Abstract
In many computational and communication contexts, products of independent Gaussian matrices are used as surrogates for single, larger Gaussian matrices. While the spectral properties of such products are well-documented, their global distributional characteristics have remained less clear. This paper provides the first rigorous bounds on the Total Variation (TV) distance between a product of $r$ independent Gaussian matrices and a single Gaussian matrix. We show that the product is nearly indistinguishable from a Gaussian when the intermediate dimensions are sufficiently large, establishing a sharp transition for distinguishability.
Introduction
Random matrices, particularly those with Gaussian entries, are fundamental tools in wireless communications (e.g., MIMO channel modeling) and randomized numerical linear algebra (e.g., sketching algorithms). In practice, researchers often encounter or intentionally construct products of independent Gaussian matrices, denoted as $A_r = G_r G_{r-1} \dots G_1$.
A critical question for the reliability of these applications is whether the resulting product $A_r$ behaves like a single Gaussian matrix $G$. If the distribution of the product deviates significantly from Gaussianity, the statistical guarantees of algorithms—such as those used in "sketch-and-solve" paradigms—may no longer hold. This paper addresses this gap by quantifying the "Gaussianity" of these products through the lens of Total Variation distance.
The Bottleneck Effect in Matrix Chains
The core finding of this research is that the intermediate dimensions ($d_i$) of the matrix chain act as a bottleneck. If any intermediate dimension is too small, it "warps" the resulting distribution, making it statistically distinguishable from a true Gaussian distribution. Conversely, as these dimensions grow relative to the input and output dimensions ($p$ and $q$), the product distribution converges to a Gaussian.
Methodology
The authors employ a dual-sided approach to establish the limits of Gaussianity, providing both upper and lower bounds on the TV distance.
Upper Bound: Proving Closeness
To prove that the product is close to Gaussian, the researchers utilize the rotational invariance of Gaussian measures. The proof strategy involves a recursive application of the Singular Value Decomposition (SVD). By decomposing a Gaussian matrix into a diagonal scaling matrix and Haar-distributed orthogonal matrices, the authors leverage recent advancements in bounding the distance between submatrices of orthogonal matrices and Gaussian matrices (specifically results from Jiang and Ma). This allows them to show that each step in the matrix product chain maintains a high degree of Gaussianity if the dimensions are large enough.
Lower Bound: Establishing Distinguishability
To determine when the distributions diverge, the authors identify a specific test statistic: the fourth power of the Schatten 4-norm, defined as $\text{tr}((X^T X)^2)$. Using Chebyshev’s inequality and Poincaré’s inequality for Gaussian measures, they demonstrate that if the intermediate dimensions are below a certain threshold, this statistic behaves differently enough to allow an observer to distinguish the product from a single Gaussian with high probability.
Key Results
The primary contribution of the paper is summarized in Theorem 1, which provides a quantitative bound on the Total Variation distance $d_{TV}(A_r, A_1)$:
-
Closeness Guarantee: The TV distance is bounded by: $$d_{TV}(A_r, A_1) \le C_1 \sum_{i=1}^{r-1} \sqrt{\frac{pq}{d_i}}$$ where $p$ and $q$ are the input/output dimensions and $d_i$ are the intermediate dimensions. This implies that as $d_i \gg pq$, the product becomes indistinguishable from a single Gaussian.
-
The Sharp Transition: For a constant number of matrices $r$, the paper establishes a near-dichotomy. When $p \approx q$, the product is close to Gaussian if $d_i \gg p^2$, but it becomes distinguishable if $d_i \ll p^2$. Specifically, if $d_i < \max\{p, q\}^{1/2} \min\{p, q\}^{3/2}$, the distributions are at least $2/3$ apart in TV distance.
Discussion and Future Work
These results provide a vital "rule of thumb" for engineers: when replacing a large Gaussian matrix with a product of smaller ones to save on memory or computation, the intermediate dimensions must be significantly larger than the product of the external dimensions to preserve statistical properties.
However, several avenues for future research remain:
- Deep Chains: The current lower bounds are primarily derived for a constant $r$. The behavior of very deep chains, where $r$ scales with the dimensions, is not yet fully characterized.
- Alternative Statistics: While the Schatten 4-norm is an effective distinguisher, other statistics might yield tighter bounds for specific non-square cases.
- Hardware Implementation: Future work could bridge the gap between this statistical theory and the actual computational gains seen on specific hardware architectures when using matrix products for sketching.
Summary updated