DDE-BIFTOOL: a Matlab package

for bifurcation analysis

of delay dierential equations

K. Engelborghs

Report TW 305, March 2000

n Katholieke Universiteit Leuven

Department of Computer Science
Celestijnenlaan 200A { B-3001 Heverlee (Belgium)
 DDE-BIFTOOL: a Matlab package
for bifurcation analysis
of delay dierential equations

K. Engelborghs

Report TW 305, March 2000

Department of Computer Science, K.U.Leuven

Abstract
DDE-BIFTOOL is a collection of Matlab routines for numer-
ical bifurcation analysis of systems of delay dierential equations
with several xed, discrete delays. The package allows to compute,
continue and analyse stability of steady state solutions and peri-
odic solutions. It further allows to compute and continue steady
state fold and Hopf bifurcations and to switch, from the latter, to
an emanating branch of periodic solutions. To analyse the stabil-
ity of steady state solutions, approximations are computed to the
rightmost, stability-determining roots of the characteristic equation
which can subsequently be used as starting values in a Newton pro-
cedure. For periodic solutions, approximations to the Floquet mul-
tipliers are computed. We describe the structure of the package, its
routines, and its data and method parameter structures. We illus-
trate its use through a step-by-step analysis of an example system.

Keywords : delay equations, periodic solutions, collocation methods.

AMS(MOS) Classication : Primary : 65J15, Secondary : 65P05.
1 Introduction
This report is a user manual for the package DDE-BIFTOOL, version 1.00. DDE-BIFTOOL con-
sists of a set of routines written in Matlab [18], a widely used environment for scientic computing.
The aim of the package is to provide a tool for numerical bifurcation analysis of steady state solu-
tions and periodic solutions of delay dierential equations with multiple xed, discrete delays (DDEs).
The package is freely available for scientic use. A list of the les which constitute DDE-BIFTOOL
is contained in appendix A. A copyright and warranty notice together with instructions on obtain-
ing the package can be found in appendix B. Up-to-date information can be found on the web page
http://www.cs.kuleuven.ac.be/~koen/delay/ddebiftool.shtml. Note that the package is typical
research software and is provided "as is" without warranty of any kind (see appendix B).
In the rest of this report we assume the reader is familiar with the notion of a delay dierential
equation and with the basic concepts of bifurcation analysis for ordinary dierential equations. The
theory on delay dierential equations and a large number of examples are described in several books.
Most notably the early [4, 8, 7, 15, 22] and the more recent [2, 20, 16, 5, 21]. Several excellent books
contain introductions to dynamical systems and bifurcation theory of ordinary dierential equations, see,
e.g., [27, 14, 1, 28, 23].
A large number of packages exist for bifurcation analysis of systems of ordinary dierential equations
as, e.g., AUTO, LocBif, DsTool and CONTENT, see [6, 19, 3, 24]. For delay dierential equations no com-
parable software is publicly available. For simulation (time integration) of delay dierential equations the
reader is, e.g., referred to the packages ARCHI, DKLAG6, XPPAUT, DDVERK and dde23, see [26, 31, 13,
12, 29]. Of these, only XPPAUT has a graphical interface (and allows limited stability analysis of steady
state solutions of DDEs along the lines of [25]). An up-to-date list of (and links to) available software for
DDEs can be found on the web page http://www.cs.kuleuven.ac.be/~koen/delay/software.shtml.
DDE-BIFTOOL allows to compute branches of steady state solutions and steady state fold and Hopf
bifurcations using continuation. Given an equilibrium it allows to approximate the rightmost, stability
determining roots of the characteristic equation which can further be corrected using a Newton iteration.
Furthermore, periodic solutions and approximations of the Floquet multipliers can be computed using
orthogonal collocation with adaptive mesh selection and branches of periodic solutions can be continued
starting from a previously computed Hopf point or an initial guess of a periodic solution prole.
The package can best be compared with one of the early versions of AUTO as it does not provide
simulation but does provide the continuation of steady state solutions and of periodic solutions using
orthogonal collocation. A large dierence is that no automatic detection of bifurcations is supported.
Instead the evolution of the eigenvalues can be computed along solution branches which allow the user
to detect and identify bifurcations using appropriate visualization.
The remainder of this paper is structured as follows. In section 2 the structure of DDE-BIFTOOL
is outlined. Some necessary notations and properties of delay dierential equations are brie y described
in section 3. Usage of the package will be illustrated by means of the example system given in section 4.
How such a system can be dened for use inside the package is described in section 5. The data structures
used to represent points, branches, stability information and method parameters are described in section
6. Usage of the code is illustrated in section 7 through a step-by-step analysis of the example system.
Input and output parameter descriptions of routines used to compute and manipulate individual points
are described in section 8. Similar descriptions are provided for routines to compute and manipulate
branches in section 9. More details on the numerical methods and the corresponding method parameters
are given in section 10. Finally, the report ends with some brief comments on limits to the package and
future plans in section 11.

2 Structure of DDE-BIFTOOL
The structure of the package is depicted in gure 1. It consists of four layers. Layer 0 contains the system
denition and consists of routines which allow to evaluate the right hand side f and its derivatives and

1
 USER
Layer 3 : BRANCH MANIPULATION

Layer 2 : POINT MANIPULATION

Layer 1 : NUMERICAL METHODS

Layer 0 : SYSTEM DEFINITION

Figure 1: The structure of DDE-BIFTOOL. Arrows indicate the calling ( ) or writing (
) of routines in a
certain layer.

to set or get the parameters and the delays. It should be provided by the user and is explained in more
detail in section 5. All le names in this layer start with "sys ".
Layer 1 forms the numerical core of the package and is (normally) not directly accessed by the user.
The numerical methods used are explained brie y in section 10, more details can be found in the papers
[25, 10, 9] and in [11]. Its functionality is hidden by and used through layers 2 and 3.
Layer 2 contains routines to manipulate individual points. Names of routines in this layer start
with "p ". A point has one of the following four types. It can be a steady state point (abbreviated
"stst"), steady state Hopf (abbreviated "hopf") or fold (abbreviated "fold") bifurcation point or a periodic
solution point (abbreviated "psol"). Furthermore a point can contain additional information concerning
its stability. Routines are provided to compute individual points, to compute and plot their stability and
to convert points from one type to another.
Layer 3 contains routines to manipulate branches. Names of routines in this layer start with "br ".
A branch is the combination of an array of (at least two) points, three sets of method parameters and
specications concerning the free parameters. The array contains points of the same type ordered along
the branch. The method parameters concern the computation of individual points, the continuation
strategy and the computation of stability. The parameter information includes specication of the free
parameters (which are allowed to vary along the branch), parameter bounds and maximal step sizes.
Routines are provided to extend a given branch (that is, to compute extra points using continuation), to
(re)compute stability along the branch and to visualize the branch and/or its stability.
Layers 2 and 3 require specic data structures, explained in section 6, to represent points, stability
information, branches, to pass method parameters and to specify plotting information. Usage of these
layers is demonstrated in section 7 through a step-by-step analysis of the example system. Descriptions
of input/output parameters and functionality of all routines in layers 2 and 3 are given in sections 8
respectively 9.

3 Delay dierential equations

Consider the system of delay dierential equations,
d
dt x(t) = f (x(t); x(t 1 ); : : : ; x(t m ); ); (1)

2
where x(t) 2 Rn , f : Rn(m+1)  Rp ! Rn is a nonlinear smooth function depending on a number of
parameters  2 Rp , and delays i > 0, i = 1; : : : ; m. Call  the maximal delay,
 = i=1max :
;:::;m i
The linearization of (1) around a solution x (t) is the variational equation, given by,
d y(t) = A (t)y(t) + X m
dt 0 Ai (t)y(t i ); (2)
i=1
where, using f  f (x0 ; x1 ; : : : ; xm ; ),

@f 
Ai (t) = @x ; i = 0; : : : ; m: (3)
i  
(x (t);x (t 1 );:::;x (t m );)
If x (t) corresponds to a steady state solution,
x (t)  x 2 Rn ; with f (x ; x ; : : : ; x ; ) = 0;
then the matrices Ai (t) are constant, Ai (t)  Ai , and the corresponding variational equation (2) leads to
a characteristic equation. Dene the n  n-dimensional matrix
as
m
X

() = I A0 Ai e i :
i=1
Then the characteristic equation reads,
det(
()) = 0: (4)
Equation (4) has an innite number of roots  2 C which determine the stability of the steady state
solution x . The steady state solution is (asymptotically) stable provided all roots of the characteristic
equation (4) have negative real part; it is unstable if there exists a root with positive real part. It is
known that the number of roots in any right half plane <() > , 2 R is nite, hence, the stability is
always determined by a nite number of roots.
Bifurcations occur whenever roots move through the imaginary axis as one or more parameters are
changed. Generically a fold bifurcation (or turning point) occurs when the root is real and a Hopf
bifurcation occurs when it is a complex pair.
A periodic solution x (t) is a solution which repeats itself after a nite time, that is,
x (t + T ) = x (t); for all t:
Here T > 0 is the period. The stability around the periodic solution is determined by the time integration
operator S (T; 0) which integrates the variational equation (2) around x (t) from time t = 0 over the
period. This operator is called the monodromy operator and its (innite number of) eigenvalues, which
are independent of the starting moment t = 0, are called the Floquet multipliers. Furthermore, if T  
then S (T; 0) is compact.
For autonomous systems there is always a trivial Floquet multiplier at 1, corresponding to a perturb-
ation around the periodic solution. The periodic solution is stable provided all multipliers (except the
trivial one) have modulus smaller than 1, it is unstable if there exists a multiplier with modulus larger
than 1.

4 The illustrative example

As an illustrative example we will use the following system of delay dierential equations, taken from
[30], 
x_1 (t) = x1 (t) +  tanh(x1 (t s )) + a12 tanh(x2 (t 2 )) (5)
x_2 (t) = x2 (t) +  tanh(x2 (t s )) + a21 tanh(x1 (t 1 )):
3
This system models two coupled neurons with time delayed connections. It has two components (x1 and
x2 ), three delays (1 , 2 and s ), and four parameters (,  , a12 and a21 ). A Matlab denition of system
(5) for use inside the package is given in section 5. Subsequent analysis of the system using the package
is demonstrated in section 7.

5 System denition
To dene a system, the user should provide the following Matlab functions, given here for system (5).
 sys init.m:
Before investigating a given system, a single call is made to a routine sys init.m which has no
arguments and returns the name and dimension n of the system under study. This routine also adds
the directory in which the package resides to the current path variable. Hence, after calling this
routine, DDE-BIFTOOL can be used from within the directory of the system (being preferably
dierent from the directory of the package). The specic directory entered into the path command
depends on the platform used (see help path in Matlab). If necessary some global variables used
in the system denition can also be declared here.
function [name,dim]=sys_init()

name='neuron';
dim=2;
path=path(path,'/home/koen/DELAY/matlab/dde_biftools/');

return;

 sys rhs.m:
The right hand side of the system is dened in sys rhs.m. It has two arguments, xx 2 Rn(m+1)
which contains the state variable(s) at the present and in the past, xx = [x(t) x(t 1 ) : : : x(t m )],
and par 2 R1p which contains the parameters, par = . The delays i , i = 1 : : : ; m are considered
to be part of the parameters (i = j(i) , i = 1; : : : ; m). This is natural since the stability of steady
solutions and the position and stability of periodic solutions depend on the values of the delays.
Furthermore delays can occur both as a 'physical' parameter and as delay, as in x_ = x(t  ).
From these inputs the right hand side f is evaluated at time t. Notice that the parameters have a
specic order in par indicated in the comment line.
function f=sys_rhs(xx,par)

% kappa beta a12 a21 tau1 tau2 tau_s

f(1,1)=-par(1)*xx(1,1)+par(2)*tanh(xx(1,4))+par(3)*tanh(xx(2,3));
f(2,1)=-par(1)*xx(2,1)+par(2)*tanh(xx(2,4))+par(4)*tanh(xx(1,2));

return;

 sys deri.m:
Several derivatives of the right hand side function f need to be evaluated and should be supplied
via a routine sys deri.m. The function sys deri has as input variables xx and par (with ordering
of state variables and parameters as before), nx, np and v. Here, v 2 C n1 or empty. The result J
is a matrix of partial derivatives of f which depends on the type of derivative requested via nx and
np multiplied with v (when nonempty), see table 1.

4
 J is informally dened as follows. Initialize J with f . If nx is nonempty take the derivative of
J with respect to each of its elements. Each element is a number between 0 and m based on
f  f (x0 ; x1 ; : : : ; xm ; ). E.g., if nx has only one element take the derivative with respect to xnx1(1) .
If it has two elements, take, of the result, the derivative with respect to xnx1(2) and so on. Similarly,
if np is nonempty take, of the resulting J, the derivative with respect to np (i) where i ranges over
all the elements of np, 1  i  p. Finally, if v is not an empty vector multiply the result with v. The
latter is used to prevent J from being a tensor if two derivatives with respect to state variables are
taken (when nx contains two elements). Not all possible combinations of these derivatives should be
provided. In the current version, nx has at most two elements and np at most one. The possibilities
are further restricted as listed in table 1.
In the last row of table 1 the elements of J are given by,
  n !
@ @ X @fi v ;
Ji;j =
@xnx(2) Anx(1) v i;j
=
@xnxj (2) nx(1) k
k=1 @xk
with Al as dened in (3).
length(nx) length(np) v J
@f = A nn
1 0 empty @xnx(1) @f nx(1) 2 R
n1
0 1 empty @np2 (1) 2 R
@f nn
1 1 empty @xnx(1) @np 2
R
2 0 2 C n1 @x
@nx(2) Anx(1) v 2 C nn
Table 1: Results of the function sys deri depending on its input parameters nx, np and v using f 
f (x0 ; x1 ; : : : ; xm ;  ).

The resulting routine is quite long, even for the small system (5). Furthermore, implementing so
many derivatives is an activity prone to a number of typing mistakes. Hence a default routine
df deriv.m is available which implements nite dierence formulas to approximate the requested
derivatives (using several calls to sys rhs). A copy of this le can be used to replace sys deri.m.
It is, however, recommended to provide at least the rst order derivatives with respect to the state
variables using analytical formulas. These derivatives occur in the determining systems for fold
and Hopf bifurcations and in the computation of characteristic roots and Floquet multipliers. All
other derivatives are only necessary in the Jacobians of the respective Newton procedures and thus
in uence only the convergence speed.

5
function J=sys_deri(xx,par,nx,np,v)

% kappa beta a12 a21 tau1 tau2 tau_s

J=[];

if length(nx)==1 & length(np)==0 & isempty(v)

% first order derivatives wrt state variables
if nx==0 % derivative wrt x(t)
J(1,1)=-par(1);
J(2,2)=-par(1);
elseif nx==1 % derivative wrt x(t-tau1)
J(2,1)=par(4)*(1-tanh(xx(1,2))^2);
J(2,2)=0;
elseif nx==2 % derivative wrt x(t-tau2)
J(1,2)=par(3)*(1-tanh(xx(2,3))^2);
J(2,2)=0;
elseif nx==3 % derivative wrt x(t-tau_s)
J(1,1)=par(2)*(1-tanh(xx(1,4))^2);
J(2,2)=par(2)*(1-tanh(xx(2,4))^2);
end;
elseif length(nx)==1 & length(np)==1 & isempty(v)
% mixed state, parameter derivatives
if nx==0 % derivative wrt x(t)
if np==1 % derivative wrt beta
J(1,1)=-1;
J(2,2)=-1;
else
J=zeros(2);
end;
elseif nx==1 % derivative wrt x(t-tau1)
if np==4 % derivative wrt a21
J(2,1)=1-tanh(xx(1,2))^2;
J(2,2)=0;
else
J=zeros(2);
end;
elseif nx==2 % derivative wrt x(t-tau2)
if np==3 % derivative wrt a12
J(1,2)=1-tanh(xx(2,3))^2;
J(2,2)=0;
else
J=zeros(2);
end;
elseif nx==3 % derivative wrt x(t-tau_s)
if np==2 % derivative wrt beta
J(1,1)=1-tanh(xx(1,4))^2;
J(2,2)=1-tanh(xx(2,4))^2;
else
J=zeros(2);
end;
end;

6
elseif length(nx)=0 & length(np)==1 & isempty(v)
% first order derivatives wrt parameters
if np==1 % derivative wrt kappa
J(1,1)=-xx(1,1);
J(2,1)=-xx(2,1);
elseif np==2 % derivative wrt beta
J(1,1)=tanh(xx(1,4));
J(2,1)=tanh(xx(2,4));
elseif np==3 % derivative wrt a12
J(1,1)=tanh(xx(2,3));
J(2,1)=0;
elseif np==4 % derivative wrt a21
J(2,1)=tanh(xx(1,2));
elseif np==5 | np==6 | np==7 % derivative wrt tau
J=zeros(2,1);
end;
elseif length(nx)==2>0 & length(np)==0 & ~isempty(v)
% second order derivatives wrt state variables
if nx(1)==0 % first derivative wrt x(t)
J=zeros(2);
elseif nx(1)==1 % first derivative wrt x(t-tau1)
if nx(2)==1
th=tanh(xx(1,2));
J(2,1)=-2*par(4)*th*(1-th*th)*v(1);
J(2,2)=0;
else
J=zeros(2);
end;
elseif nx(1)==2 % derivative wrt x(t-tau2)
if nx(2)==2
th=tanh(xx(2,3));
J(1,2)=-2*par(3)*th*(1-th*th)*v(2);
J(2,2)=0;
else
J=zeros(2);
end;
elseif nx(1)==3 % derivative wrt x(t-tau_s)
if nx(2)==3
th1=tanh(xx(1,4));
J(1,1)=-2*par(2)*th1*(1-th1*th1)*v(1);
th2=tanh(xx(1,4));
J(2,2)=-2*par(2)*th2*(1-th2*th2)*v(2);
else
J=zeros(2);
end;
end;
end;

if isempty(J)
err=[nx np size(v)]
error('SYS_DERI: requested derivative could not be computed!');
end;

return;

7
  sys tau.m:
As a last system routine a function is required which returns the position of the delays in the
parameter list. The order in this list corresponds to the order in which they appear in xx as passed
to the functions sys rhs and sys deri.
function tau=sys_tau()

% kappa beta a12 a21 tau1 tau2 tau_s

tau=[5 6 7];

return;

 sys cond.m: A system routine sys cond can be used to add extra conditions during corrections
and continuation, see section 10.2.

6 Data structures
In this section we describe the data structures used to present individual points, stability information,
branches of points, method parameters and plotting information.
The Matlab structure array is an array of elds each of which is a named variable containing some
value(s) (similar to the struct in C and the record in the Pascal programming language). The structure
allows to group variables into a combined entity using meaningful names. Individual elds are addressed
by appending a dot and the eld name to the structure array variable name. Dening for instance a
steady state point as a structure containing the elds 'type', 'parameter', 'x' and 'stability' (see also
further) can be done using the following Matlab commands.
>> stst.type='stst';
>> stst.parameter=[1 2 -0.1 5];
>> stst.x=[0 0]';
>> stst.stability=[];
>> stst
stst = type: 'stst'
parameter: [1 2 -0.1000 5]
x: [2x1 double]
stability: []

More information about the Matlab structure array can be obtained by typing help struct on the Matlab
command line.
Point structures Table 2 describes the structures used to represent a single steady state, fold, Hopf
and periodic solution point.
A steady state solution is represented by the parameter values  (which contain also the delay values,
see section 5) and x . A fold bifurcation is represented by the parameter values , its position x and a
null-vector of the characteristic matrix
(0). A Hopf bifurcation is represented by the parameter values
, its position x , a frequency ! and a (complex) null-vector of the characteristic matrix
(i!).
A periodic solution is represented by the parameter values , the period T and a time-scaled prole
x (t=T ) on a mesh in [0,1]. The mesh is an ordered collection of interval points f0 = t0 < t1 < : : : < tL =
1g and representation points ti+ dj , i = 0; : : : ; L 1, j = 1; : : : ; d 1 which need to be chosen in function
of the interval points as
ti+ dj = ti + dj (ti+1 ti ):

8
 eld content eld content eld content
type 'stst' type 'fold' type 'hopf'
parameter R1p parameter R1p parameter R1p
x Rn1 x Rn1 x Rn1
stability empty or struct v Rn1 v C n1
stability empty or struct omega R
stability empty or struct
eld content
type 'psol'
parameter R1p
mesh [0; 1] 1 (Ld +1) or empty
degree N0
prole Rn(Ld+1)
period R+0
stability empty or struct

Table 2: Field names and corresponding content for the point structures used to represent steady state solutions,
fold and Hopf points and periodic solutions. Here, n is the system dimension, p is the number of parameters, L
is the number of intervals used to represent the periodic solution and d is the degree of the polynomial on each
interval.

Note that this assumption is not checked but needs to be fullled for correct results! The
prole is a continuous piecewise polynomial on the mesh. More specically, it is a polynomial of degree
d on each subinterval [ti ; ti+1 ], i = 0; : : : ; L 1. Each of these polynomials is uniquely represented by its
values at the points fti+ dj gj=0;:::;d. Hence the complete prole is represented by its value at all the mesh
points,
x (ti+ dj ); i = 0; : : : ; L 1; j = 0; : : : ; d 1; and x (tL ):
Because polynomials on adjacent intervals share the value at the common interval point, this represent-
ation is automatically continuous (it is, however, not continuously dierentiable). (As indicated in table
2, the mesh may be empty, which indicates the use of an equidistant, xed mesh.)
The point structures are used as input to the point manipulation routines (layer 2) and are used inside
the branch structure (see further). The order of the elds in the point structures is important (because
they are used as elements of an array inside the branch structure). No such restriction holds for the other
structures (method, plot and branch) described in the rest of this section.
Stability structures Each of the point structures contains a stability eld. If no stability was com-
puted this eld is empty, otherwise, it contains the computed stability information in the form described
in table 3.
For steady state, fold and Hopf points, approximations to the rightmost roots of the characteristic
equation are provided in eld 'l0' in order of decreasing real part. The steplength that was used to obtain
the approximations is provided in eld 'h'. Corrected roots are provided in eld 'l1' and the number
of Newton iterations applied for each corrected root in a corresponding eld 'n1'. If unconverged roots
are discarded, 'n1' is empty and the roots in 'l1' are ordered with respect to real part; otherwise the
order in 'l1' corresponds to the order in 'l0' and an element 1 in 'n1' signals that no convergence was
reached for the corresponding root in 'l0' and the last computed iterate is stored in 'l1'. The collection of
uncorrected roots presents more accurate yet less robust information than the collection of approximate
roots, see section 10. For periodic solutions only (uncorrected) approximations to the Floquet multipliers
are provided in a eld 'mu' (in order of decreasing modulus).

9
 eld content eld content
h R mu C nm
l0 C nl
l1 C nc
n1 (f 1g [ N 0 )nc or empty

Table 3: Stability structures for roots of the characteristic equation (in steady state, fold and Hopf structures)
(left) and for Floquet multipliers (in the periodic solutions structure) (right). Here, nl is the number of approx-
imated roots, nc is the number of corrected roots and nm is the number of Floquet multipliers.

Method parameters To compute a single steady state, fold, Hopf or periodic solution point several
method parameters have to be passed to the appropriate routines. These parameters are collected into a
structure with the elds given in table 4.
eld content default value
newton max iterations N0 5
newton nmon iterations N 1
halting accuracy R+ (1e-10,1e-9,1e-9,1e-8)
minimal accuracy R+0 (1e-8,1e-7,1e-7,1e-6)
extra condition f0; 1g 0
print residual info f0; 1g 0

Table 4: Point method structure: elds and possible values. When dierent, default values are given in the order
('stst','fold','hopf','psol').

eld content default value

phase condition f0; 1g 1
collocation parameters [0; 1]d or empty empty
adapt mesh before correct N 0
adapt mesh after correct N 3

Table 5: Point method structure: extra elds and possible values for the computation of periodic solutions.
For the computation of periodic solutions, additional elds are necessary, see table 5. The meaning
of the dierent elds in tables 4 and 5 is explained in section 10.
Similarly, for the approximation and correction of roots of the characteristic equation respectively for
the computation of the Floquet multipliers method parameters are passed using a structure of the form
given in table 6.
Branch structures A branch consists of an ordered array of points (all of the same type), and three
method structures containing point method parameters, continuation parameters respectively stability
computation parameters, see table 7.
The branch structure has three elds. One, called 'point', which contains an array of point structures,
one, called 'method', which is itself a structure containing three subelds and a third, called 'parameter'
which contains four subelds. The three subelds of the method eld are again structures. The rst,
called 'point', contains point method parameters as described in table 4. The second, called 'stability',
contains stability method parameters as described in table 6 and the third, called 'continuation' contains
continuation method parameters as described in table 8. Hence the branch structure incorporates all
necessary method parameters which are thus automatically kept when saving a branch variable to le. The

10
 eld content default value
lms parameter alpha Rk time lms('bdf',4)
lms parameter beta Rk time lms('bdf',4)
lms parameter rho R+0 time saf(alpha,beta,0.01,0.01)
interpolation order N0 4
minimal time step R+0+ 0.01
maximal time step R0 0.1
max number of eigenvalues N0 100
minimal real part R or empty empty
max newton iterations N 6
root accuracy R+ 0 1e-6
remove unconverged roots f0; 1g 1
eld content default value
collocation parameters [0; 1]d or empty empty
max number of eigenvalues N 100
minimal modulus R+ 0.01

Table 6: Stability method structures: elds and possible values for the approximation and correction of roots of
the characteristic equation (top), or for the approximation Floquet multipliers (bottom). The LMS-parameters
are default set to the fourth order backwards dierentiation LMS-method.
eld subeld content
point array of point struct
method point point method struct
method stability stability method struct
method continuation continuation method struct
parameter free N pf
parameter min bound [N R]pi
parameter max bound [N R]pa
parameter max step [N R]ps

Table 7: Branch structure: elds and possible values. Here, pf is the number of free parameters; pi , pa and ps are
the number of minimal parameter values, maximal parameter values respectively maximal parameter steplength
values. If any of these values are zero, the corresponding subeld is empty.

parameter eld contains a list of free parameter numbers which are allowed to vary during computations,
and a list of parameter bounds and maximal steplengths. Each row of the bound and steplength subelds
consists of a parameter number (rst element) and the value for the bound or steplength limitation.
Examples are given in section 7.
A default, empty branch structure can be obtained by passing a list of free parameters and the point
type (as 'stst', 'fold', 'hopf', or 'psol') to the function df brnch. A minimal bound zero is then set for each
delay. The method contains default parameters (containing appropriate point, stability and continuation
elds) obtained from the function df mthod with as only argument the type of solution point.
Scalar measure structure After a branch has been computed some possibilities are oered to plot
its content. For this a (scalar) measure structure is used which denes what information should be taken
and how it should be processed to obtain a measure of a given point (such as the amplitude of the prole
of a periodic solution, etc...), see table 9. The result applied to a variable point is to be interpreted as
scalar_measure=func(point.field.subfield(row,col));

11
 eld content default value
steplength condition f0; 1g 1
plot f0; 1g 1
prediction f1g 1
steplength growth factor R+
0 1.2
plot progress f0; 1g 1
plot measure struct or empty empty
halt before reject f0; 1g 0

Table 8: Continuation method structure: elds and possible values.

where 'eld' presents the eld to select, 'subeld' is empty or presents the subeld to select, 'row' presents
the row number or contains one of the functions mentioned in table 9. These functions are applied
columnwise over all rows. The function 'all' species that the all rows should be returned. The meaning
of 'col' is similar to 'row' but for columns. To avoid ambiguity it is required that either 'row' or 'col'
contains a number or that both contain the function 'all'. If nonempty, the function 'func' is applied to
the result. Note that 'func' can be a standard Matlab function as well as a user written function. Note
also that, when using the value 'all' in the elds 'col' and/or 'row' it is possible to return a non-scalar
measure (possibly but not necessarily further processed by 'func').
eld content meaning
eld f'parameter','x','v','omega',... rst eld to select
'prole','period','stability'g
subeld f",'l0','l1','mu'g " or second eld to select
row N or f'min','max','mean','ampl','all'g row index
col N or f'min','max','mean','ampl','all'g column index
func f",'real','imag','abs'g function to apply

Table 9: Measure structure: elds, content and meaning of a structure describing a measure of a point.

7 An illustrative ride-through
After the system has been implemented (see section 5), bifurcation analysis can be performed using the
point and branch manipulation layers. Specication of the functions in these layers is given in sections 8
respectively 9. Here we outline an illustrative ride-through using the example (5).
The commands below are listed in the le demo1.m. The gures shown are produced during its
execution. Some of these gures have important colour coding and others are gradually built up. Hence
the reader is advised to read this section while observing the gures from the Matlab run of demo1.
After starting Matlab in the directory of the system denition, we install the system by calling its
initialization le,
>> [name,n]=sys_init
name = neuron
n = 2

It is clear that (5) has a steady state solution (x1 ; x2 ) = (0; 0) for all values of the parameters. We
dene a rst steady state solution using the parameter values  = 0:5,  = 1, a12 = 1, a21 = 2:34,
1 = 2 = 0:2 and s = 1:5.
>> stst.kind='stst';
>> stst.parameter=[1/2 -1 1 2.34 0.2 0.2 1.5];

12
>> stst.x=[0 0]'
stst = kind: 'stst'
parameter: [0.5000 -1 1 2.3400 0.2000 0.2000 1.5000]
x: [2x1 double]

25
2
20
1.5
15
1
10

0.5 5

ℑ(λ)
ℑ(λ)

0
0
−5
−0.5
−10
−1
−15

−1.5 −20

−25
−2 −2 −1.5 −1 −0.5 0 0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4
ℜ(λ) ℜ(λ)

Figure 2: Approximated () and corrected () roots of the characteristic equation of system (5) at its steady
state solution (x1 ; x2 ) = (0; 0). Real parts computed up to <()  1 (left), <()  2 (right).
We get default point method parameters and correct the point,
>> method=df_mthod('stst')
method = continuation: [1x1 struct]
point: [1x1 struct]
stability: [1x1 struct]
>> [stst,success]=p_correc(stst,[],[],method.point)
stst = kind: 'stst'
parameter: [0.5000 -1 1 2.3400 0.2000 0.2000 1.5000]
x: [2x1 double]
success = 1
>> stst.x
ans = 0
0

which, being already a correct solution, remains unchanged. Computing and plotting stability of the
corrected point reveals it has one unstable real mode, see gure 2 (left).
>> stst.stability=p_stabil(stst,method.stability);
>> figure(1); clf;
>> p_splot(stst);

Seeing that only a few characteristic roots were computed we set minimal real part to a more negative
value (it is default empty which means that roots are computed up to <()  1= ) and recompute
stability to obtain gure 2 (right).
>> method.stability.minimal_real_part=-2;
>> stst.stability=p_stabil(stst,method.stability);
>> figure(2); clf;
>> p_splot(stst);

In both gures, approximations () and corrections () are nearly indistinguishable.
We will use this point as a rst point to compute a branch of steady state solutions. First, we obtain
an empty branch with free parameter a21 limited by a21 2 [0; 5] and
a21  0:2 between points.

13
>> branch1=df_brnch(4,'stst')
branch1 = method: [1x1 struct]
parameter: [1x1 struct]
point: []
>> branch1.parameter
ans = free: 4
min_bound: [3x2 double]
max_bound: []
max_step: []
>> branch1.parameter.min_bound
ans = 5 0
6 0
7 0
>> branch1.parameter.min_bound(4,:)=[4 0];
>> branch1.parameter.max_bound(1,:)=[4 5];
>> branch1.parameter.max_step(1,:)=[4 0.2];

To obtain a second starting point we change parameter value a21 slightly and correct again.
>> branch1.point(1)=stst;
>> stst.parameter(4)=stst.parameter(4)+0.1;
>> [stst,success]=p_correc(stst,[],[],method.point);
>> branch1.point(2)=stst;

Because we know how the branch of steady state solutions continued in a21 looks like (it is constant
at (x1 ; x2 ) = (0; 0)) we disable plotting during continuation by setting the corresponding continuation
method parameter to zero.
>> branch1.method.continuation.plot=0;

With two starting points and suitable method parameters we are ready to continue the branch in
parameter a21 (number 4), allowing it to vary in the interval [0; 5] using a maximum stepsize of 0.2 and
a maximum of 100 corrections.
>> [branch1,s,f,r]=br_contn(branch1,100)
BR_CONTN warning: boundary hit.
branch1 = method: [1x1 struct]
parameter: [1x1 struct]
point: [1x16 struct]
s = 15
f = 0
r = 0

During continuation, sixteen points were successfully computed (s = 16) before the right boundary
a21 = 5 was hit (signalled by a warning). No corrections failed (f = 0) and no computed points were
later rejected (r = 0). Reversing the order of the branch points allows to continue to the left.
>> branch1=br_rvers(branch1);
>> [branch1,s,f,r]=br_contn(branch1,100);
BR_CONTN warning: boundary hit.

We compute the stability along the branch.

>> branch1.method.stability.minimal_real_part=-2;
>> branch1=br_stabl(branch1,0,1);

After obtaining suitable measure structures we plot the real part of the approximated and corrected
roots of the characteristic equation along the branch, see gure 3 (left).

14
>> [xm,ym]=df_measr(1,branch1);
>> figure(3); clf;
>> br_plot(branch1,xm,ym,'b');
>> ym
ym = field: 'stability'
subfield: 'l1'
row: 'all'
col: 1
func: 'real'
>> ym.subfield='l0';
>> br_plot(branch1,xm,ym,'c');
>> plot([0 5],[0 0],'-.');
>> axis([0 5 -2 1.5]);

Again approximations and corrections are nearly indistinguishable. From this gure alone it is not
clear which real parts correspond to real roots respectively complex pairs of roots. For this it is useful
to compare gures 2 and 3 (left). Notice the strange behaviour (coinciding of several complex pairs of
roots) at a21 = 0. At this parameter value one of the couplings between the neurons is broken. In fact,
for a21 = 0, the evolution of the second component is independent of the evolution of the rst. Where
1.5
1.5

1
1

0.5
0.5

0
0
ℜ(λ)

ℜ(λ)

−0.5
−0.5

−1
−1

−1.5
−1.5

−2
0 1 2 3 4 5 −2
a21 0 5 10 15 20 25 30
point number

Figure 3: Real parts of the approximated (left) and corrected (left,right) roots of the characteristic equation
versus a21 (left) respectively the point number along the branch (right).
lines cross the zero line, bifurcations occur. If we want to compute the Hopf bifurcation near a21  0:8
we need its point number. This is most easily obtained by plotting the stability versus the point numbers
along the branch, see gure 3 (right).
>> figure(4); clf;
>> br_plot(branch1,[],ym,'b');
>> br_plot(branch1,[],ym,'b.');
>> plot([0 30],[0 0],'-.');

We select point 24 and turn it into an (approximate) Hopf bifurcation point.

>> hopf=p_tohopf(branch1.point(24));

We correct the Hopf point using appropriate method parameters and one free parameter (a21 ). We
then copy the corrected point to keep it for later use.
>> method=df_mthod('hopf');
>> [hopf,success]=p_correc(hopf,4,[],method.point)
hopf = kind: 'hopf'
parameter: [0.5000 -1 1 0.8071 0.2000 0.2000 1.5000]
x: [2x1 double]

15
 2

1.5

0.5

ℑ(λ)
0

−0.5

−1

−1.5

−2
−0.35 −0.3 −0.25 −0.2 −0.15 −0.1 −0.05 0 0.05
ℜ(λ)

Figure 4: Characteristic roots at Hopf point: a pair of pure imaginary eigenvalues is clearly visible.

v: [2x1 double]
omega: 0.7820
success = 1
>> first_hopf=hopf;

Computing and plotting stability of the Hopf point clearly reveals the pair of pure imaginary eigen-
values, see gure 4
>> hopf.stability=p_stabil(hopf,method.stability);
>> figure(5); clf;
>> p_splot(hopf);

In order to follow a branch of Hopf bifurcations in the two parameter space (a21 ; s ) we again need
two starting points. Hence we use the Hopf point already found and one perturbed in s and corrected
in a21 , to start on a branch of Hopf bifurcations. For the free parameters, a21 and s , we provide suitable
intervals, a21 2 [0; 4] and s 2 [0; 10], and maximal stepsizes, 0:2 for a21 and 0:5 for s .
>> branch2=df_brnch([4 7],'hopf');
>> branch2.parameter.min_bound(4:5,:)=[[4 0]' [7 0]']';
>> branch2.parameter.max_bound(1:2,:)=[[4 4]' [7 10]']';
>> branch2.parameter.max_step(1:2,:)=[[4 0.2]' [7 0.5]']';
>> branch2.point(1)=hopf;
>> hopf.parameter(7)=hopf.parameter(7)+0.1;
>> [hopf,success]=p_correc(hopf,4,[],method.point);
>> branch2.point(2)=hopf;

We continue the branch on both sides by an intermediate order reversal and a second call to br contn.
>> figure(6); clf;
>> [branch2,s,f,r]=br_contn(branch2,40);
BR_CONTN warning: boundary hit.
>> branch2=br_rvers(branch2);
>> [branch2,s,f,r]=br_contn(branch2,20);

As we did not change continuation method parameters, predictions and corrections will be plotted
during continuation. The nal result is shown in gure 5 (left). At the top, the branch hits the boundary
s = 10. To the right, however, it seemingly turned back onto itself. We compute and plot stability along
the branch.
>> branch2=br_stabl(branch2,0,0);

16
 11 11

9 9

7 7
s

s
τ

τ
5 5

3 3

1 1
−0.5 0 0.5 1 1.5 2 2.5 −0.5 0 0.5 1 1.5 2 2.5
a a
21 21

Figure 5: Predictions and corrections in the (a21 ; s )-plane after computation of a rst branch of Hopf bifurcations
(left) and a second, intersecting branch of Hopf bifurcations (right).

>> figure(7); clf;

>> [xm,ym]=df_measr(1,branch2);
>> ym.subfield='l0';
>> br_plot(branch2,[],ym,'c');
>> ym.subfield='l1';
>> br_plot(branch2,[],ym,'b');

If, during these computations we would have obtained warnings of the kind,
TIME_H warning: h_min is reached.

it would indicate that the time integration step required to obtain good approximations to the requested
rightmost characteristic roots is too small. By default, characteristic roots are computed up to <() 
1= .
We also notice a double Hopf point on the left but nothing special at the right end, which could explain
the observed turning of the branch. Plotting the frequency ! versus s reveals what has happened, see
gure 6 (right). For small s , ! goes through zero, indicating the presence of a Bogdanov-Takens point.
The subsequent turning is a recomputation of the same branch with negative frequencies.
>> figure(8); clf;
>> [xm,ym]=df_measr(0,branch2);
>> ym
ym = field: 'parameter'
subfield: ''
row: 1
col: 7
func: ''
>> ym.field='omega';
>> ym.col=1;
>> xm
xm = field: 'parameter'
subfield: ''
row: 1
col: 4
func: ''
>> xm.col=7;
>> br_plot(branch2,xm,ym,'c');
>> grid;

Selecting the double Hopf point we produce an approximation of the second Hopf point.

17
 1
0.2

0.75
0.1

0.5
0

0.25
−0.1
ℜ(λ)

ω
0
−0.2

−0.25
−0.3

−0.5
−0.4

−0.5 −0.75

−0.6 −1
1 2 3 4 5 6 7 8 9
0 10 20 30 40 50 τ
point number s

Figure 6: Left: Real part of characteristic roots along the branch of Hopf bifurcations shown in gure 5 (left).
Right: The frequency of the Hopf bifurcation along the same branch.

>> hopf=p_tohopf(branch2.point(4));
>> [hopf,success]=p_correc(hopf,4,[],method.point)
hopf = kind: 'hopf'
parameter: [0.5000 -1 1 -0.0103 0.2000 0.2000 8.5530]
x: [2x1 double]
v: [2x1 double]
omega: 0.9768
success = 0

However, without success. Printing residual information gives a list of the Newton iteration number
and the norm of the residual. This reveals at least temporarily divergence of the correction process.
>> method.point.print_residual_info=1;
>> format short e;
>> hopf=p_tohopf(branch2.point(4));
>> [hopf,success]=p_correc(hopf,4,[],method.point);
norm_residual = 1.0000e+00 9.3116e-03
norm_residual = 2.0000e+00 5.4574e-01
norm_residual = 3.0000e+00 6.2629e-02
norm_residual = 4.0000e+00 1.8903e-03
norm_residual = 5.0000e+00 3.2357e-05

Or we did not allow enough Newton iterations, or the free parameter is not so appropriate. We
successfully try again using s as a free parameter.
>> hopf=p_tohopf(branch2.point(4));
>> [hopf,success]=p_correc(hopf,7,[],method.point)
norm_residual = 1.0000e+00 9.3116e-03
norm_residual = 2.0000e+00 6.8069e-04
norm_residual = 3.0000e+00 2.3169e-07
norm_residual = 4.0000e+00 4.3066e-13
hopf = kind: 'hopf'
parameter: [5.0000e-01 -1 1 2.0657e-01 2.0000e-01 2.0000e-01 8.6340e+00]
x: [2x1 double]
v: [2x1 double]
omega: 9.1581e-01
success = 1

Using the second Hopf point we compute the intersecting branch of Hopf points depicted in gure 5
(right). Setting plot progress to zero disables intermediate plotting such that we see only the end result.

18
>> branch3=df_brnch([4 7],'hopf');
>> branch3.parameter=branch2.parameter;
>> branch3.point(1)=hopf;
>> hopf.parameter(4)=hopf.parameter(4)-0.05;
>> method.point.print_residual_info=0; format short;
>> [hopf,success]=p_correc(hopf,7,[],method.point);
>> branch3.point(2)=hopf;
>> branch3.method.continuation.plot_progress=0;
>> figure(6);
>> [branch3,s,f,r]=br_contn(branch3,100);
BR_CONTN warning: boundary hit.
>> branch3=br_rvers(branch3);
>> [branch3,s,f,r]=br_contn(branch3,100);
BR_CONTN warning: boundary hit.

We use the rst Hopf point we computed (rst hopf ) to construct a small amplitude (1e 2) periodic
solution on an equidistant mesh of 18 intervals with piecewise polynomial degree 3.
>> intervals=18;
>> degree=3;
>> [psol,stepcond]=p_topsol(first_hopf,1e-2,degree,intervals);

This steplength condition returned ensures the branch switch from the Hopf to the periodic solution
as it avoids convergence of the amplitude to zero during corrections. Due to the presence of the steplength
condition we also need to free one parameter, here a21 .
>> method=df_mthod('psol');
>> [psol,success]=p_correc(psol,4,stepcond,method.point)
psol = kind: 'psol'
parameter: [0.5000 -1 1 0.8072 0.2000 0.2000 1.5000]
mesh: [1x55 double]
degree: 3
profile: [2x55 double]
period: 8.0354
success = 1

The result, along with a degenerate periodic solution with amplitude zero is used to start on the
emanating branch of periodic solutions, see gure 7 (left). We avoid adaptive mesh selection and save
memory by clearing the mesh eld. An equidistant mesh is then automatically used which is kept xed
during continuation. Simple clearing of the mesh eld is only possible if it is already equidistant. This is
the case here as p tohopf returns a solution on an equidistant mesh.
>> branch4=df_brnch(4,'psol');
>> branch4.parameter.min_bound(4,:)=[4 0];
>> branch4.parameter.max_bound(1,:)=[4 5];
>> branch4.parameter.max_step(1,:)=[4 0.1];
>> deg_psol=p_topsol(first_hopf,0,degree,intervals);
>> deg_psol.mesh=[];
>> branch4.point(1)=deg_psol;
>> psol.mesh=[];
>> branch4.point(2)=psol;
>> figure(9); clf;
>> [branch4,s,f,r]=br_contn(branch4,50);

Notice how computing periodic solution branches takes considerably more computational time. Zoom-
ing shows erratic behaviour of the last computed branch points, shortly beyond a turning point, see gure
7 (right).
>> axis([2.3 2.4 0.95 1.15]);

19
 1.4 1.15

1.2

1.1
1
max x1(t) − min x1(t)

max x (t) − min x (t)

1
0.8

1.05
0.6

1
0.4
1

0.2

0 0.95
0.6 1 1.4 1.8 2.2 2.6 2.3 2.32 2.34 2.36 2.38 2.4
a a
21 21

Figure 7: Branch of periodic solutions emanating from a Hopf point (left). The branch turns at the far right and
a zoom (right) indicates computational diculties at the end.

Plotting some of the last solution proles shows that smoothness and thus also accuracy are lost, see
gure 9 (left).
>> ll=length(branch4.point);
>> figure(10); clf;
>> subplot(3,1,1);
>> p_pplot(branch4.point(ll-10));
>> subplot(3,1,2);
>> p_pplot(branch4.point(ll-5));
>> subplot(3,1,3);
>> p_pplot(branch4.point(ll-1));

From a plot of the period along the branch we could suspect a homoclinic or heteroclinic bifurcation
scenario, see gure 8 (left).
170 700

600
140

500

110
400
T
T

300
80

200

50
100

20 0
2.2 2.24 2.28 2.32 2.36 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
a a21
21

Figure 8: Left: Period along the computed branch shown in gure 7. Right: Added period predictions and
corrections during recalculations using adaptive mesh selection.

>> figure(11); clf;

>> [xm,ym]=df_measr(0,branch4);
>> ym
ym = field: 'profile'
subfield: ''
row: 1

20
 col: 'ampl'
func: ''
>> ym.field='period';
>> ym.col=1;
>> br_plot(branch4,xm,ym,'b');
>> axis([2.2 2.36 20 170]);

1 1
x (t), x (t)

0.5
2

x1(t), x2(t)
0
1

0
−1
0 0.2 0.4 0.6 0.8 1 −0.5
t/T
1
−1
x (t), x (t)

0 0.2 0.4 0.6 0.8 1

2

0 t/T
1

x1(t+0.02T), x2(t+0.02T)
−1
0 0.2 0.4 0.6 0.8 1 0.5
t/T
1
0
x (t), x (t)
2

0 −0.5
1

−1 −1
0 0.2 0.4 0.6 0.8 1 −1 −0.5 0 0.5 1
t/T x1(t), x2(t)

Figure 9: Some solution proles using equidistant meshes (left) and adapted meshes right (right) along the branch
of periodic solutions shown in gure 7.
The result of computing and plotting stability (Floquet multipliers) just before and after the turning
point is shown in gure 10. The second spectrum is clearly unstable but no accurate trivial Floquet
multiplier is present at 1.
>> psol=branch4.point(ll-11);
>> psol.stability=p_stabil(psol,method.stability);
>> figure(12); clf;
>> subplot(2,1,1);
>> p_splot(psol);
>> axis image;
>> psol=branch4.point(ll-8);
>> psol.stability=p_stabil(psol,method.stability);
>> subplot(2,1,2);
>> p_splot(psol);

First, we recompute a point on a rened, adapted mesh.

>> psol=branch4.point(ll-12);
>> intervals=40;
>> degree=4;
>> psol=p_remesh(psol,degree,intervals);
>> method.point.adapt_mesh_after_correct=1;
>> method.point.newton_max_iterations=7;
>> method.point.newton_nmon_iterations=2;
>> [psol,success]=p_correc(psol,[],[],method.point)
psol = kind: 'psol'
parameter: [0.5000 -1 1 2.3358 0.2000 0.2000 1.5000]
mesh: [1x161 double]
degree: 4
profile: [2x161 double]
period: 38.4916
success = 1

21
 1

0.5

ℑ(µ)
0

−0.5

−1
−1 −0.5 0 0.5 1
ℜ(µ)
1

0.5

ℑ(µ)
0

−0.5

−1
−10 0 10 20 30 40 50 60
ℜ(µ)

Figure 10: Floquet multipliers for a periodic solutions before (top) and just after (bottom) the turning point
visible in gure 7.

Then we recompute the branch using adaptive mesh selection (with reinterpolation and additional
corrections) after correcting every point, see gure 8 (right).
>> branch5=df_brnch(4,'psol');
>> branch5.parameter=branch4.parameter;
>> branch5.point(1)=psol;
>> psol.parameter(4)=psol.parameter(4)+0.01;
>> [psol,success]=p_correc(psol,[],[],method.point,1);
>> branch5.point(2)=psol;
>> branch5.method=method;
>> [xm,ym]=df_measr(0,branch5);
>> ym.field='period';
>> ym.col=1;
>> figure(11); axis auto; hold on;
>> branch5.method.continuation.plot_measure.x=xm;
>> branch5.method.continuation.plot_measure.y=ym;
>> branch5=br_contn(branch5,25);

Increasing mesh sizes and using adaptive mesh selection also improves the accuracy of the computed
Floquet multipliers.
>> psol=branch5.point(6);
>> psol.stability=p_stabil(psol,method.stability);
>> psol.stability.mu
ans = 241.2300
1.0000

Plotting of a point clearly shows the (double) homoclinic nature of the solutions, see gure 9 (right).
>> figure(13); clf;
>> subplot(2,1,1);
>> ll=length(branch5.point);
>> psol=branch5.point(ll-5);
>> plot(psol.mesh,psol.profile);
>> subplot(2,1,2);
>> psol1=p_remesh(psol,degree,0:0.001:1);
>> psol2=p_remesh(psol,degree,(0:0.001:1)+0.02);
>> plot(psol1.profile',psol2.profile');
>> psol.period
ans = 399.7466

22
 In this case, using the le df deriv.m instead of the analytical derivatives le given in section 5,
yields results which are visually the same as the ones given above.

8 Point manipulation
Several of the point manipulation routines have already been used in the previous section. Here we outline
their functionality and input and output parameters. A brief description of parameters is also contained
within the source code and can be obtained in Matlab using the help command. Note that a vector of
zero elements corresponds to an empty matrix (written in Matlab as []).
function [point,success]=p_correc(point0,free_par,step_cnd,method,adapt)

Function p correc corrects a given point.

 point0: initial, approximate solution point as a point structure (see table 2).
 free par: a vector of zero, one or more free parameters.
 step cnd: a vector of zero, one or more linear steplength conditions. Each steplength condition is
assumed fullled for the initial point and hence only the coecients of the condition with respect
to all unknowns are needed. These coecients are passed as a point structure (see table 2). This
means that, for, e.g., a steady state solution point p the i-th steplength condition reads
step cnd(i):parameter(p:parameter point0:parameter)T + step cnd(i):xT (p:x point0:x) = 0;
and similar formulas hold for the other solution types.
 method: a point method structure containing the method parameters (see table 4).
 adapt (optional): if zero or absent, do not use adaptive mesh selection (for periodic solutions); if
one, correct, use adaptive mesh selection and recorrect.
 point: the result of correcting point0 using the method parameters, steplength condition(s) and free
parameter(s) given. Stability information present in point0, is not passed onto point. If divergence
occured, point contains the nal iterate.
 success: nonzero if convergence was detected (that is, if the requested accuracy has been reached).
function stability=p_stabil(point,method)

Function p stabil computes stability of a given point by approximating its stability-determining eigen-
values.
 point: a solution point as a point structure (see table 2).
 method: a stability method structure (see table 6).
 stability: the computed stability of the point through a collection of approximated eigenvalues (as
a structure described in table 3). For steady state, fold and Hopf points both approximations and
corrections to the rightmost roots of the characteristic equation are provided. For periodic solutions
approximations to the dominant Floquet multipliers are computed.
function p_splot(point)

Function p splot plots the characteristic roots respectively Floquet multipliers of a given point (which
should contain nonempty stability information). Characteristic root approximations and Floquet multi-
pliers are plotted using '', corrected characteristic roots using ''.

23
function stst_point=p_tostst(point)
function fold_point=p_tofold(point)
function hopf_point=p_tohopf(point)
function [psol_point,stepcond]=p_topsol(point,ampl,degree,nr_int)
function [psol_point,stepcond]=p_topsol(point,ampl,coll_points)

The functions p tostst, p tofold, p tohopf and p topsol convert a given point into an approximation
of a new point of the kind indicated by their name. They are used to switch from a steady state point to
a Hopf point or fold point, from a Hopf point to a fold point or vice versa, from a (nearby double) Hopf
point to the second Hopf point, from a Hopf point to the emanating branch of periodic solutions and
from a periodic solution near a period doubling bifurcation to the period-doubled branch. When starting
a periodic solution branch from a Hopf point, an equidistant mesh is produced with nr int intervals and
piecewise polynomials of degree col degree and a steplength condition stepcond is returned which should
be used (together with a corresponding free parameter) in correcting the returned point. This steplength
condition (normally) prevents convergence back to the steady state solution (as a degenerate periodic
solution of amplitude zero). When jumping to a period-doubled branch, a period-doubled solution prole
is produced using collp oints for collocation points and a mesh which is the (scaled) concatenation of two
times the original mesh. A steplength condition is returned which (normally) prevents convergence back
to the single period branch.
function rm_point=p_remesh(point,new_degree,new_mesh)

Function p remesh changes the piecewise polynomial representation of a given periodic solution point.
 point: initial point, containing old mesh, old degree and old prole.
 new degree: new degree of piecewise polynomials.
 new mesh: mesh for new representation of periodic solution prole either as a (non-scalar) row
vector of mesh points (both interval and representation points, with the latter chosen equidistant
between the former, see section 6) or as the new number of intervals. In the latter case the new
mesh is adaptively chosen based on the old prole.
 rm point: returned point containing new degree, new mesh and an appropriately interpolated (but
uncorrected!) prole.

The following routines are used within branch routines but are less interesting for the general user.
function sc_measure=p_measur(p,measure)

Function p measur computes the (scalar) measure measure of the given point p (see table 9).
function p=axpy(a,x,y)

Function p axpy performs the axpy-operation on points. That is, it computes p = ax + y where a is a
scalar, and x and y are two point structures of the same type. p is the result of the operation on all
appropriate elds of the given points. If x and y are periodic solutions on dierent meshes, interpolation
is used and the result is obtained on the mesh of x. Stability information, if present, is not passed onto
p.
function n=p_norm(point)

Function p norm computes some norm of a given point structure.

function normalized_p=p_normlz(p)

Function p normlz performs some normalization on the given point structure p. In particular, fold and
Hopf determining eigenvectors are scaled to norm 1.

24
9 Branch manipulation
Usage of most of the branch manipulation routines has already been illustrated in section 7. Here we
outline their functionality and input and output variables. As for all routines in the package, a brief
description of the parameters is also contained within the source code and can be obtained in Matlab
using the help command.
function [c_branch,succ,fail,rjct]=br_contn(branch,max_tries)

The function br contn computes (or rather extends) a branch of solution points.
 branch: initial branch containing at least two points and computation, stability and continuation
method structured and a free parameter structure as described in table 7.
 max tries: maximum number of corrections allowed.
 c branch: the branch returned is a copy of the initial branch containing further the extra points
computed (starting from the end of the point array in the initial branch).
 succ: number of successfull corrections.
 fail: number of failed corrections.
 rjct: number of rejected points.
Note also that successfully computed points are normalized using the procedure p normlz (see section
8).
function br_plot(branch,x_measure,y_measure,line_type)

Function br plot plots a branch (in the current gure).

 branch: branch to plot (see table 7).
 x measure: (scalar) measure to produce plotting quantities for the x-axis (see table 9). If empty,
the point number is used to plot against.
 y measure: (scalar) measure to produce plotting quantities for the y-axis (see table 9). If empty,
the point number is used to plot against.
 line type (optional): line type to plot with.
function [x_measure,y_measure]=br_dfmsr(stability,branch)
function [x_measure,y_measure]=br_dfmsr(stability,par_list,kind)

Function br measur returns default measures for plotting.

 stability: nonzero if measures are required to plot stability information.
 branch: a given branch (see table 7) for which default measures should be constructed.
 par list: a list of parameters for which default measures should be constructed.
 kind: a point type for which default measures should be constructed.
 x measure: default scalar measure to use for the x-axis. x measure is chosen as the rst parameter
which varies along the branch or as the rst parameter of par list.

25
  y measure: default scalar measure to use for the y-axis. If stability is zero, the following choices
are made for y measure. For steady state solutions, the rst component which varies along the
branch; for fold and Hopf bifurcations the rst parameter value (dierent from the one used for
x measure) which varies along the branch. For periodic solutions, the amplitude of the st varying
component. If stability is nonzero, y measure selects the real part of the characteristic roots (for
steady state solutions, fold and Hopf bifurcations) or the modulus of the Floquet multipliers (for
periodic solutions).
function st_branch=br_stabl(branch,skip,recompute)

Function br stabl computes stability information along a previously computed branch.

 branch: given branch (see table 7).
 skip: number of points to skip between stability computations. That is, computations are performed
and stability eld is lled in every skip + 1-th point.
 recompute: if zero, do not recompute stability information present. If nonzero, discard and recom-
pute old stability information present (for points which were not skipped).
 st branch: a copy of the given branch whose (non-skipped) points contain a non-empty stability
eld with computed stability information (using the method parameters contained in branch).
function t_branch=br_rvers(branch)

To continue a branch in the other direction (from the beginning instead of from the end of its point
array), br rvers reverses the order of the points in the branches point array.
function recmp_branch=br_recmp(branch,point_numbers)

Function br recmp recomputes part of a branch.

 branch: initial branch (see table 7).
 point numbers (optional): vector of one or more point numbers which should be recomputed. Empty
or absent if the complete point array should be recomputed.
 recmp branch: a copy of the initial branch with points who were (successfully) recomputed replaced.
If a recomputation fails, a warning message is given and the old value remains present.
This routine can, e.g., be used after changing some method parameters within the branch method struc-
tures.
function [col,lengths]=br_measr(branch,measure)

Function br selec computes a measure along a branch.

 branch: given branch (see table 7).
 measure: given measure (see table 9).
 col: the collection of measures taken along the branch (over its point array) ordered row-wise. Thus,
a column vector is returned if measure is scalar. Otherwise, col contains a matrix.
 lengths: vector of lengths of the measures along the branch. If the measure is not scalar, it is
possible that its length varies along the branch (e.g. when plotting rightmost characteristic roots).
In this situation col is a matrix with number of columns equal to the maximal length of the measures
encountered. Extra elements of col are automically put to zero by Matlab. lengths can then be used
to prevent plotting of extra zeros.

26
10 Numerical methods
This section contains short descriptions of the numerical methods and the method parameters used in
DDE-BIFTOOL. More details on the methods can be found in the articles [25, 10, 9] or in [11].

10.1 Determining systems

Below we state the determining systems used to compute and continue steady state solutions, steady
state fold and Hopf bifurcations and periodic solutions of systems of delay dierential equations.
For each determining system we mention the number of free parameters necessary to obtain (gener-
ically) isolated solutions. In the package, the necessary number of free parameters is further raised by
the number of steplength conditions plus the number of extra conditions used. This choice ensures the
use of square Jacobians during Newton iteration. If, on the other hand, the number of free paramet-
ers, steplength conditions and extra conditions are not appropriately matched Newton iteration solves
systems with a non-square Jacobian (for which Matlab uses an over- or under-determined least squares
procedure). If possible, it is better to avoid such a situation.
Steady state solutions A steady state solution x 2 Rn is determined from the following n  n-
dimensional determining system with no free parameters.
f (x ; x ; : : : ; x ; ) = 0: (6)
Steady state fold bifurcations Fold bifurcations, (x 2 Rn ; v 2 Rn ) are determined from the follow-
ing 2n + 1  2n + 1-dimensional determining system using one free parameter.
8
< f (x ; x ; : : : ; x ; ) = 0
:

(x ; ; 0)v = 0 (7)
cT v 1 = 0
Here, cT v 1 = 0 presents a suitable normalization of v. The vector c 2 Rn is chosen as c =
v(0) =(v(0) T v(0) ), where v(0) is the initial value of v.
Steady state Hopf bifurcations Hopf bifurcations, (x 2 Rn ; v 2 C n ; ! 2 R) are determined from
the following 2n + 1  2n + 2-dimensional partially complex (and by this fact more properly called a
3n + 2  3n + 2-dimensional) determining system using one free parameter.
8
< f (x ; x ; : : : ; x ; ) = 0
:

(x ; ; i!)v = 0 (8)
cT v 1 = 0
Periodic solutions Periodic solutions are found as solutions (u(s); s 2 [0; 1]; T 2 R) of the following
n(Ld + 1) + 1  n(Ld + 1) + 1-dimensional system with no free parameters.
8
>
>
<
u_ (ci;j ) = Tf (u(ci;j ); u((ci;j T1 ) mod 1); : : : ; u((ci;j Tm ) mod 1); ) = 0;
i = 0; : : : ; L 1; j = 1; : : : ; d (9)
>
>
:
u(0) = u(1)
p(u) = 0
Here, p represents the integral phase condition
Z 1
u_ (s)
u(s)ds = 0; (10)
0

27
where u is the current solution and
u its correction. The collocation points are obtained as
ci;j = ti + cj (ti+1 ti ); i = 0; : : : ; L 1; j = 1; : : : ; d;
from the interval points ti , i = 0; : : : ; L 1 and the collocation parameters cj , j = 1; : : : ; d. The prole u
is discretized as a piecewise polynomial as explained in section 6. This representation has a discontinuous
derivative at the interval points. If ci;j coincides with ti the right derivative is taken in (9), if it coincides
with ti+1 the left derivative is taken. In other words the derivative taken at ci;j is that of u restricted to
[ti ; ti+1 ].
Point method parameters The point method parameters (see table 4) specify the following options.
 newton max iterations: maximum number of Newton iterations.
 newton nmon iterations: during a rst phase of newton nmon iterations + 1 Newton iterations the
norm of the residual is allowed to increase. After these iterations, corrections are halted upon
residual increase.
 halting accuracy: corrections are halted when the norm of the last computed residual is less than or
equal to halting accuracy is reached.
 minimal accuracy: a corrected point is accepted when the norm of the last computed residual is less
than or equal to minimal accuracy.
 extra condition: this parameter is nonzero when extra conditions are provided in a routine sys cond.m
which should border the determining systems during corrections. The routine accepts the current
point as input and produces an array of condition residuals and corresponding condition derivatives
(as an array of point structures) as illustrated below (x10.2).
 print residual info: when nonzero, the Newton iteration number and resulting norm of the residual
are printed to the screen during corrections.
For periodic solutions, the extra periodic solution parameters (see table 5) provide the following inform-
ation.
 phase condition: when nonzero the integral phase condition (10) is used.
 collocation parameters: this parameter contains user given collocation parameters. When empty,
Gauss-Legendre collocation points are chosen.
 adapt mesh before correct: before correction and if the mesh inside the point is nonempty, adapt
the mesh every adapt mesh before correct points. E.g.: if zero, do not adapt; if one, adapt every
point; if two adapt the points with odd point number.
 adapt mesh after correct: similar to adapt mesh before correct but adapt mesh after successfull cor-
rections and correct again.

10.2 Extra conditions

When correcting a point or computing a branch, it is possible to add one or more extra conditions
and corresponding free parameters to the determining systems presented earlier. These extra conditions
should be implemented using a le sys cond.m in the directory of the system denition and setting the
method parameter extra condition to 1 (cf. table 4). The function sys cond accepts the current point as
input and produces a residual and corresponding condition derivatives (as a point structure) per extra
condition.

28
 As an example, suppose we want to compute a branch of periodic solutions of system (5) subject to
the following extra conditions 
T = 200;
a212 + a221 = 1;
that is, we wish to continue a branch with xed period T = 200 and parameter dependence a212 + a221 = 1.
The following routine implements these conditions by evaluating and returning each residual for the given
point and the derivatives of the conditions w.r.t. all unknowns (that is, w.r.t. to all the components of
the point structure).
function [resi,condi]=sys_cond(point)

% kappa beta a12 a21 tau1 tau2 tau_s

if point.kind=='psol'
% fix period at 200:
resi(1)=point.period-200;
% derivative of first condition wrt unknowns:
condi(1)=p_axpy(0,point,[]);
condi(1).period=1;
% parameter condition:
resi(2)=point.parameter(3)^2+point.parameter(4)^2-1;
% derivative of second condition wrt unknowns:
condi(2)=p_axpy(0,point,[]);
condi(2).parameter(3)=2*point.parameter(3);
condi(2).parameter(4)=2*point.parameter(4);
else
error('SYS_COND: point is not psol.');
end;

return;

10.3 Continuation
During continuation, a branch is extended by a combination of predictions and corrections. A new point
is predicted based on previously computed points using secant prediction over an appropriate steplength.
The prediction is then corrected using the determining systems (6), (7), (8) or (9) bordered with a
steplength condition which requires orthogonality of the correction to the secant vector. Hence one extra
free parameter is necessary compared to the numbers mentioned in the previous section.
The following continuation and steplength determination strategy is used. If the last point was
successfully computed, the steplength is multiplied with a given, constant factor greater than 1. If
corrections diverged or if the corrected point was rejected because its accuracy was not acceptable, a new
point is predicted, using linear interpolation, halfway between the last two successfully computed branch
points. If the correction of this point succeeds, it is inserted in the point array of the branch (before the
previously last computed point). If the correction of the interpolated point fails again, the last successfully
computed branch point is rejected (for fear of branch switch) and the interpolation procedure is repeated
between the (new) last two branch points. Hence, if, after a failure, the interpolation procedure succeeds,
the steplength is approximately divided by a factor two. Test results indicate that this procedure is quite
eective and proves an ecient alternative to using only (secant) extrapolation with steplength control.
The reason for this is mainly that the secant extrapolation direction is not in uenced by halving the
steplength but it is by inserting a newly computed point in between the last two computed points.
The continuation method parameters (see table 8) have the following meaning.
 plot: if nonzero, plot predictions and corrections during continuation.

29
  prediction: this parameter should be 1, indicating that secant prediction is used (being currently
the only alternative).
 steplength growth factor: grow the steplength with this factor in every step except during interpol-
ation.
 plot progress: if nonzero, plotting is visible during continuation process. If zero, only the nal result
is drawn.
 plot measure: if empty use default measures to plot. Otherwise plot measure contains two elds, 'x'
and 'y', which contain measures (see table 9) for use in plotting during continuation.
 halt before reject: If this parameter is nonzero, continuation is halted whenever (and instead of)
rejecting a previously accepted point based on the above strategy.

10.4 Roots of the characteristic equation

Roots of the characteristic equation are approximated using a linear multi-step (LMS-) method applied
to (2).
Consider the linear k-step formula
k
X k
X
j yL+j = h j fL+j : (11)
j =0 j =0
Here, 0 = 1, h is a (xed) step size and yj presents the numerical approximation of y(t) at the mesh point
tj := jh. The right hand side fj := f (yj ; y~(tj 1 ); : : : ; y~(tj m )) is computed using approximations
y~(tj 1 ) obtained from yi in the past, i < j . In particular, the use of so-called Nordsieck interpolation,
leads to s
X
y~(tj + h) = Pl ()yj+l ;  2 [0; 1): (12)
l= r
using
s
 k:
Y
Pl () :=
k= r; k6=l l k
The resulting method is explicit whenever 0 = 0 and min i > sh. That is, yL+k can then directly
be computed from (11) by evaluating
k 1
X k
X
yL+k = j yL+j + h j fL+j :
j =0 j =0
whose right hand side depends only on yj , j < L + k.
For the linear variational equation (2) around a steady state solution x (t)  x we have
m
X
fj = A0 yj + Ai y~(tj i ) (13)
i=0
where we have omitted the dependency of Ai on x . The stability of the dierence scheme (11), (13) can
be evaluated by setting yj = j Lmin , j = Lmin; : : : ; L + k where Lmin is the smallest index used, taking
the determinant of (11) and computing the roots . If the roots of the polynomial in  all have modulus
smaller than 1, the trajectories of the LMS-method converge to zero. If roots exist with modulus greater
then trajectories exist which grow unbounded.

30
 Since the LMS-method forms an approximation of the time integration operator over the time step
h, so do the roots  approximate the eigenvalues of S (h; 0). The eigenvalues of S (h; 0) are exponential
transforms of the roots  of the characteristic equation (4),
 = exp(h):
Hence, once  is found,  can be extracted using,
<() = ln(hjj) : (14)
The imaginary part of  is found modulo =h, using
arcsin( =j(j) )
=()  h (mod h ): (15)
For small h, 0 < h  1, the smallest representation in (15) is assumed the most accurate one (that is,
we let arcsin map into [ =2; =2]).
The parameters r and s (from formula (12)) are chosen such that r  s  r + 2 (see [17]). The choice
of h is based on the related heuristic outlined in [11].
Approximations for the rightmost roots  obtained from the LMS-method using (14), (15) can be
corrected using a Newton process on the system,


()v = 0 (16)
cT v 1 = 0
A starting value for v is the eigenvector of
() corresponding to its smallest eigenvalue (in modulus).
Note that the collection of successfully corrected roots presents more accurate yet less robust inform-
ation than the set of uncorrected roots. Indeed, attraction domains of roots of equations like (16) can be
very small and hence corrections may diverge or approximations of dierent roots may be corrected to a
single 'exact' root thereby missing part of the spectrum. The latter does not occur when computing the
(full) spectrum of a discretization of S (h; 0).
Stability information is kept in the structure of table 3 (left). The time step used is kept in eld
h. Approximate roots are kept in eld l0, corrected roots in eld l1. If unconverged corrected roots are
discarded, eld n1 is empty. Otherwise, the number of Newton iterations used is kept for each root in
the corresponding position of n1. Here, 1 signals that convergence to the required accuracy was not
reached. The stability method parameters (see table 6 (top)) now have the following meaning.
 lms parameter alpha: LMS-method parameters j ordered from past to present, j = 0; 1; : : :; k.
 lms parameter beta: LMS-method parameters j ordered from past to present, j = 0; 1; : : : ; k.
 lms parameter rho: safety radius LMS; of the LMS-method stability region. For a precise denition,
see [11, xIII.3.2].
 interpolation order: order of the interpolation in the past, r + s = interpolation order.
 minimal time step: minimal time step relative to maximal delay, h  minimal time step.
 maximal time step: maximal time step relative to maximal delay, h  minimal time step.
 max number of eigenvalues: maximum number of rightmost eigenvalues to keep.
 minimal real part: choose h such as to approximate eigenvalues with <()  minimal real part well,
discard eigenvalues with <() < minimal real part. If h is smaller than its minimal value, it is set
to the minimal value and a warning is uttered. If it is larger than its maximal value it is reduced
to that number without warning. If minimal and maximal value coincide, h is set to this value
without warning. If minimal real part is empty, the value minimal real part = 1 is used.

31
  max newton iterations: maximum number of Newton iterations during the correction process (16).
 root accuracy: required accuracy of the norm of the residual of (16) during corrections.
 remove unconverged roots: if this parameter is zero, unconverged roots are discarded (and stability
eld n1 is empty).

10.5 Floquet multipliers

Floquet multipliers are computed as eigenvalues of the discretized time integration operator S (T; 0).
The discretization is obtained using the collocation equations (9) without the modulo operation (and
without phase and periodicity condition). From this system a discrete, linear map is obtained between
the variables presenting the segment [ =T; 0] and those presenting the segment [ =T + 1; 1]. If these
variables overlap, part of the map is just a time shift.
Stability information is kept in the structure of table 3 (right). Approximations to the Floquet
multipliers are kept in eld mu. The stability method parameters (see table 6 (bottom)) have the
following meaning.
 collocation parameters: user given collocation parameters or empty for Gauss-Legendre collocation
points.
 max number of eigenvalues: maximum number of multipliers to keep.
 minimal modulus: discard multipliers with jj < minimal modulus.

11 Concluding comments
The rst aim of DDE-BIFTOOL is to provide a portable, user-friendly tool for numerical bifurcation
analysis of steady state solutions and periodic solutions of systems of delay equations of the kind (1).
Part of this goal was fullled through choosing the portable, programmer-friendly environment oered
by Matlab. Robustness with respect to the numerical approximation is achieved through automatic
steplength selection in approximating the rightmost characteristic roots and through collocation using
piecewise polynomials combined with adaptive mesh selection.
Although the package has been successfully tested on a number of realistic examples, a word of caution
may be appropriate. First of all, the package is essentially a research code (hence we accept no reliability)
in a quite unexplored area of current research. In our experience up to now, new examples did not fail
to produce interesting theoretical questions (e.g., concerning homoclinic or heteroclinic solutions) many
of which remain unsolved today. Unlike for ordinary dierential equations, discretization of the state
space is unavoidable during computations on delay equations. Hence the user of the package is strongly
advised to investigate the eect of discretization using tests on dierent meshes and with dierent method
parameters; and, if possible, to compare with analytical results and/or results obtained using simulation.
Although there are no 'hard' limits programmed in the package (with respect to system and/or mesh
sizes), the user will notice the rapidly increasing computation time for increasing system dimension and
mesh sizes. This exhibits itself most profoundly in the stability and periodic solution computations. In-
deed, eigenvalues are computed from large sparse matrices without exploiting sparseness and the Newton
procedure for periodic solutions is implemented using direct methods. Nevertheless the current version
is sucient to perform bifurcation analysis of systems with reasonable properties in reasonable execution
times. Furthermore we hope future versions will include routines which scale better with the size of the
problem.
Other future plans include a graphical user interface and the extension to other types of delay equations
such as state-dependent, distributed delay and neutral functional dierential equations.

32
Acknowledgements
This report presents results of the research project OT/98/16, funded by the Research Council K.U.Leuven;
of the research project G.0270.00 funded by the Fund for Scientic Research - Flanders (Belgium) and
of the research project IUAP P4/02 funded by the programme on Interuniversity Poles of Attraction,
initiated by the Belgian State, Prime Minister's Oce for Science, Technology and Culture. The author,
K. Engelborghs, is a research assistant of the Fund for Scientic Research - Flanders (Belgium).
Thanks to T. Luzyanina and B. Haegeman for being the rst users and testers of DDE-BIFTOOL.
Thanks to D. Roose for being my promoter during the research that culminated into this package.

References
[1] J. Argyris, G. Faust, and M. Haase. An Exploration of Chaos | An Introduction for Natural
Scientists and Engineers. North Holland Amsterdam, 1994.
[2] N. V. Azbelev, V. P. Maksimov, and L. F. Rakhmatullina. Introduction to the Theory of Functional
Dierential Equations. Nauka, Moscow, 1991. (in Russian).
[3] A. Back, J. Guckenheimer, M. Myers, F. Wicklin, and P. Worfolk. DsTool: Computer assisted
exploration of dynamical systems. AMS Notices, 39:303{309, 1992.
[4] R. Bellman and K. L. Cooke. Dierential-Dierence Equations, volume 6 of Mathematics in science
and engineering. Academic Press, 1963.
[5] O. Diekmann, S. A. van Gils, S. M. Verduyn Lunel, and H.-O. Walther. Delay Equations: Functional-
, Complex-, and Nonlinear Analysis, volume 110 of Applied Mathematical Sciences. Springer-Verlag,
1995.
[6] E. J. Doedel, A. R. Champneys, T. F. Fairgrieve, Y. A. Kuznetsov, B. Sandstede, and X. Wang.
AUTO97: Continuation and bifurcation software for ordinary dierential equations; available by
FTP from ftp.cs.concordia.ca in directory pub/doedel/auto.
[7] R. D. Driver. Ordinary and Delay Dierential Equations, volume 20 of Applied Mathematical Science.
Springer-Verlag, 1977.
[8] L. E. El'sgol'ts and S. B. Norkin. Introduction to the Theory and Application of Dierential Equations
with Deviating Arguments, volume 105 of Mathematics in science and engineering. Academic Press,
1973.
[9] K. Engelborghs, T. Luzyanina, K. J. in t Hout, and D. Roose. Collocation methods for the compu-
tation of periodic solutions of delay dierential equations. SIAM J. Sci. Comput., 2000. Accepted.
[10] K. Engelborghs and D. Roose. Numerical computation of stability and detection of Hopf bifurcations
of steady state solutions of delay dierential equations. Advances in Computational Mathematics,
10(3{4):271{289, 1999.
[11] Koen Engelborghs. Numerical Bifurcation Analysis of Delay Dierential Equations. PhD thesis,
Department of Computer Science, Katholieke Universiteit Leuven, Leuven, Belgium, May 2000.
[12] W. H. Enright and H. Hayashi. A delay dierential equation solver based on a continuous Runge-
Kutta method with defect control. Numer. Algorithms, 16:349{364, 1997.
[13] B. Ermentrout. XPPAUT3.91 - The dierential equations tool. University of Pittsburgh, Pittsburgh,
(http://www.pitt.edu/phase/) 1998.
[14] J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical Systems and Bifurcations of
Vector Fields. Springer-Verlag New York, 1983.

33
[15] J. K. Hale. Theory of Functional Dierential Equations, volume 3 of Applied Mathematical Sciences.
Springer-Verlag, 1977.
[16] J. K. Hale and S. M. Verduyn Lunel. Introduction to Functional Dierential Equations, volume 99
of Applied Mathematical Sciences. Springer-Verlag, 1993.
[17] T. Hong-Jiong and K. Jiao-Xun. The numerical stability of linear multistep methods for delay
dierential equations with many delays. SIAM Journal of Numerical Analysis, 33(3):883{889, June
1996.
[18] The MathWorks Inc. MATLAB 5.3. 1999.
[19] A. I. Khibnik, Yu. A. Kuznetsov, V. V. Levitin, and E. N. Nikolaev. LOCBIF: Interactive Local
Bifurcation Analyzer, version 2.2. Institute of Mathematical Problems in Biology, Russian Academy
of Sciences, Pushchino, 1992.
[20] V. Kolmanovskii and A. Myshkis. Applied Theory of Functional Dierential Equations, volume 85
of Mathematics and Its Applications. Kluwer Academic Publishers, 1992.
[21] V. B. Kolmanovskii and A. Myshkis. Introduction to the theory and application of functional dif-
ferential equations, volume 463 of Mathematics and its applications. Kluwer Academic Publishers,
1999.
[22] V. B. Kolmanovskii and V. R. Nosov. Stability of functional dierential equations, volume 180 of
Mathematics in Science and Engineering. Academic Press, 1986.
[23] Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory, volume 112 of Applied Mathematical
Sciences. Springer-Verlag, 1995.
[24] Yu. A. Kuznetsov and V. V. Levitin. CONTENT: A multiplatform environment for analyzing dy-
namical systems. Dynamical Systems Laboratory, Centrum voor Wiskunde en Informatica; available
via ftp.cwi.nl in directory pub/CONTENT.
[25] T. Luzyanina and D. Roose. Numerical stability analysis and computation of Hopf bifurcation points
for delay dierential equations. Journal of Computational and Applied Mathematics, 72:379{392,
1996.
[26] C. A. H. Paul. A user-guide to Archi - an explicit Runge-Kutta code for solving delay and neutral
dierential equations. Technical Report 283, The University of Manchester, Manchester Center for
Computational Mathematics, December 1995.
[27] J. Hale S.-N. Chow. Methods of Bifurcation Theory. Springer-Verlag, 1982.
[28] Rudiger Seydel. Practical Bifurcation and Stability Analysis | From Equilibrium to Chaos, volume 5
of Interdisciplinary Applied Mathematics. Springer-Verlag Berlin, 2 edition, 1994.
[29] L. F. Shampine and S. Thompson. Solving delay dierential equations with dde23. Submitted, 2000.
[30] L. P. Shayer and S. A. Campbell. Stability, bifurcation and multistability in a system of two coupled
neurons with multiple time delays. SIAM J. Applied Mathematics, 1999. To appear.
[31] S. Thompson, S. P. Corwin, and D. Sarafyan. DKLAG6: A code based on continuously imbedded
sixth order Runge-Kutta methods for the solution of state dependent functional dierential equations.
Applied Numerical Mathematics, 24(2{3):319{330, 1997.

34
Appendix A: List of les
Version 1.00 of DDE-BIFTOOL contains the following les.
Layer 0 Layer 1 Layer 2 Layer 3 Extra
sys cond auto cnt p axpy br contn df brnch
sys deri auto eqd p correc br measr df deriv
sys init auto msh p measur br plot df measr
sys rhs auto ord p norm br recmp df mthod
fold jac p normlz br rvers demo1
hopf jac p pplot br stabl
mult app p remesh
mult dbl p secant
mult int p splot
mult plt p stabil
poly del p tofold
poly dla p tohopf
poly elg p topsol
poly gau p tostst
poly lgr
poly lob
psol eva
psol jac
psol msh
root app
root cha
root int
root nwt
root plt
stst jac
time h
time lms
time nrd
time saf

35
Appendix B: Obtaining the package
DDE-BIFTOOL is freely available for scientic (non-commercial) use. It was written by the author as
a part of his PhD at the Computer Science Department of the K.U.Leuven under supervision of Prof. D.
Roose.
The following terms cover the use of the software package DDE-BIFTOOL:
1. The package DDE-BIFTOOL can be used only for the purpose of internal research excluding any
commercial use of the package DDE-BIFTOOL as such or as a part of a software product.
2. K.U.Leuven, Department of Computer Science shall for all purposes be considered the
owner of DDE-BIFTOOL and of all copyright, trade secret, patent or other intellectual property
rights therein.
3. The package DDE-BIFTOOL is provided on an "as is" basis and for the purposes described in
paragraph 1 only. In no circumstances can K.U.Leuven be held liable for any deciency, fault or
other mishappening with regard to the use or performance of the package DDE-BIFTOOL.
4. All scientic publications, for which the package DDE-BIFTOOL has been used, shall mention
usage of the package DDE-BIFTOOL, and shall refer to the following publication:
K. Engelborghs. DDE-BIFTOOL: a Matlab package for bifurcation analysis of delay
dierential equations. Technical Report TW-305, Department of Computer Science,
K.U.Leuven, Leuven, Belgium, 2000.
Upon acceptance of the above terms, one can obtain the package DDE-BIFTOOL (version 1.00) by
mailing your full name, aliation and address to [email protected]. The package
will then be forwarded to you.

36