MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Let A and B be positive semidefinite matrices. It is not hard to see that $(A-B)^2 \leq 2A^2 + 2B^2$. In fact, $2A^2 + 2B^2 - (A-B)^2 = (A+B)^2$ is positive semidefinite.

My question is: Is there a constant c (independent of A and B and the dimension) such that

$$(A-B)^2 \leq c (A+B)^2?$$


share|cite|improve this question
[edited only to TeXify the title] – Noam D. Elkies Jan 10 '12 at 16:14
up vote 29 down vote accepted

There is no such $c$ even if we use only $2 \times 2$ matrices. For any $c \geq 1$ let $A,B$ be the positive-semidefinite matrices $$ A = \left( \begin{array}{lc} c^2 & c \cr c & 1 \end{array} \right), \phantom\infty B = \left( \begin{array}{cc} 1 & 0 \cr 0 & 0 \end{array} \right). $$ of rank $1$. Then we calculate that the difference $$ D := c(A+B)^2 - (A-B)^2 $$ has determinant $(c-1)^2 - 4c^3 < 0$, and is thus not positive semidefinite.

In fact this counterexample works for all $c \in \bf R$: looking around $\ker A = {\rm span} \lbrace(-1,c)\rbrace$ we find the negative vector $v = (-c, c^2+1)$. To verify that $\langle v, Dv \rangle < 0$, recall that for any vector $x$ and any symmetric matrix $M$ of the same order we have $$ \langle x, M^2 x \rangle = \langle Mx, Mx \rangle = |Mx|^2. $$ Here we compute $v(A+B) = (0,1)$ and $v(A-B) = (2c,1)$, so $$ \langle v, Dv \rangle = c \phantom. |(0,1)|^2 - |(2c,1)|^2 = -4c^2 + c - 1, $$ which is negative for all real $c$. If $c$ is small enough that $\det(D) > 0$ then $D$ is negative definite.

share|cite|improve this answer
Many thanks for your answer! – Omar Jan 10 '12 at 17:51
You're welcome! I thought the minimal $c$ might grow with the dimension $n$, and tried some pairs of rank-1 matrices for $n=2$ (because they form the boundary of the positive-semidefinite cone), and when $A = ({1 \phantom+ 1 \atop 1 \phantom+ 1})$ and $B = ({1 \phantom+ 0 \atop 0 \phantom+ 0})$ produced a surprisingly large $c$ of almost $6$ I tried varying $A$ and soon found that already for $n=2$ the minimal $c$ can get arbitrarily large. – Noam D. Elkies Jan 10 '12 at 20:31

A weaker version is also not true, that is, there is no constant $c$ such that $$\lambda_j(A-B)^2 \leq c \lambda_j(A+B)^2, ~~ j=1, 2,\ldots$$ where $\lambda_j(A)$ means the $j$-th largest eigenvalue of $A$.

Nevertheless, it is trivial to see that
$$trace(A-B)^2 \leq trace(A+B)^2.$$

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.