16.3 Digression.

Some differential equations theory

Now you have some equations of motion in a mechanics problem, so solve them. Let’s prove that linear, second order
differential equations with regular coefficients (analytic functions that possess power series) have regular solutions
(no singularities).
Example. Linear eauations. This is all that you need to do quantum mechanics of the Schrödinger wave me-
chanics variety. Begin with

R1 (t)ü + R2 (t)u̇ + R3 (t)u = 0

for which we assume that Taylor series exist
X X X X
Rj = rj,n tn , u(t) = qn tn , u̇(t) = (n + 1) qn+1 tn , ü(t) = (n + 1)(n + 2) qn+2 tn
n=0 n=0 n=0 n=0

When you multiply two infinite series, you can consolidate the results into a single series
X  X  X X
n 
An t n An tn = Am Bn−m tn
n=0 n=0 n=0 m=0

For example
X X
n 
R3 u = qm r3,n−m tn
n=0 m=0
X X
n 
R2 u̇ = (m + 1) qm+1 r2,n−m tn
n=0 m=0
X X
n 
R2 ü = (m + 1)(m + 2) qm+2 r1,n−m tn
n=0 m=0

and if all three parts are combined, we obtain

X X
n  
qm r3,n−m + (m + 1) qm+1 r2,n−m + (m + 1)(m + 2) qm+2 r1,n−m tn = 0
n m=0

This is a set of simple linear equations that can be solved order by order for the unknown qm in terms of the known
rj,n . In the first equation there are three coefficients q0 , q1 , q2 , we are free to choose q0 and q1 as constants of inte-
gration, and determine q2 from them
q0 r3,0 + q1 r2,0
n = 0; q0 r3,0 + q1 r2,0 + 2q2 r1,0 = 0, q2 = −
2r1,0
(provided r1,0 6= 0). The next order will involve only one unknown
n = 1; (q1 r3,0 + q0 r3,1 ) + (2q2 r2,0 + q1 r2,1 ) + (6q3 r1,0 + 2q2 r1,1 ) = 0
from which we get q3
(q1 r3,0 + q0 r3,1 ) + (2q2 r2,0 + q1 r2,1 + 2q2 r1,1 )
q3 = −
6r1,0
and this process continues. We generate a two-parameter set of analytic series solutions to the differential equations,
no non-analytic points of any kind. If r1,0 = 0, we simply make appropriate modifications but we can still carry out
this simple-minded procedure (the method of Frobenius without indicial) for solving the equation of motion.

Many if not all of these equations are named, a very short list of some of the most common named special functions
of mathematical physics is given below, greatly expanding on the three I gave in class. There are many others, these
are some that you have worked with extensively. All of these functions are regular, having no singularities in the
t-variable, real or complex. Of course you can create rational combinations that do have poles. For example the
function coth(t) has a pole, but it is not the solution of a linear differential equation with regular coefficients.

81
 R1 (t) R2 (t) R3 (t) Name
1 0 c Trig/hyperbolic (sin(t), cos(t), sinh(t), cosh(t))
1 − t2 −2t n(n + 1) Legendre Pn (t)
1 −2t 2n Hermite Hn (t)
t (a + 1 − t) n Laguerre Lan (t)
(1 − t2 ) −t n2 Chebyshev Tn (t)
(a,b)
(1 − t2 ) (b − a − (a + b + 2)t) n(n + a + b + 1) Jacobi Pn
t2 t (t2 − ν 2 ) Bessel Jν (t), Yν (t)
(1 − t2 )2 −2t(1 − t2 ) ν(ν + 1)(1 − t2 ) − µ2 Spherical harmonic Yνµ (t)
t(1 − t) (c − (a + b + 1)t) −ab Hypergeometric F (a, b; c; t)
t (c − t) −a Confluent hypergeometric Φ(a, c; t)
But what about nonlinear equations? The differential equation
du u 1
=− , u=
dt t−c t−c
has a non-movable singularity (location is fixed by the differential equation itself). Note that R3 (t) is not regular, it
has a pole at t = c. The non-linear equation
du 1
= −u2 , u=
dt t−c
has a movable singularity (the location is an unspecified constant of integration) in the form of a pole. The non-linear
equation
du 1 √
= , u= t−c
dt 2u
has a movable singularity (point of non-analyticity) in the form of a branch point whose location is a constant of
integration.

It turns out that all non-linear differential equations whose only singularities are movable poles (Painleve
property) have been classified just as the special functions (solutions to linear equations) have been classified and
their solutions are well-known. There are only 50 differential equations with the so called Painleve property (which
is that the only movable singularities are poles). Forty-four of these equations have solutions that are expressible
in terms of common special functions such as trigonometric, exponential, elliptic, or orthogonal polynomial. For
example u̇ = c2 − u2 has solution u = c coth(c(t − t0 )), which has a movable singularity at t = t0 . Six of the 50
cannot be solved in terms of named special functions, these are the six Painleve transcendents, whose solutions
can be built up by Laurent series. The Painleve transcendents are (in order)

d2 u d2 u
P − 1; = 6u2 + z, P − 2; = 2u3 + zu + a
dz 2 dz 2
d2 u 1  du 2 1 du 1 2 3 d d2 u 1  du 2 3 3 b
P − 3; = − + (au + b) + cu + , P − 4; = + u + 4zu2 + 2(z 2 − a)u +
dz 2 u dz z dz z u dz 2 2u dz 2 u
the remaining two are a bit more complicated, and I hade to look up P-3 and P-4 as it is . You don’t see these much
in physics outside of statistical mechanics (but you see everything there); the reduced density matrix for interacting
bosons on a ring is given by an integral of the fifth Painleve’ transcendent.

Determining whether or not a system of evolution equations can be reduced to any of the 50 solvable categories is
tantamount to determining integrability, meaning that the equations have an attainable solution by quadrature. The
procedure to determine this integrability is called the Painleve test. The test begins with a theorem;

Theorem 1. A necessary condition that an autonomous system of equations (no explicit dependence on the inde-
paendent variable)
dui
= g(u1 , u2 , · · · , uN )
dz

82
with g(u) a rational function of u has the Painleve property is that there is a Laurent series
∞
X
ui (z) = (z − z0 )m ai,n (z − z0 )n
n=0

where N − 1 coefficients ai,n are arbitrary (the constants of integration). The expansion is only a valid Laurent
expansion if m is a negative integer. Therefore step number one is to determine this leading exponent m. If m
can be found integral, and the requisite N − 1 constants ai,n found, the system of equations passes the Painleve test
and it is a foregone conclusion that it can be reduced to one or more of the 50 solvable differential equations, and
integrated to a final solution.

Back to the main thread...

Mathematical physicists (first Kowalewski, then Painleve’) introduced a revolutionary technique, eventually called
singular point analysis, to the rigid-body motion problem and were the first to find new and very complex solutions
that could not be found by existing techniques. Singular point analysis is the non-linear differential equation
equivalent of the method of Frobenius for linear equations, it is however very much more sophisticated and
powerful, because it can reveal the order of polynomial conserved quantities which may be used to achieve quadra-
ture. At the core of singular point analysis one finds conditions under which solutions can be expanded in Laurent
series.

Let’s examine a somewhat artificial Euler problem (no torques) but nonphysical principal moments, cyclically permute

ω̇1 = ω2 ω3 , ω̇2 = ω2 ω3 , ω̇3 = ω1 ω2

First note that these have a particular solution

ω1 = a1 τ −1 , ω2 = a2 τ −1 , ω3 = a3 τ −1

that is singular with a movable pole singularity τ = t − t0 of weight 1. Next we find that

ω1 ω̇1 = ω2 ω˙2 = ω3 ω˙3 = ω1 ω2 ω3

and so we subtract, obtaining two constants of the motion

d 2  d 2 
ω1 − ω22 = 0, ω12 − ω22 = I1 , ω1 − ω32 = 0, ω12 − ω32 = I2
dt dt
of weight two (quadratic in weight-one variables). Insert these into the first equation of motion
q
ω̇1 = ω2 ω3 = (ω12 − I1 )(ω12 − I2 )

and we have reduced the problem to a quadrature

Z
dω1
t − t0 = p
(I1 − ω12 )(I2 − ω12 )

A simple change of variables turns this into an elliptic integral;

p p Z
dθ
ω1 = I1 sin θ, I2 (t − t0 ) = q
(1 − I1
I2 sin2 θ)

The usual question is ”What do we need to know about elliptic integrals?”. They are nice periodic functions like
trig functions, they arise in similar contexts, but they are far richer in properties.
Consider a simple pendulum with total energy
θ0
E = mgl(1 − cos θ0 ) = 2mgl sin2
2

83
and equation of motion
g
θ̈ = − sin θ
l
Multiply by θ̇ and integrate to r
Z
dθ 2g
√ = t
cos θ − cos θ0 l
and let
θ
x = sin
2
then Z r
dx 2E θ0
p = t = u, with k −1 = sin
(1 − x2 )(1 − k 2 x2 ) ml2 2
Rx R sin(t)
This defines the Jacobi elliptic function sn(u) in the same way that sin−1 (x) = 0
√ dy = t, or 0
√ dy =t
1−y 2 1−y 2

Z sn(u)
dx
p =u
0 (1 − x2 )(1 − k 2 x2 )

and its inverse, the complete elliptic integral (first and second types respectively)
Z θ
dφ
F (θ, k) = q = u, sn(u) = sin(θ), cn(u) = cos(θ)
0 (1 − k 2 sin2 φ)
Z θ q Z u Z u
2 0 0
E(θ, k) = (1 − k2 2
sin φ) dφ = dn (u ) du = u − k 2
sn2 (u0 ) du0
0 0 0

The integrand has non-analytic points at ±1, and ± k1 . We will can obtain the periodicity properties from the integral
itself, with the use of complex variables techniques. Elliptic integrals have the property that they are singular at
complex values of their variable, with a simple pole, and tis property is exploited by singular point analysis, the
method invented by Sonya Kowalewski to attack the rigid body problem.

The idea of singular point or Painleve analysis is to determine the order of polynomial conserved quantities, and by
studying the behavior of Laurent series solutions of equations of motion near singular points, to determine whether
or not it will be possible to reduce the system of equations of motion to any of the limited number of differential
equations whose solutions possess movable singularities that are simple poles, and no other singularities. Kowalewski
found a fourth case that has a high-order polynomial fourth conserved quantity for the rigid body problem by these
methods.

Painleve analysis has three stages; determining the exponent of the leading or dominant singularity, determining the
resonances, and determining the constants of integration. If any step fails, the system of equations fails the test,
and solutions are not likely to be found conforming to the 50 classes of differential equations with pole-singularities.
On the surface it looks like the process gives you a series solution, and it is hard to recognize what a series adds
up to, but actually the end result is much more. Painleve analysis can be used to determine the weights of (and
existence of) polynomial constants of the motion, and with a sufficient number of these we can reduce the system to
quadrature. You should think of this as being the primary benefit of Painleve analysis.

Let me reiterate this important point: The purpose of singular point analysis is not to obtain a series solution.
Series solutions are not practical, they are not easy to calculate with. A formal series does not need to be convergent
in order to be useful, so the fact that one can’t graph a solution from a formal series may not be relevent here, but the
fact remains that the series solution is not the goal. The real purposes are to determine orders of polynomial
invariants and the conditions under which a full quadrature is possible. The former connection between
components of the Painleve test (the dominant singularities) was fully established in 1983, so I am showing you some
fairly modern advances in theoretical physics.

84
16.4 Singular point analysis
What makes negative powers of the expansions so important? Poles of different orders cannot cancel one another.
This fact can be used to establish the Painleve test in stages. The test establishes whether or not the solutions to
an autonomous sytem of equations of motion will be describable in terms of the 50 Painleve’ types.
The first stage of the test is to find m, the order of the dominant (pole) singularity, and to do so it is sufficient
to insert only the leading summand of the Laurent series into each equation.
For example consider the Euler-like system
du1 du2 du3
= u2 u3 , = u1 u3 , = u1 u2
dt dt dt
The symmetry under interchanges ui ↔ uj indicates that all of the variables have the same leading singularity order,
so insert
ui = αi (t − t0 )m + · · ·
into each to get

mα1 (t − t0 )m−1 + · · · = α2 α3 (t − t0 )2m + · · ·

m−1
mα2 (t − t0 ) + ··· = α1 α3 (t − t0 )2m + · · ·
mα3 (t − t0 )m−1 + · · · = α1 α2 (t − t0 )2m + · · ·

indicating m−1 = 2m or m = −1. We note that the most non-linear terms and highest derivatives determine
m.
What if this stage gives m non-integral? Then the test fails at the very first stage and the equations do not possess
solutions in any of the 50 differential equations in the test category. The equations may possess more complicated
solutions, such as fractional power series (Puiseaux series, the currency of algebraic geometry), or logarithmic ψ-
series.
Notice that the monomial comprising the dominant singularity may in fact be a particular solution
of the full equations of motion, depending on whether or not the equations possess certain properties which we will
discuss in a moment. This particular solution may not appear to be useful since it is singular, but a solution to an
EOM is a solution after all and will prove to be very handy indeed.
Example. Consider the equations of motion
du1 du2
= u32 , = −u31
dt dt
The Painleve test first stage indicates m = − 12 , and so fails. The equations cannot be reduced to one of the 50
standard equations, but this information can be used to develop series
∞
X ∞
X
1 1
u1 = (t − t0 )− 2 ai (t − t0 )i , u2 = (t − t0 )− 2 bi (t − t0 )i
i=0 i=0

which are solutions. The fact that the squares of these functions have integral leading singularities suggests a route
to a full quadrature that you will explore in homework 6 .

The second stage of the Painleve test is to determine at which levels (orders) the cancellation order-by-order of
singularities in the equations of motion will lead to unspecified (free) constants. These orders are called resonances,
and will be closely connected to conserved quantities of the autonomous system. To establish a resonance r, we make
substitutions
ui = αi (t − t0 )m + βi (t − t0 )m+r
into the equations of motion, using our previously determined m and αi values from the previous stage. We then
demand cancellations of the next to highest order singularities. This leads to homogeneous linear equations in
the βi coefficients, and if the determinant vanishes for a given r, then for that exponent one of the βi coefficients
will be unspecified. This gives our required constants of integration. The Painleve test fails if a resonance r is
non-integral. You will always find that −1 is a resonance. Just knowing the resonances will prove to be valuable.

85
The third stage is to establish these constants (appearing at orders given by the resonances) and to show that
beyond a certain order all series coefficients can be shown to be functions of them. For N equations of motion we
must establish N − 1 such constants of integration in accordance with the theorem. Order by order we obtain sets
of linear equations for series coefficients.
Consider the linear homogeneous equations
    
x a b x
0 = ax + by, 0 = cx + dy or M = =0
y c d y

Solve for x;
x (ad − bc) = x det(M ) = 0
Either x or det(M ) is zero. If det(M ) 6= 0, then x must be, and so must be y, the only solution is trivial. On the
other hand if det(M ) = 0, x can be anything, and we then calculate y from the free value of x by y = − ab x = − dc x.
In other words one of the variables is a free parameter.

The Painleve test is very easy to code up in a computer algebra system, it was the first problem that I ever used a
CAS (REDUCE) for.

Example. Establish the resonances for

du1 du2 du3
= u2 u3 , = u1 u3 , = −u1 u2 ,
dt dt dt
for which
m = −1, −α1 = α2 α3 , −α2 = α1 α3 , −α3 = −α1 α2
Take principle-branch logarithms;

πi + ln α1 = ln α2 + ln α3 , πi + ln α2 = ln α1 + ln α3 , ln α3 = ln α1 + ln α2

and solve; we find that

πi πi
ln α1 = , α1 = i, ln α2 = , α2 = i, α3 = −1
2 2
Now we insert

u1 = α1 (t − t0 )−1 + β1 (t − t0 )r−1 + · · ·
u2 = α2 (t − t0 )−1 + β2 (t − t0 )r−1 + · · ·
u3 = α3 (t − t0 )−1 + β3 (t − t0 )r−1 + · · ·

into the equations of motion. As long as the αi ’s obey the three constraints from the first level, the pole of order 2
cancels. The equations are

−α1 τ −2 + (r − 1)β1 τ r−2 + · · · = α2 α3 τ −2 + β2 α3 τ −2+r + β3 α2 τ −2+r + β2 β3 τ −2+2r + · · ·

−α2 τ −2 + (r − 1)β2 τ r−2 + · · · = α1 α3 τ −2 + β1 α3 τ −2+r + β3 α1 τ −2+r + β1 β3 τ −2+2r + · · ·
 
−α3 τ −2 + (r − 1)β3 τ r−2 + · · · = − α2 α1 τ −2 + β2 α1 τ −2+r + β1 α2 τ −2+r + β2 β1 τ −2+2r + · · ·

The first terms on each side cancel by virtue of stage one, the last term on the right in each equation is not the
lowest order term in τ = t − t0 , and so the stage two equations are

(r − 1)β1 τ r−2 = β2 α3 τ −2+r + β3 α2 τ −2+r

(r − 1)β2 τ r−2 = β1 α3 τ −2+r + β3 α1 τ −2+r
 
(r − 1)β3 τ r−2 = − β2 α1 τ −2+r + β1 α2 τ −2+r

86
These are three homogeneous equations in three unknowns β1 , β2 , β3 . Vanishing of the determinant of coefficients
will determine at what r-values a βi will be unspecified (free), the matrix is not invertible and so a non-trivial
solution of the linear equations exists.
 
 1−r α3 α2 

 α3 1−r α1  = 0, (1 − r)3 − 2α1 α2 α3 − (1 − r)(−α12 − α22 + α32 ) = 0

 −α2 −α1 1 − r 

however our α-relations reduce this to

0 = λ3 − 3λ − 2 = (λ + 1)(λ + 1)(λ − 2), λ=1−r

We discover a resonance at −1 and a double at 2. There is always a resonance at r = −1. There is an unspecified
constant lurking in the set {α1 , α2 , α3 }. But didn’t we find specific values for these? Yes, but we did so by choos-
ing a branch of the logarithm, we could have chosen any branch. A concrete example carried to completion is in order.

16.5 Execution of the Painleve test for Euler’s equations

The equations of motion
L̇1 = u2 L2 L3 , L̇2 = v 2 L1 L3 , L̇3 = w2 L1 L2
with A2 + B 2 + C 2 = 0 can be completely solved by Painleve’ analysis. The result will be Laurent series into which
we can insert any initial conditions and obtain any solution, but the series solutions themselves are secondary to
what the test gives us. Begin with a Laurent series ansatz such as;

L1 = a0 (t − t0 )−1 + a1 + a2 (t − t0 ) + a3 (t − t0 )2 + · · ·
L2 = b0 (t − t0 )−1 + b1 + b2 (t − t0 ) + b3 (t − t0 )2 + · · ·
L3 = c0 (t − t0 )−1 + c1 + c2 (t − t0 ) + c3 (t − t0 )2 + · · ·

and perform the first stage; elimination of the dominant divergences. Computer symbolic manipulation is especially
useful for this; I will use REDUCE, which translates easily into C++ using GiNaC or into Mathematica. The basic
program code listing is below (using x = (t − t0 )) and u, v, w for the ”moments of inertia”. I ran this before class
today...

operator a,b,c; %1
L1:=for n:=-1:4 sum a(n)*x^n; %2
L2:=for n:=-1:4 sum b(n)*x^n; %3
L3:=for n:=-1:4 sum c(n)*x^n; %4
EOM1:=x^2*(df(L1,x)-u^2*L2*L3); %5
EOM2:=x^2*(df(L2,x)-v^2*L1*L3); %6
EOM3:=x^2*(df(L3,x)-w^2*L1*L2); %7
Lev0:=solve({coeffn(EOM1,x,0),coeffn(EOM2,x,0),
coeffn(EOM3,x,0)},{a(-1),b(-1),c(-1)}); %8
let first(Lev0); %9
Lev1:=solve({coeffn(EOM1,x,1),coeffn(EOM2,x,1),
coeffn(EOM3,x,1)},{a(0),b(0),c(0)}); %10
let first(Lev1); %11
Lev2:=solve({coeffn(EOM1,x,2),coeffn(EOM2,x,2),
coeffn(EOM3,x,2)},{a(1),b(1),c(1)}); %12
let first(Lev2); %13
Lev3:=solve({coeffn(EOM1,x,3),coeffn(EOM2,x,3),
coeffn(EOM3,x,3)},{a(2),b(2),c(2)}); %14
let first(Lev3); %15
Lev4:=solve({coeffn(EOM1,x,4),coeffn(EOM2,x,4),
coeffn(EOM3,x,4)},{a(3),b(3),c(3)}); %16

87
Step 1. Run statements #1 − #7 through REDUCE to get set up. Then run statement #8, which will find all
solutions of the nonlinear equations for the coefficints of the leading singularity. Run statement #9 to accept the
first solution
1 1 1
a(−1) = , b(−1) = , c(−1) = −
vw uw uv
and substitute the results into all occurances of these variables in all objects.

Step 2. Run statement #10 through REDUCE. You find that the determinant of the equation for the coefficients
a(0), b(0), c(0) is 2, so there is only the trivial solution; the list ”Lev1” of solutions contains obly this choice a(0) =
0, b(0) = 0, c(0) = 0 , so run statement #11 to accept it and make the stubstitutions.
Here is the equation for the coefficients (REDUCE is not case-sensitive);
  
0 U V −1 −U W −1 a(0)
 V U −1 0 −V W −1   b(0)  = 0
−1 −1
−W U −W V 0 c(0)

Step 3. Run statement #12 through REDUCE. The determinant for the equations determining a(1), b(1), c(1) is 0,
  
1 U V −1 −U W −1 a(1)
 V U −1 1 −V W −1   b(1)  = 0
−1 −1
−W U −W V 1 c(1)

so there are solutions. There is in fact a two-parameter family of solutions

u*(arbcomplex(5)*v - arbcomplex(4)*w)
lev2 := {{a(1)=---------------------------------------,
v*w

b(1)=arbcomplex(4),

c(1)=arbcomplex(5)}}
This stage of the calculation introduces two constants of integration, and 2 = 3 − 1 so this is all we need to pass the
test. I pick pick a(1), b(1) to be independent.
This is in complete accord with the fact that a resonance occurs at r = 2, which corresponds to the terms
∝ (t − t0 )−1+2 = (t − t0 )1 in the Laurent expansions. run statement #13 to accept it and make the substitutions.

That successfully concludes the Painleve test, we run statement #14 to find that a(2) = 0, b(2) = 0, c(2) = 0 because
the determinant of the homogeneous equation for them is 4
  
2 U V −1 −U W −1 a(2)
 V U −1 2 −V W −1   b(2)  = 0
−1 −1
−W U −W V 2 c(2)

and to accept the results we run the ”let clause” and make substitutions.

Step 4 and beyond. After this point we encounter only inhomogeneous equations for the coefficients, and
find all coefficients from this point on are functions of the constants gotten in step 3.
    
3 U V −1 −U W −1 a(3) a(1)b(1)U W + b2 (1)U 2 W/V
 V U −1 3 −V W −1   b(3)  =  a(1)b(1)V W + a2 (1)V 2 W/U 
−1 −1
−W U −W V 3 c(3) a(1)b(1)W 2

This process can be continued, giving all coefficients. These particular Laurent series, with a single complex pole
at t0 complex, is characteristic of elliptic functions. In fact once a few of the functions are carefully written out, we
can recognize the Laurent series of certain elliptic integrals taking shape.

88
The final structure of the Laurent series will be

ui = ai τ −m + ci,r1 τ −m+r1 + · · · + ci,rk τ −m+rk + R(ai , ci,r1 , · · · , ci,rk , τ )

in which m is the leading singularity exponent, rj are the ordered resonances, ci,rj are the free parameters (constants
of integration), and R has a Taylor series whose lowest term begins with a power of τ greater than the exponent of
the last resonance, and all of its coefficients are functions of ai , ci,r1 , · · · , ci,rk .
I can’t say this too often; the goal is not to get a series solution. It is to reveal the conditions for a
full quadrature and to establish the orders of the polynomial conserved quantities needed to achieve it.

How do the series generated by carrying out the Painleve’ test reveal the orders of the polynomial conserved quan-
tities needed to achieve quadrature?

Next we explore the following connections

• Particular solutions of the equations of motion (that may be very unphysical) that just happen to be singular
functions of pole type with a moveable singularity, and
P
• perturbations around these solutions that look like the rest of u = τ −m n=0 an τ n , and the connection between
orders n = r in the series with the weights of polynomial first integrals, and te order in the series at which an
unspecified constant is introduced that subsequent coefficients are functions of.

89