multigrid methods for markov chains - mathematicshdesterc/websitew/data/presentations/pr… ·...
TRANSCRIPT
![Page 1: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/1.jpg)
NII Tokyo
Multigrid Methods for Markov Chains
Hans De Sterck Department of Applied Mathematics, University of Waterloo
![Page 2: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/2.jpg)
NII Tokyo [email protected]
collaborators
• Killian Miller Department of Applied Mathematics, University of Waterloo, Canada
• Steve McCormick, Tom Manteuffel, John Ruge Department of Applied Mathematics, University of Colorado at Boulder, USA
• Geoff Sanders Lawrence Livermore National Laboratory, USA
![Page 3: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/3.jpg)
NII Tokyo [email protected]
1. simple Markov chain example • start in one state with
probability 1: what is the stationary probability vector after ∞ number of steps?
• stationary probability:
• this particular Markov chain is an example of a random walk on a graph
![Page 4: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/4.jpg)
NII Tokyo [email protected]
applications of Markov Chains
• information retrieval • performance
modelling of computer systems
• analysis of biological systems
• queueing theory"• Googleʼs PageRank"• ..."
![Page 5: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/5.jpg)
NII Tokyo [email protected]
2. problem statement
• B is column-stochastic
• B is irreducible (every state can be reached from every other state in the directed graph) ⇒
(no probability sinks!) probability sinks not irreducible
![Page 6: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/6.jpg)
NII Tokyo [email protected]
3. power method
• largest eigenvalue of B:
• power method:
– convergence factor: – convergence is very slow when
(slowly mixing Markov chain) (JAC, GS also slow)
![Page 7: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/7.jpg)
NII Tokyo [email protected]
some example Markov chains
• uniform 2D lattice
symmetric, real spectrum for B
![Page 10: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/10.jpg)
NII Tokyo [email protected]
numerical results: one-level (power) iteration for random graph problem
• start from random intial guess
• let
• iterate on
with
until
• W=O(n^2) method (A sparse, O(n) iterations)
![Page 11: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/11.jpg)
NII Tokyo [email protected]
when is power method slow?
• power method is slow on graphs with local links,
• power method is fast on graphs with global links, short distances
• PageRank is fast mixing: you can just do power method (PageRank is made fast by artificially adding global links from all webpages to all webpages with probability 0.15)
![Page 12: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/12.jpg)
NII Tokyo [email protected]
why/when is power method slow? why multilevel methods?
0.5
0.5 0.5 0.5 0.5 0.5
0.5 0.5 0.5 0.5 0.5
0.5
1
1
![Page 13: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/13.jpg)
NII Tokyo [email protected]
• high-frequency error is removed by relaxation (weighted Jacobi, Gauss-Seidel, ... power method)
• low-frequency-error needs to be removed by coarse-grid correction
principle of multigrid (for PDEs)
![Page 14: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/14.jpg)
NII Tokyo [email protected]
multigrid hierarchy: V-cycle
• multigrid V-cycle: relax (=smooth) on successively coarser grids transfer error using restriction (R=PT) and interpolation (P)
• W=O(n) : (optimally) scalable method
![Page 15: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/15.jpg)
NII Tokyo [email protected]
4. aggregation for Markov chains
• form three coarse, aggregated states
(Simon and Ando, 1961)
![Page 17: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/17.jpg)
NII Tokyo [email protected]
two-level aggregation method
repeat
(similar to lumping method from Takahashi, 1975) (‘iterative aggregation/disaggregation’) (note: there is a convergence proof for this two-level method, Marek and Mayer 1998, 2003)
![Page 18: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/18.jpg)
![Page 19: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/19.jpg)
NII Tokyo [email protected]
multilevel aggregation algorithm
(Krieger, Horton 1994, but no good way to build Q, convergence not good)
![Page 20: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/20.jpg)
NII Tokyo [email protected]
well-posedness: singular M-matrices
• singular M-matrix:
• our A=I-B is a singular M-matrix on all levels
![Page 22: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/22.jpg)
NII Tokyo [email protected]
aggregation strategy
• fine-level relaxation should efficiently distribute probability within aggregates (smooth out local, high-frequency errors)
• coarse-level update will efficiently distribute probability between aggregates (smooth out global, low-frequency errors)
• base aggregates on ‘strong connections’ in
![Page 23: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/23.jpg)
NII Tokyo [email protected]
aggregation strategy
scaled problem matrix:
strong connection: coefficient is large in either of rows i or j
( θ ∈ (0,1), θ=0.25 )
![Page 26: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/26.jpg)
NII Tokyo [email protected]
numerical results: aggregation multigrid for random walk problem"
does not work so well yet (not O(n) ...)
![Page 27: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/27.jpg)
our work since 2006: speed up the multilevel aggregation method
goal: W=O(n) (number of V-cycles independent of n)
1. smoothed aggregation (SIAM J. Sc. Comp., submitted 2008)
2. build P by algebraic multigrid (SIAM J. Sc. Comp., submitted 2009)
3. recursive iterant recombination (SIAM J. Sc. Comp., submitted 2009)
4. overcorrection (NLAA, submitted 2010)
(inspired by algebraic multigrid for PDEs)
NII Tokyo [email protected]
![Page 28: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/28.jpg)
NII Tokyo [email protected]
5. overlapping aggregates: we need ‘smoothed aggregation’...
after smoothing:
coarse grid correction with Q:
coarse grid correction with Qs:
(Vanek, Mandel, and Brezina, Computing, 1996)
![Page 29: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/29.jpg)
NII Tokyo [email protected]
smoothed aggregation
• smooth the columns of P with weighted Jacobi:
• smooth the rows of R with weighted Jacobi:
![Page 30: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/30.jpg)
NII Tokyo [email protected]
smoothed aggregation: a problem with signs
• smoothed coarse level operator:
• problem: Acs is not a singular M-matrix (signs wrong) • solution:
lumping approach • well-posedness of
this approach shown in our paper
![Page 31: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/31.jpg)
NII Tokyo [email protected]
numerical results: smoothed aggregation multigrid for random graph problem
![Page 32: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/32.jpg)
NII Tokyo [email protected]
6. algebraic multigrid for Markov chains
• scaled problem matrix:
• multiplicative error equation:
• we can use ‘standard’ AMG on • define AMG coarsening and interpolation
• lumping can be done as for smoothed aggregation
![Page 35: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/35.jpg)
7. recursively accelerated (pure) aggregation
• idea: recombine iterates at all levels in W cycle
NII Tokyo [email protected]
![Page 36: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/36.jpg)
recursively accelerated (pure) aggregation
• for Ax=b, use recursive Krylov acceleration • for Markov: need to impose probability constraints
NII Tokyo [email protected]
• standard quadratic programming problem
![Page 38: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/38.jpg)
8. over-correction, and ‘frozen’ additive cycles
• (with Eran Treister and Irad Yavneh)
• idea: ‘shape’ of correction is often good, but ‘amplitude’ may be too small therefore, overcorrect with factor α
• determine optimal α automatically
NII Tokyo [email protected]
![Page 39: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/39.jpg)
‘frozen’ additive cycles
• idea: replace expensive ‘multiplicative’ cycles by cheap ‘frozen’ additive cycles (as soon as good convergence)
• can do this ‘on-the-fly’ (OTF) • can lead to large speed gains
NII Tokyo [email protected]
![Page 40: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/40.jpg)
‘frozen’ additive cycles
multiplicative formulation:
additive formulation:
equivalent via:
NII Tokyo [email protected]
![Page 44: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/44.jpg)
9. conclusions
• algebraic multilevel methods can lead to W=O(n) solvers for slowly mixing Markov chains
• we have developed several ways to accelerate ‘pure’ aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid recursive iterant recombination over-correction (and frozen additive cycles) (these approaches are inspired on multigrid for
PDEs)
NII Tokyo [email protected]
![Page 45: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/45.jpg)
conclusions
• theory is very hard because the systems are non-symmetric
• our methods will not be fast for: high-dimensional lattices, queues, tensor-product
structure fast mixing Markov chains (but we can handle unstructured Markov chains)
• good results are obtained for many slowly mixing Markov chains
• multilevel methods can be very powerful
NII Tokyo [email protected]
![Page 47: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/47.jpg)
SIAM CSE 2009 [email protected]
6. Test Problems
(De Sterck et al., SISC, 2008, ‘Multilevel adaptive aggregation for Markov chains,
with application to web ranking’)
![Page 48: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/48.jpg)
SIAM CSE 2009 [email protected]
6.1 Uniform 1D Chain
• random walk on (undirected) graph • all edges have the same weight • transition probability for directed edge =
weight of edge / sum of weights of outgoing edges • solution trivial - test problem • random walk on undirected graph gives real-spectrum B
0.5
0.5 0.5 0.5 0.5 0.5
0.5 0.5 0.5 0.5 0.5
0.5
1
1
![Page 60: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/60.jpg)
SIAM CSE 2009 [email protected]
8. Conclusions
• A-SAM (Smoothed Aggregation for Markov Chains) and MCAMG (Algebraic Multigrid for Markov Chains) are scalable: they are algorithms for calculating the stationary vector of slowly mixing Markov chains with near-optimal complexity
• smoothing is essential for aggregation for many problems • appropriate theoretical framework (well-posedness) • no theory yet on (optimal) convergence (non-symmetric
matrices) • this can be done in parallel • other presentations in this mini-symposium: other multilevel
methods for the stationary Markov problem
• Questions?
![Page 61: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/61.jpg)
SIAM CSE 2009 [email protected]
Algebraic Aggregation Mechanism
(scaled problem matrix)
(strength matrix)
![Page 62: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/62.jpg)
SIAM CSE 2009 [email protected]
Error Equation
• multiplicative error: • error equation:
• coarse grid equation:
• restriction and interpolation:
• coarse grid correction:
![Page 65: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/65.jpg)
RSA 2009 [email protected]
numerical results: smoothed aggregation multigrid for periodic 2D lattice problem
![Page 66: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/66.jpg)
SIAM CSE 2009 [email protected]
We Need ‘Smoothed Aggregation’...
after smoothing:
coarse grid correction with Q:
coarse grid correction with Qs:
(Vanek, Mandel, and Brezina, Computing, 1996)
![Page 67: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/67.jpg)
SIAM CSE 2009 [email protected]
Smoothed Aggregation
• smooth the columns of P with weighted Jacobi:
• smooth the rows of R with weighted Jacobi:
![Page 68: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/68.jpg)
SIAM CSE 2009 [email protected]
Smoothed Aggregation
• smoothed coarse level operator:
• problem: Acs is not a singular M-matrix (signs wrong)
• solution: lumping approach on S in
![Page 69: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/69.jpg)
SIAM CSE 2009 [email protected]
Smoothed Aggregation
• we want as little lumping as possible • only lump ‘offending’ elements (i,j):
(we consider both off-diagonal signs and reducibility here!) • for ‘offending’ elements (i,j), add S{i,j} to S:
conserves both row and column sums
![Page 70: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/70.jpg)
SIAM CSE 2009 [email protected]
Lumped Smoothed Method is Well-posed (A-SAM: Algebraic Smoothed Aggregation for Markov Chains)
(De Sterck et al., SISC (accepted, 2009), ‘Smoothed aggregation multigrid for Markov chains’)
![Page 71: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/71.jpg)
SIAM CSE 2009 [email protected]
AMG Properties
• we can show: all elements of P >= 0 • lumping can be done as in the Smoothed
Aggregation case:
• lumping conserves row and column sums:
![Page 73: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/73.jpg)
SIAM CSE 2009 [email protected]
MCAMG Properties
(De Sterck et al., ‘Algebraic Multigrid for Markov Chains’, preprint)
![Page 74: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/74.jpg)
8. numerical results
1) random walk on 2D lattice
note: ‘+’ means additional top-level acceleration with window size 3
Copper 2010
![Page 75: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/75.jpg)
2) tandem queue
Copper 2010
![Page 76: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/76.jpg)
quadratic programming problem
efficient explicit solution for recombination of two iterates
Copper 2010
![Page 77: Multigrid Methods for Markov Chains - Mathematicshdesterc/websiteW/Data/presentations/pr… · aggregation methods such that W=O(n) is reached smoothed aggregation algebraic multigrid](https://reader036.vdocuments.mx/reader036/viewer/2022081411/60aa7afb9141bc3f796769e3/html5/thumbnails/77.jpg)
quadratic programming problem
efficient explicit solution for recombination of two iterates
Copper 2010