1

MATH3491/5492:
DISCRETE SYSTEMS AND INTEGRABILITY

Frank W NIJHOFF
University of Leeds, School of Mathematics
0

Aknowledgement:
These lecture notes are an adaptation of a text which is intended as a monograph by J.
Hietarinta, N. Joshi and F.W. Nijhoff on the same subject.
Preface

These Lectures deal with discrete systems, i.e. models of (physical, biological, economic,
etc.) phenomena that are mathematically modelled through equations that involve finite
(as opposed to infinitesimal ) operations.
When modelling continuous phenomena in nature we are often using differential equations
as the mathematical tools to encode the essence of what is going on in the system we are
studying. The differential equations are believed to encapture the main mechanisms behind
the phenomena, based on either microscopic theory or general principles relevant to the
subject. Whatever the source of the model, solving the basic differential equations is then
believed to lead to predictions on the behaviour of the system.
In many circumstances, the phenomena of nature are most adequately described not
in terms of continuous but in terms of discontinuous steps (for instance many models in
biology, economy, neural processes, decision processes, etc.). In that case we are dealing
with discrete-time systems, which can mathematically be descrided through:
• recurrence relations;
• difference equations;
• dynamical mappings.
In addition to these types of systems we will also consider:
• functional relations
which are of a slightly different character, but will also arise in connection to discrete systems.
All of the above will form the main mathematical tools in these Lectures.
To make a comparison with the continuous theory, let us recall the types of differential
equations we know. These fall into two classes:
1. Ordinary differential equations (ODEs), which are of the form:

F ( y(x), y ′ (x), y ′′ (x), . . . ; x ) = 0 , (0.0.1)

in which F is some expression (a function of its arguments) and where x is called the
independent variable, whilst y(x) is the dependent variable since it depends on x. The
primes denote differentiation w.r.t. x:
dy d2 y
y ′ (x) = , , y ′′ (x) = ,...
dx dx2

1
2

and the aim is to try and solve y as a function of x, wherever possible, from (0.0.1).
The solution, if we can find it at all, is not unique, and needs further data in order to
render it unique. For instance, this can be done by imposing, in addition to the ODE
itself, a number of initial data, e.g. by fixing values of y and some of its derivatives at
a given value of x, say at x = 0, namely:

y(0) , y ′ (0) , . . .

2. Partial differential equations (PDEs), in which case we have more than one independent
variable. For instance, if we have two independent variables, say x and t, a PDE takes
on the general form:

F ( y, yx , yt , yxx , yxt , ytt , . . . ; x, t ) = 0 , (0.0.2)

where F denotes again some expression, and we have in its argument the function
y(x, t), the dependent variable depending on both independenet variables, and its
partial derivatives

∂y ∂y ∂2y ∂2y ∂2y

yx = , yt = , yxx = , yxt = , ytt = ,...
∂x ∂t ∂x2 ∂x ∂t ∂t2
The theory of PDEs is quite different in character from the theory of ODEs. In fact,
PDEs may possess many solutions and the determination of auxiliary data is much
more complicated. These data may comprise initial values as well as boundary values,
and a finite number of auxiliary data (to fix the solution) is no longer sufficient.
Let us turn now to the discrete situation. Let us recall that the derivatives, on which
differential equations are built, come from the limit:
dy y(x + h) − y(x)
= lim , (0.0.3)
dx h→0 h
and that without taking the limit we actually first encounter the difference operation:
y(x + h) − y(x)
∆h y(x) = . (0.0.4)
h
The derivative (0.0.3) and the differential (0.0.4) are, in fact, quite distinct operations: the
first is a local operation, involving the function at one point x, whilst the difference operation
(0.0.4) is inherently nonlocal and involves two points x and x + h at a distance h. By taking
the limit h → 0, we actually throw away information and thus the difference operator ∆h is
in a sense a more general construct. There is good reason to study continuous systems in
which the derivative plays the main role: the operations are local, instantaneous, involving
smooth functions, thus apparently most suitable to describe the apparent smoothness and
continuity that we observe in Nature.
The study of difference equations seem often to come in hindsight: in those cases where
one cannot calculate solutions of differential equations by analytic means, finite-difference
schemes are used to obtain approximate solutions by numerical methods. Here the discrete
equations seem to come as an artefact and a tool rather than being fundamental to the
 3

phenomena under study. However, in view of eq. (0.0.3) this may be considered as a logical
reversal: since we get the derivative (0.0.3) by the reductive procedure of a limit from the
finite operation (0.0.4) of a difference, shouldn’t we consider the latter as being the more
fudamental?
In modern physics, this question gains even more weight. The underlying structure of
subatomic physics, based on quantum mechanics, leaves no doubt that there is a fundamental
width (given in terms of Planck’s constant ~) beyond which a “continuum” is no longer
perceptible, and even may make no longer sense. Thus, maybe the continuum aspect of
Nature is an illusion, only of validity on the scale of the macroscopic world. Rather than
considering the finite-difference to be the approximation, one actually should consider the
derivative to be the approximation (or simplification) of a more fundamental set of tools. A
famous citation by A. Einstein1 declares:

“To be sure, it has been pointed out that the introduction of a space-time con-
tinuum may be considered as contrary to nature in view of the molecular
structure of everything which happens on a small scale. It is maintained that
perhaps the success of the Heisenberg method points to a purely algebraic de-
scription of nature, that is to the elimination of continuum functions from
physics. Then, however, we must also give up, by principle, the space-time
continuum. It is not unimaginable that human ingenuity will some day find
methods which will make it possible to proceed along such a path. At the
present time, however, such a program looks like an attempt to breathe in
empty space.”

Why then haven’t we then progressed further in the theory of difference equations? It
is fair to say that in contrast to the theory of (ordinary or partial) differential equations,
the theory of difference equations is even today, in the beginning of the 21st century, still
in its infancy. One reason for this discrepancy is that in spite of the simple appearance of
the basic difference operation (0.0.4) the development of comprehensive theory has proved
to be far more difficult. Many more technical difficulties are met in the development of
difference calculus than in differential calculus, mainly due to the nonlocal nature of the
difference operator. Furthermore, it is much more difficult to classify difference equations
than is the case for differential equations: there are just too many possibilities of writing
down difference equations, whereas for differential equations one can easily distinguish them
by their order, degree, well-posedness, or amenability to boundary and initial value problems,
etc.
But there are also historical reasons. Up to the beginning of the 20th century, mathemati-
cians considered the theory of differential equations and the theory of difference equations
very much as two aspects of one and the same endeavor. It is mostly after the two world
wars that a generation of mathematicians working systematically on difference equations
(the famous school of G.D. Birkhoff) disappeared.
It is also important to recall the role of physics as an inspiration in the development of
mathematics, from Newton on. At the beginning of 20th century there were two interesting
developments in physics, which also caught the interest of mathematicians: general relativity
1 A. Einstein, Physics and Reality, 1936. (Reprinted in Essays in Physics, Philosophocal Library, New

York, 1950)
4

and quantum theory. Both relied on “new” mathematics, and attention was drawn on
developing these tools further. It was only in the 1960’s that attention returned on classical
dynamics, on discrete dynamics and the associated chaos on one hand, and on PDE’s and
the associated soliton equations on the other. Major progress was then made on both of
these areas in the 1970’s and 1980’s so that both have now developed into paradigms with
associated mature theory.
What has fundamentally changed in the last two decades is the surge of interest in inte-
grable difference equations (see, e.g, http://www.side-conferences.net). Integrability is
strongly associated with regularity of the behavior, but this regularity is not due to triviality,
but rather due to some underlying mathematical structure that restricts the dynamics in a
nontrivial way. This has already been well established in the theory of soliton equations.
The theory of integrable difference equations (both ordinary and partial) has undergone a
revolution since the early 1980’s, and has also achieved major developments, although many
problems are still open. In fact, the various notions of integrability (of which more in the
following) have proven to be a very strong guiding principle which has driven the subject
forward. The subjects which it has strongly affected include the following
• ordinary difference equations (O∆Es);
• partial difference equations (P∆Es);
• integrable discrete-time systems (mappings);
• discrete and “difference” geometry;
• linear and nonlinear special functions, orthogonal polynomials;
• representation theory of (infinite-dimensional and quantum) algebras & groups;
• combinatorics and graph theory.
Some of these developments go back to results from the 19th century and even earlier,
but it is fair to say that in combining classical results (from Leibniz, Bernoulli, Gauss,
Lagrange, etc.), via the turn of the 19/20th century (Poincaré, Birkhoff, Nörlund, etc.)
with the results of the modern age, the study of integrability in discrete systems forms at
the present time the most promising route towards a general theory of difference equations
and discrete systems.
One of the amazing facts is that in the study of integrable difference equations, we will
encounter all the above type of equations, but moreover these types turn out to be quite
interconnected. In fact, we will see that in many examples we can interpret one and the same
system either as a dynamical map or finite difference equation, or as an analytic difference
equation, or even as functional equation, without damaging the main integrability aspect of
the system.

Types of Difference Equations (∆Es)

In these Lectures we are going to consider various types of discrete systems. From a general
perspective let us briefly mention the various types of difference equations that we might
encounter.
 5

First, as with differential equations, they divide up between ordinary difference equations
(O∆Es) and partial difference equations (P∆Es), depending on whether there is one or more
than one independent variable. The independent variable is discrete, but now this can have
different interpretations:
a) the independent variable can take on only integer values, so we denote it by n ∈ Z, , and
thus the dependent variable y can be denoted by yn ;
b) the independent variable x can take on all values (real or perhaps even complex), but
in the equation the dependent variable is evaluated at integer shifts of the indepen-
dent variable. Thus, the difference equation will involve evaluations of the dependent
variable y(x) at shifts of x by multiples of a fixed quantity h:
y(x) , y(x + h) , y(x + 2h) , . . . x, h ∈ R .

Remark: These two cases are quite different in nature and the associated problems are
quite distinct. However, case a) can be viewed as specialisation of case b), by setting
yn ≡ y(x0 + nh)
and by fixing x0 as a starting value, we can consider the values of the idependent variable
yn as a stroboscope of the second case.

Remark: Before describing various types of difference equations, let us first make the
obvious observation that prior to the the difference operator (0.0.4) there is an even more
basic operation, namely that of the (lattice) shift. This is simply the operation of evaluating
the dependent variable at a shifted value of the independent variable, and it is useful to
introduce an operator associated with this operation:
Th y(x) = y(x + h) , (0.0.5)
which we will use from time to time. It is easy to see that the difference operator and its
higher orders can be simply expressed in terms of the shifts Th :
1
∆h y(x) =(Th − id)y(x) ,
h
1
∆2h y(x) = T 2 − 2Th + id)y(x) ,
h2 h
... ...
n  
n (−1)n X n
∆h y(x) = (−Th )n−j y(x)
hn j=0 j

and thus the nth order difference operator acting on y(x) can be expressed in terms of the
shifted variables y(x), y(x + h), . . . , y(x + nh) . Thus an equation of the form:
F(y(x), ∆h y(x), ∆2h y(x), . . . ; x) = 0
can be rewritten in the form:
F̄(y(x), y(x + h), y(x + 2h), . . . ; x) = 0 ,
6

where the expression F̄ can be straightforwardly be obtained from F by using the above
mentioned substitutions.

n−k n−k+1 n−k+2 n+l−1 n+l

←− h −→ → x

Taking into account this remark, we can distinguish the following types of discrete systems:
a) Ordinary finite difference equations: by which we mean recurrence relations of the form

F(yn−k , yn−k+1 , . . . , yn+l−1 , yn+l ; n) = 0 , n ∈ Z, (0.0.6)

where k and l are fixed integers (see Figure) . We can think of (0.0.6) as an iterative
system where we wish to solve yn at discrete points only. Thus, giving initial values
at a sufficient number of points, generically y0 , y1 , . . . , yk+l , we can hope to iterate
the equation and calculate step-by-step yk+l+1 and subsequent values. The equation
may depend explicitely on n, in which case the equation is nonautonomous, but if
F depends only on n through the dependent variables then the equation (0.0.6) is
autonomous. The order of the equation is generically given by k + l, i.e. the number of
initial data required to achieve a well-defined iteration process. We will sometime refer
to an equation of the form (0.0.6) as an (k + l + 1)-point map (assuming all “points”
at which y is evaluated appear in the expression F).
Finite difference equations of this type can also be viewed as a dynamical mapping
as follows. Assuming that we can solve yn+l uniquely from (0.0.6), leading to an
expression of the form
yn+l = Fn (yn−k , . . . , yn+l−1 ) ,
and introducing the k + l-component vector

y n = (yn−k−l , . . . , yn−1 )t

we can rewrite (0.0.6) as a system of equations through the dynamical map:

yn 7→ e n = y n+1 = F n (y n ) ,
y (0.0.7)

with F denoting the vector-valued function with components Fj (viewed as a function

of the vector y n ).
b) “analytic” (ordinary) difference equations: i.e. equations of the type

F (y(x − kh), y(x − (k − 1)h), · · · , f (x + lh); x) = 0 , (0.0.8)

where now, even though the equation only involves integer shifts (by an increment h) in
the arument of the dependent variable y(x), the independent variable x is meant to be
a continuous variable, and we would like to solve (0.0.10) for functions y(x) in some
appropriate class of functions. Clearly, there is indeterminacy in this problem: the
 7

solutions cannot be determined fully without additional information, and the solution
is determined only up to a shift over h. This implies that the initial data have to
be given over an entire interval on the real line, or alternatively that the role of
“integration constants” is played by periodic functions obeying:

π(x + h) = π(x) .

c) functional equations: This is a different class of equations altogether,taking the form

F (f (x), g(x), f (x + y), g(x + y), f (x + y + z), . . . ; x, y, z, . . . ) = 0 , (0.0.9)

which is supposed to hold for arbitrary values of the arguments of the functions f , g,
etc. Under general assumptions on these functions (such as continuity, differentiability,
etc.), the imposition that the eq. (0.0.9) should hold for arbitrary values for x, y, . . . , is
often sufficient to almost uniquely fix the functions that solve the functional equation.
The notion of initial values is irrelevant in this context.

Exercise # 1: Consider the functional equation:

F(f (x), f (y), f (x + y)) = f (x + y) − f (x)f (y) = 0 .

Under the assumption that f is differentiable for all real values of its argument, and setting
f ′ (0) = α 6= 0, show that the unique solution is given by f (x) = eαx .

Exercise # 2: Consider the functional equation:

F (f (x), f (y), g(x), g(y), f (x + y)) = f (x + y) − f (x)f (y) + g(x)g(y) = 0 .

Under the assumption that f , g are both differentiable for all real values of their argu-
ments, and that f is an even function of its argument, determine the general solution of the
functional equation.

c) Partial difference equations: in this case we have more than one independent variable
x, t ∈ R or in the finite difference case n, m ∈ Z. The equations can take the form:

F(y(x − k1 δ, t − k2 ε), . . . , y(x + l1 δ, t + l2 ε) ; x, t) = 0 (0.0.10)

with k1 , k2 , l1 , l2 fixed integers, in the case of partial analytic difference equations, or

in the finite difference case:

F(yn−k1 ,m−k2 , . . . , yn+l1 ,m+l2 ; n, m) = 0 . (0.0.11)

The dependent variable y(x, t) and yn,m respectively depend here on two (or more)
independent variables with discrete shifts. The second case can be reduced to the first
case by setting
yn,m = y(x0 + nδ, t0 + mε) ,
but obviously the two types of problems are quite different in nature, since in the
analytic case we want to solve for y(x, t) as a function of a continuous range of values of
its arguments. The theory of the latter type of equations is still very under-developed.
8

δ −→ n

m
↓

One of the amazing facts is that in the study of integrable difference equations, we
will encounter all these type of equations, but moreover these types turn out to be quite
interconnected. In fact, we will see that in many examnples we can interpret one and
the same system either as a dynamical map or finite difference equation, or as an analytic
difference equation, or even as functional equation, without damaging the main integrability
aspect of the system. In the latter interpretation we are often dealing with solutions in terms
of (nonlinear discrete) special functions.
The subject is also related to a new branch of geometry, called difference geometry, which
is the discrete analogue of the classical differential geometry of curves and surfaces. It turns
out that integrable P∆Es describe discrete quadrilateral surfaces like the one depicted in
the figure below:

The development of this new theory has been one of the more intriguing applications of the
subject.
 9

Outline of this Module

In this module we will go through some of the modern developments in the theory of inte-
grable difference equations. Basic calculus and algebra are all that are needed as prerequisites
to start with. Later on some knowledge of elliptic functions will be useful (which we will
treat in an elementary way at the advanced level of the module).
We will first start with rehearsing in the first lectures some aspects of the general theory
of difference equations, such as going over some elementary facts from the calculus of finite
differences, and treating some elements of the theory of linear difference equations as well
as specific basic examples of nonlinear difference equations. After that we will present
some systematic methods to obtain discrete equations from parameter-families of continuous
equations. In particular we will show how certain discrete transformations on the solutions
of differential equations for some well-known special functions lead to difference equations
for those families of functions. These, in fact, are already manifestations of structures of
integrable systems.
Next our focus will turn to the theory of partial difference equations, from which we will
develop the notions of integrability further. To motivate the special partial difference equa-
tions under consideration, we will introduce them through the consideration of Bäcklund
transformations for some of the well-known integrable PDEs. Once we have established the
P∆Es, we will study their discrete integrability aspects from a modern perspective. Subse-
quently, we will look at a variety of special solutions, which allow us to get a kaleidoscopic
view on the various techniques used in integrable systems. The soliton solutions will allow
us to introduce a powerful structure in terms of infinite recurrence relations which help
us to find the connections between different integrable equations (this will be done at the
advanced level of the module). Furthermore, these structures can be generalised to obtain
P∆Es in higher dimensions related to the famous KP (Kadomtsev-Petviashvili) hierarchy.
Other solutions are obtained through the consideration of initial-value problems on the lat-
tice, which lead to reductions in terms of integrable dynamical mappings. This brings us
to the theory of (finite-dimensional) integrable discrete-time systems. Furthermore, for the
advanced level we will make a deviation into elliptic functions which we need, not only
to paramatrise solutions of the latter, but also to find yet richer classes of integrable sys-
tems. Finally, we will discuss some important structural aspects of the theory, possibly the
Lagrangian description and/or the structure of symmetries on the lattice and the possible
special solutions we can obtain through the corresponsding reduction techniques. The lat-
ter would lead us in particular to nonlinear special functions and discrete analogues of the
famous Painlevé transcendental equations.
10

General Literature:
1. G.E. Andrews, R. Askey and R. Roy, Special Functions, (Cambridge Univ. Press,
1999).
2. C.M. Bender and S.A. Orszag, Advanced Mathematical Methods for Scientists and
Engineers, (McGraw-Hill, 1978).
3. P.R. Garabedian, Partial Differential Equations, (Chelsea, 1964).
4. F.B. Hildebrand, Introduction to Numerical Analysis, (McGraw-Hill, 1956; Dover,
1987).
5. E.L. Ince, Ordinary Differential Equations, (Dover, 1956)
6. K. Iwasaki, H. Kimura, S. Shimomura, M. Yoshida, From Gauss to Painlevé, (Vieweg,
1991)
7. L.M. Milne-Thompson, The Calculus of Finite Differences, (AMS Chelsea Publ. 2000),
first edition, 1933.
8. N.E. Nörlund, Vorlesungen über Differenzenrechnung, (Chelsea, 1954), first edition ,
1923.
9. E.T. Whittaker and G.N. Watson, A Course of Modern Analysis, (Cambridge Univ.
Press, 1927 (4th ed.))
10. R. Sauer, Differenzengeometrie, (Springer Verlag, 1970).
11. A.S. Bobenko and Yu. Suris,Discrete Differential Geometry; Integrable structure,
Grad. Studies in Math. vol 98, AMS Publs. 2008, (a preliminary version can be
found on: http:arxiv.org/math/abs/0504358).
12. J. Hietarinta, N. Joshi and F.W. Nijhoff, Discrete Systems and Integrability, (mono-
graph in preparation).
Lecture 1

Elementary Theory of
Difference Equations

In this lecture we will give a brief account some of the basic tools to deal with (ordinary)
difference equations and systems based on discrete operations. Most of the machinery is
rather elementary, as we will concentrate here primarily on linear finite-difference equations,
although some simple nonlinear equations will be dealt with as well. Some of this material
is needed in subsequent lectures dealing with integrable systems. Some analogies with and
distinctions from the theory of differential equations will be mentioned. Probably the best
text for further reading is the classic monograph by L.M.Milne-Thompson, or the relevant
chapters 2 and 5 in the famous book by C.M. Bender & S.A. Orszag. However, there
are many elementary textbooks on the basic theory of finite-difference equations, e.g. the
text by W.G. Kelley & A.C. Peterson. For partial finite-difference equations, material can
be found in the monographs by P.R. Garabedian (chapter 3 and 13) and in the book by
F.B. Hildebrand. These latter texts are motivated by the problem of numerical simulations
through finite-difference equations.

1.1 Elementary Difference Calculus

1.1.1 Difference versus differential operators
In the Preface we have already introduced the difference operator ∆h , depending on a step-
size parameter h, which was defined by its action on a function y(x), of a single independent
variable x through the formula

1
∆h y(x) = (Th − id) y(x) , (1.1.1)
h
pointing out that it essentially builds on the shift operator

Th y(x) = y(x + h) . (1.1.2)

11
12 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

It is tempting to expect that one could build a theory based on the difference operator
(1.1.1) in a fairly similar way as for the differential operator d/dx , but there are impor-
tant distinctions. Recalling, from elementary calculus, some of the main properties of the
differential operator, they can be summarised as follows:
linearity: for any two functions y(x), w(x) and constants α, β we have:
d dy dw
(αy(x) + βw(x)) = α +β
dx dx dx

product rule: For the product of any two fuctions y(x), w(x) we have the Leibniz rule:
d dy dw
(y(x)w(x)) = w(x) + y(x)
dx dx dx

composition rule: for two functions z(y) and y(x) and forming the composed function
Z(x) = z ◦ y(x) = z(y(x)) by substitution we have the chain rule

dZ
= z ′ (y(x))y ′ (x)
dx
where z ′ (y) = dz/dy , y ′ (x) = dy/dx .
For the difference operator, however, the last rule really stops making sense (at least without
making many further assumptions), whilst the product rule takes a subtly differen form,
namely by computing
1
∆h (y(x)w(x)) = [y(x + h)w(x + h) − y(x)w(x)]
h
1 1
= y(x + h)[w(x + h) − w(x)] − [y(x + h) − y(x)]w(x)
h h
= y(x + h)∆h (x) + (∆h y(x))w(x) ,

which can be expressed as:

∆h (yw) = (Th y)∆h w + (∆h y)w , (1.1.3)

which looks similar to the differential product rule, but has one crucial extra element: it
involves necessarily the shift operator Th as well!. Obviously we could write (1.1.4) as well
in the form
∆h (yw) = y∆h w + (∆h y)Th w , (1.1.4)
but in either way the shift operator Th cannot be avoided. Thus, the difference calculus
cannot be built with the difference operator ∆h alone, and we need the shift operator Th as
well, which by itself has the crucial product property:

Th (yw) = (Th y)Th w . (1.1.5)

In fact, a ”difference” calculus can be built on the shift operator alone, but we have to realise
that the latter is essentially a nonlocal operator: it connects two values of the argument,
1.1. ELEMENTARY DIFFERENCE CALCULUS 13

namely x and x+ h differeing by a fixed stepsize h. Thus, a theory of difference equations, or

equations defined by h-shifts is essentially more complicated than the continuous differential
case, as can also be argued by considering the Taylor expansion: for regular points x of an
analytic function, we have the Taylor series:
h ′ h2 h3
Th f (x) = f (x + h) = f (x) + f (x) + f ′′ (x) + f ′′′ (x) + · · · (1.1.6)
1! 2! 3!
which shows that the shift operator involves not only the derivative f ′ (x) = df /dx, but also
all higher order derivatives. In fact, the right-hand side of (4.2.1) can be formally expressed
as exp(h d/dx) f (x) , showing that the shift-operator Th , is an infinite-order operator.
Let us now consider, in this lecture, the finite-difference case, i.e. we concentrate on
functions, for the time being only depending on one single discrete variable n which takes
values in the integers n ∈ Z, and which we can denote by yn . In fact,by setting
yn = y(x0 + nh)
fixing an initial point x0 we can reduce the (analytic) difference situation above by the
finite-difference case in which we have a function of a discrete variable n shifting by units,
i.e., for which we can describe the shift by a map: yn 7→ yn+1 . This amounts to viewing the
function y as a map y : N → R : n 7→ yn , which produces an infinite or semi-infinite sequence
of values: {y−N , y−N +1 , . . . , y0 , y1 , . . . , yn , . . . } . The discrete variable n will be considered
to be the independent variable, whilst the fuction value yn is the dependent variable. The
calculus we develop in this lecture will concentrate on this particular situation.

1.1.2 The finite-difference operator

For simplicity, we shall for the time being work with the (forward) finite-difference and
finite-shift operators:
(∆y)n = yn+1 − yn := ∆yn , (T y)n = yn+1 := T yn , (1.1.7)
−1
and we set T yn = yn−1 as being the inverse of the shift operator T . Thus, we have
∆ = T − id . Obviously, we could also introduce other types of difference operators, for
instance:
∇yn = yn − yn−1 = (id − T −1 )yn , yn = yn+1 − yn−1 = (T − T −1 )yn ,
i.e., the backward difference operator and the symmetric difference operator, respectively,
but we will make little use of them. Furthermore, the notation varies in the literature and
sometimes ∇ is used for the average: ∇yn = yn+1 + yn .
Difference equations can now be written in terms of these operators, and it is indeed
often useful to write linear difference equations in terms of T . It is easy to see that
∆yn = yn+1 − yn = (T − id) yn ,
2


∆ yn = yn+2 − 2yn+1 + yn = T 2 − 2T + id yn ,
... ...
Xk   Xk  
k k
∆k yn = j
(−1) yn+k−j = (−1) k
(−T )k−j yn .
j=0
j j=0
j
14 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

Exercise 1.1.1. Derive the last formula. Show also that

∆∇yn = ∇∆yn = yn+1 − 2yn + yn−1 , 2 yn = yn+2 − 2yn + yn−2 .
As pointed out in the previous subsection, whereas the difference operator ∆ can ob-
viously be viewed as a ‘discrete derivative’, analogous to the differential operator d/dx in
differential calculus, there are some crucial differences with the differential case.
We will (almost always) assume that the functions y(x + n) = yn of the discrete variable
form a ring of functions (in the usual sense of abstract algebra), and that pointwise algebraic
operations hold (linear combinations, products, etc.). In the same way as derivatives, the
difference operations act naturally on linear combinations of functions:
∆(αxn + βyn ) = α∆xn + β∆yn , (1.1.8a)
for two functions x, y : Z → R, and arbitrary constants α, β.
However the product (Leibniz) rule has to be modified from the one for derivatives,
namely:
∆(xn yn ) = xn+1 yn+1 − xn yn = (T xn )∆yn + (∆xn )yn = (∆xn )T yn + xn ∆yn , (1.1.8b)
noting as we did earlier that one cannot write the product rule in terms of difference operators
∆ alone, but one needs to involve the shift operator E as well! This is a crucial distinction
with the continuous case of derivatives, and the reason why difference equations are usually
more complicated than differential equations.
Exercise 1.1.2. Compute ∆(yn /zn ).
Another crucial distinction is the observation that there is no natural analogue for dif-
ference operators of the chain rule for derivatives. In fact, this is because there is no natural
composition rule of functions of the discrete variable, unless we specify the target space of
the maps n 7→ yn in a very special way, namely from integers to integers. However, that is
what we usually do not want to do, and most commonly we want the functions yn to take
values in a continuous set, e.g., the reals R or the complex numbers C.
Exercise 1.1.3. Compute the ∆-derivative for the functions log(an), sin(an).

1.1.3 Difference analogues for some familiar functions

The difference operator acts in some ways similar to the derivative (e.g., linearity) but due
to the change in the Leibniz rule many things are also different. Nevertheless we would like
to have analogues to familiar functions, so that ∆ would act on them in a similar way as
d/dx.
In particular we would like to have a function on which ∆ would operate as d/dx operates
for xk , and replace the rule
d k
x = kxk−1 ,
dx
by something analogous. If we think of x being replaced by the discrete variable n, such a
function should obey the rule:
∆f (n; k) = k f (n; k − 1) . (1.1.9)
1.1. ELEMENTARY DIFFERENCE CALCULUS 15

It turns out that the corresponding function can be given in terms of the so-called Pochham-
mer symbol (a)k defined by1
(a)k = a (a − 1) (a − 2) . . . (a − k + 1) , for k > 0, (a)0 ≡ 1 , (1.1.10)
in which k is an integer, but where a could still take on any (real) value. This is also called
the falling factorial2 and it is sometimes denoted by a(k) .
Performing the following computation:
∆(n)k = (n + 1)k − (n)k
= [(n + 1)n(n − 1) · · · (n − k + 2)] − [n(n − 1) · · · (n − k + 2)(n − k + 1)]
= n(n − 1) · · · (n − k + 2)[(n + 1) − (n − k + 1)] = k[n(n − 1) · · · (n − k + 2)]
= k (n)k−1 ,
which shows that f (n; k) = (n)k is a solution of the difference equation (1.1.9) establishing
the analogy with the power function.
Exercise 1.1.4. In the case that k is a negative integer we can introduce the symbol:
(a)k = 1/ ((a + 1)(a + 2) · · · (a + |k|)) k<0, (1.1.11)
(k)
(also denoted as a ). Show that the function f (n; k) = (n)k is a solution of the difference
equation (1.1.9) for the power function also for negative k.
Since ∆ is a linear operator, we can get an analogue to any function defined by a power
series solution, simply by considering linear combinations of power functions of different
degree (the degree being teh variable k). For example, the analogue of the exponential
function exp(ax) , would be
∞
X ak (n)k
e(n; a) = = (1 + a)n . (1.1.12)
k!
k=0

Exercise 1.1.5.
i) Show that the discrete exponential function e(n; a) obeys the difference equation
∆e(n; a) = a e(n; a) ;
ii) show that the second equality in (1.1.12) holds, and that as a consequence we have the
relation
e(n; a)e(m; a) = e(n + m; a) .
1 Note that a does not need to be an integer for this definition to make sense. In fact, one can extend the

definition of (a)k to noninteger values of k by setting

Γ(a + k)
(a)k =
Γ(a)
where Γ denotes the standard Γ function, which we will refrain from defining here. For the time being
it is sufficient to state that Γ(z) is a complex-valued function, generalising the factorial, and obeying the
difference equation: Γ(z + 1) = zΓ(z).
2 In the literature appears also the raising factorial, defined by products of increasing factors instead of

(1.1.10). To distinguish the two one could introduce the products:

(a)±
k = a (a ± 1) (a ± 2) . . . (a ± (k − 1)) ,
but we will not make much use of this notation.
16 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

1.1.4 Discrete integration: the antidifference operator

In the same way as the notion of the integral arises as the inverse of the operation of the
derivative, i.e., the anti-derivative of the function f (x)):
Z x
dF (x)
f (x) = ⇒ F (x) = f (x′ ) dx′ ,
dx
where the indefinite integral denotes a primitive of the function f (x) (fixing a lower limit
of integration amounts to specifying an integration constant). The inverse operation of the
difference operator ∆, i.e., its anti-difference, is given by the summation:
n−1
X
fn = ∆Fn ⇒ Fn = ∆−1 fn := fj .
j

Note that the upper index in the summation is n − 1, since ∆ is defined by the forward
differentiation. In both cases the inverse is not unique, but is determined up to a constant:
an integration constant w.r.t. the variable x in the derivative case, and a constant w.r.t. the
variable n in the difference case. In the indefinite integration/summation the constant is
left unspecified, while in the definite integration/summation this constant can be fixed by
specifying the lower limit in the integral/sum, e.g.,
Z x n−1
X
F (x) = f (x) dx and Fn = fj .
0 j=0

This lower limit is purely a matter of choice and the choice depends on the problem at
hand (e.g., initial values when we deal with differential or difference equations). We can also
define the summation so that n is as the lower index:
M
X
Gn := fj , ∆Gn = −fn .
j=n

Exercise 1.1.6. Give the antiderivative of the function f (x) = xk , and the antidifference
of the function f (n; k) = (n)k for both k a positive and negative integer.
The usual rules of linearity apply to the antidifferential:
X X X
(αxn + βyn ) = α xn + β yn . (1.1.13)

There are also rules corresponding to integration by parts:

n−1
X n−1
X
fj ∆gj = fn gn − fm gm − ∆fj T gj (1.1.14)
j=m j=m

The summation symbol can also be used to define a generating function. Let
∞
X ∞
X zk
F (z) := fk z k , or F̄ (z) := fk ,
k!
k=0 k=0
1.1. ELEMENTARY DIFFERENCE CALCULUS 17

then yn can be recovered as follows:

 n  n
1 d d
fn = F (z) = F̄ (z) .
n! dz z=0 dz z=0

This can be useful if the summation can be done to obtain a closed form expression.

1.1.5 ∗ q-Difference operators and Jackson integrals

A case that resides in some sense inbetween is the theory of q-difference equations. The
q-shift and (forward) q-difference operators are defined in a multiplicative way:

f (qx) − f (x)
Tq f (x) = f (qx) , Dq f (x) = (1.1.15)
(q − 1)x

with the “base” q being a fixed complex number.

A correspondence with the difference case can be made by setting

f (x) = F (ln(x)) , q = eh ⇒ f (qx) = F (ln(x) + h)

but one should be warned that introducing logarithms may affect the analytic nature of the
solutions in a dramatic way.
The power function coincides here with the continuous power function, since

qk − 1
Dq xk = (k)q xk−1 with (k)q = (1.1.16)
q−1

the (k)q being called q-integers. As a consequence, the so-called q-exponential function can
be defined by
X∞
xk
expq (x) = , (k)q ! = (k)q (k − 1)q . . . (1)q (1.1.17)
(k)q !
k=0

A more common notation is given by the Andrew’s symbols:

∞
Y
(a; q)∞
(a; q)k = where (a; q)∞ = (1 − aq j ) , |q| < 1 (1.1.18)
(q k a; q)∞ j=0

which are the q-analogues of the Pochammer symbols. In terms of these symbols the q-
exponential can also be defined as eq (x) = 1/(x; q)∞ , since we have the q-binomial identity:

X xk ∞  
1 x
= = expq . (1.1.19)
(x; q)∞ (q; q)k 1−q
k=0

Exercise 1.1.7. Prove the q-binomial identity P∞(1.1.19) by deriving the difference equation
eq (qx) = (1 − x)eq (x) , and setting eq (x) = k=0 ck xk derive a recurrence relation for the
coefficients ck , whilst c0 = 1.
18 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

The inverse of the q-difference operator Dq is given by the so-called Jackson integral,
defined by
Z x ∞
X
f (x)dq x =: (1 − q)x q k f (q k x) (1.1.20)
0 k=0

Exercise 1.1.8. Show that the analogue of the fundamental theorem of integration holds for
the Jackson integral, i.e., that we have
Z x Z x
Dq f (x)dq x = (Dq f )(x)dq x = f (x)
0 0

1.2 Linear (finite-)difference equations

We note first, that by defining the anti-difference operators, we have already “solved” some
difference equations, namely equations given in the simple form:

∆Fn = fn , Dq F (x) = f (x) with fn , f (x) given.

This would correspond to the simplest linear inhomogeneous case. The next class of differ-
ence equations are more general linear difference equations. We know that linear differential
equation are far easier to treat, in general, than nonlinear differential equations, because the
linearity allows the superposition of solutions (at least in the homogeneous case). Similarly,
linear difference equation have the same property and, hence, are expected to be simpler
than nonlinear equations. However, both for differential as well as difference equations, one
has to distinguish cases:
• linear homogeneous with constant coefficients;
• linear inhomogeneous with constant coefficients
(but possibly with a nonconstant inhomogeneous term);
• linear with nonconstant coefficients.
In the first case the difference equation is autonomous, i.e., the independent variable enters in
the equation only through the dependent variable. For such equations elementary methods
exist to solve them even in explicit form. For the non-constant coefficient case, the situation
is much more complicated, and often solutions can only be found through power series, and
the latter are subject to subtle considerations of singular points of the equation where the
coefficients may blow up. In this lecture we will mostly restrict ourselves to considering the
constant coefficient case, but in the next lecture we will come back to the case of non-constant
coefficients.

1.2.1 Linear constant coefficient difference equations

The order of an ordinary differential equation can be defined to be the number of initial
data needed to specify the integration constants in the solution. Thus, a differential equation
of the form y ′ (x) = F (y, x) is of first order, since after separation of variables or other
techniques to find the solution, one integration is needed leading to one free parameter in the
1.2. LINEAR (FINITE-)DIFFERENCE EQUATIONS 19

general solution. hence, we need one initial value to fully specify the solution. An equation
of the form y ′′ = F (y, y ′ , x) would be second order, because in general we would need
two initial data to specify the solution fully, e.g. y(x0 ), y ′ (x0 ) at a given value x0 of the
independent variable. Similar considerations apply to finite-difference equations as well: the
order of such an equation would by definition equate the number of initial values needed to
specify fully the solution.

Example: A 2-point equation of the form ∆yn+1 = F (yn ) would be generically of first order,
since we could start from a given value y0 and iterate the solution as a map, thereby fixing the
solution entirely. However, it is not true that for all functions F defining the equation we would have
a first order difference equation, because if we take F (y) := c−y, then the difference equation would
reduce to: yn+1 = c , c being a constant, and the solution of this equation is simply yn ≡ c, ∀n.
The latter would effectively be a 1-point equation, and hence an equation of zeroth order (no initial
values needed to specify the solution, taking into account that c is a constant of the equation, and
hence given a priori). Similar arguments apply to a difference equation of the form yn+1 = G(yn ) ,
which is obviusly connected to the previous form by setting G(y) = F (y) + y.

Example: A 3-point equation of the form yn+1 + yn−1 = F (yn ) would be generically of second
order, since we would for generic F need two initial date, say y−1 , y0 to iterate the equation as a
map:
(yn−1 , yn ) 7→ (yn , yn+1 ) = (yn , −yn−1 + F (yn ))
However, it is not alway right to consider the digits in the shifts of a difference equation as an
indication of the order. Thus, the equation ∆2 yn = F (yn+1 , yn ) does not always need to be of
second order, even though we have a second degree difference operator involved. In fact, if we take
F (e
y, y) = f (y) − 2e
y , then the equation reduces to a 2-point form, and the solution would be of first
order, albeit on a sublattice of integers of either even or odd labeled points.

The general case of first order, constant-coefficient, linear finite-difference equations is not
very exciting. Such equations can be written in the form ayn+1 +byn = 0 , (a, b constant) in
the homogeneous case, leading to the solution yn = (−b/a)n y0 , and in the inhomogeneous
case (with given function fn ) we can follow the procedure of the example below.

Example: Solve the inhomogeneous first order difference equation

ayn+1 + byn = fn , a, b constants , fn a given function of n.

The solution of the homogeneous equation (replacing fn by 0) is given above. By a principle which
is called variation of constants, we set yn = ((−b/a)n wn (i.e. replacing the constant y0 by a
function wn of n yet to be determined). Inserting this into the equation yields:
 n    
b b 1  a n
− a − wn+1 + bwn = fn ⇒ wn+1 − wn = − − fn ,
a a b b

and the latter equation can be “integrated” by a summation, yielding:

1 X  a j
n−1
wn = w0 − − fj , w0 (constant) intial value .
b j=0 b
20 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

Thus, returning to the variable yn , the general solution is given by:

 n n−1  n−j
b 1X b
yn = − w0 − − fj ,
a b j=0 a

in which w0 plays the role of an “integration constant”.

Example: We will now consider the general case of second order constant-coefficient, linear
finite-difference equations, the general form of the equation being:

ayn+1 + byn + cyn−1 = fn , a, b, c constants , fn a given function of n. . (1.2.21)

If fn ≡ 0 then the equation is called homogeneous, otherwise it is called inhomogeneous. The

procedure for finding the solution follows closely the standard method for solving constant coefficient
differential equations, comprising the following steps:
1. Find the general solution of the corresponding homogeneous equation:

ayn+1 + byn + cyn−1 = 0 .

This requires solving the corresponding characteristic equation associated with this difference
equation, which is obtained by exploring a simple trial solution of the form yn ∼ λn , where
λ is to be determined. Inserting this into the euqation we get the characteristic equation:

aλn+1 + bλn + cλn−1 = 0 ⇒ aλ2 + bλ + c = 0 .

Then we distinguish the following possibilities:

-distinct roots, λ1 6= λ2 : the general solution in this case is ynhom = c1 λn n
1 + c2 λ2 ;

-coinciding roots, λ1 = λ2 : the general solution is given by ynhom = (c1 + c2 n)λn

1 ,

where c1 , c2 are arbitrary parameters (integration constants) of the solution3 .

2. Find a specific solution of the inhomogeneous equation, e.g. by educated guess, or, more
systematically by the method of variation of constants (see problem on the Examples # 1
sheet: this amounts to allowing c1 , c2 in the general homogeneous solution to depend on n
and by plugging this into the inhomogeneous equation to find the correpsonding c1 (n), c2 (n)).
We note that in general, a particular solution of the inhomogeneous equation (1.2.21) is given
by:
n−1
1 X λn−j − λn−j
ynpart = 2 1
fj , (1.2.22)
a j=0 λ2 − λ1

when λ1 6= λ2 .
3 Note that these solutions are the consequence of the superposition principle: in the case of distinct

roots λn n
1 and λ2 are two independent solutions of the homogeneous equation, and the general solution is an
arbitrary linear combination of these two. In the case of coinciding roots λ1 = λ2 the independent solutions
are λn n
1 and nλ1 and again a linear combination of these two solutions leads to the general solution. The
independent solutions in the latter case can be seen as arising from a limiting procedure of the former case:
λn n
2 −λ1
replacing the independent solutions λn n n
1 , λ2 w.l.o.g. by λ1 and λ2 −λ1 , (note that the latter is itself a linear
combination of the two orginal independent solutions), and taking the limit that the two roots coincide,
i.e. λ2 = λ1 + ǫ, with ǫ → 0, the second solution goes in this limit over into nλn−1 1 , which is the other
independent solution (up to a constant factor) in the coinciding roots case.
1.2. LINEAR (FINITE-)DIFFERENCE EQUATIONS 21

Exercise 1.2.1. Verify that the formula (1.2.22) yields a particular solution of (1.2.21).
Give a similar formula for the case of coinciding roots λ1 = λ2 by setting λ2 = λ1 + ǫ, and
taking the limit ǫ → 0.
3. The general solution of the inhomogeneous equation is now given by the combination yn =
ynhom + ynpart of the general solution of the homogeneous equation (as given in step # 1) and
the particular solution of the inhomogeneous equation (as given in step # 2). The solution
depends on integration constants c1 , c2 , which can subsequently be determined (but only
once we have the full solution of the inhomogeneous problem) by imposing initial conditions,
e.g. by giving y0 and y1 .

Example: The Fibonacci numbers are given by the recursion relation

Fn+1 = Fn + Fn−1 , or (T 2 − T − 1)Fn = 0,

with initial values F0 = F1 = 1. The characteristic equation is given by

1 √
λ2 − λ − 1 = 0 ⇒ λ= (1 ± 5)
2
and hence the general solution
1 √ 1 √ F0 =F1 =1 1 √ 1 √
Fn = a[ (1 + 5)]n + b[ (1 − 5)]n =⇒ a = √ (1 + 5) , b = − √ (1 − 5).
2 2 2 5 2 5
Alternatively,
P∞ we can use a generating function to find the Fibonacci numbers: Introduce F (z) =
n
n=0 Fn z , with z an indeterminate. Plug this into the equation to get:
∞
X
(Fn+2 − Fn+1 − Fn ) z n = z −2 [F (z) − F1 z − F0 ] − z −1 [F (z) − F0 ] − F (z) = 0
n=0

Use the initial conditions F0 = F1 = 1 to solve for F (z), yielding:

F0 (1 − z) + F1 z 1
F (z) = =
1 − z − z2 1 − z − z2
and expanding the r.h.s. w.r.t. powers of z yields the Fibonacci numbers as coefficients.

1.2.2 Linear first order ∆Es with nonconstant coefficients

Let us now consider the general case of non-constant coefficient first-order linear difference
equation, which can be treated to by a technique analogous to the method of integrating
factors in differential equations. Thus, let us consider a difference equation of the type:

yn+1 − gn yn = hn , (1.2.23)

where gn (6= 0), hn are some given functions. As before, the complete solution of (1.2.23)
is a sum of the general solution of the associated homogeneous problem and a particular
solution of the full inhomogeneous equation.
It is easy to see, that the homogeneous version of (1.2.23), namely

wn+1 − gn wn = 0 , (1.2.24)
22 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

can be solved by
n−1
Y
wn = w0 gj , (1.2.25)
j=0

and, thus, by setting yn = wn zn , where a new (unknown) function zn remains to be

determined, we obtain by substitution the difference equation

zn+1 − zn = hn /wn+1 . (1.2.26)

Since the left-hand side is now of the form ∆zn , the equation can be simply integrated with
as solution
n−1
X
zn := hk /wk+1 , (1.2.27)
k=k0

where k0 is an arbitrary lower limit of the sum. Thus the general solution of (1.2.23) is
n−1
! n−1
X hk Y
yn := C+ Qk gj . (1.2.28)
k=k0 j=j0 gj j=j0

Here C is the “summation constant”, associated with the homogeneous solution, its value
be determined by an initial condition. This formula (1.2.28) is really only a formal solution,
and in practice we would like to express the sums and products in closed Q form, but
P there
are no general methods for doing that. Sometimes it is useful to note that gj = e log gj .

Example: In order to solve

yn+1 − nyn = n,
we use (1.2.28) with gn = n, hn = n. For convenience we take j0 = k0 = 1. Then we get
 P 
yn = (n − 1)! C + n−1 1
k=1 (k−1)! .

1.2.3 Higher order linear difference equations and linear (in)dependence

Most of the above considerations can be extended to the higher order case as well. The
general linear equation is of the form

a(N ) (N −1)
n yn+N + an yn+N −1 + · · · a(1) (0)
n yn+1 + an yn = fn (1.2.29)
(j)
where an , fn are some given functions of n, and where we assume that the product aN 0
n an is
(j)
nonvanishing. In the case of constant coefficients (i.e., when all an = a(j) do not depend on
the independent variable n) the homogeneous problem reduces to solving the corresponding
characteristic equation:

a(N ) λN + a(N −1) λN −1 + · · · a(1) λ + a(0) = 0 ,

1.2. LINEAR (FINITE-)DIFFERENCE EQUATIONS 23

leading to roots λ1 , . . . , λN (some of which may coincide). If all roots are mutually distinct,
the solution would take the form

N
X
yn = cj λnj , cj constant coefficients
j=1

but if some roots coincide we will have to supply different independent solution. In fact, if
a root λj has multiplicity µj (i.e. µj is the power of the factor (λ − λj ) in the factorisation
of the polynomial given by the characteristic equation), then the corresponding linearly
independent solutions are λnj , nλnj , . . . , nµj −1 λnj . The general solution of (1.2.29) in the
inhomogeneous case is given by a linear combination of N linearly independent general
solutions of the homogeneous part plus a particular solution of the inhomogeneous equation.
(j)
In the case of non-constant coefficients (i.e., when the coefficients (1.2.29) an are non-
trivial functions of n) the situation is markedly more complicated, and to find basic solutions
one needs to resort to developing power series in the same way as is done in the Fuchsian
theory of differential equations. Such power series solutions often depend crucially on the
occurrence of singularities of the coefficients. We will refrain from treating this theory at
this juncture. The only issue we want to mention at this point is the nontrivial matter of
linear independency of solutions of the difference equation. In the continuous case of linear
differential equations the linear independence of solutions is expressed through the non-
vanishing of a particular determinant associated with the differential equation, the so-called
Wronski determinant or Wronskian (see Example sheet # 1). In the case of linear differ-
ence equations the role of the Wronskian is replaced by the so-called Casorati determinant
or Casoratian. This determinant is constructed as follows. Suppose that we have N discrete
(j)
functions yn , j = 1, . . . , N , then the Casoratian of these functions is given by

(1) (2) (N )
yn yn ... yn
(1) (2) (N )
yn+1 yn+1 ... yn+1
C(yn(1) , yn(2) , . . . , yn(N ) ) = .. .. .. .. , (1.2.30)
. . . .
(1) (2) (N )
yn+N −1 yn+N −1 ... yn+N −1

and these functions are independent if and only if their Casoratian is nonzero.

Exercise 1.2.2. In the homogeneous case of (1.2.29), i.e., setting fn ≡ 0, show that the
n-dependence of the Casoration is given by:

(0)
(1) (2) N) an
C(yn+1 , yn+1 , . . . , yn+1 ) = (−1)N (N )
C(yn(1) , yn(2) , . . . , ynN ) )
an

(0) (N )
which, provided an /an 6= 0 , guarantees that the indpendence of the solutions is preserved
as n progresses.
24 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

The question of solutions of linear nonconstant coefficient dif-

ference equations becomes even more pressing in the theory of
linear analytic difference equations, i.e. the case that the inde-
pendent variable can take on all real or complex values (see the
discussion in the Preface). The latter linear theory was devel-
oped in the beginning of the 20th century by the school of George
D. Birkhoff and collaborators (W.J. Trjitzinsky, R. Carmichael,
J. LeCaine). This group of mathematicians developed system-
atic methods for the Fuchsian theory of linear difference equa-
tions (some of which is represented in the monograph by N.E.
Nörlund, cited in the Preface). Unfortunately, the work in this
direction seemed to have ceased with the onset of the second
world war. It was only in recent years that aspects of the re- Figure 1.1: G.D. Birkhoff
sults of the Birkhoff school were reconsidered in modern mathematics, and various groups
are currently in the process of finalising the work of this school (e.g. the group of Prof.
Jean-Pierre Ramis and his students at Toulouse University).

∗ Variation of constants – matrix formulation

We first note that the N th order linear difference equation (1.2.29) can be cast into a system
of N linear first order equations by introducing new variables
[j]
yn[j] =: yn+j−1 ⇒ yn+1 = yn[j+1] , j = 1, . . . , N − 1

and furthermore
[N ] 1  
yn[N ] =: yn+N −1 ⇒ yn+1 = yn+N = (N )
fn − a(N
n
−1) [N ]
yn − · · · − a(1) [2] (0) [1]
n y n − an y n .
an
 T
[1] [2] [N ]
Introducing the n-component vector y n = yn , yn , . . . , yn we can now write the
system of equations above, in vector form as follows:

y n+1 = An y n + f n , (1.2.31)

with An a matrix of coefficient, and f n a vector given by:

 
  0 1 0 ... ... 0
0  0 0 1 0 ... 0 
 0   
   .. .. .. .. .. 
fn =  ..  , An =   . . . . .  .

 .   
 0 1 
(N )
fn /an a(0) a(1) a(N −2)
a(N −1)
− (N n
) − (N
n
) ... ... − n
(N )
n
(N )
an an an an

The general solution of the matrix difference equation (1.2.31) is very easily obtained, once
we assume that we know a fundamental matrix solution Y n of the homogeneous matrix
equation, i.e., a N × N matrix solution of the equation

Y n+1 = An Y n , such that det(Y n ) 6= 0.

1.2. LINEAR (FINITE-)DIFFERENCE EQUATIONS 25

Note that any matrix Ye n obtained from Y n by matrix multiplication from the right by an
invertible constant matrix C, i.e.,
Y n → Ye n = Y n C
is again a fundamental matrix solution of the same homogeneous matrix difference equation.
Whether or not the sum can be computed in closed form depends on the particular nature
of the coefficients in the equation and whether or not we can find a reasonable expression
for the fundamental matrix solution Y n .
The variation of constants approach now amounts to finding a solution of the inhomo-
geneous vector equation (1.2.31) of the form y n = Y n cn where the N -component vector
cn remains to be determined as a function of n. Plugging this form of y n into the inhomo-
geneous equation we obtain:
Y n+1 cn+1 = An Y n cn+1 = An Y n cn + f n
⇒ cn+1 − cn = Y −1
n+1 f n
n−1
X
⇒ cn = c0 + Y −1
j+1 f j
j=0
n−1
X
⇒ y n = Y n cn = Y n c0 + Y n Y −1 −1
j Aj f j .
j=0

Thus, once Y n is known, the last formula provides the solution of the vector difference
equation, and this is also the general solution, because the first term contains an arbitrary
constant N -component vector c0 which corresponds to the N integration constants of the
equation.
In the case of the N th order difference equation (1.2.29) the vector y n can be written as
y n = (yn , yn+1 , . . . , yn+N −1 )T
(1) (N )
and if we know N independent solutions of that equation yn , . . . , yn the matrix Y n is
given by  
(1) (2) (N )
yn yn ··· yn
 (1) (2) (N ) 
 yn+1 yn+1 ··· yn+1 
Yn=  .. .. .. 

 . . . 
(1) (2) (N )
yn+N −1 yn+N −1 · · · yn+N −1
from which it is clear that the determinant det(Y n ) can be identified with the Casoratian of
the N independent solutions. The explicit case of N = 2 of this construction will be studied
in the Examples # 1 sheet.

1.2.4 Further techniques

1. Factorisation method: It is sometime useful to write the equation in terms of the E
shift operator
(aN
nT
N
+ aN
n
−1 N −1
T + · · · a1n T + a0n )yn = hn , (1.2.32)
and try to factorize the operator.
26 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

Example: The difference equation

2n2
yn+1 + yn−1 = yn
n2−1

can also be written as

2n2
(T 2 − n2 −1
T + 1)yn−1 = 0.
The operator factorizes as
n n−1
(T − n−1
)(T − n
)yn−1 = 0.
First let
n−1
wn−1 := (T − n
)yn−1 ,
and solve
n
(T − n−1
)wn−1 = 0, =⇒ wn = 3Cn.
Next
n−1
(T − n
)yn−1 = 3C(n − 1)
is solved with
yn = C(n2 − 1) + D/n.

2. Reduction from homogeneous to inhomogeneous equation: If a particular solu-

tion of the inhomogeneous equation can be found that it can be used to reduce the problem
into an homogeneous one.
Another method of reducing the problem to a homogeneous one is to operate on the
equation with an operator that annihilates the inhomogeneous part. Then one solves the
resulting (higher order) homogeneous equation. The solution is then substituted into the
inhomogeneous equation and the extra coefficients determined.

3. Reduction of order: With a known solution of the homogeneous equation the order
of the equation can be reduced.

Example: Consider the second order equation

a2 yn+2 + a1 yn+1 + a0 yn = b, ∀n,

[1]
where ai , b may depend on n. Suppose yn is a solutions of the homogeneous part. Then substituting
[1]
yn = yn zn yields

[1] [1] [1]

a2 yn+2 (zn+2 − zn+1 ) + (a2 yn+2 + a1 yn+1 )(zn+1 − zn ) = b

which is first order in wn := zn+1 − zn .

1.3. A SIMPLE NONLINEAR DIFFERENCE EQUATION – DISCRETE RICCATI EQUATION27

4. z-transforms: In treatment of the Fibonacci difference equation we have introduced

the generating function of the Fibonacci numbers. This is a particular example of a z-
transform for the difference equation (reminiscent of the Laplace transform for differential
equations): by introducing suitable series expansions in terms of an auxiliary variable z,
such as
∞
X
Y (z) = yj z j ,
j=0

the linear difference equation can be often converted into an algebraic equation for Y (z) (in
the case of constant coefficients), or into an ODE (for coefficients depending polynomialy
on the independent discrete variable n). Examples are given on the Example sheet # 1.

1.3 A simple nonlinear difference equation – discrete

Riccati equation
Sometimes a nonlinear difference equation can be linearized, i.e. rendered into a linear
equation by a suitable change of variables or by applying other tricks. A case where this
happens is the important example of the discrete Riccati equation, i.e., a difference analogue
of the famous Riccati differential equation, named after the Count Jacopo Francesco Riccati
(1676-1754). Since we will often encounter both the discrete as well as the continuous Riccati
equation in the context of integrable systems, we will present here a few properties of both
equations.

1.3.1 Continuous Riccati equation

The usual Riccati differential equation is a first order nonlinear differential equation of the
form:
dy
= a(x)y 2 + b(x)y + c(x) , (1.3.33)
dx
i.e., it possesses a quadratic nonlinearity. The coefficients a(x), b(x) and c(x) can be arbitrary
functions of x. In the case they are constants, the equation (1.3.33) can be solved by
separation of variables.

Exercise 1.3.1. In the case that the coefficients a,b,c in (1.3.33) are constants, and denoting
by y1 , y2 the roots of the quadratic ay 2 + by + c , give an expression for the general solution
in terms of y1 , y2 . What happens if the roots of this quadratic coincide?

In the general case, when the coefficients are not necessarily constant, we note the fol-
lowing properties:

1. the Riccati equation is form invariant under linear fractional transformations (Möbius
transformations):

αy(x) + β δe
y (x) − β
y(x) 7→ ye(x) = ⇔ y(x) = , (1.3.34)
γy(x) + δ −γey(x) + α
28 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

where α, β, γ, δ are functions of x such that αδ − βγ 6= 0. In fact, it is easy to check

that implementing the transformation (1.3.34) turns (1.3.33) into an equation for ye of
the form
de
y
=e y 2 + eb(x)e
a(x)e y+e
c(x) .
dx
a(x), eb(x), e
with new coefficients e c(x).

2. The Riccati equation can be linearized by setting: y(x) = f (x)/g(x), substitution of

which leads to
gf ′ − f g ′ = af 2 + bf g + cg 2 .
and the latter relation can be split into two linear equations, leading to the system:
 ′
f = λf + cg ,
g ′ = −af + (λ − b)g ,

where λ is arbitrary. Choosing λ in appropriate way and eliminating g, this system

leads to a second-order linear ODE for f .

3. Any four solutions yi (x) , (i = 1, 2, 3, 4) of the same Riccati equation are related
through the relation

(y1 − y2 )(y3 − y4 )
CR[y1 , y2 , y3 , y4 ] =: = constant , (1.3.35)
(y1 − y3 )(y2 − y4 )

where CR stands for cross-ratio. In fact, for any two solutions yi , yj of the Riccati
equation, it is easy to derive:

d
ln(yi − yj ) = a(yi + yj ) + b ,
dx
and, hence, the combination

d
ln((y1 − y2 )(y3 − y4 ))
dx
is invariant under permutations of the indices.

4. Given two solution y1 (x), y2 (x) of the same Riccati equation, a new solution can be
found in the “interpolation form”:

y1 + ρy2 dρ
y(x) = provided ρ(x) solves = a(y1 − y2 )ρ . (1.3.36)
1+ρ dx

Hence, given any two solutions, we can find an interpolating solution of the Riccati
equation through a solution of the linear equation for ρ.

Exercise 1.3.2. Prove the statement in part 4 above, deriving the linear differential equation
for ρ in (1.3.36), given two solutions y1 , y2 of the Riccati equation.
1.3. A SIMPLE NONLINEAR DIFFERENCE EQUATION – DISCRETE RICCATI EQUATION29

1.3.2 Discrete Riccati equation

The Riccati equation (1.3.33) has a natural analogue, the discrete Riccati equation given by

yn yn+1 + an yn+1 + bn yn + cn = 0 , (1.3.37)

where the coefficients an , bn , cn are given functions of n. In fact, it is more natural to

introduce the form:
  
An Bn yn
R(yn , yn+1 ) = Cn yn yn+1 + Dn yn+1 − An yn − Bn = (−1, yn+1) .
Cn Dn 1
(1.3.38)
and denote by R(yn , yn+1 ) = 0 the discrete Riccati equation. It follows immediately from
this form that the discrete Riccati equation can be linearised as:
    
yn+1 An Bn yn
=λ
1 Cn Dn 1

with λ some proportionality constant. In fact, λ = (Cn yn + Dn )−1 Furthermore, solving

for yn+1 , we obtain the fractional linear form:
An yn + Bn
yn+1 = ,
Cn yn + Dn
which reminds us of the Möbius transformation. In other words, the Möbius transform,
interpreted as a map yn 7→ yn+1 is equivalent to the iteration of solutions of the Riccati
equation.
We recover the differential Riccati equation (1.3.33) from the discrete one in the form
(1.3.38) performing the limit ε → 0, setting
dy Cn An − Dn Bn
yn+1 ∼ y(x) + ε , ∼ −εa(x) , ∼ εb(x) , ∼ εc(x)
dx Dn Dn Dn
Exercise 1.3.3. Show that from the limit described above one indeed recovers the continuous
Riccati equation in the form (1.3.33).
Let us now consider again the discrete Riccati equation in the form (1.3.37). We will
now establish analogous properties of this equation to ones of the continuous equation.
1. The equation is form-invariant under fractional linear (Möbius) transformation. In
fact, under the transformation
αyn + β
yn 7→ yen = , αδ − βγ 6= 0 ,
γyn + δ
the equation (1.3.37) goes over into an equation fr yen of the same form with coefficients
an , ebn , e
e cn .
(1) (2)
2. Given two independent solutions yn , yn the combination
(1) (2)
y n + ρn y n
yn =
1 + ρn
30 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

is again a solution of the discrete Riccati equation provided that ρn solves the linear
difference equation
(1) (1) (2) (2)
(yn(2) yn+1 + an yn+1 + bn yn(2) + cn )ρn + (yn(1) yn+1 + an yn+1 + bn yn(1) + cn )ρn+1 = 0 .
(1) (2)
By using again the discrete Riccati equation for yn , yn , we can deduce from this
that this linear equation can be written in either of the two equivalent forms:
(2) (1)
y n + an yn+1 + bn
ρn+1 = (1)
ρn = (2)
ρn .
y n + an yn+1 + bn
(i) (j)
3. For any two different solutions yn and yn we have the following relation
(i) (j)
yn+1 − yn+1 cn − an b n
(i) (j)
= (i) (j)
,
yn − yn (an + yn )(an + yn )
and hence we conclude that
(1) (2) (3) (4))
(yn+1 − yn+1 )(yn+1 − yn+1 ) (cn − an bn )2
(1) (2) (3) (4))
= (1) (2) (3) (4)
,
(yn − yn )(yn − yn ) (an + yn )(an + yn )(an + yn )(an + yn )
(1) (2) (3) (4)
for any four solutions yn , yn , yn , yn . Consequently, we infer immediately that for
(1) (2) (3) (4)
any four solutions yn , yn , yn , yn , of the discrete Riccati equation we have that
the cross-ratio
(1) (2) (3) (4)
(yn − yn )(yn − yn )
CR[yn(1) , yn(2) , yn(3) , yn(4) ] = (1) (3) (2) (4)
= constant .
(yn − yn )(yn − yn )
4. By setting yn = fn /gn in (1.3.37) we can show that the equation can be linearised
leading to the linear matricial equation:
    
fn+1 bn cn fn
=κ , (1.3.39)
gn+1 −1 −an gn
in which κ can be chosen arbitrarily. It is easy to derive from this matricial equation
a second order linear difference equation for fn .
Exercise 1.3.4. Derive the matrix equation (1.3.39) from the discrete Riccati equation
(1.3.37), and by eliminating gn derive a second order difference equation for fn .

Example: An alternative, more direct way to linearize the Riccati equation (1.3.37) is by rewrit-
ing the equation as
(yn + an )(yn+1 + bn ) + cn − an bn = 0 , (1.3.40)
which suggests a substitution of the form
wn
yn + a n = ⇒ wn+1 + (bn − an+1 )wn + (cn − an bn )wn−1 = 0 ,
wn−1
which brings us directly to a linear second order difference equation for wn .

In some isolated cases a judicious grouping of terms can help to find a linearizing trans-
formation.
1.4. PARTIAL DIFFERENCE EQUATIONS 31

Example: Consider the difference equation

yn+1 yn−1 − yn2 = yn yn−1 .

Dividing by yn yn−1 yields

yn+1 yn
− = 1,
yn yn−1
yn
which is linear for wn = yn−1
.

Exercise 1.3.5. Linearize the logistic equation (Verhulst equation) dy/dx = ay(1 − y) by
setting y = 1/(w + b). Do the same for a discrete analogue of the logistic equation, namely
yn+1 = ayn (1 − yn+1 ) . Show also that the linearization trick breaks down for another
discretization of the logistic equation, namely yn+1 = ayn (1 − yn ) .

1.4 Partial Difference Equations

Partial difference equations ( P∆E) have arisen in the context of numerical methods for
PDEs, where they correspond to finite-difference schemes on lattices, cf. the monographs
by P.R. Garabedian and F.B. Hildebrand. In this section we will mention briefly some
aspects, but we will not go into these methods in any depth because the type of P∆Es we
are interested in this course arise from quite different considerations and require their own
specific methodology.

1.4.1 Discretization and boundary values

An the first example let us consider the heat equation

∂y ∂2y
= .
∂t ∂x2
If we discretize the time derivative using the forward difference ∆ and the second order
space derivative with ∆∇ we obtain
1 1
ε (y(x, t + ε) − y(x, t)) = δ 2 (y(x + δ, t) − 2y(x, t) + y(x − δ, t)),

or with discrete indices identifying yn,m = y(x0 + nδ, t0 + mε), relative to some arbitrary
origin (x0 , t0 ), we can then write this equation more simply as:

yn,m+1 = α(yn+1,m + yn−1,m ) + βyn,m , (1.4.41)

where α = ε/δ 2 , β = 1 − 2ε/δ 2 . The points that are involved in this equation are indicated
in Figure 1.2
An important problem in solving P∆E’s is the boundary or initial value question. This
often requires a careful consideration of graph of the lattice points involved in the equation.
In the present case we can see from Figure 1.2 that given data on the line t = 0, i.e., m = 0
we can propagate the solution forward in time. If instead we want to propagate into the
positive x-direction, we need initial data on two vertical lines, say n = n0 , n0 + 1. We could
32 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

t m t m

x n x n

a) b)
Figure 1.2: Points involved a) in equation (1.4.41) and b) in (1.4.42).

also use two lines on an angle, such as m − n = s0 , s0 + 1. These initial values correspond
to the equation being second order in the x derivatives.
For practical calculations it is often useful to introduce two shift operators,
T yn,m = yn+1,m , Syn,m = yn,m+1 .
For example the discrete heat-equation (1.4.41) can then be written as
Syn,m = (α + β + αT −1 )yn,m .
If we have now the initial value given as yn,0 = fn then we find the solution in the form
yn,m = (αT + β + αT −1 )m fn .
Next consider the Laplace equation
∂2y ∂ 2y
+ = 0.
∂t2 ∂x2
This could be discretize as the second derivative in the heat equation, but in both x and t.
If we keep the same central point in both discretizations we obtain
yn,m+1 + yn,m−1 + σ(yn+1,m + yn−1,m ) + ρyn,m = 0, (1.4.42)
where σ = ε2 /δ 2 , ρ = −2(1 + ε2 /δ 2 ). The points that are involved in this equation are
indicated in Figure 1.2. From the Figure it is now clear that we need can proceed if we are
given the initial values on two adjacent vertical lines or horizontal lines.

1.4.2 Separation of variables

For linear constant coefficient PDE’s the ansatz y(x, t) = f (ax + bt) often leads to a solution
if the parameters a, b are chosen suitably. If the terms in the equation have the same total
order we get a characteristic equation for the parameters, while function f can be arbitrary,
in other cases we can reduce the equation to an ODE.
In discrete case the above works much less frequently. It is useful if the points involved
fall on a line, we can then change to new variables and reduce the problem to an O∆E.
1.4. PARTIAL DIFFERENCE EQUATIONS 33

Example: Consider the equation

yn,m = ryn+1,m−2 + syn+2,m−4 .

If we now defined new variable as follows:

k = n, l = 2n + m

the equation becomes

yk,l = ryk+1,l + syk+2,l
which can be considered as an O∆E in k. Let now fi (k) be the two solutions of this equation (as
discussed in Section 1.2.1), then the solution of the original equation is

yn,m = h1 (2n + m)f1 (n) + h2 (2n + m)f2 (n),

where the hi are arbitrary functions.

Other classes of mostly nonlinear P∆Es will be considered in subsequent lectures mo-
tivated by the context of integrable systems. Their stencil on the lattice can take different
shapes, e.g. of a four-point hook form:

H(yn,m , yn+1,m , yn+1,m−1 , yn+2,m−1 ) = 0 (1.4.43)

a four-point quadrilateral form:

Q(yn,m , yn+1,m , yn,m+1 , yn+1,m+1 ) = 0 (1.4.44)

or a five-point star form:

S(yn,m , yn+1,m , yn−1,m , yn,m+1 , yn,m−1 ) = 0 (1.4.45)

or even more exotic shapes.

The points that are involved in the above two-dimensional maps are given in Figure 1.3.

ε
n
δ

a) b) c)
Figure 1.3: A difference equation may involve a different configuration of points in the
lattice. The cases a), b) and c) correspond to (1.4.43), (1.4.44) and (1.4.45), respectively
34 LECTURE 1. ELEMENTARY THEORY OF DIFFERENCE EQUATIONS

ց
?
ր

Figure 1.4: Acceptable and problematic initial values for (1.4.44)

From Figure 1.4 we can also infer that the definition of an initial value problem can be
delicate. For example in case b), which is discussed in detail later, we can propagate from
any staircase-like initial values, but any “overhang” could lead to a contradiction in the
initial-value problem (due to overdeterminedness).
The problem of establishing continuum limits is also more complicated for P∆E’s than
for O∆E’s, and in fact not unique: there are more than one way of obtaining contiuous
equations from discrete ones by limits. We return to this problem later, at this point it
suffices to note that sometimes we may have to take the continuum limit along a diagonal
of the lattice in order to obtain nontrivial equations.

Literature:
1. L.M. Milne-Thompson, The Calculus of Finite Differences, (AMS Chelsea Publ. 2000),
first edition, 1933.
2. C.M. Bender and S.A. Orszag: “Advanced Mathematical Methods for Scientists and
Engineers”, (McGraw-Hill, 1978)
3. W.G. Kelley and A.C. Peterson: “Difference Equations: An Introduction with Appli-
cations”, Second Ed. (Academic Press, 2001).
4. P.R. Garabedian, Partial Differential Equations, (Chelsea, 1964).
5. F.B. Hildebrand, Finite-Difference equations and Simulations, (Prentice-Hall, 1968).
6. G. Andrews, R. Askey and R. Roy, Special Functions, (Cambridge Univ. Press, 1999).
Lecture 2

From Continuous to Discrete

Equations via Transformations

In this part we will discuss how difference equations arise from transformations applied to
differential equations. Thus we find that some functions can be defined both by a differential
equation and by a difference equation. Of course the independent variables of these two types
of equations are not the same, but rather we find interesting duality between parameters
and independent variables.
Often functions defined by a differential equation posses special transformations. One
particularly important class of transformations are those that act on the differential equation
by changing the values of its parameters. When such transformations are iterated, we obtain
a sequence of differential equations (and often a sequence of solutions) that are strung
together by the changing sequence of parameter values.
As particularly simple example consider the differential equation

x w′ = α w. (2.0.1)

Here x is the independent (continuous) variable, the prime stands for x-derivative (w′ = dwdx )
and α is a parameter of the equation. It is easy to find the solution of this equation, it is

wα (x) = k xα , (2.0.2)

where k is required integration constant. From the form of the solution (2.0.2) we can see,
that the simple transformation of multiplying the function wα by x changes α, that is

wα+1 = x wα . (2.0.3)

Equation (2.0.3) can now be interpreted as a discrete or difference equation, where α is

the (discrete) independent variable and x is a parameter. Thus we have found, that the
equations (2.0.1) and (2.0.3) both describe the same function wα (x) of (2.0.2) but from
two different points of view. The key observation here is that of the compatibility between
the differential equation (2.0.1) and the difference equation (2.0.3) which can be explicitly

35
36LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

observed from the following litte computation:

′
wα+1 = wα + xwα′ ′
⇒ xwα+1 = xwα + x2 wα′
⇒ (α + 1)wα+1 = xwα + x(αwα ) ⇔ wα+1 = xwα ,

which could also be stated as saying that the operation of taking the derivative d/dx and
that of the shift in the parameter α, Tα , commute on solutions, i.e., (Tα w)′ = Tα (w′ ) . This
is a theme we will see coming back many times.
We will first discuss the case of such procedures for ordinary differential equations
(ODEs), using the examples of equations for certain classical special functions, such as
the Bessel, Legendre, and hypergeometric functions. These can be shown not only linear
ODEs, but also some ordinary difference equations (O∆Es) in terms of certain parameters.
There are also nonlinear special functions, such as the elliptic functions and the Painlevé
transcendents, that satisfy nonlinear differential equations, for which similar ideas apply,
but we will not treat them in this Lecture. Next, we will move from ODEs to partial differ-
ential equations (PDEs), and treat briefly some basic theory of soliton equations, and show
how similar ideas there lead to the construction of integrable partial difference equations
(P∆Es).

2.1 Special Functions and linear equations

Special functions [see, e.g., Whittaker-Watson, 1927] such as the Bessel, Parabolic Cylin-
der, and hypergeometric functions satisfy ODEs with an independent variable x and at
least one parameter α. They also satisfy recurrence relations in which the parameter α is
shifted. We will now study some of these differential equations and derive the corresponding
transformations and related difference equations.

2.1.1 Weber Functions and Hermite Polynomials

Consider the differential equation


w′′ + α + 1
2 − 41 x2 w = 0, (2.1.4)

where the primes denote differentiation in x. This equation arises in many physical applica-
tions, for example in the study of quantum mechanics in harmonic potential. A particular
special solution of this equation is the parabolic cylinder or Weber function w =: Dα (x),
uniquely specified by the asymptotic behaviour
  
α −x2 /4 α(α − 1) 1
Dα (x) = x e 1− 2
+O , x → +∞. (2.1.5)
2x x4
The simplest special case arises when α is an integer n, this yields the Hermite polynomials
Hn (x), defined by

√
Dn (x) = 2−n/2 exp −x2 /4 Hn (z), z = x/ 2. (2.1.6)
Instead of thinking of Equation (2.1.4) as one equation specified by one fixed value of
α, it is more productive to think of it as an infinite sequence of equations, each of which is
2.1. SPECIAL FUNCTIONS AND LINEAR EQUATIONS 37

specified by successive values of α = n + α0 . This alternative perspective has a wonderful

consequence: we can generate new solutions of each successive equation by knowing solutions
of an earlier equation in the sequence. To see how to do this, note that the differential
operator in Equation (2.1.4) factorizes:



∂x − x/2 ∂x + x/2 w = −α w. (2.1.7)

Assume that w = wα (x) is a general solution corresponding to a given value of α. Let us

define

e := ∂x + x/2 w.
w (2.1.8)
Then we have from (2.1.7) the system of two equations


e
∂x + x/2
w = w,
(2.1.9)
∂x − x/2 we = −α w.

e we immediately get (2.1.7), while eliminating w yields

If we use the first to eliminate w



∂x + x/2 ∂x − x/2 w e = −α w,e (2.1.10)

or after expanding

e′′ + (α − 1) +
w 1
2 − 41 x2 we = 0. (2.1.11)
which is the same as (2.1.4) except for a new parameter value α̃ = α − 1, and thus w eα̃(x) ∝
wα−1 (x).
Since the equations are linear the proportionality constant is in principle free, and is
determined by some other considerations, like normalization. In this particular case the
solutions of (2.1.4) (the parabolic cylinder functions Dα (x)) are conventionally normalized
to have the asymptotic behaviour as given in (2.1.5). Using (2.1.8) on (2.1.5) we get

2


∂x + x/2 Dα (x) = α xα−1 e−x /4
1 + O 1/x2 = αDα−1 ,

and therefore we should use the normalization

wα = Dα (x) , eα̃(x) = αDα−1 (x).

w

Equation (2.1.9) can be written in two additional ways, using this normalization. First by
solving for the term that appears alone:

Dα−1 (x) = α1 [Dα′ (x) + 21 xDα (x)],
(2.1.12)
Dα (x) = Dα−1 ′
(x) − 12 xDα−1 (x).

and secondly by solving for the derivative terms:


Dα′ (x) = − 12 xDα (x) + αDα−1 (x),
(2.1.13)
Dα−1 (x) = 12 xDα−1 (x) − Dα (x).
′

The first pair allows us to travel up and down in the chain of Dα (x)’s for different α’s
separated by integers. The second form (2.1.13) allows us to derive a fully discrete equation:
38LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

substituting α → α + 1 into the second equation of (2.1.13) and then subtracting the two
equations we get
Dα+1 (x) − x Dα (x) + α Dα−1 (x) = 0. (2.1.14)
This is a difference equation, where α is the independent variable and x the parameter.
Let us now reflect on what we have obtained. The conventional theory of ordinary
differential equations regards the parameters α as given and fixed, while x varies in some
domain. By using the Darboux transformation (2.1.8) we can change the parameter α. In
the fully dual point of view, we would keep x is fixed while α changes. At an intermediate
step also have recurrence relations that relate w(x; α), w(x; α̃) and their x-derivatives. In
the theory of integrable equations, such recurrence relations are called (auto-) Bäcklund
transformations.

Exercise 2.1.1. Note that we could

have
factorized the operator in Weber’s equation in a
different order
to get ∂x + x/2 ∂x − x/2 w = −(α + 1) w. Find the equation satisfied by
b = (∂x − x/2 w and deduce another Darboux transformation. Also show that w
w b ∝ wα+1 .

Exercise 2.1.2. Use the relationship between Weber function and Hermite polynomials
(2.1.6) to derive the difference equation for Hermite polynomials Hn+1 (x) = 2xHn (x) −
2nHn−1 (x), starting from (2.1.14).

Quantum mechanical interpretation: Equation (2.1.4) looks like a Schrödinger equa-

tion for a particle moving in a harmonic potential, so the discussion above has a quantum
mechanical interpretation: ∂x + x/2 is a lowering operator. It kills the lowest energy state:
2
(∂x + x/2)ψ = 0, from which we find ψ0 ∝ e−x /4 . ∂x − x/2 is the raising operator, which
generates excited states starting from the ground state ψ0 .

2.1.2 Darboux and Bäcklund transformation in general

Jean Gaston Darboux (1842 - 1917) was one of the pioneers of
classical differential geometry. Furthermore, he made important
contributions to many other fields of mathematics including dif-
ferential equations. In his book [Darboux, 1914] (p.210, Section
408) a theorem, which has been revived in the modern theory
of integrable systems, leading to what in the modern literature,
cf. e.g. [Matveev & Sall, 1991;Rogers & Schief, 2002], is called
a Darboux transformation. Related to it is the notion of a so-
called Bäcklund transformation, which we describe briefly below
and in the next section in more detail (for the case of PDEs).
Whereas Darboux’ theorem applies primarily to linear equations,
the Bäcklund transformation, as we shall see, is most relevant to
nonlinear equations. Figure 2.1: J.G. Darboux
Darboux’ theorem can be described as follows:
Theorem 2.1.1. (Darboux transformation) Let us consider the differential equation

y ′′ = [φ(x) + h]y , (2.1.15)

2.1. SPECIAL FUNCTIONS AND LINEAR EQUATIONS 39

with parameter h. Suppose f (x) is a particular solution of this equation for some specific
value h1 of h. Let us define a new function ye by

ye := [∂x − (log f )′ ]y. (2.1.16)

Then ye solves the equation

e
ye′′ = [φ(x) + h]e
y, where φe := φ − 2(log f )′′ . (2.1.17)

We say that φe is the Darboux transform of the potential and ye of the wave-function.
What is behind this theorem is a factorisation property. In fact, given that the “poten-
tial” function φ(x) does not depend on the (spectral) parameter h, the solution y(x, h1 ) =
f (x) of (2.1.15) allows one to express this potential as

φ(x) = −h1 + ∂x2 log f + (∂x log f )2 ,

and hence eq. (2.1.15) for general h can be rewritten as:

y ′′ = (h − h1 + v ′ + v 2 )y , where v = ∂x log f .

This can be factorised as follows:

(∂x + v)(∂x − v)y = (h − h1 )y ,

and setting ye = (∂x − v)y , we obtain by interchanging the factors:

(∂x − v)(∂x + v)e

y = (h − h1 )e
y,

e
the equation (??) with a new potential φ(x) = φ(x) − 2v ′ given by (2.1.16).
We note that the result on the Weber function of the previous section can be viewed
as an application of Darboux’ theorem: In comparison with (2.1.4) we see that we should
choose φ = x2 /4, f = ex /4 , h1 = 1/2. Then we find that φe = φ − 1 in agreement with
2

(2.1.11).
The pair of equations (2.1.12) are a special case of a con-
struction named after Albert Victor Bäcklund (1845 - 1922) who
worked on transformations of surfaces in differential geometry.
In the the modern era the connection between the latter sub-
ject and the theory of differential equations has become more
prominent. In fact, there is close relation between transforma-
tions between special surfaces in terms of coordinates on these
surfaces and transformations between solutions of (linear and
nonlinear) differential equations. As we shall see in the next
section these transforms will form the basis of the construction
of exact discretizations of those very same differential equations.
We defer the precise definition of a Bäcklund transformation to
later when we deal with PDEs, but only present here a loose Figure 2.2: A.V. Bäcklund
definition of this concept as follows:
40LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

Definition 2.1.1 (Bäcklund transformation (loose definition)). Suppose we have a pair

of equations depending on two dependent variables u and v and possibly on their partial
derivatives: 
F (u, ux , . . . , v, vx , . . . ) = 0,
(2.1.18)
G(u, ux , . . . , v, vx , . . . ) = 0.
If upon eliminating v we obtain the equation R(u, ux, uxx . . . ) = 0 and upon eliminating u
we obtain S(v, vx , vxx . . . ) = 0 then (2.1.18) is called a Bäcklund transformation (BT)
between the equations R = 0 and S = 0. If R and S differ only through some parameters
the transformation is called an auto-Bäcklund transformation (aBT).
We can consider the Darboux transformation y 7→ ye described in the theoreom 2.1.1 as
a special case of a Bäcklund transformation. However, we shall mostly employ the term in
connection with transformations between solutions of nonlinear equations, in particular of
nonlinear PDEs and partial difference equations (PδEs).

2.1.3 Another example: Bessel Functions

As an application of the general DT method let us consider the Bessel functions, which also
occur in many applications, e.g., is the study of waves in circular domains.
Bessel functions are defined as solutions of

x2 w′′ + xw′ + (x2 − ν 2 )w = 0, (2.1.19)

where ν is a parameter, with specific asymptotic behaviours: The standard Bessel function
of the first kind is defined by1
 x ν X
∞
k
−x2 /4
Jν (x) = . (2.1.20)
2 k! Γ(ν + k + 1)
k=0

We will again consider Bessel’s equation as an infinite sequence of equations in the space
of the parameter value ν. To allow us to iterate in this parameter space, we need recurrence
relations for Bessel functions. To find Darboux transformations from which recurrence
relations follow, it is easier to start by transforming equation (2.1.19) into a socalled Sturm-
Liouville form, i.e. without the first derivative term.
Exercise 2.1.3. Transforming the dependent variable w by setting w(x) = p(x)y(x), and
substituting this into eq. (2.1.19), show that we can derive a second order ODE for
√ y without
a term containing y ′ , by choosing 2x p′ (x) + p(x) = 0, which implies p(x) = 1/ x.
√
Transforming variables by taking w(x; ν) = y(x; ν)/ x, we get
 
ν 2 − 1/4
y ′′ + 1 − y=0 (2.1.21)
x2

Eq. (2.1.21) corresponds to Darboux’ equation (2.1.15) by setting φ(x) = (ν 2 −1/4)/x2 and
h = −1. Following Darboux’ theorem 2.1.1 we should first construct a suitablefunction f ,
1 Recall the definition of the Γ-function, which is used in this formula, as the function obeying the difference

equation Γ(z + 1) = zΓ(z) .

2.1. SPECIAL FUNCTIONS AND LINEAR EQUATIONS 41

which we can take in the form f := xσ which corresponds to the value h1 = 0 in (2.1.15),
and where we need to take σ = (1/2) ± ν. Taking, for convenience, the lower sign we obtain
the following transformation from (2.1.17)

ỹ := (∂x + (ν − 1/2)/x) y, (2.1.22)

ν 2 − 1/4 (ν − 1)2 − 1/4
φ̃ := 2
− 2(log f )′′ = . (2.1.23)
x x2

Thus, we see here again an example here the Darboux transformation implies an integer
step in the parameter, namely the parameter ν, and therefore we can identify

ỹ = yν−1 (x), (2.1.24)

up to an arbitrary constant factor, which (as follows from an explicit computation) we can
set equal to 1, to be consistent with the series representation (2.1.20) of the solution of the
Bessel equation.
Using (2.1.22)) we find easily the factorization

(∂x − (ν − 12 ) x1 )(∂x + (ν − 21 ) x1 ) = (∂x2 − (ν 2 − 41 ) x12 ) (2.1.25)

and the Bäcklund pair

  1

∂x + (ν − 21 )/x y = ỹ,
(2.1.26)
∂x − (ν − 2 )/x ỹ = − y,

where eliminating ỹ yields (2.1.21) for y, while eliminating y yields

 
′′ (ν − 1)2 − 1/4
ỹ + 1 − ỹ = 0 ,
x2

i.e., eq. (2.1.21) with ν → ν − 1.

√
Returning to the original variables with y = x w the above equations become
  
 ∂x + ν/x w = w̃,
(2.1.27)
∂x − (ν − 1)/x w̃ = −w.

Now solving for the derivative terms with (2.1.24) and shifting ν → ν + 1 in the second
equation we obtain

xwν′ = −ν wν + x wν−1 ,
(2.1.28)
xwν′ = ν wν − x wν+1 ,

and after subtracting we get the difference equation in ν with x a parameter:

x wν+1 − 2ν wν + x wν−1 = 0. (2.1.29)

The equations (2.1.28) and (2.1.29) hold for the Bessel functions of the first kind Jν (x).
42LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

2.2 Bäcklund transformations for non-linear PDEs

Many non-linear integrable PDEs possess Bäcklund transformations and their consistency
condition leads to integrable nonlinear partial difference equations. In this section we will
mainly discuss the Korteweg–de Vries (KdV) equation, which is given by the PDE:

ut = uxxx + 6uux . (2.2.30)

This so-called nonlinear evolution equation was derived in

1895 by the two people whose names it bears, in a study on
shallow water waves and where an exact “solitary wave” solution
was presented. The equation was a key milestone in a big con-
traversy on the nature of waves, not least following the famous
“real life” observation of a solitary wave by John Scott Russell
in 1834. It is not the place here to recite the whole history of
the soliton, which can be found in several of the existing mono-
graphs on solitons and integrable systems, see e.g. [Ablowitz &
Segur, 1982; Calogero & Degasperis, 1982; Newell, 1985; Drazin
& Johnson, 1989; Ablowitz & Clarkson, 1991]. We just mention
that the KdV equation was revived 70 years after Korteweg and
de Vries’ paper, when in a celebrated study by C. Gardner, J. Figure 2.3: D. Korteweg
Greene, M. Kruskal and R. Miura, it was shown that thenonlinear PDE (2.2.30) can be
exactly solved by an ingenious method, which is nowadays referred to as the inverse scat-
tering transform method. Although this method is only applicable to very special equations,
equations that we refer to as soliton equations or exactly integrable equations, we now know
entire infinite families of such equations to which the method can be applied to find exact
solutions of the nonlinear equations (this being in stark contrast with the generic situation
that nonlinear PDEs in the general case cannot be exactly solved and that typically we have
to resort to either qualitative studies or perturbative and numerical methods to study their
solutions).

2.2.1 Lax pair for KdV

One of the key properties of this equation is that there exists an underlying overdetermined
system of linear equations:

ψxx + uψ = λψ , (2.2.31a)
ψt = 4ψxxx + 6uψx + 3ux ψ, (2.2.31b)

whose consistency condition leads to (2.2.30). The first equation (2.2.31a) has the form of
a linear spectral problem for the differential operator

L = ∂x2 + u , (2.2.32)

where the coefficient u = u(x, t) plays the role of a potential. The parameter λ is an
eigenvalue of the operator L and can, in principle, depend on t if u depends on it, but
by the definition of an eigenvalue of a differential operator it should independent of x. The
2.2. BÄCKLUND TRANSFORMATIONS FOR NON-LINEAR PDES 43

second equation (2.2.31b) describes the (linear) time-evolution of the function ψ(x, t), where
again the same u enters in the coefficients. The system (2.2.31) is overdetermined: the two
linear equations can only be compatible with each other if additional conditions hold for the
coefficients, which are all expressed in terms of u. Furthermore, we shall assume that the
time-evolution descibes an isospectral deformation 2 of the linear spectral problem, i.e., we
assume that λ is, in fact, independent of t.
Theorem 2.2.1. Under the assumption of isospectrality, i.e., λt = 0 , the linear system
(2.2.31) is self-consistent, i.e., (ψxx )t = (ψt )xx , iff either ψ ≡ 0 or the potential u = u(x, t)
obeys the Korteweg-de Vries (KdV) equation (2.2.30).
Proof. First, using (2.2.31a) rewrite (2.2.31b) as follows:
ψt = (4λ + 2u)ψx − ux ψ .
and now consider the cross-derivatives:
(ψxx )t = ((λ − u)ψ)t = (λt − ut )ψ + (λ − u) [(4λ + 2u)ψx − ux ψ]
(ψt )xx = (4λ + 2u)ψxxx + 4ux ψxx + 2uxxψx − ux ψxx − 2uxx ψx − uxxx ψ
= (4λ + 2u) ((λ − u)ψ)x + 3ux (λ − u)ψ − uxxxψ
⇒ (ψt )xx − (ψxx )t = (ut − uxxx − 6uux) ψ − λt ψ .
Hence, under the condition of isospectrality, λt = 0 , we see that the system is compatible,
i.e. (ψt )xx = (ψxx )t provided u(x, t) obeys the KdV equation.
The linear system of equations associated with a non-linear PDE is symptomatic of its
integrability through the inverse scattering method. Such a linear system, consisting of a
spectral problem and an equation for the time evolution, is called a Lax pair, after P.D.
Lax who gave a systematic framework for describing such linear problems in his celebrated
paper [Lax,1968]. Although for virtually all soliton equations Lax pairs have been found,
there is no fully algorithmic method known to produce a Lax pair for a given equation.

2.2.2 Miura transformation

The KdV equation possesses a remarkable transformation, called the Miura transformation
after its inventor, which gives rise to many insights about its solutions, and in particular
can be used to derive a Bäcklund transformation for KdV. To find it, consider how we
can eliminate the function u from the system (2.2.31) and obtain a PDE in terms of the
“eigenfunction” ψ itself.
From (2.2.31a) we find
u = λ − ψxx /ψ,
and inserting this into (2.2.31b) we obtain the following equation for ψ
ψx ψxx
ψt = ψxxx − 3 + 6λψx . (2.2.33)
ψ
2 By ”deformation” we understand the variation of the operator L in (2.2.32) through the change of the

potential u(x, t) as t varies, which in principle may affect the spectrum of the operator, i.e., the collection of
eigenvalues. However, if the deformation is ”isospectral” the spectrum is preserved while varying t, meaning
that the eigenvalues λ do not change as t varies.
44LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

Introducing the variable:

v := ∂x log ψ , (2.2.34)
we easily obtain from (2.2.31a) the Miura transformation

u = λ − vx − v 2 . (2.2.35)

Furthermore, after taking derivatives w.r.t. x on both sides of (4.2.29) and expressing all
terms using v we get a PDE governing v:

vt = vxxx − 6v 2 vx + 6λvx . (2.2.36)

The latter equation (for λ = 0) is known as the modified KdV equation (MKdV), and
it differs from the KdV equation (2.2.30) notably in the nonlinear term. The differential
substitution (2.2.35) allows one to find a solution of the KdV equation given a solution of
the MKdV equation: if v solves the MKdV (2.2.36) and u is defined by (2.2.35), then u
solves KdV (2.2.30).

2.2.3 The Bäcklund transformation

Let us now turn to the derivation of Bäcklund transformations using the above. We start
with the Miura transformation (2.2.35), and combine it with the simple observation that
the equation (2.2.36) is invariant under the replacement v 7→ −v . The idea is to use one
sign in transforming from u to v and another in transforming from v to ue, that is, we will
have the Miura transformations

e = λ + vx − v 2 ,
u (2.2.37a)
u = λ − vx − v 2 . (2.2.37b)

It is surprising that the trivial transformation v 7→ −v implies a highly nontrivial trans-

formation u 7→ u e on the solutions of the KdV equation.
Adding and subtracting the two relations above we obtain

e + u = 2(λ − v 2 )
u (2.2.38a)
e − u = 2vx .
u (2.2.38b)

The latter can be integrated if we introduce the variable w by taking u = wx . For the KdV
equation this change of variables leads to

wt = wxxx + 3wx2 , (2.2.39)

after one integration in x. (Note that we have omitted an irrelevant integration constant.)
This equation for w is called the potential KdV equation (PKdV).
In terms of this new dependent variable the equation (2.2.38b) can be integrated to
e − w = 2v , and inserting it into the first relation (2.2.38a) we obtain
w

1 2
e + w)x = 2λ −
(w e − w) .
(w (2.2.40a)
2
2.2. BÄCKLUND TRANSFORMATIONS FOR NON-LINEAR PDES 45

written entirely in terms of w.

Equation (2.2.41) provides us with the x-dependent part of the Bäcklund transformation
e we need also a t-dependent equation, which can be
. To fully characterize the solution w
readily found by using the PKdV equation itself. Adding (2.2.39) for w and w e and using
(2.2.41) to reduce wxxx + wexxx we obtain the relation
e + w)t = (w
(w exx ) + 2(wx2 + wx w
e − w) (wxx − w ex2 ) ,
ex + w (2.2.40b)
and the relations (2.2.41), (2.2.42) together constitute the Bäcklund transformation for the
KdV equation. (Note that in practice, one could use the PKdV equation itself rather than
(2.2.42) to implement the BT.)
Supplementing the latter relation by a relation for the t derivatives, using the PKdV
itself, we obtain the following statement:
Theorem 2.2.2. The system of relations
1
e + w)x
(w = 2λ − e − w)2 ,
(w (2.2.41)
2
e + w)t
(w = (w exx ) + 2(wx2 + wx w
e − w) (wxx − w ex2 ) ,
ex + w (2.2.42)
e t)
defines a transformation from a given solution w(x, t) of the PKdV to a new solution w(x,
of the PKdV equation.
Proof. Differentiating (2.2.41) by t and (2.2.42) by x we get respectively:
(w
e + w)xt = −(w
e − w)(w
et − wt )
(w
e + w)tx = (w
ex − wx )(wxx − w
exx ) + (w
e − w)(wxxx − w
exxx )
+2 (2wx wxx + 2w
ex w
exx + w
ex wxx + wx w
exx )
= (w
e − w)(wxxx − w
exxx ) + 3(w
ex + wx )(w
exx + wxx )
and subtracting the first from the second we get, using also the x-derivative of (2.2.41),
0 = (w
e − w) [(w
et − w
exxx ) − (wt − wxxx )] + 3(wex + wx ) [−(w
e − w)(w
ex − wx )]
 
= (w
e − w) (wet − w
exxx − 3w ex2 ) − (wt − wxxx − 3wx2 ) ,
and hence, if w solves the PKdV equation then either w
e = w or w
e solves the PKdV as well.

2.2.4 Using BTs to generate multisoliton solutions

Note that the Bäcklund pair (2.2.40) is different from the one we had before in that it contains
a parameter that does not appear at all in its base equation (2.2.39). This parameter can
be used to generate more complicated solutions from simpler ones.
Suppose we know a given “seed solution” solution w of the PKdV, then inserting this
into (2.2.41) we obtain a first order nonlinear ODE for w. This ODE will always be of the
form:
1 2
ex = − w
w e + a(x)w e + b(x),
2
where the right-hand side is a quadratic in w e (with x-dependent coefficients). This is a
well-known type of differential equation called a Riccati equation. These equations are
generally solvable through a linearisation procedure. After solving this equation we have
some integration constants that may depend on t, they can be determined from (2.2.42).
46LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

Example: As a specific example, consider the simplest case where the seed solution w of the
PKdV equation (2.2.39) is the trivial solution w ≡ 0. Setting w ≡ 0 in (2.2.41) yields

1 2
w
ex = 2λ − w
e ,
2
which can be integrated by separation of variables and yields

w(x,
e t) = 2k tanh (kx + c(t)) , λ = k2 .

Substituting this expression into the (2.2.42) (with w = 0) reveals that we must take ct = 4k3 t + c0 ,
were c0 is a constant. Thus we have obtained the solution:


e t) = 2k tanh kx + 4k3 t + c0 .
w(x, (2.2.43)

This implies that for the solution of the KdV equation we obtain the solution


u ex = 2k2 sech2 kx + 4k3 t + c0 ,
e(x, t) = w (2.2.44)

which is the famous formula for the 1-soliton solution of the KdV equation.

2.2.5 Permutability property of BTs

The solution we obtained above can now be regarded as the starting point for applying
the BT once again to obtain yet another solution of the PKdV equation. Carrying this out
further we can iteratively obtain an infinite sequence of increasingly complicated solutions of
the same nonlinear PDE. The procedure of solving a Riccati equation at each stage obviously
becomes increasingly more cumbersome as we go along. However, there is a powerful new
ingredient that can be used to simplify the iteration, namely the permutability property of
the BTs.
Suppose we want to compose two different BTs, one with a parameter λ, as in (2.2.40),
and one with another parameter, say µ, given by

λ 1 2
e
BTλ : w 7→ w e + w)x = 2λ −
(w e − w) ,
(w (2.2.45a)
2
µ 1 2
b
BTµ : w 7→ w b + w)x = 2µ − (w
(w b − w) , (2.2.45b)
2
where we have used the notation w, e w
b to denote the solution obtained by applying the BT
with parameter λ, µ, respectively.
There are now two ways to compose these BTs: either start with BTλ and and subse-
quently apply BTµ , or the other way around. In this way we get iterated solutions which
we can denote by wbe and w eb respectively,

be = BTµ ◦ BTλ w,
w eb = BTλ ◦ BTµ w.
w

The highly nontrivial result is that, under certain conditions, both ways of composing BTs
lead to the same result: w be = w,
eb and hence the two BTs commute. This is the famous
permutability property of the BTs.
2.2. BÄCKLUND TRANSFORMATIONS FOR NON-LINEAR PDES 47

Theorem 2.2.3. The BTs given by (2.2.45a),(2.2.45b) for different parameters λ and µ
generate solutions (with a suitable choice of integration constants) for which we have the
following commutation diagram of BTs:

BTλ e
w BTµ

w be = w
w eb

BTµ BTλ
b
w

The proof of the permutability property is quite deep and relies on the spectral properties
that play at the background of the equations.
Proof. A direct proof is computational. Using the additional relations:

µ b
 
1 b 2
e 7→ w
BTµ : w e be + w
w e e−w
w e ,
= 2µ − (2.2.45c)
x 2
  1 e 2
λ e e
b 7→ w
BTλ : w b b+w
w b = 2λ − b−w
w b . (2.2.45d)
x 2
be as follows:
e in terms of w and w
Iterating the first Bäcklund chain we can solve w

1 b be − w)x + 2(λ − µ)
(w
e=
w e + w) +
(w ,
2 be − w
w
and reinserting this into the BT we obtain:
 2
λ+µ be + w)x + ∂ 2 log(w
= (w be − w) + 1 ∂x log(w
be − w)
x
2
1 b 2 (λ − µ)2
+ (w e − w) + 2 ,
8 be − w)2
(w

and the latter relation is symmetric under the interchange of λ and µ. A similar symmetry
can be derived for the t-part of the BT. Hence, starting from an arbitrary seed w we can
find solutions, by appropriate choice of integration constants, which are symmetric under
interchange of λ and µ.
The consequences of the permutability property are far reaching, and we will give an
explicit realization as follows. In fact, using all four Bäcklund relations (2.2.45a)-(2.2.45d),
and setting w be = w,
eb we can now eliminate all the derivatives from the four eqs. (2.2.45) we
obtain a purely algebraic equation of the form:

be − w)(w
(w b − w)
e = 4(µ − λ). (2.2.46)

be without having to
This allows us to obtain directly the iterated BT transformed variable w
derive the solution through the Riccati equations of the BT.
48LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

Exercise 2.2.1. Construct a 2-soliton solution by starting from the seed solution w ≡ 0,
and two 1-soliton solutions of the type (2.2.43) (with different parameters k and l, where
be from (2.2.46). Note that the phases c0 may
λ = k 2 , µ = l2 , respectively) and solving for w
be taken to be different for w e and w.
b Verify explicitly that the wbe so constructed actually
solves (2.2.39).

2.2.6 Bäcklund transformation for the sine-Gordon equation

BTs exist for other integrable evolution equations as well. In fact the first BT, the one
proposed by Bäcklund himself, is associated with the sine-Gordon equation,

θxt = sin θ , (2.2.47)

λ
For this equation the Bäcklund transformation θ → θe is given by the following relations:
!
  e+ θ
θ
θe − θ = 2λ sin , (2.2.48a)
x 2
!
  2 θe − θ
e
θ+θ = sin , (2.2.48b)
t λ 2

e t). By calculating the t derivative (2.2.48a)

connecting a variable θ(x, t) to a new variable θ(x,
and the x-derivative (2.2.48b) and then taking a sum or difference, one can easily derive
(2.2.47) for θe or θ, respectively. Thus (2.2.48) is a one-parameter auto-Bäcklund transfor-
mation for the sine-Gordon equation.
µ
As before, we can now introduce a second BT θ → θb of the form (2.2.48) with parameter
µ, namely
!
  b+ θ
θ
θb − θ = 2µ sin , (2.2.49a)
x 2
!
  2 b− θ
θ
θb + θ = sin . (2.2.49b)
t µ 2

We can also apply BTµ on θ, e and BTλ on θ,

b and if among these 8 equations we eliminate all
b e
derivatives (under the assumption of the permutability of the BTs, i.e., θe = θb ), we obtain
the following permutability property:
   
be e b be b e
θ + θ − θ − θ λ θ + θ − θ − θ
sin  = sin  . (2.2.50)
4 µ 4

i
If, for simplification, we denote e 2 θ = w, we can write (2.2.50) as

bew
λ(w b − ww) bew
e = µ(w e − ww).
b (2.2.51)
2.2. BÄCKLUND TRANSFORMATIONS FOR NON-LINEAR PDES 49

e
ee
w
ee
w
b
ee
e
w w

w be
w

b
w b
be
w
bb
w

b
bb
w

Figure 2.4: A lattice of BTs

Exercise 2.2.2. Starting from the trivial solution of the sine-Gordon equation θ ≡ 0, use
e t) of the same equation containing the parameter
the BT (2.2.48) to obtain a solution θ(x,
λ, namely:   
e t) = 4 tan−1 exp λx + t + ϕ0
θ(x, , (2.2.52)
λ
where ϕ0 is a (constant) phase. (Hint: Use the integral
Z 
dϕ ϕ
= ln tan + c.
sin ϕ 2

2.2.7 Transition to lattice equations

Is obvious from the above, that by iterating the BTs with two different parameters we obtain
from one seed solution w an entire lattice of solutions, see Figure 2.4. Note that building
this lattice of solutions crucially depends on the validity of the permutability property!
We have derived the permutability equations (2.2.46) and (2.2.51) from the properties
of the PKdV and SG, respectively, thus these equations are descriptive, they describe yet
another property of the sequence of functions derived using BTs. We can introduce an
enumeration of the solutions as follows:

wn,m = BTλn ◦ BTµm w, (2.2.53)

after which we can write (2.2.46) and (2.2.51) as difference equations of the form:

(wn+1,m+1 − wn,m )(wn,m+1 − wn+1,m ) = 4(µ − λ), (2.2.54)

and

λ(wn+1,m+1 wn,m+1 − wn+1,m wn,m ) = µ(wn+1,m+1 wn+1,m − wn,m+1 wn,m ), (2.2.55)

50LECTURE 2. FROM CONTINUOUS TO DISCRETE EQUATIONS VIA TRANSFORMATIONS

respectively, the shifts along the lattice wn,m 7→ wn+1,m and wn,m 7→ wn,m+1 corresponding
to the application of the Bäcklund transformations BTλ and BTµ .
We will now change our point of view: at each elementary plaquette of the above lattice
of solutions we have a relation of the form (2.2.54) or (2.2.55) (or something else for other
equations), and we will now elevate these equations as being the main equations of interest.
In a sense having reached this point we can “forget” about the original PDE’s from whence
the construction originated, and place the permutability equations at the centre of our focus.
Later we will indeed consider lattice equations, i.e., P∆E’s like (2.2.54) or (2.2.55) on
their own merit and study their remarkable properties. In particular we will show that these
equations are integrable in some precise sense.

Literature:
1. E.T. Whittaker and G.N. Watson, A course of Modern Analysis, fourth edition, 1927,
(Cambridge University Press, 2002).
2. J.G. Darboux Lecons sur la theorie generale des surfaces et les applications geometriques
du calcul infinitesimal (Gauthier-Villars, Paris, 1914)
3. V.B. Matveev and M.A. Salle, Darboux transformations and solitons (Springer, 1991)
4. C. Rogers and W.K. Schief, Bäcklund and Darboux Transformations (Cambridge Uni-
versity Press, 2002).
5. D.J. Korteweg and G. de Vries, On the change of form of long waves advancing in a
rectangular channel, and on a new type of long stationary waves, Philos. Mag. (5) 39
(1895) 422–443.
6. C.S. Gardner, J.M. Greene, M.D. Kruskal and R.M. Miura, Method for solving the
Korteweg-de Vries equation, Phys. Rev. Lett. 19 (1967) 1095–1097.
7. P.D. Lax, Integrals of nonlinear equations of evolution and solitary waves, Commun.
Math. Phys. 21 (1968) 467–490.
8. R.M. Miura, The Korteweg-de Vries Equation: A Survey of Results, SIAM Review 18
# 3 (1976) 412–459.
9. G.L. Lamb, Elements of Soliton Theory, (Wiley Interscience, 1980).
10. M.J. Ablowitz and H. Segur, Solitons and the Inverse Scattering Transform, (SIAM,
1982).
11. F. Calogero and A. Degasperis, Spectral Transform and Solitons, vol. 1, (North-
Holland Publ., Amsterdam, 1982).
12. A.C. Newell, Solitons in Mathematics and Physics, (SIAM, Philadelphia, 1985).
13. P.G. Drazin and R.S. Johnson, Solitons: An Introduction, (Cambridge University
Press, 1989).
14. M.J. Ablowitz and P.A. Clarkson, Solitons, Nonlinear Evolution Equations and Inverse
Scattering, (Cambridge Univ. Press, 1991).
Lecture 3

Integrability of P∆Es

Motivated by the lattice structure emerging from the permutability/superposition properties

of the Bäcklund transformations of the previous Lecture, we will now consider the integrabil-
ity properties of these viewed as partial difference equations (P∆Es) on the two-dimensional
space-time lattice. What we will discover is that the integrability can be given a precise,
even algorithmic, meaning. The presence of parameters (namely the Bäcklund parameters
λ and µ, which we will now reinterpret as lattice parameters) will play a crucial role in
the development of the theory. Furthermore, as we will see in subsequent Lectures, the
parameters render the P∆Es very rich: since they can be seen to represent the widths of the
underlying lattice grid they allow us to recover, through continuum limits, a great wealth
of other equations, semi-continuous (i.e. differential-difference type) as well as fully contin-
uous (i.e. partial differential type) ones. The interplay between the discrete and continuous
structures will prove to be one of the emerging features of the integrable systems that we
study.
The history of integrable difference equations goes back to seminal papers by Ablowitz
& Ladik and by Hirota in the 1970’s, cf. [1,2]. The first was motivated by the search
for integrable numerical algorithms through finite-difference approximations. This fits well
into the general problem of the analysis of finite-difference P∆Es arising from numerical
studies of PDEs, cf. e.g. the monograph by P.R. Garabedian, Ch. 13. More recently,
systematic methods for the construction of integrable nonlinear finite-difference P∆Es were
found, e.g. through the representation theory of infinite-dimensional Lie algebras, [3], or
through singular linear integral equations and connections with Bäcklund transformations,
cf. [4,5].

3.1 Quadrilateral P∆Es

We will investigate here partial difference equations (P∆E’s) of the following canonical form
(which we will call quadrilateral P∆Es)

e, u
Q(u, u b
b, u
e) = 0 , (3.1.1)

51
52 LECTURE 3. INTEGRABILITY OF P∆ES

where we adopt the canonical notation of vertices surrounding an elementary plaquette on

a rectangular lattice:
u := un,m , e = un+1,m
u
b := un,m+1
u , b
e = un+1,m+1
u
Schematically, this configuration of points is given by:

u e
u
- -

? ?

? - -?
b
e
u
b
u

The notation is inspired by the one for Bäcklund transformations, which as we have
seen in the previous Lecture, give rise to purely algebraic equation as a consequence of their
permutability property. However, here we will forget, at first instance, about this connection,
and consider lattice equations of the form (3.1.1) in their own right as a partial difference
equation (P∆E) on a two-dimensional lattice. Thus, n and m, play the role of independent
(discrete) variables, very much like x and t being the continuous variables for an equation
like the KdV equation
Even though the form (3.1.1) seems very restrictive, it will turn out that it is in a sense the
most elementary form as a model type of equations, and on the other hand remarkably rich.
In fact, as we shall see later, P∆Es of the form (3.1.1) are “rich enough” to be candidates
for discretisations of PDEs of arbitrary order in the spatial and temporal variables.
A classification of P∆Es in the same way as of PDE s, does not yet exist. Nonetheless, we
may consider a P∆E of the form (3.1.1) to have features reminiscent of hyperbolic PDEs.
In fact, if the equation Q = 0 can be solved uniquely for each elementary quadrilateral,
then one may pose initial value problems (IVPs) in ways very similar to hyperbolic type of
equations such as the KdV equation itself (i.e. as a discrete nonlinear evolution equation).
This can be done as follows.
The naive approach would be to consider IVP where we would assign values of the
dependenet variable u along horizontal array of vertices in the lattice, i.e. values for un,0 ,
for all n, considering the variable m to be the temporal discrete variable. It is easy to see,
however, that such an IVP would lead to a nonlocal problem if we want to use (3.1.1) as an
iteration scheme to find all values un,m for m > 0. In fact, to calculate any value un,1 say
for given n, we would need to involve all initial values un′ ,0 with n′ < n, and furthermore
have to assume limiting behaviour as n → −∞. This would be a complicated procedure.
However, there is nothing that tells us that we should identify the n- and m axes as the
spatial and temporal axes respectively. The lattice picture allows us to play other and more
3.1. QUADRILATERAL P∆ES 53

natural games, and it is the equation itself that gives the lead in provding is with the natural
IVP that we could impose. Thus, changing the perspective slightly, we may tilt the lattice
and rather consider initial value data to be imposed on configurations like a “sawtooth” (or
a “ladder”), such as:

  
R R R

The analogy with hyperbolic PDEs can be further seen from the consideration of the
“memory” each point in the lattice has of the initial values that are involved in its determi-
nation upon iteration. In fact, any given point has a backward shadow (the analogue of the
so-called “lightcone”) of points the values of u on which determine the value at that point,
as indicated by the picture:

the point at the bottom of the cone being fully determined by the initial values at the top
of the cone and only and exclusively by those values! This is very much reminiscent again
of what happens in the case of hyperbolic PDEs.

Multilinearity In order to have a unique iteration scheme arising from an equation of

the form (3.1.1) the equation should be linear in each of the variables around the quadrilat-
eral, i.e. the equation should multilinear. The most general form of a quadrilateral lattice
equation which is linear in each dependent variable around the quadrilateral, and which in
addition respects reversal symmetry with respect to the shifts e- and b on the lattice, takes
the following form:
   
k0 ub
uu b
eu
e + k1 ub e + ub
uu b
uue + ue
uub
e+ubu
eub
e + k2 u bu
e + uub
e
     
+k3 ue u+u b
bu
e + k4 ub u+u eub
e + k5 u + u e+u b+u b
e + k6 = 0 , (3.1.2)

where k1 , . . . , k6 are coefficients (which may depend on additional parameters).

Exercise 3.1.1. Show that, starting from a multilinear quadrilateral lattice equation (3.1.1)
with 16 general coefficients, we arrive at the form (3.1.2) by assuming that the equation
remains unchanged when we reverse the e- or bshifts, i.e. when we replace u e by u (by which
e
we mean the backward shift related to the e-shift, see picture), or when we replace b by u
u
(meaning the backward shift related to the b-shift, see picture). b
54 LECTURE 3. INTEGRABILITY OF P∆ES

u
b

u u e
u
e u

b
u

It is easy to see that the lattice equation (2.2.54) derived in Lecture 2 from the Bäcklund
transform of KdV is precisely of the form given above. However, not all equations of the
form (3.1.2) will be of interest to us. We will be interested in those equations (for specific
choices of the coefficients k1 , . . . , k6 ) which have the additional property of being integrable.
What we mean by this will be explored in the next section.

3.2 Integrability and Consistency-around-the-Cube

We will now consider a class of quadrilateral P∆Es (3.1.1) in which, apart from the inde-
pendent discrete variables n, m on which the variable un,m depends, there are parameters
which we associate with these independent variables. We can think of these parameters as
being the parameters which measure the width of the grid in the directions associated with
n and m, and we refer to them as lattice parameters. Thus, denoting these parameters by
p, q, the equations under question will take the form:

b, u
Q(u, u eub
e; p, q) = 0 , (3.2.1)

and if we demand that the values for u at each vertex can be solved uniquely, implying
multilinearity in each of these variables around the quadrilateral, then we are led again to
eq. (3.1.2) with the coefficients k0 , . . . , k6 depending on the lattice parameters p and q in a
specific way. The question is what criteria to use in order to chose that dependence!
It is here that we will restrict ourselves to quadrilateral P∆Es which we regard to be
integrable. The question of what is the proper definition integrability, and to answer that
question in general is very difficult: a one-fits-all definition (for all the possible type of
systems that we would like to regard as being integrable) is possibly not possible to give
in a precise mathematical sense. However, if we restrict ourselves here to P∆Es, and in
particular quadrilateral P∆Es, then we can aspire to be a bit more precise. We will explore
the definition of an integrable quadrilateral P∆E by means of the following example, namely
the example of the P∆E arising from the BTs for the KdV equation, eq. (2.2.54).
First, we remark that the presence of the lattice parameters p, q is crucial: whereas
normally we would like to consider the parameters to be chosen once and for all and then
remain fixed (thus specifying a specifi equation) when solving the equation on the lattice,
here we will argue that we should look at (2.2.54) as defining a whole parameter-family of
equations, and that it makes sense to look at them altogether with p and q variable. In
3.2. INTEGRABILITY AND CONSISTENCY-AROUND-THE-CUBE 55

doing this, however, we must attach each parameter to a specific diecrete variable such as
p being associated with the variable n, and q with m. This is also natural from the way in
which we derived (2.2.54) from the BT construction: each BT is attached to a parameter (λ,
µ,. . . ), and with each parameter we can build a direction in an infinite-dimensional lattice
of BTs.
The main point we want to make now is the following:

Statement: the infinite parameter-family of P∆Es represented by the lattice equation,

such as (2.2.54) is compatible, i.e. in each quadrilateral sublattice of the infinite-dimensional
lattice we can consistently impose a copy of the P∆E in terms of the relevant discrete
variable and associated with a corresponding lattice parameter.

Let us illustrate this by means of the example of (2.2.54), in which we will identify (for
later convenience) the parameters 4λ := p2 and 4µ := q 2 . What the statement above
suggests is that we can “embed” the equation in a multi-dimensional lattice by conbsidering
the dependent variable w, not only to depend on n and m (with associated lattice paameters
p and q espectively), but that w may in fact depend on an infinity of lattice variables each
associated with its parameter, as follows:

w = wn,m,h,... = w(n, m, h, . . . ; p, q, r, . . . )

and with each of these variables we have a corresponding elementary shift on the lattice:

e := wn+1,m,h
w , b := wn,m+1,h
w , w := wn,m,h+1 ...

Rewriting the equation (2.2.54) in the form (2.2.46) with the substitutions for λ and µ
investigate what happens if we impose a copy of the same equation in all three lattice
directions. This would lead to the system of equations:

b − w)(w
(w e be = p2 − q 2 ,
− w) (3.2.2a)
e
(w − w)(w −we ) = p2 − r 2 , (3.2.2b)
(w b = r2 − q 2 .
b − w)(w − w) (3.2.2c)

These equations are consistent if the evaluations along the cube are independent of the way
of calculating the final point. Imposing initial values:

w := a , e =: b
w , b =: c
w , w =: d
56 LECTURE 3. INTEGRABILITY OF P∆ES

w e
w

e
w
w

b
w be
w

b
w b
e
w
In fact:
2 2
e = (p − q )w
b ew b + (r2 − p2 )w w
b + (q 2 − r2 )ww e
w . (3.2.3)
(r2 − q 2 )we + (q 2 − p2 )w + (p2 − r2 )w
b
independent of the way in which he value at this vertex is calculated! This property, implying
that under relevant initial value problems on the 3-dimensional lattice the iteration of the
solution can be performed in an unambiguous way (namely independent of the way by which
we perform the calculation of the iterates) will be referred to as the consistency-around-the-
cube (CAC) property. It is this property that we shall consider to be the main hallmarfk of
the integrability of the equation.

Remark: We note also, for later reference, that in the above case the formula for w b
e is
independent of the value w at the opposite end of the main diagonal across the cube. This
property we will refer to as the tetrahedron property.

Other examples of integrable quadrilateral P∆Es: There are many examples of

integrable quadrilateral P∆Es that have been discovered over the years. A number of these
belong to the lattice KdV-family of equations, they comprise the following cases:
Lattice potential KdV equation:

b−u
(p − q + u b
e) = p2 − q 2 ,
e)(p + q + u − u (3.2.4)
and this is, in fact, equivalent to the equation (2.2.54) by the change of (dependent)
variable w = u − np − mq − c (c a constant with respect to n and m). We will show
in Lecture 4 that this equation reduces to the potential KdV equation after a double
continuum limit, so that we rightly regard it as a discretisation of the latter equation.
Lattice potential MKdV equation:
v − veb
p(vb v − vbb
ve) = q(ve ve) (3.2.5)
Solutions of this equation are related to the solutions of the previous one (3.2.4) via
the relations
pe
v − qb
v b pv + qb
ve
b−u
p−q+u e= , e=
p+q+u−u , (3.2.6)
v e
v
3.3. LAX PAIR FOR LATTICE POTENTIAL KDV 57

which constitute the analogues of the Miura transformation (2.2.35).

Lattice SKdV equation:

z −b
(z − zb)(e ze) p2
= 2 (3.2.7)
z −b
(z − ze)(b ze) q

which is the “Schwarzian KdV equation”. This equation is invariant under Möbius
transformations:
αz + β
z 7→ Z = .
γz + δ
Solutions of eq. (3.2.7) are related to the ones of (3.2.5) via the relations:

p(z − ze) = ve
v , q(z − zb) = vb
v. (3.2.8)

Exercise 3.2.1. Show by explicit computation that the P∆Es given by (3.2.4)-(3.2.7) possess
the consistency-around-the-cube property.

Exercise 3.2.2. Show that by using the relations (3.2.6) one can derive (3.2.4) from (3.2.5)
and vice versa.

3.3 Lax pair for lattice potential KdV

We shall next demonstrate how the CAC property explained in the previous section gives rise
to the existence of a Lax pair, i.e. an overdetenrmined linear system of difference equations,
the compatibility of which is verified iff the nonlinear lattice equation is satisfied. The idea
is the following: Having verified the consistency of the equation around the cube, this tells
us that we can add any lattice direction to the original lattice and impose simultaneously
the equation in the three two-dimensional quadrilateral sublattices. The main idea is now
to consider the additional lattice variable h ∈ Z associated with lattice parameter k as
an auxiliary “virtual” variable, whilst acknowledging only the shifts in the orignal lattice
variables n and m as the operations of interest. This then suggests that the shift in the
third direction: w 7→ w should not appear in any of the equations, implying that wherever
w appears we should treat it as a new dependent variable w := W .
Proceeding in this way, we rewrite (3.2.2b) and (3.2.2c) as follows:
2 2
f − w) = k 2 − p2 e
f = wW + (k − p − ww)
e W
(W − w)( ⇒ W , (3.3.1a)
W −we
2 2
c − w) = k 2 − q 2 e
f = wW + (k − p − ww)
b W
(W − w)( ⇒ W . (3.3.1b)
W −we

Noting that eqs. (3.3.1) are both fractional linear in W , we can linearise these equations by
the substitution:
F
W = ,
G
58 LECTURE 3. INTEGRABILITY OF P∆ES

leading to

Fe wF + (k 2 − p2 − ww)G
e
= ,
Ge F − wGe
Fb wF + (k 2 − q 2 − ww)G
b
= ,
b
G F − wGb

and since at least one of the two functions F or G can be chosen freely, this allows us to
split in each of these equations the numerator and denominator to give:
(
(

Fe = γ wF + [k 2 − p2 − ww]Ge , Fb = γ ′ wF + [k 2 − q 2 − ww]G
b ,
e = γ (F − wG) respectively b = γ ′ (F − wG)
G e , G b ,
(3.3.2)
in which γ and γ ′ are to be specified later. What happens next is obvious: we introduce the
two-component vector  
F
φ= ,
G
and write (3.3.2) as a system of two 2×2 matrix equations:

e = Lφ ,
φ b = Mφ ,
φ (3.3.3a)

with the matrices

   
w k 2 − p2 − w w
e ′ w k 2 − q 2 − ww
b
L=γ , M =γ (3.3.3b)
1 −we 1 −w b

How does this linear system work? The consistency relation of the linear problem (3.3.3a)
b
e =φ e
b , which is the condition that simply expresses
is obtained from the condition that φ
that φ must be a proper function of the lattice variables n and m. Calculating the left
hand-side of this condition we get on the one hand

b [ =L
e = (Lφ) b = LM
bφ b φ,
φ

whereas the right hand-side would be calculated as

e ^
b = (M fφe =M
f Lφ .
φ φ) = M

Equating both sides we see that a sufficient condition for the consistency is the matricial
equation:
b
LM =M fL , (3.3.4)
which we will loosely refer to as the Lax equation, but it is sometimes also referred to as
a discrete zero-curvature condition (the reason for this terminology will be explained in
(3.3.1)). Pictorially the Lax equation is illustrated by the following diagram
3.3. LAX PAIR FOR LATTICE POTENTIAL KDV 59

φ L e
φ

f
M
M

b
e
b
φ φ
b
L

where the vectors φ are located at the vertices of the quadrilateral and in which the matrices
L and M are attached to the edges linking the vertices.
Using the explicit form (3.3.3b) of the matrices L and M , and working out the condition
(3.3.4) we find
! 
′ b k 2 − p2 − w
w bwbe w k 2 − q 2 − ww b
bγ
γ be =
1 −w 1 −w b
! 
′ we k2 − q2 − w be
ew w k 2 − p2 − w w e
=γeγ be
1 −w 1 −w e

and we will chose the γ and γ ′ (which were unspecified so far) such that the relation for the
determinants of this equation:

γ b det(M ) = γ
bγ ′ det(L) f ) det(L)
e′ γ det(M (3.3.5)

is trivially satisified. Since in this example the determinants of L and M are given by

det(L) = p2 − k 2 , det(M ) = q 2 − k 2 ,

respectively, it is clear that the condition (3.3.5) is satisified by simply taking γ = γ ′ = 1.

But we will encounter other examples where a nontrivial choice of γ and γ ′ is needed in
order to satisfy the determinantal condition (3.3.5). Working out all the entries on both
sides of the above matrix products it is straightforward to see that the (1,1)- and (2,2) entry
of the matrix equation both yield the same condition on w, namely the equation (2.2.46)!
Moreover, the (2,1) entry is trivially, and the (1,2) entry we don’t even need to calculate,
because having checked three of the entries to be satisfied the final entry must also be
satisfied by virtue of the fact that we have aranged the determinantal condition (3.3.5) to
be satisfied. In conclusion, we see thus that from the Lax equation (3.3.4) we recover the
lattice equation (2.2.54). We observe furthermore, that albeit both Lax matrices L and M
depend on the auxiliary variable k the final nonlinear equation for w does not depend on k!
Exercise 3.3.1. Suppose that the Lax matrices L and M can be expanded in a power series
in a small parameter δ and ǫ respectively as follows:

L = 1 + δL1 + · · · , M = 1 + ǫM 1 + · · · ,
60 LECTURE 3. INTEGRABILITY OF P∆ES

and we expand the shifted variable by Taylor expansion

b = L + ǫ∂t L + · · ·
L , f = L + δ∂x M + · · ·
M ,
by expansion we obtain in dominant order (namely terms proportional to ǫδ) the following
matricial equation:
∂t L1 − ∂x M 1 + [ L1 , M 1 ] = 0 , (3.3.6)
where [ L , M ] = LM − M L denotes the usual matrix commutator bracket. Eq. (3.3.6)
arises also in differential geometry and it is from there that it has the interpretation as
a zero-curvature condition of differential manifolds (with the relevant interpretation of the
matrices L and M ).

3.4 ∗ Classification of Quadrilateral P∆Es

In a beautiful paper by V. Adler, A. Bobenko and Yu. Suris, cf. ref. [8], the problem of
classifying all quadrilateral lattices which are integrable in the sense of the CAC property
discussed in section 3.2 was considered. Since this is in our view an important result, it is
useful to reproduce the whole list of resulting equations here.

Theorem ([8]): Consider quadrilateral P∆Es of the general form:

e, u
Q(u, u b
b, u
e; p, q) = 0 ,
using the notation indicated in the following diagram (renaming the lattice parameters in
order to avoid confusion with previously used notation),

u p e
u
- -

q ? ?q

? - -?
b
u p b
e
u
subject to the following restrictions:
a) Linearity: Q is multilinear in its arguments, i.e., it is linear in each vertex-variable u,e
u,
u b
b,u
e;
b) Symmetry: Q is invariant under the group D4 of symmetries of the square, generated by
the interchangements:

e, u
Q(u, u b
b, u
e; p, q) = ±Q(u, u
b, u b
e, u u, u, b
e; q, p) = ±Q(e e, u
u b; p, q) ; (3.4.1)
c) Tetrahedral Condition: in the consistency check, the evaluation of the point on the cube
b
e is independent of u.
given by u
3.4. ∗ CLASSIFICATION OF QUADRILATERAL P∆ES 61

Q-list:

Q1 : b)(e
p(u − u b
e) − q(u − u
u−u e)(b b
e) = δ 2 pq(q − p)
u−u (3.4.2a)
Q2 : b)(e
p(u − u b
e) − q(u − u
u−u e)(b b
e) + pq(p − q)(u + u
u−u e+u b
b+u
e) =
= pq(p − q)(p2 − pq + q 2 ) (3.4.2b)
Q3 : p(1 − q 2 )(ub
u+u b
e) − q(1 − p2 )(ue
eu u+u bu b
e) =
 2 2

2 2 b 2 (1 − p )(1 − q )
= (p − q ) (b uue + uu
e) + δ (3.4.2c)
4pq
Q4 : p(ue
u+u bu b
e) − q(ub
u+u eub
e) =
pQ − qP  b b

= 2 2
(b
uue + uu
e) − pq(1 + ue uubu
e) (3.4.2d)
1−p q
where P 2 = p4 − γp2 + 1 , Q2 = q 4 − γq 2 + 1 .
H-list:

H1 : b
e)(b
(u − u e) = p2 − q 2
u−u (3.4.3a)
H2 : b
e)(e
(u − u b) = (p − q)(u + u
u−u e+u b
e) + p2 − q 2
b+u (3.4.3b)
H3 : p(ue bb
u+u e) − q(ub
u eu
u+u b
e) = δ 2 (p2 − q 2 ) (3.4.3c)

A-list:

A1 : p (u + ub)(e b
e) − q(u + u
u+u e)(bu+be) = δ 2 pq(p2 − q 2 )
u (3.4.4a)
A2 : 2
p(1 − q )(ub eu
u+u b 2
e) − q(1 − p )(ue
u+u b
bu 2 2
e) + (p − q ) (1 + ue
uu b
bu
e) = 0
(3.4.4b)

The Q-equations are related through the following coalescence diagram, i.e. displaying
how the equations relate to eachother through certain limits on the parameters:

Q3 (Q3 )δ=0

Q4

Q2 Q1 (Q1 )δ=0

Soliton type solutions have been constructed by now for all equations in the ABS list [Nijhoff,
Atkinson & Hietarinta, 2009; Atkinson & Nijhoff, 2010].
62 LECTURE 3. INTEGRABILITY OF P∆ES

Remarks: It is in itself remarkable that the list of equations found is so short. In fact,
all equations in the list Q are in fact special subcases of the last equation, Q4, which was
discovered by V. Adler in 1997. The latter equation can also be expressed as:
h i
A [(u − b)(bu − b) − (a − b)(c − b)] (e b
e − b) − (a − b)(c − b) +
u − b)(u
h i
+B [(u − a)(e u − a) − (b − a)(c − a)] (b
u − a)(ub
e − a) − (b − a)(c − a) =
= ABC(a − b) (3.4.5)

cf. [9]. Here the extra parameters c, C are related through C 2 = 4c3 − g2 c − g3 and to
(a, A) and (b, B) through the relations:

A(c − b) + B(c − a) = C(a − b) ,

 2
1 A+B
a+b+c= .
4 a−b

A third alternative form of Adler’s equation reads as follows:

     
sn(α) vev+b vb
ve − sn(β) vb v + veb
ve − sn(α − β) vevb + vb
ve
 
+ksn(α)sn(β)sn(α − β) 1 + ve v vbb
ve = 0 , (3.4.6)

[Hietarinta, 2005] which is somewhat simpler. In (3.4.6) the sn denote the Jacobi elliptic
functions, cf . Appendix A, of modulus k, i.e. sn(α) = sn(α; k), etc.
It should be noted that all previous examples presented in section 3.2 can be recognised
as special subcases of the equations in the lists Q,H and A, and which can be obtained by
“degeneration” of Adler’s equation in either of its forms. All these equations possess Lax
pairs, but it was only recently that explicit solutions were found for most of the new cases
including of Adler’s equation, cf. e.g. [12].
Adler’s equation is the discrete analogue of a famous soliton equation discovered in 1980
by I. Krichever and S. Novikov, which reads:

3 u2xx − (4u3 − g2 u − g3 )
ut = uxxx − , (3.4.7)
2 ux
and which generalises the Schwarzian KdV equation.

3.5 Lattice KdV Equation

We finish this Lecture by deriving the lattice analogue of the KdV equation, in contrast to
the lattice potential KdV equation (2.2.54) (or equivalently (3.2.4)):

(wn,m − wn+1,m+1 )(wn,m+1 − wn+1,m ) = p2 − q 2 . (3.5.1)

We recall that the difference between the continuous KdV equation (2.2.30) and its potential
version (2.2.39), as discussed in Lecture 2, was simply that its solutions were connected
3.6. SINGULARITY CONFINEMENT 63

through taking a derivative, u = wx . However, on the lattice there are many ways in
which can do the analogue of “taking the derivative”, for instance we can replace it by
taking a difference involving two neighbouring vertices, like ∆n wn,m ≡ wn+1,m − wn,m or
∆m wn,m ≡ wn,m+1 − wn,m , but these are not the only choices. Alternatively we can take
a difference between vertices farther away or across diagonals, such as:

∆wn,m = wn+1,m+1 − wn,m or ∆wn,m = wn,m+1 − wn+1,m ,

or a host of other choices. As long as a well-chosen continuum limit reduces these to a

derivative they can be justifiably be considered to be the discrete analogues of the operation
of taking a partial derivative. The most sensible way to make a choice is to look at the
equation at hand and then decide what would be for the given equation the most natural
choice to apply in that case.
In the case of the lattive potential KdV equation (3.5.1) the natural choice of the discrete
analogue of the derivative seems to be either one of the choices given above, namely a
difference across the diagonal. This suggest that as lattice KdV variables we would take
either one of the two choices:
be
Q=w−w b−w
or R = w e. (3.5.2)

It is then straightforward from (3.2.4) (using the notation in terms of e- and bshifts) to
derive the following equation for Q:

e= a − a
b−Q
Q ⇔
be
R−R =
a
−
a
, (3.5.3)
Q Q be b
R R e

where a ≡ p2 − q 2 . The equations for Q and R are simply related by the fact that QR = a.
A Lax pair for either of the equations (3.5.3) can be easily derived, by starting from
the Lax pair for the potential KdV equation. In fact, rather than writing it as a first order
matricial system, one may derive from the Lax pair a three-point linear equation in terms
of one of the components of the vector φ on which the Lax matrices act. In this way we
obtain the following scalar Lax pair:
be =
ϕ b+Λϕ ,
Qϕ (3.5.4a)
e =
ϕ b+ Rϕ ,
ϕ (3.5.4b)

remembering that QR = a.
Exercise 3.5.1. Derive the Lax pair (3.5.4) from the set of equations (3.3.2) (with γ =
γ ′ = 1) by setting ϕ = G and identifying the spectral variable by Λ = k 2 − q 2 .
Exercise 3.5.2. Show that the consistency condition of the Lax pair (3.5.4) viewed as an
overdetermined discrete linear system, leads to the lattice KdV equation (3.5.3).

3.6 Singularity confinement

In the seminal paper of B. Grammaticos et al. another point of view on integrability was
developed than the one discussed in section 3.2, namely based on the issue of well-posedness
64 LECTURE 3. INTEGRABILITY OF P∆ES

of initial value problems on the lattice. It is based on a phenomenon happening in integrable

discrete systems, which nowadays is called singularity confinement and which describes what
happens with singularities of the solutions of discrete system in the space of initial values.

Proposition: In an integrable lattice equation of KdV type singularities induced by initial

data do not propagate.

As an example let us consider the lattice KdV equation (3.5.3), (setting for convenience
the parameter a = 1 w.l.o.g.)

1 1
Rn+1,m+1 − Rn,m = − .
Rn+1,m Rn,m+1

We will now study how the the initial data progresses, when one hits a singularity. The
initial data is given at the solid line on Figure 3.1, a, b, 0, c, d. When one proceeds from
these initial values one obtains infinity at two places, then one 0 at the next level and finally
two ambiguities of the type ∞ − ∞.
A more detailed analysis with the initial value 01 = ε (small) yields the following values
at the subsequent iterations

1 1 1 1
∞1 = b + − , ∞2 = c + − ,
ε a d ε
at the first stage, and on the next

1 1 1 1
s=a+ − , t=d+ − ,
∞1 f g ∞2
 
1 1 1 1 2
02 = ε + − = −ε + b − c + − ε + ...
∞2 ∞1 a d

b c

a 01 d

f ∞1 ∞2 g

s 02 t

?1 ?2

Figure 3.1: Propagation of singularities in a 2D map. Here n grows in the SE direction and
m in the SW direction.
3.6. SINGULARITY CONFINEMENT 65

Then at the next step we can resolve the ambiguities:

1 1 1 1
?1 = ∞1 + − =c+ − + O(ε)
02 s d a − 1/f
1 1 1 1
?2 = ∞2 + − =b− + + O(ε)
t 02 a d + 1/g

Thus the singularity is confined.

Note that if the lattice equation is deformed, e.g. by taking:

1 λ
Rn+1,m+1 − Rn,m = −
Rn+1,m Rn,m+1

with λ 6= 1, the above fine cancellation would no longer happen, and singularities would
again occur at ?1 and ?2 , and would persist throughout! In that case the singularities are no
longer confined to a finite number of iteration steps, and we conclude that the corresponding
map is not integrable.

Remark: Another point of view on 2D singularity confinement is provided by the require-

ment of “ultra-local” singularity confinement (R. Sahadevan and H. Capel, 2003): One only
assumes a singularity at one point and requires regularity at all other points.
Referring to Figure 3.2 we have initial values are at black disks, the values at open
circles are determined from them. The initial values can be chosen so that u11 = ∞, and the
requirement is that u12 , u21 are finite and ambiguity at u22 can be resolved using ǫ-analysis.
For example for the lattice potential KdV in the form

1
wn+1,m+1 = wn,m − ,
wn,m+1 − wn+1,m

the initial value w01 = w10 implies

w11 = ∞, , w12 = w01 ,

w21 = w10 , , w22 = ∞ − ∞ ?

w02 w12 w22

w01 w11 w21

w00 w20
w10

Figure 3.2: Setting for “ultra-local” singularity confinement.

66 LECTURE 3. INTEGRABILITY OF P∆ES

and a detailed analysis reveals that if w10 = w01 + ǫ then

1
w11 = ǫ + w00 ,
w12 = w01 + ǫ + ǫ2 (w02 − w00 ) + O(ǫ3 ),
w21 = w01 + 0 + ǫ2 (−w20 + w00 ) + O(ǫ3 ),
1
w22 = w11 −
w12 − w21
1 1
= + w00 −
ǫ ǫ + ǫ2 (w02 + w02 − 2w00 ) + O(ǫ3 )
= w02 + w20 − w00 + O(ǫ)
This resolves the ∞ − ∞ singularity, and recovers initial value w00 which was temporarily
submerged to order ǫ2 .

Literature
1. M.J. Ablowitz and F.J. Ladik, A nonlinear difference scheme and inverse scattering, Stud.
Appl. Math. 55 (1976) 213—229; On the solution of a class of nonlinear partial difference
equations, ibid. 57 (1977) 1–12.
2. R. Hirota, Nonlinear partial difference equations I-III, J. Phys. Soc. Japan 43 (1977), 1424–
1433, 2074–2089.
3. E. Date, M. Jimbo and T. Miwa, Method for generating discrete soliton equations I-V, J.
Phys. Soc. Japan 51 (1982) 4116–4131, 52 (1983) 388–393, 761–771.
4. B. Grammaticos, A. Ramani and V. Papageorgiou, Do integrable mappings have the Painlevé
property?, Phys. Rev. Lett. bf 67 (1991) 1825–1828.
5. F.W. Nijhoff, G.R.W. Quispel and H.W. Capel, Direct linearization of nonlinear difference-
difference equations, Phys. Lett. 97A (1983) 125–128; G.R.W. Quispel, F.W. Nijhoff, H.W.
Capel and J. van der Linden, Linear integral equations and nonliner difference-difference
equations, Physica 125A (1984) 344–380.
6. F.W. Nijhoff and H.W. Capel, The discrete Korteweg-de Vries equation, Acta Applicandae
Mathematicae 39 (1995) 133–158.
7. F.W. Nijhoff and A.J. Walker, The discrete and continuous Painlevé VI hierarchy and the
Garnier systems, Glasgow Math. J. 43A (2001) 109–123.
8. V.E. Adler, A.I. Bobenko and Yu.B. Suris, Classification of integrable equations on quad-
graphs, Commun. Math. Phys. 233 (2003) 513–543.
9. F.W. Nijhoff, Lax pair for the Adler (lattice Krichever-Novikov) system, Phys. Lett. 297A
(2002), 49–58.
10. J. Hietarinta, Searching for CAC-maps, J. Nonlin. Math. Phys. 12 Suppl. 2 (2005) 223–230.
11. F.W. Nijhoff, J. Atkinson and J. Hietarinta, Soliton Solutions for ABS Lattice Equations: I
Cauchy Matrix Approach, J. Phys. A: Math. Theor. 42 40 (2009) 404005 (34pp).
12. J. Atkinson and F.W. Nijhoff, A constructive approach to the soliton solutions of integrable
lattice equations, Commun. Math. Phys. 299 (2010) 283–304.
Lecture 4

Special Solutions & Continuum

Limits

Integrable nonlinear equations stand out among other differential or difference equations
by the fact that one can construct infinite classes of solutions in explicit form, i.e. exact
solutions which can be explicitely written down. This distinguishes them from most other
model equations that occur in the applied mathematical sciences.
The partial difference equations (P∆Es) that we have considered in Lecture 3 admit
several types of special solutions. Such solutions include:

• rational solutions;

• soliton solutions;

• periodic solutions;

• similarity solutions.

These different solutions require different techniques to obtain them, and to write them
down in terms of explicit formulae. For example, to obtain soliton solutions we can use
the inverse scattering transform method, or alternatively the technique of Bäcklund trans-
formations, whilst period solutions rely on a method called finite-gap integration, which
relies on techniques from algebraic geometry. In most cases the problem of obtaining such
special solutions amounts to the investigation of reductions of the P∆Es: the choice of the
class of solutions restricts the parameter-space from an infinite-dimensional space to a finite-
dimensional one. Effectively this means that the original P∆E is reduced to an ordinary
difference equation (O∆E) or a system of O∆Es in the process of obtaining the solution,
which could often be expressed through the imposition of additional compatible constraints.
The reduced equations, can then be studied in their own right, and they may take different
shapes, such as:

• discrete-time equations of motion (of a many-body system);

• dynamical mappings;

67
68 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

• ordinary difference equations of Painlevé type,

which then can subsequently solved from that perspective. Thus, we get a solution scheme
that looks like:
P∆E → O∆E system → explicit solution .
In this lecture we will concentrate on two types of special solutions: i) single soliton solutions
obtainable through BTs; ii) simple periodic solutions as dynamical mappings.

4.1 Solutions from discrete Bäcklund transforms

There are several ways to obtain soliton solutions: the inverse scattering method, the Hirota’s
direct method and the use of Bäcklund transformations, see e.g. the textbooks [Lamb, 1980;
Ablowitz & Segur, 1982; Dodd, Eilbeck, Gibbon & Morris, 1982; Newell, 1983; Drazin &
Johnson, 1989]. Here we will focuss primarily on the latter method, which in the discrete
case is particularly natural, because in some sense the equations constituting the BTs are,
in fact, copies of the lattice equations themselves! There are actually two types of BTs:

• auto-Bäcklund tranforms, i.e. nonlinear transformations that bring you from one so-
lution, of the P∆E under consideration, to another solution of the same equation;

• non-auto-Bäcklund tranforms, which bring you from a given solution of a P∆E to a

solution of a different P∆E. In fact, the Miura transformations we have encountered
previously, fall in this class.

In either case, one of the first problems we encounter is to find an initial solution, which
could be a trivial solution (like the zero solution, if the equation admits it), but sometimes
it might be problematic to find such a simple solution. Whenever we have such a solution,
we can use it as a ”seed solution” on which we can implement the BT. Here we will give
some examples, both in the case of a non-auto-BT, as well as in the case of an auto-BT,
how this works in the discrete case (the continuous case was already treated in Lecture 2).

4.1.1 Non-auto-BTs for P∆Es

We have encountered such non-auto-BTs as ”Miura transformations” before, between various
members of the KdV class of differential equations, as well as of the difference equations,
e.g. the relations (3.2.6) between the lattice potential KdV and MKdV equations, and
the relations (3.2.8) between the lattice potential MKdV and the lattice Schwarzian KdV
equation.
We will give another example of such a non-auto-BT between two P∆Es, and show how
it can be used to generate a solution of one of the two equations provided one knows a
solution of the other. The example we will use is between the equations H1 (which coincides
with our standard example (2.2.46)) and H2 of the ABS list presented in subsection 3.4, i.e.
(3.4.3b). Consider the relations:

e
−2ww = p2 + v + ve , (4.1.1a)
b
−2ww = q 2 + v + vb , (4.1.1b)
4.1. SOLUTIONS FROM DISCRETE BÄCKLUND TRANSFORMS 69

which was first given in [Atkinson, 2008]. By eliminating the variable w from both equations
(4.1.1a) and (4.1.1b), using shifts in the variables n and m on the lattice, it is easy to show
that v obeys the lattice equation
 
v − vb)(v − b
(e v ) = (p2 − q 2 ) v + ve + vb + b
e v + p2 + q 2 .
e (4.1.2)

Similarly, by eliminating v from the eqs. (4.1.1) it follows that w obeys the lattice potential
KdV equation
b − w)(w
(w e be = (p2 − q 2 ) .
− w) (4.1.3)

We can now take a known solution of (4.1.3), for instance the solution

w = wn,m = w0 + np + mq , w0 = constant .r.t. n and m , (4.1.4)

and insert this into the relations (4.1.1). Thus, making explicit the dependence on the
variables n and m, we obtain from (4.1.1a)

p2 + vn,m + vn+1,m = −2(w0 + np + mq)(w0 + (n + 1)p + mq)

⇒ vn,m + vn+1,m = −(w0 + np + mq)2 − (w0 + (n + 1)p + mq)2

where the latter equality follows from direct computation. The latter equation being an
O∆E of the form yn + yn+1 = 0 , where yn = vn,m + (w0 + np + mq)2 , we can directly
integrate to obtain a solution of the form yn = c(−1)n , with c a constant w.r.t. the variable
n. Thus, we obtain
vn,m = cm (−1)n − (w0 + np + mq)2 ,

where cm is constant w.r.t. n (but not necessarily w.r.t. m!) as a solution of the first relation
of the non-auto-BT (4.1.1). Subsequently, we have to satisfy also the second relation of the
non-auto-BT, but in this case, due to the symmetry of the relations as well as of the initial
solution, it is easy to see that the above expression for vn,m also yields a solution to the
second relation, i.e. (4.1.1b), provided we take cm = c(−1)m with now c to be constant also
w.r.t. the variable m Hence, we arrive at the solution of H2, (4.1.2), namely

vn,m = c(−1)n+m − (w0 + np + mq)2 . (4.1.5)

Note that the solutions (4.1.4) and (4.1.5) can be extended to solutions on a multidimensional
lattice by assuming that the initial value w0 depends on additional variables associated with
additional lattice directions. Hence, we can etend the solutions easily to a 3-dimensional
lattice by setting

wn,m,h = u0 + np + mq + hr , vn,m,h = c(−1)n+m+h − (w0 + np + mq + hr)2 , (4.1.6)

and it is automatic that this solution will obey copies of the equations (4.1.3) and (4.1.2)
on each 2-dimensional sublattice of the 3-dimensional lattice of variables n, m and h. Thus,
we could call the solution (4.1.6) to be a covariantly extended solution of the quadrilateral
lattice equation.
70 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

4.1.2 Bäcklund transformations for P∆Es

Let us consider the lattice equation

be − w)(w
(w b = p2 − q 2 ,
e − w) (4.1.7)

for which we have seen in Lecture 3 that it possesses the C.A.C. property. This means that
we can impose the same equation in any number of lattice directions of a multidmensional
lattice, each direction of which carries its own lattice parameter. Thus, we can impose on
the same function w of independent discrete variables the additional equations;

e − w)(w
(w e − w) = p2 − k 2 , (4.1.8a)
b b − w) =
(w − w)(w 2
q −k . 2
(4.1.8b)

and we interpret these equations as a BT u 7→ u of a given solution of (4.1.7) to a new

soution, in much the same way as eqs. (2.2.40) of Lecture 2 form the BT of the continuous
equation, namely the ordinary potential KdV equation.

Exercise 4.1.1. By using the compatibility condition w b

e =w e
b arising from shift on the two
equations (4.1.8), show that if w is a solution of (4.1.7) it follows that w is a solution of the
same equation.

In the same way that, once the inital solution w is given, the spatial part of the BT
(2.2.41) is a Riccati equation (and hence linearisable), in the discrete case both parts of
the discrete BT (4.1.8) are discrete Riccati equations. In a similar way as we have done
in subsection 2.2.4 we shall now show how to obtain the 1-soliton solution of (4.1.7) by
implementation of the discrete BT.

Explicit example: one-soltion solution from discrete BT

The solutiom we start off with as a seed solution for the BT is the simple one we have
encountered before (see Lecture 3):
(0)
wn,m = ξ0 + np + mq .

In connection with the C.A.C. property which establishes the multidimensional consistency
of the equation, the main thing to recognise now is that we need also multidimensional
consistency of the solution. This means that additional variables associated with other
lattice directions, may be hidden in the initial value ξ0 , in such a way that, for instance,
ξ 0 = ξ0 + k . Taking this into account, we can now search for a form of the solution wn,m of
the BT (4.1.8), which we assume to be in the form

fn,m
w n,m = w (0)
n,m + , (4.1.9)
gn,m

where the factorisation into the functions f and g wll be taken such that we get a linear set
of equations for these two functions (based on the fact that eqs. (4.1.8) is a set of discrete
4.1. SOLUTIONS FROM DISCRETE BÄCKLUND TRANSFORMS 71

Riccati equations). Substituting (4.1.9) into (4.1.8a) we get

! 
(0) fe f
e
w + −w (0) (0)
e −w −
w (0)
= p2 − k 2
e
g g
 (0)    fe f fe
⇒ e − w(0) f + w
− w e(0) − w (0) =
g e
g g ge
 (0) 
e −w
w (0)
f
fe
⇒ =
,
e
g −f + we(0) − w(0) g

and this can be split into the following (matricial) linear system
  !    
fe e (0) − w(0) ,
w 0 f p+k 0 f
= = ,
ge −1 , w (0)
e −w (0) g −1 p−k g

and a simlar computation holds for the equation for the vector (f, g)T in the other lattice
direction following from (4.1.8b). Thus, we obtain in both lattice directions the linear
matricial equations
φn+1,m = P φn,m , φn,m+1 = Qφn,m , (4.1.10)
with    
p+k 0 q+k 0
P = , Q= .
−1 p−k −1 q−k
Solving this system would lead to the solution of the BT, and this can be done on the basis
of the following observations:
1. the two equations in (4.1.10) can be simultaneously solved, because in fact the two
b e
e = φ.
b
matrices P and Q commute, P Q = QP (check this!), which implies that φ
n m
Thus, the solution of the vector φ is simply given by φn,m = P Q φ0,0 in terms of
an initial vector φ0,0 .
2. crucially the matrices P and Q are lower triangular constant matrices, which implies
that it is easy to compute arbitrary powers of such matrices, following the general rule:
  !
a 0 N aN , 0
A= ⇒ A = N
−bN .
c b c a a−b , bN

(Prove this statement by induction!).

Evaluating the product P n Qm we, thus, obtain the following result:
    
fn,m (p + k)n (q + k)m , 0 f0,0
= 1 .
gn,m − 2k [(p + k)n (q + k)m − (p − k)n (q − k)m ] , (p − k)n (q − k)m g0,0
(4.1.11)
The Bäcklund transformed solution un,m can now be extracted from (4.1.11), simply reading
off the functions fn,m , gn,m and expressing their ratio in terms of the initial value ratio
72 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

f0.0 /g0,0 and thus reconstituting (4.1.9). The end result, after some simple algebra, is the
(1)
solution wn,m ≡ wn,m given by
 n  m
ρn,m (w 0,0 − ξ0 − k) p+k q+k
w n,m = ξ0 +k+np+mq+ 1 , ρn,m ≡ ,
1 − 2k (w 0,0 − ξ0 − k)(ρn,m − 1) p−k q−k
(4.1.12)
and this is the 1-soliton solution of the lattice equation (4.1.7). Note that we see the return
of the plane-wave factors ρn,m which is the ingredient in this exact solution through which
the main dependence on the independent discrete variables n, m enters in the solution.

Exercise 4.1.2. Compare the 1-soliton solution (4.1.12) of the lattice potential KdV equa-
tion with the 1-soliton solution (2.2.43) of the continuous equation, by recalling that the
discrete plane wave factors ρ become continuous exponential functions after a continuum
limit, (see section 4.4).

Finally, the solution of the actual lattice kdV equation, which according to section 3.5 is
the equation (3.5.3) in terms of either Qn,m = wn,m −wn+1,m+1 or Rn,m = wn,m+1 −wn+1,m ,
cf. (3.5.2), can obviously be obtained directly from (4.1.12) in explicit terms.

Remark: Higher soliton solutions for the lattice equation (4.1.7) can be obtained by it-
erating this procedure, or more simply, by applying the lattice equation once again, but
now viewed as a permutability condition from three known solutions, namely the seed solu-
tion w(0) , the 1-soliton solution w(1) with parameter k = k1 and another 1-soliton solution
w(2) with prameter k = k2 , to yield a new 2-soliton solution w(12) (depending on the two
parameters k1 and k2 ) according to the diagram:

k1 w(1) k2

w(0) w(12)

k2 k1
w(2)
This procedure is, however, not the most effective one to obtain higher soliton solutions,
and alternative direct methods can yield these solutions in closed form, using special types
of determinants. (See, e.g., the results in [12]).

4.2 Continuum limits

One interest in difference equations arise from the fact that they appear as finite-difference
approximations to differential equations, e.g. in Numerical Analysis. In that case one would
be intereted to make sure that a well-defined continuum limit yields the continuous equation
that one is interested in studying. Even if discretization of given differential equations is not
the main motivation for investigating many of the equations we have presented so far, one
may be interested in the questions: what continuous equation corresponds to the discrete
4.2. CONTINUUM LIMITS 73

equation we are looking at? Thus, studying systematically continuum limits of given P∆Es
is of interest anyway. In doing this there a two important features to take on board: i)
Continuum limits are not unique – one can have more than one differential equation arising
from one and the same difference equation; ii) in performing a continuum limits one should
respect the solution structure – making sure that the continuum limit of the equations
reflects a limit on the solutions as well. To make sure this is the case, we will motivate the
limits by investigating first what happens to the corresponding linear equations associated
witht he nonlinear equations of interest.

4.2.1 Plane-wave factors and linearisation

As an example for our ideas we will study the continuum limits of (2.2.46), i.e. the P∆E as-
sociated with the Bäcklund transformations of the KdV equation, which we constructed in
Lecture 2. In Lecture 3 we pointed out that the Bäcklund parameters λ and µ, which can be
associated with the directions of the lattice, could be interpreted as parameters measuring
somehow the grid size in each direction. It is, thus, these parameters that can be used as
“tuning parameters” by which the lattice can be shrunk or expanded in a certain direction,
eventually allowing us to shrink the lattice points together to create a continuum of points.
However, there may be many different ways in which this can happen and, thus, in
principle there may be various limits that we could perform on a given lattice equation. In a
continuum limit involving a lattice parameter (or step-size parameter), say h, the operation
of a difference such as (0.0.4) will tend to a derivative, namely by (0.0.3), as was explained in
Lecture 1. Thus, it is by doing Taylor expansions on the shift operators like (0.0.5), namely
by inserting in the equations expansions of the form:

h dy h2 d2 y
y(x + h) = y(x) + + + ··· (4.2.1)
1! dx 2! dx2
into the difference equations, and then by expanding power-by-power in the lattice param-
eter h that we obtain the transition from difference to differential equation. Typically the
equation that emerges as “the continuum limit” of the discrete equation is the coefficient of
the dominant term in this expansion as h → 0.
Performing this sequence of steps on a given lattice equation there are two questions to
answer, namely
• among the various parameters present in the lattice equation, how do we identify the
parameter (or combination of parameters) to take as the one tending to zero in order
to shrink the lattice?
• how do we determine the behaviour of the independent and dependent variables under
the limit on this chosen parameter?
In general a brute force or naive continuum limit may easily lead to a total collapse of the
equation, where the Taylor expansions applied to the equation lead to mismatch of orders,
and hence to a situation where the limit results in no equation at all (due to conflicting
constraints emerging from the expansion) or to a trivial equation in leading order in the
lattice parameter. To avoid this problem and to answer the questions above, it turns out it
is useful to first study the continuum limits of the linearised equation before we attack the
74 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

full nonlinear equation. We will, thus, first derive the linearised form of the lattice equations
under consideration and derive a special class of solutions of these linear equations to use
these as a guidance on how to take nontrivial and consistent limits.
By a linearisation of a nonlinear lattice equation such as (2.2.46) we mean the linear
equation obtained by expanding the dependent variable around a specific known solution
of the nonlinear equation and taking the dominant term in the expansion. The simplest
linearisations are obtained by taking a trivial solution, such as the zero solution (if it exists).
In the case of the example (2.2.46), taking w ≡ 0 is not allowed since it does not lead to a
solution of that equation, but we can modify the equation slightly by changing it into

(p − q + un,m+1 − un+1,m )(p + q + un,m − un+1,m+1 ) = p2 − q 2 , (4.2.2)

which is (3.2.4) in explicit form, by setting as before 4λ = p2 , 4µ = q 2 , and wn,m =

un,m − np − mq . It is easy to see that (4.2.2) admits the solution un,m ≡ 0, i.e. u vanishes
for all n, m. By setting next
un,m = ǫρn,m ,
and expanding up to linear terms in the small parameter ǫ, we obtain the following linear
equation for ρ:

(p + q)(ρn,m+1 − ρn+1,m ) = (p − q)(ρn+1,m+1 − ρn,m ) . (4.2.3)

It is easily verified that the linear lattice equation (4.2.3) obeys the consistency-around-the-
cube property of section 3.2 in the same way as the full nonlinear equation, and hence we can
consistently embed this equation in a higher dimensional lattice by writing the compatible
system:

ρ − ρe) = (p − q)(b
(p + q)(b ρe − ρ) , (4.2.4a)
(p + k)(ρ − ρe) = (p − k)(eρ − ρ) , (4.2.4b)
(q + k)(ρ − ρb) = (q − k)(bρ − ρ) , (4.2.4c)

where ¯ denotes the shift in the third direction associated with lattice parameter k. The
linear system (4.2.4) has many solutions, but we will fix a specific class of solution by
demanding that the specific variable k is associated with a shift ρ 7→ ρ such that ρ = 0.
(One way of thinking of what this means is that the solution ρ with this constraint is
k
obtained from an (inverse) Bäcklund transformation ρ 7→ ρ with seed solution ρ = 0.)
Solving the two relations (4.2.4b) and (4.2.4c) with ρ = 0:
p−k q−k
ρe = ρ , ρb = ρ,
p+k q+k
we obtain the solution:  n  m
p−k q−k
ρn,m = ρ0,0 , (4.2.5)
p+k q+k
with ρ0,0 some arbitrary initial value. It is straightforward to show by direct computation
that (4.2.5) is a solution of (4.2.3).
The solutions (4.2.5) we will refer to as lattice plane-wave factors and have already seen
how they appeared in section 4.1. Thus, not only do they play a role as approximated
4.2. CONTINUUM LIMITS 75

solutions (namely as solutions of the linearised version of the nonlinear lattice equation
(4.2.2)) but they also are the key ingredient in the exact solution of the full nonlinear
equation. It is these solutions that we will exploit in the next section to formulate the
precise limits on parameters and discrete variables to get nontrivial limiting equations.

4.2.2 The semi-continuous limits

We will now study various limits we can perform on the plane wave factors ρ of the specific
form found in (4.2.5). The guiding principle would be to seek ways in which this solution
will approach exponential factors with continuous variables in the exponents. The “trick”
to be used is the following limit which is well-known from basic analysis
 α n
lim 1 + = eα .
n→∞ n

Straight continuum limit

Focusing on one of the factors in (4.2.5) we can easily see that in the limit
m
m→∞ , q → ∞ such that finite (4.2.6)
q

we obtain:
 m  m  m
q−k −2k ξ (−2k)
lim = lim 1 + = lim 1 + = e−2kξ ,
m→∞
q→∞ q + k m→∞
q→∞ q + k m→∞ m (1 − kξ/m)
m=ξq m=ξq

(4.2.7)
using the fact the extra term in the denominator within the brackets becomes negligeable
as m → ∞.
Let us now investigate the effect of this limit on the equations, e.g. the linear equation
(4.2.4a). The idea is to re-interpret the dependent variable as
m
ρ = ρn,m =: ρn (ξ) , ξ = ξ0 + , (4.2.8)
q

where ξ0 is some initial value. This means that any shift in the discrete variable m increments
in the argument of the function ρ by 1/q. The next thing is for 1/q small, to consider Taylor
series expansions of the form:
1 1 1 1 2
ρn,m+1 = ρn (ξ + ) = ρn (ξ) + ∂ξ ρn (ξ) + ∂ ρn (ξ) + · · · ,
q q 2 q2 ξ

and inserting this into the eq. (4.2.4a) we get:

  
p 1 1 1 2
1+ ρn − ρn+1 + ∂ξ ρn + ∂ ρ n + · · ·
q q 2 q2
  
p 1 1 1 2
= −1 + ρn+1 − ρn + ∂ξ ρn+1 + ∂ ρ n+1 + · · · ,
q q 2 q2
76 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

leading in dominant order (in terms of 1q ) to the differential-difference equation (D∆E):

∂ξ (ρn+1 + ρn ) = 2p(ρn+1 − ρn ) . (4.2.9)

This linear equation is a mixed form of differential equation (w.r.t variable ξ) and discrete
(w.r.t. variable n, and is hence a semi-discrete equation. By construction, it can be directly
verified that the form:  n
p−k
ρn (ξ) = e−2kξ ρ0 (0) ,
p+k
is a solution of this equation.
Inspired by this result, let us now turn to the nonlinear equation for u, (4.2.2), and
perform exactly the same limit there. Thus, introducing in a similar was as for the linear
equation the reinterpretation of the discrete variables
m
un,m =: un (ξ) , ξ = ξ0 + , (4.2.10)
q

and we derive by similar Taylor expansions as before:

   
1 1 1 2
p2 − q 2 = p − q + un + ∂ξ un + ∂ u n + · · · − u n+1 ×
q 2 q2
  
1 1 1 2
× p + q + un − un+1 + ∂ξ un+1 + ∂ u n+1 + · · · ,
q 2 q2

Expanding in powers of 1/q, noting that the dominant terms of order O(q 2 ) and order O(q)
cancel identically, we obtain as coefficient of the leading term of order O(1) the following
equation
∂ξ (un + un+1 ) = 2p(un+1 − un ) − (un+1 − un )2 , (4.2.11)
which is consequently the continuum limit of the lattice equation (4.2.2) under this limit as
q → ∞. Eq. (4.2.11) is a nonlinear D∆E, which is of first order in the derivative w.r.t. the
continuous variable ξ and of first order in the discrete variable n as well.

Remark: By comparing eq. (4.2.11) with eq. (2.2.41) we recognise that by the change of
variables x = 2ξ , p2 = 4λ2 , w
b − w = un+1 − un − p , ∂ξ un = 2wx , we recover the spatial
part of the BT of the KdV equation.

Other lattice equations: We can perform simlar limits on the other members of the
KdV family of lattice equations, namely on (3.2.5) amd (3.2.7), leading to
2
∂ξ (vn+1 vn ) = p(vn+1 − vn2 ) (4.2.12)

and
(∂ξ zn )(∂ξ zn+1 ) = p2 (zn − zn+1 )2 , (4.2.13)
respectively. We shall show in section 4.3 that all these D∆Es are integrable by virtue of
the existence of semi-discrete analogues of the Lax pairs in all cases.
4.2. CONTINUUM LIMITS 77

Skew continuum limit

The limit described above is not the only continum limit that we can perform on the lattice
equation. Instead of taking a limit on one of the variables n and m separately, one can
also first mix them up, by means of a change of independent variables on the lattice, before
taking a limit. As we shall see this will lead to quite different semi-continuous equation as
a result of the limit.
To describe this, let us first consider the linearised equation (4.2.3) and the following
change of variables:

ρn,m = Rn+m,m ⇒ ρn+1,m = RN +1,m , ρn,m+1 = RN +1,m+1 , ρn+1,m+1 = RN +2,m ,

(4.2.14)
where N = n + m which can be visualised in the diagram:

R=ρ e = ρe
R

b be b
ee b
R R = ρb R = ρe

This change of independent variables (n, m) 7→ (N = n + m, m) bring the linear equation

(4.2.4a) in the form

(p + q)(RN +1,m+1 − RN +1,m ) = (p − q)(RN +2,m+1 − RN,m ) . (4.2.15)

Let us investigate what happens on the level of the plane-wave factors ρ given by (4.2.5)
with this change of variables (n, m) 7→ (N = n + m, m) . Rearranging factors we get:
 n+m  m  N  m
p−k q−k p+k p−k 2(q − p)k
ρn,m = ρ0,0 = 1+ ρ0,0 := RN,m .
p+k q+k p−k p+k (q + k)(p − k)

Keeping N fixed and setting δ = q − p , we can now perform the limit

n → −∞ , m→∞ , δ→0 such that N fixed , δm finite

and focusing on what happens with the second factor in this limit we observe:
 m  m
2δk 2δk
lim 1 + = lim 1 +
m→∞
δ→0 (p + δ + k)(p − k) m→∞
δ→0 (p2 − k 2 ) + (p − k)δ
δm=τ δm=τ
 m  
τ 2k 2kτ
= lim 1 + = exp , (4.2.16)
m→∞ m (p2 − k 2 ) + (p − k)τ /m) p2 − k 2
78 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

using the fact the extra term in the denominator within the brackets becomes negligeable as
m → ∞. Thus, the limit (4.2.16) makes good sense on the level of the plane-wave factors.
To investigate what happens with the linear lattice equation (4.2.15) under the limit
(4.2.16), we set
R = RN,m =: RN (τ ) , τ = τ0 + mδ , (4.2.17)
(allowing for some constan background value τ0 of the continuous variable), and applying
the Taylor expansion:
1 2 2
RN,m+1 = RN (τ + δ) = RN (τ ) + δ ∂τ RN (τ ) + δ ∂τ RN (τ ) + · · · .
2
Inserting this into the eq. (4.2.15) we get:
  
1 2
(2p + δ) RN +1 + δ ṘN +1 + δ R̈N +1 + · · · − RN +1
2
  
1 2
= −δ RN +2 + δ ṘN +2 + δ R̈N +2 + · · · − RN ,
2

which in leading order O(δ) yields as dominant term in the expansion the equation:

2pṘN = RN −1 − RN +1 . (4.2.18)

It is straightforward to check that

 N  
p−k 2kτ
RN (τ ) = exp R0 (0) , (4.2.19)
p+k p − k2
2

provides a solution of eq. (4.2.18).

Let us now move to the nonlinear equation and perform a similar limit there. In fact,
applying the change of variables (n, m) 7→ (N = n + m, m) in (4.2.2) with the changes:

un,m = Un+m,m ⇒ un+1,m = UN +1,m , un,m+1 = UN +1,m+1 , un+1,m+1 = UN +2,m ,

(4.2.20)
we obtain

(p − q + UN +1,m+1 − UN +1,m )(p + q + UN,m − UN +2,m ) = p2 − q 2 , (4.2.21)

and then reinterpreting the variable U as

U = UN,m =: UN (τ ) , τ = τ0 + mδ , (4.2.22)

we can perform a similar Taylor expansion on U as we did for R. Inserting this into (4.2.21)
we get
   
1
−δ + UN +1 + δ U̇N +1 + δ 2 ÜN +1 + · · · − UN +1 ×
2
  
1
× 2p + δ + UN − UN +2 + δ U̇N +2 + δ 2 ÜN +2 + · · · = −(2p + δ)δ ,
2
4.2. CONTINUUM LIMITS 79

which in leading order yields

(U̇N +1 − 1)(2p + UN − UN +2 ) = −2p ,

or equivalently
2p
U̇N = 1 − . (4.2.23)
2p + UN −1 − UN +1

In spite of the fact that eq. (4.2.23) was derived starting from the same lattice equation,
it is quite a different nonlinear D∆E from (4.2.11), as is evident by inspecting the orders:
this equation is first order in the derivative w.r.t. the continuous variable τ , and second
order w.r.t. the discrete variable N , whereas (4.2.11) is first order in both the discrete and
continuous variables but the derivative w.r.t. ξ acts on both un and un+1 . We will assess in
section 4.3 that both equations (4.2.11) and (4.2.23) are integrable in the sense that there
exists an associated Lax pair in both cases.

Other examples: The process described above to obtain a mixed continuum limit we
refer to as the skew continuum limit of the lattice equation. It can be applied also to the
other members of the lattice KdV family, namely to eqs. (3.2.5) and (3.2.7). In those cases
we obtain
VN −1 − VN +1
p∂τ log VN = , (4.2.24)
VN −1 + VN +1

and
2 (ZN −1 − ZN )(ZN − ZN +1 )
ŻN = (4.2.25)
p ZN −1 − ZN +1

respectively.

4.2.3 Full continuum limit

Finding the full continuum limit of lattice equations such as eq. (4.2.2) is a two-stage process:
first we must establish a semi-discrete (semi-continuous) limit, as we have established in the
previous section, and second to find the full continuum limit from the latter in the next step.
In this section we will perform the second step of the process starting from the skew limit
of the equation, and do a second limit in order to turn the remaining discrete variable into
a continuous one. In fact, in order to obtain a nontrivial limit, we shall see that the process
is slightly more involved , and that we need to mix once again both spatial and temporal
variables in order to obtain a nonlinear PDE, retaining the integrability.
As before we shall use the plane-wave factors to guide us in finding the limit we need
to impose in order to get nontrivial equations from the semi-discrete one. Taking the skew
limit of the plane-wave factor, i.e. the form of the variable RN (τ ) as given in (4.2.19), we
80 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

have
 N         
p−k 2kτ 2kτ k k
exp = exp + N ln 1 − − ln 1 +
p+k p2 − k 2 p2 − k 2 p p
    
2kτ k2 k4 k 1 k3 1 k5
= exp 1 + 2 + 4 + · · · − 2N + + + ···
p2 p p p 3 p3 5 p5
       
τ N 3 τ 1 N 5 τ 1 N
= exp 2k − + 2k − + 2k − + ···
p2 p p4 3 p3 p6 5 p5
3
t+k5 t′ +···
−→ ekx+k .

This leads to the identification of the variables of the full continuum limit as the coefficients
of the various powers of k in the expansion of the exponents of the plane-wave factors,
namely    
τ τ 1 ξ
x=2 −ξ , t=2 − , (4.2.26)
p2 p4 3 p2
in which as before ξ = N/p . Thus, this calculation suggests, first, that the solution (4.2.19)
can be identified as follows:
1 1 1 1
RN (τ ) := R(ξ, τ ) ⇒ RN ±1 (τ ) = R(ξ ± , τ ) = R ± Rξ + 2 Rξξ ± 3 Rξξξ + · · · ,
p p 2p 6p

and, second, to perform the change of variables (4.2.26), i.e.

R(ξ, τ ) := R(x, t) ,

which implies by chain rule:

∂R ∂R ∂x ∂R ∂t 2
= + = −2Rx − Rt , (4.2.27a)
∂ξ ∂x ∂ξ ∂t ∂ξ 3p2
∂R ∂R ∂x ∂R ∂t 2 2
= + = 2 Rx + Rt . (4.2.27b)
∂τ ∂x ∂τ ∂t ∂τ p p4

Since the two seps (i.e. Taylor expansion of the shift in N and change of variables (ξ, τ ) 7→
(x, t)) involve the parameter p, the expansion in powers of 1/p and the selection of the
dominant term can only be done after having done both of these steps. Thus, starting from
the semi-discrete equation (4.2.18) we get
2 1
2p∂τ R = − ∂ξ R − 3 ∂ξ3 R − · · · ⇒
p 3p
     3
2 2 2 2 1 2
⇒ 2p ∂x + ∂t R = − −2∂x − ∂t R − −2∂ x − ∂t R − ··· ,
p2 p4 p 3p2 3p3 3p2

and we observe that while the leading term of order O(1/p) cancels identically, the next
term of order O(1/p3 ) yields precisely the equation:

Rt = Rxxx ,
4.2. CONTINUUM LIMITS 81

as expected.
Turning now to the nonlinear equation (4.2.23) and performing the same continuum limit
there, setting first
1 1 1 1
UN (τ ) := U (ξ, τ ) ⇒ UN ±1 (τ ) = U (ξ ± , τ ) = U ± U ξ + 2 U ξξ ± 3 U ξξξ + · · · ,
p p 2p 6p

which by insertion in (4.2.23) yields:

  −1
1 2 1
∂τ U = 1 − 1 − U ξ + 3 U ξξξ + · · · ,
2p p 3p

and, second, to use similar expressions as (4.2.27) to implement the change of variables
(ξ, τ ) 7→ (x, t) to obtain:
  "    3 #−1
2 2 1 2 1 2
∂x + 4 ∂t U = 1 − 1 − 2 −2∂x − 2 ∂t U − 4 −2∂x − 2 ∂t U − ···
p2 p p 3p 6p 3p
"    3 #    2
1 2 1 2 1 2
= − 2 −2∂x − 2 ∂t U + 4 −2∂x − 2 ∂t U + · · · − 2 −2∂x − 2 ∂t U + · · · − · · ·
p 3p 6p 3p p 3p

expanding the demoninator on the right-hand side in a power series. Again we observe
that the terms of order O(1/p) cancel identically, and that the next dominant order
O(1/p3 ) yields the equation:
Ut = Uxxx − 3Ux2 , (4.2.28)
which is the potential KdV equation (coinciding with (2.2.39) in terms of the variable w up
to a change of sign in the independent variables, i.e. after x 7→ −x, t 7→ −t we recover the
equation for w).

We see that we have come now full circle: we started out with the continuous KdV
equation, derived its Bäcklund transformations, which using the permutability theorem led
to the construction of a lattice of solutions, the relations between these solutions being
reinterpreted as a partial difference equation on the twodimensional lattice. As a dynamical
equation the latter was seen as a discretisation of some continuum equations, both semi-
discrete as well as fully discrete, and the full continuum limit now turns out to be the
equation we started out from, namely the KdV itself. Perhaps this is not so surprising,
but what is remarkable is that the equations at all levels are compatible with eachother:
the continuous and discrete equations can be imposed simultaneously on one and the same
variable u, and, in fact, the variable U of (2.2.39) is nothing else than the variable w of
Lecture # 2.

Remark: Performing similar continuum limits on the semi-discrete equations (4.2.24) and
(4.2.25) we recover the following fully continuous equations, namely

Vx Vxx
Vt = Vxxx − 3 , (4.2.29)
V
82 LECTURE 4. SPECIAL SOLUTIONS & CONTINUUM LIMITS

which is the potential MKdV equation, and

2
3 Zxx
Zt = Zxxx − , (4.2.30)
2 Zx
which is the Schwarzian KdV equation.

Literature
1. G.L. Lamb, Elements of Soliton Theory, (Wiley Interscience, 1980).

2. M.J. Ablowitz and H. Segur, Solitons and the Inverse Scattering Transform, (SIAM, 1982).

3. R.K. Dodd, J.C. Eilbeck, J.D. Gibbon and H.C. Morris, Solitons and Nonlinear Wave Equa-
tions, (Academic Press, 1982).

4. A.C. Newell, Solitons in Mathematics and Physics, (SIAM, 1983).

5. P.G. Drazin and R.S. Johnson, Solitons. An Introduction, (Cambridge Univ. Press, 1989).

6. T. Miwa, M. Jimbo and E. Date, Solitons, Differential Equations, Symmetries and Infinite
Dimensional Lie Algebras, (Cambridge University Press, 2000).

7. M.J. Ablowitz and P.A. Clarkson, Solitons, Nonlinear Evolution Equations and Inverse Scat-
tering, LMS Series (Cambridge Univ. Press, 1991).

8. R. Hirota, Exact solution of the Korteweg-de Vries equation for multiple collisions of solitons,
Phys. Rev. Lett. 27 (1971), 1192.

9. R. Hirota, Direct Method for finding exact solutins of nonlinear evolution equations, Lect.
Notes in Math. 515 (Springer Verlag, 1976), p. 40.

10. R. Hirota, Direct Methods in Soliton Theory, in: R.K. Bullough and P.J. Caudrey, eds.,
Solitons, (Springer Verlag, 1980).

11. J. Atkinson, Bäcklund transformations for integrable lattice equations, J. Phys. A: Math.
Theor. 41 (2008) 135202 (8 pp.).

12. J. Atkinson, J. Hietarinta and F.W. Nijhoff, Seed and soliton solutions for Adler’s lattice
equation, J. Phys. A: Math. Theor. 40 1 (2007) F1–F8.

13. J. Atkinson, J. Hietarinta and F.W. Nijhoff, Soliton solutions for Q3, J. Phys. A: Math
Theor. 41 14 (2008) 142001 (11pp)

14. V.G. Papageorgiou, F.W. Nijhoff and H.W. Capel, Integrable Mappings and Nonlinear Inte-
grable Lattice Equations, Physics Letters 147A (1990), 106–114.

15. H.W. Capel, F.W. Nijhoff, H.W. Capel, Complete Integrability of Lagrangian Mappings and
Lattices of KDV Type, Physics Letters 155A (1991), 377–387.

16. G.R.W. Quispel, H.W. Capel, V.G. Papageorgiou and F.W. Nijhoff, Integrable Mappings
derived from Soliton Equations, Physica 173A (1991), 243–266.
Contents

1 Elementary Theory of Difference Equations 11

1.1 Elementary Difference Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.1 Difference versus differential operators . . . . . . . . . . . . . . . . . . 11
1.1.2 The finite-difference operator . . . . . . . . . . . . . . . . . . . . . . . 13
1.1.3 Difference analogues for some familiar functions . . . . . . . . . . . . . 14
1.1.4 Discrete integration: the antidifference operator . . . . . . . . . . . . . 16
1.1.5 ∗ q-Difference operators and Jackson integrals . . . . . . . . . . . . . . 17
1.2 Linear (finite-)difference equations . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2.1 Linear constant coefficient difference equations . . . . . . . . . . . . . 18
1.2.2 Linear first order ∆Es with nonconstant coefficients . . . . . . . . . . 21
1.2.3 Higher order linear difference equations and linear (in)dependence . . 22
1.2.4 Further techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3 A simple nonlinear difference equation – discrete Riccati equation . . . . . . . 27
1.3.1 Continuous Riccati equation . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.2 Discrete Riccati equation . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4 Partial Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.4.1 Discretization and boundary values . . . . . . . . . . . . . . . . . . . . 31
1.4.2 Separation of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2 From Continuous to Discrete Equations via Transformations 35

2.1 Special Functions and linear equations . . . . . . . . . . . . . . . . . . . . . . 36
2.1.1 Weber Functions and Hermite Polynomials . . . . . . . . . . . . . . . 36
2.1.2 Darboux and Bäcklund transformation in general . . . . . . . . . . . . 38
2.1.3 Another example: Bessel Functions . . . . . . . . . . . . . . . . . . . . 40
2.2 Bäcklund transformations for non-linear PDEs . . . . . . . . . . . . . . . . . 42
2.2.1 Lax pair for KdV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.2.2 Miura transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.3 The Bäcklund transformation . . . . . . . . . . . . . . . . . . . . . . 44
2.2.4 Using BTs to generate multisoliton solutions . . . . . . . . . . . . . . 45
2.2.5 Permutability property of BTs . . . . . . . . . . . . . . . . . . . . . . 46
2.2.6 Bäcklund transformation for the sine-Gordon equation . . . . . . . . . 48
2.2.7 Transition to lattice equations . . . . . . . . . . . . . . . . . . . . . . 49

83
84 CONTENTS

3 Integrability of P∆Es 51
3.1 Quadrilateral P∆Es . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Integrability and Consistency-around-the-Cube . . . . . . . . . . . . . . . . . 54
3.3 Lax pair for lattice potential KdV . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4 ∗ Classification of Quadrilateral P∆Es . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Lattice KdV Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.6 Singularity confinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Special Solutions & Continuum Limits 67

4.1 Solutions from discrete Bäcklund transforms . . . . . . . . . . . . . . . . . . . 68
4.1.1 Non-auto-BTs for P∆Es . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1.2 Bäcklund transformations for P∆Es . . . . . . . . . . . . . . . . . . . 70
4.2 Continuum limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.1 Plane-wave factors and linearisation . . . . . . . . . . . . . . . . . . . 73
4.2.2 The semi-continuous limits . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2.3 Full continuum limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79