Is it random sampling?

It can be randomized, but it imposes stratification and is not a set of independent ordinary random draws.

Does it eliminate the curse of dimensionality?

No. It improves marginal coverage while high-dimensional interactions can remain difficult.

Can it use nonuniform distributions?

Yes. Equal-probability strata can be transformed through suitable inverse distribution functions, with additional care for dependence and discrete values.

AI Glossary / Statistics and probability / Intermediate

Latin hypercube design

Definition

A Latin hypercube design is a sampling arrangement that spreads a chosen number of runs across every input variable's range or probability distribution. For n runs, each input is divided into n strata, and every stratum contributes exactly one value. Values from different inputs are paired through permutations to form the runs. The defining balance is along each individual input axis; it does not require every multidimensional combination to be represented. The design is widely used in computer experiments, uncertainty propagation and surrogate-model training.

What it means

An expensive simulation may have many inputs but a limited evaluation budget. Sampling independently can leave large gaps along some inputs, while a full grid grows rapidly with dimension. Latin hypercube sampling provides deliberate marginal coverage without requiring a full grid. In a probability setting, strata usually contain equal probability rather than equal physical width. For an exploratory engineering design, inputs may instead be scaled across specified bounds. These interpretations must be distinguished: uniform coverage over an allowable range does not establish that all values are equally likely in reality.

How it works

Choose the number of runs, input definitions and distributions or bounds. Partition each unit-interval coordinate into equal strata. Select one point within each stratum, often randomly or at its midpoint, then independently permute the coordinates to assemble runs. Transform coordinates into physical inputs using scaling or inverse distribution functions. Inspect projections, constraints and dependence before evaluating the model. Additional optimization can improve separation or discrepancy, but it can also alter sampling properties. Correlated inputs and forbidden combinations need deliberate treatment; arbitrary rejection or coordinate movement can destroy the original Latin structure.

Three practical examples

Simulation planning: distribute a limited set of thermal-model evaluations over uncertain material and operating inputs.
Uncertainty analysis: sample input probability distributions before studying variation in a model's output.
Surrogate training: create an initial collection of model runs for fitting an emulator, then add evaluations where validation reveals weak coverage. None of these uses makes the original simulation an accurate representation of the physical system.

Think of it this way

Think of dealing one card from every numbered band of each input, then shuffling which cards travel together. Every band is represented, so one variable cannot accidentally be ignored over a wide interval. However, the shuffled combinations may still leave important joint regions unexplored. Seeing every temperature band and every pressure band does not mean seeing every temperature-pressure pairing.

Worked example

Consider four runs with two normalized inputs. Set u = 0.125, 0.375, 0.625, 0.875 and pair them with v = 0.625, 0.125, 0.375, 0.875. Both coordinates contain one midpoint from each quarter of the unit interval, so this is a centered Latin hypercube. The mean of the additive response f(u,v) = 2u + v is exactly 1.5, matching its average over the unit square. Now consider g(u,v) = uv. Its four values sum to 1.125, so the sample mean is 0.28125. The exact uniform-square average is 0.25, giving an error of 0.03125. This original calculation shows both marginal balance and the absence of a guarantee for interactions. Four points cover only four of the sixteen quarter-by-quarter cells. Because these points were chosen deterministically at centers, this example is not a claim of unbiasedness under a randomized sampling scheme.

Illustrative case study

In a fictional case, Meridian Cooling has enough computing budget for forty initial equipment simulations. Its analysts define six inputs and generate a Latin hypercube within physically feasible ranges. They discover that some combinations violate a known pressure constraint, so they redesign the sampling procedure and document how the feasible domain changes. After fitting an emulator, they test additional locations and add runs near regions with high prediction error. The company describes the design as an efficient starting sample, not proof that all failure conditions have been covered.

Common misconceptions and mistakes

A Latin hypercube is not a full factorial design, and marginal stratification does not ensure uniformly good two-dimensional or higher-dimensional coverage. It does not automatically preserve a desired correlation structure. Centered points and randomly jittered points have different inferential properties. A single sample's ordinary independent-observation standard error may be inappropriate because the sampling scheme introduces dependence. Adding arbitrary extra points usually breaks exact stratification for the enlarged sample. Strong interactions, discontinuities and rare failure regions can require specialized designs or further targeted evaluation.

Business use

Businesses use the approach when each simulation or assessment is costly and input coverage matters. The decision should begin with the purpose: integration, sensitivity analysis, optimization or emulator training. These objectives can favor different design refinements. Record input transformations, constraints, randomization settings, run identifiers and failed evaluations. Compare alternatives using output-relevant validation, not only a visually attractive scatterplot. A design that covers assumed inputs well still inherits errors in those assumptions, the simulator and the subsequent analysis.

Check your understanding

In the four-run example, does one value in each quarter of both inputs imply coverage of all sixteen joint cells?

Answer: No. Only four combinations are evaluated. The additive response happens to average exactly, while the interaction response averages 0.28125 instead of 0.25. Marginal balance alone does not guarantee exact integration or comprehensive joint coverage.

Related terms

Learn next

Learn how randomized stratification changes variance estimates, then compare Latin hypercubes with scrambled low-discrepancy sequences. Study projection quality and sequential design before extending a sample, and examine how constraints and correlated inputs affect the target distribution.

Sources

SciPy, LatinHypercube documentation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.qmc.LatinHypercube.html . Provides the defining stratification, construction options and implementation limitations. The worked example and fictional case above are original illustrations; no published performance result is claimed for them.

Previous: Lateral sound Next: Latin hypercube sampling