Traffic jams move. A slowdown that starts at an on-ramp travels backward up the freeway at roughly 15 to 20 km/h, while in free-flowing traffic information travels forward with the cars. To estimate speeds everywhere on a road from the few vehicles that report their GPS positions, which is often only 3–5% of the fleet, a model has to know which direction in space and time correlation runs. The best probabilistic method, a Gaussian process (a model that predicts unobserved values as a weighted blend of nearby observations, with weights set by a "covariance kernel" encoding how similar two points are), tilts its kernel along one wave direction learned for the whole road. Real roads are not uniform. Wave speed changes near merges, lane drops and bottlenecks, and the road switches between free flow and congestion. A single global tilt therefore blurs exactly the shock fronts that matter.
We propose to replace the global tilt with a smooth, continuous change of coordinates. Time is re-measured along the local characteristic lines of traffic, with a wave speed that varies along the corridor. A regime field decides, at each point, whether the forward or the backward wave governs correlation. Because the kernel stays stationary in the warped coordinates, it remains valid as a covariance without special constructions. It also needs no grid, so the resolution at which speeds are reported no longer trades off against how many observations fall in each cell. The question the work answers is concrete: on sparse probe data, does letting wave speed vary in space produce better and better-calibrated reconstructions than a two-regime model with fixed wave speeds, and if so, where does the gain come from?
Core Claim
Two existing methods each solve half of the problem. The rotated anisotropic Gaussian process of Wu et al. rotates the space-time kernel by one learned angle and scales up with inducing points (a small set of summary locations that make Gaussian process inference tractable on large data). It reports MAE/RMSE of 4.48/6.27 m/s on HighD at 5% connected-vehicle penetration, but its angle and lengthscales are global. The oblique-grid method TW-LSMC of He et al. shears a discrete space-time grid along the backward wave speed so that shocks line up with matrix columns, and then completes the matrix as low-rank plus sparse. Its shear is also global, and it works on 3 m × 5 s cells that are mostly empty at 3% penetration.
Our claim is that the continuous analogue of the oblique grid, with a spatially varying shear, is a valid Gaussian process input warping. Gating that warp with a latent congestion field gives a non-stationary, directional covariance. That covariance is identifiable from sparse probes because it adds only a few smooth one-dimensional functions rather than a free lengthscale tensor at every point.
Concretely, let $w(x) < 0$ be the local backward wave speed at road position $x$, parameterized as a low-dimensional spline with a prior centered on the empirical −15 to −20 km/h range. The congestion-characteristic coordinate is
$$\tau_c(x,t) = t - \int_0^x \frac{dx'}{w(x')},$$
which reduces to the oblique shear of He et al. when $w$ is constant. The free-flow coordinate $\tau_f(x,t) = t - x/v_f$ uses the free-flow speed $v_f$. The covariance between space-time points $z=(x,t)$ and $z'$ is
$$k(z,z') = \pi_f(z)\pi_f(z')\,k_f\big((x,\tau_f),(x',\tau_f')\big) + \pi_c(z)\pi_c(z')\,k_c\big((x,\tau_c),(x',\tau_c')\big).$$
Here $k_f$ and $k_c$ are stationary Matérn kernels with long correlation along their characteristic coordinate and short correlation across it. The weights $\pi_c(z) = \sigma(h(z))$ and $\pi_f = 1-\pi_c$ come from a latent Gaussian process $h$ passed through a logistic sigmoid $\sigma$, so $\pi_c$ is the local probability of congestion. Each term is a stationary kernel on warped inputs scaled by an amplitude function, so the sum is positive semi-definite by construction. We therefore avoid the validity issues that motivated Paciorek et al..
Research Program
The first component is inference. The speed field, the gate $h$ and the spline coefficients of $w$ are learned jointly by variational inference with inducing points, following the scaling recipe of Wu et al. so that cost is linear in the number of probe points. The integral defining $\tau_c$ is evaluated in closed form for a piecewise-linear $1/w$.
The second component is identifiability. We will test whether $w(x)$ is recoverable from sparse probes by fitting on SUMO and Vissim simulations, where the true local wave speed is known from the simulator. Recovery error will be measured against penetration rate. Simulation lets us place a bottleneck and know the ground-truth wave field, which real data cannot provide.
The third component is the discretization question. Because predictions are continuous, we can evaluate our model and TW-LSMC at the same set of output resolutions. We expect TW-LSMC's error to follow a U-shaped curve in cell size, with empty cells at fine resolution and smeared shocks at coarse resolution, as the trade-off described by Fink et al. predicts. We expect the Gaussian process error to stay flat until the target resolution falls below the posterior's own correlation length.
This program draws on several prior lines of work in distinct roles. Paciorek et al. and the Gibbs kernels used by Lalchand et al. for Himalayan precipitation motivate local covariance, but they learn free lengthscale fields with MCMC on fewer than 1,000 points. Paciorek et al. also note that smoothly varying lengthscales handle sharp jumps poorly. That result constrains our design: we represent shocks by the gate and the warp rather than by shrinking lengthscales. Input warping as a route to non-stationarity comes from Hansen et al. and Dexheimer et al., who warp spectra and pixel distances respectively, but neither uses a physically defined warp. The evidence that wave speed is not constant comes from Zhang et al., whose time-varying shockwave model reconstructs trajectories deterministically and gives no field posterior.
Viewed through the spatial non-stationarity literature, $w(x)$ is a spatially varying coefficient, as in geographically weighted regression. The difference is that it is pooled through a smooth prior rather than fitted window by window. Window-by-window fitting becomes ill-conditioned in data deserts, the failure mode noted for local methods such as PyGRF.
Baselines
The naive reference point is a stationary isotropic Matérn Gaussian process with inducing points, trained on the same probe points. It shows how much any directional structure is worth.
The strongest existing probabilistic competitor is the rotated anisotropic Gaussian process of Wu et al., including their multi-lane variant where lanes are available. It is the closest method in kind: it is Bayesian, uses the same data and has the same scaling machinery. It reports 4.48/6.27 m/s MAE/RMSE on HighD at 5% penetration, against 55.9/121.5 m/s for the tensor-completion method STH-LRTC. We will rerun it with the same number of inducing points, the same optimizer budget and the same penetration samples. Beating it means lower RMSE and lower CRPS (continuous ranked probability score, a single score for both accuracy and calibration of a predictive distribution) on the same splits, with the gap concentrated near bottlenecks. Wu et al. do not report a calibration result on Lacuna, so we will compute coverage of 90% intervals ourselves for both models.
The strongest deterministic competitor is TW-LSMC. He et al. report a 12% RMSE improvement over prior methods on NGSIM US-101, 18% on corrupted data, and more than 20× speedup; no absolute RMSE for it is reported on Lacuna. We will run it at its native 3 m × 5 s grid and at a sweep of coarser grids, giving it its best resolution per penetration rate. We will also give it the global wave speed our model learns, so that it is not penalized for a poor shear.
The physics-informed networks of Shi et al., which use second-order ARZ physics, and Thodi et al., which use kinematic-wave CNN kernels, enter as point-estimate baselines on NGSIM and HighD under matched wall-clock budgets. The Adaptive Smoothing Method enters as the engineering reference; no reported result was found on Lacuna for its absolute error on these splits.
Three ablations isolate the key ingredient. The first fixes $w$ to a constant, giving a Bayesian two-regime model that is the honest version of adaptive smoothing. The second removes the gate, leaving a plain sum of the two warped kernels. The third replaces the warp with a fully free Gibbs-style orientation field $\theta(x,t)$ to test whether extra flexibility helps or overfits.
Why not just…?
Why not just use a gated two-regime model with fixed wave speeds? This is the most serious threat, and it is the first ablation. If wave speed varies little along a corridor, it may match the full model. NGSIM's 600 m segment might well be such a case. The experiments settle this by comparing errors stratified by distance to the bottleneck in SUMO and Vissim, where we control how much $w$ varies. If the constant-$w$ ablation matches the full model even when the simulated wave speed varies by a large fraction, we will report that spatial variation of $w$ does not matter at realistic penetration. That result would be useful in itself.
Why not just add a forward and a backward rotated kernel without gating? A plain sum lets every point correlate along both directions everywhere. Free-flow points would then borrow strength from backward-propagating shocks they never experienced. The ungated ablation measures this directly, and we expect the damage to appear as overconfident intervals at the edges of congestion.
Why not just fit Wu et al.'s model in separate road segments? Segment-wise fitting is the geographically weighted regression strategy, and it fails for the same reason. At 3–5% penetration the short segment around a bottleneck has too few probe points to fit an angle, and predictions jump at segment boundaries. We will include a segment-wise variant with segment length tuned on validation data.
Why not just let the orientation vary freely with a Gibbs kernel? Paciorek et al. report that such models are limited to small data and that latent lengthscales become unidentifiable where data are sparse. The free-$\theta$ ablation tests whether that holds here. We expect it to fit dense recordings well and degrade fastest as penetration drops.
Proposed Experiments
We use NGSIM US-101 lane 2 (600 m over 2500 s) and HighD recordings at 3%, 5%, 10%, 15% and 50% probe penetration, with 10 random probe draws per rate. We add the SUMO bottleneck scenario and the Vissim E-22 corridor for controlled wave-speed heterogeneity with known ground truth. Metrics are MAE, RMSE, CRPS, negative log predictive density and 90% interval coverage on held-out ground truth, plus recovery error for $w(x)$ in simulation.
The decisive checks are three. First, the full model must beat both Wu et al.'s rotated Gaussian process and the constant-$w$ ablation near bottlenecks at 3–5% penetration, not merely on average. Second, TW-LSMC's error versus cell size must show the predicted U-shape while ours stays flat. Third, the recovered $w(x)$ in simulation must track the true wave speed better than chance at 5% penetration. If the first check fails while the second holds, the contribution narrows to a grid-free, calibrated version of the oblique grid, and we will report it as that.
References
- Traffic state estimation from vehicle trajectories with anisotropic Gaussian processes
- Efficient and Robust Freeway Traffic Speed Estimation Under Oblique Grid Using Vehicle Trajectory Data
- Nonstationary Covariance Functions for Gaussian Process Regression
- Adaptive Spatio-Temporal Exploratory Models: Hemisphere-wide species distributions from massively crowdsourced eBird data
- Kernel Learning for Explainable Climate Science
- Probabilistic signal estimation for vibrational spectroscopy with a flexible non-stationary Gaussian process baseline model
- Learning a Depth Covariance Function
- A time-varying shockwave speed model for reconstructing trajectories on freeways using Lagrangian and Eulerian observations
- PyGRF: An Improved Python Geographical Random Forest Model and Case Studies in Public Health and Natural Disasters
- Physics-Informed Deep Learning for Traffic State Estimation: A Hybrid Paradigm Informed By Second-Order Traffic Models
- Incorporating Kinematic Wave Theory Into a Deep Learning Method for High-Resolution Traffic Speed Estimation