Functional Calculation 0 : Numerical Apéritif
Neville Holmes
School of Computing and Information Systems, University of Tasmania
Launceston 7250, Australia
Neville.Holmes@utas.edu.au
Abstract
This article is the first in a series expounding the joys of functional calculation. Functional calculation does with operations applied to functions and numbers what numerical calculation does with functions applied to numbers. The functional notation used as the vehicle in this series is provided by a freely available calculation tool called J.
This article gives a detailed illustration of a few numerical uses of the notation supported by the J interpreter. This explanation is preliminary to following articles, which will review some of that tool’s support for numerical calculation. The illustration takes a simple function and shows how J, or any other tool for functional calculation, can be used to investigate special values of the function.
Calculation
Calculation is an honourable and satisfying occupation with a long history and a rich tradition [1].
Relatively recent developments in electronic computation have led to the adoption of three classes of aids to calculation, namely
- electronic or programmed calculators based directly on traditional calculation,
- spreadsheet programs, and
- complex mathematical systems.
Electronic calculators are unreliable and very poorly designed[2]. Spreadsheet programs are very powerful and useful for the kind of calculation that suits them, but clumsy and limited otherwise. The complex mathematical systems are extremely powerful, but in an elephantine way.
The kind of functional calculation described in following articles is like an enormous but sympathetic extension of traditional calculation. The extension is particularly effective in dealing with the kinds of tables to which spreadsheet programs are suited but allowing conveniently for tables of many dimensions. It is also useful for collections of tables of different kinds. The extension is of surprising expressive power so that problems which would otherwise require the might of a complex mathematical system can be tackled simply.
This article will go through some simple numerical calculations step by step, explaining them in detail. Readers will get most benefit out of the description if they get their own interpreter and themselves carry out the calculations (and any others suggested by the examples) as they are being described.
The notation used here is called simply J. Programs which evaluate J expressions as they are keyed in (because they seem to interpret the notation) are called interpreters. Interpreters for the more popular computing platforms can be downloaded free from Jsoftware[3]. The J mailing lists are active and helpful to new users, and much valuable material can be found on the J Wiki[4].
The primitive functions, those functions with symbols for names, used in this article are few, and very many more such are provided by the interpreter, and are listed and briefly explained by the interpreter’s help facilities.
The purpose of this article is to show how simple functions can be used simply to get useful results, the kind of results that show the interpreter in the light of a utility calculator. To allow this to be better appreciated, examples which involve the use of lists of numbers have been chosen.
The calculations used here as an example are based on an exploration
of a function named f in the description. Any reader carrying out
the example calculations can define f by keying in
f =: -o.
but this definition will not be further explained here because the
expression used to define f is a functional expression
and, as such, will be explained in later articles of this series.
Finding Zeroes
Given the function f and a need to explore its behaviour,
the first step is to name a useful list of test values, and then to
apply the function to those values. Keying in the expression
nn =: i.11
uses the primitive function i. (pronounced
integers) to produce the first eleven non-negative integers,
to which the copula =: (pronounced gets or
becomes) gives the name nn.
The copula is used only for conferring names, and it is not a function as such. No result is shown because the interpreter is just naming. However any named result can be looked at just by keying in its name, in this case
nn 0 1 2 3 4 5 6 7 8 9 10
shows nn to be a list of eleven integers.
Some general remarks are appropriate at this point. The J interpreter
takes in the expressions it evaluates indented, but the results are
shown unindented. For the first expression shown above, the spaces
between the nn and the =:, and between the
=: and the i., are not required by the
interpreter, but have been put there to make the expression more
readable. More than one space could have been used with same result,
and spaces could have been put between the i. and the
11 without affecting the result. However, putting a space
within the =: or the i. would have affected
the result just as putting a space within the 11 would
have, because the notation uses suffixed dots and colons to extend the
ASCII character set, so that in effect = and
=. and =: are three different characters.
The function f can now be evaluated for the argument
nn, and its result displayed, by
f nn 1 _0.423311 1.10587 10.6608 42.0318 132.705 384.579 1074.64 2955.83 8074.81 21995
which shows a quite distinctive behaviour indeed.
However, the inconveniently small font forced on the printing above of
the results of f nn makes it difficult here to appreciate
fully the nature of that behaviour, so it is useful to define with
scant comment
show =: ".&(0.1&":)
which names a function that can be used as in
show f nn 1 _0.4 1.1 10.7 42 132.7 384.6 1074.6 2955.8 8074.8 21995
to force each value to be displayed with at most one decimal place. This is perhaps not so useful here, but will be so later. [The preceding discussion springs from the formatting of the class handout.]
One of the values in the result list is negative. Notice that the
notation disambiguates the hyphen by representing the property of
negativity by the sign _ so that the hyphen can be kept
to stand for the negation and subtraction functions. To investigate
how f behaves for negative arguments, the test values can
have ten subtracted from them before applying f by keying
in
show f nn-10 31.4 28.3 25.1 22 18.9 15.7 12.6 9.4 6.4 3.5 1
where the values displayed have been rounded to one decimal place by
show. To make completely plain what the argument value is for each
result value, the laminate function ,: can be
used in
a ,: show f a =: nn-10 _10 _9 _8 _7 _6 _5 _4 _3 _2 _1 0 31.4 28.3 25.1 22 18.9 15.7 12.6 9.5 6.4 3.5 1
where the test values are temporarily named a so that
they can be stacked by ,: on top of the results. Care
must be taken in putting together expressions like the one keyed in
above. For example, the copula =: gives the name to its
left to its predicate, which is everything to its right, just as in
English sentences the predicate of a verb is everything to its right.
Providing there is a space between the f and the
a, then there are two names, and the copula only uses the
name next to it. If there’s no space, the name is fa and
there is no function f any more. Similarly, the predicate
of the laminate function is everything to its right, which is the
result of applying f to the values of a.
The same technique can be used to stack the values of nn
on the result of f nn by
nn ,: show f nn 0 1 2 3 4 5 6 7 8 9 10 1 _0.4 1.1 10.7 42 132.7 384.6 1074.6 2955.8 8074.8 21995
which shows quite different behaviour from that for negative argument values.
The most interesting feature here is that the function has two zeroes,
one between arguments of 0 and 1, and one
between 1 and 2. These zeroes can now be
explored one at a time, though the exploration of the second zero is
here left as an exercise for the reader.
The First Zero
The simplest zero to explore is the one between 0 and
1, and by
a ,: show f a=:nn%10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 0.8 0.6 0.4 0.2 0.1 _0.1 _0.2 _0.3 _0.4 _0.4
the eleven values between and including 0 and
1 are displayed. Because the
conventional divide character ÷ is not usually mapped to a key,
the expression nn%10
asks for nn to be divided by ten. Thus, the expression
nn%10 gives a list of values between 0 and
1 incrementing by tenths, and from the result of applying
f to nn%10 it can easily be seen that the
zero lies between argument values of 0.5 0.6, about
halfway between.
The limitations of the function show are evident here,
but it can easily be redefined as
show =: ".&(0.2&":)
which then returns a second decimal place to sight in
a ,: show f a=:nn%10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1 0.79 0.59 0.41 0.24 0.08 _0.06 _0.19 _0.29 _0.37 _0.42
so that the zero can be predicted to be slightly closer to
0.6 than to 0.5. Not only are fractional
values like 0.5 and 0.6 always displayed
with a leading zero, but they must also be keyed in with one. Hence
the next step in finding the zero, going down to hundredths, is
a ,: show f a=: 0.5+nn%100 0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59 0.6 0.08 0.06 0.05 0.03 0.02 0.01 _0.01 _0.02 _0.04 _0.05 _0.06
where the leading zero of the 0.5 of the expression being
keyed in is required, not optional. The new display of the results
shows that the zero is between argument values 0.55 and
0.56.
The result values are becoming rather too close to zero, and will now need to be scaled up, so in
a ,: show 1000*f a=: 0.55+nn%1000 0.55 0.551 0.552 0.553 0.554 0.555 0.556 0.557 0.558 0.559 0.56 5.38 3.97 2.56 1.16 _0.24 _1.64 _3.04 _4.44 _5.83 _7.23 _8.62
the results are multiplied by a thousand before being displayed, the
* symbol being used to stand for the multiplication
function
because the × symbol is not usually mapped to a key.
In the absence of parentheses, the
right argument of a function is the entire expression to the right of
the function symbol, so that the * in 1000*
takes the result of the application of f as its right
argument. By now, the argument values are getting a bit too long to
print out comfortably, and the way they correspond should be getting
to be clear, so they won’t be shown stacked any longer. The argument
of f now being from 0.55 to
0.56, and the zero being located between
0.553 and 0.554, the next narrowing down can
be done by
show 10000*f 0.553+nn%10000 11.6 10.2 8.79 7.39 5.99 4.58 3.18 1.78 0.38 _1.02 _2.42
which now places the zero between arguments 0.5538 and
0.5539. The J interpreter makes it easy enough to bring
down the previous line and simply insert the three extra digits
needed at each stage of the calculation. But by now the strings of
zero digits are becoming a bit hard to keep visual track of, so, using
e-notation for the large scaling integer and the small interval size,
the next step becomes
show 1e5*f 0.5538+nn%1e5 3.79 2.39 0.99 _0.42 _1.82 _3.22 _4.62 _6.02 _7.42 _8.83 _10.23
This process of narrowing down has now become somewhat repetitive. Of
course, the whole process could be made automatic, but by doing it
step by step the succession of numbers gives a better feel for how the
function f is behaving. Nevertheless, the repeated part
of the expression used just above could relatively easily be
encapsulated (by an experienced J user) as in
g =: ]*f@(+nn&%)
so that the investigation can now proceed as follows.
show 0.55382 g 1e6 9.86 8.46 7.06 5.66 4.26 2.85 1.45 0.05 _1.35 _2.75 _4.15 show 0.553827 g 1e7 0.51 _0.89 _2.29 _3.69 _5.09 _6.49 _7.9 _9.3 _10.7 _12.1 _13.5 show 0.5538270 g 1e8 5.14 3.73 2.33 0.93 _0.47 _1.87 _3.27 _4.68 _6.08 _7.48 _8.88 show 0.55382703 g 1e9 9.31 7.91 6.51 5.11 3.71 2.31 0.9 _0.5 _1.9 _3.3 _4.7
For some time now the results have been showing a strange uniformity. It might be suspected either that the arithmetic is beginning to give out, or that the variation between successive values has become uniform. This can be checked by taking the result values in pairs and subtracting them. This is done (without explanation here) by
show 2-/\0.55382703 g 1e9 1.4 1.4 1.4 1.4 1.4 1.4 1.4 1.4 1.4 1.4
which immediately shows the postulated uniformity. Indeed, keying in
the expression without the show gives ten values 1.40169
so it is very uniform. This could allow a fairly accurate
interpolation by hand of just where the zero is, but, for the purposes
of exploration, the estimation process is continued here as before,
except that now
show =: ".&(0.1&":)
because only one decimal place is now needed to be shown.
show 0.553827036 g 1e10 9 7.6 6.2 4.8 3.4 2 0.6 _0.8 _2.2 _3.6 _5 show 0.5538270366 g 1e11 6.2 4.8 3.4 2 0.6 _0.8 _2.2 _3.6 _5 _6.4 _7.8 show 0.55382703664 g 1e12 6.3 4.9 3.5 2.1 0.7 _0.7 _2.1 _3.5 _4.9 _6.3 _7.7 show 0.553827036644 g 1e13 7.2 5.8 4.4 3 1.6 0.2 _1.2 _2.6 _4 _5.4 _6.8 show 0.5538270366445 g 1e14 1.9 0.5 _0.9 _2.3 _3.7 _5.1 _6.5 _7.9 _9.3 _10.7 _12.1 show 0.55382703664451 g 1e15 4.9 3.3 2 0.7 _0.7 _2 _3.3 _4.7 _6.2 _7.5 _8.9 show 0.553827036644513 g 1e16 6.7 4.4 6.7 2.2 0 0 0 _2.2 _6.7 _2.2 _6.7
The values have become erratic, so much so that doubt must attach to any attempt to add further digits to the value found for the first zero. However, checking with
f 0.55382703664451 0.553827036644513 0.5538270366445135 4.88498e_15 6.66134e_16 0
suggests that at least the first fifteen decimal places should be reliable, provided the calculation built into the interpreter is reliable to such precision. It is important to realise that the barrier that has been encountered here is one of precision, not range, since
4*4e32 4e_32 1.6e33 1.6e_31
shows that a much greater range can be satisfactorily used.
Finding The Lowest Value
While the value of the second zero can be as easily found as the
first, between the two zeroes all the values of f are
negative, it might be presumed. This can be explored by
show 0.5 g 5 0.4 _0.9 _1.8 _2.3 _2.1 _1.2 0.7 3.6 7.8 13.7 21.6
where the interval between arguments is 0.2 and starting
from 0.5. This suggests a minimum value for
f of about _2.3%5 when the argument is about
1.1, probably just a bit larger than 1.1.
Before going further, it is convenient to redefine f as
its negation, thus
f =: -&(-o.)
so that the values being looked at will be free of the negative sign, though now what was lowest value will have become the highest. Consequently,
show 1.1 g 100 45.2 45.3 45.4 45.4 45.5 45.5 45.4 45.4 45.3 45.1 45
suggests that, while the now-highest value of f is
0.455, the argument corresponding to this value is
round about 1.14. But this is now estimation, and one
decimal place for show is too few.
The next steps in calculation are therefore
show =: ".&(0.2&":) show 1.14 g 1000 454.65 454.66 454.67 454.68 454.68 454.68 454.68 454.67 ...
which verifies the highest value to be about 0.45468 but
it is mainly a guess that the argument for that value is about
1.144. Issuing
show=:".&(0.3&":)
to get three decimal places then allows
show _45468+10*1.144 g 10000 0.151 0.172 0.191 0.206 0.218 0.226 0.232 0.234 0.234 0.23 0.223
where the _45468+10* reduces the size of numbers
displayed and gets an extra decimal place at the same time. The
procedure can be continued, thus
show _4546823+100*1.1447 g 100000 0.447 0.455 0.46 0.461 0.46 0.455 0.447 0.436 0.422 0.405 0.384
which adds an extra digit (a bit uncertainly – is it 2,
or 3?) to the value of the argument, and two extra digits
to the value of the function. The procedure goes on, here splitting
the difference between the 2 and the 3.
show _454682346+1000*1.144725 g 1000000 0.102 0.116 0.126 0.134 0.138 0.139 0.137 0.132 0.124 0.113 0.098
which confirms 2, and its successor 9, and
again adds a couple of digits, 13, maybe even
138, to the value of the function. Now
show _454682346138+100000*1.144729 g 1e7 0.132 0.395 0.626 0.825 0.994 1.131 1.236 1.311 1.353 1.364 1.344
shows that the last digit of the previous function value should have
been 9 rather than 8, but still gives
another two function-value digits for the single argument digit. In
the next step
show _45468234613935+1e6*1.1447298 g 1e8 0.273 0.547 0.766 0.984 1.117 1.297 1.344 1.391 1.43 1.477 1.477
leads immediately to
show _4546823461393647+1e7*1.14472989 g 1e9 10 5 _4 5 1 1 _4 1 1 _4
where everything now blows up at a point where the function value is known to about 16 digits, but the argument value to only 9.
This mismatch between the accuracy of the result and that of its argument is a good illustration of the inherent unreliability of simple computer arithmetic. More extreme examples of mismatch could easily be concocted.
Conclusion
The procedures described in this article show the way, or the style, in which calculations can be carried out interactively using a tool like J. If the calculation tool is mastered, most calculations can be carried out in the same style. The barrier is attaining that mastery.
Indeed the repeated searching for an answer by a kind of trial and error is not characteristic of functional calculation, since mastery will enable functions to be defined which can do the searching automatically. Such virtuosity would not be appropriately demonstrated in an apéritif.
The interactive style of calculation depends on two factors illustrated in this article. The first is simply the ability to key in an expression and have it evaluated ‘on the fly’. The second is the ability to give a name to an expression then use that name from then on as though it were that expression.
This article is only an apéritif to the main meal of getting to grips with a ramified functional calculation tool.
References
- Menninger, K. (1958) Zahlwort und Ziffer, Vandenhoeck & Ruprecht, Gottingen, 2nd edition (as Number Words and Number Symbols by MIT Press in 1969).
- Thimbleby, H.W. (1995) “A New Calculator and Why it is Necessary”, The Computer Journal, 38(6), 418–433.
- Jsoftware: http://www.jsoftware.com
- The J Wiki: http://www.jsoftware.com/jwiki