The presence of chaos (or lack thereof) in relativistic binary inspiral systems has received intense attention recently due to the implications for gravitational-wave detection [1–8], especially regarding the generation of theoretical templates for use in matched filters. There is concern that the sensitive dependence to initial conditions that characterizes chaos may make the calculation of such templates difficult or impossible [8]. In particular, in the presence of chaos the number of templates would increase exponentially with the number of wave cycles to be fitted. In addition to this important concern, the problem of chaos in general relativity has inherent interest, as the dynamical behavior

of general relativistic systems is poorly understood.

Several authors have reported the presence of chaos for systems of two point masses in which one or both particles are spinning [1, 3, 6]. Our work follows up on [3], which studies the dynamics of a spinning test particle orbiting a nonrotating (Schwarzschild) black hole using the Papapetrou equa- tions [Eqs. (2.9)]. We extend this work to a rotating (Kerr) black hole, motivated by the expectation that many astrophysically relevant black holes have nonzero angular momentum. Furthermore, the potential for chaos may be greater in Kerr spacetime since the Kerr metric has less symmetry and hence fewer integrals of the motion than Schwarzschild. In addition, the decision to focus on test particles is motivated partially by the LISA gravitational wave detector [9], which will be sensitive to radiation from spinning compact objects orbiting supermassive black holes in galactic nuclei. Using the Kerr metric is appropriate since such supermassive black holes will in general have nonzero spin. There are many techniques for investigating chaos in dynamical systems, but for the case at hand we favor the use of Lyapunov exponents to quantify chaos. Informally, if 0 is the phase- space distance between two nearby initial conditions in phase space, then for chaotic systems the separation grows exponentially (sensitive dependence on initial conditions): (τ) =0eλτ, whereλ is the Lyapunov exponent. (See Sec. 2.3.1 for a discussion of issues related to the choice of metric used to determine the distance in phase space.) The value of Lyapunov exponents lies not only in establishing chaos, but also in providing a characteristic timescaleτλ = 1/λfor the exponential

separation.

By definition, chaotic orbits are bounded phase space flows with at least one nonzero Lyapunov exponent. There are additional technical requirements for chaos that rule out periodic or quasiperi- odic orbits, equilibria, and other types of patterned behavior [10]. For example, unstable circular orbits in Schwarzschild spacetime can have positive Lyapunov exponents [5], but such orbits are completely integrable (see Sec. 2.6) and hence not chaotic. In practice, we restrict ourselves to generic orbits, avoiding the specialized initial conditions that lead to positive Lyapunov exponents in the absence of chaos.

The use of Lyapunov exponents is potentially dangerous in general relativity because of the
freedom to redefine the time coordinate. Chaos can seemingly be removed by a coordinate trans-
formation: simply letτ0 _{= log}_{τ} _{and the chaos disappears. Fortunately, in our case there is a fixed}

background spacetime with a time coordinate that is not dynamical but rather is simply a reparam- eterization of the proper time. As a result, we will not encounter this time coordinate redefinition ambiguity (which plagued, for example, attempts to establish chaos in mixmaster cosmological mod- els, until coordinate-invariant methods were developed [11]). Furthermore, we can compare times in different coordinate systems using ratios: if tp is the period of a periodic orbit in some coordinate

conversion factor between times in different coordinate systems [5]: t τ = tp τp . (2.1)

For chaotic orbits, which are not periodic, we use the average value ofdt/dτ over the orbit, so that

tλ
τλ
=
_{dt}
dτ
(2.2)

as discussed in Sec. 2.7.4. [This more general formula reduces to Eq. (2.1) in the case of periodic orbits.] Since we want to measure the local divergence of trajectories, the natural definition is to use the divergence in local Lorentz frames, which suggests that we use the proper timeτas our time parameter. The Lyapunov timescale in any coordinate system can then be obtained using Eq. (2.2). Lyapunov exponents provide a quantitative definition of chaos, but there are several common qualitative methods as well, none of which we use in the present case, for reasons explained below. Perhaps the most common qualitative tool in the analysis of dynamical systems is the use of Poincar´e surfaces of section. Poincar´e sections reduce the phase space by one dimension by considering the intersection of the phase space trajectory with some fixed surface, typically taken to be a plane. Plotting momentum vs. position for intersections of the trajectory with this surface then gives a qualitative view of the dynamics. As noted in [4], such sections are most useful when the number of degrees of freedom minus the number of constraints (including integrals of the motion) is not greater than two, since in this case the resulting points fall on a one-dimensional curve for non-chaotic orbits, but are “dusty” for chaotic orbits (and in the case of dissipative dynamical systems lie on fractal attractors). Unfortunately, the system we consider has too many degrees of freedom for Poincar´e sections to be useful. It is possible to plot momentum vs. position when the trajectory intersects a section that is a plane in physical space (sayx= 0) [3], but this is not in general a true Poincar´e section.1

Other qualitative methods include power spectra and chaotic attractors. The power spectra for regular orbits have a finite number of discrete frequencies, whereas their chaotic counterparts are continuous. Unfortunately, it is difficult to differentiate between complicated regular orbits, quasiperiodic orbits, and chaotic orbits, so we have avoided their use. Chaotic attractors, which typically involve orbits asymptotically attracted to a fractal structure, are powerful tools for explor- ing chaos, but their use is limited to dissipative systems [10]. Nondissipative systems, including test particles in general relativity, do not possess attractors [12].

Following Suzuki and Maeda [3], we use the Papapetrou equations to model the dynamics

1_{In [3], they are aided by the symmetry of Schwarzschild spacetime, which guarantees that one component of the}

spin tensor (Sec. 2.2.1 below) is zero in the equatorial plane. As a result, it turns out that all but two of their variables are determined on the surface, and thus their sections are valid. Unfortunately, the reduced symmetry of the Kerr metric makes this method unsuitable for the system we consider in this chapter.

of a spinning test particle in the absence of gravitational radiation. We extend their work in a Schwarzschild background by considering orbits in Kerr spacetime, and we also improve on their methods for calculating Lyapunov exponents. The most significant improvement is the use of a rigorous method for determining Lyapunov exponents using the linearized equations of motion for each trajectory in phase space (Sec. 2.3.1), which requires knowledge of the Jacobian matrix for the Papapetrou system (Sec. 2.5.2). We augment this method with an implementation of an informal deviation vector approach, which tracks the size of an initial deviation of size0and uses the relation

(τ) = 0eλτ discussed above. We are careful in all cases to incorporate the constrained nature of the Papapetrou equations (Sec. 2.2.1) in the calculation of Lyapunov exponents (Sec. 2.4.2).

We use units where G=c= 1 and sign conventions as in MTW [13]. We use vector arrows for
4-vectors (e.g., ~pfor the 4-momentum) and boldface for Euclidean vectors (e.g., ξfor a Euclidean
tangent vector). The symbol log refers in all cases to the natural logarithm log_{e}.