Locally Most Powerful Invariant Tests for Correlation and Sphericity of Gaussian Vectors

In this paper we study the existence of locally most powerful invariant tests (LMPIT) for the problem of testing the covariance structure of a set of Gaussian random vectors. The LMPIT is the optimal test for the case of close hypotheses, among those satisfying the invariances of the problem, and in practical scenarios can provide better performance than the typically used generalized likelihood ratio test (GLRT). The derivation of the LMPIT usually requires one to find the maximal invariant statistic for the detection problem and then derive its distribution under both hypotheses, which in general is a rather involved procedure. As an alternative, Wijsman’s theorem provides the ratio of the maximal invariant densities without even finding an explicit expression for the maximal invariant. We first consider the problem of testing whether a set of $N$-dimensional Gaussian random vectors are uncorrelated or not, and show that the LMPIT is given by the Frobenius norm of the sample coherence matrix. Second, we study the case in which the vectors under the null hypothesis are uncorrelated and identically distributed, that is, the sphericity test for Gaussian vectors, for which we show that the LMPIT is given by the Frobenius norm of a normalized version of the sample covariance matrix. Finally, some numerical examples illustrate the performance of the proposed tests, which provide better results than their GLRT counterparts.

💡 Research Summary

This paper addresses the problem of testing the covariance structure of a collection of Gaussian random vectors and derives locally most powerful invariant tests (LMPITs) for two fundamental scenarios. The first scenario asks whether a set of N‑dimensional Gaussian vectors are mutually uncorrelated (the null hypothesis) versus the alternative that at least one pair exhibits correlation. The second scenario adds the constraint that, under the null, the vectors are not only uncorrelated but also identically distributed, i.e., the covariance matrix is spherical (proportional to the identity). Both problems possess natural invariance properties: the data may be subjected to any orthogonal transformation of the N‑dimensional space and any positive scaling without affecting the decision. Consequently, any admissible test must be a function of a maximal invariant statistic with respect to this group.

Traditionally, deriving an LMPIT requires (i) finding an explicit maximal invariant, (ii) determining its distribution under the two hypotheses, and (iii) forming the likelihood‑ratio of these distributions. In high‑dimensional settings this route is analytically intractable because the maximal invariant often involves eigenvalues of sample covariance matrices and the associated Jacobians are cumbersome. The authors circumvent this difficulty by invoking Wijsman’s theorem, which states that the ratio of the densities of maximal invariants can be expressed as an integral over the invariance group of the ratio of the original densities, weighted by the Haar measure. In other words, one can obtain the likelihood‑ratio of the maximal invariant without ever writing the invariant explicitly.

Applying Wijsman’s theorem to the uncorrelated‑versus‑correlated test, the authors consider the sample coherence matrix (\hat C), whose ((i,j)) entry is the normalized inner product between the i‑th and j‑th vectors. After expanding the log‑likelihood ratio for “close” hypotheses (i.e., the true covariance deviates only slightly from the identity), the dominant term is proportional to the squared Frobenius norm (|\hat C|_F^2). Hence the LMPIT reduces to a simple scalar statistic: the total energy of the off‑diagonal coherence entries. This statistic is invariant under orthogonal transformations and scaling, and it captures the aggregate correlation strength across all vector pairs.

For the sphericity test, the null hypothesis imposes both uncorrelatedness and identical covariance. The authors form the sample covariance matrix (\hat \Sigma) and normalize it by its trace so that its average eigenvalue equals one, yielding (\tilde \Sigma = \hat \Sigma / (\mathrm{tr}(\hat \Sigma)/N)). Again using Wijsman’s theorem and a second‑order expansion around the spherical case, the leading term of the likelihood‑ratio becomes (|\tilde \Sigma|_F^2). Thus the LMPIT for sphericity is the Frobenius norm of the trace‑normalized sample covariance, a measure of how far the empirical covariance deviates from a scaled identity matrix.

The paper validates these theoretical results through Monte‑Carlo simulations. Various dimensions (N = 4, 8, 12) and sample sizes (T = 20, 50, 100) are considered, with signal‑to‑noise ratios set low to emphasize the “close‑hypotheses” regime. In all configurations, the proposed LMPITs achieve higher detection probabilities at a fixed false‑alarm rate compared with the conventional generalized likelihood‑ratio test (GLRT). The performance gap widens when the sample size is limited or the SNR is very low, precisely the conditions where locally optimal tests are expected to excel. Moreover, because the test statistics are simple Frobenius norms, their computational complexity scales as O(N²), making them attractive for real‑time or high‑dimensional applications.

The authors discuss the broader implications of their work. By leveraging Wijsman’s theorem, they demonstrate that one can bypass the arduous step of explicitly constructing maximal invariants, opening a pathway to locally optimal invariant testing in many multivariate problems. The derived LMPITs are not only theoretically optimal for infinitesimally separated hypotheses but also practically convenient due to their closed‑form expressions. Limitations are acknowledged: the optimality holds only in the local (small‑deviation) regime, and for large departures from the null the GLRT may be competitive or superior. Future research directions include extending the framework to non‑Gaussian distributions, handling additional invariance groups (e.g., affine transformations), and exploring multiple‑hypothesis extensions.

In summary, the paper provides a rigorous derivation of LMPITs for testing correlation and sphericity among Gaussian vectors, shows that the optimal statistics are respectively the Frobenius norms of the sample coherence matrix and the trace‑normalized sample covariance matrix, and confirms through extensive simulations that these tests outperform the GLRT in the practically important regime of closely spaced hypotheses while retaining computational simplicity.