Numerical Method
Matrix exponential evaluation algorithms compute exact discrete-time process noise covariance and state transition matrices for linear dynamic systems. Published by Charles Van Loan, van loan integration solves the continuous matrix differential equations of state estimation by constructing an augmented block matrix of dynamic models and continuous spectral densities. Exponentiating this composite block matrix yields the discrete-time state transition matrix alongside the discrete process noise covariance in a single numerical step.
Navigation filters rely on this method to integrate continuous kinematic models into discrete kalman filter algorithms without deriving analytical integrals for complex noise formulations. The technique applies equally to time-invariant system dynamics of arbitrary dimension.
Matrix Mechanics
Implementation begins by assembling a block matrix twice the size of the system state, placing the system dynamics matrix in the upper left, the continuous noise spectral density scaled by the input coupling matrix in the upper right, zeros in the lower left, and the negative transpose of the dynamics matrix in the lower right. Applying matrix exponentiation to this block multiplied by the time step produces a partitioned result from which the discrete matrices emerge directly. In van loan integration, the upper right block of the exponentiated matrix represents the product of the state transition matrix and the discrete noise covariance matrix.
Multiplying this block by the transpose of the state transition matrix extracts the discrete process noise covariance matrix. Padé approximations with scaling and squaring provide the standard numerical engine for the core matrix exponential step.
Computational Verification
Algorithm validation checks the resulting covariance matrix for symmetry and positive definiteness across extreme integration intervals. Metrology engineers verify integration integrity by comparing output matrices against analytical solutions derived from low-order kinematic models. High-order vehicle dynamics models, such as integrated inertial navigation with flexible structural modes, validate numerical stability under varying sample rates.
Proper integration avoids numerical instability in downstream filter iterations.
Numerical Constraint
Computational costs scale with the cube of the state vector dimension due to the doubled size of the augmented block matrix. Real-time microcontrollers executing high-rate navigation updates experience significant execution overhead when running large block matrix exponentials. When integration time intervals become large relative to the highest system natural frequencies, scaling and squaring procedures encounter floating-point roundoff issues that degrade matrix precision.
Extremely stiff dynamic systems containing widely separated time constants require dedicated matrix conditioning to avoid catastrophic loss of precision during matrix inversion steps.