Satellite estimates of how much carbon plants absorb, called gross primary productivity or GPP, rest on a simple physical product. GPP is the sunlight a canopy absorbs, which is measurable from space, multiplied by a light-use efficiency. That efficiency is the fraction of absorbed light turned into carbon, and plants lower it under heat, dry air, or dry soil. Machine-learning upscaling increasingly replaces this product with large, flexible regressors trained on flux-tower records. The overparameterization paradox says such models can fit their training data exactly and still generalize. In that literature, "benign overfitting" names the case where exact fitting of noisy labels costs little on new data.
This note argues that for flux data, that promise hinges on two numbers practitioners rarely compute. The first is how many inputs the efficiency term actually depends on. The second is how many statistically independent sites and climate regimes the training set contains. The raw count of half-hourly records does not matter much. We propose to measure where the error spike near exact fitting falls, and whether a second descent follows it, under honest spatial holdouts. We also test whether keeping the physical product and learning only the efficiency term moves that spike to a harmless place.
Core Claim
Consider a network that predicts GPP directly from all satellite and weather inputs. Its generalization risk scales with the full input dimension and with the number of independent sites. Now factor the model as absorbed light times a learned efficiency, where the efficiency is bounded between zero and a maximum. Its capacity is then governed only by the few stress drivers that feed the efficiency term. The standard contraction argument gives this directly. Write $f(x)=A(x)\,g(z)$, where $A(x)\le A_{\max}$ is absorbed light computed from satellite canopy absorption and incoming radiation, $g$ is a learned efficiency from a class $\mathcal{G}$, and $z\in\mathbb{R}^k$ holds the $k$ stress drivers. Scaling each prediction by a known factor of at most $A_{\max}$ gives
$$\mathfrak{R}_m(\mathcal{F}) \le A_{\max}\,\mathfrak{R}_m(\mathcal{G}),$$
where $\mathfrak{R}_m$ is the Rademacher complexity, a measure of how well a function class can fit random signs, over $m$ samples. Flux records are strongly autocorrelated within a site. So under a blocking argument, $m$ is the number of independent site blocks rather than the number of records.
The empirical prediction follows from double-descent theory. The error spike appears where model width matches the number of fitted constraints. Under leave-site-out evaluation, that spike and the regime after it should be set by $k$ and by the site count, not by the full dimension and the record count. The spikes seen under random splits should mostly reflect temporal leakage.
The open question is whether exact fitting is benign here at all. We treat the answer as unknown and do not assume it.
Why the Three Directions Meet Here
The overparameterization theory supplies the hypotheses. Deep learning: a statistical viewpoint shows that exact fitting is benign when the input covariance has a small number of strong directions plus a long, heavy tail that absorbs noise. It states this through two effective-rank conditions. Risk Bounds for Over-parameterized Maximum Margin Classification on Sub-Gaussian Mixtures locates an eigenvalue-decay threshold that separates dimension-dependent from dimension-free behaviour. Both results assume independent, roughly Gaussian data. Neither has been checked on collinear Earth-observation inputs, and doing that check is part of our work.
Memorize to generalize proves that near-exact fitting is necessary for optimal risk in overparameterized linear regression. Its result is strongest as label noise shrinks. Eddy-covariance GPP carries large noise, including noise from the step that partitions net flux into photosynthesis and respiration. That result therefore motivates the test but does not settle it, and it may simply fail to apply here.
Exact Gap between Generalization Error and Uniform Convergence in Random Feature Models shows that bounds over a whole norm ball can become vacuous even while true risk falls. This constrains how we use the bound above. We use it to rank model families by predicted capacity, not to certify accuracy. The dimension-bound papers play a similar constraining role. Zhang et al. show that guarantees which ignore the training algorithm cannot escape dimension. That is why we compare the actual solvers used in practice, minimum-norm fitting and ridge regression, rather than all possible fitted models.
The flux literature supplies both the structure and the reason to believe $k$ is small. Physics Informed Modeling of Ecosystem Respiration via Dynamic Mode Decomposition with Control Input finds that respiration follows four to six dominant dynamic modes on FLUXNET2015. Our proposal treats that finding as evidence for low intrinsic dimension, not as a modelling method. Observed strong atmospheric water constraints on forest photosynthesis shows that the dominant stress, dry air versus dry soil, switches with climate. The switch depends on whether rainfall and sunlight peak together in the season. This gives a physically meaningful shift on which to test out-of-regime generalization.
Research Program
We build two model families on the same FLUXNET2015 daily data. The inputs are MODIS canopy absorption and vegetation indices, ERA-5 meteorology, and GOSIF fluorescence, all as used in Improved Global Gross Primary Productivity Estimation by Considering Canopy Nitrogen Concentrations.
The first family is unstructured. It is a random-feature regressor, plus a two-layer network trained end to end, that maps all roughly 14 to 20 inputs to GPP.
The second family is structured. It computes absorbed light physically and learns a bounded efficiency from $k$ stress drivers. The drivers are air temperature, vapor pressure deficit, soil water, and incoming radiation to allow for light saturation. Canopy nitrogen can be added optionally through the green chlorophyll index. Plants can also use diffuse light more efficiently than direct light. If that effect breaks the pure product, it will show up as structured residuals, and we will report it.
For each family we sweep width from far below to far above the number of training blocks. At each width we fit both the minimum-norm exact fit and ridge regression with a cross-validated penalty.
We also compute the eigenvalue spectrum of the feature covariance at each width. From it we evaluate the Bartlett effective ranks. This tells us whether the spectral conditions for benign fitting are predicted to hold before we look at test error.
On the theory side, we write out the bound for the product class with block-dependent data. We report the bound's value and its ranking alongside measured risk, following the risk-versus-bound comparison of Yang, Bai and Mei.
Baselines
The naive reference is a climatological mean seasonal cycle per plant functional type. Any model must beat this under leave-site-out splits to count as useful.
The classical physical baseline is NI-LUE from Improved Global Gross Primary Productivity Estimation. It is a light-use-efficiency model with nitrogen-adjusted maximum efficiency, calibrated by a global optimizer. It reports daily $R^2 = 0.63 \pm 0.01$ and RMSE $= 2.27 \pm 0.05$ gC m⁻² d⁻¹ under 10-fold cross-validation on FLUXNET. We rerun it on our splits, since its folds may not hold out whole sites.
The strongest data-driven competitor is LightGBM, following Assessment of Six Machine Learning Methods for Predicting Gross Primary Productivity in Grassland. It reports $R^2 = 0.971$ and RMSE $= 3.149$ on Mongolian Plateau grassland. That number comes from a single regional setting, apparently without site holdout, so we treat it as an upper reference and not a target. LightGBM gets the same inputs and the same tuning budget of 100 hyperparameter trials, selected by inner leave-site-out folds.
The key ablation separates the physical factorization from mere dimension reduction. We replace the physics-chosen stress drivers with the top $k$ principal components of all inputs, keeping the same width sweep.
The proposal succeeds only under three conditions. The structured family must match or beat LightGBM on leave-site-out RMSE. It must beat the principal-component ablation. Its interpolation behaviour must be the one predicted by the spectral diagnostic.
Why not just…?
Why not just use LightGBM, which already fits grassland GPP very well? The reported accuracy of Li et al. does not come from a site holdout, and it says nothing about behaviour under climate-regime shift. LightGBM is a genuine threat on in-distribution accuracy, and we expect it to win on random splits. The question is whether it keeps that lead on the synchrony-regime and drought-year holdouts. Our experiments measure exactly that, and a LightGBM win there would be reported as the main finding.
Why not just use ridge regression and skip the interpolation question? This is the most likely winner. Flux noise is large, which is the regime where Cheng et al.'s case for exact fitting is weakest. We therefore make tuned ridge the reference point. The outcome variable becomes the gap between ridge and minimum-norm fitting at each width and split type. If that gap is small only in the structured family, the result tells practitioners when interpolating upscalers are safe. If ridge dominates everywhere, that is a clean negative result for benign overfitting on flux data.
Why not just reduce inputs with principal components or dynamic modes? Bartlett et al. argue that truncating the covariance tail removes the part of the spectrum that absorbs noise. That argument assumes independent Gaussian data, so it may not transfer. The principal-component ablation tests directly whether physical choice of inputs matters beyond low dimension.
Decisive Checks
We use three splits on FLUXNET2015: random days, leave-site-out, and leave-regime-out, which trains on synchronous-climate forests and tests on asynchronous ones following Su et al. We add a fourth holdout of heat and drought years. Metrics are daily RMSE and $R^2$, with separate RMSE on the drought subset.
The central hypothesis fails if either of two things happens under leave-site-out evaluation. The first is that the error spike stays tied to record count rather than block count. The second is that the structured family shows no shift in its spike or second descent relative to the unstructured one. Either outcome would mean the physical product buys no statistical capacity.
A second check compares the spectral diagnostic with measured behaviour. If the effective-rank conditions predict benign fitting and minimum-norm fitting still loses badly to ridge, the linear theory does not transfer to flux data. We will report that as a limit of the theory.
References
- Deep learning: a statistical viewpoint
- Risk Bounds for Over-parameterized Maximum Margin Classification on Sub-Gaussian Mixtures
- Memorize to generalize: on the necessity of interpolation in high dimensional linear regression
- Exact Gap between Generalization Error and Uniform Convergence in Random Feature Models
- Uniform Convergence and Generalization for Nonconvex Stochastic Minimax Problems
- Physics Informed Modeling of Ecosystem Respiration via Dynamic Mode Decomposition with Control Input
- Observed strong atmospheric water constraints on forest photosynthesis using eddy covariance and satellite-based data across the Northern Hemisphere
- Improved Global Gross Primary Productivity Estimation by Considering Canopy Nitrogen Concentrations and Multiple Environmental Factors
- Assessment of Six Machine Learning Methods for Predicting Gross Primary Productivity in Grassland