Inexact spectral projected gradient methods on convex sets
read more
Citations
Probing the Pareto Frontier for Basis Pursuit Solutions
On Augmented Lagrangian Methods with General Lower-Level Constraints
Projected Barzilai-Borwein methods for large-scale box-constrained quadratic programming
Spectral residual method without gradient information for solving large-scale nonlinear systems of equations
A scaled gradient projection method for constrained image deblurring
References
Two-Point Step Size Gradient Methods
A nonmonotone line search technique for Newton's method
Nonmonotone Spectral Projected Gradient Methods on Convex Sets
Related Papers (5)
Nonmonotone Spectral Projected Gradient Methods on Convex Sets
A nonmonotone line search technique for Newton's method
Frequently Asked Questions (15)
Q2. What is the inner product space of symmetric matrices?
In this inner product space, the set S of symmetric matrices form a closed subspace and SDD+ is a closed and convex polyhedral cone [1, 18, 16].
Q3. How is the solution of the problem solved?
An effective way of solving (41) is by means of alternating projection methods combined with a geometrical understanding of the feasible region.
Q4. What is the bounded set of indices?
Since the set {dk}k∈K4 is bounded, there exists a sequence of indices K5 ⊂ K4 such that limk∈K5 dk = d and B ∈ B such that limk∈K5 Bk = B.
Q5. How many variables can be solved in the SPG method?
The SPG method was able to solve problems of this family with up to 96254 variables and up to 578648 constraints in very few seconds of computer time.
Q6. What is the simplest way to find the ncolsncols?
For the feasible region, L and U are given ncols×ncols real matrices, and SDD+ represents the cone of symmetric and diagonally dominant matrices with positive diagonal, i.e.,SDD+ = {X ∈ IRncols×ncols | XT = X and xii ≥ ∑j 6=i|xij | for all i}.
Q7. What is the function of the alternating projection algorithm?
Let us recall that for a given nonempty closed and convex set Ω of IRn, and any y0 ∈ IRn, there exists a unique solution y∗ to the problemmin y ∈ Ω‖y0 − y‖, (21)which is called the projection of y0 onto Ω and is denoted by PΩ(y0).
Q8. What are the good features of the ISPG method?
Numerical experiments were presented concerning constrained least-squares rectangular matrix problems to illustrate the good features of the ISPG method.
Q9. What is the dk of the algorithm?
Let dk be such that xk + dk ∈ Ω andQk(dk) ≤ η Qk(d̄k). (4)If dk = 0, stop the execution of the algorithm declaring that xk is a stationary point.
Q10. What is the Frobenius norm of a real matrix?
‖A‖F denotes the Frobenius norm of a real matrix A, defined as‖A‖2F = 〈A,A〉 = ∑i,j(aij) 2 ,where the inner product is given by 〈A,B〉 = trace(AT B).
Q11. What is the spectral choice of Bk?
In this case,Bk = 1λspgk Iwhereλspgk ={min(λmax,max(λmin, s T k sk/s T k yk)), if s T k yk > 0, λmax, otherwise,sk = xk − xk−1 and yk = gk − gk−1; so thatQk(d) = ‖d‖22λspgk + gTk d. (20)When Bk = (1/λ spg k )I (spectral choice) the optimal direction d̄k is obtained by projecting xk − λ spg k gk onto Ω, with respect to the Euclidean norm.
Q12. What is the simplest way to find a symmetric ncolsncol?
In particular, the authors consider the following problem:Minimize ‖AX −B‖2F subject toX ∈ SDD+0 ≤ L ≤ X ≤ U,(41)where A and B are given nrows× ncols real matrices, nrows ≥ ncols, rank(A) = ncols, and X is the symmetric ncols×ncols matrix that the authors wish to find.
Q13. What was the performance of the experiment?
All the experiments were run on a Sun Ultra 60 Workstation with 2 UltraSPARC-II processors at 296-Mhz, 512 Mb of main memory, and SunOS 5.7 operating system.
Q14. What is the main drawback of the SPG method?
In Section 3 the authors introduce the ISPG method and the authors describe the use of Dykstra’s alternating projection method for obtaining inexact projections onto closed and convex sets.
Q15. What is the way to solve the problemmin y y0?
the authors assume that for all y ∈ IRn, the calculation of PΩ(y) is a difficult task, whereas, for each Ωi, PΩi(y) is easy to obtain.