SCIENCE
Kitt Peak Spectrograph Pipelines Reject Faint Planet Signals
Kitt Peak National Observatory hosts more than twenty optical telescopes on a mountain in the Quinlan Mountains of the Tohono O'odham Nation, about 88 kilometers west-southwest of Tucson. Some of those instruments hunt planets by watching stars wobble. The wobble is tiny, and the software that turns raw spectra into radial velocities routinely throws it away. What follows is a breakdown of where in the pipeline that happens, and what a researcher can do about it.
Faint Signals Lost in Software
The radial velocity method detects a planet by measuring the line-of-sight motion of its host star. A star with a close-in giant planet might move by tens of meters per second; a star with a small planet moves by a meter per second or less. The signal is a periodic shift in the star's spectrum, and it is buried under photon noise, telluric absorption, and instrumental drift.
Pipelines are built to survive that noise. They flag bad pixels, reject cosmic ray hits, and discard exposures whose signal-to-noise ratio falls below a threshold. Those choices are defensible on their own terms. They also mean that a genuine 0.5 m/s wobble can look, to the software, exactly like the kind of junk the software was written to remove.
The result is a selection effect that operates before any astronomer looks at a plot. Faint planet signals go missing not because the telescope failed, but because the reduction code decided they were not worth keeping.
How the Pipeline Actually Works
A spectrograph splits starlight into wavelength channels and records the pattern of absorption lines. A shift in those lines toward the blue or red end of the spectrum corresponds to motion toward or away from the observer. Cross-correlation with a template turns that shift into a radial velocity number, usually with an uncertainty attached.
Before that number exists, the pipeline does a lot of housekeeping. It subtracts a bias frame, divides by a flat field, identifies the spectral orders, and fits a wavelength solution using a calibration lamp or a laser frequency comb. Each step can introduce or remove a few tenths of a meter per second of apparent motion.
Outlier rejection sits at the end of this chain. The pipeline compares each measurement to a running model of the star and discards points that deviate too far. The threshold is usually set in units of the expected scatter. A point three standard deviations from the model is a candidate for deletion.
Those thresholds are tuned on bright, stable stars, because that is where the calibration data come from. A faint star with a small companion produces a different statistical profile, and the same threshold behaves differently on it.
The Cost of Cleaning Data
Aggressive cuts do real work. They remove cosmic ray hits, bad columns, and nights when the spectrograph drifted. Without them, a single bad exposure can dominate a fit and produce a spurious detection. The astrophysical false positive rate drops when the pipeline is strict.
The problem is that faint planet signals and instrumental noise overlap in the same region of the data. A 0.5 m/s wobble and a 0.5 m/s systematic error are not distinguishable from a single measurement. Only the periodicity separates them, and periodicity is exactly what a per-point rejection scheme cannot see.
Rejection rates vary across spectrographs and across stars. A bright, quiet star might lose a few percent of its exposures; a faint, active star might lose a third. The lost points are not random. They cluster at the faint end, which is where small planets live.
Occurrence statistics inherit that bias. If the pipeline discards faint signals, the resulting catalog undercounts small planets, and the correction applied afterward is a model, not a measurement. This site has argued in a related piece on telescope time allocation that the infrastructure around detection shapes which claims get tested at all.
A Worked Example on Kitt Peak
Consider a hypothetical star observed on a Kitt Peak spectrograph. It has a planet that induces a 0.5 m/s wobble with a period of a few days. The star is faint enough that each exposure carries roughly 1 m/s of photon noise, so the signal is below the per-point uncertainty.
The pipeline fits a model to the velocity time series and rejects points more than three standard deviations from that model. With 1 m/s scatter and a 0.5 m/s signal, many of the planet's contribution falls inside the rejection boundary, but the tails do not. Some points that carry the wobble get cut, and the fit converges on a flatter, quieter solution.
Lower the threshold to two sigma and the signal survives. The cost is that more noise points survive too, and the false positive rate rises. The trade-off is not a failure of engineering. It is a choice about which kind of error the observer is willing to accept.
A related piece on this site, on consortia spending on data pipelines, makes a similar point in a different field: the processing layer absorbs resources and shapes conclusions long before anyone interprets a result.
What the Data Can and Cannot Say
Pipelines are not neutral filters. Every threshold encodes an assumption about what a real signal looks like and what noise looks like. When those assumptions are tuned on bright stars, they carry a brightness bias into the faint end of the sample.
Detection claims depend on processing choices. Two teams analyzing the same spectra with different rejection thresholds can reach different conclusions about whether a planet exists. Neither team is necessarily wrong; they have made different bets about the noise model.
There is no consensus on the optimal rejection threshold. A threshold low enough to recover every faint signal will also admit enough noise to make individual detections unreliable. A threshold high enough to keep the catalog clean will systematically remove the smallest planets.
The practical response is transparency. Report the rejection threshold, the cut criteria, and the fraction of exposures discarded. Without those numbers, a reader cannot tell whether a non-detection means the planet is absent or the pipeline removed it.
Practical Steps for Researchers
- Publish the rejection threshold and cut criteria alongside the radial velocity table, including the fraction of exposures discarded per star.
- Run at least one alternative pipeline on the same spectra and report whether the detection survives the change.
- Flag marginal detections explicitly, with the signal-to-noise ratio and the number of points near the rejection boundary.
- Share raw spectra, not just reduced velocities, so that other groups can reanalyze the data with their own thresholds.
- Model the noise sources explicitly, including telluric contamination and instrumental drift, rather than absorbing them into a single rejection cut.
The cost of doing this is real. Raw spectra are large, and alternative pipelines take time to run. A survey that publishes every threshold invites reanalysis that may contradict its own catalog. The alternative is a literature where faint planet signals disappear into a software decision that no reader can inspect.
Instrument-Specific Thresholds and Their Consequences
Different spectrographs on Kitt Peak employ different rejection strategies. The Mayall 4-meter telescope's DESI spectrograph, for instance, is designed for galaxy surveys and has a pipeline optimized for faint, extended sources; its rejection thresholds are tuned to that use case, not to the search for small planets around bright stars. The WIYN 3.5-meter telescope's NEID spectrograph, by contrast, is built for precision radial velocities and uses a laser frequency comb for calibration. Its pipeline is designed to preserve as much signal as possible, but it still has to reject bad exposures. The trade-off is that NEID's rejection thresholds are tighter, which means it may discard more data from active stars, potentially losing small planet signals.
The point is that thresholds are not universal. They are tuned to the instrument's primary science case. When a spectrograph is repurposed for a different kind of observation, the thresholds may no longer be appropriate. This is a form of instrumental bias that can shape survey results.
Researchers should be aware of this when comparing results across instruments. A non-detection from one spectrograph may simply reflect a stricter rejection threshold, not the absence of a planet. The only way to know is to report the thresholds and the discarded data fractions.
