Let $X \in \mathbb{R}^{n \times p}$ consist of iid $\mathcal{N}(0,1)$. Assume that $n/p$ converges to a positive constant. Denote by $\sigma_1 \ge \sigma_2 \ge \cdots \ge \sigma_{\min(n,p)} \ge 0$ the singular values. What's the joint distribution of $$\sigma_1 - \sigma_2, \sigma_2 - \sigma_3, \sigma_3 - \sigma_4, \ldots?$$


The joint distribution of the spacings $x_i=\sigma_i-\sigma_{i+1}$, $i=1,2,\ldots m-1$, $m=\min(n,p)$, follows from the joint distribution of the singular values $\sigma_i$ ($i=1,2,\ldots m$, ordered from large to small):

$$P(\sigma_1,\sigma_2,\ldots\sigma_m)\propto \prod_{i=1}^m e^{-\sigma_i/2}\sigma_i^{\alpha/2}\prod_{i<j}^m (\sigma_i-\sigma_j)$$

with the definition $\alpha=\max(n,p)-\min(n,p)-1$. To obtain the distribution of the spacings you substitute

$$\sigma_i=\sigma_m+\sum_{k=i}^{m-1}x_i,\;\;i=1,2,\ldots m-1$$

The determinant of this transformation is unity, so you immediately arrive at the joint distribution $P(x_1,x_2,x_{m-1},\sigma_m)$ of the $m-1$ spacings and the smallest eigenvalue $\sigma_m$. You may or may not want to integrate out the remaining $\sigma_m$. (I would think that setting $\sigma_m\mapsto 0$ should be quite accurate an approximation.)


Your Answer

By clicking "Post Your Answer", you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.