A signal against a background

The record begins with a mixture

A faint star does not arrive at a camera on a perfectly empty stage. The same pixels may collect light from the sky, unresolved objects and scattered light, as well as the target. The detector also contributes electronic effects. The original image records their mixture. A measurement has to define which contribution answers the question.

Signal means the contribution of interest in a stated measurement. Background means a contribution we need to estimate or model to extract it. These names depend on the question. Diffuse light that is a background for a star measurement might itself be the signal in a study of a nebula.

Noise describes variability that limits the estimate. A background can have a nonzero average and fluctuating counts. An instrument can also have a bias or a patterned defect. Calling everything unwanted “noise” hides the difference between a removable average, a random fluctuation and a systematic error.

Estimate what lies underneath the target

In simple aperture photometry, we add the values in a defined region around a source. Here aperture means the region selected in the image, not the physical opening of the telescope. A neighbouring region without the target helps estimate the local background per pixel. An annulus is a ring sometimes used for this purpose.

The nearby region must represent the background under the target. A second star, a bright gradient or a defect in that region can spoil the estimate. Software can reject some contaminated pixels, but the observer still needs to inspect the field. Crowded or structured fields may require more advanced methods than a simple ring.

Use values with compatible units and calibration. Raw camera values are often digital units, whose conversion to detected electrons depends on the instrument. Do not label them photons without establishing that relationship. The following arithmetic uses explicitly defined, synthetic calibrated count units.

A simple subtraction, with every quantity defined

Let C be the sum of the calibrated counts in the target aperture, n the number of pixels in that aperture, and b the estimated background count per pixel. A simple estimated source total is S = C − n × b. The product n × b has count units, like C; subtracting b alone would mix a per-pixel quantity with a total.

In an invented exposure, C = 15,000 count units, n = 25 pixels and b = 400 count units per pixel. The estimated background within the aperture is 25 × 400 = 10,000 count units. The remaining source estimate is S = 15,000 − 10,000 = 5,000 count units.

This calculation assumes the background estimate applies across the aperture and relevant instrumental offsets have already been treated. It estimates the source contribution inside that chosen region. It is not automatically the source’s total emitted light, its calibrated flux at Earth, or even all the light in the star’s blurred image.

Subtracting the average leaves an uncertain estimate

The two synthetic exposures have the same estimated source total even though the raw aperture total changed. Their measurement quality need not be equal. A brighter background usually brings larger count fluctuations. Its estimated average can be subtracted, but the particular fluctuations that occurred underneath the target are not known exactly.

There is also uncertainty in b itself. If the estimated background in the first exposure were 410 rather than 400 count units per pixel, its source estimate would be 15,000 − 25 × 410 = 4,750. A difference of 10 per pixel became 250 across the aperture. Background choice matters especially when a faint source is extracted from a large total.

This is why a darkened image is not the same as a precise measurement. Moving a display slider can make the background look black while leaving the underlying data and their uncertainties unchanged. Save the original data, the chosen measuring regions and the processing steps.

Collecting more light helps under stated conditions

For an ideal independent photon-counting process, random counting fluctuations are often modelled with Poisson statistics: if the expected count is N, its standard deviation is √N. The fractional fluctuation is then 1/√N. This is a model of repeated counts, not a statement that every individual exposure differs from its expectation by exactly √N.

As a source-only illustration, an expected 100 detected events has a standard deviation of 10, or 10%; 10,000 has a standard deviation of 100, or 1%. The absolute fluctuation grew, but the fractional fluctuation shrank. Real measurements also include background, detector noise, calibration and atmospheric effects. Applying √N to the background-subtracted source alone can underestimate the uncertainty.

Longer exposures can collect more source light, but also more background and possibly smear a changing target. Saturation can destroy a detector’s simple proportional response. A wider measuring aperture can include more of a source’s spread-out image while collecting additional background or neighbouring light. More is useful only when the measurement remains suitable for the question.

A weak estimate can constrain a claim without proving absence

If the source estimate is comparable to its uncertainty, treat a detection claim cautiously. A subtraction can even return a negative value for a very weak source because the aperture and background estimates fluctuate. That is not negative physical light. It can be an ordinary outcome of an uncertain estimator, or a reason to investigate the background model.

A non-detection states that the chosen method did not reliably distinguish the source under the specified conditions. A useful upper limit requires a defined statistical procedure, sensitivity and assumptions. Do not invent a limit by replacing a negative estimate with zero or by taking the faintest pixel on a displayed image.

Before asking what a faint source is doing, ask whether it was identified correctly, whether the background was representative and how the uncertainty was estimated. The next lesson uses repeated measurements and comparison objects to test changes that a single extracted number cannot settle.

Worked example

Try the reasoning

1. Synthetic C = 12,000 count units, n = 40 pixels and b = 250 count units per pixel. Find S.

   Show the answer and reasoning

   2,000 count units.

   The aperture background is 40 × 250 = 10,000; subtract it from 12,000.

2. A display adjustment makes the background black. Has its noise disappeared?

   Show the answer and reasoning

   No. Display appearance does not remove uncertainty from the recorded measurement.

   The original counts and the background estimation still determine the measurement.

3. A very faint target gives a background-subtracted estimate slightly below zero. Did it emit negative light?

   Show the answer and reasoning

   No. The estimated difference can be negative because the measurement and background estimate fluctuate.

   Inspect the model and uncertainty; do not relabel an uncertain estimate as a physical negative brightness.

What to carry forward

• Define the signal and estimate the background in compatible units.
• Background subtraction removes an estimated contribution, not every fluctuation or systematic error.
• A detection, a non-detection and a statistically defined limit make different claims.