Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. 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 $A, X\in\mathbb{R}^{n\times n}$. The scalar objective function is $$J=\mathrm{tr}(AX)$$ If no constraints, let the derivative of $J$ with respect to $X$ be zeros, then we have $$A=0$$ Suppose $A$ is also a complex function, from $A=0$ I can further calculate something.

My question is: what if $X$ is an orthogonal matrix, i.e., $X^TX=I$? Then it becomes an constrained optimization problem. Can I still use matrix differential techniques to derive the minimizer? Thanks.

share|cite|improve this question
up vote 3 down vote accepted

The method of Lagrange multipliers yields the solution.

One studies the (real-valued) function $F(X)$ with the (matrix) constraint that $C(X)=I$ for $F(X)=\text{trace}(AX)$ and $C(X)=X^TX$, hence $$ F(X)=\sum\limits_{ij}A_{ji}X_{ij},\quad C_{ij}(X)=\sum\limits_{k}X_{ki}X_{kj}. $$ The gradients of $F$ and $C$ are given by $\partial_{ij}F(X)=A_{ji}$ and $$ \partial_{ij}C_{k\ell}(X)=X_{i\ell}[k=j]+X_{ik}[\ell=j]. $$ One wants that there exists some multipliers $\lambda_{ij}$ such that, for every $i$ and $j$, $$ \partial_{ij}F(X)=\sum\limits_{k\ell}\lambda_{k\ell}\partial_{ij}C_{k\ell}(X). $$ This condition reads $$ A_{ji}=\sum\limits_{k,\ell}\lambda_{k\ell}\left(X_{i\ell}[k=j]+X_{ik}[\ell=j]\right)=\sum\limits_{\ell}\lambda_{j\ell}X_{i\ell}+\sum\limits_{k}\lambda_{kj}X_{ik}, $$ or, equivalently, introducing the matrix $\Lambda=(\lambda_{ij})$, $$ A^T=X\Lambda^T+X\Lambda. $$ The matrix $M=\Lambda^T+\Lambda$ is such that $M^T=M$ and $XM=A^T$, hence $X^TX=I$ implies that $M$ should be such that $M^TM=M^2=AA^T$.

Using pseudo-inverse matrices if $A$ is not invertible and the usual definition of the square root of a symmetric matrix, one sees that the maximizing and minimizing matrices $X$ and the maximum and minimum values of $F$ are

When $A$ is invertible, one can use the usual definition of the square root of a symmetric matrix to find $M$. The maximizing and minimizing matrices $X$ and the maximum and minimum values of $F$ are$$ X_0=A^TM^{-1}=\pm A^T(AA^T)^{-1/2},\quad F(X_0)=\pm\mbox{trace}((AA^T)^{1/2}). $$ When $A$ is not invertible, the formula $X_0=A^TM^{-1}$, whichever meaning one gives to the notation $M^{-1}$, cannot yield an orthogonal matrix $X_0$ since, if $A$ is not invertible, neither are $A^T$, nor $A^T$ times any matrix. On the other hand, the formula $F(X_0)=\pm\mbox{trace}((AA^T)^{1/2})$ might still be valid.

share|cite|improve this answer
Nice Solution! Thanks a lot! – Shiyu Sep 20 '11 at 12:19
Hi Didier -- I need this solution for the answer to this question, but I don't understand the part about the pseudoinverses. In that case, $AA^T$ (in your notation, not the $A$ from the other question) may not be invertible; e.g. for one example I have $$A=\pmatrix{0&1\\0&0}\;,$$ and it seems that in that case I can't get $X_0$ in the way you've written it, even using the pseudoinverse, because the resulting matrix isn't orthogonal? – joriki Oct 30 '12 at 17:59
I found another solution and used it in the answer but also linked to this question here; I'd be interested to know if what you had in mind with the pseudoinverse is equivalent to the singular-value decomposition used there. – joriki Oct 30 '12 at 18:42
@joriki It has been a long time since I wrote this... It seems that here $AA^T$ is symmetric nonnegative hence diagonalizable as $AA^T=PDP^T$ with $PP^T=I$ and $D$ a diagonal matrix with a diagonal $(d_k)_k$, $d_k\geqslant0$. Then $(AA^T)^{-1/2}=PFP^T$ where $F$ is the diagonal matrix with the diagonal $(f_k)_k$, $f_k=1/\sqrt{d_k}$ if $d_k\gt0$, $f_k=0$ if $d_k=0$. But maybe this was not your question? – Did Oct 30 '12 at 18:42
@joriki Anyway, $X_0=\pm A^T\cdot$anything cannot be orthogonal when $A$ is not invertible, hence this case of the proof does not hold. Hmmmm... :-( – Did Oct 30 '12 at 20:02

Another way to get did's answer is to use the SVD to reduce to the case of $A$ being diagonal: We have $A = USV'$, where $U$ and $V$ are orthogonal and $S$ is diagonal, and so

$$\def\Tr{\operatorname{Tr}}\Tr(AX) = \Tr(SV'XU)\;.$$

Since $X\to V'XU$ is a bijection on the set of orthogonal matrices,

$$\min\{\Tr(AX)\mid X\text{ orthogonal}\} = \min\{\Tr(SW)\mid W\text{ orthogonal}\}\;.$$

But $\Tr(SW) = \sum_iS_{ii}W_{ii}$, and all the $S_{ii}$ are non-negative and $-1\le W_{ii}\le1$, so that $\Tr(SW)\ge\Tr(S(-I))$, and so the minimum occurs at $W=-I$ (or $X=-VU'$) and is $-\sum_iS_{ii}$.

share|cite|improve this answer
I formatted your answer because I thought it's important to have this solution worked out here; but it was quite a bit of work; I encourage you to take a look at the formatting FAQ; it's not as hard as it might seem. – joriki Oct 30 '12 at 23:49
Thanks joriki, will do. – dmuir Nov 3 '12 at 10:40

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.