When I did mathematics, I used to think that I work in very esotheric field of secant varieties[1]. Funny how I now see it pop up everywhere. Anyway, here's some geometric interpretation of his approach:
In space of all matrices, the set of rank one matrices is called a Segre variety[2]. Segre variety is cut out by 2x2 minors of a generic matrix, which is what the author's post is all about. What he's effectively trying to do is given a plane (which is the set of possible matrices having some given values), finding a point in the intersection of the Segre variety and that plane. The way he does it is by intersecting that plane one by one with large varieties cut out by a single minor, and noting that all points in the intersection must lie in a hyperplane not already containing the plane of possible matrices. This reduces the dimension of the plane of possible matrices, as they must lie in both the original plane and the constraint hyperplane. After enough steps we're left with a single point. The clever thing here is finding a good order of intersections, so that at each step we can find that hyperplane.
Now, rank k matrices form a k-th secant variety to the Segre variety, and they are cut out by k x k minors, and so for their part 2, I expect using larger minors.
Results should be the same as the convex optimization technique in the "well-defined" entries, though numerical stability is a real concern that I haven't looked at yet. Empirically I haven't had any issues, but if you have entries in the singular vectors sufficiently close to zero you can definitely imagine some problems. I don't think that's unique to this approach, either.
In space of all matrices, the set of rank one matrices is called a Segre variety[2]. Segre variety is cut out by 2x2 minors of a generic matrix, which is what the author's post is all about. What he's effectively trying to do is given a plane (which is the set of possible matrices having some given values), finding a point in the intersection of the Segre variety and that plane. The way he does it is by intersecting that plane one by one with large varieties cut out by a single minor, and noting that all points in the intersection must lie in a hyperplane not already containing the plane of possible matrices. This reduces the dimension of the plane of possible matrices, as they must lie in both the original plane and the constraint hyperplane. After enough steps we're left with a single point. The clever thing here is finding a good order of intersections, so that at each step we can find that hyperplane.
Now, rank k matrices form a k-th secant variety to the Segre variety, and they are cut out by k x k minors, and so for their part 2, I expect using larger minors.
[1] - https://en.wikipedia.org/wiki/Secant_variety [2] - https://en.wikipedia.org/wiki/Segre_embedding