Incremental Multi-Hypothesis Smoothing (iMHS)

• Incremental Multi-Hypothesis Smoothing (iMHS) is a smoothing algorithm that updates state trajectories in hybrid systems by maintaining multiple backward hypotheses to mitigate path degeneracy.
• It combines particle filtering techniques (PaRIS) and factor-graph Bayes trees to achieve linear per-update time complexity and controlled variance growth in online estimation.
• Experimental validations demonstrate that iMHS enhances real-time detection in applications such as lane change, aircraft maneuver, and legged robot contact, while keeping memory requirements efficient.

Incremental Multi-Hypothesis Smoothing (iMHS) is a class of online smoothing algorithms designed for efficient estimation in latent-variable dynamical models, with particular efficacy in hybrid discrete-continuous systems and hidden Markov models (HMMs). iMHS algorithms operate incrementally, updating estimates as new data arrives while maintaining multiple backward hypotheses per trajectory, thereby controlling the path degeneracy inherent in single-hypothesis filters. Theoretical analyses establish statistical consistency, linear time complexity (per update), and controlled variance growth under reasonable model assumptions. iMHS encompasses both particle-based (e.g., PaRIS) and factor-graph-based (e.g., multi-hypothesis Bayes trees) implementations, unifying concepts from sequential Monte Carlo, graphical models, and hybrid system inference (Olsson et al., 2014, Jiang et al., 2021).

1. Problem Formulations: Smoothing in State-Space and Hybrid Models

iMHS addresses the estimation of latent state trajectories given a sequence of observations in models where the system can switch between multiple discrete modes, often with continuous-valued state dynamics and stochastic emissions. The general state-space HMM formulation is:

• Latent process: $X_0 \sim \pi_0$, $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$
• Emissions: $Y_t \mid X_t = x$ have density $g(x, y)$

The joint smoothing objective is to approximate smoothed expectations of additive functionals:

$$S_t := \mathbb{E}\Bigg[ \sum_{\ell=0}^{t-1} f_\ell(X_\ell, X_{\ell+1}) \Big| Y_{0:t} = y_{0:t} \Bigg ]$$

In hybrid systems, the state has both continuous ($x_k \in \mathbb{R}^n$) and discrete ($m_k \in \{1,...,|M|\}$) components. The posterior factorizes under hybrid dynamic Bayesian network assumptions:

$$p(X^K, M^{K-1} | Z^K ) \propto \phi(m_0)\phi(x_0)\phi^z_0(x_0) \prod_{k=1}^{K-1} [\phi(m_{k-1}, m_k)\phi^m_k(x_{k-1}, x_k, m_{k-1})\phi^z_k(x_k)]$$

where $\phi^m_k$ and $\phi^z_k$ are process and measurement factors, respectively (Jiang et al., 2021).

2. Core Algorithmic Structure: Multi-Hypothesis Updating

The central feature of iMHS is the propagation of $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$0 backward “hypotheses” (i.e., sampled parental histories or mode assignments) per new forward-sampled state at each time step, thus inducing a "multi-hypothesis" support sufficient to mitigate the rapid path degeneracy that occurs in conventional single backward-sample smoothers.

Particle-Based Online Smoothing (PaRIS)

The PaRIS algorithm proceeds as follows for $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$1 particles and $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$2 backward draws:

1. Forward particle filter update: Resample ancestors via weighted multinomial selection, propagate by the transition kernel, evaluate the new likelihood.
2. On-the-fly multi-hypothesis backward sampling: For each new particle, draw $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$3 backward indices from the filter-induced backward kernel:

$X_{t+1} \mid X_t = x \sim Q(x, \cdot)$4

1. Auxiliary statistics update: Update additive functionals using the $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$5 sampled backward hypotheses via:

$X_{t+1} \mid X_t = x \sim Q(x, \cdot)$6

2. Output: Propagate surviving particles, weights, and auxiliary statistics. Optional: Estimate smoothed statistics by weighted averages (Olsson et al., 2014).

Factor-Graph and Bayes Tree Incremental Smoothing

For hybrid-state estimation, iMHS eliminates factor graphs into a multi-hypothesis Bayes tree, branching on discrete mode histories. Key steps:

• Factor-graph construction: At time $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$7, introduce new process, measurement, and mode-transition factors.
• Batch elimination: Sequentially eliminate variables to construct a chain/tree of conditionals; structure is distributed across the multi-hypothesis Bayes tree, with each leaf representing a unique discrete mode history.
• Incremental update (iSAM-style):
  • For each surviving hypothesis, expand all possible mode branches for the next time, perform local QR elimination, update branch weights.
  • Prune hypotheses with posterior weight below threshold $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$8, ensuring tractable representation.
  • Only the last $X_{t+1} \mid X_t = x \sim Q(x, \cdot)$9 layers (fixed lag) need retention, maintaining efficient memory use (Jiang et al., 2021).

3. Numerical Stability, Path Degeneracy, and Variance Analysis

A defining property of iMHS is its control over estimator degeneracy:

• Single-hypothesis smoothing ($Y_t \mid X_t = x$0): Rapid collapse—all backward paths coalesce, yielding quadratic growth of estimator variance with time.
• Multi-hypothesis ($Y_t \mid X_t = x$1): By independently sampling $Y_t \mid X_t = x$2 backward hypotheses per particle, the algorithm maintains a rich, non-degenerate representation. This breaks the collapse of ancestral lineages.
• Variance scaling: Rigorous results establish that the asymptotic variance growth is linear in time:

$Y_t \mid X_t = x$3

where $Y_t \mid X_t = x$4 is a model-dependent constant. In the limit $Y_t \mid X_t = x$5, the penalty vanishes relative to the standard forward-backward sampler. Marginal smoothing admits even tighter, uniform-in-$Y_t \mid X_t = x$6 variance bounds (Olsson et al., 2014).

4. Computational Complexity and Memory Requirements

iMHS algorithms achieve online efficiency suitable for real-time applications:

• PaRIS: Each backward draw can be implemented in $Y_t \mid X_t = x$7 time (via accept-reject sampling under mild assumptions). Each step is $Y_t \mid X_t = x$8 in time, with $Y_t \mid X_t = x$9 memory requirement as only the current layer of particles and auxiliary statistics is retained.
• Multi-hypothesis Bayes tree (factor-graph): Complexity per step is $g(x, y)$0, for $g(x, y)$1 active hypotheses, $g(x, y)$2 modes, and state size $g(x, y)$3 per clique. Memory is $g(x, y)$4, but in practice, aggressive pruning ($g(x, y)$5) and fixed-lag truncation keep $g(x, y)$6 bounded and resource use near-constant per step (Jiang et al., 2021).

5. Experimental Validation and Benchmarking

Validation across multiple domains demonstrates iMHS achieves real-time, statistically robust mode inference in hybrid systems:

• Lane change detection: On the NGSIM I-80 dataset (10 Hz), using constant-position and constant-velocity modes, immediate identification of mode switches is achieved.
• Aircraft maneuver detection: In synthetic simulations, the smoother detects maneuver changes with zero lag, outperforming IMM filter variants.
• Legged robot contact detection: On both simulated and real quadruped data, iMHS accurately recovers foot contact modes, as reflected in cross-entropy errors. Pruning keeps the number of active hypotheses ($g(x, y)$7) below 10 in practice, enabling sustained real-time performance (Jiang et al., 2021).

Table 1: Per-leg Cross-entropy Error in Legged-Robot Applications

Platform	LH	LF	RF	RH
A1 (sim)	1.013	1.089	0.377	1.242
ANYmal (real)	0.819	0.639	1.362	1.284

6. Connections with Related Methods

iMHS generalizes and connects several themes:

• Sequential Monte Carlo (SMC): iMHS subsumes standard particle smoothers but uniquely leverages incremental, multi-hypothesis backward sampling for improved stability.
• Rao-Blackwellized Smoothing and IMM: Unlike IMM and marginalization-based approaches (prone to lag and degeneracy), iMHS achieves near zero-lag state and mode inference by explicitly representing distinct mode and trajectory hypotheses.
• Incremental Smoothing and Mapping (iSAM): The hybrid Bayes-tree iMHS fuses iSAM's efficient incremental inference with explicit preservation of multimodal mode histories, automating shared clique management for repeated prefix histories (Jiang et al., 2021, Olsson et al., 2014).

7. Practical Implications, Limitations, and Outlook

iMHS frameworks deliver statistically principled, scalable smoothing with robust handling of mode transitions and nonlinearity in hybrid and HMM settings. Aggressive hypothesis pruning, fixed-lag management, and constant-factor per-step complexity confer practical tractability. A plausible implication is the feasibility of deploying iMHS algorithms in embedded and low-latency robotics systems where both real-time constraints and accurate mode estimation are critical. Challenges persist in scaling to very large mode sets or highly nonlocal dependencies, where the exponential growth of unpruned hypotheses may transiently breach resource limits depending on the informativeness of the model and measurement sequences (Jiang et al., 2021, Olsson et al., 2014).