Learning the climate of dynamical systems with state-space systems

James Murray Louw; Juan-Pablo Ortega

arxiv: 2512.15530 · v2 · submitted 2025-12-17 · 🌊 nlin.CD · math.DS

Learning the climate of dynamical systems with state-space systems

James Murray Louw , Juan-Pablo Ortega This is my paper

Pith reviewed 2026-05-16 22:06 UTC · model grok-4.3

classification 🌊 nlin.CD math.DS

keywords state-space systemsdynamical systemsclimate learningdistribution forecastingstructural stabilitymixing measurestime series modelingchaotic dynamics

0 comments

The pith

If a dynamical system is structurally stable with a mixing or attracting measure, C1-close state-space proxies keep forecasted probability distributions close to the true ones at arbitrarily long horizons.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that state-space systems can replicate the long-term statistical behavior of certain dynamical systems even when they only approximate the true dynamics. When the underlying process is structurally stable and has a mixing or attracting measure, a proxy close in the C1 sense produces distribution forecasts that stay near the true future distributions for all time, provided the initial distribution is regular enough. This matters because point forecasts in chaotic systems diverge quickly due to sensitivity to initial conditions, yet the overall climate remains predictable in a statistical sense. The result establishes that universal families of state-space models can achieve this replication with high accuracy under the stated conditions.

Core claim

If the underlying data-generating process is structurally stable and possesses a mixing or an attracting measure, then a sufficiently regular initial probability distribution remains close to the true future distribution at arbitrarily long time horizons when forecasted by a C1-close state-space proxy. Thus, under these conditions, learning the climate of a dynamic process with a universal family of state-space systems is feasible with arbitrarily high accuracy.

What carries the argument

The C1-close state-space proxy, which approximates the true map closely enough in the C1 topology to preserve structural stability and the mixing or attracting property under iteration.

If this is right

Distribution forecasts remain stable at long horizons while point forecasts stay sensitive to initial conditions.
The long-term statistical climate of the system can be replicated with arbitrary accuracy by sufficiently close proxies.
Universal families of state-space systems become viable for high-accuracy climate learning under the stated conditions.
The result separates the feasibility of statistical replication from the impossibility of exact trajectory prediction in chaotic regimes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This framework could apply directly to approximate models used in weather or ocean forecasting, where exact trajectories are unattainable but statistical properties matter.
Numerical checks on standard examples such as the logistic map at parameter 4 or the Lorenz attractor would provide immediate verification of the distribution closeness.
The result highlights a boundary case: systems lacking structural stability or mixing measures may lose this long-horizon distribution guarantee even under small model errors.

Load-bearing premise

The data-generating process must be structurally stable and possess a mixing or attracting measure.

What would settle it

A concrete counterexample of a structurally stable system with a mixing measure where a C1-close state-space proxy produces iterated distributions that diverge from the true ones at long horizons.

Figures

Figures reproduced from arXiv: 2512.15530 by James Murray Louw, Juan-Pablo Ortega.

**Figure 2.** Figure 2: (a) Mean distance between points in µ 1 t and µ 2 t (blue) and between points in µ 1 t and ˆµ 1 t (orange). The error in point predictions (orange) settles at a value comparable to the distance between unrelated initial conditions (blue) after 18 Lyapunov times of prediction. (b) Squared distance in the MMD metric between µ 1 t and µ 2 t (blue) and between µ 1 t and ˆµ 1 t (orange). The error in distributi… view at source ↗

read the original abstract

State-space systems encompass a broad class of algorithms used for modeling and forecasting time series. For such systems to be effective, two objectives must be met: (i) accurate point forecasts of the time series must be produced, and (ii) the long-term statistical behaviour of the underlying data-generating process must be replicated. The latter objective, often referred to as learning the climate, is closely related to the task of producing accurate distribution forecasts. Empirical evidence shows that distribution forecasts are far more stable than point forecasts, which are sensitive to initial conditions. In this work, we rigorously study this phenomenon for state-space systems. The main result shows that, if the underlying data-generating process is structurally stable and possesses a mixing or an attracting measure, then a sufficiently regular initial probability distribution remains close to the true future distribution at arbitrarily long time horizons when forecasted by a $C^1-$close state-space proxy. Thus, under these conditions, learning the climate of a dynamic process with a universal family of state-space systems is feasible with arbitrarily high accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper proves that C1-close state-space proxies preserve long-horizon distributional closeness for structurally stable systems with mixing measures, which is a clean but conditional result.

read the letter

The main thing to know is that this paper gives a theorem showing state-space systems can learn the climate of a dynamical system: if the true process is structurally stable and has a mixing or attracting measure, then a sufficiently C1-close proxy keeps the forecasted distribution close to the true one at all future times for regular initial measures. This directly addresses why distribution forecasts tend to be more stable than point forecasts in practice. It does a good job connecting that empirical pattern to standard facts from dynamical systems, like continuous dependence of invariant measures under topological conjugacy. The argument appears to rest on those classical tools without obvious circularity. The result is new as a specific statement for state-space proxies rather than a restatement of prior work. The main limitation is that the conditions are strong. Structural stability and the existence of a mixing measure do not hold exactly for many systems of interest, including climate models, so the theorem's reach depends on how well real data approximate them. The abstract gives no proof details or numerical checks, which leaves open whether the regularity assumptions are easy to verify or overly restrictive in applications. Without those, it's difficult to judge how tight the closeness bound is or how sensitive it is to small violations. This is for readers focused on rigorous foundations for nonlinear time series and chaotic forecasting. Someone working on theoretical justification for state-space methods in climate or similar fields would find the theorem useful. It deserves peer review because the claim is stated clearly and grounded in established mathematics, even if the paper will likely need added examples and scope discussion.

Referee Report

2 major / 2 minor

Summary. The paper claims that if the underlying data-generating dynamical system is structurally stable and possesses a mixing or attracting measure, then forecasts produced by a C¹-close state-space proxy keep a sufficiently regular initial probability distribution close to the true future distribution at arbitrarily long time horizons. This conditional result is presented as the main theorem and is used to conclude that learning the climate is feasible with arbitrarily high accuracy for universal families of state-space systems.

Significance. If the theorem holds under the stated hypotheses, the work supplies a rigorous dynamical-systems justification for the empirical stability of distributional forecasts relative to point forecasts. It connects standard facts about structural stability, topological conjugacy, and continuous dependence of invariant measures to the practical task of climate learning, thereby strengthening the theoretical basis for state-space modeling in nonlinear time-series analysis.

major comments (2)

[Main Theorem] Main Theorem (presumably §3): the proof sketch in the manuscript must explicitly invoke and cite the precise result on continuous dependence of invariant measures under C¹ perturbations of structurally stable maps; without this reference the dependence of the distributional closeness on the mixing/attracting assumption remains opaque.
[§2] §2 (Definitions): the class of 'sufficiently regular' initial probability distributions is left imprecise; the argument requires at minimum absolute continuity with respect to the reference measure on the attractor, and this must be stated as a hypothesis to avoid the result holding only for a null set of measures.

minor comments (2)

[Abstract] Abstract: the phrase 'universal family of state-space systems' is used without prior definition; a brief clarification of the function class (e.g., recurrent networks with sufficient width) should appear in the introduction.
[§2] Notation: the precise meaning of 'C¹-close' (norm on the space of maps and their derivatives) should be fixed in §2 with an explicit reference to the C¹ topology.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the careful reading and constructive comments. We will incorporate the suggested clarifications to strengthen the presentation of the main theorem.

read point-by-point responses

Referee: [Main Theorem] Main Theorem (presumably §3): the proof sketch in the manuscript must explicitly invoke and cite the precise result on continuous dependence of invariant measures under C¹ perturbations of structurally stable maps; without this reference the dependence of the distributional closeness on the mixing/attracting assumption remains opaque.

Authors: We agree with this observation. In the revised manuscript, we will explicitly invoke and cite the relevant theorem on the continuous dependence of invariant measures for C¹-close perturbations of structurally stable diffeomorphisms (for example, results following from the work of Ruelle or standard references such as Katok and Hasselblatt). We will expand the proof sketch in §3 to clearly demonstrate how the mixing or attracting property of the measure ensures the distributional closeness at long horizons. revision: yes
Referee: [§2] §2 (Definitions): the class of 'sufficiently regular' initial probability distributions is left imprecise; the argument requires at minimum absolute continuity with respect to the reference measure on the attractor, and this must be stated as a hypothesis to avoid the result holding only for a null set of measures.

Authors: We thank the referee for pointing this out. We will revise §2 to explicitly state that the initial probability distributions are assumed to be absolutely continuous with respect to the reference (Lebesgue or SRB) measure on the attractor. This hypothesis will be added to ensure the result applies to a positive-measure set of distributions. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The central result is a conditional theorem: structural stability plus existence of a mixing or attracting measure implies that C¹-close state-space proxies preserve distributional closeness at arbitrary future times for regular initial measures. This follows directly from standard facts in dynamical systems (topological conjugacy, continuous dependence of invariant measures) without any reduction to the paper's own fitted parameters, self-definitions, or unverified self-citations. The hypotheses are stated explicitly in the abstract and no load-bearing step collapses to an input by construction. The derivation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions from ergodic theory and dynamical systems; no free parameters or new entities are introduced in the abstract.

axioms (2)

domain assumption The underlying data-generating process is structurally stable
Invoked in the main result for the closeness of distributions.
domain assumption The process possesses a mixing or an attracting measure
Required for the long-term behavior to be well-defined and for the proxy to stay close.

pith-pipeline@v0.9.0 · 5477 in / 1240 out tokens · 38168 ms · 2026-05-16T22:06:49.774247+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

if the underlying data-generating process is structurally stable and possesses a mixing or an attracting measure, then a sufficiently regular initial probability distribution remains close to the true future distribution at arbitrarily long time horizons when forecasted by a C¹-close state-space proxy
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

mixing and attracting measures are asymptotically stable fixed points of the Perron-Frobenius operator

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages · 2 internal anchors

[1]

Learning theory for dynamical systems

[Berr 23] T. Berry and S. Das. “Learning theory for dynamical systems”.SIAM Journal on Applied Dynamical Systems, Vol. 22, No. 3, pp. 2082–2122,

work page 2082
[2]

Echoes of the Past: A Unified Perspective on Fading memory and Echo States

[Orte 25a] J.-P. Ortega and F. Rossmannek. “Echoes of the past: a unified perspective on fading memory and echo states”.Preprint. arXiv.2508.19145,

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Stochastic dynamics learning with state-space systems

[Orte 25d] J.-P. Ortega and F. Rossmannek. “Stochastic dynamics learning with state-space systems”.Preprint. arXiv.2508.07876,

work page internal anchor Pith review Pith/arXiv arXiv

[1] [1]

Learning theory for dynamical systems

[Berr 23] T. Berry and S. Das. “Learning theory for dynamical systems”.SIAM Journal on Applied Dynamical Systems, Vol. 22, No. 3, pp. 2082–2122,

work page 2082

[2] [2]

Echoes of the Past: A Unified Perspective on Fading memory and Echo States

[Orte 25a] J.-P. Ortega and F. Rossmannek. “Echoes of the past: a unified perspective on fading memory and echo states”.Preprint. arXiv.2508.19145,

work page internal anchor Pith review Pith/arXiv arXiv

[3] [3]

Stochastic dynamics learning with state-space systems

[Orte 25d] J.-P. Ortega and F. Rossmannek. “Stochastic dynamics learning with state-space systems”.Preprint. arXiv.2508.07876,

work page internal anchor Pith review Pith/arXiv arXiv