MST-Direct: Matching via Sinkhorn Transport for Multivariate Geostatistical Simulation with Complex Non-Linear Dependencies

Tchalies Bachmann Schmitz · 2026

Read MST-Direct: Matching via Sinkhorn Transport for Multivariate Geostatistical Simulation with Complex Non-Linear Dependencies by Tchalies Bachmann Schmitz (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-18036.txt.

Introduction

Schmitz proposes MST-Direct, a method for multivariate geostatistical simulation based on optimal transport. Geological variables often link in bent, bimodal, or step-like ways. Gaussian copula and LU methods assume straight-line links and lose those patterns. MST-Direct uses the Sinkhorn algorithm to match full joint distributions at once. All variables move as one vector, so spatial structure and cross-variable links are kept together.

Catalog record

CREATORTchalies Bachmann Schmitz
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22675
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.18036

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MST-Direct: Matching via Sinkhorn Transport for Multivariate Geostatistical Simulation with Complex Non-Linear Dependencies Tchalies Bachmann Schmitz Abstract Multivariate geostatistical simulation requires the faithful reproduction of complex non-linear dependencies between geological variables, including bimodal distributions, step functions, and heteroscedastic relationships. Traditional methods such as Gaussian Copula and LU Decomposition assume linear correlation structures and fail to preserve these complex joint distribution shapes. We propose MST-Direct (Matching via Sinkhorn Transport), a novel algorithm that leverages Optimal Transport theory with the Sinkhorn algorithm to directly match multivariate distributions while preserving spatial correlation structures. The method processes all variables simultaneously as a single multidimensional vector, using relational matching with k-nearest neighbor adjacency to maintain spatial coherence. We validate our approach through comprehensive experiments comparing MST-Direct with Gaussian Copula and LU Decomposition on synthetic data with five types of complex bivariate relationships: step function, Gaussian mixture, sinusoidal, random branching, and heteroscedastic. Results demonstrate that MST-Direct achieves perfect shape preservation (100% histogram similarity across all relationship types) while maintaining competitive variogram reproduction. The method represents a significant advancement for applications requiring accurate modeling of non-linear geological dependencies. I Introduction Geostatistical simulation is a fundamental technique for generating multiple realizations of geological models that replicate spatial features observed in experimental data [ 1 ] . These techniques have been extensively applied in geological, hydrological, and environmental applications to simulate physical properties such as mineral grades, porosity, and permeability. The statistical analysis of multiple stochastic realizations is crucial for uncertainty quantification and risk management in decision-making processes [ 2 ] . Traditional multivariate simulation methods, such as Sequential Gaussian Simulation (SGS) with collocated cokriging [ 1 ] and Gaussian Copula approaches [ 3 ] , rely on linear correlation coefficients to characterize dependencies between variables. However, many real-world geological phenomena exhibit complex non-linear relationships that cannot be adequately described by linear correlation. Examples include: • Bimodal distributions : Where two distinct populations exist with different correlation patterns, common in mixed lithological environments. • Step functions : Where the relationship between variables changes abruptly at threshold values, typical of mineralization boundaries. • Heteroscedastic relationships : Where variance changes as a function of the predictor variable, observed in grade-thickness relationships. • Sinusoidal patterns : Exhibiting periodic behavior in the correlation structure, common in cyclical depositional environments. The Stepwise Conditional Transformation (SCT) proposed by Leuangthong and Deutsch [ 3 ] and the Projection Pursuit Multivariate Transform (PPMT) by Barnett et al. [ 4 ] have addressed some limitations of traditional methods. More recently, Direct Multivariate Simulation [ 5 ] extended these concepts to handle complex joint distributions. However, these approaches still rely on transformations to Gaussian space, which may not fully preserve complex joint distribution shapes. Optimal Transport (OT) theory provides a mathematical framework for comparing and mapping probability distributions [ 6 , 7 ] . The Sinkhorn algorithm [ 8 , 9 ] enables efficient computation of entropy-regularized optimal transport, making it practical for large-scale applications. Recent advances have demonstrated the power of OT in machine learning tasks including domain adaptation, generative modeling, and distribution matching. In this work, we propose MST-Direct (Matching via Sinkhorn Transport - Direct), a novel algorithm that applies optimal transport theory to multivariate geostatistical simulation. Our key contributions are: 1. A novel application of optimal transport to geostatistical simulation that preserves complex non-linear dependencies. 2. A relational matching approach using k-nearest neighbor adjacency to maintain spatial structure. 3. Direct multivariate processing that handles all variables simultaneously without iterative sequential updates. 4. Comprehensive experimental validation demonstrating 100% shape preservation across diverse relationship types. II Related Work II-A Traditional Multivariate Geostatistics The foundational work in geostatistics by Matheron established the theoretical basis for spatial interpolation and simulation [ 1 ] . Multivariate extensions typically rely on the multivariate Gaussian distribution assumption, which enables analytical tractability but limits the representation of complex dependencies [ 2 ] . Sequential Gaussian Simulation (SGS) with collocated cokriging remains the standard approach for multivariate problems. However, it requires iterative simulation of each variable conditional on previously simulated values, potentially accumulating errors and biases. II-B Copula Methods Copula theory, introduced by Sklar [ 10 ] , separates the marginal distributions from the dependence structure. Gaussian copulas have been applied to geostatistical simulation [ 3 ] , but their reliance on linear correlation limits their ability to capture non-linear dependencies. II-C Multivariate Transforms The Stepwise Conditional Transformation (SCT) [ 3 ] decorrelates variables through conditional distributions. The Projection Pursuit Multivariate Transform (PPMT) [ 4 ] uses iterative projections to achieve multivariate Gaussianity. Direct Multivariate Simulation [ 5 ] generalizes these concepts but faces computational challenges in high dimensions. II-D Optimal Transport Optimal transport theory, dating back to Monge in 1781 and formalized by Kantorovich in 1942, provides a principled framework for mapping between probability distributions [ 6 ] . The Sinkhorn algorithm [ 8 ] for computing doubly stochastic matrices was recognized by Cuturi [ 9 ] as an efficient solver for entropy-regularized optimal transport. Recent work [ 7 ] has established OT as a fundamental tool in machine learning and data science. III Methodology III-A Problem Formulation Consider two spatial random fields X(\mathbf{u}) and Y(\mathbf{u}) defined over a domain D\subset\mathbb{R}^{2} , where \mathbf{u}=(u_{x},u_{y}) represents spatial coordinates. Each field has its own spatial correlation structure characterized by a variogram model: \gamma_{X}(h)=\frac{1}{2}\mathbb{E}[(X(\mathbf{u})-X(\mathbf{u}+\mathbf{h}))^{2}] (1) \gamma_{Y}(h)=\frac{1}{2}\mathbb{E}[(Y(\mathbf{u})-Y(\mathbf{u}+\mathbf{h}))^{2}] (2) where h=\|\mathbf{h}\| is the lag distance. The joint distribution p(X,Y) exhibits complex non-linear dependencies that cannot be characterized by linear correlation alone. The objective is to generate realizations that simultaneously preserve: 1. The marginal variogram structures \gamma_{X}(h) and \gamma_{Y}(h) 2. The complex joint distribution shape p(X,Y) III-B Optimal Transport Framework Optimal transport seeks a coupling between source distribution P and target distribution Q that minimizes the total transport cost. Given n source points and n target points, the discrete optimal transport problem finds a coupling matrix \mathbf{M}\in\mathbb{R}^{n\times n} solving: \min_{\mathbf{M}}\sum_{i,j}C_{ij}M_{ij} (3) subject to the marginal constraints: \sum_{j}M_{ij}=\frac{1}{n},\quad\sum_{i}M_{ij}=\frac{1}{n},\quad M_{ij}\geq 0 (4) where \mathbf{C} is the cost matrix, typically based on distance or similarity. III-C Entropy-Regularized Sinkhorn Algorithm The Sinkhorn algorithm [ 9 ] solves an entropy-regularized version of the optimal transport problem: \min_{\mathbf{M}}\sum_{i,j}C_{ij}M_{ij}-\frac{1}{\beta}H(\mathbf{M}) (5) where H(\mathbf{M})=-\sum_{i,j}M_{ij}\log M_{ij} is the entropy of \mathbf{M} and \beta>0 controls the regularization strength. Larger \beta values yield solutions closer to the unregularized optimal transport. The solution is obtained through iterative row and column normalization: \mathbf{M}^{(t+1)}=\text{diag}(\mathbf{r}^{(t)})\,\mathbf{K}\,\text{diag}(\mathbf{c}^{(t)}) (6) where \mathbf{K}=\exp(-\beta\mathbf{C}) is the Gibbs kernel, and \mathbf{r} , \mathbf{c} are the row and column scaling vectors updated alternately to satisfy the marginal constraints. III-D Relational Matching To preserve spatial structure, we extend the standard Sinkhorn algorithm with relational matching. Given adjacency matrices \mathbf{A}_{V} and \mathbf{A}_{U} representing k-nearest neighbor relationships, the cost matrix incorporates both feature similarity and relational consistency: C_{ij}=-\mathbf{v}_{i}^{\top}\mathbf{u}_{j}-\lambda\sum_{k}A_{V}(i,k)\cdot M_{kl}\cdot A_{U}(j,l) (7) where \mathbf{v}_{i} , \mathbf{u}_{j} are the normalized feature vectors, \lambda is the relational weight parameter, and \mathbf{M} is the current coupling estimate. The relational term encourages that if point i is a spatial neighbor of point k in the source, then their matched points should also be neighbors in the target. The combined objective becomes: M_{ij}\propto\exp\left(\beta\cdot\mathbf{v}_{i}^{\top}\mathbf{u}_{j}+\lambda\cdot(\mathbf{A}_{V}\mathbf{M}\mathbf{A}_{U}^{\top})_{ij}\right) (8) This is solved iteratively, alternating between Sinkhorn normalization and relational term updates. III-E MST-Direct Algorithm The complete MST-Direct algorithm processes all variables simultaneously. Algorithm 1 presents the pseudocode. Algorithm

Read the complete text