![Page 1: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/1.jpg)
Sparse Matrices
and Large Data Issues
Workshop ENAR
March 15, 2009
Reinhard Furrer, CSM
![Page 2: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/2.jpg)
Outline
2
What are sparse matrices?
How to work with sparse matrices?
Sparse positive definite matrices in statistics.
Sparse matrices and fields.
Thanks NSF!
DMS-0621118
![Page 3: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/3.jpg)
Outline
3
What are sparse matrices?
How to work with sparse matrices?
Sparse positive definite matrices in statistics.
Sparse matrices and fields.
![Page 4: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/4.jpg)
Sparse Matrices
4
What is “sparse” or a sparse matrix?
According to Wiktionary/Wikipedia:
Sparse: (Adjective)
1. Having widely spaced intervals
2. Not dense; meager
Sparse matrix:
a matrix populated primarily with zeros.
![Page 5: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/5.jpg)
Sparse Matrices
5
R> n <- 15
R> A <- array( runif(n^2), c(n,n)) + diag(n)
R> A[A < 0.75] <- 0
column
row
2 4 6 8 10 12 14
1412
108
64
2
![Page 6: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/6.jpg)
Sparse Matrices
6
R> n <- 15
R> A <- array( runif(n^2), c(n,n)) + diag(n)
R> A[A < 0.75] <- 0
R> AtA <- t(A) %*% A
column
row
2 4 6 8 10 12 14
1412
108
64
2
![Page 7: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/7.jpg)
Sparse Matrices
7
Why should we use sparse matrices?
![Page 8: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/8.jpg)
Sparse Matrices
7
Why should we use sparse matrices?
1. Savings in storage
2. Savings in computing time
![Page 9: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/9.jpg)
Sparse Matrices
8
Why should we use sparse matrices?
1. Savings in storage
nonzeros vs total elements
2. Savings in computing time
0.066s vs 0.003 for 1,000× 1,000 matrix multiplication
![Page 10: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/10.jpg)
Sparse Matrices
8
Why should we use sparse matrices?
1. Savings in storage
nonzeros vs total elements
2. Savings in computing time
0.066s vs 0.003 for 1,000× 1,000 matrix multiplication
To exploit the savings need to exploit the sparsity.
![Page 11: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/11.jpg)
Sparse Matrices
8
Why should we use sparse matrices?
1. Savings in storage
nonzeros vs total elements
2. Savings in computing time
0.066s vs 0.003 for 1,000× 1,000 matrix multiplication
To exploit the savings need to exploit the sparsity.
We need a clever storage format and fast algorithms.
![Page 12: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/12.jpg)
Storage Formats
9
Let A = (aij) ∈ Rn×m and z the number of its nonzero elements.
1. Naive/“traditional”/classic format:
one vector of length n×m and a dimension attribute.
![Page 13: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/13.jpg)
Storage Formats
9
Let A = (aij) ∈ Rn×m and z the number of its nonzero elements.
1. Naive/“traditional”/classic format:
one vector of length n×m and a dimension attribute.
2. Triplet format:
three vectors of length z, (i, j, aij) and a dimension attribute.
![Page 14: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/14.jpg)
Storage Formats
9
Let A = (aij) ∈ Rn×m and z the number of its nonzero elements.
1. Naive/“traditional”/classic format:
one vector of length n×m and a dimension attribute.
2. Triplet format:
three vectors of length z, (i, j, aij) and a dimension attribute.
3. Compressed sparse row (CSR) format:
eliminate redundant row indices.
![Page 15: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/15.jpg)
Storage Formats
9
Let A = (aij) ∈ Rn×m and z the number of its nonzero elements.
1. Naive/“traditional”/classic format:
one vector of length n×m and a dimension attribute.
2. Triplet format:
three vectors of length z, (i, j, aij) and a dimension attribute.
3. Compressed sparse row (CSR) format:
eliminate redundant row indices.
4. and about 10 more . . .
![Page 16: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/16.jpg)
Storage Formats, Example
10
A =
1 0.1 0 0.2 0.3
0.4 2 0 0.5 00 0 3 0 0.6
0.7 0.8 0 4 00.9 0 0.0 0 5
column
row
1 2 3 4 5
54
32
1
![Page 17: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/17.jpg)
Storage Formats, Example
10
A =
1 0.1 0 0.2 0.3
0.4 2 0 0.5 00 0 3 0 0.6
0.7 0.8 0 4 00.9 0 0.0 0 5
column
row
1 2 3 4 5
54
32
1
Naive/traditional/classic:
1, .4, 0, .7, .9, .1, 2, 0, .8, 0, 0, 0, 3, 0, .0, .2, .5, 0, 4, 0, .3, 0, .6, 0, 5
![Page 18: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/18.jpg)
Storage Formats, Example
11
A =
1 0.1 0 0.2 0.3
0.4 2 0 0.5 00 0 3 0 0.6
0.7 0.8 0 4 00.9 0 0.0 0 5
column
row
1 2 3 4 5
54
32
1
Triplet:i : 1 1 1 1 2 2 2 3 3 4 4 4 5 5 5j : 1 2 4 5 1 2 4 2 3 1 2 4 1 3 5aij : 1 .1 .2 .3 .4 2 .5 3 .6 .7 .8 4 .9 .0 5
![Page 19: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/19.jpg)
Storage Formats, Example
12
A =
1 0.1 0 0.2 0.3
0.4 2 0 0.5 00 0 3 0 0.6
0.7 0.8 0 4 00.9 0 0.0 0 5
column
row
1 2 3 4 5
54
32
1
i : 1 2 3 4 5j : 1 2 4 5 1 2 4 2 3 1 2 4 1 3 5aij : 1 .1 .2 .3 .4 2 .5 3 .6 .7 .8 4 .9 .0 5
![Page 20: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/20.jpg)
Storage Formats, Example
13
A =
1 0.1 0 0.2 0.3
0.4 2 0 0.5 00 0 3 0 0.6
0.7 0.8 0 4 00.9 0 0.0 0 5
column
row
1 2 3 4 5
54
32
1
CSR:ptr : 1 5 8 10 13 16j : 1 2 4 5 1 2 4 2 3 1 2 4 1 3 5aij : 1 .1 .2 .3 .4 2 .5 3 .6 .7 .8 4 .9 .0 5
![Page 21: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/21.jpg)
Compressed Sparse Row Format
14
1. the nonzero values row by row
2. the (ordered) column indices of nonzero values
3. the position in the previous two vectors corresponding to new
rows, given as pointers
4. the column dimension of the matrix.
CSR:ptr : 1 5 8 10 13 16j : 1 2 4 5 1 2 4 2 3 1 2 4 1 3 5aij : 1 .1 .2 .3 .4 2 .5 3 .6 .7 .8 4 .9 .0 5
![Page 22: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/22.jpg)
(Dis)Advantages
15
1. Naive format:
No savings in storage and computation (for sparse matrices)
Status quo
![Page 23: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/23.jpg)
(Dis)Advantages
15
1. Naive format:
No savings in storage and computation (for sparse matrices)
Status quo
2. Triplet format:
Savings in storage and computation for sparse matrices
Loss in storage and computation for full matrices
Intuitive
![Page 24: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/24.jpg)
(Dis)Advantages
15
1. Naive format:
No savings in storage and computation (for sparse matrices)
Status quo
2. Triplet format:
Savings in storage and computation for sparse matrices
Loss in storage and computation for full matrices
Intuitive
3. Compressed sparse row (CSR) format:
Apart from intuitive, same as triplet
![Page 25: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/25.jpg)
(Dis)Advantages
15
1. Naive format:
No savings in storage and computation (for sparse matrices)
Status quo
2. Triplet format:
Savings in storage and computation for sparse matrices
Loss in storage and computation for full matrices
Intuitive
3. Compressed sparse row (CSR) format:
Apart from intuitive, same as triplet
Faster element access
Many available algorithms
Arbitrary choice for “row” vs “column” format (CSC)
![Page 26: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/26.jpg)
Implications
16
With a new storage format new “algorithms” are required . . .
Is it worthwhile???
![Page 27: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/27.jpg)
Timing
17
Setup:
R> timing <- function(expr)
+ as.vector( system.time( for (i in 1:N) expr)[1])
R> N <- 1000 # how many operations
R> n <- 999 # matrix dimension
R> cutoff <- 0.9 # what will be set to 0
R> A <- array( runif(n^2), c(n,n))
R> A[A < cutoff] <- 0
R> S <- somecalltomagicfunctiontogetsparseformat( A)
Compare timing for different operations on A and S.
![Page 28: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/28.jpg)
Timing
18
R> timing( A + sqrt(A) )
[1] 0.058
R> timing( S + sqrt(S) )
[1] 0.061
R> timing( AtA <- t(A) %*% A )
[1] 0.467
R> timing( StS <- t(S) %*% S )
[1] 4.222
![Page 29: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/29.jpg)
Timing
19
R> timing( A[1,2] <- .5 )
[1] 0.007
R> timing( S[1,2] <- .5 )
[1] 0.018
R> timing( A[n,n-1] <- .5 )
[1] 0.001
R> timing( S[n,n-1] <- .5 )
[1] 0.012
![Page 30: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/30.jpg)
Timing
20
R> timing( xA <- solve(AtA, rep(1,n)) )
[1] 1.116
R> timing( xS <- solve(StS, rep(1,n)) )
[1] 1.51
R> timing( RA <- chol(AtA) )
[1] 0.488
R> timing( RS <- chol(StS) )
[1] 1.504
![Page 31: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/31.jpg)
Timing
20
R> timing( xA <- solve(AtA, rep(1,n)) )
[1] 1.116
R> timing( xS <- solve(StS, rep(1,n)) )
[1] 1.51
R> timing( RA <- chol(AtA) )
[1] 0.488
R> timing( RS <- chol(StS) )
[1] 1.504
Is it really worthwhile? What is going on?
![Page 32: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/32.jpg)
Timing
21
Matrix S
![Page 33: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/33.jpg)
Timing
22
Matrix S Matrix StS
![Page 34: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/34.jpg)
Timing
23
With cutoff 0.99:
R> timing( AtA <- t(A) %*% A )
[1] 0.106
R> timing( StS <- t(S) %*% S )
[1] 0.089
R> timing( RA <- chol(AtA) )
[1] 0.494
R> timing( RS <- chol(StS) )
[1] 0.451
![Page 35: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/35.jpg)
Timing
24
Matrix StS Matrix chol(StS)
Density of the factor is 41% with fill-in ratio 7.2.
![Page 36: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/36.jpg)
Timing
25
With cutoff 0.999:
R> timing( AtA <- t(A) %*% A )
[1] 0.059
R> timing( StS <- t(S) %*% S )
[1] 0.002
R> timing( RA <- chol(AtA) )
[1] 0.466
R> timing( RS <- chol(StS) )
[1] 0.007
![Page 37: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/37.jpg)
Timing
26
Matrix StS Matrix chol(StS)
Density of the factor is .6% with fill-in ratio 2.3.
![Page 38: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/38.jpg)
Implications
27
With a new storage format new “algorithms” are required . . .
Is it worthwhile??? Yes!
![Page 39: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/39.jpg)
Implications
27
With a new storage format new “algorithms” are required . . .
Is it worthwhile??? Yes!
Especially since
spam: R package for sparse matrix algebra.
![Page 40: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/40.jpg)
Outline
28
What are sparse matrices?
How to work with sparse matrices?
Sparse positive definite matrices in statistics.
Sparse matrices and fields.
![Page 41: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/41.jpg)
What is spam?
29
• an R package for sparse matrix algebra
– publicly available from CRAN, 0.15-3
– platform independent and documented
![Page 42: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/42.jpg)
What is spam?
29
• an R package for sparse matrix algebra
– publicly available from CRAN, 0.15-3
– platform independent and documented
• storage economical and fast
– uses “old Yale sparse format”
– most routines are in Fortran, adapted for spam
– balance between readability and overhead
– flags for “expert” users
![Page 43: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/43.jpg)
What is spam?
29
• an R package for sparse matrix algebra
– publicly available from CRAN, 0.15-3
– platform independent and documented
• storage economical and fast
– uses “old Yale sparse format”
– most routines are in Fortran, adapted for spam
– balance between readability and overhead
– flags for “expert” users
• versatile, intuitive and simple
– wrap an as.spam( ) and go
– S4 and S3 syntax
![Page 44: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/44.jpg)
What is spam?
29
• an R package for sparse matrix algebra
– publicly available from CRAN, 0.15-3
– platform independent and documented
• storage economical and fast
– uses “old Yale sparse format”
– most routines are in Fortran, adapted for spam
– balance between readability and overhead
– flags for “expert” users
• versatile, intuitive and simple
– wrap an as.spam( ) and go
– S4 and S3 syntax
• situated between SparseM and Matrix
![Page 45: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/45.jpg)
Representation of Sparse Matrices
30
spam defines a S4 class spam containing the vectors:
“entries”, “colindices”, “rowpointers” and ”dimension”.
R> slotNames( "spam")
[1] "entries" "colindices" "rowpointers" "dimension"
R> getSlots( "spam")
entries colindices rowpointers dimension
"numeric" "integer" "integer" "integer"
![Page 46: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/46.jpg)
Representation of Sparse Matrices
31
R> A
[,1] [,2] [,3] [,4] [,5]
[1,] 1.0 0.1 0 0.2 0.3
[2,] 0.6 2.0 0 0.5 0.0
[3,] 0.0 0.0 3 0.0 0.6
[4,] 0.7 0.8 0 4.0 0.0
[5,] 0.9 0.0 1 0.0 5.0
Class ’spam’
R> slotNames(A)
[1] "entries" "colindices" "rowpointers" "dimension"
R> A@entries
[1] 1.0 0.1 0.2 0.3 0.6 2.0 0.5 3.0 0.6 0.7 0.8 4.0 0.9 1.0 5.0
R> A@colindices
[1] 1 2 4 5 1 2 4 3 5 1 2 4 1 3 5
R> A@rowpointers
[1] 1 5 8 10 13 16
R> A@dimension
[1] 5 5
![Page 47: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/47.jpg)
Creating Sparse Matrices
32
Similar coercion techniques as with matrix:
• spam(...)
• as.spam(...)
Special functions:
• diag.spam(...)
• nearest.dist(...)
![Page 48: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/48.jpg)
Methods for spam
33
• Similar behavior as with matrices
plot; dim; determinant; %*%; +; ...
• Slightly enhanced behavior
print; dim<-; chol;
• Specific behavior
Math; Math2; Summary; ...
• New methods
display; ordering;
![Page 49: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/49.jpg)
Create Covariance Matrices
34
Covariance matrix:
nearest.dist and applying a covariance function:
R> C <- nearest.dist(x, diag=TRUE, upper=NULL)
R> C@entries <- Wendland( C@entries, dim=2, k=1)
![Page 50: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/50.jpg)
Create Covariance Matrices
34
Covariance matrix:
nearest.dist and applying a covariance function:
R> C <- nearest.dist(x, diag=TRUE, upper=NULL)
R> C@entries <- Wendland( C@entries, dim=2, k=1)
Precision matrix (GMRF):
— regular grids: nearest.dist with different cutoffs
R> diag.spam(n) +
+ (b1-b2) * nearest.dist(x, delta=1, upper=NULL) +
+ b2 * nearest.dist(x, delta=sqrt(2), upper=NULL)
— irregular grids: using incidence list and spam
R> incidence <- list( i=..., j=..., values)
R> C <- spam( incidence, n, n)
![Page 51: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/51.jpg)
Solving Linear Systems
35
A key feature of spam is to solve efficiently linear systems.
To solve the system Ax = b, we
• perform a Cholesky factorisation A = UTU
• solve two triangular systems UTz = b and Ux = z
But we need to “ensure” that U is as sparse as possible!
![Page 52: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/52.jpg)
Solving Linear Systems
35
A key feature of spam is to solve efficiently linear systems.
To solve the system Ax = b, we
• perform a Cholesky factorisation A = UTU
• solve two triangular systems UTz = b and Ux = z
But we need to “ensure” that U is as sparse as possible!
Permute the rows and columns of A: PTAP = UTU.
![Page 53: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/53.jpg)
Cholesky
36
Some technical details about a Cholesky decomposition:
[1] Determine permutation and
permute the input matrix A to obtain PTAP
[2] Symbolic factorization:
the sparsity structure of U is constructed
[3] Numeric factorization:
the elements of U are computed
![Page 54: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/54.jpg)
Cholesky
37
spam knows Cholesky!
— Several methods to construct permutation matrices P
— update to perform only ‘partial’ Cholesky factors
— Flags for avoiding sanity checks
![Page 55: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/55.jpg)
Cholesky
38
Cholesky factor with MMD RCM ordering
Density: 1.5%, fill-in: 4.7 Density: 2.7%, fill-in: 8.1
![Page 56: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/56.jpg)
Cholesky
39
5 10 20 50 100 200
L (log scale)
seco
nd
s (lo
g s
cale
)
.01
.11
10
10
0
5 10 20 50 100 200
L (log scale)
Mb
yte
s (lo
g s
cale
)
11
01
00
10
00
Time and memory usage for 101 Cholesky factorizations (solid)
and one factorization and 100 updates (dashed) of a precision
matrix from different sizes L of regular L×L grids with a second
order neighbor structure.
(The precision matrix from L = 200 has L4 = 1.6 ·109 elements).
See also demo( "article-jss").
![Page 57: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/57.jpg)
Cholesky
40
Gain of time and memory usage with different options and ar-
guments in the case of a second order neighbor structure of a
regular 50×50 grid and of the US counties. The time and mem-
ory usage for the generic call chol are 6.2 seconds, 174.5 Mbytes
and 15.1 seconds, 416.6 Mbytes, respectively.
Regular grid US countiesOptions or arguments time memory time memoryUsing the specific call chol.spam 1.001 0.992 0.954 1.004Option safemode=c(FALSE,FALSE,FALSE) 0.961 1.002 0.988 0.997Option cholsymmetrycheck=FALSE 0.568 0.524 0.646 0.493Passing memory=list(nnzR=..., nnzcolindices=...) 0.969 0.979 0.928 0.972All of the above 0.561 0.508 0.618 0.490All of the above and passing pivot=... to chol.spam 0.542 0.528 0.572 0.496All of the above and option cholpivotcheck=FALSE 0.510 0.511 0.557 0.489Numeric update only using update 0.132 0.070 0.170 0.063
![Page 58: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/58.jpg)
Cholesky
41
BTW:
efficient Cholesky factorization ⇐⇒ efficient determinant
calculation:
det(C) = det(UT) det(U) =n∏
i=1
U2ii
![Page 59: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/59.jpg)
Options in spam
42
For “experts”, flags to speed up the code:
R> powerboost() # in spam_0.15-4
R> noquote( format( spam.options()) )
eps drop printsize
2.220446e-16 FALSE 100
imagesize trivalues cex
10000 FALSE 1200
safemode dopivoting cholsymmetrycheck
TRUE, TRUE, TRUE TRUE TRUE
cholpivotcheck cholupdatesingular cholincreasefactor
TRUE warning 1.25, 1.25
nearestdistincreasefactor nearestdistnnz
1.25 160000, 400
![Page 60: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/60.jpg)
Limits of spam
43
What can spam not do (yet)?
• LU decompositions
• SVD/eigendecompositions
• Non-double elements
• . . .
But, please, comments to [email protected]!
![Page 61: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/61.jpg)
Outline
44
What are sparse matrices?
How to work with sparse matrices?
Sparse positive definite matrices in statistics.
Sparse matrices and fields.
![Page 62: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/62.jpg)
Sparse Matrices in Statistics
45
Where do large matrices occur?
• Location matrices
• Design matrices
![Page 63: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/63.jpg)
Sparse Matrices in Statistics
45
Where do large matrices occur?
• Location matrices
• Design matrices
• Covariance matrices
• Precision matrices
![Page 64: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/64.jpg)
Sparse Matrices in Statistics
46
• Covariance matrices:
Compactly supported covariance functions
Tapering
• Precision matrices:
(Gaussian) Markov random fields
(Tapering???)
We have symmetric positive definite (spd) matrices.
![Page 65: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/65.jpg)
Motivation
47
Precipitation anomaly along 40◦ lat.
*****
*
**
**
* ***** ***
**
*******
* ********
***
****
**********
*
***************
*******
*******
−120 −110 −100 −90 −80
−4
−2
02
4
Longitude
Ano
mal
y
Boulder
Lafayette (15km)
Woodrow (156km)
Indianapolis (1632km)
***
*
![Page 66: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/66.jpg)
Motivation
47
Precipitation anomaly along 40◦ lat.
Ordinary kriging
−120 −110 −100 −90 −80
−4
−2
02
4
01
Longitude
Ano
mal
y
Boulder
Lafayette (15km)
Woodrow (156km)
Indianapolis (1632km)
** ****
** ***********
*********
* ***********
*************
**
*****************
************
![Page 67: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/67.jpg)
Motivation
47
Precipitation anomaly along 40◦ lat.
Nearest neighbor kriging with 8 observations
*****
*
**
**
* ***** ***
**
*******
* ********
***
****
**********
*
***************
*******
*******
−120 −110 −100 −90 −80
−4
−2
02
4
01
Longitude
Ano
mal
y
Boulder
Lafayette (15km)
Woodrow (156km)
Indianapolis (1632km)
******
* **
![Page 68: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/68.jpg)
Motivation
47
Precipitation anomaly along 40◦ lat.
Tapering
*****
*
**
**
* ***** ***
**
*******
* ********
***
****
**********
*
***************
*******
*******
−120 −110 −100 −90 −80
−4
−2
02
4
01
Longitude
Ano
mal
y
Boulder
Lafayette (15km)
Woodrow (156km)
Indianapolis (1632km)
************
* *********
![Page 69: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/69.jpg)
Examples
48
0.0 0.2 0.4 0.6 0.8
Cov
aria
nce
Lag
ExponentialSpheric
![Page 70: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/70.jpg)
Examples
48
0.0 0.2 0.4 0.6 0.8
Cov
aria
nce
Lag
ExponentialSphericExp * Spheric
![Page 71: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/71.jpg)
Examples
48
0.0 0.2 0.4 0.6 0.8
Cov
aria
nce
Lag
ExponentialSphericExp * Spheric
0.0 0.2 0.4 0.6 0.8
Cov
aria
nce
Lag
Matern (1.5)Wendland
![Page 72: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/72.jpg)
Examples
48
0.0 0.2 0.4 0.6 0.8
Cov
aria
nce
Lag
ExponentialSphericExp * Spheric
0.0 0.2 0.4 0.6 0.8
Cov
aria
nce
Lag
Matern (1.5)WendlandMatern * Wendland
![Page 73: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/73.jpg)
Objective
49
For an isotropic and stationary process
with Matern covariance C0(h),
find a taper Cθ(h),
such that kriging with the product C0(h)Cθ(h)
is asymptotically optimal.
![Page 74: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/74.jpg)
Objective
49
For an isotropic and stationary process
with Matern covariance C0(h),
find a taper Cθ(h),
such that kriging with the product C0(h)Cθ(h)
is asymptotically optimal.
MSE(x∗, C0Cθ)
MSE(x∗, C0)→ 1
%(x∗, C0Cθ)
MSE(x∗, C0)→ γ
%(x∗, C) = C(0)− c∗TC−1c∗
![Page 75: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/75.jpg)
Misspecified Covariances
50
In a series of (Annals) papers, Stein gives asymptotic results for
misspecified covariances.
Under appropriate conditions, tapering is a form of misspecifica-
tion.
The taper has to be
• as differentiable at the origin as the original covariance
• more differentiable throughout the domain than at the origin
![Page 76: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/76.jpg)
Tapering
51
Tapering is an (asymptotically and computationally) efficient tech-
nique to create sparse covariance matrices.
Taper range can be justified by computing resources. However,
20–30 locations within the taper range is often sufficient.
“Classical” tapers are:
• spherical: Cθ(h) =(1− |h|
θ
)2
+
(1 + |h|
θ
)
• Wendland-type: Cθ(h) =(1− |h|
θ
)`+k
+· polynomial in |h|
θ of deg k
![Page 77: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/77.jpg)
Positive Definite Matrices
52
A (large) covariance (often) appears in:
• drawing from a multivariate normal distribution
• calculating/maximizing the (log-)likelihood
• linear/quadratic discrimination analysis
• PCA, EOF, . . .
But all boils down to solving a linear system and
possibly calculating the determinant . . .
‘Sparse PCA’ is sparse in a different sense . . .
![Page 78: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/78.jpg)
Outline
53
What are sparse matrices?
How to work with sparse matrices?
Sparse positive definite matrices in statistics.
Sparse matrices and fields.
![Page 79: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/79.jpg)
Sparse Matrices and fields
54
• fields is not bound to a specific sparse matrix format
• All heavy lifting is done in mKrig or Krig.engine.fixed
• For a specific sparse format, requires the methods:
chol, backsolve, forwardsolve and diag
as well as elementary matrix operations
need to exist
• If available uses operators to handle diagonal matrices quickly
The covariance matrix has to stem from particular class.
![Page 80: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/80.jpg)
Sparse Matrices and fields
54
• fields is not bound to a specific sparse matrix format
• All heavy lifting is done in mKrig or Krig.engine.fixed
• For a specific sparse format, requires the methods:
chol, backsolve, forwardsolve and diag
as well as elementary matrix operations
need to exist
• If available uses operators to handle diagonal matrices quickly
The covariance matrix has to stem from particular class.
fields uses spam as default package!
![Page 81: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/81.jpg)
Example mKrig
55
With appropriate covariance function:
R> x <- USprecip[ precipsubset, 1:2] # locations
R> Y <- USprecip[ precipsubset, 4] # anomaly
R> out <- mKrig(x,Y, m=1, cov.function="wendland.cov",theta=1.5)
R> out.p <- predict.surface( out, nx=220, ny=120)
R> surface(out.p, type=’I’)
R> US(add=T)
R> points(x,pch=’.’)
![Page 82: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/82.jpg)
Example mKrig
56
![Page 83: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/83.jpg)
Example Krig
57
R> out <- Krig( x,Y, m=1, cov.function="wendland.cov",theta=1.5,
+ lambda=0)
R> out.q <- predict.surface( out, nx=220, ny=120)
R> sum( ( out.q$z-out.p$z)^2, na.rm=T)
[1] 1.616783e-20
Krig/predict is slower (here 2.1/3.7 vs 10.4/3.9 seconds).
![Page 84: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/84.jpg)
Wendland family
58
wendland.cov (based on Wendland) produces a spam matrix.
All matrix functions are appropriately overloaded . . .
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.4
0.8
d
Cov
aria
nce
k=0k=1k=2k=3
Spherical
To create sparse covariance matrices based on other covariance
functions, use wendland.cov as skeleton.
![Page 85: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/85.jpg)
Tapering
59
Tapering can be performed with stationary.taper.cov.
Arguments are (selection):
Covariance = "Exponential"
Taper = "Wendland"
Taper.args = NULL : arguments for the taper function
Dist.args = NULL : arguments passed to nearest.dist
... : arguments passed to covariance function
All arguments can also be passed from mKrig/Krig
![Page 86: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/86.jpg)
Tapering
60
Compare the predicted surfaces without and with tapering:
R> out1 <- mKrig( xr,Yr, m=1, theta=1.5) )
R> out1.p <- predict.surface( out1, nx=220, ny=120) )
R> out2 <- mKrig( xr,Yr, m=1, theta=1.5,
+ cov.function="stationary.taper.cov",
+ Taper.args = list(k=0, theta=3)))
R> out2.p <- predict.surface( out2, nx=220, ny=120))
(timing yields 4/22 and 1/9 seconds)
![Page 87: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/87.jpg)
Tapering
61
Exponential covariance and with tapering
![Page 88: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/88.jpg)
Tapering
62
Exponential covariance and with tapering
![Page 89: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/89.jpg)
Asides . . .
63
The following arguments of mKrig/Krig are linked to spam:
Dist.args : arguments passed to nearest.dist
chol.args : arguments passed to chol
Use their help for fine tuning.
predict.se.Krig, predict.surface.se.Krig are very inefficient
because nrow(x) equations need to be solved.
![Page 90: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/90.jpg)
How Big is Big?
64
Upper limit to create a large matrix is the minimum of:
(1) available memory (machine and OS/shell dependent)
Error: ’cannot allocate vector of size’
(2) addressing capacity (231 − 1)
Error: ’cannot allocate vector of length’
However, R is based on passing by value, calls create local copies
(often 3–4 times the space of the object is used).
R> help("Memory-limits")
![Page 91: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/91.jpg)
And Beyond?
65
Parallelization:
nws, snow, Rmpi, . . .
Memory “Outsourcing”:
Matrices are not (entirely) kept in memory:
ff, filehash, biglm, . . .
(S+ has the library BufferedMatrix)
![Page 92: Sparse Matrices and Large Data Issues · – most routines are in Fortran, adapted for spam – balance between readability and overhead – flags for “expert” users. What is](https://reader030.vdocuments.site/reader030/viewer/2022013007/5c65090109d3f2a86e8c0ba3/html5/thumbnails/92.jpg)
References
66
Furrer, R., Genton, M. G. and Nychka, D. (2006). Covariance
Tapering for Interpolation of Large Spatial Datasets. Journal of
Computational and Graphical Statistics, 15(3), 502–523.
Furrer, R. and Sain, S. R. (2009). Spatial Model Fitting for Large
Datasets with Applications to Climate and Microarray Problems.
Statistics and Computing, 19(2), 113–128, doi:10.1007/s11222-
008-9075-x.
Furrer, R. and Sain, S. R. (2008). spam: A Sparse Matrix R
Package with Emphasis on MCMC Methods for Gaussian Markov
Random Fields. Submitted.