Seventy-one resistant colonies do not necessarily represent seventy-one separate mutations. One early mutation can leave many descendants. A fluctuation test uses variation between independently grown cultures to investigate that hidden history.
This guide follows the reasoning rather than a laboratory protocol. We will compare two invented datasets with the same mean, build a simple family tree and examine what a zero-count culture can tell us. The central distinction is between a genetic event and the cells descended from it.
Continue from The Gold Standard of Genetics, within the Gold Standard of Education: learning that transfers into life and the How X Works learning library. All numerical datasets below are invented teaching examples.
Find the explanation you need
Start with the comparison design if the word “independent” feels vague. Use the family tree to understand jackpots, then test the assumptions behind estimating a mutation rate.
Full chapter index · Practice and checking · Genetics subject guide
Chapter contents
Frame the timing question · Chapters 1–2
Compare the evidence · Chapters 3–4
Recover the hidden history · Chapters 5–6
Inspect the assumptions · Chapters 7–8
Transfer and practise · Chapters 9–10
Questions, sources and connected reading · Chapters 11–12
The classic question is whether resistant variants arise before a selective challenge or only as a response to that challenge. Seeing survivors after selection does not, by itself, reveal when the relevant change occurred.
Luria and Delbrück investigated bacterial resistance to a virus in their 1943 study, Mutations of Bacteria from Virus Sensitivity to Virus Resistance. The important design idea was to compare outcomes across independently grown cultures.
An early event can be amplified by growth before the challenge is applied. A later event has fewer opportunities to leave descendants. If you only look at the final plate, those different histories are hidden.
Our question is deliberately narrow: how can the distribution of final counts help distinguish models of mutation timing? It is not a claim that every survival mechanism in every organism follows one simple model.
Separate the historical question from a universal slogan
The timing question has a useful everyday form: did an observation become visible because the process began then, or because a later condition revealed something already present? In the mutation setting, growth can preserve a history that the final count no longer displays directly. A comparison design tries to recover information about that history through patterns across cultures.
Keep three moments distinct: the beginning of the growth history under study, the selective challenge, and the observation of colonies. “Before selection” identifies the event’s position relative to the challenge. It does not mean the mutation existed forever, nor that it necessarily existed in the original founder. An event arising during the unselected growth period can still precede selection.
The classic evidence should not be inflated into the claim that environments can never influence mutation processes. Our task is narrower: compare specified models and identify why an early heritable event can generate variable descendant counts. A successful explanation names the alternatives tested and the conditions supporting the inference. This is stronger science than a memorable sentence that tries to settle every possible biological setting.
A mutation is a heritable genetic change. A mutant cell carries the relevant change. A descendant lineage may contain many mutant cells from one ancestral event. A counted colony is an observation filtered through growth, sampling and detection.
Those distinctions stop an easy mistake. Dividing resistant colonies by the final population can describe an observed mutant frequency under a specified assay. It does not automatically estimate the probability of a new mutation per cell division.
Selection is another process. A challenge can make a previously rare variant observable because cells differ in survival or growth under that condition. Observing the variant after a challenge does not establish that the challenge created it.
Write a chain on paper: event → descendants → sample → detectable colonies. Mark which links were observed and which were modelled. This is the evidence route the rest of the article will examine.
Give each quantity its own denominator
Suppose a fictional final population contains 1,000,000 cells, of which 100 are detectably mutant under a fully specified observation system. The final mutant frequency is 100/1,000,000 = 0.0001. That fraction alone does not tell us whether one early event or many late events produced the 100 cells. The same endpoint can conceal different event histories.
Now suppose someone divides 100 by ten generations and reports ten mutations per generation. This mixes a descendant count with a time-like index. Generations in a growing tree do not represent equal numbers of cell divisions: later generations contain more dividing cells. A valid event-rate denominator must correspond to the actual opportunity count in the chosen model.
Even a colony count needs an observation convention. Some cells may not yield a detectable colony under a particular assay, and a sampled fraction may omit relevant lineages. For our mathematical examples, the stated conditions make the count interpretable. Outside those conditions, record the detection process rather than treating the printed number as an unfiltered census of mutation events.
CHAPTER 3 OF 12 · Compare the evidence
3. Compare independent cultures with a shared culture
Imagine two designs with equal final population sizes and comparable detection conditions. In the first, eight cultures grow separately from small starting populations. In the second, one culture grows and then provides eight equal aliquots.
The first design has eight growth histories. The second samples one shared history eight times. Equal numbers of plates do not make the designs equivalent.
| Replicate | Independent growth histories | Aliquots from one grown culture |
|---|---|---|
| 1 | 0 | 4 |
| 2 | 0 | 7 |
| 3 | 1 | 8 |
| 4 | 1 | 9 |
| 5 | 2 | 10 |
| 6 | 2 | 12 |
| 7 | 3 | 14 |
| 8 | 71 | 16 |
| Total | 80 | 80 |
| Mean | 10 | 10 |
Both columns contain 80 colonies in total, so both means are 10. A report containing only those means would conceal the striking difference in spread.
The independent-history column contains two zeros and one large count of 71. The shared-history column varies more modestly. This is an invented pattern chosen to make the comparison visible. It is not a statistical test result and does not reproduce the original experiment’s data.
Locate where the histories separate
Imagine labelling the common starting material as the trunk of a tree. Separate cultures create branches whose later histories can differ. Aliquots drawn after one culture has grown create several observations of a later branch. They do not move the branching point back to the beginning. The chronology, rather than the number of containers at the end, determines which variation the comparison can reveal.
This also clarifies the role of the starting material. If a mutant lineage is already present before the separation being studied, several later cultures may inherit it. Independence is a property to justify at the relevant stage, not an adjective conferred by assigning different replicate numbers. The statistical model and the ancestry model must describe compatible experimental units.
Aliquots remain useful. They can help examine how sampling and observation vary within a shared grown population. That is a worthwhile question, just a different one from variation accumulated across separate growth histories. A well-designed report can retain both kinds of information while labelling them clearly. Combining them into one undifferentiated replicate count would hide the distinction the fluctuation comparison depends on.
CHAPTER 4 OF 12 · Compare the evidence
4. Calculate the spread without treating it as a verdict
For a transparent descriptive comparison, calculate the mean squared distance from each column’s mean. Here we divide by eight to describe these eight values, rather than using the seven-denominator sample-variance estimator.
For the independent-history column, the squared deviations sum to 4,260. Its descriptive variance is 4,260/8 = 532.5. For the shared-history column, they sum to 106, giving 106/8 = 13.25.
The arithmetic shows that equal means can coexist with very different fluctuation patterns. It does not establish a universal threshold for diagnosing mutation timing. A real analysis needs an appropriate probability model, enough independent cultures and information about growth and detection.
A simple post-challenge model with equal independent chances for many comparable cells can give approximately Poisson-like counts under restrictive conditions. That provides one possible comparison. The scientific job is to test the model’s conditions, not declare every irregular count a genetic discovery.
Check what the mean leaves out
Sort the independent-history values: 0, 0, 1, 1, 2, 2, 3, 71. Their median is 1.5, while their mean is 10. In the shared-history column the middle pair is 9 and 10, giving a median of 9.5. These descriptive summaries expose how strongly the large independent-history value pulls its mean upward. Neither the median nor the mean, by itself, estimates the mutation-event rate here.
If you use the usual seven-denominator sample variance instead, the values become 4,260/7 ≈ 608.57 and 106/7 ≈ 15.14. The numerical convention changes, but the large contrast in spread remains. Always say which convention you used. Otherwise two correct calculations can appear to disagree because the denominator was left implicit.
Deleting 71 because it looks inconvenient would erase the feature the timing model aims to explain. Keeping it without checking its identity and measurement quality would also be careless. The appropriate response is to preserve the observation, examine possible errors and analyse it with a model that can accommodate the process under investigation. An unusual value is neither automatically an error nor automatically a discovery.
CHAPTER 5 OF 12 · Recover the hidden history
5. Explain the jackpot through descendants
Consider a deliberately simplified growth tree. One cell doubles each generation, all cells divide equally, and no cells die. At generation 10 there are 1,024 cells.
| Mutation present in one cell at generation | Further doublings to generation 10 | Mutant descendants at generation 10 |
|---|---|---|
| 3 | 7 | 128 |
| 9 | 1 | 2 |
A mutation already present in one of the eight cells at generation 3 can lead to 128 descendants by generation 10. A mutation present in one cell at generation 9 leads to only two by generation 10.
Each history contains one mutation event. The final mutant counts differ sixty-four-fold. The multiplier comes from when the event entered the growing tree.
The word “jackpot” describes a large descendant count generated by a favourable early timing in this model. It does not imply that the organism intended the mutation. Nor does the largest final count necessarily identify the culture with the most independent mutation events.
Count divisions as well as generations
In the ideal binary tree, the first generation contains two cells, the second four, and generation ten contains 1,024. Starting from one cell and allowing no death, the total number of division events is 1 + 2 + 4 + … + 512 = 1,023. Each division increases the population by one, giving the same check: final population minus starting population.
Ten generations and 1,023 divisions are different denominators. A model that assigns a small event probability per division must account for many more opportunities late in growth, while an event occurring early has more time to amplify. These two effects coexist: many late opportunities and potentially very large early lineages. They help explain why simply counting final mutants does not count the original events.
For a concrete counterexample, 32 final mutant cells could come from one lineage present at generation five under the ideal assumptions. Another constructed history could supply several later lineages whose descendants also total 32. The endpoint alone cannot choose between them. The fluctuation approach gains information through a distribution across histories, rather than pretending that each endpoint has a uniquely recoverable family tree.
CHAPTER 6 OF 12 · Recover the hidden history
6. Use zero cultures for a different inference
A culture with zero detected mutants is informative under a model that connects absence of detected descendants with absence of successful mutation events.
Suppose the number of relevant successful events per culture is Poisson with mean m, and every such event would produce at least one detected mutant. Then the probability of a zero culture is P₀ = e−m. Rearranging gives m = −ln(P₀).
In our independent-history teaching data, two of eight cultures are zero. Treating 2/8 = 0.25 as an estimate gives m ≈ −ln(0.25) ≈ 1.39 events per culture under those assumptions.
That number is not 1.39 events per cell division. A division-based rate also needs the relevant number of divisions and a compatible growth model. Eight cultures provide a deliberately small illustration, with substantial sampling uncertainty. This calculation explains an estimator’s logic; it is not a recommended analysis of an actual experiment.
Understand what the logarithm recovers
The zero-class equation begins with an event model, not with the mutant-count mean. If successful event counts are Poisson with mean m, the chance of no successful event is e−m. Under complete detection, a zero culture corresponds to that no-event outcome. The logarithm then reverses the exponential relationship. Its usefulness comes from the match between those two meanings of zero.
The small dataset makes the sensitivity easy to inspect. One zero among eight gives m ≈ −ln(1/8) = 2.079. Two zeros give 1.386, and three give about 0.981. Changing the observed zero count by one therefore produces a substantial change in the estimate. Reporting many decimal places from eight cultures would hide how coarse the underlying evidence is.
Two boundary outcomes need care. If every culture is zero, the plug-in estimate is zero, but that is not proof that the true event rate is exactly zero. If none is zero, −ln(0) has no finite value, so this simple plug-in method cannot provide a finite estimate. Neither boundary justifies replacing an inconvenient zero fraction with an invented small number. A suitable inferential method must respect the actual information and design.
CHAPTER 7 OF 12 · Inspect the assumptions
7. Recognise when zero no longer means no event
Suppose only a small fraction of each culture is sampled. A culture can contain mutant descendants while the tested fraction contains none. A zero observation now has more than one possible explanation.
The same problem arises when a lineage is lost, grows poorly, fails to express the detectable phenotype in time, or is not recovered by the assay. The link between event and observed colony must be modelled.
Jones, Thomas and Rogers examine detection and estimation assumptions in Luria–Delbrück Fluctuation Experiments: Design and Analysis. Their work is a useful reminder that an estimator and its experimental design belong together.
For the classroom calculation, keep the assumption explicit: all successful events are detected. For actual data, investigate whether that is justified. Correcting the logarithm cannot repair an incorrect observation model.
Detection probability changes the equation
Here is a separate mathematical illustration. Suppose successful events are Poisson with mean m, and each event independently has probability d of leaving at least one detected descendant. Under this deliberately simplified event-detection model, detected events are Poisson with mean md, so the probability of an observed zero is e−md. A zero fraction of one quarter identifies md ≈ 1.386, rather than m alone.
If d were known to be one half in that model, the corresponding m would be about 2.773. This is a conditional demonstration of the ambiguity, not a correction formula for discarding half a real culture. A lineage can contain many descendants, and its chance of being detected after cell sampling depends on its size. Sampling half the cells does not generally make every mutation event’s detection probability one half.
The distinction prevents a tempting shortcut: doubling a mutation-rate estimate because half the volume was tested. A valid adjustment needs the sampling and lineage model appropriate to the data. The broader habit is simple: when an observation is zero, identify whether that means no event, no surviving lineage, no sampled descendant or no detectable response.
CHAPTER 8 OF 12 · Inspect the assumptions
8. Audit the comparison before interpreting the spread
Independent cultures should be comparable in the features the proposed model treats as constant. Different final population sizes can create different numbers of opportunities for mutation. Different mutant growth rates can change descendant counts. An initially present resistant lineage can also affect interpretation.
Retain the culture identity through every measurement. Technical repeats of one culture measure a different source of variation from separate biological growth histories. More measurements of the same history do not create more independent histories.
Consider alternative survival explanations too. A resistant phenotype and a transient survival state are not automatically the same biological phenomenon. The relevant genetic and phenotypic evidence depends on the organism and question.
The responsible conclusion follows the whole evidence chain. A fluctuation pattern can support pre-existing heritable variation under a suitable design without proving that every mutation rate is constant in every environment.
Compare models that could actually be distinguished
Suppose one group finishes with a much larger cell population than another. More mutation opportunities and different lineage sizes can change the outcome even if an underlying event probability per division is unchanged. A raw difference in final mutant counts would not isolate that probability. Retaining final population information helps expose this alternative explanation.
Likewise, a slower-growing mutant lineage can leave fewer descendants than the equal-growth tree predicts. That does not reverse the distinction between mutants and mutations; it adds another reason why the count-to-event relationship needs modelling. The question is which biological and observation assumptions matter enough to change the comparison.
A useful evidence table has one row per culture and separate fields for its growth history, relevant final-size information, observation fraction, count and data-quality notes. These fields need not all be known in a classroom exercise, but unknowns should remain visible. A statistical calculation can be exact for its supplied values while the larger inference remains conditional on quantities that were never measured. Identifying that limit is part of the finished answer.
CHAPTER 9 OF 12 · Transfer and practise
9. Transfer the reasoning to a changed design
A student proposes to grow one large culture, divide it into twenty portions and use the twenty results as independent mutation histories. What went wrong?
Division after growth changes the sampling stage, not the earlier ancestry. All twenty portions share the history before division. An early mutant lineage can be distributed among several portions.
Now consider cultures separated before the growth period of interest. They can accumulate distinct subsequent histories, although shared starting material still needs attention. The meaningful independence depends on when the histories separate.
This reasoning transfers beyond Genetics. Repeated measurements of one object and independent objects answer different questions about variability. The biological details differ, but the learning move stays the same: identify the unit that experienced the process before counting the number of observations.
Transfer the experimental-unit idea
Consider ten readings of one instrument exposed to one changing signal. Those readings can describe repeatability within that setting, but they do not provide ten independently manufactured instruments. Now consider ten instruments, each tested once. The second design addresses between-instrument variation more directly, although other design details still matter. The comparison is an analogy for experimental units, not a claim that instruments behave like cell lineages.
Return to Genetics and change the question. If the purpose is to understand counting variability from a fixed grown culture, aliquots may be exactly the appropriate unit. If the purpose is to study independent mutation histories during growth, separation must occur before the period whose histories are being compared. A good design begins with the question, not a universal preference for one type of replicate.
Ask a learner to draw the branching point and label when the selective condition enters. Then ask what changes if the same final counts are reassigned to a different branching diagram. The arithmetic remains fixed, but its interpretation changes. This is a demanding transfer check because it requires connecting a number to the process that generated it.
CHAPTER 10 OF 12 · Transfer and practise
10. Practise with a family tree and a zero fraction
First, use the ideal binary tree again. One mutant cell is present at generation 5; the population is followed to generation 10. How many descendants does that lineage have? There are five doublings, so 2⁵ = 32, assuming equal growth and no loss.
Second, imagine 20 genuinely independent, comparable cultures, of which 10 have no detected mutants. Under the complete-detection zero-class model, P₀ ≈ 0.50 and m ≈ −ln(0.50) ≈ 0.693 successful events per culture.
Finally, change only the observation condition: half of each culture is discarded before testing. Can you reuse 0.693 without modification? No. A zero in the tested half need not mean no event in the whole culture. The detection model has changed.
A strong solution gives the calculation and names the assumption that permits it. That is more valuable than recalling one mutation-rate formula without knowing what its zeros represent.
Draft a conclusion with an explicit scope
For the eight-culture illustration, a careful answer can say: the datasets have equal means but different spreads; early mutation followed by growth provides a mechanism for large between-culture fluctuations; the result is informative only within a suitable comparison design and observation model. It should not claim a precise division-based mutation rate from the descriptive variance alone.
Try a second independent exercise. Six of twelve comparable cultures have zero detected mutants under the complete-detection Poisson event model. The plug-in estimate is again −ln(0.5) ≈ 0.693 events per culture. The equality with the twenty-culture exercise comes from the same zero fraction. It does not mean the two datasets contain the same amount of information about that fraction.
Finally, imagine a report that gives only “mean resistant count: ten” and omits every individual culture count. List what cannot now be reconstructed: the zero fraction, the median, the jackpot pattern and the descriptive variance. The missing records prevent several checks that an extra decimal place on the mean could never restore. Preserving individual observations is therefore part of preserving the scientific question.
CHAPTER 11 OF 12 · Questions and continuation
11. Questions about counts and causation
Is the average mutant count the mutation rate?
No. Descendant amplification makes counts sensitive to event timing. Frequency, events per culture and events per division are different quantities.
Does one jackpot prove pre-selection mutation?
No. A single large count needs technical and biological investigation. The inference rests on a designed comparison and its distribution across independent histories.
Are aliquots useless?
No. They can reveal sampling and assay variation. They simply do not replace independent growth histories for the same question.
Does “random mutation” mean all DNA changes are equally likely?
No. The timing argument here does not require every position or change to have equal probability. State which form of randomness the model actually uses.
Should learners perform this with unknown microbes?
The learning tasks here use invented counts and a mathematical tree. They require no culture handling. Actual microbial work belongs within the relevant supervised laboratory arrangements.
Can two experiments have the same zero fraction but different counts?
Yes. Under the simple zero-class estimator, the plug-in m depends on the zero fraction, so rearranging nonzero counts without changing zeros leaves that estimate unchanged. It can still change other features relevant to a fuller analysis. An estimator intentionally uses particular information; it does not make the rest of the observations irrelevant to checking the model.
Does a successful event mean every mutation is detected?
No. The phrase defines the event class used in the illustrative model. A mutation that leaves no surviving detectable lineage lies outside that observation pathway. If the intended scientific rate includes those events, the model must connect them to what can be observed. Naming a selected class carefully prevents a result about observable resistance from becoming a claim about all genetic changes.
How should a small dataset be reported?
Keep the individual counts, experimental-unit definitions and assumptions available. Give descriptive summaries without treating their precision as evidence of certainty. If a rate is estimated, distinguish its units from mutant frequency and provide an uncertainty analysis suited to the actual design. Our eight values teach why those requirements matter; they do not establish a minimum sample size for every investigation.
Original question and analytical limits
Luria and Delbrück, 1943: the original study record identifies the historical paper and its DOI. Jones, Thomas and Rogers, 1994: design and analysis discusses estimation and incomplete counting. These links were checked on 21 October 2026. Our count arrays, family tree and exercises are original illustrations, not historical experimental results.
Understand the biological foundation
Return to The Gold Standard of Genetics. Use How Science Works: Genetics for the connection between inheritance, variation and causal evidence.
The Gold Standard of Microbiology places the organism and laboratory question in a wider subject. The Gold Standard of Evolution connects variation with population change.
Follow samples and interpretation
Laboratory Technicians connects sample identity with traceable results. Biotechnology follows discovery into reproducible evidence. Mitochondrial Bottlenecks and Sampling Variation examines another case where a mean conceals variation; its biological process is different from this growth model.
Keep learning transferable
The next useful task is to explain which stage created each source of variability. Continue through the Gold Standard of Education and the How X Works library. “Gold standard” here is an educational benchmark rather than an official laboratory certification.
Compare a different sampling problem
To see how combining samples can change an expected proportion without changing any organism’s genotype, continue with The Gold Standard of Genetics | The Wahlund Effect and the Population Hidden Inside a Sample.
Extend the evidence check
Selection and detection sit between a genetic event and the observed counts. Compare that evidence pathway with Segregation Distortion and the Survival Filter Behind an Inheritance Ratio, where transmission and later survival must also be distinguished.