In this paper
AbstractIntroductionConstant biasSetupAsymptotic forecast biasThe time-averaged analysis incrementBias correction using the averaged analysis incrementIterated correctionWhat cannot be identifiedRelation to adaptive bias correctionState-dependent biasSetupAveraged incrementsLinear biasNonlinear biasCorrection with the averaged incrementDecomposition of the incrementsIncrements averaged by slow stateIndividual incrementsLinear dynamicsNonlinear dynamicsIterating the learned correctionNumerical experimentsLinear test systemConstant biasAveraged incrementIterated correctionNonlinear dynamicsState-dependent biasAveraged incrementsLearned correctionsConclusionsAcknowledgementsDeclaration of Generative AI and AI-assisted technologies in the writing processProofsExperimental settingsReferencesWorksPhysics 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:
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).
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).
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).
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.
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 , , and denote the true, forecast, and analysis states respectively.
We consider a linear state-space model. The true state evolves as
where is the model and is zero-mean model noise. Observations are given by
with and positive definite. Throughout this paper we assume is time-invariant.
The state is estimated with a Kalman filter at steady state. Its forecast model is imperfect, introducing a constant bias at each forecast step, which represents systematic model error. The forecast and analysis steps are
where and are the means of the filter’s forecast and analysis distributions. The Kalman gain is
where 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 , so and are those of the filter with a perfect model. The analysis error covariance is , where is the identity matrix, and
Throughout, we assume that is detectable and is stabilisable (the latter holds, for instance, if 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 satisfies , where 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 is positive definite. This holds if is positive definite, since by (4).
Throughout, denotes expectation over realisations of the model noise, the observation noise, and the initial condition. The filter’s estimates and are computed from the observations and are therefore random variables in this sense.
Asymptotic forecast bias
Define the forecast bias at step as . Taking expectations of the analysis step (3b) gives the analysis bias:
This equation was also derived in Dee and da Silva [17]; Dee [16].
The analysis bias is thus the forecast bias multiplied by . Propagating through the model to the next forecast:
The bias therefore satisfies the linear recurrence relation
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
and the analysis bias converges to .
Proof. Since , the recurrence relation (6) converges to the unique fixed point satisfying , i.e., (see, e.g., Theorem 4.13 in Elaydi [20]). Since implies 1 is not an eigenvalue of , the matrix is invertible and . 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 may be unstable.
The time-averaged analysis increment
The analysis increment at step is , where is the innovation. Taking expectations:
Up to sign, the asymptotic mean of the innovations is therefore , 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 as
and write . Write each increment as a deterministic part plus a fluctuation,
Since the dynamics are linear, the bias only shifts the mean of the innovations. The centred innovations are those the steady-state filter would produce with a perfect model, which, at stationarity, are independent and identically distributed as , with (Anderson and Moore [2]). The time average of the fluctuations therefore tends to zero, by the law of large numbers. Since , the time average of the deterministic part tends to . Hence
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 .
The averaged increment does not see all of the bias. The matrix has the same null space as : if then, since is invertible, , so . By (9), only the component of in the row space of is therefore detectable from the averaged analysis increment. The component in the null space , that is, in state directions not seen by the observations, is invisible to it; such directions exist whenever , in particular when . This does not mean that unobserved variables are left uncorrected. The averaged increment 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 are, at stationarity, independent with covariance , their time average over cycles has covariance , so the error in estimating from cycles is . In addition, the initial transient decays geometrically, since , and so biases the average by ; the decay is slow when 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., . If the observations have a systematic bias , the innovation satisfies
and the averaged analysis increment becomes
with now the modified fixed point given below, conflating the forecast bias with the observation bias projected through . The fixed-point equation (8) is modified to
so that observation bias acts as an additional effective per-step model bias , since the analysis bias is propagated by 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),
and the forecast is corrected by subtracting it at each cycle:
Here is the actual per-step bias of the forecast model and its estimate. Using (8), the effective per-step bias after correction is:
The new asymptotic bias satisfies , giving:
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,
where is the Mahalanobis norm with respect to the innovation covariance . The inequality is strict whenever .
Proof. In Appendix A. □
Iterated correction
The correction (12) can be applied iteratively as follows: at iteration , 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 , the filter is run with the forecast model
and the per-step bias is updated by subtracting the bias estimate of that run, :
where is the asymptotic forecast bias of iteration . Equivalently, : the forecast model at iteration subtracts the sum of all previous bias estimates. The new asymptotic bias satisfies:
Unrolling the recurrence (14) from :
which does not in general converge to zero. The matrix vanishes on the null space of , so leaves any 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 obeys a closed recurrence. Since with
applying to (14) gives , and hence
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: for every of the form , and in particular as . If has full row rank , that is, no observation is a linear combination of the others), the rate is
.
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 . When has full row rank, a single-step reduction for every is equivalent to , and this fails for some systems in which is sufficiently non-normal. The iterates 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 is invertible on vectors of the form , and we write for this inverse.1 It is the usual inverse when 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
and the asymptotic bias converges to , where
The limit is invisible to the observations (), and the remaining per-step bias is of the form with .
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 with , as in Corollary 1: . The shifted sequence obeys
which is the forecast model with per-step bias : a model with this bias reproduces the true trajectory shifted by . Since , the shift is invisible in the observations at every time: for all . We write 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 and , with , 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 differs from the set of biases in unobserved variables (), which are in general identifiable, since 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 , the shift argument does not apply in general: shifting the state by corresponds to the state-dependent bias , with the Jacobian of at , which a constant bias can reproduce only if does not depend on . Biases in these directions then becomes identifiable through the variation of 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
where the analysis step is (3b) as before and is a parameter that controls how much recent information is weighted in the running bias estimate . The forecast model thus has a per-step bias that is updated after every cycle by 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 to be sufficiently small), it reaches the same limit, , and so likewise leaves only the unidentifiable component of the bias.
To see this, let as before and be the mean of the (random) estimate. Taking expectations gives
which is the recurrence (6) with the per-step bias replaced by the corrected per-step bias . The fixed points are those with , i.e., , since has the same null space as (§ 2.3), so that the innovation bias vanishes, as for the iterated correction. The fixed point is moreover unique: since , the update gives with of the form . At a fixed point, , and applying gives , so requires , that is, , since is invertible on vectors of the form . Hence , 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:
where is an unknown function; the constant bias of Section 2 is the case .
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 . For the linear truth (1), this requires , 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 , with , and the expected increment . Since , the forecast bias satisfies
Averaging the increments over a long record therefore gives , where . 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 depends on the full distribution of , not just its mean, which is set by . Closing it requires either linearity of or an approximation, which we discuss below. The resulting fixed-point equations for the asymptotic forecast bias are a tool for understanding what the increments measure, not something to be solved in practice.
Linear bias
Suppose for and .2 Then , and since , the recurrence (19) closes:
If , then, since (§ 3.1), the forecast bias converges to
which reduces to (7) when . 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 , closing (19) requires an approximation. Since the true state has zero mean at stationarity (§ 3.1), the stationary mean of the analysis is .
The model bias depends on , which varies across realisations (equivalently, at stationarity, over a long time series). Unless is affine (§ 3.2.1), the average of over these varying analyses, , differs from its value at the average analysis, . Ignoring this variation, that is, replacing by , gives the fixed-point equation
We call this the zeroth-order approximation, because it keeps only the leading term of a Taylor expansion of about .
Accounting for the variation of the analysis requires its stationary covariance, , which differs from . We approximate 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 , the analysis error has covariance asymptotically and is uncorrelated with (Anderson and Moore, 1979, Section 5.2, Theorem 2.5), so that . Here is the stationary covariance of the true state, which exists by the ergodicity assumption of § 3.1 and solves . For the biased model, the approximation 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 . Unlike , does not shrink as the observations improve: it approaches , the natural variability of the system. When this is not small relative to the scale on which varies, the zeroth-order approximation is poor.
The analysis of the filter with a perfect model, , is Gaussian, as a linear function of Gaussian variables. If the bias does not strongly distort this distribution, we can approximate for large in (19), giving the Gaussian closure
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 gives a corrected forecast bias governed by rather than , 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, . Let 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 ,
and each increment can be decomposed as
Here generalises the fluctuation of the same name in § 2.3.
Proof. Both filters apply the analysis update to the same observations, so , and the forecast step gives ; unrolling gives (24a). The increment of the biased filter is , which is (24b).
The decomposition involves no expectation. Its second term, , is the increment of the filter with a perfect model, that is, times its innovation. These innovations are asymptotically distributed as and independent of all earlier observations (Anderson and Moore, 1979, Section 5.3, Theorem 3.1), and is a function of those observations. Hence is asymptotically distributed as , as in § 2.3, and independent of . The first term, , is the model error along the recent trajectory, weighted by , and projected by . For constant , is linear in , 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 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 . By (24a), the contribution of the bias to the forecast error, , is a weighted sum of past values of the model error along the trajectory, with weights that decay over about cycles. If changes little over cycles, that is, if it varies on a timescale of cycles, we can make a quasi-static approximation by noting that stays close, on average, to , the value it would settle at if were held fixed at . Here denotes the slow components of the forecast.
Averaging the increments over cycles with similar values of then gives
so that the averaged increment, as a function of , 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 , it averages the increments over roughly the last cycles, and so tracks the part of the model error that changes slowly compared with this window and with the memory 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 . The machine learning approaches cited in the introduction estimate it by least-squares regression: given a record of forecasts and analyses, they fit a correction function over a class of functions, such as neural networks,
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 is added to the forecast . 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 , since has zero mean and is independent of (Section 3.3). Each increment is therefore an unbiased but noisy measurement of . If changes little along the trajectory over the memory time , then , and each increment approximates . Such a regression therefore recovers the model error as filtered by the assimilation: accumulated over the memory of the filter and projected by , which, as in Section 2.3, cannot see the part of the accumulated bias in .
Nonlinear dynamics
For nonlinear dynamics, the truth evolves as , the forecast is , and the filter with a perfect model uses , so that 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
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 : 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 , although for nonlinear dynamics bias in the unobserved directions need not be invisible (Remark 4). The second term is the conditional mean of , 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 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 and the observation dimension is . The model consists of two damped rotations, with , and each observation is the sum of two state components, so that is two-dimensional. The observation and model error covariances, and , 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 . The EnKF uses multiplicative covariance inflation, and model noise 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 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 , and the averaged increment with , 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 , but differs slightly more because of sampling error with a finite ensemble.
Figure 2. Time-averaged analysis increment minus its theoretical value from (9), for each of the four state components (the four points in each group). Here 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 standard error.
| Model | , rad |
| Observations | ; , |
| Constant bias (Sections 4.2.1 and 4.2.2) | |
| Linear bias (Section 4.3.1) | , , |
| Nonlinear bias (Section 4.3.1) | (elementwise), |
| EnKF (Section 4.2.1) | stochastic, , anomaly inflation 1.02 |
| Run lengths | Sections 4.2.1 and 4.3.1: , 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. is the rotation matrix by angle , and the cyclic shift, with .
Iterated correction
We verify the iterated correction convergence (§ 2.5) numerically, with the KF on the same system. At each iteration , the filter is run with the forecast model with per-step bias , the bias estimate is computed from the time-averaged analysis increment, and subtracted from the per-step bias to give . Figure 3 shows both the full-state bias norm and the innovation bias norm 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 . Orange squares: innovation bias norm . 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 drops sharply on the first iteration then plateaus at , in agreement with the limit of Corollary 1. This residual lies in , and it cannot be removed by any correction. By contrast, the innovation bias 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 . 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),
with , , and periodic boundary conditions, integrated with a fourth-order Runge–Kutta scheme. Every other variable is observed (), 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 plus a small random part, whose largest component, expressed as a tendency, is about 15% of the forcing . The parameter values are listed in Table 2 (Appendix § 9).
| Model | , ; RK4 with |
| Observations | (); , |
| Constant bias (Section 4.2.3) | , with drawn once |
| State-dependent bias (Section 4.3.2) | (elementwise) |
| EnKF | stochastic, covariance inflation 1.04, Gaspari–Cohn localisation of half-width 5; Section 4.2.3: ; Section 4.3.2: |
| Run lengths | 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 ( 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 , 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 decreases geometrically, and approaches the level of sampling noise. The full-state bias norm 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 . Orange squares: innovation bias norm . The shading along the curves shows 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 , with and given in Table 1. The theoretical fixed point is given by (21).
Nonlinear bias. We take 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 at every cycle. Its conditional mean given the forecast is estimated by fitting a network of the same form to 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 , the systematic increment that the filter would make with a perfect model. Iterating the correction as in Section 3.5 brings it closer to (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 in the observed components, for individual cycles (points) and averaged in bins of (squares). The solid line is the theory, the conditional mean of given the forecast, and the dashed line is . (b) Same as (a), but after three iterations of the correction. (c) Norm of the analysis bias in the observed components, , 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, , , , and . Recall that , so that exists, and that and are positive definite. For symmetric matrices, means that is a positive semidefinite. Proposition 2 uses the following identities, in which .
Lemma 1 (Key identities).
Proof. Since , we have , that is, . Hence
Since , this equals , using . The second identity is Anderson and Moore (1979, Section 4.5, Problem 5.1) at . □
Proof of Proposition 2. From (12), . By the second identity in (27), , so satisfies , and hence , that is, . By the first identity,
Moreover, . If is in , so that for some , then , since . Then, . Hence, for every nonzero ,
that is,
Taking gives the result, with strict inequality whenever . □
Proof of Proposition 3. By (28), shrinks the -norm of every nonzero vector in , which it maps into itself, since . Since is finite-dimensional, shrinks all such vectors by at least a common factor :
for all . Hence geometrically for every , and in particular decreases monotonically to zero. If has full row rank, , so
For an eigenvalue of with eigenvector ,
so , and . □
Proof of Corollary 1. Each iteration subtracts from the per-step bias, by (15). Since , on which geometrically (Proposition 3), the Neumann series gives , so and . Since , we have . Hence satisfies , and , since when . □
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]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]Anderson, B.D.O. and Moore, J.B. (1979) Optimal Filtering. Englewood Cliffs, NJ: Prentice-Hall.
- [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]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]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]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]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]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
- [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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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




