5
$\begingroup$

Let $A\in\mathbb{R}^{r\times m}$ be a matrix of full row rank, and let $\cdot^+$ denote the Moore-Penrose inverse.

Consider a sequence of matrices $\{B_n\}_{n>1}$, $B_n\in\mathbb{R}^{m\times n}$, whose entries are i.i.d. random Gaussian variables with zero mean and finite variance. Numerical simulations suggest that $$ \|(AB_n)^+ - B_n^+ A^+\| \to 0\ \ \text{ as }\ \ n\to\infty $$ where $\|\cdot\|$ denotes the 2-norm of a matrix.

Is it possible to provide a formal proof of this claim?

My question could be either very silly (if so, please close it) or a well-known fact (if so, I will be glad if you can provide pointers to the literature). In any case, the observed numerical behavior looks quite surprising to me, since it is rather well-known that for $A$ and $B$ both of full row rank it holds $(AB)^+ \ne B^+ A^+$ (except for some very special cases).

$\endgroup$
  • 4
    $\begingroup$ I guess it's because $B$ has 'almost orthogonal' rows (by LLN/CLT-type properties), and $(AB)^+=B^+A^+$ is a property that holds true when $B$ has orthogonal rows. Probably someone more versed than me with random matrices knows how to make this more formal. $\endgroup$ – Federico Poloni Jan 2 at 13:10
  • $\begingroup$ I do not know much about random matrix theory, but is $B_n$ not of full rank with probability 1 independent of $n$? $\endgroup$ – student Jan 3 at 15:25
  • $\begingroup$ @N.T.: Yes, I would say that $B_n$ is of full (row or column, depending on $m$) rank with probability 1 for every $n$. Why? $\endgroup$ – Ludwig Jan 3 at 15:46
  • $\begingroup$ @Ludwig Ah, there is my mistake. $A$ has full row rank. Apologies! $\endgroup$ – student Jan 3 at 16:07
  • $\begingroup$ @FedericoPoloni: Thanks for your comment! I was wondering whether the rows of $B$ need to be orthonormal for $(AB)^+=B^+ A^+$ to hold true. $\endgroup$ – Ludwig Jan 3 at 16:53
2
$\begingroup$

As @FedericoPoloni pointed out, this must hinge on the fact that the rows of $B_n$ tend to be orthogonal as $n$ increases. In fact, $$\mathrm{E}[(B_n B_n^*)_{ij}] = n \sigma^2 \delta_{ij} \\ \mathrm{Cov}[(B_n B_n^*)_{ij}, (B_n B_n^*)_{i'j'}] = n \sigma^4 \, (\delta_{ii'} \delta_{jj'} + \delta_{ij'} \delta_{i'j})$$ so that we might as well write $$B_n B_n^* = n \sigma^2 I + \sqrt{n} \sigma^2 R$$ where $R$ is a random matrix with entries of zero mean and $O(1)$ variance. The near-orthogonality comes into play as $$\lim_{n \to \infty} \frac{1}{n} B_n B_n^* = \sigma^2 I$$ which should feature somehow in showing the suggestion.

To introduce this into the M-P inverse, the only thing that comes to mind right now is to take the limit definition $$A^+ = \lim_{\delta \searrow 0} A^* (A A^* + \delta I)^{-1}$$ to define a sequence $$(A B_n)^+_k = B_n^* A^* (A B_n B_n^* A^* + \tfrac{1}{k} I)^{-1}$$ that converges to $(A B_n)^+$. We also have $B_n^+ = B_n^* (B_n B_n^*)^{-1}$ a.s. for $n \ge m$ and $A^+ = A^* (A A^*)^{-1}$. Hence for a fixed $k$ we can evaluate $$\Vert (A B_n)^+_k - B_n^+ A^+ \Vert = \frac{1}{n} \left\Vert B_n^* A^* \left(A \, (\tfrac{1}{n} B_n B_n^*) \, A^* + \tfrac{1}{n k} I\right)^{-1} - B_n^* (\tfrac{1}{n} B_n B_n^*)^{-1} A^* (A A^*)^{-1} \right\Vert$$ which looks harmless enough to be pushed below any desired bound with a suitably large $n$.

If that's the case, pick $\epsilon > 0$ and use the triangle inequality $$\Vert (A B_n)^+ - B_n^+ A^+ \Vert \le \Vert (A B_n)^+ - (A B_n)^+_k \Vert + \Vert (A B_n)^+_k - B_n^+ A^+ \Vert$$ to look at each term on the rhs. individually: first find $k$ s.t. the first term is less than $\epsilon/2$, and then find $n$ s.t. the second term is less than $\epsilon/2$.


Moments of the entries of $BB^*$

For iid. normal entries $B_{ij}$ with zero mean and variance $\sigma^2$: $$\mathrm{E}[B_{ij} B_{kl}] = \sigma^2 \delta_{ik} \delta_{jl}$$ Hence the expectation: $$\mathrm{E}[(B B^*)_{ij}] = \sum_k \mathrm{E}[B_{ik} B_{jk}] = \sum_k \sigma^2 \delta_{ij} \delta_{kk} = n \sigma^2 \delta_{ij}$$ For jointly normal $X_1, X_2, X_3, X_4$ with zero mean: $$\mathrm{E}[X_1 X_2 X_3 X_4] = \mathrm{E}[X_1 X_2] \mathrm{E}[X_3 X_4] + \mathrm{E}[X_1 X_3] \mathrm{E}[X_2 X_4] + \mathrm{E}[X_1 X_4] \mathrm{E}[X_2 X_3]$$ Hence the covariance: $$ \mathrm{Cov}[(B B^*)_{ij}, (B B^*)_{i'j'}] = \sum_{k,k'} \mathrm{Cov}[B_{ik} B_{jk}, B_{i'k'} B_{j'k'}] \\ = \sum_{k,k'} \left\{\mathrm{E}[B_{ik} B_{jk} B_{i'k'} B_{j'k'}] - \mathrm{E}[B_{ik} B_{jk}] \mathrm{E}[B_{i'k'} B_{j'k'}]\right\} \\ = \sum_{k,k'} \left\{\mathrm{E}[B_{ik} B_{i'k'}] \mathrm{E}[B_{jk} B_{j'k'}] + \mathrm{E}[B_{ik} B_{j'k'}] \mathrm{E}[B_{jk} B_{i'k'}]\right\} \\ = \sum_{k,k'} \left\{\sigma^4 \delta_{ii'} \delta_{kk'} \delta_{jj'} \delta_{kk'} + \sigma^4 \delta_{ij'} \delta_{kk'} \delta_{i'j} \delta_{kk'}\right\} \\ = n \sigma^4 (\delta_{ii'} \delta_{jj'} + \delta_{ij'} \delta_{i'j})$$

$\endgroup$
  • $\begingroup$ Thanks for your answer. Could you please elaborate a little more (or provide a reference) on the derivation of the first two displayed equations? $\endgroup$ – Ludwig Jan 11 at 9:26
  • $\begingroup$ You get the expectation and covariance from the fact the entries of $B_n$ are independent and normal. I have added the steps to the answer, below the line. $\endgroup$ – student Jan 11 at 10:07
  • $\begingroup$ Oh I see, sorry for the silly question. One last comment: which type of convergence are you considering when you write $\lim_{n\to \infty} \frac{1}{n} B_n B_n^*=\sigma^2 I$? $\endgroup$ – Ludwig Jan 11 at 10:15
  • $\begingroup$ Note that this limit is just a heuristic and not used; the important step is to find a suitable $n$ later on, which you would have to make rigorous. Nevertheless, if you want to know the limit, I would expect (at least) convergence in probability relatively straightforwardly from Markov's inequality for $\Pr[\frac{\sigma^2}{\sqrt{n}} \Vert R \Vert > \epsilon]$ plus some variance-type inequality for $\mathrm{E}[\Vert R \Vert^k]$ in the assumed norm, since $R$ is of fixed size $m \times m$ and does not grow with $n$. $\endgroup$ – student Jan 11 at 12:30
  • $\begingroup$ I see, thank you for clarifying! So the missing step is to show that $\|(AB_n)^+-B_n^+A^+\|$ tends to zero (in probability?) as $n$ goes to infinity, and this should follow from the fact that $B_n$ is almost orthogonal as $n$ tends to infinity. Am I correct? $\endgroup$ – Ludwig Jan 15 at 16:44

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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