The paper proposes a new method for learning a stable linear dynamical system from data, based on a recent paper  that shows that a matrix is stable if and only if it can be written as a product involving positive definite and orthogonal matrices. The proposed algorithm uses O(n^2) space in contrast to O(n^4) space used by previous similar methods, where n is the state dimension. The authors show that the new algorithm gets lower reconstruction error compared to baselines. Reviewers recommend acceptance and weren't concerned that the paper relies heavily on . I agree and I suggest it be accepted as a poster.