Sliced Latin hypercube designs with arbitrary run sizes

Jin Xu, Xu He, Xiaojun Duan, Zhengming WangarXiv 2019 · DOI · arXiv

Abstract

Latin hypercube designs (LHDs) are a cornerstone of computer experiments, providing optimal univariate stratification for sampling high-dimensional spaces. Sliced Latin hypercube designs (SLHDs) extend this utility by allowing a large LHD to be partitioned into smaller, valid LHD sub-designs. Historically, SLHDs have been constrained by rigid requirements for slice sizes, often requiring them to be equal or follow specific ratios. This paper presents the first construction method for SLHDs that allows for arbitrarily chosen run sizes for each slice. Additionally, an algorithm is provided to reduce spurious correlations within these designs, ensuring high statistical efficiency across diverse computational workflows.

Introduction

In fields ranging from climate modeling to aerospace engineering, researchers rely on complex computer simulations to predict real-world outcomes. Because these simulations are often computationally expensive, researchers use Latin Hypercube Designs (LHDs) to select the most informative sample points. An LHD ensures that each input variable is sampled evenly across its entire range, a property known as univariate stratification.

Modern computational environments, however, often involve heterogeneous resources—such as a mix of high-performance clusters and local workstations—or multi-fidelity modeling where some simulations are faster than others. Sliced Latin Hypercube Designs (SLHDs) were developed to address these needs by allowing a global experiment to be divided into batches (slices). Yet, until now, the mathematical frameworks for SLHDs required slices to be of equal size, limiting their practical application in unbalanced or asynchronous environments.

The Challenge of Rigid Slicing

Standard SLHDs are designed such that the union of all slices forms a large LHD, and each individual slice is also an LHD. The primary bottleneck in existing literature has been the "equal run size" constraint. In a real-world scenario, a researcher might have three available computing nodes with different capacities, requiring batches of, for example, 50, 30, and 20 runs. Previous construction methods could not guarantee that these unequal batches would maintain the necessary Latin Hypercube properties. This paper fills this gap by providing a mathematical recipe for slices of any size, enabling "hardware-aware" experimental design.

Methodology

1. Midpoint Construction Algorithm

The authors propose a two-stage approach centered on "midpoint" designs. While many LHDs use randomized points within bins, the authors found that targeting the exact centers of bins makes the design more mathematically tractable for arbitrary slicing.

The construction uses a specific rounding function, $\delta_i$, to assign these midpoint values to different slices ($G_j$). This ensures that both the individual slices and the combined aggregate design achieve optimal one-dimensional uniformity. This deterministic approach to bin assignment allows the researcher to specify any set of integers for slice sizes, provided they sum to the total number of runs.

2. Correlation Reduction

A common pitfall in LHDs is the accidental correlation between columns (e.g., variable X increasing alongside variable Y), which can bias results. To mitigate this, the authors adapted an iterative algorithm that utilizes linear regression residuals to "clean" the columns. This process significantly reduces correlations while strictly preserving the LHD structure. The algorithm is computationally efficient, typically converging within ten iterations.

Experimental Evaluation

To validate the proposed designs, the researchers conducted numerical integration tests using standard mathematical benchmark functions:

  • $f_1(x)$: A log-product function involving five variables.
  • $f_2(x)$: A function involving the log of the sum of inverse square roots of variables.

These functions serve as proxies for complex computer codes. The proposed designs (SLH and the correlation-reduced CSLH) were compared against standard randomized LHDs and Flexible Sliced Designs (FSD).

Key Findings

  • Superior Accuracy: The CSLH design consistently produced the lowest Root-Mean-Square Error (RMSE) in numerical integration, demonstrating that the combination of arbitrary slicing and correlation reduction does not sacrifice statistical power.
  • Robustness to Failure: A critical advantage of this method is its resilience. In a simulated "computer failure" scenario—where a random slice of data is lost—the remaining data from an SLH design maintains its balance and stratification. In contrast, losing a random chunk of a standard LHD often results in "clumpy" data and a significant loss of accuracy.
  • Efficiency: The method allows for asynchronous experimentation, where sample sizes can be tailored to a specific computational budget or hardware availability on the fly without compromising the integrity of the statistical design.

Conclusion and Future Work

This work introduces a flexible framework for designing computer experiments that can adapt to the realities of modern computing. By allowing for arbitrary slice sizes, the proposed SLHDs provide a robust solution for multi-fidelity modeling and distributed computing.

While the midpoint design strategy offers significant advantages in tractability and variance reduction, future research may explore incorporating jittered randomization for applications that require it. Additionally, while the current algorithm optimizes for column-wise correlations, further extensions could incorporate "space-filling" properties, such as Maximin distance, to further enhance performance in high-dimensional spaces.

Summary updated

Abstract

Latin hypercube designs achieve optimal univariate stratifications and are useful for computer experiments. Sliced Latin hypercube designs are Latin hypercube designs that can be partitioned into smaller Latin hypercube designs. In this work, we give, to the best of our knowledge, the first construction of sliced Latin hypercube designs that allow arbitrarily chosen run sizes for the slices. We also provide an algorithm to reduce correlations of our proposed designs.
BibTeX
@misc{xu2019sliced,
  author = {Jin Xu and Xu He and Xiaojun Duan and Zhengming Wang},
  title = {{Sliced Latin hypercube designs with arbitrary run sizes}},
  year = {2019},
  eprint = {1905.02721},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.02721},
}

Loading related content...

Figure