The MASS
package contains two small helper functions called fractions
and rational
. Originally these were included to allow patterns in patterned matrices to be more easily exhibited and hence, detected. This happens when the entries are expressible in the form of “vulgar” fractions, i.e. numerator/denominator, rather than as recurring decimals.
For example a Hilbert matrix:
C1 | C2 | C3 | C4 | C5 | |
---|---|---|---|---|---|
R1 | 1.0000000 | 0.5000000 | 0.3333333 | 0.2500000 | 0.2000000 |
R2 | 0.5000000 | 0.3333333 | 0.2500000 | 0.2000000 | 0.1666667 |
R3 | 0.3333333 | 0.2500000 | 0.2000000 | 0.1666667 | 0.1428571 |
R4 | 0.2500000 | 0.2000000 | 0.1666667 | 0.1428571 | 0.1250000 |
R5 | 0.2000000 | 0.1666667 | 0.1428571 | 0.1250000 | 0.1111111 |
has a pattern much easier to appreciate if the entries are expressed as vulgar fractions:
C1 | C2 | C3 | C4 | C5 | |
---|---|---|---|---|---|
R1 | 1 | 1/2 | 1/3 | 1/4 | 1/5 |
R2 | 1/2 | 1/3 | 1/4 | 1/5 | 1/6 |
R3 | 1/3 | 1/4 | 1/5 | 1/6 | 1/7 |
R4 | 1/4 | 1/5 | 1/6 | 1/7 | 1/8 |
R5 | 1/5 | 1/6 | 1/7 | 1/8 | 1/9 |
This package facilitates the calculation and presentation of numerical quantities as rational approximations.
To be more specific, a rational approximation to a real number is a ratio of two integers, the numerator and denominator, with the denominator positive. For uniqueness of representation we assume that the numerator and denominator are relatively prime, that is, have no common factor, or in other words the fraction is expressed in its lowest terms.
A second vignette, Theory, outlines, for the curious reader, the standard method used to find them, using continued fractions, and some basic mathematical properties. It is of mathematical interest, but not necessary to use the package itself.
This small package offers merely a way to present numerical values by means of close rational approximations, displayed as vulgar fractions. It does not provide exact rational arithmetic. Such a facility would require unlimited precision rational arithmetic, which is well beyond the scope of this package. In fact, integer overflow is a constant limitation in our use of fractions in the present package.
The package provides two main functions: fractional
and numerical
, and several auxiliaries. We discuss these in sequence now.
Readers interested in the details of the algorithm used and its implementation in R
should refer to the Theory vignette.
fractional
and unfractional
fractional
is the main function of the package. The argument list is as follows:
function (x, eps = 1e-06, maxConv = 20, sync = FALSE)
x
A numeric vector, matrix or array.eps = 1e-06
An absolute error tolerance. The rational approximation is the first convergent for which the absolute error falls below this bound.maxConv = 20
A secondary stopping criterion. An upper limit on the number of convergents (See Theory vignette) to consider.sync = FALSE
Should the numerical value be “synchronized” with the displayed numerical approximation? See below.The value returned is a numeric vector like x
, but inheriting from class fractional
. It may be used in arithmetic operations in the normal way and retains full floating point accuracy, unless sync = TRUE
. When printed, or coerced to character, however, the rational approximation is used as the representation, which as a numerical value may not precisely equal the full floating point value.
If sync = TRUE
is set, then the numerical value is changed to agree with the numerical value and this agreement is maintained in synchrony, as much as possible with floating point arithmetic, during future numerical operations.
unfractional
The function unfractional
is a convenience function that strips the class and some additional attributes from an object of class "fractional"
, thus demoting it back to a numerical object like the one from which it was generated.
fractional
objects in arithmetic operations.Methods are provided for the S3
group generic functions Ops
and Math
, allowing the mathematical operators and simple mathematical functions to be used with them. The normal rules apply. In an expression e1 + e2
(where +
can be any of the operators) the fractional
method is dispatched if either e1
or e2
(or both) is/are of class fractional
, and the result is of class fractional
.
For the Math
group generic, operations are conducted but treating the objects as a normal numeric object, that is, ignoring the fractional
class.
as.character
The package provides a method for the primitive generic function base::as.character
for objects of class "fractional"
. The arguments are as follows:
function (x, eps = attr(x, "eps"), maxConv = attr(x, "maxConv"), ...)
That is, the values for eps
and maxConv
are carried forward from the original call. This is because it is only when the object is displayed or coerced to character that the computation of the continued fraction is carried out.
The result is a character string object of primary class charFrac
. This primary class membership essentially ensures that the object is printed without quotation marks.
numerators
and denominators
These two S3
generic functions will extract the numerators and denominators, respectively.
The argument is a single object, x
, and the result is an object of type integer
, but carrying attributes such as dim
or dimnames
from the original.
Methods are provided for objects of S3
class fractional
or charFrac
numerical
This is essentially a method function for base::as.numeric
providing coercion to numeric for objects of S3
class fractional
or charFrac
.
It is written as an S3 generic function, and the main methods uses numerators
and denominators
in an obvious way.
The default
method passes the object to base::as.numeric
.
rat
, ratr
and .ratr
The function rat
is a front-end to the C++
function that carries out the continued fraction computations. It would normally not be called directly by users.
The functions ratr
and .ratr
, in combination, provide the identical computation but done in pure R
. They are provided for pedagogical and timing purposes only.
vfractional
This is a variant on fractional
, without the sync
argument, but vectorized with respect to the three arguments x
, eps
and maxConv
. It is mainly used for examples.
The Theory vignette provides many example in context, but the following very elementary examples may give the flavour of how the functions are used.
library(dplyr)
(x <- matrix(1:9/12, 3, 3) %>% fractional)
[,1] [,2] [,3]
[1,] 1/12 1/3 7/12
[2,] 1/6 5/12 2/3
[3,] 1/4 1/2 3/4
(xr <- 1/x)
[,1] [,2] [,3]
[1,] 12 3 12/7
[2,] 6 12/5 3/2
[3,] 4 2 4/3
x + 10
[,1] [,2] [,3]
[1,] 121/12 31/3 127/12
[2,] 61/6 125/12 32/3
[3,] 41/4 21/2 43/4
(x2 <- x^2)
[,1] [,2] [,3]
[1,] 1/144 1/9 49/144
[2,] 1/36 25/144 4/9
[3,] 1/16 1/4 9/16
(sx2 <- sqrt(x2)) ## demoted back to numeric by sqrt()
[,1] [,2] [,3]
[1,] 0.08333333 0.3333333 0.5833333
[2,] 0.16666667 0.4166667 0.6666667
[3,] 0.25000000 0.5000000 0.7500000
fractional(sx2) ## reinstated
[,1] [,2] [,3]
[1,] 1/12 1/3 7/12
[2,] 1/6 5/12 2/3
[3,] 1/4 1/2 3/4
solve(xr) %>% fractional ## xr is non-singular with a simple inverse
[,1] [,2] [,3]
[1,] 7/18 -10/9 3/4
[2,] -35/9 160/9 -15
[3,] 14/3 -70/3 21
numerators(x2) ## numerators and denominators are available
[,1] [,2] [,3]
[1,] 1 1 49
[2,] 1 25 4
[3,] 1 1 9
denominators(x2)
[,1] [,2] [,3]
[1,] 144 9 144
[2,] 36 144 9
[3,] 16 4 16
The golden ratio of the ancient Greeks, \(\phi = \left(\sqrt{5}+1\right)/2 = 1.61803398874989490\dots\) is well-known to be the limit of the ratio of successive Fibonacci numbers. As a partial indication of this result, the following demonstration is at least suggestive:
F <- c(1,1,numeric(15))
for(i in 3:17) ## Fibonacci sequence by recurrence
F[i] <- F[i-1] + F[i-2]
F
[1] 1 1 2 3 5 8 13 21 34 55 89 144 233 377 610
[16] 987 1597
(phi <- (sqrt(5) + 1)/2)
[1] 1.618034
The function vfractional
may be used to provide a sequence of optimal1 rational approximations with non-decreasing denominators:
vfractional(phi, eps = 0, maxConv = 1:16)
[1] 1 2 3/2 5/3 8/5 13/8 21/13 34/21
[9] 55/34 89/55 144/89 233/144 377/233 610/377 987/610 1597/987
With random numbers it is mildly interesting from a theoretical point of view to look at the distribution of the number of convergents in the continued fraction expansion needed to achieve a specific error tolerance. We have the tools to explore this somewhat arcane issue:
library(ggplot2)
N <- 500000
set.seed(3210)
rat(rnorm(N), eps = 1.0e-07)[, "n"] %>%
table(x = .) %>%
as.data.frame(responseName = "f") %>%
ggplot(.) + aes(x = x, y = f/N) +
geom_bar(stat = "identity", fill = "steel blue") +
ylab("Relative frequency") + xlab("Convergents")
In particular this appears to justify the choice of maxConv = 20
as a suitable default2.