9
$\begingroup$

Define the following cost functions $f_1, f_2 :SU(n) \rightarrow \mathbb{R}$ by $f_1(U) = Re \left( \text{Tr}\left(G^{\dagger} U \right) \right)$ and $f_2(U) = \left| \left( \text{Tr}\left(G^{\dagger} U \right) \right) \right|^2$ for some fixed $G \in SU(n)$.

What are the gradients $\nabla f_1, \nabla f_2$ of these functions and how can they be expressed?

$\endgroup$
4
  • $\begingroup$ Where is $W$ used? $\endgroup$ – Robin Goodfellow Jul 11 '15 at 16:25
  • $\begingroup$ Fixed it, thanks $\endgroup$ – Benjamin Jul 11 '15 at 17:44
  • $\begingroup$ Your first sentence gets an endorsement from the Redundancy Department of Redundancy :-) $\endgroup$ – Mariano Suárez-Álvarez Jul 11 '15 at 17:49
  • $\begingroup$ Is part of your question what the gradient of a function on $SU(n)$ is? $\endgroup$ – Mariano Suárez-Álvarez Jul 11 '15 at 17:53
5
$\begingroup$

So let's do $f_2$. What you need to do is to first work out the derivative in $M_n$, and then project onto $SU(n)$. That is, we identify $\mathfrak{su}(n)$ with its dual via the inner product $\langle U,V \rangle = \text{Re}\bigl(\text{Tr}(U^\dagger V)\bigr)$. (You need to do this identification if you want to figure out gradient flow, because $U^{-1} \frac{dU}{dt}$ has to be in $\mathfrak{su}(n)$.) Note that I am thinking of $SU(n)$ and $\mathfrak{su}(n)$ as a Lie group or Lie algebra respectively over the real numbers, not the complex numbers.

So if $V \in \mathfrak{su}(n)$, then at $U \in SU(n)$ we have $$ f_2(U(I + \epsilon V)) = f_2(U) + \epsilon \langle \nabla f_2,V \rangle + O(\epsilon^2) .$$ Multiplying out, we obtain $$ \langle \nabla f_2,V \rangle = 2 \text{Re} \bigl(\overline{\text{Tr}(G^\dagger U)} \, \text{Tr}(G^\dagger U V) \bigr) = 2 \text{Re} \bigl(\text{Tr}(H^\dagger U V) \bigr) ,$$ where $$ K = \text{Tr}(G^\dagger U) \, G . $$ So we see that $$ \nabla f_2 = 2 U^\dagger K + \Lambda ,$$ where $\Lambda \in \mathfrak{su}(n)^\perp$ is the "Lagrange multiplier" such that $\nabla f_2 \in \mathfrak{su}(n)$. Here $$\mathfrak{su}(n)^\perp = \{\Lambda \in M_n : \text{$\langle \Lambda,W\rangle = 0$ for all $W \in \mathfrak{su}(n)$}\} ,$$ After some reflection, you realize that $\mathfrak{su}(n)^\perp$ is the space of Hermitian matrices, and that $\nabla f_2$ is the anti-Hermitian part of $2 U^\dagger K$, that is $$ \nabla f_2 = U^\dagger K - K^\dagger U .$$ So the gradient flow is $$ \frac{dU}{dt} = U \, \nabla f_2 = K - U K^\dagger U .$$

$\endgroup$
18
  • 1
    $\begingroup$ $\nabla f_2$ is going to be an element of the dual space of $\mathfrak{su}(n)$. One construction of the dual of $\mathfrak{su}(n)$ is to think of it as a quotient of $M_n$ via the duality $\langle \nabla f_2, V\rangle = \text{Tr}((\nabla f_2)^T V) = \sum_{ij} (\nabla f_2)_{ij} V_{ij} $. $\endgroup$ – Stephen Montgomery-Smith Jul 12 '15 at 18:21
  • 1
    $\begingroup$ Except actually I am looking at the dual of the lie algebra translated by left multiplication of $U$. The whole thing is so much easier if you think of it in terms of matrices than abstract objects. $\endgroup$ – Stephen Montgomery-Smith Jul 12 '15 at 18:24
  • 1
    $\begingroup$ There are quite a few. The key words are gradient and double bracket flow. They are mostly easily accessible with no pay wall etc.. $\endgroup$ – Benjamin Jul 12 '15 at 23:58
  • 1
    $\begingroup$ This is very interesting to me, because I have been studying this paper: "Gradient descent of the frame potential," by Peter G. Casazza and Matthew Fickus, framerc.org/papers/Gradientdescentoftheframepotential.pdf (see also framerc.org/papers/RoadToEqualNormParsevalFrames.pdf) which use very similar gradient flows to solve problems in the communications industry. I think the connection between the work of Brockett and these papers could be quite substantial. In particular, they seem to have no reference to double bracket formulas. $\endgroup$ – Stephen Montgomery-Smith Jul 13 '15 at 16:29
  • 1
    $\begingroup$ OK, how about this rewrite? $\endgroup$ – Stephen Montgomery-Smith Jul 13 '15 at 22:12

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.