Trajectory Variance Metric Overview

• Trajectory Variance Metric is a quantitative measure that captures dispersion and inconsistency among trajectories using statistical, geometric, and information-theoretic approaches.
• It is applied across diverse domains—such as human movement science, reinforcement learning, quantum mechanics, and multi-object tracking—to evaluate performance and risk.
• Different formulations like PVP, LTVE, MOM, and assignment-based metrics demonstrate its adaptability in addressing variability in time-series, dynamical systems, and control tasks.

A trajectory variance metric is a quantitative measure characterizing the dispersion, variability, or inconsistency among trajectories in a given space, or across repeated executions of a task, as a function of time or other structural factors. The form of the metric, its theoretical underpinnings, and practical applications vary by domain: human movement science, dynamical systems, computer vision, reinforcement learning, quantum mechanics, and multi-object tracking all employ specialized trajectory variance metrics, each backed by distinct statistical, information-theoretic, or geometric frameworks.

1. Core Concepts and Formal Definitions

At the foundation, trajectory variance metrics are predicated on formalizations of statistical variance, geometric or functional distances between paths, or state-dependent moments of trajectory-dependent stochastic processes.

• Positional Variance Profile (PVP): For $N$ time-aligned spatial trajectories $x_j(t)$, $j=1,\ldots,N$, the positional variance at time $t$ is

$$\sigma^2(t) = \frac{1}{N-1} \sum_{j=1}^N \big(x_j(t) - \mu(t)\big)^2, \quad \mu(t)=\frac{1}{N}\sum_{j=1}^N x_j(t).$$

The temporal profile $\sigma^2(t)$—the PVP—exposes phases of increasing and decreasing variance characteristic of stochastic or feedback-driven processes (Gori et al., 2018).

• Trajectory Metric in Dynamical Systems: Let $\Psi_1,\Psi_2$ be two curves in $\mathbb{R}^n$, with a metric $d(\Psi_1,\Psi_2)$ such as discrete $L^2$:

$$d_{l^2}(\Psi_{1},\Psi_{2}) = \sqrt{\frac{1}{k} \sum_{j=0}^{k-1} \|\Psi_1(j) - \Psi_2(j)\|_2^2}$$

and alternatives including Fréchet and Hausdorff metrics (Tsai et al., 2024).

• Variance of Trajectory Returns in RL: For policy $\pi$, the variance $\Var_{\pi}(R)$ of trajectory return $R$ is

$\Var_{\pi}(R) = \mathbb{E}_\pi[R^2] - \big(\mathbb{E}_\pi[R]\big)^2$

with trajectory variance serving as a risk-sensitive objective (Jain et al., 2021).

• No-Reference Variance (MOM) for Registered Point Clouds: MOM evaluates the averaged direction-wise variance of point clouds partitioned by surface normals, strongly correlating with the Relative Pose Error (RPE), and is defined as

$$\text{MOM} = \frac{1}{3} \sum_{k=1}^3 V_k,\quad V_k = \frac{1}{|P_k|} \sum_{p\in P_k} (n_k^\intercal p - \mu_k)^2$$

where $n_k$ are mutually orthogonal directions (Kornilova et al., 2021).

• Variance Decomposition in Bohmian Mechanics: The quantum variance is split into the ensemble variance of the weak actual value field and a quantum term:

$\text{Var}_Q[A] = \text{Var}_B[A_w] + 2A$

with $$A_w(x) = \Re(\psi^*(x) (\hat{A}\psi)(x))/|\psi(x)|^2$$, and $2A$ quantifying phase-amplitude coupling (Ye, 31 Dec 2025).

• Set-Based Multi-Object Tracking Metrics: Metrics like the time-weighted multidimensional assignment metric encode trajectory-level distances via dynamically weighted combinations of localization, false/missed, and switch costs (García-Fernández et al., 2021).

2. Information-Theoretic and Statistical Trajectory Variance Metrics

In movement science, the feedback information-theoretic transmission scheme (FITTS) provides a rigorous model for decomposing and quantifying trajectory variance in aimed movements. Here, movements are viewed as information transmission over a noisy feedback channel. The principal metric is the exponential variance decay rate in the homing-in phase, identified as the channel capacity $C$. For $t \geq \tau$ (the homing phase onset), the profile follows: $\sigma^2(t) = \sigma_0^2 \exp[-2C(t-\tau)]$ $C$ is estimated empirically via a log-linear fit to $\log_2 \sigma(t)$ on $[\tau, \Omega]$. This parameter is invariant to experimental manipulations of movement distance and width, robustly characterizing individual performance across tasks (Gori et al., 2018).

3. Geometric and Metric-Based Trajectory Variation

The quantification of trajectory variance in the context of dynamical flows leverages metrics induced by the full path geometry rather than endpoints:

• Local Trajectory Variation Exponent (LTVE): Given a vector field $v(x,t)$ and initial condition $x$, LTVE at $x$ is

$$\text{LTVE}(x; \delta, T) = \frac{1}{T} \ln \left( \frac{1}{\delta} \max_{u: \|u\|_2 = 1} d(\Psi_{x+\delta u}, \Psi_x) \right)$$

capturing maximal separation under a chosen trajectory metric $d$ across the interval $[t_0,t_0+T]$. With specific choices of $d$, LTVE generalizes the finite time Lyapunov exponent (FTLE) to pathwise separation, thus providing enhanced sensitivity to structure throughout the trajectory, rather than only at endpoints (Tsai et al., 2024).

Metric choices (e.g., $L^2$, Fréchet, Hausdorff) substantially affect sensitivity to temporal alignment and path deviations.

4. Variance Metrics in Learning and Control

In reinforcement learning, trajectory variance is central to risk-sensitive objectives. Actor-critic algorithms explicitly penalize return variance by augmenting the objective,

$J_\lambda(\pi) = \mathbb{E}[R] - \lambda \Var_\pi(R),$

and employ direct variance estimation via temporal-difference updates. Critic estimates for first and second moments ($V(s)=\mathbb{E}[R|s]$, $U(s)=\mathbb{E}[R^2|s]$) provide incremental updates to trajectory variance estimates at each state, enabling variance-penalized policy optimization. Empirically, such algorithms reduce trajectory return variance by $15$–$50\%$ with minimal loss in mean performance, enhancing reliability in task execution (Jain et al., 2021).

5. Physical, Functional, and No-Reference Trajectory Variance Technologies

In robotics and mapping, trajectory variance metrics must operate without ground truth:

• MOM (Mutually Orthogonal Metric): MOM leverages the variance of point-cloud projections along mutually orthogonal surface normals, extracted via point cloud clustering and graph-based selection. This no-reference metric correlates near-linearly with traditional RPE on both synthetic and real datasets, delivering efficient and robust benchmarking in point-based mapping applications. MOM can be computed online or offline and is robust in settings with abundant orthogonal planar geometry (Kornilova et al., 2021).

In quantum mechanics, the "trajectory variance metric" as defined by ensemble variance in Bohmian mechanics isolates classical-like fluctuations of weak actual values from quantum fluctuations (the latter quantified precisely in terms of the average quantum potential for canonical observables). For position, the Bohmian variance matches the quantum variance, while for momentum, the decomposition exposes non-classicality via quantum potential (Ye, 31 Dec 2025).

6. Set and Assignment-Based Trajectory Variance in Multi-Object Tracking

For multi-object tracking evaluation, trajectory variance is embedded in assignment-based metrics that capture geometric, labeling, and temporal discrepancies. The time-weighted trajectory metric aggregates costs over localization, false and missed targets, and track switching, with weight sequences $\{w_1^k\},\{w_2^k\}$ enabling emphasis on specific time intervals. The assignment is formulated via multidimensional assignment or LP-relaxed versions, ensuring metric properties under all parameterizations. The metric accommodates random sets of trajectories, allowing empirical risk estimation and ranking of algorithms across diverse scenarios (García-Fernández et al., 2021).

Setting	Representative Metric	Underlying Principle
Human movement	FITTS/PVP, Capacity $C$	Exponential variance decay, info theory
Fluid flows	LTVE	Pathwise maximal metric separation
3D mapping	MOM	Direction-wise variance in point clouds
RL/control	Return variance, $J_\lambda$	Moment-based risk-sensitive optimization
Quantum mechanics	Bohmian Var	Ensemble and quantum decomposed variance
Multi-object track	Assignment metrics, LP	Temporal, spatial, label switch variance

A plausible implication is that trajectory variance metrics are highly contextual—parameterizations, data aggregation, and their interpretability depend sensitively on the domain, targeted application, and the available measurement modality.

7. Domain-Specific Implications and Open Directions

Trajectory variance metrics yield interpretable, sometimes physically grounded, criteria for comparing system reliability, control quality, or inference performance. In movement science, the stability of information capacity $C$ supports robust individual-level calibration. In dynamical systems, dependence on full-path distance allows nuanced detection of coherent structures. No-reference metrics such as MOM simplify benchmarking in large-scale mapping where ground-truth is infeasible.

Notably, in quantum mechanics, trajectory variance decompositions fail to grant physical meaning to spin observables, highlighting foundational distinctions across observables. In tracking, assignment-based metrics facilitate inter-method comparison with flexible time-sensitivity.

This suggests an ongoing need for rigorously defined, context-aware trajectory variance metrics, with continued research into their invariance, computational tractability, and domain-specific significance.