Astronomy · Finding change in the sky

Artifacts, classification and uncertain events

Distinguish a reliable change from its desired physical class, and use a declared parent population and flagging model to calculate the class probability of a uniformly chosen flagged item.

Continue in the approved chapter order. Necessary vocabulary, quantities and assumptions are explained in the lesson.

Back to the chapter

After reading, you can

Separate artifacts, real but irrelevant sources and uncertain desired events.
Reverse a conditional flagging probability using the supplied base rates.
Explain why reliability scores, class scores, p-values and calibrated probabilities answer different questions.

Real can still mean outside the chosen class

A search needs a target definition. A moving Solar System object or an ordinary variable star can be astrophysically real while outside a particular transient class. Calling it a false positive for that search does not imply that its detected photons or physical source are fictitious.

Instrumental and processing effects can also create candidate changes. Particle hits on a detector, bad pixels, internal reflections, unmatched image blur, registration errors and changing calibration can leave features that need to be distinguished from the desired signal.

Shapes, position, repetition and the science/reference context offer diagnostic evidence, rather than an infallible checklist. A real source can be blended or poorly sampled, while a recurrent artifact can survive several similar observations.

The image-processing papers discuss models and quality checks for these alternatives. Their historical system performance is not a universal reliability claim. A candidate should retain competing interpretations when the available data do not distinguish them.

Ask which conditional question a number answers

A detection-reliability score can help rank apparently real changes against processing artifacts. A physical-class score addresses membership in a source category. A period-search false-alarm probability is calibrated under a specified noise null and search. Those numbers do not answer the same question.

A score between zero and one is not automatically a calibrated posterior probability for the current candidate population. Calibration asks whether stated probabilities agree with outcomes in an appropriately defined and selected population under the relevant conditions.

Training samples, observing conditions, calibration and reference versions can change how a classifier behaves. Masci and collaborators explicitly discuss re-evaluating learned reliability when instrumental signatures or astrophysical environments differ from the training material.

The following original model supplies both prevalence and conditional flag rates. It demonstrates the reversal of a probability under those assumptions, rather than reporting measured classifier accuracy or recommending a real selection threshold.

A high target flag rate does not determine the flagged mixture

Stipulate a parent population of 1000 items containing 50 desired-class events and 950 other items. The other class may include artifacts and real sources outside the target. Assume a desired item is flagged with probability 4/5 and an other item with probability 1/50.

Expected desired flags are 50 × 4/5 = 40. Expected other flags are 950 × 1/50 = 19. The expected total is 59, so the ratio of expected desired flags to expected total flags is 40/59 ≈ 0.677966102.

Now define the probability question precisely: choose one parent item uniformly first, then condition on its being flagged under this model. P(desired and flagged) = 0.05 × 0.80 = 0.04 and P(flagged) = 0.04 + 0.95 × 0.02 = 0.059.

Bayes’ rule gives P(desired | flagged) = 0.04/0.059 = 40/59. This equality belongs to the stated uniform-item probability model. It does not assert that every realized flagged catalogue has that proportion or that the expectation of a random catalogue ratio equals a ratio of expectations.

Original parent and flagging model. The desired class and other class partition 1000 supplied items. Flagging rates are invented conditional probabilities, not measured classifier performance. Flag counts below are expectations.
Parent class	Parent count	Conditional flag probability	Expected flags
Desired event	50	4/5 = 0.80	40
Other item	950	1/50 = 0.02	19
Total	1000	Marginal 0.059	59

Eighty percent and forty over fifty-nine are different answers

The supplied 0.80 is the chance of a flag given that an item is desired. The supplied 40/59 is the chance that a uniformly chosen parent item is desired after conditioning on its flag. Reversing the condition requires the parent composition and the flag probability for the other class.

The large other population contributes 19 expected flags even with its smaller conditional rate. A low contamination rate per other item can therefore matter when those items greatly outnumber the target. The toy’s result would change if its prevalence or conditional flagging model changed.

The posterior is also only as specific as the class definition. If desired means one broad candidate family, 40/59 does not distinguish every physical mechanism inside that family. A reliability score about artifacts cannot automatically be promoted into confidence in one stellar process.

The calculation supplies no universal threshold or actual source prevalence. Applying a similar inference to real data requires justified class boundaries, a relevant population, tested conditional performance and an uncertainty treatment.

A classification can be provisional and still informative

A useful candidate record identifies the current hypotheses, the observations supporting them, the quality and version of relevant scores, and the evidence still missing. An uncertain category can guide further work without being presented as a confirmed cause.

Additional bands, spectra, motion information and independent measurements can test different alternatives. Repeated use of the same detector, reference or calibration can repeat a failure mode as well as a true signal, so dependencies remain part of the interpretation.

A classification update should retain its evidential basis and processing history. New conditions or populations can change the applicability of a trained score. More significant detection evidence does not by itself resolve every ambiguity in the physical model.

The next chapter investigates variable stars. It will connect pulsation, eclipses, rotation and eruptions to measurements while preserving the difference between a recurring light-curve pattern and a uniquely identified mechanism.

Worked examples

Reverse the flagging condition explicitly

For a uniformly chosen parent item conditioned on being flagged, what is the desired-class probability?

The desired-and-flagged probability is (50/1000)(4/5) = 0.04.
The total flagged probability is 0.04 + (950/1000)(1/50) = 0.059.
Divide 0.04 by 0.059 to obtain 40/59 ≈ 0.677966102.

It is 40/59 under the supplied model, rather than the desired-class flag rate 4/5.

Include the numerous other items

Why do 19 expected other flags matter despite the small other-item flag probability?

There are 950 other items in the supplied parent.
Their conditional rate 1/50 produces 950/50 = 19 expected flags.
Those join 40 expected desired flags, changing the flagged mixture from the desired-only view.

Base rates matter: the larger other population contributes substantially to the flagged sample even at a lower per-item rate.

Try the reasoning

Can a candidate be astrophysically real while remaining a false positive for a narrowly defined desired class?
Show the answer and reasoning
Yes.
The search target and existence of a real variable or moving source are different questions.
Does P(flag | desired) alone determine P(desired | flag)?
Show the answer and reasoning
No.
The parent prevalence and flagging of other classes are also needed.
Is every score between zero and one automatically a calibrated class probability?
Show the answer and reasoning
No.
Its definition, training and calibration on the relevant selected population must be established.
Must a particular realized flagged catalogue contain exactly 40/59 desired events?
Show the answer and reasoning
No.
The counts are expectations, and the posterior is defined for the stated uniform-parent-item conditioning model.

What to carry forward

Artifact reliability and desired physical classification are separate inferences.
Reversing a conditional flag probability requires base rates and the other-class response.
Scores, null tests and provisional classes retain their population, version and evidence limits.

References

Teaching resources and research supporting this lesson’s explanations and examples.