# Finding SVD efficiently for $AB^T$

I posted this on cs theory yesterday but did not get an answer and hence I am posting here.

I have a low rank matrix given as $AB^T$ where $A,B \in \mathbb{R}^{n \times p}$ and $p \ll n$. (I know $A$ and $B$ separately)

One efficient way to get the Singular Value Decomposition is to do a reduced QR on $A$ and $B$ i.e.

$A = Q_A R_A$ and $B = Q_B R_B$.

The above can be done in $\mathcal{O}(p^2n)$ cost.

I could then compute the svd of $R_A R_B^T = U_1 \Sigma V_1^T$ which is $\mathcal{O}(p^3)$.

Hence this will give me $$AB^T = (Q_A U_1) \Sigma (Q_B V_1)^T$$

The total cost is $\mathcal{O}(np^2)$.

However, I am wondering if there are other efficient ways to go about doing this to reduce the coefficient infront of $p^2n$. If I am right, the coefficient infront of $np^2$ is $3$ if we go about doing QR. The reason why I am interested in minimizing the coefficient is that my $n$ is really ginormous. So if I were to implement it, a cost cutting on the coefficient of the $\mathcal{O}(n)$ could be significant.

I am wondering if we could exploit the fact that the left singular vectors must span $A$ and the right singular vectors span $B$

i.e. we know that if $AB^T = U \Sigma V^T$, then

$$A = U \alpha \text{ and } B = V \beta$$

and we want $\alpha \beta^T$ to be diagonal and $U$,$V$ needs to be unitary.

EDIT What I essentially want is an exact rank $r<p$ approximation to the product $AB^T$ in the form $U_r V_r^T$. I am not looking for approximate or probabilistic algorithms.

-
Any sparsity structure to your matrix? Have you looked at Fast SVD? That can be handy, in particular, if $r \ll p$. – cardinal Mar 18 '11 at 3:34
@cardinal: There is no structure in $A$ and $B$. And what Fast SVD are you talking about? could you provide a link? I am not looking for approximate or randomized algorithms... – user17762 Mar 18 '11 at 3:37
I recall a MATLAB package for doing fast exact truncated SVDs. Don't know if this or this is helpful at all. – cardinal Mar 18 '11 at 3:41
@cardinal: I don't think it would help my problem since my problem has the matrix in the low rank form $AB^T$ where $A,B$ are thin matrices. – user17762 Mar 18 '11 at 4:11
Have you seen this? – J. M. Jul 21 '11 at 17:54

I presume you're already aware that just the mere fact of multiplying $\mathbf A$ and $\mathbf B^T$ already exposes you to numerical instability, similar to the reasons why one is better off computing the singular value decomposition of a matrix $\mathbf A$ instead of the eigendecomposition of $\mathbf A^T \mathbf A$.
The paper doesn't answer the question. Even in the above procedure I have, I never multiply out $A$ and $B^{T}$. The paper has the essentially same algorithm but with taking into account the numerical stability. – user17762 Mar 18 '11 at 7:16