Kernel Mean Embedding of Distributions: A Review and Beyond

1.1 Purpose and Scope

• The survey reviews kernel mean embeddings comprehensively, emphasizing theoretical results, practical applications, related areas, and challenging issues that may motivate research directions.It presents key theory and new applications while focusing on intuition and proof sketches, with full proofs left to cited papers.Source: Introduction · p. 6, §1.1 · p. 11, §1.2 · p. 12

Introduction

• Characteristic kernels make the mean map injective, enabling distribution metrics, two-sample testing, and extensions of learning algorithms to probability distributions.The embedding can capture all information about P and supports unique distribution characterization.Source: Introduction · p. 8
• Conditional mean embeddings represent conditional distributions as RKHS elements or operators, enabling nonparametric sum, product, and Bayes’ rules.These representations support probabilistic inference, graphical models, reinforcement learning, and causal discovery without parametric distributional assumptions.Source: Introduction · pp. 9–11

1.2. Outline of the Survey

• The survey covers RKHS foundations, marginal and conditional embeddings, estimation, distributional learning, probabilistic inference, causal discovery, related methods, and future research.It also discusses applications including MMD, dependence measures, graphical models, reinforcement learning, and conditional mean embedding challenges.Source: §1.2 · p. 13

1.3 Notation

• The notation section defines symbols for spaces, kernels, feature maps, Gram matrices, tensor products, and covariance and cross-covariance operators.Table 1.1 summarizes the commonly used notation and symbols throughout the survey.Source: §1.3 · p. 14, Table 1.1 · p. 14

2.1 Learning with Kernels

• Positive definite kernels guarantee feature maps into inner-product spaces, enabling nonlinear algorithms through the kernel trick without explicitly constructing high-dimensional features.The same framework supports kernels for Euclidean, structured, and functional data, while random Fourier features approximate infinite-dimensional kernels to reduce Gram-matrix costs.Source: §2.1 · pp. 17, 18, 19, 22, 23
• Bochner’s theorem characterizes translation-invariant positive definite kernels through non-negative spectral measures, whose frequency support can determine whether a kernel mean embedding uniquely identifies a distribution.The spectral measure also acts as a filter selecting characteristic-function frequencies in distributional similarity computations.Source: §2.1 · pp. 21–22

2.2 Reproducing Kernel Hilbert Spaces

• An RKHS is a Hilbert space of functions with bounded evaluation functionals, giving kernels a reproducing property and regularity that supports well-behaved learning and generalization.Each point has a representer k_x, the canonical feature map; positive definite kernels and RKHSs uniquely determine one another.Source: §2.2 · pp. 24–27

Background

• The background establishes Hilbert spaces as complete inner-product spaces and notes that the survey focuses on embedding distributions in such spaces rather than more general Banach spaces.Examples include Euclidean spaces, square-summable sequences, and square-integrable functions.Source: §2.2 · pp. 24–25

2.3 Hilbert-Schmidt Operators

• A separable Hilbert space is defined as one possessing a countable basis, providing the final foundational notion introduced in this part.The survey identifies Hilbert-Schmidt operators as ubiquitous in RKHS applications before introducing separability.Source: §2.3 · p. 27
• For separable Hilbert spaces H and F, a Hilbert-Schmidt operator A:F→H is bounded and has finite Hilbert-Schmidt norm, independent of the chosen orthonormal bases.Hilbert-Schmidt operators form the Hilbert space HS(F,H) with inner product ⟨A,B⟩HS = P⟨Af_j,Bf_j⟩H.Source: §2.3 · pp. 27–28
• The identity operator Id:H→H is Hilbert-Schmidt if and only if H is finite dimensional, while every separable infinite-dimensional Hilbert space is isometric to ℓ2.For infinite-dimensional H, the identity’s squared Hilbert-Schmidt norm diverges.Source: §2.3 · pp. 27–28
• Hilbert-Schmidt operators underpin kernel applications including dependence measurement, sufficient dimension reduction, kernel CCA, kernel PCA, kernel Bayes’ rule, and covariance-operator distances for multicategory image classification.The cited applications use the Hilbert-Schmidt norm of cross-covariance operators, extensions to random processes, or a Log-Hilbert-Schmidt metric.Source: §2.3 · p. 28

Distributions

• Kernel mean embeddings generalize feature maps from data points and Dirac measures to probability measures, mapping each distribution into an RKHS through an expectation operation.The construction extends the data-point feature map measure-theoretically and is summarized in Figures 3.1 and 3.2.Source: Distributions · p. 29, Figure 3.1 · p. 30, §3.1 · pp. 32–33

3.1 From Data Point to Probability Measure

• Dirac-measure embeddings reproduce ordinary kernel evaluations and are therefore equivalent to pointwise feature maps, motivating the move to non-trivial probability measures.Finite-support signed measures include empirical measures, while related constructions also cover Donsker measures and point processes.Source: §3.1 · pp. 30–32
• The embedding satisfies E_P[f(X)] = ⟨f, µP⟩_H for RKHS functions, turning expectation evaluation into an inner product useful in graphical models and probabilistic inference.This expectation reproducing property also extends to conditional distributions.Source: §3.1 · p. 34

3.1.1 Explicit Representation of Mean Embeddings

• Polynomial-kernel embeddings retain moments up to the kernel degree, so increasing the degree stores more information about the distribution.For the exponential inner-product kernel, the embedding is the moment-generating function when that function exists.Source: §3.1.1 · pp. 34–35
• Characteristic-function representations always exist, and for translation-invariant kernels the RKHS inner product of embeddings equals the L2(Λ) inner product of characteristic functions.The Fourier kernel itself is not positive definite, but characteristic functions represent embeddings for translation-invariant kernels on R^d.Source: §3.1.1 · pp. 35–36

3.1.2 Empirical Estimation of Mean Embeddings

• The empirical embedding averages kernel features over an i.i.d. sample and is unbiased, converging to the population embedding as sample size tends to infinity.The same expression is the mean embedding of the sample’s empirical distribution.Source: §3.1.2 · pp. 36–37
• Under suitable kernel assumptions, Hilbert-space distribution embeddings allow RKHS methods to operate on marginal probability measures, while conditional distributions are deferred to Section 4.The section focuses on P(X) rather than P(Y |X).Source: §3.1.2 · p. 37

3.2 Covariance Operators

• Covariance and cross-covariance operators extend their Euclidean matrix counterparts to infinite-dimensional RKHS elements.The section reviews these operators as important components of Hilbert space embeddings of distributions.Source: §3.2 · p. 37
• The uncentered cross-covariance operator CYX maps H to G through the tensor-product feature element ϕ(Y)⊗φ(X) under the joint distribution PXY.Its centered form is ˜CYX := EYX[ϕ(Y) ⊗φ(X)] −µPY ⊗ µPX = µPYX −µPY ⊗PX.Source: §3.2 · pp. 37–38
• Hilbert-Schmidt operators HS(H, G) are isomorphic to the tensor-product space H ⊗G, and this correspondence is an isometry.The identification uses the product kernel and the orthonormal basis {φa ⊗ϕb}a,b of H ⊗G.Source: §3.2 · p. 38
• The cross-covariance operator is well-defined as a Hilbert-Schmidt operator when the expected feature-kernel product is finite, as ensured by the stated moment assumption.The bound uses EYX[k(X, X)l(Y, Y )] ≤ {EX[k(X, X)]EY [l(Y, Y )]}1/2.Source: §3.2 · p. 38
• When X = Y, CXX is a positive self-adjoint covariance operator, while CXY is the adjoint of CYX.The operators also have integral expressions involving the joint distribution PXY and the marginal distribution PX.Source: §3.2 · p. 39
• If EYX[g(Y )|X = ·] ∈H for g ∈G, Theorem 3.2 establishes the key operator relationship used to define conditional mean embeddings.The section explicitly identifies this result as crucial for the conditional mean embedding introduced in Section 4.Source: §3.2 · p. 39
• Given i.i.d. samples from PXY, centered cross-covariance and covariance operators can be estimated empirically using the centering matrix and feature matrices.The centered estimate uses H = In −n−11n, with Φ and Ψ formed from the sampled feature maps.Source: §3.2 · p. 39

3.2. Covariance Operators

• The cross-covariance operator decomposes through marginal covariance operators and a unique bounded normalized cross-covariance operator V_YX with norm at most 1.V_YX captures dependence with less influence from marginal distributions and relates to canonical correlation and conditional dependence measures.Source: §3.2 · p. 40
• Covariance operators underpin kernel PCA, kernel Fisher discriminant, kernel CCA, kernel ICA, and independence or conditional-independence measures, while sharing eigenvalues with an associated integral operator.The covariance operator C_XX coincides with T_k when defined on L2(X, P_X).Source: §3.2 · p. 40

3.3 Properties of the Mean Embedding

• The empirical mean embedding converges to the population embedding at rate n^-1/2 under bounded continuous positive definite kernels on separable topological spaces.The guarantee holds with probability at least 1 − δ, under the stated boundedness assumption sup_x k(x, x) ≤ C_k < ∞.Source: §3.3 · p. 41
• The empirical mean embedding is minimax optimal, and √n(μ̂_P − μ_P) converges to a zero-mean Gaussian process on the RKHS.Tolstikhin et al. (2016) established minimax optimality, while the Gaussian-process limit is attributed to Berlinet and Thomas-Agnan (2004; Section 9.1).Source: §3.3 · p. 41

3.3.1 Characteristic and Universal Kernels

• A characteristic kernel makes the mean-embedding map injective, so ||μ_P − μ_Q||_H = 0 if and only if P = Q.Characteristic kernels were introduced in Fukumizu et al. (2004), and Gaussian and Laplacian kernels are characteristic on R^d (Fukumizu et al. 2008).Source: §3.3.1 · p. 41, §3.3 · p. 42

3.3. Properties of the Mean Embedding

• A universal kernel has an RKHS dense in C_b(X) on compact metric spaces, and universal kernels are therefore characteristic because they are dense in L2(X, P).Universality is defined by Steinwart (2002).Source: §3.3 · p. 42, Characterization of Characteristic… · p. 42

Characterization of Characteristic Kernels

• Characteristic kernels arise when H_k + R is dense in L_q(X, P), and translation-invariant kernels satisfy sufficient Fourier-transform conditions, including Gaussian and Laplacian kernels.For translation-invariant kernels on R^d, Fourier support conditions and strict positivity of the transform yield characteristicness (Fukumizu et al. 2008; Sriperumbudur et al. 2008; 2010; 2011a).Source: Characterization of Characteristic… · pp. 42–43
• Characteristic kernels remain characteristic after the conformal transformation k̃(x, y) = f(x)k(x, y)f(y) when f is bounded, continuous, strictly positive, and preserves bounded diagonals.This construction is given by Fukumizu et al. (2009b; Lemma 2).Source: Characterization of Characteristic… · p. 43

3.3. Properties of the Mean Embedding

• Universal kernels are characteristic, while characteristic kernels uniquely identify distributions; universality and characteristicness coincide for continuous translation-invariant kernels decaying at infinity.Gaussian and Laplace kernels are universal on compact domains, and strictly positive definite kernels are universal on discrete domains.Source: §3.3 · p. 45
• Characteristic kernels are essential for asymptotically valid hypothesis-testing statistics, but distributional prediction can use non-characteristic kernels as frequency-selective filters.For compactly supported distributions on R^d, translation-invariant kernels whose spectral support has non-empty interior can still be characteristic.Source: §3.3 · pp. 45–46
• Kernel choice can be less restrictive when embedding uniqueness is unnecessary, because the kernel then determines which characteristic-function frequencies appear in the representation.This makes the Fourier-domain shape of the kernel informative for predictive learning on distributional data (Muandet et al. 2012, Muandet and Schölkopf 2013, Oliva et al. 2014, Szabó et al. 2016).Source: §3.3 · pp. 45–46

3.4 Kernel Mean Estimation and Approximation

• The empirical kernel mean converges to the population embedding at rate O_p(n^-1/2), and is minimax when no information about the underlying distribution is available.This motivates seeking improved estimators despite the empirical estimator’s optimality in the unrestricted setting.Source: §3.4 · p. 47

3.4.1 Kernel Mean Shrinkage Estimators

• Muandet et al. (2014a; 2016) propose shrinking the empirical embedding toward a fixed RKHS element, reducing variance for α ∈ (0, 2) at the cost of increased bias.The shrinkage parameter can be selected by minimizing estimated mean-squared error, and Muandet et al. (2016) use leave-one-out cross-validation for a data-dependent regularization choice.Source: §3.4.1 · pp. 47–49
• Spectral shrinkage estimators project data onto the KPCA basis, shrink components independently according to a filter function, and reconstruct the embedding.Efficient procedures such as Landweber iteration and iterated Tikhonov avoid explicit eigendecomposition of the kernel matrix.Source: §3.4.1 · p. 49

3.4.2 Approximating the Kernel Mean

• Kernel mean approximation addresses quadratic scaling and infinite-dimensional feature maps through either sparse subsets of samples or finite-dimensional representations.Sparse methods can have linear sample-size complexity and benefit repeated evaluations, but require solving a non-trivial subsampling optimization problem.Source: §3.4.2 · p. 51
• Random feature maps produce compact finite-dimensional kernel mean estimates, accelerate downstream algorithms, and preserve similarity when sampled from the Fourier distribution of translation-invariant kernels.Their limitation is that random features apply only to certain kernel classes.Source: §3.4.2 · pp. 51–52

3.5 Maximum Mean Discrepancy

• MMD is the integral probability metric over an RKHS unit ball and equals the RKHS distance between the mean embeddings of P and Q.It also admits a kernel-expectation expression involving independent samples from P and Q.Source: §3.5 · pp. 53–54
• For characteristic kernels, MMD[H, P, Q] = 0 if and only if P = Q; for translation-invariant kernels, it is an L2(Λ) distance between characteristic functions.MMD is also bounded by the Wasserstein and total variation distances under the stated bounded-kernel condition.Source: §3.5 · p. 54
• MMD estimators use U- and V-statistics: U-statistics are unbiased, whereas V-statistics are generally biased, and quadratic-time estimates have established asymptotic behavior.The unbiased empirical MMD is formed from kernel evaluations within and between the two samples.Source: §3.5 · pp. 54–56
• MMD converges at order 1/√n independently of dimension d, but dimension-dependent constants and reduced test power can still cause a curse of dimensionality.Under the stated bound, |MMD_u^2[H, X, Y] − MMD^2[H, P, Q]| ≤ c_d(m^-1/2 + n^-1/2), with c_d depending on d but not m or n.Source: §3.5 · p. 57
• MMD two-sample tests are distribution-free, support structured domains with positive definite kernels, and can use asymptotic test-statistic distributions for improved empirical critical values.Large-deviation tests are generally more conservative because their critical values are distribution independent (Gretton et al. 2012a).Source: §3.5 · pp. 57–58

3.5.1 Scaling up MMD

• Scaling methods reduce quadratic cost O(n^2d): linear-time subsampling reaches O(nd) with higher variance, B-tests use O(n^3/2d), and Fastfood can reach O(n log d).Random-feature and optimized-location tests can retain or improve power while providing interpretable distribution-discriminating features (Zhao and Meng 2015; Chwialkowski et al. 2015; Jitkrittum et al. 2016).Source: §3.5.1 · pp. 58–60

3.5.2 Application: Training Deep Generative Models via MMD

• MMD replaces GAN discriminators with an RKHS-based two-sample test, yielding a closed-form discriminator, exact distribution matching for characteristic kernels, and a more computationally efficient objective.The MMD test avoids the hard minimax objective used by traditional GANs (Li et al. 2015; Dziugaite et al. 2015).Source: Figure 3.3 · p. 62, §3.5.2 · p. 63

3.6 Kernel Dependency Measures

• HSIC (Gretton et al. 2005a) measures nonlinear dependence as the squared Hilbert–Schmidt norm of an RKHS cross-covariance operator.It is motivated by limitations of rank-based criteria that detect only linear dependence and of parametric nonlinear measures requiring restrictive distributional assumptions.Source: §3.6 · p. 64
• With a characteristic product kernel, HSIC(X, Y, k, l) = 0 if and only if X and Y are independent.The result follows because zero cross-covariance equates the joint and product kernel mean embeddings, which characteristicness then identifies with equality of the corresponding distributions.Source: §3.6 · p. 64
• For i.i.d. samples, HSIC has an empirical unbiased estimator equivalent to (1/n^2)tr(˜K˜L), using centered Gram matrices.The estimator is also equivalent to Achard et al.’s (2003) quadratic dependence measure, although a zero value in their test does not imply independence.Source: §3.6 · pp. 64–65
• The empirical HSIC estimate can be computed in O(n^2) time and converges to population HSIC at rate 1/√n when bias is negligible.Song et al. (2007b) later provided a slight improvement for an unbiased HSIC estimator.Source: §3.6 · p. 65
• HSIC supports independence testing through a critical region derived from a large-deviation bound at significance level α, with constants depending only on the kernels.The supplied passage specifies the bound’s dependence on α, n, and a kernel-dependent constant C, but its displayed statistic is incomplete.Source: §3.6 · p. 65
• Dependence maximization with HSIC enables supervised feature selection and CLUHSIC clustering, which includes k-means, weighted k-means, and hierarchical clustering as special cases.The framework also makes MaxMMD equivalent to dependence maximization and extends to applications such as kernelized sorting.Source: §3.6 · p. 65

3.6. Kernel Dependency Measures

• Independence is tested by the MMD between PXY and PX ⊗ PY, estimated by pairing observed samples with randomly permuted counterparts.The permutation approximates samples from PX ⊗ PY and becomes accurate as n →∞, although it introduces negligible dependence among permuted variables.Source: §3.6 · p. 66
• Extensions of HSIC cover conditional dependence, Lancaster three-variable interaction, relative dependency, and arbitrary-variable testing, with correlated-statistic tests improving power and consistency.As the number of variables grows, estimator convergence may become arbitrarily slow; conditional dependence measures are discussed in §4.7.Source: §3.6 · pp. 66–67

3.6.1 Extensions to Non-i.i.d. Random Variables

• For non-i.i.d. variables, kernel dependence tests incorporate structural or mixing assumptions, including clique-based Structured-HSIC and spectral or shifted-signal tests for time series.Structured-HSIC applies to graphical-model dependencies, while time-series methods use the kernel cross-spectral density operator or shifted null estimates under mixing conditions.Source: §3.6.1 · p. 67

3.7 Learning on Distributional Data

• Kernel mean embeddings represent probability distributions as features, enabling learning for noisy data, multiple-instance learning, anomaly detection, statistical estimation, and causal inference while reducing effective sample size and concealing individual identities.This representation supports distributional learning without requiring the data to be treated as individual points.Source: §3.7 · p. 68

3.7.1 Kernels on Probability Distributions

• Distributional kernels operate on embedded measures, can be approximated with finite features for efficiency, and inherit consistency, reduced dimensionality issues, and kernel-method generalization tools without parametric distribution assumptions.The kernel can be formed by integrating an input-space kernel over two distributions, with empirical sums providing its sample-based counterpart.Source: §3.7.1 · pp. 68–69

3.7.2 Properties of Distributional Kernels

• For compact X and universal embedding kernel k, the induced RKHS is dense in the associated function class, while an injective map P 7→µP yields density in Cb(M1+(X)).The basic distributional kernel is linear over probability measures, so it is not universal on distributions in the usual Steinwart (2002) sense; nonlinear kernels address this limitation.Source: §3.7.2 · pp. 69–70

3.7.3 Distributional Risk Minimization

• Distributional risk minimization expresses any regularized solution through the kernel mean embeddings of the training distributions, generalizing the classical representer theorem and supporting outputs that may be binary, real-valued, structured, or distributional.The framework underlies support measure machines, including multiscale SVMs for Gaussian inputs and applications such as natural image categorization and group anomaly detection.Source: §3.7.3 · pp. 71–73
• The distributional classification bound converges to zero as both the number of distributions n and within-distribution sample size n_i grow, requiring √(log n/n_i) = o(1) and only n_i on the order of √log n for one bound term.This analysis captures the two-stage sampling structure in which samples, rather than complete distributions, are observed.Source: §3.7.3 · pp. 73–74, Related Works on Learning from… · p. 77

Related Works on Learning from Distributions

• Compared with generative-model, KDE, and uncertainty-aware kernels, kernel mean embeddings learn directly from distributions without strong parametric assumptions and can handle structured data, although KDE convergence is slow in high dimensions.Related equivalences connect GMMK, PPK with ρ = 1, and the linear embedding kernel for k(x, x′) = δ(x − x′), while expected-kernel SOCP methods can be more computationally demanding than QP.Source: Related Works on Learning from… · pp. 76–77

3.8 Recovering Information from Mean Embeddings

• The section relates distributional pre-images and kernel herding because both seek input-space objects corresponding to a specified kernel mean embedding.The distributional pre-image extends the classical pre-image problem from points to probability distributions, while herding obtains samples from the represented distribution.Source: §3.8 · p. 78

3.8.1 Distributional Pre-Image Problem

• Distributional pre-images approximate an underlying distribution by optimizing a parameterized family against an estimated embedding, but the recovered pre-image need not represent the MAP estimate.For point-mass distributions, the problem reduces to the classical pre-image problem; mixture optimization can be non-convex in component means and variances.Source: §3.8.1 · pp. 78–79
• Weighted averages from a consistent embedding estimator can recover moments, interval measures, and density information, but an error in Kanagawa and Fukumizu (2014) invalidated its original guarantees.Kanagawa et al. (2016b) later proved convergence for functions in a suitable power of the RKHS, including misspecified cases where f ∉ H.Source: §3.8.1 · pp. 80–81

3.8.2 Kernel Herding

• Kernel herding deterministically generates pseudo-samples whose empirical moments asymptotically match the data, and its RKHS expectation error decreases at O(1/T) versus O(1/√T) for random samples.The rate remains guaranteed with some herding error and is attributed to negative autocorrelation that favors unexplored high-density regions.Source: §3.8.2 · pp. 81–82
• Herding connects to Bayesian quadrature, determinantal point processes, convex optimization, and quasi-Monte Carlo, while high-dimensional or infinite-dimensional settings can undermine its fast-rate guarantees.Convex-optimization and Frank–Wolfe interpretations address some computational and convergence concerns.Source: §3.8.2 · pp. 82–83

Distributions

• Conditional mean embedding represents P(Y|X) as a family of RKHS embeddings indexed by X, equivalently an operator mapping H to G, thereby modeling functional dependence.It supports applications including dynamical systems, reinforcement learning, latent-variable models, kernel Bayes’ rule, and causal discovery (Song et al. 2009; 2013).Source: Distributions · p. 84, Figure 4.1 · p. 85

4.1.1 Conditional Mean Embeddings

• The conditional embedding U_Y|X maps k(x,·) to U_Y|x, while U_Y|x represents conditional expectations through ⟨g,U_Y|x⟩_G for every g∈G.Under E_Y|X[g(Y)|X]∈H, the defined embedding satisfies these operator and reproducing properties (Song et al. 2009; Fukumizu et al. 2004).Source: §4.1.1 · p. 87

4.1. From Marginal to Conditional Distributions

• Regularization replaces the potentially nonexistent C_YX C_XX^-1 with C_YX(C_XX+λI)^-1, whose empirical estimator is consistent under mild conditions.The unregularized operator may fail in continuous domains; for Gaussian RKHSs, even constant conditional expectations can lie outside the required RKHS, whereas discrete characteristic-kernel settings satisfy the condition.Source: §4.1 · p. 88, §4.1.2 · p. 88

4.1.2 Empirical Estimation of Conditional Mean Embeddings

• The empirical conditional embedding uses β=(K+nλI_n)^-1k_x, so coefficients depend on the conditioning value, need not be positive, and need not sum to one.Regularization is required because the empirical covariance operator can have arbitrarily small positive eigenvalues in infinite-dimensional RKHSs; spectral filtering offers alternatives.Source: §4.1.2 · p. 89, §4.2 · p. 90

4.2. Regression Interpretation

• O(n^-b/(4b+1)) is the convergence rate under polynomial eigenvalue decay, improving the general Op((nλ)^-1/2+λ^1/2) rate for conditional embeddings.Conditional embeddings are harder to estimate than marginal embeddings, which converge at Op(n^-1/2).Source: §4.2 · p. 90

4.2 Regression Interpretation

• O(log(n)/n) is the minimax convergence rate derived from vector-valued regression when G is finite dimensional, compared with O(n^-1/4) for Song et al. (2009).The regression formulation supplies a well-defined loss for cross-validation and enables performance analysis using vector-valued regression theory.Source: §4.2 · pp. 91–92

4.3 Basic Operations: Sum, Product, and Bayes’ Rules

• Kernel mean embeddings implement sum, product, and Bayes’ rules directly in RKHSs without parametric assumptions on the involved distributions.These operations manipulate marginal and conditional embeddings to obtain an embedding of the resulting distribution (Song et al. 2009; 2013; Fukumizu et al. 2013; Schölkopf et al. 2015).Source: §4.3 · pp. 92–93

4.3.1 Kernel Sum Rule

• The kernel sum rule gives µ_X=U_X|Yµ_Y, whereas the product rule factors the joint embedding through conditional embeddings and tensor-product feature representations.Unlike the sum rule, the product rule takes E_X[φ(X)⊗φ(X)] as input and can be expressed using Hilbert-Schmidt operators.Source: §4.3.1 · p. 93, §4.3.2 · p. 94

4.3.2 Kernel Product Rule

• Kernel product and sum operations require no parametric assumptions but can be statistically difficult and computationally costly in some applications.These operations are implemented through regularized operator and coefficient-matrix formulas.Source: §4.3.2 · p. 95

4.3.3 Kernel Bayes’ Rule

• Kernel Bayes’ rule obtains posterior embeddings nonparametrically from prior and likelihood embeddings using kernel sum and product rules, with consistency guarantees under stated assumptions.The posterior embedding and posterior expectations converge in probability; reported rates include O_p(n^-1/3α) in L2(P), while practical convergence may be faster.Source: §4.3.3 · pp. 95–97

4.3.4 Functional Operations on Mean Embeddings

• Mean embeddings estimate distributions of arbitrary measurable functional operations on independent random variables directly, avoiding density estimation and extending beyond simple arithmetic transforms.Consistency and finite-sample guarantees extend to continuous functions of multiple variables, though dependent variables require a consistent joint embedding.Source: §4.3.4 · pp. 97–98
• Regularization replaces potentially ill-defined inverse covariance operators in conditional embeddings and converges under sufficiently slow decay of λ_n, with cross-validation commonly selecting λ_n.Strong range conditions can be impractical; Gaussian kernels fail such assumptions for a wide class of distributions.Source: §4.3.4 · pp. 98–99

4.4 Graphical Models and Probabilistic Inference

• Graphical-model inference applies kernel sum, product, and Bayes operations to embeddings nonparametrically, accommodating structured non-Gaussian continuous variables without density estimation.Filtering uses embedding-based prediction and conditioning, while embedding-based HMM learning requires only unlabeled observation sequences and avoids local search.Source: §4.4 · pp. 99–102
• Embedding-based graphical inference remains sensitive to kernel choices, may require pre-image recovery for full posteriors, and can be computationally costly, often prompting low-rank approximations.Conditional-operator estimation can also be statistically difficult and computationally costly during filtering.Source: §4.4 · pp. 100–101

4.5 Markov Chain Monte Carlo Methods

• Kernel adaptive Metropolis-Hastings uses RKHS covariance information to adapt proposals to nonlinear target support, while kernel Hamiltonian Monte Carlo addresses doubly intractable targets without energy gradients.The RKHS proposal is oriented toward nearby high-density regions and avoids wrongly scaled proposals; RKHS exponential-family surrogates are learned by score matching.Source: §4.5 · p. 103, §4.6 · p. 104

4.6 Markov Decision Processes and Reinforcement Learning

• Conditional mean embeddings represent MDP transition dynamics and estimate Bellman operators, but unnormalized coefficients can destroy contraction and isotonicity, motivating normalized estimates.Kernelized value iteration extends to POMDPs through kernel Bayes’ rule, although the POMDP algorithm requires labeled latent-state training data.Source: §4.6 · p. 106

4.6. Markov Decision Processes and Reinforcement Learning

• Kernel mean embeddings represent predictive states nonparametrically, support continuous actions and observations, update states via kernel Bayes’ rule, and can model nonlinear dynamics despite linear PSR prediction functions.They avoid kernel density estimation’s slow convergence and difficult numerical integration, while expectation computation becomes an RKHS inner product with linear complexity in the number of training points.Source: §4.6 · p. 107

4.7 Conditional Dependency Measures

• Kernel conditional-independence tests address nonlinear and non-Gaussian settings where classical tests rely on Gaussianity, linearity, or discretization and remain vulnerable to the conditioning variable’s curse of dimensionality.Conditional independence is central to Bayesian-network learning, gene-expression analysis, and causal discovery.Source: §4.7 · pp. 108–109
• Kernel conditional-dependence measures use conditional cross-covariance operators whose vanishing characterizes X and Y as conditionally independent given Z, but their null distributions may require computationally expensive bootstrapping.Alternative approaches derive the null distribution asymptotically or convert conditional independence into unconditional testing through residualization or Z-aware permutations.Source: §4.7 · pp. 108–109

4.8 Causal Discovery

• Kernel embeddings support causal discovery by enabling conditional-independence-based DAG recovery, bivariate causal inference through kernel probabilistic programming and distribution classification, and causal-direction criteria based on distributional complexity.The approaches include embedding additive-noise structural equation models and classifying empirical cause-effect distributions without handcrafted features.Source: §4.8 · pp. 110–111
• Kernel mean embeddings detect finite hidden confounders by decomposing a joint distribution into mixture components, with the number of components identified exactly and confounders then found by clustering.The method addresses a major challenge for causal inference algorithms: hidden confounding.Source: §4.8 · p. 111

5.1 Beyond Density Estimation and Characteristic Function

• KME reframes density-based comparisons as biased MMD, bypassing difficult density estimation, especially in high dimensions.The same perspective supports direct divergence approximation through density-ratio estimation and regression.Source: §5.1 · p. 113

5.2 Classical Kernel Machines

• Classical classifiers, anomaly detectors, and centering operations can be reinterpreted through kernel means of distributions, linking MMD to empirical risk minimization and large-margin classification.Class-conditional means correspond to embeddings of P(X|Y = +1) and P(X|Y = −1).Source: §5.2 · pp. 114–115

5.3 Distance-based Two-Sample Test

• Energy distance and distance covariance equal distances between KME distributions for suitable kernels, while distance-kernel HSIC generalizes dependence measures beyond Euclidean distances.Using semi-metrics of negative type gives kernel-based methods a more flexible choice for dependence measures.Source: §5.3 · p. 115

5.4 Fourier Optics

• Fraunhofer diffraction is a special case of KME, and bounded-support objects can in theory be completely recovered from diffraction-limited images because the embedding is injective.Harmeling et al. (2013) also provides a practical inversion approach.Source: §5.4 · p. 116

5.5 Probabilistic Perspective

• Under a Gaussian-process prior, the kernel mean is an expected covariance of induced functions, connecting KME to probabilistic distance constructions.Despite this connection, a fully Bayesian interpretation of kernel mean embedding remains an open question.Source: §5.5 · pp. 116–117

Future Directions

• The survey frames these unresolved questions as important directions for future research in kernel mean embedding.It presents the field as impactful while emphasizing that numerous open problems remain.Source: Future Directions · p. 118

6.1 Kernel Choice Problem

• Choosing the “right” kernel and its parameter values remains a fundamental open problem because finite samples may prevent characteristic kernels from preserving all practically relevant distributional information.Prior knowledge may therefore be important for kernel selection, and related issues have been studied in Sriperumbudur et al. (2009) and Gretton et al. (2012b).Source: §6.1 · p. 118

116 Future Directions

• Future work should strengthen distribution representations because they are essential for several statistical applications.Source: 116 Future Directions · p. 119

6.2 Stochastic Processes and Bayesian Interpretation

• Extending kernel mean embeddings to stochastic processes and Bayesian nonparametric frameworks could support probabilistic inference and Bayesian interpretations, including extensions related to Gaussian processes.Stochastic processes involve distributions over infinite-dimensional objects, unlike the finite-dimensional objects typically handled by current formulations.Source: §6.2 · p. 119

6.3 Scalability

• Scalable kernel-learning advances are needed so kernel mean embedding algorithms can handle increasingly complex and large-scale data, including through recent MMD methods.Source: §6.3 · p. 119

6.4 High-dimensional Inference

• High-dimensional inference requires theory clarifying how ambient dimension affects kernel methods, shrinkage estimators, MMD, and related tests.Kernel methods are less prone to the curse of dimensionality than classical approaches such as kernel density estimation, but their underlying theory remains limited; see Reddi et al. (2015) and Ramdas et al. (2015a;b).Source: §6.4 · pp. 119–120

6.5 Causality

• Kernel mean embeddings may represent counterfactual outcome distributions, but causal inference remains challenging because treatment and control potential outcomes cannot both be observed.Existing directions include bivariate causal inference as classification on joint cause-effect distributions (Lopez-Paz et al. 2015).Source: §6.5 · p. 120

6.6 Privacy-Preserving Embedding

• Future work must determine whether kernel mean embedding methods can preserve privacy and comparable performance while avoiding direct access to individual data and protecting distribution-level information.Applications such as two-sample testing and dependency measures can expose individual data points to adversarial attack.Source: §6.6 · pp. 120–121

6.7 Statistics, Machine Learning, and Invariant Distribution Embeddings

• A key direction is to characterize distribution-regression estimators and connect them to classical statistical estimators while developing invariant representations across domains.Using kernel mean embeddings, Szabó et al. (2016) analyzed distribution regression theoretically, while Flaxman et al. (2015b) and Lopez-Paz et al. (2015) applied it to ecological and causal problems.Source: §6.7 · p. 121

• Kernel mean embedding maps distributions into an RKHS, extending kernel methods to probability measures and supporting applications across statistics and machine learning.The survey identifies it as a powerful tool for probabilistic modeling, statistical inference, machine learning, and causal discovery.Source: Conclusions · p. 122
• The survey establishes RKHSs, positive definite kernels, and Hilbert-Schmidt operators as foundations, then relates kernel mean embedding to other areas and identifies open problems for future research.Source: Conclusions · pp. 122–123
• Marginal distribution embeddings enable RKHS methods for probability measures, supporting two-sample and independence testing, anomaly detection, MCMC, predictive learning, and causal inference.Source: Conclusions · p. 123
• Conditional mean embeddings provide non-parametric sum, product, and Bayes’ rules for graphical models, probabilistic inference, and reinforcement learning, and correspond naturally to vector-valued regression.Source: Conclusions · p. 123