A probability mass function can be represented by a multi-dimensional array. However, for high-dimensional distributions where each variable may have a large state space, lack of computer memory can become a problem. For example, an 80-dimensional random vector in which each variable has 10 levels will lead to a state space with 1080 cells. Such a distribution can not be stored in a computer; in fact, 1080 is one of the estimates of the number of atoms in the universe. However, if the array consists of only a few non-zero values, we need only store these values along with information about their location. That is, a sparse representation of a table. Sparta was created for efficient multiplication and marginalization of sparse tables.
Consider two arrays f
and g
:
dn <- function(x) setNames(lapply(x, paste0, 1:2), toupper(x))
d <- c(2, 2, 2)
f <- array(c(5, 4, 0, 7, 0, 9, 0, 0), d, dn(c("x", "y", "z")))
g <- array(c(7, 6, 0, 6, 0, 0, 9, 0), d, dn(c("y", "z", "w")))
with flat layouts
ftable(f, row.vars = "X")
#> Y y1 y2
#> Z z1 z2 z1 z2
#> X
#> x1 5 0 0 0
#> x2 4 9 7 0
ftable(g, row.vars = "W")
#> Y y1 y2
#> Z z1 z2 z1 z2
#> W
#> w1 7 0 6 6
#> w2 0 9 0 0
We can convert these to their equivalent sparta versions as
Printing the object by the default printing method yields
print.default(sf)
#> [,1] [,2] [,3] [,4]
#> X 1 2 2 2
#> Y 1 1 2 1
#> Z 1 1 1 2
#> attr(,"vals")
#> [1] 5 4 7 9
#> attr(,"dim_names")
#> attr(,"dim_names")$X
#> [1] "x1" "x2"
#>
#> attr(,"dim_names")$Y
#> [1] "y1" "y2"
#>
#> attr(,"dim_names")$Z
#> [1] "z1" "z2"
#>
#> attr(,"class")
#> [1] "sparta" "matrix"
The columns are the cells in the sparse matrix and the
vals
attribute are the corresponding values which can be
extracted with the vals
function. Furthermore, the domain
resides in the dim_names
attribute, which can also be
extracted using the dim_names
function. From the output, we
see that (x2
, y2
, z1
) has a value
of 2. Using the sparta
print method prettifies things:
where row i corresponds to
column i in the sparse matrix.
The product of sf
and sg
mfg <- mult(sf, sg); mfg
#> X Y Z W val
#> 1 2 1 2 2 81
#> 2 2 2 1 1 42
#> 3 1 1 1 1 35
#> 4 2 1 1 1 28
Converting sf
into a conditional probability table (CPT)
with conditioning variable Z
:
sf_cpt <- as_cpt(sf, y = "Z"); sf_cpt
#> X Y Z val
#> 1 1 1 1 0.312
#> 2 2 1 1 0.250
#> 3 2 2 1 0.438
#> 4 2 1 2 1.000
Slicing sf
on X1 = x1
and dropping the
X
dimension
reduces sf
to a single non-zero element, whereas the
equivalent dense case would result in a (Y,Z)
table with
one non-zero element and three zero-elements.
Marginalizing (or summing) out Y
in sg
yields
Finally, we mention that a sparse table can be created using the
constructor sparta_struct
, which can be necessary to use if
the corresponding dense table is too large to have in memory.
Function name | Description |
---|---|
as_<sparta> |
Convert -like object to a sparta |
as_<array/df/cpt> |
Convert sparta object to an
array/data.frame/CPT |
sparta_struct |
Constructor for sparta objects |
mult, div, marg, slice |
Multiply/divide/marginalize/slice |
normalize |
Normalize (the values of the result sum to one) |
get_val |
Extract the value for a specific named cell |
get_cell_name |
Extract the named cell |
get_values |
Extract the values |
dim_names |
Extract the domain |
names |
Extract the variable names |
max/min |
The maximum/minimum value |
which_<max/min>_cell |
The column index referring to the max/min value |
which_<max/min>_idx |
The configuration corresponding to the max/min value |
sum |
Sum the values |
equiv |
Test if two tables are identical up to permutations of the columns |