Source-linked AI summary
A Dynamic Likelihood Approach to Filtering for Advection-Diffusion Dynamics
Johannes Krotz, Juan M. Restrepo, Jorge Ramirez
TL;DR
The paper addresses posterior estimation for noisy advection-diffusion dynamics, extending a Dynamic Likelihood approach developed for hyperbolic problems. It propagates observations and their uncertainties along characteristics while treating diffusion through the model and observations. Simulations show better DLF estimates than the standard Kalman Filter when advection dominates, observations are sparse, and precision is high.
Problem
The paper studies posterior covariance estimation for noisy advection-diffusion states, extending a method previously developed for hyperbolic dynamics to this setting.
Method
The DLF uses observations and their statistics to evolve likelihoods along characteristics, with a split treatment in which diffusion is evaluated alongside model dynamics.
Results
The DLF outperformed the standard KF on several accuracy metrics, especially when advection dominates diffusion and observations are sparse and high-precision.
Takeaways & Limitations
DLF estimates more accurately capture phase and amplitude, center of mass, local RMS values, and overall mass in the reported simulations.
Takeaways & Limitations
DLF requires special time integrators and can have higher computational complexity than the KF, so pseudo-observation retention must be managed.
Abstract
from arXiv · showhide
A Bayesian data assimilation scheme is formulated for advection-dominated advective and diffusive evolutionary problems, based upon the Dynamic Likelihood (DLF) approach to filtering. The DLF was developed specifically for hyperbolic problems -waves-, and in this paper, it is extended via a split step formulation, to handle advection-diffusion problems. In the dynamic likelihood approach, observations and their statistics are used to propagate probabilities along characteristics, evolving the likelihood in time. The estimate posterior thus inherits phase information. For advection-diffusion the advective part of the time evolution is handled on the basis of observations alone, while the diffusive part is informed through the model as well as observations. We expect, and indeed show here, that in advection-dominated problems, the DLF approach produces better estimates than other assimilation approaches, particularly when the observations are sparse and have low uncertainty. The added computational expense of the method is cubic in the total number of observations over time, which is on the same order of magnitude as a standard Kalman filter and can be mitigated by bounding the number of forward propagated observations, discarding the least informative data.
1. Introduction
The paper extends the Dynamic Likelihood approach, originally developed for hyperbolic wave problems, to forced advection-diffusion data assimilation. It uses observations to evolve likelihoods along characteristics and compares the resulting estimator with the Kalman Filter.
- Dynamic Likelihood framework: The Dynamic Likelihood approach modifies posterior distributions by exploiting finite-time information propagation in wave dynamics.It is described as an approach applicable in principle to linear or nonlinear assimilation methodologies.
- Advection-diffusion extension: The paper develops DLF data assimilation for forced advection-diffusion equations, expanding its application beyond hyperbolic problems.The focus is on estimating quantities of interest when uncertainty arises from advection and forcing.
- Advection-diffusion extension: DLF evolves likelihoods along characteristics by generating pseudo-observations between actual observation times.Each pseudo-observation is derived from a real observation at an earlier time.
- Evaluation design: The advection-diffusion DLF formulation uses techniques and ideas similar to a Kalman Filter.The paper compares its outcomes with Kalman Filter estimates for this problem.
2. Statement of the Problem
The study estimates posterior uncertainty for a noisy scalar advection-diffusion state from localized, noisy observations. It focuses on linear dynamics to compare DLF with optimal filtering schemes while describing a sequential model-and-observation update.
- Problem formulation: The central objective is minimizing the trace of the posterior covariance for a noisy scalar state variable given noisy observations.The state evolves according to a noisy advection-diffusion initial value problem.
- Motivation: The DLF is tested because earlier hyperbolic-problem results reported better estimates when observations were sparse and had low uncertainty.The paper examines whether this advantage persists for advection-dominated advection-diffusion dynamics.
- Sequential estimation: The model provides the prior, while observations determine the likelihood at observation times.At times without observations, the posterior is propagated through the model alone.
- Observations: Observations are temporally and spatially localized measurements available only at selected observation times.Their measurement errors are modeled as mean-zero normal vectors with known covariance, and the number of locations is time-independent.
- Sequential estimation: The sequential filtering procedure forecasts the state with the model and performs an analysis update when observations are available.The analysis uses prior moments together with observation moments and error covariance.
4. Analysis of Computational Complexity of the DLF
DLF adds computational cost through pseudo-observation propagation and gain calculations. Its matrix inversions scale cubically with the number of concurrently used observations, motivating bounded retention of older pseudo-observations.
- Sources of overhead: The DLF overhead relative to standard filtering comes from calculating pseudo-observations and calculating and applying multiple gains.These are the two additional computational sources identified by the paper.
- Pseudo-observations: Pseudo-observation calculations are linear in the number of time steps for an explicit ODE solver and involve covariance calculations over observation locations.The text gives an intermediate bound of O(MNI^2) for these calculations.
- Gain calculations: O(NM^3I^3) bounds DLF gain-computation cost when all pseudo-observations are retained, compared with O(MI^3) for the Kalman Filter.The cubic dependence comes from inverting matrices whose size grows with the number of retained observations.
- Complexity control: As pseudo-observations age, their uncertainty increases, supporting selective discarding to limit computational complexity.The paper proposes capping their number or discarding relatively more uncertain observations and pseudo-observations.
- Complexity control: Capping pseudo-observations at pI retains only the last p observation times and reduces concurrent data from MI to pI.At observation times, the oldest retained pseudo-observations are discarded after the multi-analysis step.
5. Numerical Results and Comparisons
Across advection-dominated experiments, the DLF generally improves estimation over the KF, especially with sparse observations and uncertain initial conditions or model parameters. Its advantage weakens as diffusion or observation density increases, with some metric-specific exceptions.
- Experimental setting: α = D/(c_0L) measures the relative influence of diffusion versus advection, framing the regimes compared in the experiments.The study focuses on whether a wave-oriented approach remains effective when advection dominates and diffusion is present.
- Known initial conditions: The DLF preserves observation-driven adjustments over time, whereas KF adjustments quickly vanish under diffusion.The DLF reinforces observations through pseudo-observations propagated along characteristics.
- Observation-density sweep: With sparse observations, the DLF outperforms the KF on Mass error and generally on RMS, CoM, and Calibration; the KF gains an advantage when observations are plentiful.The sparse-observation cases are I = 10 and I = 20 in the parameter sweep.
- Uncertain initial conditions: Under uncertain initial amplitude, the DLF corrects the error rapidly and retains lower RMS and Mass errors with better Calibration than the KF.The model starts with σ = 0.7 while the truth has σ = 1; the DLF reaches the known-initial-data error levels after a brief adjustment period.
- Uncertain initial conditions: When initial phase is wrong, the DLF suppresses the incorrect mode within a few steps, while the KF takes nearly the full simulation to remove the phase error.The DLF can temporarily represent two modes, causing an initial Mass-error disadvantage before the incorrect mode is removed.
- Parameter sensitivity: Across broader parameter sweeps, the DLF remains strongest for RMS, CoM, and Calibration, but its Mass-error advantage can reverse briefly with few observations because of early two-mode estimates.The DLF’s advantage generally decreases as diffusion increases, although it remains useful when observations are sparse.
6. Discussion and Conclusions
The paper extends the Dynamic Likelihood approach from hyperbolic dynamics to advection-diffusion problems and finds improved estimation accuracy, especially when advection dominates and observations are sparse and precise.
- Contribution: The DLF extends dynamic-likelihood data assimilation from hyperbolic wave dynamics to forced advection-diffusion problems.The extension addresses diffusion along characteristic paths using model derivatives as an estimator, while observations evolve through time along characteristics.
- Method: The implementation combines observations and model information as observations are evolved through time along characteristics, including extrapolation to off-grid points.This formulation uses data and the model to drive observation evolution when grid points and characteristic paths do not coincide.
- Results: Numerical simulations show that the DLF outperforms the standard KF on several accuracy metrics.The comparison evaluates posterior estimates using RMS, Mass, CoM, and Calibration measures.
- Scope of findings: The DLF is superior when advection dominates diffusion and observations are sparse and highly precise, with the reported advantages persisting even when diffusion and advection are comparable.The paper also reports better variance estimation, particularly with uncertain initial data or an ill-informed model.
- Results: The DLF produces lower RMS error than the KF in all experiments and more accurately predicts phase, amplitude, center of mass, and overall mass under initial-data uncertainty.Its estimates are less sensitive to uncertain initial data and reach the correct phase and amplitude more quickly.
- Practical implication: The approach retains the ability to make Bayesian forecasts beyond the latest observation time by propagating model- and pseudo-observation-informed likelihoods into the future.This can produce future posterior means that differ from forecasts based only on the model-informed prior.