Optimal sensor placement under model uncertainty in the weak-constraint 4D-Var framework

Alen Alexanderian, Hugo Díaz, Vishwas Rao and Arvind K. Saibaba · 2025

Read Optimal sensor placement under model uncertainty in the weak-constraint 4D-Var framework by Alen Alexanderian, Hugo Díaz, Vishwas Rao and Arvind K. Saibaba (2025) 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-2502-00150.txt.

Introduction

Alen Alexanderian, Hugo Díaz, Vishwas Rao, and Arvind K. Saibaba ask where to place sensors when the model feeding on their data is itself imperfect. Data assimilation merges sensor readings with a model to track a changing system. The weak-constraint version of the 4D-Var method admits that the model's dynamics carry error. The authors build a placement rule on expected information gain, measured as the distance between the forecast distribution and the updated one.

Under Gaussian errors and linear dynamics they derive an exact formula for that gain. Randomized shortcuts make the formula cheap to evaluate on large problems, and alternate forms add flexibility. A column-selection algorithm then picks near-optimal sensor sets without ever forming the model's adjoint. Tests in one and two spatial dimensions show the placements work well.

Catalog record

CREATORAlen Alexanderian, Hugo Díaz, Vishwas Rao and Arvind K. Saibaba
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-00150

Browse the full library · About the Tower of Babel

Full text

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

Optimal sensor placement under model uncertainty in the weak-constraint 4D-Var framework Thanks: The work of AA was supported in part by the US National Science Foundation (NSF) grant DMS-2111044. HD and AKS were supported by the Department of Energy, Office of Science, Advanced Scientific Computing Research (ASCR) Program through the award DE-SC0023188. AKS was also supported by the NSF, in part, through the award DMS-1845406. VR was supported by the U.S. Department of Energy, Office of Science, ASCR Program under contracts DE-AC02-06CH11357 and DE-SC0023188. Alen Alexanderian, Hugo Díaz, Vishwas Rao and Arvind K. Saibaba Abstract . In data assimilation, the model may be subject to uncertainties and errors. The weak-constraint data assimilation framework enables incorporating model uncertainty in the dynamics of the governing equations. We propose a new framework for near-optimal sensor placement in the weak-constrained setting. This is achieved by first deriving a design criterion based on the expected information gain, which involves the Kullback-Leibler divergence from the forecast prior to the posterior distribution. An explicit formula for this criterion is provided, assuming that the model error and background are independent and Gaussian and the dynamics are linear. We discuss algorithmic approaches to efficiently evaluate this criterion through randomized approximations. To provide further insight and flexibility in computations, we also provide alternative expressions for the criteria. We provide an algorithm to find near-optimal experimental designs using column subset selection, including a randomized algorithm that avoids computing the adjoint of the forward operator. Through numerical experiments in one and two spatial dimensions, we show the effectiveness of our proposed methods. 1991 Mathematics Subject Classification Primary: 65M32, 62K05; Secondary: 65F40, 65F60. Alen Alexanderian {}^{{\hrefmailto:[email protected]}*1} Hugo Díaz {}^{{\hrefmailto:[email protected]}1} Vishwas Rao {}^{{\hrefmailto:[email protected]}2} and Arvind K. Saibaba {}^{{\hrefmailto:[email protected]}1} 1 Department of Mathematics, North Carolina State University, USA 2 Argonne National Laboratory, USA (Communicated by Handling Editor) 1. Introduction Accurately estimating the state of complex dynamical systems is essential in fields such as meteorology, oceanography, and climate modeling. Data assimilation techniques, which combine noisy observations with mathematical models, play a critical role in reconstructing and predicting the state of these systems. These methods have been extensively studied and developed, with notable contributions in both theoretical and applied contexts and advancements in geophysical and climate-related applications [ 20 , 22 , 24 , 26 , 34 ] . The Strong-Constraint Four-Dimensional Variational (SC4D-Var) method assumes perfect model dynamics. However, in many applications, the mathematical model does not perfectly represent the underlying physics, and a failure to account for this can be detrimental to the data assimilation results [ 41 ] . The Weak-Constraint 4D-Var (WC4D-Var) method [ 38 , 46 , 53 ] addresses these shortcomings by accounting for uncertainties in both the model and the observations. This uncertainty is variously referred to as model error, model misspecification, or model uncertainty, depending on the field of study, though they all reflect limitations or inaccuracies in the forward model. However, throughout this paper we refer to these as model errors. In many applications of interest, the data is in the form of sensor measurements. A question related to data assimilation is: How should one optimally collect data under physical or budgetary constraints? This falls under the purview of optimal experimental design (OED). To enable optimal sensor placement, we assume that we have identified n_{s} candidate sensor locations and we have to choose k optimal sensor locations. There are many challenges that need to be addressed to determine the optimal placement of these sensors. First, we need to consider a notion of optimality—a design criterion—that accounts for the model errors. Second, even a single evaluation of the design criterion is computationally expensive. And finally, exploring the space of \binom{n_{s}}{k} possibilities is computationally infeasible in practical settings. The goal of this article is to address all these aforementioned challenges. To the first point, while standard approaches for OED ignore model error, we develop a notion of optimality in the WC4D-Var framework and study the behavior of the optimality criterion as the model error vanishes. To the second point, we develop several efficient methods for evaluating the objective function, and to the third point, we build on our recent work on OED using column subset selection [ 19 ] and adapt it to the specific criterion that we derive and time-dependent problems. The specific contributions of this article are listed next, followed by a discussion on related work in the literature. Contributions This article introduces a novel framework for near-optimal sensor placement within the context of the WC4D-Var method. By explicitly incorporating model errors, we develop a method to optimize sensor locations for enhanced data assimilation. Our key contributions include: (1) New optimality criterion : In Section 3 , we derive a novel OED criterion for the WC4D-Var framework that quantifies the expected information gain (EIG) from the forecast to the posterior distribution. We provide a closed-form expression for this criterion when the forecast and the posterior distributions are Gaussian, and which involves the log-determinant of the prior and posterior covariance matrices. (2) Convergence to standard SC4D-Var criterion : We show that as the model error vanishes, the optimality criterion for WC4D-Var converges to the criterion used in the standard SC4D-Var formulation, ensuring consistency between the methods, see Section 3.4 . (3) Alternative forms of the criterion : In Section 3.2 , we derive alternative formulations of the WC4D-Var EIG criterion (each formulation is equivalent, up to an additive constant, independent of the data). Notably, all these formulations are interconnected through appropriate Schur complements and provide different computational benefits. (4) Matrix-free method for evaluating the design criterion : In Section 4 , we propose a matrix-free approach for evaluating the WC4D-Var design criterion, based on stochastic trace estimators, and the Lanczos method to enable efficient computation of the different formulations of the criterion. (5) Near-optimal sensor placement : In Section 5 , we propose a novel approach for near-optimal sensor placement by combining ideas from low-rank tensor approximations with the column subset selection problem, specifically leveraging the Golub-Klema-Stewart (GKS) method. We introduce a randomized variant of the GKS method that is computationally efficient and can be implemented in a matrix-free and adjoint-free method. This makes the methods applicable to a wide variety of scenarios, especially involving legacy implementations. (6) Numerical experiments : In Section 6 , we conduct comprehensive numerical experiments in both one and two spatial dimensions to evaluate the performance of our proposed methods. The one-dimensional setting enables a direct comparison with the optimal sensor placement, which is computationally intractable in higher dimensions due to its combinatorial nature. The numerical experiments show that the algorithms are accurate and computationally efficient. Related work Model errors arises in various contexts, including statistical inference, Bayesian inverse problems, regression analysis, and data assimilation. In this section, we first review approaches for handling model errors in general, then focus on strategies that specifically address model errors in the context of OED. In statistical inference, model errors have been extensively studied. Specifically, within the framework of Maximum Likelihood Estimation (MLE), several studies have identified conditions under which the estimator converges to a well-defined limit, even when the assumed probability model is incorrect, see [ 52 ] and references therein. A related approach involves the

Read the complete text