I have to show that $\sup_{n \in \mathbb{N}}|Tr(A_nB_n)| < \infty$. I know that $\sup_{n \in \mathbb{N}}||A_n||, \sup_{n \in \mathbb{N}}||B_n|| < \infty$, where $A_n,B_n$ are $2 \times 2$ random matrices and $ ||X|| = Tr(XX^T)^{1/2}$. Is there any inequality which I can use?

  • $\begingroup$ I'm not sure of the exact matrix norm you're using. But can you use the definition of trace directly? I.e. $tr(AB) = \sum_{i=1}^n(AB)_{ii} = \sum_{i=1}^n \mathbf{A}_{row\,i}^\rm{T} \; \mathbf{B}_{col\,i} = \sum_{i=1}^n \sum_j (A)_{ij} (B)_{ji}$ $\endgroup$ – TooTone Jul 19 '13 at 17:58
  • 2
    $\begingroup$ Then they have finite entries and obviously the trace of their product is finite! You can obtain a much stronger result, though, using the Cauchy-Schwarz Inequality. (Whether the entries are random or not is immaterial.) $\endgroup$ – whuber Jul 19 '13 at 18:36
  • 1
    $\begingroup$ @TooTone: en.wikipedia.org/wiki/Supremum. Kolibris: the inequality applies directly by viewing the matrices as $4$-dimensional vectors. There's little more to be said; the rest is an application of the definition of supremum. $\endgroup$ – whuber Jul 19 '13 at 18:51
  • 1
    $\begingroup$ @TooTone Another approach which you might like is tantamount to the Holder Inequality: from the axiomatic relation $||A_n-B_n||^2\ge 0$ it is straightforward to deduce $2Tr(A_nB_n')\le||A_n||^2+||B_n'||^2$, whence upper bounds $\alpha$ for $\{||A_n||\}$ and $\beta$ for $\{||B_n'||\}$ give an upper bound of $(\alpha^2+\beta^2)/2$ for $\{Tr(A_nB_n)\}$, etc. $\endgroup$ – whuber Jul 19 '13 at 19:25
  • 1
    $\begingroup$ @TooTone The supremum is applied here to sequences of numbers, not matrices: the numbers are $Tr(A_nB_n)$, $||A_n||$, and $||B_n||.$ It might help to interpret the question in words as "when there is an upper bound to the norms of the $A_n$ and an upper bound to the norms of the $B_n$, then there is also an upper bound to the norms of the inner products $Tr(A_nB_n)$. So, if all the $A$'s can't get too big and all the $B$'s can't get too big, then there's a (finite) limit to how big all their inner products can get." That should make the truth of this assertion obvious. $\endgroup$ – whuber Jul 19 '13 at 23:31

Following @whuber's suggestion, starting with

$$\begin{eqnarray*} A &=& \left(\begin{array}{rrrr} a & b \\ c & d \end{array}\right)\\ P &=& \left(\begin{array}{rrrr} p & q \\ r & s \end{array}\right)\\ \end{eqnarray*}_.$$

then the matrix product is

$$\begin{eqnarray*} AP^T &=& \left(\begin{array}{rrrr} a & b \\ c & d \end{array}\right) \left(\begin{array}{rrrr} p & r \\ q & s \end{array}\right)\\ &=& \left(\begin{array}{llll} ap+bq & \ldots \\ \ldots & cr + ds \end{array}\right)\\ \end{eqnarray*}_.$$

The norm of a matrix $$\begin{eqnarray*} ||A|| &=& Tr(AA^T)^{1/2} \\ &=& (a^2 + b^2 + c^2 + d^2)^{1/2} \end{eqnarray*}_,$$

is just the Euclidian norm of the matrix treated as a vector of its coefficients $\mathbf{a} = (a, b, c, d)^T$.

The trace of the product $$\begin{eqnarray*} Tr(AP^T) &=& ap + bq + cr + ds \end{eqnarray*},$$

is just the dot product (inner product) of the matrices treated as vectors of their coefficients.

The Cauchy-Schwarz inequality states that

$$ |\mathbf{x}\cdot\mathbf{y}| \le ||\mathbf{x}||\,||\mathbf{y}||,$$

where the vector norms on the right are Euclidian. This inequality can then be applied to the vectors $\mathbf{a}$ and $\mathbf{p}$ of coefficients of the matrices $A$ and $P$.

$$ |Tr(AP^T)| \le ||A||\,||P|| $$

We are given $\sup_{n\in N}||A_n|| < \infty$, and $\sup_{n\in N}||B_n|| < \infty$. Hence

$$ \sup_{n\in N}|Tr(AP^T)| \le \sup_{n\in N}(||A||\,||P||) < \infty, \;\;\mbox{and}$$ $$\sup_{n\in N}|Tr(AP^T)| < \infty .$$

However this is not the form required in the question: it gives an inequality for the trace of the product $AP^T$ rather than $AP$. But $||X||=||X^T||$, which follows from the fact that for Euclidian vector norms $||\mathbf{x}^T|| = ||\mathbf{x}||$, or equivalently, for any matrix $X$, $Tr(XX^T)=Tr(X^TX)$. Setting $B=P^T$ $$\sup_{n\in N}|Tr(AB)| < \infty .$$

| cite | improve this answer | |
  • 1
    $\begingroup$ I think that you can start with $AP$ instead of $AP^T$. In that case there is no problems with a transpose. whuber: an idea to vectorize matrices is very nice. $\endgroup$ – Kolibris Jul 20 '13 at 13:31
  • $\begingroup$ (On the spelling of Cauchy-Schwarz: it's Hermann Amandus Schwarz, rather than Schwartz.) $\endgroup$ – Silverfish Oct 3 '15 at 0:14

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.