10

Let $A$ and $B$ be two arbitrary matrix with proper dimension for multiplication.

Consider this trace inequlaty which is trace of multiplication of two matrices versus their individual traces

$$\text{tr}(AB) \leq \text{tr(A)} \text{tr(B)}$$

1- Do we have result for rectangular matrix that satisfy this inequality?

2- If they were square matrices what are the conditions?

3- Is there any specific name for this inequality?

  • 2
    How do you define the trace of a matrix which is not a square? In case you take $A=B$ this equivalent to $$\sum_{i=1}^n \lambda_i^2 \leq \left(\sum_{i=1}^n \lambda_i \right)^2$$ – Severin Schraven Jan 04 '19 at 17:39
  • If we have a $m\times n$-matrix, and $k:=\min(m,n)$, the definition $$tr(A)=\sum_{j=1}^k a_{jj}$$ would make sense. – Peter Jan 04 '19 at 17:43
  • @Peter You are right, I was thinking about something which is coordinate-free. – Severin Schraven Jan 04 '19 at 17:45
  • @Severin Schraven: peter answered that. But to get into the problem, let they be square first and focus on 2 and 3. –  Jan 04 '19 at 17:46

1 Answers1

7

Can't think of anything deep, but if both $A$ and $B$ are positive semidefinite, the inequality is true: when $a=\operatorname{tr}(A)$, we have $A\preceq aI$ and hence $$ \operatorname{tr}(AB) =\operatorname{tr}(B^{1/2}AB^{1/2}) \le\operatorname{tr}(B^{1/2}(aI)B^{1/2}) =\operatorname{tr}(A)\operatorname{tr}(B). $$ This also follows from (and hence is weaker than) von Neumann's trace inequality, which in this context says that $$ \operatorname{tr}(AB)\le\sum_i\lambda_i(A)\lambda_i(B) $$ when the eigenvalues of $A$ and $B$ are arranged in the same (ascending or descending) order.

user1551
  • 139,064
  • is there any name for this inequality? –  Jan 04 '19 at 17:51
  • Can you explain what $A^{1/2}$ is? When they are psd we have $A=U_A\Lambda_AU_A^T$ and $B=U_B\Lambda_BU_B^T$. So $\text{tr}(U_A\Lambda_AU_A^TU_B\Lambda_BU_B^T)=\text{tr}(U_B^TU_A\Lambda_AU_A^TU_B\Lambda_B)$, how can I get what you have? –  Jan 04 '19 at 17:54
  • What's that $F$ norm? – Yanko Jan 04 '19 at 17:57
  • @Yanko $|\cdot|_F$ denotes Frobenius norm. – user1551 Jan 04 '19 at 18:00
  • @Saeed $A^{1/2}$ is the (unique) positive semidefinite square root of $A$. If you orthogonally diagonalise $A$ as $QDQ^T$, then $A^{1/2}=QD^{1/2}Q^T$ where $D^{1/2}$ is the entrywise square root of $D$ (i.e. $D^{1/2}$ is the diagonal matrix whose diagonal entries are the square roots of their correspondents in $D$). – user1551 Jan 04 '19 at 18:03
  • @ user1551: Thnak you so much for your quick and easy way. I was thinking of $\text{tr}(U_B^TU_A\Lambda_A^{1/2}\Lambda_A^{1/2}U_A^TU_B\Lambda_B^{1/2}\Lambda_B^{1/2})=\text{tr}(\Lambda_B^{1/2}U_B^TU_A\Lambda_A^{1/2}\Lambda_A^{1/2}U_A^TU_B\Lambda_B^{1/2})$ which has the same result but yours is more beautiful. –  Jan 04 '19 at 18:07
  • Could you explain what you mean by $A\preceq aI$ and how this is used to generate the trace inequality? – user1936752 Jun 22 '20 at 13:49
  • 1
    @user1936752 It's the positive semidefinite partial ordering. – user1551 Jun 22 '20 at 18:35