Switching Linear Dynamical Systems

• Switching Linear Dynamical Systems (SLDS) are probabilistic models combining discrete Markovian regime changes with continuous linear dynamics to capture non-stationary behaviors.
• SLDS employs advanced inference techniques like block Gibbs sampling and Polya–Gamma augmentation to manage the complexity of latent state trajectories.
• Extensions such as rSLDS and REDSLDS enhance segmentation quality and interpretability by integrating explicit-duration and state-dependent transitions.

A Switching Linear Dynamical System (SLDS) is a probabilistic generative model for time series in which the system dynamics are governed by both a discrete latent regime (switch) process and continuous latent linear dynamical processes, with both components influencing the observations in structured ways. The regime sequence typically evolves as a Markov process, modulating the local parameters of the linear-Gaussian state space model to capture heterogeneous or non-stationary periods in multivariate sequence data. SLDS and their extensions—including recurrent, explicit-duration, tree-structured, and nonparametric variants—are widely used in time series analysis, signal segmentation, neuroscience, robotics, and beyond for their statistical expressivity in modeling complex, piecewise-linear or regime-switching dynamical phenomena.

1. Generative Structure of Switching Linear Dynamical Systems

At each time step $t$, the SLDS comprises:

• A discrete latent variable $z_t \in \{1,\ldots, K\}$ (“regime” or “mode”), typically evolving via a Markov chain with transition matrix $\pi$: $p(z_t \mid z_{t-1}) = \pi_{z_{t-1},z_t}$.
• A continuous latent state $x_t \in \mathbb{R}^M$, whose dynamics are modulated by the current regime:

$$x_t = A_{z_t} x_{t-1} + a_{z_t} + w_t, \quad w_t \sim \mathcal{N}(0, Q_{z_t}).$$

• An observation model, often linear-Gaussian and regime-dependent:

$$y_t = C_{z_t} x_t + c_{z_t} + f_t, \quad f_t \sim \mathcal{N}(0, S_{z_t}).$$

• The joint likelihood factorizes as:

$$p(y_{1:T}, x_{1:T}, z_{1:T}) = p(z_1) p(x_1|z_1) p(y_1|x_1, z_1) \prod_{t=2}^T p(z_t|z_{t-1}) p(x_t|x_{t-1},z_t) p(y_t|x_t,z_t).$$

The above structure enables the SLDS to capture abrupt changes between locally linear regimes as commonly observed in physical, biological, and engineered systems (Nassar et al., 2018, Słupiński et al., 2024).

2. Advanced Extensions: rSLDS and REDSLDS

The recurrent Switching Linear Dynamical System (rSLDS) introduces dependence of the switching probabilities on the continuous state, typically through a stick-breaking logistic regression: $p(z_t = k | z_{t-1}, x_{t-1}) = \pi_{SB}(v_t^S)_k,$ where $v_t^S = R^{S}_{z_{t-1}} x_{t-1} + r^{S}_{z_{t-1}}$ and the stick-breaking transform constructs probabilities for $K$ regimes.

The Recurrent Explicit Duration SLDS (REDSLDS) further augments the model with an explicit duration variable $d_t \in \{1,\ldots, D_{\text{max}}\}$ at each step, supporting state-dependent non-geometric sojourn times. The full joint probability becomes: $$\begin{align*} p(y_{1:T}, x_{1:T}, z_{1:T}, d_{1:T}) =\ &p(z_1)\,p(d_1|z_1)\,p(x_1|z_1)\,p(y_1|x_1,z_1) \ &\times \prod_{t=2}^T p(z_t|z_{t-1},d_{t-1},x_{t-1})\,p(d_t|z_t,d_{t-1},x_{t-1})\,p(x_t|x_{t-1},z_t)\,p(y_t|x_t,z_t). \end{align*}$$ The explicit duration mechanism prevents unrealistically rapid switching, improves segmentation quality, and enables temporally coherent state sequences (Słupiński et al., 2024).

3. Inference Methodologies: Block Gibbs and Polya–Gamma Augmentation

Inference in SLDS and its generalizations is typically intractable for exact closed-form computation due to the exponential growth in discrete-state trajectories. Instead, the following structured approaches are used:

• Block Gibbs Sampling: Alternately sample blocks (discrete states, durations, continuous states, Polya–Gamma variables, model parameters) conditionally using analytic posteriors. Conditioning on the discrete latent trajectory and durations, the continuous latent states $x_{1:T}$ follow a linear-Gaussian state-space model amenable to Kalman smoothing.
• Polya–Gamma Augmentation: To facilitate efficient inference for models with logistic or multinomial links (state-dependent transitions, explicit durations), Polya–Gamma random variables linearize the logistic terms, yielding conditionally conjugate (Gaussian) updates:

$$\frac{e^{a\psi}}{(1+e^\psi)^b} = 2^{-b}e^{\kappa\psi} \int_0^\infty e^{-\frac12 \omega \psi^2} p_{\mathrm{PG}}(\omega|b,0) d\omega, \quad \kappa = a - b/2.$$

This transformation renders the likelihood quadratic in the regression weights and continuous state, allowing for tractable block sampling of all parameters and latent variables (Słupiński et al., 2024, Nassar et al., 2018).

• Forward–Backward Recursions for (z, d): In models with explicit durations, the joint discrete process $s_t = (z_t, d_t)$ is handled using standard forward–backward inference, suitably adapted to the combinatorially expanded state space.
• Conjugate Analytical Updates: For model parameters (e.g., transition matrices, dynamics, emission parameters), conjugate priors such as Dirichlet (for transitions) and Matrix-Normal–Inverse-Wishart (for dynamics) enable efficient closed-form updates.

4. Empirical Performance and Segmentation Quality

Experimental results on diverse benchmarks demonstrate the segmentation and predictive advantages of explicit-duration and recurrent extensions.

• On the simulated NASCAR® task, REDSLDS attains higher segmentation accuracy (≈0.65) and weighted F₁-score (≈0.68) relative to the baseline rSLDS (accuracy ≈0.48, F₁ ≈0.49). Model log-likelihood is similarly improved (log L ≈9.04×10⁴ vs. 9.13×10⁴).
• In honey-bee waggle-dance segmentation, REDSLDS achieves accuracy and weighted F₁-score ≈0.85, whereas rSLDS yields ≈0.37 and ≈0.40, respectively.
• On high-dimensional BehaveNet mouse video embeddings, REDSLDS recovers more persistent, interpretable regime partitions, while rSLDS tends to collapse to degenerate solutions or a single state.

These results uniformly indicate that the explicit-duration mechanism substantially enhances temporal coherence and prevents unrealistic switching artifacts (Słupiński et al., 2024).

5. Explicit Duration and Recurrence Mechanisms

Explicit duration modeling allows the duration distribution in each regime to deviate from the implicit geometric distribution of Markov models. In REDSLDS:

• If $d_{t-1} > 1$, state and regime remain unchanged and $d_t = d_{t-1}-1$.
• Upon expiration ($d_{t-1} = 1$), new duration and possibly new regime are sampled, with duration transitions dependent on the previous continuous state $x_{t-1}$.
• Duration draws employ stick-breaking logistic-categorical links enabling flexible, potentially state-dependent dwell-time distributions.

This construction generalizes both classic Markov and (non-recurrent) explicit-duration HMMs, combining the advantages of both (preventing rapid switching, permitting context-dependent durations, and supporting rich segmentation behavior) (Słupiński et al., 2024).

6. Model Variants and Parameter Learning

Several variants exist within the SLDS paradigm:

• Classic SLDS: Markovian transitions, piecewise-linear dynamics (no state-dependent transitions, geometric dwell times).
• rSLDS: Discrete switches use stick-breaking logistic regression on the continuous latent state, partitioning state-space with regime-specific hyperplanes (Nassar et al., 2018, Linderman et al., 2016).
• REDSLDS: Augments rSLDS with explicit duration variables and stick-breaking logistic-categorical duration models (Słupiński et al., 2024).
• TrSLDS (Tree-Structured): Employs a tree-structured hierarchy of locally linear regimes for multi-scale decomposition (Nassar et al., 2018).

Parameter estimation proceeds via Gibbs or EM, exploiting conjugacy for transition and emission/dynamics blocks, with duration and recurrent transition weights sampled from Gaussian posteriors given Polya–Gamma auxiliary variables.

7. Practical Implications, Significance, and Limitations

The SLDS family, especially with explicit-duration and recurrent extensions, provides a structured approach to modeling and segmenting multivariate time series exhibiting abrupt, context-dependent regime changes. Their interpretability, flexibility in capturing dwell-time statistics, and efficient Bayesian learning schemes make them particularly attractive for sequence segmentation, dynamical system discovery, and interpretable time series analysis.

A key limitation is the increased computational cost associated with explicit-duration and high-dimensional discrete state spaces, necessitating careful use of augmentation and pruning methods for scalability in long or high-frequency sequences. The combinatorics of the explicit-duration process also present modeling and inference challenges for long-duration tasks.

Empirically, the addition of explicit durations in a recurrent framework consistently improves segmentation over Markovian and naive recurrent SLDS baselines, reflected in higher accuracy, F₁-scores, and log-likelihoods across applications ranging from controlled simulations to real animal behavior data (Słupiński et al., 2024).