Astronomy · Planning observations around a question

Turning a question into measurable quantities

Translate a scientific question into predictions for an observing design, then compare repeating one degenerate measurement with adding a complementary band.

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 detecting a signal, estimating a parameter and distinguishing proposed models.
Choose an observable after accounting for free parameters and nuisance quantities.
Connect required precision, calibration and observing conditions to the scientific question.

Decide what a possible observation must distinguish

“Observe the source more carefully” does not yet define a scientific investigation. Is the aim to detect emission, estimate a period, measure a spectral ratio, distinguish two physical proposals, or test whether a prediction fails? Each goal can require different data even for the same object.

Begin with the quantity that would appear in the record: a count in a channel, a band-integrated flux, an angular separation, a timing difference or a series of positions. Specify its units, region, time window and resolution. Then state how the competing scientific interpretations predict it.

Detection asks whether the recorded signal meets a defined comparison with background. Estimation asks which parameter values are compatible with a model and data. Discrimination asks which proposed predictions the measurements can distinguish. A strong detection need not answer the other two questions.

The observation model must also include nuisance quantities such as an unknown overall amplitude, background, extinction or calibration scale. A proposed difference that disappears when those quantities are allowed to vary is not a difference the chosen observation can resolve.

More precision in the same quantity may leave the question unchanged

Consider two original abstract band templates, H1 = (1, 1) and H2 = (1, 2), in two declared disjoint bands. Each has its own freely fitted positive amplitude a. Its predicted band-integrated fluxes are aH. Express each flux numerically in the fixed unit F0 = 10⁻¹⁶ W m⁻². The templates are teaching proposals, not identified source materials or measured spectra.

Suppose only band 1 is measured and its numerical flux is 10. Both templates fit that value with amplitude a = 10. H1 then predicts 10 in band 2 and H2 predicts 20, but band 2 has not been observed.

Making the band-1 uncertainty smaller does not remove this exact degeneracy under the declared free-amplitude model: both proposals still predict the same available measurement after fitting their amplitudes. More repeated measurements can improve precision on that combination without distinguishing the templates.

This conclusion depends on the allowed nuisance freedom. An independent constraint on amplitude, another parameter relation or a different physical model could change it. Observational design must therefore state what is held fixed and what can adjust before claiming that one measurement will discriminate explanations.

A complementary observable can test the template shape

Now stipulate a synthetic two-band outcome y = (10, 20), with independent Gaussian measurement uncertainties of one numerical flux unit in each band, known calibration and no other noise or nuisance parameters. Fit a separate positive amplitude for each template by minimizing χ²(a) = Σi(yi − aHi)².

For these equal uncertainties, differentiating gives a = (Σi Hi yi)/(Σi Hi²). For H1 this is (10 + 20)/(1 + 1) = 15, yielding predictions (15, 15) and residuals (−5, +5). The minimum χ² is 25 + 25 = 50.

For H2 the fitted amplitude is (10 + 2 × 20)/(1 + 4) = 10. Its predictions are (10, 20), with both residuals zero and minimum χ² = 0. The added band distinguishes these specified template shapes under the supplied error model.

The difference of 50 is not a posterior probability that H2 is true, a universal significance level or identification of a unique physical cause. Other templates, uncertain cross-band calibration, correlated errors, source variability between exposures or an unknown extra component could change the comparison.

Original two-band outcome in numerical units F0 = 10⁻¹⁶ W m⁻². Each model has its own fitted amplitude; both band uncertainties are 1 F0 and independent.
Template	Best amplitude	Predicted band 1	Predicted band 2	Minimum χ²
H1 = (1, 1)	15	15	15	50
H2 = (1, 2)	10	10	20	0

An answerable question needs a feasible measurement

Having chosen the second band, work back through the response model. Is the source visible in that band and time? Does the instrument cover the required wavelength range? Are its resolution and field adequate to separate the source from neighbours or background? Can the calibration compare the two bands at the required precision?

If the source varies, sequential measurements can mix spectral shape with temporal change. The design may need simultaneous coverage, a justified variability model or repeated measurements that constrain the change. If the desired quantity is a ratio, its uncertainty must retain relevant shared calibration and background effects.

ESO’s observing guidance illustrates how conditions follow the measurement: sky brightness matters for faint signals, while image quality matters when a scientific question needs compact structure resolved. Its observing-block description also separates acquisition, science and possible attached calibration steps. Finding the source correctly and measuring the intended region are prerequisites to interpreting the science exposure.

These are planning principles, not a recommendation to buy equipment or a promise of telescope access. Actual modes, visibility, overhead and constraints must be checked in the documentation for the instrument and proposal being considered. A feasible exposure that measures the wrong quantity does not become useful merely by running longer.

State what each outcome would teach

An observing plan concerns possible future data, rather than a known future outcome. Forecast the expected signal and uncertainty under the relevant alternatives and nuisance ranges. Ask whether the difference survives realistic backgrounds, calibration uncertainty, observing gaps and the available exposure.

Loredo’s adaptive-design framework separates inference from the decision about what to observe next. Future measurements are predicted using the current model and uncertainty; their expected scientific value can then be compared under a declared objective and resource constraints. The preferred design remains conditional on those models and that objective.

A useful concise plan states the question, competing predictions, observable and units, allowed nuisance freedom, instrument response, precision target, controls and what each plausible outcome would imply. Include outcomes that would leave both proposals plausible or challenge both. Designing only around the currently favoured outcome can conceal a weak test.

The two-band example shows why an additional kind of measurement can be more useful than a more precise repetition of an uninformative one. The next lesson quantifies another part of the design: how integration time, detector readouts and observing overhead combine to limit precision.

Worked examples

Check whether improved precision answers the question

With only band 1 equal to 10, can reducing its uncertainty distinguish the free-amplitude templates (1, 1) and (1, 2)?

Both have a first template component of one.
Each can independently choose a = 10 and fit the first-band value exactly.
The missing second-band prediction is where the templates differ.

No, under this free-amplitude model. Greater precision in band 1 does not lift the exact degeneracy.

Use both measured bands

For y = (10, 20) with independent unit Gaussian uncertainties, what amplitude and minimum χ² follow for H2 = (1, 2)?

Use a = (Σ Hi yi)/(Σ Hi²).
Calculate a = (10 + 40)/(1 + 4) = 10.
The predicted pair (10, 20) equals the supplied synthetic outcome.

a = 10 and χ² = 0. This compares the declared template with these data; it does not establish that every aspect of a physical source model is correct.

Try the reasoning

Does detecting a source with high signal-to-noise guarantee discrimination between two models that predict the same measured signal?
Show the answer and reasoning
No.
Detection and model discrimination are different goals.
Why fit amplitudes separately for the two template alternatives?
Show the answer and reasoning
Each alternative was defined with its own unconstrained amplitude.
Holding one arbitrarily fixed would change the stated comparison.
What additional issue arises if a variable source is observed in the two bands at different times?
Show the answer and reasoning
Temporal variability can be confused with spectral shape.
The plan needs simultaneous coverage or a justified treatment of the change.
Is an observable useful solely because an instrument can measure it very precisely?
Show the answer and reasoning
No.
Its possible predictions must address the intended scientific question under the allowed nuisance freedom.

What to carry forward

Choose the observable by asking which possible predictions it can distinguish.
Account for nuisance parameters before promising that improved precision will resolve a question.
Connect the desired inference to calibration, timing, response and practical observing conditions.

References

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