# solving trace norm equality [closed]

## Problem Formulation

under what conditions can we solve $\mathrm{trace}(\mathbf{AB})=0$ ? or more specifically, when will $\mathrm{trace}(\mathbf{AB})=0$ implies that $\mathrm{trace}(\mathbf{B})=0$.

## Related Art

This occurs to me when $\mathbf{A}$ was a Positive semidefinite matrix, and $\mathbf{B}$ was a symmetric matrix with a parameter $\lambda$. I want to solve a partial derivatives equations which constrains that $\mathrm{trace}[\mathbf{AB}(\lambda)] = 0$, i.e., $$\frac{\partial \mathcal{L}(i)}{\partial \gamma_i} = \mathrm{Tr}\left[ (\mathbf{I}_{d_i} + \gamma_i\mathbf{B}_i\mathbf{s}_i)^{-1}\mathbf{B}_i\mathbf{s}_i\right] - \mathbf{q}_i^T(\gamma_i^{-1}\mathbf{B}_i^{-1} + \mathbf{s}_i)^{-1}\frac{\mathbf{B}_i^{-1}}{\gamma_i^2}(\gamma_i^{-1}\mathbf{B}_i^{-1} + \mathbf{s}_i)^{-1}\mathbf{q}_i,$$ where $\mathbf{s}_i$ and $\mathbf{B}_i$ are symmetric PSD matrix, $\mathbf{q}_i$ is a column vector, $\gamma_i$ is a scalar, $\mathbf{I}_d$ is an identity matrix with size $d$. I want to solve $$\frac{\partial \mathcal{L}(i)}{\partial \gamma_i} = 0.$$ with $\gamma_i$ derived in a closed-form.

Using cyclic product properties of the trace norm, I could only go this far, $$\frac{\partial \mathcal{L}(i)}{\partial \gamma_i} = \mathrm{Tr}\left[ (\mathbf{B}_i^{-1} + \gamma_i\mathbf{s}_i)^{-1}\mathbf{B}_i^{-1}(\mathbf{B}_i^{-1} + \gamma_i\mathbf{s}_i)^{-1}\left( \mathbf{s}_i + \gamma_i\mathbf{s}_i\mathbf{B}_i\mathbf{s}_i - \mathbf{q}_i\mathbf{q}_i^T \right) \right]$$ I sensed that the closed form solution of $\gamma_i$ might be related with $$\mathrm{trace}\left( \mathbf{s}_i + \gamma_i\mathbf{s}_i\mathbf{B}_i\mathbf{s}_i - \mathbf{q}_i\mathbf{q}_i^T \right) = 0.$$

## Counter Example

plus, I can also give a counter example. Let $$\mathbf{A} = \begin{pmatrix} 2 & 1 \\\ 1 & 2 \end{pmatrix}, \quad \mathbf{B} = \begin{pmatrix} 1+\lambda & -2 \\\ -2 & \lambda - 3 \end{pmatrix},$$ where $\mathrm{trace}(\mathbf{AB})=0$ implies $\lambda=2$, $\mathrm{trace}(\mathbf{B})=0$ implies $\lambda=1$.

-

## closed as off topic by Chris Godsil, Benoît Kloeckner, Yemon Choi, Denis Serre, Neil StricklandJun 5 '13 at 9:49

Questions on MathOverflow are expected to relate to research level mathematics within the scope defined by the community. Consider editing the question or leaving comments for improvement if you believe the question can be reworded to fit within the scope. Read more about reopening questions here.If this question can be reworded to fit the rules in the help center, please edit the question.

This is not the right site for this question. See the FAQ for alternatives. – Chris Godsil Jun 4 '13 at 8:56
Dear Chris, I have modified my problem. I think it related to the problem of linalg and the optimization of logdet function, this problem happened mostly in $\ell_1$ minimization solvers. L – liubenyuan Jun 4 '13 at 9:09
$\text{Tr}(AB)$ is a symmetric non-degenerate inner product on the space of all matrices, with signature $(\frac{(n+1)n}2,\frac{(n-1)n}2)$. Over $mathbb R$, the symmetric bilinear form $\text{Tr}(AB^\top)$ is positive definite. – Peter Michor Jun 4 '13 at 11:55
Dear Peter, you mean that $\mathrm{Tr}(\mathbf{AB^T})$ is PD and therefore could not equal $0$ ? – liubenyuan Jun 4 '13 at 12:52
Note that $A=0$, $B$ any matrix with nonzero trace, is a pretty good counterexample too. – 1015 Jun 4 '13 at 14:34

Theorem: Let $V$ be a finite-dimensional vector space over a field $K$ of characteristic zero. Let $E\subseteq F$ be two linear subspaces of $End(V)$ and $M=\lbrace A \in End(V) \mid [A,F] \subseteq E\rbrace$. Assume that $tr (AB)=0$ for all $A\in M$. Then $B$ is nilpotent with $tr(B)=tr(B^2)=\cdots = tr(B^n)=0$ for all $n$.
Edit: $[A,B]=AB-BA$ in $End(V)$.
Sorry, dear Dietrich, I could not quite follow your answer. and what does End(V) mean ? Which book should I refer to ? However thanks for your answer, it seem like I could obtain $\mathrm{Tr}(B)=0$ with some additional constrains imposed on $\mathbf{A}$. – liubenyuan Jun 4 '13 at 12:43
$End(V)$ is the space of all linear maps $A\colon V\rightarrow V$ (square matrices, if you want). – Dietrich Burde Jun 4 '13 at 13:04
@Dietrich: Do you really mean $AF-FA$, not $\left[A,F\right]$ (which are different things, e. g., because $AF-FA$ is the same as $AF+FA$) ? – darij grinberg Jun 4 '13 at 17:42
(If it is $\left[A,F\right]$ that Dietrich wanted to say, the result appears in C. Brookes' "Lectures on Lie Algebras", scribed by Eva Belmont people.ds.cam.ac.uk/eb525/lie-notes.pdf . It is Lemma 9.4 (3.25). The former number apparently can change while the latter is invariant. $x_s$ and $x_n$ are defined in Lemma 9.1 (3.22).) – darij grinberg Jun 4 '13 at 17:47