bdb106a0560c4e46ccc488ef010af787-Reviews.html

The key result shows for n samples drawn from some underlying distribution, the quality of subspace estimation improves at a rate O(n -r), where r related to the decay rate of the spectrum of the underlying distribution. Review: I am not familiar with the previous literature on PAC-style analysis of subspace learning, or if properties of the spectrum of the covariance was previously considered for subspace learning; so assuming that the work is novel, I believe authors have done a good job in relating these concepts. I do have a few suggestions that the authors should consider adding to the current text: Although authors have focused on the theoretical aspects of subspace learning, it would be nice to see how well the condition of'polynomial decay' holds on real world data. This would help with the significance of this work to the larger machine learning audience. Going a step further, it would be very instructive to see what the rates look like when the covariance C is unknown.