Private Matrix Approximation and Geometry of Unitary Orbits

被引:0
|
作者
Mangoubi, Oren [1 ]
Wu, Yikai [2 ]
Kale, Satyen [3 ]
Thakurta, Abhradeep Guha [3 ]
Vishnoi, Nisheeth K. [2 ]
机构
[1] Worcester Polytech Inst, Worcester, MA 01609 USA
[2] Yale Univ, New Haven, CT 06520 USA
[3] Google Res, Brooklyn, NY USA
来源
关键词
Unitary orbits; differential privacy; packing number; PCA; rank-k approximation;
D O I
暂无
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Consider the following optimization problem: Given n x n matrices A and., maximize < A, U Lambda U degrees > where U varies over the unitary group Upnq. This problem seeks to approximate A by a matrix whose spectrum is the same as Lambda and, by setting Lambda to be appropriate diagonal matrices, one can recover matrix approximation problems such as PCA and rank-k approximation. We study the problem of designing differentially private algorithms for this optimization problem in settings where the matrix A is constructed using users' private data. We give efficient and private algorithms that come with upper and lower bounds on the approximation error. Our results unify and improve upon several prior works on private matrix approximation problems. They rely on extensions of packing/covering number bounds for Grassmannians to unitary orbits which should be of independent interest.
引用
收藏
页数:42
相关论文
共 50 条