The Gold Standard of Randomness

A shuffled list, an unpredictable measurement and a simulated waiting time can all involve randomness, but their requirements differ. Begin by naming the possible outcomes, the probabilities assigned to them and what information is available before an outcome is observed.

A random model specifies uncertainty; it need not claim that the physical world has no causes. A deterministic process can be hard to predict because its starting conditions are poorly known. Conversely, an irregular-looking sequence can be completely predictable to someone who knows its generating rule.

The gold standard is a learner who can explain the model, calculate within it and recognise which conclusions depend on independence, equal probabilities or limited information. This root connects school probability to sampling, experimental allocation, algorithms and simulation.

Contents · Next chapter

For one ordinary six-sided die under a fair-face model, the elementary outcomes are 1 through 6, each with probability 1/6. The event ‘even’ contains 2,4,6 and has probability 3/6 = 1/2.

For two independent fair dice, use the 36 ordered pairs as equally likely outcomes. A total of 7 has six supporting pairs, so its probability is 6/36 = 1/6. The totals 2 through 12 are not themselves equally likely.

The representation matters. Counting eleven possible totals and assigning 1/11 to each would silently change the model. Before using ‘favourable outcomes divided by possible outcomes’, establish that the elementary outcomes being counted have equal probabilities.

Contents · Previous chapter · Next chapter

A random variable can have unequal outcome probabilities. An invented bag containing five red, three blue and two green tokens gives probabilities 0.5, 0.3 and 0.2 for the colour of a uniformly selected token. The token selection is uniform over ten distinct tokens, while colour is not uniform over three labels.

Fairness depends on the task. A teaching activity may require equal probabilities for each student; a survey may deliberately sample some groups more heavily and later use design weights. State the intended distribution instead of treating equal chances as the definition of every valid random process.

Probabilities must be non-negative and sum to one over a complete discrete sample space. A missing category or an overlapping classification can break that accounting even when the drawing or software appears plausible.

Contents · Previous chapter · Next chapter

Independence means that learning one event occurred does not change the probability of the other. For events A and B, P(A∩B) = P(A)P(B) is the corresponding product condition. Equal marginal probabilities alone do not establish independence.

Choose one fair bit B, then output the pair (B,B). Each position individually is zero or one with probability 1/2. Yet the only pairs are 00 and 11, each with probability 1/2; observing the first bit determines the second. Independent fair bits would give each of 00,01,10,11 probability 1/4.

Sampling without replacement also changes conditional chances. From three red and two blue tokens, the first red probability is 3/5; after a red is removed, the second red probability is 2/4. Thus the chance of two reds is (3/5)(2/4) = 3/10, rather than (3/5)².

Contents · Previous chapter · Next chapter

For n independent trials with constant success probability p, the count X has expectation np and variance np(1 − p). The expectation describes an average over repetitions of the experiment, not a compulsory result in one run.

In an invented batch of 100 independent trials with p = 0.60, the expected count is 60 and standard deviation is √24 ≈ 4.90. Counts such as 57 or 64 can occur under the model. A count does not become an error merely because it differs from 60.

The proportion X/n has variance p(1 − p)/n under the same assumptions. State those conditions before reusing the formula for correlated observations or changing probabilities. The expected-frequency explanation develops this school-level distinction.

Contents · Previous chapter · Next chapter

For repeated independent Bernoulli trials with fixed p, the law of large numbers says the observed proportion converges to p. More precisely, for any fixed positive tolerance, the probability of exceeding that tolerance tends to zero as the trial count grows. This does not promise that every longer prefix is closer than the preceding one.

Suppose the first ten tosses of an independent fair coin all happen to be heads. The next toss still has head probability 1/2 under that model. Future tails are not owed as repayment. A long-run proportion can approach one half because an early finite imbalance becomes small relative to the growing total.

A striking run may motivate checking whether the fair independent model is appropriate. That is a different question from predicting compensation while retaining the model. Keep ‘what the assumed model predicts’ separate from ‘whether the observations challenge the model’.

Contents · Previous chapter · Next chapter

Random selection operates on a defined list or frame. If eligible people are absent from that frame, a perfectly implemented draw cannot select them. If someone appears twice, their inclusion probability may change.

For an invented list of 40 distinct eligible names, a uniform draw of one name gives each probability 1/40. If five eligible names were never entered, those five have probability zero in that procedure. Randomness inside the incomplete list does not repair its coverage.

Sampling design also determines how group representation is handled. Stratified random sampling draws within defined groups according to a stated plan; a quota filled by conveniently available people uses a different selection mechanism. Continue through the sampling foundation article and the stratified-sampling guide.

Contents · Previous chapter · Next chapter

Random sampling selects units from a population. Random assignment allocates selected units to conditions. A study can use one, both or neither. The first supports generalisation under its design; the second supports a treatment comparison under the relevant experimental assumptions.

For 12 selected participants assigned six per group, one transparent plan is to choose six of the 12 uniformly for group A and assign the remainder to B. There are 12 choose 6 = 924 possible A-groups, each equally likely under this plan.

Random assignment balances characteristics in expectation across possible allocations, not necessarily exactly in one realised allocation. Record the allocation rule, preserve it in analysis and distinguish it from concealment of upcoming assignments or blinding after assignment. The randomisation branch develops the design.

Contents · Previous chapter · Next chapter

The modern in-place Fisher–Yates procedure fixes one position at a time. With positions numbered 0 to n−1, start at i = n−1, choose j uniformly from 0 through i inclusive, swap those positions, then decrease i until the process finishes.

For three distinct cards, there are three choices for the last position and two for the next. Each final ordering has probability (1/3)(1/2) = 1/6 when the required choices are uniform and independent. A self-swap is a valid possibility; forbidding it changes the distribution.

A loop that always swaps with any of all three positions uses a different process. With three such choices it has 27 equally likely choice paths, which cannot divide equally among six orderings. The Fisher–Yates specialist article supplies a longer trace and proof.

Contents · Previous chapter · Next chapter

A uniform input does not make every transformation uniform. Suppose a source produces each integer 0 through 9 with equal probability. Taking the remainder on division by 4 maps 0,4,8 to 0; 1,5,9 to 1; 2,6 to 2; and 3,7 to 3.

The resulting probabilities are 0.3,0.3,0.2,0.2. The remainder operation created unequal output counts because ten equally likely inputs do not divide evenly into four output classes.

For this teaching example, reject 8 and 9, redraw from the same independent uniform source, and take the accepted value modulo 4. Among the eight accepted values each remainder has two representatives, so the accepted result is uniform. The proof concerns the accepted distribution, not merely the appearance of a few outputs.

Contents · Previous chapter · Next chapter

A uniform number U on [0,1) can select the token colours from chapter 3: red for U < 0.5, blue for 0.5 ≤ U < 0.8 and green for 0.8 ≤ U < 1. Each interval’s length gives its probability in the ideal continuous model.

Endpoint conventions should leave no gap and no overlap. Real software has finite precision, so an implementation must respect its generator’s supported values and integer-selection rules. Rounding a continuous value to a category can distort edge probabilities.

More specialised methods solve different generation tasks. Alias tables support repeated draws from a discrete weighted distribution; Box–Muller transforms ideal independent uniform draws into normal variables. Knuth–Yao methods organise fair random bits for discrete sampling. Use the directory to study their assumptions and costs rather than assuming one method generates every desired distribution.

Contents · Previous chapter · Next chapter

A pseudorandom generator follows a deterministic update from stored state. Given the same algorithm, state and sequence of calls, its future outputs are determined. This reproducibility is useful for simulations and debugging.

A deliberately small classroom recurrence is x_next = (3x + 1) mod 8. Starting at 0 gives 1,4,5,0, then repeats. Its cycle has only four states even though eight possible state labels exist. A complicated-looking update does not guarantee a long period.

Finite deterministic state eventually repeats along an indefinitely continued trajectory. Period, output distribution and predictability are different properties. This toy recurrence is a counting example, not a recommendation for practical simulation or security.

Contents · Previous chapter · Next chapter

A seed initialises generator state through a particular implementation. It can allow a run to be repeated, but the result also depends on generator choice, software version, data, call order and sometimes parallel execution details.

Record those details alongside the seed. If an extra random draw is inserted early in a procedure, later outputs can change because the stream has advanced. Reusing one seed at the beginning of every replication can instead reproduce the same replication repeatedly.

Use a documented stream-management plan for repeated or parallel simulations. Reproducibility makes a calculation inspectable; it does not establish that its distribution, code or scientific model is correct. The research random-seeds article explores what a rerun can and cannot preserve.

Contents · Previous chapter · Next chapter

A physical source can produce variable measurements while retaining bias, dependence or predictable environmental effects. Quantifying uncertainty requires a source model and a statement of what an observer knows; visible noise amplitude is not an entropy measurement.

For a simple known binary distribution with P(1) = 0.75, the best fixed single-symbol guess is 1 and succeeds with probability 0.75. Its min-entropy is −log₂0.75 ≈ 0.415 bits. This describes that model, not a certification of a measured device.

Shannon entropy answers a different averaging question: −Σp log₂p is about 0.811 bits for probabilities 0.75 and 0.25. Neither number can simply be multiplied by a sample count when dependence changes the joint distribution. A deterministic transformation can reorganise uncertainty but cannot create it from a completely known constant input.

Contents · Previous chapter · Next chapter

A frequency test can inspect the proportion of ones; a runs test inspects consecutive groups; autocorrelation checks dependence at selected lags. Each targets particular departures from a specified null model. A sequence can pass one and fail another.

A small p-value measures how unusual the chosen statistic is under the null model. It is not the probability that a generator is faulty or the probability that the null is true. Many tests also create multiple opportunities for an apparently unusual result by chance.

Passing a test suite does not establish unpredictability or suitability for every task. NIST explicitly distinguishes statistical testing from a full security argument. Use the generator, entropy and testing branch for that boundary, and retain raw sequences and a stated test plan when investigating classroom examples.

Contents · Previous chapter · Next chapter

Monte Carlo uses sampled outcomes to estimate a quantity under a specified model. To estimate event probability p, simulate n relevant trials, count k occurrences and compute p̂ = k/n. The simulation must actually generate the intended process.

If a constructed run records 300 events in 1,000 independent trials, p̂ = 0.30. Its approximate Monte Carlo standard error is √[p̂(1−p̂)/n] ≈ 0.0145 under the independent Bernoulli model. This estimates computational sampling variation, not uncertainty from every possible source.

For the same underlying p, quadrupling n approximately halves that standard error because it scales as 1/√n. It does not eliminate bias caused by a wrong model or faulty implementation. Extremely rare events may require much more careful methods; observing zero occurrences does not prove their probability is zero.

The Monte Carlo foundation article develops numerical integration, uncertainty propagation and rare-event limitations.

Contents · Previous chapter · Next chapter

Successive samples from a Markov-chain method can be dependent by design. A series of readings from one slowly changing system can also be strongly correlated. Counting them as independent observations can overstate information and understate uncertainty.

The copied-bit pair in chapter 4 supplies two recorded values but only one fresh fair-bit choice. Similar reasoning explains why repeating an identical simulation with the same initial state adds no independent replication.

Low-discrepancy sequences deliberately spread points through a domain and support quasi-Monte Carlo integration. Their structure is useful for that task, but they are not interchangeable with independent random draws. Error assessment must match the actual construction. Continue with the Sobol’ and quasi-Monte Carlo branch.

Contents · Previous chapter · Next chapter

Reservoir sampling maintains a uniform sample from a stream whose final size may be unknown. Bootstrap procedures resample observed data to study variability under a resampling model. Permutation tests use allowed relabellings justified by a design or exchangeability condition.

Randomised algorithms use chance for a different purpose again. Freivalds’ method checks matrix products with random vectors; random contraction methods search for graph cuts; RANSAC samples candidate subsets for model fitting. Their guarantees depend on the precise selection rule and repeated-trial conditions.

Read each guarantee in its own language: inclusion probability, resampling validity, error probability or computational cost. The shared word ‘random’ does not make the methods equivalent. Use the directory to move into the specialist branch, then restate the sample space and assumptions before applying it.

Contents · Previous chapter · Next chapter

Try five checks. A: a fair coin has produced five heads; under independent fair tosses, find the next-head probability. B: a uniform integer 0 through 5 is reduced modulo 4; find each remainder probability. C: a random fair bit is copied ten times; decide whether those ten values are independent. D: distinguish randomly selecting students from randomly allocating already selected students. E: explain whether more simulations repair an omitted real-world mechanism.

Answers: A remains 1/2. In B, remainders 0 and 1 each occur with probability 2/6; remainders 2 and 3 each occur with probability 1/6. C has complete dependence once the first bit is known. D separates sampling from assignment. E reduces computational sampling error under the model, not the model’s omission.

Then change B’s source to a uniform integer 0 through 7. All four remainders now have probability 1/4 because each has two source values. The calculation should change for a reason the learner can explain.

This follows The Gold Standard of Education and the immutable Secondary 1 Mathematics foundation reference: define quantities, show the reasoning, correct a specific error and test a changed situation independently.

Contents · Previous chapter · Next chapter

Must random outcomes have equal probabilities? No. Uniformity is one particular distribution; weighted random selection can be intentional.

Does a random-looking sequence prove randomness? No. Appearance does not establish its generating mechanism, independence or unpredictability.

Does long-run convergence mean a short run must balance? No. Finite counts fluctuate, and independent future trials do not compensate deliberately for earlier results.

Does a fixed seed make a simulation invalid? No. Reproducible pseudorandom streams are useful when their properties and management suit the calculation. Reproducing the same run repeatedly is different from obtaining independent replications.

Does random selection guarantee a representative realised sample? No. Coverage, inclusion probabilities, sample size, non-response and chance imbalance still matter.

Can a classroom generator protect secrets? A teaching recurrence or balanced output sequence supplies no security guarantee. Security applications require their own appropriate construction and implementation evidence.

Where should I continue? Use the probability branches for concepts, sampling and randomisation for study design, shuffling and distribution-generation branches for algorithms, and Monte Carlo or resampling for computational inference. Each route begins with a stated model.

Contents · Previous chapter · Specialist reading directory