Order Reduction of Nonlinear Delay-Differential Equations with Periodic Coefficients

Eric A. Butcher
Department of Mechanical Engineering
New Mexico State University
Las Cruces, NM 88003, U.S.A.
Ph. 1-575-646-6179; Fax. 1-575-646-6111; Email: [email protected]

Abstract

A technique for order reduction of nonlinear delay differential equations with time-
periodic coefficients is presented. The DDEs considered here have at most cubic nonlinearities
multiplied by a perturbation parameter. The periodic terms and matrices are not assumed to have
predetermined norm bounds, thus making the method applicable to systems with strong
parametric excitation. Perturbation expansion converts the nonlinear response problem into
solutions of a series of non-homogenous linear ordinary differential equations with time periodic
coefficients. One set of linear non-homogenous ODEs is solved for each power of the
perturbation parameter. Each ODE is solved by a Chebyshev spectral collocation method. Thus
we compute a finite approximation to the nonlinear infinite-dimensional map for the DDE. The
accuracy of the method is demonstrated with a nonlinear delayed Mathieu equation, a milling
model, and a single inverted pendulum with a periodic retarded follower force and nonlinear
restoring force in which the amplitude of the limit cycle associated with a flip bifurcation is
found analytically and compared to that obtained from direct numerical simulation.

1. Introduction

Dimensional reduction of nonlinear DDEs has been considered by researchers in the past
using different approaches. A center manifold algorithm for constant coefficient DDEs near
Hopf bifurcation points was first formulated in [1]. Unlike the case for center manifold
reduction of ODEs [2], the algorithm for DDEs is necessarily stated in the language of functional
analysis and requires the description of the adjoint system. The method was first applied to a
practical system (Hopf bifurcation in machine tool vibrations) much later [3,4]. Recently,
alternative techniques for dimensional reduction of nonlinear DDEs have been proposed. These
include the use of stiff and soft substructures [5], the method of multiple scales [6], and a
Galerkin projection technique which reduces a DDE to a small number of ODEs [7]. It should be
noted that all of these techniques are designed for autonomous nonlinear DDEs.
The study of nonlinear time-periodic DDEs has included the computation of the response
[8] and the use of normal forms and continuation algorithms. Much of the practical work in the
engineering community has focused on the milling problem which, unlike the case of turning, is
modeled by DDEs with time-periodic coefficients. It has been observed in numerous studies
that, in addition to the secondary Hopf or Neimark-Sacker bifurcation, such a system can also
exhibit flip or period-doubling bifurcations. The center manifold algorithms in [1,3,4] cannot be
directly applied to the time-periodic case without additional complexities associated with
redefining the adjoint equation. Nevertheless a center manifold computation for the case of flip
bifurcation in single DOF low immersion milling was carried out in [9] by assuming that the
system can be accurately modeled using a 2-dimensional map of the cutting tool’s position and
velocity from one instantaneous cut to the next. For such a system, the coefficients are
essentially periodically-spaced Dirac delta functions. The problem of milling with a finite
cutting time, on the other hand, requires a calculation using functional analytic tools to
decompose the dynamics on the center manifold [10]. This algorithm is based on previous efforts
in the mathematical literature to generalize the adjoint equation and formalize the center
manifold theory for the case of periodic DDEs. The motivation for the present paper is in a
possible alternative method which does not involve the explicit computation of the adjoint.
The method here for dimensional or center manifold reduction of periodic DDEs makes
use of a nonlinear extension of the infinite dimensional monodromy operator defined in [11] for
linear periodic DDEs. The nonlinear operator maps the initial function onto each subsequent
delay interval. Iteration of the nonlinear operator produces the solution via the ‘method of
steps’. The linear part of this operator has been previously approximated by a finite dimensional
matrix using a variety of numerical techniques, including orthogonal polynomial expansion [12]
and collocation [13] for the purpose of stability analysis. In this paper we use the method of
Chebyshev spectral collocation which is discussed at length by Trefethen [15] and was applied to
study the stability of linear periodic DDEs in [13-14]. It is assumed that the periodically
modulated nonlinearities are multiplied by a perturbation parameter. Using a classical
perturbation technique, the solution in any interval is expanded using the solution from the
previous interval in the in the zeroth order part. The expansion proceeds in a usual way by
powers of perturbation parameter. Collecting terms for like powers of the perturbation parameter
leads to a series of linear non-homogenous ordinary differential equations which are then solved
using Chebyshev spectral collocation. Quadratic and cubic nonlinearities result in similar
nonlinearities in the resulting approximate nonlinear DDE solution map. Increased accuracy is
obtained by increasing the number of collocation points and by adding more terms to the
perturbation expansion. Next, center manifold reduction is applied to the nonlinear map. An
algorithm that performs center manifold reduction on large nonlinear maps in MATLAB© is
developed for the purpose of making the problem computationally tractable.

2. Approximation of Functional Mapping

Consider a time periodic nonlinear DDE given by

x
= A(t ) x + B (t ) x(t − τ ) + εf ( x(t ), x(t − τ ), t )

(1)
x( s) = φ ( s) ∀ s ∈[−τ ,0]

where x(t ) is an n -dimensional state vector and A(t ) = A(t + T ) , B (t ) = B (t + T ) are n × n

coefficient matrices with principal period T . Here τ is a fixed delay and ε is the perturbation
parameter. The nonlinear term f (α , β , t ) = f ( x(t ), x(t − τ ), t ) is a vector-valued function, with n
components, which is periodic with period T in its third input. For the purpose of simplicity, it is
assumed that the delay τ is the same as the principal period T . (Our method, however,
generalizes to τ ≠ T case.) We assume that f (α , β , t ) is a polynomial of degree 3 in α and β .
For the linear system obtained by setting ε = 0 in (1), one may analyze stability by
considering the infinite dimensional monodromy operator which maps the solution x ( p−1) (t ) in
[( p − 1)τ , pτ ] onto the solution x ( p ) (t ) in [ pτ , ( p + 1)τ ] . The eigenvalues of the monodromy
operator, the Floquet multipliers, determine the stability of the linear system. An accurate finite
dimensional approximation of the monodromy operator, that is, a “monodromy matrix”, can be
computed using Chebyshev spectral collocation [13-15] and other techniques. For the nonlinear
DDE, we will approximate the corresponding nonlinear map x ( p −1) (t )  x ( p ) (t ) by a
perturbation series in ε . Note that (1) typically has parameters but these are not explicitly
included in the analysis for now. Obtaining the nonlinear map in a parameter dependent manner
would allow for bifurcation analysis at the critical points. Later we re-introduce the bifurcation
parameter via versal deformation theory as in [16] for the ODE case. Suppose the solution x(t )
in the interval [0,τ ] = [0, T ] for (1) is written in the form
x(t ) = φ (t ) + ξ 0 (t ) + εξ1 (t ) + ε 2ξ 2 (t ) + ε 3ξ 3 (t ) + .... (2)

Here φ (t ) is a vector of known functions defined over [0,τ ] by translation from [−τ ,0] . The
functions ξ i (t ), i = 0,1,2,.. are unknown n -dimensional vectors for each power of ε . The initial
conditions for these unknown vectors are
ξ 0 (0) = φ (τ ) − φ (0), ξ1 (0) = 0, ξ 2 (0) = 0, ξ 3 (0) = 0,... so that x(0)=φ(τ). Substituting (2) in
(1) and retaining terms up to third order, a sequence of linear time-periodic ODEs is obtained
after collecting like powers of ε .
We now consider a sequence of infinite-dimensional maps

ε 0 : mξ 0 = Vmφ
ε 1 : mξ 1 = P21 (mφ ,Vmφ ) + P31 (mφ ,Vmφ ) = P1 (mφ )
ε 2 : mξ 2 = P22 (mφ ,Vmφ , P1 (mφ )) + P32 (mφ ,Vmφ , P1 (mφ )) = P 2 (mφ ) (3)
ε 3 : mξ 3 = P23 (mφ ,Vmφ , P1 (mφ ), P 2 (mφ )) + P33 (mφ ,Vmφ , P1 (mφ ), P 2 (mφ ))
= P 3 (mφ )
which solve the time-periodic ODEs thus produced from the perturbation expansion. Here Pi j
denotes a polynomial of homogeneous degree i in its j+1 coefficients. Equation (3) is in an
abstract function space until one chooses a particular finite representation for the infinite-
dimensional vector m w corresponding to an arbitrary function w(t ) . Using equations (2-3), we
express the solution x(t ) = x ( p ) (t ) in an interval [ pτ , ( p + 1)τ ] as a nonlinear transformation of
the solution in the previous interval φ (t ) = x ( p −1) (t ) in [( p − 1)τ , pτ ] as

m x ( p) = ( I + V )mφ + εP1 (mφ ) + ε 2 P 2 (mφ ) + ε 3 P 3 (mφ ) (4)

While we use the method of Chebyshev collocation to approximate the map in (4), in this paper
we do not discuss the procedure in detail, but instead refer the reader to [13-15].

3. Order Reduction

The nonlinear map in equation (4) can be written for a particular value of ε as

mx (k + 1) = Umx (k ) + CQ [mx (k )]2 + CC [mx (k )]3 (5)

where mx (k + 1) is the collocation vector in a particular interval, mx (k ) is the collocation vector
in the preceding interval, [mx (k )]i is the vector of all the possible independent nonlinear terms of
the order i = 2,3 of the collocation vector mx (k ) and, C Q and C C are the quadratic and cubic
coefficient matrices. The order of nonlinearities in [mx (k )]i is a lexicographic order. The
coefficient matrix C Q will have appropriate entries for each of the nonlinearities appearing in
the map. In equation (5), only up to cubic powers of nonlinearities are retained, even though the
powers in equation (4) are of much higher order. If n is the dimension of the state space in
equation (1) and N is the number of Chebyshev points used in collocation, then there are
nN nN

∑ i = o(nN 2 ) independent quadratic and

i =1
∑ i(i + 1) / 2 = o(nN
i =1
3
) independent cubic terms.

Equation (5) can be written in the modal coordinates by means of a modal transformation
such that the states can be partitioned according to the eigenvalues of the diagonal JU into
master and slave states as
mzc (k + 1)  µ c 0  mzc (k ) [CQ ]  [CCM ]c 
M c
2 3
=
 m (k + 1)   s 
+ 
 [C M ]s [ m zc , m zs ] +  M s [mzc , mzs ] (6)
 zs   0 µ   zs   Q 
m ( k ) [C
 C  ]

A modified method of center manifold reduction for nonlinear maps [30] is applied to equation
(6). A nonlinear transformation is defined of the form

mzs (k ) = U 2 [mzc (k )]2 + U 3[mzc (k )]3 (7)

where U 2 and U 3 are undetermined coefficient matrices of quadratic and cubic terms.
Substituting equation (7) into equation (6), a nonlinear algebraic equation of the form

U 2 ( µ c m zc + [C QM ]c [m zc , m zs ] 2 + [C CM ]c [m zc , m zs ]3 ) 2 +
U 3 ( µ s m zs + [C QM ]c [m zc , m zs ] 2 + [C CM ]c [m zc , m zs ]3 ) 3 (8)
= µ s m zs + [C QM ] s [m zc , m zs ]2 + [C CM ] s [m zc , m zs ]3

is obtained with the independent variable k suppressed for brevity. Only quadratic and cubic
terms are retained in equation (8) and then coefficients of the like powers are collected for
determining U 2 and U 3 . The resulting coefficient equations for quadratic and cubic terms are

[m zc ] 2 : U 2 µ c = µ sU 2 + s ([C QM ] s ) (9)

with

1. s ([C QM ] s ) = surviving coefficients from matrix [C QM ] s due to order 2 truncation

 2. µ c = a coefficient matrix resulting from squaring each element of vector µ c m zc

and

[m zc ]3 : U 3 µˆ s + U 2 cs ( µ c m zc × [C QM ]c [m zc , m zs ]2 ) = µ sU 3 + s ([C CM ] s )
(10)
+ cs ([C QM ]c [m zc , m zs ]2 )

with

3. s ([C QM ] s ) = surviving coefficients of matrix [C QM ] s due to order 3 truncation

4. cs ( µ c m zc × [C QM ]c [m zc , m zs ] 2 ) = surviving coefficients of the coefficient matrix

resulting from the product in the parentheses, due to order 3 truncation

5. cs ([C QM ]c [m zc , m zs ] 2 ) = surviving coefficients of the coefficient matrix resulting from

the term in the parentheses due to order 3 truncation

6. µ̂ s = a coefficient matrix resulting from cubing each element of vector µ s m zc

Equations (9-10) are linear generalized Lyapunov equations which are solved for unknown
coefficient matrices U 2 and U 3 by using Kronecker products. This yields a map on the
subspace corresponding to master coordinates given by

mzc (k + 1) = µ c mzc (k ) + a2 [mzc (k )]2 + a3[mzc (k )]3 (11)

4. Example

Consider the single inverted pendulum in the horizontal plane with a linear and a
quadratic torsional spring at the base, a linear torsional damper at the base, and acted upon by a
T-periodic follower force proportional to the delayed (with delay period T) angular displacement.
The equation of motion is given as

ml 2 q

+ cq
+ k l q + k n q 2 + [ P1 + P2 cos(ωt )]l sin(q (t ) − ηq (t − T )) = 0 (12)

where T = 2π / ω and m is kilograms, l is in meters, c is torsional damping in N.m.s/rad, k l is

a linear torsional stiffness and k n is nonlinear torsional stiffness in N.m/rad. Expanding in a
Taylor series about the zero equilibrium position and retaining up to cubic terms, equation (12)
can be written in the state space form with P = ( P1 + P2 cos(ωt )) / ml and x(t ) = q (t ) . Also,
(k n / ml 2 ) = (η 3 / 6ml ) can be designated as ε , which conforms to the form of equation (1). At a
particular parameter set given by
 k l / ml 2 = 1.75, c / ml 2 = 0.0482979, P1 / ml = 0.4025,
(13)
P2 / ml = 0.734, k n / ml 2 = η 3 / 6ml = 1 / 6, η = 1, T = 2.4

the system undergoes a flip bifurcation which is evidenced by a -1 eigenvalue of the monodromy
matrix ( I + V ) in equation (4). We choose c / ml 2 as the bifurcation parameter.
The numerical center manifold algorithm is programmed in MATLAB and applied to
system (12) with parameter set (13) with N=32 (collocation points) and n=2 (states) which
translates into a map having a 64-dimensional collocation vector, 2080 quadratic terms and
45760 cubic terms. The size of the coefficient matrices are 64 x 2080 and 64 x 45760, for
quadratic and cubic terms, respectively. The reduced scalar map on the center manifold takes the
form of equation (11) with µ c = −1 + γ where γ is a versal deformation parameter and
a 2 = −0.3310, a 3 = −0.0511 . The post-bifurcation limit cycle amplitude is derived using the
second iterate of equation (11) as

mzc (k + 2) = (1 − 2γ )mzc (k ) − 2δ (mzc (k ))3 (14)

where δ = a2 2 + a3 and terms higher than cubic have been dropped. The fixed points of equation
(14) (from which the amplitude of the periodic orbit can be recovered) are obtained by setting
mzc (k + 2) = mzc (k ) as

γ
mzc (k ) = ± − 2
(15)
a2 + a3

where γ = µ c + 1 . Now the bifurcation parameter c / ml 2 (which was temporarily omitted to

allow for the numerical computations) is re-introduced (compare with [33] for the ODE case) by
computing the gradient g c of γ with respect to c / ml 2 . This is found to be g c = 0.5195 and
hence δ = 0.0583 which implies that the bifurcation is super-critical. Using equation (15), with
γ = g c ∆(c / ml 2 ) , the amplitude of the limit cycle is computed and compared with the one
obtained from the direct numerical simulation. Figure 1 plots the displacement and velocity
variables for equation (12) for half the doubled period after steady state is reached computed via
the MATLAB DDE23 routine and the proposed algorithm. The center manifold (order reduction)
calculations confirm the amplitude and the doubled period.

5. Conclusions

A method to compute center manifold reductions of delay differential equations with periodic
coefficients has been proposed and illustrated. The method uses classical perturbation analysis
assuming that the nonlinearities are multiplied by the perturbation parameter. An approximate
nonlinear map is constructed using Chebyshev spectral collocation which takes the solution in
each interval of length equal to the principal period to the solution in the next interval. The
method can be extended to the analysis of Hopf bifurcation and to order reduction problems
associated with DDEs with periodic coefficients. The scope of the systems considered can also
be widened by considering DDEs with a rational relationship between the delay and the principal
periods. The use of versal deformation theory allows a bifurcation analysis in terms of a
parameter as shown in the example, although the nonlinear map is numerically constructed
(using MATLAB) to allow for computational tractability.

Figure 1. Displacement and velocity for equation (12)

for half of the doubled period with DDE23 (+) and proposed algorithm (continuous)

References
[1] Hassard, B. and Wan, Y. H.:Bifurcation formulae derived from center manifold, Journal of Mathematical
Analysis and Applications, 63, 297-312 (1978).
[2] Carr, J.: Applications of Centre Manifold Theory, Springer, New York (1981).
[3] Kalmar-Nagy, T., Stepan, G., and Moon F. C.: Subcritical Hopf bifucration in the delay equation model for
machine tool vibrations, Nonlinear Dynamics, 26, 121-142 (2001).
[4] Gilsinn, D. E.: Estimating critical Hopf bifurcation parameters for a second order delay differential equation
with application to machine tool chatter, Nonlinear Dynamics, 30, 103-154 (2002).
[5] Wang, Z. H. and Hu, H. Y.: Dimensional reduction for nonlinear time-delayed systems composed of stiff and
soft substructures, Nonlinear Dynamics, 25, 317-331 (2001).
[6] Das, S. L. and Chaterjee, A.: Multiple scales without center manifold reductions for delay differential equations
near Hopf bifurcations, Nonlinear Dynamics, 30, 323-335 (2002).
[7] Wahi, P. and Chaterjee, A.: Galerkin projections for delay differential equations, Proceedings of ASME
DETC’03/VIB-48570, Chicago, IL, Sept. 2-6, 2003.
[8] Ma, H., Butcher E. A. and Bueler, E.: Chebyshev expansion of linear and piecewise linear dynamic systems
with time delay and periodic coefficients under control excitations, ASME Journal of Dynamic Systems
Measurement and Control, 125, 236-243 (2003).
[9] Szalai, R., Stepan, G., and Hogan, S. J.: Global dynamics of low-immersion high-speed milling, Chaos 14, 1-9
(2004).
[10] Szalai, R., and Stepan, G.: Period doubling bifurcation and center manifold reduction in a time-periodic and
time-delayed model of machining, J. Nonlinear Science, in press.
[11] Halanay, S., Differential Equations: Stability, Oscillations, Time Lags, Academic Press, New York (1966).
[12] Butcher, E. A., Ma, H., Bueler, E., Averina, V., and Szabo, Z.: Stability of linear time-periodic delay-
differential equations via Chebyshev polynomials, International Journal of Numerical Methods in Engineering 59,
895-922 (2004).
[13] Bueler, E.: A Chebyshev collocation for linear, periodic ordinary and delay differential equations: a posteriori
estimates, SIAM J. Numerical Analysis, in press.
[14] Butcher, E. A., Nindujarla, P., and Bueler, E.: Stability of up- and down-milling using Chebyshev collocation
method, Proc. ASME DETC 2005-84880, Sept. 24-28, 2005, Long Beach, CA (2005).
[15] Trefethen, L. N.: Spectral Methods in MATLAB, Software, Environments, and Tools, Society for Industrial and
Applied Mathematics (SIAM), Philadelphia, PA (2000).
[16] David, A. and Sinha, S. C., Versal Deformation and Local Bifurcation Analysis of Time-Periodic Nonlinear
Systems, Nonlinear Dynamics 21, 317-336 (2000).