The problem of integration in finite terms

1. The Freshman Calculus Problem

2. The Artificial Intelligence Problem
Symbolic Integration 3. The Algebraic (Rational Function) Problem
4. The Decision Problem (Liouville-Risch...)
5. The Computer Algebra System problem.
Lecture 14
6. Extensions to definite integration, numerical
integration, approximation by Taylor series,
Laurent series, asymptotic series.

Richard Fateman CS 282 Lecture 14 1 Richard Fateman CS 282 Lecture 14 2

The Freshman Calculus Problem. Find the Freshman Calc Æ AI: James Slagle’s
flaws. approach
1. Humans are intelligent • AND-OR trees for evaluation of difficulty of
2. MIT freshman are especially intelligent humans. a path (do we abandon this route or plow on?)
3. Freshman study integral calculus (this is in 1960’s
before Advanced Placement).
4. Therefore solving freshman calculus problems or or
requires intelligence.
5. If a computer can solve freshman calculus problems, and
then it is intelligent (An example of Artificial
Intelligence).
6. To solve these problems, imitate freshman students.
7. Future work: do the rest of math, and the rest of • Game playing: could be any complex task.
AI.
Richard Fateman CS 282 Lecture 14 3 Richard Fateman CS 282 Lecture 14 4

Slagle’s program SAINT Symbolic Automatic

The Algebraic (Rational Function) Problem
INTegrator
• ELINST: elementary instance expression • Many attempts to do this right, e.g. Theses of
pattern matching (in assembler). R.Tobey, E. Horowitz, and with extensions, M.
• Simplification program (in assembler). Rothstein, B. Trager.
• Used lisp prefix expressions (* x (sin x)) • Main idea is to take a rational expression
• First major lisp program (in Lisp 1.5). q(x)/r(x), express it as a polynomial + “proper”
• Took about a minute per problem (about the fraction P+Q/R with deg(Q)<= deg(R).
same time as a student!) running uncompiled on
Integrate P, a polynomial, trivially.
an IBM 7090 (32k word x 36 bits/word).
• Could not do rational function integration (out Expand Q/R in partial fractions.
of space.)
Richard Fateman CS 282 Lecture 14 5 Richard Fateman CS 282 Lecture 14 6

1
Partial fractions “trivialized” an example of the general cubic...

• From Q/R Find all the complex roots of R,

a0, ...,an.
Express Q/R as ∑ ci/(x-ai) which integrates to
∑ ci log(x-ai).

Can we always find {ai}? Complex roots can be

found numerically. Can we find ci?
It turns out that we can expand the summation
and solve everywhere... with a big problem..
Richard Fateman CS 282 Lecture 14 7 Richard Fateman CS 282 Lecture 14 8

The answer is unmanageable even if we can

What’s the problem?
compute it.

• If the denominator R= s3*x3+s2*x2+s1*x+s0, • For the cubic, each ci looks something like this
then the each of the roots ai look like this:

• and so it is ugly. Often enough these expressions can

be simplified, and so the search has been to find a
minimal algebraic extension in which the ci can be
expressed. We DON’T need to see the ai, but can use
facts like a1a2a3 is constant term of polynomial R; we
have neat forms for all symmetric functions of roots.

Richard Fateman CS 282 Lecture 14 9 Richard Fateman CS 282 Lecture 14 10

This trick works for numeric denominators

For a general quartic, it gets worse.
too
• And for order 5 or more there is in general no
formula “in radicals” for the roots. That is, we
can’t compute the ai in general for symbolic
polynomials
• So we can find approximate numerical roots in
C if the problem has a denominator solely in
Q[x], all bets are off “symbolically”.
• The solution is “RootSum(f(x),poly,x)”
expressions where we have ∑ f(ai) where {ai}
are the roots of the polynomial poly.

Richard Fateman CS 282 Lecture 14 11 Richard Fateman CS 282 Lecture 14 12

2
Rothstein (etal), find the ci easier Squarefree isn’t a problem

• Let B(x) be square free and of degree >= A(x), • Let the problem be ∫A(x)/B(x)n with deg(A)<=deg(B).
• Since B itself is square free, its GCD with B’=dB/dx is
• then ∫ (A/B)dx = ci log(vi) 1. Thus we can compute c,e such that c*B+e*B’=1.
• where ci, are the distinct roots of the • Now multiply through by A to get cAB+eB’A=A, and
polynomial R(c)=Resultant(x,A(x)-cB’(x),B(x)) substitute in the original ∫A(x)/B(x)n to get
∫(cAB+eB’A)/B(x)n . Now dividing through we get
and vi=gcd(A(x)-ciB’(x),B(x)) for i = 1,...n.
∫ cA/Bn-1 + ∫ eA* (B’/Bn) where the second term can be
integrated by parts to lower the power of Bn in the
• We need Square-free computation and denominator.
(reminder: ∫ u dv = uv-∫ v du. Let u =eA, dv=B’/Bn,
Resultant wrt x
v=1/((1-n)Bn-1), du=(eA)’, so ∫ v du comes out as
<something>/Bn-1+ ∫ <something>/Bn-1 . Iterate until Bn =
B1 .)
Richard Fateman CS 282 Lecture 14 13 Richard Fateman CS 282 Lecture 14 14

Resultant isn’t a problem The decision problem

• Well, actually it is something we’d prefer not • Liouville [1809-1882] During the period (1833-1841)
presented a theory of integration; proved elliptic
to do since it takes a while, but we have integrals cannot have elementary expressions.
algorithms for it. • Various other writers advanced the subject in late
1800’s
• J. Ritt (1948) Integration in Finite Terms Columbia
• Conclusion: we can do rational function Univ. Press
integration pretty well. But it took us into the • M. Rosenlicht (AMM. 1972) Integration in Finite
mid-1980s to do it right. Terms

• R. Risch (1968) [unreadable]

• B. Trager, J. Davenport, M.Bronstein [implementation]
Richard Fateman CS 282 Lecture 14 15 Richard Fateman CS 282 Lecture 14 16

Risch’s result What’s a Differential Field?

Theorem: Let K be a differential field and f be from K .

Then an elementary extension of the field K which has the • Equip a field with an additional operator D
same field of constants as K and contains an element g which satisfies identities parallel to the rules
such that g’ = f exists if and only if there exist constants
c1..cn from K and functions u0,..,un from K such that for differentiation of functions. These
f= u0’+ ∑i=1,...,nci(ui’/ui)
structures obviously include various fields of
or
functions (e.g. the field of rational functions
in one variable over the real field).
g=∫ fdx = u0+ ∑ i=1..,n ci log(ui)
• The goal is to provide a setting to answer
concretely such questions as, "Does this
note: this allows additional logs in the answer, but that’s
all. The structure of the integral is specified. function have an elementary antiderivative?”
Richard Fateman CS 282 Lecture 14 17 Richard Fateman CS 282 Lecture 14 18

3
Formally, start with a field F and a
Also
derivation map

• Derivation is a map of F into itself • If a, b are in a differential field F with a

a # a’ such that being nonzero, we call a an exponential of b or
(a+b)’ = a’ + b’ b a logarithm of a if b’ = a’/a.
(ab)’=a’b+ab’
CONSEQUENTLY • For algebraic extensions of F, there is a
theorem that says that F is of characteristic
(a/b)’= (a’b-ab’)/b2 if a,b, ∈ F and b≠ 0 zero and K is an algebraic extension field of F
(an)’ = nan-1a’ for all integers n then the derivation on F can be extended
1’ =0; the constants F are a subfield of F uniquely to K.

Richard Fateman CS 282 Lecture 14 19 Richard Fateman CS 282 Lecture 14 20

We must also deal with algebraic pieces examples

• sometimes we prefer arctan (if a<0 below), sometimes

logs.
• Algebraic extensions z where p(z)=0 in F
• Monomial extensions exp() and log().
Independent extensions are required e.g.
log(x) and log(x2) don’t cut it since (log(x2))- • Risch converts ∫ xnsin x to exponentials
(2log(x)) is a constant (perhaps 0)

• and then can’t do it..

Richard Fateman CS 282 Lecture 14 21 Richard Fateman CS 282 Lecture 14 22

Extensions of the Risch algorithm The Risch Algorithm: two steps back.

• Allow certain functions beyond log, exp such • There is a fallacy in claiming the Risch "algorithm" is
an algorithm at all: it depends, for solution of
that the form of the integral still holds. For subproblems, on heuristics to tell if certain
example, erf(x) = (2/π)∫0x exp(-t2)dt has the expressions are equivalent to zero.
appropriate structure. • We have gone over Richardson’s arguments to show
that the zero-equivalence problem over a class of
expressions much smaller than that of interest for
integration is recursively unsolvable.
• But that is not the problem with most
implementations of the Risch algorithm. They mostly
have not been programmed completely, because, even
assuming you can solve the zero-equivalence problem,
the procedure is hard to program.
Richard Fateman CS 282 Lecture 14 23 Richard Fateman CS 282 Lecture 14 24

4
The Risch Algorithm: answers are not The Risch Algorithm: answering wrong
ncessarily what you expect problem
• It is not nearly as useful as you might think, because • Also, most people are interested in definite,
it returns algebraic antiderivatives whose validity not indefinite integrals, at least once they've
may be on a set of measure zero. Work by D. finished with Freshman calculus.
Jeffrey A. Rich (among others) on removing
• And in cases where approximate solutions are
gratuitous discontinuities, is helpful.
easily obtained for the corresponding
• The Risch algorithm may also, in the vast majority of definite integral, the Risch algorithm may
problems, simply say, after an impressive pause, grind on for a long while and then say there
nope. can't do it. There is no reasonable complexity
is no closed form. Or if there is a closed
analysis for the process, which is probably why
form, if it is to be ultimately evaluated
certain authors ignore it.
numerically, the closed form may be less
useful than the quadrature formula!
Richard Fateman CS 282 Lecture 14 25 Richard Fateman CS 282 Lecture 14 26

Nevertheless, the Risch Algorithm fascinates

It also eludes us.
us
• The theory of algebraic integration, and its It is in some sense at the pinnacle of CAS
corresponding history is interesting. algorithms: it uses more mechanisms of a
• The “solution” to calculus problems can probably be computer algebra system than nearly any
presented this way. other program.
• It’s a good advertising slogan. – Rational function manipulation (GCD, partial
• The definite integration problem is also a classic in fractions, factoring)
the applied mathematics literature: – Simplification in a differential field (algebraic,
exponential/log, complex extensions)
• It would seem that vast tables (20,000 entries)
– Solving certain ODEs
could be replaced with a computer program. (Actually
Risch alg. is almost irrelevant here) – J. Moses, J. Davenport, B. Trager, D. Lazard, R.
Rioboo, A. Norman, G. Cherry, M. Bronstein,

Richard Fateman CS 282 Lecture 14 27 Richard Fateman CS 282 Lecture 14 28

How should one write an integration

Go over examples of Risch in Moses’ paper
program? (Moses)

• Detailed understanding of Risch’s original • Quick solution to easy problems

presentation is a challenge. • A collection of methods and transformations
• Radical methods
• (less known, table lookup)
• consider approximate or numerical approaches
expand integrand in taylor series, orthogonal polynomials,
fourier series, or asymptotic series or even approximate as
rational functions.
(for rational functions), approximate the roots of the
denominator and do partial fraction expansion
do the whole task by quadrature (a well-studied area) treating
the integrand as a “black box” capable only of returning a
value at a point. [difficulties with infinite range or nasty
behaviors.
Richard Fateman CS 282 Lecture 14 29 Richard Fateman CS 282 Lecture 14 30

5
Losing numerically Losing symbolically

• Try integrating x*sin(x) between 0 and 5000. • Sometimes the answer is possible but ugly. Consider
∫1/(z64+1) dz integrated to
• Numerically, it’s tough.
• Symbolically it is –5000 cos (5000)+sin(5000)

• Computer algebra systems get forms that are worse

than this (try them!) and often contribute other
monstrosities more easily integrated numerically

Richard Fateman CS 282 Lecture 14 31 Richard Fateman CS 282 Lecture 14 32