In reading a paper I came across this expression which I don't quite understand: $$ \frac{\lambda_1}{2N}\operatorname{tr}\left((\mathbf{H}^M\mathbf{H}^M)^T\right) $$ For context, $\lambda_1$ and $N$ are scalars and $\mathbf{H}^M$ ($H$ henceforth) is a $MxN$ matrix.

The paper claims that this value is related to the variance of the column vectors which make up H, but either there is a typo or (quite likely) a linear algebra concept I don't know. The reason I am confused is that trace requires square input, and if H were to be made square then the transpose would have to apply to one of the two $H$s, not both after multiplying them (which itself doesn't make sense to me because only a square matrix can be multiplied by itself in the first place).

The paper is here, the mentioned expression is at the bottom of the third page in equation (5).

  • $\begingroup$ Could it be $\mathbf{H}^M (\mathbf{H}^M)^T$ instead? That would be square for any matrix $\mathbf{H}^M$. $\endgroup$ Jul 25 '18 at 18:36
  • $\begingroup$ That is what I had been thinking, but assuming that I haven't been able to figure out what the relationship to the variance of the column vectors would be. At that point it becomes a stats question more than a linalg one. If that were a typo then the trace results in the sum of the element-wise square of H, which to me seems more correlated to magnitude of values than variance. $\endgroup$ Jul 25 '18 at 18:40
  • $\begingroup$ Good point, added. $\endgroup$ Jul 25 '18 at 18:44

It looks like a typo. In the paper that you have linked, the equation(12) on page(4) rewrites

$$ arg.min .J = ... -\frac{\lambda_1}{2}\Bigl(tr\bigl(\frac{1}{N}H^M(H^M)^T\bigr)+\alpha tr(\Sigma_B-\Sigma_W)\Bigr) + ... $$

with the transpose correctly placed.


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.