PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Fidelity Particle Flow Filters based on Log Homotopy

Abhinab Bhattacharjee, Andrey A. Popov, Amit N. Subrahmanya, Adrian Sandu

arXiv:2610.35712v1cs.CEcs.NAmath.NA

TL;DR

High-dimensional particle filters can suffer weight degeneracy, while sampling-based data assimilation can require costly model runs. This work combines particle flows with a multi-fidelity control-variate filter, and reports comparable or lower computational cost alongside reduced estimator variance. Its conclusions are demonstrated in the stated bi-fidelity setting, with future extensions identified.

  • Problem

    Particle-filter weights can collapse in high dimensions, and sampling-based data assimilation requires many potentially costly model runs.

    Source: §1 · p. 3

  • Method

    The MFPFF augments a small FOM ensemble with a larger ROM ensemble as a control variate, using Exact and Gromov log-homotopy flows.

    Source: §1 · p. 4

  • Results

    The MFPFF reduces estimator variance, and experiments report comparable or lower computational cost using fewer expensive high-fidelity runs.

    Source: §1 · p. 4, §5 · p. 23

  • Takeaways & Limitations

    The reported advantage also holds in the QG test problem with localization and nonlinear observations, within the studied bi-fidelity setting.

    Source: §5 · p. 23

    The study uses intrusive Galerkin ROMs; replacing them with data-driven ROMs is identified as future work for applications without access to governing PDEs.

    Source: §5 · p. 23

Abstract

from arXiv · show

This work proposes a new Multi-Fidelity Particle Flow Filter (MFPFF) based on a linear control variate framework and multilevel Monte Carlo methods. Conventional particle filters often suffer from weight collapse and particle degeneracy in high-dimensional spaces. Particle flow filters address this by smoothly transporting prior particles toward the posterior distribution, thereby avoiding the need for resampling. This work constructs multi-fidelity particle flow filters by augmenting a small number of expensive high-fidelity particles (full-order model runs) with a large number of inexpensive low-resolution particles (reduced-order surrogates constructed via intrusive Galerkin projection). These surrogates act as control variates to enhance the statistical estimate of the full-order posterior, yielding enhanced analysis accuracy at comparable computational cost. The work focuses on two particle flows based on log homotopy, the stochastic Gromov flow and deterministic Exact flow. Numerical experiments with the Korteweg-De Vries (KdV) and Quasi-Geostrophic (QG) test problems demonstrate that the proposed MFPFF achieves accuracy comparable to high-fidelity simulations while significantly reducing computational expense.

1 Introduction

Particle flows address particle-filter weight degeneracy by moving particles from prior to posterior, while multi-fidelity methods combine inexpensive models with costly accurate runs. This work introduces a bi-fidelity particle-flow filter using ROMs as control variates and investigates Exact and Gromov flows, localization, and nonlinear observations.

1 Introduction

  • Particle flows keep weights constant while moving particles from prior toward posterior, but may still sample a distribution different from the target posterior.Conventional particle-filter weights can concentrate on very few particles when observations are informative or system dimension is high.Source: §1 · p. 3
  • Low-fidelity models capture bulk statistical information, while costly high-fidelity runs correct bias and reduce estimator variance [30].The study uses a large inexpensive ensemble and a small expensive ensemble, and restricts its experiments to two fidelity levels.Source: §1 · pp. 3–4
  • POD ROMs project FOM dynamics onto a low-dimensional subspace retaining dominant behavior, providing an inexpensive surrogate.POD extracts energetically significant modes from FOM simulation snapshots [46].Source: §1 · p. 3
  • Earlier multilevel particle filters use optimal-transport coupling but assume equal fine- and coarse-ensemble sizes, unsuitable when FOM runs cost substantially more [31] [46].A multilevel Kalman-filter approach can yield nonpositive covariance and stability issues in Kalman-gain calculation [32] [46].Source: §1 · p. 4
  • The MFPFF uses a linear control-variate framework, augmenting a limited ensemble of expensive FOM simulations with a significantly larger, cheaper, lower-accuracy ROM ensemble, and achieves significant estimator-variance reduction versus standard single-fidelity filters.Source: §1 · p. 4
  • The study investigates deterministic Exact and stochastic Gromov flows, and explores localization and nonlinear observations in the MFPFF setting.The paper presents the method as an extension of prior work [46].Source: §1 · p. 4

2 Background

The background establishes Bayesian data assimilation and particle-flow filtering, then introduces the Exact and Gromov flows, linear control variates, and POD reduced-order models used by the paper.

2.1 Data Assimilation

  • Data assimilation combines prior information and observation likelihoods to obtain a posterior analysis; this study assumes deterministic model dynamics.Observation errors are assumed independent and identically distributed Gaussian, although Gaussianity is not required in general.Source: §2.1 · p. 5

2.2 Particle Flow Filters

  • Particle flows use drift and diffusion to transport forecast particles toward the posterior, and log-homotopy flows gradually incorporate the observation likelihood over synthetic time.The cited flows transform prior particles into equally weighted posterior samples over a finite synthetic-time interval.Source: §2.2 · pp. 6–7

2.2.1 Exact Flow

  • Exact flow uses a zero-diffusion, linear-drift particle equation; nonlinear observations are handled by subtracting the observation-model linearization.Particles are advanced with Euler steps; more sophisticated time-stepping methods are deferred to future work.Source: §2.2.1 · p. 7

2.2.2 Gromov Flow

  • Gromov flow uses nonzero diffusion, and a Sherman–Morrison–Woodbury reformulation avoids direct inversion of the prior covariance.The diffusion square root is approximated, and the flow expression is further manipulated to avoid numerical issues at λ=0.Source: §2.2.2 · p. 8

2.3 Linear Control Variates

  • A linear control variate forms an unbiased total variate from a principal variate and correlated control variate; its optimized gain reduces variance when the needed covariance is full rank.Empirical covariance undersampling can make the gain inversion ill-conditioned, motivating a gain robust to sampling errors.Source: §2.3 · pp. 8–9

2.4 Reduced Order Modeling

  • POD extracts dominant modes from snapshots and uses Galerkin projection to construct a low-dimensional, computationally inexpensive surrogate for the original dynamics.DEIM is noted as an option for nonlinear projection terms that avoids reconstructing the full state when dimensions are very high.Source: §2.4 · pp. 9–10

3 Methodology

The MFPFF combines FOM and ROM ensembles through a projected, three-variate framework, then applies particle flow updates to the analysis ensembles. The methodology also extends the bi-fidelity construction to multiple fidelity levels.

3 Methodology

  • The framework combines FOM principal and ROM ancillary variates with a highly correlated ROM control variate, using an optimal projection between full and reduced spaces.The projection operator is problem-dependent and remains constant across assimilation steps when the ROM is unchanged.Source: §3 · pp. 10–11

3.1 Multifidelity Forecast Step

  • At each forecast step, FOM particles are projected into ROM space to strengthen principal–control correlation before the FOM and ROM dynamics propagate their ensembles.The control and ancillary variates are assumed independent and the control variate is discarded after assimilation; the principal and ancillary variates are treated as independent throughout sequential assimilation.Source: §3.1 · pp. 11–12

3.2 Multifidelity Analysis Step

  • The analysis flow transforms FOM- and ROM-space samples using cross-covariance information, then applies bias correction and rebuilds a correlated control variate for the next forecast.The control variate is discarded during the analysis step.Source: §3.2 · pp. 12–13

3.3 Statistical Analysis

  • Under the stated assumptions, the principal-variate update is equivalent to a perturbed EnKF update, and the control-variate framework helps reduce uncertainty and sampling errors.The statistical analysis considers linear observations and a one-pseudo-step Gromov flow.Source: §3.3 · pp. 13, 15

3.4 Extension to Multiple Fidelity Levels

  • The multilevel extension makes each level’s ancillary variate the next level’s principal variate, with model cost and accuracy decreasing at lower levels.The construction assumes access to an optimal projection operator at each level.Source: §3.4 · pp. 15–16

4 Numerical Experiments

Experiments test the filters on KdV and QG, with additional QG localization and nonlinear-observation cases. Across these settings, the reported comparisons show accuracy gains or comparable performance at lower model-propagation cost, with an instability caveat for Exact flow under higher inflation and nonlinear observations.

4 Numerical Experiments

  • The experiments average spatio-temporal RMSE over ten initializations for KdV and QG twin experiments.Failed heatmap cases indicate that at least one initialization failed for that ensemble and inflation combination.Source: §4 · p. 16

4.1 Korteweg-De Vries (KdV)

  • In KdV, bi-fidelity Gromov and Exact flow show an advantage over single-fidelity FOM runs with just two FOM ensembles; a single FOM propagation costs about 20 ROM propagations.The comparison also notes that higher inflation can introduce undissipated spurious modes and higher errors.Source: §4.1 · pp. 17–18, Figure 3 · p. 18, Figure 4 · p. 18

4.2 Quasi-Geostrophic Equation (QGE)

  • For QG without localization, bi-fidelity filters achieve comparable accuracy at substantially lower cost, including after accounting for augmented ROMs.A single FOM propagation is about 40 times more expensive than one ROM propagation.Source: §4.2 · p. 19, Figure 5 · p. 20, Figure 6 · p. 20

4.2.1 Localization

  • With covariance localization at radius 15 grid points, the filter remains compatible with localization and bi-fidelity runs remain cheaper with as few as two ensemble members.The covariance tapering targets spurious correlations and increases covariance-matrix rank.Source: §4.2.1 · pp. 19, 21, Figure 7 · p. 21, Figure 8 · p. 21

4.2.2 Nonlinear Observations Operators

  • For sparse nonlinear velocity-magnitude observations, Gromov flow performs well, while Exact flow is unstable at higher inflation under the assumptions of equation (3.11).Source: Figure 9 · p. 22, §4.2.2 · p. 22

4.2.3 Cost vs RMSE

  • Exact flow shows a slight RMSE advantage and Gromov flow a significant advantage when one FOM is replaced by computationally equivalent ROMs, comparing model-propagation costs only.The experiments ignore assimilation costs.Source: §4.2.3 · p. 22

5 Conclusions

The paper proposes a multi-fidelity data-assimilation algorithm combining log-homotopy particle flows with control variates. Its experiments report computational advantages, including with localization and nonlinear observations in the QG test problem.

5 Conclusions

  • The method combines a small number of accurate, expensive high-fidelity runs with many cheaper, potentially biased low-fidelity runs.The authors report comparable or lower computational cost and similar advantages with localization and nonlinear observations in the QG test problem.Source: §5 · pp. 22–23
  • The framework uses an optimal projection or interpolation operator to map variates between spaces, and can extend to multiple fidelity levels.Models at different levels may have distinct dimensions and inner products.Source: §5 · p. 23
  • Future directions include data-driven ROMs, multi-level importance-sampling filters, nonlinear-observation Exact flow, and more stable time integration schemes.The authors also propose applying the work to state and parameter estimation in large realistic digital-twin models.Source: §5 · p. 23

A ROM for KdV

The KdV ROM is built from snapshots of a two-soliton simulation and a POD–Galerkin projection. The passage describes the reduced evolution equations and their finite-difference representation.

A ROM for KdV

  • The two-soliton dynamics challenge POD ROMs because accurate capture requires many modes, and snapshot singular values decay slowly.The forward simulation uses 1000 time steps of size 0.01; the singular values are shown in Figure A.1.Source: A ROM for KdV · p. 24, Figure A.1 · p. 24
  • The ROM substitutes a POD expansion into the governing equation and projects onto basis functions using inner products and orthogonality.The passage introduces time-dependent coefficients and basis functions before deriving the reduced equation.Source: A ROM for KdV · p. 24
  • The spatially discretized model uses finite-difference derivative operators, with a linear operator and quadratic tensor defined through the basis and derivative matrices.The passage identifies the Hadamard product in the discretization and the optimal operator used for this test problem.Source: A ROM for KdV · pp. 24–25

B ROM for QG

The QG ROM constructs a vorticity basis from simulation snapshots using a correlation matrix and POD eigenpairs. It truncates that basis and derives reduced evolution equations for the approximations.

B ROM for QG

  • The vorticity basis is built from 2000 snapshots spanning nearly a month of simulation, using a quadrature rule to evaluate the correlation matrix.The eigenvalues of the correlation matrix are shown in Figure B.1.Source: B ROM for QG · p. 25, Figure B.1 · p. 25
  • Solving the correlation-matrix eigenproblem yields ordered eigenvalues and corresponding eigenvectors, from which the first r informative modes are retained for the vorticity basis.The ROM streamfunction basis is derived from the vorticity basis using the stated Laplacian relationship.Source: B ROM for QG · p. 25
  • The reduced QG evolution is derived by using shared time coefficients for the vorticity and streamfunction approximations, then projecting the governing equation onto the vorticity basis.The final condensed form defines coefficients using the Rossby and Reynolds numbers and inner products.Source: B ROM for QG · p. 26
Loading 2610.35712v1…