Scientific computing · data analysis · interactive signal lab

Noise: The Thing Between You and a Discovery

The signal is rarely missing. It is usually buried under enough nonsense to make a patient scientist question their life choices.

Welcome to data analysis: where the universe whispers, your detector coughs, and your smoothing window quietly decides whether you become famous or wrong.
2026-05-08 · Hertford, London, UK 18 min read Interactive signal/noise lab
Noise and signal recovery visual
Signal recovery is not magic. It is statistics, assumptions, and occasionally resisting the urge to overfit the universe.

Signal Recovery Lab

Change noise, smoothing, and signal frequency. Watch how quickly a real pattern can become either obvious or suspiciously invented.

Detection Feed

Estimated SNR--
Recovered amplitude--
Detection mood--
Data · uncertainty · discovery

Noise is not the enemy. Unexamined noise is.

Every measurement arrives carrying baggage. Some of it is random. Some of it is instrumental. Some of it is the universe being unhelpful. The scientific task is not to remove everything inconvenient; it is to understand what remains.

A good detection is not a pretty bump on a graph. It is a bump with a story: where the photons came from, how the detector behaved, what the calibration did, what alternatives were tested, and why the same feature should still exist after someone else pokes it with a different pipeline.

The universe is allowed to be subtle. Your analysis is not allowed to be vague.
Chapter 01 · Signal

The signal is the story you are trying to hear.

A signal is any repeatable or physically meaningful structure in the data. In astronomy, it might be a spectral line, a transit dip, a radial-velocity wobble, a pulsar timing pattern, a gravitational-wave chirp, or a faint source in an image.

The important word is not only "repeatable"; it is also "physically meaningful." A signal has a shape that should make sense before the plot seduces you. A planet transit should have a duration, depth, ingress, and egress that fit an orbit. A spectral line should land at a plausible wavelength after redshift and calibration. A pulsar should keep time with a precision that embarrasses ordinary clocks.

Simple measurement model
\[y(t)=s(t)+n(t)\]

The measured data \(y(t)\) contain the true signal \(s(t)\) plus noise \(n(t)\). Real life adds systematics, calibration drift, sampling gaps, detector quirks, and deadlines.

A more honest version writes the model as a set of parameters. The data are not just "signal plus fuzz"; they are a measurement of something through an instrument.

Model plus residuals
\[y_i=m(x_i;\theta)+\epsilon_i\]

Here \(m(x_i;\theta)\) is a model evaluated at observation \(x_i\), \(\theta\) is the parameter set, and \(\epsilon_i\) is what remains. The residuals are where confidence grows teeth or politely collapses.

Transit signalA planet blocks a fraction of starlight. The useful structure is not just a dip, but a repeatable dip with the right geometry.
Spectral signalAn atom or molecule removes light at specific wavelengths. The pattern of lines matters more than a single tempting spike.
Timing signalPulsars, eclipses, and transits encode physics in when events arrive, not only how bright they are.
Image signalA faint source must beat the background, the point-spread function, detector artifacts, and the human urge to see shapes.
Chapter 02 · Noise

Noise is everything that makes the plot less obedient.

Noise can come from photon statistics, electronics, background light, imperfect calibration, atmospheric effects, detector temperature, readout patterns, cosmic rays, sampling choices, or the star itself. Treating all noise as simple random scatter is how you accidentally write fiction with error bars.

The friendliest kind of noise is independent and random. If you repeat the measurement, it tends to average down. The mean becomes more stable, the uncertainty shrinks, and the universe starts sounding less like static. The less friendly kind is correlated or systematic: it has memory, structure, direction, and a suspicious talent for looking like the signal you wanted.

Photon counting noise
\[\sigma_N \simeq \sqrt{N}, \qquad \mathrm{SNR}\simeq \frac{N}{\sqrt{N}}=\sqrt{N}\]

If you count \(N\) photons, the Poisson uncertainty is roughly \(\sqrt{N}\). That is why collecting four times as many photons improves SNR by about a factor of two, not four. Nature charges interest.

Random noiseScatter that averages down when independent. Lovely when it behaves.
Systematic noiseStructured errors. The plot looks meaningful because the instrument is lying consistently.
Correlated noiseNoise with memory. Nearby points move together, so pretending every point is independent inflates confidence.
Astrophysical noiseStars rotate, flare, pulse, granulate, spot, and generally refuse to be clean light bulbs.

In precision astronomy, noise is not a single villain. It is a budget. Shot noise, read noise, dark current, sky background, flat-field errors, and calibration residuals all get a line item. If you do not know which term dominates, you do not yet know what improvement would actually help.

Chapter 03 · Smoothing

Smoothing is useful. It is also how people accidentally invent planets.

A smoothing window can reveal broad trends by suppressing high-frequency noise. But too much smoothing can erase real features, shift peaks, broaden dips, or create the illusion of structure. A beautiful curve is not automatically a truthful curve. Annoying, but important.

Smoothing is a filter. It decides which scales in the data are allowed to survive. That is not morally bad; every instrument has a resolution limit anyway. The danger begins when the smoothing scale is chosen after staring at the plot until it looks convincing.

Moving average filter
\[\bar{y}_i=\frac{1}{2k+1}\sum_{j=-k}^{k}y_{i+j}\]

A moving average replaces each point with nearby points. It reduces jaggedness, but it also makes adjacent points correlated. The smoothed curve has fewer independent pieces of information than it visually appears to have.

Signal-to-noise ratio
\[\mathrm{SNR}=\frac{\mu_{\mathrm{signal}}}{\sigma_{\mathrm{noise}}}\]

In this simplified form, the signal strength is compared to the spread of the noise. Different fields define SNR in more specific ways depending on the measurement, the model, and the statistics of the residuals.

Under-smoothingThe real trend is still buried. You preserve detail, but the human eye has to fight the noise field.
Over-smoothingThe curve becomes elegant and possibly wrong. Small real features can vanish while broad fake ones appear.
ResolutionA spectrograph, telescope, or cadence already smooths reality before your code touches the data.
Pre-registrationChoosing the smoothing rule before looking at the result is boring in exactly the way good science often is.
Chapter 04 · False positives

A false positive is a discovery wearing a fake moustache.

The danger is not only missing real signals. It is believing in signals that are not there. That is why independent checks, null tests, injection-recovery tests, and physical plausibility matter.

If the detection disappears when you change a reasonable analysis choice, it was probably not the universe revealing itself. It was your pipeline doing interpretive dance.

False positives are especially good at hiding inside large searches. If you inspect one light curve, one spectrum, or one image, a rare fluctuation is rare. If you inspect a million windows, channels, targets, trial periods, and parameter combinations, a rare fluctuation starts acting like it paid rent.

Look-elsewhere effect
\[P(\mathrm{one\ or\ more})=1-(1-p)^M \approx Mp \quad \mathrm{for}\ Mp\ll1\]

If one test has false-alarm probability \(p\), then \(M\) independent trials raise the chance that at least one of them produces a tempting accident. Big searches need stricter standards.

Null testRun the same analysis where no signal should exist. If it finds one anyway, the pipeline is confessing.
Injection recoveryAdd fake signals of known strength, run the pipeline blindly, and measure what it can actually recover.
Control channelCheck detector regions, comparison stars, off-line spectral bands, or scrambled time order.
Independent pipelineA result is stronger when a different analysis path lands on the same physics.
Chapter 05 · Thresholds

A threshold is not a magic wand. It is a negotiated truce with uncertainty.

Scientists love thresholds because they turn messy evidence into a sentence: detected or not detected. But the threshold is not the discovery. It is a rule for deciding how much risk you are willing to tolerate.

A high threshold reduces false alarms but misses faint real signals. A low threshold catches more real signals but invites more impostors. The right choice depends on the cost of being wrong, the size of the search, the prior plausibility, and whether follow-up observations are possible.

A common Gaussian intuition
\[z=\frac{x-\mu}{\sigma}\]

The \(z\)-score says how many standard deviations a measurement \(x\) is from an expected mean \(\mu\). It is useful only when the noise model is honest. Non-Gaussian tails, correlated residuals, or underestimated uncertainties can make a glamorous \(z\)-score very fragile.

This is why astronomy papers often separate "candidate" from "confirmed." A candidate is interesting enough to deserve attention. A confirmed result has survived enough independent pressure that alternative explanations become less convincing.

Chapter 06 · Discovery workflow

A discovery should survive being treated with suspicion.

The mature version of data analysis is not cynicism. It is structured doubt. You want to give a real signal every fair chance to reveal itself while giving every fake signal several excellent opportunities to embarrass you privately before publication.

1. CalibrateSubtract bias, dark current, sky, background, flat-field structure, and known instrumental patterns before asking cosmic questions.
2. ModelWrite down the signal shape you expect and the noise process you are assuming. Hidden assumptions still count.
3. ResidualsAfter fitting, inspect what is left. Residual structure is often the first clue that the model is too confident.
4. Stress testChange reasonable choices: aperture, binning, baseline, smoothing, priors, comparison sample, and time cuts.
5. InjectAdd fake signals. If the pipeline cannot recover known truth, it has not earned the right to declare unknown truth.
6. ReplicateUse another night, another instrument, another wavelength, another team, or another method whenever possible.

That workflow is slower than a dramatic claim. It is also how a dramatic claim becomes durable. The goal is not to make every result boring; the goal is to make the excitement expensive enough that only the sturdy signals can afford it.

Takeaway

Good science is not noise-free. It is noise-aware.

Data analysis is the discipline of being excited and suspicious at the same time. That tension is not a weakness. It is the reason discoveries survive contact with reality.

The cleanest plots are not always the most honest plots. The honest plot tells you what was measured, what was assumed, how uncertainty was handled, and what would make the claim fail. That is the kind of plot a discovery can stand on.

A discovery is not the moment the signal appears. It is the moment the signal keeps appearing after your best attempts to make it go away.