betaFeedback

WorksPhysics and Astronomy

Analysis increments for model bias diagnosis and correction: Theoretical foundations

Abstract

A common method for diagnosing and correcting model bias in weather and climate prediction is based on analysis increments: the difference between the analysis and the forecast at each data assimilation cycle. Their time average is used to estimate and correct the bias, and machine learning is increasingly used to map the forecast state to the increments, giving a state-dependent correction. Despite its widespread use, the method lacks theoretical foundations. We first analyse a Kalman filter whose forecast model has a constant bias. We show what the averaged increment measures, and prove that iterating the correction removes all of the bias seen through the observations; for linear dynamics, no method can identify the rest. We then extend the analysis to state-dependent biases. For linear dynamics, each increment records the model error over the filter's memory, as seen through the observations. We use this to determine what the increments measure when averaged over a long period, averaged over periods with similar slowly varying conditions, and used without averaging, as in machine-learned corrections. Numerical experiments confirm the theory for a linear model and suggest that its main conclusions carry over to the nonlinear Lorenz '96 model, including for a correction learned by a neural network.

Introduction

Data assimilation (DA) combines model forecasts with observations to produce state estimates of dynamical systems (Kalnay, 2002; Sanz-Alonso et al., 2023). A central assumption of most DA methods is that model errors, if present, have zero mean and are uncorrelated in time. The Kalman filter is optimal under this assumption [33], and ensemble Kalman filters are built on it [21]. In practice, however, forecast models are invariably imperfect, and their errors often have a systematic component, a model bias, which in general depends on the state of the system [30, 16].

The presence of model bias degrades the quality of the analysis and can cause filters to diverge in severe cases. Considerable effort has been devoted to detecting and correcting model bias in DA systems [25, 17, 16]. A key diagnostic tool is the analysis increment: the difference between the analysis and the forecast at each cycle. Persistent nonzero mean increments are a clear symptom of model or observation bias [16], and time-averaged increments are a standard diagnostic of model error in operational numerical weather prediction.

The idea behind using analysis increments as a diagnostic tool is the following. If the observations are unbiased, the analysis pulls the forecast towards them, so the analysis increment should point against the model bias; see Figure 1 for an illustration.

Figure 1. A schematic showing analysis increments in the presence of a systematic model bias. In this case, the forecast model is biased upwards, and the analysis increments point in the opposite direction. Note that the analysis still retains some of the bias. Error bars on the observations show the observation error standard deviation.

Once diagnosed, the estimated bias can be subtracted from the forecast to correct it. Saha (1992) corrected model bias by adding the mean one-day forecast error, relative to the verifying analyses, as a constant forcing in an operational forecast model. Dee and da Silva [17]; Dee [16] analysed forecast bias in cycled data assimilation mathematically, and proposed estimating the bias adaptively during the assimilation. Danforth et al. [15] estimated the mean, periodic, and linear state-dependent components of model error from forecast errors relative to a reanalysis, and used them for correction. Bhargava et al. [8] used the analysis increments of an operational system to diagnose its mean and periodic model biases. Li et al. (2009); Mitchell and Carrassi (2015) considered correcting the bias estimated from analysis increments in conjunction with accounting for residual random errors, through additive noise or an estimated model error covariance, and Carrassi and Vannitsem [12] estimated the model error covariance itself from analysis increments.

More recently, machine learning models have been trained to predict analysis increments from the model state, giving a nonlinear, state-dependent correction to the forecast model ([10]; [13]; [1]; [28]; [22, 23]). More broadly, the intersection of DA and machine learning is a rapidly growing field, reviewed by [7]; [4]; [3]; Levine and Stuart (2022) give a general framework for learning model error from data.

The popularity of the analysis increment method can be explained by the fact that it is an offline one: the bias (whether constant or state-dependent) can be estimated from a time series of analysis increments of an existing assimilation system, such as an operational archive or a reanalysis, without modifying or rerunning that system. Online alternatives instead estimate the bias during the assimilation, which requires changing the assimilation system and, for some methods, running it repeatedly; we relate these methods to the increment method in § 2.6 and § 2.7. The distinction matters particularly for learned corrections. Offline, the machine learning model can be trained by ordinary regression on a fixed data set of states and increments; online, training it within the assimilation requires either differentiating through the assimilation ([14]) or estimating its parameters jointly with the state ([27]), which precludes standard gradient-based training and is difficult for models with many parameters.

Despite the widespread practical use of analysis increments as a bias diagnostic and correction method, what the increments measure, and what corrections based on them achieve, have not been established mathematically. Even for a constant bias, the relationship between the time-averaged analysis increment and the asymptotic forecast bias does not appear to have been stated explicitly in the data assimilation literature. How state-dependent biases appear in the increments, and how repeated corrections behave, do not appear to have been analysed either. The method is often justified by the analyses being a proxy for the true state (e.g., [19]; [15]; Rodwell and Palmer, 2007), despite their being contaminated by the bias themselves. Dee [16] derives how a forecast bias is transferred to the analysis in a single analysis step, and Dee and da Silva [17] give the recursion for the forecast bias over many cycles of a filter that ignores the bias, but do not relate it to the averaged increment. In the control literature, Lee and Athans (1979) derived the steady-state mean error of a Kalman filter with a mismatched model, but did not use it to interpret the filter’s increments. The present work fills these gaps.

Our main contributions are:

  1. For a constant additive bias, we derive the fixed point of the asymptotic forecast bias (Proposition 1), and show that the time-averaged analysis increment converges to minus the Kalman gain applied to the innovation bias (9).

  1. We prove that correcting the forecast model with the averaged increment reduces the innovation bias in the norm defined by the innovation covariance (Proposition 2), and that iterating the correction drives the innovation bias to zero geometrically (Proposition 3).

  1. We show that the iterated correction removes all of the bias that is identifiable from the observations: what remains is equivalent to a constant offset of the state in unobserved directions, which, for linear dynamics, no method can detect from the observations alone, including augmented-state filters (Corollary 1).

  1. For state-dependent biases, we show what the increments measure at three levels of averaging. For the averaged increment, we give the exact fixed point for linear biases, and approximate fixed points for nonlinear biases. We then study how increments averaged by a slowly varying state diagnose the slowly varying part of the model error. Finally, we consider individual increments, and what learned corrections regressed on them capture.

  1. We test the theory numerically with a Kalman filter and an ensemble Kalman filter, applied to both linear and nonlinear models, and constant and state-dependent biases.

The paper is organised as follows. Section 2 analyses the constant bias case, which admits a complete analytical treatment and provides the foundation and baseline for what follows. Section 3 extends the analysis to state-dependent biases, considering in turn the averaged increments, increments averaged by a slowly varying state, and learned corrections on individual increments. Section 4 presents numerical experiments. Section 5 concludes.

Constant bias

Setup

We follow the notation of Ide et al. [32] throughout, in which superscripts tt, ff, and aa denote the true, forecast, and analysis states respectively.

We consider a linear state-space model. The true state evolves as

xi+1t=Mxit+ηi,(1)x^{t}_{i+1} = Mx^{t}_{i} + \eta_i, \tag*{(1)}

where M∈Rn×nM \in\mathbb{R}^{n \times n} is the model and ηi∼N(0,Q)\eta_i \sim\mathcal{N}(0,Q) is zero-mean model noise. Observations are given by

yi=Hxit+ϵi,ϵi∼N(0,R),(2)y_i = Hx^{t}_i + \epsilon_i,\qquad\epsilon_i \sim\mathcal{N}(0,R), \tag*{(2)}

with H∈Rp×nH \in\mathbb{R}^{p \times n} and RR positive definite. Throughout this paper we assume HH is time-invariant.

The state is estimated with a Kalman filter at steady state. Its forecast model is imperfect, introducing a constant bias b∈Rnb \in\mathbb{R}^n at each forecast step, which represents systematic model error. The forecast and analysis steps are

xi+1f=Mxia+b,(3)x^{f}_{i+1} = Mx^{a}_i + b, \tag*{(3)}
xia=xif+K(yi−Hxif).(4)x^{a}_i = x^{f}_i + K(y_i - Hx^{f}_i). \tag*{(4)}

where xifx^f_i and xiax^a_i are the means of the filter’s forecast and analysis distributions. The Kalman gain is

K:=PfHT(HPfHT+R)−1,K := P^fH^{T}(HP^fH^{T}+R)^{-1},

where PfP^f is the steady-state forecast error covariance.

The filter is unaware of the bias. Its forecast error covariance satisfies the Riccati equation (4) below, which does not involve bb, so PfP^f and KK are those of the filter with a perfect model. The analysis error covariance is Pa:=(In−KH)PfP^a := (I_n - KH)P^f, where InI_n is the n×nn \times n identity matrix, and

Pf=MPaMT+Q=M(Pf−PfHT(HPfHT+R)−1HPf)MT+Q.(5)P^f = MP^aM^{T} + Q = M(P^f - P^fH^{T}(HP^fH^{T}+R)^{-1}HP^f)M^{T} + Q. \tag*{(5)}

Throughout, we assume that (M,H)(M,H) is detectable and (M,Q1/2)(M,Q^{1/2}) is stabilisable (the latter holds, for instance, if QQ is positive definite). Then (4) has a unique positive semidefinite solution, the filter’s covariance converges to it from any initial covariance, and the steady-state filter is stable: the matrix Φ:=M(In−KH)\Phi:= M(I_n - KH) satisfies ρ(Φ)<1\rho(\Phi) < 1, where ρ(Φ)\rho(\Phi) denotes the spectral radius of $\Phi {citet}[2 Section 4.4]. Since the gain converges to its steady-state value from any initial covariance, assuming the filter to be at steady state, as above, does not affect the asymptotic results that follow. With its gain constant, the filter is the same as cycled three-dimensional variational assimilation (3D-Var) with a static background error covariance equal to $P^f$ ({citet}Kalnay (2002, Sections 5.4.1 and 5.5.1)`). We also assume that PfP^f is positive definite. This holds if QQ is positive definite, since Pf=MPaMT+QP^f = MP^aM^T + Q by (4).

Throughout, E\mathbb{E} denotes expectation over realisations of the model noise, the observation noise, and the initial condition. The filter’s estimates xifx_i^f and xiax_i^a are computed from the observations and are therefore random variables in this sense.

Asymptotic forecast bias

Define the forecast bias at step ii as bif:=E[xif−xit]b_i^f := \mathbb{E}[x_i^f - x_i^t]. Taking expectations of the analysis step (3b) gives the analysis bias:

bia:=E[xia−xit]=bif−KHbif=(In−KH)bif.(6)b_i^a := \mathbb{E}[x_i^a - x_i^t] = b_i^f - KHb_i^f = (I_n - KH)b_i^f. \tag*{(6)}

This equation was also derived in Dee and da Silva [17]; Dee [16].

The analysis bias is thus the forecast bias multiplied by (In−KH)(I_n-KH). Propagating through the model to the next forecast:

bi+1f=E[xi+1f−xi+1t]=Mbia+b=M(In−KH)bif+b.b_{i+1}^{f}=\mathbb{E}[x_{i+1}^{f}-x_{i+1}^{t}]=Mb_i^{a}+b=M(I_n-KH)b_i^{f}+b.

The bias therefore satisfies the linear recurrence relation

bi+1f=Φbif+b,Φ=M(In−KH).(7)b_{i+1}^{f}=\Phi b_i^{f}+b,\qquad\Phi=M(I_n-KH). \tag*{(7)}

We can now obtain the fixed point of this recurrence relation.

Proposition 1 (Asymptotic forecast bias). Under the assumptions of Section 2.1, the forecast bias converges to the fixed point

b∗=(In−Φ)−1b=(In−M(In−KH))−1b,(8)b^*=(I_n-\Phi)^{-1}b=(I_n-M(I_n-KH))^{-1}b, \tag*{(8)}

and the analysis bias converges to (In−KH)b∗(I_n-KH)b^*.

Proof. Since ρ(Φ)<1\rho(\Phi)<1, the recurrence relation (6) converges to the unique fixed point b∗b^* satisfying b∗=Φb∗+bb^*=\Phi b^*+b, i.e., (In−Φ)b∗=b(I_n-\Phi)b^*=b (see, e.g., Theorem 4.13 in Elaydi [20]). Since ρ(Φ)<1\rho(\Phi)<1 implies 1 is not an eigenvalue of Φ\Phi, the matrix (In−Φ)(I_n-\Phi) is invertible and b∗=(In−Φ)−1bb^*=(I_n-\Phi)^{-1}b. The analysis bias follows from (5). □

Dee and da Silva [17] give the recursion (6), for a time-varying gain, in their eq. (79). Their eq. (87), the condition on the per-step bias under which the forecast bias is constant, is (7) rearranged, although they do not use it to obtain the asymptotic forecast bias or show that the forecast bias converges to it. Lee and Athans (1979, eq. (29)) derive an equivalent expression for the steady-state mean error of a mismatched steady-state Kalman filter, assuming that both the true and the model dynamics are stable; here the assumptions of Section 2.1 suffice, and MM may be unstable.

The time-averaged analysis increment

The analysis increment at step ii is xia−xif=Kdix_i^a-x_i^f=K d_i, where di:=yi−Hxifd_i:=y_i-Hx_i^f is the innovation. Taking expectations:

E[xia−xif]=KE[di]=KHE[xit−xif]=−KHbif.\mathbb{E}[x_i^a-x_i^f]=K\mathbb{E}[d_i]=KH\mathbb{E}[x_i^t-x_i^f]=-KHb_i^f.

Up to sign, the asymptotic mean of the innovations is therefore Hb∗Hb^*, which we call the innovation bias.

We now connect the expected analysis increment to the time-averaged one. Define the average of the analysis increments from time 1 to TT as

[xa−xf]T:=1T∑i=1T(xia−xif),[x^a-x^f]_T:=\frac{1}{T}\sum_{i=1}^{T}(x_i^a-x_i^f),

and write xa−xf:=lim⁡T→∞[xa−xf]Tx^a-x^f:=\lim_{T\to\infty}[x^a-x^f]_T. Write each increment as a deterministic part plus a fluctuation,

xia−xif=−KHbif+νi,νi:=K(di−E[di]).(9)x_i^a-x_i^f=-KHb_i^f+\nu_i,\qquad\nu_i:=K(d_i-\mathbb{E}[d_i]). \tag*{(9)}

Since the dynamics are linear, the bias only shifts the mean of the innovations. The centred innovations di−E[di]d_i-\mathbb{E}[d_i] are those the steady-state filter would produce with a perfect model, which, at stationarity, are independent and identically distributed as N(0,S)\mathcal{N}(0,S), with S=HPfHT+RS=HP^fH^{T}+R (Anderson and Moore [2]). The time average of the fluctuations νi\nu_i therefore tends to zero, by the law of large numbers. Since bif→b∗b_i^f\to b^*, the time average of the deterministic part −KHbif-KHb_i^f tends to −KHb∗-KHb^*. Hence

xa−xf=−KHb∗.(10)x^a-x^f=-KHb^*. \tag*{(10)}

For a single analysis, Dee and da Silva [17] noted that the mean increment underestimates the forecast bias, since the gain coefficients are generally less than one. Equation (9) makes this precise for the cycled filter, in which the forecast bias itself has accumulated to b∗b^*.

The averaged increment does not see all of the bias. The matrix KHKH has the same null space as HH: if KHz=0KHz=0 then, since PfP^f is invertible, (Hz)TS−1(Hz)=0(Hz)^TS^{-1}(Hz)=0, so Hz=0Hz=0. By (9), only the component of b∗b^* in the row space of HH is therefore detectable from the averaged analysis increment. The component in the null space ker⁡(H)={z:Hz=0}\ker(H)=\{z:Hz=0\}, that is, in state directions not seen by the observations, is invisible to it; such directions exist whenever rank⁡(H)<n\operatorname{rank}(H)<n, in particular when p<np<n. This does not mean that unobserved variables are left uncorrected. The averaged increment −KHb∗-KHb^* is generally nonzero in unobserved variables, because the gain spreads the observed bias into them through the forecast-error covariances, just as the analysis update unobserved state variables.

Remark 1 (Rate of convergence of the time average). Since the fluctuations νi\nu_i are, at stationarity, independent with covariance KSKTKSK^T, their time average over TT cycles has covariance KSKT/TKSK^T/T, so the error in estimating −KHb∗-KHb^* from TT cycles is O(T−1/2)O(T^{-1/2}). In addition, the initial transient bif−b∗=Φi(b0f−b∗)b_i^f-b^*=\Phi^i(b_0^f-b^*) decays geometrically, since ρ(Φ)<1\rho(\Phi)<1, and so biases the average by O(1/T)O(1/T); the decay is slow when ρ(Φ)\rho(\Phi) is close to one. For suboptimal filters, the innovations are serially correlated, and the asymptotic covariance will differ.

Remark 2 (Unbiasedness of observations). A critical assumption of the above analysis is that the observations are unbiased, i.e., E[ϵi]=0\mathbb{E}[\epsilon_i]=0. If the observations have a systematic bias bo=E[ϵi]≠0b^o=\mathbb{E}[\epsilon_i]\ne0, the innovation satisfies

E[di]=E[yi−Hxif]=HE[xit−xif]+bo=−Hbif+bo,\mathbb{E}[d_i]=\mathbb{E}[y_i-Hx_i^f]=H\mathbb{E}[x_i^t-x_i^f]+b^o=-Hb_i^f+b^o,

and the averaged analysis increment becomes

xa−xf=−KHb∗+Kbo.x^a-x^f=-KHb^*+Kb^o.

with b∗b^{*} now the modified fixed point given below, conflating the forecast bias b∗b^{*} with the observation bias bob^{o} projected through KK. The fixed-point equation (8) is modified to

b∗=(In−Φ)−1(b+MKbo),(11)b^{*} = (I_n-\Phi)^{-1}(b+MKb^{o}), \tag*{(11)}

so that observation bias acts as an additional effective per-step model bias MKboMKb^{o}, since the analysis bias (In−KH)bif+Kbo(I_n-KH)b_i^{f}+Kb^{o} is propagated by MM to the next forecast. In this case the averaged increment correction does not distinguish between model and observation bias, and applying it to the forecast corrects the combined effect rather than the model bias alone. Fowler [24] studies how observation bias affects the performance of bias correction methods.

Bias correction using the averaged analysis increment

A natural bias correction scheme is to use the time-averaged analysis increment as an estimate of the bias. Since the increment points against the bias, the estimate is, by (10),

b^:=−x‾a−x‾f=KHb∗,\hat{b}:=-\overline{x}^{a}-\overline{x}^{f}=KHb^{*},

and the forecast is corrected by subtracting it at each cycle:

x~i+1f=Mxia+b−b^=Mxia+b−KHb∗.(12)\tilde{x}_{i+1}^{f}=Mx_i^{a}+b-\hat{b}=Mx_i^{a}+b-KHb^{*}. \tag*{(12)}

Here bb is the actual per-step bias of the forecast model and b^\hat{b} its estimate. Using (8), the effective per-step bias after correction is:

b~=b−b^=(In−KH(In−Φ)−1)b.\tilde{b}=b-\hat{b}=\left(I_n-KH(I_n-\Phi)^{-1}\right)b.

The new asymptotic bias satisfies b~∗=(In−Φ)−1b~\tilde{b}^{*}=(I_n-\Phi)^{-1}\tilde{b}, giving:

b~∗=(In−(In−Φ)−1KH)b∗.(13)\tilde{b}^{*}=\left(I_n-(I_n-\Phi)^{-1}KH\right)b^{*}. \tag*{(13)}

Is the asymptotic bias improved after correction? The following proposition shows that the correction does in fact reduce the innovation bias in the Mahalanobis norm with respect to the innovation covariance.

Proposition 2 (Mahalanobis norm reduction). Under the assumptions of Section 2.1,

∥Hb~∗∥S≤∥Hb∗∥S,\lVert H\tilde{b}^{*}\rVert_{S}\leq\lVert Hb^{*}\rVert_{S},

where ∥v∥S2:=vTS−1v\lVert v\rVert_{S}^{2}:=v^{T}S^{-1}v is the Mahalanobis norm with respect to the innovation covariance S=HPfHT+RS=HP^{f}H^{T}+R. The inequality is strict whenever Hb∗≠0Hb^{*}\ne0.

Proof. In Appendix A. □

Iterated correction

The correction (12) can be applied iteratively as follows: at iteration jj, run the filter to convergence, compute the time-averaged analysis increment, subtract the corresponding bias estimate from the already-corrected forecast model, and repeat. Unlike a single correction, this is no longer offline, since each iteration requires a new assimilation run. To our knowledge, this procedure has only been applied empirically in the climate modelling literature (Krinner et al. (2026)), and, with learned state-dependent corrections, by Gregory et al. [29] (see § 3.5).

At iteration jj, the filter is run with the forecast model

xi+1f,(j)=Mxia,(j)+b(j),b(0)=b,x_{i+1}^{f,(j)}=Mx_i^{a,(j)}+b^{(j)},\qquad b^{(0)}=b,

and the per-step bias is updated by subtracting the bias estimate of that run, b^(j):=−x‾a,(j)−x‾f,(j)=KHb∗(j)\hat{b}^{(j)}:=-\overline{x}^{a,(j)}-\overline{x}^{f,(j)}=KHb^{*(j)}:

b(j+1)=b(j)−b^(j)=b(j)−KHb∗(j),b^{(j+1)}=b^{(j)}-\hat{b}^{(j)}=b^{(j)}-KHb^{*(j)},

where b∗(j)=(In−Φ)−1b(j)b^{*(j)}=(I_n-\Phi)^{-1}b^{(j)} is the asymptotic forecast bias of iteration jj. Equivalently, b(j)=b−∑k=0j−1b^(k)b^{(j)}=b-\sum_{k=0}^{j-1}\hat{b}^{(k)}: the forecast model at iteration jj subtracts the sum of all previous bias estimates. The new asymptotic bias satisfies:

b∗(j+1)=(In−Φ)−1b(j+1)=(In−(In−Φ)−1KH)b∗(j).(14)b^{*(j+1)}=(I_n-\Phi)^{-1}b^{(j+1)}=\left(I_n-(I_n-\Phi)^{-1}KH\right)b^{*(j)}. \tag*{(14)}

Unrolling the recurrence (14) from b∗(0)=b∗b^{*(0)}=b^{*}:

b∗(j)=(In−(In−Φ)−1KH)jb∗,b^{*(j)}=\left(I_n-(I_n-\Phi)^{-1}KH\right)^{j}b^{*},

which does not in general converge to zero. The matrix (In−Φ)−1KH(I_n-\Phi)^{-1}KH vanishes on the null space of HH, so In−(In−Φ)−1KHI_n-(I_n-\Phi)^{-1}KH leaves any b∗∈ker⁡(H)b^{*}\in\ker(H) unchanged: as noted in § 2.3, bias that is invisible in the observations is not removed, no matter how many iterations are performed.

The innovation bias Hb∗(j)Hb^{*(j)} obeys a closed recurrence. Since H(In−Φ)−1KH=CHH(I_n-\Phi)^{-1}KH=CH with

C:=H(In−Φ)−1K∈Rp×p,(15)C:=H(I_n-\Phi)^{-1}K\in\mathbb{R}^{p\times p}, \tag*{(15)}

applying HH to (14) gives Hb∗(j+1)=(Ip−C)Hb∗(j)Hb^{*(j+1)}=(I_p-C)Hb^{*(j)}, and hence

Hb∗(j)=(Ip−C)jHb∗.(16)Hb^{*(j)}=(I_p-C)^{j}Hb^{*}. \tag*{(16)}

The following proposition characterises the convergence of the iterated correction in observation space.

Proposition 3 (Convergence of the iterated correction). *Under the assumptions of Section 2.1, the iterated correction drives the innovation bias to zero geometrically: (Ip−C)jv→0(I_p-C)^jv\to0 for every vv of the form HzHz, and in particular ∥Hb∗(j)∥2→0\lVert Hb^{*(j)}\rVert_{2}\to0 as j→∞j\to\infty. If HH has full row rank (rank⁡(H)=p(\operatorname{rank}(H)=p, that is, no observation is a linear combination of the others), the rate is

ρ(Ip−C)<1\rho(I_p-C)<1.

Proof. In Appendix A.

Remark 3 (Single-step norm reduction and transient growth). Proposition 2 does not extend to the Euclidean norm: a single correction can increase ∥Hb∗∥2\lVert H b^*\rVert_2. When HH has full row rank, a single-step reduction for every b∗b^* is equivalent to σmax⁡(Ip−C)≤1\sigma_{\max}(I_p-C)\leq1, and this fails for some systems in which CC is sufficiently non-normal. The iterates (Ip−C)jHb∗(I_p-C)^j H b^* still converge to zero by Proposition 3, but not necessarily monotonically, a phenomenon known as transient growth (Trefethen and Embree, 2005).

Proposition 3 also implies that CC is invertible on vectors of the form HzH z, and we write C−1C^{-1} for this inverse.1 It is the usual inverse when HH has full row rank.

Corollary 1 (Limit of the iterated correction). Under the assumptions of § 2.1, the per-step bias of the iterated correction converges to

b(∞)=b−KC−1Hb∗,(17)b^{(\infty)}=b-KC^{-1}Hb^*, \tag*{(17)}

and the asymptotic bias converges to Πb∗\Pi b^*, where

Π:=In−(In−Φ)−1KC−1H.(18)\Pi:=I_n-(I_n-\Phi)^{-1}KC^{-1}H. \tag*{(18)}

The limit Πb∗\Pi b^* is invisible to the observations (HΠb∗=0H\Pi b^*=0), and the remaining per-step bias is of the form b(∞)=(In−M)zb^{(\infty)}=(I_n-M)z with Hz=0H z=0.

Proof. In Appendix A.

What cannot be identified

What does the remaining bias of Corollary 1 mean? Take the true trajectory and shift every state by a constant vector zz with Hz=0H z=0, as in Corollary 1: xi:=xit+zx_i:=x_i^t+z. The shifted sequence obeys

xi+1i=xi+1t+z=Mxit+ηi+z=M(xi−z)+ηi+z=Mxi+(In−M)z+ηi,x_{i+1}^i=x_{i+1}^t+z=Mx_i^t+\eta_i+z =M(x_i-z)+\eta_i+z=Mx_i+(I_n-M)z+\eta_i,

which is the forecast model with per-step bias (In−M)z(I_n-M)z: a model with this bias reproduces the true trajectory shifted by zz. Since Hz=0H z=0, the shift is invisible in the observations at every time: Hxi=H(xit+z)=HxitH x_i=H(x_i^t+z)=H x_i^t for all ii. We write (In−M)ker⁡(H)(I_n-M)\ker(H) for the set of such per-step biases.

For linear dynamics, no method that uses only the observations can therefore detect a bias in this set: the models with per-step biases bb and b+(In−M)zb+(I_n-M)z, with Hz=0H z=0, give the same distribution of observations, apart from the effect of the initial condition. In control-theoretic terms, these biases are undetectable inputs, whose effect on the outputs can be reproduced by a change of initial state (Pasqualetti et al., 2013, Lemma 3.1). The iterated correction thus removes all of the bias that is identifiable from the observations. Note that (In−M)ker⁡(H)(I_n-M)\ker(H) differs from the set of biases in unobserved variables (Hb=0H b=0), which are in general identifiable, since MM propagates them into the observed directions.

Online methods, which estimate the bias during the assimilation, meet the same limit. In augmented-state filters, which treat the bias as an additional state variable [25, 17], these directions are unobservable modes of the augmented system (Pannocchia and Rawlings, 2003, Lemma 1); and for likelihood-based methods [18, 14] (Tandeo et al. 2020), all biases that differ by an element of this set are equally likely given the observations, apart from a contribution from the prior on the initial state that does not grow with the length of the record, so maximising the likelihood cannot determine that part of the bias.

Remark 4 (Nonlinear dynamics). For a nonlinear model M\mathcal{M}, the shift argument does not apply in general: shifting the state by zz corresponds to the state-dependent bias z+M(x)−M(x+z)≈(In−J(x))zz+\mathcal{M}(x)-\mathcal{M}(x+z)\approx(I_n-J(x))z, with J(x)J(x) the Jacobian of M\mathcal{M} at xx, which a constant bias can reproduce only if J(x)zJ(x)z does not depend on xx. Biases in these directions then becomes identifiable through the variation of J(x)J(x) over the attractor.

Relation to adaptive bias correction

The iterated correction of Section 2.5 is closely related to the adaptive bias correction scheme of Dee (2005, Section 3b), which can be written as

xi+1f=Mxia+b−b^i,(19)x_{i+1}^{f}=Mx_i^a+b-\hat b_i, \tag*{(19)}
b^i=b^i−1−γ(xia−xif),b^0=0,(20)\hat b_i=\hat b_{i-1}-\gamma\left(x_i^a-x_i^f\right),\qquad\hat b_0=0, \tag*{(20)}

where the analysis step is (3b) as before and γ>0\gamma>0 is a parameter that controls how much recent information is weighted in the running bias estimate b^i\hat b_i. The forecast model thus has a per-step bias b−b^ib-\hat b_i that is updated after every cycle by γ\gamma times the analysis increment, whereas the iterated correction updates it after a whole run by the averaged increment: the scheme is an online counterpart of the iterated correction. If it converges (which requires γ\gamma to be sufficiently small), it reaches the same limit, Πb∗\Pi b^*, and so likewise leaves only the unidentifiable component of the bias.

To see this, let bif:=E[xif−xit]b_i^f:=\mathbb{E}[x_i^f-x_i^t] as before and βi:=E[b^i]\beta_i:=\mathbb{E}[\hat b_i] be the mean of the (random) estimate. Taking expectations gives

bi+1f=Φbif+b−βi,βi=βi−1+γKHbif,b_{i+1}^f=\Phi b_i^f+b-\beta_i,\qquad\beta_i=\beta_{i-1}+\gamma K H b_i^f,

which is the recurrence (6) with the per-step bias bb replaced by the corrected per-step bias b−βib-\beta_i. The fixed points are those with KHbf=0K H b^f=0, i.e., Hbf=0H b^f=0, since KHK H has the same null space as HH (§ 2.3), so that the innovation bias vanishes, as for the iterated correction. The fixed point is moreover unique: since b^0=0\hat{b}_0 = 0, the update gives βi=Kvi\beta_i = K v_i with vi=H(∑k≤ibkf)v_i = H(\sum_{k \leq i} b_k^{f}) of the form HzHz. At a fixed point, bf=(In−Φ)−1(b−Kv)b^f = (I_n - \Phi)^{-1}(b - Kv), and applying HH gives Hbf=Hb∗−CvHb^f = Hb^* - Cv, so Hbf=0Hb^f = 0 requires Cv=Hb∗Cv = Hb^*, that is, v=C−1Hb∗v = C^{-1}Hb^*, since CC is invertible on vectors of the form HzHz. Hence bf=Πb∗b^f = \Pi b^*, as in Corollary 1.

State-dependent bias

When the model error depends on the state, the increments can be used at three levels of averaging: averaged over the whole record, as in a constant correction (§ 3.2); averaged over states with similar slowly varying components (§ 3.4); or without averaging, as training data for a learned state-dependent correction (§ 3.5). The first follows from the mean of the filter equations, as for a constant bias; the other two rest on an exact decomposition of each increment (§ 3.3).

Setup

We keep the setup of § 2.1. The only change is that the bias of the forecast model now depends on the state at the start of the forecast step:

xi+1f=Mxia+b(xia),(21)x_{i+1}^{f} = M x_i^{a} + b(x_i^{a}), \tag*{(21)}

where b:Rn→Rnb : \mathbb{R}^{n} \to\mathbb{R}^{n} is an unknown function; the constant bias of Section 2 is the case b(x)≡bb(x) \equiv b.

When relating averages over a long record to expectations, we assume that the joint process of the truth and the filter converges to a stationary distribution and is ergodic, so that averages of the increments over a long record converge to the corresponding expectations as i→∞i \to\infty. For the linear truth (1), this requires ρ(M)<1\rho(M) < 1, so the true state has zero mean at stationarity.

Averaged increments

Taking expectations of the analysis step gives, as in Section 2.2, the analysis bias (In−KH)bif(I_n - KH)b_i^f, with bif=E[xif−xit]b_i^f = \mathbb{E}[x_i^f - x_i^t], and the expected increment −KHbif-KHb_i^f. Since xi+1f−xi+1t=M(xia−xit)+b(xia)−ηix_{i+1}^f - x_{i+1}^t = M(x_i^a - x_i^t) + b(x_i^a) - \eta_i, the forecast bias satisfies

bi+1f=Φbif+E[b(xia)].(22)b_{i+1}^{f} = \Phi b_i^{f} + \mathbb{E}[b(x_i^a)]. \tag*{(22)}

Averaging the increments over a long record therefore gives −KHb∗-KHb^*, where b∗=lim⁡i→∞bif=(In−Φ)−1lim⁡i→∞E[b(xia)]b^* = \lim_{i\to\infty} b_i^f = (I_n - \Phi)^{-1}\lim_{i\to\infty}\mathbb{E}[b(x_i^a)]. This limit exists by the assumption of Section 3.1, and it is the asymptotic forecast bias, that is, the expected forecast error at stationarity; by ergodicity, it is also the time average of the forecast error over a long record. The averaged increment thus reflects the model error averaged over the states that the analysis visits, which can differ substantially from the model error at the mean state. A constant correction based on it therefore removes only this climatological average; where the model error changes sign over the visited states, much of it averages out, and only a state-dependent correction can remove it.

Unlike in the constant bias case, the recurrence (19) is not closed: the term E[b(xia)]\mathbb{E}[b(x_i^a)] depends on the full distribution of xiax_i^a, not just its mean, which is set by bifb_i^f. Closing it requires either linearity of bb or an approximation, which we discuss below. The resulting fixed-point equations for the asymptotic forecast bias b∗b^* are a tool for understanding what the increments measure, not something to be solved in practice.

Linear bias

Suppose b(x)=Bx+cb(x) = Bx + c for B∈Rn×nB \in\mathbb{R}^{n\times n} and c∈Rnc \in\mathbb{R}^{n}.2 Then E[b(xia)]=BE[xia]+c\mathbb{E}[b(x_i^a)] = B\mathbb{E}[x_i^a] + c, and since E[xia]=E[xit]+bia=E[xit]+(In−KH)bif\mathbb{E}[x_i^a] = \mathbb{E}[x_i^t] + b_i^a = \mathbb{E}[x_i^t] + (I_n - KH)b_i^f, the recurrence (19) closes:

bi+1f=(Φ+B(In−KH))bif+BE[xit]+c.(23)b_{i+1}^{f} = \left(\Phi+ B(I_n - KH)\right)b_i^{f} + B\mathbb{E}[x_i^t] + c. \tag*{(23)}

If ρ(Φ+B(In−KH))<1\rho(\Phi+ B(I_n - KH)) < 1, then, since E[xit]→0\mathbb{E}[x_i^t] \to0 (§ 3.1), the forecast bias converges to

b∗=(In−Φ−B(In−KH))−1c,(24)b^* = \left(I_n - \Phi- B(I_n - KH)\right)^{-1}c, \tag*{(24)}

which reduces to (7) when B=0B = 0. This is the steady-state mean error of a Kalman filter with a mismatched model ((Lee and Athans, 1979, eq. (29)).

Nonlinear bias

For a general nonlinear bb, closing (19) requires an approximation. Since the true state has zero mean at stationarity (§ 3.1), the stationary mean of the analysis is xˉa:=lim⁡i→∞E[xia]=(In−KH)b∗\bar{x}^a := \lim_{i\to\infty}\mathbb{E}[x_i^a] = (I_n - KH)b^*.

The model bias depends on xiax_i^a, which varies across realisations (equivalently, at stationarity, over a long time series). Unless bb is affine (§ 3.2.1), the average of bb over these varying analyses, E[b(xia)]\mathbb{E}[b(x_i^a)], differs from its value at the average analysis, b(xˉa)b(\bar{x}^a). Ignoring this variation, that is, replacing E[b(xia)]\mathbb{E}[b(x_i^a)] by b(xˉa)b(\bar{x}^a), gives the fixed-point equation

b∗=(In−Φ)−1b((In−KH)b∗).(25)b^* = (I_n - \Phi)^{-1}b\left((I_n - KH)b^*\right). \tag*{(25)}

We call this the zeroth-order approximation, because it keeps only the leading term of a Taylor expansion of bb about xˉa\bar{x}^a.

Accounting for the variation of the analysis requires its stationary covariance, Σa:=lim⁡i→∞Cov⁡(xia)\Sigma^a := \lim_{i\to\infty}\operatorname{Cov}(x_i^a), which differs from PaP^a. We approximate Σa\Sigma^a of the filter with a biased model by the covariance of the analy sis of the filter with a perfect model. Denoting the perfect-model analysis by xia,0x_i^{a,0}, the analysis error xia,0−xitx_i^{a,0}-x_i^t has covariance PaP^a asymptotically and is uncorrelated with xia,0x_i^{a,0} (Anderson and Moore, 1979, Section 5.2, Theorem 2.5), so that lim⁡i→∞Cov⁡(xia,0)=Σt−Pa\lim_{i\to\infty}\operatorname{Cov}(x_i^{a,0})=\Sigma^t-P^a. Here Σt\Sigma^t is the stationary covariance of the true state, which exists by the ergodicity assumption of § 3.1 and solves Σt=MΣtMT+Q\Sigma^t=M\Sigma^tM^T+Q. For the biased model, the approximation Σa≈Σt−Pa\Sigma^a\approx\Sigma^t-P^a is exact for a constant bias, which only shifts the analysis; a state-dependent bias adds a random component to the forecast error, which changes Σa\Sigma^a. Unlike PaP^a, Σa\Sigma^a does not shrink as the observations improve: it approaches Σt\Sigma^t, the natural variability of the system. When this is not small relative to the scale on which bb varies, the zeroth-order approximation is poor.

The analysis of the filter with a perfect model, xia,0x_i^{a,0}, is Gaussian, as a linear function of Gaussian variables. If the bias does not strongly distort this distribution, we can approximate xia∼N(xˉa,Σa)x_i^a\sim\mathcal{N}(\bar{x}^a,\Sigma^a) for large ii in (19), giving the Gaussian closure

b∗=(In−Φ)−1EX∼N((In−KH)b∗,Σa)[b(X)].(26)b^*=(I_n-\Phi)^{-1}\mathbb{E}_{X\sim\mathcal{N}((I_n-KH)b^*,\Sigma^a)}[b(X)]. \tag*{(26)}

Correction with the averaged increment

Correcting the model with the averaged increment, as in § 2.4 for a constant bias, can fail when the bias is state-dependent. Even for a linear bias, Propositions 2 and 3 need not hold: subtracting a constant correction from cc gives a corrected forecast bias governed by (In−Φ−B(In−KH))−1(I_n-\Phi-B(I_n-KH))^{-1} rather than (In−Φ)−1(I_n-\Phi)^{-1}, to which the identities of Lemma 1 do not apply. This means that a constant correction can increase the innovation bias in this setting, and the iterated correction can diverge. Consistent with this, in experiments with a quasi-geostrophic model, DelSole and Hou [19] found that a constant correction did not improve forecast skill when the model error was mainly state-dependent.

Decomposition of the increments

Averaging increments over the whole record precludes state-dependent corrections. To understand the state-dependent information contained in the increments, we need a decomposition of each increment, not only its mean. Consider a second filter with the same gain and the same observations, but with a perfect forecast model, xi+1f,0=Mxia,0x_{i+1}^{f,0}=Mx_i^{a,0}. Let Δi:=xif−xif,0\Delta_i:=x_i^f-x_i^{f,0} be the difference between the two forecasts, that is, the contribution of the bias to the forecast error.

Proposition 4 (Decomposition of the increments). Under the assumptions of Section 2.1, for any b(⋅)b(\cdot),

Δi+1=ΦΔi+b(xia),so thatΔi=ΦiΔ0+∑k=0i−1Φkb(xi−1−ka),(27)\Delta_{i+1}=\Phi\Delta_i+b(x_i^a),\qquad\text{so that}\qquad\Delta_i=\Phi^i\Delta_0+\sum_{k=0}^{i-1}\Phi^k b(x_{i-1-k}^a), \tag*{(27)}

and each increment can be decomposed as

xia−xif,0=−KHΔi+νi,νi:=xia,0−xif,0.(28)x_i^a-x_i^{f,0}=-KH\Delta_i+\nu_i,\qquad\nu_i:=x_i^{a,0}-x_i^{f,0}. \tag*{(28)}

Here νi\nu_i generalises the fluctuation of the same name in § 2.3.

Proof. Both filters apply the analysis update xia=(In−KH)xif+Kyix_i^a=(I_n-KH)x_i^f+Ky_i to the same observations, so xia−xia,0=(In−KH)Δix_i^a-x_i^{a,0}=(I_n-KH)\Delta_i, and the forecast step gives Δi+1=M(In−KH)Δi+b(xia)\Delta_{i+1}=M(I_n-KH)\Delta_i+b(x_i^a); unrolling gives (24a). The increment of the biased filter is K(yi−Hxif)=K(yi−Hxif,0)−KHΔiK(y_i-Hx_i^f)=K(y_i-Hx_i^{f,0})-KH\Delta_i, which is (24b). □\square

The decomposition involves no expectation. Its second term, νi\nu_i, is the increment of the filter with a perfect model, that is, KK times its innovation. These innovations are asymptotically distributed as N(0,S)\mathcal{N}(0,S) and independent of all earlier observations (Anderson and Moore, 1979, Section 5.3, Theorem 3.1), and xif,0x_i^{f,0} is a function of those observations. Hence νi\nu_i is asymptotically distributed as N(0,KSKT)\mathcal{N}(0,KS K^T), as in § 2.3, and independent of xifx_i^f. The first term, −KHΔi-KH\Delta_i, is the model error along the recent trajectory, weighted by Φk\Phi^k, and projected by KHKH. For constant bb, Δi\Delta_i is linear in bb, and (24b) is the decomposition used by Friedland (1969, Section III) to compute the optimal estimate of a constant bias from the innovations of a filter that ignores it; it then reduces to the split (8), with Δi=bif\Delta_i=b_i^f asymptotically.

Increments averaged by slow state

Between averaging the increments over the whole record and the individual increments lies an intermediate case. Suppose that the model error depends on the state mainly through components that vary slowly, such as those carrying a seasonal cycle or a regime that persists for many assimilation cycles, and denote these slow components by ss. By (24a), the contribution of the bias to the forecast error, Δi\Delta_i, is a weighted sum of past values of the model error along the trajectory, with weights Φk\Phi^k that decay over about τ:=1/(1−ρ(Φ))\tau:=1/(1-\rho(\Phi)) cycles. If ss changes little over τ\tau cycles, that is, if it varies on a timescale of L≫τL\gg\tau cycles, we can make a quasi-static approximation by noting that Δi\Delta_i stays close, on average, to (In−Φ)−1E[b(xi−1)∣si](I_n-\Phi)^{-1}\mathbb{E}[b(x_{i-1})\mid s_i], the value it would settle at if ss were held fixed at sis_i. Here si:=s(xif)s_i:=s(x_i^f) denotes the slow components of the forecast.

Averaging the increments over cycles with similar values of sis_i then gives

E[xia−xif∣si]≈−KH(In−Φ)−1E[b(xi−1)∣si],(29)\mathbb{E}[x_i^a-x_i^f\mid s_i]\approx-KH(I_n-\Phi)^{-1}\mathbb{E}[b(x_{i-1})\mid s_i], \tag*{(29)}

so that the averaged increment, as a function of sis_i, diagnoses the slowly varying part of the model error.

An adaptive scheme such as that of Dee [16] (§ 2.7) can also leverage this idea: with small γ\gamma, it averages the increments over roughly the last 1/γ1/\gamma cycles, and so tracks the part of the model error that changes slowly compared with this window and with the memory τ\tau of the filter, but not its dependence on the fast-changing state.

Individual increments

We now move beyond the slow-state averaging to learning the full state-dependent correction. The natural target is the part of the increment that can be predicted from the forecast state, the conditional mean E[xia−xif∣xif]\mathbb{E}[x_i^{a}-x_i^{f}\mid x_i^{f}]. The machine learning approaches cited in the introduction estimate it by least-squares regression: given a record of TT forecasts and analyses, they fit a correction function f^:Rn→Rn\hat{f}:\mathbb{R}^n\to\mathbb{R}^n over a class F\mathcal{F} of functions, such as neural networks,

f^=arg⁡min⁡f∈F 1T∑i=1T∥(xia−xif)−f(xif)∥22,\hat{f}=\underset{f\in\mathcal{F}}{\arg\min}\ \frac{1}{T}\sum_{i=1}^{T}\left\|(x_i^{a}-x_i^{f})-f(x_i^{f})\right\|_2^2,

which approaches the conditional mean for a long record and a sufficiently flexible class. This follows from the ergodicity assumption of Section 3.1 and the classical result that the conditional mean minimises the expected squared error [2 Section 2.3, Theorem 3.2]. Once learned, the predicted increment f^(xif)\hat{f}(x_i^{f}) is added to the forecast xifx_i^{f}. We now ask what this conditional mean, and each increment, represent, first for linear and then for nonlinear dynamics.

Linear dynamics

Taking the conditional mean of (24b) gives E[xia−xif∣xif]=−KHE[Δi∣xif]\mathbb{E}[x_i^{a}-x_i^{f}\mid x_i^{f}]=-KH\mathbb{E}[\Delta_i\mid x_i^{f}], since νi\nu_i has zero mean and is independent of xifx_i^{f} (Section 3.3). Each increment is therefore an unbiased but noisy measurement of −KHΔi-KH\Delta_i. If bb changes little along the trajectory over the memory time τ\tau, then Δi≈(In−Φ)−1b(xi−1a)\Delta_i\approx(I_n-\Phi)^{-1}b(x_{i-1}^{a}), and each increment approximates −KH(In−Φ)−1b(xi−1a)-KH(I_n-\Phi)^{-1}b(x_{i-1}^{a}). Such a regression therefore recovers the model error as filtered by the assimilation: accumulated over the memory of the filter and projected by KHKH, which, as in Section 2.3, cannot see the part of the accumulated bias in ker⁡(H)\ker(H).

Nonlinear dynamics

For nonlinear dynamics, the truth evolves as xi+1t=M(xit)+ηix_{i+1}^{t}=M(x_i^{t})+\eta_i, the forecast is xi+1f=M(xia)+b(xia)x_{i+1}^{f}=M(x_i^{a})+b(x_i^{a}), and the filter with a perfect model uses xi+1f,0=M(xia,0)x_{i+1}^{f,0}=M(x_i^{a,0}), so that M(x)=MxM(x)=Mx recovers the setting above. With a fixed gain, the split (24b) still holds exactly, since the analysis step is linear, and taking its conditional mean gives

E[xia−xif∣xif]=−KHE[Δi∣xif]+E[νi∣xif].(30)\mathbb{E}[x_i^{a}-x_i^{f}\mid x_i^{f}]=-KH\mathbb{E}[\Delta_i\mid x_i^{f}]+\mathbb{E}[\nu_i\mid x_i^{f}]. \tag*{(30)}

The first term is the contribution of the model error, as in the linear case. The contribution of the bias to the forecast error now satisfies Δi+1=b(xia)+M(xia)−M(xia,0)\Delta_{i+1}=b(x_i^{a})+M(x_i^{a})-M(x_i^{a,0}): the model error of the latest step plus that of earlier steps, propagated by the model, in place of the explicit form (24a). As before, the model error is seen only through KHKH, although for nonlinear dynamics bias in the unobserved directions need not be invisible (Remark 4). The second term is the conditional mean of νi=xia,0−xif,0\nu_i=x_i^{a,0}-x_i^{f,0}, and it represents the mean increment that the filter would make with a perfect model, given the forecast. For linear dynamics with Gaussian noise it vanishes, but in general it does not. A regression on the increments learns this term along with the model error.

Iterating the learned correction

As for a constant bias in Section 2.5, the learned correction can be iterated. After the correction is learned, it is applied to the forecast model. The filter is then rerun with the corrected model, and the correction is retrained on the new increments plus the current correction. This can start from the previous iteration, so that the network is fine-tuned rather than retrained from scratch. Most applications, such as those cited in the introduction and Brajard et al. [11], stop after one step; Gregory et al. [29] iterate three times.

If the iteration converges, it follows from Proposition 4 that, for linear dynamics, it stops when the remaining model error, accumulated over the memory of the filter, is on average invisible to the observations for each forecast state. This is the analogue of the condition Hb∗=0Hb^*=0 for a constant bias and, as there, does not imply that the remaining model error is zero. Whether the iteration converges does not appear to have been studied.

Numerical experiments

Linear test system

The experiments of Sections 4.2.1, 4.2.2, and 4.3.1 use a Kalman filter (KF) operating at steady state, and those of Section 4.2.1 also a stochastic ensemble Kalman filter (EnKF) with perturbed observations [21 Section 3.4], on a linear system. The state dimension is n=4n=4 and the observation dimension is p=2p=2. The model MM consists of two damped rotations, with ρ(M)=0.95\rho(M)=0.95, and each observation is the sum of two state components, so that ker⁡(H)\ker(H) is two-dimensional. The observation and model error covariances, RR and QQ, are multiples of the identity. The parameter values, including those of the biases below, are listed in Table 1 (Appendix B).

The steady-state KF gain is computed from the discrete algebraic Riccati equation (4). This gives ρ(Φ)=0.805\rho(\Phi)=0.805. The EnKF uses multiplicative covariance inflation, and model noise ηi∼N(0,Q)\eta_i\sim\mathcal{N}(0,Q) is drawn independently for each member’s forecast.

Constant bias

Averaged increment

We verify the fixed-point expressions (7) and (9) numerically for both filters on the linear system of Section 4.1, with the bias bb given in Table 1. Figure 2 compares the time-averaged analysis increment with theory for each filter. For the KF, the time-averaged forecast error agrees with the asymptotic forecast bias b∗b^*, and the averaged increment with −KHb∗-KHb^*, with one standard error in every component. The standard error is estimated from the spread of the means over equal, non-overlapping segments of the record, each much longer than the memory of the filter, which accounts for the correlation between successive cycles. The averaged EnKF increment also agrees with −KHb∗-K H b^{*}, but differs slightly more because of sampling error with a finite ensemble.

Figure 2. Time-averaged analysis increment minus its theoretical value −KHb∗-K H b^{*} from (9), for each of the four state components (the four points in each group). Here bb is given by (7) for the constant bias of Section 4.2.1 (KF and EnKF), by (21) for the linear bias, and by the zeroth-order approximation (22) or the Gaussian closure (23) for the nonlinear bias of Section 4.3.1 (KF). Error bars show ±1\pm1 standard error.

ModelM=0.95 diag⁡(R(θ),R(2θ))M = 0.95\,\operatorname{diag}(R(\theta), R(2\theta)), θ=0.4\theta= 0.4 rad
ObservationsH=(I2 I2)H = (I_2\ I_2); Q=0.5 I4Q = 0.5\,I_4, R=0.5 I2R = 0.5\,I_2
Constant bias (Sections 4.2.1 and 4.2.2)b=(0.5, −0.3, 0.2, −0.4)Tb = (0.5,\,-0.3,\,0.2,\,-0.4)^{\mathsf{T}}
Linear bias (Section 4.3.1)b(x)=Bx+cb(x) = Bx + c, B=0.05 (I4−P)B = 0.05\,(I_4 - P), c=(0.3, −0.2, 0.1, −0.15)Tc = (0.3,\,-0.2,\,0.1,\,-0.15)^{\mathsf{T}}
Nonlinear bias (Section 4.3.1)b(x)=c+αsin⁡(x)b(x) = c + \alpha\sin(x) (elementwise), α=0.3\alpha= 0.3
EnKF (Section 4.2.1)stochastic, N=50N = 50, anomaly inflation 1.02
Run lengthsSections 4.2.1 and 4.3.1: T=30 000T = 30\,000, spin-up 500; Section 4.2.2: 8 000 cycles per iteration, spin-up 200

Table 1. Parameters of the experiments with the linear system. R(θ)R(\theta) is the 2×22 \times2 rotation matrix by angle θ\theta, and PP the cyclic shift, (Px)k=xk+1(Px)_k = x_{k+1} with x5:=x1x_5 := x_1.

Iterated correction

We verify the iterated correction convergence (§ 2.5) numerically, with the KF on the same system. At each iteration jj, the filter is run with the forecast model with per-step bias b(j)b^{(j)}, the bias estimate b^(j)=−xa,(j)−xf,(j)\hat{b}^{(j)}=-x^{a,(j)}-x^{f,(j)} is computed from the time-averaged analysis increment, and subtracted from the per-step bias to give b(j+1)b^{(j+1)}. Figure 3 shows both the full-state bias norm ∥b∗(j)∥2\lVert b^{*(j)}\rVert_2 and the innovation bias norm ∥Hb∗(j)∥2\lVert H b^{*(j)}\rVert_2 over five iterations, compared to the theoretical predictions obtained by iterating (13) and from (15).

Figure 3. Convergence of the iterated analysis increment correction. Green circles: full-state bias norm ∥b∗(j)∥2\lVert b^{*(j)}\rVert_2. Orange squares: innovation bias norm ∥Hb∗(j)∥2\lVert H b^{*(j)}\rVert_2. Markers show numerical results. Dashed lines show the median, and the shading the central 68% range, of the norms predicted by (13) and (15), with the sampling noise of the finite-time averages included.

The full state bias ∥b∗(j)∥2\lVert b^{*(j)}\rVert_2 drops sharply on the first iteration then plateaus at ≈0.39\approx0.39, in agreement with the limit ∥Πb∥2=0.38\lVert\Pi b\rVert_2=0.38 of Corollary 1. This residual lies in ker⁡(H)\operatorname{ker}(H), and it cannot be removed by any correction. By contrast, the innovation bias ∥Hb∗(j)∥2\lVert H b^{*(j)}\rVert_2 converges to zero geometrically, as guaranteed by Proposition 3: it decreases by a factor of about 0.16 in the first iteration, matching the rate ρ(Ip−C)=0.162\rho(I_p-C)=0.162. From iteration 3 onwards it remains at the level of sampling noise from the finite averaging window.

Nonlinear dynamics

The theory of Section 2 assumes linear dynamics and a fixed gain. To check whether the iterated correction still works in this setting, we repeat the experiment of Section 4.2.2 with the Lorenz ’96 model (Lorenz, 1996),

dxkdt=(xk+1−xk−2)xk−1−xk+F,k=1,…,n,\frac{d x_k}{d t}=(x_{k+1}-x_{k-2})x_{k-1}-x_k+F,\qquad k=1,\ldots,n,

with n=20n=20, F=8F=8, and periodic boundary conditions, integrated with a fourth-order Runge–Kutta scheme. Every other variable is observed (p=10p=10), and a localised EnKF [31] with multiplicative inflation and Gaspari–Cohn localisation [26] is used. The forecast model has a constant, spatially varying bias, a sinusoid in kk plus a small random part, whose largest component, expressed as a tendency, is about 15% of the forcing FF. The parameter values are listed in Table 2 (Appendix § 9).

Modeln=20n = 20, F=8F = 8; RK4 with Δt=0.05\Delta t = 0.05
Observationsx1,x3,…,x19x_1, x_3, \ldots, x_{19} (p=10p = 10); Q=0.01 I20Q = 0.01\,I_{20}, R=0.5 I10R = 0.5\,I_{10}
Constant bias (Section 4.2.3)bk=0.04sin⁡(2π(k−1)/n)+0.016 ξkb_k = 0.04\sin(2\pi(k-1)/n) + 0.016\,\xi_k, with ξk∼N(0,1)\xi_k \sim\mathcal{N}(0,1) drawn once
State-dependent bias (Section 4.3.2)b(x)=0.15sin⁡(0.6x+0.5)b(x) = 0.15\sin(0.6x + 0.5) (elementwise)
EnKFstochastic, covariance inflation 1.04, Gaspari–Cohn localisation of half-width 5; Section 4.2.3: N=100N = 100; Section 4.3.2: N=40N = 40
Run lengthsT=30 000T = 30\,000 cycles; Section 4.2.3: spin-up 1 000; Section 4.3.2: spin-up 2 000
Network (Section 4.3.2)one input, standardised, two hidden layers of 32 tanh units, one output; Adam with learning rate 3×10−33 \times10^{-3} (10−310^{-3} when fine-tuning) and cosine decay, 40 epochs, batches of 4 096; output layer refitted by least squares

Table 2. Parameters of the Lorenz ’96 experiments.

Here b∗b^{*}, as in Section 3.2, the expected forecast error at stationarity, estimated by the time-averaged forecast error. Figure 4 shows that the behaviour of Section 4.2.2 carries over. The innovation bias norm ∥Hb∗∥2\lVert H b^{*}\rVert_2 decreases geometrically, and approaches the level of sampling noise. The full-state bias norm ∥b∗∥2\lVert b^{*}\rVert_2 also decreases at first, but then levels off above the perfect-model level. A residual bias, almost entirely in the unobserved components, thus remains.

Figure 4. Iterated increment correction for the Lorenz ’96 model with a constant, spatially varying bias. Green circles: full-state bias norm ∥b∗(j)∥2\lVert b^{*(j)} \rVert_2. Orange squares: innovation bias norm ∥Hb∗(j)∥2\lVert H b^{*(j)} \rVert_2. The shading along the curves shows ±1\pm1 standard error, and the horizontal bands of the same colours the central 68% range of each norm expected from sampling noise alone with a perfect model.

State-dependent bias

Averaged increments

We verify the fixed-point expressions of Section 3.2 numerically for both a linear and a nonlinear bias function, using the system of Section 4.1.

Linear bias. We take b(x)=Bx+cb(x) = Bx + c, with BB and cc given in Table 1. The theoretical fixed point is given by (21).

Nonlinear bias. We take b(x)=c+αsin⁡(x)b(x) = c + \alpha\sin(x) elementwise. Then, in the forecast step, the filter applies (18), but it keeps the steady-state Kalman gain of Section 4.1.

Figure 2 compares the time-averaged analysis increment with theory for both bias functions. For the linear bias, the theory is exact, and it agrees with the averaged increment to within one standard error in every component. For the nonlinear bias, the Gaussian closure also agrees to within one standard error, whereas the zeroth-order approximation, which ignores the spread of the analysis state, does not.

Learned corrections

To illustrate Section 3.5, we learn a state-dependent correction from the increments of the Lorenz ’96 system of Section 4.2.3. Here the bias depends on the state through the same sinusoidal function in every component. The other settings are as in Section 4.2.3, except for the bias, the ensemble size, and the spin-up (Table 2). The correction is a neural network whose input is the forecast value of a single component and whose output is the predicted increment of that component, with two hidden layers. It is fitted by least squares, as in Section 3.5, to the increments of the ensemble mean in the observed components, with the ensemble-mean forecast as input, and is applied to those components only. In the filter, the network is evaluated at the forecast of each ensemble member, and its output is added to that forecast.

To compare with the theory (26), we run a second filter with a perfect model and the same gains and observations, which gives −KiHΔi-K_i H\Delta_i at every cycle. Its conditional mean given the forecast is estimated by fitting a network of the same form to −KiHΔi-K_i H\Delta_i in place of the increment.

Figure 5(a) shows the first learned correction in the observed components. It follows the theory closely and demonstrates that, as per Section 3.5, the regression learns the model error as filtered by the assimilation, not the model error itself. The small remaining difference is the conditional mean of νi\nu_i, the systematic increment that the filter would make with a perfect model. Iterating the correction as in Section 3.5 brings it closer to −b-b (Figure 5(b)). The analysis bias in the observed components falls about fivefold (Figure 5(c)).

Conclusions

Analysis increments are a widely used and inexpensive way to diagnose and correct model bias. This paper has provided a theoretical foundation for their use.

For a linear model with a constant bias, and observations assimilated with a Kalman filter, we showed that the forecast bias settles to a steady-state value. The averaged increment provides an estimate of this bias, seeing the part of it that is visible in the observations, spread into unobserved variables by the Kalman gain. Correcting the model with the averaged increment reduces the bias but does not remove it. Repeating the correction, each time using the increments of the corrected model, removes the bias seen by the observations completely, at a geometric rate. What remains is equivalent to a constant offset of the state in directions that the observations do not see, and no method can detect it. The repeated correction provides an alternative to online methods, such as augmented-state filters and likelihood-based methods, and does not require changing the assimilation system.

For state-dependent biases, which machine-learned corrections target, the picture is similar but less complete. For linear dynamics, each increment records the model error accumulated over the memory of the filter, plus noise. As for a constant bias, it sees only the part visible in the observations, spread into unobserved variables by the Kalman gain. Averaged over a long record, the increments measure the mean of this filtered error over the states the system visits. Unlike for a constant bias, correcting with this mean can increase the bias seen by the observations. Averaged separately for each season or weather regime, the increments recover the slowly varying part of the model error, provided it changes slowly compared with the memory of the filter. The individual increments can be used for learning statedependent corrections; however, they learn the filtered model error, not the model error itself.

Figure 5. Learned corrections in the Lorenz ’96 experiment with a state-dependent bias. (a) First learned correction against the model error −b-b in the observed components, for individual cycles (points) and averaged in bins of −b-b (squares). The solid line is the theory, the conditional mean of −KiHΔi-K_iH\Delta_i given the forecast, and the dashed line is y=xy=x. (b) Same as (a), but after three iterations of the correction. (c) Norm of the analysis bias in the observed components, ∥Hba∥2\lVert H b^a\rVert_2, estimated from the time-mean analysis error, against the number of iterations. The line shows the mean of two independent runs, which differ by at most 0.005.

Numerical experiments confirm the theory for linear systems. Experiments with the Lorenz ’96 model and an ensemble Kalman filter suggest that the main conclusions carry over to nonlinear systems.

Several questions remain open. Our results assume a filter that is unaware of the bias and has a fixed gain; extending them to filters that adapt their gain to the bias ([17]) is an area for future work. Most of the analysis in this work focuses on linear systems, and extending it to the fully nonlinear setting would bring the theory closer to operational systems and to machine-learned corrections ([10, 28, 7, 3] (Mojgani et al. 2024)). For nonlinear dynamics, part of the bias that is undetectable for linear models can become detectable (Remark 4). How much can be recovered is an open question. Statistics beyond the mean, such as lagged innovation covariances ([9]), or learned corrections with memory, which use the recent history of the forecasts rather than only the current one ((Levine and Stuart, 2022)), may help. A similar analysis could be applied to increments from smoothers such as 4D-Var or the Kalman smoother. Finally, the analysis could be extended to observations of statistics of the state rather than the state itself ([6]) and to multi-model data assimilation ([5]).

Acknowledgements

I am grateful to the late Eugenia Kalnay, whose work with her students and collaborators on estimating and correcting model error from analysis increments motivated this work.

Declaration of Generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the author used Claude Sonnet 4.6, Opus 5, and Opus 5.5 (Anthropic) to assist with mathematical exposition and proof development, as well as implementation of numerical experiments. After using these tools, the author reviewed and edited all content, verified all mathematical arguments and numerical results, and takes full responsibility for the content of the publication.

Proofs

Throughout, S=HPfHT+RS = H P^{f} H^{T} + R, K=PfHTS−1K = P^{f} H^{T} S^{-1}, Φ=M(In−KH)\Phi= M(I_{n} - KH), and C=H(In−Φ)KC = H(I_{n} - \Phi)K. Recall that ρ(Φ)<1\rho(\Phi) < 1, so that X:=(In−Φ)−1X := (I_{n} - \Phi)^{-1} exists, and that PfP^{f} and RR are positive definite. For symmetric matrices, A⪯BA \preceq B means that B−AB - A is a positive semidefinite. Proposition 2 uses the following identities, in which W:=Ip−HXMKW := I_{p} - HXM K.

Lemma 1 (Key identities).

Ip−C=WRS−1,S=WRWT+HXQXTHT.(31)I_{p} - C = WRS^{-1}, \qquad S = WRW^{T} + HXQX^{T}H^{T}. \tag*{(31)}

Proof. Since KS=PfHTKS = P^{f}H^{T}, we have HK=(S−R)S−1HK = (S - R)S^{-1}, that is, RS−1=Ip−HKRS^{-1} = I_{p} - HK. Hence

WRS−1=(Ip−HXMK)(Ip−HK)=Ip−HK−HXMK(Ip−HK).WRS^{-1} = (I_{p} - HXM K)(I_{p} - HK) = I_{p} - HK - HXM K(I_{p} - HK).

Since MK(Ip−HK)=M(K−KHK)=M(In−KH)K=ΦKMK(I_{p} - HK) = M(K - KHK) = M(I_{n} - KH)K = \Phi K, this equals Ip−H(In+XΦ)K=Ip−HXK=Ip−CI_{p} - H(I_{n} + X\Phi)K = I_{p} - HXK = I_{p} - C, using In+XΦ=XI_{n} + X\Phi= X. The second identity is Anderson and Moore (1979, Section 4.5, Problem 5.1) at z=1z = 1. □

Proof of Proposition 2. From (12), Hbˉ∗=H(In−(In−Φ)−1KH)b∗=(Ip−C)Hb∗H\bar b^{*} = H(I_{n} - (I_{n} - \Phi)^{-1}KH)b^{*} = (I_{p} - C)Hb^{*}. By the second identity in (27), WRWT⪯SWRW^{T} \preceq S, so Y:=S−1/2WR1/2Y := S^{-1/2}WR^{1/2} satisfies YYT⪯IpYY^{T} \preceq I_{p}, and hence YTY⪯IpY^{T}Y \preceq I_{p}, that is, R1/2WTS−1WR1/2⪯IpR^{1/2}W^{T}S^{-1}WR^{1/2} \preceq I_{p}. By the first identity,

(Ip−C)TS−1(Ip−C)=S−1R1/2(R1/2WTS−1WR1/2)R1/2S−1⪯S−1RS−1.(I_{p} - C)^{T}S^{-1}(I_{p} - C) = S^{-1}R^{1/2}(R^{1/2}W^{T}S^{-1}WR^{1/2})R^{1/2}S^{-1} \preceq S^{-1}RS^{-1}.

Moreover, S−1−S−1RS−1=S−1(S−R)S−1=S−1HPfHTS−1⪰0S^{-1} - S^{-1}RS^{-1} = S^{-1}(S - R)S^{-1} = S^{-1}HP^{f}H^{T}S^{-1} \succeq0. If ν≠0\nu\ne0 is in range⁡(H)\operatorname{range}(H), so that ν=Hz\nu= Hz for some z∈Rnz \in\mathbb{R}^{n}, then u:=HTS−1ν≠0u := H^{T}S^{-1}\nu\ne0, since zTu=(Hz)TS−1ν=νTS−1ν>0z^{T}u = (Hz)^{T}S^{-1}\nu= \nu^{T}S^{-1}\nu> 0. Then, νT(S−1−S−1RS−1)ν=uTPfu>0\nu^{T}(S^{-1} - S^{-1}RS^{-1})\nu= u^{T}P^{f}u > 0. Hence, for every nonzero ν∈range⁡(H)\nu\in\operatorname{range}(H),

∥(Ip−C)ν∥S2≤νTS−1RS−1ν=∥ν∥S2−uTPfu<∥ν∥S2,\lVert(I_{p} - C)\nu\rVert_{S}^{2} \le\nu^{T}S^{-1}RS^{-1}\nu= \lVert\nu\rVert_{S}^{2} - u^{T}P^{f}u < \lVert\nu\rVert_{S}^{2},

that is,

∥(Ip−C)ν∥S<∥ν∥Sfor every nonzero ν∈range⁡(H).(32)\lVert(I_{p} - C)\nu\rVert_{S} < \lVert\nu\rVert_{S} \quad\text{for every nonzero } \nu\in\operatorname{range}(H). \tag*{(32)}

Taking v=Hb∗v = Hb^* gives the result, with strict inequality whenever Hb∗≠0Hb^* \ne0. □

Proof of Proposition 3. By (28), Ip−CI_p - C shrinks the SS-norm of every nonzero vector in range⁡(H)\operatorname{range}(H), which it maps into itself, since (Ip−C)Hz=H(z−XKHz)(I_p - C)Hz = H(z - XKH z). Since range⁡(H)\operatorname{range}(H) is finite-dimensional, Ip−CI_p - C shrinks all such vectors by at least a common factor ϑ<1\vartheta< 1:

∥(Ip−C)v∥S≤ϑ∥v∥S\lVert(I_p - C)v\rVert_S \le\vartheta\lVert v\rVert_S

for all v∈range⁡(H)v \in\operatorname{range}(H). Hence (Ip−C)jv→0(I_p - C)^jv \to0 geometrically for every v∈range⁡(H)v \in\operatorname{range}(H), and in particular ∥Hb∗(j)∥S\lVert Hb^{*(j)}\rVert_S decreases monotonically to zero. If HH has full row rank, range⁡(H)=Rp\operatorname{range}(H) = \mathbb{R}^p, so

(Ip−C)TS−1(Ip−C)⪯ϑ2S−1.(I_p - C)^{\mathsf{T}}S^{-1}(I_p - C) \preceq\vartheta^2 S^{-1}.

For an eigenvalue μ\mu of Ip−CI_p - C with eigenvector x∈Cpx \in\mathbb{C}^p,

∣μ∣2x∗S−1x=x∗(Ip−C)TS−1(Ip−C)x≤ϑ2x∗S−1x,|\mu|^2 x^*S^{-1}x = x^*(I_p - C)^{\mathsf{T}}S^{-1}(I_p - C)x \le\vartheta^2 x^*S^{-1}x,

so ∣μ∣≤ϑ|\mu| \le\vartheta, and ρ(Ip−C)≤ϑ<1\rho(I_p - C) \le\vartheta< 1. □

Proof of Corollary 1. Each iteration subtracts KHb∗(j)=K(Ip−C)jHb∗KHb^{*(j)} = K(I_p - C)^jHb^* from the per-step bias, by (15). Since Hb∗∈range⁡(H)Hb^* \in\operatorname{range}(H), on which (Ip−C)j→0(I_p - C)^j \to0 geometrically (Proposition 3), the Neumann series gives ∑j≥0(Ip−C)jHb∗=C−1Hb∗\sum_{j\ge0}(I_p - C)^jHb^* = C^{-1}Hb^*, so b(j)→b(∞)b^{(j)} \to b^{(\infty)} and b∗(j)=(In−Φ)−1b(j)→Πb∗b^{*(j)} = (I_n - \Phi)^{-1}b^{(j)} \to\Pi b^*. Since H(In−Φ)−1KC−1H=CC−1H=HH(I_n - \Phi)^{-1}KC^{-1}H = CC^{-1}H = H, we have HΠ=0H\Pi= 0. Hence z:=Πb∗z := \Pi b^* satisfies Hz=0Hz = 0, and b(∞)=(In−Φ)z=(In−M)zb^{(\infty)} = (I_n - \Phi)z = (I_n - M)z, since Φz=Mz\Phi z = Mz when Hz=0Hz = 0. □

Experimental settings

Table 1 lists the parameters of the experiments with the linear system (Sections 4.2.1, 4.2.2, and 4.3.1), and Table 2 lists the parameters of the experiments with the Lorenz ’96 model (Sections 4.2.3 and 4.3.2).

References

  1. [1]Amemiya, A., Shlok, M. and Miyoshi, T. (2023) Application of recurrent neural networks to model bias correction: Idealized experiments with the Lorenz-96 model. Journal of Advances in Modeling Earth Systems, 15(2), e2022MS003164. Available from: https://doi.org/10.1029/2022MS003164
  2. [2]Anderson, B.D.O. and Moore, J.B. (1979) Optimal Filtering. Englewood Cliffs, NJ: Prentice-Hall.
  3. [3]Arcucci, R., Healy, S., Dance, S., Lei, L., Bach, E., Weaver, A.T., Miyoshi, T. et al. (2026) The convergence of machine learning and data assimilation in Earth system science. npj Artificial Intelligence, 2(1), 48. Available from: https://doi.org/10.1038/s44387-026-00107-0
  4. [4]Bach, E. (2027) The interface of data assimilation and machine learning. In: Weather and Climate: Applications of Machine Learning and Artificial Intelligence, number 13 in Developments in Weather and Climate Science. Elsevier. Available from: https://doi.org/10.48550/arXiv.2610.07496
  5. [5]Bach, E. and Ghil, M. (2023) A multi-model ensemble Kalman filter for data assimilation and forecasting. Journal of Advances in Modeling Earth Systems, 15(1), e2022MS003123. Available from: https://doi.org/10.1029/2022MS003123
  6. [6]Bach, E., Colonius, T., Scherl, I. and Stuart, A. (2024) Filtering dynamical systems using observations of statistics. Chaos: An Interdisciplinary Journal of Nonlinear Science, 34(3), 033119. Available from: https://doi.org/10.1063/5.0171827
  7. [7]Bach, E., Baptista, R., Sanz-Alonso, D. and Stuart, A. (2026) Machine Learning for Inverse Problems and Data Assimilation. Available from: https://doi.org/10.48550/arXiv.2410.10523. arXiv:2410.10523v3 [stat]
  8. [8]Bhargava, K., Kalnay, E., Carton, J.A. and Yang, F. (2018) Estimation of systematic errors in the GFS using analysis increments. Journal of Geophysical Research: Atmospheres, 123(3), 1626–1637. Available from: https://doi.org/10.1002/2017JD027423
  9. [9]Bonavita, M. (2021) Exploring the structure of time-correlated model errors in the ECMWF data assimilation system. Quarterly Journal of the Royal Meteorological Society, 147(739), 3454–3471. Available from: https://doi.org/10.1002/qj.4137
  10. [10]Bonavita, M. and Laloyaux, P. (2020) Machine learning for model error inference and correction. Journal of Advances in Modeling Earth Systems, 12(12), e2020MS002232. Available from: https://doi.org/10.1029/2020MS002232
  11. [11]Brajard, J., Carrassi, A., Bocquet, M. and Bertino, L. (2021) Combining data assimilation and machine learning to infer unresolved scale parametrization. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 379(2194), 20200086. Available from: https://doi.org/10.1098/rsta.2020.0086
  12. [12]Carrassi, A. and Vanitssem, S. (2011) Treatment of the error due to unresolved scales in sequential data assimilation. International Journal of Bifurcation and Chaos, 21(12), 3619–3626. Available from: https://doi.org/10.1142/S0218127411030775
  13. [13]Chen, T.C., Penny, S.G., Whitaker, J.S., Frolov, S., Pincus, R. and Tulich, S. (2022a) Correcting systematic and state-dependent errors in the NOAA FV3-GFS using neural networks. Journal of Advances in Modeling Earth Systems, 14(11), e2022MS003309. Available from: https://doi.org/10.1029/2022MS003309
  14. [14]Chen, Y., Sanz-Alonso, D. and Willett, R. (2022b) Autodifferentiable ensemble Kalman filters. SIAM Journal on Mathematics of Data Science, 4(2), 801–833. Available from: https://doi.org/10.1137/21M1434477
  15. [15]Danforth, C.M., Kalnay, E. and Miyoshi, T. (2007) Estimating and correcting global weather model error. Monthly Weather Review, 135(2), 281–299. Available from: https://doi.org/10.1175/MWR3289.1
  16. [16]Dee, D.P. (2005) Bias and data assimilation. Quarterly Journal of the Royal Meteorological Society, 131(613), 3323–3343. Available from: https://doi.org/10.1256/qj.05.137
  17. [17]Dee, D.P. and da Silva, A.M. (1998) Data assimilation in the presence of forecast bias. Quarterly Journal of the Royal Meteorological Society, 124(545), 269–295. Available from: https://doi.org/10.1002/qj.49712454512
  18. [18]Dee, D.P. and da Silva, A.M. (1999) Maximum-likelihood estimation of forecast and observation error covariance parameters. Part I: Methodology. Monthly Weather Review, 127(8), 1822–1834. Available from: https://doi.org/10.1175/1520-0493(1999)127<1822:MLEOFA>2.0.CO;2
  19. [19]DelSole, T. and Hou, A.Y. (1999) Empirical correction of a dynamical model. Part I: Fundamental issues. Monthly Weather Review, 127(11), 2533–2545. Available from: https://doi.org/10.1175/1520-0493(1999)127<2533:ECOADM>2.0.CO;2
  20. [20]Elaydi, S. (2005) An Introduction to Difference Equations, 3rd edition. Undergraduate Texts in Mathematics. New York, NY: Springer-Verlag. Available from: https://doi.org/10.1007/0-387-27602-5
  21. [21]Evensen, G. (2003) The ensemble Kalman filter: Theoretical formulation and practical implementation. Ocean Dynamics, 53(4), 343–367. Available from: https://doi.org/10.1007/s10236-003-0036-9
  22. [22]Farchi, A., Chrust, M., Bocquet, M. and Bonavita, M. (2025) Development of an offline and online hybrid model for the Integrated Forecasting System. Quarterly Journal of the Royal Meteorological Society, 151(768), e4934. Available from: https://doi.org/10.1002/qj.4934
  23. [23]Farchi, A., Chrust, M., Bonavita, M., Bocquet, M., Pan, W. and Laloyaux, P. (2026) Batched NN 4D-Var: A performant hybrid model framework for the Integrated Forecasting System. Available from: https://doi.org/10.48550/arXiv.2610.32582. arXiv:2610.32582
  24. [24]Fowler, A.M. (2024) On the robustness of methods to account for background bias in data assimilation to uncertainties in the bias estimates. Quarterly Journal of the Royal Meteorological Society, 150(763), 3657–3671. Available from: https://doi.org/10.1002/qj.4790
  25. [25]Friedland, B. (1969) Treatment of bias in recursive filtering. IEEE Transactions on Automatic Control, 14(4), 359–367. Available from: https://doi.org/10.1109/TAC.1969.1099223
  26. [26]Gaspari, G. and Cohn, S.E. (1999) Construction of correlation functions in two and three dimensions. Quarterly Journal of the Royal Meteorological Society, 125(554), 723–757. Available from: https://doi.org/10.1002/qj.49712555417
  27. [27]Gottwald, G.A. and Reich, S. (2021) Combining machine learning and data assimilation to forecast dynamical systems from noisy partial observations. Chaos: An Interdisciplinary Journal of Nonlinear Science, 31(10), 101103. Available from: https://doi.org/10.1063/5.0066080
  28. [28]Gregory, W., Bushuk, M., Adcroft, A., Zhang, Y. and Zanna, L. (2023) Deep learning of systematic sea ice model errors from data assimilation increments. Journal of Advances in Modeling Earth Systems, 15(10), e2023MS003757. Available from: https://doi.org/10.1029/2023MS003757
  29. [29]Gregory, W., Bushuk, M., Zhang, Y., Adcroft, A. and Zanna, L. (2024) Machine learning for online sea ice bias correction within global ice-ocean simulations. Geophysical Research Letters, 51(3), e2023GL106776. Available from: https://doi.org/10.1029/2023GL106776
  30. [30]Griffith, A.K. and Nichols, N.K. (2000) Adjoint methods in data assimilation for estimating model error. Flow, Turbulence and Combustion, 65(3), 469–488. Available from: https://doi.org/10.1023/A:1014541092093
  31. [31]Houtekamer, P.L. and Mitchell, H.L. (2001) A sequential ensemble Kalman filter for atmospheric data assimilation. Monthly Weather Review, 129(1), 123–137. Available from: https://doi.org/10.1175/1520-0493(2001)129<0123:ASEKFF>2.0.CO;2
  32. [32]Ide, K., Courtier, P., Ghil, M. and Lorenc, A.C. (1997) Unified notation for data assimilation: Operational, sequential and variational. Journal of the Meteorological Society of Japan. Ser. II, 75(1B), 181–189. Available from: https://doi.org/10.2151/jmsj1965.75.1B_181
  33. [33]Kalman, R.E. (1960) A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1), 35–45. Available from: https://doi.org/10.1115/1.3662552

Paper details

vlnc:W_jhfb76s8
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.

  • Physics and Astronomy

Claude Sonnet 4.6 (bare), Claude Opus 5 (bare), Opus 5.5

View statement
Looking for related papers…
Contents