Behavioral Heterogeneity as Quantum-Inspired Representation

Abstract: Driver heterogeneity is often reduced to labels or discrete regimes, compressing what is inherently dynamic into static categories. We introduce quantum-inspired representation that models each driver as an evolving latent state, presented as a density matrix with structured mathematical properties. Behavioral observations are embedded via non-linear Random Fourier Features, while state evolution blends temporal persistence of behavior with context-dependent profile activation. We evaluate our approach on empirical driving data, Third Generation Simulation Data (TGSIM), showing how driving profiles are extracted and analyzed.

• The paper introduces a quantum-inspired driver representation in which Random Fourier Features and evolving density matrices preserve continuous mixtures and transitions between behavioral modes.
• The model identifies four distinct profiles on 3.2 million TGSIM observations, including three nearly rank-1 regimes and one genuinely multimodal profile, while selecting four profiles over three or five using likelihood.
• The framework links behavioral profiles to driving context, such as traffic density and headway, but its predictive advantage over sequential baselines and transferability beyond two TGSIM sites remain untested.

Overview and motivation

"Behavioral Heterogeneity as Quantum-Inspired Representation" (2603.22729) by Elayan and Kontar (University of Nebraska–Lincoln) proposes a framework in which driver heterogeneity is modeled as a continuously evolving latent state rather than a static label. The authors argue that prevailing approaches—parametric car-following extensions with driver-specific parameters, and end-to-end data-driven models that learn an average representation—either constrain variability to predefined functional forms or compress it into discrete semantic categories such as "aggressive" or "timid." Their central claim is that the information discarded by such compression lies in the transition logic between behavioral modes, and that this logic can be preserved by representing each driver as a density matrix that evolves over time.

The framework is explicitly "quantum-inspired" rather than quantum: it satisfies three structural criteria adopted from density-matrix learning with random features—(i) data encoded as normalized states in a Hilbert space, (ii) quadratic measurement rules consistent with the Born rule, and (iii) uncertainty represented by valid density matrices (symmetric, positive semidefinite, trace-normalized). No claim of physical quantum computation is made; the density matrix serves as a mathematically constrained representation of a probabilistic mixture over behavioral modes.

Model formulation

Each observation $(i,t)$ is a behavioral vector $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$ comprising relative speed to the leader, signed acceleration, and headway. This compact three-dimensional state is mapped through Random Fourier Features (RFF), $\phi_j(x) = \cos(w_j^\top x + b_j)$ with $w_j \sim \mathcal{N}(0, \sigma^{-2}I_d)$ and $b_j \sim \mathrm{Uniform}(0, 2\pi)$, into a $D$-dimensional normalized representation $\tilde{\phi}(x)$, approximating a Gaussian RBF kernel and allowing nonlinear interactions (e.g., headway-dependent sensitivity) to be handled linearly in feature space.

The model assumes $K$ population-level behavioral profiles, each a density matrix $\rho_k$ satisfying $\rho_k = \rho_k^\top \succeq 0$, $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$0. Diagonal entries weight individual RFF components; off-diagonal entries encode interactions among them. Context—pedestrian proximity, stop-sign proximity, traffic density within an omni-directional perception region, and corridor-average speed—modulates profile activation through a softmax weighting $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$1 with learned coefficients $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$2. The driver state is then predicted by blending temporal persistence with the context-weighted mixture,

$x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$3

and the likelihood of the observed behavioral vector under the predicted state is the quadratic form $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$4, consistent with the Born rule. After observation, the state is corrected via convex combination with the rank-one outer product $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$5, with adaptation rate $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$6. All parameters are fit by minimizing negative log-likelihood under the density-matrix constraints, using automatic differentiation in PyTorch. The authors note a practical computational result: for their configuration ($x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$7, $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$8, $x_{it} = [\Delta v_{it}, a_{it}, h_{it}]^\top$9), with more than 40,000 parameters, autodiff reduced per-epoch training from roughly 3–4 hours (coordinate-wise finite differences) to under 15 minutes.

Empirical analysis on TGSIM

The framework is evaluated on the FHWA Third Generation Simulation (TGSIM) datasets from two deployments: the Foggy Bottom urban intersection and the I-395 freeway, sampled at 0.1 s. After Gaussian smoothing and filtering, the dataset comprises 3,200,397 observations from 4,360 human-driven vehicle trajectories (1,277 urban, 3,083 freeway).

Profile selection. Comparing $\phi_j(x) = \cos(w_j^\top x + b_j)$0, mean negative log-likelihood per observation drops from 0.658 ($\phi_j(x) = \cos(w_j^\top x + b_j)$1) to 0.629 ($\phi_j(x) = \cos(w_j^\top x + b_j)$2) and shows no improvement at $\phi_j(x) = \cos(w_j^\top x + b_j)$3 (0.629). The authors select $\phi_j(x) = \cos(w_j^\top x + b_j)$4 on the basis of likelihood, spectral richness, and parsimony.

Spectral structure. A key finding is that three of the four learned profiles converge to effectively rank-1 spectra (leading eigenvalue 0.9999), while Profile 3 is genuinely mixed, with leading eigenvalues of 0.7151 and 0.2783 (71.5% and 27.8% of spectral mass). The authors present this as a structural advantage of the density-matrix parameterization: each profile learns its own effective dimensionality without an imposed rank constraint, so a single model can contain both sharply identified regimes and a regime that is irreducibly a probabilistic mixture of two behavioral tendencies.

Context activation and behavioral interpretation. The learned $\phi_j(x) = \cos(w_j^\top x + b_j)$5 coefficients produce interpretable and internally consistent activations. Profile 2 is the most context-sensitive, with a large positive coefficient on density ($\phi_j(x) = \cos(w_j^\top x + b_j)$6), and corresponds to dense, interaction-heavy following (mean headway ≈ 44 m). Profile 1 loads negatively on all spatial context variables (strongest suppression on density, $\phi_j(x) = \cos(w_j^\top x + b_j)$7) and corresponds to free-flow driving with large headways (mean ≈ 83 m) and near-neutral acceleration. Profile 4 captures responsive speed recovery (mean $\phi_j(x) = \cos(w_j^\top x + b_j)$8 m/s with positive acceleration) in faster, moderately interactive streams. Profile 3, the multimodal regime, is context-insensitive (negative coefficients on density and average speed, $\phi_j(x) = \cos(w_j^\top x + b_j)$9 and $w_j \sim \mathcal{N}(0, \sigma^{-2}I_d)$0), consistent with transitional driving spanning multiple conditions; its two modes separate a controlled-progression pattern (mean headway ≈ 53 m) from a more assertive following pattern (≈ 41 m).

Geometric separation. Pairwise Frobenius distances between profile density matrices are all well above zero, confirming distinct profile geometries. The largest separations are between Profiles 1–4 ($w_j \sim \mathcal{N}(0, \sigma^{-2}I_d)$1) and 1–2 ($w_j \sim \mathcal{N}(0, \sigma^{-2}I_d)$2), contrasting free-flow with interaction-heavy regimes. Profile 3 is nearest Profile 1 ($w_j \sim \mathcal{N}(0, \sigma^{-2}I_d)$3), consistent with its dominant mode sharing large headways with the low-constraint regime.

The coherence between spectral structure, context activation, and behavioral statistics across both urban and freeway deployments is the paper's strongest empirical support: the activation mechanism distinguishes environment-specific behavior without suppressing either setting.

Limitations and open questions

The authors are explicit that the framework is intentionally compact, and several concessions qualify the results. The behavioral state omits lateral behavior and richer interaction cues; the context vector is hand-defined, and the model's behavior depends on how context is constructed; no ablation isolates the contribution of individual contextual variables; and no comparison against sequential baselines (e.g., regime-switching or Bayesian dynamic regression car-following models) is provided, so the claim that evolving density-matrix states improve predictive performance over simpler alternatives remains untested in this paper. The evaluation is also confined to two TGSIM sites, leaving transferability across sites and conditions open. The paper accordingly identifies as open questions: whether more flexible but constrained context mappings improve fit, and whether evolving-state representations improve prediction of system-level outcomes such as traffic stability, capacity drop, and mixed-autonomy effects.

Conclusion

This paper contributes a mathematically principled, interpretable representation of driver heterogeneity in which behavioral profiles are density matrices over a nonlinear feature space, driver states evolve through persistence–context mixing and observation-based correction, and likelihoods follow quadratic measurement rules. On 3.2 million TGSIM observations, it recovers four parsimonious, geometrically separated profiles—including one irreducibly multimodal regime—whose context activations align coherently with urban and freeway conditions. The contribution is primarily representational and interpretive; establishing predictive superiority and system-level utility relative to sequential baselines remains an open empirical question.