

last edited 3 years ago by test1 
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15  
Editor: test1
Time: 2015/02/23 17:25:57 GMT+0 

Note: 
removed:  Important Note   There is a bug (#8) in the version of Axiom currently running on this server that messes up the output by missing some parenthesis. A preliminary  though a little unsatisfactory  patch is available. We hope that a proper fix will soon be applied. changed: Please add other bugs you find to IssueTracker by clicking on "Bug reports":http://wiki.axiomdeveloper.org/IssueTracker on the top left of any page and filling out the appropriate forms. If you find a bug please report in FriCAS BugTracker. changed: this program at your own computer, you need the FriCAS version of axiom. this program at your own computer, you need to install FriCAS. removed:  Fortunately, with the right tool, it is a matter of a moment to figure out formulas for all of these sequences. In this article we describe a computer program that encompasses well known techniques and adds new ideas that we hope to be very effective.
Author: Martin Rubey
A more thorough discussion of this package can be found at http://arxiv.org/abs/math.CO/0702086
)set output tex off
)set output algebra on
If you find a bug please report in FriCAS BugTracker.
Finally, please feel free to try this package in the SandBox! If you would like to use this program at your own computer, you need to install FriCAS.
If you find the package useful, please let me know!
Abstract
We present a software package that guesses formulas for sequences of, for
example, rational numbers or rational functions, given the first few terms.
Thereby we extend and complement Christian Krattenthaler's program Rate
and
the relevant parts of Bruno Salvy and Paul Zimmermann's GFUN
.
This research was partially supported by the Austrian Science Foundation FWF, grant S8302MAT.
For some a brainteaser, for others one step in proving their next theorem: given the first few terms of a sequence of, say, integers, what is the next term, what is the general formula? Of course, no unique solution exists, however, by Occam's razor, we will prefer a "simple" formula over a more "complicated" one.
Some sequences are very easy to "guess", like
(1) 
(2) 
Others are a little harder, for example
(3) 
Of course, at times we might want to guess a formula for a sequence of polynomials, too:
(4) 
(5) 
Fortunately, with the right tool, it is a matter of a moment to figure out
formulas for all of these sequences. In this article we describe a computer
program that encompasses well known techniques and adds new ideas that we hope
to be very effective. In particular, we generalize both Christian
Krattenthaler's program Rate
, and the guessing functions present
in GFUN
written by Bruno Salvy and Paul Zimmermann. With a little
manual aid, we can guess multivariate formulas as well, along the lines of
Doron Zeilberger's programs GuessRat
and GuessHolo
.
We would also like to mention The online encyclopedia of integer sequences of Neil Sloane. There, you can enter a sequence of integers and chances are good that the website will respond with one or more likely matches. However, the approach taken is quite different from ours: the encyclopedia keeps a list of currently sequences, entered more or less manually, and it compares the given sequence with each one of those. Besides that, it tries some simple transformations on the given sequence to find a match. Furthermore it tries some simple programs we will describe below to find a formula, although with a time limit, i.e., it gives up when too much time has elapsed.
Thus, the two approaches complement each other: For example, there are sequences where no simple formula is likely to exist, and which can thus be found only in the encyclopedia. On the other hand, there are many sequences that have not yet found their way into the encyclopedia, but can be guessed in a few minutes by your computer.
On the historical side, we remark that already in 1966 Paul W. Abrahams implemented a program to identify sequences given their first few terms...
A formula for Sequence (1) is almost trivial to guess: it seems obvious that it is . More generally, if we believe that the sequence in question is generated by a polynomial, we can simply apply interpolation. However, how can we know that a polynomial formula is appropriate? The answer is quite simple: we use all but the last few terms of the sequence to derive the formula. After this, the last terms are compared with the values predicted by the polynomial. If they coincide, we can be confident that the guessed formula is correct. We call the number of terms used for checking the formula the safety of the result.
Apart from safety, the main problem we have to solve is about efficiency. For example, maybe we would like to test whether the term of the sequence is given by a formula of the form
(6) 
Thus, we need to find efficient algorithms that test for large classes of formulas. Obviously, such algorithms exist for interpolation and Pade approximation. For the present package, we implemented an efficient algorithm for a far reaching generalization of interpolation, proposed by Bernhard Beckermann and George Labahn, see FractionFreeFastGaussianElimination. Furthermore, we show that there is also a way to guess sequences generated by Formula (6).
Using these algorithms our package clearly outperforms both Rate
and GFUN
,
in terms of speed as well as in the range of formulas that can be guessed.
In the following section we outline the capabilities of our package. In the Section therafter we describe the most important options that modify the behaviour of the functions.
In this section we briefly present the function classes which are covered by our package. Throughout this section, is the function we would like to guess, and is its generating function. The values are supposed to be elements of some field , usually the field of rationals or rational functions. We alert the reader that the first value in the given sequence always corresponds to the value .
guessRec
finds recurrences of the form
(7) 
guessRec([1,1, 0, 1,  1, 2,  1, 5,  4, 29,  13, 854,  685])
2 (1) [[f(n): f(n + 2) + f(n + 1)  f(n) = 0,f(0) = 1, f(1) = 1]]
Note that, at least in the current implementation, we do not exclude solutions that do not determine the function completely. For example, given a list containing only zeros and ones, one result will be
guessPRec
only looks for recurrences with linear , i.e., it
recognizes Precursive sequences. As an example,
guessPRec([0,1, 0, 1/6, 0, 1/120, 0, 1/5040, 0, 1/362880, 0, 1/39916800])
(2) []
Concerning analogues, guessRec(q)
finds recurrences of the
form (7), where is a polynomial with coefficients in .
Similarly, we provide analogues for guessPRec
and
guessRat
. Finally, guessExpRat(q)
recognizes functions of the form
For Sequence (5), we enter
guessExpRat(q)([(12*q)/(1q),12*q, (1q)*(12*q)^3, (1q)^2*(12*q)*(12*q2*q^2)^3], [])
n n (2 q  1)(2 q  3 q + 1) (4) [] q  1
guessADE
finds an algebraic differential equation for ,
i.e., an equation of the form
(8) 
guessADE([1,1, 2, 9/2, 32/3, 625/24, 324/5, 117649/720, 131072/315])
(5) []
guessHolo
only looks for equations of the form (11) with
linear , that is, it recognizes holonomic or differentiallyfinite
functions. It is well known that the class of holonomic functions coincides
with the class of functions having Precursive Taylor coefficients. However,
the number of terms necessary to find the differential equation often differs
greatly from the number of terms necessary to find the recurrence. Returning
to the example given for guessPRec
, we find that already the first 6 terms
are sufficient to guess a generating function:
guessHolo([0,1, 0, 1/6, 0, 1/120])
3 n ,, x 4 (6) [[[x ]f(x): f (x) + f(x) = 0, f(x) = x   + O(x )]] 6
Moreover, now we immediately recognise the coefficients as being those of the
sine function.
guessHolo
is also the function provided by GFUN
. Here is a comparison
of average running times in seconds over several runs on the same machine for
a list of elements
guessAlg
looks for an algebraic equation satisfied by ,
i.e., an equation of the form
first guessAlg [1,1, 2, 5, 14, 42]
n 2 2 3 4 (7) [[x ]f(x): x f(x)  f(x) + 1 = 0,f(x) = 1 + x + 2 x + 5 x + O(x )]
guessPade
recognises rational generating functions. For the
Fibonacci sequence given in Equation (2), we find as likely
solution
We provide analogues, replacing differentiation with dilation:
guessADE(q)
finds differential equations of the form
(9) 
guessHolo
, guessAlg
, and guessPade
.
To guess a formula for Sequence (4), we enter
guessRat(q)([1,1+q+q^2, (1+q+q^2)*(1+q^2), (1+q^2)*(1+q+q^2+q^3+q^4)], [])
3 2 n 2 n q q + ( q  q)q + 1 (8) [] 3 2 q  q  q + 1
The main observation made by Christian Krattenthaler in designing his program
Rate
is the following: it occurs frequently that although a sequence of numbers
is not generated by a rational function, the sequence of successive quotients is.
We slightly extend upon this idea, and apply recursively one or both of the two following operators:
guessSum
 the differencing operator, transforming
into .guessProduct
 the operator that transforms into
.For example, to guess a formula for Sequence (3), we enter
guess([0,1, 3, 9, 33], [guessRat], [guessSum, guessProduct])
s  1 n  1 5 + ++++ (9) [ >   p + 2] +   4 s = 0 p = 0 5 4
The second argument to guess
indicates which of the functions of the
previous section to apply to each of the generated sequence, while the third
argument indicates which operators to use to generate new sequences.
In the case where only the operator is applied, our package is directly
comparable to Rate
. In this case the standard example is the number of
alternating sign matrices
guess([1,1, 2, 7, 42, 429, 7436, 218348], [guessRat], [guessProduct])
p  1 2 n  1 8 27 p + 54 p + 24 ++++ ++++ 7 7 (10) [     ]     2 p = 0 p = 0 16 p + 32 p + 12 8 7 7 7
Here are the average running times in seconds for our package and Rate
over
several runs on the same machine for a list of elements:
: & 14 & 15 & 16 & 17 & 18\ Rate: & 1.0 & 3.3 & 29.7 & 44.9 & 398\ Guess: & 0.9 & 2.3 & 6.6 & 22.4 & 74 " title="{lrrrrrrrrrr} $n$: & 14 & 15 & 16 & 17 & 18\ Rate: & 1.0 & 3.3 & 29.7 & 44.9 & 398\ Guess: & 0.9 & 2.3 & 6.6 & 22.4 & 74 " class="equation" src="images/225703948598163504516.0px.png" align="bottom" Style="verticalalign:textbottom" width="204" height="52"/> 
To give you the maximum flexibility in guessing a formula for your favourite sequence, we provide options that modify the behaviour of the functions as described in Section~\ref{sec:functionclasses}. The options are appended, separated by commas, to the guessing function in the form \spad{option==value}. See below for some examples.
debug
specifies whether informations about progress should be
reported.safety
specifies, as explained at the beginning of
Section 2, the number of values reserved for testing any
solutions found. The default setting is 1. Experiments seem to indicate that for guessADE
higher settings are
appropriate than for guessRat
. I.e., if a rational function
interpolates the given list of terms, where the final term is used for
testing, we can be pretty sure that the formula found is correct. By
contrast, we recommend setting safety
to 3 or 4 when using
guessADE
. For all algorithms except guessExpRat
we recommend to
omit trailing zeros.
one
specifies whether the guessing function should return as soon
as at least one solution is found. By default, this option is set to true
.maxDegree
specifies the maximum degree of the coefficient
polynomials in an algebraic differential equation or a recursion with
polynomial coefficients. For rational functions with an exponential term,
maxDegree
bounds the degree of the denominator polynomial. This option
is especially interesting if trying rather long sequences where it is unclear
whether a solution will be found or not. Setting maxDegree
to 1, which is the default, specifies that the
maximum degree can be arbitrary.allDegrees
specifies whether all possibilities of the degree
vector  taking into account maxDegree
 should be tried. The
default is true
for guessPade
and guessRat
and false
for all other functions.homogeneous
specifies whether the search space should be
restricted to homogeneous algebraic differential equations or homogeneous
recurrences. By default, it is set to false
.maxDerivative
 maxShift
specify the maximum derivative in
an algebraic differential equation, or, in a recurrence relation, the maximum
shift. Setting the option to 1 specifies that the maximum derivative 
the maximum shift  may be arbitrary.maxPower
specifies the maximum total degree in an algebraic
differential equation or recurrence: for example, the degree of
is 4. Setting the option to 1 specifies that the maximum total degree
may be arbitrary. For example,
l := [1,1, 1, 1, 2, 3, 7, 23, 59, 314, 1529, 8209, 83313, 620297, 7869898, 126742987, 1687054711, 47301104551, 1123424582771, 32606721084786, 1662315215971057];
guessRec(l,maxPower==2)
(12) [ 2 [f(n):  f(n)f(n + 4) + f(n + 1)f(n + 3) + f(n + 2) = 0,f(0) = 1, f(1) = 1, f(2) = 1, f(3) = 1] ]
returns the Somos4 recurrence, whereas without limiting the power to 2, we need the first 33 values, and instead of roughly one second half a minute of computing time.
maxLevel
specifies how many levels of recursion are tried when
applying operators. Note that, applying either of the two operators results in a sequence which is by one
shorter than the original sequence. Therefore, in case both guessSum
and guessProduct
are specified, the number of times a guessing
algorithm from the given list of functions is applied is roughly , where
$n$ is the number of terms in the given sequence. Thus, especially when the
list of terms is long, it is important to set maxLevel
to a low value.Still, the default value is 1, which means that the number of levels is only restricted by the number of terms given in the sequence.
indexName
, variableName
, functionName
specify
symbols to be used for the output. The defaults are n
, x
and f
respectively.The output of any function described in Section 3 is a list of formulae which seem to fit, along with an integer that states from which term on the formula is correct. The latter is necessary, because rational interpolation features sometimes unattainable points, as the following example shows:
guessRat([3,4, 7/2, 18/5, 11/3, 26/7])
2 2 (13) [[f(n): ( n  n + 2)f(n) + 4 n + 2 n  6 = 0]]
indicates that the first two terms of the sequence might not coincide with the value predicted by the returned function. A similar situation occurs, if the function generating the sequence has a singular point at , where and is the number of given values. We would like to stress that this is rather a feature than a bug: most terms will be correct, just as in the example above, where the value at is indeed 3.
)version
Value = "FriCAS 1.3.4 compiled at Wed Jun 27 14:53:01 UTC 2018"