Random Field Representations of Kernel Distances

Ian Langmore

REVIEW 3 major objections 5 minor 22 references

Random Field Representations of Kernel Distances

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper establishes that every kernel distance between probability measures can be rewritten as the expected squared inner product of a random field with the difference of the two measures.

desk verdict Useful random-field reformulation of kernel distances with a promising new GFF distance, but the Dirichlet energy identity on R^d is asserted without proof and Lemma 6.4 contains a genuine subadditivity gap; both are fixable. read the letter →

arxiv 2505.23141 v1 pith:7JH5CNLQ submitted 2025-05-29 math.PR math.FAmath.STstat.TH

Ian Langmore This is my paper

classification math.PRmath.FAmath.STstat.TH MSC 60G6060G1562G10

keywords kerneldistancesrandomfieldrepresentationenergydistancefractionalBrownianmotionGaussianfreeDirichletcharacteristickernelsmaximummeandiscrepancy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper establishes that every distance between probability measures induced by a positive semi-definite kernel can be rewritten as an expected squared inner product between a random field and the difference of the two measures. The author argues this random-field viewpoint is not a curiosity: the standard “energy distance” is exactly the distance induced by fractional Brownian motion, and the common generalization that preserves continuity is obtained by replacing Brownian motion with fractional Brownian fields at Hurst index $H$. The paper then proposes a different generalization, replacing Brownian motion by a Gaussian free field, which yields a new “Dirichlet energy distance” that is formally a multidimensional Cramér–von Mises distance. A characteristic-field support theorem gives a criterion for when such distances separate measures, and signal-to-noise analysis shows how path properties determine which differences in moments the distance is sensitive to.

What carries the argument

The load-bearing object is the identity $D^2(\mu-\nu) = \mathbb{E}_U\langle U, \mu-\nu\rangle^2$, read as an expected quadratic variation or expected inner product. It converts questions about kernel distances into questions about random fields: stationarity of increments becomes shift invariance, scale covariance becomes a Hurst index, and path roughness controls which moments the distance weighs. For the new distance, the machinery is the Gaussian free field, defined as the centered Gaussian field whose covariance is the Green's function of the Laplacian, with pairings controlled through Sobolev-type spaces $K^s$ and with the Dirichlet-energy representation following from Fubini and Green's identities.

What would settle it

In dimension $d=2$, take two probability measures whose density difference $q$ behaves like $1/(\|x\|^2 \log \|x\|)$ near the origin; the Green-identity argument predicts $D^2(\mu-\nu; G)$ diverges, so a careful Monte-Carlo evaluation of the expected squared Gaussian-free-field pairing that returns a finite, consistent value would show the claimed integral representation fails outside the stated integrability conditions.

Watch

Extended reading notes

Core claim

The central claim is that for a random field $U$ with covariance kernel $k$, the kernel distance $D^2(\mu-\nu) = \mathbb{E}[k(X,X')] + \mathbb{E}[k(Y,Y')] - 2\mathbb{E}[k(X,Y)]$ is identical to $\mathbb{E}_U[\langle U, \mu-\nu\rangle^2]$, where $\langle U, \mu-\nu\rangle = \int U\,d(\mu-\nu)$. From this viewpoint, the energy distance corresponds to fractional Brownian motion with Hurst index $H=1/2$, and the widely used generalization that preserves continuity is the family of fractional Brownian fields with spectral density $\varphi(\omega) \propto \|\omega\|^{-(d+2H)}$. The paper's new object is the Dirichlet energy distance, obtained by using a Gaussian free field as $U$; for measures with a density $q$ and potential $\phi$ solving $\Delta\phi = q$ with $\phi \to 0$ at infinity, $D^2(\mu-\nu; G) = \int \|\nabla\phi(x)\|^2\,dx$, which in one dimension reduces to the classical Cramér–von Mises distance. The paper also proves that a Gaussian field is characteristic over the dual of a Banach space exactly when its support is the whole space, giving a support-theoretic criterion that links path properties to distance properties.

Load-bearing premise

The Dirichlet energy distance is only defined when the difference $\mu-\nu$ has a density $q$ satisfying dimension-dependent integrability conditions and when Fubini and Green's identities apply; if $q$ is too heavy-tailed in dimension two, or not in $L^1 \cap L^2$ in higher dimensions, the claimed integral is infinite or undefined.

Editorial extensions

If this is right

The random-field identity holds for every positive semi-definite kernel, so any kernel distance can be studied through the path properties of its inducing field.
Energy distance is the $H=1/2$ fractional Brownian field distance; choosing $H>1/2$ makes the distance more sensitive to odd moments such as mean shifts, while $H<1/2$ makes it more sensitive to even moments such as variance shifts.
The Gaussian free field distance is well defined for heavy-tailed distributions such as multivariate Student's $t$ with fewer than one finite moment, where the kernel form of energy distance fails in finite precision.
A Gaussian field is characteristic over a dual space if and only if its support is the whole space, yielding concrete support theorems for fractional Brownian fields and Gaussian free fields.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the identity suggests a recipe for designing new distances: choose any Gaussian field whose paths are dense in a suitable Banach space, and its expected squared inner product defines a metric between measures.
The Dirichlet energy distance may inherit useful properties from the Gaussian free field's conformal invariance, making it a candidate for two-sample testing on domains with nontrivial geometry, though the paper does not develop this link.
A testable extension is to estimate the Dirichlet energy distance with a spectral or finite-element discretization of the Laplacian Green's function, which could bypass the analytic integrability conditions and extend the distance to densities outside $L^2$.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear Colleague,

The headline: this paper is worth a serious referee, but the main new object—the Dirichlet energy distance—is not rigorously established on R^d. The representation D^2 = E^2 is the standard RKHS/MMD view, so the novelty is not the representation itself. What is new is the choice of the Gaussian free field as the inducing field, giving a distance that formally extends Cramér-von Mises and, unlike energy distance, stays finite for heavy-tailed multivariate t distributions. The support-theorem perspective is also clean: a Gaussian field is characteristic over a dual space iff its support is the whole space. That part is right and useful.

The Fourier representation and the fBm generalization are handled carefully, and the SNR analysis gives practical guidance on choosing H. The section on natural assumptions is a good synthesis of earlier ideas.

Now the soft spots. The proof of Lemma 6.4 is wrong. It claims that if a set A has positive measure and contains no support point, then for each a there is an open ball of zero probability, and concludes P(A) is bounded by a sum over an uncountable family. That step is invalid: uncountable subadditivity fails, and the union of those balls need not be measurable. The theorem may still be true, but this proof does not establish it. A different argument is needed.

The bigger gap is the Dirichlet energy identity, equation (16). The paper asserts that under dimension-dependent integrability conditions, Fubini and Green's identities give D^2 = integral ||Q||^2. The stress-test note is correct: the boundary terms are never shown to vanish, and the conditions are not shown to place the density in the dual of the GFF's Gaussian Hilbert space. For d=2 the Green's function is unbounded at infinity, so the double integral is only conditionally meaningful; the log-moment condition is asserted, not derived. For d=1, the stated first-moment condition is stricter than necessary—a zero-mean density with |x|^{-2} tail has finite Brownian distance. So the advertised class is not characterized. This gap matters because the practical advantage of the GFF distance is precisely that it handles heavy tails. The compact-domain case is treated properly via Sobolev spaces, so the fix likely involves doing the analysis on R^d more carefully. As it stands, the central claim is unsupported.

The numerics illustrate the point but are not reproducible from shipped artifacts; that alone is a minor issue.

Who this is for: researchers in MMD, energy statistics, or proper scoring rules, especially those concerned with heavy-tailed data. The paper deserves peer review, not a desk reject. I would send it out and ask the referee to focus on the GFF integrability conditions and the proof of Theorem 6.1. I would not cite the R^d Dirichlet energy result in its current form, though the support theorem and the fBm discussion are worth keeping in mind.

Referee Report

3 major / 5 minor

Summary. The paper proposes representing positive semi-definite kernel distances between probability measures as E_U[⟨U, μ−ν⟩²] for a random field U, proves equivalence of the quadratic-variation, kernel, and Fourier-spectral forms, and uses this viewpoint to interpret energy distance as induced by fractional Brownian motion and to introduce a new 'Dirichlet energy distance' induced by the Gaussian free field. It also derives support theorems for characteristic Gaussian fields and discusses practical implications for signal-to-noise ratios.

Significance. The random-field representation is a useful unifying perspective with some genuinely original elements, notably the identification of a continuity-preserving generalization of energy distance with fractional Brownian motion and the proposal of a Gaussian-free-field distance that formally extends the Cramér–von Mises distance to multiple dimensions. The paper gives explicit constructive equivalences and a concrete spectral characterization (Eq. 10, Prop. 2.1) rather than fitting parameters to data. However, the central GFF identity (16) is asserted under integrability conditions that are not proved sufficient for the double integral, Fubini, or boundary terms, and the support theorem relies on an invalid uncountable-subadditivity step; both need to be repaired before the claims are fully supported.

major comments (3)

[§6.2, Lemma 6.4] The proof of Lemma 6.4 performs an uncountable sum: P[U ∈ ∪_{a∈A} O_a] ≤ Σ_{a∈A} P[U ∈ O_a] = 0. Countable subadditivity does not extend to arbitrary uncountable families, so this inequality is invalid. Since the lemma is used in the proof of Theorem 6.1 to conclude Span(S_{U,B}) is dense when U is characteristic, the support theorem is not established as written. The argument needs to be replaced, for example by exploiting separability of B and a countable dense subset, or by an alternative argument that avoids uncountable subadditivity.
[§3.3, Eq. (16)] The identity D²(μ−ν;G) = ∫∫ q(x)q(y)N_d(x,y) dxdy = ∫ ||Q(x)||² dx is the paper's headline novelty, but it is justified only by 'so long as Green's identities hold' followed by dimension-dependent integrability conditions. The text does not prove that the stated conditions (d=1: ∫|q(x)x|<∞; d=2: ∫|q(x)| log ||x|| <∞; d≥3: q∈L²∩L¹) make the double integral well-defined, allow Fubini, place q in the dual of the GFF's Gaussian Hilbert space, or ensure that the boundary terms in Green's identity vanish at infinity. For d=2, N_2 is unbounded at infinity, so the double integral is at best conditionally defined, and for d=1 the first-moment condition is not necessary, so the advertised class is not characterized. Without a rigorous statement and proof of these conditions, the central claim that the GFF induces a characteristic distance equal to the Dirichlet energy on the stated class is unsupported.
[§3.3, non-compact case] For non-compact X=R^d, the GFF is a generalized Gaussian field, but the paper does not define the space of test functions q for which ⟨G,q⟩ is meaningful on all of R^d, nor prove that the stated integrability conditions imply q lies in the dual of the GFF. The compact-X discussion following Eq. (16) is rigorous, but the non-compact case, which is the one used for the claimed Dirichlet energy distance, is not. Please provide a precise definition of the GFF on R^d and a proof of the pairing identity for the stated class of q, or restrict the claims to a class where such a proof can be supplied.

minor comments (5)

[§3.3] The text says 'let ϕ be the potential solving Δϕ = 0'; for the Dirichlet energy identity with Q=∇ϕ to hold, ϕ must solve −Δϕ = q (up to sign convention). Please correct this typo.
[§3.4] The text says 'fractional GFF distance with α = −5', which conflicts with the stated range 0<α<d for the fractional free fields; Figure 3's caption says α=5. Please clarify the intended value.
[§3.5, Eq. (17)] The signal in the SNR expression is written as (E{⟨U,µ−ν⟩})², but earlier the signal is defined as E{⟨U,µ−ν⟩²}. Please check and correct the expression.
[§6] The notation Span(S_{U,B}) is used both for the linear span and its closure; the proof of Theorem 6.1 relies on this distinction, so it would help to denote the closure explicitly, e.g., overline{Span(S_{U,B})}.
[§3.2] In Eqs. (14) and (15), the limits as H→1 and H→0 are stated informally for 'bounded sets' and 'continuous q∈L1∩L2'; please specify the precise mode of convergence (distribution of the field or convergence of the quadratic form) and the required conditions on the measures.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the field, kernel, Fourier, and Gaussian-free-field identities are direct equivalences; the main caveats are unproved analytic conditions, not circular reasoning.

full rationale

The paper's derivation chain is a set of explicit equivalence representations rather than a fitted or self-referential prediction. Equation (1)-(3) rewrite the kernel distance as the expected square of a field-measure pairing using Fubini and the covariance identity; equation (10) derives the Fourier representation from Wiener integrals and Parseval; equation (13) is the covariance calculation for fractional Brownian motion. The Gaussian free field distance in (16) is obtained by substituting the Green's function covariance and applying Fubini and Green's identities, with the stated dimension-dependent integrability conditions. The characteristic and support statements in Sections 3 and 6 follow from the same pairing formula and standard functional-analytic arguments, not from assuming the desired conclusion. The desiderata in Section 5 restrict to fractional Brownian motion by citing the external textbook [1], not by a self-citation chain. The only self-citation is [6], used as an example of an additive-Brownian loss; it is not load-bearing for any central claim. The genuine weakness is mathematical rigor rather than circularity: the identity D^2 = integral ||Q||^2 dx in (16) is asserted subject to 'so long as Green's identities hold' and conditions that are stated but not fully proved, and the d=2 case is delicate because the Green's function is unbounded at infinity. These are correctness and completeness concerns, not instances of a result being equivalent to its inputs by construction. Therefore no circular step is present.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

The paper introduces no fitted constants or new entities; it relies on standard objects (fBm, GFF) and standard functional analysis. The main unstated assumption is the separability/regularity needed for the proof of Lemma 6.4, and the modeling desiderata in Section 5.

assumptions (8)

standard math Gaussian process existence with a prescribed covariance kernel
Used in Section 2.2 to equate any PSD kernel with the covariance of a centered field U.
standard math Fractional Brownian field B_H on R^d exists with the stated covariance and spectral density
Invoked in Section 3.2 (citing [1]) to define continuous fractional fields.
standard math Gaussian free field exists as a generalized Gaussian field with Green's function covariance
Invoked in Section 3.3 and Section 6.1.3; the GFF is a known object.
standard math Green's identities and Fubini apply on unbounded domains under stated integrability conditions
Used to establish equation (16) and to connect GFF pairing to Dirichlet energy.
standard math Riesz-Markov and duality of quotient spaces identify the duals
Used in Sections 6.1.1 and 6.1.2 to identify (C_x0(X))* = F0(X) and (C_0,alpha)* with finite moment measures.
standard math The support of a Gaussian process is a closed linear subspace
Used in Corollary 6.2, citing [19].
domain assumption The Banach space B is separable or the finite Borel measure is regular, so zero-measure neighborhoods can be made countable/finite
Not stated; the proof of Lemma 6.4 as written uses uncountable subadditivity, which is invalid without such a property.
domain assumption The desiderata of stationary increments, scale invariance, and path regularity restrict allowable fields
Section 5 treats these as natural assumptions; they are modeling choices, not mathematical axioms.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Random Field Representations of Kernel Distances." pith.science (2026). https://pith.science/paper/7JH5CNLQ

@misc{pith2026250523141,
  author       = {Pith},
  title        = {Pith review of: Random Field Representations of Kernel Distances},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7JH5CNLQ}},
  note         = {Machine review of arXiv:2505.23141}
}

read the original abstract

Positive semi-definite kernels are used to induce pseudo-metrics, or ``distances'', between measures. We write these as an expected quadratic variation of, or expected inner product between, a random field and the difference of measures. This alternate viewpoint offers important intuition and interesting connections to existing forms. Metric distances leading to convenient finite sample estimates are shown to be induced by fields with dense support, stationary increments, and scale invariance. The main example of this is energy distance. We show that the common generalization preserving continuity is induced by fractional Brownian motion. We induce an alternate generalization with the Gaussian free field, formally extending the Cram\'er-von Mises distance. Pathwise properties give intuition about practical aspects of each. This is demonstrated through signal to noise ratio studies.

Figures

Figures reproduced from arXiv: 2505.23141 by the authors.

**Figure 1.** Traces of fBM: Smaller Hurst index H results in a field that decorrelates quickly, and approximates a (random) constant plus stationary noise. H = 0.5 corresponds to Brownian motion. As H → 1, the traces become linear with random slope. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗

**Figure 2.** Moments of fractional motion: Shows that, for smaller H < 1/2, ⟨BH, tk ⟩ is larger for even k. For H > 1/2, ⟨BH, tk ⟩ is larger for odd k [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗

**Figure 3.** Signal to noise sweep comparing continuous and Dirichlet energies: SNR as a function of multivariate Student’s T dof f. Computed with 4M samples using the kernel forms of the scores. Specfically, H = 1/2 and α = 5 from section 3.4. The Dirichlet energy is well defined even when the distribution has non-finite mean. The kernel form of continuous energy requires degrees of freedom f > 2H = 1, otherwise results became … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: SNR Sweep: Shows that the signal, D2 (µ−ν; BH) is the differentiating factor in SNR calculations, since the standard deviation is mostly perturbation independent. The signal shape (with respect to H) is as expected due to the path-wise heuristics from section 3.5. 22 …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Fractional Fields and Applications

Serge Cohen and Jacques Istas. Fractional Fields and Applications. Mathe- matiques & Applications (Berlin). Springer, 2013 edition, 30 May 2013

work page 2013
[2]

Probability: Theory and Examples

Rick Durrett. Probability: Theory and Examples . Cambridge University Press, April 2019

work page 2019
[3]

Strictly Proper Scoring Rules, Prediction, and Estimation

Tilmann Gneiting and Adrian E Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. J. Am. Stat. Assoc., 102(477):359–378, 1 March 2007

work page 2007
[4]

A kernel method for the two-sample problem

Arthur Gretton, Karsten Borgwardt, Malte J Rasch, Bernhard Scholkopf, and Alexander J Smola. A kernel method for the two-sample problem. arXiv [cs.LG], 15 May 2008

work page 2008
[5]

A Kernel Two- Sample Test

Arthur Gretton, Karsten M Borgwardt, Tuebingen Mpg De, Malte J Rasch, Xinjiekouwai St, Bernhard Sch¨ olkopf, and Alexander Smola. A Kernel Two- Sample Test. J. Mach. Learn. Res. , 13(25):723–773, 2012

work page 2012
[6]

Neural general circulation models for weather and climate

Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Kl¨ ower, James Lottes, Stephan Rasp, Peter D¨ uben, Sam Hatfield, Peter Battaglia, Alvaro Sanchez-Gonzalez, Matthew Willson, Michael P Brenner, and Stephan Hoyer. Neural general circulation models for weather and climate. Nature, pages 1–7, 22 July 2024

work page 2024
[7]

AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score

Simon Lang, Mihai Alexe, Mariana C A Clare, Christopher Roberts, Ril- wan Adewoyin, Zied Ben Bouall` egue, Matthew Chantry, Jesper Dramsch, Peter D Dueben, Sara Hahner, Pedro Maciel, Ana Prieto-Nemesio, Cathal O’Brien, Florian Pinault, Jan Polster, Baudouin Raoult, Steffen Tietsche, and Martin Leutbecher. AIFS-CRPS: Ensemble forecasting using a model trai...

work page 2024
[8]

Sharp constants in the hardy-Littlewood-Sobolev and related inequalities

Elliott H Lieb. Sharp constants in the hardy-Littlewood-Sobolev and related inequalities. Ann. Math., 118(2):349, September 1983

work page 1983

Show all 22 references

[9]

Characterization of translation invariant MMD on R d and connections with Wasserstein distances

Thibault Modeste and Cl´ ement Dombry. Characterization of translation invariant MMD on R d and connections with Wasserstein distances. J. Mach. Learn. Res., 25(237), May 2024

2024
[10]

Cambridge series in statistical and proba- bilistic mathematics: Brownian motion series number 30

Peter Morters and Yuval Peres. Cambridge series in statistical and proba- bilistic mathematics: Brownian motion series number 30 . Cambridge Uni- versity Press, 25 March 2010. 18

2010
[11]

Growth rate of certain Gaussian processes

Steven Orey. Growth rate of certain Gaussian processes. In Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability, Volume 2: Probability Theory , volume 6.2, pages 443–452. University of California Press, 1 January 1972

1972
[12]

Methods of Modern Mathematical Physics

M Reed and B Simon. Methods of Modern Mathematical Physics. I Func- tional Analysis. Academic Press, second edition, 1980

1980
[13]

Real and complex analysis, 3rd ed

Walter Rudin. Real and complex analysis, 3rd ed . McGraw-Hill, Inc., 1987

1987
[14]

Fractional Brownian motion with variable Hurst parameter: Definition and properties

Jelena Ryvkina. Fractional Brownian motion with variable Hurst parameter: Definition and properties. arXiv [math.PR] , 12 June 2013

2013
[15]

Fractional Brownian motion in a nutshell

Georgiy Shevchenko. Fractional Brownian motion in a nutshell. Int. J. Mod. Phys. Conf. Ser. , 36:1560002, January 2015

2015
[16]

Hilbert space embeddings and metrics on probability mea- sures

Bharath K Sriperumbudur, A Gretton, K Fukumizu, B Scholkopf, and Gert R G Lanckriet. Hilbert space embeddings and metrics on probability mea- sures. J. Mach. Learn. Res. , 11:1517–1561, 30 July 2009

2009
[17]

Brownian distance covariance

G´ abor J Sz´ ekely and Maria L Rizzo. Brownian distance covariance. Ann. Appl. Stat. , 3(4):1236–1265, December 2009

2009
[18]

Energy statistics: A class of statistics based on distances

G´ abor J Sz´ ekely and Maria L Rizzo. Energy statistics: A class of statistics based on distances. J. Stat. Plan. Inference , 143(8):1249–1272, 1 August 2013

2013
[19]

The topological support of Gaussian measure in Banach space

N Vakhania. The topological support of Gaussian measure in Banach space. Nagoya Math. J. , 57(none):59–63, 1 June 1975

1975
[20]

Reproducing kernel Hilbert spaces of Gaussian priors

A W van der Vaart and J H van Zanten. Reproducing kernel Hilbert spaces of Gaussian priors. In Institute of Mathematical Statistics Collections , vol- ume 3, pages 200–222. Institute of Mathematical Statistics, 1 January 2008

2008
[21]

Controlling Wasserstein Distances by Kernel Norms with Application to Compressive Statistical Learning

Titouan Vayer and R´ emi Gribonval. Controlling Wasserstein Distances by Kernel Norms with Application to Compressive Statistical Learning. J. Mach. Learn. Res., 24(149):1–51, 2023

2023
[22]

Lecture notes on the Gaussian Free Field

Wendelin Werner and Ellen Powell. Lecture notes on the Gaussian Free Field. arXiv [math.PR] , 9 April 2020. 19 Figure 1: Traces of fBM: Smaller Hurst index H results in a field that decorrelates quickly, and approximates a (random) constant plus stationary noise. H = 0.5 corre...

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Fractional Fields and Applications

Serge Cohen and Jacques Istas. Fractional Fields and Applications. Mathe- matiques & Applications (Berlin). Springer, 2013 edition, 30 May 2013

work page 2013

[2] [2]

Probability: Theory and Examples

Rick Durrett. Probability: Theory and Examples . Cambridge University Press, April 2019

work page 2019

[3] [3]

Strictly Proper Scoring Rules, Prediction, and Estimation

Tilmann Gneiting and Adrian E Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. J. Am. Stat. Assoc., 102(477):359–378, 1 March 2007

work page 2007

[4] [4]

A kernel method for the two-sample problem

Arthur Gretton, Karsten Borgwardt, Malte J Rasch, Bernhard Scholkopf, and Alexander J Smola. A kernel method for the two-sample problem. arXiv [cs.LG], 15 May 2008

work page 2008

[5] [5]

A Kernel Two- Sample Test

Arthur Gretton, Karsten M Borgwardt, Tuebingen Mpg De, Malte J Rasch, Xinjiekouwai St, Bernhard Sch¨ olkopf, and Alexander Smola. A Kernel Two- Sample Test. J. Mach. Learn. Res. , 13(25):723–773, 2012

work page 2012

[6] [6]

Neural general circulation models for weather and climate

Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Kl¨ ower, James Lottes, Stephan Rasp, Peter D¨ uben, Sam Hatfield, Peter Battaglia, Alvaro Sanchez-Gonzalez, Matthew Willson, Michael P Brenner, and Stephan Hoyer. Neural general circulation models for weather and climate. Nature, pages 1–7, 22 July 2024

work page 2024

[7] [7]

AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score

Simon Lang, Mihai Alexe, Mariana C A Clare, Christopher Roberts, Ril- wan Adewoyin, Zied Ben Bouall` egue, Matthew Chantry, Jesper Dramsch, Peter D Dueben, Sara Hahner, Pedro Maciel, Ana Prieto-Nemesio, Cathal O’Brien, Florian Pinault, Jan Polster, Baudouin Raoult, Steffen Tietsche, and Martin Leutbecher. AIFS-CRPS: Ensemble forecasting using a model trai...

work page 2024

[8] [8]

Sharp constants in the hardy-Littlewood-Sobolev and related inequalities

Elliott H Lieb. Sharp constants in the hardy-Littlewood-Sobolev and related inequalities. Ann. Math., 118(2):349, September 1983

work page 1983