Flush-Clock Blocking for Honest Uncertainty on City CO2 Trends

Cities increasingly want to check their reported CO2 cuts against the atmosphere. Sensors placed around a city measure how much CO2 the city adds to the air passing over it, and a transport model, which simulates how wind and vertical mixing carry gas from sources to sensors, turns that added CO2 back into an emissions estimate. The Paris system reports a long-term decline of 2% ± 0.6% per year and a 13% ± 1% drop during the 2020 lockdown (Lessons from a 6-year atmospheric inversion over Paris). Those ± values are what a city would cite in a verification claim. They are only trustworthy if the errors behind them are close to independent. They are not. When the transport model misjudges the boundary layer on a calm evening, it stays wrong for many hours, sometimes for days. Treating each hour or day as a fresh error makes the reported uncertainty too narrow.

This proposal addresses that problem with an idea from statistics on dependent data: split the record into blocks that are nearly independent and compute uncertainty from block-level statistics. The new ingredient is where the block boundaries go. They are not placed on the wall clock. They are placed on a "flush clock" that advances as the wind replaces the city's air. A city's CO2 dome resets once the air over it has been ventilated. So one calm day may form a single long block, while a windy afternoon may contain several short ones. The study measures whether this physically timed blocking gives intervals on emission trends and anomalies that actually cover the truth at their stated rate. It compares against the cheaper fixes a sceptic would reach for first.

Core Claim

The central hypothesis is that errors in city-scale inversion are much closer to stationary, fading-memory noise when time is measured in airshed flushes than when it is measured in hours. Stationary here means the error behaves the same way over time; fading memory means its dependence on the past dies out. Let $u(t)$ be the observed wind speed at the city, $L$ the urban fetch (the distance air travels across the city), and $h(t)$ the observed boundary-layer height, the depth of air in which surface emissions mix. Define the flush clock as

$$\theta(t) = \int_0^t \frac{u(s)}{L}\, g\big(h(s)\big)\, ds,$$

where $g$ is a monotone ventilation weight with $g=1$ in a well-mixed afternoon layer. One unit of $\theta$ is roughly one replacement of the urban air. Regeneration points are the times at which $\theta$ has advanced by a threshold $c$ since the previous one. The blocks run between regeneration points, with a buffer of one flush unit discarded at each boundary.

The claim has two testable parts. First, residual autocorrelation functions from different weather regimes should collapse onto a single decay curve when plotted against $\theta$, but not when plotted against hours. Second, trend and anomaly intervals built from flush-clock blocks should reach nominal coverage in controlled tests, while fixed-length blocks with the same average length should not. Fixed-length blocks are too short in calm periods and waste data in windy ones.

Where This Sits

Three strands supply the method, and none has met atmospheric inversion. The first is the method of blocks. Meir's structural risk minimization for β-mixing series (Structural Risk Minimization for Nonparametric Time Series Prediction) uses gapped independent blocks to extend learning guarantees to dependent data. Lasso guarantees under β-mixing (Regularized Estimation in High Dimensional Time Series under Mixing Conditions.) shows that such blocking stays valid even when the underlying dynamics are misspecified. The empirical Bernstein inequality for dependent Hilbert-space data builds data-dependent bounds from interlaced blocks separated by a gap $\tau$. All three assume a single fixed decay rate for dependence. This proposal replaces that fixed gap with a gap set by the weather, and it tests whether the change is needed.

The second strand is thinning, which keeps only observations spaced far enough apart to be nearly independent. Least squares with Markovian data proves that dropping data at the mixing time is minimax optimal in the worst case but wasteful for structured systems. Fast GP approximations for autocorrelated data thins sensor streams at an interval chosen from the partial autocorrelation function and reports a 16% RMSE gain on autocorrelated data. That interval is static and purely statistical. The flush clock is the dynamic, physically grounded version of it, and the Chokhachian code (resource) supplies a direct baseline.

The third strand is decomposition, which splits a time series into trend, repeating cycles and residual. Here it serves as the signal model rather than the main contribution. N-BEATS separates trend and seasonality with constrained basis blocks, and CoST and STEM-LTS enforce separation between components with contrastive and covariance penalties. None of them produces calibrated uncertainty, and N-BEATS gives point forecasts only. In this proposal, decomposition isolates the trend and weekly or heating-driven cycles of the daily emission estimates, so that the residual left over is the thing the blocks must handle.

Two Paris studies motivate the work. The uncertainty-source study documents nighttime misfits of up to 12 ppm and background disagreements between CAMS and CarbonTracker of up to 5 ppm under easterly flow. Both are structured, persistent errors. The 6-year inversion, as the research file summarizes it, uses simplified diagonal or stationary error covariances. One study also constrains the design. The 6-year inversion gained accuracy by adding morning data, so any method that simply throws observations away conflicts with what the field has learned. That is why the approach blocks the record rather than thinning it.

Research Program

The pipeline has three stages.

The first stage is the signal model. The daily posterior emission series from the inversion is decomposed into a smooth trend, a weekly activity cycle, a heating cycle whose amplitude is tied to temperature, and a residual. Tying the heating amplitude to temperature answers the known failure of fixed Fourier bases when seasonal amplitude varies. Two versions are compared: an N-BEATS-style interpretable network and classical STL (seasonal-trend decomposition by local regression) plus regression. Whichever gives cleaner residuals under the controlled tests below is kept.

The second stage is the flush clock. $\theta(t)$ is computed from observed wind and from the boundary-layer heights measured at SIRTA, the observatory the Paris inversion already relies on for these heights. The regeneration threshold $c$ is chosen by one rule, fixed in advance: the smallest $c$ at which the residual lag-one correlation between adjacent blocks falls below 0.1 on a calibration year that is not used for evaluation.

The third stage is certification. Trend slopes and event anomalies, such as the lockdown drop, are estimated with block-level resampling and with an empirical Bernstein bound over blocks, both computed on the flush-clock blocks.

Background error from continental air masses may have longer memory than the urban flush time. The analysis checks for this directly. If the residual autocorrelation in $\theta$ does not decay, a second, slower clock tied to synoptic regime changes is added. Needing that second clock is itself a finding worth reporting.

Baselines

The naive reference is the current treatment. Daily posterior estimates are taken as independent, and the trend interval comes from ordinary regression standard errors or from a diagonal-error posterior. This is the setting behind the reported 2% ± 0.6% per year (Paris inversion). No reported result was found on Lacuna for the empirical coverage of that interval. Measuring that coverage is the first result of this study.

The strongest existing parametric competitor is an inversion whose error covariance decays exponentially in wall-clock time, with the correlation length fitted by maximum likelihood on the calibration year. This is the standard atmospheric fix and a serious threat.

The strongest nonparametric competitor is a fixed-length moving-block bootstrap, with block length selected automatically from the autocorrelation function, together with PACF-selected thinning run through the Chokhachian code (paper). That paper reports a 16% RMSE gain for thinned Scaled Vecchia over standard Scaled Vecchia. No reported result was found on Lacuna for either method on CO2 inversion coverage.

The key ablation uses fixed wall-clock blocks whose mean length equals the mean flush-clock block length. This isolates the effect of placing boundaries by physics rather than simply using longer blocks.

All methods see the same daily posterior series, the same decomposition, and the same calibration year for choosing tuning parameters. Each gets a single tuning pass. Beating a baseline means two things: coverage of nominal 90% intervals within 85–95% in every weather regime, where the baseline misses that range in at least one regime; and a median interval width no more than 20% larger than that of the best baseline that also reaches nominal coverage.

Why not just…?

Why not just use an exponential temporal error covariance in the inversion? This is the most serious alternative. The Lasso β-mixing analysis (paper; Regularized Estimation in High Dimensional Time Series under Mixing Conditions.) argues that parametric error models fail under misspecification, and a single correlation length cannot represent both calm stagnation and windy flushing. That argument is not proof, though. If an exponential kernel whose correlation length depends on wind speed matches flush-clock coverage, then the flush clock should be used inside the covariance instead of for blocking. The experiments include exactly that variant, and the paper would report it as the winner if it wins.

Why not just average to daily values and treat the days as independent? Daily averaging removes dependence between hours but not between days. Background errors of up to 5 ppm tied to multi-day easterly episodes (uncertainty study) persist across days. The regime-stratified coverage test shows directly whether daily independence fails during stagnant spells.

Why not just use a standard block bootstrap with automatically chosen block length? Automatic length selection averages dependence over the whole record. The ablation with matched mean block length is designed to show whether physical placement of the boundaries matters. If it does not, the conclusion becomes simply "use longer blocks," and that is the result the paper would report.

Why not just use STL instead of a neural decomposition? STL may well be enough. The decomposition is not the main contribution. It is chosen empirically by which method leaves residuals with less leftover structure in the controlled tests.

Proposed Experiments

The decisive test is a synthetic-truth experiment on the Paris configuration, using the WRF-Chem setup of the uncertainty study. The inversion is linear in emissions for a fixed transport model. The sensitivities of each sensor to each source can therefore be computed once and reused for hundreds of injected emission scenarios, built from Origins.earth priors with known trends between −4% and 0% per year and step changes of 5–15%. Observations are generated with one transport configuration and inverted with another, using a different boundary-layer scheme and CAMS versus CarbonTracker backgrounds. This produces realistic structured errors without an invented noise model. The outcomes are coverage, interval width, and minimum detectable trend, stratified by wind regime.

The second test checks whether the autocorrelation curves collapse, by comparing the spread of residual ACFs across regimes in $\theta$ against the spread in hours.

A sanity check on synthetic AR processes, with a time-varying mixing rate in the style of the Markovian least-squares setting, confirms that the estimator behaves as expected where the truth is known.

On real Paris data from 2016 to 2021, the evaluation uses the 2020 lockdown as a natural experiment and leave-one-year-out prediction coverage. External comparison is against Carbon Monitor Cities, keeping in mind its own ±21.7% uncertainty.

The finding that would sink the idea is either of two results. If flush-clock blocks give no coverage gain over matched fixed blocks in the synthetic-truth experiment, physical placement adds nothing. If a wind-dependent exponential covariance matches flush-clock blocking at lower cost, the blocking step is unnecessary.