The Marchenko-Pastur law is a fundamental result in random matrix theory. It describes the distribution of eigenvalues of a sample covariance matrix when you have many data points in high dimensions. Classically, this law was derived assuming the data coordinates are statistically independent of one another. The paper asks whether the same law still holds when the coordinates are allowed to have certain dependencies.
The authors study a specific but rich class of random vectors where the coordinates, while not independent, satisfy two symmetry conditions: exchangeability (the distribution does not change if you permute the coordinates) and unconditionality (the distribution does not change if you flip the signs of any coordinates). They then build higher-dimensional data by taking tensor powers of such vectors, which is a structured way of combining a vector with itself repeatedly to create a much larger object encoding interactions among the original coordinates. This kind of construction appears naturally in machine learning and statistics when you want to capture higher-order feature interactions.
The main result is that even with these dependencies among coordinates, the Marchenko-Pastur law still governs the eigenvalue distribution of the resulting sample covariance matrix. This matters because it extends a previously known result that required full coordinate independence, which is rarely realistic. The authors show their framework applies to a wide variety of practically relevant random vectors, suggesting that the Marchenko-Pastur law is more robust to dependence structures than previously understood, as long as the right symmetry conditions hold.