An extensible test matrix collection for Julia.
Give access to a wealth of sample and test matrices and accompanying data.
A set of matrices is generated locally (with arguments controlling the special case).
Another set is loaded from one of the publicly accessible matrix collections
SuiteSparse Matrix Collection
(formerly University of Florida Matrix Collection
)
and the Matrix Market Collection
, the latter being obsolescent.
Access is like
using MatrixDepot
?MatrixDepot # display package help info
A = matrixdepot("hilb", 10) # locally generated hilbert matrix dimensions (10,10)
A = matrixdepot("HB/1138_bus") # named matrix of the SuiteSparse Collection
or
md = mdopen("*/bfly") # named matrix with some extra data
A = md.A
co = md.coord
tx = md("Gname_10.txt")
md.<tab><tab> # overview of the "fields" of md returning like
# A m n dnz nnz coord Gname_10.txt G_10 Gcoord_10
or also
mdinfo("gravity") # text info about the selected matrix
md = mdopen("gravity", 10, false) # locally generated example with rhs and solution
A = md.A
b = md.b
x = md.x
NOTE: If you use Windows, you need to install MinGW/MSYS or Cygwin in order to use the SuiteSparse sparse and MatrixMarket matrix collection interface.
To install the release version, type
julia> Pkg.add("MatrixDepot")
Every Matrix type has a unique name, which is a string of one of the forms:
"name"
 used for matrices, which are generated locally."dir/name"
 for all matrices of theSuiteSparse
collection."dir/subdir/name"
 for all matrices of theMatrixMarket
collection.
The names are similar to relative path names, separated by a slash character.
The components of the name must not contain any of the characters "/*[]"
.
a set of matrices may be assigned to predefined or userdefined groups.
The group names are represented as Julia
symbols in the form :symmetric
.
The group names are therefore restricted to valid Julia
identifiers, that means
start with a letter and contain only letters, digits, and '_'
.
Every matrix has a numeric identifier, which is unique for its area:

builtin(id)
 one of the builtin matrix generators  currentlyid ∈ 1:59
. 
user(id)
 a userdefined matrix generator  starting with1
. 
sp(id)
 one of theSuiteSparse
collection. The integer ids are the 'official' ident numbers assigned by the collection. Currentlyid ∈ 1:3000
. 
mm(id)
 one of theMatrixMarket
collection. Here id follows the ordering of the index file of the collection.
For some functions it makes sense to have lists of matrix names to operate on, for example to select a set matrices with certain properties. These sets are described by 'Patterns', which are applied to matrix names and also to other matrix properties.
The following pattern types are supported:

"name"
 a string matching exactly a matrix name 
"shellpattern"
 a string with shell wildcards'?', '*', "[...]"
included. 
r"egular expression"
 a regular expression to match the matrix name. 
:group
 one of the defined group names; match all matrices in the group 
qualified numeric identifiers
 examplesbuiltin(10)
,sp(1:5, 7)
,mm(1), sp(:)

predicate_function
 the name of a predefined or userdefined boolean function of the internal data typeMatrixData
. Example:issymmetric
. 
abstract vector of subpatterns
OR
 any of the subpattern matches 
tuple of subpatterns
AND
 all of the subpatterns match 
~pattern
 negation of a pattern the \neg  operator ~ may be applied to all patterns
To express OR
and AND
, the binary operators 
and &
and (
/ )
are preferred.
Examples:
"gravity"  "HB/*" & ~(ishermitian & iscomplex) & ~sp(20:30)
The set of all known matrices can be expressed as empty tuple ()
. In a shell
pattern the double **
matches also slash characters, in contrast to the single *
.
A convenient form of a predicategenerator is
@pred(expression)
where expression is a valid Julia
boolean expression, which may access all
properties of MatrixData
as literal variable names.
Examples:
@pred(author == "J. Brown")
is translated to:
d > :author in propertynames(d) && d.author == "J. Brown"
@pred(500_000 <= n * m < 1_000_000)
restricts the size of matched matrices.
@pred(10^4 <= n <= 2*10^4 && n == m && nnz / n > 10 )
in average more than 10 entries per row
There is s set of predefined predicate functions including:
(issymmetric, ishermitian, isgeneral, isskew, isreal, iscomplex, isboolean, islocal, isremote, isloaded, isunloaded, isbuiltin, isuser, issparse)
Special predicate generators keyword(word...)
and hasdata(symbol...)
allow to
support keywordsearch and check for the existence of metadata.
For example: hasdata(:x) & ~keyword("fluid"
provides solution (x) and does not mention "fluid".
Beware that some sparse matrices contain nonstructural zeros, that is, coefficients stored explicitly but whose value is 0
.
In this case a discrepancy between nnz(A) and sum(!iszero, A) will be observed.
mdinfo() # overview
listgroups() # list all defined group names
mdlist(pattern) # array of matrix names according to pattern
listdata(pattern) # array of `MatrixData`objects according to pattern
listnames(pattern) # MDformatted listing of all names according to pattern
listdir("*//*") # MDformatted  group over part before `//`  count matching
mdinfo() # overview over database
mdinfo(pattern) # individual documentation about matrix(es) matching pattern
A = matrixdepot("kahan", 10)
generates a matrix using one of the builtin generators
md = mdopen("kahan", 10)
returns a handle md
; matrix can be obtained by
A = md.A
In general the first form is preferable, if only the pure matrix is required. For remote collections no arguments are used.
The second form allows to access all types of 'metadata', which may be available for some local or remote matrices.
Examples:
md = mdopen("spikes", 5, false); A = md.A; b = md.b; x = md.x
md = mdopen("Rommes/bips07_1998"); A = md.A; v = md.iv; title = md.data.title; nodenames = md("nodename.txt")
The last example shows, how to access textual metadata, when the name contains
Julia
nonword characters. Also if the metadataname is stored in a variable,
the last form has to be used.
meta = metasymbols(md)[2]; sec_matrix = md(meta)
The function metasymbols
returns a list of all symbols denoting metadata
provided by md
. Whether expressed as symbols or strings does not matter.
The system function propertynames(md)
returns all data of md
. That includes
size information and metadata.
propertynames(md.data)
gives an overview about all attributes of the
md.data::MatrixData
, which can for example be used in the @pred
definitions.
The remote data are originally stored at the remote website of one of the matrix collections. Before they are presented to the user, they are downloaded to local disk storage, which serves as a permanent cache.
By default, the data directory is a scratchspace managed by Scratch.jl
, but can be changed by setting the MATRIXDEPOT_DATA
environment variable.
The data directory can be queried by
julia> MatrixDepot.data_dir()
"/home/.../.julia/scratchspaces/b51810bbc9f355daae3c350fc1fbce05/data
The occasional user needs not bother about downloads, because that is done in the background if matrix files are missing on the local disk.
The same is true for the data required by mdinfo(pattern)
. Actually these are
stored in separate files if the full matrix files (which may be huge) are not yet loaded.
A download job to transmit a subset of remote matrix files may be started to
load header data for all files. Header data always include the matrix type
according to the matrixmarketformat and the size values m
rownumber,
n
= columnsnumber, and dnz
number of stored data of the main sparse matrix.
MatrixDepot.loadinfo(pattern)
where pattern
defines the subset.
That is possible for the SuiteSparse collection and the
NIST MatrixMarket collection.
The patterns can always refer to matrix names and id numbers.
In the case of SuiteSparse
collection, also the metadata
"date"
, "kind"
, "m"
, "n"
, "nnz"
are available and can be used, before individual matrix data
have been loaded. They are contained in a data file obtained from the remote site.
For MatrixMarket
collection, patterns are restricted to names and id numbers.
In general it would be possible by loadinfo("**")
to load all header data. That
would last maybe an hour and generate some traffic for the remote sites.
Nevertheless it is not necessary to do so, if you don't need the header data
for the following task.
MatrixDepot.load(pattern)
loads all data files for the patterns.
Patterns can only refer to attributes, which are already available.
In the case of SuiteSparse
that includes the size info "date"
, "kind"
,
"m"
, "n"
, and "nnz"
and all additional attributes loaded in the previous step,
which include "author"
, "title"
, "notes"
, and keywords.
In the case of MatrixMarket
you can only refer to "m"
, "n"
, and "dnz"
,
if previously loaded with the header data.
Please do not:
MatrixDepot.load("**")
. That would require some day(s) to finish and include
some really big data files (~100GB), which could be more than your disks can hold.
Make a reasonable selection, before you start a bulk download. Local and already loaded matrices are skipped automatically.
Example:
MatrixDepot.load(sp(:) & @pred(nnz < 100_000))
to download only problems with given
number of stored entries in the main matrix.
To see an overview of the matrices in the collection, type
julia> using MatrixDepot
julia> mdinfo()
Currently loaded Matrices
–––––––––––––––––––––––––––
builtin(#)
–––––––––– ––––––––––– ––––––––––– ––––––––––– –––––––––– –––––––––––– ––––––––––– ––––––––––– ––––––––––––– ––––––––––––
1 baart 7 circul 13 fiedler 19 gravity 25 invhilb 31 magic 37 parter 43 randcorr 49 shaw 55 ursell
2 binomial 8 clement 14 forsythe 20 grcar 26 invol 32 minij 38 pascal 44 rando 50 smallworld 56 vand
3 blur 9 companion 15 foxgood 21 hadamard 27 kahan 33 moler 39 pei 45 randsvd 51 spikes 57 wathen
4 cauchy 10 deriv2 16 frank 22 hankel 28 kms 34 neumann 40 phillips 46 rohess 52 toeplitz 58 wilkinson
5 chebspec 11 dingdong 17 gilbert 23 heat 29 lehmer 35 oscillate 41 poisson 47 rosser 53 tridiag 59 wing
6 chow 12 erdrey 18 golub 24 hilb 30 lotkin 36 parallax 42 prolate 48 sampling 54 triw
user(#)
–––––––––
1 randsym
Groups
–––––– ––––––– ––––– –––– ––––– ––––– ––––––– ––––––– –––––– –––––– ––––––– –––––– –––––––––
all builtin local user eigen graph illcond inverse posdef random regprob sparse symmetric
Suite Sparse of
–––––––––––– ––––
2770 2833
MatrixMarket of
–––––––––––– –––
488 498
We can generate a 4by4 Hilbert matrix by typing
julia> matrixdepot("hilb", 4)
4x4 Array{Float64,2}:
1.0 0.5 0.333333 0.25
0.5 0.333333 0.25 0.2
0.333333 0.25 0.2 0.166667
0.25 0.2 0.166667 0.142857
We can type the matrix name to get documentation about the matrix.
julia> mdinfo("hilb")
Hilbert matrix
≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡
The Hilbert matrix has (i,j) element 1/(i+j1). It is notorious for being
ill conditioned. It is symmetric positive definite and totally positive.
Input options:
• [type,] dim: the dimension of the matrix;
• [type,] row_dim, col_dim: the row and column dimensions.
Groups: ["inverse", "illcond", "symmetric", "posdef"]
References:
M. D. Choi, Tricks or treats with the Hilbert matrix, Amer. Math. Monthly,
90 (1983), pp. 301312.
N. J. Higham, Accuracy and Stability of Numerical Algorithms, second
edition, Society for Industrial and Applied Mathematics, Philadelphia, PA,
USA, 2002; sec. 28.1.
We can also specify the data type for locally generated matrices.
julia> matrixdepot("hilb", Float16, 5, 3)
5x3 Array{Float16,2}:
1.0 0.5 0.33325
0.5 0.33325 0.25
0.33325 0.25 0.19995
0.25 0.19995 0.16663
0.19995 0.16663 0.14282
julia> matrixdepot("hilb", Rational{Int}, 4)
4x4 Array{Rational{T<:Integer},2}:
1//1 1//2 1//3 1//4
1//2 1//3 1//4 1//5
1//3 1//4 1//5 1//6
1//4 1//5 1//6 1//7
Matrices can be accessed by a variety of patterns and composed patterns.
Integer numbers i
refer to the ident numbers in sp(i)
, mm(i)
, builtin(i)
, user(i)
.
Here sp
... denote the supported matrix collections SuiteSparse (formerly UFL),
Matrix Market, builtin, userdefined.
julia> mdlist(sp(1)) # here sp(1) is the ident number of the SuiteSparse collection
list(1)
–––––––––––
HB/1138_bus
julia> listnames(builtin(1, 5:10)) # the internal numbering of the builtinfunctions
list(7)
––––––– –––––––– –––– –––––– ––––––– ––––––––– ––––––
baart chebspec chow circul clement companion deriv2
julia> mdlist(builtin(1:4, 6, 10:15)  user(1:10) )
12element Array{String,1}:
"baart"
"binomial"
"blur"
"cauchy"
"chow"
"deriv2"
"dingdong"
"erdrey"
"fiedler"
"forsythe"
"foxgood"
"randsym"
While the listnames
command renders the output as markdown table, the internal
mdlist
produces an array of valid matrix names.
We can type a group name to see all the matrices in that group. Group names are always written as symbols to distinguish them form matrix names and pattern, which are always strings.
julia> listnames(:symmetric)
list(22)
–––––––– –––––––– ––––––– –––––– ––––––––– –––––––– ––––––– –––––––––
cauchy dingdong hilb lehmer oscillate poisson randsym wilkinson
circul fiedler invhilb minij pascal prolate tridiag
clement hankel kms moler pei randcorr wathen
It is possible to extend the builtin local problems with user defined generators and groups. We can add new matrix generators and define new groups of matrices.

Weijian Zhang and Nicholas J. Higham, "Matrix Depot: An Extensible Test Matrix Collection for Julia", PeerJ Comput. Sci., 2:e58 (2016), [pdf]

Nicholas J. Higham, "Algorithm 694, A Collection of Test Matrices in MATLAB", ACM Trans. Math. Software, vol. 17. (1991), pp 289305 [pdf] [doi]

T.A. Davis and Y. Hu, "The University of Florida Sparse Matrix Collection", ACM Transaction on Mathematical Software, vol. 38, Issue 1, (2011), pp 1:11:25 [pdf]

R.F. Boisvert, R. Pozo, K. A. Remington, R. F. Barrett, & J. Dongarra, " Matrix Market: a web resource for test matrix collections", Quality of Numerical Software (1996) (pp. 125137). [pdf]

Per Christian Hansen, "Test Matrices for Regularization Methods", SIAM Journal on Scientific Computing, vol. 16, 2, (1995) pp.506512. [pdf] [doi]