It’s not quite this simple, the same issue arises if every PSD completion of the known-diagonal minor has zero determinant (e.g. ((?, 1, 2), (1, 1, 1), (2, 1, 1))). But I think in that case making the remaining diagonal entries large enough still makes the eigenvalues at least −ε, which is good enough.
It’s not quite this simple, the same issue arises if every PSD completion of the known-diagonal minor has zero determinant (e.g. ((?, 1, 2), (1, 1, 1), (2, 1, 1))). But I think in that case making the remaining diagonal entries large enough still makes the eigenvalues at least −ε, which is good enough.