Quantifying internal variability in large-ensemble climate data with the Wasserstein distance

arXiv:2608.14455 · physics.ao-ph, physics.comp-ph · Submitted 2026-08-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Quantifying internal variability in large-ensemble climate data with the Wasserstein distance".

Mira: Here's the gist of the paper. Quantifying internal variability in large-ensemble climate data with the Wasserstein distance proposes a metric based on optimal transport to quantify the relative magnitude of internal variability in single-model initial-condition large ensembles (SMILEs) without requiring additional parameters or depending on distribution shape.

Kai: Metric Proposal and Theoretical Foundation The proposed metric, denoted as γW1, is constructed from the 1-Wasserstein distance to quantify internal variability and the sum of two transport costs to quantify total variability.

Mira: This construction requires only sorting and differencing, involving no additional parameters and applying to non-Gaussian variables because the 1-Wasserstein distance quantifies the difference between any two probability distributions.

Kai: The magnitude of internal variability is given by wint, which is defined as the 1-Wasserstein distance between the distribution of all samples xn,t and that of the ensemble mean mt.

Title and authors: Kai: So we're talking about this paper, "Quantifying internal variability in large-ensemble climate data with the Wasserstein distance." It sounds complicated, but at its core, they are proposing a new way to measure how much of the climate change signal is due to internal chaos versus what's just external forcing.

Mira: Yeah, and it's using something called the Wasserstein distance. That’s a form of optimal transport cost, which basically lets you compare how different probability distributions are related, even if they aren't normally shaped.

Lev: So what does this mean for us when we think about running actual climate models? Is this going to be computationally heavy?

Kai: Well, the paper suggests that calculating this ratio requires only sorting the samples and differencing the resulting quantiles without any extra parameters. That makes it way easier than some other metrics.

Mira: Exactly, and what they're saying is that this new metric reflects the entire distribution shape, including non-Gaussianity, which is something variance-based metrics often miss.

Lev: If it only needs sorting and differencing, that sounds promising for real hardware because we don't have a lot of memory or extra calculation steps to worry about.

Kai: Right. So the big idea here is that we can quantify internal variability in those single-model initial-condition large ensembles, or SMILEs, without needing to guess the shape of the data distribution.

The paper's summary: Mira: The paper breaks down climate variability into forced variability and internal variability. Forced is what comes from external forcing like radiative forcing, and internal is what comes from the chaotic dynamics of the atmosphere and ocean system itself.

Kai: They set up these SMILEs where many simulations start with the same model and forcing but have different initial conditions, so they can compare those different starting points.

Lev: And in a SMILE, they assume that internal variability averages out over enough ensemble members, so the ensemble mean is considered forced variability and the leftover stuff is internal variability.

Mira: They look at several existing metrics for comparing these two components—like the variance-based metric gamma var and the information-theoretic metric gamma info—but they find that their utility for analyzing SMILEs hasn't been fully explored yet.

Kai: This study proposes a new metric, which is based on the Wasserstein distance, to quantify the relative magnitude of internal variability. They define it as the one-Wasserstein distance between the distribution of all samples and that of the ensemble mean, divided by the sum of two transport costs <ref:2608.14455#pg1,divided by the sum of>.

Mira: It’s essentially quantifying how far apart these two distributions are in terms of optimal transport cost, which they claim is very effective because it captures non-Gaussianity.

Lev: So what does this metric actually do for us? It helps us assess uncertainty in climate projections by giving a better measure of the internal versus external signal.

Kai: The paper shows that this ratio is computed without needing any additional parameters, which simplifies things when you're trying to apply it across different datasets.

The paper's improvements: Lev: So what are the specific advantages they claim for this Wasserstein distance metric over the existing ones, like gamma var or gamma info?

Mira: They point out that gamma var is defined as the ratio of internal variability p sigma squared int to total variability p sigma squared int plus sigma squared ext, which is straightforward but it relies on assumptions about variance.

Kai: And then they have gamma info, which uses mutual information and Shannon entropy, but that might be harder to compute reliably sometimes depending on how you model the system.

Mira: The Wasserstein distance metric they propose is better because it incorporates the entire distribution shape, including non-Gaussianity, which means it's more robust when dealing with real climate data.

Lev: They also claim this metric is nearly as easy to compute as gamma var, requiring only sorting and differencing without needing extra parameters. That’s a huge operational win for large-scale analysis.

Kai: And they validated it using synthetic data drawn from Gaussian, uniform, and lognormal distributions, showing that it gives stable estimates regardless of the distribution shape or if there are outliers present in the ensemble.

Mira: They specifically note that for ensembles with about forty members or more, this metric gives stable estimates whether or not outliers are present. That suggests a threshold for reliability in using this method.

Conclusion: Kai: So to wrap up, the paper on "Quantifying internal variability in large-ensemble climate data with the Wasserstein distance" is proposing a way to measure internal variability that uses optimal transport costs instead of relying on simple variance or information theory.

Mira: The main points are that this new metric captures the full distribution shape, handles non-Gaussianity well, and it’s easier to compute than some alternatives because it only requires sorting and differencing.

Lev: It seems like a practical tool for assessing uncertainty in climate projections because it's stable when you have a decent ensemble size, around forty members or more.

Kai: They applied this to the Community Earth System Model Large Ensemble, CESM-LE, looking at two meters air temperature and total precipitation under different forcing scenarios. The finding there is that this metric shows the response to external forcing most clearly for variables with non-Gaussian, heavy-tailed distributions like precipitation.

Mira: That's a key result because it means we can better see how the internal variability contribution changes as external forcing increases, especially in complex variables.

Lev: So for real hardware running simulations, this suggests that if we use a large enough ensemble to get those stable estimates, this Wasserstein distance is a solid way to analyze the results.

Kai: Yeah, it’s certainly a method that reflects the whole distribution rather than just one number derived from it. That's what they are building towards with this work on "Quantifying internal variability in large-ensemble climate data with the Wasserstein distance."

Research Institute for Earth and Information Sciences, Japan Agency for Marine-Earth Science and Technology

physics.ao-ph, physics.comp-ph

Submitted: 2026-08-14

Updated: 2026-11-16

Code: https://github.com/YukiYasuda2718/wasserstein

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Quantifying internal variability in large-ensemble climate data with the Wasserstein distance proposes a metric based on optimal transport to quantify the relative magnitude of internal variability…

Key concepts

1-Wasserstein distance
This distance measures the minimum cost to transform one probability distribution into another. Unlike simple variance, it captures the entire shape of the distribution, making it suitable for non-Gaussian climate data. It quantifies how different two distributions are by finding the most efficient way to move mass from one to the other.
Internal Variability (wint)
This metric calculates the variability within a single model's initial conditions compared to its ensemble mean. It represents how much a single model's starting state fluctuates around the average of all ensemble members, providing a measure of internal uncertainty without needing distribution shape assumptions.
Forced Variability (wext)
This measures the variability introduced by external forcing, comparing the ensemble mean to the total mean. It quantifies how much the system's overall state changes due to external influences like changing greenhouse gas concentrations or solar activity, helping isolate these effects.
Gamma W1 (γW1)
This proposed metric combines internal and forced variability using Wasserstein distances. It is designed to be parameter-free and robust across different distribution shapes, making it a stable tool for assessing relative variability in large climate ensembles.

Terminology

Summary

Quantifying internal variability in large-ensemble climate data with the Wasserstein distance proposes a metric based on optimal transport to quantify the relative magnitude of internal variability in single-model initial-condition large ensembles (SMILEs) without requiring additional parameters or depending on distribution shape. p. 0

Where the paper says it

jp; Abstract Quantifying forced and internal variability in climate data is fundamental to detecting the climate change signal and assessing uncertainty in climate projections. We propose a metric that quantifies the relative magnitude of internal variability in single-model initial-condition large ensembles (SMILEs).

Metric Proposal and Theoretical Foundation

The proposed metric, denoted as γW1, is constructed from the 1-Wasserstein distance to quantify internal variability and the sum of two transport costs to quantify total variability. This construction requires only sorting and differencing, involving no additional parameters and applying to non-Gaussian variables because the 1-Wasserstein distance quantifies the difference between any two probability distributions. p. 0

Where the paper says it

Computing this ratio requires only sorting the samples and differencing the resulting quantiles, with no additional parameters. The metric reflects the entire distribution shape and applies to non-Gaussian variables, because the 1-Wasserstein distance quantifies the difference between any two probability distributions.

The magnitude of internal variability is given by wint, which is defined as the 1-Wasserstein distance between the distribution of all samples xn,t and that of the ensemble mean mt. p. 5
Where the paper says it

(4)], we define quantities that characterize internal and forced variability and take their ratio as the metric (Fig. The magnitude of internal variability is given by the Wasserstein distance wint between the distribution of all samples xn,t and that of the ensemble mean mt: wint := 1 NensNT Nens X NT i=1 |x(i) − m(i) |.

The magnitude of forced variability is given by wext, which is defined as the 1-Wasserstein distance between the distribution of mt and that of the total mean m. p. 5
Where the paper says it

, the absolute difference), and thus wint is the 1-Wasserstein distance. The magnitude of forced variability is likewise given by the 1-Wasserstein distance between the distribution of mt and that of the total mean m: wext := 1 NensNT Nens X NT i=1 |m(i) − m|.

Comparison with Existing Metrics

The paper compares γW1 against two existing metrics: the variance-based metric γvar and the information-theoretic metric γinfo. The variance-based metricγvar is defined as the ratio of the magnitude of internal variability pσ2int to that of the total variability pσ2int + σ2ext. p. 1

Where the paper says it

(2024) defined an information-theoretic metric that uses the Shannon entropy and mutual information to evaluate the magnitudes of forced and internal variability as distribution-based uncertainties. They applied the metric to non-Gaussian variables and showed that it is more robust to outliers than the variance-based metric.

The information-theoretic metricγinfo is defined as 1 − I(X; M) / H(X), where I(X; M) denotes the mutual information and H(X) the Shannon entropy. p. 4
Where the paper says it

The information-theoretic metric γinfo is defined as follows (Sane et al. 2024): γinfo := 1 − I(X; M) H(X) = H(X | M) H(X) , (7) H(X | M) = H(X) − I(X; M), (8) where I(X; M) denotes the mutual information and H(X) the Shannon entropy (e.

The authors argue that γW1 is nearly as easy to compute as γvar, requiring only sorting and differencing without additional parameters. p. 6
Where the paper says it

Constructing histograms is cumbersome and makes the metric depend on how the bins are chosen. The proposed γW1 is nearly as easy to compute as γvar, requiring only sorting and differencing and no additional parameters.

Validation with Synthetic Data

The metric γW1 was validated using synthetic climate data drawn from Gaussian, uniform, and lognormal distributions. The results show that γW1 gives stable estimates irrespective of the distribution shape and the presence of outliers provided the ensemble has about 40 members or more. p. 0

Where the paper says it

We validate the metric with synthetic climate data from Gaussian, uniform, and lognormal distributions. Unlike the two existing metrics, the proposed metric gives stable estimates irrespective of the distribution shape and the presence of outliers, provided that the ensemble has about 40 members or more.

Specifically, for Nens ∼ 40 or more, γW1 gives stable estimates whether or not outliers are present. p. 8
Where the paper says it

It becomes large only for Nens ≲ 20. Therefore, for Nens ∼ 40 or more, γW1 gives stable estimates whether or not outliers are present.

Application to Climate Model Data

The proposed metric was applied to the 2 m air temperature and total precipitation of the Community Earth System Model Large Ensemble (CESM-LE) under two different forcing scenarios. The analysis indicated that all metrics show that the relative contribution of internal variability decreases as the forcing increases, but the proposed metric shows this response most clearly. p. 0

Where the paper says it

We then apply the proposed and existing metrics to the 2 m air temperature and total precipitation of the Community Earth System Model Large Ensemble (CESM-LE) under two different forcing scenarios. All the metrics indicate that the relative contribution of internal variability decreases as the forcing increases, but the proposed metric shows this response most clearly.

For variables with non-Gaussian, heavy-tailed probability distributions such as precipitation, γW1 is therefore the most effective of the three for detecting the forced response. p. 6
Where the paper says it

The random numbers were drawn from the Gaussian and uniform distributions used in their study and, in addition, from a lognormal distribution intended to represent heavy-tailed variables such as precipitation. The time series are independent and identically distributed in time, and each experiment uses one of these three distributions.

Conclusion and Future Directions

The authors conclude that γW1 is well suited to analyzing the relative magnitude of internal variability in large-ensemble datasets because it is stable and reflects the entire distribution shape. The study suggests three directions for extending the research: defining barycenters as an alternative to the arithmetic mean, incorporating model uncertainty through multi-model ensembles, and performing a Wasserstein stability analysis of probability distributions. p. 0

Where the paper says it

jp; Abstract Quantifying forced and internal variability in climate data is fundamental to detecting the climate change signal and assessing uncertainty in climate projections. We propose a metric that quantifies the relative magnitude of internal variability in single-model initial-condition large ensembles (SMILEs).

The authors expect the Wasserstein distance to develop further as a practical tool for analyzing large-ensemble datasets. p. 0
Where the paper says it

All the metrics indicate that the relative contribution of internal variability decreases as the forcing increases, but the proposed metric shows this response most clearly. The 1-Wasserstein distance thus provides a simple and useful tool for analyzing large-ensemble datasets.

Abbreviations

CESM-LE Community Earth System Model Large Ensemble.

ITCZ Intertropical Convergence Zone.

SMILE Single-Model Initial-condition Large Ensemble. p. 0

Where the paper says it

The 1-Wasserstein distance thus provides a simple and useful tool for analyzing large-ensemble datasets. Keywords: Climate Variability, Single-Model Initial-condition Large Ensemble (SMILE), Optimal Transport, Wasserstein Distance 1 arXiv:2608.

RCP8.5 Representative Concentration Pathway 8. p. 7

Where the paper says it

The trend associated with external forcing is part of the forced response, so we did not apply linear detrending (Sane et al. We analyzed the model outputs from two forcing experiments: one under 20th-century historical forcing (20C) for 1920–2005 and the other under representative concentration pathway 8.

5.

20C 20th-Century historical forcing. p. 7

Where the paper says it

The trend associated with external forcing is part of the forced response, so we did not apply linear detrending (Sane et al. We analyzed the model outputs from two forcing experiments: one under 20th-century historical forcing (20C) for 1920–2005 and the other under representative concentration pathway 8.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Integration of Wasserstein distance into variability metrics allows AI systems to quantify internal variability using a form of optimal transport cost which reflects the entire distribution shape and applies to non-Gaussian variables. This enables AI to create more robust uncertainty assessments for climate projections by moving beyond variance limitations.

  2. The proposed metric gamma W1 can provide stable estimates for internal variability when an ensemble has about 40 members or more, suggesting a threshold of Nens reache[s] about 40 or more for reliable results. This feature allows AI models to select appropriate ensemble sizes for accurate uncertainty quantification without relying on distribution shape assumptions.

  3. The metric gamma W1 shows the largest decrease when comparing forcing scenarios in CESM-LE, indicating it gives the strongest signal of the forced response. This specific sensitivity allows AI to more clearly detect how internal variability's relative contribution changes as external forcing increases.

Abstract

Estimating forced and internal variability in climate data is fundamental to detecting the climate change signal and assessing uncertainty in climate projections. We propose a metric that quantifies the relative magnitude of internal variability in single-model initial-condition large ensembles (SMILEs). We measure forced and internal variability by two 1-Wasserstein distances, a form of optimal transport cost. The proposed metric is the distance for internal variability divided by the sum of the two. Computing this ratio requires only sorting the samples and differencing the resulting quantiles, with no additional parameters. The metric reflects the entire distribution shape and applies to non-Gaussian variables, because the 1-Wasserstein distance measures the difference between any two probability distributions. We validate the metric with synthetic climate data from Gaussian, uniform, and lognormal distributions. Unlike the two existing metrics, the proposed metric gives stable estimates in the presence of outliers irrespective of the distribution shape, provided that the ensemble has about 40 members or more. We then apply the proposed and existing metrics to the 2 m air temperature and total precipitation of the Community Earth System Model Large Ensemble (CESM-LE) under two different forcing scenarios. All three metrics agree in the spatial pattern of the difference between the scenarios but not in amplitude. The amplitude of the difference is largest for the proposed metric, consistent with the synthetic-data result that this metric changes most rapidly when internal variability dominates. The 1-Wasserstein distance thus provides a simple and useful tool for analyzing large-ensemble datasets.

Sources

Related papers