Planetary evolution models couple a slowly cooling interior to an atmosphere that insulates it. Volcanoes release carbon dioxide and water, and the resulting greenhouse effect sets how fast the interior cools and melts. These models contain sharp tipping points. A stagnant-lid Venus either keeps its oceans for hundreds of millions of years or tips into an irreversible runaway greenhouse, depending on its starting carbon inventory. Researchers who want to scan thousands of hypothetical exoplanets, or fit these models to spectra, increasingly replace the simulator with a cheap learned approximation, called a surrogate. Nobody can currently say whether a surrogate that fails near a tipping point failed because it was badly trained or because its model class cannot represent the jump.
This note proposes to answer that question with a certificate. From theoretical computer science we borrow the dual polynomial: a weighting of input points that provably defeats every polynomial below a given degree, obtained from linear-programming duality. From the study of function composition we borrow the fact that polynomial degree multiplies when functions are nested. The claim is that the tipping points of coupled interior–atmosphere models are manufactured by the feedback loop. Each one-step physical update (interior cooling, outgassing, climate response) has low certified degree. The multi-step coupled evolution map has high certified degree, and that degree is concentrated on the threshold surface. If true, the result gives a concrete design rule: approximate the components and iterate them in the loop rather than emulating the end-to-end outcome. It also gives a way to certify, rather than guess, how many simulations and how much surrogate capacity a given planetary question needs.
Core Claim
Take a coupled evolution model with input parameters $x$ in a box $B$ (initial CO₂ inventory, initial mantle temperature, oxygen fugacity, orbital distance). Let $F(x)$ be an outcome of interest, such as the time liquid water survives or the final atmospheric CO₂. For any finite grid $X \subset B$ and degree $d$, the best achievable worst-case error of a degree-$d$ polynomial on $X$ equals the value of a dual linear program:
$$\epsilon_d(X) \;=\; \max_{\psi} \sum_{x\in X}\psi(x)F(x) \quad\text{s.t.}\quad \sum_{x\in X}|\psi(x)|=1,\;\; \sum_{x\in X}\psi(x)p(x)=0 \text{ for all } p \text{ with } \deg p\le d.$$
The weighting $\psi$ is the dual witness. Hölder's inequality shows that no degree-$d$ polynomial can approximate $F$ on $B$ to within worst-case error below $\epsilon_d(X)$, because $X$ is a subset of $B$. So a finite set of simulator runs yields a rigorous lower bound on approximation error over the whole domain.
We make three claims. First, the support of $\psi$ concentrates on the runaway threshold, so the certificate locates the difficulty. Second, the one-step component maps have small certified degree $d_c$. Third, the certified degree of the $T$-step coupled map grows roughly like the composition bound $\deg(g^{\circ T}) \le d_c^{\,T}$ until it saturates at the degree forced by the threshold's sharpness. The composition bound holds because substituting a degree-$d_c$ polynomial into itself multiplies degrees. The practical consequence is that modular surrogates, which learn low-degree components and are then iterated, should beat end-to-end surrogates at equal simulation budget near tipping points, and match them away from those points.
Positioning
The mechanism comes from the complexity-theory literature. Dual lower bounds for approximate degree and Markov-Bernstein inequalities builds explicit dual witnesses and ties them to Markov–Bernstein inequalities, the classical fact that a low-degree polynomial cannot be steep. That fact is exactly why a sharp tipping point forces high degree, so we reuse the paper's continuous-domain perspective. Hardness Amplification and the Approximate Degree of Constant-Depth Circuits proves that one-sided hardness amplifies under OR-composition (error at least $1-2^{-t}$ for $t$ copies). A Nearly Optimal Lower Bound on the Approximate Degree of AC0 and Classical Lower Bounds from Quantum Upper Bounds show how degree behaves for nested compositions, the latter proving $\widetilde\Theta(\sqrt n\,\widetilde{\deg}(f))$ for $\mathrm{OR}_n\circ f$. All of these work on the Boolean hypercube with hand-built witnesses for worst-case functions. We instead solve the dual program numerically on grids of physical simulator outputs. We do not need explicit constructions, and we give up asymptotic statements in exchange for concrete, instance-level certificates.
Composition work motivates the loop hypothesis. The composition complexity of majority shows that assembling majority from local pieces costs an unavoidable $\log k$ overhead. Composition limits and separating examples for some boolean function complexity measures studies what happens when a function is composed with itself indefinitely, which is the structure of a time-stepped coupled model. How Compositional is a Model? proves that fixed computation graphs pay approximation penalties on data-dependent structure. None of these touches physical feedback loops. The research file found no paper connecting either direction to planetary modeling.
The application regime comes from coupled planetary models. Early Habitability and Crustal Decarbonation of a Stagnant‐Lid Venus couples stagnant-lid convection to a grey radiative model. It reports water retention of up to 900 Myr with weathering versus about 400 Myr without. It finds that 5×10¹⁹ mol CO₂ allows climate regulation while 1×10²⁰ mol triggers ocean evaporation, and that final CO₂ is bimodal by at least an order of magnitude. That is a documented, one-sided tipping point, which makes it our main target. Evolution and Spectral Response of a Steam Atmosphere for Early Earth with a Coupled Climate–Interior Model couples interior cooling to line-by-line radiative transfer and high-temperature chemistry. Its magma-ocean lifetimes (about 1 to several Myr) supply a second target that is smoother but more expensive per run.
On the surrogate side, SamudrACE is the closest precedent for modular coupled emulation. It couples separately trained atmosphere and ocean emulators, but it needed two-stage fine-tuning to stop drift and it smooths extremes. It offers no account of which pieces are intrinsically hard. Deep Polynomial Chaos Expansion is the strongest polynomial surrogate family in the file. It fits orthogonal-polynomial expansions and can only provide upper bounds, which is the gap the dual certificate fills.
Research Program
The first phase is a reference simulator. We re-implement a stagnant-lid interior–climate model in the style of the Venus paper above, with parameterized convection, melt-driven outgassing, weathering, carbonate burial and decarbonation, and grey radiative balance. We expose both the end-to-end outcome map and the one-step component maps. The first component is the interior-plus-outgassing update given surface temperature. The second is the climate response given the atmospheric inventory. We validate against the published 900/400 Myr and 5×10¹⁹/1×10²⁰ mol figures before running any analysis. A second, smaller study uses a magma-ocean steam model with tabulated outgoing longwave radiation, meaning the infrared energy the planet radiates to space. This checks whether the pattern holds in a smoother regime.
The second phase is the certificate computation. For each map we solve the primal and dual Chebyshev programs in a Chebyshev polynomial basis, for degrees up to about 30 in four dimensions. That gives roughly 46,000 basis functions, a large but tractable sparse linear program. We use nested grids so that $\epsilon_d(X)$ can be seen to stabilize. The outputs are a certified error-versus-degree curve for each map, the certified degree $d_\epsilon$ at the error target, and the support of $\psi$. We quantify concentration of that support by the fraction of $|\psi|$ mass lying within a band around the runaway surface, where the band is identified independently from simulator bimodality.
The third phase tests the composition hypothesis. We measure $d_\epsilon$ for the $T$-step map as $T$ increases, and compare it with $d_c^{\,T}$ and with the saturation level implied by the threshold's measured steepness through the Markov–Bernstein argument. We also test one-sidedness. Irreversible runaway resembles the one-sided structure that Hardness Amplification and the Approximate Degree of Constant-Depth Circuits exploits, so we check whether one-sided error (never falsely predicting habitability) is certifiably costlier than two-sided error.
The fourth phase applies the results. We use the dual support as a sampling rule that concentrates new simulator runs where $|\psi|$ is large. We compare end-to-end surrogates against modular surrogates that learn the low-degree components and iterate them. Everything is certified for polynomial surrogates. For neural surrogates the certificate is a diagnostic only, and we say so throughout.
Baselines
The naive reference point is the simulator itself. It sets the cost scale: surrogates only matter where repeated runs are expensive, which holds for the line-by-line magma-ocean model more than for the grey Venus model. The naive surrogate reference is an end-to-end degree-$d$ least-squares polynomial fit on uniformly sampled runs.
The strongest polynomial competitor is Deep Polynomial Chaos Expansion trained end-to-end. That paper reports a relative mean squared error of about 0.0001 on its Steady-state Diffusion task, a smooth problem. There is no reported result on Lacuna for any surrogate on the stagnant-lid Venus or magma-ocean outcome maps, so we establish those numbers ourselves. The strongest neural competitor is an end-to-end multilayer perceptron. PDEBench reports that the Fourier Neural Operator generally outperformed U-Nets and PINNs on forward prediction, but its targets are fields, not parameter-to-outcome maps, so we use it only as a sanity check on method behavior near shocks.
All surrogates receive identical simulator budgets (500, 2,000 and 10,000 runs), identical held-out dense test grids, and equal hyperparameter search budgets measured in trials. Beating a baseline means lower worst-case and threshold-band error at equal budget, or equal error with fewer runs.
The ablations isolate each ingredient. Modular-versus-end-to-end surrogates at matched capacity isolate the composition claim. Dual-support sampling is compared with uniform sampling and with residual-driven active sampling, which picks the points where the current fit errs most; this isolates the certificate's sampling value. A classify-then-regress piecewise surrogate tests whether knowing the threshold is all that matters.
Why not just…?
Why not just fit a polynomial chaos expansion and report held-out error? Held-out error is an upper bound on what the model class can do. When it is large near the threshold, it cannot separate bad optimization from impossibility, and DeepPCE's own authors note sensitivity to initialization. The dual program gives the matching lower bound on the same grid. Where the two curves meet, the question is settled. Where they do not, the gap itself tells us the fit is under-optimized.
Why not just detect the regime with a classifier and fit smooth models on each side? This is a real threat. If the threshold surface is itself low-degree, a piecewise surrogate may match the modular surrogate at lower cost. The composition hypothesis actually predicts that this surface is smooth, since it is generated by smooth components. We include the piecewise model as a baseline and will report plainly if it ties. The certificate still contributes in that case, by locating the surface for the classifier and certifying that the unsplit problem is hard.
Why not just use residual-driven active learning to find the hard region? Residual-driven sampling chases the current model's mistakes, which depend on the model. The dual support is a property of the target map and the degree alone. The sampling ablation measures whether this difference matters in practice. If residual sampling matches dual sampling, we drop the sampling claim and keep the certification claim.
Why not just run the 1D code, since it is cheap? For the grey stagnant-lid model, surrogates buy little speed, and we treat that model mainly as a clean testbed for the certificates. The line-by-line radiative transfer and chemistry in the magma-ocean model are far more expensive per run. They are where the modular-versus-end-to-end rule has practical payoff for inference sweeps.
Decisive Experiments
The composition claim fails if the certified degree of the one-step components is comparable to that of the end-to-end map. It also fails if the degree does not grow with the number of coupling steps before saturation. Either result would mean the sharpness lives inside a component, for example the evaporation step, rather than in the loop, and the modular design rule would lose its justification.
The localization claim fails if under half of the dual weight lies in a narrow band around the independently identified runaway surface. The practical claim needs modular surrogates to cut threshold-band worst-case error relative to end-to-end DeepPCE and the multilayer perceptron at 2,000 runs, and to reach end-to-end accuracy with substantially fewer runs, on both the Venus and magma-ocean models. A positive magma-ocean result, with its smoother outgoing-radiation-limited dynamics, would show that the effect is not an artifact of one bimodal system. A negative result there would bound the claim to systems with irreversible tipping points, which is still a useful and honest scope.
References
- Dual Lower Bounds for Approximate Degree and Markov-Bernstein Inequalities
- Hardness Amplification and the Approximate Degree of Constant-Depth Circuits
- A Nearly Optimal Lower Bound on the Approximate Degree of AC0
- Classical Lower Bounds from Quantum Upper Bounds
- The Composition Complexity of Majority
- Composition limits and separating examples for some boolean function complexity measures
- How Compositional is a Model?
- Early Habitability and Crustal Decarbonation of a Stagnant‐Lid Venus
- Evolution and Spectral Response of a Steam Atmosphere for Early Earth with a Coupled Climate–Interior Model
- SamudrACE: Fast and Accurate Coupled Climate Modeling with 3D Ocean and Atmosphere Emulators
- Deep Polynomial Chaos Expansion
- PDEBench: An Extensive Benchmark for Scientific Machine Learning