Finding the complex SDP matrix maximizing an eigenvalue-dependent criterion

157 Views Asked by At

Let $v_1,\cdots,v_k \in \mathbb{C}^n$ with $||v_i||=1$ for all $1\leq i\leq k$ be a certain discretization of the n-sphere of $\mathbb{C}^n$.

In some optimization problem, I have to work with a complex semi-definite matrix variable $W$ of size $n$ for which I want to approximate linearly (by relying essentially on the trace operator) the maximum eigenvalue, noted $\lambda_{W,\text{max}}$ .

Since $\lambda_{W,\text{max}}=\max_{v\in\mathbb{C}^n\text{ s.t. }||v||=1} \ \ \ v^HWv$ (not linear in $W$), my idea (which may be not that new, although I have not find any article going in that direction so far) is to approximate this maximum eigenvalue of $W$ by

$\max_{1\leq i \leq k} \ \ \ v_i^HWv_i$

which I believe could be expressed linearly in semi-definite programming for instance.

My question is the following : since the $v_i$ are known, is there a way (either analytically or algorithmically) to find, before performing the optimization I am actually interested in, the complex SDP matrix $W$ that maximizes

$\dfrac{\lambda_{W,\text{max}}}{\max_{1\leq i \leq k} v_i^HWv_i}$

along with the value of that maximum ratio ?

Note 1 : in my particular case, it could be sufficient to determine that maximum ratio by considering only rank one $W$ matrices: I'm not sure it would simplify the analysis, but if it does, let's exploit that.

Note 2 : this ratio would be a way to know in advance by how much the n-sphere discretization can miss the actual value of the maximum eigenvalue (therefore the approximation used would be quantified which can be very valuable in my application)

1

There are 1 best solutions below

4
On

For any matrix SDP matrix $W$ of rank 1 there exists a unitary matrix $U$ and a positive $\lambda$ such that $W = U^*diag(\lambda,0,...)U$, and conversely. Hence the ratio you want to maximise can be maximised for $U$ unitary and $\lambda$ positive. Now for any $U$ and $\lambda$ we have $$ v_i^*U^*diag(\lambda,0,...)Uv_i=\lambda |(Uv_i)_1|^2 $$ so your maximal ratio is $$ \max_{U^*U = Id, \lambda >0}\frac{\lambda}{\max_{1\leq i \leq k}\lambda |(Uv_i)_1|^2} = \left( \min_{U^*U=Id} \max_{1\leq i \leq k} |(Uv_i)_1|\right)^{-2}. $$ Also for any unity vector $u$, its image by the application $U \mapsto Uu$ is the unity sphere so you are left with $$ \left( \min_{u \in \mathbb{S}} \max_{1\leq i \leq k} |\langle u,v_i\rangle|\right)^{-2}. $$