Data Assimilation

The Central Idea

Suppose a system has an unknown state

at time step \(k\).

A dynamical model predicts how that state evolves:

where \(\mathbf{w}_k\) represents model or process uncertainty.

Observations satisfy

where \(\mathbf{v}_k\) represents measurement error.

Data assimilation repeatedly combines these two sources of information.

Forecast and Analysis

Most data-assimilation methods alternate between two stages:

1. Forecast: propagate the current estimate through the model.
2. Analysis: update the forecast using new observations.

This creates the recurring cycle

The Forecast State

Let

denote the analyzed state at time \(k-1\).

The model produces the forecast

The superscript \(f\) denotes forecast.

The Analysis State

Once observation

arrives, the forecast is updated to produce

The superscript \(a\) denotes analysis.

The analysis combines model prediction and observational information.

The Innovation

The difference between the observation and the predicted observation is called the innovation:

The innovation measures the new information supplied by the observation relative to the forecast.

Why the Innovation Matters

If the forecast already predicts the observation accurately, the innovation is small.

If the forecast and observation disagree substantially, the innovation is large.

But the size of the correction should depend not only on the innovation but also on the relative uncertainty of the forecast and measurement.

Uncertainty Is Central

Data assimilation does not merely average model and data.

It weights them according to uncertainty.

If observations are highly precise, the analysis should move strongly toward the data.

If observations are noisy but the model forecast is reliable, the analysis should remain closer to the forecast.

Scalar Weighted Update

Suppose a scalar forecast is

with variance

and the measurement is

with variance

A Bayesian Gaussian update has the form

where \(K\) is a gain.

The Scalar Kalman Gain

The gain is

If forecast uncertainty \(P^f\) is large relative to measurement uncertainty \(R\), then \(K\) is close to \(1\).

If \(R\) is large relative to \(P^f\), then \(K\) is close to \(0\).

Interpreting the Gain

The update

can be read as

The gain determines how strongly the new data alter the forecast.

Analysis Variance in the Scalar Case

The analysis variance is

Assimilating an informative observation generally reduces uncertainty.

A Numerical Scalar Example

Suppose

while the observation is

Then

The analysis becomes

The Observation Operator

Observations may not measure the entire state directly.

A linear observation model is

The matrix \(H_k\) maps the full model state into the quantities observed by the instrument.

Partial Observation

Suppose

but only \(x_2\) is measured.

Then

The predicted observation is

State Covariance

Uncertainty in a vector state is represented by a covariance matrix

Diagonal entries measure uncertainty in individual state components.

Off-diagonal entries measure how state errors vary together.

Why State Correlation Matters

If unobserved state variable \(x_1\) is strongly correlated with observed variable \(x_2\), then measuring \(x_2\) can provide information about \(x_1\).

Data assimilation can therefore update unobserved state components through covariance relationships.

Linear State-Space Model

A classical linear system is

with observations

Assume

and

The Kalman Filter

For linear Gaussian state-space models, the Kalman filter provides an exact recursive Bayesian update of the state mean and covariance.

It alternates between:

1. forecasting the mean and covariance;
2. assimilating the new measurement.

Kalman Forecast Mean

If

is the previous analysis mean, then

Kalman Forecast Covariance

The forecast covariance is

The first term propagates existing uncertainty.

The second term adds new process uncertainty.

The Innovation Covariance

The predicted observation uncertainty is

This combines uncertainty in the forecasted observation with measurement uncertainty.

The Matrix Kalman Gain

The Kalman gain is

It determines how observation-space discrepancies are mapped back into state corrections.

The Kalman Analysis Mean

The state update is

This is the vector version of forecast plus gain times innovation.

The Kalman Analysis Covariance

One common form is

More numerically robust covariance-update formulas are often used in practice, but this expression captures the central structure.

Process Noise

The covariance matrix

represents uncertainty in the dynamical model itself.

Sources may include:

• unmodeled forcing;
• unresolved small-scale dynamics;
• uncertain inputs;
• numerical approximations;
• genuine stochastic evolution.

Measurement Noise

The covariance matrix

represents uncertainty in the observations.

It may include:

• sensor precision;
• calibration error;
• sampling noise;
• instrument resolution;
• observation preprocessing.

Why \(Q\) and \(R\) Must Be Distinguished

If uncertainty is incorrectly assigned to the model or the measurements, the filter can become overconfident or overly reactive.

Large \(Q\) tells the filter to distrust long model forecasts.

Large \(R\) tells the filter to distrust observations.

Filter Overconfidence

If the forecast covariance is underestimated, the analysis may give too little weight to observations.

The filter can then become overly confident in an inaccurate model trajectory.

Observation Overweighting

If measurement uncertainty is underestimated, noisy observations can produce excessively large corrections.

The resulting state estimate may track noise instead of the underlying system.

The Bayesian Interpretation

The forecast distribution acts as a prior:

The new observation contributes a likelihood:

Bayes' rule produces the analysis posterior:

Filtering

Filtering estimates the current state using observations up through the current time:

It is designed for real-time or sequential estimation.

Prediction

Prediction estimates a future state using observations only up to the present:

Prediction uncertainty generally grows as the forecast horizon increases.

Smoothing

Smoothing estimates a past state using observations that may occur after that state:

Future data can improve estimates of earlier states.

Filtering Versus Smoothing

Filtering is causal and suitable for online estimation.

Smoothing is retrospective and can use future observations to refine past estimates.

Smoothing is common in post-processing and scientific reconstruction.

Nonlinear Systems

Many real systems obey

and

where \(f\) and \(h\) are nonlinear.

The standard Kalman filter is no longer exact.

Extended Kalman Filter

The extended Kalman filter, or EKF, linearizes nonlinear models around the current estimate.

The state-transition Jacobian is

and the observation Jacobian is

EKF Covariance Propagation

The covariance is propagated using the local linear approximation:

The remaining update then resembles the linear Kalman filter.

Limitations of Linearization

If the system is strongly nonlinear or uncertainty is broad, a local linearization may poorly approximate the true probability distribution.

The EKF can then become biased, overconfident, or unstable.

Unscented Kalman Filter Idea

The unscented Kalman filter represents uncertainty using a carefully chosen set of sample points.

These points are propagated through the nonlinear model directly.

The resulting transformed mean and covariance can approximate nonlinear effects without explicit Jacobian derivatives.

Ensemble Kalman Filter

The ensemble Kalman filter, or EnKF, represents state uncertainty using an ensemble:

Each ensemble member is propagated through the model.

Ensemble Mean

The ensemble mean is

It approximates the state mean.

Ensemble Covariance

The sample covariance is

The ensemble therefore provides both a state estimate and an uncertainty estimate.

Why Ensemble Methods Are Useful

Large dynamical systems may have millions of state variables.

Storing and propagating a full covariance matrix can be impossible.

Ensemble methods approximate covariance using a manageable collection of model states.

Sampling Error in Ensembles

A finite ensemble gives only an approximate covariance.

If \(N\) is small relative to state dimension, sample correlations can be noisy or spurious.

This is one of the major practical challenges of ensemble assimilation.

Covariance Inflation

Ensemble spread may become too small.

Covariance inflation deliberately increases ensemble deviations from the mean:

with

This helps prevent overconfidence.

Localization

In high-dimensional systems, small ensembles can create unrealistic long-range correlations.

Localization reduces or removes covariance relationships between state variables and observations that are too far apart or otherwise implausible.

Particle Filters

Particle filters represent the posterior using weighted samples:

Unlike Kalman-type methods, particle filters do not require the posterior to be approximately Gaussian.

Particle Weighting

After an observation arrives, particle weights are updated according to the likelihood:

Particles that better explain the observation receive more weight.

Weight Degeneracy

After repeated updates, a few particles may carry nearly all the weight.

This is called weight degeneracy.

Effective sample size decreases even though the nominal particle count remains unchanged.

Resampling

Particle filters often resample the ensemble so that high-weight particles are replicated and very low-weight particles are discarded.

Resampling combats weight degeneracy but can reduce diversity.

The Curse of Dimensionality

Particle filters can become extremely expensive in high-dimensional systems.

As state dimension grows, a very large number of particles may be required to represent the posterior adequately.

Variational Data Assimilation

Data assimilation can also be formulated as an optimization problem.

A typical objective combines:

• departure from a background state;
• misfit to observations;
• dynamical-model constraints.

A Simple Variational Objective

Let

be a background state with covariance \(B\).

One objective is

Three-Dimensional Variational Assimilation

In 3D-Var, observations within an assimilation window are used to find an analyzed state that balances background and observation information.

The model dynamics do not explicitly connect every observation time inside the optimization window.

Four-Dimensional Variational Assimilation

In 4D-Var, the dynamical model connects observations across time.

An initial state is adjusted so that its model trajectory fits observations throughout an assimilation window.

Why 4D-Var Is an Inverse Problem

The unknown may be the initial state

The forward model maps this initial state to predicted observations over time.

The optimization then infers the initial condition from the entire observation sequence.

Adjoint Methods

In large variational problems, computing gradients with respect to millions of state variables can be expensive.

Adjoint methods efficiently propagate sensitivity information backward through the model.

They are central to many large-scale 4D-Var systems.

State Estimation Versus Parameter Estimation

Data assimilation is often introduced as state estimation.

But unknown parameters can also be inferred sequentially.

One strategy is to augment the state:

State Augmentation

If parameters are assumed constant,

The augmented state can then be propagated and updated by the assimilation method.

Cross-covariance between states and parameters allows observations to modify parameter estimates.

Slowly Varying Parameters

A parameter may drift over time:

The process noise

represents allowed parameter evolution.

Identifiability in Sequential Estimation

Sequential data do not automatically make parameters identifiable.

Parameters can be estimated only if their effects become visible in the observed state trajectory.

Persistent parameter-state correlation alone is not enough if the measured outputs cannot distinguish competing parameter values.

Observability

Observability asks whether the internal state of a dynamical system can be inferred from its observations over time.

It is the dynamic counterpart of identifiability.

A system may contain states that never affect measured outputs and are therefore unobservable.

Linear Observability

For a linear time-invariant system

the observability matrix is

Full rank indicates that the full state can, in principle, be reconstructed from a sufficiently informative observation sequence.

Assimilation Frequency

Observations may arrive:

• continuously;
• at fixed intervals;
• irregularly;
• from multiple asynchronous sensors.

Long intervals between updates allow model uncertainty to grow.

Observation Timing Matters

Measurements are most informative when taken where system behavior is sensitive to uncertain states or parameters.

Poorly timed observations can add little information even when they are precise.

Multiple Observation Types

Different sensors may measure different functions of the state:

Combining complementary observations can improve observability and parameter identifiability.

Asynchronous Data

Real observation systems may report data at different times.

Assimilation frameworks can update whenever new measurements become available, provided the model state is propagated to the correct observation time.

Missing Observations

If no measurement is available at a particular time, the system can continue forecasting using the dynamical model.

Uncertainty generally grows until observations resume.

Outliers in Sequential Assimilation

A single extreme measurement can create a large innovation.

Before applying a large state correction, the observation should be checked against expected measurement uncertainty and physical plausibility.

Innovation Statistics

Under a correctly specified filter,

should have covariance approximately

Innovation diagnostics can therefore reveal misestimated uncertainty.

Normalized Innovations

A scalar normalized innovation is

Persistent values much larger than expected suggest underestimated error, outliers, or model misspecification.

Filter Divergence

A filter diverges when its estimated state becomes increasingly inconsistent with the true system and new observations.

Causes include:

• underestimated covariance;
• model bias;
• incorrect observation operators;
• strong nonlinearity;
• numerical instability.

Model Bias

If the dynamical model has persistent bias, repeated assimilation may correct the same error at every observation time.

The innovation sequence may then exhibit a nonzero mean.

Bias correction or model improvement may be necessary.

Assimilation and Regularization

Data assimilation is naturally regularized by forecast uncertainty and model structure.

The analysis does not fit observations arbitrarily; it balances them against the prior forecast and its covariance.

Assimilation and Bayesian Estimation

Sequential Bayesian updating is the conceptual foundation of many data-assimilation methods.

The forecast distribution is the prior, the observation model supplies the likelihood, and the analysis is the posterior.

Assimilation and Least Squares

For Gaussian errors, many assimilation updates can also be derived by minimizing a weighted quadratic objective.

For example,

The minimizer is the Gaussian analysis state.

Assimilation and Inverse Problems

Each observation update is a small inverse problem.

The unknown state is inferred from:

• the forecast model;
• the observation operator;
• measurement uncertainty;
• forecast uncertainty.

Sequential assimilation solves these inverse problems repeatedly over time.

Real-Time Applications

Data assimilation is central to:

• weather forecasting;
• ocean modeling;
• navigation;
• tracking;
• industrial monitoring;
• robotics;
• energy systems;
• biological state estimation.

Offline Reconstruction

Assimilation methods are also used after all observations have been collected.

Smoothing and variational methods can reconstruct an entire trajectory from a complete historical data set.

Computational Cost

Large-scale data assimilation can be expensive because it may require:

• many forward model runs;
• large ensembles;
• matrix factorizations;
• adjoint calculations;
• repeated optimization over time windows.

Parallelism

Ensemble members and particles can often be propagated independently.

This makes ensemble and particle methods naturally suitable for parallel computation.

Model Reduction

Reduced-order models can lower assimilation cost by representing the state in a lower-dimensional basis:

where

contains fewer coordinates than the full state.

Common Error: Treating Model and Data as Equally Reliable

Data assimilation should not average model and observations without regard to uncertainty.

Relative covariance determines how strongly each source influences the analysis.

Common Error: Ignoring the Observation Operator

A sensor rarely measures the complete model state directly.

The transformation from state to observation space must be modeled explicitly.

Common Error: Confusing Process and Measurement Noise

Process noise describes uncertainty in system evolution.

Measurement noise describes uncertainty in observing the state.

Assigning one source to the other changes the filter behavior.

Common Error: Underestimating Covariance

An overconfident forecast or observation model can cause the assimilation system to ignore useful information or overreact to noise.

Covariance calibration is therefore a central practical task.

Common Error: Assuming Sequential Data Guarantee Identifiability

Repeated observations help only when parameter or state effects are actually visible in the measured outputs.

Unobservable directions remain unresolved regardless of the number of update steps.

Common Error: Ignoring Ensemble Sampling Error

A small ensemble can produce noisy covariance estimates and spurious correlations.

Inflation, localization, or larger ensembles may be needed.

Common Error: Trusting a Filter Without Innovation Diagnostics

A filter can produce smooth-looking estimates while using inconsistent uncertainty assumptions.

Innovation statistics help test whether forecast and measurement covariances are plausible.

A Reliable Data-Assimilation Workflow

1. Define the dynamical state model.
2. Define the observation operator.
3. Separate process noise from measurement noise.
4. Specify or estimate \(Q\) and \(R\).
5. Initialize the state estimate and covariance.
6. Forecast the state forward.
7. Forecast the uncertainty forward.
8. Compute predicted observations and innovations.
9. Update the state using the appropriate gain or posterior method.
10. Update uncertainty.
11. Inspect innovation statistics and residual structure.
12. Check observability and parameter identifiability.
13. Use inflation, localization, or robust methods when needed.
14. Validate trajectories and uncertainty against withheld or synthetic data.

Practice 1: Innovation

A scalar forecast predicts

and the observation is

Compute the innovation if the observation operator is the identity.

Practice 2: Scalar Kalman Gain

Suppose

Compute

Practice 3: Scalar Analysis

Using

compute the analysis state.

Practice 4: Forecast Covariance

In a scalar system,

Compute

Practice 5: Observation Operator

If

and only \(x_3\) is observed, write a linear observation matrix \(H\).

Practice 6: Process Noise

What does the covariance matrix \(Q\) represent in a state-space model?

Practice 7: Measurement Noise

What happens qualitatively to the Kalman gain when measurement variance \(R\) becomes very large?

Practice 8: Filtering Versus Smoothing

Which method can use observations from future times to improve an estimate of a past state?

Practice 9: Ensemble Covariance

Why can a small ensemble create spurious correlations?

Practice 10: Parameter Estimation

How can an unknown constant parameter be included in a sequential state-estimation method?

Practice Answers

Practice 1

Practice 2

Practice 3

Practice 4

Practice 5

Practice 6

\(Q\) represents uncertainty in the system evolution, including unmodeled dynamics, uncertain forcing, stochastic variation, and other process errors.

Practice 7

As \(R\) becomes very large, the observation is treated as unreliable and the Kalman gain decreases toward zero. The analysis remains closer to the forecast.

Practice 8

Smoothing can use future observations to improve estimates of earlier states.

Practice 9

A small ensemble provides a noisy sample estimate of covariance. Random finite-sample relationships can appear between variables even when the true correlation is weak or zero.

Practice 10

Augment the state vector:

The parameter can then be updated through its covariance with observed state behavior.

What to Remember

Data assimilation combines dynamical forecasts and observations to estimate an evolving state and its uncertainty.

• The forecast step propagates the current estimate through the model.
• The analysis step incorporates new observations.
• The innovation is the difference between observed and predicted measurements.
• The Kalman gain balances forecast uncertainty against measurement uncertainty.
• State covariance allows observations to update unmeasured but correlated variables.
• \(Q\) represents process uncertainty; \(R\) represents measurement uncertainty.
• The Kalman filter is exact for linear Gaussian state-space models.
• EKF, unscented, ensemble, particle, and variational methods extend the basic idea.
• Filtering estimates the current state; smoothing can revise past states using future data.
• Parameters can be estimated sequentially through state augmentation.
• Observability and identifiability remain essential.
• Innovation diagnostics help test whether uncertainty assumptions are realistic.

The central principle is:

Where We Go Next

Data assimilation uses observations to update states and parameters in evolving systems. The same inverse-problem logic also appears when the unknown is a spatial field or image reconstructed from indirect measurements.

The next lesson, Imaging and Reconstruction, develops that setting through forward operators, projection data, blur, deconvolution, regularization, resolution, artifacts, and computational reconstruction.