The Scaling and Squaring Method for the Matrix Exponential Revisited
read more
Citations
Multiconfiguration Self-Consistent Field and Multireference Configuration Interaction Methods and Applications
Computing the Action of the Matrix Exponential, with an Application to Exponential Integrators
Functions of matrices
A New Scaling and Squaring Algorithm for the Matrix Exponential
A Fast and Log-Euclidean Polyaffine Framework for Locally Linear Registration
References
Digital control of dynamic systems
Benchmarking optimization software with performance profiles
Nineteen Dubious Ways to Compute the Exponential of a Matrix, Twenty-Five Years Later ⁄
Related Papers (5)
Nineteen Dubious Ways to Compute the Exponential of a Matrix, Twenty-Five Years Later ⁄
Frequently Asked Questions (18)
Q2. What is the purpose of the subgraph centrality?
By combining walks of all possible lengths connecting node i to itself, and applying a weighting that decreases rapidly with the walk length, the subgraph centrality aims to capture the participation of the node in question in all subgraphs in the network.
Q3. How can the squaring phase be affected?
it is well known that rounding errors can significantly affect the scaling and squaring method, because the squaring phase can suffer from severe numerical cancellation.
Q4. What is the developed method for eH?
Of the available methods the most developed are Krylov methods, and these require the evaluation of eH for a much smaller Hessenberg matrix H related to A, for which the scaling and squaring method is well suited.
Q5. What is the software implementing the algorithm described here?
Software implementing the algorithm described here is available in the function expm in MATLAB (Version 7.2, R2006a, onwards), the function MatrixExp in Mathematica (Version 5.1 onwards), and routine F01ECF in the NAG Library (from Mark 22).
Q6. What is the accurate method on this set of test problems?
Since the curve for expm lies below all the other curves, expm is the least accurate method on this set of test matrices, as measured by the performance profile.
Q7. What is the effect of rounding errors on the evaluation of qm(A)?
To obtain rm the authors solve a multiple right-hand side linear system with qm(A) as coefficient matrix, so to be sure that this system is solved accurately the authors need to check that qm(A) is well conditioned.
Q8. What is the main problem of the scaling and squaring method?
A weakness of the scaling and squaring method that was first pointed out by Kenney and Laub [24] is that a choice of scaling based on ‖A‖ sometimes produces a much stronger scaling than is necessary in order to achieve the desired accuracy—with potentially detrimental consequences for numerical stability.
Q9. What is the cost of evaluating rm(A)?
which can be evaluated with m + 1 matrix multiplications by forming A2, A4, . . . , A2m. Thenq2m(A) = U − V is available at no extra cost.
Q10. How can the authors derive a forward error bound for the scaling and squaring method?
By using standard error analysis techniques it is possible to derive a forward error bound for the scaling and squaring method, as has been done by Ward [38].
Q11. What was the first reference to the matrix exponential in solving differential equations?
One of the first references to emphasize the important role of the matrix exponential in solving differential equations was the 1938 book Elementary Matrices and Some Applications to Dynamics and Differential Equations [13] by Frazer, Duncan, and Collar, who were in the Aerodynamics Department at the National Physical Laboratory in England.
Q12. what is the effect of rounding errors on the final squaring phase?
The effect of rounding errors on the final squaring phase remains an open question, but in their experiments the overall algorithm has performed in a numerically stable way throughout.
Q13. What was the first definition of the exponential of a matrix?
The exponential of a matrix was first introduced by Laguerre [26] in 1867, who gave the now standard power series definition: for A ∈ Cn×n,(0.1) eA = The author+A+ A22! +A33! + · · · .
Q14. What is the common error in the Exp(13) curve?
For α = 1, the Exp(13) curve is the highest: it intersects the y-axis at p = 0.52, which means that this method has the smallest error in 52% of the examples—more often than any other method.
Q15. What is the reason for the resulting bound?
it is designed to provide an explicit and easily computable error bound, and the resulting bound is far from being sharp.
Q16. How many squarings does Algorithm 2.3 require?
Algorithm 2.3 uses one to three fewer squarings than the algorithms with which the authors have compared it, and hence it has a potential advantage in accuracy.
Q17. what is the truncation error in the padé approximant?
MATRIX EXPONENTIAL REVISITED 753Theorem 2.1 is a backward error result: it interprets the truncation errors in the Padé approximant as equivalent to a perturbation in the original matrix A. (The result holds, in fact, for any rational approximation rm, as the authors have not yet used specific properties of a Padé approximant.)
Q18. What is the difference between expm and padm?
The authors note that in padm, pm and qm are evaluated in increasing order of norms of the terms, as in Algorithm 2.3, whereas expm uses the reverse ordering.