Detailed Description
class gtsam::AcceleratedPowerMethod< Operator >
Compute maximum Eigenpair with accelerated power method.
References : 1) G. Golub and C. V. Loan, Matrix Computations, 3rd ed. Baltimore, Johns Hopkins University Press, 1996, pp.405-411 2) Rosen, D. and Carlone, L., 2017, September. Computational enhancements for certifiably correct SLAM. In Proceedings of the International Conference on Intelligent Robots and Systems. 3) Yulun Tian and Kasra Khosoussi and David M. Rosen and Jonathan P. How, 2020, Aug, Distributed Certifiably Correct Pose-Graph Optimization, Arxiv 4) C. de Sa, B. He, I. Mitliagkas, C. Ré, and P. Xu, “Accelerated stochastic power iteration,” in Proc. Mach. Learn. Res., no. 84, 2018, pp. 58–67
It performs the following iteration: \( x_{k+1} = A * x_k - \beta * x_{k-1} \) where A is the aim matrix we want to get eigenpair of, x is the Ritz vector
Template argument Operator just needs multiplication operator
Inheritance diagram for gtsam::AcceleratedPowerMethod< Operator >:
Public Member Functions | |
| AcceleratedPowerMethod (const Operator &A, const boost::optional< Vector > initial=boost::none, double initialBeta=0.0) | |
| Constructor from aim matrix A (given as Matrix or Sparse), optional intial vector as ritzVector. | |
| Vector | acceleratedPowerIteration (const Vector &x1, const Vector &x0, const double beta) const |
| Run accelerated power iteration to get ritzVector with beta and previous two ritzVector x0 and x00, and return y = (A * x0 - \beta * x00) / || A * x0. More... | |
| Vector | acceleratedPowerIteration () const |
| Run accelerated power iteration to get ritzVector with beta and previous two ritzVector x0 and x00, and return y = (A * x0 - \beta * x00) / || A * x0. More... | |
| double | estimateBeta (const size_t T=10) const |
| Tuning the momentum beta using the Best Heavy Ball algorithm in Ref(3), T is the iteration time to find beta with largest Rayleigh quotient. | |
| bool | compute (size_t maxIterations, double tol) |
| Start the accelerated iteration, after performing the accelerated iteration, calculate the ritz error, repeat this operation until the ritz error converge. More... | |
| Public Member Functions inherited from gtsam::PowerMethod< Operator > | |
| PowerMethod (const Operator &A, const boost::optional< Vector > initial=boost::none) | |
| Construct from the aim matrix and intial ritz vector. | |
| Vector | powerIteration (const Vector &x) const |
| Run power iteration to get ritzVector with previous ritzVector x, and return A * x / || A * x ||. | |
| Vector | powerIteration () const |
| Run power iteration to get ritzVector with previous ritzVector x, and return A * x / || A * x ||. | |
| bool | converged (double tol) const |
| After Perform power iteration on a single Ritz value, check if the Ritz residual for the current Ritz pair is less than the required convergence tol, return true if yes, else false. | |
| size_t | nrIterations () const |
| Return the number of iterations. | |
| bool | compute (size_t maxIterations, double tol) |
| Start the power/accelerated iteration, after performing the power/accelerated iteration, calculate the ritz error, repeat this operation until the ritz error converge. More... | |
| double | eigenvalue () const |
| Return the eigenvalue. | |
| Vector | eigenvector () const |
| Return the eigenvector. | |
Additional Inherited Members | |
| Protected Attributes inherited from gtsam::PowerMethod< Operator > | |
| const Operator & | A_ |
| Const reference to an externally-held matrix whose minimum-eigenvalue we want to compute. | |
| const int | dim_ |
| size_t | nrIterations_ |
| double | ritzValue_ |
| Vector | ritzVector_ |
Member Function Documentation
◆ acceleratedPowerIteration() [1/2]
|
inline |
Run accelerated power iteration to get ritzVector with beta and previous two ritzVector x0 and x00, and return y = (A * x0 - \beta * x00) / || A * x0.
- \beta * x00 ||
◆ acceleratedPowerIteration() [2/2]
|
inline |
Run accelerated power iteration to get ritzVector with beta and previous two ritzVector x0 and x00, and return y = (A * x0 - \beta * x00) / || A * x0.
- \beta * x00 ||
◆ compute()
|
inline |
Start the accelerated iteration, after performing the accelerated iteration, calculate the ritz error, repeat this operation until the ritz error converge.
If converged return true, else false.
The documentation for this class was generated from the following file:
- /Users/dellaert/git/github/gtsam/linear/AcceleratedPowerMethod.h
