Stochastic nonlocal traffic flow models with Markovian noise

1. Problem Statement

The paper addresses the modeling of traffic flow using nonlocal scalar conservation laws, where driver behavior depends on traffic conditions ahead (look-ahead distance) rather than just local density. While deterministic nonlocal models are well-established, real-world traffic is subject to stochastic influences (e.g., human reaction variability, sensor noise, or incomplete information).

Previous work by the authors ([5]) introduced a stochastic nonlocal velocity (sNV) model driven by white noise (time-independent, uncorrelated random perturbations). However, this initial framework had limitations:

1. Theoretical Gaps: The measurability of the random weak entropy solutions was not rigorously established, leaving the definition of expected values (e.g., mean traffic density) mathematically ambiguous.
2. Modeling Limitations: White noise fails to capture temporal correlations (autocorrelation) inherent in real driving behavior, where drivers react to conditions over a duration rather than instantaneously.
3. Computational Cost: Estimating the expected traffic state via standard Monte Carlo simulations is computationally expensive and suffers from numerical diffusion (smoothing of shocks) when averaging discontinuous density profiles.

The paper aims to extend the sNV framework to include Markovian noise (specifically Jacobi-type processes), prove the measurability of solutions to ensure well-defined expectations, and derive a deterministic mean-value PDE that serves as a computationally efficient proxy for the expected traffic state.

2. Methodology

A. Mathematical Framework

The authors extend the deterministic nonlocal velocity (NV) model:
$$\partial_t \rho + \partial_x (\rho (W_\eta * v(\rho))) = 0$$
to a stochastic version (sNV) where the velocity function $v(\rho)$ is perturbed by a time-dependent error term $\epsilon(t)$:
$$v_\epsilon(\rho, t) = \max\{0, v(\rho) + \epsilon(t)\}$$

Key Theoretical Steps:

1. Measurability Proof: Using a deterministic stability estimate (Lemma 1.5) which bounds the $L^1$-distance between two solutions based on the difference in their noise realizations, the authors prove that the random weak entropy solution $\rho(t, x)$ is a measurable function with respect to the probability space. This ensures that quantities like the expectation $E[\rho(t, x)]$ are well-defined.
2. Local Solution Operator: They define a local solution operator $S^a_t$ that maps an initial density to the solution at time $t$ for a constant noise realization $a$. This operator is shown to be locally Lipschitz.

B. Noise Modeling: Markovian vs. White Noise

The paper replaces the white noise assumption with Markovian error processes.

• White Noise: Independent, uniformly distributed increments (as in [5]).
• Jacobi Process (Markovian): The authors introduce a stochastic differential equation (SDE) of Jacobi type to generate the noise:
  $$d\epsilon(t) = -\alpha \epsilon(t) dt + \sigma \sqrt{(\epsilon(t) + \tau)(\tau - \epsilon(t))} dW_t$$
  • Properties: This process is uniformly bounded within $[-\tau, \tau]$ (crucial for physical interpretability, ensuring velocity remains non-negative and bounded), possesses autocorrelation (temporal memory), and has a known Fokker-Planck equation for its probability density evolution.
  • Implementation: An acceptance-rejection Euler-Maruyama scheme is used for numerical simulation to maintain the bounds.

C. Derivation of the Mean-Value Proxy (EsNV)

To avoid expensive Monte Carlo averaging of density (which smears shocks), the authors derive a deterministic Mean-Value Hyperbolic PDE (EsNV):
$$\partial_t \bar{m} + \partial_x (\bar{m} (W_\eta * \bar{v}(\bar{m}, t))) = 0$$
where $\bar{v}$ is the expected velocity:
$$\bar{v}(\bar{m}, t) = \int_{-\tau}^{\tau} \max\{0, v(\bar{m}) + a\} \nu_t(da)$$
Here, $\nu_t$ is the probability distribution of the noise at time $t$. This approach decouples the noise propagation from the density, propagating a single expected density using an averaged velocity function.

D. Numerical Schemes

• Stochastic Scheme: A Godunov-type finite volume scheme is used for the sNV model. The noise is sampled at each time step according to the Markov chain (White or Jacobi).
• Mean-Value Scheme: The same Godunov scheme is applied to the EsNV model, using the pre-computed expected velocity $\bar{v}$.
• Validation Metric: Instead of comparing density averages (which smooth shocks), the authors validate the proxy by comparing characteristics (particle trajectories). They define a "Characteristic Monte Carlo Average" and compare it against the characteristics of the EsNV solution.

3. Key Contributions

1. Theoretical Rigor: Proved the measurability of random weak entropy solutions for nonlocal traffic models under general bounded noise assumptions, establishing the mathematical foundation for defining expectations.
2. Novel Noise Model: Introduced Jacobi-type Markovian noise into traffic flow modeling. This provides:
   • Physical Interpretability: Bounded noise ensures velocities remain realistic.
   • Autocorrelation: Captures the temporal memory of driver behavior, leading to significantly stronger and more realistic perturbations than white noise.
3. Mean-Value Proxy (EsNV): Derived a deterministic PDE that approximates the expected traffic state. Unlike standard Monte Carlo averages of density, this proxy preserves shock structures and satisfies a conservation law.
4. Validation via Characteristics: Demonstrated that while density averages smooth out shocks, the average of characteristics converges to the solution of the EsNV model, validating the proxy's accuracy in capturing the "most likely" traffic evolution.

4. Results

• Impact of Noise Type: Simulations show that Jacobi-type noise induces significantly stronger fluctuations in traffic density compared to white noise. White noise tends to average out locally due to lack of correlation, whereas Markovian noise creates persistent perturbations that significantly alter traffic waves.
• Proxy Accuracy:
  • In smooth traffic regions, the EsNV solution closely matches the Monte Carlo average.
  • Near shocks (traffic jams), the Monte Carlo average of densities becomes artificially smooth (diffusive).
  • However, the EsNV solution aligns with the median of the stochastic realizations and, crucially, matches the average of the characteristics. This confirms that EsNV correctly captures the propagation of the expected traffic wave without artificial diffusion.
• Nonlocality Benefit: The study highlights that the nonlocal look-ahead parameter ($\eta$) stabilizes the system. As $\eta$ decreases (approaching local models), the bias between the proxy and the true expectation increases, validating the importance of nonlocality in stochastic settings.

5. Significance

This work bridges a critical gap between stochastic conservation laws and practical traffic modeling:

• Computational Efficiency: The EsNV proxy offers a deterministic alternative to Monte Carlo simulations for predicting expected traffic conditions, drastically reducing computational cost while preserving the physics of shock propagation.
• Realistic Modeling: By moving from white noise to Markovian (Jacobi) noise, the model better reflects the reality of connected and autonomous vehicles, where drivers (or algorithms) react to information over time intervals, not just instantaneously.
• Theoretical Foundation: The proof of measurability resolves technical ambiguities in previous literature, allowing for rigorous statistical analysis of stochastic traffic flow.

In summary, the paper provides a robust, mathematically sound, and computationally efficient framework for simulating and predicting traffic flow under uncertainty, specifically tailored for the era of connected and autonomous vehicles.