Random projections beyond zero overlap

Timothy L.H. Wee,Sekhar Tatikonda

2023-12-03

Abstract:A random vector whose norm and overlap (inner product with an independent copy) concentrates is shown to have random low-dimensional projections that are approximately random Gaussians. Conversely, asymptotically random Gaussian projections imply these hypotheses. This extends and unites several existing results in geometric functional analysis and spin glasses. Applications include a large-system characterization of the joint law of cavity fields in the Sherrington-Kirkpatrick model.

Probability

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is about the projection distribution characteristics of high - dimensional vectors in low - dimensional space in the case of non - zero overlap (inner product with an independent copy) in random projection. Specifically, the article shows that if the norm and overlap of a random vector are concentrated, its low - dimensional random projection is approximately a random Gaussian distribution. Conversely, if the asymptotically random Gaussian projection holds, these assumptions also hold. This extends and unifies some existing results in the fields of geometric functional analysis and spin glass.

### Main problem summary:

1. **Properties of random projection**: Research on the distribution characteristics of the low - dimensional random projection of a random vector when its norm and overlap are concentrated.
2. **Impact of non - zero overlap**: Explore the impact of non - zero overlap on the results of random projection, and prove that even in the case of non - zero overlap, the low - dimensional projection can still be approximated as a Gaussian distribution.
3. **Proof of the converse proposition**: Not only prove the projection distribution characteristics under the above conditions, but also prove that if the projection distribution is Gaussian, the norm and overlap of the original vector must also be concentrated.

### Formula representation:

Let \(x\) be a random vector, and \(\Theta = (\theta_1,\ldots,\theta_k)\) be an \(N\times k\) projection matrix, where \(\theta_i\) are independent and identically distributed standard Gaussian vectors. For two independent copies \(x_1\) and \(x_2\), assume that there are constants \(0\leq q <\rho\) such that:

\[
\frac{1}{N}\|x\|^2\simeq\rho,\quad\frac{1}{N}x_1\cdot x_2\simeq q
\]

Then:

\[
L(\Theta^{\top}x|\Theta)\simeq L(\sqrt{q}z + \sqrt{\rho - q}\xi|z)
\]

where \(z\) and \(\xi\) are independent standard Gaussian random vectors.

### Application examples:

- **Cavity field distribution in spin glass models**: This result can be used to describe the joint distribution of cavity fields in the Sherrington - Kirkpatrick model, which is of great significance for understanding the statistical properties of complex systems.
- **Central limit theorem in geometric functional analysis**: This result can be applied to the projection central limit theorem and large deviation principle of high - dimensional bodies, further expanding the theoretical framework of geometric functional analysis.

Through these studies, the paper not only solves the problem of distribution characteristics in the case of non - zero overlap in random projection, but also provides a theoretical basis for further research in related fields.