betaFeedback

WorksMathematics

Enhancing Exploration in Ensemble Kalman Inversion via Covariance Inflation and Projected Dropout

Abstract

Ensemble Kalman inversion (EKI) is a derivative-free and parallelizable method for parameter estimation in Bayesian inverse problems involving expensive forward models. It adapts Kalman filtering by repeatedly assimilating the observation data through an artificial dynamical system with identity state evolution and the parameter-to-observation map as its observation operator. However, its performance can be limited by ensemble collapse and confinement to the affine span of the initial ensemble. We propose an exploration-enhanced EKI framework that combines multiplicative covariance inflation with a projected-dropout correction of the ensemble mean. Inflation helps maintain ensemble spread, while projected dropout enables the mean to explore directions beyond the initial affine span. For linear inverse problems, we provide a discrete-time, finite-ensemble analysis of the resulting method and prove the least-squares objective converges exponentially to its minimum in expectation. Within the initial ensemble subspace, the contraction factor is independent of the conditioning of the inverse problem, and the limiting ensemble covariance provides a low-rank representation of the posterior covariance. Experiments on ill-conditioned linear systems, 100-dimensional nonlinear benchmarks, and a PDE-constrained barotropic flow inverse problem demonstrate the effectiveness of the proposed method.

Submitted to: Inverse Problems. The authors are in alphabetical order.

Introduction

Many problems arising in scientific computing, data assimilation, and uncertainty quantification can be formulated as inverse problems, in which one seeks to recover an unknown parameter from indirect and noisy observations. A common formulation is the weighted nonlinear least-squares problem

min⁡θ∈RNθ12(x−F(θ))TΣν−1(x−F(θ))=min⁡θ∈RNθ12∥Σν−1/2(x−F(θ))∥2(1)\min_{\theta\in\mathbb{R}^{N_\theta}} \frac{1}{2}(x-\mathcal{F}(\theta))^T\Sigma_\nu^{-1}(x-\mathcal{F}(\theta)) = \min_{\theta\in\mathbb{R}^{N_\theta}} \frac{1}{2}\left\lVert\Sigma_\nu^{-1/2}(x-\mathcal{F}(\theta))\right\rVert^2 \tag*{(1)}

where θ∈RNθ\theta\in\mathbb{R}^{N_\theta} is the unknown parameter, x∈RNxx \in\mathbb{R}^{N_x} is the observation, F:RNθ→RNx\mathcal{F} : \mathbb{R}^{N_\theta} \to\mathbb{R}^{N_x} is the forward map, and Σν∈RNx×Nx\Sigma_\nu\in\mathbb{R}^{N_x \times N_x} is symmetric positive definite. This framework covers both unregularized data-fitting problems and Tikhonov-regularized inverse problems.

In many large-scale applications, the parameter dimension is high, the forward model is computationally expensive, and derivative information with respect to θ\theta is either unavailable or unreliable. Ensemble Kalman inversion (EKI) [1, 2, 3] provides a derivative-free and ensemble-based approach for approximately solving (1). The method introduces the artificial filtering model

evolution:θn+1=θn,observation:xn+1=F(θn+1)+νn+1,(2)\begin{aligned} \text{evolution:}\quad&\theta_{n+1} = \theta_n,\\ \text{observation:}\quad&x_{n+1} = \mathcal{F}(\theta_{n+1}) + \nu_{n+1}, \tag*{(2)} \end{aligned}

where νn+1∼N(0,Σν)\nu_{n+1} \sim\mathcal{N}(0,\Sigma_\nu), and applies ensemble Kalman filtering ideas [4, 5] to the repeated observation xn+1=xx_{n+1}=x. In this way, EKI inherits the computational simplicity, scalability, and parallel structure of ensemble Kalman methods while avoiding adjoint or gradient computations.

Despite these advantages, standard EKI has two structural limitations. First, the static artificial evolution in (2) does not replenish ensemble spread, while the Kalman analysis step continuously reduces covariance. This leads to ensemble collapse and algebraically slow convergence: in the linear setting, continuous-time EKI has L2L^2 mean error of order O(t−1/2)O(t^{-1/2}) [6, 7, 8], and discrete-time EKI has corresponding order O(n−1/2)O(n^{-1/2}) [9]. Second, all ensemble particles remain confined to the affine span of the initial ensemble [3]. This subspace property can be beneficial as an implicit regularizer, but it also prevents recovery of directions not represented by the initial ensemble, which is restrictive when the ensemble size J≪NθJ \ll N_\theta.

Contributions

Our main contributions are as follows:

(i) We propose an inflated projected-dropout EKI method based on two modifications of standard EKI that preserve its parallel structure. Covariance inflation sustains ensemble spread, whereas projected dropout enables the ensemble mean to move beyond the initial affine span.

(ii) We provide a discrete-time, finite-ensemble analysis for linear inverse problems. The error in initial ensemble subspace decays exponentially at a rate independent of the condition number of the restricted Hessian, while the expected error in the complementary subspace also contracts exponentially. Meanwhile, the limiting covariance also provides a low-rank representation of posterior uncertainty.

(iii) Experiments on ill-conditioned linear systems, nonlinear benchmarks with curved or multimodal objective landscapes, and a PDE-constrained inverse problem for barotropic flow demonstrate accelerated convergence, sustained ensemble spread, and improved out-of-subspace recovery.

Literature Review

Ensemble Kalman inversion

From an optimization perspective, ensemble Kalman inversion [1, 2, 3] seeks to minimize a least-squares objective through repeated ensemble Kalman updates. Its convergence is studied through the long-time behavior, as t→∞t \to\infty, of the associated continuous-time EKI dynamics [6, 10]. These methods are particularly attractive for expensive black-box forward models whose derivatives are difficult to obtain or unreliable [11, 12]. Applications include geophysics and reservoir history matching [13, 14, 15], turbulent flows [16, 17, 18], fluid-structure interaction [19], and climate models [20]. In fluid applications, ensemble averaging can also provide a smoothing effect that mitigates sensitivity to small-scale oscillations [21 Lemma 2][22]. Adaptive surrogate models have also been combined with Kalman inversion to reduce the cost of repeated forward evaluations while controlling surrogate error [23, 24]. The present work focuses on the deterministic square-root implementations [25], including ensemble adjustment Kalman inversion (EAKI) and ensemble transform Kalman inversion (ETKI) [26, 27], based on the ensemble adjustment and ensemble transform Kalman filters [28, 29], respectively. Both reproduce the Kalman analysis covariance exactly, which facilitates the discrete-time, finite-ensemble analysis developed here.

From a sampling perspective, Kalman inversion methods have also been developed for Bayesian inverse problems. In that setting, the artificial time is often normalized to [0,1][0, 1], so that each update incorporates only a fraction of the likelihood. This gives rise to a homotopy or transport interpretation, in which the prior is deformed into the posterior over one unit of artificial time [30, 31, 32]. More generally, these methods are related to measure transport [33, 34, 35, 36, 37], sequential Monte Carlo methods [38, 39], and non-Gaussian extensions [40]. The present work instead considers a covariance-inflated artificial dynamical system whose long-time stationary behavior yields a Gaussian approximation of the posterior. This viewpoint is closely related to Fisher–Rao gradient flows under Gaussian [41, 42][26 Remark 1] or Gaussian mixture [43, 44, 45, 46] approximations.

Ensemble collapse and subspace restriction

Standard EKI suffers from ensemble collapse and algebraically slow convergence [6, 10, 47]. Existing approaches to acceleration include optimization-inspired updates [48, 49], covariance inflation, adaptive step sizes, and stabilized continuous-time dynamics [27, 50, 51, 26]. The covariance-inflated dynamics studied here follow this direction.

A separate limitation is that standard EKI remains confined to the affine span of the initial ensemble [3]. Although this restriction provides implicit regularization [52], it prevents recovery of components outside that span. Strategies for overcoming it include localization [53, 54] and dropout perturbations [47], the latter admitting exponential convergence guarantees in expectation under suitable assumptions. Our projected dropout correction combines exploration beyond the initial affine span with the covariance update of inflated EAKI. It modifies only the ensemble mean, using observation responses orthogonal to those represented by the original ensemble, while preserving the analysis covariance for the current predicted ensemble.

Organization

The remainder of the paper is organized as follows. In § 2, we introduce the inverse problem, its Bayesian formulation, and the weighted nonlinear least-squares objective. In § 3, we introduce the methodology. In § 4, we present a discrete-time, finite-ensemble convergence analysis in the linear setting. Numerical experiments are presented in § 5, followed by concluding remarks in § 6.

Inverse Problem and Weighted Nonlinear Least Squares

We consider the inverse problem of recovering an unknown parameter θ∈RNθ\theta\in\mathbb{R}^{N_\theta} from noisy observations y∈RNyy \in\mathbb{R}^{N_y} related by [55]

y=G(θ)+η.y = G(\theta) + \eta.

Here GG denotes the forward model mapping parameters to observables, and η\eta denotes observational noise. We assume that η\eta is a realization of a mean-zero Gaussian random variable with covariance Ση\Sigma_\eta, namely, η∼N(0,Ση)\eta\sim\mathcal{N}(0,\Sigma_\eta). The corresponding data-misfit function is

Φ(θ,y)=12(y−G(θ))TΣη−1(y−G(θ))=12∥Ση−1/2(y−G(θ))∥2.(3)\Phi(\theta, y) = \frac{1}{2}(y - G(\theta))^{T}\Sigma_\eta^{-1}(y - G(\theta)) = \frac{1}{2}\left\|\Sigma_\eta^{-1/2}(y - G(\theta))\right\|^{2}. \tag*{(3)}

This is a weighted nonlinear least-squares objective, and minimizing Φ\Phi yields an estimate of the unknown parameter θ\theta.

Inverse problems are often ill posed and therefore require regularization. In the Bayesian framework, regularization is introduced by incorporating prior information on the unknown parameter. Specifically, θ\theta and yy are treated as random variables. Given a prior density ρprior(θ)\rho_{\mathrm{prior}}(\theta), the inverse problem is formulated in terms of the posterior density ρpost(θ)\rho_{\mathrm{post}}(\theta) conditioned on the observation yy [56, 57, 58]:

ρpost(θ)=1Z(y)e−Φ(θ,y)ρprior(θ),\rho_{\mathrm{post}}(\theta) = \frac{1}{Z(y)}e^{-\Phi(\theta,y)}\rho_{\mathrm{prior}}(\theta),

where Z(y)=∫e−Φ(θ,y)ρprior(θ)dθZ(y) = \int e^{-\Phi(\theta,y)}\rho_{\mathrm{prior}}(\theta)\mathrm{d}\theta is the normalization constant. We focus on the case in which the prior ρprior\rho_{\mathrm{prior}} is Gaussian, or is approximated by a Gaussian distribution with mean r0r_0 and covariance Σ0\Sigma_0. In this case, the posterior density can be written as

ρpost(θ)∝e−ΦR(θ,y),ΦR(θ,y)=Φ(θ,y)+12∥Σ0−1/2(θ−r0)∥2.(4)\rho_{\mathrm{post}}(\theta) \propto e^{-\Phi_R(\theta,y)}, \qquad\Phi_R(\theta,y) = \Phi(\theta,y) + \frac{1}{2}\left\|\Sigma_0^{-1/2}(\theta-r_0)\right\|^2. \tag*{(4)}

The function ΦR\Phi_R is the Tikhonov-regularized version of the data misfit Φ\Phi [59], and minimizing ΦR\Phi_R yields the maximum a posteriori estimate. This observation shows that both the unregularized inverse problem (3) and the regularized Bayesian inverse problem (4) can be treated within a unified weighted nonlinear least-squares framework. Accordingly, in the remainder of the paper we work with the general formulation

min⁡θ∈RNθ12(x−F(θ))TΣν−1(x−F(θ))=min⁡θ∈RNθ12∥Σν−1/2(x−F(θ))∥2,(5)\min_{\theta\in\mathbb{R}^{N_\theta}} \frac{1}{2}(x-\mathcal{F}(\theta))^{\mathsf{T}}\Sigma_\nu^{-1}(x-\mathcal{F}(\theta)) = \min_{\theta\in\mathbb{R}^{N_\theta}} \frac{1}{2}\left\|\Sigma_\nu^{-1/2}(x-\mathcal{F}(\theta))\right\|^2, \tag*{(5)}

where (x,F,Σν)(x,\mathcal{F},\Sigma_\nu) denotes either the original inverse problem or its augmented regularized formulation.

Inflated Ensemble Kalman Inversion with Projected Dropout

In this section, we present our proposed ensemble Kalman inversion method, which combines covariance inflation with a projected dropout mean correction. We first introduce covariance-inflated mean-field dynamics and derive the corresponding ensemble adjustment Kalman inversion (EAKI) scheme. We then develop the projected dropout correction, which enables the ensemble mean to explore beyond the affine span of the initial ensemble without changing the covariance produced by the EAKI analysis step.

Inflated Mean-Field Dynamics and EAKI

Ensemble Kalman inversion addresses the inverse problem (5) by introducing an artificial dynamical system and repeatedly assimilating the same observed data using ensemble Kalman filtering. To mitigate ensemble collapse and improve exploration within the ensemble subspace, we replace the identity evolution in (2) with the following covariance-inflated mean-field dynamics:

evolution:θn+1=mn+11−Δτ(θn−mn),observation:xn+1=F(θn+1)+νn+1.(6)\begin{aligned} \text{evolution:}\quad&\theta_{n+1}=m_n+\sqrt{\frac{1}{1-\Delta\tau}}(\theta_n-m_n), && \\ \text{observation:}\quad&x_{n+1}=\mathcal{F}(\theta_{n+1})+\nu_{n+1}. && \tag*{(6)} \end{aligned}

Here Δτ∈(0,1)\Delta\tau\in(0,1) denotes the artificial time step. Let Yn:=(x1†,x2†,…,xn†)Y_n:=(x_1^\dagger,x_2^\dagger,\ldots,x_n^\dagger) denote the observation history, with xk†=xx_k^\dagger=x for every kk. The conditional law of θn\theta_n given YnY_n is the filtering distribution at iteration nn, and mn=E[θn∣Yn]m_n=\mathbb{E}[\theta_n\mid Y_n] denotes its mean. Equation (6a) is therefore of mean-field type, since it depends on the filtering distribution through its mean. The artificial observation noises are independent Gaussian random variables with

νn+1∼N(0,Σν,n+1),Σν,n+1=1ΔτΣν.(7)\nu_{n+1}\sim\mathcal{N}(0,\Sigma_{\nu,n+1}),\qquad\Sigma_{\nu,n+1}=\frac{1}{\Delta\tau}\Sigma_\nu. \tag*{(7)}

The scaling with respect to Δτ\Delta\tau is crucial for ensuring that the resulting filtering distribution provides a consistent Gaussian approximation of the posterior uncertainty; see Theorem 1.

We approximate the filtering distribution of θn\theta_n given YnY_n by an ensemble of JJ particles {θnj}j=1J\{\theta_n^j\}_{j=1}^J. At the ensemble level, we use mnm_n to denote the empirical mean J−1∑j=1JθnjJ^{-1}\sum_{j=1}^J\theta_n^j, which approximates the conditional mean E[θn∣Yn]\mathbb{E}[\theta_n\mid Y_n]. Applying the inflated evolution (6a) to each particle yields a predicted ensemble that approximates the conditional distribution of θn+1\theta_{n+1} given YnY_n:

θ^n+1j=mn+11−Δτ(θnj−mn),j=1,…,J,(8)\widehat{\theta}_{n+1}^j=m_n+\sqrt{\frac{1}{1-\Delta\tau}}(\theta_n^j-m_n),\qquad j=1,\ldots,J, \tag*{(8)}

with empirical mean m^n+1\widehat{m}_{n+1} and covariance C^n+1\widehat{C}_{n+1}.

In the analysis step, we consider the predicted observation xn+1=F(θn+1)+νn+1x_{n+1}=\mathcal{F}(\theta_{n+1})+\nu_{n+1} defined by (6b). We approximate the conditional joint distribution of (θn+1,xn+1)(\theta_{n+1},x_{n+1}) given YnY_n by the Gaussian distribution

N([m^n+1x^n+1],[C^n+1C^n+1θx(C^n+1θx)TC^n+1xx]),(9)\mathcal{N}\left( \begin{bmatrix} \widehat{m}_{n+1}\\ \widehat{x}_{n+1} \end{bmatrix}, \begin{bmatrix} \widehat{C}_{n+1} & \widehat{C}_{n+1}^{\theta x}\\ (\widehat{C}_{n+1}^{\theta x})^{T} & \widehat{C}_{n+1}^{xx} \end{bmatrix} \right), \tag*{(9)}

where the predictive moments satisfy

x^n+1=E[xn+1∣Yn]=E[F(θn+1)∣Yn],\widehat{x}_{n+1}=\mathbb{E}[x_{n+1}\mid Y_n]=\mathbb{E}[\mathcal{F}(\theta_{n+1})\mid Y_n],
C^n+1θx=Cov⁡[θn+1,xn+1∣Yn]=Cov⁡[θn+1,F(θn+1)∣Yn],\widehat{C}_{n+1}^{\theta x}=\operatorname{Cov}[\theta_{n+1},x_{n+1}\mid Y_n]=\operatorname{Cov}[\theta_{n+1},\mathcal{F}(\theta_{n+1})\mid Y_n],
C^n+1xx=Cov⁡[xn+1∣Yn]=Cov⁡[F(θn+1)∣Yn]+Σν,n+1.(10)\widehat{C}_{n+1}^{xx}=\operatorname{Cov}[x_{n+1}\mid Y_n]=\operatorname{Cov}[\mathcal{F}(\theta_{n+1})\mid Y_n]+\Sigma_{\nu,n+1}. \tag*{(10)}

Conditioning the Gaussian approximation (9) on xn+1=xx_{n+1}=x yields a Gaussian approximation to the filtering distribution of θn+1\theta_{n+1} given Yn+1Y_{n+1}. The mean and covariance of this approximation are given by the Kalman update:

mn+1=m^n+1+C^n+1θx(C^n+1xx)−1(x−x^n+1),(11)m_{n+1}=\widehat{m}_{n+1}+\widehat{C}_{n+1}^{\theta x}(\widehat{C}_{n+1}^{xx})^{-1}(x-\widehat{x}_{n+1}), \tag*{(11)}
Cn+1=C^n+1−C^n+1θx(C^n+1xx)−1(C^n+1θx)T.(12)C_{n+1}=\widehat{C}_{n+1}-\widehat{C}_{n+1}^{\theta x}(\widehat{C}_{n+1}^{xx})^{-1}(\widehat{C}_{n+1}^{\theta x})^{T}. \tag*{(12)}

In ensemble Kalman methods, the predictive moments in (10) are estimated empirically, and the filtering distribution of θn+1\theta_{n+1} given Yn+1Y_{n+1} is approximated by an updated ensemble. For the predicted ensemble (8), we evaluate the forward map and compute the empirical means:

x^n+1j=F(θ^n+1j),x^n+1=1J∑j=1Jx^n+1j.(13)\widehat{x}_{n+1}^{j} = \mathcal{F}(\widehat{\theta}_{n+1}^{j}), \qquad\widehat{x}_{n+1} = \frac{1}{J}\sum_{j=1}^{J}\widehat{x}_{n+1}^{j}. \tag*{(13)}

We then form the normalized ensemble anomaly matrices

Z^n+1=1J−1[θ^n+11−mn ⋯ θ^n+1J−mn]∈RNθ×J,(14)\widehat{Z}_{n+1} = \frac{1}{\sqrt{J-1}}\left[\widehat{\theta}_{n+1}^{1}-m_n \ \cdots\ \widehat{\theta}_{n+1}^{J}-m_n\right]\in\mathbb{R}^{N_\theta\times J}, \tag*{(14)}
Y^n+1=1J−1[x^n+11−x^n+1 ⋯ x^n+1J−x^n+1]∈RNx×J.(15)\widehat{Y}_{n+1} = \frac{1}{\sqrt{J-1}}\left[\widehat{x}_{n+1}^{1}-\widehat{x}_{n+1} \ \cdots\ \widehat{x}_{n+1}^{J}-\widehat{x}_{n+1}\right]\in\mathbb{R}^{N_x\times J}. \tag*{(15)}

Here we used m^n+1=mn\widehat{m}_{n+1}=m_n. These matrices yield the empirical covariance estimates in (10)

C^n+1=Z^n+1Z^n+1T,C^n+1θx=Z^n+1Y^n+1T,C^n+1xx=Y^n+1Y^n+1T+Σν,n+1,(16)\widehat{C}_{n+1}=\widehat{Z}_{n+1}\widehat{Z}_{n+1}^{T}, \qquad\widehat{C}^{\theta x}_{n+1}=\widehat{Z}_{n+1}\widehat{Y}_{n+1}^{T}, \qquad\widehat{C}^{xx}_{n+1}=\widehat{Y}_{n+1}\widehat{Y}_{n+1}^{T}+\Sigma_{\nu,n+1}, \tag*{(16)}

where the observation noise covariance is included analytically.

The analysis step constructs an updated ensemble {θn+1j}j=1J\{\theta_{n+1}^{j}\}_{j=1}^{J} to approximate the filtering distribution. A key computational advantage of ensemble Kalman methods is that the particle updates can be implemented using ensemble anomaly matrices, without explicitly forming or storing the full Nθ×NθN_\theta\times N_\theta parameter covariance matrix. This is particularly useful in high-dimensional problems with J≪NθJ\ll N_\theta. We focus on ensemble adjustment Kalman inversion (EAKI), which is based on the ensemble adjustment Kalman filter [28]. EAKI updates the mean according to (11) and transforms the ensemble anomalies to reproduce the Kalman analysis covariance (12). Using the Woodbury identity, this covariance can be written as

Cn+1=Z^n+1(I+Y^n+1TΣν,n+1−1Y^n+1)−1Z^n+1T.(17)C_{n+1}=\widehat{Z}_{n+1}\left(I+\widehat{Y}_{n+1}^{T}\Sigma_{\nu,n+1}^{-1}\widehat{Y}_{n+1}\right)^{-1}\widehat{Z}_{n+1}^{T}. \tag*{(17)}

EAKI realizes this covariance through a linear transformation acting on the ensemble anomalies from the left. To construct this transformation, consider the compact singular value decomposition of Z^n+1\widehat{Z}_{n+1} and the associated spectral decomposition:

Compact SVD:Z^n+1=UZS^Z1/2VZT,Spectral decomposition:VZT(I+Y^n+1TΣν,n+1−1Y^n+1)−1VZ=WZSAWZT.(18)\begin{aligned} \text{Compact SVD:}\quad&\widehat{Z}_{n+1}=U_Z\widehat{S}_Z^{1/2}V_Z^{T},\\ \text{Spectral decomposition:}\quad&V_Z^{T}\left(I+\widehat{Y}_{n+1}^{T}\Sigma_{\nu,n+1}^{-1}\widehat{Y}_{n+1}\right)^{-1}V_Z=W_ZS_AW_Z^{T}. \tag*{(18)} \end{aligned}

Here S^Z\widehat{S}_Z and SAS_A are nonsingular diagonal matrices of size rank⁡(Z^n+1)\operatorname{rank}(\widehat{Z}_{n+1}). Define the transformation matrix

An+1=UZS^Z1/2WZSA1/2S^Z−1/2WZTUZT.(19)A_{n+1}=U_Z\widehat{S}_Z^{1/2}W_ZS_A^{1/2}\widehat{S}_Z^{-1/2}W_Z^{T}U_Z^{T}. \tag*{(19)}

The EAKI analysis step is then

mn+1=mn+C^n+1θx(C^n+1xx)−1(x−x^n+1),(20)m_{n+1}=m_n+\widehat{C}^{\theta x}_{n+1}\left(\widehat{C}^{xx}_{n+1}\right)^{-1}\left(x-\widehat{x}_{n+1}\right), \tag*{(20)}
θn+1j=mn+1+An+1(θ^n+1j−m^n+1),j=1,…,J.(21)\theta^j_{n+1}=m_{n+1}+A_{n+1}\left(\widehat{\theta}^j_{n+1}-\widehat{m}_{n+1}\right),\qquad j=1,\ldots,J. \tag*{(21)}

By construction, the normalized anomaly matrix Zn+1Z_{n+1} of the updated ensemble satisfies

Zn+1Zn+1T=Z^n+1(I+Y^n+1TΣ^ν,n+1−1Y^n+1)−1Z^n+1T=C^n+1−C^n+1θx(C^n+1xx)−1(C^n+1θx)T.(22)Z_{n+1}Z_{n+1}^{T}=\widehat{Z}_{n+1}\left(I+\widehat{Y}_{n+1}^{T}\widehat{\Sigma}_{\nu,n+1}^{-1}\widehat{Y}_{n+1}\right)^{-1}\widehat{Z}_{n+1}^{T}=\widehat{C}_{n+1}-\widehat{C}^{\theta x}_{n+1}\left(\widehat{C}^{xx}_{n+1}\right)^{-1}\left(\widehat{C}^{\theta x}_{n+1}\right)^{T}. \tag*{(22)}

Thus EAKI is a deterministic ensemble implementation of the Gaussian conditioning update (11).

Subspace property. Both standard EAKI and its covariance-inflated variant remain confined to the affine span of the initial ensemble. Let Q:=Range⁡(Z0)Q := \operatorname{Range}(Z_0). For both methods, the prediction step (8) preserves the anomaly subspace: Range⁡(Z^n+1)=Range⁡(Zn)\operatorname{Range}(\widehat{Z}_{n+1})=\operatorname{Range}(Z_n), since it either leaves the anomalies unchanged or rescales them by a nonzero factor. Moreover, the EAKI mean increment (18a) lies in Range⁡(Z^n+1)\operatorname{Range}(\widehat{Z}_{n+1}). By (17), An+1A_{n+1} has the same column space as Z^n+1\widehat{Z}_{n+1}, and its restriction to this subspace is invertible. Consequently, Range⁡(Zn+1)=Range⁡(An+1Z^n+1)=Range⁡(Z^n+1)\operatorname{Range}(Z_{n+1})=\operatorname{Range}(A_{n+1}\widehat{Z}_{n+1})=\operatorname{Range}(\widehat{Z}_{n+1}). It follows by induction that θnj∈m0+Q\theta^j_n\in m_0+Q and Range⁡(Zn)=Q\operatorname{Range}(Z_n)=Q for all n≥0n\ge0. Inflation therefore increases ensemble spread without enlarging the affine span of the initial ensemble.

Projected Dropout Mean Correction

Covariance inflation expands the predicted anomalies within QQ, but the ensemble mean remains in m0+Qm_0+Q. To allow the mean to explore beyond this affine space, we augment its update with directions generated by coordinate-wise dropout [47]. Setting selected coordinates of an anomaly (13a) to zero can change its direction and move it outside QQ. We evaluate these candidate directions through the forward map and retain the parts of their observation responses that are not represented by the original ensemble.

Specifically, at iteration nn, the diagonal mask Rn∈RNθ×NθR_n\in\mathbb{R}^{N_\theta\times N_\theta} is drawn independently of earlier masks from independent Bernoulli(λ\lambda) entries, where 0<λ<10<\lambda<1, with the entire mask resampled if it is zero. Applying the same mask to all predicted anomalies (13a) preserves their zero sum and gives the centered perturbations τ~n+1j=Rn(θ^n+1j−mn)\widetilde{\tau}^j_{n+1}=R_n(\widehat{\theta}^j_{n+1}-m_n). Their forward responses x~n+1j=F(mn+τ~n+1j)\widetilde{x}^j_{n+1}=\mathcal{F}(m_n+\widetilde{\tau}^j_{n+1}) have mean x~n+1=J−1∑j=1Jx~n+1j\widetilde{x}_{n+1}=J^{-1}\sum_{j=1}^{J}\widetilde{x}^j_{n+1}. The normalized parameter and observation anomaly matrices are then

Z~n+1=RnZ^n+1=1J−1[τ~n+11,…,τ~n+1J],(23)\widetilde{Z}_{n+1}=R_n\widehat{Z}_{n+1}=\frac{1}{\sqrt{J-1}}\left[\widetilde{\tau}^1_{n+1},\ldots,\widetilde{\tau}^J_{n+1}\right], \tag*{(23)}
Y~n+1=1J−1[x~n+11−x~n+1,…,x~n+1J−x~n+1].(24)\widetilde{Y}_{n+1}=\frac{1}{\sqrt{J-1}}\left[\widetilde{x}^1_{n+1}-\widetilde{x}_{n+1},\ldots,\widetilde{x}^J_{n+1}-\widetilde{x}_{n+1}\right]. \tag*{(24)}

The columns of Y^n+1\widehat{Y}_{n+1} in (13b) and Y~n+1\widetilde{Y}_{n+1} describe the observation changes sampled by the original and perturbed ensembles, respectively. To isolate complementary observation responses, we project Y~n+1\widetilde{Y}_{n+1} onto the orthogonal complement of Range⁡(Y^n+1)\operatorname{Range}(\widehat{Y}_{n+1}) in the observation-noise metric. This metric agrees with that of the least-squares objective, since Σν,n+1−1=ΔτΣν−1\Sigma_{\nu,n+1}^{-1}=\Delta\tau\Sigma_{\nu}^{-1}. The projected observation anomalies and the associated parameter directions are

Ln+1=(Y^n+1TΣν,n+1−1Y^n+1)†Y^n+1TΣν,n+1−1Y~n+1,(25)L_{n+1}=\left(\widehat{Y}_{n+1}^{T}\Sigma_{\nu,n+1}^{-1}\widehat{Y}_{n+1}\right)^{\dagger}\widehat{Y}_{n+1}^{T}\Sigma_{\nu,n+1}^{-1}\widetilde{Y}_{n+1}, \tag*{(25)}
Y~n+1⊥=Y~n+1−Y^n+1Ln+1,(26)\widetilde{Y}_{n+1}^{\perp}=\widetilde{Y}_{n+1}-\widehat{Y}_{n+1}L_{n+1}, \tag*{(26)}
Z~n+1⊥=Z~n+1−Z^n+1Ln+1.(27)\widetilde{Z}_{n+1}^{\perp}=\widetilde{Z}_{n+1}-\widehat{Z}_{n+1}L_{n+1}. \tag*{(27)}

Here (⋅)†(\cdot)^{\dagger} denotes the Moore–Penrose pseudoinverse. The common coefficients Ln+1L_{n+1} keep the parameter directions Z~n+1⊥\widetilde{Z}_{n+1}^{\perp} paired with the observation anomalies Y~n+1⊥\widetilde{Y}_{n+1}^{\perp}. Moreover, after whitening by Σν,n+1−1/2\Sigma_{\nu,n+1}^{-1/2}, the original and projected observation blocks are orthogonal:

Y^n+1TΣν,n+1−1Y~n+1⊥=0.(28)\widehat{Y}_{n+1}^{T}\Sigma_{\nu,n+1}^{-1}\widetilde{Y}_{n+1}^{\perp}=0. \tag*{(28)}

We construct the augmented anomaly matrices by combining the original and projected blocks:

Zn+1aug=[Z^n+1,Z~n+1⊥],Yn+1aug=[Y^n+1,Y~n+1⊥].(29)Z_{n+1}^{\mathrm{aug}}=\left[\widehat{Z}_{n+1},\widetilde{Z}_{n+1}^{\perp}\right],\qquad Y_{n+1}^{\mathrm{aug}}=\left[\widehat{Y}_{n+1},\widetilde{Y}_{n+1}^{\perp}\right]. \tag*{(29)}

In whitened observation coordinates, orthogonality (28) separates the two contributions to the joint Kalman gain

Kn+1aug:=Zn+1aug(Yn+1aug)T(Yn+1aug(Yn+1aug)T+Σν,n+1)−1=Z^n+1Y^n+1T(Y^n+1Y^n+1T+Σν,n+1)−1+Z~n+1⊥(Y~n+1⊥)T(Y~n+1⊥(Y~n+1⊥)T+Σν,n+1)−1(30)\begin{aligned} K_{n+1}^{\mathrm{aug}}:=Z_{n+1}^{\mathrm{aug}}(Y_{n+1}^{\mathrm{aug}})^{T}\left(Y_{n+1}^{\mathrm{aug}}(Y_{n+1}^{\mathrm{aug}})^{T}+\Sigma_{\nu,n+1}\right)^{-1} \\ &=\widehat{Z}_{n+1}\widehat{Y}_{n+1}^{T}\left(\widehat{Y}_{n+1}\widehat{Y}_{n+1}^{T}+\Sigma_{\nu,n+1}\right)^{-1} +\widetilde{Z}_{n+1}^{\perp}(\widetilde{Y}_{n+1}^{\perp})^{T}\left(\widetilde{Y}_{n+1}^{\perp}(\widetilde{Y}_{n+1}^{\perp})^{T}+\Sigma_{\nu,n+1}\right)^{-1} \tag*{(30)} \end{aligned}

original

projected

The original block responds to residual components in its observation span, while the projected block responds to components in the complementary span sampled by the projected anomalies. Applying this gain to the residual at the current mean mnm_n gives

mn+1=mn+Zn+1aug(Yn+1aug)T(Yn+1aug(Yn+1aug)T+Σν,n+1)−1(x−F(mn)).(31)m_{n+1}=m_n+Z_{n+1}^{\mathrm{aug}}(Y_{n+1}^{\mathrm{aug}})^{T}\left(Y_{n+1}^{\mathrm{aug}}(Y_{n+1}^{\mathrm{aug}})^{T}+\Sigma_{\nu,n+1}\right)^{-1}(x-\mathcal{F}(m_n)). \tag*{(31)}

For a linear forward map, F(mn)=x^n+1\mathcal{F}(m_n)=\widehat{x}_{n+1}, so this residual coincides with that used in the ordinary EAKI mean update (20).

Finally, for the ensemble particle update, we use the augmented blocks only to update the mean, allowing it to move beyond the initial affine span. The anomaly transform An+1A_{n+1} is computed from the original anomalies via (19), preserving the EAKI analysis covariance for the current predicted ensemble. The resulting particle update is

θn+1j=mn+1+An+1(θ^n+1j−mn),j=1,…,J.(32)\theta_{n+1}^{j}=m_{n+1}+A_{n+1}\left(\widehat{\theta}_{n+1}^{j}-m_n\right),\qquad j=1,\ldots,J. \tag*{(32)}

The linear theory in § 4.2 shows that, under the coordinate-coverage condition (46), this mean correction preserves EAKI’s exponential convergence within the initial ensemble subspace Q\mathcal{Q} and yields exponential mean-square decay of the complementary observable error.

Algorithm 1 Inflated projected-dropout EAKI

  1. Input: initial ensemble {θ0j}j=1J\{\theta_0^j\}_{j=1}^J, data xx, forward map F\mathcal{F}, covariance Σν\Sigma_\nu, step size Δτ\Delta\tau, dropout rate 1−λ1-\lambda, and iteration count NN.

  2. for n=0,1,…,N−1n=0,1,\ldots,N-1 do

  3. Apply the inflated prediction (8) and form Z^n+1\widehat{Z}_{n+1} and Y^n+1\widehat{Y}_{n+1}.

  4. Compute An+1A_{n+1} from the original predicted anomalies using (16) and (17).

  5. Generate RnR_n, resampling if the resulting mask is all zeros, and compute the masked anomalies via (20).

  6. Compute the projected and augmented anomalies by (21) and (23).

  7. Update the mean by (25) and the particles by (26).

  8. end for

  9. Output: corrected ensemble {θNj}j=1J\{\theta_N^j\}_{j=1}^J and mean mNm_N.

Remark 1 (Geometry of the projection). The observation projection (21) measures parameter directions through their noise-weighted effects on the forward response. If F(θ)=Fθ\mathcal{F}(\theta)=F\theta, then Y^n+1=FZ^n+1\widehat{Y}_{n+1}=F\widehat{Z}_{n+1} and Y~n+1⊥=FZ~n+1⊥\widetilde{Y}_{n+1}^{\perp}=F\widetilde{Z}_{n+1}^{\perp}. Since Σν,n+1−1=ΔτΣν−1\Sigma_{\nu,n+1}^{-1}=\Delta\tau\Sigma_\nu^{-1}, the weighted orthogonality above becomes

Z^n+1TFTΣν−1FZ~n+1⊥=0.\widehat{Z}_{n+1}^{T}F^{T}\Sigma_\nu^{-1}F\widetilde{Z}_{n+1}^{\perp}=0.

Here FTΣν−1FF^{T}\Sigma_\nu^{-1}F is the Hessian of the weighted least-squares objective. Thus the projected parameter anomalies are orthogonal to the original ones in the Hessian form, which accounts for forward sensitivity and observation noise.

The proposed method requires 2J+12J+1 evaluations of the forward map per iteration. Algorithm 1 summarizes the complete algorithm.

Theoretical Analysis

In this section, we provide a discrete-time, finite-ensemble analysis of the proposed EAKI method for linear inverse problems to clarify the roles of covariance inflation and projected dropout. We assume that the forward map takes the form

F(θ)=Fθ,(33)\mathcal{F}(\theta)=F\theta, \tag*{(33)}

and the associated weighted least-squares objective is ΦR(m)=12∥Σν−1/2(x−Fm)∥2\Phi_R(m)=\frac{1}{2}\left\lVert\Sigma_\nu^{-1/2}(x-Fm)\right\rVert^2. For an initial ensemble of size J≥2J\ge2, let m0∈RNθm_0\in\mathbb{R}^{N_\theta}, C0∈RNθ×NθC_0\in\mathbb{R}^{N_\theta\times N_\theta}, and Z0∈RNθ×JZ_0\in\mathbb{R}^{N_\theta\times J} denote its empirical mean, empirical covariance, and normalized anomaly matrix, respectively, so that C0=Z0Z0TC_0=Z_0Z_0^T. Let r=rank⁡(C0)r=\operatorname{rank}(C_0). Since the ensemble anomalies are centered, r≤min⁡{Nθ,J−1}r\le\min\{N_\theta,J-1\}. The covariance C0C_0 admits the compact spectral decomposition

C0=QD0QT,(34)C_0=QD_0Q^T, \tag*{(34)}

where Q∈RNθ×rQ \in\mathbb{R}^{N_\theta\times r} satisfies QTQ=IrQ^TQ = I_r, and D0∈Rr×rD_0 \in\mathbb{R}^{r \times r} is diagonal with strictly positive entries. The columns of QQ form an orthonormal basis of the initial anomaly subspace.

In Section 4.1, we show that covariance inflation yields exponential convergence of the ensemble mean within the initial affine span, with a convergence factor independent of conditioning. In Section 4.2, we analyze the full method and show that projected dropout enables exploration beyond the initial affine span and yields exponential convergence of the expected least-squares objective value to its minimum, at a rate that may depend on conditioning.

Covariance inflation

Without the dropout correction, the inflated EAKI updates in Eqs. (8) and (18) remain confined to the initial affine span m0+Range⁡(Q)m_0 + \operatorname{Range}(Q). To account for the initial covariance, we introduce the normalized coordinates θ=m0+QD01/2z\theta= m_0 + QD_0^{1/2}z. In the zz-coordinates, the initial ensemble has zero empirical mean and identity empirical covariance.

The least-squares problem restricted to this affine space need not have a unique minimizer, since Range⁡(Q)\operatorname{Range}(Q) may contain nonzero directions in ker⁡(F)\ker(F). To separate observable and unobservable directions, we consider the compact spectral decomposition of the least-squares Hessian in the zz-coordinates:

D01/2QTFTΣν−1FQD01/2=P1ΓP1T,(35)D_0^{1/2}Q^TF^T\Sigma_\nu^{-1}FQD_0^{1/2} = P_1\Gamma P_1^T, \tag*{(35)}

where r′=rank⁡(FQ)r' = \operatorname{rank}(FQ) and P1∈Rr×r′P_1 \in\mathbb{R}^{r \times r'}, P1TP1=Ir′P_1^T P_1 = I_{r'}, Γ=diag⁡(γ1,…,γr′)\Gamma= \operatorname{diag}(\gamma_1,\ldots,\gamma_{r'}), γi>0\gamma_i > 0. Let P2∈Rr×(r−r′)P_2 \in\mathbb{R}^{r \times(r-r')} have orthonormal columns spanning the null space of this Hessian, and define the orthogonal matrix P=[P1 P2]∈Rr×rP = [P_1\ P_2] \in\mathbb{R}^{r \times r}. This decomposition separates the normalized coordinates into observable and unobservable components, in the sense that

FQD01/2P2=0.(36)FQD_0^{1/2}P_2 = 0. \tag*{(36)}

Thus, writing m=m0+QD01/2(P1b+P2c)m = m_0 + QD_0^{1/2}(P_1b + P_2c), with b∈Rr′b \in\mathbb{R}^{r'}, c∈Rr−r′c \in\mathbb{R}^{r-r'}, the data misfit depends only on bb. The full and reduced problems therefore have the same minimum value:

min⁡z∈Rr12∥Σν−1/2(x−Fm0−FQD01/2z)∥2=min⁡b∈Rr′12∥Σν−1/2(x−Fm0−FQD01/2P1b)∥2.\min_{z \in\mathbb{R}^r} \frac{1}{2}\left\|\Sigma_\nu^{-1/2}\left(x-Fm_0-FQD_0^{1/2}z\right)\right\|^2 = \min_{b \in\mathbb{R}^{r'}} \frac{1}{2}\left\|\Sigma_\nu^{-1/2}\left(x-Fm_0-FQD_0^{1/2}P_1b\right)\right\|^2.

The reduced problem in bb has Hessian Γ≻0\Gamma\succ0 and hence the unique minimizer

b⋆=Γ−1P1TD01/2QTFTΣν−1(x−Fm0).(37)b_\star= \Gamma^{-1}P_1^TD_0^{1/2}Q^TF^T\Sigma_\nu^{-1}(x-Fm_0). \tag*{(37)}

We select the minimum-norm representative z⋆=P1b⋆z_\star= P_1b_\star, and define the distinguished minimizer over the initial affine span by m⋆=m0+QD01/2P1b⋆m_\star= m_0 + QD_0^{1/2}P_1b_\star.

Theorem 1 (Inflated EAKI in the linear case). Assume that 0<Δτ<10 < \Delta\tau< 1, Σν≻0\Sigma_\nu\succ0, and F(θ)=FθF(\theta) = F\theta. Then, for the inflated EAKI, the following statements hold:

(i) Mean and covariance recursions. The empirical mean and covariance satisfy

mn+1=mn+CnFT(FCnFT+1−ΔτΔτΣν)−1(x−Fmn),(38)m_{n+1}=m_n+C_nF^T\left(FC_nF^T+\frac{1-\Delta\tau}{\Delta\tau}\Sigma_\nu\right)^{-1}(x-Fm_n), \tag*{(38)}
Cn+1=11−Δτ[Cn−CnFT(FCnFT+1−ΔτΔτΣν)−1FCn].(39)C_{n+1}=\frac{1}{1-\Delta\tau}\left[C_n-C_nF^T\left(FC_nF^T+\frac{1-\Delta\tau}{\Delta\tau}\Sigma_\nu\right)^{-1}FC_n\right]. \tag*{(39)}

(ii) Exponential covariance convergence in observable directions. For every n≥0n\ge0, Cn=QDnQTC_n=QD_nQ^T with Dn≻0D_n\succ0 satisfying

Dn−1=QTFTΣν−1FQ+(1−Δτ)n(D0−1−QTFTΣν−1FQ).(40)D_n^{-1}=Q^TF^T\Sigma_\nu^{-1}FQ+(1-\Delta\tau)^n\left(D_0^{-1}-Q^TF^T\Sigma_\nu^{-1}FQ\right). \tag*{(40)}

Hence the pseudoinverse Cn†=QDn−1QTC_n^\dagger=QD_n^{-1}Q^T converges exponentially to the projected Hessian QQTFTΣν−1FQQTQQ^TF^T\Sigma_\nu^{-1}FQQ^T. Moreover,

Dn=D01/2(P1ΛnP1T+(1−Δτ)−nP2P2T)D01/2,(41)D_n=D_0^{1/2}\left(P_1\Lambda_nP_1^T+(1-\Delta\tau)^{-n}P_2P_2^T\right)D_0^{1/2}, \tag*{(41)}

where Λn:=[(1−Δτ)nIr+(1−(1−Δτ)n)Γ]−1\Lambda_n:=\left[(1-\Delta\tau)^nI_{r}+\left(1-(1-\Delta\tau)^n\right)\Gamma\right]^{-1}.

(iii) Exponential mean convergence in the initial ensemble subspace.

mn=m0+QD01/2P1bn,bn−b⋆=−(1−Δτ)nΛnb⋆,(42)m_n=m_0+QD_0^{1/2}P_1b_n,\qquad b_n-b_\star=-(1-\Delta\tau)^n\Lambda_nb_\star, \tag*{(42)}

and

∥bn−b⋆∥2≤∥b⋆∥2min⁡{1,λmin⁡(Γ)}(1−Δτ)n.(43)\lVert b_n-b_\star\rVert_2\le\frac{\lVert b_\star\rVert_2}{\min\{1,\lambda_{\min}(\Gamma)\}}(1-\Delta\tau)^n. \tag*{(43)}

Thus the geometric convergence factor is 1−Δτ1-\Delta\tau, independently of the condition number of Γ\Gamma, although the multiplicative constants may depend on Γ\Gamma.

Proof. For (i), the prediction step (8) gives, in the linear case,

m^n+1=mn,x^n+1=Fmn,Z^n+1=(1−Δτ)−1/2Zn,Y^n+1=FZ^n+1.\widehat{m}_{n+1}=m_n,\qquad\widehat{x}_{n+1}=Fm_n,\qquad\widehat{Z}_{n+1}=(1-\Delta\tau)^{-1/2}Z_n,\qquad\widehat{Y}_{n+1}=F\widehat{Z}_{n+1}.

Substituting these identities and the noise scaling (7) into the EAKI mean and covariance updates (20) and (22) gives (32). Since rank⁡(Cn)≤J−1\operatorname{rank}(C_n)\le J-1, the covariance is singular whenever J≤NθJ\le N_\theta.

For (ii), the base case follows from (34). Suppose Cn=QDnQTC_n=QD_nQ^T with Dn≻0D_n\succ0. Applying the Woodbury identity to the covariance update (39) gives

Cn+1=Q((1−Δτ)Dn−1+ΔτQTFTΣν−1FQ)−1QT.C_{n+1}=Q\left((1-\Delta\tau)D_n^{-1}+\Delta\tau Q^TF^T\Sigma_\nu^{-1}FQ\right)^{-1}Q^T.

Thus Cn+1=QDn+1QTC_{n+1}=QD_{n+1}Q^T with Dn+1−1=(1−Δτ)Dn−1+ΔτQTFTΣν−1FQ≻0D_{n+1}^{-1}=(1-\Delta\tau)D_n^{-1}+\Delta\tau Q^TF^T\Sigma_\nu^{-1}FQ\succ0. Solving the recursion gives (40). Multiplying that identity on the left and right by D01/2D_0^{1/2} and using (35), we obtain

D01/2Dn−1D01/2=(1−Δτ)nIr+(1−(1−Δτ)n)P1ΓP1T=P1Λn−1P1T+(1−Δτ)nP2P2T.D_0^{1/2}D_n^{-1}D_0^{1/2}=(1-\Delta\tau)^nI_r+\left(1-(1-\Delta\tau)^n\right)P_1\Gamma P_1^T=P_1\Lambda_n^{-1}P_1^T+(1-\Delta\tau)^nP_2P_2^T.

Inverting this identity yields (41). Since CnT=QDn−1QTC_n^T = QD_n^{-1}Q^T, its exponential convergence follows directly from (40).

For (iii)(iii), we rewrite the increment in the mean update (38) using the following identity:

CnFT(FCnFT+1−ΔτΔτΣν)−1=ΔτCn+1FTΣν−1=ΔτQDn+1QTFTΣν−1=ΔτQD01/2P1Λn+1P1TD01/2QTFTΣν−1.(44)\begin{aligned} C_nF^T\left(FC_nF^T+\frac{1-\Delta\tau}{\Delta\tau}\Sigma_\nu\right)^{-1} =\Delta\tau C_{n+1}F^T\Sigma_\nu^{-1} =\Delta\tau QD_{n+1}Q^TF^T\Sigma_\nu^{-1} \\ &=\Delta\tau QD_0^{1/2}P_1\Lambda_{n+1}P_1^TD_0^{1/2}Q^TF^T\Sigma_\nu^{-1}. \tag*{(44)} \end{aligned}

Here the first equality uses the covariance update (39), the second uses Cn+1=QDn+1QTC_{n+1}=QD_{n+1}Q^T, and the last follows from (41) and the null-space identity (36) of P2P_2. Thus every mean increment lies in Range⁡(QD01/2P1)\operatorname{Range}(QD_0^{1/2}P_1), so mn=m0+QD01/2P1bnm_n=m_0+QD_0^{1/2}P_1b_n with b0=0b_0=0. Substituting this representation into (38) and using the definitions of Γ\Gamma and b⋆b_\star yields

bn+1=bn+ΔτΛn+1P1TD01/2QTFTΣν−1(x−Fm0−FQD01/2P1bn)=bn+ΔτΛn+1Γ(b⋆−bn)=b⋆+(1−Δτ)Λn+1Λn−1(bn−b⋆).(45)\begin{aligned} b_{n+1}=b_n+\Delta\tau\Lambda_{n+1}P_1^TD_0^{1/2}Q^TF^T\Sigma_\nu^{-1}(x-Fm_0-FQD_0^{1/2}P_1b_n) \\ &=b_n+\Delta\tau\Lambda_{n+1}\Gamma(b_\star-b_n) \\ &=b_\star+(1-\Delta\tau)\Lambda_{n+1}\Lambda_n^{-1}(b_n-b_\star). \tag*{(45)} \end{aligned}

The last equality uses the recursion Λn+1−1=(1−Δτ)Λn−1+ΔτΓ\Lambda_{n+1}^{-1}=(1-\Delta\tau)\Lambda_n^{-1}+\Delta\tau\Gamma, from part (ii), which implies I−ΔτΛn+1Γ=(1−Δτ)Λn+1Λn−1I-\Delta\tau\Lambda_{n+1}\Gamma=(1-\Delta\tau)\Lambda_{n+1}\Lambda_n^{-1}. Since Λ0=Ir′\Lambda_0=I_{r'} and b0=0b_0=0, iteration gives

bn−b⋆=−(1−Δτ)nΛnb⋆,b_n-b_\star=-(1-\Delta\tau)^n\Lambda_n b_\star,

proving (42). Finally, every eigenvalue of Λn−1=(1−Δτ)nIr′+(1−(1−Δτ)n)Γ\Lambda_n^{-1}=(1-\Delta\tau)^nI_{r'}+(1-(1-\Delta\tau)^n)\Gamma is bounded below by min⁡{1,λmin⁡(Γ)}\min\{1,\lambda_{\min}(\Gamma)\}. Combining this bound with (42) yields (43).

Remark 2 (Unobservable covariance directions). If the initial ensemble subspace contains nonzero directions in the null space of FF, then r′<rr'<r in (35). The second term in the covariance formula (41) grows like (1−Δτ)−n(1-\Delta\tau)^{-n}. Consequently, the uncertainty represented by the ensemble in these directions tends to infinity. This behavior emphasizes the importance of incorporating the prior distribution through the augmented formulation (4), which regularizes these otherwise unobservable directions.

Projected Dropout Correction

We next analyze the projected dropout mean correction introduced in § 3.2.

Theorem 2 (Inflated Projected Dropout EAKI in the linear case). Assume the linear setting and notation of Theorem 1. For each nn, independently draw a vector with i.i.d. Bernoulli(λ\lambda) entries, where 0<λ<10<\lambda<1, resampling the entire vector whenever it is zero, and let RnR_n be the diagonal matrix formed from the accepted vector. Recall that the columns of QQ form an orthonormal basis of the initial anomaly subspace. Assume r′=rank⁡(FQ)≥1r'=\operatorname{rank}(FQ)\ge1, and define

S:=Range⁡(Σν−1/2FQ),R:=S⊥∩Range⁡(Σν−1/2F),\mathcal{S}:=\operatorname{Range}(\Sigma_\nu^{-1/2}FQ),\qquad\mathcal{R}:=\mathcal{S}^{\perp}\cap\operatorname{Range}(\Sigma_\nu^{-1/2}F),

where S⊥S^{\perp} is the orthogonal complement of SS in RNx\mathbb{R}^{N_x}. Let P⊥P_{\perp} denote the Euclidean projection onto S⊥S^{\perp}. Suppose that the initial ensemble subspace satisfies the coordinate-coverage condition

ℓQ:=min⁡1≤i≤NθeiTQQTei>0.(46)\ell_Q := \min_{1\le i\le N_{\theta}} e_i^{T}QQ^{T}e_i > 0. \tag*{(46)}

Let m∗m^{*} be a minimizer of

min⁡m∈RNθΦR(m),ΦR(m)=12∥Σν−1/2(x−Fm)∥2,\min_{m\in\mathbb{R}^{N_{\theta}}}\Phi_R(m), \qquad\Phi_R(m)=\frac{1}{2}\left\|\Sigma_{\nu}^{-1/2}(x-Fm)\right\|^{2},

and let b⋆b_{\star} be defined by (37). Then, for every n≥0n\ge0, the following statements hold. Throughout, all expectations are taken only over the dropout masks.

(i) Covariance preservation. The projected dropout mean correction leaves the covariance update unchanged. Consequently, all identities and convergence estimates for the covariance in Theorem 1 remain valid.

(ii) In-subspace mean convergence. There exist bn∈Rr′b_n\in\mathbb{R}^{r'} and un∈RNθu_n\in\mathbb{R}^{N_{\theta}} such that

mn=m0+QD01/2P1bn+un,Σν−1/2Fun∈S⊥.(47)m_n=m_0+QD_0^{1/2}P_1b_n+u_n,\qquad\Sigma_{\nu}^{-1/2}Fu_n\in S^{\perp}. \tag*{(47)}

The reduced sequence bnb_n is exactly the same as in the no-dropout iteration:

bn−b⋆=−(1−Δτ)nΛnb⋆.(48)b_n-b_{\star}=-(1-\Delta\tau)^n\Lambda_n b_{\star}. \tag*{(48)}

In particular, the same geometric convergence bound holds:

∥bn−b⋆∥2≤∥b⋆∥2min⁡{1,λmin⁡(Γ)}(1−Δτ)n.(49)\left\|b_n-b_{\star}\right\|_2\le\frac{\left\|b_{\star}\right\|_2}{\min\{1,\lambda_{\min}(\Gamma)\}}(1-\Delta\tau)^n. \tag*{(49)}

(iii) Complementary-space contraction. If R≠{0}\mathcal{R}\ne\{0\}, then there exists 0<β<10<\beta<1, depending only on C0C_0, FF, Σν\Sigma_{\nu}, Δτ\Delta\tau, λ\lambda, such that the orthogonal observation-space error satisfies

E[∥P⊥Σν−1/2F(mn−m∗)∥2]≤(1−β)n∥P⊥Σν−1/2F(m0−m∗)∥2.(50)\mathbb{E}\left[\left\|P_{\perp}\Sigma_{\nu}^{-1/2}F(m_n-m^{*})\right\|^{2}\right]\le(1-\beta)^n\left\|P_{\perp}\Sigma_{\nu}^{-1/2}F(m_0-m^{*})\right\|^{2}. \tag*{(50)}

If R={0}\mathcal{R}=\{0\}, the orthogonal error is identically zero.

(iv) Objective convergence. Furthermore, the objective error satisfies

E[ΦR(mn)−ΦR(m∗)]≤12∥Γ∥2(∥b⋆∥2min⁡{1,λmin⁡(Γ)})2(1−Δτ)2n+12(1−β)n∥P⊥Σν−1/2F(m0−m∗)∥2.(51)\mathbb{E}[\Phi_R(m_n)-\Phi_R(m^{*})]\le\frac{1}{2}\|\Gamma\|_2\left(\frac{\|b_{\star}\|_2}{\min\{1,\lambda_{\min}(\Gamma)\}}\right)^2(1-\Delta\tau)^{2n} +\frac{1}{2}(1-\beta)^n\left\|P_{\perp}\Sigma_{\nu}^{-1/2}F(m_0-m^{*})\right\|^2. \tag*{(51)}

When R={0}\mathcal{R}=\{0\}, the second term is zero.

Remark 3 (Observation subspaces and convergence mechanism). In the linear setting, minimizing ΦR\Phi_R is equivalent to eliminating the whitened observable error, since

ΦR(mn)−ΦR(m∗)=12∥Σν−1/2F(mn−m∗)∥2.\Phi_R(m_n)-\Phi_R(m^{*})=\frac{1}{2}\left\|\Sigma_{\nu}^{-1/2}F(m_n-m^{*})\right\|^2.

The observable range decomposes as Range⁡(Σν−1/2F)=S⊕R\operatorname{Range}(\Sigma_{\nu}^{-1/2}F)=\mathcal{S}\oplus\mathcal{R}, where S\mathcal{S} contains the observation directions represented by the initial ensemble and R\mathcal{R} contains the remaining observable directions. On this range, P⊥P_{\perp} extracts the component in R\mathcal{R}. The projected dropout correction preserves the EAKI update in S\mathcal{S} and supplies an additional correction in R\mathcal{R}.

Proof of Theorem 2. For (i), the particle update (32) gives

θn+1j−mn+1=An+1(θ^n+1j−mn),j=1,…,J.(52)\theta_{n+1}^{j}-m_{n+1}=A_{n+1}\left(\widehat{\theta}_{n+1}^{j}-m_n\right),\qquad j=1,\ldots,J. \tag*{(52)}

Hence Zn+1=An+1Z^n+1Z_{n+1}=A_{n+1}\widehat{Z}_{n+1}, exactly the anomaly update of EAKI; translating the particles to the corrected mean does not alter their centered anomalies. The same translation preserves the EAKI covariance. In the linear case, the covariance recursion depends only on the covariance, so all covariance identities in Theorem 1 remain valid. In particular, Range⁡(Z^n+1)=Range⁡(Q)\operatorname{Range}(\widehat{Z}_{n+1})=\operatorname{Range}(Q), since Dn≻0D_n\succ0.

For (ii), the decomposition (30) expresses the Kalman gain as the sum of the original EAKI gain and the projected dropout gain. The former equals ΔτCn+1FTΣν−1\Delta\tau C_{n+1}F^{T}\Sigma_{\nu}^{-1} by (44). Thus, the mean update (31) takes the form

mn+1=mn+ΔτCn+1FTΣν−1(x−Fmn)+dn+1,(53)m_{n+1}=m_n+\Delta\tau C_{n+1}F^{T}\Sigma_{\nu}^{-1}(x-Fm_n)+d_{n+1}, \tag*{(53)}
dn+1:=Z~n+1⊥(Y~n+1⊥)T(Y~n+1⊥(Y~n+1⊥)T+Σν,n+1)−1(x−Fmn).(54)d_{n+1}:=\widetilde{Z}_{n+1}^{\perp}(\widetilde{Y}_{n+1}^{\perp})^{T}\left(\widetilde{Y}_{n+1}^{\perp}(\widetilde{Y}_{n+1}^{\perp})^{T}+\Sigma_{\nu,n+1}\right)^{-1}(x-Fm_n). \tag*{(54)}

By part (i), Range⁡(Σν−1/2Y^n+1)=S\operatorname{Range}(\Sigma_{\nu}^{-1/2}\widehat{Y}_{n+1})=\mathcal{S}. Weighted orthogonality and FZ~n+1⊥=Y~n+1⊥F\widetilde{Z}_{n+1}^{\perp}=\widetilde{Y}_{n+1}^{\perp} therefore give Range⁡(Σν−1/2FZ~n+1⊥)⊆R\operatorname{Range}(\Sigma_{\nu}^{-1/2}F\widetilde{Z}_{n+1}^{\perp})\subseteq\mathcal{R}, so Σν−1/2Fdn+1∈R\Sigma_{\nu}^{-1/2}Fd_{n+1}\in\mathcal{R}.

The decomposition (47) holds at n=0n=0 with b0=0b_0=0 and u0=0u_0=0. Suppose it holds at step nn. The accumulated dropout displacement unu_n does not affect the original EAKI increment, since

QTFTΣν−1Fun=(Σν−1/2FQ)T(Σν−1/2Fun)=0.Q^{T}F^{T}\Sigma_{\nu}^{-1}Fu_n=(\Sigma_{\nu}^{-1/2}FQ)^{T}(\Sigma_{\nu}^{-1/2}Fu_n)=0.

Substituting (47) at nn into (53) and using the gain identity (44) together with the definition of b⋆b_{\star} in (37) yields

bn+1=bn+ΔτΛn+1Γ(b⋆−bn),un+1=un+dn+1.b_{n+1}=b_n+\Delta\tau\Lambda_{n+1}\Gamma(b_{\star}-b_n),\qquad u_{n+1}=u_n+d_{n+1}.

The second update preserves Σν−1/2Fun+1∈S⊥\Sigma_{\nu}^{-1/2}Fu_{n+1}\in\mathcal{S}^{\perp}, completing the induction. The first update is exactly (45), yielding (48) and (49) for every sequence of masks.

For (iii), let Hn:=σ(R0,…,Rn−1)\mathcal{H}_n:=\sigma(R_0,\ldots,R_{n-1}) denote the mask history. Conditional expectations given Hn\mathcal{H}_n average over the newly sampled mask RnR_n. Define the orthogonal observation-space error

hn:=P⊥Σν−1/2F(mn−m∗)∈R.(55)h_n:=P_{\perp}\Sigma_{\nu}^{-1/2}F(m_n-m^{\ast})\in\mathcal{R}. \tag*{(55)}

The inclusion hn∈Rh_n\in\mathcal{R} follows from S⊆Range⁡(Σν−1/2F)\mathcal{S}\subseteq\operatorname{Range}(\Sigma_{\nu}^{-1/2}F): projecting a vector in this observable range onto S⊥\mathcal{S}^{\perp} yields a vector in R\mathcal{R}. If R={0}\mathcal{R}=\{0\}, then (55) gives hn=0h_n = 0 for all nn, and we take β=0\beta= 0. In the following, assume R≠{0}\mathcal{R} \ne\{0\}. To derive a recursion for hn+1h_{n+1}, we first show that the original EAKI increment in (46a) does not contribute to hn+1h_{n+1}. Indeed, Cn+1=QDn+1QTC_{n+1} = QD_{n+1}Q^{T} and P⊥Σν−1/2FQ=0P_{\perp}\Sigma_{\nu}^{-1/2}FQ = 0, so P⊥Σν−1/2F[ΔτCn+1FTΣν−1(x−Fmn)]=0P_{\perp}\Sigma_{\nu}^{-1/2}F[\Delta\tau C_{n+1}F^{T}\Sigma_{\nu}^{-1}(x-Fm_n)] = 0. Next, to express the contribution of dropout increment dn+1d_{n+1} in (46a) to hn+1h_{n+1}, we introduce whitened anomalies, and the Gram matrix of whitened anomalies:

En+1:=Σν−1/2Y~n+1⊥,Gn+1:=En+1En+1T.(56)E_{n+1} := \Sigma_{\nu}^{-1/2}\widetilde{Y}_{n+1}^{\perp}, \qquad G_{n+1} := E_{n+1}E_{n+1}^{T}. \tag*{(56)}

Part (ii)(ii) gives Range⁡(En+1)⊆R\operatorname{Range}(E_{n+1}) \subseteq\mathcal{R}. The first-order optimality condition FTΣν−1(x−Fm∗)=0F^{T}\Sigma_{\nu}^{-1}(x-Fm^{*}) = 0 and Range⁡(Gn+1)=Range⁡(En+1)⊆R\operatorname{Range}(G_{n+1}) = \operatorname{Range}(E_{n+1}) \subseteq\mathcal{R} imply Gn+1Σν−1/2(x−Fmn)=−Gn+1hnG_{n+1}\Sigma_{\nu}^{-1/2}(x-Fm_n) = -G_{n+1}h_n. Since Y~n+1⊥=Σν1/2En+1\widetilde{Y}_{n+1}^{\perp} = \Sigma_{\nu}^{1/2}E_{n+1} and Σν,n+1=1ΔτΣν\Sigma_{\nu,n+1} = \frac{1}{\Delta\tau}\Sigma_{\nu}, the covariance in the dropout gain factors as

Y~n+1⊥(Y~n+1⊥)T+Σν,n+1=Σν1/2(Gn+1+Δτ−1I)Σν1/2.\widetilde{Y}_{n+1}^{\perp}(\widetilde{Y}_{n+1}^{\perp})^{T}+\Sigma_{\nu,n+1}=\Sigma_{\nu}^{1/2}(G_{n+1}+\Delta\tau^{-1}I)\Sigma_{\nu}^{1/2}.

Inverting this factorization and using P⊥Σν−1/2FZ~n+1⊥=P⊥En+1=En+1P_{\perp}\Sigma_{\nu}^{-1/2}F\widetilde{Z}_{n+1}^{\perp}=P_{\perp}E_{n+1}=E_{n+1}, we can left-multiply (46a) by P⊥Σν−1/2FP_{\perp}\Sigma_{\nu}^{-1/2}F and then obtain

hn+1=hn+P⊥Σν−1/2Fdn+1=hn+En+1(Y~n+1⊥)TΣν−1/2(Gn+1+Δτ−1I)−1Σν−1/2(x−Fmn)=hn+Gn+1(Gn+1+Δτ−1I)−1Σν−1/2(x−Fmn)=hn−Gn+1(Gn+1+Δτ−1I)−1hn=(ΔτGn+1+I)−1hn.(57)\begin{aligned} h_{n+1} &= h_n + P_{\perp}\Sigma_{\nu}^{-1/2}Fd_{n+1}\\ &= h_n + E_{n+1}(\widetilde{Y}_{n+1}^{\perp})^{T}\Sigma_{\nu}^{-1/2}(G_{n+1}+\Delta\tau^{-1}I)^{-1}\Sigma_{\nu}^{-1/2}(x-Fm_n)\\ &= h_n + G_{n+1}(G_{n+1}+\Delta\tau^{-1}I)^{-1}\Sigma_{\nu}^{-1/2}(x-Fm_n)\\ &= h_n - G_{n+1}(G_{n+1}+\Delta\tau^{-1}I)^{-1}h_n\\ &= (\Delta\tau G_{n+1}+I)^{-1}h_n. \tag*{(57)} \end{aligned}

By (49), the change in squared error is

∥hn+1∥2−∥hn∥2=−hnT[I−(I+ΔτGn+1)−2]hn.(58)\lVert h_{n+1}\rVert^{2}-\lVert h_n\rVert^{2}=-h_n^{T}[I-(I+\Delta\tau G_{n+1})^{-2}]h_n. \tag*{(58)}

Next we will estimate the operator I−(I+ΔτGn+1)−2I-(I+\Delta\tau G_{n+1})^{-2} on R\mathcal{R} by comparing Gn+1G_{n+1} with a bounded reference matrix. Since the projection P⊥P_{\perp} only removes the original observation component, (21) yields

En+1=P⊥Σν−1/2F(RnZ^n+1−Z^n+1Ln+1)=P⊥Σν−1/2FRnZ^n+1.(59)E_{n+1}=P_{\perp}\Sigma_{\nu}^{-1/2}F\left(R_n\widehat{Z}_{n+1}-\widehat{Z}_{n+1}L_{n+1}\right)=P_{\perp}\Sigma_{\nu}^{-1/2}FR_n\widehat{Z}_{n+1}. \tag*{(59)}

and thus we can rewrite Gn+1=En+1En+1TG_{n+1}=E_{n+1}E_{n+1}^{T} as

Gn+1=P⊥Σν−1/2FRn(Z^n+1Z^n+1T)RnFTΣν−1/2P⊥.(60)G_{n+1}=P_{\perp}\Sigma_{\nu}^{-1/2}FR_n(\widehat{Z}_{n+1}\widehat{Z}_{n+1}^{T})R_nF^{T}\Sigma_{\nu}^{-1/2}P_{\perp}. \tag*{(60)}

Introduce the reference matrix

G‾n+1:=P⊥Σν−1/2FRnQQTRnFTΣν−1/2P⊥.(61)\overline{G}_{n+1}:=P_{\perp}\Sigma_{\nu}^{-1/2}FR_nQQ^{T}R_nF^{T}\Sigma_{\nu}^{-1/2}P_{\perp}. \tag*{(61)}

To compare it with Gn+1G_{n+1}, set γ2:=1max⁡{∥Dn−1∥2,∥FTΣν−1F∥2}\gamma^{2}:=\frac{1}{\max\{\lVert D_n^{-1}\rVert_{2},\lVert F^{T}\Sigma_{\nu}^{-1}F\rVert_{2}\}}. The convex-combination form of (33) and QTQ=IrQ^{T}Q=I_r imply Dn−1⪰γ2IrD_n^{-1}\succeq\gamma^{2}I_r. Hence the inflated prediction satisfies

Z^n+1Z^n+1T=QDn1−ΔτQT⪰γ21−ΔτQQT⪰γ2QQT.(62)\widehat{Z}_{n+1}\widehat{Z}_{n+1}^{T}=Q\frac{D_n}{1-\Delta\tau}Q^{T}\succeq\frac{\gamma^{2}}{1-\Delta\tau}QQ^{T}\succeq\gamma^{2}QQ^{T}. \tag*{(62)}

Combining this bound with (52) gives Gn+1⪰γ2G‾n+1G_{n+1} \succeq\gamma^2\overline{G}_{n+1}. Moreover, since ∥Q∥2=1\lVert Q\rVert_2 = 1 and ∥Rn∥2,∥P⊥∥2≤1\lVert R_n\rVert_2, \lVert P_{\perp}\rVert_2 \le1,

0⪯G‾n+1⪯M2I,M:=∥Σν−1/2F∥2.0 \preceq\overline{G}_{n+1} \preceq M^2 I,\qquad M := \lVert\Sigma_{\nu}^{-1/2}F\rVert_2.

Spectral calculus and order reversal under matrix inversion then give

I−(I+ΔτGn+1)−2⪰ΔτGn+1(I+ΔτGn+1)−1⪰γ2Δτ1+γ2M2ΔτG‾n+1.(63)I - (I + \Delta\tau G_{n+1})^{-2} \succeq\Delta\tau G_{n+1}(I + \Delta\tau G_{n+1})^{-1} \succeq\frac{\gamma^2\Delta\tau}{1+\gamma^2M^2\Delta\tau}\overline{G}_{n+1}. \tag*{(63)}

It remains to show that the reference matrix G‾n+1\overline{G}_{n+1} is uniformly lower-bounded over all directions in R\mathcal{R}, in expectation. Fix h∈Rh \in\mathcal{R} and let v:=FTΣν−1/2hv := F^T\Sigma_{\nu}^{-1/2}h. Since P⊥h=hP_{\perp}h = h, (53) gives hTG‾n+1h=vTRnQQTRnvh^T\overline{G}_{n+1}h = v^T R_nQQ^TR_nv. Moreover, independence of the Bernoulli coordinates gives E[RnQQTRn]=(λ2QQT+λ(1−λ)diag⁡(QQT))(1−(1−λ)Nθ)−1\mathbb{E}[R_nQQ^TR_n] = \left(\lambda^2QQ^T + \lambda(1-\lambda)\operatorname{diag}(QQ^T)\right)\left(1-(1-\lambda)^{N_\theta}\right)^{-1}, where the divisor 1−(1−λ)Nθ1-(1-\lambda)^{N_\theta} arises from the condition that Rn≠0R_n \ne0. Combining with the coordinate-coverage condition (39),

hTE[G‾n+1∣Hn]h=E[vTRnQQTRnv∣Hn]=λ2∥QTv∥2+λ(1−λ)∑i=1Nθ(eiTQQTei)vi21−(1−λ)Nθ≥λ(1−λ)ℓQ∥v∥2.(64)\begin{aligned} h^T\mathbb{E}[\overline{G}_{n+1}\mid\mathcal{H}_n]h &= \mathbb{E}[v^TR_nQQ^TR_nv\mid\mathcal{H}_n]\\ &= \frac{\lambda^2\lVert Q^Tv\rVert^2+\lambda(1-\lambda)\sum_{i=1}^{N_\theta}(e_i^TQQ^Te_i)v_i^2}{1-(1-\lambda)^{N_\theta}}\\ &\ge\lambda(1-\lambda)\ell_Q\lVert v\rVert^2. \tag*{(64)} \end{aligned}

To bound ∥v∥\lVert v\rVert from below by ∥h∥\lVert h\rVert, let σF>0\sigma_F > 0 be the smallest positive singular value of Σν−1/2F\Sigma_{\nu}^{-1/2}F. The matrix Σν−1/2FFTΣν−1/2\Sigma_{\nu}^{-1/2}FF^T\Sigma_{\nu}^{-1/2} is positive definite on Range⁡(Σν−1/2F)\operatorname{Range}(\Sigma_{\nu}^{-1/2}F), with smallest eigenvalue σF2\sigma_F^2 on that range. Hence

zTΣν−1/2FFTΣν−1/2z≥σF2∥z∥2,∀z∈Range⁡(Σν−1/2F).(65)z^T\Sigma_{\nu}^{-1/2}FF^T\Sigma_{\nu}^{-1/2}z \ge\sigma_F^2\lVert z\rVert^2,\qquad\forall z\in\operatorname{Range}(\Sigma_{\nu}^{-1/2}F). \tag*{(65)}

Applying this to h∈R⊂Range⁡(Σν−1/2F)h\in\mathcal{R}\subset\operatorname{Range}(\Sigma_{\nu}^{-1/2}F) gives ∥v∥2≥σF2∥h∥2\lVert v\rVert^2 \ge\sigma_F^2\lVert h\rVert^2. Since the preceding quadratic-form estimate (56) holds for every h∈Rh\in\mathcal{R}, we obtain

hTE[G‾n+1∣Hn]h≥λ(1−λ)ℓQσF2∥h∥2∀h∈R.(66)h^T\mathbb{E}[\overline{G}_{n+1}\mid\mathcal{H}_n]h \ge\lambda(1-\lambda)\ell_Q\sigma_F^2\lVert h\rVert^2\qquad\forall h\in\mathcal{R}. \tag*{(66)}

By (50), the squared-error decrease is hnT[I−(I+ΔτGn+1)−2]hnh_n^T[I-(I+\Delta\tau G_{n+1})^{-2}]h_n. Since hn∈Rh_n\in\mathcal{R} is Hn\mathcal{H}_n-measurable, we can apply the preceding matrix bound (55) inside the conditional expectation and then use (58):

E[∥hn+1∥2∣Hn]=∥hn∥2−hnTE[I−(I+ΔτGn+1)−2∣Hn]hn≤∥hn∥2−γ2Δτ1+γ2M2ΔτhnTE[G‾n+1∣Hn]hn≤(1−β)∥hn∥2.(67)\begin{aligned} \mathbb{E}[\lVert h_{n+1}\rVert^2\mid\mathcal{H}_n] &=\lVert h_n\rVert^2-h_n^T\mathbb{E}[I-(I+\Delta\tau G_{n+1})^{-2}\mid\mathcal{H}_n]h_n\\ &\le\lVert h_n\rVert^2-\frac{\gamma^2\Delta\tau}{1+\gamma^2M^2\Delta\tau}h_n^T\mathbb{E}[\overline{G}_{n+1}\mid\mathcal{H}_n]h_n\\ &\le(1-\beta)\lVert h_n\rVert^2. \tag*{(67)} \end{aligned}

where

β:=λ(1−λ)ℓQσF2γ2Δτ1+γ2M2Δτ.(68)\beta:= \lambda(1-\lambda)\ell_Q\sigma_F^2\frac{\gamma^2\Delta\tau}{1+\gamma^2M^2\Delta\tau}. \tag*{(68)}

Taking expectations in (59) and using the tower property gives E∥hn+1∥2≤(1−β)E∥hn∥2\mathbb{E}\lVert h_{n+1}\rVert^2 \le(1-\beta)\mathbb{E}\lVert h_n\rVert^2. Iterating yields E∥hn∥2≤(1−β)n∥h0∥2\mathbb{E}\lVert h_n\rVert^2 \le(1-\beta)^n\lVert h_0\rVert^2, which is (43) by the definition of hnh_n in (47).

For (iv)(iv), the first-order optimality condition at m∗m^{*} gives ΦR(m)−ΦR(m∗)=12∥Σν−1/2F(m−m∗)∥2\Phi_{R}(m)-\Phi_{R}(m^{*})=\frac{1}{2}\lVert\Sigma_{\nu}^{-1/2}F(m-m^{*})\rVert^{2}. Combining this condition with the optimality condition for the restricted minimizer m⋆=m0+QD01/2P1b⋆m_{\star}=m_{0}+QD_{0}^{1/2}P_{1}b_{\star} yields Σν−1/2F(m⋆−m∗)⊥S\Sigma_{\nu}^{-1/2}F(m_{\star}-m^{*})\perp\mathcal{S}. Using (40) and the definition of m⋆m_{\star}, we write

mn−m∗=QD01/2P1(bn−b⋆)+un+(m⋆−m∗).(69)m_{n}-m^{*}=QD_{0}^{1/2}P_{1}(b_{n}-b_{\star})+u_{n}+(m_{\star}-m^{*}). \tag*{(69)}

Since Σν−1/2FQD01/2P1(bn−b⋆)∈S\Sigma_{\nu}^{-1/2}FQD_{0}^{1/2}P_{1}(b_{n}-b_{\star})\in\mathcal{S} and Σν−1/2Fun, Σν−1/2F(m⋆−m∗)∈S⊥\Sigma_{\nu}^{-1/2}Fu_{n},\ \Sigma_{\nu}^{-1/2}F(m_{\star}-m^{*})\in\mathcal{S}^{\perp}, applying P⊥Σν−1/2FP_{\perp}\Sigma_{\nu}^{-1/2}F to (61) and using the definition of hnh_{n} in (47) gives hn=Σν−1/2Fun+Σν−1/2F(m⋆−m∗)h_{n}=\Sigma_{\nu}^{-1/2}Fu_{n}+\Sigma_{\nu}^{-1/2}F(m_{\star}-m^{*}). Multiplying (61) by Σν−1/2F\Sigma_{\nu}^{-1/2}F and replacing the last two terms by hnh_{n} yields

Σν−1/2F(mn−m∗)=Σν−1/2FQD01/2P1(bn−b⋆)+hn.\Sigma_{\nu}^{-1/2}F(m_{n}-m^{*})=\Sigma_{\nu}^{-1/2}FQD_{0}^{1/2}P_{1}(b_{n}-b_{\star})+h_{n}.

The two terms lie in S\mathcal{S} and R\mathcal{R}, respectively. By orthogonality and (29),

ΦR(mn)−ΦR(m∗)=12(bn−b⋆)TΓ(bn−b⋆)+12∥hn∥2.(70)\Phi_{R}(m_{n})-\Phi_{R}(m^{*})=\frac{1}{2}(b_{n}-b_{\star})^{T}\Gamma(b_{n}-b_{\star})+\frac{1}{2}\lVert h_{n}\rVert^{2}. \tag*{(70)}

Taking expectations and applying the bounds for bn−b⋆b_{n}-b_{\star} and hnh_{n} in (42) and (43), respectively, proves (44).

Numerical Study

In this section, we evaluate the proposed inflated projected-dropout ensemble adjustment Kalman inversion (IPD-EAKI) method on three classes of inverse problems1. We compare IPD-EAKI with inflated EAKI (I-EAKI), described in § 3.1, dropout ensemble Kalman inversion (DEKI) [47], and the covariance matrix adaptation evolution strategy (CMA-ES) [60]. I-EAKI provides a baseline without the projected dropout correction, while DEKI provides a comparison with an existing dropout method. CMA-ES is a stochastic, derivative-free optimizer that adapts a Gaussian search distribution.

  • (i) Ill-conditioned linear inverse problems: We vary the condition number of the forward operator and the number of observable directions missing from the initial ensemble subspace. These experiments demonstrate how covariance inflation accelerates convergence within that subspace and how projected dropout recovers the missing observable components.

  • (ii) 100-dimensional nonlinear problems: We consider the monotone cubic, Rosenbrock, and Rastrigin problems, representing smooth nonlinear distortion, curved and ill-conditioned valleys, and highly oscillatory landscapes with multiple local minima, respectively. We assess both optimization performance and the evolution of ensemble spread.

(iii) Barotropic flow inverse problem: We recover the initial vorticity field on the sphere from noisy zonal wind observations at two subsequent times. This example assesses the applicability of IPD-EAKI to a nonlinear PDE-constrained inverse problem.

Throughout the experiments, I-EAKI and IPD-EAKI use a fixed artificial time step Δτ=0.5\Delta\tau= 0.5. IPD-EAKI uses a dropout rate of 1−λ=0.51-\lambda= 0.5. Ensemble sizes and other method-specific settings are specified for each example.

Ill-conditioned linear inverse problems

We consider the noiseless linear inverse problem x=Fθ†x = F\theta^\dagger with Nθ=Nx=100N_\theta= N_x = 100, identity weighting Σν=I\Sigma_\nu= I, and no prior regularization. For each condition number κ∈{1,10,102,103,104}\kappa\in\{1, 10, 10^2, 10^3, 10^4\}, we construct the forward matrix as

F=UFdiag⁡(σ1,…,σ100)VFT,σi=κ−(i−1)/99,(71)F = U_F\operatorname{diag}(\sigma_1,\ldots,\sigma_{100})V_F^T,\qquad\sigma_i = \kappa^{-(i-1)/99}, \tag*{(71)}

where UFU_F and VFV_F are random orthogonal matrices. The prescribed singular values ensure that cond⁡2(F)=κ\operatorname{cond}_2(F) = \kappa. We initialize J=21J = 21 particles in the subspace spanned by the first 20 right singular vectors of FF. Let Qsub=[v1,…,v20]Q_{\mathrm{sub}} = [v_1,\ldots,v_{20}], where viv_i denotes the iith column of VFV_F. We draw independent coefficients c0(j)∼N(0,I20)c_0^{(j)} \sim\mathcal{N}(0,I_{20}) and set θ0(j)=Qsubc0(j)\theta_0^{(j)} = Q_{\mathrm{sub}}c_0^{(j)}. The initial ensemble therefore has zero empirical mean, and its empirical covariance has rank 20 almost surely. All directions in this initial subspace are observable, since FQsubFQ_{\mathrm{sub}} has full column rank. We construct the truth as

θ†=30∑i=120+Nmissingviσi.(72)\theta^\dagger= 30\sum_{i=1}^{20+N_{\mathrm{missing}}}\frac{v_i}{\sigma_i}. \tag*{(72)}

where Nmissing∈{0,2,10}N_{\mathrm{missing}} \in\{0,2,10\} is the number of active singular directions outside the initial ensemble subspace. This scaling gives an observation-space amplitude of 30 in each active left singular direction.

We compare DEKI, I-EAKI, IPD-EAKI, and CMA-ES over 100 iterations and ten random seeds. DEKI uses the same dropout rate and a step-size prefactor of 0.5 for both the mean and ensemble-deviation updates. CMA-ES starts at the initial ensemble mean with step size σ0=1\sigma_0 = 1 and population size 21. Convergence is measured by the relative data misfit misfit⁡n=∥Fmn−x∥2/∥x∥2\operatorname{misfit}_n = \lVert Fm_n-x\rVert_2/\lVert x\rVert_2, where mnm_n is the ensemble mean for the ensemble Kalman methods and the best point found so far for CMA-ES. DEKI and IPD-EAKI require approximately twice as many forward evaluations per iteration as I-EAKI and CMA-ES.

Figure 1 shows the relative data misfit as a function of iteration. For Nmissing=0N_{\mathrm{missing}} = 0, I-EAKI and IPD-EAKI exhibit nearly identical decay to numerical precision, consistent with the asymptotic factor 1−Δτ1-\Delta\tau, which is independent of the condition number κ\kappa; see Theorem 1. For Nmissing>0N_{\mathrm{missing}} > 0, I-EAKI approaches the subspace misfit floor, whereas the projected dropout mean correction enables IPD-EAKI to reduce the misfit further, although this additional reduction slows as κ\kappa increases. Under the same iteration budget, IPD-EAKI achieves lower final misfits than DEKI and CMA-ES, both of which remain sensitive to conditioning.

Figure 1. Condition-number sensitivity for the noiseless linear inverse problem. Columns show DEKI, I-EAKI, IPD-EAKI, and CMA-ES; rows correspond to Nmissing=0,2,10N_{\mathrm{missing}} = 0, 2, 10. Colors indicate κ∈{1,10,102,103,104}\kappa\in\{1, 10, 10^{2}, 10^{3}, 10^{4}\}. Each curve shows the pointwise median relative data misfit over ten random seeds, with shaded regions indicating the interquartile range. The diagnostic uses the ensemble mean for the ensemble Kalman methods and the best point found so far for CMA-ES. All vertical axes are logarithmic, with different limits across rows.

100-dimensional nonlinear problems

We consider three nonlinear least-squares problems in Nθ=100N_{\theta} = 100 dimensions: min⁡Φ(θ)\min\Phi(\theta), Φ(θ)=12∥F(θ)∥22\Phi(\theta) = \frac{1}{2}\lVert\mathcal{F}(\theta)\rVert_{2}^{2}. Each problem has a known global minimizer θ∗\theta^{\ast} satisfying F(θ∗)=0\mathcal{F}(\theta^{\ast}) = 0.

(1) Monotone cubic problem. The forward map is

F(θ)=g(Bθ)−g(Bθ∗)∈RNθ,(73)\mathcal{F}(\theta) = g(B\theta) - g(B\theta^{\ast}) \in\mathbb{R}^{N_{\theta}}, \tag*{(73)}

where g(z)=z+2z3g(z) = z + 2z^{3} is applied componentwise, B∈RNθ×NθB \in\mathbb{R}^{N_{\theta}\times N_{\theta}} is a random orthogonal matrix, and θ∗\theta^{\ast} is randomly generated. Since g′(z)>0g'(z) > 0, this example has a unique global minimizer and tests smooth nonlinear distortion without spurious local minima.

(2) Rosenbrock problem. The classical Rosenbrock objective admits the following residual representation:

F(θ)=[f1(θ),⋯ ,f2Nθ−2(θ)]Tf2i−1(θ)=200(θi+1−θi2),f2i(θ)=2(1−θi),\mathcal{F}(\theta) = [f_1(\theta), \cdots, f_{2N_\theta-2}(\theta)]^T \qquad f_{2i-1}(\theta) = \sqrt{200}(\theta_{i+1}-\theta_i^2), \qquad f_{2i}(\theta) = \sqrt{2}(1-\theta_i),

The unique global minimizer is θ∗=(1,…,1)T\theta^*=(1,\ldots,1)^T. This problem tests convergence in curved, ill-conditioned valleys.

(3) Rastrigin problem. The classical Rastrigin objective admits the residual representation

F(θ)=2[θ20sin⁡(πθ)]∈R2Nθ,\mathcal{F}(\theta)=\sqrt{2}\begin{bmatrix}\theta\\ \sqrt{20}\sin(\pi\theta)\end{bmatrix}\in\mathbb{R}^{2N_\theta},

where the sine function is applied componentwise. The unique global minimizer is θ∗=0\theta^*=0. This problem tests robustness on a highly oscillatory landscape with many local minima.

We compare DEKI, IPD-EAKI, and CMA-ES using the same budget of forward evaluations and ten random seeds. For all experiments, we use J∈{20,50}J\in\{20,50\} particles. Following the standard setup, we sample the initial ensembles from Gaussian distributions N(0,I)\mathcal{N}(0,I) for the monotone cubic and Rosenbrock problems, and uniformly from [−3,5]Nθ[-3,5]^{N_\theta} for the Rastrigin problem.

Figure 2 shows representative two-dimensional objective slices with projected final particle locations, and the best evaluated objective and ensemble covariance norm. The slices are obtained by varying the first two parameter coordinates θ1,θ2\theta_1,\theta_2 while fixing all remaining coordinates at their reference minimizer values θ3∗,θ4∗,…,θNθ∗\theta_3^*,\theta_4^*,\ldots,\theta_{N_\theta}^*. Across all three problems, IPD-EAKI and DEKI initially reduce the objective faster than CMA-ES, but DEKI subsequently stagnates, coinciding with rapid covariance collapse. IPD-EAKI achieves lower median best objective values while maintaining ensemble diversity. On Rosenbrock, IPD-EAKI rapidly reaches objective values comparable to CMA-ES, although both methods subsequently progress slowly, consistent with the difficulty of navigating the narrow, curved valley. On Rastrigin, IPD-EAKI also outperforms CMA-ES, suggesting that sustained ensemble spread may facilitate exploration of the multimodal landscape.

Figure 2. Nonlinear results on the monotone cubic, Rosenbrock, and Rastrigin problems, shown from top to bottom. The first two columns show two-dimensional objective slices with the final particle positions for J=20J = 20 and J=50J = 50, respectively, from one representative run. The third and fourth columns show the best evaluated objective value and covariance norm against forward model evaluations. Colors distinguish IPD-EAKI, DEKI, and CMA-ES; solid and dashed curves correspond to J=20J = 20 and J=50J = 50. Curves show the median over 10 independent random seeds, with shaded regions indicating the interquartile range.

Barotropic flow inverse problem

Finally, we consider the incompressible Euler equations on a rotating sphere, written in vorticity form as

∂tζ=−v⋅∇(ζ+f),∇2ψ=ζ,v=k×∇ψ,(74)\partial_t\zeta=-v\cdot\nabla(\zeta+f), \qquad\nabla^2\psi=\zeta, \qquad v=k\times\nabla\psi, \tag*{(74)}

where ζ\zeta is relative vorticity, ψ\psi is the streamfunction, kk is the radial unit vector, and f=2Ωsin⁡φf=2\Omega\sin\varphi is the Coriolis parameter at latitude φ\varphi. We use Earth’s radius R=6.3712×106 mR=6.3712\times10^6\ \mathrm{m} and rotation rate Ω=7.292×10−5 s−1\Omega=7.292\times10^{-5}\ \mathrm{s}^{-1}. Following an idealized configuration for Northern Hemisphere winter flow in the upper troposphere [61, 62], we construct the reference initial vorticity as ζref=ζb+ζ′\zeta_{\mathrm{ref}} = \zeta_b + \zeta'. Here, ζb\zeta_b is the background vorticity associated with the zonal wind

ub(φ)=25cos⁡φ−30cos⁡3φ+300sin⁡2φcos⁡6φ m s−1,(75)u_b(\varphi) = 25 \cos\varphi- 30 \cos^3 \varphi+ 300 \sin^2 \varphi\cos^6 \varphi\ \mathrm{m}\,\mathrm{s}^{-1}, \tag*{(75)}

and the disturbance is given by ζ′(ℓ,φ)=A2cos⁡φexp⁡[−(φ−φ0φW)2]cos⁡(4ℓ)\zeta'(\ell, \varphi) = \frac{A}{2} \cos\varphi\exp\left[-\left(\frac{\varphi-\varphi_0}{\varphi_W}\right)^2\right]\cos(4\ell), where ℓ\ell is longitude, φ0=45∘N\varphi_0 = 45^\circ\mathrm{N}, φW=15∘\varphi_W = 15^\circ, and A=8.0×10−5 s−1A = 8.0 \times10^{-5}\,\mathrm{s}^{-1}. The initial condition is shown in Figure 3.

Background and initial vorticity, and zonal velocity at two times

Figure 3. The upper panels show the background vorticity (left) and the full initial vorticity (right). The lower panels show the zonal wind at T=12.0T = 12.0 h (left) and T=24.0T = 24.0 h (right). Black dots mark the same 50 observation locations in the Northern Hemisphere.

The forward problem is solved at T85 spectral resolution on a 512×256512 \times256 longitude–latitude grid, using a filtered leapfrog integrator [63] with a 1800 s time step.

For the inverse problem, we recover the initial vorticity field ζ0\zeta_0 from noisy pointwise observations of the zonal wind velocity. Observations are collected at the same 50 randomly selected locations in the Northern Hemisphere at T=12T = 12 h and T=24T = 24 h and perturbed by Gaussian noise with covariance matrix Ση\Sigma_\eta determined by componentwise standard deviations equal to 5% of the absolute values of the corresponding reference data, yielding the observation vector yobsy_{\mathrm{obs}} with Ny=100N_y = 100 entries (see the bottom row of Figure 3). The background vorticity ζb\zeta_b is used as the prior mean of the full initial field. The zero-mean real vorticity field on the sphere can be approximated with Ntrunc=7N_{\mathrm{trunc}} = 7 truncated wavenumber with triangular truncation, as follows,

ζ0(ℓ,φ)=ζb+∑n=1Ntrunca0,nP0,n(φ)+∑m=1Ntrunc∑n=mNtrunc2Pm,n(φ)(am,nccos⁡(mℓ)+am,nssin⁡(mℓ)).(76)\zeta_0(\ell,\varphi) = \zeta_b + \sum_{n=1}^{N_{\mathrm{trunc}}} a_{0,n}P_{0,n}(\varphi) + \sum_{m=1}^{N_{\mathrm{trunc}}} \sum_{n=m}^{N_{\mathrm{trunc}}} 2P_{m,n}(\varphi)\left(a^{c}_{m,n}\cos(m\ell) + a^{s}_{m,n}\sin(m\ell)\right). \tag*{(76)}

Collecting these coefficients gives a∈R63a \in\mathbb{R}^{63}. Solving the (74) gives the forward model mapping G:a→yobs\mathcal{G}: a \to y_{\mathrm{obs}}. The Gaussian prior N(0,Σ0)\mathcal{N}(0,\Sigma_0) with Σ0=9I63\Sigma_0 = 9I_{63} is incorporated through the augmented map

F(a)=[G(a)a],x=[yobs0],Σν=[Ση00Σ0].(77)\mathcal{F}(a) = \begin{bmatrix} \mathcal{G}(a) \\ a \end{bmatrix}, \qquad x = \begin{bmatrix} y_{\mathrm{obs}} \\ 0 \end{bmatrix}, \qquad\Sigma_\nu= \begin{bmatrix} \Sigma_\eta& 0 \\ 0 & \Sigma_0 \end{bmatrix}. \tag*{(77)}

We compare two IPD-EAKI configurations, (J,Niter)=(30,20)(J, N_{\mathrm{iter}}) = (30, 20) and (60,10)(60, 10), using comparable budgets of forward evaluations. Each configuration is run 10 times with different dropout seeds, keeping the problem, observations, and initialization fixed. Figure 4 shows the relative vorticity error and the regularized objective ΦR\Phi_R associated with the augmented formulation (69). Both configurations reduce these quantities, with the 60-member ensemble achieving lower values. This improvement may reflect the larger initial ensemble subspace, which provides more directions for direct EAKI updates, potentially accelerating convergence and reducing reliance on projected dropout for exploration beyond that subspace. The reconstructed initial vorticity fields in Figure 5-top also show that the 60-member ensemble provides a more accurate reconstruction. The final ensemble standard-deviation fields in Figure 5-bottom highlight estimated uncertainty in the Southern Hemisphere, where no direct observations are available. The higher standard deviations obtained with J=60J = 60 may reflect the broader range of parameter directions represented by its ensemble covariance.

Relative vorticity error and regularized objective for two IPD-EAKI configurations

Figure 4. Relative vorticity error ∥ζ0(mn)−ζref∥2∥ζref∥2\frac{\lVert\zeta_0(m_n)-\zeta_{\mathrm{ref}}\rVert_2}{\lVert\zeta_{\mathrm{ref}}\rVert_2} (left) and regularized objective ΦR(mn)\Phi_R(m_n) (right) for the two IPD-EAKI configurations. Each curve shows the mean over 10 runs with different dropout seeds, keeping the problem, observations, and initialization fixed within each configuration. Shaded regions indicate the mean ±\pm sample standard deviation.

Reconstructed initial vorticity fields from the final ensemble means and pointwise standard deviations

Figure 5. Reconstructed initial vorticity fields from the final ensemble means (top) and pointwise standard deviations over the final ensemble members (bottom), using the first of 10 trials for each configuration: (J,Niter)=(30,20)(J, N_{\mathrm{iter}}) = (30, 20) (left) and (60,10)(60, 10) (right). The reference initial vorticity is shown in Figure 3.

Conclusion

We enhance the EKI framework by combining multiplicative covariance inflation with projected dropout to improve exploration and convergence while retaining its derivative-free, parallel structure. Numerical experiments on ill-conditioned linear systems, high-dimensional nonlinear benchmarks, and barotropic flow inversion support the complementary roles of the two mechanisms. Future work will investigate adaptive parameter selection and line-search or trust-region acceptance criteria to further improve robustness in nonlinear problems.

Acknowledgments We acknowledge funding support from National Key R&D Program of China 2025YFA1018700, National Natural Science Foundation of China (No.12471403, 62595771, and 12288101), the Fundamental Research Funds for the Central Universities of China, and the support of the high-performance computing platform of Peking University.

AI disclosure. We used GPT-6 (OpenAI) to assist in developing theoretical arguments for the linear case. We subsequently verified, simplified, and organized these arguments into the proofs of Theorems 1 and 2. GPT-6 was also used to improve the language. The authors take full responsibility for the final content.

References

  1. [1]Yan Chen and Dean S Oliver. Ensemble randomized maximum likelihood method as an iterative ensemble smoother. Mathematical Geosciences, 44(1):1–26, 2012.DOI
  2. [2]Alexandre A Emerick and Albert C Reynolds. Investigation of the sampling performance of ensemble-based methods with a simple reservoir model. Computational Geosciences, 17(2):325–350, 2013.DOI
  3. [3]Marco A Iglesias, Kody JH Law, and Andrew M Stuart. Ensemble Kalman methods for inverse problems. Inverse Problems, 29(4):045001, 2013.DOI
  4. [4]Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. J. Basic Eng. Mar, 82(1):35–45, 1960.DOI
  5. [5]Geir Evensen. Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics. Journal of Geophysical Research: Oceans, 99(C5):10143–10162, 1994.DOI
  6. [6]Claudia Schillings and Andrew M Stuart. Analysis of the ensemble Kalman filter for inverse problems. SIAM Journal on Numerical Analysis, 55(3):1264–1290, 2017.arxiv.org/abs/1602.02020
  7. [7]Claudia Schillings and Andrew M Stuart. Convergence analysis of ensemble Kalman inversion: The linear, noisy case. Applicable Analysis, 97(1):107–123, 2018.arxiv.org/abs/1702.07894
  8. [8]Dirk Blömker, Claudia Schillings, Philipp Wacker, and Simon Weissmann. Well posedness and convergence analysis of the ensemble kalman inversion. Inverse Problems, 35(8):085007, 2019.DOI
  9. [9]Elizabeth Qian and Christopher Beattie. The fundamental subspaces of ensemble kalman inversion. SIAM Review, 67(4):771–798, 2025.arxiv.org/abs/2409.08862
  10. [10]Leon Bungert and Philipp Wacker. Complete deterministic dynamics and spectral decomposition of the linear ensemble kalman inversion. SIAM/ASA Journal on Uncertainty Quantification, 11(1):320–357, 2023.
  11. [11]Daniel J Lea, Myles R Allen, and Thomas WN Haine. Sensitivity analysis of the climate of a chaotic system. Tellus A: Dynamic Meteorology and Oceanography, 52(5):523–532, 2000.
  12. [12]Qiqi Wang, Rui Hu, and Patrick Blonigan. Least squares shadowing sensitivity analysis of chaotic limit cycle oscillations. Journal of Computational Physics, 267:210–224, 2014.
  13. [13]Yaqing Gu and Dean S Oliver. An iterative ensemble kalman filter for multiphase fluid flow data assimilation. Spe Journal, 12(04):438–446, 2007.
  14. [14]Harrie-Jan Hendricks Franssen and W Kinzelbach. Real-time groundwater flow modeling with the ensemble kalman filter: Joint estimation of states and parameters and the filter inbreeding problem. Water resources research, 44(9), 2008.
  15. [15]Sigurd I Aanonsen, Geir Nævdal, Dean S Oliver, Albert C Reynolds, and Brice Vallès. The ensemble kalman filter in reservoir engineering—a review. Spe Journal, 14(03):393–412, 2009.
  16. [16]Han Gao and Jian-Xun Wang. A bi-fidelity ensemble Kalman method for PDE-constrained inverse problems in computational mechanics. Computational Mechanics, 67(4):1115–1131, 2021.
  17. [17]Xin-Lei Zhang, Heng Xiao, Xiaodong Luo, and Guowei He. Ensemble kalman method for learning turbulence models from indirect observation data. Journal of Fluid Mechanics, 949:A26, 2022.
  18. [18]YoungIn Shin, Miles Chan, Jianyu Wang, Tony Zahtila, Catherine Gorle, Gianluca Iaccarino, and Michael Howland. Multi-fidelity modeling and uncertainty quantification of heterogeneous roughness. In Center for Turbulence Research Proceedings of the Summer Program, 2024.
  19. [19]Shunxiang Cao and Daniel Zhengyu Huang. Bayesian calibration for large-scale fluid structure interaction problems under embedded/immersed boundary framework. International Journal for Numerical Methods in Engineering, 123(8):1791–1812, 2022.
  20. [20]Oliver RA Dunbar, Alfredo Garbuno-Inigo, Tapio Schneider, and Andrew M Stuart. Calibration and uncertainty quantification of convective parameters in an idealized gcm. Journal of Advances in Modeling Earth Systems, 13(9):e2020MS002454, 2021.
  21. [21]Daniel Zhengyu Huang, Tapio Schneider, and Andrew M Stuart. Iterated kalman methodology for inverse problems. Journal of Computational Physics, page 111262, 2022.
  22. [22]Oliver RA Dunbar, Andrew B Duncan, Andrew M Stuart, and Marie-Therese Wolfram. Ensemble inference methods for models with noisy and expensive likelihoods. SIAM Journal on Applied Dynamical Systems, 21(2):1539–1572, 2022.
  23. [23]Liang Yan and Tao Zhou. An adaptive multifidelity pc-based ensemble kalman inversion for inverse problems. International Journal for Uncertainty Quantification, 9(3), 2019.
  24. [24]Zhiwei Gao, Liang Yan, and Tao Zhou. Adaptive operator learning for infinite-dimensional bayesian inverse problems. SIAM/ASA Journal on Uncertainty Quantification, 12(4):1389–1423, 2024.
  25. [25]Michael K Tippett, Jeffrey L Anderson, Craig H Bishop, Thomas M Hamill, and Jeffrey S Whitaker. Ensemble square root filters. Monthly Weather Review, 131(7):1485–1490, 2003.
  26. [26]Daniel Zhengyu Huang, Jiaoyang Huang, Sebastian Reich, and Andrew M Stuart. Efficient derivative-free bayesian inference for large-scale inverse problems. Inverse Problems, 38(12):125006, 2022.
  27. [27]Neil Chada and Xin Tong. Convergence acceleration of ensemble kalman inversion in nonlinear settings. Mathematics of Computation, 91(335):1247–1280, 2022.
  28. [28]Jeffrey L Anderson. An ensemble adjustment Kalman filter for data assimilation. Monthly weather review, 129(12):2884–2903, 2001.
  29. [29]Craig H Bishop, Brian J Etherton, and Sharanya J Majumdar. Adaptive sampling with the ensemble transform Kalman filter. Part I: Theoretical aspects. Monthly weather review, 129(3):420–436, 2001.
  30. [30]Alexandre A Emerick and Albert C Reynolds. Ensemble smoother with multiple data assimilation. Computers & Geosciences, 55:3–15, 2013.DOI
  31. [31]Zhiyan Ding and Qin Li. Ensemble kalman inversion: mean-field limit and convergence analysis. Statistics and computing, 31(1):9, 2021.DOI
  32. [32]Marco Iglesias and Yuchen Yang. Adaptive regularisation for ensemble Kalman inversion. Inverse Problems, 37(2):025008, 2021.arxiv.org/abs/2006.14980
  33. [33]Sebastian Reich. A nonparametric ensemble transform method for Bayesian inference. SIAM Journal on Scientific Computing, 35(4):A2013–A2024, 2013.DOI
  34. [34]Kody Law, Andrew Stuart, and Kostas Zygalakis. Data assimilation. Cham, Switzerland: Springer, 2015.arxiv.org/abs/1506.07825
  35. [35]Youssef Marzouk, Tarek Moselhy, Matthew Parno, and Alessio Spantini. Sampling via measure transport: An introduction. Springer Books, pages 785–825, 2017.
  36. [36]Sangeetika Ruchi, Svetlana Dubinkina, and Marco A Iglesias. Transform-based particle filtering for elliptic bayesian inverse problems. Inverse Problems, 35(11):115005, 2019.DOI
  37. [37]Jakiw Pistrigach and Sebastian Reich. Affine-invariant ensemble transform methods for logistic regression. Foundations of Computational Mathematics, 23(2):675–708, 2023.
  38. [38]Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential monte carlo samplers. Journal of the Royal Statistical Society Series B: Statistical Methodology, 68(3):411–436, 2006.
  39. [39]Arnaud Doucet, Nando de Freitas, and Neil Gordon, editors. Sequential Monte Carlo Methods in Practice. Springer, New York, NY, 2001.
  40. [40]Ricardo Baptista, Youssef Marzouk, and Olivier Zahm. On the representation and learning of monotone triangular transport maps. Foundations of Computational Mathematics, 24(6):2063–2108, 2024.
  41. [41]Nicolas Chopin, Francesca Crucinio, and Anna Korba. A connection between tempering and entropic mirror descent. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 8782–8800. PMLR, 2024.
  42. [42]Yifan Chen, Daniel Zhengyu Huang, Jiaoyang Huang, Sebastian Reich, and Andrew M Stuart. Sampling via gradient flows in the space of probability measures. Mathematics of Computation, 2026.arxiv.org/abs/2310.03597
  43. [43]Yifan Chen, Daniel Zhengyu Huang, Jiaoyang Huang, Sebastian Reich, and Andrew M Stuart. Efficient, multimodal, and derivative-free bayesian inference with fisher–rao gradient flows. Inverse Problems, 40(12):125001, 2024.arxiv.org/abs/2406.17263
  44. [44]Baojun Che, Yifan Chen, Zhenghao Huan, Daniel Zhengyu Huang, and Weijie Wang. Stable derivative free gaussian mixture variational inference for bayesian inverse problems. SIAM Journal on Scientific Computing, 47(5):A2583–A2608, 2025.arxiv.org/abs/2501.04259
  45. [45]Baojun Che, Yifan Chen, Daniel Zhengyu Huang, Xinying Mao, and Weijie Wang. Adaptive exponential integration for stable gaussian mixture black-box variational inference. arXiv preprint arXiv:2601.14855, 2026.arxiv.org/abs/2601.14855
  46. [46]Huchen Yang and Jin-Long Wu. Polarized ensemble kalman inversion: A derivative-free method for multimodal inverse problems with non-asymptotic analysis. arXiv preprint arXiv:2610.31907, 2026.arxiv.org/abs/2610.31907
  47. [47]Shuigen Liu, Sebastian Reich, and Xin T Tong. Dropout ensemble kalman inversion for high dimensional inverse problems. SIAM Journal on Numerical Analysis, 63(2):685–715, 2025.DOI
  48. [48]Nikola B Kovachki and Andrew M Stuart. Ensemble Kalman inversion: A derivative-free technique for machine learning tasks. Inverse Problems, 35(9):095005, 2019.DOI
  49. [49]Sydney Vernon, Eviatar Bach, and Oliver R. A. Dunbar. Nesterov acceleration for ensemble Kalman inversion and variants. Journal of Computational Physics, 535:114063, 2025.DOI
  50. [50]Simon Weissmann. Gradient flow structure and convergence analysis of the ensemble kalman inversion for nonlinear forward models. Inverse Problems, 38(10):105011, 2022.
  51. [51]Dieter Armbruster, Michael Herty, and Giuseppe Visconti. A stabilization of a continuous limit of the ensemble kalman inversion. SIAM Journal on Numerical Analysis, 60(3):1494–1515, 2022.arxiv.org/abs/2006.15390
  52. [52]Fabian Parzer and Otmar Scherzer. On convergence rates of adaptive ensemble kalman inversion for linear ill-posed problems. Numerische Mathematik, 152(2):371–409, 2022.DOI
  53. [53]Xin T Tong and Matthias Morzfeld. Localized ensemble kalman inversion. Inverse Problems, 39(6):064002, 2023.DOI
  54. [54]Omar Al-Ghattas and Daniel Sanz-Alonso. Non-asymptotic analysis of ensemble kalman updates: effective dimension and localization. Information and Inference: A Journal of the IMA, 13(1):iaad043, 2024.DOI
  55. [55]Andrew M Stuart. Inverse problems: A Bayesian perspective. Acta numerica, 19:451–559, 2010.DOI
  56. [56]Jari Kaipio and Erkki Somersalo. Statistical and computational inverse problems, volume 160. Springer Science & Business Media, 2006.
  57. [57]Masoumeh Dashti and Andrew M Stuart. The Bayesian approach to inverse problems. In Roger Ghanem, David Higdon, and Houman Owadi, editors, Handbook of Uncertainty Quantification, pages 311–428. Springer, Cham, 2017.
  58. [58]Masoumeh Dashti, Kody JH Law, Andrew M Stuart, and Jochen Voss. MAP estimators and their consistency in Bayesian nonparametric inverse problems. Inverse Problems, 29(9):095017, 2013.
  59. [59]Neil K Chada, Andrew M Stuart, and Xin T Tong. Tikhonov regularization within ensemble Kalman inversion. SIAM Journal on Numerical Analysis, 58(2):1263–1294, 2020.arxiv.org/abs/1901.10382
  60. [60]Nikolaus Hansen, Sibylle D Müller, and Petros Koumoutsakos. Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (cma-es). Evolutionary computation, 11(1):1–18, 2003.pubmed.ncbi.nlm.nih.gov/12804094
  61. [61]Isaac M. Held. Pseudomomentum and the orthogonality of modes in shear flows. Journal of Atmospheric Sciences, 42(21):2280 – 2288, 1985.doi.org/10.1175/1520-0469(1985)042<2280:patoom>2.0.co;2
  62. [62]Isaac M. Held and Peter J. Phillips. Linear and nonlinear barotropic decay on the sphere. Journal of Atmospheric Sciences, 44(1):200 – 207, 1987.doi.org/10.1175/1520-0469(1987)044<0200:lanbdo>2.0.co;2
  63. [63]Isaac M Held and Max J Suarez. A proposal for the intercomparison of the dynamical cores of atmospheric general circulation models. Bulletin of the American Meteorological Society, 75(10):1825–1830, 1994.doi.org/10.1175/1520-0477(1994)075<1825:apftio>2.0.co;2

Paper details

vlnc:W_35vysxtp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents