Latin hypercube sampling to reduce the variance in Monte Carlo integration: a practical guide and an application to gravitational-wave astronomy
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: We're looking at the paper: "Latin hypercube sampling to reduce the variance in Monte Carlo integration".
Jocelyn: Here's the gist of the paper. Latin hypercube sampling to reduce the variance in Monte Carlo integration: a practical guide and an application to gravitational-wave astronomy provides a practical introduction to Latin hypercube sampling (LHS) and its extension based on orthogonal arrays (OA-LHS) as strategies to reduce the variance in Monte Carlo integration by stratifying the sampling distribution In applied statistics, likelihoods involving high-dimensional integrals are often approximated via Monte Carlo summation Controlling the resulting numerical variance is crucial for reliable inference.
Vera: Monte Carlo summation is a powerful technique for approximating high-dimensional integrals, often scaling more favorably than conventional grid-based quadrature.
Jocelyn: In this context, Latin hypercube sampling (LHS) is a popular space-filling strategy for drawing randomized multidimensional samples with uniform stratification in each one-dimensional marginal.
Vera: Generalizations such as orthogonal-array Latin hypercube sampling (OA-LHS) extend this stratification to higher-dimensional marginals.
Title and authors: Vera: The authors are Matteo Boschini and Davide Gerosa from the University of Milano-Bicocca, and they’re showing how these sampling strategies work in practice using gravitational-wave astronomy as an example.
Jocelyn: It’s interesting because it takes something theoretical about Latin hypercube sampling and applies it directly to a real problem, which is estimating black hole population properties.
Subrahmanyan: It shows how these mathematical tools can actually help us deal with the complexity of those likelihood integrals, which are often way too big for standard Monte Carlo summation to handle efficiently.
The paper's summary: Vera: In essence, the paper explains that Latin hypercube sampling is a good way to draw random samples in a high-dimensional space while making sure each dimension gets a fair representation across the whole range.
Jocelyn: They also talk about orthogonal array Latin hypercube sampling as an extension of this, which helps further reduce variance by stratifying the distribution in higher dimensions.
Subrahmanyan: The core idea they are pushing here is that under certain conditions, this LHS approach is mathematically guaranteed to have a variance no greater than a standard Monte Carlo estimator.
The paper's improvements: Vera: They show that for a square-integrable function, the variance of the LHS estimator can be written as one over N times sigma squared minus X sub j times sigma squared j.
Jocelyn: And they point out that using an orthogonal array Latin hypercube sampling with strength two gives us an extra benefit because it can remove those two-factor interactions in the variance calculation.
Subrahmanyan: That removal of pairwise interactions is where the real potential is, because those are often what make the standard Monte Carlo method struggle when you have complicated dependencies between parameters.
Conclusion: Vera: So to wrap up, they find that even though the gains they found in this specific gravitational-wave astronomy application were small—only about one point seven percent of the variance was additive—the mathematical guarantee that LHS has an asymptotic variance no greater than standard Monte Carlo is important.
Jocelyn: And because constructing these designs doesn't cost much time compared to doing actual simulations, they recommend using the orthogonal array Latin hypercube sampling for drawing those next set of LIGO or Virgo injections.
Subrahmanyan: From my side, it reinforces that space-filling designs are a useful tool for population inference; it means that when we are modeling things like merging black holes, we can use these stratification methods to get a more precise view of the selection effects without needing extra data.
Matteo Boschini, Davide Gerosa
Dipartimento di Fisica ’G. Occhialini, University of Milano-Bicocca · INFN Sezione di Milano-Bicocca
astro-ph.IM, astro-ph.CO, astro-ph.HE, gr-qc
Submitted: 2026-11-02
Updated: 2026-11-02
Comments: 10 pages, 4 figures, 1 table
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 91/100
The gist: Latin hypercube sampling to reduce the variance in Monte Carlo integration: a practical guide and an application to gravitational-wave astronomy provides a practical introduction to Latin hypercube…
Key concepts
- Latin Hypercube Sampling (LHS)
- LHS is a space-filling sampling strategy that draws points in a d-dimensional unit hypercube while ensuring uniform distribution across each one-dimensional projection. It achieves this by dividing each dimension into N bins and selecting one bin uniformly at random for each point, leading to statistically better coverage than simple random sampling.
- Orthogonal Array Latin Hypercube Sampling (OA-LHS)
- OA-LHS extends LHS by using orthogonal arrays to stratify the sampling distribution across higher-dimensional marginals. This technique is powerful because it can remove two-factor interactions in the variance decomposition of the integral, leading to potentially much greater variance reduction than standard LHS.
- Monte Carlo Integration Variance
- Monte Carlo integration approximates high-dimensional integrals by summing random samples. The resulting numerical error, or variance, determines how reliable the estimate is. Stratified sampling methods like LHS and OA-LHS are used to mathematically control and reduce this variance compared to standard Monte Carlo methods.
- Gravitational-Wave Astronomy Application
- The paper applies these sampling techniques to estimate selection functions in gravitational-wave astronomy, specifically for merging black holes and neutron stars. By reducing the variance in likelihood estimates, researchers can obtain more precise population properties from the same number of simulations.
Terminology
Summary
Latin hypercube sampling to reduce the variance in Monte Carlo integration: a practical guide and an application to gravitational-wave astronomy provides a practical introduction to Latin hypercube sampling (LHS) and its extension based on orthogonal arrays (OA-LHS) as strategies to reduce the variance in Monte Carlo integration by stratifying the sampling distribution In applied statistics, likelihoods involving high-dimensional integrals are often approximated via Monte Carlo summation Controlling the resulting numerical variance is crucial for reliable inference. p. 0
Introduction to Sampling Strategies
Monte Carlo summation is a powerful technique for approximating high-dimensional integrals, often scaling more favorably than conventional grid-based quadrature. In this context, Latin hypercube sampling (LHS) is a popular space-filling strategy for drawing randomized multidimensional samples with uniform stratification in each one-dimensional marginal. p. 0 Generalizations such as orthogonal-array Latin hypercube sampling (OA-LHS) extend this stratification to higher-dimensional marginals. p. 0 Crucially, under appropriate regularity conditions, the asymptotic variance of an LHS-based Monte Carlo estimator is mathematically guaranteed to be no greater than that of a standard Monte Carlo estimator. p. 0
Mechanism of Latin Hypercube Sampling
LHS is a technique that allows drawing N points in the unit hypercube in d dimensions [0, 1)d while guaranteeing uniform distributions in each one-dimensional projection. A practical construction of an LHS design with N points in the unit hypercube [0, 1]d is as follows. p. 1 The construction involves dividing each dimension into N disjoint bins and selecting one bin uniformly at random in each dimension, independently across dimensions, and drawing a point uniformly within the resulting d-dimensional cell. p. 1 The resulting LHS elements are given by ui j = πj(i) + Vi j / N, where Vi j ∼ U[0, 1), i ∈ R. p. 1
Variance Reduction in Integrals
The variance reduction of an LHS-based Monte Carlo estimator depends on the structure of the integrand. For a square-integrable function in d variables f(x), the variance can be decomposed into additive and interaction terms. p. 4 The asymptotic variance for LHS is Varh ˆILHSi ≃ 1/N (σ squared − X j σ 2j). p. 4 Using an OA-LHS of strength 2 offers additional gain as it can remove the two-factor interactions. p. 4
Application to Gravitational-Wave Astronomy
In gravitational-wave astronomy, hierarchical Bayesian statistics is used to estimate the population properties of merging black holes and neutron stars. Evaluating this likelihood requires the computation of Nobs + 1 integrals. p. 0 The selection function ξ(λ) is given by ξ(λ) = Z ppop(θλ) p(detθ) dθ. p. 1 The paper investigates the effect of LHS and OA-LHS on the variance of this selection-function estimate. p. 1
Conclusion on Practicality
Although the gains found are marginal, LHS-based estimators are mathematically guaranteed to have an asymptotic variance no greater than that of standard Monte Carlo. The computational cost of constructing these designs is negligible relative to evaluating gravitational-wave injections. p. 0 Therefore, the authors recommend drawing the next set of LIGO/Virgo injections using OA-LHS. p. 7 The limited gains can be understood through a variance-based sensitivity analysis showing that only 1. p. 77% of the variance is additive, while the remaining 85. p. 79% resides in interactions of third and higher order. p. 7
Acknowledgements
We thank Tristan Bruel, Jack Heinzel, Matthew Mould, Alexandre Toubiana, and Salvatore Vitale for discussions. M.B. and D.G. are supported by ERC Starting Grant No. p. 8 945155–GWmining. p. 8 Computational work was performed at CINECA with allocations through INFN and the University of Milano-Bicocca. p. 8
References
The paper cites numerous works, including those by Caflisch (1998), Stein (1987), and Owen (1992b). The references also include recent work on gravitational-wave astronomy, such as Abac et al. p. 0 (2025) and Essick et al. p. 1 (2023). The authors also cite Finn and Chernoff (1993) for semi-analytic approaches. p. 1 The references cover the mathematical foundations of orthogonal arrays, such as Hedayat et al. p. 2 (1999). The paper also references work on machine-learning emulators and posterior sampling designs. p. 1 The references include the LIGO/Virgo/KAGRA Collaboration documentation and papers detailing waveform models like IMRPhenomXPHM. p. 6 The list of references is extensive, including authors like Zhou et al. (2013) and Yeung et al. p. 7 (2026). The paper concludes with a comprehensive list of cited literature.
Summary
Latin hypercube sampling to reduce the variance in Monte Carlo integration provides a practical guide and an application to gravitational-wave astronomy, showing that strength-2 orthogonal arrays can reduce variance by orders of magnitude when the integrand is dominated by additive contributions or pairwise interactions between coordinates. This method is particularly useful in GW population inference where selection effects are a bottleneck. p. 8
How the training signal is built
The selection function integral is approximated using Monte Carlo sums, where the estimator is ξ(λ) ∼ ppop(θiλ) pinj(θi) p(detθi) + θi∼pinj. The injections are drawn from a reference distribution pinj used during LIGO/Virgo’s fourth observing run. p. 6 Each of the d = 15 components of θ corresponds to one coordinate of the unit hypercube. p. 6 The physical parameters are obtained using inverse cumulative distribution functions, as summarized in Table 1. p. 6
Where the gains come from
The variance reduction is quantified by comparing the variances of the estimators with a hypothesis test. The ratio R = Var cLHS,OA-LHS/Var cMC follows an F distribution with (R−1, R−1) degrees of freedom. p. 7 For the selection function estimates shown in Figure 4, the ratios RLHS = 0. p. 796 and ROA-LHS = 0. p. 789 at N = 994009 injected binaries. p. 7 The authors find no statistically significant difference between the variances of the three sampling methods for this application. p. 7
Conclusions
When applied to Monte Carlo integration, space-filling designs can reduce the asymptotic variance of the resulting estimator. LHS enforces one-dimensional stratification, while orthogonal arrays extend this property to higher-dimensional marginals. p. 8 Strength-2 orthogonal arrays can be constructed algorithmically and used to stratify all two-dimensional projections. p. 8 The resulting OA-LHS estimator can reduce variance by orders of magnitude when the integrand is dominated by additive contributions or pairwise interactions between coordinates. p. 8 These methods offer potential improvements “for free,” without requiring additional injections. p. 8
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved sampling for Monte Carlo integration by using Latin hypercube sampling (LHS) to reduce variance in likelihood estimates, as LHS is
mathematically guaranteed to be no greater than that of a standard Monte Carlo estimator
when the integrand is monotone. This allows AI systems performing hierarchical Bayesian inference to obtain more reliable posterior structures from the same number of samples. -
Improved selection function estimation in gravitational-wave astronomy by using orthogonal array Latin hypercube sampling (OA-LHS) to reduce variance, as OA-LHS can
remove those from pairwise interactions
yieldingorders-of-magnitude improvements over plain Monte Carlo estimators for suitable integrands.
This means AI systems modeling population properties of merging black holes and neutron stars can achieve more precise estimates of selection effects. -
Enhanced uncertainty quantification in hierarchical inference by using LHS or OA-LHS to strategically stratify sampling, as this
removes variance contributions from individual coordinates
(LHS) orpairwise interactions
(OA-LHS). This capability allows AI systems to identify which parameter dependencies contribute most to the numerical noise in likelihood estimates and focus computational effort where it is most needed.
Sources
- Advanced LIGO
- GWTC-4.0: Population Properties of Merging Compact Binaries
- GWTC-5.0: Population Properties of Merging Compact Binaries
- Prospects for Observing and Localizing Gravitational-Wave Transients with Advanced LIGO, Advanced Virgo and KAGRA
- Advanced Virgo: a 2nd generation interferometric gravitational wave detector
- Planck 2015 results. XIII. Cosmological parameters
- LHS in LHS: A new expansion strategy for Latin hypercube sampling in simulation design
- A neural network emulator of the Advanced LIGO and Advanced Virgo selection function
- Black Hole Coagulation: Modeling Hierarchical Mergers in Black Hole Populations
- Semianalytic Sensitivity Estimates for Catalogs of Gravitational-Wave Transients
- Compact Binary Coalescence Sensitivity Estimates with Injection Campaigns during the LIGO-Virgo-KAGRA Collaborations' Fourth Observing Run
- Precision Requirements for Monte Carlo Sums within Hierarchical Bayesian Inference
- Accuracy Requirements for Empirically-Measured Selection Functions
- Observing binary inspiral in gravitational radiation: One interferometer
- Normalizing Flow Emulator for Gravitational Wave Detection Probability in the LIGO/Virgo/KAGRA O4 Observing Run
- Quick recipes for gravitational-wave selection effects
- Gravitational-wave selection effects using neural-network classifiers
- Hierarchical Inference of Binary Neutron Star Mass Distribution and Equation of State with Gravitational Waves
- When (not) to trust Monte Carlo approximations for hierarchical Bayesian inference
- A physically modelled selection function for compact binary mergers in the LIGO-Virgo O3 run and beyond
Related papers
- On-sky demonstration of a vector Zernike wavefront sensor in a cascaded adaptive optics system
- Euclid Data Release 1 (DR1): Calibration strategy and observations
- The General Antiparticle Spectrometer (GAPS) Antarctic Balloon Payload
- Detection and excision of radio-frequency interference in radio arrays
- Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
- Ushering in sub-arcsecond angular resolution observatories in the far-infrared: The case for a double-Fourier interferometer