acceptodds
Under review as a conference paper at ICLR 2027

One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators

Abstract

Probabilistic conditioning is concerned with the identification of a distribution of a random variable given a random variable . It is a cornerstone of scientific and engineering applications where modeling uncertainty is key. This problem has traditionally been addressed in machine learning by directly learning the conditional distribution of a fixed joint distribution. This paper introduces a novel perspective: we propose to solve the conditioning problem by identifying a single operator that maps any joint density to its conditional, thus amortizing over joint-conditional pairs. We establish that the conditioning operator can be approximated to arbitrary accuracy by neural operators. Our proof relies on new results establishing continuity of the conditioning operator over suitable classes of densities, including densities defined on unbounded domains. Last, we learn the conditioning map for a class of Gaussian mixtures using neural operators, illustrating the promise of our framework. This work provides the theoretical underpinnings for general-purpose, amortized methods for probabilistic conditioning.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Neural operators, probabilistic conditioning, amortized inference, operator learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_azX9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a novel operator-theoretic framework for probabilistic conditioning, formalizing the conditioning operation as a nonlinear operator that maps joint probability densities to their conditional densities or kernels. The authors establish rigorous theoretical foundations by proving continuity, quantitative Hölder stability, and universal approximation theorems using neural operators (such as Fourier Neural Operators and Transformer Neural Operators) over both compact subsets of continuous densities and unbounded L1L^1 domains. Empirically, the work provides a proof-of-concept demonstration on 2D Gaussian mixture models and kernel density estimation (KDE) settings, showing that trained neural operators can accurately emulate conditioning and outperform traditional plug-in numerical integration estimators.

Critical strengths
  1. Framing probabilistic conditioning as an infinite-dimensional mapping (ρ,y)↦ρ(⋅∣y)(\rho, y) \mapsto \rho(\cdot\vert{}y) offers a powerful paradigm for amortizing conditional inference across diverse joint distributions without retraining.
  2. The theoretical analysis is exceptionally thorough, establishing local/global Lipschitz stability, maximal continuous extensions, and FullSpace Neural Operator (FullNO) universal approximation guarantees on unbounded domains.
  3. The empirical evaluation successfully demonstrates that FNOs and TNOs can approximate conditional kernels on synthetic Gaussian mixtures and achieve lower median errors than standard KDE plug-in baselines.
Critical weaknesses
  1. Theoretical stability bounds rely heavily on a strict uniform lower bound on data marginals (inf⁡ymρ(y)≥δ\inf_{y} m_\rho(y) \ge \delta), with constants that blow up as δ→0\delta \to 0, limiting robustness in realistic settings where marginals approach zero.
  2. The reliance on grid-based function representations (64264^2 grids in 2D) incurs the curse of dimensionality (O(Dd)O(D^d)), rendering grid discretizations intractable for higher-dimensional problems.
  3. The framework requires explicit evaluation access to joint probability densities, whereas modern machine learning applications almost exclusively provide empirical point clouds of samples.
  4. The evaluation lacks comparisons against modern amortized conditional generative models (e.g., conditional normalizing flows or conditional diffusion models), and the architecture lacks explicit enforcement of probability density constraints (non-negativity and integration to one).
Questions for authors
  1. How does the required training sample size and model parameter scale as the dimensions of the state space DD and query space rr increase beyond 2, given the exponential scaling cost of grid-based discretizations?
  2. How robust are trained neural operators when input joint densities violate the strict uniform lower bound condition (inf⁡ymρ(y)≥δ\inf_y m_\rho(y) \ge \delta), such as near support boundaries or deterministic relationships where marginals vanish?
  3. How do you ensure during training and inference that predicted conditional kernels satisfy fundamental probability axioms (non-negativity and integrating to 1 for each fixed yy) without explicit architectural normalization layers?
  4. Can the proposed operator-learning framework be extended to take empirical point clouds (samples) directly as input—thereby bypassing intermediate grid estimation—and what theoretical adjustments would be required?
4 comments
AuthorsOpenReview ↗

Response to Reviewer azX9

We thank the reviewer for the thoughtful and valuable feedback. Below we address their comments.

Marginal lower bounds and robustness (W1/Q2).

The dependence on the marginal lower bound reflects the stability of conditioning in the chosen error metric. In particular, uniform approximation of the entire conditional kernel requires accuracy even at arbitrarily rare queries y∈Ey \in E.

The quantitative stability bounds do, indeed, deteriorate as δ→0\delta \to 0. However, we view this as a fundamental aspect of the problem of conditioning densities in the supremum norm. Indeed, in Theorem 4.6 we extend conditioning to all joints with positive marginals without requiring a uniform lower bound, a set we name X+\mathcal{X}_+. Then, in Theorem 4.7 we show that this set is maximal meaning that no continuous extension of the conditioning operator exists on a larger set. As the limit δ→0\delta \to 0 effectively exposes densities in the boundary of this maximal set X+\mathcal{X}_+ it is necessary that the stability estimates deteriorate with this limit. That being said, we wish to note that class of densities admitting such a uniform lower bound is considerably large. For a parametric family pθθ∈Θ\\{ p_\theta \\}_{\theta \in \Theta} of continuous densities with positive marginals mpθ>0m_{p_{\theta}} > 0 defined on a compact domain D×ED \times E and with compact parameter space Θ\Theta, if the parametrization θ↦pθ\theta \mapsto p_\theta is also continuous then δ:=min⁡θ∈Θmin⁡y∈Empθ(y)>0\delta := \min_{\theta \in \Theta} \min_{y \in E} m_{p_\theta}(y) > 0 and thus pθθ∈Θ⊆Xδ\\{ p_\theta \\}_{\theta \in \Theta} \subseteq \mathcal{X}_\delta in the notation of the paper. Many standard parametric families, on compact parameter ranges bounded away from degeneracy, satisfy these assumptions.

Lastly, we point the reviewer to Figures 1 and 2: Neural Operators (NOs) exhibit low error even in regions of low probability mass, although there does appear to be a mild deterioration in performance as one approaches the boundary of the domain, see right-most plot in Figure 1. Moreover, as can be in the left-most plot of Figure 2, one actually observes a widening of the performance gap between the plug-in KDE to the Neural Operator (NO) as the uniform lower bound δ\delta decreases. One might say that the NO relies on the joint-conditional example pairs it has seen to do inference on regions of low marginal probability 0<mρ(y)≪10 < m_\rho(y) \ll 1 that result in few observations.

Dimensional scaling (W2/Q1).

We agree that a grid based approach used in the proof-of-concept experiments scales exponentially with dimension. However, neither the conditioning formulation nor the neural-operator architectures underlying our approximation theorems require a tensor-grid representation. Indeed, a closer look at Section 3.1 reveals that all components of the considered Neural Operator architectures are agnostic to the representation of the underlying function. For example, the lifting and activation operations act pointwise and the nonlocal operations in Definition 3.2 are integral features $ a_j(v)=\int_\Omega (T_jv(z))\cdot\psi_j(z),dz . $ In general, all of these operations could be carried out on an input function represented in terms of arbitrary bases, for example polynomials, wavelets, kernel functions or some other sparse basis expansions or low-rank representation. Ultimately, the effectiveness of such alternate function representations would depend on appropriate regularity or compressibility properties of the underlying family of probability densities. Section 6 identifies scalable function representation as an open problem. In the final manuscript, we will discuss these possibilities and the additional analysis needed to establish their efficiency.

Lastly, we wish to note that we view our work as foundational rather than algorithmic. As such, we view the experiments as demonstrating consistency with the theory we have developed and not as proposing a practical scheme for conditioning in high dimensions. We view this paper as the first step in a longer research program that aims to build amortized conditional operators from samples rather than densities. Providing such an architecture with approximation guarantees is a formidable challenge; please see our answer below for further discussion of this.

AuthorsOpenReview ↗

Response to Reviewer azX9 (Part 2)

Sample-based inputs and the formulation of conditioning (W3/Q4).

We agree that building an architecture, with approximation guarantees, that directly accepts samples is a key objective. However, such an extension raises fundamental mathematical questions. Indeed, if one attempts to formulate conditioning as an operator acting on probability measures, including discrete measures, endowed with the topology of weak convergence, one quickly runs into problems. Indeed, the conditioning operator is not continuous in such a space. This is discussed in Example E.3 in the Appendix, briefly recounted below: Write δ(a,b)\delta_{(a,b)} for the point mass at (a,b)(a,b) and consider the sequence

\quad , \quad P=\tfrac12\delta_{(0,0)}+\tfrac12\delta_{(1,0)}.$$ Note that $P_n \to P$ in a weak sense. However, their conditional distributions at $y=0$ satisfy $$P_n(\cdot\mid 0)=\delta_0, \quad , \quad P(\cdot\mid 0)=\tfrac12\delta_0+\tfrac12\delta_1$$ and it clear that $P_n(\cdot\mid 0) \nrightarrow P(\cdot\mid 0)$. This shows that extending our framework requires identifying appropriate distribution classes, topologies, and approximation criteria where conditioning is continuous and approximable by Neural Operators. Section 6 identifies the sample-based extension as an open direction. **Probability axiom satisfaction (Q3).** The architectures used in our experiments do not impose non-negativity and unit integral as hard constraints. Our results concern approximation of conditional densities as functions. Density estimators that are not bona fide probability densities have a long history in statistics . For example, higher-order kernel density estimators can take negative values. Tsybakov [1, Section 1.2.1, p. 10, "Positivity constraint"] explicitly discusses this issue and shows that taking the positive part cannot increase pointwise squared error. Orthogonal series density estimators provide another classical example [1, Section 1.7]. More generally, correcting estimators that are not bona fide densities has a long history; see, for example, Gajek [2]. For our $L^1$ setting, validity can be enforced by a simple post-processing step. For a fixed query $y$, let $f_y$ be an integrable raw prediction, define $f_y^+(x)=\max\\{f_y(x),0\\}$ and $Z_y=\int f_y^+(x)\,dx$, and set $$\widetilde{f}_y(x)=\frac{f_y^+(x)}{Z_y}$$ when $Z_y>0$, using a fixed probability density otherwise. For any true conditional density $\kappa_y$, $$\|\widetilde{f}_y-\kappa_y\|_{L^1} \le 2\|f_y-\kappa_y\|_{L^1}.$$ Indeed, the change caused by normalization is bounded by $\|f^+_y - \kappa_y \|_{L^1}$ and clipping cannot increase the $L^1$ error, namely $\|f^+_y - \kappa_y \|_{L^1} \leq \|f_y - \kappa_y \|_{L^1}$. Thus, exact density constraints can be enforced while preserving the $L^1$ approximation guarantee up to a factor of two. In the final manuscript we will clarify the fact that we are not enforcing hard probability distribution constraints and we will discuss ways to amend this. **References** [1] Alexandre B. Tsybakov. *Introduction to Nonparametric Estimation*. Springer, 2009. Section 1.2.1, p. 10, and Section 1.7, pp. 48–49. [DOI](https://doi.org/10.1007/b13794). [2] Lesław Gajek. "On Improving Density Estimators which are not Bona Fide Functions." *The Annals of Statistics*, 14(4):1612–1618, 1986. [DOI](https://doi.org/10.1214/aos/1176350182). **Comparisons with conditional generative models (W4).** Amortization in conditional flows and diffusion models is relevant to the broader goal of conditional inference. However, such methods typically do not provide amortization across problem instances but rather across query points $y \in E$. Indeed, in their standard fixed-joint formulation, such methods learn $y\mapsto\rho(\cdot\mid y)$ for a given data-generating joint $\rho$ which at inference can be queried at any $y \in E$, hence the amortization. For us, amortization is a stronger notion: we propose a conditional operator that maps any joint $\rho$ to the collection of its conditions, e.g. to the conditional kernel $\kappa_\rho$ that contains all conditions $\rho(\cdot \mid y) = \kappa_\rho(\cdot \mid y)$ for any $y \in E$. In other words, we amortize over both the problem instance $\rho$ and all query points $y \in E$. Therefore, a fair comparison would account for the cost of re-training a conditional flow model for every new joint $\rho$ whereas a trained Neural Operator would just condition $\rho$ with no extra training cost. An accuracy / cost trade-off would likely arise, whereby problem instance specific methods might achieve better accuracy but at a much greater compute cost. Such comparisons would be useful and will be certainly pursued in future work. We will discuss these considerations further in both the related work and open questions sections of the paper.

I thank the authors for taking their time for the detailed and mathematically thoughtful responses to all four of the questions.

Given that the paper's primary contribution is a rigorous operator-theoretic foundation for amortized conditioning and that the theoretical questions have been satisfactorily addressed, I am raising my score from 2 to 3.

We thank the reviewer for their kind words and constructive feedback.

Reviewer_TcCZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper treats probabilistic conditioning as an operator taking joint densities to conditional densities, allowing one model to be used across different input distributions. The authors establish continuity and stability results and use these to obtain neural-operator approximation guarantees on suitable compact classes. They also consider unbounded domains. Gaussian-mixture experiments provide a proof of concept.

I think the idea is cool, makes sense, and I've wondered it myself.

Critical strengths

The operator formulation is appealing, and the continuity analysis makes clear which assumptions support amortization across distributions. I found this connection particulary useful.

Critical weaknesses

I was not certain how their results compare to other results which I know of on approximation on unbounded domains, or generative neural operators.

Questions for authors
  1. How is the paper related to the generative neural operator of: "Neural Operators Which Certifiably Simultaneously Solve Infinitely Many Convex Programs using Logarithmic Rank and Linear Complexity" by Anastasis Kratsios, Ariel Neufeld, Philipp Schmocker

  2. I wonder how your result compare to the unbounded domain results of "Global universal approximation of functional input maps on weighted spaces" by Christa Cuchiero, Philipp Schmocker, Josef Teichmann

1 comment
AuthorsOpenReview ↗

Response to Reviewer TcCZ

We thank the reviewer for their kind words and suggesting relevant literature. We discuss the points raised below.

Question 1

We thank the reviewer for asking about generative neural operators. The cited paper "Neural Operators Which Certifiably Simultaneously Solve Infinitely Many Convex Programs using Logarithmic Rank and Linear Complexity" by Anastasis Kratsios, Ariel Neufeld, Philipp Schmocker does not anymore concern generative neural operators, but rather neural operators that deterministically approximate the loss-to-minimizer operator for optimization problems. An earlier version of that paper was generative, but not anymore. Regardless, we understand "generative neural operator" to mean a neural operator that parametrizes a single fixed probability measure supported on a function space. Most works (e.g. [RFARA]) on generative neural operators are not amortized. They concern the transport-based approximation of a fixed probability measure P∈P(H)P\in\mathcal{P}(H) supported on a function space H; the network input is a realization of noise, not a probability measure. Every time PP changes, the neural operator must be re-trained. There are conditional versions of this approach that amortize over the conditioning variable yy; but the network must still be re-trained every time the joint distribution changes. These works deliver HH-valued samples as output. In contrast, our work centers on the theoretical density-to-density problem in finite dimensions, in which the probability density is over Rd\mathbb{R}^d rather than function space HH. The input to our neural operator is the joint density itself (and possibly yy), and the output is a function that approximates the conditional (kernel) in the uniform or L1L^1 sense. This directly addresses amortization over many probability densities in comparison to other work in amortized inference that fix the joint or conditional distribution and only vary the data yy.

Question 2

Our work has some connections to “Global universal approximation of functional input maps on weighted spaces" by Christa Cuchiero, Philipp Schmocker, Josef Teichmann" [CST] that we will include in the revision of our manuscript. At the level of function-to-function operator approximation, our paper works with neural operators, while [CST] works with a generalization of DeepONet encoder-decoder architectures. Both works consider function spaces over unbounded spatial domains. Theorem 4.13 of [CST] establishes global approximation in the weighted norm sup⁡u∈X∣F(u)−Φ(u)∣Yψ(u)<ε\sup_{u\in X}\frac{|F(u)-\Phi(u)|_Y}{\psi(u)}<\varepsilon under their weighted-space, feature, and activation hypotheses. Its output space YY can be any Banach space, including L1(Rd′)L^1(\mathbb R^{d'}), so it already permits outputs on unbounded spatial domains. Similarly for XX, which can be a (strict) subset of L1(Rd)L^1(\mathbb{R}^d).

The FullNO result in our paper establishes sup⁡u∈K∣F(u)−Ψ(u)∣L1(Rd′)<ε\sup_{u\in K}|F(u)-\Psi(u)|_{L^1(\mathbb R^{d'})}<\varepsilon for every L1L^1 continuous FF and L1L^1-compact KK. Thus, its approximation is uniform over a compact family of inputs, while the error integrates over the entire unbounded spatial domain. One can specialize Theorem 4.13 of [CST] to X=KX=K and ψ≡1\psi\equiv 1, which would yield a result comparable to how we measure error.

So, the main difference is as follows. Our paper makes the distinct contribution of FullNO via spatial restriction, clipping, a bounded-domain neural operator (such as standard FNO), and zero extension, with compactness controlling the truncation errors. This allows us to apply our approach to many off-the-shelf neural operator architectures, like FNOs.

References:

[RFARA] Md Ashiqur Rahman, Manuel A. Florez, Anima Anandkumar, Zachary E. Ross, and Kamyar Azizzadenesheli. Generative adversarial neural operators. Transactions on Machine Learning Research, 2022.

Reviewer_LdhtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigatss probabilistic conditioning as an operator mapping joint densities (eg obtained though KDE) to conditional densities. The paper then studies two variants in the settings, 1) kernel operator which produces all conditional densities and 2) the in-context operator which produces only the target conditional density based on the provided query y. The authors derive universal approximation guarantees for both operators and show numerical experiments with operators learned on Gaussian mixture densities.

Critical strengths

S1) The main strength of the paper is the formalization and theoretical investigation of amortised inference for conditioning with joints as distribution-valued inputs. As a general concept I find this very interesting and think that this is also of interest to a broader audience. I should add that I did not go through the proofs but I could follow the sketches in main text and was able to follow the main arguments.

S2) The paper is in general very well written and easuy to follow.

Critical weaknesses

W1) The empirical evaluation is not very thorough and I think it does not convincingly convey why one should learn the conditioning operation in the first place. If we already have access to the full joint, numeric integration will give the exact (or approximate) conditional. A main reason for learning this operation would be to fully leverage the „amortized“ idea, ie, at deployment time it iss significantly faster. Now I do believe that this could be the case but the current 2D experiments do not provide any kind of run-time comparison. I think also a small „accuracy vs computational time required“ would be very interesting even for the small scale and controlled experiments here. Lastly, I think just looking at Gaussians is not very interesting and having more complicated distributions would be more convincing. In summary, I see the empirical evaluation as only a rough proof of concept and with this it is the main weakness of the paper.

Questions for authors

Q1) In table 2, the FNO yield a better accuracy compared to the plugin KDE. I find this a bit surprising and can you give more intuition on this? Could it be that the network just learns a denoising of the KDE input and then with this yield more accurate conditioning? If that is the case, I think that is interesting and I could be worth a few sentences and maybe a small empirical test varying the number of samples to construct the KDE.

Q2) To yield a valid conditional density, one requires p(x | y) >=0 and the interval over x to be 1. If I understand the current architecture correctly, this is currently not enforced right? Have you checked if also without enforcing this you actually satisfy both conditions? Or how large are deviations here?

7 comments
AuthorsOpenReview ↗

Response to Reviewer Ldht

We thank the reviewer for their kind words and astute observations. We are conducting further numerical experiments to thoroughly address the points raised. For now, we note the following:

Probability axiom satisfaction (Q2).

The reviewer is correct that the constraints defining a probability measure (non-negativity and normalization) are not enforced explicitly in our architecture. We will provide numerics demonstrating the extent to which these constraints are violated.

We would also like to note that density estimators that are not bona fide probability densities have a long history in statistics, specifically in non-parametric density estimation. For example, higher-order kernel density estimators can take negative values. Tsybakov [1, Section 1.2.1, p. 10, "Positivity constraint"] explicitly discusses this issue and shows that taking the positive part cannot increase pointwise squared error. Orthogonal series density estimators provide another classical example [1, Section 1.7]. More generally, correcting estimators that are not bona fide densities has a long history; see, for example, Gajek [2]. For our L1L^1 setting, validity can be enforced by a simple post-processing step. For a fixed query yy, let fyf_y be an integrable raw prediction, define fy+(x)=max⁡fy(x),0f_y^+(x)=\max\\{f_y(x),0\\} and Zy=∫fy+(x) dxZ_y=\int f_y^+(x)\,dx, and set f~y(x)=fy+(x)Zy\widetilde{f}_y(x)=\frac{f_y^+(x)}{Z_y} when Zy>0Z_y>0, using a fixed probability density otherwise. For any true conditional density κy\kappa_y,

\le 2\|f_y-\kappa_y\|_{L^1}.$$ Indeed, the change caused by normalization is bounded by $\|f^+_y - \kappa_y \|_{L^1}$ and clipping cannot increase the $L^1$ error, namely $\|f^+_y - \kappa_y \|_{L^1} \leq \|f_y - \kappa_y \|_{L^1}$. Thus, exact density constraints can be enforced while preserving the $L^1$ approximation guarantee up to a factor of two. In the final manuscript we will clarify the fact that we are not enforcing hard probability distribution constraints and we will discuss ways to amend this. **References** [1] Alexandre B. Tsybakov. *Introduction to Nonparametric Estimation*. Springer, 2009. Section 1.2.1, p. 10, and Section 1.7, pp. 48–49. [DOI](https://doi.org/10.1007/b13794). [2] Lesław Gajek. "On Improving Density Estimators which are not Bona Fide Functions." *The Annals of Statistics*, 14(4):1612–1618, 1986. [DOI](https://doi.org/10.1214/aos/1176350182).

Thank you for providing more context around Q2. With the discussion deadline approaching, I wanted ask if you could give some further comments/intuition on Q1 potentially with the experiments that you have ran already?

AuthorsOpenReview ↗

Response to Reviewer Ldht (Part 2)

We thank the reviewer for the careful reading and constructive suggestions. We ran new experiments to address their points. All new results use freshly generated data following Appendix F.2.

We apologize for the delayed response and we will be monitoring this channel closely for the remainder of the discussion period to address any questions or concerns.

W1: Why learn conditioning & runtime considerations.

We agree that, when both methods use the same precomputed grid of joint-density values, the plug-in conditional requires only summation and normalization. Our experiments show no runtime advantage for the neural operator in this setting, and we will clarify this in the revision.

Learning across joint–conditional pairs can still improve accuracy when the joint is estimated from limited samples. At queries with very low marginal probability, few samples inform the KDE conditional; the operator can use structure learned from the training family to predict it. Additionally, when performing inference on a KDE input built from samples, a trained operator can learn to correct the KDE's bias, as the reviewer suggested. Following this intuition, we re-trained various neural operators on KDE inputs and obtained a further error decrease, 40–50% relative to the plug-in for n≥500n\ge 500.

A separate possibility is that a speed advantage may arise on larger problems with expensive joint-density evaluations if a learned operator reaches a given accuracy from fewer evaluations than quadrature. Our 64264^2 experiments do not test that possibility; they show an accuracy gain at similar runtime.

As a first step toward the sample-based setting, we report accuracy and wall-clock time of FNO and TNO Neural Operator architectures trained on KDE inputs as well as of the KDE plug-in. We consider 200 held-out joints from the 3-component Gaussian mixture family, we draw nn samples, build a Gaussian KDE on the 64×64 grid, and compute the conditional either by the plug-in formula or with an FNO or TNO trained on KDE inputs. Times are per instance, include the KDE construction, and are measured on an A100 GPU with batched inference.

Median relative L1L^1 error (median time per instance), 3-component mixture:

Methodn=500n=500n=2,000n=2{,}000n=20,000n=20{,}000
KDE + plug-in0.569 (0.7 ms)0.497 (2.5 ms)0.407 (24 ms)
KDE + FNO0.343 (0.9 ms)0.255 (2.8 ms)0.223 (25 ms)
KDE + TNO0.347 (0.7 ms)0.260 (2.6 ms)0.219 (24 ms)

We notice that

  1. The operator costs about the same as the plug-in. Its forward pass adds under 0.3 ms per instance on top of the KDE for n∈500,2,000n \in \\{500, 2{,}000\\} and 1ms for n=20,000n=20{,}000.
  2. The operators' error sits around 40–50% below the plug-in KDE at every nn.
  3. The operators improve only slowly at large nn (FNO: 0.255 →\to 0.223 from n=2,000n=2{,}000 to 20,00020{,}000). Our KDE-trained models saturate there: their training error kept decreasing while validation error plateaued, which suggests overfitting to the fixed set of KDE realizations seen in training.

W1: Beyond Gaussian joints.

Please look at updated PDF for plots. We re-train on a non-Gaussian "banana" family: x∼N(μ,σ2)x\sim\mathcal N(\mu,\sigma^2) and y=c+a(x−μ)2+b(x−μ)+εy=c+a(x-\mu)^2+b(x-\mu)+\varepsilon. Its conditionals p(x∣y)p(x\mid y) are bimodal and non-Gaussian, and the reference kernel is computed by quadrature. Detailed results and visualizations are presented in Section F.4 and in blue text in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.