Radioactive Decay & Counting Statistics · Guide · glidepathscience.org
GuideRadioactive Decay & Counting Statistics
Contents
This Guide explains the ideas and equations behind every Walkthrough step, Concept Challenge question and Predict round, in the order a textbook takes them. Part 1 prepares the Intro track and Part 2 the Advanced track; each takes about a quarter of an hour to read. Each section ends with a worked example that is not one of the sim’s own problems.
Part 1 · The essentials
Math you’ll need: algebra, powers and powers of ten, and natural logarithms (ln) on a calculator; one optional box uses a derivative.
Unstable nuclei
Most nuclei around us are stable: the carbon in your body and the iron in a nail have been the same for billions of years. Some nuclei, though, have too many neutrons, too few, or simply too much energy, and sooner or later each one changes into something more stable by throwing out a particle: an alpha particle, an electron (beta decay), or a gamma ray. That change is radioactive decay. The nucleus left behind, the daughter, may be stable or may decay in turn.
This Guide is about when decays happen and how we count them. Two facts sit side by side, and the sim shows both. A single decay is completely unpredictable. A large sample decays by a law as steady as a clock. Between them sits the counter, which sees only some of the decays, clicks even when no source is present, and gives a different number every time you count. Learning to read those clicks honestly is the second half of the topic.
Random nuclei, predictable samples
Watch a handful of identical nuclei and you cannot say which will decay next. Nothing inside a nucleus counts down; it does not age. Each second, every nucleus of an isotope has the same small chance of decaying, whether it was made a moment ago or has existed since the Earth formed. A nucleus that has lasted a long time is not “due”.
Now watch ten thousand of them. Because each nucleus has the same chance per second, the number that decay in a second is proportional to the number present: when there are many, many decay; when few remain, few decay. The count of survivors therefore falls along a smooth, predictable curve. Run the same ten thousand again and the curve is almost exactly the same, although no single nucleus does the same thing twice.
The time for half of a sample to decay is the half-life, $T_{1/2}$. It is a property of the isotope: it does not depend on how many nuclei there are, how old the sample is, or where you start the clock. From 10,000 nuclei to 5,000 takes one half-life; so does 5,000 to 2,500, or 3,000 to 1,500.
Half-life and halves of halves
If one half-life leaves half, two leave half of a half, a quarter, and three leave an eighth. After $n$ half-lives the fraction left is $(\tfrac12)^n$, so
$N = N_0(\tfrac12)^{t/T_{1/2}}$
where $N_0$ is the number at the start and $t/T_{1/2}$ the number of half-lives, which need not be a whole number. The usual mistake is to think a sample is gone after two half-lives (“half, then the other half”). It is not: each half-life takes half of what is left, never all of it. In a real sample the halving goes on until only a handful of nuclei remain, and then the last few go at random.
The decay constant and activity
The chance per second that one nucleus decays is the decay constant, $λ$, measured in s⁻¹. A sample of $N$ nuclei then loses about $λN$ nuclei each second. A quantity that falls at a rate proportional to itself is an exponential, so $N = N_0e^{−λt}$, the same curve as the halving law written with $e$. Setting $N = N_0/2$ links the two: $e^{−λT_{1/2}} = \tfrac12$, so
$λ = \ln 2/T_{1/2}$
with $\ln 2$ = 0.693. A short half-life means a large $λ$: each nucleus is likely to go soon. The mean life, $τ = 1/λ$, is the average lifetime of a nucleus; it is longer than the half-life by a factor 1/ln 2 = 1.44.
The number of decays per second is the activity,
$A = λN$
measured in becquerels (1 Bq = one decay per second). An older unit, the curie, is 3.7 × 10¹⁰ Bq; school sources are about 0.1 µCi, or 3,700 Bq. Because $A$ is proportional to $N$, the activity halves every half-life too. To find $N$ in a sample, convert its mass to moles and multiply by Avogadro’s number; to find $λ$, the half-life must be in seconds, since a becquerel is per second. Forgetting that conversion is the commonest slip.
What a counter counts
A Geiger counter does not measure the activity. Radiation leaves a source in every direction, and only the part aimed at the tube can enter it; of that, some passes through without being counted. The counter records a fraction $ε$ of the decays, its efficiency for this source in this position, so the count rate is
$R = εA$
Moving the tube closer raises $ε$ and the count rate, while the activity stays exactly the same. Rates on a classroom counter are usually quoted in counts per minute.
A counter also clicks with no source near it. Cosmic rays, and the natural radioactivity of rocks, bricks, concrete and air, make a background of some tens of counts per minute. It is measured on its own, with the source taken away, and subtracted. Subtract rates, not counts, so the two readings may last different times:
$R_n = C/t − C_b/t_b$
where $C$ counts in time $t$ are the gross reading with the source and $C_b$ in $t_b$ the background. Background matters most when the source is weak: a sample counting little more than the background is mostly background.
Why counts scatter, and how long to count
Count a steady source for ten seconds, then again: you will almost never get the same number twice. The decays are random, so the number that falls in any window scatters about its average. Neither count is wrong; each is an honest sample. The surprise is how large and how predictable the scatter is. For counts of random events, the spread (the standard deviation) of a count $C$ is about its square root:
$σ_C ≈ \sqrt C$
So a single count carries its own error bar, $C ± \sqrt C$, and about two counts in three land within one $\sqrt C$ of the true average. The relative uncertainty is $\sqrt C/C = 1/\sqrt C$: 100 counts give 10 %, 10,000 give 1 %. Only the number of counts matters, not how long they took, so a weak source must simply be counted for longer.
That gives a rule for planning. To reach a relative uncertainty $f$ you need $1/\sqrt C = f$, that is
$C = 1/f^2$
counts, and at a count rate $R$ that takes a time $t = C/R$. Precision is expensive: halving the uncertainty takes four times as long. With a background to subtract, both counts contribute their scatter, and the time grows further.
Measuring a half-life
To measure a half-life in the lab, count a short-lived source in a series of equal time bins (say 30 s each), subtract the background from each, and plot the net counts against time. On an ordinary axis the points fall along a curve; on a logarithmic axis they fall along a straight line, because $\ln N = \ln N_0 − λt$ is linear in $t$. The line’s slope is $−λ$, and then $T_{1/2} = \ln 2/λ$. You can also read the half-life straight off the line: the time for the net counts to halve.
The points do not sit exactly on the line: each has its $\sqrt C$ scatter, drawn as an error bar. The early bins, with hundreds of counts, are known to a few percent; the late bins, with a handful, scatter widely. A good fit leans on the precise points, and ignores none. This sim’s fit treats each bin as a Poisson count and finds the decay that makes the whole set most likely.
Carbon-14 dating
Cosmic rays turn a little nitrogen in the upper atmosphere into carbon-14, which mixes into the carbon dioxide that plants take in and animals eat. While an organism lives it keeps the same tiny share of carbon-14 as the air. When it dies, no new carbon arrives, and its carbon-14 decays with $T_{1/2}$ = 5,700 y. The activity per gram of carbon then falls as $A = A_0e^{−λt}$, where $A_0$ is the activity per gram of living material. Solving for the time,
$t = (T_{1/2}/\ln 2)\ln(A_0/A)$
A sample at half the modern activity is one half-life old, 5,700 y; at a quarter, two half-lives. The natural logarithm is essential: decay is not a straight line in time, so “70 % gone” is not 70 % of anything simple.
The method has limits. It dates only things that were once alive. It assumes the air’s share of carbon-14 was the same in the past as now, which is only roughly true, so real laboratories correct their ages with calibration curves (this sim leaves calibration out), and by convention they quote “radiocarbon ages” with an older half-life, 5,568 y. And after about ten half-lives too little carbon-14 is left to count above the background: a dinosaur bone has none at all.
Part 2 · Going further
Math you’ll need: probability distributions (binomial, Poisson, Gaussian), adding variances, simple derivatives for propagating uncertainty and minimizing, and logarithms.
Decay as a Poisson process
A constant chance per unit time, the same for every nucleus and independent of all the others, defines a Poisson process. Its first consequence is that it has no memory. The probability that a nucleus survives a further time $t$ is $e^{−λt}$, whatever its age: one that has already survived three half-lives survives the next with probability ½, exactly like a new one.
Next, the survivors. Each of $N_0$ nuclei survives to time $t$ independently with probability $p = e^{−λt}$, so the number left is binomial, with mean $N_0p$ and standard deviation
$σ = \sqrt{N_0p(1 − p)}$
When $p$ is small, either because most nuclei have decayed or because we count the few that decay in a short window, $1 − p ≈ 1$ and the binomial becomes the Poisson distribution, whose variance equals its mean. The probability of exactly $k$ counts when the mean is $μ$ is
$P(k) = μ^ke^{−μ}/k!$
For large $μ$ this is close to a Gaussian with $σ = \sqrt μ$, which is where the $\sqrt C$ rule comes from. For small $μ$ it is lopsided. At $μ$ = 1 a third of all windows (36.8 %) hold no counts, so a single zero says little about whether a source is there, and $0 ± \sqrt 0$ is not an honest error bar. Below about 20 counts use the Poisson probabilities themselves.
Finally, the gaps between clicks. In a Poisson process at rate $R$ the waiting time to the next click is exponential: the chance that a gap is shorter than $τ$ is $1 − e^{−Rτ}$, and the mean gap is $1/R$. The most likely gap is the shortest, so random clicks bunch together, and a Geiger counter sounds irregular rather than steady. At 200 counts/s the mean gap is 5 ms, yet 1.8 % of gaps exceed 20 ms ($e^{−4}$).
The net rate and the best split
A net rate is a difference of two independent counts, and the variances of independent quantities add, even when the quantities subtract. A rate $C/t$ has variance $C/t^2$, so
$σ_{R_n} = \sqrt{C/t^2 + C_b/t_b^2}$
Subtracting a background always costs precision; it never cancels scatter. For a weak source on a large background the net rate can be lost entirely in the background’s own scatter.
Given a fixed total time $T = t + t_b$, how should it be shared? Write the variance as $R/t + R_b/t_b$ and minimize it: the derivative gives $R/t^2 = R_b/t_b^2$, so
$t/t_b = \sqrt{R/R_b}$
The busier count gets more time, but only as the square root of the ratio. Equal times are best only when the two rates are equal.
Fitting half-lives honestly
A least-squares line through ln(net counts) needs weights: the uncertainty of $\ln C$ is about $1/\sqrt C$, so each point should count in proportion to $C$. Equal weights let the noisy tail tilt the line. But two problems remain at low counts. Bins whose net count is zero or negative cannot be logged at all and must be dropped, which keeps only the upward fluctuations in the tail; and weights taken from the measured counts give more weight to bins that happened to fluctuate up. Both flatten the line, so the half-life comes out too long. How much depends on how far the tail sinks into the background: in the sim’s Monte Carlo (500 runs each), a strong Ba-137m sample of 40,000 nuclei comes out about 4 % long, but the Walkthrough’s weak sample of 3,000 nuclei, 10 % counted over 15 background counts a bin, comes out 59 % long on average.
The cure is to fit the counts themselves. In a Poisson maximum-likelihood fit, the expected counts in bin $k$ are $μ_k$ = (decays expected in the bin) + background, and the fit chooses the amplitude and $λ$ that make the observed counts most probable, maximizing $Σ[C_k\ln μ_k − μ_k]$. No logs are taken and no bin is dropped, and on that same weak sample it averages within 1 % of the true half-life. Its uncertainty comes from how sharply the likelihood peaks, and the sim’s Monte Carlo checks show it covers the true value about two times in three, as a ±1σ should. A χ² test on repeated counts asks the related question of whether the scatter is Poisson at all: for $n$ counts, $χ^2 = Σ(C − C̄)^2/C̄$ should be near $n − 1$; much larger suggests a drifting source or counter.
Dead time
After each count a Geiger tube is blind for a dead time $τ$, typically around 100 µs, while its discharge clears. Any particle arriving then is lost. In the nonparalyzable model the counter simply ignores arrivals during $τ$: it is live for a fraction $1 − mτ$ of the time, so the measured rate $m$ and true rate $n$ are related by $m = n/(1 + nτ)$, or
$n = m/(1 − mτ)$
The measured rate levels off at $1/τ$. In the paralyzable model every arrival, counted or not, restarts the dead time, so $m = ne^{−nτ}$. That rate peaks at $n = 1/τ$ and then falls: a very hot source can read lower than a moderate one. Real counters lie between the two models. Below $mτ$ ≈ 0.01 the correction is negligible; above 0.1 it matters, and the models begin to disagree.
Detection limits
Is a weak sample radioactive at all? Count it and a blank for the same time. If the blank gives $B$ counts, the net count of a sample with nothing in it scatters by about $\sqrt{2B}$. Currie’s rules, with a 5 % chance of a false alarm, set a critical level $L_C = 2.33\sqrt B$: a net count above it is declared a detection. But a sample whose true net count is exactly $L_C$ is missed half the time. The detection limit, the true net count detected 95 % of the time, is about twice as large:
$L_D = 2.71 + 4.65\sqrt B$
These are for a blank counted as long as the sample. If the blank runs for a time $t_b$ and the sample for $t$, the empty sample’s net count scatters by $\sqrt{B(1 + t/t_b)}$, so $L_C = 1.645\sqrt{B(1 + t/t_b)}$ and $L_D = 2.71 + 2L_C$: a long blank lowers the floor, a short one raises it. The Dating Lab uses this form.
A measurement below $L_C$ is reported as “not detected, less than” an upper limit, never as a negative or zero activity. For dating, the same limit sets the oldest sample a counter can date in a given counting time.
Chains and equilibrium
When a radioactive parent (1) decays into a radioactive daughter (2), the daughter is made at the rate $λ_1N_1$ and decays at $λ_2N_2$. Bateman’s solution for a pure parent at $t$ = 0 is $N_2 = λ_1N_1(0)(e^{−λ_1t} − e^{−λ_2t})/(λ_2 − λ_1)$. Three cases follow.
If the parent lives far longer than the daughter, the daughter’s activity rises until it equals the parent’s: secular equilibrium. Wash the daughter out and it grows back as
$A_2 = A_1(1 − e^{−λ_2t})$
reaching half the parent’s activity in one daughter half-life. This is a generator: a long-lived parent on a column, “milked” for its short-lived daughter, as cesium-137 gives barium-137m every few minutes. If the parent lives only a few times longer, the daughter overtakes it and then the two fall together with the parent’s half-life, in transient equilibrium, $A_2/A_1 = λ_2/(λ_2 − λ_1)$, slightly above 1. The medical Mo-99/Tc-99m generator is of this kind (65.94 h and 6.0067 h, ratio 1.10, reduced by Mo-99’s branching fraction to Tc-99m). If the daughter outlives the parent, there is no equilibrium: the parent dies out and the daughter decays on its own.
Ages and their uncertainty
A measured age inherits the uncertainty of the activity ratio $r = A/A_0$. Since $t = (T_{1/2}/\ln 2)\ln(1/r)$, a small change in $r$ changes $t$ by $(T_{1/2}/\ln 2)\,dr/r$, so to first order
$σ_t = (T_{1/2}/\ln 2)(σ_r/r)$
For carbon-14 the factor is the mean life, 8,223 y: every 1 % in the ratio is 82 y in the age, at any age. But $σ_r/r$ grows as the activity falls, because an old sample gives few counts over the same background, and counting longer helps only as the square root of the time. Eventually the net count drops below the detection limit and the honest result is “older than”: with a ratio limit of 0.01 that is 8,223 y × ln 100 = 37,900 y.
Reference
Common mistakes
Part 1
- Thinking a nucleus ages or becomes “due”: each has the same chance per second at any age.
- Thinking a sample is gone after two half-lives: each half-life halves what is left.
- Dividing by $2n$ instead of $2^n$ after $n$ half-lives.
- Using $1/T_{1/2}$ for $λ$, or a half-life in minutes or years with becquerels.
- Taking the count rate for the activity: a counter sees only a fraction $ε$.
- Forgetting the background, or subtracting counts taken over different times.
- Using $1/C$, or $\sqrt C$ alone, for the relative uncertainty instead of $1/\sqrt C$.
- Using a linear proportion, or log₁₀, for an age.
Part 2
- Using the Poisson $\sqrt{\text{mean}}$ for survivors when $p$ is not small.
- Trusting $±\sqrt C$ or a Gaussian at low counts, and reading a zero as “no source”.
- Subtracting variances when rates are subtracted.
- Splitting counting time in the ratio of the rates instead of its square root.
- Fitting ln(counts) at low counts, which runs long.
- Taking the critical level for the detection limit.
The equation sheet
| Part 1 | |
|---|---|
| decay law | $N = N_0(\tfrac12)^{t/T_{1/2}} = N_0e^{−λt}$ |
| half-life | $λ = \ln 2/T_{1/2}$ |
| activity | $A = λN$ |
| count rate | $R = εA$ |
| net rate | $R_n = C/t − C_b/t_b$ |
| counting scatter | $σ_C ≈ \sqrt C$ |
| counts for a precision $f$ | $C = 1/f^2$ |
| an age | $t = (T_{1/2}/\ln 2)\ln(A_0/A)$ |
| Part 2 | |
|---|---|
| survivors | $σ = \sqrt{N_0p(1 − p)}$ |
| Poisson | $P(k) = μ^ke^{−μ}/k!$ |
| a net rate’s σ | $σ_{R_n} = \sqrt{C/t^2 + C_b/t_b^2}$ |
| the best split | $t/t_b = \sqrt{R/R_b}$ |
| dead time | $n = m/(1 − mτ)$ |
| detection limit | $L_D = 2.71 + 4.65\sqrt B$ |
| regrowth | $A_2 = A_1(1 − e^{−λ_2t})$ |
| an age’s σ | $σ_t = (T_{1/2}/\ln 2)(σ_r/r)$ |
Symbols: $N$, $N_0$ nuclei now and at the start; $t$ time; $T_{1/2}$ half-life; $λ$ decay constant; $τ$ mean life (in Part 2, also the dead time); $A$ activity in Bq ($A_0$ modern, for dating); $ε$ efficiency; $R$ a count rate ($R_b$ background, $R_n$ net); $C$, $C_b$ counts in times $t$, $t_b$; $f$ a relative uncertainty; $p$ a survival probability; $μ$ a mean count; $k$ a number of counts; $m$, $n$ measured and true rates; $B$ blank counts; $L_C$, $L_D$ the critical level and detection limit; $A_1$, $A_2$ and $λ_1$, $λ_2$ a parent and its daughter; $r$ an activity ratio.
Sources and further reading
- OpenStax, University Physics Volume 3, chapter 10, “Nuclear Physics,” §10.3 Radioactive Decay, CC BY-NC-SA 4.0 (openstax.org): the Intro track follows its order and notation.
- G. F. Knoll, Radiation Detection and Measurement (Wiley), the chapter on counting statistics: net rates, the best split, dead time and χ² tests.
- L. A. Currie, “Limits for qualitative detection and quantitative determination,” Analytical Chemistry 40, 586 (1968).
- Half-lives: NUBASE2020 (Kondev et al., Chinese Physics C 45, 030001, 2021) and the NNDC’s NuDat.
- This sim’s Field Manual (H) for the model and its limits; Atomic Structure & Isotopes for isotopes and their notation.
Radioactive Decay & Counting Statistics v1.1 · © 2026 Dr. David Meier · CC BY-NC-SA 4.0 · glidepathscience.org