A Gaussian Mixture Model (GMM) represents a random variable’s density as a convex
combination of K unimodal Gaussian components (“modes”), p∗(z)=∑k=1KπkN(μk,Σk) with mixing weights πk≥0, ∑kπk=1. It is the
natural model for multimodal uncertainty — distributions with several distinct plausible
outcomes that a single Gaussian cannot capture. In ren2022chance
the GMM models the predicted position of a moving obstacle whose discrete high-level intents
(e.g. go-straight vs. turn) come from a trajectory-prediction network; this is a risk-layer construct and is regime-agnostic (the cited application is ground autonomous
driving, not a space system — no free-flying or free-floating dynamics are assumed).
GMM density over the uncertain parameter δ∼p∗ (component k is N(μk,Σk)): p∗(z)=∑k=1KπkN(μk,Σk),∑k=1Kπk=1.
A single chance constraint under a GMM decomposes into one per-mode Gaussian chance constraint
plus a risk-budget split across the modes (the key property that makes GMM tractable):
\ge 1-\epsilon_k\ \ \forall k\in\mathbb{Z}_{1:K},
\qquad \sum_{k=1}^{K} \pi_k \epsilon_k = \epsilon.$$
> **Notation flags.** (1) The source writes the per-mode SOC tightening constant as $\Gamma_k$, a
> *scalar* inverse-normal/CVaR quantile. This **collides with the load-bearing matrix
> $\boldsymbol\Gamma$** (the $12\times(6+n)$ coordinate transform) in [notation.md](../notation.md);
> the SOC form is therefore omitted here and the scalar lives on
> [chance_constraints](chance_constraints.md) / [conditional_value_at_risk](conditional_value_at_risk.md).
> (2) The source uses $\epsilon$ for the per-mode risk bound and $\alpha$ as the CVaR auxiliary
> (infimum) variable; [notation.md](../notation.md) reserves $\alpha$ for the CVaR *confidence
> level*. Symbols $\pi_k,\mu_k,\Sigma_k,K,\epsilon_k$ are not yet in the registry — added here
> faithfully to the source, flagged for central registration.
> (3) The source's $\tilde{\boldsymbol x}=[\,\boldsymbol x^\top\ 1\,]^\top$ is the *augmented constraint
> vector* of the affine collision constraint $\boldsymbol\delta^\top\tilde{\boldsymbol x}\le0$ — **not**
> the stacked outer-loop tracking error $\tilde{\boldsymbol x}\in\mathbb{R}^9$ that
> [notation.md](../notation.md) reserves under the same glyph; and $\boldsymbol\delta$ here is the
> source's uncertain constraint parameter (not yet in the registry). Both are reproduced
> source-faithfully and flagged so the collision is not silent.
## Source Support
- [ren2022chance](../sources/ren2022chance.md) — primary and sole source: defines the GMM, decomposes a
GMM chance constraint into per-mode Gaussian chance constraints with a weight-split risk budget, and
shows (on nuScenes/Trajectron++) that GMM modelling yields *less conservative* feasible trajectories
than a single unimodal Gaussian, which was infeasible. Moments may be known or estimated from samples
(clustered, e.g. by expectation–maximization).
## Related Topics
- [chance_constraints](chance_constraints.md) — a GMM chance constraint reduces to one Gaussian chance
constraint per mode; the GMM is the *uncertainty model* that those constraints are taken over.
- [conditional_value_at_risk](conditional_value_at_risk.md) — the source's risk-aware tightening of each
per-mode chance constraint; under an affine constraint both share one SOC form, differing only by a scalar.
- [covariance_propagation](covariance_propagation.md) — each GMM mode carries a covariance $\Sigma_k$;
propagating those per-mode covariances through the system is what feeds the chance/CVaR constraints.
- motion_planning_under_uncertainty — the GMM is the uncertainty
representation plugged into the chance-/CVaR-constrained trajectory planner.
- measurement_uncertainty — the GMM moments are estimated from finite samples
(sample-to-mode assignment + per-mode mean/covariance), so estimation error must be robustified; see
also [moment_concentration_bound](moment_concentration_bound.md).
## Open Questions
- The source assumes the sample-to-mode assignment and the mode weights $\pi_k$ are known a priori (from a
predictor's latent variables, Assumptions 2–3); how does the per-mode chance-constraint decomposition
degrade when the number of modes $K$ or the assignment is itself uncertain?
- For risk-aware inspection of a tumbling target, would multimodal *target-pose* uncertainty (several
plausible tumble modes) be well-captured by a GMM, and does the affine-constraint reduction survive once
a free-flying manipulator's collision constraint is linearized to $\boldsymbol\delta^\top\tilde{\boldsymbol x}\le0$?