对稀疏矩阵逆协方差矩阵矩阵的二次逼近法估计.pptxVIP

Sparse Inverse Covariance Matrix Estimation ---Using Quadratic Approximation 1.IntroductionReviews LASSO with Gauss Graph Model Stepwise Regression LARS(least angle regression solution)  2.Problem Setup 2.Problem Setup 3.Quadratic Approximation Method Notes: g(X) is twice differentiable, and strictly convex h(X) is convex but non-differentiable. Set Newton direction: 3.Quadratic Approximation Method 3.1 Identifying variables to update 3.2 Computing the Step Size 3.2 Computing the Step Size 3.3 Computing the Newton Direction 3.3 Computing the Newton Direction 2.The solution of the one-variable problem corresponding to Dt yields ??; We expand the terms and omit the term not dependent on ?? Letting The minimum is achieved for: Where 3.3 Computing the Newton Direction 4.Convergence analysis In this section, we show that our algorithm has strong convergence guarantees. Our first main result shows that our algorithm does converge to the optimum solutions. Our second result then shows that the asymptotic convergence rate is actually superlinear, specifically quadratic. 4.1 convergence guarantee 4.1 convergence guarantee Lemma 3 There exists a unique minimizer X* for the object function Lemma 4 X* is the optimal solution of object function if and only if Where the minimum-norm sub-gradient is defined by 4.1 convergence guarantee 4.1 convergence guarantee 4.2 Proof of the quadratic convergence 5.Experiment performance In this section, we compare our method QUIC with other state-of-the-art methods on both synthetic and real datasets. We have implemented QUIC in C++, and all the experiments were executed on 2.83 GHz Xeon X5440 machines with 32G RAM and Linux OS. The comparisons on synthetic datasets. p stands for dimension, indicates the number of nonzeros in ground truth inverse covariance matrix, is the number of nonzeros in the solution, and ε is a specified relative error of objective value. ? ind