J.

Symbolic Computation (1996) 22, 459468

Symbolic Solutions for a Class of Partial Dierential
Equations
BRAM DE JAGER

AND BRAM VAN ASCH

Faculty of Mechanical Engineering,

Faculty of Mathematics and Computer Science,

Eindhoven University of Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands
(Received 12 June 1995)
An algorithm to generate solutions for members of a class of completely integrable partial
dierential equations has been derived from a constructive proof of Frobenius Theorem.
The algorithm is implemented as a procedure in the computer algebra system Maple.
Because the implementation uses the facilities of Maple for solving sets of ordinary
dierential equations and for sets of nonlinear equations, and these facilities are limited,
the problems that actually can be solved are restricted in size and complexity. Several
examples, some derived from industrial practice, are presented to illustrate the use of the
algorithm and to demonstrate the advantages and shortcomings of the implementation.
c 1996 Academic Press Limited
1. Introduction
Solving systems of dierential equations is a dicult problem, except for some special
classes. General purpose CAS (computer algebra systems) do not provide facilities that
can be applied generally. Maples command dsolve, for instance, will provide full so-
lutions only to a restricted class of ODEs (ordinary dierential equations). The same
holds for other CAS.
A facility provided by Maple is the liesym package that computes Lie point symmetries
of dierential equations. This package is not integrated in the dsolve command. Most
other CAS provide equivalent facilities. For ODEs these Lie point symmetries, or if
desired other symmetries, can be used to reduce the order of the ODE. The following
example is taken from Stephani (1989, p. 30).
Consider the second order dierential equation
y

= (x y)(y

)
3
.
By considering the Lie point symmetry
x = x + (x y), y = y

E-mail: A.G.de.Jager@wfw.wtb.tue.nl

E-mail: wsinaa@win.tue.nl
07477171/96/100459 + 10 $25.00/0 c 1996 Academic Press Limited
460 B. de Jager and B. van Asch
we arrive at a change of coordinates
t = y, s = ln(x y)
which transforms the ODE into
s

+ (s

)
2
+ 1 = 0.
So, we have reduced the order by 1, because this is a rst order ODE in s

.
In general the following holds. If we have an ODE of order n, and there are n Lie point
symmetries, we can reduce the problem to order 0, provided that the Lie group generated
by the symmetries is solvable.
For PDEs (partial dierential equations) the situation is dierent. Here, Lie point
symmetries cannot be used to lower the order of the dierential equation. Instead, they
might be used to reduce the number of independent variables, and thus gain access to
special classes of solutions. Eventually this might even lead to an ODE. In Stephani
(1989), for example, it is shown how the wave equation

2
u
x
2
+

2
u
y
2
+

2
u
z
2


2
u
t
2
= 0
can be reduced to the ODE

d
2
w
d
2
+
dw
d
= 0, =
x
2
+ y
2
t
2
z
2
, w = u
using symmetries. However, in general the problem of nding Lie point symmetries for
PDEs is of equal diculty to solving the dierential equation.
Having demonstrated the procedure to use symmetries in solving (partial) dierential
equations, the question arises if this technique has any practical value. As answers to this
question several dierent points of view have been encountered in the literature. Stephani
is rather pessimistic. In his book (Stephani, 1989) it is noted that Lie symmetries have
hardly ever any value for obtaining explicit solutions for ODEs or PDEs. However, since
the publication of Stephanis book, a number of articles appeared using certain kinds of
symmetries in nding exact solutions of PDEs. In MacCallum (1995) it is remarked that
Lie symmetries are at least an ordering principle for obtaining solutions, and results of
a computer algebra program are summarized that could solve a large part of a set of
reference ODEs using Lie symmetries. Also, Schwarz and Wolf (MacCallum, 1995) use
Lie symmetries as an integral part of their programs to compute explicit solutions of
ODEs and PDEs. A lot of other work in this area has been performed. Furthermore,
Lies symmetry method is the method that is generally applicable to generate exact
solutions, so it certainly has its value.
We do not delve into this further, but conclude that other approaches than Lie sym-
metries to solve certain classes of PDEs are worthwhile avenues, so other ways to obtain
solutions have to be searched for. This comes down, in most cases, to restrict the class
of PDEs one wants to consider, and to develop rather specic methods for this class.
For special types of PDEs, as the ones to be considered here, the method of character-
istics is an alternative, because it is particularly well suited for coupled systems of rst
order (Kevorkian, 1993; Sneddon, 1957). In this paper, however, yet another method,
based on a constructive proof of Frobenius Theorem, is discussed. The method is suit-
able for sets of rst order PDEs that are completely integrable, or, in other words, for
which a certain distribution is involutive, as stated by Frobenius Theorem. Note that
Symbolic Solutions for a Class of Partial Dierential Equations 461
this type of PDE is frequently encountered in, at least, nonlinear system theory in a dif-
ferential geometry setting, and has therefore great practical signicance. For applications
in this area we are content with nding a single closed form solution, but without user
intervention.
2. Frobenius Theorem
This section and the next one draws from Isidori (1989). Starting with sets of rst
order PDEs of the following form
h(x)
x
_
f
1
(x) f
d
(x)

=
_
0 0

(2.1)
with x R
n
,
h(x)
x
the row vector [
h
x
1

h
x
n
], and f
i
a smooth vector eld (i = 1, . . . , d),
we want n d independent solutions h(x) of this set of PDEs. This is equivalent to the
complete integrability of the distribution (x) spanned by the columns f
i
, i = 1, . . . , d,
with assumed to be nonsingular. Frobenius Theorem now states
A nonsingular distribution is completely integrable if and only if it is involutive.
The constructive proof of the suciency part of this Theorem in Isidori (1989, pp. 26
30) is used to solve the PDEs.
3. Algorithm
The solution algorithm used can be described by the following steps.
1. Test the involutiveness of the distribution (x) = span{f
1
(x), . . . , f
d
(x)}.
2. Extend the distribution with vector elds f
d+1
, . . . , f
n
such that the n vector
elds f
1
, . . . , f
n
span R
n
.
3. Compute the ows
f
i
t
(x

) for the vector elds f

i
, i = n, . . . , 1. The ow x(t) =

f
t
(x

) of a vector eld f is by denition the solution of the initial value problem

dx
dt
= f with initial condition x(0) = x

.
4. Determine the mapping F : R
n
R
n
dened by composing the ows:
F(z
1
, . . . , z
n
) =
f
1
z
1

f
n
z
n
(x

).
Here denotes composition with respect to x. So

f
n1
z
n1

f
n
z
n
(x

) =
f
n1
z
n1
(
f
n
z
n
(x

))
etc.
5. The last n d components of the inverse mapping F
1
are independent solutions
of the PDEs.
4. Implementation
The steps of the solution algorithm are implemented in the new procedure psolve as
follows.
462 B. de Jager and B. van Asch
1. The involutiveness of the distribution is tested by checking if the rank of the ma-
trices
_
f
1
f
d
[f
i
, f
j
]

for 1 i < j d, where [f
i
, f
j
] =
f
j
x
f
i

f
i
x
f
j
denotes the Lie bracket of f
i
and f
j
, does not increase and is still equal to rank [ f
1
f
d
] = d.
2. The extension is performed iteratively, by augmenting the matrix [ f
1
f
d
] with a
single unit column (the simplest vector eld), testing the rank, if the rank increases
this column is kept and the next unit column is added until the rank is n.
3. The ow of the vector elds is computed with the dsolve procedure.
4. The construction of the mapping F, i.e., the composition of the ows, is performed
by backward substitution of the computed ows.
5. The inverse mapping F
1
(the solution of a set of nonlinear algebraic equations)
is computed with the solve procedure and the solution of the PDEs is extracted
from this mapping.
Several checks are built in to test if the input data is correct, e.g., if the distribution
is nonsingular, to test if the computations are performed without errors, and nally to
test the solution h(x). The rank determination in the input check and in steps 1 and 2
is done with Gauss elimination techniques and thus gives a generic rank, i.e., for a lower
dimensional subset in R
n
the rank may drop.
We remark that the solution for (2.1) may not be unique. In general, psolve will only
compute a single closed form solution, which is all we require in our use of this procedure,
that may not be the most simple or insightful one. This is illustrated in the examples.
It is also evident from this sketch of psolve that critical parts in the implementation
are the dsolve and solve procedures that should perform some of the hard work. The
dsolve function is (partly) based on an implementation of the Risch integration algo-
rithm (Geddes et al., 1992). The main reasons to set up the implementation in this way
are (1) ample use is made of existing code, (2) any improvement made to the dsolve
and solve procedures will also improve psolve.
The procedures dsolve and solve sometimes fail to give an appropriate answer. The
capabilities and limitations of the dsolve procedure are, for instance, discussed in Postel
and Zimmermann (1996). This review also shows that, although not perfect, the dsolve
procedure of Maple is reasonably powerful compared with the facilities of other CAS.
The limitation of the solve procedure to solve conveniently sets of nonlinear equations
is illustrated in example 5 in the next section.
5. Examples
This section presents ve examples. Four for which an explicit solution can be found,
and a last one for which an explicit solution is known to exist but cannot be computed
directly by psolve. The rst and last example are computed with Maple V Release 3,
the other examples with Maple V. For all examples it is straightforward to check the
answers by hand.
Symbolic Solutions for a Class of Partial Dierential Equations 463
Example 1. The single PDE that has to be solved for a nontrivial h = h(x
1
, x
2
)
is (Isidori, 1989, Example 1.4.1)
h(x)
x
_
e
x
2
1
_
= 0. (5.1)
Note that the partial dierential equation is linear in h(x) but the coecients of the
derivatives are not all polynomial in x. The solution h(x) is computed by Maple as
follows.
> n := 2:
> x := array(1..n):
> f := array(1..n,1..1,[ [exp(x[2])], [1] ]):
> xinit := {seq (x[i]=0,i=1..n)}:
> psolve(f,x,xinit);
[ x[1] - exp(x[2]) + 1 ]
Note that the command psolve needs a description of the dierential equation in f,
the name of the independent variable x in x, and the initial condition x

in xinit. The
last line of the output contains the desired solution,
h(x) = x
1
e
x
2
, (5.2)
where the constant 1 can be neglected. The correctness of the solution can be veried
directly by substitution of (5.2) in the dierential equation (5.1).
Example 2. The set of PDEs is (Isidori, 1989, Example 1.4.3)
h(x)
x
_
_
2x
3
x
1
1 2x
2
0 x
3
_
_
=
_
0 0

. (5.3)
The solution h(x) is computed by psolve as follows.
> n := 3:
> x := array(1..n):
> f := array(1..n,1..2,[[2*x[3], -x[1]], [-1, -2*x[2]], [0, x[3]]]):
> xinit := {seq (x[i]=0,i=1..n-1), x[n]=1}:
> psolve(f,x,xinit);
2
[ x[1] x[3] + 2 x[3] x[2] ]
The solution generated is correct, as can be checked by substitution in (5.3) of the
solution h(x) = x
1
x
3
+ 2x
2
x
2
3
.
For arbitrary initial conditions, generated by
> xinit := {seq (x[i]=x0[i],i=1..n)}:
the computed solution becomes
2 2
x[1] x[3] + 2 x[3] x[2] - 2 x0[3] x0[2] - x0[3] x0[1]
[ ------------------------------------------------------- ]
x0[3]
464 B. de Jager and B. van Asch
which shows that for x

3
= 0 this solution is not valid.
If in the set xinit the condition for x

3
is set to 0 the following error results.
Warning, (in psolve) MAPLE failed in testing the inverse mapping of F,
[ locz3 exp(- locz2), - locz1, 0 ], for component, 1, 2, 3
although this inverse was computed by solve.
This is a limitation of the implementation, that does not account for restrictions on the
initial conditions. It is easy to see that in fact a unique inverse does not exist, because
the Jacobian of the mapping F is singular for x

3
= 0.
Example 3. The problem considered stems from the analysis of the exact linearizability
by state feedback and coordinate change of the dynamics of a DC motor driving a load
with moment of inertia J. The relevant set of PDEs is given by
h(x)
x
_
_
1/L
s
R
s
/L
2
s
0 Kx
3
/L
r
0 Kx
2
/J
_
_
=
_
0 0

that has to be solved for h(x). The indexed L, R, and the constant K are the inductance,
resistance, and motor constant respectively. The motor inertia is lumped with the load
inertia J. The indices
s
and
r
stand for stator and rotor. The components of the state x
are stator and rotor currents and rotor speed. The solution can be computed by psolve
according to the following session log.
> n := 3:
> x := array(1..n):
> f := array(1..n,1..2,[[1/Ls, Rs/Ls^2], [0, K*x[3]/Lr], [0, -K*x[2]/J]]):
> xinit := {seq (x[i]=0,i=1..n)}:
> psolve(f,x,xinit);
2 2 1/2
(J x[3] + Lr x[2] )
[ ------------------------ ]
1/2
J
It is easy to see that, e.g., h(x) = L
r
x
2
2
+ Jx
2
3
, an energy type function, is also a
solution, so the simplest form for h is not necessarily found.
Example 4. This set of PDEs is derived from the input matrix of a model for a robot
with one translational and one rotational joint and occurs in computing the transforma-
tion to a normal form of the equations of motion of the robot.
h(x)
x
_
_
1 0
sinx
2
0
0 1
_
_
=
_
0 0

.
Note that here a non-polynomial function in x is involved. Initially, a solution cannot be
computed, as shown by the following log.
> n := 3:
> x := array(1..n):
Symbolic Solutions for a Class of Partial Dierential Equations 465
> f := array(1..n,1..2,[[1, 0], [sin(x[2]), 0], [0, 1]]):
> xinit := {seq (x[i]=0,i=1..n)}:
> psolve(f,x,xinit);
psolve:, unable to compute coeff,
MAPLE procedure dsolve failed in solving the flow of: ,
[ 1, sin(locx2(locz1)), 0 ], continuing with extdsolve
The message is due to limitations of the dsolve procedure, used to compute the ows,
to solve sets of ODEs. To remedy this, an extension named extdsolve was written, that
is able to solve the ow in this case (this extension was also needed to solve Example 1).
Now the following appears later on in the computation.
Warning, (in psolve) MAPLE failed in testing the inverse mapping of F,
exp(locz1) (- 1 + cos(x0[2] + locz3))
locz1+x0[1], -2 arctan(-------------------------------------), x0[3]+locz2
sin(x0[2] + locz3)
for component, 2, although this inverse was computed by solve.
(- 1 + cos(x[2])) exp(- x[1] + x0[1])
[ - x0[2] - 2 arctan(-------------------------------------) ]
sin(x[2])
The last line contains the desired answer
h(x) = 2 arctan
_
tan
x
2
2
e
x
1
+x

1
_
x

2
because
1 + cos x
2
sinx
2
= tan
x
2
2
.
The warning is due to limitations in testing if the inverse mapping F
1
as computed by
the solve procedure is correct. In this case it has no consequences, but the warning has
the intention to generate some caution with respect to the correctness of the answer.
Example 5. Consider the single PDE which is a generalization of the rst PDE in
Example 4
(x
1
)
h
x
1
+ (x
2
)
h
x
2
= 0
i.e.,
f
1
(x) =
_
(x
1
)
(x
2
)
_
.
Choose
f
2
(x) =
_
1
0
_
.
It is obvious that

f
2
t
=
_
t + x

1
x

2
_
.
466 B. de Jager and B. van Asch
The computation of
f
1
t
(x

) comes down to solving the two ODEs

dx
1
dt
= (x
1
) and
dx
2
dt
= (x
2
).
In both cases the variables can be separated. A rst problem may arise when we try to
express
_
1
(x
1
)
dx
1
and
_
1
(x
2
)
dx
2
in terms of elementary functions. This may not be possible, but if we are using CAS this
problem can be solved by adding the integrals of 1/(x
1
) and 1/(x
2
) to the class of
functions we are using. So, let us assume we can nd functions (x
1
) and (x
2
) which
are the integrals of 1/(x
1
) and 1/(x
2
) respectively. Let us assume furthermore that
and have inverses
1
and
1
. Then it follows that

f
1
t
(x

) =
_

1
(t + (x

1
))

1
(t + (x

2
))
_
.
An easy calculation, following steps 4 and 5 of the solution algorithm, yields
h(x
1
, x
2
) =
1
((x
1
) (x
2
) (x

2
)) x

1
(5.4)
and the problem seems to be solved.
However, if we take (x
1
) = 1 and (x
2
) = 1/(1 + x
2
2
) and use psolve we get
> n := 2:
> x := array(1..n):
> f := array(1..n,1..1,[ [1], [1/(1+x[2]**2)] ]):
> xinit := {seq (x[i]=x0[i],i=1..n)}:
> psolve(f,x,xinit);
Warning, (in psolve) MAPLE failed in testing the inverse mapping of F,
...
for component, 2, although this inverse was computed by solve.
3 2 2 3
[ RootOf(_Z + 3 _Z x0[2] + (3 + 3 x0[2] ) _Z + x0[2] + 3 x[1]
3
+ 3 x0[2] - 3 x0[1] - 3 x[2] - x[2] ) ]
The last two lines form the answer, in a kind of implicit form containing the dummy
variable _Z instead of in the desired explicit form.
It is easy to see where things go awry. The solution of
dx
2
dt
=
1
1 + x
2
2
, x
2
(0) = x

2
is given by
1
3
x
3
2
+ x
2
= t +
1
3
(x

2
)
3
+ x

2
.
To compute
1
the cubic equation
1
3
x
3
2
+ x
2
(t +
1
3
(x

2
)
3
+ x

2
) = 0
Symbolic Solutions for a Class of Partial Dierential Equations 467
has to be solved for x
2
. It has three solutions, and only one of those is real, namely
x
2
=
_
_
4 + 9p
2
+ 3p
2
_1
3

_
_
4 + 9p
2
3p
2
_1
3
,
where
p = t +
1
3
(x

2
)
3
+ x

2
.
The cubic equation for
1
leads to the cubic root in the answer generated by Maple.
The program does not choose one particular root of the equation and therefore does
not generate an explicit answer. The obstruction can be removed in this particular case.
One could add a few extra statements to select the real solution, using the allvalues
command to get the three explicit solutions. This would be an ad hoc remedy, enlarging
the class of solvable problems only slightly. Also, although the real root of the answer
given above by Maple is correct, it is more complicated than the answer, h(x
1
, x
2
) =
x
1

1
3
x
3
2
x
2

1
3
(x

2
)
3
x

2
x

1
, given by (5.4). Here too, the simplest answer is not
generated.
6. Conclusions and Recommendations
An algorithm has been devised and implemented as a procedure in Maple to solve
some partial dierential equations.
The main conclusion is that rather elementary and some more involved partial dieren-
tial equations can be solved, but larger and more complicated ones cannot be completely
tackled yet.
Possible approaches to remedy this situation, at least for the users of CAS, are to:
1. improve Maples dsolve and solve procedures;
2. investigate if other algorithms for solving the PDEs, e.g., based on the method
of characteristics, are amenable to implementation and check if they provide more
powerful tools or are less taxing for the CAS;
3. provide a standard Maple procedure for solving as many PDEs as possible.
After a quick look into the capabilities of some other CAS we expect that our expe-
riences and main conclusion do hold for those systems as well. Probably, they show the
same type of shortcomings and may need the same remedies as suggested for Maple. This
is also evident in Postel and Zimmermann (1996).
Acknowledgments
Harm van Essen implemented an initial version of psolve in the Non L
i
n
Con package

(van Essen, 1992; van Essen and de Jager, 1993; de Jager, 1995). Some crucial points in
the implementation have been modied signicantly, however, since then.

Available on request from the rst author.

468 B. de Jager and B. van Asch
References
de Jager, B. (1995). The use of symbolic computation in nonlinear control: Is it viable?, IEEE Trans.
Automat. Control 40(1), 8489.
Geddes, K.O., Czapor, S.R., Labahn, G. (1992). Algorithms for Computer Algebra, Boston, Kluwer Aca-
demic Publishers.
Isidori, A. (1989). Nonlinear Control Systems: An Introduction, 2nd edn, Berlin, Springer-Verlag.
Kevorkian, J. (1993). Partial Dierential Equations: Analytical Solution Techniques, New York, Chapman
& Hall. Second printing.
MacCallum, M.A.H. (1995). Using computer algebra to solve ordinary dierential equations. In Cohen,
A.M., van Gastel, L. and Verduyn Lunel, S. (eds), Computer Algebra in Industry 2: Problem Solving
in Practice, Chichester, John Wiley & Sons, pp. 1941.
Postel, F., Zimmermann, P. (1996). A review of the ODE solvers of Axiom, Derive, Macsyma, Maple,
Mathematica, MuPAD and Reduce. Preprint, presented at the 5th Rhine Workshop on Computer
Algebra.
Sneddon, I. (1957). Elements of Partial Dierential Equations, New York, McGraw-Hill Book Company.
Stephani, H. (1989). Dierential equations: Their solution using symmetries, Cambridge, Cambridge
University Press.
van Essen, H. (1992). Symbols speak louder than numbers: Analysis and design of nonlinear control
systems with the symbolic computation system MAPLE, MSc thesis, Eindhoven University of Tech-
nology, Fac. of Mechanical Engineering. Report WFW 92.061.
van Essen, H., de Jager, B. (1993). Analysis and design of nonlinear control systems with the symbolic
computation system Maple. In Nieuwenhuis, J.W., Praagman, C. and Trentelman, H.L. (eds), Proc.
of the second European Control Conf., Vol. 4, Groningen, The Netherlands, pp. 20812085.