Joint Asymptotic Normality

Updated 26 November 2025

Joint asymptotic normality is defined as the convergence in distribution of vector-valued statistics to a Gaussian limit under averaged correlation conditions.
It employs characteristic function analysis and the Cesàro mean of cross-correlations to ensure decaying dependence and achieve joint independence.
The framework extends classical central limit theory to high-dimensional and dependent data, enabling robust statistical inference in diverse applications.

Joint asymptotic normality describes the emergence of joint multivariate Gaussian limiting distributions for collections of suitably normalized statistics—typically means, functionals, or other estimators—constructed from independent or dependent random structures. The term is most precisely defined via convergence in distribution of vector-valued statistics to a multivariate normal law, with both covariance structure and independence or dependence precisely characterized in terms of underlying probabilistic quantities such as covariances, correlations, or high-moment asymptotics. Recent research has clarified necessary and sufficient conditions for joint convergence to various forms of multivariate normality, extending classical central limit theory to high-dimensional, dependent, and heterogeneous data regimes.

1. Foundational Setting: Joint Asymptotic Normality of Standardized Sample Means

Consider an infinite sequence of square-integrable independent random vectors $(X_i,Y_i)$ in $\mathbb{R}^2$ , with means $\mu_X,\mu_Y$ and nonzero finite variances $\sigma_X^2,\sigma_Y^2$ . Define sample means and standardized versions: $\overline X_n = \frac{1}{n}\sum_{i=1}^n X_i, \quad \overline Y_n = \frac{1}{n}\sum_{i=1}^n Y_i,$

$S_n^{(X)} = \frac{\sum_{i=1}^n (X_i-\mu_X)}{\sigma_X\sqrt{n}}, \quad S_n^{(Y)} = \frac{\sum_{i=1}^n (Y_i-\mu_Y)}{\sigma_Y\sqrt{n}}.$

The focus is on the joint distribution of $(S_{n_1}^{(X)}, S_{n_2}^{(Y)})$ and its convergence as $n_1,n_2\to\infty$ .

A key quantitative tool is the sequence of “instantaneous” pairwise correlation coefficients $\rho_k = \mathrm{Corr}(X_k, Y_k)$ , and their Cesàro mean: $C_n = \frac{1}{n}\sum_{k=1}^n \rho_k.$ The principal result is the following:

Theorem (Majumdar & Majumdar). The pair $(S_{n_1}^{(X)}, S_{n_2}^{(Y)})$ converges in distribution to a jointly standard normal vector with independent components if and only if $C_n \to 0$ as $n \to \infty$ (Majumdar et al., 2017).

This necessary and sufficient condition directly connects the vanishing of averaged cross-correlations to asymptotic independence and joint normality.

2. Characteristic Function Analysis and Lévy Continuity for Nets

The proof utilizes factorization properties of the joint characteristic function: $\varphi_{n_1,n_2}(s,t) = E\left[e^{i(s S_{n_1}^{(X)} + t S_{n_2}^{(Y)})}\right].$ Under independence of the individual vectors, this decomposes, and the “cross-term” involves $C_m$ where $m = \min(n_1, n_2)$ . Lévy’s continuity theorem for nets is applied to deduce convergence in the double-indexed (net) sense, even when $n_1 \ne n_2$ . Uniform control over the cross-terms, which decay under $C_n \to 0$ , ensures joint convergence to the bivariate standard normal distribution.

Necessity follows from observing that

$\mathrm{Cov}(S_n^{(X)}, S_n^{(Y)}) = C_n,$

so joint convergence to $\mathcal{N}_2(0, I)$ requires $C_n \to 0$ . Sufficiency is proved via uniform vanishing of the cross-term in the characteristic function on compacts (Majumdar et al., 2017).

3. Connections to Multivariate CLT and Generalized Structures

Classically, the multivariate central limit theorem for i.i.d. random vectors gives a multivariate normal limit with fixed covariance, determined by the stationary cross-correlations $\rho$ : $(S_n^{(X)}, S_n^{(Y)}) \xrightarrow{d} \mathcal{N}_2(0, \Sigma), \;\; \Sigma = \begin{pmatrix} 1 & \rho \ \rho & 1 \end{pmatrix}.$ The result above generalizes this: with time-varying correlations, as in heterogeneous or dependent samples, asymptotic (average) decorrelation ( $C_n \to 0$ ) guarantees independent normal limits, while convergence of $C_n \to \rho$ gives a bivariate normal with correlation $\rho$ .

This framework is robust against various complications: the Cesàro mean criterion applies regardless of whether the cross-correlation pattern is deterministic, periodic, or exhibits alternating signs (e.g., $\rho_k = (-1)^k$ ). In such cases, if the average vanishes, asymptotic independence and joint normality are assured.

The Cesàro mean approach is conceptually and technically distinct from classical mixing or strong independence conditions. In the context of two-sample problems, neither cross-sample independence nor sample-size ratio conditions are necessary: Majumdar & Majumdar establish that Cesàro convergence of the cross-sample correlation coefficients to zero suffices, even if sample sizes grow at different rates and the dependence structure is nontrivial (Majumdar et al., 2016). The spherically symmetric limit law characterization further reinforces that only bivariate standard normality is possible under spherical symmetry, blocking the appearance of other symmetric stable laws within this framework (Majumdar et al., 2017).

Applications of the joint asymptotic normality criterion extend to:

Multivariate extensions: joint convergence of vectors of standardized means in higher dimensions can be reduced to block-structured versions of the Cesàro mean criterion (Majumdar et al., 2017, Hitczenko et al., 2023).
Combinatorial models: joint normality of fringe subtree counts in random trees, or multinomial occupancy counts, similarly hinges on verifying joint mean, covariance, and high-moment asymptotics within this framework (Janson, 2013, Hitczenko et al., 2023).
Random matrix theory: joint CLTs for largest eigenvalues and associated statistics in high-dimensional spiked models, where block-diagonal joint normality reflects asymptotic independence of spectral components when spike-related correlations are suitably separated (Wang et al., 2014, Li et al., 2019, Dharmawansa et al., 2014).

5. Illustrative Special Cases and Contrasts

Specific examples clarify the scope and sharpness of the result:

IID with constant correlation: if $\mathrm{Corr}(X_i,Y_i)=\rho$ for all $i$ , then $C_n \to \rho$ and the joint limiting distribution is the standard bivariate normal with correlation $\rho$ .
Alternating correlation: $\rho_k=1$ for odd $k$ , $\rho_k=-1$ for even $k$ yields $C_n \to 0$ , so that the standardized means become jointly asymptotically independent.
Degenerate or singular limits: The machinery detects cases where the covariance matrix of the limit would be singular (e.g., linear deterministic relationships among the statistics) and excludes joint normality in such pathological cases unless the high-moment/tightness conditions are adapted.

6. Broader Impact, Methodological Integration, and Open Directions

Joint asymptotic normality criteria featuring explicit Cesàro average (or high-moment) conditions provide powerful, verifiable, and often minimal requirements for Gaussian joint limits in a broad range of statistical, combinatorial, and random matrix contexts. These results enable unified analyses across settings—e.g., subsampling, functional data analysis, and high-dimensional regression—where direct independence or stationary assumptions are unattainable or undesirable. They further inform the design of inference procedures, the interpretation of limiting independence phenomena, and the development of robust resampling and bootstrap methods for joint distributions of complex statistics (Volgushev et al., 2013, Berkes et al., 2015).

Current research directions include generalizations to dependent sequences with weak dependence, higher-dimensional or Banach-space generalizations, and identification of universality regimes for joint Gaussianity in nonlinear or non-Euclidean settings.

Key Reference:

Majumdar, R. and Majumdar, S., "Necessary and Sufficient Condition for Asymptotic Normality of Standardized Sample Means" (Majumdar et al., 2017)