Cosmology · Cosmic maps, uncertainty and fair comparisons

Signal, foreground and instrumental noise

Separate two chosen frequency channels algebraically to see how a foreground can cancel while measurement noise remains.

Preparation: ratios, powers of ten and ordinary arithmetic. Physical quantities and notation are introduced as they are needed.

Back to the chapter

After reading, you can

Distinguish a desired signal, foreground emission and instrumental noise.
Explain how frequency information can help separate components.
Calculate a simple separation residual and state what the model assumes.

A foreground is defined relative to a question

A telescope looking for the cosmic microwave background also receives radiation from our Galaxy and other sources. For that question, Galactic emission is a foreground that must be accounted for. For a researcher studying Galactic dust, the same emission may be the desired signal. Foreground does not mean unreal or scientifically useless.

Instrumental noise is different: it describes unwanted fluctuations or uncertainty introduced by measurement, rather than another astrophysical component along the sightline. Calibration errors and imperfect instrument-response modelling can supply further effects. A single colour in a displayed map does not tell us how much each contributed.

An observed frequency channel combines those contributions through its instrument response. Different astrophysical components can change differently with observing frequency, which gives a way to distinguish them. This requires comparable units and a model of those responses; a set of differently coloured images is not automatically a component separation.

Real analyses also account for angular response, calibration and the variation of emission properties across the sky. The small example here deliberately leaves those complications out so that the logic of one cancellation can be followed exactly.

Choose two channels with known responses

Let s be a desired signal and f a foreground amplitude, both in one shared arbitrary signal unit. Choose two calibrated channel values d₁ = s + f + n₁ and d₂ = s + 2f + n₂. The noise offsets are n₁ and n₂. The dimensionless factor 2 is supplied as part of this teaching model; it is not a measured dust spectrum.

The signal has equal response in the two channels, while the foreground response doubles in the second. Multiply the first channel by 2 and subtract the second: 2d₁ − d₂ = 2s + 2f + 2n₁ − s − 2f − n₂ = s + 2n₁ − n₂.

The foreground terms cancel because their specified responses were known and different from the signal’s. Without noise, this combination would recover s exactly. With noise, the combination retains a residual 2n₁ − n₂.

If the foreground instead had the same response as the signal in both channels, these two numbers would not distinguish them by this method. Separation needs information that makes the components distinguishable; more channels with identical information do not solve that ambiguity.

Cleaning does not mean removing every error

Choose s = 2, f = 3, n₁ = 0.1 and n₂ = −0.2. Then d₁ = 2 + 3 + 0.1 = 5.1 and d₂ = 2 + 6 − 0.2 = 7.8. The separated value is 2 × 5.1 − 7.8 = 2.4.

The result differs from the chosen true signal 2 by 0.4. That residual equals 2 × 0.1 − (−0.2) = 0.4, just as the algebra predicted. The foreground cancellation is exact in this supplied model, yet the recovered signal is still not exact.

The particular noise offsets were chosen numbers for one example. They do not estimate a noise distribution by themselves. If a separate model specifies independent noise with standard uncertainties 0.2 in each channel, the combination’s standard uncertainty is √(4 × 0.2² + 0.2²) = √5 × 0.2 ≈ 0.447 in the shared unit.

The factor 4 comes from squaring the coefficient 2. This calculation assumes the specified noises are uncorrelated. Correlation would add a covariance term; uncertainties in the response coefficients or calibration would need their own treatment. A clean-looking map cannot remove these assumptions.

One chosen two-channel example in shared arbitrary signal units
Quantity	Value	Role
s	2	Desired signal
f	3	Foreground amplitude
n₁, n₂	0.1, −0.2	Chosen noise offsets
d₁, d₂	5.1, 7.8	Observed-channel values in this toy
2d₁ − d₂	2.4	Separated value
Separated value minus s	0.4	Residual for these chosen offsets

Two chosen channels have matched arbitrary signal units and exact responses d₁=s+f+n₁ and d₂=s+2f+n₂. With s=2, f=3, n₁=+0.1 and n₂=−0.2, the first stacked sum is 5.1. The second has signal 2 plus foreground 6 before a negative offset −0.2, marked by the hatched end strip and leftward arrow from 8 to 7.8. The coloured positive contributions sum to 8 before that subtraction; the reported second channel value is 7.8. On the same 0–8 linear contribution scale, the lower combination 2d₁−d₂ equals 2.4: desired signal 2 plus residual 2n₁−n₂=+0.4. Direct labels accompany colour. The particular noise offsets are separate from a separately specified zero-covariance noise model with standard uncertainties u₁=u₂=0.2, for which the separated standard uncertainty is √5×0.2≈0.447. No actual dust spectrum, map measurement or response fit is represented. Original Ark teaching model.

Real component maps carry processing decisions

A real analysis may exclude directions where the foreground or its modelling uncertainty is too large. The resulting mask belongs to the analysis, not to the physical boundary of the background. A masked pixel has not been measured to contain no cosmic signal.

Different component-separation methods can combine channels using different assumptions. Comparing their outputs can reveal sensitivity to those choices, but agreement is not proof that every systematic effect is absent. Methods using the same input maps may share calibration or modelling limitations.

The historical Planck 2018 component-separation analysis explicitly discusses its data model, masks, noise and foreground residuals. It provides an example of this documented process, rather than a claim that the sky has been reduced to a perfectly isolated component.

An inpainted presentation region contains a model-based replacement, not a new direct observation. Distinguish the information used for a scientific calculation from the processing used to display a continuous image. A colourful full-sky picture alone does not establish which regions entered a parameter fit.

Ask which ambiguity the new information resolves

In the example, different frequency responses supplied the leverage to cancel the foreground. Better instrument calibration would address another uncertainty. More independent measurements could reduce some noise. None of these is interchangeable with simply changing a colour scale.

For a first reading, follow the arithmetic from the two channel sums to 2.4. That single residual explains why signal extraction is an inference with uncertainty. Later analyses extend this idea to many components, imperfectly known spectra and correlated maps.

Worked reasoning

Recover the signal plus its residual

With d₁ = 5.1 and d₂ = 7.8 in the stated model, what does 2d₁ − d₂ give? How far is it from the chosen signal s = 2?

Twice the first channel is 10.2. Subtract the second: 10.2 − 7.8 = 2.4.
Subtract the chosen signal: 2.4 − 2 = 0.4.
The algebraic residual is 2n₁ − n₂ = 0.2 + 0.2 = 0.4. Foreground cancellation has not cancelled the noise.

The separated value is 2.4, with residual +0.4 for these chosen offsets.

Try the reasoning

If n₁ = n₂ = 0 while the stated responses remain exact, what is 2d₁ − d₂?
Show the answer and reasoning
s.
The terms 2f cancel and the signal terms leave 2s − s = s.
Would the same two-channel trick distinguish signal and foreground if their channel responses were identical?
Show the answer and reasoning
No.
The two components would enter these channels in the same proportion; extra discriminating information would be needed.
Is Galactic dust emission always unwanted noise?
Show the answer and reasoning
No. Its role depends on the scientific target.
It is an astrophysical foreground for some background studies and a desired signal for studies of the Galaxy.

What to carry forward

Signal and foreground are roles in a scientific question; instrumental noise is a different contribution.
Distinct known responses can make component separation possible.
A separated signal retains noise and modelling uncertainty even when one foreground term cancels.

References

Sources supporting the explanations, physical distinctions and examples in this lesson.