Sei sulla pagina 1di 16

Closed Forms: What They

Are and Why We Care


Jonathan M. Borwein and Richard E. Crandall

Closed Forms: What They Are not be considered closed-form. However,


Mathematics abounds in terms that are in frequent the choice of what to call closed-form
use yet are rarely made precise. Two such are rigor- and what not is rather arbitrary since a
ous proof and closed form (absent the technical use new closed-form function could simply
within differential algebra). If a rigorous proof is be defined in terms of the infinite sum.
that which convinces the appropriate audience, Eric Weisstein
then a closed form is that which looks funda- There is not much to disagree with in this, but
mental to the requisite consumer. In both cases, it is far from rigorous.
this is a community-varying and epoch-dependent
notion. What was a compelling proof in 1810 may Second Approach. The next attempt follows a
well not be now; what is a fine closed form in 2010 September 16, 1997, question to the long-operating
may have been anathema a century ago. In this Dr. Math site1 and is a good model of what
article we are intentionally informal as befits a interested students are likely to be told.
topic that intrinsically has no one right answer.
Subject: Closed form solutions
Let us begin by sampling the Web for various
Dear Dr. Math: What is the exact mathe-
approaches to informal definitions of closed
matical definition of a closed form solution?
form.
Is a solution in closed form simply if an
expression relating all of the variables can
Definitions
be derived for a problem solution, as op-
First Approach to a Definition of Closed Form. The posed to some higher-level problems where
first comes from MathWorld [56] and so may well there is either no solution, or the problem
be the first and last definition a student or other can only be solved incrementally or numeri-
seeker-after-easy-truth finds. cally?
An equation is said to be a closed-form Sincerely, .
solution if it solves a given problem in terms The answer followed on September 22:
of functions and mathematical operations
This is a very good question! This matter
from a given generally accepted set. For
has been debated by mathematicians for
example, an infinite sum would generally
some time, but without a good resolution.
Jonathan M. Borwein is professor of mathematics at the Some formulas are agreed by all to be
University of Newcastle. His email address is jonathan. in closed form. Those are the ones which
borwein@newcastle.edu.au. contain only a finite number of symbols,
Richard E. Crandall is professor of Mathematics at Reed and include only the operators +, , , /,
College. His email address is crandall@reed.edu.
1
DOI: http://dx.doi.org/10.1090/noti936 Available at http://mathforum.org/dr/math/.

50 Notices of the AMS Volume 60, Number 1


and a small list of commonly occurring Tim Chow explains nicely why he eschews
functions such as nth roots, exponentials, capturing all algebraic numbers in his definition,
logarithms, trigonometric functions, inverse why he wishes only elementary quantities to have
trigonometric functions, greatest integer closed forms, whence he prefers E to Ritts 1948
functions, factorials, and the like. definition of elementary numbers as the smallest
More controversial would be formulas algebraically closed subfield L of C that is closed
that include infinite summations or prod- under exp and log. His reasons include that:
ucts, or more exotic functions, such as Intuitively, closed-form implies explicit,
the Riemann zeta function, functions ex- and most algebraic functions have no simple
pressed as integrals of other functions explicit expression.
that cannot be performed symbolically,
Assuming Shanuels conjecture that given n
functions that are solutions of differential complex numbers z1 , . . . , zn which are linearly in-
equations (such as Bessel functions or hy- dependent over the rational numbers, the extension
pergeometric functions), or some functions field
defined recursively. Some functions whose
Q (z1 , . . . , zn , ez1 , . . . , ezn )
values are impossible to compute at some
specific points would probably be agreed has transcendence degree of at least n over the
not to be in closed form (example: f (x) = 0 rationals, then the algebraic members of E are
if x is an algebraic number, but f (x) = 1 if exactly those solvable in radicals [31]. We may
x is transcendental. For most numbers, we thence think of Chows class as the smallest
do not know if they are transcendental or plausible class of closed forms. Only a mad version
not). I hope this is what you wanted. of Markov would want to further circumscribe the
class.
No more formal but representative of many
dictionary definitions is:
Special Functions

Third Approach. A coauthor of the current article In an increasingly computational world, an ex-
is at least in part responsible for the following brief plicit/implicit dichotomy is occasionally useful,
definition from a recent mathematics dictionary but not very frequently. Often we will prefer com-
[17]: putationally the numerical implicit value of an
algebraic number to its explicit tower of radicals,
closed form n. an expression for a given and it seems increasingly perverse to distinguish
function or quantity, especially an integral, the root of 2x5 10x + 5 from that of 2x4 10x + 5
in terms of known and well understood or to prefer arctan( /7) to arctan(1). We illustrate
quantities, such as the evaluation of these issues further in Examples 7, 9, and 14.
Z
We would prefer to view all values of classical
exp(x2 ) dx
special functions of mathematical physics [54] at
algebraic arguments as being closed forms. Again
as .
there is no generally accepted class of special
Collins Dictionary functions, but most practitioners would agree that
And of course one cares more for a closed form the solutions to the classical second-order algebraic
when the object under study is important, such as differential equations (linear or, say, Painlev) are
when it engages the normal distribution as above. included. But so are various hypertranscendental
With that selection recorded, let us turn to some functions, such as , B, and , which do not arise
more formal proposals. in that way.2
Hence, we do not wish to accept any definition
Fourth Approach. Various notions of elementary of special function which relies on the underlying
numbers have been proposed. functions satisfying natural differential equations.
Definition [31]. A subfield F of C is closed The class must be extensible; new special functions
under exp and log if (1) exp(x) F for all are always being discovered.
x F and (2) log(x) F for all nonzero A recent American Mathematical Monthly re-
x F, where log is the branch of the view3 of [47] says:
natural logarithm function such that <
2
Im(log x) for all x. The field E of EL Of course, a value of a hypertranscendental function
at algebraic argument may be very well behaved; see
numbers is the intersection of all subfields
Example 4.
of C that are closed under exp and log. 3
Available at http://www.maa.org/maa%20reviews/
Tim Chow 4221.html.

January 2013 Notices of the AMS 51


Theres no rigorous definition of special
functions, but the following definition is in
line with the general consensus: functions
that are commonly used in applications have
many nice properties and are not typically
available on a calculator. Obviously this is
a sloppy definition, and yet it works fairly
well in practice. Most people would agree,
for example, that the Gamma function is
included in the list of special functions and
that trigonometric functions are not.
Once again, there is much to agree with and
much to quibble about in this reprise. That said,
most serious books on the topic are little more spe-
cific. For us, special functions are nonelementary
(a) modulus of W
functions about which a significant literature has
developed because of their importance in either
mathematical theory or in practice. We certainly
prefer that this literature include the existence of
excellent algorithms for their computation. This is
all consonant withif somewhat more ecumenical
thanTemmes description in the preface of his
lovely book [54, Preface, p. xi]:
[W]e call a function special when the func-
tion, just like the logarithm, the exponential
and trigonometric functions (the elemen-
tary transcendental functions), belongs to
the toolbox of the applied mathematician,
the physicist or engineer.
Even more economically, Andrews, Askey, and
Roy start the preface to their important book (b) W on real line
Special Functions [1] by writing:
Figure 1. The Lambert W function.
Paul Turan once remarked that special
functions would be more appropriately
labeled useful functions. We consider this to be a splendid closed form
With little more ado, they then start to work on even though, again assuming Shanuels conjecture,
the Gamma and Beta functions; indeed, the term W (1) 6 E [31]. Additionally, it has only recently
special function is not in their index. Near the rigorously been proven that W is not an elementary
end of their preface, they also write: function in Liouvilles precise sense [26]. We also
[W]e suggest that the day of formulas may note that successful simplification in a modern
be experiencing a new dawn. CAS [29] requires a great deal of knowledge of
special functions. 
Example 1 (Lamberts W). The Lambert W func-
tion, W (x), is defined by appropriate solution of Further Approaches
y ey = x [21, pp. 277279]. This function has been Fifth Approach. PlanetMaths offering, as of Feb-
implemented in computer algebra systems (CAS) ruary 15, 2010,4 is certainly in the elementary
and has many uses despite being unknown to most number corner.
scientists and only relatively recently named [41]. expressible in closed form (Definition) An
It is now embedded as a primitive in Maple and expression is expressible in a closed form
Mathematica with the same status as any other if it can be converted (simplified) into
well-studied special or elementary function. (See, an expression containing only elementary
for example, the tome [26].) The CAS know its functions, combined by a finite amount of
power series and much more. For instance, in Maple rational operations and compositions.
entering:
PlanetMath
> fsolve(exp(x)*x=1);identify(%);
returns 4
Available at http://planetmath.org/encyclopedia/
0.5671432904, LambertW(1) ClosedForm4.html.

52 Notices of the AMS Volume 60, Number 1


This reflects both much of what is best and Now we are prepared to posit
what is worst about the mathematical wisdom Definition [7]. The ring of hyperclosure H is
of crowds. For the reasons adduced above, we the smallest subring of C containing the set
wish to distinguishbut admit boththose closed X. Elements of H are deemed hyperclosed.
forms that give analytic insight from those which In other words, the ring H is generated by all
are sufficient and prerequisite for effective com- general hypergeometric evaluations under the , +
putation. Our own current preferred class [7] is operators, all symbolized by
described next.
H = hXi,+ .
Sixth Approach. We wish to establish a set X of
H will contain a great many interesting closed
generalized hypergeometric evaluations; see [7] for
forms from modern research. Note that H contains
an initial, rudimentary definition which we shall
all closed forms in the sense of Wilf and Zeilberger
refine presently. First, we want any convergent
[48, Ch. 8], wherein only finite linear combinations
sum
X of hypergeometric evaluations are allowed.
(1) x= cn z n So what numbers are in the ring H? First off,
n0 almost no complex numbers belong to this ring!
to be an element of our set X, where z is algebraic, This is easily seen by noting that the set of general
c0 is rational, and for n > 0, hypergeometric evaluations is countable, so the
A(n) generated ring must also be countable. Still, a
cn = cn1 great many fundamental numbers are provably
B(n)
hyperclosed. Examples follow, in which we let
for integer polynomials A, B with deg A deg B. Un-
denote an arbitrary algebraic number and n any
der these conditions the expansion for x converges
positive integer:
absolutely on the open disk |z| < 1. However, we
also allow x to be any finite analytic-continuation , log , e , ;
value of such a series. Moreover, when z lies on the dilogarithmic combination
a branch cut, we presume both branch limits to
1
 
be elements of X. (See ensuing examples for some Li2 + (log 2)(log 3);
5
clarification.) It is important to note that our set
the elliptic integral K();
X is closed under rational multiplication due to
freedom of choice for c0 . the zeta function values (n);

Example 2 (First members of X). The generalized special functions such as


hypergeometric function evaluation the Bessel evaluations Jn ().
!
a1 , . . . ap+1 Incidentally, it occurs in some modern experi-
p+1 Fp
z
b1 , . . . , b p mental developments that the real or imaginary part
of a hypergeometric evaluation is under scrutiny.
for rational ai , bj with all bj positive has branch
Generally, <, = operations preserve hyperclosure
cut z (0, ), and the evaluation is an element of
simply because the series (or continuations) at z
X for complex z not on the cut (and the evaluation
and z can be linearly combined in the ring H. Re-
on each side of said cut is alsoPin X).
ferring to Example 2, we see that for algebraic z, the
The trilogarithm Li3 (z) := n1 z n /n3 offers a
number < (Li3 (z)) is hyperclosed, and even on the
canonical instance. Formally, 7 2
cut, < (Li3 (2)) = 16 (3) + 8 log 2 is hyperclosed.
1 1, 1, 1, 1
 
  
Li3 (z) = 4 F3 z ,
In general, < p+1 Fp

z 2, 2, 2 z is hyperclosed.

and for z = 1/2 the hypergeometric series con- We are not claiming that hyperclosure is any
verges absolutely, with kind of final definition for closed forms, but we
do believe that any defining paradigm for closed
1 7 1 2
 
Li3 = (3) + log3 2 log 2. forms must include this ring of hyperclosure H.
2 8 6 12
One way to reach further is to define a ring of
Continuation values at z = 2 on the branch cut can superclosure as the closure
be inferred as
S := hHH i,+ .
7 2
lim+ Li3 (2 i) = (3) + log 2 i log 2,
0 16 8 4 This ring contains numbers such as
so both complex numbers on the right here are 1
 el-
 e + e , ,
ements of X. The quadralogarithmic value Li4 21 (3)(5)
is thought not to be similarly decomposable but and of course a vast collection of numbers that may
likewise belongs to X.  not belong to H itself. If we say that an element of

January 2013 Notices of the AMS 53


S is superclosed, we still preserve the countability is an L-series value over an elliptic curve E with
of all superclosed numbers. Again, any good conductor 15. Rogers [50] recasts (2) as
definition of closed form should incorporate
!2
whatever is in the ring S. ? 15
X 2n (1/16)2n+1
(3) F(3, 5) = 2 ,
n=0 n 2n + 1
Seventh Approach. In a more algebraic topological
setting, it might make sense to define closed where
forms to be those arising as periods, that is, F(b, c) := (1 + b)(1 + c)
as integrals of rational functions (with integer X (1)n+m+j+k
2
parameters) in n variables over domains defined n,m,j,k (6n + 1)2 + b(6m + 1)2 + c(6j + 1)2 + bc(6k + 1)2
by algebraic equations. These ideas originate in is a four-dimensional lattice sum.
the theory of elliptic and abelian integrals and While (3) remains a conjecture,5 Rogers is able to
are deeply studied [42]. Periods form an algebra evaluate many values of F(b, c) in terms of Meijer-
and certainly capture many constants. They are G or hypergeometric functions. We shall consider
especially well suited to the study of L-series, the most famous crystal sum, the Madelung con-
multizeta values, polylogarithms, and the like but stant, in Example 15. 
again will not capture all that we wish. For example,
e is conjectured not to be a period, as is Eulers It is striking how beautiful combinatorial games
constant (see the section Profound Curiosities). can be when played under the rubric of hyper- or
Moreover, while many periods have nice series, it superclosure.
is not clear that all do.
As this takes us well outside our domain of Example 4 (Superclosure of at rational argu-
expertise, we content ourselves with two examples ments). Let us begin with the Beta function
originating in the study of Mahler measures. We (r ) (s)
B(r , s) := ,
refer to a fundamental paper by Deninger [39] and (r + s)
a very recent paper of Rogers [50] for details.
with
R s1 (s) defined, if one wishes, as (s) :=
t
Example 3 (Periods and Mahler measures [39]). 0 t e dt. It turns out that for any rationals r , s
The logarithmic Mahler measure of a polynomial the Beta function is hyperclosed. This is immediate
P in n-variables can be defined as from the hypergeometric identities
Z1Z1 Z1
1 rs r , s
   
(P ) := log P e2 i1 , . . . , e2 in = 2 F1
1 ,

0 0 0 B(r , s) r +s 1
d1 dn . sin (r + s) (1 r )M (1 s)M
B(r , s) =
sin r sin s M!(1 r s)M
Then (P ) turns out to be an example of a period,
r, s
 
and its exponential, M(P ) := exp((P )), is a mean 2 F1 1 ,
of the values of P on the unit n-torus. When M + 1
n = 1 and P has integer coefficients, M(P ) is where M is any integer chosen such that the hy-
always an algebraic integer. An excellent online pergeometric series converges, say M = d1 + r + se.
synopsis can be found in Dave Boyds article, (Each of these Beta relations is a variant of the cele-
http://eom.springer.de/m/m120070.htm. In- brated Gauss evaluation of 2 F1 at 1 [1], [54] and is
deed, Boyd has been one of the driving forces in also the reason B is a period.)
the field. A brief introduction to the univariate We did not seize upon the Beta function arbitrar-
case is also given in [22, pp. 358359]. ily, for, remarkably, the hyperclosure of B(r , s)1
There is a remarkable series of recent results leads to compelling results on the Gamma function
many more discovered experimentally than itself. Indeed, consider for example this product
provenexpressing various multidimensional of four Beta-function evaluations:
(P ) as arithmetic quantities. Boyd observes that (1/5) (1/5) (2/5) (1/5)
there appears to be a tight connection to K-theory.
(2/5) (3/5)
An early result due to Smyth (see [51], also [53]) is
(3/5) (1/5) (4/5) (1/5)
that (1 + x + y) = L03 (1). Here L3 is the Dirichlet .
L-series modulo three. A partner result of Smyths (4/5) (5/5)
is that (1 + x + y + z) = 7 (3)/ 2 , a number We know this product is hyperclosed. But upon
that is certainly hyperclosed since both (3) and inspection we see that the product is just 5 (1/5).
1/ are. A conjecture of Deninger [39], confirmed
to over fifty places, is that 5
Equation (2) is now proven, see M. Rogers and W. Zudilin,
? 15 On the Mahler measure of 1 + X + 1/X + Y + 1/Y , preprint
(2) (1 + x + y + 1/x + 1/y) = 2 LE (2) (2011), http://arxiv.org/abs/1102.1153.

54 Notices of the AMS Volume 60, Number 1


Along such lines one can prove that for any posi- disk, convergence is geometric, requiring O(D 1+ )
tive rational a/b (in lowest terms), we have hyper- operations to achieve D good digits. However, in
closure of powers of the Gamma function in the very many cases this can be genuinely enhanced
form to O(D 1/2+ ) [22].
b (a/b) H.
Perforce, we therefore have a superclosure result Closed Forms: Why We Care
for any (rational) and its reciprocal: In many optimization problems, simple,
approximate solutions are more useful than
1 (a/b) S.
complex exact solutions.
Again, like calculations show that b (a/b) is a pe-
Steve Wright
riod [42].
 One fundamental consequence is thus
2 1 As Steve Wright observed in a recent lecture on
2 = 1 is hyperclosed; thus every integer
sparse optimization, it may well be that a compli-

power of is hyperclosed. cated analytic solution is practically intractable,


Incidentally, deeper combinatorial
  analysis but a simplifying assumption leads to a very practi-
5 1
shows that in spite of our 5 Beta-chain above, cal closed form approximation (e.g., in compressed
it really takes only logarithmically many (i.e., sensing). In addition to appealing to Occams razor,
O(log b)) hypergeometric evaluations to write Wright instances that:
Gamma-powers. For example, (a) the data quality may not justify exactness,
!4 (b) the simple solution may be more robust,
7 1 1 17 , 17
 
= 3 6 2 F1 1
(c) it may be easier to explain/actuate/
7 2 7 1

!2 implement/store,
27 , 7
2 4 4
!
7,7 (d) and it may conform better to prior knowl-
2 F1 1 F
2 1 .
1
1 1 edge.
We note also that for (n/24) with n integer, elliptic As mathematical discovery more and more
integral algorithms are known that converge as fast involves extensive computation, the premium
as those for [27], [22].  on having a closed form increases. The insight
provided by discovering a closed form ideally
The above remarks on superclosure of (a/b) comes at the top of the list, but efficiency of
lead to the property of superclosure for special computation will run a good second.
functions such as J () for algebraic and
rational and for many of the mighty Meijer- Example 5 (The amplitude of a pendulum).
G functions, as the latter can frequently be Wikipedia,6 after giving the classical small angle
written by Slaters theorem [15] as superpositions (simple harmonic) approximation
of hypergeometric evaluations with composite-
s
L
gamma products as coefficients. (See Example 8 p 2
g
for instances of Meijer-G in current research.)
There is an interesting alternative way to en- for the period p of a pendulum of length L and
vision hyperclosure or at least something very amplitude , develops the exact solution in a form
close to our above definition. This is an idea equivalent to
of J. Carette [28] to the effect that solutions at
s
L
 
algebraic end-points and algebraic initial points p=4 K sin
g 2
for holonomic ODEsi.e., differential-equation sys-
tems with integer-polynomial coefficientscould and then says:
be considered closed. One might say diffeoclosed. This integral cannot be evaluated in terms
An example of a diffeoclosed number is J1 (1), i.e., of elementary functions. It can be rewrit-
from the Bessel differential equation for J1 (z) with ten in the form of the elliptic function of
z [0, 1]; it suffices without loss of generality the first kind (also see Jacobis elliptic func-
to consider topologically clean trajectories of the tions), which gives little advantage since
variable over [0, 1]. There is a formal ring of that form is also insoluble.
diffeoclosure, which ring is very similar to our H; True, an elliptic-integral solution is not elementary,
however, there is the caution that trajectory solu- yet the notion of insolubility is misleading for two
tions can sometimes have nontrivial topology, so reasons: First, it is known that for some special
precise ring definitions would need to be effected angles the pendulum period can be given a closed
carefully. form. As discussed in [33], one exact solution is, for
It is natural to ask, what is the complexity
of hypergeometric evaluations? Certainly for the 6
Available at http://en.wikipedia.org/wiki/
converging forms with variable z on the open unit Pendulum_(mathematics).

January 2013 Notices of the AMS 55


= /2 (so pendulum is released from horizontal- these computations are performed using conven-
rod position), tional double precision (64-bit IEEE) arithmetic.
s ! Then if a particular phase space point is deemed
L
p = 2 . numerically unstable, it is recomputed with double-
g 2 (3/4)
double precision. These researchers expect that
It is readily measurable in even a rudimentary lab- further optimization of the procedure for identify-

oratory that the excess factor here, 2 (3/4) ing unstable points may be required to arrive at an
1.18034, looks just right; i.e., a horizontal-release optimal compromise between numerical accuracy
pendulum takes 18 percent longer to fall. More- and speed of the code. Thus they plan to incor-
over, there is an exact dynamical solution for the porate arbitrary precision arithmetic into these
time-dependent angle (t), namely, for a pendu- calculations. Their objective is to design a pro-
lum with () = and (0) = 0; i.e., the bob cedure where instead of using fixed double or
crosses angle zero (hanging straight down) at time
quadruple precision for unstable points, the num-
zero, but in the limits of time the bob ends
ber of digits in the higher precision calculation is
up straight vertical. We have period p = , yet the
dynamically set according to the instability of the
exact angle (t) for given t can be written in terms
point. Any subroutine which uses a closed form
of elementary functions!
symbolic solution (exact or approximate) is likely
The second misleading aspect is this: K is, for
to prove much more robust and efficient. 
any , remarkably tractable in a computational
sense. Indeed K admits a quadratic transformation
Detailed Examples
1 1 k2 We start with three examples originating in [16].
(4) K (k) = (1 + k1 ) K (k1 ) , k1 := ,
1 + 1 k2 In the January 2002 issue of SIAM News, Nick
as was known already to Landen, Legendre, and Trefethen presented ten diverse problems used
Gauss. in teaching modern graduate numerical analysis
In fact all elementary functions to very high students at Oxford University, the answer to
precision are well computed via K [22]. So the com- each being a certain real number. Readers were
ment was roughly accurate in the world of slide challenged to compute ten digits of each answer,
rules or pocket calculators; it is misleading today if with a $100 prize to the best entrant. Trefethen
one has access to any computer package. Neverthe- wrote,
less, both deserve to be called closed forms: one
exact and the other an elegant approximate closed If anyone gets 50 digits in total, I will be
form (excellent in its domain of applicability, much impressed.
as with Newtonian mechanics) that is equivalent to To his surprise, a total of ninety-four teams,



representing twenty-five different nations, sub-
K sin mitted results. Twenty of these teams received
2 2
for small initial amplitude . To compute a full one hundred points (ten correct digits for
K( /6) = 1.699075885 . . . to five places re- each problem). The problems and solutions are
quires using (4) only twice and then estimating the dissected most entertainingly in [16]. One of the
resultant integral by /2. A third step gives the current authors wrote the following in a review
ten-digit precision shown.  [19] of [16]:
Success in solving these problems required
It is now the case that much mathematical
computation is hybrid: mixing numeric and sym- a broad knowledge of mathematics and nu-
bolic computation. Indeed, which is which may not merical analysis, together with significant
be clear to the user if, say, numeric techniques computational effort, to obtain solutions
have been used to return a symbolic answer or if and ensure correctness of the results. As
a symbolic closed form has been used to make described in [16] the strengths and lim-
possible a numerical integration. Moving from itations of Maple, Mathematica, Matlab
classical to modern physics, both understand- (The 3Ms), and other software tools such
ing and effectiveness frequently demand hybrid as PARI or GAP, were strikingly revealed in
computation. these ventures. Almost all of the solvers
relied in large part on one or more of these
Example 6 (Scattering amplitudes [2]). An interna- three packages, and while most solvers
tional team of physicists, in preparation for the
attempted to confirm their results, there
Large Hadron Collider (LHC), is computing scat-
was no explicit requirement for proofs to
tering amplitudes involving quarks, gluons, and
be provided.
gauge vector bosons in order to predict what re-
sults could be expected on the LHC. By default, Example 7 (Trefethen problem #2 [16], [19]).

56 Notices of the AMS Volume 60, Number 1


A photon moving at speed 1 in the x-y plane
4
starts at t = 0 at (x, y) = (1/2, 1/10) head-
ing due east. Around every integer lattice 3

point (i, j) in the plane, a circular mirror of 2


radius 1/3 has been erected. How far from 1
the origin is the photon at t = 10?
Using interval arithmetic with starting intervals 6 4 2 2
1
of size smaller than 105000 , one can actually find
the position of the particle at time 2000, not just 2
at time ten. This makes a fine exercise in very high- 3
precision interval computation, but in the absence (a) W3
of any closed form, one is driven to such numerical
gymnastics to deal with error propagation.  4

Example 8 (Trefethens problem #9 [16], [19]). 3


R2 2
The integral I(a) = 0 [2 + sin(10)]x
sin(/(2 x)) dx depends on the param- 1
eter . What is the value [0, 5] at which
6 4 2 2
I() achieves its maximum?
1
The maximum parameter is expressible in terms
2
of a Meijer-G function, which is a special func-
tion with a solid history. While knowledge of this 3
function was not common among the contestants, (b) W4
Mathematica and Maple both will figure this out
Figure 2. Moments of n -step walks in the plane.
[15], and then the help files or a Web search will
W3 , W4 analytically continued to the real line.
quickly inform the scientist.
This is another measure of the changing envi-
ronment. It is usually a good ideaand not at all Figure 2 and in the complex plane in Figure 3.
immoralto data-mine. These Meijer-G functions, The latter highlights the utility of the Meijer-G
first introduced in 1936, also occur in quantum representations. Note the poles and removable
field theory and many other places [8]. For exam- singularities.
ple, the moments of an n-step random walk in the The Meijer-G functions are now described in the
plane are given for s > 0 by newly completed Digital Library of Mathematical
s
Functions 7 and as such are now full, indeed central,

n 2 x i
Z X
(5) Wn (s) := e k
dx. members of the family of special functions. 
[0,1]n k=1


Example 9 (Trefethens problem #10 [16], [19]).
It transpires [24], [36] that for all complex s
A particle at the center of a 10 1 rectangle
(6)
! undergoes Brownian motion (i.e., 2-D ran-
(1 + s/2) 2,1 1, 1, 1 1 .
W3 (s) = G 1 s s dom walk with infinitesimal step lengths)
(1/2) (s/2) 3,3 2,2,2
4
till it hits the boundary. What is the proba-
Moreover, for s not an odd integer, we have bility that it hits at one of the ends rather
than at one of the sides?
!2
1 1 1
1 s s , 2, 2 1
 
2
W3 (s) = 2s+1 tan s1 3 F2

s+3 s+3

2 2 4 Hitting the Ends. Bornemann [16] starts his remark-
2 2 , 2
! s s s ! able solution by exploring Monte-Carlo methods,
s 2,2,2 1
+ s 3 F2 which are shown to be impracticable. He then
4 .

s1
2 1, 2 reformulates the problem deterministically as
We have not given the somewhat technical defini- the value at the center of a 10 1 rectangle of an
tion of MeijerG, but Maple, Mathematica, Google appropriate harmonic measure [57] of the ends,
searches, Wikipedia, the DLMF, or many other tools arising from a five-point discretization of Laplaces
will. equation with Dirichlet boundary conditions. This
There are two corresponding formulae for W4 . is then solved by a well-chosen sparse Cholesky
We thus know from our Sixth Approach section, 7
A massive revision of Abramowitz and Stegun, with
in regard to superclosure of -evaluations, that the now redundant tables removed, is available at
both W3 (q), W4 (q) are superclosed for rational ar- http://dlmf.nist.gov. The hard copy version is
gument q for q not an odd integer. We illustrate also now out [45]. It is not entirely a substitute for the
by showing graphs of W3 , W4 on the real line in original version, as coverage has changed.

January 2013 Notices of the AMS 57


answer is
2
(9) p= arcsin (k100 ) ,

where
   
k100 := 3 2 2 2 + 5 3 + 10
 2 2
4
2+ 5 .

No one (except harmonic analysts perhaps) antici-


pated a closed form, let alone one like this.

Where Does This Come From? In fact, [22, (3.2.29)]


shows that

(1)n (2n + 1) 1
 
(a) W3
X
(10) sech = arcsin k
n=0
2n + 1 2 2
exactly when k2 is parameterized by theta func-
tions in terms of the so-called nome, q = exp( ),
as Jacobi discovered. We have
P 2
2 (q) q (n+1/2)
(11) k2 = 22 = n=
P
n2
, q := e .
3 (q) n= q
Comparing (10) and (7) we see that the solution is
k100 = 6.02806910155971082882540712292 . . . 107 ,
as asserted in (9).
The explicit form now follows from classical
nineteenth-century theory, as discussed, say, in
[16], [22]. In fact, k210 is the singular value sent
by Ramanujan to Hardy in his famous letter of
(b) W4 introduction
[21], [22]. If Trefethen
had asked
for
a 210 1 box or, even better, a 15 14 one,
Figure 3. W3 via (6) and W4 in the complex plane.
this would have shown up in the answer, since in
general
2 
solver. At this point a reliable numerical value of (12) p(a, b) = arcsin ka2 /b2 .

3.837587979 107 is obtained, and the posed
Alternatively, armed only with the knowledge
problem is solved numerically to the requisite ten
that the singular values of rational parameters
places.
are always algebraic, we may finish entirely
This is the warm-up. We may proceed to develop
computationally as described in [19]. 
two analytic solutions, the first using separation
We finish this section with two attractive ap-
of variables on the underlying PDE on a general plied examples from optics and astrophysics
2a 2b rectangle. We learn that respectively.

4 X (1)n (2n + 1)
 
(7) p(a, b) = sech , Example 10 (Mirages [46]). In [46] the authors,
n=0 2n + 1 2 using geometric methods, develop an exact but im-
plicit formula for the path followed by a light ray
where := a/b. A second method using conformal
propagating over the earth with radial variations in
mappings yields
the refractive index. By suitably simplifying, they
 
(8) arccot = p(a, b) + arg K eip(a,b) , are able to provide an explicit integral closed form.
2 They then expand it asymptotically. This is done
where K is again the complete elliptic integral of with the knowledge that the approximation is good
the first kind. It will not be apparent to a reader to six or seven places, more than enough to use it
unfamiliar with inversion of elliptic integrals that on optically realistic scales. Moreover, in the case
(7) and (8) encode the same solutionthough they of quadratic or linear refractive indices, these steps
must, as the solution is unique in (0, 1)and each may be done analytically.
can now be used to solve for = 10 to arbitrary In other words, as advanced by Wright, a
precision. Bornemann ultimately shows that the tractable and elegant approximate closed form is

58 Notices of the AMS Volume 60, Number 1


trajectories for which indices n being respectively
1
(14) 0 (r ) = r 3 + r ,
6
(15) 1 (r ) = sin r ,
r
(16) 5 (r ) = p .
1 + r 2 /3
The respective
star radii are thus closed forms
z0 = 6 and z1 = , while for (16), with index
n = 5 we have infinite star radius (no positive zero
for the pressure 5 ).
In the spirit of our previous optics example,
the Lane-Emden equation is a simplification of
(a) A superior mirage a complicated underlying theoryin this astro-
physics case, hydrodynamicsand one is rewarded
by some closed-form star radii. But what about, say,
index n = 2? We do not know a closed-form func-
tion for the trajectory in any convenient sense.
What the present authors have calculated (in 2005)
is the n = 2 star radius, as a high-precision number
z2 = 4.352874595946124676973570061526142628112365363213008835302151 . . . .

If only we could gain a closed form for this special


radius, we might be able to guess the nature of the
whole trajectory! 
(b) An inferior mirage (Photo Ctein.)

Figure 4. Two impressive mirages.


Recent Examples Relating to Our Own Work
Example 12 (Ising integrals [5], [8]). We recently
studied the following classes of integrals [5]. The
obtained to replace a problematic exact solution. Dn integrals arise in the Ising model of mathemat-
From these forms interesting qualitative conse- ical physics (showing ferromagnetic temperature-
quences follow. With a quadratic index, images driven phase shifts; see Figure 5 and [32]), and the
are uniformly magnified in the vertical direction; Cn have tight connections to quantum field theory
only with higher-order indices can nonuniform [8]:
Z Z
vertical distortion occur. This sort of knowledge 4 1 du1 dun
Cn = 2 ,
allows one, for example, to correct distortions of n! u1 un
P
0 0 n
j=1 (uj + 1/uj )
photographic images as in Figure 4 with confidence  2
ui uj
and efficiency. 
Q
Z Z i<j
4 ui +uj du1 dun
Dn = 2 ,
n! u1 un
P
0 0 n
Example 11 (Structure of stars). The celebrated j=1 (uj + 1/uj )
Lane-Emden equation, presumed to describe the Z1 Z1
2
uk uj
pressure at radius r within a star, can be put in
Y
En = 2 dt2 dt3 dtn ,
0 0 uk + uj
the form 1j<kn

d2 Qk
where (in the last line) uk = i=1 ti .
(13) r n1 = n ,
dr 2 Needless to say, evaluating these multidimen-
with boundary conditions (0) = 0, 0 (0) = 1, and sional integrals to high precision presents a daunt-
positive real constant n, all of this giving rise to ing computational challenge. Fortunately, in the
a unique trajectory n (r ) on r [0, ). (Some au- first case, the Cn integrals can be written as one-
thors invoke the substitution (r ) := r (r ) to get dimensional integrals:
an equivalent ODE for temperature ; see [30].) The
2n
Z
beautiful thing is, where this pressure trajectory Cn = pK0n (p) dp,
n! 0
crosses zero for positive radius r is supposed to
be the star radius ; call that zero zn . where K0 is the modified Bessel function. After com-
Amazingly, the Lane-Emden equation has puting Cn to 1000-digit accuracy for various n, we
known exact solutions for n = 0, 1, 5, the pressure were able to identify the first few instances of Cn

January 2013 Notices of the AMS 59


numerical value as
lim Cn = 2e2 ,
n
where is the Euler constant ; see the section Pro-
found Curiosities. We later were able to prove this
factthis is merely the first term of an asymptotic
expansionand thus showed that the Cn integrals
are fundamental in this context [5].
The integrals Dn and En are much more diffi-
cult to evaluate, since they are not reducible to
one-dimensional integrals (as far as we can tell);
but with certain symmetry transformations and
symbolic integration we were able to symbolically
reduce the dimension in each case by one or two.
In the case of D5 and E5 , the resulting 3-D inte-
(a) Critical temperature
grands are extremely complicated (see Figure 6),
but we were nonetheless able to numerically evalu-
ate these to at least 240-digit precision on a highly
parallel computer system. This would have been
impossible without the symbolic reduction. We
give the integral in extenso to show the difference
between a humanly accessible answer and one a
computer finds useful.
In this way, we produced the following evalua-
tions, all of which, except the last, we subsequently
were able to prove:
D2 = 1/3,
D3 = 8 + 4 2 /3 27 L3 (2),
D4 = 4 2 /9 1/6 7(3)/2,
E2 = 6 8 log 2,
E3 = 10 2 2 8 log 2 + 32 log2 2,
E4 = 22 82(3) 24 log 2 + 176 log2 2,
(b) Wolfram Player Demonstration
256(log3 2)/3 + 16 2 log 2 22 2 /3,
Figure 5. The 2-dimensional Ising Model of
Ferromagnetism (a) plotting magnetization C and
(blue, with peak) and specific heat M (red, (17)
decaying) per site against absolute temperature ?
E5 = 42 1984 Li4 (1/2) + 189 4 /10 74(3)
T (image provided by Jacques Perk) [44,
pp. 9193, 245]. 1272(3) log 2 + 40 2 log2 2
62 2 /3 + 40( 2 log 2)/3 + 88 log4 2

in terms of well-known constants, e.g., + 464 log2 2 40 log 2,


X 1 1
 where Li denotes the polylogarithm function.
C3 = L3 (2) := , In the case of D2 , D3 , and D4 , these are confir-
n0
(3n + 1)2 (3n + 2)2
mations of known results. We tried but failed to
7 recognize D5 in terms of similar constants (the
C4 = (3),
12 500-digit numerical value is accessible9 if anyone
where denotes the Riemann zeta function. When wishes to try to find a closed form or, in the manner
we computed Cn for fairly large n, for instance, of the hard sciences, to confirm our data values).
C1024 = 0.63047350337438679612204019271087890435458707871273234 . . . , The conjectured identity shown here for E5 was
confirmed to 240-digit accuracy, which is 180 digits
we found that these values rather quickly ap-
beyond the level that could reasonably be ascribed
proached a limit. By using the new edition of the
to numerical round-off error. Thus we are quite
Inverse Symbolic Calculator,8 we identified this
9
Available at http://crd.lbl.gov/~dhbailey/
8
Available at http://isc2.carma.newcastle.edu.au/. dhbpapers/ising-data.pdf.

60 Notices of the AMS Volume 60, Number 1


E5 =
Z1 Z1 Z1 h
2(1 x)2 (1 y)2 (1 xy)2 (1 z)2 (1 yz)2 (1 xyz)2
with experimental estimate c1 = 0.605443657 . . . .
0 0 0
In [6] we were able to derive the exact theoretical
4(x + 1)(xy + 1) log(2) y 5 z 3 x7 y 4 z 2 (4(y + 1)z + 3)x6 y 3 z y 2 + 1 z 2
 h   

value of this QRS constant c1 as the unique zero


+4(y + 1)z + 5) x5 + y 2 4y(y + 1)z 3 + 3 y 2 + 1 z 2 + 4(y + 1)z 1 x4
   
of the Hurwitz zeta (1/2, z/2) on z (0, 2). In
+y z z 2 + 4z + 5 y 2 + 4 z 2 + 1 y + 5z + 4 x3 + 3z 2 4z + 1 y 2
       
so doing, we were able to prove the conjectured
4zy + 1) x2 (y(5z + 4) + 4)x 1 / (x 1)3 (xy 1)3 (xyz 1)3
i h i
behavior. Moreover, we were able to sketch the
+ 3(y 1)2 y 4 (z 1)2 z 2 (yz 1)2 x6 + 2y 3 z 3(z 1)2 z 3 y 5 + z 2 5z 3 + 3z 2
h  
higher-order asymptotic behavior, something that
+3z + 5) y 4 + (z 1)2 z 5z 2 + 16z + 5 y 3 + 3z 5 + 3z 4 22z 3 22z 2 would have been impossible without discovery of
  

+3z + 3) y 2 + 3 2z 4 + z 3 + 2z 2 + z 2 y + 3z 3 + 5z 2 + 5z + 3 x5 an analytic formula.


  

+ y 2 7(z 1)2 z 4 y 6 2z 3 z 3 + 15z 2 + 15z + 1 y 5 + 2z 2 21z 4 + 6z 3


   
Does this deserve to be called a closed form?
+14z 2 + 6z 21 y 4 2z z 5 6z 4 27z 3 27z 2 6z + 1 y 3 + 7z 6
   
In our opinion, resoundingly yes unless all inverse
30z 5 + 28z 4 + 54z 3 + 28z 2 30z + 7 y 2 2 7z 5 + 15z 4 6z 3 6z 2 + 15z
  functions such as that in Bornemanns (12) are to be
eschewed. Such constants are especially interesting
+7) y + 7z 4 2z 3 42z 2 2z + 7 x4 2y z 3 z 3 9z 2 9z + 1 y 6
   

in light of even more recent work by Steve Strogatz


+ z 2 7z 4 14z 3 18z 2 14z + 7 y 5 + z 7z 5 + 14z 4 + 3z 3 + 3z 2 + 14z
  

and his collaborators on chimera, coupled systems


+7) y 4 + z 6 14z 5 + 3z 4 + 84z 3 + 3z 2 14z + 1 y 3 3 3z 5 + 6z 4 z 3
  

which can self-organize in parts of their domain


z 2 + 6z + 3 y 2 9z 4 + 14z 3 14z 2 + 14z + 9 y + z 3 + 7z 2 + 7z + 1 x3
   
and remain disorganized elsewhere; see Figure 7
+ z 2 11z 4 + 6z 3 66z 2 + 6z + 11 y 6 + 2z 5z 5 + 13z 4 2z 3 2z 2 + 13z
   
taken from [43]. In this case, observed numerical
+5) y 5 + 11z 6 + 26z 5 + 44z 4 66z 3 + 44z 2 + 26z + 11 y 4 + 6z 5 4z 4
  
limits still need to be put in closed form. 
66z 3 66z 2 4z + 6 y 3 2 33z 4 + 2z 3 22z 2 + 2z + 33 y 2 + 6z 3
   

It is a frequent experience of ours that, as in


+26z + 6) y + 11z 2 + 10z + 11 x2 2 z 2 5z 3 + 3z 2 + 3z + 5 y 5
   

Example 13, the need for high-accuracy computa-


+ z 22z 4 + 5z 3 22z 2 + 5z + 22 y 4 + 5z 5 + 5z 4 26z 3 26z 2 + 5z + 5 y 3
   

tion drives the development of effective analytic


+ 3z 4 22z 3 26z 2 22z + 3 y 2 + 3z 3 + 5z 2 + 5z + 3 y + 5z 2 + 22z + 5 x
    
expressions (closed forms?), which in turn typically
+ 15z 2 + 2z + 2y(z 1)2 (z + 1) + 2y 3 (z 1)2 z(z + 1) + y 4 z 2 15z 2 + 2z + 15
 
shed substantial light on the subject being studied.
+y 2 15z 4 2z 3 90z 2 2z + 15 + 15 / (x 1)2 (y 1)2 (xy 1)2 (z 1)2
  i h

(yz 1)2 (xyz 1)2 4(x + 1)(y + 1)(yz + 1) z 2 y 4 + 4z(z + 1)y 3


i h  Example 14 (Box integrals [3], [7], [23]). There has
been recent research on the calculation of the ex-
+ z 2 + 1 y 2 4(z + 1)y + 4x y 2 1 y 2 z 2 1 + x2 z 2 y 4 4z(z + 1)y 3
     

pected distance of points inside a hypercube to the


z 2 + 1 y 2 + 4(z + 1)y + 1 1 log(x + 1) / (x 1)3 x(y 1)3 (yz 1)3
    i h i

hypercube. Such expectations are also called box


4(y + 1)(xy + 1)(z + 1) x2 z 2 4z 1 y 4 + 4x(x + 1) z 2 1 y 3
h     

integrals [23]. So, for example, the expectation


x2 + 1 z 2 4z 1 y 2 4(x + 1) z 2 1 y + z 2 4z 1 log(xy + 1) /
      i
r |i for random r~ [0, 1]3 has the closed form
h|~
x(y 1)3 y(xy 1)3 (z 1)3 4(z + 1)(yz + 1) x3 y 5 z 7 + x2 y 4 (4x(y + 1)
h i h 

1 1 1  
+ 5)z 6 xy 3 y 2 + 1 x2 4(y + 1)x 3 z 5 y 2 4y(y + 1)x3 + 5 y 2 + 1 x2 3 + log 2 + 3 .
     

4 24 2
+4(y + 1)x + 1) z 4 + y y 2 x3 4y(y + 1)x2 3 y 2 + 1 x 4(y + 1) z 3
   

Incidentally, box integrals are not just a math-


+ 5x2 y 2 + y 2 + 4x(y + 1)y + 1 z 2 + ((3x + 4)y + 4)z 1 log(xyz + 1) /
   i

ematicians curiosity; the integrals have been


xy(z 1)3 z(yz 1)3 (xyz 1)3
h ii
used recently to assess the randomness of brain
/ (x + 1)2 (y + 1)2 (xy + 1)2 (z + 1)2 (yz + 1)2 (xyz + 1)2 dx dy dz.
h i
synapses positioned within a parallelepiped [38].
Indeed, we had cognate results for
Z Z
Figure 6. The reduced multidimensional integral
d (s) := kx yks2 dxdy,
for E5 , which integral has led via [0,1]d [0,1]d
extreme-precision numerical quadrature and
which gives the moments of the distance between
PSLQ to the conjectured closed form given in
two points in the hypercube.
(17).
In a lovely recent paper [52] Stephan Steiner-
berger has shown that in the limit, as the dimension
goes to infinity,
confident in this result, even though we do not  s/p Z
1
Z
have a formal proof [5]. (18) lim kx yksp dxdy
Note that every one of the D, E forms above, d d [0,1]d [0,1]d
!s/p
including the conjectured last one, is hyperclosed 2
in the sense of our Sixth Approach section.  =
(p + 1)(p + 2)
Example 13 (Weakly coupling oscillators [49], [6]). for any s, p > 0. In particular, with p = 2, this
In an important analysis of coupled Winfree oscil- gives a first-order answer to our earlier published
lators, Quinn, Rand, and Strogatz [49] developed request for the asymptotic behavior of d (s).
a certain N-oscillator scenario whose bifurca- A quite recent result is that all box integrals
tion phase offset is implicitly defined, with a r |n i for integer n and dimensions 1, 2, 3, 4, 5 are
h|~
conjectured asymptotic behavior: sin 1 c1 /N, hyperclosed, in the sense of our Sixth Approach

January 2013 Notices of the AMS 61


Figure 7. Simulated chimera (figures and parameters from [43]).

section. It turns out that five-dimensional box in- Profound Curiosities


tegrals have been especially difficult, depending In our treatment of numbers enjoying hyperclosure
on knowledge of a hyperclosed form for a single or superclosure, we admitted that such numbers
definite integral J(3), where are countable, and so almost all complex num-
log(t + x2 + y 2 )
Z
bers cannot be given a closed form along such
(19) J(t) := 2 2
dx dy.
[0,1] (1 + x )(1 + y )
2 lines. What is stultifying is: How do we identify an
A proof of hyperclosure of J(t) for algebraic t 0 explicit number lying outside such countable sets?
r |2 i for r~
is established in [23, Thm. 5.1]. Thus h|~ The situation is tantamount to the modern bind
5
[0, 1] can be written in explicit hyperclosed form in regard to normal numbers, numbers which to
involving a 105 -character symbolic J(3); the au- some base have statistically random digit structure
thors of [23] were able to reduce the 5-dimensional in a certain technical sense. The bind is, though
box integral down to only 104 characters. A com- almost all numbers are absolutely normal (i.e.,
panion integral J(2) also starts out with about 105 normal to every base 2, 3, . . . ), we do not know
characters but reduces stunningly to only a few
a single fundamental constant that is provably
dozen characters, namely,
absolutely normal. (We do know some artificial
2 7 11
 
(20) J(2) = log 2 (3) + Cl2 normal numbers; see [14].)
8 48 24 6 Here is one possible way out of the dilemma.
29 5
 
Cl2 , In the theory of computability, the existence of
24 6 noncomputable real numbers, such as an encoded
where Cl2 is the Clausen function Cl2 () := list of halting Turing machines, is well established.
2
P
n1 sin(n)/n (Cl2 is the simplest nonelemen- The celebrated Chaitin constant is a well-known
tary Fourier series). noncomputable. So a folk argument goes: Since
Automating such reductions will require a
every element of the ring of hyperclosure H can be
sophisticated simplification scheme with a very
computed via converging series, it should be that
large and extensible knowledge base. With a
current research assistant, Alex Kaiser at Berke- 6 H. A good research problem would be to make
ley, we have started to design software to refine this heuristic rigorous.
and automate this process and to run it before Let us focus on some constants that might
submission of any equation-rich paper (see [9]). not be hyperclosed (nor superclosed). One
This semi-automated integrity checking becomes such constant is the celebrated Euler constant
Pn
pressing when, as above, verifiable output from := limn k=1 1/k log n. We know of no
a symbolic manipulation can be the length of a hypergeometric form for ; said constant may well
Salinger novella.  lie outside H (or even S). There are expansions for

62 Notices of the AMS Volume 60, Number 1


said crystal is NaCl (salt), as illustrated in Figure
8). Image (b) is of a Helaman Ferguson sculpture
based on M which is awarded biannually by the
Canadian Mathematical Society as part of the David
Borwein Career Award) and is given by the formal
(conditionally convergent [18]) sum
X (1)x+y+z
(21) M := p
(x,y,z)6=(0,0,0)
x2 + y 2 + z 2
= 1.747564594633...
and has never been given what a reasonable
observer would call a closed form. Nature plays
an interesting trick here: There are other crystal
structures that are tractable, yet somehow this
(a) NaCl nearest neighbors
exquisitely symmetrical salt structure remains
elusive. In general, even-dimensional crystal sums
are more tractable than odd for the same modular
function reasons that the number of representa-
tions of a number as the sum of an even number
of squares is. But this does not make them easy,
as illustrated by Example 2.
Here we have another example of a constant
having no known closed form yet is rapidly calcu-
lable. A classical rapid expansion for the Madelung
constant is due to Benson:
(22)

 q 
sech2
X
M = 12 (2m + 1)2 +(2n + 1)2 ,
m,n0
2
in which convergence is exponential. Summing for
m, n 3 produces 1.747564594 . . . , correct to
(b) CMS Prize Sculpture
eight digits. There are a great many other such
Figure 8. Two representations of salt. formulae for M (see [22], [35]).
Through the analytic methods of Buhler, Cran-
dall, Tyagi, and Zucker since 1999 (see [35], [37],
the Euler constant, such as [55], [58]), we now know approximations such as
3 4 X log(2k + 1) ( 1 ) ( 3 )
 
= log 4 log + (1)k+1 , 1 log 2 8 1 k24
4 k1 2k + 1 M + + + 83/2 8 +log ,
8 4 3 8 2 16k04
and even more exotic series (see [13]). But in the where k4 := ((21/4 1)/(21/4 + 1))2 . Two remark-
spirit of the present treatment, we do not want able things: First, this approximation is good to the
to call the infinite series closed, because it is not same thirteen decimals we give in (21); the miss-
hypergeometric per se. Relatedly, the classical ing O(1014 ) error here is a rapidly, exponentially
Bessel expansion is convergingbut alas infinitesum in this modern
Pn1 1
!n approximation theory. Second, this six-term ap-
z z2
   
k=1 k
X
K0 (z) = ln + I0 (z) + . proximation itself is indeed hyperclosed, the only
2 n=1
(n!)2 4
problematic term being the -function part, but we
Now K0 (z) has a (degenerate) Meijer-G did establish in our Sixth Approach section that
representationso potentially is superclosed for B(1/8, 3/8) and also 1/ are hyperclosed, which is
algebraic zand I0 (z) is accordingly hyperclosed, enough. Moreover, the work of Borwein and Zucker
but the nested harmonic series on the right is [27] also settles hyperclosure for that term. 
again problematic. Again, is conjectured not to
Certainly we have nothing like a proof, or even
be a period [42].
the beginnings of one, that M (or ) lies outside H
Example 15 (Madelung constant [22], [37], [58]). (or even S), but we ask on an intuitive basis, Is a
Another fascinating number is the Madelung con- constant such as the mighty M telling us that it
stant, M, of chemistry and physics [22, Section is not hyperclosed, in that our toil only seems to
9.3]. This is the potential energy at the origin of bring more closed-form terms into play, with no
an oscillating-charge crystal structure (most often exact resolution in sight?

January 2013 Notices of the AMS 63


Concluding Remarks and Open Problems have illustrated this. Moreover, we belong to the
We have posited several approaches to the elusive subset of mathematicians that finds fun in finding
notion of closed form, but what are the intersec- unanticipated closed forms.
tions and interrelations of said approaches? For
example, can our Fourth Approach be precisely Acknowledgements
absorbed into the evidently more general Sixth Thanks are due to David Bailey and Richard Brent
Approach (hyperclosure and superclosure)? for many relevant conversations and to Armin
How do we find a single number that is provably Straub for the complex plots of W3 and W4 .
not in the ring of hyperclosure H? (Though no such
number is yet known, almost all numbers are, as References
noted, not in said ring!) The same question persists [1] G. E. Andrews, R. Askey, and R. Roy, Special Functions,
for the ring of superclosure, S. Furthermore, how Cambridge University Press, 1999.
precisely can one create a field out of HH via [2] David Bailey, Roberto Barrio, and Jonathan
appropriate operator extension? Borwein, High-precision computation: Mathemati-
cal physics and dynamics, Applied Mathematics
Though H is a subset of S, how might one prove
and Computation 218 (2012), 1010610121,
that H 6= S? (Is the inequality even true?) Likewise, http://dx.doi.org/10.1016/j.amc.2012.03.087.
is the set of closed forms in the sense of [48, Ch. 8] [3] D. Bailey, J. Borwein, N. Calkin, R. Girgensohn,
(only finite linear combinations of hypergeometric R. Luke, and V. Moll, Experimental Mathematics in
evaluations) properly contained in our H? And Action, A K Peters, Natick, MA, 2007.
H
what about a construct such as HH ? Should such [4] D. H. Bailey and J. M. Borwein, Computer-assisted
an entity be anything really new? Lest one remark discovery and proof, Tapas in Experimental Mathe-
matics, 2152, in Contemporary Mathematics, vol. 457,
on the folly of such constructions, we observe that
American Mathematical Society, Providence, RI, 2008.
most everyone would say is a closed form! [5] D. H. Bailey, D. Borwein, J. M. Borwein, and
Having established the property of hyperclosure R. E. Crandall, Hypergeometric forms for Ising-
for b (a/b), are there any cases where the power class integrals, Experimental Mathematics 16 (2007),
b may be brought down? For example, 1/ is 257276.
[6] D. H. Bailey, J. M. Borwein, and R. E. Crandall,
hyperclosed, but what about 1/ ?
Resolution of the QuinnRandStrogatz constant
What is a precise connection between the ring of nonlinear physics, Experimental Mathematics 18
of hyperclosure (or superclosure) and the set of (2008), 107116.
periods or of Mahler measures (as in Example 3)? [7] D. H. Bailey, J. M. Borwein, and R. E. Crandall,
There is expounded in reference [23] a theory of Advances in the theory of box integral, Mathematics
of Computation, 79 (2010), 18391866.
expression entropy, whereby some fundamental
[8] D. H. Bailey, J. M. Borwein, D. M. Broadhurst, and
entropy estimate gives the true complexity of L. Glasser, Elliptic integral representation of Bessel
an expression. So, for example, an expression moments, J. Phys. A: Math. Theory 41 (2008), 5203
having one thousand instances of the polylog 5231.
token Li3 might really involve only about 1,000 [9] D. H. Bailey, J. M. Borwein, and A. Kaiser, Automated
characters, with that polylogarithm token encoded simplification of large symbolic expressions, preprint,
as a single character, say. (In fact, during the August 2012.
[10] D. H. Bailey, J. M. Borwein, V. Kapoor, and E. Weis-
research for [23] it was noted that the entropy
stein, Ten problems in experimental mathematics,
of Maple and Mathematica expressions of the American Mathematical Monthly 113 (2006), 481409.
same entity often had widely varying text-character [11] D. H. Bailey, P. B. Borwein, and S. Plouffe, On
counts but similar entropy assessments.) the rapid computation of various polylogarithmic
On the other hand, one basic notion of closed constants, Mathematics of Computation 66 (1997),
form is that explicitly infinite sums not be allowed. 903913.
Can these two concepts be reconciled? Meaning, [12] R. Baillie, D. Borwein, and J. Borwein, Some sinc
sums and integrals, American Math. Monthly 115
can we develop a theory of expression entropy
(2008), 888901.
by which an explicit, infinite sum is given infinite [13] D. Bailey and R. Crandall, On the random character
entropy? This might be difficult, as, for example, of fundamental constant expansions, Experimental
P 1
a sum n=1 n3/2 takes only a few characters to Mathematics 10 (2001), 175190.
symbolize, as we just did! If one can succeed, [14] , Random generators and normal number,
though, in thus resolving the entropy business for Experimental Mathematics 11 (2002), 527547.
expressions, closed form might be rephrased as [15] Folkmar Bornemann, How Mathematica
and Maple Get Meijers G-function into Prob-
finite entropy.
lem 9, preprint, 2005. Available at http://
In any event, we feel strongly that the value www-m3.ma.tum.de/bornemann/
of closed forms increases as the complexity of Numerikstreifzug/Chapter9/MeijerG.pdf.
the objects we manipulate computationally and [16] F. Bornemann, D. Laurie, S. Wagon, and J. Wald-
inspect mathematically grows, and we hope we vogel, The SIAM 100-Digit Challenge: A Study

64 Notices of the AMS Volume 60, Number 1


in High-Accuracy Numerical Computing, SIAM, [40] H. R. P. Ferguson, D. H. Bailey, and S. Arno, Anal-
Philadelphia, 2004. ysis of PSLQ, an integer relation finding algorithm,
[17] E. J. Borowski and J. M. Borwein, Web-Linked Dic- Mathematics of Computation 68, no. 225 (1999),
tionary of Mathematics, Smithsonian/Collins Edition, 351369.
2006. [41] Brian Hayes, Why W? American Scientist 93 (2005),
[18] D. Borwein, J. M. Borwein, and K. F. Taylor, Con- 10041008.
vergence of lattice sums and Madelungs constant, [42] M. Kontsevich and D. Zagier, Periods. In Mathe-
J. Math. Phys. 26 (1985), 29993009. matics Unlimited2001 and Beyond, Springer-Verlag,
[19] J. M. Borwein, The SIAM 100 Digits Challenge, ex- 2005, pp. 771808.
tended review in the Mathematical Intelligencer 27 [43] E. A. Martens, C. R. Laing, and S. H. Strogatz, Solv-
(2005), 4048. able model of spiral wave chimeras, Physical Review
[20] J. M. Borwein and D. H. Bailey, Mathematics by Ex- Letters 104 (2010), 044101.
periment: Plausible Reasoning in the 21st Century, [44] B. McCoy and Tai Tsun Wu, The Two-Dimensional
extended second edition, A K Peters, Natick, MA, 2008. Ising Model, Harvard Univ. Press, 1973.
[21] J. M. Borwein, D. H. Bailey, and R. Girgensohn, Ex- [45] F. W. J. Olver, D. W. Lozier, R. F. Boisvert,
perimentation in Mathematics: Computational Roads and C. W. Clark, NIST Handbook of Mathematical
to Discovery, A K Peters, Natick, MA, 2004. Functions, Cambridge University Press, 2010.
[22] J. M. Borwein and P. B. Borwein, Pi and the AGM, [46] B. D. Nener, N. Fowkes, and L. Borredon, Analytical
John Wiley, 1987. models of optical refraction in the troposphere, JOSA
[23] J. Borwein, D. Nuyens, A. Straub, and James Wan, 20 (2003), 867875.
Some arithmetic properties of short random walk [47] K. Oldham, J. Myland, and J. Spanier, An Atlas of
integrals, Ramanujan Journal 26 (2011), 109132. Functions: With Equator, the Atlas Function Calculator,
[24] Jonathan M. Borwein, Armin Straub, and James Springer-Verlag, 2009.
Wan, Three-step and four-step random walks, [48] M. Petkovsek, H. Wilf, and D. Zeilberger, A = B,
Experimental Mathematics, in press, January 2013. A K Peters, Ltd., 1996. Available at http://www.math.
[25] J. M. Borwein and K. Devlin, The Computer as upenn.edu/~wilf/AeqB.html.
Crucible, A K Peters, Natick, MA, 2008. [49] D. Quinn, R. Rand, and S. Strogatz, Singular un-
[26] M. Bronstein, R. M. Corless, J. H. Davenport, and locking transition in the Winfree model of coupled
D. J. Jeffrey, Algebraic properties of the Lambert W oscillators, Phys. Rev. E 75 (2007), 036218-1-10.
function from a result of Rosenlicht and of Liouville, [50] M. Rogers, Hypergeometric formulas for lattice sums
Integral Transforms and Special Functions 19 (2008), and Mahler measures, preprint, 2010.
709712. [51] C. J. Smyth, Explicit formulas for the Mahler mea-
[27] J. M. Borwein and I. J. Zucker, Fast evaluation of sures of families of multivariable polynomials, preprint,
the gamma function for small rational fractions using 2003.
complete elliptic integrals of the first kind, IMA J. [52] S. Steinerberger, Extremal uniform distribution and
Numerical Analysis 12 (1991), 519526. random chord lengths, Acta Mathematica Hungarica
[28] J. Carette, private communication, 2010. 130 (2011), 321339.
[29] , Understanding expression simplification, Pro- [53] C. J. Smyth, The Mahler measure of algebraic num-
ceedings of International Conference on Symbolic and bers: A survey, Conference Proceedings, University of
Algebraic Computation, ACM, 2004, pp.7279. Bristol, 37 April 2006. LMS Lecture Note Series, vol.
[30] S. Chandrasekhar, An Introduction to the Study of 352, Cambridge University Press, Cambridge, 2008,
Stellar Structures, Dover, New York, 1967. pp. 322349.
[31] T. Y. Chow, What is a closed-form number? American [54] N. Temme, Special Functions, an Introduction to the
Mathematical Monthly 106 (1999), 440448. Classical Functions of Mathematical Physics, John Wiley
[32] B. A. Cipra, An introduction to the Ising model, Amer. & Sons, New York, 1996.
Math. Monthly 94 (1987), no. 10, 937959. [55] S. Tyagi, New series representation for the Madelung
[33] R. E. Crandall, Topics in Advanced Scientific constant, Progr. Theoret. Phys. 114 (2005), 517521.
Computation, Springer, New York, 1996. [56] E. W. Weisstein, Closed-Form Solution,
[34] , Theory of ROOF Walks, 2007. Available from MathWorldA Wolfram Web Resource.
at http://people.reed.edu/~crandall/papers/ http://mathworld.wolfram.com/Closed-
ROOF.pdf. FormSolution.html.
[35] , New representations for the Madelung con- [57] B. L. Walden and L. A. Ward, A harmonic mea-
stant, Experimental Mathematics 8 (1999), 367379. sure interpretation of the arithmetic-geometric mean,
[36] , Analytic representations for circle-jump mo- American Mathematical Monthly 114 (2007), 610622.
ments, PSIpress, 21 November 2009; http:// [58] I. J. Zucker, Some infinite series of exponential and
www.perfscipress.com/papers/ hyperbolic functions, SIAM J. Math. Anal. 15 (1984),
AnalyticWn_psipress.pdf 406413.
[37] R. E. Crandall and J. E. Buhler, Elementary expan-
sions for Madelung constants, J. Phys. A: Math. Gen.
20 (1987), 54975510.
[38] R. Crandall, On the fractal distribution of brain
synapses, preprint, 2010.
[39] C. Deninger, Deligne periods of mixed motives, theory
and the entropy of certain actions, J. Amer. Math. Soc.
10 (1997), 259281.

January 2013 Notices of the AMS 65

Potrebbero piacerti anche