# Get rid of tr() in SVM kernel trick

I designed a kernel function (to be used within SVM) which has the expression $tr(AB)$ in it. For efficient implementation of this, I was wondering if I could write $tr(AB)$ as an inner product: $\phi(A)^T \phi(B)$? What is the function $\phi()$?

-
I'm not sure I understand your question (it would help if you eliminated or explained your jargon), but the obvious optimization is that you don't explicitly need to compute AB: tr(AB) is the sum of A_{ij}*B_{ji} over all pairs i,j. You can do this with a single pass over your matrices. –  Darsh Ranjan Jan 14 '10 at 9:25
@darsh you beat me by 3 minutes! I can explain some of the terminology. Imagine you have data in $\mathbb{R}^d$ (here it seems to be in $\mathbb{R}^{d\times d}$??) but want to work with it in $\mathbb{R}^D$, which you can accomplish by passing everything through a mapping $\Phi$. If $D \gg d$, this can be expensive. But in some cases your algorithm need only compute inner products $\langle \Phi(a), \Phi(b)\rangle$, in which case there may be a functin $k(\cdot,\cdot)$ which computes the same thing but with much less work than the explicit mapping. this is called the 'kernel trick'. –  Matus Telgarsky Jan 14 '10 at 9:34
Note that you don't really need the explicit mapping $\phi$. What you want is a kernel $k(A, B)$ which has the property that $k(A, B) = \phi(A)^T \phi(B)$. Generally, you compute the kernel directly, instead of calculating the mapping and computing the dot product -- that's why it's called kernel `trick'. In your case, it seems you want the kernel to be $k(A, B) = tr(AB)$. So it seems you want to know how to compute the trace efficiently, rather than what the mapping is. –  user3035 Jan 14 '10 at 10:02
The non-efficiency-related part of this question is basic linear algebra: tr(AB) is an inner product on the space of nxn matrices and so you can find isometric isomorphisms from M_{nxn} to R^{n^2}. Any of these will do for the map phi. I do not know whether or not any of these will improve the efficiency in calculating the trace, but I doubt it. –  Loop Space Jan 14 '10 at 10:20
Sorry for being ambiguous. I know that I can implement $k(A,B)$ directly and avoid needing $\phi()$. But it would be nice if I have $\phi()$ because then, I can use a regular linear SVM on my $\phi()$ mapped vectors. This is much simpler and faster compared to implementing $k(A,B)$ and plugging it into SVM. –  andinos Jan 14 '10 at 21:39

Matus is right. But if the matrices $A$, and $B$ have certain properties like being symmetric, or diagonal, then simply just vectorizing the matrices and taking their inner product would be equal to the $tr(AB)$.

-

If $A,B$ are arbitrary $n\times n$ matrices, by definition of trace, $\textrm{tr}(AB) = \sum_{i,j} A_{ij}B_{ji}$. This is $O(n^2)$, but just reading the entries of $A$ is $\Omega(n^2)$. Without any special structure on $A,B$, you probably can't do better.

If $A,B$ are (column) vectors, you probably mean the outer product $\textrm{tr}(AB^T) = \sum_i A_i B_i$.

Edit: andinos clarified to say he wants to know about the implicit mapping of the kernel function. Well I have bad news: It does not exist!! The proof works by showing there exist matrices $A,B$ such that the corresponding kernel matrix is not positive semi-definite. To finish, apply Mercer's theorem.

In particular, set $A = \left(\begin{array}{cc}1 & 1 \\\\ -1 & 1\end{array}\right)$ and $B = A^T = \left(\begin{array}{cc}1 & -1 \\\\ 1 & 1\end{array}\right)$. Therefore $\textrm{tr}(AB) = \textrm{tr}(AA^T) = 4$, and $\textrm{tr}(BA)$ is identical. On the other hand, $\textrm{tr}(AA) = \textrm{tr}(BB) = 0$. therefore, the kernel matrix $K$ is $\left(\begin{array}{cc}0 & 4 \\\\ 4 & 0\end{array}\right)$. Set $x = \left(\begin{array}{c} 1 \\\\ -1\end{array}\right)$, and observe that $x^T K x = -8 < 0$, and therefore $K$ is not PSD, so the kernel $k(A,B) = \textrm{tr}(AB)$ is not PSD.

On the other hand! If you had instead defined your kernel to be $k'(A,B) = \textrm{tr}(AB^T)$, notice that $k'(A,B) = \sum_{i,j}A_{ij}B_{ij} = \Phi(A)^T\Phi(B)$ where $\Phi$ simply takes its input matrix and outputs it as a column vector.

-