Sky-background limit
A Socratic walk-through of the sky-background limit — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why does leaving the shutter open twice as long not let a telescope reach twice as faint?
The naive arithmetic is irresistible. A star half as bright delivers half the photons per second, so leave the shutter open twice as long and you have collected the same number as before. The detector cannot tell the difference. Doubling the exposure should therefore double the faintness you can reach, and a night's patience should buy a factor of thousands.
It does not. Doubling the exposure buys about four-tenths of a magnitude — a factor of roughly 1.4 in flux, not 2. To reach genuinely twice as faint takes four times the time, and two magnitudes takes forty. Something is being collected alongside the star, and it is growing too.
Reasoning it through
REASONING #Ask what lands in the little patch of detector where the star's image sits. Three things: photons from the star, photons from the sky itself, and whatever noise the electronics contribute at readout. The first two accumulate with time; the third does not.
Now the step the naive argument skips. Photons arrive at random and independently, so the count is Poisson-distributed, and a Poisson count of N has standard deviation sqrt(N). The sky is therefore not merely a pedestal to be subtracted — its mean certainly can be, and every observer does subtract it. What cannot be subtracted is its graininess, sqrt of however many sky photons you collected.
Write it out. Let the star deliver f counts per second into the aperture and the sky b counts per second into the same aperture. After time t:
SNR = f·t / sqrt(f·t + b·t + read noise terms)
For a faint source, b·t dwarfs f·t, and if the exposure is long enough the read-noise term is negligible too. Then the expression collapses:
SNR ≈ (f/sqrt(b)) · sqrt(t)
Signal grows linearly with time; noise grows as the square root; the ratio grows as the square root. Hold SNR fixed at the detection threshold and the faintest detectable f falls only as 1/sqrt(t).
Convert to magnitudes, since that is how depth is quoted. A flux ratio r is 2.5 log₁₀ r magnitudes, and here r = sqrt(t₂/t₁), so the gain is 1.25 log₁₀(t₂/t₁). Doubling gives 1.25 x 0.301 ≈ 0.38 magnitudes. One magnitude needs 10^0.8 ≈ 6.3 times the exposure; two needs 40 times, three needs 250. This is why deep fields run to hundreds of hours and still stop somewhere.
Where does the background come from? At a good ground site, mostly airglow — hydroxyl and oxygen emission from the upper atmosphere, produced by the atmosphere itself and so immune to every remedy except leaving. Then scattered moonlight, zodiacal light from interplanetary dust, unresolved faint galaxies, and near cities artificial light. A dark site runs near 21 to 22 magnitudes per square arcsecond in visible light — a figure I am recalling, and one that varies with site, band and solar activity.
That points at the second and more powerful lever. Sky counts scale with the solid angle you must integrate over, set by the size of the star's blurred image. So b ∝ θ², and since SNR ∝ 1/sqrt(b), SNR ∝ 1/θ. Halve the image width and you double the signal-to-noise — exactly what four times the exposure would have bought. Sharpness is depth. That relation is why adaptive optics and space telescopes gain far more than their mirror areas suggest, and why a night of poor seeing is not merely blurry but shallow.
The refuting observation: stack a long series of identical sub-exposures of one field and measure the scatter among blank apertures against total time. The account above demands it fall as t^(-1/2). If it fell as 1/t, photon statistics would not be governing; if it flattened early, systematics — flat-field error, scattered light, confusion — would be binding rather than sky photons. In practice the t^(-1/2) law holds a long way and then does flatten, which is an honest caveat rather than a refutation.
The analogy
THE ANALOGY #Trying to hear a whisper across a room while rain drums steadily on the roof. Listening twice as long does give you more of the whisper — but twice as much rain too, and the rain's randomness only averages down as the square root. Patience helps, at a punishing exchange rate.
the analogy invites the idea that the problem is the rain's loudness, when in fact the steady average is removed cleanly by any observer and what defeats you is only its unevenness — the sky is subtracted every night, and it is the fluctuation about the subtracted value that sets the floor.
Clarifying the model
THE MODEL #Three refinements make the model honest.
The first is that the square-root law describes one regime, not all. When the background is very faint — a narrowband filter, or a detector in space — read noise can dominate instead, and then noise is fixed per exposure while signal grows linearly, so depth improves far faster than the square root. That regime ends the moment sky photons overtake read noise, which for a broad filter on the ground takes seconds. Knowing which regime you are in is what determines how to split a night into sub-exposures.
The second concerns what is being asked. This collection already explains why the night sky is dark at all — Olbers' paradox, answered by the finite age of the universe. That question asks why the background is not blindingly bright. This one starts where it finishes: granted a nearly dark sky, the residual is still overwhelmingly brighter than the sources we most want, and its shot noise stops us. The fixed point of difference is that Olbers concerns the background's mean while this concerns its variance.
The third is the design consequence. Because SNR ∝ 1/θ and, at fixed image size, SNR ∝ D for mirror diameter D, the levers are not equally priced. Doubling exposure time spends a factor of two of a scarce resource for 0.38 magnitudes; halving the image width costs money once and buys 0.75. That asymmetry is most of the argument for telescopes that are sharp rather than merely large, and for the considerable difficulty of putting them above the atmosphere.
A picture of it
THE PICTURE #How to readThe horizontal axis doubles at every step, so equal spacing means equal doublings of telescope time; the vertical axis is how much fainter you can see. The curve is straight against doubling exposures, so each doubling buys the same fixed 0.38 magnitudes — the plotted values are 1.25 log₁₀(t) computed above, not measurements. Read the far right as the sobering result: sixty-four times the observing time for barely two magnitudes.
What became clearer
WHAT CLEARED #The sky is not a curtain to be subtracted but a source of noise that grows alongside the signal. Since photon counts fluctuate as their square root, doubling the exposure improves the ratio only by sqrt(2), so depth advances as the logarithm of patience. The escape route is not more time but less sky in the aperture — a sharper image or a darker sky — because those attack the background at its source instead of averaging it down. Every deep survey is that trade written out: a fixed budget of nights, spent where the exchange rate is best.
Where to go next
ONWARD #- Why atmospheric turbulence sets the image width on most nights regardless of how big the mirror is.
- How confusion — overlapping faint sources — becomes the hard floor once photon noise has been beaten down.
- What makes the infrared sky so much darker from space, and why that matters more than mirror size.
Key terms
TERMS #| Term | What it means |
|---|---|
| Poisson noise | the intrinsic fluctuation of a random count, equal to the square root of the count. |
| Sky background | diffuse light in the aperture from airglow, scattered moonlight, zodiacal dust and unresolved sources. |
| Sky-limited regime | the condition where background shot noise dominates read noise and source noise, so depth grows as the square root of time. |
| Seeing | the angular width to which the atmosphere blurs a point source, setting how much sky enters the measurement aperture. |
Every term the collection defines is gathered in the glossary.