Initial data for an asymptotically flat black string
by
Shannon Potter
A Thesis
Presented to
The University of Guelph
In partial fulfilment of requirements
for the degree of
Master of Science
in
Physics
Guelph, Ontario, Canada
c Shannon Potter, September, 2014 ABSTRACT
Initial data for an asymptotically flat black string
Shannon Potter Co-advisors: University of Guelph, 2014 Dr. Luis Lehner and Dr. Eric Poisson
It has been argued that a black string in a compact fifth dimension is unstable and that this instability can lead to a bifurcation of the event horizon. Such an event would expose a naked singularity and – if observable by an asymptotic observer – would constitute a violation of cosmic censorship. While the previously studied black string appears to bifurcate, strictly there are no asymptotic observers when one dimension is compact. With this motivation, we consider an asymp- totically flat black string in a non-compact fifth dimension, as an instability in this black string may lead to a true violation of cosmic censorship — an event of theoretical interest. We limit ourselves to the construction of initial data — a metric representing the black string at a moment in time.
While we do not consider the time evolution, and hence do not observe the instability, we construct initial conditions which we expect to be unstable. To assess this, we find the apparent horizon and evaluate curvature invariants along the horizon. To my family, my friends and to Vincent:
for forgiving me when I wasn’t there.
To Linda Allen:
who gave advice, and made exceptions, indiscriminately.
And to everyone else in the library, in our office, and in the coffee shops:
you have been the borders of my life.
iii Acknowledgements
Thank you to my advisor Luis Lehner for his unwaverable patience and support. And to the rest of my advisory committee: Martin Williams — who gave always needed encouragement, and Eric
Poisson for reviewing my thesis. And thank you to Reggi Vallillee, for her help with so many things over the last few years.
This work was supported in part by a Queen Elizabeth II Graduate Scholarship in Science and
Technology, in part by an NSERC Discovery Grant to Luis Lehner, and through departmental bursaries.
iv Contents
Acknowledgements iv
1 Introduction 1
1.1 Cosmic censorship ...... 2
1.2 5D Schwarzschild solutions ...... 3
1.3 Gregory-Laflamme Instability ...... 6
1.4 Asymptotically flat black string ...... 8
2 Notation and Definitions 10
2.1 Discretisation ...... 11
2.2 Matrix Operations ...... 13
2.3 Derivatives ...... 14
2.4 Newton’s Method ...... 17
2.5 Sylvester’s Equation ...... 19
2.6 Derivatives of matrix products ...... 22
2.7 Separating a Hadamard product ...... 25
2.7.1 Hadamard product as Kronecker product ...... 25
2.7.2 Hadamard product as matrix product ...... 27
2.7.3 Hadamard product as matrix product for a separable matrix ...... 27
2.8 Separating matrix triple product ...... 30
3 Initial data for an asymptotically flat black string 34
3.1 Cauchy decomposition and constraint equations ...... 34
3.2 Obtaining consistent data from the constraint equations ...... 38
v 3.2.1 Analytic Method ...... 40
3.2.2 Numerical Method ...... 44
3.2.3 Solutions common to both ...... 52
4 Physical properties 55
4.1 Mass ...... 55
4.2 Apparent Horizon ...... 58
4.2.1 Method to solve for apparent horizon ...... 65
4.3 Curvature Invariant ...... 70
5 Tests and Applications 71
5.1 Conformal factor ...... 71
5.1.1 Analytic solutions ...... 71
5.1.2 Numerical Solutions ...... 77
5.2 Physical properties ...... 79
5.3 Concluding Remarks ...... 87
Bibliography 88
Appendices 89
A Numerical Methods 90
A.1 Gauss-Seidel Newton ...... 90
A.2 Linear Least Squares Regression ...... 92
A.3 Interpolation ...... 93
vi List of Tables
2.1 Derivatives and J(X˜ )H˜ for differnt forms of F(X)...... 24
vii List of Figures
1.1 Gregory-Laflamme Instability ...... 8
5.1 Analytic solutions for conformal factor and density ...... 73
5.2 Convergence of analytic solutions for conformal factor ...... 74
5.3 Examples of analytic solutions for an AFBS of varrying magnitudes ...... 75
5.4 Examples of analytic solutions for a deformed AFBS ...... 76
5.5 Numerical conformal factor and unphysical density ...... 79
5.6 Estimate of ADM mass ...... 81
5.7 Analytical conformal factor, unphysical density and Kretchmann scalar ...... 82
5.8 Apparent horizon and residue of conformal factor for analytical solution ...... 83
5.9 Kretchmann scalar evaluated on the horizon for numerical solutions ...... 84
5.10 Numerical conformal factor and unphysical density ...... 85
5.11 Apparent horizon and residue of conformal factor for numerical solution ...... 86
viii Chapter 1
Introduction
In this thesis, we are going to construct a model of a black string, for simplicity, at a single moment in time. A black string is a type of black hole which extends in a hypothetical extra fifth dimension; each cross-section of the string is a four dimensional black hole. We do this because, independent of the plausibility of this black string or a fifth dimension, the black string is known to be possibly unstable and the instability is of theoretical interest. It has been argued that an unstable black string, initially described by a single, continuous horizon, can separate into multiple black holes — an event which exposes a naked singularity.
We work within the framework of General Relativity (GR). The defining equations of GR are the Einstein Field Equations (EFE) which relate the energy of a spacetime — mostly contained in the stress-energy tensor Tαβ with its curvature. The curvature is encoded in the Riemann tensor
α Rβγδ, and in quantities derived from it Rαβ and R. By a spacetime we mean a collection of points which (except within a black hole) form a smooth convex manifold, along with a metric tensor — a mathematical object which specifies how far apart neighbouring points are in the manifold. The
EFE are, 1 R − Rg = 8πT . (1.1) αβ 2 αβ αβ
For a given stress-energy tensor, if there is a unique solution, it is given by a manifold and a metric tensor, which converts coordinate displacements (dxα) to physical distance (ds) by,
2 α β ds = gαβdx dx .
1 α The repeated indices above (α and β) imply summation over the coordinates x . The value of gαβ depends on the coordinate system chosen and the curvature of the spacetime; the curvature tensor
α Rβγδ can be obtained from gαβ. By equation (1.1), if a spacetime is curved, it must contain energy, and likewise the spacetime around an object with energy must be curved. According to GR, massive objects do not exert any force on one another, but rather they curve the spacetime in their vicinity, and it is this curvature which influences the paths of bodies, such as the orbit of the moon or the planets.
A distinct feature arising in GR is the existence of black holes. Owing to the highly curved spacetime where there is a black hole, and the limited speed of light, it is impossible for anything inside to escape the event horizon (even light — which, while massless, is still constrained to follow paths in spacetime).
The interface between the outer region — where the path of an outgoing observer, particle, or ray of light can continue outward from the black hole, and an interior region where it cannot, defines the event horizon. Outside the event horizon, the physics of the black hole may be similar to other massive objects – like a star, other bodies may orbit it etc. Inside the horizon, all paths are converging and focused at the interior. It is not realistic for multiple things to occupy the exact same location, and so at this point — the singularity — GR breaks down. While the theory can predict physical phenomena outside the horizon well, the future of anything which crosses the event horizon to the interior of the black hole is uncertain, as the path inevitably leads to a point which
GR cannot say anything about.
1.1 Cosmic censorship
In four dimensions (4D) — the three familiar spatial dimensions, plus time — the concern that
GR is unable to predict the fate of things entering the black hole is mitigated by the existence of an event horizon. Despite their being a point where GR loses predictability, nothing from the singularity can affect anything outside the horizon — as, by definition, all of the paths are converging to, and terminating at, the singularity — and so anything happening outside the horizon (orbits and paths of observers, particles or light) can be correctly predicted. It is thought that for any
2 spacetime singularity a horizon exists surrounding it and this is formalised by the Cosmic Censorship
Conjecture (CCC). It is in this sense that GR can be safely used: despite having a point (or region) of unpredictability, this region is censored by the event horizon and the external spacetime can still be described by GR.
In 4D, the CCC appears to hold under generic conditions. Stationary black objects are very simple and as a result, only a few different types exist. A non-rotating black hole is a Schwarzshild black hole, and by Birkoff’s theorem this only has a single parameter, mass. A black hole can also rotate, and possess an electric charge – giving rise to Kerr and Reisner-Nordstrom black holes. The fact that there is a unique solution for a given set of parameters is consistent with these black holes being stable. If any of these black objects were found to be unstable — so that a small perturbation to the horizon, which could be caused by in-falling matter, grows in time — then it is plausible that the black hole may undergo changes which expose the singularity. One approach to test the stability of a black object is take the known solution which describes it, perturb it by a small quantity, and determine if the perturbations will grow – indicating it is unstable — or radiate away, leaving a stable black hole (need not radiate). If it were found that a small variation of the initial conditions resulted in a wildly different solution, the black hole would be seen to be unstable. A viable way to search for realistic scenarios of naked singularities, is to appeal to perturbations of the Schwarzschild/Kerr solutions. If a singular solution were found this way, then it would suggest that cosmic censorship is an artifact of spherical symmetry. However no solutions of perturbed Schwarzschild or Kerr spacetimes have been found which have naked singularities. And it has been shown that, consistent with the CCC, that static black holes are stable so that small perturbations to the black hole are dampened, the event horizon remains intact, and the singularity censored. While some violations have been found, they require delicate conditions and in this sense are not realistic. [18].
1.2 5D Schwarzschild solutions
Although classical GR is able to correctly predict physical phenomena on a macroscopic scale, several modifications to the theory have been proposed, with the aim of having a theory consistent with quantum mechanics (a theory of quantum gravity). Some of these such as string theory include more than three spatial dimensions, and these additional spatial dimensions are often taken to be
3 compact — small, toroidal (doughnut-shaped) dimensions which, while not easily observed, are desirable for the theory. The inclusion of higher dimensions give rise to black objects with distinct physical features, not present in 4D. This is a consequence of the different rotational dynamics afforded by new dimensions, and the freedom to extend 4D horizons in higher dimensions. Several higher dimensional black objects are possible mathematically and for this larger class of solutions, it is not clear that stability and cosmic censorship, present in 4D, hold. Next, we give the simplest higher dimensional black hole solution — the extension of the Schwarzschild black hole – and subsequently review how it has been shown to be unstable.
In 4D outside of a spherically symmetric distribution of energy of total mass M, the metric is given by the Schwarzschild solution,
r ds2 = −V (r)dt2 + V (r)−1dr2 + r2dΩ2,V = 1 − Sch , (1.2) Sch Sch 2 Sch r
2 2 2 2 where rSch = 2M4 is the location of the event horizon, dΩ2 = dθ + sin (θ)dφ , and units are such that c = G4 = 1. For a given mass, M4, the solution is unique.
The static solution to the vacuum EFE which maintains the 4D symmetry gives one of the simplest higher dimensional black holes, and it is an extension of the Schwarzschild black hole to one additional spatial dimension, 5D. It is symmetric in the 4D radial coordinate but, unlike the
Schwarzschild solution, it is not unique. The spacetime can be either spherically symmetric in the 5th dimension — a hyperspherical black hole — or it can be taken as independent of the 5th dimension
— a black string.
In 5D we can either use spherical or cylindrical coordinates. For the black string we simply add the dz2 to the line element in equation (1.2) and have an extension of the Schwarzschild metric which is translationally invariant in the fifth dimension z. In this case cylindrical coordinates (t, r,
θ, φ and z) are more convenient.
r ds2 = −V (r)dt2 + V (r)−1dr2 + r2dΩ2 + dz2,V = 1 − Sch . (1.3) Sch Sch 2 Sch r
4 The event horizon is the same as the Schwarzschild black hole, rSch = 2M4, where M4 is the mass in a z = constant slice of the spacetime. As the metric is independent of z this trivially satisfies the vacuum field equations because they are satisfied on each 4D (z = constant) slice. Because the hyperspherical black hole is symmetric in the 5th dimension, it is better expressed in spherical coordinates: an angular coordinate χ and a radial coordinate R, which are related to the cylindrical coordinates by r = Rsin(χ), z = Rcos(χ) and R2 = r2 + z2. The metric of the hyperspherical black hole is, R2 ds2 = −V (R)dt2 + V (R)−1dR2 + R2dΩ2,V = 1 − BH , (1.4) BH BH 3 BH R2
2 2 2 2 2 2 where dΩ3 = dθ + sin (θ)dφ sin (χ) + dχ .
This non-uniqueness leads to the question of stability: if two static configurations are possible, and a black object in one state is perturbed, what is there to prevent it from transitioning to the other state? As we will discuss next, the stability of the 5D Schwarzschild black hole (string) depends on the ratio of the length of the black object to its mass, L/M. This has been studied in a compact 5th dimension, where it is supposed that the black string extends the entire 5th dimension, forming a closed loop. In this context a long, thin black string may become unstable (large L/M) as may a heavy, hyperspherical black hole in a small compact dimension (small L/M)). The former is analogous to a thinning stream of tap water; the evidence of the instability is the transition of the continuous stream to beads. The later is as a large sphere constrained to a small torus, the two poles are forced closer to one another, inevitably pulling together, with this deformation illustrating that the spherical configuration is unstable.
For later use, we note a few aspects about the properties of the metric in equation (1.4). It is a special case of a Schwarzschild-Tangherlini metric, a generalisation of the Schwarzschild metric into
D dimensions. The relationship between the mass of the black hole, M5, and the horizon radius is
[15], 2 3πRBH M5 = . (1.5) 8G5
The metric in equation (1.4) can also be written in a conformally flat form as,
R2 ds2 = −V (R)dt2 + Ψ2 dR2 + R2dΩ2 ,V = 1 − BH . (1.6) BH 3 BH R2
5 Where Ψ is the conformal factor, which by symmetry of the spacetime, depends only on R. To find the conformal factor we follow [12] and seek a Ψ(R) so that,
−1 2 2 2 2 2 2 2 VBH (r) dr + r dΩ3 = Ψ(R) dR + R dΩ3 .
−1 2 2 2 2 2 2 r For this to hold it must be that VBH (r) dr = Ψ(R) dR and r = Ψ(R) R , or Ψ(R) = R where
dr r p r and R are related by, dR = R VBH (r). Integrating and choosing the integration constant so that Ψ → 1 as R → ∞, the solution is, R2 Ψ = 1 + BH . (1.7) BH 4R2
For the black string metric in equation (1.3) we cannot find a conformal factor ΨBH in the same way.
1.3 Gregory-Laflamme Instability
As shown by the metrics given in equations (1.3) and (1.4), for a black object having the same mass (and no charge or angular momentum) there are two distinct solutions possible. However it could be that one solution is omitted on physical grounds: given a black hole and a black string of the same mass M, is it expected that the one that would actually result from gravitational collapse would be the one with greater entropy. The comparison of the entropy for a black hole and a black string was done by Gregory and Laflamme[7][6], and they found a critical value of L/M, above which a hyperspherical black hole is entropically preferred, and below which a black string is. Because black strings above the critical length could attain higher entropy as their black hole counterpart, it was conjectured and shown perturbatively that black strings above this length are unstable. The entropy of a black hole is proportional to the surface area of the event horizon, which can be found from integrating the invariant volume element, dV = p|g|dxn, where g is the determinant of the metric. For a surface of constant radius (the horizon radius: rBS or RBH ) at constant time t,
4 2 6 2 4 |gBS| = r sin (θ) and |gBH | = R sin (θ)sin (χ) correspond to the areas,
Z π Z 2π Z L 2 2 ABS = rBS sin(θ)dθdφdz = 4πrBSL, 0 0 0 Z π Z π Z 2π 3 2 2 3 ABH = RBH sin(θ)sin (χ)dθdχdφ = 2π RBH . 0 0 0
6 The location of the horizon of the black string is the same as in 4D (except with G = G5): rBS = 2G5M4, as each z = constant slice is just a 4D Schwarzschild black hole. M4 is the mass contained in the 4D slice, so for a black string of length L the mass would be M = M4L. For the black hole, the mass is M = M5 with M5 that in equation (1.5). Thus the areas are:
M 2 A = 16π G2, BS L 5 r 8 A = 16πM 2G3/2 . BH 5 27πM
The black hole is entropically preferred (ABH > ABS) if,
L2 27πG > 5 . M 8
In the compact case, G5 = L and the inequality above is that in [6]. This suggests that a black string is unstable (since a black hole is entropically preferred) if the length of the string is large with respect to the mass or equivalently—as the mass is proportional to the radius —with respect to the radius so that long, thin black strings are unstable. This was shown in 1991 by Ruth Gregory and
Raymond Laflamme and is known as the Gregory-Laflamme instability. They showed the instability explicitly in by perturbing the metric in equation (1.3) by adding γαβdλ, where dλ is small[7, 8].
Requiring that gαβ also satisfy the vacuum Einstein Equations and solving for the perturbation,
Ωt+iµixi they found that γαβdλ ≈ e . If Ω > 0 the perturbation grows in time (t), and they found that Ω > 0 corresponds to long wavelength perturbations in the 5th dimension (see Figure 1.1). This is consistent with the entropy argument where strings above a critical length are unstable, as long wavelength perturbations can only propagate on correspondingly long strings.
While Gregory and Laflamme showed that an instability exists they only speculated on the end state of the unstable black string, and they conjectured it would bifurcate. In conjunction with their argument based on the entropy of the black string, their analysis was convincing and it was supposed that such a scenario would arise, and that the details of the bifurcation—and any implications for cosmic censorship—were waiting on a full theory of quantum gravity.
7 Interest in the classical problem was revived in 2001 when Horowitz and Maeda showed that bifurcation cannot occur in finite proper time along the horizon[10]. They argued that the more likely end state is a non-uniform string, but did not rule out that the horizon pinches off in infinite proper time along the horizon. Infinite proper time on the horizon may correspond to finite proper time for an asymptotic observer. If the horizon pinches off there will be a momentarily naked singularity, and if this is observed asymptotically, violates cosmic censorship: the event horizon ensures that the predictability of the asymptotic spactime is not compromised by the singularity, and if this is lost — even momentarily — GR fails to model the spacetime.
Since then the full non-linear evolution of the Einstein Field Equations showed that an unstable black string evolves into a chain of hyperspherical black holes attached by thing string-like necks[14].
The thin necks are subject to the Gregory-Laflamme instability as well, and a fractal structure develops with successive generations of unstable necks. By looking at the change in string thickness from generation to generation, it was argued that the thickness does reach zero, and the horizon bifurcates. [14]
Figure 1.1: Plot of Ω versus wavenumber scaled by the horizon radius (r+). This shows that there are unstable modes (Ω > 0) for large wavelength (small m). Figure reproduced from [6].
1.4 Asymptotically flat black string
The evolution in [14] was done in a compact fifth dimension. Strictly, the CCC applies to asymp- totic observers — those infinitely far away. Because the dimension is compact, it is not possible to have an asymptotic observer along the z-direction. In this sense, a black string bifurcating in a compact dimension could be argued to not constitute a violation of cosmic censorship. Here we consider the non-compact scenario: an elongated, but asymptotically flat, black string extended in
8 the 5th dimension. The expectation is that the fractal structure found in [14] is not an artifact of the compactness but a generic outcome of the instability. If this is the case, it should arise in the non-compact dimension as well.
Our objective in the next chapters will be to construct consistent data on a given hypersurface of the spacetime which represents an AFBS. In our case, the spacetime will have four spatial dimensions and one time, and we construct data on a slice of the spacetime which is a hypersurface of constant time. If such data is sufficiently similar to that for a black string in a compact dimension, our results will suggest that the end state which has been conjectured for the compact case is likely to hold for the AFBS. Unlike the compact black string, the AFBS is (topologically) a S(3) black hole. We aim to find data for the AFBS which, far from the ends, will resemble a black string.
However, the AFBS is intuitively not a static solution to the EFE: because it is topologically a highly distorted hyperspherical black hole, it is expected to contract to a hyperspherical black hole.
However if the AFBS is sufficiently long, we expect that the black string will be unstable and tend more rapidly towards a bifurcation of the horizon rather than a contraction to a hyperspherical black hole. That this is the case can be seen by nothing that the time for the AFBS to contract would be proportional to L (linear), but the time for the instability is exponential since the perturbation goes like eΩt. Therefore, while we do not know the exact length at which the instability precedes the contraction it is natural to expect that, in principle, a long enough string can be constructed.
In Chapter 2, we discuss some notation and definitions which underly some of the methods used in Chapters 3 and 4. In Chapter 3, the Cauchy decomposition (Section 3.1) is used to establish constraint equations that the gravitational field must satisfy on the hypersurface, and from these we obtain initial data for the metric of the AFBS. Simplifying assumptions are made along the way
(which are described in Section 3.2) so that the constraints reduce to a single equation which, when supplied with an energy density, determines the metric of the AFBS. While analytic solutions can be constructed for a few specific density distributions, ultimately we rely on numerical methods to solve for generic cases. The methods we use are described in Sections 3.2.1 and 3.2.2. Finally, in order to assess whether the data found resembles a black string, we find the shape of the apparent horizon (Section 4.2) and determine the characteristics of curvature invariants evaluated on the horizon (Section 4.3).
9 Chapter 2
Notation and Definitions
In Chapter 1 we motivated our interest in an asymptotically flat black string — a black object in a noncompact 5th dimension which is analogous to the previously studied black string in a compact
5th dimension. Our objective is to study asymptotically flat black strings (AFBS) by constructing consistent data on a spacelike hypersurface in the spacetime. A novel extension of standard methods are introduced to solve numerically for the constraint on such hypersurface, and for the apparent horizon. In this chapter we describe the notation, general methods, and some useful relationships which are used in Chapters 3 and 4. These chapters describe the problems we address in this work. In particular, we will show in Section 3.2 that, under certain conditions, the constraint on the hypersurface reduces to a second order nonlinear equation for a function of two variables – a radial coordinate, r, and a coordinate along the 5th dimension, z. Similarly, when looking for an apparent horizon (Section 4.2) we solve another second order equation for a different function of r and z.
Here we define some notation and operations which we will use in both methods, and we do so in terms of a general function which we denote by χ(r, z). We introduce a discrete approximation to χ(r, z) (Section 2.1) which allows us to write the PDEs as algebraic equations. General matrix operations which we use are defined in Section 2.2, and differentiation is done using discrete derivative operators (Section 2.3). By writing the PDEs algebraically we are able to solve them using Newton’s
Method (Section 2.4). We find that, under some conditions, the equations used in implementing
Newton’s method are of the form of the Sylvester equation (Section 2.5). In order to take advantage of existing methods developed for solving the Sylvester equation, additional relationships are worked
10 out which help us to see how, and under what conditions, the equations can be written this way
(Sections 2.6–2.8).
2.1 Discretisation
We formulate an algebraic problem by considering the unknown function χ(r, z) over a discretised domain. We are interested in finding χ over a real domain (rmin, rmax) × (zmin, zmax) and we solve for the discrete values of χ represented by an nr × nz grid. Explicitly, the discrete approximation to χ(r, z) is the nr × nz matrix which we denote by X,
χ(r , z ) χ(r , z + ∆z ) ··· χ(r , z ) min min min min j min max χ(rmin + ∆ri, zmin) χ(rmin + ∆ri, zmin + ∆zj) χ(rmin + ∆ri, zmax) X = . . . . .. . . . . χ(rmax, zmin) χ(rmax, zmin + ∆zj) ··· χ(rmax, zmax)
We define the coordinates so that the origin (r = 0, z = 0) lies at the centre of the distribution of density. As a result, both the functions that χ(r, z) will represent have most variation in the vicinity of the origin. We do this because otherwise the points near the central boundary of X are subject to larger numerical error than the bulk points. This is a result of having to use less accurate approximations to derivatives — forward and backward difference equations instead of central, as shown below— at the boundary. To avoid having the part of the domain where χ(r, z) varies most correspond to where the derivatives of X are determined with less accuracy, we extend r to negative values and let rmin = −rmax. One additional, but important, added benefit is that this automatically preserves regularity of the solution at r = 0. We also adopt zmin = −zmax, and so the origin corresponds to the maximum of the density. This increases our domain, and the number of points representing X that we need to solve for, by a factor of 4. While it would be quicker to solve on the smaller domain, we use the larger one because we can be more confident that numerical issues associated with the boundary are not influencing the solution at the interior. In the case of the
Hamiltonian constraint, far from the source the solution χ(r, z) ≈ r−p for some positive power p, and so by placing boundaries far from the source we ensure that the forward and backwards derivatives are not introducing significant error. For the function used in finding the horizon we are not able to
11 use any boundary conditions, and so it is particularly important that the issues associated with the boundary are far removed from the region of interest.
For the same reason — anticipating the region of interest to be near the origin — we adopt a grid spacing which is non-uniform and gives higher resolution (smaller spacing) near the origin. The spacing ∆ri and ∆zj is such that within some region around the origin a fine spacing is used, and outside a coarse spacing is used:
fine ∗ ∆r |r| ≤ r ∆ri = coarse ∗ ∆r |r| > r
fine ∗ ∆z |z| ≤ z ∆zj = coarse ∗ ∆z |z| > z
This allows us to obtain high resolution where it is most needed, while otherwise keeping nr and nz as small as possible (as compared to the alternative of using ∆rfine and ∆zfine everywhere because we need it in the interior). The derivative matrices, defined below, are modified accordingly around these cutoffs to account for the abrupt change in ∆r or ∆z. With this in mind, we employ a coarse spacing which is an integer multiple of the fine spacing as this ensures that points on the fine grid can be treated as an extension of the coarse grid as necessary.
In formulating the algebraic problem we need to arrange the unknowns — the nr × nz elements of X — in vector form. We define the vector form of X as the nr · nz × 1 vector,
T X˜ = . X1,1, X2,1, ··· , Xnr ,1, X1,2, X2,2, ··· , Xnr ,2, ··· , X1,nz , X2,nz , ··· , Xnr ,nz
In what follows X˜ denotes the vector form of a matrix X defined as above — effectively by “stacking columns” instead of rows.
12 2.2 Matrix Operations
The operation A◦B represents the Hadamard (element-wise) product of matrices A and B (pro- vided they are of the same size), and A B Hadamard (element-wise) division. For an m×n matrix
A, A ⊗ B is the Kronecker product,
a B ··· a B 11 1n . .. . A ⊗ B = . . . . am1B ··· amnB.
Basic properties of the Kronecker product are available in [19], and a more extensive discussion, with some proofs, is given in [2]. For later reference we list a few properties here. The inverse of a
Kronecker product is the Kronecker product of the inverses:
(A ⊗ B)−1 = A−1 ⊗ B−1. (2.1)
The transpose of a Kronecker product is the Kronecker product of the transposes:
(A ⊗ B)T = AT ⊗ BT . (2.2)
A function applied to a Kronecker product (of this form) may be applied to the argument instead
[2],
f (A ⊗ 1) = f (A) ⊗ 1. (2.3)
Similarly, f (1 ⊗ B) = 1 ⊗ f (B), but this does not hold generally for f (A ⊗ B). For matrices of compatible sizes,
(A ⊗ B)(C ⊗ D) = (AC ⊗ BD) . (2.4)
When formulating the PDEs as algebraic problems, the Kronecker product arises naturally due to the block structure of the matrix needed to differentiate the vector X˜ (shown in the next section).
A very useful property of the Kronecker product, which we use frequently, is the following:
Y = AXB ↔ Y˜ = BT ⊗ A X˜ . (2.5)
13 Note that the first equation contains all matrices, whereas in the second X and Y are in their vector forms, X˜ and Y˜ . Equation (2.5) is used to switch between an algebraic problem with respect to the vector X˜ , and the equivalent one with respect to the matrix X. That equation (2.5) holds is easily understood in the context of differentiation, as shown below, but a general proof is given in [2]. The
Kronecker sum is an mn × mn matrix defined in terms of an n × n matrix P and m × m matrix Q as,
P ⊕ Q = 1n ⊗ Q + P ⊗ 1m. (2.6)
The Kronecker sum satisfies1,
exp(P ⊕ Q) = exp(P) ⊗ exp(Q). (2.7)
It has been shown that the eigenvalues of P ⊕ Q can be expressed in terms of those for P and Q.
From theorem T2.14 in [2], “βk ⊗ αi is an eigenvector of P ⊗ Q with eigenvalue αiβk and is also
2 an eigenvector of [P ⊕ Q] with eigenvalue αi + βk.” where βk is an eigenvalue of P and αk is an eigenvalue of Q.
2.3 Derivatives
To differentiate, derivative operators composed of the appropriate finite difference coefficients are defined [20]. Recall that we have defined the matrix representing χ(r, z) as having r along the
first dimension and z along the second. While this choice is arbitrary, it is the case that the first dimension of the matrix represents a particular spatial dimension and the choice must be kept in mind when taking derivatives.
As an example of what we mean by a derivative operator or matrix, the fourth-order accurate,
first derivative operator for a general n × q matrix — with the variable being differentiated changing
1by exponentiating (2.6) and using (2.3) and (2.4) 2There is a typo in [2], where the bracketed term was P ⊗ Q but should read, as written above, P ⊕ Q[3].
14 1 along the first dimension — is the n × n matrix dn,
− 25 4 −3 4 − 1 0 ··· 0 12 3 4 . 1 1 . − 0 0 ··· . 2 2 1 2 2 1 12 − 3 0 3 − 12 0 0 1 − 2 0 2 − 1 0 ··· 0 12 3 3 12 1 −1 . .. . dn = diag(∆) × . . . . ··· 0 1 − 2 0 2 − 1 0 12 3 3 12 1 2 2 1 0 − 0 − 12 3 3 12 . . 1 1 . 0 0 0 − 2 0 2 1 4 25 0 ··· 0 4 − 3 3 −4 12 where diag of a vector denotes a diagonal matrix which has the vector along the diagonal and × is regular matrix multiplication. In this case, diag(∆)−1 is the matrix with the inverse grid spacing
1 , 1 , ..., 1 on the diagonal. To see that this operator will work, note that when applied to a ∆1 ∆2 ∆n vector X˜ , − 25 X + 4X − 3X + 4 X − 1 X /∆ 12 1 2 3 3 4 4 5 1 1 − X1 + X3 /∆ 2 2 1 2 2 1 12 X1 − 3 X2 + 3 X4 − 12 X5 /∆ 1 . d X˜ = . . n 1 2 2 1 Xn−4 − Xn−3 + Xn−1 − Xn /∆ 12 3 3 12 − 1 X + 1 X /∆ 2 n−2 2 n−1 1 4 25 4 Xn−4 − 3 Xn−3 + 3Xn−2 − 4Xn−1 + 12 Xn /∆
The result is a vector containing all of the derivatives at each point, as represented by the finite difference approximations with forward (backward) coefficients being used for the first (last) element, and lower order accuracy used for the second and second-to-end points (second order accurate). We have left the spacing (∆) general in the above expressions, but, as mentioned, we will use only four distinct values: ∆rfine, ∆rcoarse, ∆zfine and ∆zcoarse. Notice that, as defined, the derivative matrices require the same spacing for each row – for example, in the first row above ∆ must be the spacing between X1 and X2 as well as the spacing between X2 and X3 and so on — and when the
15 th ∆coarse coarse and find grids meet we will use the coarse derivatives and evaluate every ∆fine point where necessary to ensure the same spacing is used for each row. We are careful to ensure that where the grids overlap corresponds to a region that we expect to be relatively flat, as to minimise any error introduced by the abrupt transition of grid spacing. When solving for the conformal factor this expectation is based on the magnitude of the density distribution, and when solving for the apparent horizon it is based on the magnitude of the gradient of the conformal factor.
2 2 −1 A second derivative operator dn is defined similarly, using appropriate coefficients and diag(∆ ) in place of diag(∆)−1, as well as operators for second order accurate derivatives. Because X is de-
fined with r along the first dimension, left multiplying a derivative matrix with X approximates the
first (d1 X) or second (d2 X) derivatives with respect to r. The first and second derivatives with nr nr respect to z are (d1 XT )T = Xd1 T and (d2 XT )T = Xd2 T, as the derivative operators must be nz nz nz nz applied to the transpose of X, XT, and then the result transposed to be have consistent coordinates with X.
Derivative operators used for the vector X˜ are defined differently, as X˜ has elements along both r and z in the first dimension. The second derivative along r is,
2 d 0n ······ 0n nr r r . 0 d2 . nr nr 2 2 . .. . D = 1nz ⊗ d = . . . , nr nr . . 2 . dn 0nr r 0 ······ 0 d2 nr nr nr where 0 is an n × n matrix of zeros, and the first derivative is D1 = 1 ⊗ d1 . nr r r nr nz nr
To differentiate along z, X˜ first must be permuted (just as, in the matrix case, X was transposed).
The permutation matrix which achieves this is the sparse matrix P with non zero elements as
16 indicated schematically below.
1 2 ··· nr +1 nr +2 ··· (nz −1)nr +1 (nz −1)nr +2 ··· nr ∗nz 1 1 0 ········· 0 2 0 0 1 . . . . . . nz 0 1 nz +1 0 1 P = . nz +2 0 1 . . . . . . 2n 0 1 z . . .. . . . . . nr nz 0 ······ 0 ··· 1
The first and second derivatives with respect to z are PTD1 P X˜ and PTD2 P X˜ . While the nz nz same result could be achieved by transformations between the desired vector and the matrix repre- sentation of the z-derivatives of X, the advantage of explicitly defining everything with respect to X˜ is that to implement Newton’s method we will need to identify a Jacobian. This is straightforward if we can distribute all derivative operators to the left of X˜ .
2.4 Newton’s Method
Newton’s method is an iterative method to solve for the roots of an equation or a system of equations[9]. It also enables a nonlinear system of equations to be replaced by a sequence of linear equations. For any function F : Rn → Rm of an n × 1 vector x, the system of m equations,
F(x) = 0, is amenable to Newton’s method. It is not generally possible to solve for x in one step (unless it is linear), and this method requires multiple iterations yielding an approximate or trial solution xk at
∗ ∗ each step k, for k = 1, 2, ..., kmax. Letting x be the true solution (in the sense that F (x ) = 0), the
∗ ∗ solution converges to x if the trial solutions xk approach x as the iteration number k increases.
17 The method requires an initial guess x0, and whether or not the solution converges can depend on
∗ how close x0 is to x .
Given an initial guess x0, Newton’s method provides the incremental change, ∆x, that should be added to the guessed value to obtain a better approximation of x∗. It is based on the Taylor expansion of F around x,
F (x + ∆x) ≈ F (x) + J(x)∆x, where J(x) is the Jacobian of F (x),
∂F1 ∂F1 ... ∂F1 ∂x1 ∂x2 ∂xn ∂F ∂F ∂F 2 2 ... 2 ∂x1 ∂x2 ∂xn J(x) = . . . . . . ... . ∂Fm ∂Fm ... ∂Fm ∂x1 ∂x2 ∂xn
With the goal of finding the solution x∗, for any trial solution xk, Newton’s method prescribes that a better estimate will be x = xk + ∆x for which F (xk + ∆x) = 0, or, from above,
x = xk + ∆x, ∆x = −J(x)−1F (x). (2.8)
If the function F is linear, then the Jacobian above does not depend on x and it will be the case that x∗ = x0 + ∆x and the equation is solved in one iteration. Otherwise, for a nonlinear equation, this is repeated until an x is found which is close enough to the unknown x∗, in the sense that F (x) is sufficiently close to zero. We assess this convergence by monitoring the norm of F at each iteration k, 1 m 2 X 2 L2 F (xk) = Fi(xk) , i=1
∗ and consider an xk to be the solution x when LF (xk) < for a set error tolerance .
Notice however that to solve for the step ∆x by inverting the Jacobian in equation (2.8) would be computationally costly and another method for solving the linear system,
J(x)∆x = F (x),
18 is often used. These include iterative methods (such as Gauss-Seidel) in which the Jacobian is decomposed in a convenient way so as to only require solving triangular matrices, requiring a much lower computational cost. Here, we use Newton’s method but solve J(x)∆x = F (x) by implementing a method used to solve Sylvester’s equation[21] which entails an even lower computational cost, particularly for large matrices.
2.5 Sylvester’s Equation
Sylvester’s equation (also known as a Lyapunov equation in the case B = AT) is a linear equation of the form,
AX + XB = −F, (2.9) for an m × m matrix A, n × n matrix B, and X being m × n. Rewriting equation (2.9) in vector form and using equation (2.5),
AX˜ + XB˜ = −F˜,
˜ ˜ ˜ AX1nz + 1nr XB = −F,
T ˜ ˜ 1nz ⊗ A + B ⊗ 1nr X = −F.
Recognising that the last line is the Kronecker sum, equation (2.9) is equivalent to,
BT ⊕ A X˜ = −F˜. (2.10)
To check that equation (2.9) has a unique solution X, we can check that the equation (2.10) has a unique solution X˜ . Inspecting equation (2.10), it is evident that the condition for a unique solution is that the matrix BT ⊕ A is invertible. Following [2] as in Section 2.2 the system in equation (2.10)
T has a unique solution if the eigenvalues of B ⊕ A are all non-zero. Letting βk denote the eigenvalue
T of B and αi denote those of A, we conclude that the solution exists as long as:
αi + βk 6= 0, ∀ i = 1..nr, k = 1..nz; (2.11)
T where nr and nz are the number of (possibly repeated) eigenvalues of A and B .
19 While equation (2.10) could be solved by inverting BT ⊕A, it is much quicker to solve in the form of equation (2.9) using the Bartels-Stewart algorithm [21, 1]. Decompose A and B using a Shur decomposition[22],
A = QUQ−1,
B = RVR−1, where U, V are upper triangular matrices and R, Q are orthogonal matrices. Substituting these expressions into equation (2.9) yields,
QUQ−1X + XRVR−1 = −F.
We can now multiply from the left by Q−1 and from the right by R,
UQ−1XR + Q−1XRV = −Q−1FR.
Defining Y = Q−1XR, this can be rewritten as,
UY + YV = −Q−1FR.
The left-hand side is equivalent to the Kronecker sum and so with respect to Y˜ the above is,
VT ⊕ U Y˜ = −Q−1˜FR. (2.12)
Equation (2.12) is similar to equation (2.10) but, since U and V are triangular, equation (2.12) is much faster to solve than equation (2.10). This algorithm is used by Matlab [4], which in turn employs SLICOT [11] and LAPACK [13] routines.
Notice that the same approach can not be taken for a general equation of the form,
AX + XB + CXD = −F, (2.13)
20 because there are three terms on the left hand side, except in the special cases below. However, under some conditions, the system can be made triangular. This is the advantage of the Bartels-Stewart algorithm, and so (in principle) may be solved as quickly as the system in equation (2.12).
In the case where C and D share orthogonal matrices with A and B (C = QWQ−1, D = QZQ−1 where W and Z are upper triangular), equation (2.13) becomes,
UY + YV + WYZ = −Q−1FR, Y = Q−1XR.
Or, in vector form, VT ⊕ U + ZT ⊗ W Y˜ = −Q−1˜FR. (2.14)
Since Z and W are upper triangular, ZT ⊗ W is triangular (or very nearly), and so a similar algorithm to the Bartels-Stewart would apply to equation (2.14). However, the Matlab algorithm which solves the Sylvester equation is substantially quicker than what we wrote ourselves, and so it would follow that an extension to solve (2.14) would be slow as well. This only works in the case that
C and D have the same orthogonal matrices in their Schur decompositions as A and B respectively.
While the Schur decomposition is not unique and so it is possible this could be achieved through the choice of Z and W, it is certainly not guaranteed for any C and D.
Another case of C and D, which is discussed more later, is that they are diagonal. In this case DT ⊗ C is guaranteed to be diagonal and the matrix VT ⊕ U + DT ⊗ C is triangular (it is redundant to decompose C and D in this case). Still, implementing the Bartels-Stewart algorithm ourselves turned out much more computationally expensive than the Matlab algorithm.
Lastly we note that the Bartels-Stewart algorithm, while quick, does not take advantage of the repeated structure of A ⊗ B. Rather it solves an mn × mn system by putting it into a triangular form. Equation (2.1) shows how easily a Kronecker product can be inverted but the same does not hold for a Kronecker sum. Note that if we could rewrite the Kronecker sum as a Kronecker product, we would simply have an m × m and n × n matrix to invert instead of the full nm × nm system to
21 solve. This would require solving for C and D where:
AX + XB = CXD,
1 ⊗ BT + A ⊗ 1 = CT ⊗ D.
However this has m2 + n2 unknowns and m2n2 equations and so it not generally solvable (although a least squares solution could be used). The opposite problem — solving for A and B given C and
D — is solvable (with some freedom in the choice of A and B) and this related problem is discussed at the end of Section 2.8.
Finally, while an explicit solution does not appear to exist in term of A, B and their inverses,
A ⊕ B has an inverse in terms of the eigenvectors of AB and their transposes: