Dynamic Linear Systems Model

• Dynamic linear systems are mathematical formalisms that use linear transformations and noise models to describe system state evolution.
• They underpin techniques such as Kalman filtering, EM, and Bayesian inference, offering efficient solutions for time series analysis and system identification.
• Advanced models extend to sparse graphs and functional dynamics, improving interpretability and robustness in a variety of applications.

A dynamic linear systems model is a mathematical formalism for describing the evolution of a system’s state when both the dynamics and the observation processes are linear, typically in discrete or continuous time. These models provide a rigorous foundation for time series analysis, control, system identification, network inference, and probabilistic modeling, often leveraging the state-space representation and exploiting conditional independence structure for efficient estimation. They are central across engineering, statistics, and machine learning, underpinning Kalman filtering, dynamic mode decomposition, recurrent neural architectures, Bayesian time series models, and graphical inference schemes.

1. State-Space Formulation and Classes of Dynamic Linear Models

A dynamic linear system is specified by latent state variables and observed outputs, evolving via linear maps with Gaussian or more general noise. The canonical discrete-time state-space model is:

$$x_{t+1}=A\,x_t+B\,u_t+w_t,\quad w_t\sim \mathcal N(0,Q)$$

$y_t=H\,x_t+v_t,\quad v_t\sim \mathcal N(0,R)$

Here, $x_t\in\mathbb R^n$ is the state, $u_t$ is a control/input, $y_t$ is observation, $A$ and $H$ are system/observation matrices, $Q$ and $R$ are noise covariances. Extensions include time-varying parameters and functional data models, as in the Bayesian Multivariate Functional Dynamic Linear Model (MFDLM), where the observation process has high-dimensional and smooth functional structure via spline bases (Kowal et al., 2014).

Classical dynamic linear models (DLMs) generalize ARMA and time-varying regression as special cases, with the underlying dynamic parameter vector $\theta_t$ evolving by a possibly time-dependent $G_t$ and driven by system/observation noises. In continuous time, dynamics may be given by linear stochastic differential equations:

$dx(t) = A x(t) dt + dw(t)$

$y_j = x(t_j) + v_j$

with $w(t)$ a Brownian process (Aalto et al., 2018).

Dimension-varying linear systems have been modeled on quotient spaces, allowing smooth transitions between linear systems of different state dimensions by defining cross-dimensional pseudo-distance and equivalence classes on states and matrices (see quotient space $\Omega$ and semigroup $\Xi$) (Cheng et al., 2018).

2. System Identification, Learning, and Inference Methods

Several inference strategies have been developed for dynamic linear systems:

• Kalman Filtering efficiently computes filtered/posterior state estimates using forward recursions, exploiting the Gaussian–Markov structure (Laine, 2019).
• Expectation–Maximization (EM) estimates unknown parameters by alternating between Kalman smoothing and M-step updates (e.g., for $A$ and covariances), though convergence and local optimals remain concerns (Jadbabaie et al., 2021).
• Covariance–Method Estimator gives closed-form solutions for hidden linear dynamics when stability holds, by solving two OLS equations for sample moment identities (regarding stationary covariances and cross-covariances) (Jadbabaie et al., 2021).
• Bayesian Variable Selection enables network inference from low-sampling time series via a sparsity-promoting prior on $A$, sampling both topology and continuous trajectories with MCMC—in particular, handling underdetermined inverse problems and non-parametricity (Aalto et al., 2018).
• Dynamic Graphical Lasso (DGLASSO) jointly models static graphical structure on process innovations (precision matrix $P$) and dynamic causal structure in the transition matrix ($A$), optimizing a penalized likelihood via block-alternating majorization-minimization and providing convex updates for sparse $A$ and $P$ (Chouzenoux et al., 2023).
• Model-Embedded Gaussian Process Regression (ME–GPR) for joint inference of ODE parameters and solution trajectories, embedding the linear system as a constraint on the GP prior covariance (Zhou et al., 2024).

System identification based on observed data matrices is exemplified by Dynamic Mode Decomposition (DMD). For discrete LTI systems $x_{k+1} = A x_k$, DMD finds $A$ by projecting future snapshot matrices onto past, exploiting SVD; for full-rank state snapshots, recovery is exact and invariant to similarity transforms (Heiland et al., 2021).

3. Advanced Modeling: Sparsity, Nonparametricity, Switching, and Functional Extensions

• Nonparametric Bayesian Sparse Graph Linear Dynamical Systems (SGLDS) induce infinite-dimensional, sparsity-promoting structures on the state transitions using Bernoulli–Poisson and gamma processes; states are classified (non-dynamic, dynamic: live/absorbing/noise-injection) according to graph affinity patterns, with normal–gamma shrinkage for non-dynamic components (Kalantari et al., 2018).
• Tree-Structured Recurrent Switching Linear Dynamical Systems (TrSLDS) approximate nonlinear time series by hierarchies of locally linear models, where the switching mechanism is a tree-based stick-breaking process. Local dynamics are governed by path-inherited matrices and biases; full Bayesian inference is implemented with Polya–Gamma augmentation yielding conjugate Gibbs updates (Nassar et al., 2018).
• Functional Dynamic Linear Models represent each observational data as a smooth function (e.g., yield curves, time–frequency spectra) using orthonormal spline bases, with factor dynamics governed by time series models (AR(1), random walk, VAR) and multivariate structure enforced via block-diagonal basis matrices. Identification and inference use Gibbs sampling with conditional conjugacy and constrained spline parameters (Kowal et al., 2014).

4. Optimization, Stability, Uncertainty, and Learning Guarantees

• Stochastic Optimization under Parametric Uncertainty treats $A$ and $B$ as random matrices described by higher-order tensors (means, covariances, cross-covariances), propagating parameter uncertainty through extended Kalman-type recursions, optimal feedback (via Riccati–tensor equations), and quantifying the precautionary control cost (0909.2542).
• Learning Guarantees for Linear RNNs: For stable LTI systems ($\rho(C)<1$), gradient descent on linear RNN architectures provably identifies the dynamic system in polynomial time/sample complexity, independent of horizon length, provided sufficient hidden width. The training loss landscape locally possesses NTK-like nearly-linear structure for overparameterized networks, supporting convergence to near-optimality (Wang et al., 2022).
• Stability and Spectral Radius: Lyapunov stability (spectral radius $< 1$) is central in ensuring model identifiability, exponential forgetting, mixing, and controlling error bounds in estimation (Wang et al., 2022, Jadbabaie et al., 2021).

5. Empirical Applications and Performance Benchmarks

Dynamic linear systems have been employed and benchmarked in diverse domains:

• Gene regulatory network inference, leveraging Bayesian variable selection in ill-posed, low-frequency settings, yielding high AUROC/AUPREC compared to EM–Lasso (Aalto et al., 2018).
• Multi-economy yield curves and brain-local field potentials, modeled via functional DLMs for smooth, interpretable, multivariate dependencies with posterior credible intervals and strong Gibbs mixing (Kowal et al., 2014).
• Hierarchical multi-scale modeling of nonlinear oscillators (FitzHugh–Nagumo, Lorenz, V1 spike trains) using TrSLDS, where tree depth enables trade-off between interpretability and predictive accuracy (Nassar et al., 2018).
• Sparse network and causality recovery in Kalman-based LG–SSMs using DGLASSO, with enhanced precision, recall, and accuracy over classical EM and Granger-baseline methods (Chouzenoux et al., 2023).
• System identification (Dynamic Mode Decomposition) for high-dimensional, data-driven models, with recovery guarantees and invariance properties under basis transformations (Heiland et al., 2021).
• Stochastic optimal control tasks (interception/landing) under tensor-based parametric uncertainty, demonstrating improved robustness and learning–precision phase separation over classical approaches (0909.2542).

6. Interpretability, Generalization, and Extensions

Model interpretability is advanced via transparent graph priors, functional constraints, and hierarchical switching structures (as in TrSLDS stick-breaking trees). Functional DLMs provide smooth decomposition into factor curves and time-varying covariate effects. Graphical approaches such as DGLASSO bridge static dependence and dynamic causality—enabling generalized inference for nonstationary systems and multivariate time series.

Nonparametricity is achieved in SGLDS via infinite-dimensional gamma process priors, allowing dynamic complexity scaling and sparse representation. Stability assumptions, high-dimensional sparse inference, and tensor formalism facilitate generalization to irregular sampling, missing data, and heteroskedastic errors. Model-embedded GP regression (ME–GPR) offers a unified “one-step” learning and uncertainty quantification for parameter inference in linear DS, enforcing exact physical constraints in the prior (Zhou et al., 2024).

7. Summary Table: Representative Model Classes

Model Type	Key Structure	Inference Algorithm
State-space DLM	Linear dynamical plus obs noise	Kalman filter, EM, MCMC (Laine, 2019)
Bayesian Sparse Graph	Bernoulli–Poisson, gamma proc.	MCMC, shrinkage (non-parametric) (Kalantari et al., 2018)
Tree-switching LDS	Hierarchical local dynamics	Polya–Gamma augmented Gibbs (Nassar et al., 2018)
Functional DLM	Spline basis, multivariate AR	Gibbs sampler, constrained splines (Kowal et al., 2014)
DGLASSO	Graphical Lasso + Granger	Block-Alt MM, logpenalty–$\ell_1$(Chouzenoux et al., 2023)
Dynamic Mode Decomp.	Data-driven linear algebra	SVD, operator projection (Heiland et al., 2021)

All dynamic linear systems models exploit the tractability of linear structure but differ in their approach to complexity control (sparsity, nonparametricity), inferential efficiency, functional representation, and integration of domain knowledge. The field continually evolves to address scalability, uncertainty quantification, multiscale phenomena, and interpretability in increasingly high-dimensional, noisy, and partially observed settings.