Chaos, Solitons and Fractals 32 (2007) 1038–1052

www.elsevier.com/locate/chaos

Generalized Lorenz models and their routes to chaos.

I. Energy-conserving vertical mode truncations
D. Roy, Z.E. Musielak *

Department of Physics, The University of Texas at Arlington, Arlington, TX 76019, USA

Accepted 22 February 2006

Abstract

A two-dimensional and dissipative Rayleigh–Bénard convection can be approximated by the Lorenz model, which
was originally derived by taking into account only three Fourier modes. Numerous attempts have been made to gener-
alize this 3D model to higher dimensions and several different methods of selecting Fourier modes have been proposed.
In this paper, generalized Lorenz models with dimensions ranging from four to nine are constructed by selecting vertical
modes that conserve energy in the dissipationless limit and lead to systems that have bounded solutions. An interesting
result is that the lowest-order generalized Lorenz model, which satisfies these criteria, is a 9D model and that its route to
chaos is the same as that observed in the original 3D Lorenz model. The latter is in contradiction to some previous results
that imply different routes to chaos in several generalized Lorenz systems. This discrepancy is explained by the fact that
previously obtained generalized Lorenz models were derived by using different methods of selecting Fourier modes and,
as a result, most of them did not obey the principle of conservation of energy in dissipationless limit.
 2006 Elsevier Ltd. All rights reserved.

1. Introduction

The Lorenz model [1] has been one of the most celebrated non-linear dynamical systems. The model approximates a
2D dissipative Rayleigh–Bénard convection in a fluid that is treated in the Boussinesq approximation [2]. The descrip-
tion of convection given by this model is simple because only three low-order modes in the double Fourier expansions
of stream function and temperature variations are taken into account [1]. Since the considered problem is two-dimen-
sional (horizontal and vertical), the modes of the Fourier expansions are labelled by two integers m and n that corre-
spond to the horizontal and vertical directions, respectively. Lorenz selected two modes with m = n = 1 and one mode
with m = 0 and n = 2 (see Section 2 for details). Based on this selection, he derived a set of three non-linear and first-
order differential equations, which have bounded solutions and also conserve energy in the dissipationless limit.
Extensive studies of this 3D Lorenz model have shown that the transition to chaos and the formation of strange
attractor in this system occurs via chaotic transients with both homoclinic and heteroclinic bifurcations as well as
homoclinic orbits being present [3–5]. Chaotic transients have been identified as one of the five basic routes to chaos;
the others are: period-doubling, quasi-periodicity, intermittency and crisis [4]. An interesting question is whether a

*
Corresponding author. Tel.: +1 817 272 2513; fax: +1 817 272 3637.
E-mail addresses: dxr3778@exchange.uta.edu (D. Roy), zmusielak@uta.edu (Z.E. Musielak).

0960-0779/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.chaos.2006.02.013
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1039

Lorenz system with a larger number of Fourier modes would have the same route to chaos as the original 3D Lorenz
model. The main goal of our paper is to answer this question by constructing generalized Lorenz systems and studying
routes to chaos in these systems.
Numerous attempts have been made to extend the 3D Lorenz system to higher dimensions by selecting higher-order
Fourier modes. Different authors have selected different modes and in most cases no justification for their selection has
been given [6–12]. In general, three basic methods of constructing generalized Lorenz models can be identified. In the first
method, one keeps m (horizontal modes) fixed and changes n (vertical modes), which is essentially the vertical mode trun-
cation. In the second method, one changes m and keeps n fixed, which means that this is the horizontal mode truncation.
Finally, in the third method, both m and n are being changed and, therefore, we call this method the horizontal and ver-
tical mode truncation. The models constructed by using one of these methods are called here generalized Lorenz models.
In previous work, 4D, 5D and 6D systems constructed by Kennamer [6] were developed by using vertical mode trun-
cation. It was shown that the 4D and 5D models did not form new systems because the added modes did not couple to
the original Lorenz system. However, the 6D model did form a new system and its route to chaos was the same as that
observed in the original 3D Lorenz model [7]. Another 6D model was developed by Howard and Krishnamurti [8] who
basically followed the first method but also included a shear flow. Thiffeault and Horton [9] showed that the model did
not conserve energy in the dissipationless limit. In order to satisfy the latter criterion, they had to add one extra mode
and construct a 7D system.
The third method was used by Humi [10], who constructed a 6D system and showed that period-doubling was the
route to chaos in this system. So far, the most prominent application of the horizontal and vertical mode truncation was
done by Curry [11,12], who constructed a 14D system and showed that his system had bounded solutions and its route
to chaos was via bifurcation by a family of invariant two tori. Another way of constructing generalized Lorenz systems
has been recently suggested by Chen and Price [13]. In neither of these studies was the principle of the conservation of
energy considered.
The previous work described above has concerned 2D Rayleigh–Bénard convection. However, there are also many
published papers dealing with 3D Rayleigh–Bénard convection; full references to those papers are given by Tong and
Gluhovsky [14] who claimed that with exception of their model and that described in [15] all other models cited by them
did not conserve energy in the dissipationless limit. One model that has not been discussed by Tong and Gluhovsky is a
9D model of 3D convection developed by Reiterer et al. [16], who showed that period-doubling was a route to chaos in
their model. Because of some relevance of this model to our results, we shall discuss it in our paper (see Section 3).
The fact that routes to chaos in the generalized Lorenz systems discussed above range from chaotic transients (orig-
inally observed in the 3D Lorenz system) to period-doubling (6D and 9D systems) and to decaying tori (14D system) is
rather surprising. Since different Fourier modes were used to construct different generalized systems, the route to chaos
seems to be very sensitive to the choice of the modes. On the other hand, each one of these generalized Lorenz models
has the 3D Lorenz model as its subset, so one would expect that the dominant route to chaos would remain the same as
in the original Lorenz system. One of the main goals of this paper is to explain these contradictory results.
Generalized Lorenz systems constructed in this paper are based exclusively on the first method and our selection of
vertical modes has been done by applying two basic criteria, namely, the conservation of energy in the dissipationless
limit [9,14] and the existence of bounded solutions [12]. We investigate routes to chaos in these systems and explain the
apparent inconsistency regarding the routes to chaos in the previously developed generalized Lorenz models. Our
choice of the vertical mode truncation to select higher-order Fourier modes can be physically justified by the fact that
fluid motions in thermal convection are primarily in the vertical direction and, therefore, the vertical modes should play
a dominant role in this description. We shall explore generalized Lorenz models constructed by using the second and
third methods in our next papers of this series.
Our paper is organized as follows. In Section 2, we derive generalized Lorenz systems by identifying energy-conserv-
ing Fourier modes in Saltzman’s equations and show that the constructed systems have bounded solutions. The results
of our studies of the onset of chaos and routes to chaos are presented and discussed in Section 3. A brief summary of
our results is given in Section 4.

2. Generalized Lorenz models

2.1. Basic equations

To describe our method of constructing generalized Lorenz systems, let us consider a 2D model of Rayleigh–Bénard
convection in a fluid that is treated in the Boussinesq approximation and described by the following set of hydrody-
namic equations:
1040 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

oV x oV z
þ ¼ 0; ð1Þ
ox oz
 2 
oV x oV x oV x 1 op o V x o2 V x
þVx þVz þ m þ ¼ 0; ð2Þ
ot ox oz q ox ox2 oz2
 2 
oV z oV z oV z 1 op o V z o2 V z
þVx þVz þ  gaT  m þ ¼ 0; ð3Þ
ot ox oz q oz ox2 oz2
 2 
oT oT oT o T o2 T
þVx þVz j þ ¼ 0; ð4Þ
ot ox oz ox2 oz2

where Vx and Vz are the horizontal and vertical components of the fluid velocity, and the fluid physical parameters are
g ¼ g^z is gravity, a is the coefficient of thermal expansion, m is
the density q, pressure p and temperature T. In addition, ~
the kinematic viscosity and j is the coefficient of thermal diffusivity.
We assume that the fluid is confined between two horizontal surfaces located at z = 0 and z = h with T(z = 0) =
T0 + DT0 and T(z = h) = T0, and that the temperature varies between the surfaces as T(z) = T0 + DT0(1  z/h). We
denote the temperature changes due to convection by h(x, z, t) and write T(x, z, t) = T0 + DT0(1  z/h) + h(x, z, t). Using
the continuity equation (Eq. (1)), we introduce the stream function w, which is defined by Vx = ow/oz and Vz =
ow/ox. To express the hydrodynamic equations in terms of h and w, we differentiate Eqs. (2) and (3) with respect to
z and x, respectively, and then subtract the former equation from the latter. This gives
o ow o ow o oh
ðr2 wÞ  ðr2 wÞ þ ðr2 wÞ  ga  mr4 w ¼ 0; ð5Þ
ot oz ox ox oz ox
where $4 = o4/ox4 + o4/ox4. The energy equation (Eq. (4)) can also be expressed in terms of w and h, and we obtain

oh DT 0 ow ow oh ow oh
  þ  jr2 h ¼ 0. ð6Þ
ot h ox oz ox ox oz
We follow Saltzman [2] and introduce the following dimensionless quantities: x* = x/h, z* = z/h, t* = tj/h2, w* = w/j,
h* = hag h3/jm, $* = $/h. Then, Eqs. (5) and (6) can be written as
o ow o ow o oh

ðr2 w Þ    ðr2 w Þ þ   ðr2 w Þ  r   rr4 w ¼ 0 ð7Þ
ot oz ox ox oz ox
and
oh ow ow oh ow oh

R    þ  jr2 h ¼ 0; ð8Þ
ot ox oz ox ox oz
where r = m/j is the Prandtl number and R = agh3DT0/jm is the Rayleigh number. We refer to this set of equations as
Saltzman’s equations. According to Saltzman [2], one may impose the following boundary conditions: w* = 0,
$*2w* = 0 and h* = 0 at both surfaces located at z = 0 and z = h, and write the solutions as double Fourier expansions
of w* and h*
X1 X1 h m n i
wðx ; z ; t Þ ¼ Wðm; n; t Þ exp 2phi x þ z ð9Þ
m¼1 n¼1
L 2h

and
X
1 X
1 h m n i
hðx ; z ; t Þ ¼ Hðm; n; t Þ exp 2phi x þ z ; ð10Þ
m¼1 n¼1
L 2h

where L is the characteristic scale representing periodicity 2L in the horizontal direction. Saltzman [2] expressed W and
H in terms of their real and imaginary parts, W(m, n) = W1(m, n)  iW2(m, n) and H(m, n) = H1(m, n)  iH2(m, n), which
do not show explicitly time-dependence, and substituted the above solutions into Eqs. (7) and (8). The general result
was a set of first-order differential equations for the Fourier coefficients W1, W2, H1 and H2. However, when the theory
was applied to describe cellular convective motions originating from small perturbations, Saltzman fixed the vertical
nodal surfaces of the convective cells by excluding all W2(m, n) and H1(m, n) modes. This Saltzman’s rule was used
by Lorenz to construct his 3D model [1] and we shall also use it here to construct our generalized Lorenz systems.
In addition, both Saltzman and Lorenz did not consider shear flows by excluding the W1 modes with m = 0, so the same
assumption will be made in our derivations.
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1041

2.2. From 3D to 6D Lorenz models

Lorenz [1] selected the following three Fourier modes: W1(1, 1), which describes the circulation of the convective roll,
and H2(1, 1) and H2(0, 2), which approximate the horizontal and vertical temperature differences in the convective roll,
respectively. He introduced the new variables X(t) = W1(1, 1), Y(t) = H2(1, 1) and Z(t) = H2(0, 2), and derived three
well-known, non-linear and first-order differential equations that describe his model. The modes selected by Lorenz
are such that the resulting 3D dynamical system conserves energy in the dissipationless limit [10]. In addition, the solu-
tions of Lorenz’s equations are bounded [12].
To extend the original 3D Lorenz model to higher dimensions, we must select higher-order Fourier modes. As
already discussed in Section 1, there are at least three different methods of selecting the modes. In this paper, we con-
struct generalized Lorenz models by using the vertical mode truncation, which means that we fix the value of m by tak-
ing m = 1 and add the vertical modes in both the stream function w and the temperature variations h (see Eqs. (9) and
(10)). Following this procedure, we select the W1(1, 2) and H2(1, 2) modes, and add the mode H2(0, 4); the fact that the
latter mode must also be considered together with the two former modes has been demonstrated by Thiffeault and Hor-
ton [9,17] based on the energy-conserving principle. Hence, we introduce X1 = W1(1, 2), Y1 = H2(1, 2) and Z1 = H2(0, 4),
and obtain the following set of equations:
dX
¼ rX þ rY ; ð11Þ
ds
dY
¼ XZ þ rX  Y ; ð12Þ
ds
dZ
¼ XY  bZ; ð13Þ
ds
dX 1 r
¼ c1 rX 1 þ Y 1 ; ð14Þ
ds c1
dY 1
¼ rX 1  c1 Y 1  2X 1 Z 1 ; ð15Þ
ds
dZ 1
¼ 2X 1 Y 1  4bZ 1 ; ð16Þ
ds
where s = p (1 + a )t is dimensionless time, a = h/L is the aspect ratio, b = 4/(1 + a ), r = R/Rc with Rc = p (1 + a2)3/
2 2 * 2 4

a2 and c1 = (4 + a2)/(1 + a2).

Since the original Lorenz variables (X, Y, Z) are decoupled from the new variables (X1, Y1, Z1), the obtained set of
equations does not describe a 6D Lorenz system but instead it represents two independent systems. One may attempt to
construct a 5D model by taking H2(0, 2) = 0 and a 4D model with H2(1, 2) = H2(0, 2) = 0, but in both cases the addi-
tional equations are not coupled to the original Lorenz equations.

2.3. From 6D to 9D Lorenz models

Since the models discussed in the previous subsection do not form new generalized Lorenz systems, we must con-
tinue adding modes to determine the lowest-order generalized Lorenz model. Using the method of the vertical mode
truncation, we select the X2 = W1(1, 3), Y2 = H2(1, 3) and Z2 = H2(0, 6) modes and derive the following set of first-order
differential equations:
dX
¼ rX þ rY ; ð17Þ
ds
dY
¼ XZ þ rX  Y þ ZX 2  2X 2 Z 1 ; ð18Þ
ds
dZ
¼ XY  bZ  X 2 Y  XY 2 ; ð19Þ
ds
dX 1 r
¼ c1 rX 1 þ Y 1 ; ð20Þ
ds c1
dY 1
¼ 2X 1 Z 1 þ rX 1  c1 Y 1 ; ð21Þ
ds
dZ 1
¼ 2XY 2 þ 2YX 2 þ 2X 1 Y 1  4bZ 1 ; ð22Þ
ds
dX 2 r
¼ rc2 X 2 þ Y 2 ; ð23Þ
ds c2
1042 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

dY 2
¼ XZ þ rX 2  c2 Y 2  2XZ 1  3X 2 Z 2 ; ð24Þ
ds
dZ 2
¼ 3X 2 Y 2  9bZ 2 ; ð25Þ
ds
where s = p2(1 + a2)t* is dimensionless time, a = h/L is the aspect ratio, b = 4/(1 + a2), r = R/Rc with Rc = p4(1 + a2)3/
a2, c1 = (4 + a2)/(1 + a2) and c2 = (9 + a2)/(1 + a2).
It is seen that in the derived 9D system the modes given by the variables X1, X2, Y1, Y2, Z1 and Z2 are coupled to the
original X, Y and Z Lorenz modes, which means that this is a new generalized Lorenz system. In order to determine
whether this is the lowest-order generalized Lorenz system, we must now consider 8D and 7D systems that are subsets
of the 9D model. To obtain an 8D system, we assume that Z2 = 0 and use it to reduce the above set of equations. The
last term in Eq. (24) becomes zero and, in addition, Eq. (25) yields the condition X2Y2 = 0, which clearly indicates that
the 8D is too limited to represent a new generalized Lorenz system. Even more severe restrictions occur when a 7D
system is derived by taking Y2 = 0. Since neither 8D nor 7D system represents a new system, we conclude that the
9D model described by Eqs. (17)–(25) is indeed the lowest-order generalized Lorenz model that can be obtained by
the method of the vertical mode truncation. Now, it remains to be checked whether our new 9D model conserves energy
in the dissipationless limit and has bounded solutions.

2.4. Validity criteria

There are two validity criteria that our new 9D system must satisfy in order to be considered a physically meaningful
system. The first criterion requires that energy is conserved in the dissipationless limit; from now on when we refer to
the conservation of energy we would always mean its conservation in the dissipationless limit. The fact that some higher
dimensional models of Rayleigh–Bénard convection based on finite Fourier mode truncations do not conserve energy in
the dissipationless limit has already been recognized in the literature [9,14,17]. There are many reasons to have energy-
conserving generalized Lorenz systems. Among these reasons, the most important are: the effects of non-conservation
energy could be large and they are relevant to the energy flow in the dissipative regime, the thermal flux in steady-state is
described correctly only by energy-conserving systems, and energy conserving truncations represent the whole system
more accurately and they reduce unphysical numerical instabilities. The second requirement is that our system has
bounded solutions; systems with unbounded solutions are treated as unphysical. This criterion has been used by many
authors to validate their generalized Lorenz systems [6,9–16].
According to Saltzman [2], the dimensionless kinetic, K*, and potential, U*, energy can be expanded into spectral
components as
1 X 1 X1
K ¼ d2 ðm; nÞjWðm; nÞj2 ð26Þ
2 m¼1 n¼1

and
1r X 1 X1
U ¼  jHðm; nÞj2 ; ð27Þ
2 R m¼1 n¼1

where d2(m, n) = (2pa)2m2 + p2n2.

We now apply these formulas to our 9D model and obtain
1
K  ¼ ½d2 ð1; 1ÞX 2 þ d2 ð1; 2ÞX 21 þ d2 ð1; 3ÞX 22  ð28Þ
2
and
1r 2
U ¼  ðY þ Y 21 þ Y 22 þ Z 2 þ Z 21 þ Z 22 Þ; ð29Þ
2R
where d2(1, 1), d2(1, 2) and d2(1, 3) are constants. To verify that the total energy, E* = K* + U*, is conserved, we write
dE d r
¼ ðK  þ U  Þ ¼ d2 ð1; 1ÞX_ þ d2 ð1; 2ÞX_ 1 þ d2 ð1; 3ÞX_ 2  ðY_ þ Y_ 1 þ Y_ 2 þ Z_ þ Z_ 1 þ Z_ 2 Þ. ð30Þ
ds ds R
Using Eqs. (17), (20) and (23), we obtain
    
dE 1 1 r
¼ r d2 ð1; 1ÞðX  Y Þ þ d2 ð1; 2Þ c1 X 1  Y 1 þ d2 ð1; 3Þ c2 X 2  Y 2  ðY_ þ Y_ 1 þ Y_ 2 þ Z_ þ Z_ 1 þ Z_ 2 Þ.
ds c1 c2 R
ð31Þ
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1043

In the dissipationless limit r ! 0 and Eq. (31) becomes

dE
lim ¼ 0; ð32Þ
r!0 ds
which gives E* = K* + U* = const. Hence, our 9D system conserves energy in the dissipationless limit.
To show that our 9D system has bounded solutions, we introduce the quantity Q [9] defined as

X
3 3 
X 2
2
Q ¼ K  þ 2 jY i j2 þ Zj  ð33Þ
i¼1 j¼1
n

Fig. 1. (X, Y, Z) phase plots of our 9D Lorenz system with r = 24 (upper panel) and r = 35 (lower panel).
1044 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

and note that all selected modes are included in Q, so if one of them diverges, then Q would also diverge. To obtain the
condition for bounded solutions, we take the time derivative of Q and write
dQ
6 ½ minð2m; jÞQ þ 4jn0 ; ð34Þ
ds
where n0 is the number of Z modes. To have dQ*/ds < 0, we must have Q > 4jn0/min(2m, j), which is the condition for
bounded solutions. This condition was checked numerically for our 9D model (see Section 3) and we found it to be
always satisfied. Hence, the model has bounded solutions.

Fig. 2. (X, Y, Z) phase plots of our 9D Lorenz system with r = 40 (upper panel) and r = 43 (lower panel).
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1045

2.5. Comparison to previous work

The most important previous work on generalized Lorenz systems has already been described in Section 1. Here, we
make a comparison between our results and those obtained by Kennamer [6], Humi [10], Curry [11] and Reiterer et al.
[16]. The main reason for this selective comparison is that only generalized Lorenz systems constructed by these authors
are directly relevant to our models. Since in all these cases higher-order modes were added to the original three modes
selected by Lorenz, in the following, we only list those extra modes.
An interesting result has been found by Kennamer who selected the W1(1, 3), H1(1, 3) and H2(0, 4) modes and
obtained an extension of the Lorenz model to six dimensions. He showed that his three extra equations were coupled
to Lorenz’s equations (see Appendix A) and that the solutions to these equations were bounded. The results presented

Fig. 3. (X2, Y2, Z2) phase plots of our 9D Lorenz system with r = 24 (upper panel) and r = 35 (lower panel).
1046 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

in Appendix A also demonstrate that this 6D Lorenz model does conserves energy. Comparison of this model to our
uncoupled 6D model given by Eqs. (11)–(16) clearly shows that Kennamer omitted the modes W1(1, 2) and H2(1, 2)
without giving any physical justification. In addition, his selection of the H2(0, 4) seems to be inconsistent with a general
principle of selecting these modes formulated by Thiffeault and Horton [9].
Another 6D Lorenz model was constructed by Humi [10] who selected the W1(1, 2), W1(2, 1) and H2(1, 2) modes and
showed that the solutions to this new 6D system were bounded. Comparison to our uncoupled 6D system shows that
the H2(0, 4) mode in our model was replaced by the W1(2, 1) mode. This replacement is not consistent with Thiffeault

Fig. 4. (X2, Y2, Z2) phase plots of our 9D Lorenz system with r = 40 (upper panel) and r = 43 (lower panel).
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1047

and Horton’s principle and, therefore, it does not conserve energy. The latter is easy to demonstrate by using the results
given in Section 2.4.
A generalized model developed by Curry [11] has 14 dimensions and it was constructed by using six W1 modes and
six H2 modes with 1 6 m 6 3 and 1 6 n 6 4, and in addition the H2(0, 2) and H2(0, 6) modes. Curry demonstrated that

Fig. 5. Power spectra of our 9D Lorenz system with r = 24 (upper panel) and r = 35 (lower panel).
1048 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

his 14D model had bounded solutions, however, he did not check for conservation of energy. The results of Section 2.4.
can be used to show that Curry’s system does not conserve energy. Note that the energy-conserving 6D Lorenz model
constructed by Kennamer [6] is a subset of the 14D system but neither Humi’s 6D model nor our 9D model are subsets
of this more general system.

Fig. 6. Power spectra of our 9D Lorenz system with r = 40 (upper panel) and r = 43 (lower panel).
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1049

Finally, we want to compare the 9D Lorenz system derived in Section 2.3 to that developed by Reiterer et al. [16]. To
describe 3D square convection cells, the authors have expanded the x, y and z components of a vector potential A, with
V = $ · A, and the temperature variations h into triple Fourier expansions, with l, m and n representing the modes in
the y, x and z direction (see Eqs. (9) and (10)), respectively. These Fourier expansions have been truncated up to the
second-order and a 9D Lorenz system was derived. The authors strongly emphasized that they used a mathematically
consistent approach to select the modes and that all second-order modes have been included in the derivation. The
model is presented in Appendix B, where we also show that this mathematically consistent approach leads a 9D system
that violates the principle of conservation of energy.

3. Routes to chaos

After demonstrating that our new 9D model conserves energy in the dissipationless limit and has bounded solutions,
we may now investigate the onset of chaos in this system and determine its route to chaos. To achieve this, we solved
numerically the set of Eqs. (17)–(25) by fixing the parameters b = 8/3 and r = 10, and varying the control parameter r
over the range 0 6 r 6 50. The main purpose of these calculations was to determine the value of r for which fully devel-
oped chaos is observed and a route that leads to this chaotic regime. The obtained results are presented by using phase
portraits, power spectra and Lyapunov spectra.
The phase portraits for the sets of variables (X, Y, Z) and (X2, Y2, Z2) are given in Figs. 1–4. The results presented in
these figures show time evolution of the system from given initial conditions to a periodic or strange attractor. The pan-
els of Figs. 1 and 3 clearly show periodic behavior of the system for the values of r ranging from 24 to 34. Periodic
behavior is also seen in the upper panels of Figs. 2 and 4, where the phase portraits for r = 40 are presented. Prominent
chaotic behavior of the system is seen in the lower panels of Figs. 2 and 4, where system’s strange attractor is displayed
for r = 43. The strange attractor has the same properties as the Lorenz strange attractor, and these properties are best
displayed in either the (X, Y, Z) or (X2, Y2, Z2) variables. Changes in the behavior of the system with the increasing value
of r are also seen in Figs. 5 and 6, which present the power spectra. The broad band power spectrum shown in the lower
panel of Fig. 6 represents additional strong evidence that the system reached fully a developed chaotic regime.
To determine the route to chaos of this system and the precise value of r at which the system enters the chaotic regime,
we calculated all nine Lyapunov exponents for the system. The three leading Lyapunov exponents are plotted versus r in
Fig. 7. The plots show that one Lyapunov exponent becomes positive when r = 40.43; however, further increase of r leads
to several spikes that quickly disappear at r = 40.49. The same Lyapunov exponent becomes positive again at r = 41.10
and it shows many spikes that imply that the system is entering full chaos via chaotic transients, which is the same route

Fig. 7. Three leading Lyapunov exponents are plotted as a function of the parameter r.
1050 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

as that observed in the original 3D Lorenz system. After the last negative spike at r = 41.44, fully developed chaos is
observed when r P 41.54. These results are consistent with those shown in Figs. 1–6.
There are similarities between our results obtained for the 9D Lorenz model, the original 3D Lorenz model [1], and
the 6D Lorenz model constructed by Kennamer [6]; note that all three models are energy conserving systems. Both the
9D and 6D systems have strange attractors that are very similar to the Lorenz strange attractor. In addition, the route to
chaos via chaotic transients observed in the 9D and 6D [7] systems is the same as that identified for the 3D Lorenz system
[3,4]. The only important difference between these three models is that each one of them exhibits fully developed chaos for
a different value of the parameter r. For the 3D Lorenz system, this critical value is r = 24.75, for the 6D system it is
r = 40.15, and finally for our 9D system r = 41.54 is required. This implies that the larger number of Fourier modes that
is taken into account the higher the value of r required for the system to enter the fully developed chaotic regime.
It becomes now clear that the period-doubling route to chaos discovered by Humi [10] and Reiterer et al. [16] in their
generalized Lorenz models was caused by their selection of modes that do not conserve energy. The same is true for the
14D system constructed by Curry [11] as this system also violates the principle of the conservation of energy. Our results
show that energy-conserving generalized Lorenz models have the same route to chaos as the original 3D Lorenz model,
which is a subset of these models.

4. Summary

We have used the method of the vertical mode truncation to construct generalized Lorenz models that describe a 2D
Rayleigh–Bénard convection. The original 3D Lorenz model is a subset of these models, which means that the general-
ized systems have been derived by adding higher-order Fourier modes to the original three modes selected by Lorenz [1].
An interesting result of our study is that the lowest-order generalized Lorenz model, which conserves energy in the dis-
sipationless limit and has bounded solutions, is a 9D system and that its route to chaos is via chaotic transients, first
observed in the 3D Lorenz system. In addition, we show that more thermal energy is required for our 9D system to
transition to chaos than for the 3D Lorenz system. The critical value of r for which our system exhibits fully developed
chaos is r = 41.54, however, it is only r = 24.75 for the 3D system. This implies that the larger number of Fourier modes
that is taken into account, the higher the value of r required for the system to enter the fully developed chaotic regime.
The method of the vertical mode truncation did not allow us to construct any physically consistent system with
dimensions ranging from four to eight. There are, however, some examples of such models constructed by others [6–
10]. One example is a 6D system developed by Kennamer [6]. There are similarities between this energy-conserving
model and our 9D system, namely, both models have strange attractors that have properties very similar to the Lorenz
strange attractor and their routes to chaos are via chaotic transients. However, the problem is that the 6D model was
derived by skipping the vertical modes with m = 1 and n = 2 and selecting a temperature mode with m = 0 and n = 4 in
an inconsistent way [9,17]. Other 6D and 9D models [10,16], which predicted period-doubling as their routes to chaos,
happened to be energy non-conserving systems. Similarly, it has been shown that a 14D model constructed by Curry
[11], with its route to chaos via decaying tori, does not conserve energy in the dissipationless limit.
Based on our results and their comparison to those previously obtained, it is now clear that the period-doubling route
to chaos discovered by Humi [10] and Reiterer et al. [16] in their generalized Lorenz models was caused by their selection
of the modes that do not conserve energy. The same is true for the 14D system constructed by Curry [11] as this system
also violates the principle of conservation of energy. On the other hand, our results clearly show that when generalized
Lorenz models are constructed in such a way that the energy-conserving Fourier modes are taken into account, then the
route to chaos in these models is the same (chaotic transients) as that observed in the original 3D Lorenz model.

Acknowledgments

We thank J.L. Fry and L.D. Swift for discussions and comments on our paper. Z.E.M. acknowledges the support of
this work by the Alexander von Humboldt Foundation.

Appendix A. Another 6D Lorenz system

Kennamer [6] constructed a new 6D Lorenz system by selecting X1(t) = W1(1, 3), Y1(t) = H2(1, 3) and
Z1(t) = H2(0, 4) in addition to the three modes originally chosen by Lorenz [1]. The 6D system is described by the fol-
lowing set of non-linear and first-order differential equations:
 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052 1051

dX
¼ rX þ rY ; ð35Þ
ds
dY
¼ XZ þ rX  Y þ ZX 1  2X 1 Z 1 ; ð36Þ
ds
dZ
¼ XY  bZ  XY 1  YX 1 ; ð37Þ
ds
dX 1 r
¼ c1 rX 1 þ Y 1 ; ð38Þ
ds c1
dY 1
¼ XZ  2XZ 1 þ rX 1  c1 Y 1 ; ð39Þ
ds
dZ 1
¼ 2XY 1 þ 2YX 1  4bZ 1 ; ð40Þ
ds
where s = p2(1 + a2)t* is dimensionless time, a = h/L is the aspect ratio, b = 4/(1 + a2), r = R/Rc with Rc = p4(1 + a2)3/
a2 and c1 = (9 + a2)/(1 + a2). Kennamer already demonstrated that this system has bounded solutions and its route to
chaos is via chaotic transients [6,7]. We now show that the system conserves energy in the dissipationless limit.
Based on the results obtained in Section 2.4, we know that the potential energy U* approaches zero as r ! 0, thus,
we only calculate the kinetic energy K* of the system and obtain
1
K  ¼ ½d2 ð1; 1ÞX 2 þ d2 ð1; 3ÞX 21 ; ð41Þ
2
where d2(1, 1) and d2(1, 2) are constants. The time derivative of K* is
dK 
¼ d2 ð1; 1ÞX_ þ d2 ð1; 3ÞX_ 1 . ð42Þ
ds
We replace X_ and X_ 1 by the RHS of Eqs. (35) and (38), respectively, and take the limit of r ! 0. Then, we have
dE
lim ¼ 0; ð43Þ
r!0 ds
which gives E* = K* + U* = const and shows that the total energy of this 6D system is conserved.

Appendix B. Another 9D Lorenz system

A system that is of our special interest here is a 9D model developed by Reiterer et al. [16]. To describe 3D square
convection cells, the authors have expanded the x, y and z components of a vector potential A, with V = $ · A, and the
temperature variations h into triple Fourier expansions, with l, m and n representing the modes in the y, x and z direc-
tion (see Eqs. (9) and (10)), respectively. These Fourier expansions have been truncated up to the second-order and the
following Fourier modes have been used: X(t) = A1(0, 2, 2), X1(t) = A1(1, 1, 1), X2(t) = A2(2, 0, 2), Y(t) = A2(1, 1, 1),
Y1(t) = A3(2, 2, 0), Y2(t) = H(0, 0, 2), Z(t) = H(0, 2, 2), Z1(t) = H(2, 0, 2) and Z2(t) = H(1, 1, 1), where A1, A2, A3 and
H are the Fourier coefficients of the Ax, Ay, Az and h expansions, respectively. This selection of Fourier modes leads
to the following equations:
dX
¼ rb1 X  X 1 Y þ b4 Y 2 þ b3 X 2 Y 1  rb2 Z; ð44Þ
ds
dX 1 1
¼ rX 1 þ XY  X 1 Y 1 þ YY 1  rZ 2 ; ð45Þ
ds 2
dX 2 2
¼ rb1 X 2 þ X 1 Y  b4 X 1  b3 XY 1 þ rb2 Z 1 ; ð46Þ
ds
dY 1
¼ rY  X 1 X 2  X 1 Y 1 þ YY 1 þ rZ 2 ; ð47Þ
ds 2
dY 1 1 2 1 2
¼ rb5 Y 1 þ X  Y ; ð48Þ
ds 2 2
dY 2
¼ b6 Y 2 þ X 1 Z 2  YZ 2 ; ð49Þ
ds
dZ
¼ b1 Z  rX þ 2Y 1 Z 1  YZ 2 ; ð50Þ
ds
1052 D. Roy, Z.E. Musielak / Chaos, Solitons and Fractals 32 (2007) 1038–1052

dZ 1
¼ b1 Z 1 þ rX 2  2Y 1 Z þ X 1 Z 2 ; ð51Þ
ds
dZ 2
¼ Z 2  rX 1 þ rY  2X 1 Y 2 þ 2YY 2 þ YZ  X 1 Z 1 ; ð52Þ
ds
where s = (1 + 2k2)t, with k = kx = ky, b1 = 4(1 + k2)/(1 + 2k2), b2 = (1 + k2)/2(1 + k2), b3 = 2(1  k2)/(1 + k2),
b4 = k2/(1 + k2), b5 = 8k2/(1 + 2k2) and b6 = 4/(1 + 2k2). Detailed numerical studies of this 9D system have been done
by Reiterer et al. [16], who showed that the system had bounded solutions and its route to chaos was via period-
doubling.
Comparison of this set of equations to that given in Section 2 (see Eqs. (17)–(25)) shows that these two 9D models
are completely different. One obvious reason for this difference is the fact that our and their 9D models describe 2D and
3D Rayleigh–Bénard convection, respectively. Another more important reason for the difference is that our 9D model
does conserve energy in the dissipationless limit, however, their 9D model does not. To demonstrate this, we use Eqs.
(26) and (27), and extend them by adding summation over l. Since the potential energy approaches zero as r ! 0, we
only calculate the kinetic energy of the system and obtain
1
K ¼ ðC 1 X 2 þ C 2 X 21 þ C 3 X 22 þ C 4 Y 2 þ C 5 Y 21 Þ; ð53Þ
2
where C1, C2, C3, C4 and C5 are constants that depend on l, m and n (see Eq. (26)). Taking derivative with respected to
time, we get
dK
¼ C 1 X_ þ C 2 X_ 1 þ C 3 X_ 2 þ C 4 Y_ þ C 5 Y_ 1 ð54Þ
ds
and use Eqs. (44)–(48) to calculate the time derivatives of X, X1, X2, Y and Y1. Since there are several terms on the RHS
of Eqs. (44)–(48) that do not explicitly depend on r, those terms will remain non-zero, and they cannot cancel each
other, when the limit r ! 0 is applied to Eq. (54). Hence, we obtain
dE
lim 6¼ 0; ð55Þ
r!0 ds
where E = K + U. This shows that the total energy of the 9D system constructed by Reiterer et al. [16] is not conserved
and, therefore, their results on the onset of chaos and route to chaos in this system are not valid. The energy-conserving
9D Lorenz model for a 2D Rayleigh–Bénard convection is presented in Section 2 of this paper.

References

[1] Lorenz EN. J Atmos Sci 1963;20:130.

[2] Saltzman B. J Atmos Sci 1962;19:329.
[3] Thompson JMT, Stewart HB. Nonlinear dynamics and chaos. New York: Wiley & Sons Ltd; 1986.
[4] Hilbron RC. Chaos and nonlinear dynamics. Oxford: Oxford University Press; 1994.
[5] Jackson EA. Perspectives of nonlinear dynamics, vol. 2. Cambridge: Cambridge University Press; 1990. p. 138.
[6] Kennamer KS. MS thesis, The University of Alabama in Huntsville, 1995.
[7] Musielak DE, Musielak ZE, Kennamer KS. Fractals 2005;13:19.
[8] Howard LN, Krishnamurti RK. J Fluid Mech 1986;170:385.
[9] Thiffeault JL, Horton W. Phys Fluids 1996;8:1715.
[10] Humi M. 2004. arXiv:nlin.CD/0409025 v1.
[11] Curry JH. Commun Math Phys 1978;60:193.
[12] Curry JH. SIAM J Math Anal 1979;10:71.
[13] Chen Z-M, Price WG. Chaos, Solitons & Fractals 2006;28:571.
[14] Tong C, Gluhovsky A. Phys Rev E 2002;65:046306-1–046306-11.
[15] Shirer HN. Beitr Phys Atmos 1986;59:126.
[16] Reiterer P, Lainscsek C, Schurrer F, Letellier C, Maquet J. J Phys A: Math Gen 1998;31:7121.
[17] Thiffeault JL. MS thesis, The University of Texas at Austin, 1995.