Reformulation of LMI

120 Views Asked by At

In a paper I have read, the authors reformulated the following LMI

$$X_{t} \succeq \left[\begin{array}{cc}\alpha_{i}\left(B_{t} U+C_{t}\right)^{\top} E_{i}\left(B_{t} U+C_{t}\right) & \left( B_{t} U+C_{t}\right)^{\top} e_{i} \alpha_{i} \\\\ \alpha_{i} e_{i}^{\top}\left( B_{t} U+C_{t}\right) & e_{i}^{0} \alpha_{i}-\beta_{t}\end{array}\right]$$

into two LMIs

$$X_{t} - \left[\begin{array}{cc}\alpha_{i}P_{t}^{i} & \left( B_{t} U+C_{t}\right)^{\top} e_{i} \alpha_{i} \\\\ \alpha_{i} e_{i}^{\top}\left( B_{t} U+C_{t}\right) & e_{i}^{0} \alpha_{i}-\beta_{t}\end{array}\right] \succeq 0$$

and

$$\left[\begin{array}{cc}P_{t}^{i} & \left(B_{t} U+C_{t}\right)^{\top} E_{i}^{1 / 2} \\\\ E_{i}^{1 / 2}\left(B_{t} U+C_{t}\right) & \mathbb{I}_{n}\end{array}\right] \succeq 0 $$

where $\alpha_i$ and $e_i^0$ are scalar and $e_i$ is vector. The decision variables are $X_{t} \in S_{+}^{N d+2}$ and $P_{t}^{i} \in S_{+}^{N d+1}$.

I know the second LMI could be reformulated into $$P_{t}^{i} - \left(B_{t} U + C_{t}\right)^{\top} E_{i} \left(B_{t} U + C_{t}\right) \succeq 0$$ according to Schur complements. Just wondering why the two LMIs are equivalent to the first LMI and why the reformulation is necessary.

Any hint would be appreciated!

1

There are 1 best solutions below

3
On BEST ANSWER

The first semidefinite constraint is not an LMI as they have quadratic term in the top-left block (assuming $U$, $B_t$ or $C_t$ is a decision variable). However, it is convex in the standard form $A-B^TCB$ (generic variable names here) so they introduce an intermediate variable to upper bound the nonlinear part, and then apply a Schur complement to arrive at two LMIs. An alternative could be to apply a Schur complement directly on the first semidefinite constraint arriving at one larger LMI but without any extra variable.