Why is mathematics important in particle counting? A detector may register clicks, light pulses, droplets, aerosol particles, cells or radioactive events one by one. The display produces an integer, but the scientific question is rarely “what number appeared once?” We need a rate, an uncertainty, a background correction and a careful statement about what the detector could miss or falsely count.
For independent events occurring at a stable average rate, the Poisson distribution gives a powerful starting model. If the expected count is lambda, both its mean and variance are lambda, so the standard deviation is square root of lambda. Relative random uncertainty falls roughly as one divided by the square root of the number counted.
That square-root rule is useful, but it is not magic. Dead time, coincidence, changing concentration, clustering, calibration, sample preparation and background can violate a simple Poisson model. This article explains the mathematics and its limits for students; it is not a radiation-safety procedure, clinical interpretation or instrument certification.
- Understand what is counted
- Build the Poisson model
- Turn counts into rates
- Subtract background honestly
- Use confidence intervals
- Work through an example
- Study detector limits
- Build student skills
- Read the FAQs
A Count Is Not Yet a Conclusion
Define the event
A counter needs a rule for deciding what constitutes one event. In a particle detector, an event might be a voltage pulse crossing a threshold. In an optical particle counter, scattered light may place a particle into a size bin. In image analysis, connected pixels may be labelled as one object.
The event rule can merge two real particles into one count or split one signal into several. Noise can create false events. A reported count must therefore identify the detector, threshold, observation time, sample volume or area, and any corrections.
The number 420 might mean 420 pulses in 10 seconds, 420 particles in one litre, or 420 objects in an image. These support different rates and claims. Units and boundaries come before statistics.
Counts are discrete
Counts take nonnegative integer values: 0, 1, 2 and so on. A model may have a noninteger expected value, such as lambda = 3.7, because expectation is a long-run average. One observation can never contain exactly 3.7 events.
This distinction helps students understand probability. Expected count is not a prediction that every interval will contain the same number. It is the centre of a distribution across repeated comparable intervals.
Random and systematic uncertainty
Random counting variation changes from repeat to repeat. Poisson statistics may describe it when assumptions hold. Systematic effects shift the result in a consistent or condition-dependent way: calibration bias, missed counts, wrong sample volume, threshold choice or contamination.
Counting longer can reduce relative random variation, but it does not automatically correct a biased detector. A million precisely counted events can still support a wrong concentration if the sampled volume is wrong by ten percent.
Rate and concentration
Count rate R = N/t, where N events are observed in time t. If sampling volume V is known and detection efficiency eta applies, a simplified concentration estimate may be C = N/(eta V).
Each denominator matters. If a pump samples 2.0 litres per minute for 5 minutes, volume is 10 litres only if flow stays at that value and conditions are defined consistently. A flow calibration uncertainty contributes to concentration uncertainty.
Why a statistical model is needed
Suppose a stable source is measured for equal intervals and gives 96, 111, 103, 88 and 107 counts. Variation does not necessarily imply the source changed. A statistical model asks whether the spread is compatible with random event timing.
The Poisson distribution provides a baseline when events occur independently at a constant average rate and two events in an infinitesimally short interval are negligibly likely. These are assumptions to test, not automatic truths about every counter.
Did You Know? Counting four times as long roughly halves the relative Poisson standard uncertainty, not quarters it. Precision improves with the square root of total counts, so each extra improvement becomes more demanding.
The Poisson Distribution
Probability mass function
If N follows a Poisson distribution with expected count lambda, then:
P(N = n) = exp(-lambda) lambda to the power n divided by n factorial,
for n = 0, 1, 2 and so on. Lambda must be nonnegative. The probabilities sum to one.
For lambda = 2, probability of zero counts is exp(-2) = about 0.1353. Probability of exactly one is exp(-2) times 2 = about 0.2707. Probability of exactly two is exp(-2) times 4/2 = also about 0.2707.
Seeing zero events does not prove the rate is zero. Even with expected count two, a zero result occurs about 13.5 percent of the time. This is why small-count reasoning needs intervals rather than a casual “none detected means none exists” conclusion.
The US National Institute of Standards and Technology’s Poisson distribution reference gives the probability form and core properties in its engineering statistics handbook.
Mean and variance
For a Poisson random variable, E[N] = lambda and Var(N) = lambda. Therefore standard deviation sigma_N = square root of lambda.
If expected count is 100, standard deviation is 10. If expected count is 10,000, standard deviation is 100. Absolute variation increases with count, but relative variation sigma/lambda = 1/square root of lambda decreases.
In practice lambda is unknown and observed N is used to estimate it. For moderate or large N, square root of N is a familiar approximate standard uncertainty for the count. For very small N, symmetric N plus or minus square root N intervals are poor and can even imply negative counts.
Additivity
The sum of independent Poisson counts is Poisson, with expected values added. If detector A expects 12 events and independent detector B expects 8 in the same interval, total expects 20.
This property also allows time intervals to be combined when the rate is stable. Ten independent one-minute counts with mean five each have a total Poisson mean 50. However, if the rate changes during the ten minutes, the total may still be Poisson under certain nonhomogeneous-process conditions, but one constant-rate interpretation would be misleading.
Waiting times
In a homogeneous Poisson process with rate r, waiting times between events follow an exponential distribution with mean 1/r. The count view asks how many events occur in a fixed time; the waiting-time view asks how long until the next event.
These are linked models. A rate of 2 events per second gives expected waiting time 0.5 second, but actual gaps vary widely. Short gaps and long gaps can occur without the source changing.
Variance-to-mean diagnostic
Across repeated equal intervals, sample variance near sample mean is consistent with a Poisson model, though not proof. Variance much larger than mean is overdispersion; possible causes include changing rate, clustering, mixed populations or unmodelled conditions.
Variance smaller than mean is underdispersion; dead time, regular spacing or constrained counts may contribute. Sampling variation means the ratio will not equal one exactly, so a formal dispersion test or simulation may be appropriate.
Binomial origin
The Poisson distribution can arise as a limit of many independent opportunities, each with small event probability. If N is binomial with M trials and probability p, while M grows, p shrinks and Mp approaches lambda, the distribution approaches Poisson.
This helps explain both its usefulness and limits. If a finite population is sampled without replacement or event probabilities strongly interact, binomial, hypergeometric or other models may be better.
Count Rates, Time and Relative Uncertainty
Rate estimate
For N counts in live time t, rate estimate R = N/t. If t is treated as exact and N is Poisson, approximate standard uncertainty uR = square root N divided by t.
Observe 900 counts in 60 seconds. R = 15 counts per second. uR = 30/60 = 0.5 count per second. Relative standard uncertainty is 0.5/15 = 3.33 percent, the same as 1/square root 900.
If measuring time has uncertainty, include it. Digital timers are often much more precise than counting variation, but trigger delays or live-time corrections can matter.
Scaling time
At stable rate r, expected count lambda = rt. Relative Poisson uncertainty is 1/square root(rt). To cut this uncertainty by half, multiply expected count and time by four. To cut it by ten, multiply them by one hundred.
This square-root law helps plan experiments. If a one-minute measurement yields 100 counts and about 10 percent relative counting uncertainty, reaching about 2 percent requires roughly 2,500 counts, or 25 minutes at the same rate, before background and systematic uncertainty.
Comparing two rates
For independent counts N1 in time t1 and N2 in time t2, compare rates R1 and R2 with uncertainties. Difference D = N1/t1 – N2/t2 has approximate variance N1/t1 squared + N2/t2 squared when observed counts estimate Poisson means.
A raw difference is not enough. Rates 100 and 110 counts per second may be clearly different if based on millions of counts, or indistinguishable if based on very short observations.
For a ratio, log-scale methods are often useful. But zero counts make simple ratios undefined. Exact Poisson rate-ratio intervals or Bayesian methods can handle small counts more appropriately.
Pooling versus averaging
If repeated observations have different durations but the same stable rate, total count divided by total time is the natural pooled estimator. The arithmetic average of individual rates gives each interval equal weight even if durations differ.
Example: 10 counts in 1 minute and 90 counts in 9 minutes both estimate 10 per minute. No issue arises. But 2 counts in 1 minute and 90 in 9 give rates 2 and 10. Their simple average is 6, while pooled rate is 92/10 = 9.2. The longer observation contains more information.
Count normalisation
Measurements may be normalised per second, per litre, per square centimetre, per unit energy bin or per unit mass. Denominators can also be uncertain. State whether the value is a count, rate, density or concentration.
Changing bin width changes expected counts per bin. A spectrum with finer bins looks noisier because each bin receives fewer events, even though total data are unchanged. Smoothing may improve appearance but creates correlation and can hide narrow features.
Detection efficiency
If each true event is detected independently with probability eta, Poisson thinning produces a Poisson observed count with mean eta times the true mean. Correcting N/eta estimates the true count, but uncertainty in eta adds systematic and calibration components.
Efficiency may depend on particle size, energy, position, composition and threshold. One constant correction is justified only for a defined range and calibration.
Background Subtraction and Net Counts
Signal-plus-background measurement
A detector may count desired signal plus unrelated background. Take a gross measurement G for time tg and a separate background measurement B for time tb. Estimated net signal rate is:
S = G/tg – B/tb.
If G and B are independent Poisson counts and times are treated as exact, variance of S is approximately G/tg squared + B/tb squared. The variances add even though the rates are subtracted.
This is a crucial rule. Subtracting background does not subtract uncertainty. Both measurements contribute random variation.
Equal-time example
Suppose gross count is 560 in 10 minutes and background is 200 in 10 minutes. Net count equivalent is 360, or 36 per minute.
Standard uncertainty of gross count is square root 560 = 23.7. Background is square root 200 = 14.1. Net-count standard uncertainty is square root(560 + 200) = square root 760 = 27.6 counts for the equal interval, or 2.76 per minute.
It would be wrong to use square root 360 = 19.0, because net count is a difference of two noisy observations, not a directly observed Poisson count.
Unequal background time
Measuring background longer can improve its rate estimate. Suppose the same gross G = 560 over 10 minutes, while background B = 1,200 over 60 minutes, still 20 per minute. Net rate remains 36 per minute.
Gross-rate variance contribution is 560/10 squared = 5.6 (counts/minute) squared. Background-rate contribution is 1,200/60 squared = 0.333. Standard uncertainty of net rate is square root 5.933 = 2.44 per minute, smaller than 2.76.
The result improves because background rate is measured with more counts. But if background drifts over time, a very long separate measurement may not represent the gross interval. Statistical precision and temporal relevance must be balanced.
Negative net estimates
Random variation can make gross rate lower than measured background rate, yielding a negative net estimate. The physical signal rate cannot be negative, but the estimator can. Replacing it silently with zero biases results upward.
Report the measurement and use an interval method that respects the nonnegative parameter. Decisions about detection should follow a defined statistical protocol, not whether the point estimate “looks positive.”
Blank, background and contamination
Background may include instrument noise, environmental events, reagent blanks or sample-handling contamination. The correct background control should match the signal measurement except for the target contribution. A background taken in a different geometry or condition may introduce bias even if counted for a long time.
NIST documentation for reference-material preparation has discussed Poisson statistical counting error within a wider measurement process. The broader lesson is that counting statistics is one component of traceable measurement, not the complete uncertainty budget.
Confidence Intervals, Especially for Small Counts
Why plus or minus square root N can fail
For N = 1, the symmetric interval 1 plus or minus 1 reaches zero. For N = 0, it gives zero width, falsely implying perfect certainty. Poisson distributions are skewed at small lambda, so normal approximations are poor.
Exact frequentist intervals use Poisson tail probabilities, often expressed through chi-square quantiles. A two-sided confidence interval for lambda after observing n over a fixed interval can be constructed with different lower and upper limits. The exact numeric choice depends on confidence level and interval convention.
For zero observed events, the upper bound is not zero. At a one-sided 95 percent level, solve P(N=0) = exp(-lambda) = 0.05. Then lambda = -ln 0.05 = about 3.00. Thus zero observed events is compatible with an expected count up to about three under that convention.
Confidence level is not probability of the fixed parameter
In a frequentist interval, the procedure covers the fixed true parameter in the chosen proportion of repeated experiments. After data are observed, the interval either covers it or does not. Saying “there is a 95 percent probability lambda lies in this particular interval” is a Bayesian interpretation unless a prior model is specified.
Students do not need to turn this into a philosophical dispute. They should name the interval method and interpret it consistently.
Bayesian intervals
A Bayesian analysis combines a prior distribution with the Poisson likelihood to produce a posterior distribution for lambda. Gamma priors are mathematically convenient because they are conjugate. Different reasonable priors can matter strongly with few counts.
Bayesian intervals can directly express posterior probability, but the prior and model must be reported. “Bayesian” is not a licence to hide assumptions.
Detection decisions
A detection threshold, decision limit and detection limit are related but not identical concepts. They depend on false-positive and false-negative criteria, background uncertainty and the measurement procedure. A rule such as “signal exceeds three square roots of background” may be a rough heuristic in limited settings, not a universal official standard.
Small-count measurements should use methods appropriate to their field and governing standards. Medical, environmental and radiation decisions can be high stakes; this educational article cannot replace current professional protocols.
Multiple bins and multiple testing
If a spectrum has hundreds of bins, some will fluctuate high by chance. Looking only at the highest bin and applying a single-bin probability exaggerates evidence. Multiple-testing correction, pre-specified regions or model-based peak fitting may be needed.
The same issue appears when repeatedly watching a counter and stopping after an unusually high interval. The stopping rule affects inference. Pre-plan observation time and analysis where practical.
Worked Example: Signal Plus Background
An educational optical counter measures a sample for 15 minutes and records G = 930 events. A matched blank is measured for 45 minutes and records B = 1,800 events. Assume independent Poisson counts, stable rates, exact live times and no efficiency correction for the first calculation.
Gross and background rates
Gross rate Rg = 930/15 = 62.0 events per minute. Background rate Rb = 1,800/45 = 40.0 events per minute. Net signal rate S = 22.0 events per minute.
Counting uncertainty
Gross-rate variance estimate = 930/15 squared = 930/225 = 4.133 (events/minute) squared.
Background-rate variance estimate = 1,800/45 squared = 1,800/2,025 = 0.889.
Net-rate variance = 5.022, so standard uncertainty uS = square root 5.022 = 2.24 events per minute.
An approximate large-count summary is 22.0 plus or minus 2.24 events per minute at one standard uncertainty. An approximate two-standard-uncertainty range is about 17.5 to 26.5, though a formal interval should state method and level.
Signal-to-uncertainty ratio
The net estimate divided by its standard uncertainty is 22.0/2.24 = about 9.8. This describes separation from zero under the stated approximate model. It does not prove the events are the intended particle type or that systematics are absent.
Concentration estimate
Suppose the counter samples 0.50 litre per minute during the 15-minute gross measurement, so gross sample volume is 7.5 litres. If net rate is 22 events per minute, net observed concentration is 22/0.50 = 44 events per litre.
If calibrated detection efficiency is eta = 0.80 for the relevant particles, corrected concentration is 44/0.80 = 55 particles per litre. This assumes background is on the same detected-count basis and efficiency is applicable.
Add efficiency uncertainty
Suppose relative standard uncertainty of net counting rate is 2.24/22 = 10.2 percent, flow rate has 3 percent relative uncertainty and efficiency has 5 percent, treated as independent. Relative uncertainty of corrected concentration is approximately square root(0.102 squared + 0.03 squared + 0.05 squared) = 11.8 percent.
Standard uncertainty is about 55 times 0.118 = 6.5 particles per litre. The random count contribution is largest, but efficiency and flow prevent the result from improving exactly as 1/square root time forever.
What if measurement time doubles?
If both signal and background conditions remain stable and only gross time doubles from 15 to 30 minutes, expected gross count doubles. Gross-rate variance contribution halves. Background contribution stays the same if the same 45-minute blank is used.
Eventually background uncertainty, efficiency, flow calibration and other systematics form a floor. Planning should allocate time where it reduces the dominant uncertainty, not simply extend every measurement equally.
Conservation-style cross-check
Expected gross rate should equal signal rate plus background rate: 22 + 40 = 62. This arithmetic is exact for the point estimates. The uncertainty is not conserved through subtraction; independent variances add.
Results table
| Quantity | Estimate | Approximate standard uncertainty from counts |
|---|---|---|
| Gross rate | 62.0 per min | sqrt(930)/15 = 2.03 per min |
| Background rate | 40.0 per min | sqrt(1800)/45 = 0.943 per min |
| Net signal rate | 22.0 per min | sqrt(2.03² + 0.943²) = 2.24 per min |
| Corrected concentration | 55 per L | About 6.5 per L including stated flow and efficiency terms |
This table reports the model, not every possible uncertainty. Sample representativeness, particle losses, threshold calibration and coincidence may dominate in a real instrument.
Dead Time, Pile-Up and Efficiency
Dead time
After registering an event, a detector may be unable to record another for a short dead time tau. At high rates, missed events make observed rate lower than true rate and the count distribution non-Poisson.
In a simplified nonparalyzable model, observed rate m and true rate n relate by m = n/(1 + n tau), or n = m/(1 – m tau) when m tau is below one. In a paralyzable model, events during dead time extend the dead interval and the relationship differs.
Choosing the wrong dead-time model can bias correction. Instrument documentation and validation are essential. Near saturation, small measurement errors can produce huge corrected-rate uncertainty.
Pulse pile-up
If two signals overlap, electronics may record one larger pulse, distort energy estimates or reject the event. Pile-up changes counts and spectra. It cannot be fixed by applying square root N to the final count.
Coincidence
Some systems intentionally require events in multiple detectors within a time window. Random coincidences depend on rates and window width. Widening the window may capture more true coincidences but also more accidental ones.
This becomes a timing and probability problem. The event definition now includes a logical condition, so independence assumptions must be reconsidered.
Size-bin misclassification
Optical particle counters infer size from signal amplitude using calibration assumptions. Two small particles passing together may resemble one larger particle. Refractive index and shape affect scattering. Count uncertainty within a bin therefore includes classification and coincidence effects, not only Poisson variation.
Sampling losses
Particles may deposit in tubing, fail to enter the inlet or be unevenly mixed. Larger particles can settle. An apparently precise count from a nonrepresentative sample does not estimate the environment accurately.
Time-varying rates
If concentration changes, counts in long intervals mix states. A nonhomogeneous Poisson process can model a time-dependent rate r(t), with expected count equal to the integral of r(t) over the interval. But estimating that function requires temporal resolution.
Short bins improve time resolution but contain fewer counts and more relative noise. This is a bias-variance and resolution trade-off.
When Counts Stop Being Poisson
Clustering and contagion
Events may arrive in bursts. Biological particles may clump; cosmic or radioactive processes can create correlated showers; manufacturing defects may occur in batches. Variance then exceeds the mean.
Negative binomial or compound Poisson models can represent certain overdispersed counts, but choosing one should follow mechanism and diagnostics, not fashion.
Finite populations
Counting sampled items without replacement from a finite lot may follow a hypergeometric model. Counting successes among a fixed number of independent trials fits a binomial model. Poisson can approximate rare binomial events, but the approximation should be checked.
Detector inhibition
Dead time and capacity limits make closely spaced events less likely to be recorded separately, causing underdispersion or saturation. A linear concentration-to-count calibration will eventually fail.
Changing exposure
If sample volume, area or live time varies, raw counts are not identically distributed. Normalise correctly and propagate denominator uncertainty. Comparing counts without exposure is like comparing distances without units.
Mixed sources
A dataset combining days, instruments or locations with different rates can look overdispersed even if each subset is Poisson. Stratify or use a hierarchical model. Variation between groups is information, not noise to erase.
Threshold selection
Changing a pulse threshold changes what counts. If the threshold is chosen after viewing data to maximise a difference, reported significance is optimistic. Pre-specification or independent validation reduces this bias.
Goodness-of-fit
Plot the count histogram, mean and variance across comparable intervals. Use residuals and an appropriate goodness-of-fit test. Ensure expected frequencies are adequate for chi-square approximations or use simulation/exact methods.
Failure of a Poisson model is not failure of mathematics. It is evidence that event independence, constant rate or detector response needs a richer explanation.
For a physical example of pulses becoming evidence, read Why Science? | Scintillation Detectors, Light Pulses and Radiation Evidence. That article owns the detector-mechanism intent; this one focuses on the mathematics of counts and uncertainty.
Misconceptions Worth Correcting
“The uncertainty is square root of the net count”
Not after independent background subtraction. Gross and background variances add. A net difference is generally not itself Poisson.
“Zero counts means zero true events”
No. Zero is an observation compatible with a range of expected counts. An upper confidence limit remains positive.
“More counts remove all uncertainty”
More counts reduce relative Poisson variation, but calibration, efficiency, exposure and sampling biases remain. Systematic terms can dominate.
“Variance equal to mean proves Poisson behaviour”
It is a useful diagnostic, not proof. Different processes can have similar first two moments. Event timing, conditions and goodness-of-fit matter.
“A negative background-corrected value is impossible, so replace it with zero”
The physical rate is nonnegative, but an unbiased noisy estimate can be negative. Silently truncating changes its statistical properties. Use an appropriate constrained interval or model.
“Ten-minute and one-minute rates can be averaged equally”
Not when estimating one stable rate from different exposures. Pool counts and times or use inverse-variance methods suited to the model.
“A display with six digits is extremely accurate”
Resolution is not accuracy. Detector efficiency, flow, dead time, calibration and sample representativeness may limit knowledge far more than display digits.
How Students Can Learn This Mathematics
Simulate rare events
Roll many dice and define an event, such as two sixes in a pair. Count events in equal blocks. Compare the histogram with a binomial model and a Poisson approximation when probability is small.
Use arrival times
Generate random exponential waiting times or record harmless events such as drops from a controlled dripper. Sum waiting times into fixed bins. Connect rate, expected count and variability.
Verify the square-root law
Simulate Poisson counts at lambda = 10, 100 and 1,000. Calculate empirical mean, standard deviation and relative standard deviation. Observe that absolute spread grows while relative spread shrinks.
Practise zero-count reasoning
For zero observed events, derive the one-sided 95 percent upper value from exp(-lambda) = 0.05. Explain why the answer is about three rather than zero.
Build a background worksheet
Enter gross count and time, background count and time, then calculate net rate and its uncertainty. Add unit checks and reject negative times or impossible efficiencies.
Allocate measurement time
Given a fixed total time, explore how much to allocate to gross and background measurements. The optimal split depends on rates and goals. This is a practical optimisation problem, not always a fifty-fifty rule.
Check dispersion
Create repeated equal-time counts. Compare sample variance with mean and inspect a time plot. Deliberately simulate a changing rate to see how overdispersion appears.
Add efficiency and volume
Convert net rate to concentration using flow or sampled volume and efficiency. Propagate uncertainties. Identify when systematic terms dominate the counting term.
Use safe and ethical data
Students can count light pulses from a safe classroom sensor, drops, seeds in images or publicly provided datasets. Do not improvise radiation sources, clinical sampling or hazardous aerosols. Follow laboratory rules and privacy requirements.
Connect to broader mathematics
Particle counting links probability, factorials, exponential functions, confidence intervals, error propagation, calibration and time series. It also connects with signal detection, quality control and data science.
Keep pathways open
Students who enjoy this topic may explore physics, chemistry, environmental monitoring, biomedical engineering, statistics, electronics, computing, metrology or laboratory science. Mathematics supports those pathways but does not guarantee admission or professional practice. Consult current official institution requirements.
For Singapore cohorts, verify MOE and SEAB terminology for Posting Groups, Full Subject-Based Banding, G1/G2/G3 subjects and the Singapore-Cambridge Secondary Education Certificate. When Additional Mathematics is relevant, use the current SEC Additional Mathematics Examination terminology for G2/G3 cohorts rather than assuming older labels.
Guidance for Parents and Teachers
Begin with the event definition. Ask what the instrument calls one count, what time or volume was observed and what could create false or missed events.
Use repeated measurements. One count teaches arithmetic; a sequence teaches variation. Let students see that stable conditions do not produce identical integers.
Distinguish random and systematic uncertainty. Longer counting improves one but may leave efficiency or volume bias unchanged. This is a central scientific habit.
Avoid the square-root shortcut for tiny counts and background differences. Introduce exact or simulation-based intervals and explain why asymmetry is natural.
Encourage graphs. Plot counts over time, histogram repeated intervals and compare variance with mean. Patterns often reveal rate changes or detector issues before a formal test.
Reward honest non-detections. A careful upper bound is more informative than claiming “zero.” Negative net estimates can be reported without embarrassment when the method is clear.
Keep high-stakes boundaries firm. Radiation, medical, environmental compliance and product-release decisions require current domain procedures, calibrated equipment and qualified professionals. A classroom Poisson calculation is not authorisation.
Frequently Asked Questions
What conditions support a Poisson count model?
Events should occur independently with a stable average rate over the interval, and simultaneous events should be negligible at the model’s time scale. Detector response must not strongly distort the count.
Why does a Poisson distribution have mean equal to variance?
It is a mathematical property of the distribution derived from its probability-generating function or moment calculations. It gives the square-root standard-deviation rule.
Why is relative uncertainty one divided by square root N?
Approximate standard uncertainty is square root N, divided by estimate N. The ratio is 1/square root N for moderate or large counts.
Does counting four times longer always halve uncertainty?
It halves the relative Poisson component when rate remains stable and counts scale with time. Background, drift and systematic uncertainties may prevent the total uncertainty from halving.
How do I subtract background?
Subtract background rate from gross rate after accounting for their observation times. Add their independent variance contributions; do not subtract uncertainties.
Why can the net count be negative?
Gross and background are random. A low gross fluctuation and high background fluctuation can produce a negative difference even when the physical signal rate is nonnegative.
What does zero observed events tell us?
It suggests the rate may be low but does not prove it is zero. The upper bound depends on confidence level, exposure and interval method.
Is the Poisson model suitable for particle counters at high rates?
Not automatically. Dead time, coincidence, pile-up and saturation can produce missed or merged events and non-Poisson behaviour.
Is a count rate the same as concentration?
No. Concentration requires sampled volume or flow and often efficiency corrections. Rate alone is instrument events per time.
Can a Poisson calculation prove what caused the pulses?
No. Statistics quantifies count variation under a model. Identifying the source requires detector response, calibration, controls and physical evidence.
Useful Next Reading
- NIST Engineering Statistics Handbook: Poisson distribution
- NIST reference-material report discussing Poisson counting error
- NIST technical report on calibration and Poisson statistics
- Why Science? | Scintillation Detectors, Light Pulses and Radiation Evidence
- Why Mathematics? | Thermometer Calibration, Interpolation and Measurement Uncertainty
- Why Mathematics? | Queueing Theory, Arrival Rates and Waiting Times
Counting mathematics teaches a wonderfully transferable discipline. Define the event, measure the exposure, choose a probability model, carry background uncertainty and challenge systematic effects. A counter’s integer is only the beginning. The real benefit of mathematics is turning that integer into a conclusion whose strength and limitations are visible.