Matrix trace derivatives

  • Thread starter em12
  • Start date
  • #1
em12
3
0
Hope this is the right section. I'm having trouble ironing out an apparent inconsistency in matrix trace derivative rules.

Two particular rules for matrix trace derivatives are

[tex]\frac{\partial}{\partial\mathbf{X}} Tr(\mathbf{X}^2\mathbf{A})=(\mathbf{X} \mathbf{A}+\mathbf{A} \mathbf{X})^T[/tex]

and

[tex] \frac{\partial}{\partial\mathbf{X}} Tr(\mathbf{X}\mathbf{A}\mathbf{X}^T)=\mathbf{X} \mathbf{A}^T+\mathbf{X}\mathbf{A}[/tex]

Now assume that [tex]\mathbf{A}[/tex] is diagonal (or maybe even just symmetric) and [tex] \mathbf{X}[/tex] is anti-symmetric. Then by the cyclic property of the trace, [tex]-Tr(\mathbf{X}^2\mathbf{A})=Tr(\mathbf{X}\mathbf{A}\mathbf{X}^T)[/tex]. So the two derivatives should be equal up to a minus sign, no?

However, the first rule returns the derivative

[tex]- (\mathbf{X}\mathbf{A}+\mathbf{A}\mathbf{X})[/tex]

and the second returns

[tex] 2\mathbf{X}\mathbf{A}[/tex].


Am I missing something?
 

Answers and Replies

  • #2
Stephen Tashi
Science Advisor
7,781
1,540
I don't know the answer to your question, but it prompted me to look at the "Matrix calculus" article in the Wikipedia. If you look at the "discussion" page for that article, you see some interesting comments that say (to me) that the notation for taking the derivative with respect to a matrix is not completely standardized. If you explain the system of notation that you are using, perhaps someone will answer your question.
 

Suggested for: Matrix trace derivatives

Replies
15
Views
695
  • Last Post
Replies
32
Views
2K
Replies
2
Views
478
  • Last Post
Replies
2
Views
682
  • Last Post
Replies
5
Views
912
  • Last Post
Replies
2
Views
828
  • Last Post
Replies
9
Views
838
Replies
10
Views
657
  • Last Post
Replies
17
Views
984
  • Last Post
Replies
2
Views
717
Top