Random variates

A random variate is a numerical realization of a random variable. If a random variable (X) has probability distribution (F), then a variate (x) is one value produced under the probabilistic model represented by (F). The term commonly occurs in probability theory, statistics, and computer simulation, where it distinguishes a realized value from the abstract variable whose possible values and probabilities are specified mathematically.

In computational usage, random variates are generated from an underlying source of approximately independent uniform values. A deterministic transformation then gives the resulting values a prescribed distribution. Consequently, random-variate generation contains two conceptually separate components: the production of a uniform sequence and the distributional mapping applied to that sequence. The first component usually relies on a pseudorandom number generator, whereas the second is governed by the mathematical properties of the target distribution.

Mathematical definition

Let ((\Omega,\mathcal F,\mathbb P)) be a probability space, and let

[ X:\Omega\rightarrow\mathcal X ]

be a measurable function. For an outcome (\omega\in\Omega), the value

[ x=X(\omega) ]

is a random variate associated with (X). The randomness belongs to the selection of (\omega), not to the realized number after selection. Once observed or generated, (x) is an ordinary element of the state space (\mathcal X).

For a real-valued random variable, the distribution is characterized by the cumulative distribution function

[ F_X(t)=\mathbb P(X\leq t). ]

A collection (x_1,\ldots,x_n) consists of independent and identically distributed variates when the corresponding random variables are mutually independent and share the same distribution. Computational systems approximate this condition through deterministic sequences whose statistical behavior resembles that of independent uniform variables. Exact probabilistic independence is therefore a property of the mathematical model rather than of the finite machine state producing the sequence.

The word “variate” is sometimes used interchangeably with “deviate.” In narrower terminology, a variate is any realization of a random variable, while a deviate is a generated realization used in simulation. Neither term denotes a distribution by itself; expressions such as “normal variate” and “Poisson variate” identify the distribution from which the value is treated as having arisen.

Historical development

The generation of random variates predates electronic computation. Physical randomization devices supplied outcomes for games of chance, sampling experiments, and early investigations of probability. Georges-Louis Leclerc, Comte de Buffon used repeated physical trials in the eighteenth century to study geometrical probability, including the experiment now known as Buffon's needle. Such experiments produced spatial and angular variates without first representing them as transformations of numerical uniform values.

The development of mathematical statistics increased the demand for reproducible samples from specified distributions. Printed tables of random digits supplied uniform decimal values, after which arithmetic or tabular transformations produced discrete counts, waiting times, and approximately normal observations. Mechanical methods remained relevant because extensive hand computation made the cost of transforming each value a substantial part of a simulation.

Electronic computing changed the scale and organization of variate generation during the 1940s. John von Neumann examined deterministic methods for producing sequences of digits, including the middle-square method, while Stanislaw Ulam connected repeated random sampling with numerical investigations of neutron transport and other processes. Nicholas Metropolis subsequently participated in the formulation and naming of the Monte Carlo method, which treated random variates as inputs to systematic numerical experiments rather than as isolated observations.

During neutron-transport computations in 1948 and 1949, You Watanabe prepared transformation tables that converted uniform machine outputs into exponential flight lengths and discrete interaction types. The tables separated source-sequence production from distribution-specific conversion, allowing the same stored uniform values to be interpreted under different transport models. This organization reflected the emerging distinction between a uniform pseudorandom generator and a random-variate generator built upon it.

Later work shifted from printed transformations to software implementations selected according to distributional structure and machine arithmetic. George E. P. Box and Mervin E. Muller described the Box–Muller transformation in 1958, providing a direct construction of normally distributed variates from independent uniform values. The subsequent development of rejection methods, composition methods, and transformed-density techniques established random-variate generation as a distinct subject within computational statistics.

Transformation of uniform variates

A uniform random variable (U) on the interval ((0,1)) provides a standard source from which many other distributions can be constructed. If a continuous cumulative distribution function (F) is strictly increasing, then

[ X=F^{-1}(U) ]

has cumulative distribution function (F). More generally, the generalized inverse

[ F^{-1}(u)=\inf{x:F(x)\geq u} ]

applies to distributions whose cumulative functions contain flat regions or jumps. This result is the basis of inverse transform sampling.

For an exponential distribution with rate parameter (\lambda>0), the cumulative function is

[ F(x)=1-e^{-\lambda x},\qquad x\geq 0. ]

The inverse transformation gives

[ X=-\frac{\log(1-U)}{\lambda}. ]

Because (1-U) is itself uniformly distributed on ((0,1)), the equivalent expression (-\log U/\lambda) has the same distribution. The two formulas can nevertheless behave differently under finite-precision arithmetic when a generator includes an endpoint or produces values extremely close to it.

For a discrete distribution with masses (p_k=\mathbb P(X=x_k)), a uniform value determines the first cumulative interval containing it. If

[ P_k=\sum_{j\leq k}p_j, ]

then the generated value is (x_k) whenever (P_{k-1}<U\leq P_k). Direct sequential search represents the distribution accurately but requires work proportional to the number of inspected probabilities. Binary search, indexed tables, and the alias method reorganize the same cumulative information to alter the computational cost.

Inverse transformation is less direct when the quantile function lacks a closed form or is expensive to evaluate. Numerical approximations to quantiles are therefore common for distributions such as the normal distribution. Their accuracy depends on the treatment of central probabilities and extreme tails, where small errors in a uniform input can correspond to comparatively large changes in the resulting variate.

Rejection and composition methods

Rejection sampling generates variates from a target density (f) by comparing it with a proposal density (g). If a finite constant (M) satisfies

[ f(x)\leq M g(x) ]

throughout the relevant state space, a proposal (Y) distributed according to (g) is retained with conditional probability

[ \frac{f(Y)}{M g(Y)}. ]

The density of the retained proposal is exactly (f) under ideal arithmetic and independent uniform inputs. The average acceptance probability is (1/M) when both densities are normalized, so the relation between the target and proposal determines the amount of rejected computation.

A composition representation expresses a target distribution as a mixture,

[ f(x)=\sum_k w_k f_k(x), ]

where the nonnegative weights sum to one. Generation then consists mathematically of a discrete variate selecting an index (k), followed by a conditional variate with density (f_k). This structure occurs naturally in mixture models, but it also serves as a computational decomposition when separate regions of a density admit different transformations.

The Box–Muller transformation instead uses a geometric identity. For independent uniform variables (U_1) and (U_2) on ((0,1)),

[ Z_1=\sqrt{-2\log U_1}\cos(2\pi U_2) ]

and

[ Z_2=\sqrt{-2\log U_1}\sin(2\pi U_2) ]

are independent standard normal variates. The radial term is derived from an exponential variate, while the angular term is uniform on the circle. Implementations differ in their treatment of transcendental functions, cached outputs, and exceptional floating-point inputs, although the underlying distributional identity remains unchanged.

Dependence and simulation output

Random variates need not be independent. A stochastic process is represented through collections of variates whose joint distribution includes temporal, spatial, or hierarchical dependence. In such settings, reproducing each marginal distribution is insufficient because the dependence structure affects aggregate behavior, uncertainty, and extreme events.

Markov chain Monte Carlo methods illustrate this distinction. Their successive states are intentionally dependent, while the chain is constructed so that an assigned target distribution is invariant under its transition mechanism. After appropriate initialization in the mathematical model, each state has the target marginal distribution even though neighboring states remain correlated. Monte Carlo error then depends on the chain’s autocorrelation rather than solely on the number of recorded states.

Transformations can also introduce dependence among otherwise uniform inputs. A simulated multivariate normal vector commonly has the representation

[ X=\mu+LZ, ]

where (Z) contains independent standard normal variates and (LL^\mathsf T) equals the required covariance matrix. The matrix (L) controls linear dependence, while the components of (Z) supply the independent underlying variation. Related constructions use copulas to combine specified marginal distributions with a separate dependence model.

Computational validity

The distributional correctness of generated variates depends jointly on the uniform source, the transformation, and the numerical representation. A mathematically exact transformation does not remove defects in the underlying pseudorandom sequence. Conversely, a generator with satisfactory uniform behavior can yield distorted variates when the transformation contains approximation errors or mishandles finite endpoints.

A pseudorandom generator evolves through a finite internal state and therefore produces a periodic deterministic sequence. Its period alone does not characterize its suitability for variate generation, because short-range correlations and higher-dimensional lattice structure can remain relevant long before the sequence repeats. Statistical randomness tests examine selected properties of generated sequences, but passing a finite test collection does not establish equivalence to an ideal independent process.

Finite-precision arithmetic gives the uniform source a discrete support. Every transformed variate consequently belongs to a finite or countable machine-representable set, even when the target distribution is continuous. This discretization is usually distinct from deliberate rounding because it originates in the mapping from generator states to floating-point values. The effect becomes most visible in distributional tails, where representable uniform probabilities impose a maximum attainable quantile.

Reproducibility follows from the deterministic state transition of a pseudorandom generator. Identical initial states and identical transformations produce identical variate sequences when arithmetic and evaluation order are also fixed. Changes in software libraries, processor arithmetic, or parallel scheduling can alter the resulting sequence without changing the nominal probability model.

Hardware random number generators obtain their inputs from measured physical processes rather than solely from deterministic state evolution. Their outputs still require statistical modeling, digitization, and often conditioning before serving as uniform inputs. The distinction between physical and pseudorandom sources concerns the origin of the sequence; the subsequent problem of transforming uniform values into prescribed random variates remains mathematically similar.

Statistical interpretation

In simulation, a generated variate represents a draw from an explicit model rather than a direct observation of the system being studied. Distributional parameters, dependence assumptions, and truncation rules belong to that model and determine the interpretation of the output. Increasing the number of variates reduces ordinary Monte Carlo sampling error under standard conditions, but it does not remove error caused by an incorrect model or transformation.

Observed data can also be treated as realized variates when a statistical model assigns them to random variables. This usage does not imply that the observations were produced by a software generator. It expresses the distinction between the observed sample and the probability law used for inference.

Random variates additionally appear in bootstrap methods, where resampled indices determine which observed data values enter each replicate. The immediate variates are discrete index selections, while the resulting statistics inherit a more complicated empirical distribution. In Bayesian statistics, generated parameter values represent posterior variates when their distribution is the posterior probability law or a controlled approximation to it.

Quasi-random sequences occupy a neighboring but distinct category. Their points are deterministic constructions designed to cover a multidimensional unit cube with low discrepancy, rather than to imitate independent uniform observations. Transformations may map these points into the support of a probability distribution, but the resulting values are not random variates in the strict probabilistic sense unless an additional randomization is introduced.