3+1 Presentación

Lecture I: Einstein equations in 3+1
formalism
Luciano Rezzolla
Albert Einstein Institute, Potsdam, Germany
IMPRS Block lecture course, Wandlitz, Nov. 2008

Plan of the lectures
The goal of this set of lectures is to introduce the
mathematical and numerical methods that are used
when modelling sources of gravitational waves.
Lecture I: 3+1 decomposition
Lecture II: gauges, initial data and GW-extraction
Lecture III: relativistic hydrodynamics
Lecture IV: modelling gravitational collapse
Lecture V: modelling compact binaries
Note:
! there is more to be said on each of these topics: just ask!
! no question is a stupid question!
Plan of lectures I, II
Numerical relativity: Why?
Numerical relativity: Why so hard?
Numerical relativity: How?
3+1 decomposition (lecture I)
vacuum spacetimes (lecture I)
Einstein equations (lecture II)
Gauge conditions (lecture II)
Initial data (lecture II)
GW-extraction (lecture II)
Numerical Relativity: why?
Among other things, numerical relativity aims at:
solve Einstein equations without approximations(!)

solve the binary problem(s)
investigate the complex physics of gravitational collapse
investigate the formation and dynamics of horizons
investigate structure and stability of NSs
modelling sources of gravitational waves
Modelling source of GWs
A simple, back-of-the-envelope calculation in the
Newtonian quadrupole approximation shows that the
luminosity in gravitational waves (energy emitted in gws
per unit time)!is " ! " ! " ! "
2 2 5
G M !v " G M
Lgw = #
c5 c5 R
i.e. intense sources are compact, massive and move at
relativistic speeds: general relativity is indispensable.
What makes GW-astronomy
! " challenging is
G
! 3.8 10 60
erg s 1
c5
i.e. even the GWs from the most intense sources will
statistically reach us as very weak
Not just an academic exercise
The calculation of the waveforms is not just an academic
achievement. Several millions !s and thousands man-hours are
dedicated to one of the most challenging physical experiments.
Knowledge of the
waveforms can
compensate for the
very small S/N
(matched-filtering).
enhance detection
and allow for source-
characterization
possible.
Numerical Relativity: why so hard?
" No obviously better formulation of the Einstein equations
ADM, conformal traceless decomposition, first-order hyperbolic, harmonic, ???
" Coordinates (spatial and time) do not have a specific meaning
this gauge freedom needs to be handled with care!
Which part of the spacetime to cover?
Consider the simplest black-hole spacetime in Cartesian
coordinates
future timelike infinity infinitely away
t timelike (finite) slice
singularity
horizon
spacelike (finite) slice
spacelike infinity infinitely away

0 2M x
Consider the simplest black-hole spacetime: its conformal
representation is given by a Carter-Penrose diagram
+
singularity i future timelike infinity
+
I future null infinity
n
izo
r
ho
0
0 i spacelike infinity

I past null infinity

i past timelike infinity
Consider the simplest black-hole spacetime: its conformal
representation is given by a Carter-Penrose diagram
+
singularity i
timelike (infinite) slice
+
I
0
0 i

I
spacelike (infinite) slice
i
Spacelike finite slices
Most common discretization of the spacetime. Reminescent of
fluid dynamics; introduces complications from outer boundary
+
singularity i timelike outer
boundary
+
I
0
0 i
very natural choice
to define initial data

and interpret results; I
outer boundary is
placed as far out as
possible i
Spacelike infinite (conformal) slices
Not common discretization of the spacetime: spacelike infinity
is included in the grid; requires suitable coord transformations
+
singularity i
outer boundary
always at
+ spacelike infinity
I
0
0 i
care needed for
treatment of
outgoing radiation;
I
removes need for
outer boundary

conditions i
Null (ingoing-outgoing) slices
Not common discretization of the spacetime. Works well in
1D but not employed in higher dimensions
+
singularity i
+
I future null infinity
0
0 i
very natural to
study radiation
(exact answer);
I past null infinity
specification of
initial data highly

non trivial i
spacelike-characteristic slices (CCE)
Combines advantages of spacelike slices with accurate
description of outgoing radiation; tested in 3D linear regimes
+ Cauchy-characteristic
singularity i
extraction (CCE)
+
I
0
0 i
very natural to study
radiation (exact

answer); specification I
of initial data simple;
matching can be important progress
cumbersome i done at AEI
ADM, conformal decomposition, first-order hyperbolic form, harmonic, ???
" Coordinates (spatial and time) do not have a special meaning
this gauge freedom need to be handled with care!
gauge conditions must avoid singularities
gauge conditions must counteract grid stretching
Choosing the right temporal gauge
Suppose you want to follow the
gravitational collapse to a bh and assume a
simplistic gauge choice (geodesic slicing):
That would lead rapidly ( ) to a

code crash! No chance of measuring gws!
Need to use smarter gauges!

Time must advance at different rates at
different positions in the grid: singularity
avoiding slicing (e.g. maximal slicing).
Some chance of measuring gws!

" Einstein field equations are highly nonlinear
essentially unknown in these regimes (well-posedeness not enough!...)
" Physical singularities are the butter-and-bread of NR
delicate techniques are needed to excise the troublesome region
excising parts of the spacetime with singularities
apparent horizon found on a given !t
In principle, the yellow region is

causally disconnected from the blue
one (ligth cones are tilted in); no
boundar y conditions would be
needed at the apparent horizon.
In practice, the actual excision region
(legosphere: black region) carved
well inside the horizon.
NOTE:
o the Einstein equations are highly nonlinear in
the yellow region! All sorts of numerical
problems...
o the (apparent) horizon must be found; this is
an expensive operation
Images by D. Pollney o the excised region has to move on the grid

" Einstein field equations are highly nonlinear
essentially unknown in these regimes (well-posedeness not enough!...)
" Physical singularities are the butter-and-bread of NR
delicate techniques are needed to excise the troublesome region
" Simply more equations to solve: stretching supercomputers resources!

large turn-around times make progress slow (2-3 weeks/simulation)
implementations of AMR techniques is extremely problematic
Numerical Relativity: how?
Lets recall the equations we are dealing with:
In the simplest scenarios it is possible to consider the spacetime

as vacuum: T = 0
In this case the above eqs reduce to:
First step: foliate the 4D spacetime
Given a manifold describing a spacetime with 4-metric we
want to foliate it via space-like, three-dimensional hypersurfaces,
i.e., leveled by a scalar function. The time coordinate t
is an obvious good choice.
Define therefore
such that
This defines the lapse function which is

strictly positive for spacelike hypersurfaces
The lapse function allows then to do two important things:
i) define the unit normal vector to the hypersurface
where
ii) define the spatial metric

Second step: decompose 4-dim tensors
provide two useful tools to decompose any 4-dim.
tensor into a purely spatial part (hence in ) and a purely timelike
part (hence orthogonal to and aligned with ).
The spatial part is obtained after contracting with the spatial
projection operator
while the timelike part is obtained after contracting with the timelike
projection operator
where the two projectors are obviously orthogonal

It is now possible to define the 3-dim covariant derivative of a
spatial tensor. This is simply the projection on of all the indices
of the the 4-dim. covariant derivative
which, as expected, is compatible with spatial metric
All of the 4-dim tensor algebra can be extended straightforwardly

to the 3-dim. spatial slice, so that the 3-dim covariant derivative
can be expressed in terms of the 3-dimensional connection
coefficients:
Similarly, the 3-dim Riemann tensor associated with is defined
via the double 3-dimensional covariant derivative of any spatial
vector , ie
where
and
More explicitely, the 3-dim Riemann tensor can be written in
terms of the 3-dim connection coefficients as
Also, the 3-dim contractions of the 3-dim Riemann tensor, i.e. the
3-dim Ricci tensor the 3-dim Ricci scalar are respectively given by
It is important not to confuse the 3-dim Riemann tensor
with the corresponding 4-dim one
is purely spatial (spatial derivatives of spatial metric )
is a full 4-dimensional object containing also time

derivatives of the full 4-dim metric
The information present in and missing in can be

found in another spatial tensor: the extrinsic curvature.
As we shall see, this information is indeed describing the time
evolution of the spatial metric
Several equivalent definitions of the extrinsic curvature are possible:
i) in terms of the unit normal to
ii) in terms of the acceleration of normal observers
iii) in terms of the acceleration of normal observers
where is the Lie derivative along

Properties of the Lie derivative
I recall that the Lie derivative can be thought of as a geometrical
generalization of a directional derivative.
For a scalar function this is given by:
For a vector field , this is given by the commutator:
For a 1-form ,this is given by:
As a result, for a generic tensor of rank this is given by:

More geometrically, the extrinsic curvature measures the changes in
the normal vector under parallel transport
Hence it measures how the 3-dim
hypersurface is bent with respect to
the 4-dim spacetime
Later on we will discuss also a
kinematical interpretation of the
extrinsic curvature in terms of the
spatial metric
Consider a vector at one position P and then

parallel parallel-transport it to a new location P + P
transport
The difference in the two vectors is proportional to
the extrinsic curvature and this can either be positive
or negative
Some properties of the extrinsic curvature
i) it is a spatial tensor by construction (obtained with contractions

of the spatial metric)
ii) it is a symmetric tensor by construction (obtained with

symmetric derivatives of the unit normal
iii) its trace is given by
since the acceleration of the normal observer is orthogonal to

the unit normal:
Third step: decompose Einstein equations
Next, we need to decompose the Einstein equations in the spatial
and timelike parts.
and to do this we need to define a few identities

First we decompose the 4-dim Riemann tensor
projecting all indices to obtain the Gauss equations
Next, we make 3 spatial projections and a timelike one to obtain

the Codazzi equations
Finally we take 2 spatial projections and 2 timelike ones to
obtain the Ricci equations
where the second derivative of the lapse has been introduced

via the identity
Another important identity which will be used in the following is
and which holds for any spatial vector

Next we consider the projections of the stress-energy tensor
since this will be relevant when considering the right-hand-side
of the Einstein equations.
We recall that for a perfect fluid this is given by
where e, p, h, and are the energy density, the pressure, the

specific enthalpy h =e+p and the rest-mass density of the fluid.
Since , (the two vectors are parallel and unit vectors)
the energy density measured by the normal observers will be
given by the double timelike projection
Similarly, the momentum density (i.e. mass current) will be given by

the mixed time and spatial projection
Note that so far we have just learnt how to decompose the 4-
dim Riemann and stress-energy tensors and have not yet
discussed how to deal with the equations that regulate the
evolution of Riemann tensor, i.e. the Einstein equations.
As we will see soon, the 3+1 splitting naturally splits the

Einstein equations in a set which is fully defined on each spatial
hypersurfaces (and does not involve therefore time derivatives)
and a set which instead relates quantities (i.e. the spatial metric
and the extrinsic curvature) between two adjacent
hypersurfaces.
The first set is usually referred to as the constraint equations,

while the second one as the evolution equations
The constraint equations (I)
We first time-project twice the left-hand-side of the Einstein
equations to obtain
Doing the same for the right-hand-side, using the Gauss eqs
contracted twice with the spatial metric and the definition of the
energy density we finally reach the form of the Hamiltonian
constraint equation
Note that this is a single elliptic equation (hence not containing

time derivative) which should be satisfied everywhere on the
spatial hypersurface
The constraint equations (II)
Similarly, with a mixed time-space projection of the left-hand-
side of the Einstein equations we obtain
Doing the same for the right-hand-side, using the contracted

Codazzi equations and the definition of the momentum density
we finally reach the form of the momentum constraint equations
which are also 3 elliptic equations.

The 4 constraint equations are the necessary and sufficient
integrability conditions for the embedding of the spacelike
hypersurfaces in the 4-dim. spacetime
Fourth step: find the direction for evolutions
Note that the unit normal to a spacelike hypersurface is not
the natural time derivative. This is because is not dual to the
surface 1-norm , i.e.
We need therefore to find a new vector along which to carry out

the time evolutions and that is dual to the surface 1-norm.
Such a vector is easily defined as
where is any spatial shift vector. Clearly now

Because the vector is dual to the 1-form , we are guaranteed
that the integral curves of are naturally parametrized by the time
coordinate.
Stated differently, all
infinitesimal vectors
originating on one
hypersurface would end up
on the same hypersurface
This is not guaranteed for
translations along
A more intuitive description of the lapse function and of the shift

vector will be presented once we introduce a coordinate basis
Note that is not necessarily timelike if the shift is superluminal
With this definition we can revise the Lie derivative along the unit
normal . Since
the Ricci equation we have encountered before:

can now be rewritten as
(*)
Once again, this a clear expression that the extrinsic curvature can
be seen as the rate of change of the spatial metric, i.e.
Finally, note that the Ricci equations (*) are definitions and not
pieces of the Einstein eqs, although this is sometimes confused
the evolution part of the Einstein equations
We are now ready to express the missing piece of the 3+1
decomposition and derive the evolution part of the Einstein eqs.
As for the constraints, we need suitable projections of the two sides
of the Einstein equations and in particular the two spatial ones, ie
Using the Ricci eqs one then obtains
where
Fifth step: select a coordinate basis
So far we have dealt with tensor eqs and not specified a coordinate
basis with unit vectors . Doing so can be useful to simplify
equations and to highlight the spatial nature of and
The choice in this case in very simple. We want:

i) three of them have to be purely spatial, i.e.
ii) the fourth one has to be along the vector , i.e.

As a result:
i.e. the Lie derivative along is a simple partial derivative
i.e. the space covariant components of a timelike vector are

zero; only the time component survives
i.e. the zeroth contravariant component of a spacelike vector

are zero; only the space components survive
Putting things together and bearing in mind that
Because for any spatial tensor the contravariant
components of the metric in a 3+1 split are
Similarly, since the covariant components are
Note that (i.e. are inverses) and thus they

can be used to raise/lower the indices of spatial tensors
We can now have a more intuitive interpretation of the lapse,
shift and spatial metric. Using the expression for the covariant 4-
dim covariant metric, the line element is given
Hence:
the lapse measures proper time
between two adjacent hypersurfaces
the shift relates spatial coordinates

the spatial metric measures distances between points on every

hypersurface
We can now have a more intuitive interpretation of the lapse,
shift and spatial metric. Using the expression for the covariant 4-
dim covariant metric, the line element is given
coordinate line
normal line Hence:
the lapse measures proper time
the shift relates spatial coordinates

the spatial metric measures distances between points on every

hypersurface
Recap: the steps taken so far...
Let us recap briefly what done so far:
First step: foliate the 4-dim spacetime in 3-dim spacelike
hypersurfaces leveled by a scalar function: the time coordinate.
This determines a normal unit vector to the hypersurfaces
Second step: decompose 4-dim spacetime tensors in spatial and

timelike parts using the normal vector and the spatial metric
Third step: rewrite Einstein equations using such decomposed

tensors
Fourth step: select a natural direction for the time evolution
Fifth step: select a coordinate basis and express all eqs

Summary of lecture I
The 3+1 splitting of the 4-dim spacetime represents an
effective way to perform numerical solutions of the Einstein eqs
In essence such a splitting amounts to projecting all 4-dim.
tensors either on spatial hypersurfaces or along directions
orthogonal to such hypersurfaces.
The 3-metric and the extrinsic curvature describe the
properties of each slice.
Two functions, the lapse and the shift, tell how to relate
coordinates between two slices: the lapse measures the proper
time, while the shift measures changes in the spatial coords.
Einstein equations naturally split into evolution equations
and constraint equations
Lecture II: gauge conditions, initial data
and gravitational-wave extraction
Luciano Rezzolla

The (ADM) Einstein eqs in 3+1
In such a foliation, we can write the Einstein eqs in the 3+1
splitting of spacetime in a set of evolution and constraint
equations as:
[6]
[6]
These are 12 hyperbolic, first-order in time, second-order
in space, nonlinear partial differential equations: evolution
equations
Similarly
Hamiltonian
Constraint (HC) [1]
Momentum
Constraints (MC) [3]
These are 6 elliptic (second-order in space), nonlinear

partial differential equations: constraint equations
All together we have:
[6]
[6]
[1]
[3]
These 6+6 (+3+1) eqs are also known as the ADM equations. In
practice only the evolution eqs are solved and the constraints are
instead monitored (more later)
ADM vs Maxwell
The ADM eqs may appear as rather cryptic and simply complicated.
However, it is easy to see analogies with the Maxwell eqs. and make
the equations less cryptic.
The relevant quantities in this case are the electric and magnetic
fields , the charge density and the charge current density
Then also the Maxwell equations split into evolution equations
and constraint equations

Also for the Maxwell eqs it is possible to show that if the constraints
are satisfied initially, then the evolution eqs preserve this property.
To further highlight the analogies lets introduce the vector potential
and so that the

Maxwell evolution
equations become
to be compared
with the ADM
evolution eqs
It is then possible to make the associations
and realize that the RHSs of the evolution equation of

involve a field variable and the spatial derivatives of a
gauge quantity
Similarly, the RHS of the evolution equation of involve
matter sources as well as second spatial derivatives of the
second field variable
Indeed, the similarities between the ADM eqs and the Maxwell
eqs written in terms of the vector potential (i.e. as in previous
slide) are so large that they suffer of the same problems/
instabilities (more later)
In practice, the ADM are essentially never used!
These equations are perfectly alright mathematically but not in

a form that is well suited for numerical implementation.
Indeed the system can be shown to be weakly hyperbolic and

hence ill-posed
In practice, numerical instabilities rapidly appear that destroy

the solution exponentially
However, the stability properties of numerical implementations

can be improved by introducing certain new auxiliary functions
and rewriting the ADM equations in terms of these functions.
To illustrate how to reach a form of the eqs that is suitable for a
numerical solution let us consider again the Maxwell eqs: they
are simpler and the logic is very similar. Lets start from
take a time derivative and use the evolution eq. for to

obtain
This would be a wave equation if not for the spoiling

term with mixed derivatives
In GR, the situation is very similar since contains mixed
derivatives in addition to a Laplace operator acting on
Without such mixed derivatives the 3+1 ADM equations
could be written with a principal part behaving as a wave
equation for
Why do we care to have wave equations?
Why do we care to have wave equations?
Wave equations are manifestly hyperbolic and mathematical

theorems guarantee the existence and uniqueness of the
solutions (more on this in lecture III).
Several numerical techniques, as those that we will
encounter in hydrodynamic, have been developed to solve
manifestly hyperbolic equations, e.g. 1st-order in time, 2nd-
in space reduction
Stated differently: we know how to solve wave equations

and what to expect.
How do we make the Maxwell eqs manifestly hyperbolic?
i) use a specific gauge, e.g. Lorentz gauge: so that

the eqs become simply
This can be done also in GR by introducing harmonic

coordinates and a generalized harmonic formulation of the
Einstein eqs. (more later)
ii) implement a gauge-invariant approach by taking a time
derivative of rather than of . This leads to
which, using the constraint reduces to

This approach is aesthetically attractive but may lead to
complications because the matter source term is
proportional to spatial derivatives of the charge density .
In GR this would correspond to derivatives of the rest-mass
density and may be divergent if a shock is present
iii) introduce a new variable to remove the mix-derivative
term, i.e. define so that the evolution eq. becomes
Clearly an evolution equations is required for which is a

new variable like all the others (just no physical meaning):
We have gained one more eq. but the system is hyperbolic!

The same is done for the ADM eqs and new evolution
variables are introduced to obtain a set of eqs that is
strongly hyperbolic
: conformal factor
ij: conformal 3-metric
K : trace of extrinsic curvature
Aij : trace-free conformal
extrinsic curvature
:Gammas
i
are our new evolution variables
The ADM equations are then rewritten as

Dt ij = 2Aij , where Dt t L
1
Dt = K ,
6
! "
TF
Dt Aij = e
4
[i j + (Rij Sij )] + K Aij 2Ail Alj ,
# $
1 2 1
Dt K = i j + Aij A + K + ( + S) ,
ij ij
3 2
! 2 ij "
Dt = 2A j + 2 jk A j K ij Sj + 6Aij j
i ij i kj
3
! 2 "
j l l ij 2 m(j m i) + ij l l .
3
These equations are also known as the BSSNOK equations
or more simply the conformal traceless formulation of the
Einstein equations.
Although not self evident, the BSSNOK equations are
strongly hyperbolic with a structure which is resembling the
1st-order in time, 2nd-order in space formulation
scalar wave equation
Conformal traceless
formulation
The BSSNOK equations are the most widely used form of

the Einstein eqs and have demonstrated to lead to stable
and accurate evolution of vacuum (binary black holes) and
non-vacuum (neutron stars) spacetimes
Recap
The hyperbolic evolution equations to be solved from one
time slice to the next are therefore: 6+6+(3+1+1)=17 (12).
We also compute 3+1=4 elliptic constraint eqs
NOTE: these eqs are not solved but only monitored to verify
5 additional constraints are introduced by new variables:

Generalized Harmonic formulation (GHC)
The conformal traceless (or BSSNOK) formulation is not the
only successful formulation of the Einstein eqs. Another
popular choice is the generalized harmonic formulation.
The Generalized harmonic formulation is based on a

generalization of the harmonic coordinates.
When such condition in enforced in the Einstein eqs, the

principal part of the equations for each metric element
becomes a scalar wave equation, with all nonlinearities and
couplings between the equations relegated to lower order
terms.
The harmonic condition is known to suffer from
pathologies (such as focussing and caustics). However, these
are gauge conditions and alternatives can be found that do
not spoil the nice properties of the harmonic formulation.
The generalized harmonic coordinates were recently proposed

by Friedrich in the form
where the a are a set of harmonic source functions.

Of course such a choice still provides Einstein equations with a
principal part that maintains a scalar-wave equation character
and can in principle remove the drawbacks of the harmonic
conditions if suitable source functions are found
Using the definition of GHC and its first derivative, the
Einstein equations in trace-reversed form
can be rewritten as
wave-like nonlinearities and couplings

principal part among different equations
isolated matter sources
This is very appealing but it does not show stable evolutions

unless a proper damping of the constraint is enforced
To compensate this, one needs to add adjustable constraint
damping terms of the type
so that the equation can be rewritten as
wave-like nonlinearities and couplings

principal part among different equations
isolated matter sources

constraint damping terms
This approach helps the solution to go back towards a

harmonic condition if it departs away from it.
NOTE: The source functions are effectively new variables
that require evolution equations of their own
Of course such evolution equations are arbitrary and the
complication (and this is a major one in the GHC formulation)
is to find the appropriate evolution equations for that yield
well-behaved coordinates.
A well-experimented choice is:
forcing term damping term

This is a damped wave equation with a forcing function
designed to prevent the lapse from deviating excessively
from the asymptotic value
All of this may sound heuristic but the choice of hyperbolic
gauge conditions is pretty much heuristic!
Gauge conditions
NOTE: the lapse, and shift are not solutions of the Einstein
equations but represent our gauge freedom, namely the
freedom (arbitrariness) in which we choose to foliate the
spacetime.
Any prescribed choice for the lapse is usually referred to as a

slicing condition, while any choice for the shift is usually
referred to as spatial gauge condition
While there are infinite possible choices, not all of them are
equally useful to carry out numerical simulations. Indeed,
there is a whole branch of numerical relativity that is
dedicated to finding suitable gauge conditions.
Several possible routes are possible
i) make a guess (i.e. prescribe a functional form) for the lapse,

and shift and hope for the best: eg geodesic slicing
obviously not a good idea
ii) fix the lapse, and shift by requiring they satisfy some
condition: eg maximal slicing for the lapse
which has the desired singularity-avoiding properties. Similarly,

the minimal distortion shift condition guarantees minimizes the
changes in the conformally related metric
Good idea mathematically (the coordinates do exactly what
they should). Unfortunately this leads to elliptic eqs which are
computationally too expensive to solve at each time
iii) fix the lapse, and shift dynamically by requiring they satisfy
comparatively simple evolution equations
This is the common solution. The advantage is the eqs for

the lapse and shift are simple time evolution eqs.
A family of slicing conditions that works very well to obtain

both a strongly hyperbolic evolution eqs. and for stable
numerical evolutions is the Bona-Masso slicing
where and is positive but otherwise

arbitrary.
Indeed the condition represents a family of
slicing conditions such that:
The 1+log slicing condition

also has excellent singularity
avoiding properties since
and hence the lapse remains
very small in those regions
where it has collapsed to
small values
Similarly, a popular choice for the shift is the hyperbolic
Gamma-driver condition
where acts as a restoring force to avoid large oscillations in

the shift and the driver tends to keep the Gammas constant
(reminiscent of minimal distortion)
Overall, the 1+log slicing condition and the Gamma-

driver shift condition are the most widely used both in
vacuum and non-vacuum spacetimes
Initial data
Einstein equations represent an initial-value problem (IVP).
Stated differently, once the solution is known/specified at any
initial time, the hyperbolic nature of the equations completely
determines the space of future solutions
t continuous spacetime t discretized spacetime
d d

space of relevant solutions
with initial data L n+1
d d
n
n1
L
L x j1 j j+1 x
Initial data
This is a concept you are surely familiar with as you have
encountered, for instance, in Newtonian dynamics/kinematics.
For any time evolution of the Einstein equations we need
therefore to specify initial data and immediately two distinct
problems arise:
i) determine a set of physical conditions that represent

physically realistic/plausible configurations
ii) solve the Einstein equations to represent such solutions. In

the vast majority of cases this amounts to a numerical solution
Computing the initial data
Exploiting what you know from other fields, you will not be
surprised if the specification of the initial data is made by
considering the system as under some equilibrium conditions.
This has the important advantage that the system of equations
needed to describe the system does not have a time
dependence, although the spacetime can be stationary.
A good example is the static equilibrium of a spherical star. Take
the Euler and Poissons equations in Newtonian gravity:
Assuming one obtains the simple differential eq.
pressure gradients balance

gravitational forces
The same logic can be applied pretty much also in GR.
One models initial data as describing a system in equilibrium,
thus removing any time dependence.
What does this mean in terms of the Einstein equations?
The same logic can be applied pretty much also in GR.
One models initial data as describing a system in equilibrium,
thus removing any time dependence.
What does this mean in terms of the Einstein equations?
Indeed the equations to be solved are the Hamiltonian and
momentum constraints:
In practice one does not solve them in this form but rather
use a conformally related metric which is, in particular flat
This introduces a number of simplifications in the equations

but also removes the gravitational-wave content from the ID
Twenty years and more have been spent on discussing whether
or not this is a good approximation
Numerical simulations show this is indeed so and the initial
error is rapidly dissipated in less than one orbit.
The calculation of initial data is a vast area of research

combining advanced methods (waveless approximation,
conformal thin sandwich, postNewtonian approximation and
matches).
All revolve around the solution of the constraint eqs and the
hydrostatic equilibrium equations for the matter sources. More
information in the Living Review by Cook (2004).
http://relativity.livingreviews.org/Articles/lrr-2004-5/
Extraction of gravitational waves
Wave-extraction techniques
Computing the waveforms is the ultimate goal of most
numerical relativity and there are several ways of extracting
GWs from numerical relativity codes:
- asymptotic measurements
null slicing
conformal compactification
- non-asymptotic measurements (finite-size extraction worldtube)
Weyl scalars
perturbative matching to a Schwarzschild background
All have different degrees of success and this depends on
the efficiency of the process which is very different for
different sources
Wave-extraction techniques
In both approaches, observers are placed on nested 2-spheres
and calculate there either the Weyl scalars or decompose the
metric into tensor spherical-harmonics to calculate the gauge-
invariant perturbations of a Schwarschild black hole
Once the waveforms

are calculated, all the
related quantities:
energy, momentum and
angular momentum
radiated can be derived
simply.
Lects. IV and V will
show some examples
Gauge-invariant pertubations
Similarly, at a sufficiently large distance from the source and
assuming the spacetime resembles that of a Schwarzschild BH
where are the odd and even-parity gauge-invariant

perturbations of a Schwarschild bh. The projection of the
momenutm flux on the equatorial plane is, for instance,
This quantity can be used, for instance, to calculate the recoil.

Weyl scalars
The Newman-Penrose formalism provides a convenient
representation for radiation-related quantities as spin-weighted
scalars. In particular, the component of the Weyl tensor
can be associated with the gravitational-radiation content of the

spacetime because it falls like ~1/r. Here is a null
frame and in practice we define an orthonormal polar
coordinate basis centred on the Cartesian grid origin
and with poles along .
Using then the normal to the slice as the components
of the frame are
Weyl scalars
We then calculate in terms of ADM related quantities as
where
Then at a sufficiently large distance from the source the GWs in

the two polarizations can be written as
Then, eg, the projection of the momentum flux on the equatorial

plane as
This quantity can be used, for instance, to calculate the recoil.

Lecture III: an introduction to
relativistic hydrodynamics
Luciano Rezzolla

Plan of this lecture
Calculations vs Simulations
The relativistic hydrodynamics eqs in 3+1
Nonlinear nature of the equations
The issue of the formulation
Numerical approaches
HRSC methods
Lets recall the equations we are dealing with:
In the simplest scenarios it is possible to consider the spacetime

as vacuum: T = 0
In this case the above eqs reduce to:
In non-vacuum spacetimes (like ours!) we need to solve for the
full set of equations:
where
Before entering in the details or relativistic hydrodynamics it
is useful and important to underline a fundamental
distinction
The description of vacuum spacetimes is mathematically and
physically complete. The full set of eqs is known and there is
no approximation whatsoever. This reflects the fact that black
holes are the simplest macroscopic objects in the universe:
fully determined in terms of their mass, spin, (possibly)
charge.
The only error is the numerical one and one could speak of
numerical relativity calculations (e.g. calculation of )
The description of non-vacuum spacetimes is mathematically

complete but physically limited. The full set of eqs is known
but approximations are made in description of matter
(stress-energy tensor, EOS, etc.). Its an approximation to
reality.
There is a numerical error but also a phenomenological

one. In this case one should speak of numerical relativity
simulations
Solving the hydrodynamics
equations
3+1 splitting also for the matter
We start again with a 3+1 split and thus a line element
And introduce a fluid (collection

of particles) with 4-velocity
And introduce a fluid (collection Note the difference between the

of particles) with 4-velocity normal unit vector to the slice
and the fluid 4-velocity . They are
both unit and timelike, i.e.
But they are really different: one is

tracks the normal to the slice the
other is the worldline of a fluid
particle

normal line
fluid worldlines both unit and timelike, i.e.

particle.

coordinate line
normal line fluid worldlines both unit and timelike, i.e.

particle. In addition there is always
the natural time vector
Overall, there are three 4-vectors to bear in mind:

What we are really interested in, however, is not the
4-velocity but rather its projection on the spatial
slice, ie the 3-velocity
Those observers with parallel
normal line
fluid worldlines to move from one slice to
the next along the normal to
the slice and are therefore
Eulerian observers.
They measure a fluid 3-velocity
Remember in fact that in special relativity

The contravariant (upstairs) components of this vector are
fluid worldlines
normal line
while the covariant (downstairs)

components are
Using the normalization condition

one obtains
Thus recognizing in the

Lorentz factor
The relativistic hydrodynamics(MHD) eqs simply express the
conservation of energy, momentum, baryon number
conservation of momentum
conservation of energy
conservation of baryon number

thermodynamics
where
Some representative examples: advection equation
Before looking at the solution of the hydrodynamical equations
there are some fundamental aspects of their nonlinear properties
which must be clarified.
The simplest linear hyperbolic
equation is the advection
equation
time
The solution is the initial data

simply translated in space and
time.
The propagation speeds are
constant everywhere (linear space
nature of the equation)
Some representative examples: Burgers equation
The simplest nonlinear hyperbolic equation is Burgers equation
where the RHS is assumed zero in the inviscid limit. Despite the
remarkable similarity, the solution to this eq. is remarkably different
time
space
Some representative examples: Burgers equation
This behaviour is referred to as shock steepening and is the
consequence that the propagation speeds are not constant
as for the advection equation but are a function of space and
time (nonlinear nature of the equation).
Stated differently, the maxima time

of the waves move faster
than the minima and tend to
catch-up.
NOTE: this is a property of
the equations and not of the
initial data. Even smooth initial
data will (eventually) shock in
space
inviscid fluids
Why should we care?
One may think that the shock steeping may happen

but on a timescale which is much longer than the
one we are interested in
Furthermore, we start from continuous initial data...
The problem of discretization
A generic problem arises when a Cauchy problem described by
a set of continuous PDEs is solved in a discretized form: the
numerical solution is, at best, piecewise constant
This is problematic when
discretizing hydrodynamical
eqs in compressible fluids.
As for Burgers eq., the
nonlinear properties
generically produce, in a
finite time, nonlinear waves
with discontinuities (ie
shocks, rarefaction waves,
etc) even from smooth data!
Discontinuous initial data
Burgers inviscid equation with discontinuous initial data offers a well-
known example of the importance of a proper writing of the equation.
Consider
with
The equation can then be

written as (light blue line)
or as (green line)
Mathematically equivalent but the numerical difference is between the

right answer (blue line) or a completely wrong one!
Conservative form of the equations
More generally, the homogeneous partial differential equation
is said to be in flux-conservative (FC) form if written as
Theorems (Lax, Wendroff; Hou, LeFloch)

FC formulation converges to the weak solution of the
problem (ie a solution of the integral form of the fc form)
NFC converges to the wrong weak solution of the problem
In conservative systems (as the hydrodynamic eqs) one usually
deals with a set of equations in FC form. Hence, the function
and the flux are replaced by a state vector and a flux
vector
The importance of a conservative formulation
If the problem is linear, the importance of a conservative
formulation is clear as it allows for analytic solutions.
Rewrite the flux conservative equations
as
where is the Jacobian matrix of constant

coefficients (the problem is linear).
Note that the state vector is indicated in bold because a
vector of components
We next diagonalize so that is the
diagonal matrix of eigenvalues of the N linear
equations, i.e.
are the set of right eigenvectors of or, equivalently,

the columns of the matrix .
Note that both steps are guaranteed to be possible by the
assumption we are dealing with a set of hyperbolic equations
Indeed, the set of eqs
is said to be (strongly) hyperbolic iff is diagonalizable with

a set of real (distinct) eigenvalues and correspondingly a
set of linearly independent (right) eigenvectors
Next we can define the characteristic variables
so that the original set of equations can be

written as
Because is diagonal, this is effectively a set of decoupled

ODEs along a set of specific curves in the plane
Stated differently
so that the characteristic variables are constant along those

curves in the plane having slope
Such curves are called characteristic curves and their slope is
locally given by the characteristics speeds
t
the characteristic speeds
x are not equal and
characteristics focus
characteristic curves
Because they remain constant along characteristics, the value

the characteristic variables at any time is known once the initial
one is determined, i.e.
Once the solution is known in terms of the characteristic
variables , it is simple to go back to the original state
vector
and hence
Stated differently, the solution at any time can be seen as the

linear superposition of waves, each propagating
independently at the speed given by the corresponding
eigenvalue
The Valencia (conservative) formulation
Consider for simplicity an non-magnetized ideal fluid with stress
energy tensor
where
As mentioned before, the hydrodynamics eqs are then given by

The Valencia (conservative) formulation
The first step in rewriting the above equations in a FC form
requires the identification of suitable conserved quantities in
place of the primitive variables . A little algebra
shows that these are:
where is the Lorentz factor

NOTE: while the conversion primitive-to-conserved is
algebraic, the inverse one is not and needs an expensive
numerical solution.
In this way one obtains the Valencia formulation (Banyuls et
al. 97) of the relativistic hydrodynamics equations
where and
NOTE: the source terms do not contain derivatives of the

hydrodynamical quantities (leaving intact the principal part)
and vanish in a flat spacetime
general relativistic
special relativistic
special relativistic
Newtonian
non self-
gravitating fluid
Discretizing the relativistic
hydrodynamics equations
Discretizing the equations in general
Lets restrict to a simpler but instructive problem: a homogeneous,
flux- conservative differential equation for the scalar function
in one dimension
Its generic, discretized form is (after volume and time integration)
where and
some approximation to
the average flux at j+1/2
Any finite-difference form of (1) must represent in the most accurate

way. Different ways of calculating lead to different evolution schemes
(FTCS, Lax, Runge-Kutta, etc, see www.aei.mpg.de/~rezzolla)
Possible solutions to the discontinuities problem:
! 1st order accurate schemes
generally fine, but very inaccurate (e.g. excessive diffusion,
with Lax method) or across discontinuities (e.g. upwind)
! 2nd order accurate schemes

generally introduce oscillations across discontinuities (not
monotonic or TVD; see dedicated slide)
! 2nd order accurate schemes with artificial viscosity

mimic Nature but not good in relativistic regimes
! Godunov Methods
good compromise between accuracy (2nd order with
smooth data, 1st-order at discontinuities) but monotonic.
Most importantly: discontinuities are exploited
Finite-Volume methods
Godunov methods are tightly related with finite-volume
methods. For simplicity, assume a 1-dim. uniform grid.
Finite-Volume Methods are based on subdividing the spatial
domain into intervals (finite volumes or grid cells) and
on keeping track of an approximation to the integral
over each of these volumes.

If is smooth, then this integral agrees with at the
midpoint of the interval to .
At each time step, we update these values using approximations
to the flux through the endpoints of the intervals.
Advantages of finite-volume methods
In terms of finite-volumes, it is easier to use important properties
of the conservation laws in deriving numerical methods.
In particular, we can ensure that the numerical method is
conservative in a way that mimics the true solution and this is
important for correctly calculating shock waves.
The quantity
approximates the integral of over the entire interval

Using a method in conservative form, the discrete sum will change
only due to the fluxes at the boundaries and . In this
way conservation (eg of mass) is guaranteed provided that the
boundary conditions are properly imposed.
High-Resolution Shock-Capturing (Godunov) Methods
Based on a simple, yet brilliant idea by Godunov (59).
Example of how basic physics can boost research in
computational physics.
Basic idea: a piecewise constant description of
hydrodynamical quantities produce a collection of local
Riemann problems whose solution can be found exactly.
The solution at time tn+1 can be constructed by piecing

together the Riemann solutions, provided that the time
step is short enough that the waves from two adjacent
Riemann problems have not yet started to interact.
What is a Riemann problem?
Its the evolution of a fluid initially composed of two states with
different and constant values of velocity, pressure and density, i.e.
where and are the two constant left and right states
A typical example of a Riemann
problem is a shock-tube where
there is a right-moving shock and a
left-moving rarefaction wave.
Not the development of a constant
state between the two waves and the
presence of a contact discontinuity
where the density is discontinuous
but pressure and velocity are not
High-Resolution Shock-Capturing (Godunov) Methods
In other words, the flux discussed in the previous

slides and needed to evolve the state vector to the new
timelevel is in this case given by
where
is the exact solution of the Riemann problem with initial

data
Solving the Riemann problem
The solution of such a problem is particularly simple because
the states are constant. In particular, given initial data
the two constant left and right states can be decomposed in

terms of the characteristic variables as
The initial data for

the i-th characteristic
variable will be:
and each of the i waves
will propagate as
Solving the Riemann problem
The solution is written as a linear superposition of waves
In other words, the solution at any point (x,t) is the sum of the
left states of all the waves which are to the right of (x,t) and of
the right states of all the waves which are to the left of (x,t)
Once changes sign, the corresponding conserved
variable changes from to while all the other with
remain constant.
Finally, because , the jump
in the fluxes for the i-th wave are
rarefaction wave shock front
Solution at the time n+1 of the

two Riemann problems at the
cell boundaries xj+1/2 and xj-1/2
Spacetime evolution of the

cell boundaries xj+1/2 and xj-1/2.
Each problem leads to a shock
wave and a rarefaction wave
moving in opposite directions
Initial data at the time n for the

cell boundaries xj+1/2 and xj-1/2
cell boundaries where fluxes are required

A quick primer of HRSC methods...
The numerical solution of a Riemann problem is based on three
basic steps:
1) reconstruct a piecewise polynomial function from the cell
averages:
A piecewise constant is the simplest reconstruction
2) evolve the hyperbolic eq. (exactly or approximately) using

these initial data to obtain after a time
3) average this solution over each grid cell to obtain a new

cell average
1) First reconstruct...
Higher accuracy is reached with a
piecewise linear reconstruction
better representation of the
solution.
Reconstructing the initial data for
the Riemann problem at the cell
boundaries can be made with a
number of algorithms
The most interesting are the TVDs
(minmod, MC, Superbee) for which
the solution is
Here is a coefficient based linear reconst.
on the slope of and varying
from (near a discontinuity) up
to (in smooth regions of the
solution). slope limiter
1) First reconstruct...
Higher accuracy is reached with a
states for a Riemann problem better representation of the
solution.
Reconstructing the initial data for
the Riemann problem at the cell
boundaries can be made with a
number of algorithms
The most interesting are the TVDs
(minmod, MC, Superbee) for which
the solution is
Here is a coefficient based linear reconst.
on the slope of and varying
from (near a discontinuity) up
to (in smooth regions of the
solution). slope limiter
A quick primer of HRSC methods...
The numerical solution of a Riemann problem is based on three
basic steps:
1) reconstruct a piecewise polynomial function from the cell
averages:
A piecewise constant is the simplest reconstruction
2) evolve the hyperbolic eq. (exactly or approximately) using

these initial data to obtain after a time
3) average this solution over each grid cell to obtain a new

cell average
2) then evolve...
As an example, Roes approximate Riemann solver can be
calculated as
where are the values of the primitive variables at the

right/left sides of the i-th interface and are the
eigenvalues and right eigenvectors of the Jacobian matrix.
The coefficients measure the jumps of the
characteristic variables across the characteristic field
3) finally: average, convert and build...
Once the solution in terms of the conserved variables
has been obtained, it is necessary to return to the primitive
variables after inverting numerically the set of equations
Note: this conversion cannot be done analytically and requires

the solution of a set of coupled eqs. This root-finding operation
is very expensive computationally.
With the primitive variables calculated, the stress-energy tensor
can be reconstructed and used on the RHS of the Einstein eqs.
This series of operations is repeated at each grid point and for
each time level
Summary
Modelling gravitational wave sources in vacuum spacetimes
leads to calculations in contrast to simulations in nonvacuum
spacetimes
The solution of the hydrodynamics equations requires
special care because of the nonlinear nature of the equations
Even smooth initial data tends to steepen and shock; in
addition any discretization leads to small discontinuities
Using a flux-conservative formulation is essential if
modelling discontinuities
HRSC methods are particularly suited to study
discontinuities since they are based on them via Riemann
problems

3+1 Presentación

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

3+1 Presentación

Uploaded by

Copyright:

Available Formats

Lecture I: Einstein equations in 3+1

Albert Einstein Institute, Potsdam, Germany

IMPRS Block lecture course, Wandlitz, Nov. 2008

solve Einstein equations without approximations(!)

spacelike (finite) slice

spacelike infinity infinitely away

That would lead rapidly ( ) to a

Need to use smarter gauges!

Some chance of measuring gws!

In principle, the yellow region is

Images by D. Pollney o the excised region has to move on the grid

" Simply more equations to solve: stretching supercomputers resources!

In the simplest scenarios it is possible to consider the spacetime

This defines the lapse function which is

i) define the unit normal vector to the hypersurface

ii) define the spatial metric

where the two projectors are obviously orthogonal

which, as expected, is compatible with spatial metric

All of the 4-dim tensor algebra can be extended straightforwardly

is purely spatial (spatial derivatives of spatial metric )

is a full 4-dimensional object containing also time

The information present in and missing in can be

i) in terms of the unit normal to

ii) in terms of the acceleration of normal observers

iii) in terms of the acceleration of normal observers

where is the Lie derivative along

For a vector field , this is given by the commutator:

For a 1-form ,this is given by:

As a result, for a generic tensor of rank this is given by:

Consider a vector at one position P and then

i) it is a spatial tensor by construction (obtained with contractions

ii) it is a symmetric tensor by construction (obtained with

iii) its trace is given by

since the acceleration of the normal observer is orthogonal to

and to do this we need to define a few identities

Next, we make 3 spatial projections and a timelike one to obtain

where the second derivative of the lapse has been introduced

Another important identity which will be used in the following is

and which holds for any spatial vector

where e, p, h, and are the energy density, the pressure, the

Similarly, the momentum density (i.e. mass current) will be given by

As we will see soon, the 3+1 splitting naturally splits the

The first set is usually referred to as the constraint equations,

Note that this is a single elliptic equation (hence not containing

Doing the same for the right-hand-side, using the contracted

which are also 3 elliptic equations.

We need therefore to find a new vector along which to carry out

where is any spatial shift vector. Clearly now

A more intuitive description of the lapse function and of the shift

the Ricci equation we have encountered before:

Using the Ricci eqs one then obtains

The choice in this case in very simple. We want:

ii) the fourth one has to be along the vector , i.e.

i.e. the Lie derivative along is a simple partial derivative

i.e. the space covariant components of a timelike vector are

i.e. the zeroth contravariant component of a spacelike vector

Similarly, since the covariant components are

Note that (i.e. are inverses) and thus they

the shift relates spatial coordinates

the spatial metric measures distances between points on every

the shift relates spatial coordinates

the spatial metric measures distances between points on every

Second step: decompose 4-dim spacetime tensors in spatial and