Showing 1–6 of 6 results for author: Mosegaard, K

Inconsistency and Acausality of Model Selection in Bayesian Inverse Problems

Authors: Klaus Mosegaard

Abstract: Bayesian inference paradigms are regarded as powerful tools for solution of inverse problems. However, when applied to inverse problems in physical sciences, Bayesian formulations suffer from a number of inconsistencies that are often overlooked. A well known, but mostly neglected, difficulty is connected to the notion of conditional probability densities. Borel, and later Kolmogorov's (1933/1956)… ▽ More Bayesian inference paradigms are regarded as powerful tools for solution of inverse problems. However, when applied to inverse problems in physical sciences, Bayesian formulations suffer from a number of inconsistencies that are often overlooked. A well known, but mostly neglected, difficulty is connected to the notion of conditional probability densities. Borel, and later Kolmogorov's (1933/1956), found that the traditional definition of conditional densities is incomplete: In different parameterizations it leads to different results. We will show an example where two apparently correct procedures applied to the same problem lead to two widely different results. Another type of inconsistency involves violation of causality. This problem is found in model selection strategies in Bayesian inversion, such as Hierarchical Bayes and Trans-Dimensional Inversion where so-called hyperparameters are included as variables to control either the number (or type) of unknowns, or the prior uncertainties on data or model parameters. For Hierarchical Bayes we demonstrate that the calculated 'prior' distributions of data or model parameters are not prior-, but posterior information. In fact, the calculated 'standard deviations' of the data are a measure of the inability of the forward function to model the data, rather than uncertainties of the data. For trans-dimensional inverse problems we show that the so-called evidence is, in fact, not a measure of the success of fitting the data for the given choice (or number) of parameters, as often claimed. We also find that the notion of Natural Parsimony is ill-defined, because of its dependence on the parameter prior. Based on this study, we find that careful rethinking of Bayesian inversion practices is required, with special emphasis on ways of avoiding the Borel-Kolmogorov inconsistency, and on the way we interpret model selection results. △ Less

Submitted 23 October, 2023; v1 submitted 10 August, 2023; originally announced August 2023.

Comments: The paper replaces arXiv:2308.05858v1 which contained incorrectly normalized distributions in two important counterexamples on hierarchical Bayes and trans-dimensional inversion. In the new, corrected version of the paper (where a key counter-example on transdimensional inversion is further expanded) the conclusions remain the same as in the original paper

Evolution of the Stress and Strain field in the Tyra field during the Post-Chalk Deposition and Seismic Inversion of fault zone using Informed-Proposal Monte Carlo

Authors: Sarouyeh Khoshkholgh, Ivanka Orozova-Bekkevold, Klaus Mosegaard

Abstract: When hydrocarbon reservoirs are used as a CO2 storage facility, an accurate uncertainty analysis and risk assessment is essential. An integration of information from geological knowledge, geological modelling, well log data, and geophysical data provides the basis for this analysis. Modelling the time development of stress/strain changes in the overburden provides prior knowledge about fault and f… ▽ More When hydrocarbon reservoirs are used as a CO2 storage facility, an accurate uncertainty analysis and risk assessment is essential. An integration of information from geological knowledge, geological modelling, well log data, and geophysical data provides the basis for this analysis. Modelling the time development of stress/strain changes in the overburden provides prior knowledge about fault and fracture probability in the reservoir, which in turn is used in seismic inversion to constrain models of faulting and fracturing. One main problem in solving large scale seismic inverse problems is high computational cost and inefficiency. We use a newly introduced methodology -- Informed-proposal Monte Carlo (IPMC) -- to deal with this problem, and to carry out a conceptual study based on real data from the Danish North Sea. The result outlines a methodology for evaluating the risk of having subseismic faulting in the overburden that potentially compromises the CO2 storage of the reservoir. △ Less

Submitted 31 August, 2021; originally announced September 2021.

Informed Proposal Monte Carlo

Authors: Sarouyeh Khoshkholgh, Andrea Zunino, Klaus Mosegaard

Abstract: Any search or sampling algorithm for solution of inverse problems needs guidance to be efficient. Many algorithms collect and apply information about the problem on the fly, and much improvement has been made in this way. However, as a consequence of the the No-Free-Lunch Theorem, the only way we can ensure a significantly better performance of search and sampling algorithms is to build in as much… ▽ More Any search or sampling algorithm for solution of inverse problems needs guidance to be efficient. Many algorithms collect and apply information about the problem on the fly, and much improvement has been made in this way. However, as a consequence of the the No-Free-Lunch Theorem, the only way we can ensure a significantly better performance of search and sampling algorithms is to build in as much information about the problem as possible. In the special case of Markov Chain Monte Carlo sampling (MCMC) we review how this is done through the choice of proposal distribution, and we show how this way of adding more information about the problem can be made particularly efficient when based on an approximate physics model of the problem. A highly nonlinear inverse scattering problem with a high-dimensional model space serves as an illustration of the gain of efficiency through this approach. △ Less

Submitted 29 May, 2020; originally announced May 2020.

Moving Target Monte Carlo

Authors: Haoyun Ying, Keheng Mao, Klaus Mosegaard

Abstract: The Markov Chain Monte Carlo (MCMC) methods are popular when considering sampling from a high-dimensional random variable $\mathbf{x}$ with possibly unnormalised probability density $p$ and observed data $\mathbf{d}$. However, MCMC requires evaluating the posterior distribution $p(\mathbf{x}|\mathbf{d})$ of the proposed candidate $\mathbf{x}$ at each iteration when constructing the acceptance rate… ▽ More The Markov Chain Monte Carlo (MCMC) methods are popular when considering sampling from a high-dimensional random variable $\mathbf{x}$ with possibly unnormalised probability density $p$ and observed data $\mathbf{d}$. However, MCMC requires evaluating the posterior distribution $p(\mathbf{x}|\mathbf{d})$ of the proposed candidate $\mathbf{x}$ at each iteration when constructing the acceptance rate. This is costly when such evaluations are intractable. In this paper, we introduce a new non-Markovian sampling algorithm called Moving Target Monte Carlo (MTMC). The acceptance rate at $n$-th iteration is constructed using an iteratively updated approximation of the posterior distribution $a_n(\mathbf{x})$ instead of $p(\mathbf{x}|\mathbf{d})$. The true value of the posterior $p(\mathbf{x}|\mathbf{d})$ is only calculated if the candidate $\mathbf{x}$ is accepted. The approximation $a_n$ utilises these evaluations and converges to $p$ as $n \rightarrow \infty$. A proof of convergence and estimation of convergence rate in different situations are given. △ Less

Submitted 10 March, 2020; originally announced March 2020.

Interior characterization in multiplanetary systems: TRAPPIST-1

Authors: Caroline Dorn, Klaus Mosegaard, Simon L Grimm, Yann Alibert

Abstract: Interior characterization traditionally relies on individual planetary properties, ignoring correlations between different planets of the same system. For multi-planetary systems, planetary data are generally correlated. This is because, the differential masses and radii are better constrained than absolute planetary masses and radii. We explore such correlations and data specific to the multiplan… ▽ More Interior characterization traditionally relies on individual planetary properties, ignoring correlations between different planets of the same system. For multi-planetary systems, planetary data are generally correlated. This is because, the differential masses and radii are better constrained than absolute planetary masses and radii. We explore such correlations and data specific to the multiplanetary-system of TRAPPIST-1 and study their value for our understanding of planet interiors. Furthermore, we demonstrate that the rocky interior of planets in a multi-planetary system can be preferentially probed by studying the most dense planet representing a rocky interior analogue. Our methodology includes a Bayesian inference analysis that uses a Markov chain Monte Carlo scheme. Our interior estimates account for the anticipated variability in the compositions and layer thicknesses of core, mantle, water oceans and ice layers, and a gas envelope. Our results show that (1) interior estimates significantly depend on available abundance proxies and (2) that the importance of inter-dependent planetary data for interior characterization is comparable to changes in data precision by 30 %. For the interiors of TRAPPIST-1 planets, we find that possible water mass fractions generally range from 0-25 %. The lack of a clear trend of water budgets with orbital period or planet mass challenges possible formation scenarios. While our estimates change relatively little with data precision, they critically depend on data accuracy. If planetary masses varied within ~24 %, interiors would be consistent with uniform (~7 %) or an increasing water mass fractions with orbital period (~2-12 %). △ Less

Submitted 6 August, 2018; originally announced August 2018.

Comments: Accepted for publication in ApJ, 20 pages, 14 figures

Mathematical Basis for Physical Inference

Authors: Albert Tarantola, Klaus Mosegaard

Abstract: While the axiomatic introduction of a probability distribution over a space is common, its use for making predictions, using physical theories and prior knowledge, suffers from a lack of formalization. We propose to introduce, in the space of all probability distributions, two operations, the OR and the AND operation, that bring to the space the necessary structure for making inferences on possi… ▽ More While the axiomatic introduction of a probability distribution over a space is common, its use for making predictions, using physical theories and prior knowledge, suffers from a lack of formalization. We propose to introduce, in the space of all probability distributions, two operations, the OR and the AND operation, that bring to the space the necessary structure for making inferences on possible values of physical parameters. While physical theories are often asumed to be analytical, we argue that consistent inference needs to replace analytical theories by probability distributions over the parameter space, and we propose a systematic way of obtaining such "theoretical correlations", using the OR operation on the results of physical experiments. Predicting the outcome of an experiment or solving "inverse problems" are then examples of the use of the AND operation. This leads to a simple and complete mathematical basis for general physical inference. △ Less

Submitted 19 September, 2000; originally announced September 2000.

Comments: 24 pages, 4 figures