You are on page 1of 6

Normal Form Transformation Around the Fold-Hopf

Bifurcation

September 14, 2021

Science abounds with examples of systems governed by simple rules yet exhibiting marvelously
complex behaviours. A broad class of instances of this phenomenon is the occurrence of chaos
in dynamical systems. By a dynamical system we mean a system of autonomous first-order
differential equations or discrete maps:
(n+1) (n)
ġi = βi (gj ) , gi = Ri (gj ) , (1)

where the variables gi are either real- or complex-valued. The study of chaos in such systems dates
back to the work of Henri Poincaré on the three-body problem. In the Hamiltonian formalism, the
trajectories of particles in phase space are described precisely by the kind of first-order differential
equations listed in (1). In his investigations of the equations of motion, Poincaré was startled to
discover that the solution space was vastly more perplex than he had anticipated, encompassing
ever-wiggling curves and an infinitude of periodic orbits dispersed unevenly in phase space.
Since the time when Poincaré caught his first glimpse of chaos, the hallmarks by which to identify
chaos have become much better understood. In addition to the presence of an infinite num-
ber of periodic orbits with an infinite range of periodicities, sometimes forming intricate fractal
structures, chaotic systems are characterized by an extreme sensitivity to initial conditions — a
sensitivity often given a precise formulation in terms of positive Lyapunov exponents — and the
property that open sets of initial states evolve in time to spread out densely in the space of all
possible states.
Dynamical systems have a wide range of applications in science and technology, and the emergence
of chaos is a commonplace occurrence in these applications, be they planetary orbits, atmospheric
convection, or population dynamics. But one class of dynamical systems in which no system has
previously been identifie as chaotic are the renormalization group (RG) flows of quantum field
theories (QFTs). Rather than trajectories of particles in phase space, these systems describe the
flow of coupling constants in a given theory as we vary the length or energy scale at which we
view the theory, but the flow equations remain of the form (1). It is therefore natural to expect
that QFTs should admit chaotic RG flows, and already Wilson and Kogut entertained this pos-
sibility in their classic review. But quite generally the theories studied in the literature exhibit
an RG flow of a simpler kind, namely flow between fixed points. Associated herewith is the idea
of universality: we can modify the details of the UV theory and still flow to the same IR theory,
all that is required is that we remain in the same basin of attraction. Contrariwise, chaos would
spell the doom of universality, with even the tiniest change to the UV theory drastically altering
the IR theory. In two, three, and four dimensions it is known that unitarity prevents this kind
of behaviour by guaranteeing the existence of c-, F -, or a-functions that change monotonically
under RG flow, and the same may be true in higher dimensions. It has also been suggested in the
literature that universality may extend beyond the realm of unitary theories.

1
In the present paper, we present two examples of chaotic RG flows: one discrete and one con-
tinuous. These examples arise on analytic continuation of conventional theories into non-unitary
regimes, in one case by allowing imaginary values of a coupling constant, in the other by allowing
symmetry groups of matrices of non-integer ranks. While many cases are known of non-unitary
theories that describe real-life physics, analytic continuation is such a commonplace tool in the-
oretical physics that we also deem the study of the full scope of its implications a worthwhile
endeavour in itself. Finally, we observe that, allowing for these analytic continuations, a wide
and varied range of infinite families of dynamical systems can be realized as RG flows, suggesting
that just as the mathematical literature on dynamical systems illuminates the study of QFTS,
so physical insight may enhance the current understanding of dynamical systems. As a case in
point, we demonstrate that one of the simplest, most extensively studied chaotic maps in the
math literature can be obtained as a special case of the Ising model.
In general it is very difficult to conclusively prove the presence of chaos in a system. But there are
a limited number of tools available. One method is to map a system onto one the few well-studied
systems that are known to be chaotic. And so in section ???, we show that the discrete RG flow of
the one-dimensional Ising model with imaginary coupling is equivalent to the chaotic map known
as the dyadic transformation, and we discuss the implications of this equivalence. For continuous
dynamics, a set of necessary conditions for the onset of chaotic dynamics, involving the presence
of a homoclinic orbit, was put forward by Shilnikov in the 1960’ies. We review his construction in
section ???. Subsequent to Shilnikov’s discovery, mathmaticians were able to show that his condi-
tions are met generically in the vicinity of certain kinds of codimension-four bifurcations, results
which were later extended to the vicinity of special codimension-three bifurcations, and finally in
2019 it was shown that chaotic dynamics occur generically near codimension-two bifurcations of a
certain class. In section ???, we present a tensor model with O(m) × O(M ) symmetry whose beta
functions undergo the above-mentioned bifurcation at the special values m = 2.521, M = 1.972,
and we provide numerical evidence that there exists a Shilnikov homoclinic orbit among the RG
trajectories of the model, thereby establishing that this QFT exhibits chaotic RG flow.

qfts encode fractals

2
Consider our tensor model at the special values of m and M given by

m∗ = 2.520527996
(2)
M∗ = 1.972260742 .

At these values the beta functions have a fixed point at

g1∗ = 31.01364337
g2∗ = 14.90146693
g3∗ = 8.879972225
g4∗ = 136.1544857
(3)
g5∗ = 3.810504207
g6∗ = −143.4849701
g7∗ = −18.64425623
g8∗ = 18.15961700 .

At this fixed point, the stability matrix has eigenvalues λi given by



1, −0.48780, 0.40014, −0.22070, −0.11374, 0.027555i, −0.027555i, 0 , (4)

whose respective eigenvectors we denote vi . The simultaneous presence of a zero eigenvalue and a
conjugate pair of imaginary eigenvalues signifies that a zero-Hopf bifurcation occurs at (m∗ , M ∗ ).
To study this bifurcation more closely, we can perform a change of variables so as to bring the
beta functions into the so-called normal form. First, we construct a transformation matrix given
in terms of the eigenvectors vi as
 
T = v1 , v2 , v3 , v4 , v5 , Re[v6 ], Im[v6 ], v8 . (5)

Using this matrix, we introduce new couplings ei defined by

ei = (T −1 )ij (gj − gj∗ ) . (6)

In the new basis, the directions e6 , e7 , and e8 are tangential to the center manifold, which, still
assuming (m, M ) = (m∗ , M ∗ ), to linear order decouples from the other directions:

βei = λi ei + O(e2 ) i ∈ {1, 2, 3, 4, 5, 8}


2
βe6 = ω e7 + O(e ) (7)
βe7 = − ω e6 + O(e2 ) ,

where ω = |λ6 | = 0.027555. This means that to second order in the dynamical variables, the flow
on the center manifold is described by the beta functions βe6 , βe7 , and βe8 evaluated on the surface
e1 = e2 = e3 = e4 = e5 = 0. Defining t ≡ log µ where µ is the RG scale, the center manifold flow
equations are given to order O(e3 ) by

de6
= ω e7 + A66 e26 + A67 e6 e7 + A68 e6 e8 + A77 e27 + A78 e7 e8 + A88 e28
dt
de7
= − ω e8 + B66 e26 + B67 e6 e7 + B68 e6 e8 + B77 e27 + B78 e7 e8 + B88 e28 (8)
dt
de6
= C66 e26 + C67 e6 e7 + C68 e6 e8 + C77 e27 + C78 e7 e8 + C88 e28 .
dt

3
where the coefficients are given by

A66 = −0.033218 B66 = −0.037662 C66 = −0.037662


A67 = 0.0022881 B67 = 0.0018789 C67 = 0.0023456
A68 = 0.064437 B68 = 0.071545 C68 = 0.073120
(9)
A77 = 0.00017549 B77 = 0.000075176 C77 = 0.00016218
A78 = −0.0023731 B78 = −0.0018765 C78 = −0.0024316
A88 = −0.031205 B88 = −0.034794 C88 = −0.035441 .

We can simplify the description of the flow greatly by a sequence of variable transformations. By
a reparametrization of the independent variable from t to τ given by
dt 1 B68 − A78 
= 1+ e8 , (10)
dτ ω 2ω
we obtain equations of the form (8) but with ω equal to one and the coefficients A78 and B68
equal. Subsequently, we perform a rotation in the (e6 , e7 )-plane so as to equate A13 and B23 . This
is achieved by changing to variables ee6 and ee7 given by

ee6 = cos(θ)e6 + sin(θ)e7 (11)


ee7 = cos(θ)e7 − sin(θ)e6 , (12)

with θ = 1.1886. Lastly, we introduce yet new variables, defined via rescalings and quadratic
shifts of the previous variables:
 
2 2 2
x = R ee6 + a66 ee6 + a67 ee6 ee7 + a68 ee6 e8 + a77 ee7 + a88 e8 (13)
 
2 2 2
y = R ee7 + b66 ee6 + b67 ee6 ee7 + b77 ee7 − a68 ee7 e8 + b88 e8 (14)
 
2 2
z = − r e8 + c66 (e e6 − ee7 ) + c67 ee6 ee7 + c68 ee6 e8 + c78 ee7 e8 , (15)

where for the various constants we have chosen the special values

a66 = −0.75702
b66 = −0.92404 c66 = −0.22212
a67 = 0.72615
b67 = −0.69275 c67 = 0.52488
a68 = 0.86938 (16)
b77 = −0.75886 c68 = 2.4950
a77 = 0.14105
b88 = −1.5939 c78 = 0.90769
a88 = −0.57986
R = 9.93551 r = −1.2862 . (17)

The result of this final change of variables is to bring the dynamical equations into the simple
form:
dx
= y − axz

dy
= − x − ayz (18)

dz
= z 2 + x2 + y 2 ,

dz
where a = 0.882592. Because a > 0 and the coefficient of x2 + y 2 in the equation for dτ is plus
one rather than minus one, the zero-fold bifurcation is of the kind that Guckenheimer and Holmes

4
refer to as type III.

Consider now the RG flow of the variables ei when m and M are not precisely equal to m∗
and M ∗ but deviate by some small amounts δm = m − m∗ and δM = M − M ∗ . In this case,
the beta functions contain constant terms and the directions e1 , ..., e5 couple linearly to e6 , e7 ,
and e8 . But we can decouple them and get rid of the constant terms by changing to variables wi ,
i ∈ {1, 2, 3, 4, 5}, given by
(0) (6) (7) (8)
wi = ei + Ai + Ai e6 + Ai e7 + Ai e8 , (19)
(k)
where Ai are 20 constants whose values we choose such that in the five beta functions βwi , the
constant terms and the linear terms proportional to e6 , e7 , and e8 all vanish. This means that

βwi = O(e2 ) . (20)


{w}=0

In other words, up to second order in the dynamical variables, the surface with wi = 0 for all
i ∈ {1, 2, 3, 4, 5} is an invariant manifold, whose flow is described by βe6 , βe7 , and βe8 . As we
did above, we can simplify the description of the flow by carrying out an analogous sequence
of variable changes and reparametrization. But because we are no longer working right at the
bifurcation point (m∗ , M ∗ ), we cannot get rid of as many terms as previously. We arrive at the
following simplified equations governing the flow:
dx
= y + νx − axz

dy
= − x + νy − ayz (21)

dz
= − µ + z 2 + x2 + y 2 ,

where a, µ, and ν to leading order in δm and δM are given by:

ν = − 40588.5 δm − 208279.2 δM (22)


µ = − 89048.7 δm − 456856.8 δM .

A property of the system (21) is that it exhibits a rotational symmetry in the (x, y)-plane. This
symmetry is not a symmetry of the full system; rather it is an artifact of omitting terms of
cubic and higher order. Consequently, the approximation (21) misrepresents essential qualitative
properties of the RG flow around the bifurcation point. To remedy this situation we will determine
the cubic corrections to (21), contenting ourselves to so doing at zeroth order in δm and δM . To
attain this end we must decouple the beta functions βe1 , ..., βe5 from the center manifold to second
order in the dynamical variables. We therefore introduce, for i ∈ {1, 2, 3, 4, 5}, yet other variables
hi given by
(66) 2 (67) (68) (77) 2 (78) (88) 2
hi = ei + Ai e6 + Ai e6 e7 + Ai e6 e8 + Ai e7 + Ai e7 e8 + Ai e8 , (23)
(jk)
where Ai are a set of 30 constants whose values we choose such that the 30 coefficients of the
monomials ej ek with j, k ∈ {6, 7, 8} in the five beta functions βhi are all equal to zero. This entails
that

βhi {h}=0
= O(e3 ) . (24)

5
Consequently, for i ∈ {6, 7, 8}, the beta functions βei {h}=0 describe the flow on the center manifold
up to third order in e6 , e7 , and e8 . If we now reparameterize t and change to variables x, y, z as
above, we arrive at the flow equations
dx X
= y + νx − axz + Aijk ri rj rk
dτ i≤j≤k
dy X
= − x + νy − ayz + Bijk ri rj rk (25)
dτ i≤j≤k
dz X
= − µ + z 2 + x2 + y 2 + Cijk ri rj rk ,
dτ i≤j≤k

*
where r = {x, y, z} and the coefficients are given by

Axxx = 0.52880 Bxxx = 0.36717 Cxxx = 2.0375


Ayyy = −1.0863 Byyy = 1.1396 Cyyy = 4.2162
Azzz = 0.61562 Bzzz = −2.7228 Czzz = −3.1416
Axxy = −2.1729 Bxxy = −1.5311 Cxxy = −1.6296
Axxz = 3.6758 Bxxz = −0.56682 Cxxz = 0.93114
(26)
Axyy = 2.3277 Bxyy = 0.94937 Cxyy = 2.5140
Ayyz = 2.5491 Byyz = −4.6653 Cyyz = −9.3359
Axzz = 1.3303 Bxzz = 0.37703 Cxzz = −4.9165
Ayzz = −2.3099 Byzz = 6.5568 Cyzz = 6.8966
Axyz = −0.98669 Bxyz = 1.3121 Cxyz = 0.84014 .

The cubic terms in (25) are non-resonant, meaning that one could get rid of them by a suitable
change of variables and thereby restore the (x, y) rotation symmetry at cubic order. But in doing
so one generates new symmetry-breaking terms at higher order. In principle one could retain
the symmetry at any desired order by iterating this procedure, but retaining the symmetry at all
orders requires a singular change of variables.

You might also like