Privacy Noise as a Known Term in the Emission Inversion Prior

Cities and satellite missions estimate greenhouse-gas emissions in two steps. First, a "bottom-up" inventory turns activity data such as fuel sales, heating use and traffic into a gridded map of emissions. Second, an atmospheric inversion corrects that map using measured CO₂ and a model of how air moves. The inversion only works if it is told how wrong the inventory is likely to be, cell by cell and sector by sector. Today that error description is mostly guessed. The finest-grained activity data, from smart meters, vehicle telematics and household fuel deliveries, is also the most privacy-sensitive, so inventories increasingly face pressure to be compiled from data that individuals or firms have randomized before handing it over.

This proposal makes one observation and builds on it. If the activity data is collected under the shuffle model of differential privacy, part of the inventory's error stops being a guess. In the shuffle model, each contributor adds a small, calibrated amount of noise locally, and an intermediary then strips identities and mixes the reports, so that the group's anonymity multiplies the privacy protection. The noise this adds is known exactly, is independent across shuffle groups, and has a variance we can compute from the privacy parameters. We propose to write this noise into the inversion's prior error model explicitly, and then to use the inversion's own information content to decide how reports are grouped and clipped. The aim is to state, before deployment, how much privacy can be granted before atmospheric data can no longer detect a given emission change, such as a lockdown-scale traffic drop.

Core Claim

The prior error covariance $\mathbf{B}$, the matrix telling the inversion how uncertain and how spatially correlated the inventory errors are, should be split into two parts:

$$\mathbf{B} = \mathbf{B}_{\text{phys}} + \mathbf{D}_{\text{priv}}(\epsilon, \delta, \{n_g\}, C).$$

Here $\mathbf{B}_{\text{phys}}$ is the usual heuristic description of ignorance about emission factors and activity, for example isotropic decay with a 50 km correlation length. $\mathbf{D}_{\text{priv}}$ is the privacy-induced term. It is block-diagonal over shuffle groups $g$, each containing $n_g$ contributors, and depends on the central privacy level $(\epsilon,\delta)$ and on the per-contributor clipping bound $C$. Its entries are computed exactly from the mechanism and a numerical privacy accountant.

The quantity we care about is the posterior variance of a target budget $q = \mathbf{w}^T\mathbf{x}$, such as Paris residential emissions for March–May. Here $\mathbf{x}$ is the vector of cell-and-sector emissions, $\mathbf{w}$ selects and sums the cells and sectors that make up the target, $\mathbf{H}$ is the transport operator linking emissions to observations, and $\mathbf{R}$ is the observation error covariance:

$$\sigma_q^2 = \mathbf{w}^T\left(\mathbf{B}^{-1} + \mathbf{H}^T\mathbf{R}^{-1}\mathbf{H}\right)^{-1}\mathbf{w}.$$

We make two claims.

The first is about calibration. Inversions that include the exact $\mathbf{D}_{\text{priv}}$ give calibrated posterior uncertainties. Inversions that absorb privacy noise by inflating $\mathbf{B}_{\text{phys}}$ ad hoc are miscalibrated in a predictable direction. Privacy noise is spatially white at the scale of a shuffle group, so treating it as correlated physical error spreads corrections across neighbouring cells that should not receive them. This is the fragmentation-versus-smearing failure that Carouge et al. showed dominates accuracy at 50 km.

The second is about design. Because $\sigma_q^2$ is differentiable in the shuffle-group layout and in the clipping bound, these choices can be optimized for the inversion rather than for raw inventory error. Two rules follow. Sparse cells should be pooled into larger shuffle groups up to roughly the physical correlation length, because merging at that scale costs little in the inversion but gains amplification. Clipping should be set per sector, because heavy-tailed emitters make naive clipping a source of bias rather than variance.

Prior Work and Its Role

The motivation comes from atmospheric work that identifies the prior error description as the binding constraint. Nalini et al. ran a 1 km Lagrangian inversion over Paris during the 2020 lockdown and named poorly known error magnitudes and correlation lengths as the main limit on separating traffic from non-traffic emissions. Carouge et al. found fine-scale accuracy highly sensitive to the assumed error structure. Santaren et al. showed that observations constrain diffuse, high-uncertainty sources poorly. The PMIF studies (Wang et al., Lespinas et al.) showed that posterior uncertainty from satellite snapshots hinges on assumed error correlations.

All of these treat $\mathbf{B}$ as purely epistemic. None considers a component of inventory error that is known by construction.

The privacy side supplies the method. Feldman et al. and Hiding Among the Clones give the amplification bounds that tie local noise to central privacy. Koskela et al. give tight numerical accounting, which we need so that $\mathbf{D}_{\text{priv}}$ is not inflated by loose bounds. Takagi and Liew give the Blanket-Mixed Gaussian mechanism. In it, each client sends pure noise with some probability and noise centred on its true vector otherwise, and its error matches the central Gaussian mechanism for vector means. We reuse it for the six-sector reports. Their analysis assumes unconstrained vectors in the unit ball and stops at the mean estimate; we carry the error through to a physical inverse problem.

Two papers constrain the design. Privacy Implications of Shuffling shows that uniform shuffling destroys spatial trends, which rules out one global shuffle and motivates per-group shuffling with public group labels. Enhancing Metric Privacy With a Shuffler shows privacy degrading sharply under a compromised shuffler for some mechanisms. We therefore also report the fallback case, in which only local guarantees hold.

Research Program

We scope privatization to diffuse sectors: residential heating, tertiary buildings and private road transport. Large point sources are typically publicly reported or satellite-visible, since the PMIF studies place the satellite-constrainable regime above roughly 0.5–2 MtC/yr. This scoping is deliberate. It puts privacy noise exactly where the atmosphere constrains least, which is the hard case.

The program has four steps.

First, we derive $\mathbf{D}_{\text{priv}}$ for per-group shuffling with three components: scalar mechanisms for single-sector totals, the Blanket-Mixed Gaussian mechanism for sector vectors, and clipping. The derivation separates the clipping bias, which is a mean shift growing with the heavy tail, from the variance term. It also handles non-negativity by post-hoc projection, with the projection-induced bias tracked explicitly.

Second, we build a grouping optimizer. It merges cells into shuffle groups (a quadtree over the grid) and chooses per-sector clipping bounds to minimize $\sigma_q^2$ for a set of target budgets at a fixed central $(\epsilon,\delta)$. Group labels are public, so the analyzer learns group totals, and the optimizer is constrained to a minimum group size.

Third, we treat longitudinal reporting under its own budget. Each contributor reports monthly, privacy composes across reports, and accounting uses the numerical privacy-loss methods. We measure how the resulting per-report noise interacts with the temporal error correlations that satellite extrapolation depends on.

Fourth, we test calibration. We check whether posterior intervals cover truth at nominal rates.

Baselines

The do-nothing reference uses only public data. The prior is ODIAC spatial emissions with TIMES temporal profiles, with a heuristic $\mathbf{B}$, as in the PMIF setups. It shows what the private activity data buys at all; if privatized data cannot beat it, the program fails.

The non-private upper bound uses the true activity-derived inventory, the synthetic truth plus physical error only, with no privacy noise. This is the configuration of Nalini et al.. They report a 2–10% uncertainty reduction near central-Paris sites, a 43% posterior versus 37% prior reduction in fossil CO₂, and 68% versus 29% traffic versus non-traffic reductions. These numbers serve as the strongest existing reference for what a 1 km urban inversion achieves without privacy.

The strongest privacy competitor is central DP. A trusted agency adds Gaussian noise to cell-sector totals at the same $(\epsilon,\delta)$, and the inversion receives the same exact $\mathbf{D}_{\text{priv}}$. No reported result was found on Lacuna for any privatized-inventory inversion, so we generate this baseline ourselves.

The key ablation keeps everything identical, including the mechanism, the privacy level and the data. It replaces the exact $\mathbf{D}_{\text{priv}}$ with the practitioner's move of inflating $\mathbf{B}_{\text{phys}}$ variances by a factor tuned on held-out OSSE months. The tuning gets a grid search of the same size we spend on the grouping optimizer. A second ablation fixes grouping to one shuffle group per cell, which isolates the contribution of inversion-aware pooling.

All configurations share the same transport operator, observation errors, synthetic truth, random seeds and privacy accountant. The proposal beats a baseline if it does at least one of two things at matched $\epsilon$. One is to cut target-budget RMSE relative to the inflated-$\mathbf{B}$ ablation and the per-cell grouping. The other is to bring 90% posterior-interval coverage within a few points of nominal where the ablation misses by more. Beating central DP is not required. Matching it within noise is the success criterion for the shuffle variant, since Takagi and Liew show that the Blanket-Mixed Gaussian mechanism tracks the central Gaussian curve. They report $\text{MSE} = d\chi^2(1+o(1))$ against a lower bound of $d\chi^2$, but no numeric RMSE value appears in the file.

Why not just…?

Why not just let a trusted agency run central DP? The inversion-side contributions, the exact $\mathbf{D}_{\text{priv}}$ and the grouping optimizer, apply equally to central DP, and we say so; central DP is included as a reference. The case for shuffling is about trust, not accuracy. Firms and households in cross-jurisdictional verification may refuse to hand raw activity to any curator. If shuffle results match central results within noise, the claim is that this trust requirement can be dropped at no cost to the inversion. If they do not match, the experiment tells us that too.

Why not just inflate the prior variances and keep the usual isotropic $\mathbf{B}$? This is the real threat. Carouge et al. found that simple isotropic structures beat elaborate ones, so a tuned inflation could plausibly match us on RMSE. Our argument rests on structure: privacy noise is white across groups and its size varies with group population, while the physical error is correlated. A single inflation factor gets both the spatial pattern and the rural–urban contrast wrong. The ablation settles this. If tuned inflation matches both RMSE and coverage, the exact term adds nothing, and we would report that as the finding.

Why not just coarsen the inventory instead of adding noise? Coarsening gives no guarantee against facility re-identification when combined with auxiliary location data, as Privacy Implications of Shuffling illustrates. It also discards the 1 km gradients that Nalini et al. needed for sector attribution. Our grouping optimizer coarsens only the shuffle groups, where the inversion's correlation length says this is cheap. We compare against uniform coarsening matched on target-budget RMSE, and report the re-identification success of a linkage attack on known facility locations under each.

Why not just use local DP with no shuffler? Error grows by roughly $\sqrt{n}$ over the central level (Feldman et al.). We run local DP anyway because it is exactly the fallback when the shuffler is compromised, and we report the posterior $\sigma_q$ it yields.

Proposed Experiments

The primary testbed is an observing system simulation experiment, in which a known synthetic truth is fed through a simulated measurement system so that errors can be scored exactly. It is built on the PMIF setup (PMIF v1.0), chosen because its Gaussian-plume transport is cheap enough for privacy sweeps. It uses ODIAC/TIMES priors, CCMP winds and a simulated CO2M imager with a 300 km swath at 2 km resolution. Synthetic contributors are generated by disaggregating cell-sector totals into households, vehicles and buildings with heavy-tailed sizes. We sweep central $\epsilon$ from 0.1 to 10, contributor density, and single-imager versus constellation sampling.

The headline output is a curve showing, for each $\epsilon$, the posterior uncertainty of city-scale diffuse-sector budgets, and the $\epsilon$ at which they cross the 20% threshold PMIF uses.

The secondary testbed is a Paris-style 1 km configuration. It uses the Origins.Earth sectoral inventory as truth, if Lagrangian footprints can be obtained from the authors, and asks whether a lockdown-size traffic drop remains detectable as $\epsilon$ shrinks.

The third testbed is a European network configuration modelled on Wang et al.. It uses a PKU-CO2 prior, IER-EDG truth and the NET17/NET43 station sets, and tests whether conclusions hold under coarse transport. That paper reports more than 30% monthly uncertainty reduction in western Germany with NET17 and 40–70% in high-emitting areas with NET43, which serve as non-private reference points.

Three checks decide the outcome. The calibration check compares coverage under exact $\mathbf{D}_{\text{priv}}$ versus tuned inflation; if they tie, the core claim fails. The pooling check measures RMSE gains from the optimizer over per-cell groups in sparse suburban and rural cells, where the payoff should concentrate. The longitudinal check tests whether monthly composition forces enough per-report noise to erase the temporal correlation that satellite extrapolation requires. This is the tension the research file flags, and a negative result there would set a hard limit on how often private activity can be reported.