Unbiased Footprints for Stagnant Air: Control-Variate Particle Transport for Urban CO2 Inversion

City-scale CO2 monitoring infers how much carbon a metropolis emits by reading small concentration enhancements of 2 to 10 ppm on top of a 420 ppm background, then running atmospheric transport backward to the surface. The weak point is weather in which the air barely moves. On calm nights, winter mornings, and days with winds under 2 m/s, vertical turbulent mixing collapses and horizontal wind stalls. Operational inversions throw these observations away. One Paris study explicitly excludes all data below 2 m/s (An attempt at estimating Paris area CO 2 emissions from atmospheric concentration measurements), and grid-based models are known to mis-simulate the shallow, stable layers in which emissions pool (Analysis of temporal and spatial variability of atmospheric CO 2 concentration within Paris from the GreenLITE™ laser imaging experiment). Mathematically, these are the conditions under which the transport equation becomes degenerate: its diffusion coefficient falls to zero in some directions and regions. Smoothing then stops and sharp fronts appear. Grid models handle this badly, because the numerical diffusion they add smears exactly those fronts.

We propose to compute the quantity an inversion actually needs, the sensitivity of each tower reading to emissions in each surface cell (the "footprint"), with a stochastic particle estimator whose correctness does not depend on smoothness. Particles are released backward in time from the sensor and follow wind plus turbulent noise. A neural network trained across an ensemble of plausible weather states supplies a control variate, a correction term with exactly zero mean that cancels most of the Monte Carlo noise. The estimator stays unbiased even where the network is wrong, and the network is most likely to be wrong at the sharp top of a collapsed boundary layer. The combination is meant to make stagnant-air observations usable at a cost that allows transport uncertainty to be integrated rather than ignored.

Core Claim

The claim has three parts. First, in degenerate atmospheric regimes, a backward stochastic representation of transport (the Feynman–Kac formula, which writes a PDE solution as an average over random paths) has no smoothing bias, unlike grid models. Second, its variance can be cut by one to two orders of magnitude with a learned martingale control variate. Third, that control variate remains a valid correction precisely where neural approximations fail. Together these let an urban inversion assimilate low-wind and nocturnal data while sampling a high-dimensional space of meteorological uncertainty. Grid models cannot afford that sampling.

To be precise about what "degenerate" means here: CO2 is a passive tracer, so the diffusion coefficient is not solution-dependent as in the porous medium equation. It depends on space and time. It is the vertical eddy diffusivity $K_z(z,t;\xi)$, which approaches zero inside a stable surface layer, combined with mean winds that approach zero. Here $\xi$ is a vector of uncertain meteorological parameters such as boundary-layer height, the stability profile, and wind perturbations. The result is a linear but directionally degenerate Kolmogorov equation. The literature on degenerate diffusion warns that it loses regularity: solutions develop kinks and fronts (Full discretization of the porous medium/fast diffusion equation based on its very weak formulation), and directional degeneracy can actively roughen solutions (Hairer et al.). For a passive tracer this loss of smoothness does not hurt path simulation, since a particle in a direction without noise simply follows the wind. It hurts two other things: grids, which smooth the fronts, and smooth function approximators, which cannot represent kinks. Our design is built around that asymmetry.

The Estimator

Write the receptor concentration as a linear functional of the emissions. The emissions are expressed in a reduced basis $\phi_1,\dots,\phi_K$ of spatial-temporal patterns, for example principal components of the Origins.earth hourly 1 km inventory, following the reduced-basis approach of Efficient PDE-Constrained optimization under high-dimensional uncertainty using derivative-informed neural operators. The Jacobian entry $G_k(\xi)$ is the expected time a backward particle released from the sensor spends in pattern $k$. Let $X_s$ be that particle path, $\sigma$ the turbulent noise matrix (zero in degenerate directions and regions), $W$ a Brownian motion, and $v_k(x,s;\xi)$ a network approximating the expected remaining residence. The estimator is

$$\hat G_k(\xi) = \frac{1}{M}\sum_{i=1}^{M}\left[\int \phi_k(X^i_s)\,ds \;-\; \int \nabla v_k(X^i_s,s;\xi)^{\top}\sigma(X^i_s,s;\xi)\,dW^i_s\right],$$

where $M$ is the number of particles. The stochastic integral has zero mean for any $v_k$, so $\hat G_k$ is unbiased whatever the network quality. Its variance comes only from the error in $\nabla v_k$ projected onto directions where noise acts. That variance vanishes when $v_k$ is exact, which is the martingale representation argument. In fully degenerate directions there is no noise and therefore no variance to remove. In the transition zone, where $K_z$ jumps from near zero to large values at the top of a stable layer, the network will err, and that error raises variance locally without biasing the footprint.

The construction reuses the method of Unbiased Deep Solvers for Linear Parametric PDEs. That paper builds such control variates for 100-dimensional, non-degenerate financial PDEs. We change three things: the operator is degenerate, the output is vector-valued (one entry per emission pattern), and the parameters are meteorological. The parametric input $\xi$ is where the high dimensionality enters. Its dimension is a few dozen, the flux basis adds another few dozen outputs, and integrating footprints over $\xi$ is what turns transport error into a quantified covariance in the inversion.

Research Program

The work proceeds in three stages. In the first stage we build a degenerate test bed where the answer is known. This is a two-dimensional (horizontal plus vertical) column-and-plume model with a prescribed $K_z$ profile that collapses below a moving inversion height and a mean wind that ramps to zero. A very large particle reference serves as ground truth. We measure the bias of an Eulerian finite-volume solver at 1 km and 100 m resolution against the reference, so that the "smoothing penalty" is a measured quantity rather than an assertion. We also measure the variance reduction of the control variate as the degeneracy sharpens. Theory supplies a hypothesis: Posterior consistency and convergence rates for Bayesian inversion with hypoelliptic operators proves, for linear hypoelliptic inverse problems, that a smoother forward operator slows posterior contraction. We predict that artificial numerical diffusion produces the same effect on the inversion, in the form of wider and biased posteriors on near-sensor fluxes. That paper's theory is linear and Gaussian. We test the prediction empirically rather than claim it extends.

In the second stage we move to Paris. We drive the particles with the 1 km WRF meteorology used in Can we use atmospheric CO 2 measurements to verify emission trends reported by cities? Lessons from a 6-year atmospheric inversion over Paris, perturbed into an ensemble over $\xi$. The perturbations are anchored to SIRTA lidar boundary-layer heights. We train $v_k$ on particle rollouts across the ensemble, using a global-in-time training objective in the spirit of Deep Feynman-Kac Methods for High-dimensional Semilinear Parabolic Equations: Revisit. That paper notes that solution-dependent or degenerate diffusion lies outside its scope, which is why we use its training scheme only for the control variate and not as the estimator.

In the third stage we run a Bayesian inversion over 2016–2021 tower data with the Origins.earth prior and CAMS background. Unlike the reference system, this inversion keeps low-wind and nocturnal observations. Their transport error covariance is estimated from the spread of footprints over $\xi$ rather than set by hand.

Baselines

The naive reference point is the current operational practice: the Eulerian inversion of Can we use atmospheric CO 2 measurements to verify emission trends reported by cities? Lessons from a 6-year atmospheric inversion over Paris, which assimilates well-mixed morning and afternoon data, combined with the 2 m/s exclusion rule of An attempt at estimating Paris area CO 2 emissions from atmospheric concentration measurements. It reports agreement within 10% with independent city inventories, an annual decrease of 2% ± 0.6% per year, and a 2020 drop of 13% ± 1%. The 2015 gradient study reports an uncertainty reduction factor of about 2. We rerun this configuration on the same years, prior, and stations. To count as beating it, our system must assimilate more hours of data, reach a larger uncertainty reduction factor, and show lower held-out residuals in low-wind conditions. It must do this while not significantly contradicting the reported 2020 drop or the 2016–2021 trend.

The strongest direct competitor is plain backward-particle Monte Carlo with no control variate. This is the standard Lagrangian treatment of receptor footprints. It shares our lack of grid smoothing, so it isolates the contribution of the learned correction. We give it the same total compute as our method, counting network training against our budget, and the same meteorological ensemble. Beating it means a lower mean-squared footprint error at matched cost, and therefore either more ensemble members or tighter posteriors.

The ablation that isolates the key ingredient uses the trained network $v_k$ directly as a surrogate footprint, with no particles. This is a biased neural emulator, analogous to the forward transport surrogates of Atmospheric Transport Modeling of CO2 with Neural Networks. That work reports global $R^2 > 0.99$ for 90-day forward transport but its largest errors in the surface layer, and it performs no inversion. We expect the surrogate-only footprint to be biased specifically in stable regimes, and the debiased estimator to remove that bias. On the inverse task itself, no reported result was found on Lacuna for any neural transport surrogate. A second ablation trains the control variate on a single meteorological state instead of the ensemble, which tests whether conditioning on the parameters $\xi$ matters.

Why not just…?

Why not just keep discarding low-wind data? This is the cheapest option, and it is defensible if stagnant periods carry little information. The evidence points the other way. The Paris six-year study found that adding morning data, taken while mixing is still developing, materially reduced flux uncertainty (Can we use atmospheric CO 2 measurements to verify emission trends reported by cities? Lessons from a 6-year atmospheric inversion over Paris). Stagnant periods are also when urban enhancements are largest. We settle the question directly: if adding low-wind data under our transport does not increase the uncertainty reduction factor in the synthetic test bed with known fluxes, the proposal fails on its own terms.

Why not just run more backward particles? This is the real threat. Plain particle Monte Carlo is already unbiased and free of numerical diffusion, so the gain from our method rests entirely on variance reduction at fixed cost. The 130.39 variance reduction factor reported by Unbiased Deep Solvers for Linear Parametric PDEs was obtained for non-degenerate dynamics, and we expect a smaller factor near collapsing layers. If the factor in our setting falls below roughly the ratio of training cost to per-run particle cost amortized over the ensemble, we will recommend plain particles. We would then report the degenerate-regime finding and the transport-uncertainty covariance as the contribution instead.

Why not just refine the Eulerian grid or run a WRF ensemble? Refining the grid reduces numerical diffusion but does not remove it at fronts where diffusivity is near zero, and each ensemble member costs a full mesoscale run. The test bed measures front bias as a function of grid spacing. If 100 m grids eliminate the stable-layer bias at an ensemble size we can afford, the grid wins on simplicity, and we will say so.

Why not just train a physics-informed network on the inverse problem? Physics-informed networks fit PDE residuals with smooth derivatives. Their known difficulty with steep and singular solutions is what motivated adaptive sampling in Annealed adaptive importance sampling method in PINNs for solving high dimensional partial differential equations. The kinked footprint at a stable-layer top is the hard case, and such a network offers no guarantee that its errors average out. We include an importance-sampled variant on the test bed only, as a check.

Proposed Experiments

The first decisive check is on the synthetic degenerate column. For each method we report footprint bias and variance against the particle reference as the boundary-layer collapse sharpens. We also report posterior coverage and flux error for known true emissions when low-wind data are included versus excluded.

The second check is on real Paris data from 2016–2021. We hold out towers in turn and score prediction residuals binned by wind speed and stability. This tests whether the inversion's data-fit gain in low-wind bins is genuine rather than an artifact. As an independent spatial check, we compare predicted chord-integrated concentrations with the GreenLITE laser measurements from 2015–2016 over central Paris; that dataset has no reported error numbers on Lacuna, so we establish the WRF-Chem error on it ourselves. We also report the uncertainty reduction factor against the factor of about 2 in An attempt at estimating Paris area CO 2 emissions from atmospheric concentration measurements. Finally, we check consistency with the 13% ± 1% drop in 2020 reported by the Eulerian inversion.

The outcome that would most change our mind is a variance reduction factor near one in stable regimes. That result would mean the learned correction contributes nothing exactly where the method was meant to help.