt, F., & Schoelkopf, R. J. (2010). Introduction to quantum noise, measurement, and amplification. Reviews of Modern Physics, 82(2), 1155–1208. https://doi.org/10.1103/RevModPhys.82.1155

Gardiner, C. W. (2009). Stochastic Methods: A Handbook for the Natural and Social Sciences (4th ed.). Springer.

Johnson, J. B. (1928). Thermal agitation of electricity in conductors. Physical Review, 32(1), 97–109. https://doi.org/10.1103/PhysRev.32.97

Nyquist, H. (1928). Thermal agitation of electric charge in conductors. Physical Review, 32(1), 110–113. https://doi.org/10.1103/PhysRev.32.110

Taylor, J. R. (1997). An Introduction to Error Analysis: The Study of Uncertainties in Physical Measurements (2nd ed.). University Science Books.