Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

A TWO-DEGREES-OF-FREEDOM HAMILTONIAN MODEL:

AN ANALYTICAL AND NUMERICAL STUDY


RAFFAELLA PAVANI

A well-studied Hamiltonian model is revisited. It is shown that known numerical results


are to be considered unreliable, because they were obtained by means of numerical meth-
ods unsuitable for Hamiltonian systems. Moreover, some analytical results are added.

Copyright © 2006 Raffaella Pavani. This is an open access article distributed under the
Creative Commons Attribution License, which permits unrestricted use, distribution,
and reproduction in any medium, provided the original work is properly cited.

1. Introduction
Our aim is to study a well-known structural engineering problem about anomalous
elastic-plastic responses of a two-degrees-of-freedom model of a fixed ended beam with
short pulse loading (e.g., [2] and references therein). In particular, the resulting elastic
vibrations may be chaotic. We will tackle this problem mainly from the point of view of
numerical analysis, but we provide even some new theoretical results.
This system was already extensively studied using Runge-Kutta methods with variable
stepsize and many results can be found in [2] and references therein, but here we want to
show that some other numerical methods can be more effective in order to understand
the qualitative behavior of the orbits, in particular when chaotic behavior is detected.
Moreover, some new analytical results support our conclusions.
In Section 2, the problem is described and equations of the used mathematical model
are provided. In Section 3, we present some theoretical results about the behavior of so-
lutions close to the equilibrium point. In Section 4, numerical results are reported and
comparisons with already known results are shown. At last, Section 5 is devoted to a final
discussion.

2. Beam mathematical model


The two-degrees-of-freedom model of a fixed ended beam is provided by Carini et al. [2].
This model was deeply studied in the field of structural engineering and enjoys a large lit-
erature, which we do not cite here for the sake of brevity (e.g., see references in [2]). In
particular it is known that a beam, deformed into the plastic range by a short transverse

Hindawi Publishing Corporation


Proceedings of the Conference on Differential & Difference Equations and Applications, pp. 905–913
906 A two-degrees-of-freedom Hamiltonian model

force pulse, can exhibit anomalous behavior when its fixed ends prohibit axial displace-
ments. Indeed, the resulting elastic vibrations may be chaotic. Here we consider the “gen-
eralized” problem proposed in [2], where the plastic strains in the beam are regarded as
given and the “loading” is taken as the imposition of initial conditions of displacement
and velocity. Damping is neglected. So the system becomes a conservative Hamiltonian
system.
The beam model is provided with two cells B and C with two flanges as in the sand-
wich beam, each exhibiting elastic-perfectly plastic behavior. Assuming symmetrical de-
flections with respect to the midsection, there are two unknown transverse displacements,
that is, w1 at the quarterpoint and w2 at the midsection, and an axial displacement u at
the quarterpoint. The axial force is assumed to be constant over the span, therefore the
axial displacement can be found in terms of the transverse displacements. Consequently,
the configuration is defined by two transverse displacements w1 and w2 , which serve as
generalized coordinates of the configuration.
The nonlinear system modeling the given beam is given by Carini et al. [2] as follows:
·· ··  
4w1 + w2 = −β 4w13 − 6w12 w2 + 4w1 w22 − w23 − 5kw1 + 3kw2 ,
·· ··   (2.1)
w1 + 2w2 = −β w23 − 3w22 w1 + 4w2 w12 − 2w13 + 3kw1 − 2kw2 ,

where β = 3.5555556 × 1012 , k = 2.61123556 × 107 .


Here we neglect the four plastic strains and assume that the stress σαi (α = B,C; i = 1,2)
are given by
   2 
w2   −w12 w1 − w2
σB1 = C 2 1 − 0.0271 w2 − 2w1 + + ,
0.1 0.1 0.1
   2 
w2   −w12 w1 − w2
σB2 = C 2 1 + 0.0271 w2 − 2w1 + + ,
0.1 0.1 0.1
  2   2  (2.2)
w1 − w2   −w12 w1 − w2
σC1 = C 2 + 0.0271 w2 − w1 − + ,
0.1 0.1 0.1
   2 
(w − w2 )2   −w12 w1 − w2
σC2 = C 2 1 − 0.0271 w2 − w1 − + ,
0.1 0.1 0.1
where C = 4e + 9.
· ·
Kinetic energy T(w1 , w2 ) and potential energy V (w1 ,w2 ), relevant to half beam, have
the following expressions:

· 2 · · · 2

T = 1.8e − 3 2w1 + w1 w2 + w2 ,
  (2.3)
V = 2.5e − 17 σB1
2
+ σB2
2
+ σC1
2
+ σC2
2
,

and the Hamiltonian function is given by H = T + V .


Alternatively, Hamiltonian function can be written as
   
H = 79.365 p12 − p1 p2 + 2p22 + V w1 ,w2 . (2.4)
Raffaella Pavani 907

The dynamical equations in canonical Hamiltonian form can be derived from the Hamil-
tonian function H; so the generalized momenta (p1 , p2 ) are defined by
∂T  · ·  ∂T  · · 
p1 = · = 0.0018 4w1 + w2 , p2 = · = 0.0018 w1 + 2w2 . (2.5)
∂w1 ∂w2
Therefore the nonlinear model of the given beam can be written as the following first-
order system:
· ·  
q1 = w1 = 79.365 2p1 − p2 ,
· ·  
q2 = w2 = 79.365 − p1 + 4p2 ,
·   (2.6)
p1 = −3.2e + 9 8w13 − 12w12 w2 + 8w1 w22 − 2w23 + 0.0073w1 − 0.0044w2 ,
·  
p2 = −3.2e + 9 2w23 − 6w22 w1 + 8w2 w12 − 4w13 + 0.0029w2 − 0.0044w1 .

The system is a two-degrees-of-freedom conservative Hamiltonian system, which turns out


to be nonintegrable, but it can be considered nearly integrable for sufficiently small values
of Hamiltonian H.

3. Analytical results
It is easy to check that the origin is the only equilibrium point of the nonlinear system
(2.6); no saddles are present. Here we will show that, in spite of the fact that the given
Hamiltonian system is nonintegrable, an analytical approximation of the solution can
be found in a neighborhood of the equilibrium point and such approximation can be
as good as we want. Our results are founded on [7, 8], where the basic theorems are
established for one degree of freedom and n (n ≥ 2) degrees of freedom.
At first we recall the following definition.
We call semitrigonometric polynomial (hereinafter STP) every function

f (t) = ar,h,k t r (sinht + coskt) (3.1)

with h,k ∈ R, r ≥ 0, r integer. The class of STPs is closed with respect to integration and
derivation. We remark that a primitive of an STP can always be expressed in closed form
through elementary integrations, giving rise to functions of the same kind. We consider
these polynomials for 0 ≤ t ≤ 1. This is not a restriction, because, if the considered inte-
gration time interval is longer, we can use a union of closed bounded time intervals with
length equal to 1.
We notice that system (2.1) can be written in a more general form as follows:
··  
x1 + ω12 x1 = P1 x1 ,x2 ,
··   (3.2)
x2 + ω22 x2 = P2 x1 ,x2 ,

where P1 (x1 ,x2 ), P2 (x1 ,x2 ) are polynomials of degree 3. Then we can apply to this system
our general results reported in [7], which can be summarized for two degrees of freedom
in the following way.
908 A two-degrees-of-freedom Hamiltonian model

If we assume that the following hypotheses are satisfied:


(1) P1 (x1 ,x2 ) and P2 (x1 ,x2 ) are polynomials in x1 (t), x2 (t) without terms of degree
< 2, with coefficients such that the sum of their absolute values is less than or
equal √ to 1;
(2) ωi > 3 2, i = 1,2;
(3) initial conditions are such that solution x0 (t) of the homogeneous system satisfies
the condition |x0 (t)| ≤ 0.25 − 0.25r , where r is the minimum degree of Pi (x1 ,x2 ),
i = 1,2,
then there exists a domain D ⊂ R2n such that for each initial condition belonging to D,
the solution x(t) = (x1 ,x2 ) of (3.2) can be expressed in the following way:


xi (t) = si, j (t), i = 1,2, (3.3)
j =1

where the series converges uniformly in [0,1], and si, j (t) are semitrigonometric functions
in [0,1], which can be computed in closed form from the initial data.
Observe that, according to our previous definition of STP, every function si, j , i = 1,2,
j ≥ 1 can be explicitly written as
Nj
 
si, j (t) = an t rn sinhn t + coskn t , (3.4)
n =1

where the indices i, j are fixed and the summand exhibits N j terms, N j ≥ 0, N j integer,
hn ,kn ∈ R, r ≥ 0, r integer.
It is clear that system (2.1) does not satisfy the above assumptions. Therefore we have
to rearrange it so that our results can be applied. To this end we write system (2.1) in
matricial form
··
M w + Nw = −βY , (3.5)
 4w3 −6w2 w +4w w2 −w3 
where w = [ ww12 ], M = [ 41 12 ], N = [ −5k3k −2k3k ], Y = w3 −1 3w2 w1 1 +4w
2 1 2 2
3 .
2 w1 −2w1
2
2 2
We change the time scale, so that s = αt, with α = 10 , and the space scale, setting w =
2

εw∗ , with ε = 10−5 .


Then using simultaneous diagonalization by congruencies, we find a unique nonsin-
gular matrix T such that T T MT = I and T T NT = D, where I is the identity matrix and D
is a diagonal matrix. At last using the change of variables w∗ (s) = Tq(s), q = (q1 , q2 ) ∈ R2 ,
from (3.5) we obtain
d2 q
+ Dq = −Z, (3.6)
ds2
Z1
where D = [ 110.16
0 8842.6 ] and Z = [ Z2 ] with
0

Z1 = −5.0572e − 4q13 + 8.0311e − 3q12 q2 − 1.9344e − 2q1 q22 + 4.5443e − 2q23 ,


(3.7)
Z2 = 1.0564e − 3q13 − 6.1083e − 3q12 q2 + 2.6003e − 2q1 q22 − 1.6526e − 3q23 .
Now hypotheses (1) and (2) are satisfied.
Raffaella Pavani 909

About the initial conditions, since they have to be close enough to the origin, we chose:
· ·
w1 (0) = w2 (0) = 0, w1 (0) = 0.56343, w2 (0) = −0.28172, corresponding to the energy level
H = 1e − 3. Changing these initial conditions according to the previous transformation,
we applied our method presented in [7] by the following steps.
(1) We solved the homogeneous form of system (3.6) and we obtained q01 (s) =
−53.749sin10.4959s, q02 (s) = −9.4692sin94.035s. We notice that, as here r = 3,
the condition |q0 (s)| ≤ 0.25 − 0.253 was fulfilled only for s ≤ 4e − 4, that is, t ≤
4e − 6.
(2) We computed the first iterate q1 = (q11 , q12 ), solving system (3.6), where in Z we
replaced q1 (s) with q01 (s) and q2 (s) with q02 (s).
(3) We recovered the original variables (with the first approximated solutions de-
noted w0 and w1 ).
We observe that step 2 can be repeated as many times as it suffices, substituting the kth
approximated solution in Z, in order to get the (k + 1)th approximated solution.
For the sake of brevity we report just the second element in both vectors w0 and w1 :
w02 = 2.4417e − 4sin1049.6t − 5.7212e − 5sin9403.5t,
w12 = −5.7269e − 5sin9403.5t + 2.4226e − 4sin1049.6t
− 6.9642e − 8sin3148.7t + 5.8135e − 10sin28210t (3.8)
+ 1.1712e − 8sin17757t + 7.6531e − 8sin11502t
− 2.4214e − 8sin7304.4t − 8.8818e − 9sin19856t.

As we expected, this orbit is nearly quasiperiodic and the fundamental frequencies are
given by the following (rounded) values:

ω1 = 1049.6, ω2 = 9403.5. (3.9)

We remark that our proposed method belongs to the class of iterative methods; indeed
it is not a series development method. So the accuracy of the approximation can be
estimated in terms of the difference between two successive iterates |xk+1 (t) − xk (t)|,
k = 0,1,2,.... In this case, |w1 (t) − w0 (t)| < 2e − 6 for t ≤ 6e − 3, where 6e − 3 is about the
length of the first cycle of the solution w12 . This enlightens the fact that convergence hap-
pens in a longer interval than expected. Actually, convergence is guaranteed for t ≤ 4e − 6
only, as pointed above, but this condition is just sufficient.

4. Numerical results
In order to carry out a reliable numerical study of our mathematical model (2.1), we
compared results obtained by the following different numerical methods:
(1) an explicit Runge-Kutta method of order 4 (RK4 in public domain);
(2) an explicit Runge-Kutta method with variable stepsize of order 4 (MATLAB rou-
tine ODE45);
(3) an implicit symplectic Runge-Kutta method of order 12 (gni irk2 in public do-
main by E. Hairer);
(4) a conservative method of order 10, which numerically preserves energy (TOMG).
910 A two-degrees-of-freedom Hamiltonian model

Table 4.1

RK4 ODE45 gni irk2 TOMG


h 1e − 5 2.9e − 5 1e − 5 9e − 6
Δs 9.032e − 6 9.528e − 6 9.035e − 6 9.035e − 6
ΔH 4.1e − 9 1.7e − 6 1e − 18 1e − 17

The symplectic method is described in [5], whereas the conservative method will be
available as soon as possible, since at the present it is just a preliminary version (by F.
Mazzia and the author) of a new program inspired by both TOM methods for BVM
problems and GAM methods for IVP problems [1].
In particular, we considered two types of orbits: a close-to-equilibrium orbit and a
far-from-equilibrium orbit. As the considered system is nonintegrable, but nearly inte-
grable for small values of Hamiltonian, we should obtain different qualitative behaviors
of computed orbits.

4.1. A close-to-equilibrium orbit. We refer to the initial conditions chosen in Section 3.


Within the first cycle of the considered orbit, that is, for t ≤ 6e − 3, we report for each
method the maximum difference Δs between the computed solution and the analytical
one, that is, w12 reported in Section 3; the corresponding integration step h; the max-
imum differences ΔH between the computed Hamiltonian and the analytical one (see
Table 4.1).
Here the value of h for ODE45 has to be read as the maximum of the used integration
steps. Moreover, we recall that here the used machine precision is 2.2e − 16.
It is clear that all the numerical methods are equivalent about the approximation of the
solution, but they behave in a different way about the conservation of energy. Indeed, in
this short interval of time, the symplectic and the conservative methods preserve exactly
energy, but the Runge-Kutta methods do not. Now a fundamental question arises: what
happens over long time, when the analytical solution is not easily feasible?
The conservation of energy exhibits a completely different behavior, as Figure 4.1
shows; here the maximum error in Hamiltonian is reported versus increasing intervals
of time.
We observe that the conservative method TOMG goes on conserving the energy with
high constant accuracy, as requested. Actually, the Hamiltonian error always retains the
order 10−12 .
On the contrary, gni irk2 loses an order of magnitude in Hamiltonian error within
420 seconds; this means that in a couple of hours the system stops, instead of oscillating
indefinitely.
Even ODE45 exhibits the same linearly increasing behavior of Hamiltonian error (not
reported in Figure 4.2). This means that again the considered system is numerically sim-
ulated as it were dissipative. This is in accordance with the fact that Hamiltonian systems
are not structurally stable against non Hamiltonian perturbations, such as those introduced
by classical explicit Runge-Kutta methods (with both fixed and variable stepsizes).
Raffaella Pavani 911

×10−11 H = 1e − 3
6

Maximum error in Hamiltonian


5
Gni
4

1
TOMG
0
0 50 100 150 200 250 300 350 400 450
Time (s)
Figure 4.1

H = 1000, p2 = 1.733
102
TOM
100
Power spectrum density

10−2

10−4

10−6

10−8

10−10
0 5 10 15 20 25 30 35 40 45 50
Frequency (kHz)
Figure 4.2

About gni irk2, we carried out many other different tests and we concluded that the
behavior of gni irk2 does not change decreasing the stepsize, but it does change if we
solve the scaled system (3.6) instead of system (2.1). In this case, the symplectic method
preserves energy with a constant maximum error ΔH, which turns out to be ᏻ(10−12 ),
exactly as TOMG does. It is clear that for the highly nonlinear Hamiltonian model (2.1),
gni irk2 experiences stability problems, which are well known indeed (e.g., [3]) for gen-
eral symplectic methods.
Consequently, we suggest to choose the numerical method with great care, even for
close-to-equilibrium orbits.

4.2. A far-from-equilibrium orbit. We chose a second set of initial conditions: w1 (0) =


· ·
w2 (0) = 0, w1 (0) = 430.16, w2 (0) = 266.31, corresponding to the energy level H = 1000.
Unfortunately, no analytical approximation of the solution is available for this case. Here
the oscillation cycle is long about 4e − 4 seconds.
912 A two-degrees-of-freedom Hamiltonian model

H = 1000
10−1
10−2

Power spectrum density


10−3
10−4
10−5
10−6
10−7
10−8
10−9
0 5 10 15 20 25 30 35 40 45 50
Frequency (kHz)
Figure 4.3

For t ∈ [0,10], we compare solutions computed by TOMG and ODE45, since all the
results presented in literature are obtained by explicit Runge-Kutta methods of order 4
with variable stepsize, such as the method implemented by ODE45. There is no robust
way to distinguish chaotic from quasiperiodic solutions on the basis of the computed
waveform. So different tools (e.g., power spectra, Lyapunov exponents, Poincaré maps,
phase diagrams, etc.) should be used to detect a chaotic behavior. Here we resort just to
power spectrum, because it suffices to show how explicit Runge-Kutta methods can be
misleading in the study of Hamiltonian systems.
In Figure 4.2, we report the power spectrum computed using the MATLAB FFT rou-
tine applied to the data provided by TOMG; the semilogarithmic scale is used. It is clear
that the main frequencies are still two, as in the case of close-to-equilibrium orbit given
in Section 4.1. However here the main frequencies are perturbed by others; this means
that the nearly quasiperiodic orbit is going to become chaotic; indeed, the orbit exhibits
the so-called “Nekhoroshev” regime (e.g., [4]). The same qualitative results are obtained
by gni irk2, but with larger perturbations of the main frequencies, as expected.
In Figure 4.3, we report the power spectrum computed using the MATLAB FFT rou-
tine applied to the data provided by ODE45 routine; again the semilogarithmic scale is
used. Here the power spectrum suggests a chaotic behavior, which means a completely dif-
ferent conclusion. Actually, it does not present any dominant frequency, but a large band
of frequencies very close one to another. It is obvious that any other tool, such as, for
example, a Poincaré map should exhibit an analogous chaotic behavior, whenever it uses
the same discrete set of numerical data, provided by ODE45.

5. Conclusions
Conservative dynamical systems are not structurally stable against nonconservative per-
turbations. Actually, the popular explicit Runge-Kutta methods (with both fixed and vari-
able stepsizes) introduce such kind of perturbations. Indeed, mathematical models are
often discretized according to algorithms that have little to do with the original problem,
Raffaella Pavani 913

whereas computational methods should reflect known structural features of the problem
under consideration, in particular they should preserve Hamiltonian for Hamiltonian
problems (e.g., [6] and references therein).
As well known, the most basic structural property of Hamiltonian systems is that they
are symplectic (i.e., they conserve phase-space volume); so symplectic numerical meth-
ods seem to be the first choice. However, according to the seminal paper by Ge Zhong and
Marsed [9], if an integrator is both symplectic and conservative, it must be exact. Nor-
mally the luxury of an exact discretization is not available, therefore in general symplectic
integration does not conserve Hamiltonian, in particular for highly nonlinear noninte-
grable Hamiltonian systems. Here we suggest to use conservative methods, which guaran-
tee the conservation of Hamiltonian. Then the drawback is that Hamiltonian phase-space
structure will not be preserved. Indeed, the problem of the choice between symplectic and
conservative methods remains open.
Actually, the integration method that is most suitable for a given Hamiltonian system
ultimately depends on the nature of the physical problem, the integration time scale, and
the kinds of questions addressed by the numerical simulation.
However, for systems which are to remain essentially nondissipative through numeri-
cal simulations (such as systems in structural engineering), conservative numerical meth-
ods seem to be the best choice.

References
[1] L. Brugnano and D. Trigiante, Solving Differential Problems by Multistep Initial and Boundary
Value Methods, Stability and Control: Theory, Methods and Applications, vol. 6, Gordon and
Breach Science, Amsterdam, 1998.
[2] A. Carini, L. Castiglioni, and P. S. Symonds, Regular and chaotic responses of a Hamiltonian beam
model, European Journal of Mechanics. A Solids 16 (1997), no. 2, 341–368.
[3] O. Gonzalez and J. C. Simo, On the stability of symplectic and energy-momentum algorithms for
non-linear Hamiltonian systems with symmetry, Computer Methods in Applied Mechanics and
Engineering 134 (1996), no. 3-4, 197–222.
[4] M. Guzzo, Long-term stability analysis of quasi integrable degenerate systems through the spectral
formulation of the Nekhoroshev theorem, Celestial Mechanics & Dynamical Astronomy 83 (2002),
no. 1–4, 303–323.
[5] E. Hairer and M. Hairer, GniCodes-Matlab Programs for Geometric Integration, under the item
“software”, 2002, http://www.unige.ch/math/folks/hairer.
[6] R. Pavani, Numerical Hamiltonian chaos, Proceedings of 10th Jubilee National Congress on The-
oretical and Applied Mathematics, Varna (Ya. Ivanov, E. Manoach, and R. Kazandjiev, eds.),
Prof. Marin Drinov Academic Publishing House, Sofia, September 2005.
[7] , On the representation of close-to-equilibrium solutions of n-dimensional conservative os-
cillators, preprint in Quad. Dip. Mat. - Politecnico di Milano N. 644/P (2005).
[8] R. Pavani and R. Talamo, On the representation of periodic solutions of Newtonian systems, Math-
ematical and Computer Modelling 42 (2005), 1255–1262.
[9] G. Zhong and J. E. Marsden, Lie-Poisson Hamilton-Jacobi theory and Lie-Poisson integrators,
Physics Letters. A 133 (1988), no. 3, 134–139.

Raffaella Pavani: Department of Mathematics, Politecnico di Milano, 20133 Milano, Italy


E-mail address: rafpav@mate.polimi.it

You might also like