A Lens That Sees the Skew: Nonlinear Climate Models ✛ ✎

Original: "Beyond Gaussian Assumptions: A Nonlinear Generalization of Linear Inverse Modeling"

arXiv:2503.24234 · 2025-03-31 · arXiv non-exclusive · independent analysis · 2 min

Standard climate models only see symmetric oscillations, but a new quadratic lens focuses the skew between El Niño and La Niña.

Glasses with flat lenses show a symmetric world: rise and fall, heat and cold look the same. For decades, climatologists have viewed El Niño through such glasses, using linear inverse models. These captured the rhythm of oscillations but missed that the warm phase of El Niño is typically sharper and stronger than the cold La Niña. The skew remained beyond the visible.

The new work proposes bending the lens: adding quadratic terms to the stochastic differential equation. If a linear model is a flat pane, a quadratic one is a curved lens that focuses not only the mean but also the asymmetry. The parameters of such a lens are estimated from correlation functions of observed data up to fourth order. It is like determining the shape of a lens by how it refracts light from different angles.

Key trick: for white noise, the dynamics parameters are expressed in terms of the expectation and the second-, third-, and fourth-order correlation functions at zero, along with their first derivatives. For colored noise, second derivatives and cross-correlations between noise and state are added.

The method was tested on a two-dimensional synthetic system without chaos and on the three-dimensional Lorenz system with chaos. On real El Niño 3.4 data from 1979 to 2021, the new lens produced skewed distributions: the Wasserstein distance for sea surface temperature dropped from 0.0949 for linear models to 0.0555 for the nonlinear one with colored noise. The skew that was once invisible has now emerged.

Climatologists can now examine the asymmetry of other atmospheric and oceanic regimes, no longer settling for symmetric pictures. The lens has become more precise, and the world through it is closer to reality.

Whole graph →

The minimum dimension for chaos in an autonomous differentiable system is three, so the authors first test the method on a two-dimensional system (where chaos is absent) and then on the three-dimensional Lorenz system.

\frac{d}{dt}x_i = \sum_{j,k} B_{ijk} x_j x_k + \sum_j A_{ij} x_j + C_i + \sum_j (\sqrt{2Q})_{ij} \cdot \text{noise}_j

The rate of change of each variable depends on quadratic interactions, linear damping, a constant shift, and random forcing.

0 = \text{Sym}\left( B \times_2 M(0) + A \times_1 K(0) + C \times_0 E^T + Q \right)

In the stationary state, deterministic damping and stochastic forcing balance each other, which allows the noise matrix to be determined.

Q_{C-nLIM} = \arg\min_Q \left( \|R\|_F + \|K(0) - K_{obs}(0)\|_F + \|M(0) - M_{obs}(0)\|_F \right)

The noise matrix is chosen so that synthetic correlations and the balance residual are as close as possible to observed ones.

Concepts

linear inverse model nonlinear laplacian spectral analysis normalizing flows probabilistic model Stochastic modeling gravitational lensing modeling nonlinear dynamics gaussian process Contrastive Learning linear regression quantum-like models Large Language Models Gaussian distribution weighted mean Weakly Supervised Learning matched filter simulated data nonlinearity physics informed neural network lattice simulation

I am the author

A Lens That Sees the Skew: Nonlinear Climate Models ✛ ✎