Multivariate Conditional Expectation (MUCE)

• Multivariate Conditional Expectation (MUCE) is the generalization of classical conditional expectation to vector-valued functions, essential for probability, statistics, and machine learning.
• Operator-theoretic and RKHS-based estimation methods enable practical spectral algorithms and regularized solutions for complex, high-dimensional data.
• Extensions to Banach spaces and measure-theoretic formulations facilitate robust optimization, risk assessment, and improved model interpretability in multivariate settings.

Multivariate Conditional Expectation (MUCE) is a foundational concept generalizing the classical conditional expectation to random vectors and functions of multiple random variables, with central applications across probability theory, functional analysis, statistics, machine learning, and robust optimization. The MUCE, in its most basic form, is the mapping $x \mapsto \mathbb{E}[f(X,Y)\mid X=x]$, where $(X,Y)$ is a random pair (or more generally a random vector), and $f$ is an integrable function. Recent advances extend MUCE to measure-theoretic, operator-theoretic, and structural modeling frameworks, embracing Banach space-valued functions, count vectors, truncated distributions, robust bounds, and machine learning explainability.

1. Fundamental Concepts and Definitions

The classical MUCE is defined for $(X,Y)\sim\mu$ (with $X$ taking values in a topological space and $Y$ in a measurable space) as

$$\bar{f}(x) := \mathbb{E}^\mu[f(X,Y)\mid X=x] = \int_Y f(x, y) \, d\mu(y|x),$$

where $d\mu(y|x)$ is the disintegration of $\mu$ given $X=x$ (Das, 2023). Under mild continuity assumptions on $f$ and the kernel $x\mapsto\mu(\cdot|x)$, $\bar{f}$ admits a continuous representative.

For vector-valued random variables $Y\in\mathbb{R}^n$, the MUCE becomes $\mathbb{E}[Y\mid X]$ with all components, orderings, and constraints generalized; in conditional expectation theory for Banach spaces, multifuntional extensions deliver set-valued expectations (Musial, 2023). In the context of generalized skew-elliptical distributions, MUCE under truncation yields closed-form expressions for doubly truncated and tail-conditional moments (Zuo et al., 2022).

2. Operator-Theoretic and RKHS-Based Estimation

The RKHS-based compactification operator approach estimates MUCE as the solution of a regularized linear inverse problem:

• Choose a symmetric positive-definite kernel $k$ yielding RKHS $H\subset C(X)$.
• Define the kernel integral operator $K^\nu$ and the compactification operator $P^{\mu_X}$ (Das, 2023).
• The exact operator equation is

$P^{\mu_X}K^{\mu_X}\bar{f} = Q^\mu_\delta f,$

where $Q^\mu_\delta$ is a $\delta$-smoothed smoothing operator.

Given empirical proxies for the data, this reduces to solving

$P^{\alpha_X}K^\nu \phi = Q^\alpha_\delta f,$

for $\phi$ via Tikhonov regularization, leading to practical spectral algorithms. Vector-valued and manifold cases extend via product kernels and geometry-aware smoothing, preserving convergence rates even in high dimensions.

Data-driven solutions use kernels and smoothing operators with spectral filtering, yielding convergence guarantees in both infinite- and finite-sample regimes. Numerical algorithms involve kernel matrices, Markov matrices, and regularized least squares solutions, with accuracy governed by regularization, smoothing bandwidth, and landmark count.

3. Structural Models and Compatibility for Count Data

The MUCE for count-valued random vectors $Y\in\mathbb{N}_0^d$ is addressed through linear conditional specification. The central problem is characterizing when a system of linear conditional expectations

$$\mathbb{E}[Y_i \mid Y_{-i}] = \alpha_i + \sum_{j\ne i}\beta_{ij} Y_j$$

is compatible, i.e., arises from a bona fide joint probability law (Lu et al., 27 Feb 2025). Key results include:

• Compound Autoregressive (CAR) and Random Coefficient Integer Autoregressive models deliver exact, limited solutions in bivariate and low-dimensional cases, such as Poisson-Gamma and Beta-NB conjugacy.
• General semi-parametric MUCE models—imposing only the linear conditional expectation—admit a wide solution class, contingent on spectral constraints ($\rho(B)<1$ for regression matrix $B$).
• Existence results leverage Farkas' Lemma and M-matrix conditions; estimation is computationally efficient via composite/quasi-likelihoods, with no demanding Markov chain Monte Carlo required.

4. Measure-Theoretic, Banach-Space, and Functional Extensions

MUCE generalizes to Banach space-valued multifunctions under the Radon-Nikodym Property (RNP), supporting set-valued conditional expectations and multimeasures (Musial, 2023). Let $\mathcal{G}:\Omega\to\mathrm{cb}(X)$ map into the space of nonempty, closed, convex, bounded subsets of Banach space $X$.

• Scalar measurability and Pettis integrability enable definition of MUCE as an $\mathscr{X}$-measurable multifunction.
• Main representation result:

$$\mathbb{E}[\mathcal{G}\mid\mathscr{X}](\omega) = \overline{\left\{\mathbb{E}[\xi\mid\mathscr{X}](\omega) : \xi\in S(\mathcal{G})\right\}}$$

where $S(\mathcal{G})$ is the family of measurable selections.

• Effros measurability, extremal selection theorems, and lifting techniques apply in nonseparable spaces, giving a robust framework for MUCE in vector- and set-valued function spaces.

Multidimensional Fatou lemmas extend the analytic backbone to conditional expectation with cone constraints, providing order-monotonicity and lower semicontinuity. These properties are key for multivariate risk measures, decision models, and constrained stochastic optimization (Babaei et al., 2018).

5. Statistical Estimation, Moment-Based Bounds, and Robust Optimization

In statistical and stochastic programming contexts, MUCE plays a central role in distributionally robust optimization (DRO). Distribution-free bounds are sharp when only moment constraints and conditional events are specified (Eekelen, 2023). The fundamental problem is bounding

$$\sup_{P\in\mathcal{M}_+(\Omega)} \frac{\mathbb{E}_P[g(X)1_C(X)]}{\mathbb{E}_P[1_C(X)]}$$

under moment constraints. The solution involves:

• Charnes–Cooper transformations to linearize fractional problems.
• Semi-infinite LP and conic duality yields extremal distributions supported on $m+1$ atoms (with $m$ moments), reducible to tractable SDPs for polynomial data.
• Strong duality delivers sharp attainability; robust bounds apply directly in DRO models with side information (contextual newsvendor, mean–covariance ambiguity sets).

Monte Carlo estimation of MUCE is formalized via L²-projection minimization; numerical error bounds are available for regression approximations (linear, polynomial, neural network families) (Cheridito et al., 2021). Confidence intervals for error rates exploit CLT or Chebyshev bounds.

6. MUCE in Explainability and Machine Learning

MUCE is uniquely suited for explainable AI in black-box predictive modeling (Ruiz-España et al., 12 Jan 2026).

• Generalizes Individual Conditional Expectation (ICE) by varying multiple features jointly, sampling multi-dimensional grids.
• Quantifies local feature interaction effects on predictions; surfaces and heatmaps expose nonlinear behaviors missed by univariate profiling.
• Key metrics:
  • Stability index: measures prediction sensitivity to feature perturbation.
  • Uncertainty index (and asymmetric uncertainty$^+$/$^-$): quantifies the model's local irreducibility.
• Applied to XGBoost classifiers/regressors, MUCE demonstrates enhanced interpretability near decision boundaries in synthetic and mixed-type real datasets. Complementary use with ICE enables rapid ranking of brittle features.

7. MUCE for Truncated, Skewed, and Heavy-Tailed Multivariate Laws

For generalized skew-elliptical (GSE) distributions, explicit formulae for doubly-truncated expectation and covariance extend MUCE to skew-normal, skew-t, skew-Laplace, and skew-logistic cases (Zuo et al., 2022):

• The doubly-truncated MUCE is

$$\mathbb{E}[Y \mid a < Y \le b] = \mu + \Sigma \frac{\mathbf{d}}{F_Z(\xi_a, \xi_b)},$$

with all quantities specified via generator and skew CDFs.

• Tail-conditional MUCE (MTCE) is defined analogously, replacing the upper bound with $+\infty$, thus accommodating tail risk assessment in high-dimensional, heavy-tailed settings.

Practical existence and computation hinge on appropriate choices of generator functions and bounding conditions. Applications include tail-risk measurement, elliptical mixture modeling, and multivariate insurance.

---

MUCE encompasses a broad spectrum of theory and practice. Its operator-theoretic, measure-theoretic, and structural model formulations connect pure probability with applied statistics, functional analysis, stochastic optimization, machine learning interpretability, and multivariate risk assessment.