More collected signal can improve relative precision. Brightening a measurement is not the same operation.
By the end: Explain photon shot noise and the limits of simple ISO or sensor-size comparisons.
Even under steady illumination, the count detected in an interval varies. For independent photon arrivals in an ideal shot-noise model, the standard deviation is the square root of the expected count. Standard deviation describes a typical scale of fluctuation, not a guaranteed maximum error.
At 100 expected detected electrons, shot-noise standard deviation is 10 and signal-to-noise ratio is 10. At 400, deviation is 20 but the ratio is 20. Absolute variation grows while relative variation shrinks. More light does not eliminate randomness; it makes the signal larger relative to it.
Treat the electron count as a measurement of one fixed patch during one exposure. An expected count of 100 does not mean that every exposure records 100. It means that many repetitions under the same conditions would average toward 100. In the ideal independent-arrival model, the variance is also 100. Variance measures average squared departures from the mean. Taking its square root puts the variation back into electrons, giving a standard deviation of 10. These are statistical expectations, not a sequence a camera must follow.
Now compare a patch expected to supply 400 electrons. Its variance is 400, so its standard deviation is 20. The absolute fluctuation has doubled, but the signal has quadrupled. Divide deviation by signal to compare precision. Ten divided by 100 is 0.10, or 10 percent. Twenty divided by 400 is 0.05, or 5 percent. This explains why an image with more detected light can look smoother even though the underlying counts fluctuate by a larger number of electrons.
For another doubling of relative precision, start at 400 and require a deviation equal to one fortieth of the signal rather than one twentieth. A count of 1,600 has square root 40, and 40 divided by 1,600 is 0.025. Its relative deviation is 2.5 percent. Four times the collection halves the relative deviation. Merely doubling collection to 800 gives a square root of about 28.3 and a relative deviation of about 3.54 percent. Two times the light is useful, but it does not halve this variation.
| Expected electrons | Shot deviation, electrons | Deviation / signal | Signal / deviation |
|---|---|---|---|
| 100 | 10 | 10% | 10 |
| 400 | 20 | 5% | 20 |
| 800 | 28.28 | 3.54% | 28.28 |
| 1,600 | 40 | 2.5% | 40 |
A dark patch and a bright patch in the same photograph can therefore have different relative noise. Suppose the dark surface gives 100 electrons and a brighter surface gives 1,600 over equal sensing areas. Their absolute deviations are 10 and 40 electrons, but their relative deviations are 10 and 2.5 percent. Brightening the dark patch to resemble the bright one cannot change its original counting precision. Display contrast, texture, and surrounding tones also affect visibility, so relative deviation alone is not a complete prediction of appearance.
To see the distinction without taking a photograph, compare constructed repeated counts of 90, 100, and 110 with counts of 380, 400, and 420. The second set has larger departures in electrons, yet smaller departures as a fraction of its middle value. These three values illustrate arithmetic only. They are not a random sample with the claimed model variance. Nor does the standard deviation promise that a real measurement stays inside those endpoints. Occasional larger departures are consistent with random arrivals.
For optional camera practice, choose a still matte subject containing both bright and dark texture. Keep its position and illumination stable while comparing repeated images, if a device is available. Look for fine changes within patches rather than treating every difference as noise. Automatic exposure, focus, white balance, or merged frames may change too. Without a camera, predict which patch in the numerical example has the larger absolute fluctuation and which has the poorer relative precision. The answers differ because the comparison asks two different questions.
Read noise comes from measuring the stored charge and related electronics. Thermal effects, fixed-pattern errors, and processing can also matter. A simple independent-noise model adds variances, so read-noise standard deviation is squared before combining with shot-noise variance. Not every real error is independent or random.
ISO can change readout gain and effective read noise. Some cameras have multiple gain regimes. Consequently, “always use the lowest ISO” is not a complete rule when time and aperture are constrained. First collect enough light without unacceptable blur, depth loss, or clipping; then use an appropriate ISO and inspect device behavior.
Read noise becomes easier to reason about when expressed as an equivalent uncertainty in detected electrons. Suppose the expected signal is 100 electrons and independent read noise has a standard deviation of 10 electrons. Shot-noise variance is 100. Read-noise variance is 10 squared, also 100. Add the variances to get 200, then take the square root to get about 14.14 electrons of total deviation. Adding the two deviations directly would give 20 and would overstate this particular independent-noise model.
At 1,600 expected electrons with the same read noise, the total variance is 1,600 plus 100, or 1,700. The deviation is about 41.23 electrons. Without read noise it would be 40, so read noise makes a relatively small difference here. At only 25 expected electrons, the variance is 25 plus 100, or 125, and the deviation is about 11.18. Signal divided by deviation is about 2.24. Electronics now contributes most of the variance. The importance of read noise depends on the collected signal.
These numbers describe a paper comparison, not an extra slider in the experiment. The displayed experiment uses shot noise alone. Its arithmetic should agree with the zero-read-noise case, not with this combined case. Real cameras can also show stripes, hot pixels, or correlated processing errors. Adding independent variances does not describe every visible defect. A stripe affecting many neighboring samples is not interchangeable with independent fluctuations that average away when samples are combined.
| Expected signal, e− | Shot variance, e−² | Read variance, e−² | Total deviation, e− | Signal / deviation |
|---|---|---|---|---|
| 25 | 25 | 100 | 11.18 | 2.24 |
| 100 | 100 | 100 | 14.14 | 7.07 |
| 1,600 | 1,600 | 100 | 41.23 | 38.81 |
Consider a linear measurement with signal 100 and deviation 14.14. Multiplying every stored value by four makes its signal 400 and its deviation about 56.57. The ratio remains about 7.07. Collecting four times the photons instead gives expected signal 400, shot variance 400, and the same read variance 100. Its deviation becomes the square root of 500, about 22.36, and its ratio is about 17.89. Equal final signal values can therefore represent very different measurement precision. Neither calculation includes clipping or nonlinear rendering.
An ISO change is not necessarily the same as that ideal multiplication. Camera electronics may amplify before some later noise enters the chain. In an invented camera, suppose a gain change reduces input-referred read-noise deviation from 10 to 2 electrons while collection stays at 100. The total deviation changes from square root 200 to square root 104, about 10.20. The ratio rises from 7.07 to 9.81, but remains below the shot-only ratio of 10. No photons were added. This is a conditional mechanism, not a specification for your device.
A useful comparison must therefore name what stayed fixed. If an automatic camera raises ISO and shortens exposure, it changes collection as well as gain. If it also merges frames, one displayed exposure time may not describe all the evidence used. When controls are unavailable, compare a still subject in stronger and weaker light and note that the camera may compensate automatically. Do not infer a sensor's gain behavior from appearance alone. Separating the physical exposure, the readout, and the rendering identifies what further evidence would be needed.
| Operation | Signal after operation | Deviation after operation | Ratio |
|---|---|---|---|
| 100 e−, read deviation 10 e− | 100 e− | 14.14 e− | 7.07 |
| Multiply that measurement by four | 400 scaled units | 56.57 scaled units | 7.07 |
| Collect 400 e−, same read noise | 400 e− | 22.36 e− | 17.89 |
| Keep 100 e−, reduce read deviation to 2 e− | 100 e− | 10.20 e− | 9.81 |
S is expected detected electrons; r is read-noise standard deviation in electrons. Shot noise is modeled as Poisson and read noise as independent.
Here r = 0. The ratio is dimensionless, not a bound on every individual measurement.
Expected signal 400; noise deviation 20 electrons; relative deviation 5%.
| Expected signal | Shot-noise deviation | Signal/noise |
|---|---|---|
| 100 electrons | 10 | 10 |
| 400 electrons | 20 | 20 |
| 1,600 electrons | 40 | 40 |
Imagine two ideal sensors with the same shape, but one has twice the width and twice the height. Its area is four times as large. Frame the same scene from the same viewpoint, use the same f-number and time, and assume equal transmission and detection efficiency. Light exposure per unit sensor area is then approximately equal for ordinary distant scenes. A corresponding patch occupies four times the area on the larger sensor, so it can collect four times the electrons in total.
Suppose that patch collects 100 electrons in the smaller image and 400 in the larger. Under shot noise alone their signal-to-noise ratios are 10 and 20. To compare those numbers fairly, display both complete images at the same dimensions. Comparing one sensor pixel from each is a different question if their pixel counts or pixel areas differ. A larger sensor with more pixels can distribute its extra collected signal among more samples.
Consider four neighboring ideal sensing cells, each expected to record 100 electrons. Summing their signals gives 400. If their shot noise is independent, their variances add to 400 and the deviation of the sum is 20. Averaging rather than summing divides both signal and deviation by four, giving a mean of 100 and deviation 5. Relative variation is still 5 percent. Averaging does not create more recorded photons. It uses measurements over a larger area and sacrifices the ability to distinguish differences within that area.
This loss matters at edges. Suppose four cells cover values of 0, 0, 200, and 200 across a sharp boundary. Their average is 100, identical to the average of four cells each measuring a uniform 100. Once only the average remains, those two arrangements cannot be distinguished. Noise reduction that combines neighbors must decide which differences belong to noise and which belong to the subject. A broad smooth wall tolerates more averaging than fine lettering, hair, or fabric at the same output scale.
Equal framing and f-number do not guarantee equal depth of field across sensor formats. In the ideal comparison, the larger format needs about twice the focal length to preserve field of view at the same viewpoint. At f/4 its entrance pupil is then twice the diameter. If it instead uses f/8, its pupil diameter returns to the smaller setup's value. Exposure per area falls to one quarter. Multiplying that quarter by four times the sensor area gives approximately equal total collection, at unchanged time and transmission.
That last comparison explains why a format advantage must state its constraints. A fourfold area increase can provide fourfold total light at equal f-number, but stopping down two stops can spend that advantage on a more comparable depth requirement. These are simplified distant-scene relationships, not complete equivalence guarantees. Diffraction, lens aberrations, pixel layout, read noise, and processing remain relevant. For a paper exercise, decide whether a comparison holds f-number fixed or holds a similar depth requirement fixed before predicting its noise result. Otherwise the claim has no single numerical answer.
At equal scene, framing, f-number, time, and transmission, larger sensor area can collect more total light even when exposure per unit area is similar. That can help at matched final output, but depth of field, lens size, resolution, and processing may differ. Equal pixel-level crops are not always equal viewing conditions.
Denoising estimates underlying structure. Strong filtering may erase small texture or invent plausible detail, especially in computational tools. Keep an original and compare at the intended display size. The model below shows expected signal and noise, not a randomly generated sample photograph or a prediction for a named camera.
A smooth result is not proof of an accurate measurement. Suppose a small fabric pattern alternates between expected values of 90 and 110, with random deviation around each. A filter that replaces every adjacent pair by its average tends toward 100. It suppresses random departures, but also removes the real alternating pattern. If the intended output cannot resolve that pattern, the trade may be acceptable. If the photograph documents the weave, the same operation removes the subject information that matters. Specify the task before choosing the smoothing strength.
Compare three kinds of evidence in a noisy photograph. A broad tone change across a wall may survive substantial averaging because it spans many samples. Thin writing may disappear when neighboring letters and gaps merge. A repeated grid may turn into a false pattern after sampling or resizing. Increasing sharpness can strengthen both true edges and false structure. A cleaner-looking edge is therefore not enough to establish that a particular letter or texture was present.
Output size changes this judgment without changing capture. Suppose an image is 4,000 pixels wide and is delivered 1,000 pixels wide. Each output pixel covers roughly four input pixels across and four vertically, or sixteen input positions. In an ideal uniform patch with independent equal noise, averaging sixteen samples divides deviation by four while leaving the average signal unchanged. Real resizing filters do not all use a simple square average. Demosaicing and prior denoising also correlate nearby samples. The fourfold result is a bounded calculation, not a promised software outcome.
Combining separate frames differs from combining neighboring locations. Four perfectly aligned exposures of a stationary scene can measure the same location repeatedly. If each supplies signal 100 with independent shot deviation 10, their average has signal 100 and deviation 5. Spatial detail need not be sacrificed in the same way as averaging four adjacent subject positions. However, the four exposures collect four times the total photons over their combined capture time. Comparing their result with one equally short exposure does not hold total collection fixed.
Movement limits that benefit. A face that shifts between frames places its eye edge at different locations. A simple average makes a wider or doubled edge rather than a more precise eye. Alignment can compensate for some movement, but moving parts, changing expressions, occlusions, and flickering lights complicate it. Multi-frame software makes additional decisions about which measurements to combine. No unnamed phone should be assumed to perform the ideal four-frame calculation.
Try an evidence-based review on paper or with an existing image. Choose one smooth patch and one fine feature that should remain distinguishable. Predict whether stronger smoothing helps each, then compare at the intended delivery size rather than only at extreme magnification. If an editor is available, preserve the original and change only noise reduction for this comparison. Without one, use the alternating 90 and 110 example to explain what averaging retains and removes. Report smoothness and retained structure separately instead of declaring the smoother version universally better.
You want to double shot-noise-limited signal-to-noise ratio from an expected count of 400. What count is needed?
The ratio grows with the square root of count.
| Count | SNR |
|---|---|
| 400 | 20 |
| 800 | About 28.3 |
| 1,600 | 40 |
You need 1,600 expected electrons, four times the count or two stops more collection. Doubling to 800 gives only a square-root-of-two improvement. This assumes shot noise dominates and the sensor does not saturate.
An editor doubles every value in a linear image whose signal is 100 and noise deviation is 10. What happens to their ratio?
Scale the deviation as well as the signal.
| Stage | Signal | Deviation | Ratio |
|---|---|---|---|
| Before | 100 | 10 | 10 |
| After ×2 | 200 | 20 | 10 |
Signal becomes 200 and deviation becomes 20, so the ratio stays 10. This ideal multiplication makes the image brighter but adds no information. Clipping, quantization, and later nonlinear processing can make real editing less reversible.
The lessons and answers stand alone. These optional secondary references explain the optics and imaging concepts. Numerical examples are constructed, not measured camera specifications. Visual judgments depend on the intended image; the course does not claim a universal aesthetic score.
Shot noise, read noise, and signal-to-noise concepts. The count model omits sensor-specific electronics and spatial processing.
Secondary overview of exposure, time, aperture, and sensitivity conventions. Our light-collection and brightness ratios are simplified teaching models, not camera calibrations.