Cosmology For Developers#

Cosmologies in Functions#

It is often useful to assume a default cosmology so that the exact cosmology does not have to be specified every time a function or method is called. In this case, it is possible to specify a “default” cosmology.

You can set the default cosmology to a predefined value by using the “default_cosmology” option in the [cosmology.core] section of the configuration file (see Configuration System (astropy.config)). Alternatively, you can use the set() function of default_cosmology to set a cosmology for the current Python session. If you have not set a default cosmology using one of the methods described above, then the cosmology module will default to using the default_cosmology._value parameters.

You can override the default cosmology through the default_cosmology science state object, using something like the following:

from astropy.cosmology import default_cosmology

def myfunc(..., cosmo=None):
    if cosmo is None:
        cosmo = default_cosmology.get()

    ... function code here ...

This ensures that all code consistently uses the default cosmology unless explicitly overridden.


If you are preparing a paper and thus need to ensure your code provides reproducible results, it is better to use an explicit cosmology (for example WMAP9.H(0) instead of default_cosmology.get().H(0)). Use of the default cosmology should generally be reserved for code that allows for the global cosmology state to be changed; e.g. code included in astropy core or an affiliated package.

Custom Cosmologies#

In astropy.cosmology cosmologies are classes, so custom cosmologies may be implemented by subclassing Cosmology (or more likely FLRW) and adding details specific to that cosmology. Here we will review some of those details and tips and tricks to building a performant cosmology class.

from astropy.cosmology import FLRW

@dataclass(frozen=True, eq=False)
class CustomCosmology(FLRW):
    ...  # [details discussed below]


An astropy Cosmology is characterized by 1) its class, which encodes the physics, and 2) its free parameter(s), which specify a cosmological realization. When defining the former, all parameters must be declared using Parameter and should have values assigned at instantiation.

A Parameter is a descriptor. When accessed from a class it transparently stores information, like the units and accepted equivalencies, that might be opaquely contained in the constructor signature or more deeply in the code. On a cosmology instance, the descriptor will return the parameter value.

There are a number of best practices. For a reference, this is excerpted from the definition of FLRW.

@dataclass(frozen=True, eq=False)
class FLRW(Cosmology):

    H0: Parameter = Parameter(doc="Hubble constant as an `~astropy.units.Quantity` at z=0",
                              unit="km/(s Mpc)", fvalidate="scalar")
    Om0: Parameter = Parameter(doc="Omega matter; matter density/critical density at z=0",
    Ode0: Parameter = Parameter(doc="Omega dark energy; dark energy density/critical density at z=0.",
    Tcmb0: Parameter = Parameter(doc="Temperature of the CMB as `~astropy.units.Quantity` at z=0.",
                                 default=0.0 * u.K, unit="Kelvin", fmt="0.4g", fvalidate="scalar")
    Neff: Parameter = Parameter(doc="Number of effective neutrino species.",
                                default=3.04, fvalidate="non-negative")
    m_nu: Parameter = Parameter(doc="Mass of neutrino species.",
                                default=0.0*u.eV, unit="eV", equivalencies=u.mass_energy(), fmt="")
    Ob0: Parameter = Parameter(doc="Omega baryon; baryonic matter density/critical density at z=0.",

    def Ob0(self, param, value):
        """Validate baryon density to None or positive float > matter density."""
        if value is None:
            return value
        value = _validate_non_negative(self, param, value)
        if value > self.Om0:
            raise ValueError("baryonic density can not be larger than total matter density.")
        return value

First note that all the parameters are also arguments in __init__(). This is not strictly necessary, but is good practice. If the parameter has units (and related equivalencies) these must be specified on the Parameter, as seen in The “H0” item in parameters.

The next important thing to note is how the parameter value is set, in __init__. Parameter allows for a value to be set once (before auto-locking), so self.H0 = H0 will use this setter and put the value on “._H0”. The advantage of this method over direct assignment to the private attribute is the use of validators. Parameter allows for custom value validators, using the method-decorator validator, that can check a value’s validity and modify the value, e.g to assign units. If no custom validator is specified the default is to check if the Parameter has defined units and if so, return the value as a Quantity with those units, using all enabled and the parameter’s unit equivalencies.

The last thing to note is pretty formatting for the Cosmology. Each Parameter defaults to the format specification “.3g”, but this may be overridden, like Tcmb0 does.

If a new cosmology modifies an existing Parameter, then the clone() method is useful to deep-copy the parameter and change any constructor argument. For example, see FlatFLRWMixin in astropy.cosmology.flrw (also shown below).

@dataclass(frozen=True, eq=False)
class FlatFLRWMixin(FlatCosmologyMixin):

    Ode0: Parameter = FLRW.parameters["Ode0"].clone(derived=True)


Mixins are used in cosmology to reuse code across multiple classes in different inheritance lines. We use the term loosely as mixins are meant to be strictly orthogonal, but may not be, particularly in __init__.

Currently the only mixin is FlatCosmologyMixin and its FLRW-specific subclass FlatFLRWMixin. “Flat” cosmologies should use this mixin. FlatFLRWMixin must precede the base class in the multiple-inheritance so that this mixin’s __init__ proceeds the base class’.

Speeding up Integrals in Custom Cosmologies#

The supplied cosmology classes use a few tricks to speed up distance and time integrals. It is not necessary for anyone subclassing FLRW to use these tricks – but if they do, such calculations may be a lot faster.

The first, more basic, idea is that, in many cases, it’s a big deal to provide explicit formulae for inv_efunc() rather than simply setting up de_energy_scale – assuming there is a nice expression. As noted above, almost all of the provided classes do this, and that template can pretty much be followed directly with the appropriate formula changes.

The second, and more advanced, option is to also explicitly provide a scalar only version of inv_efunc(). This results in a fairly large speedup (>10x in most cases) in the distance and age integrals, even if only done in python, because testing whether the inputs are iterable or pure scalars turns out to be rather expensive. To take advantage of this, the key thing is to explicitly set the instance variables self._inv_efunc_scalar and self._inv_efunc_scalar_args in the constructor for the subclass, where the latter are all the arguments except z to _inv_efunc_scalar. The provided classes do use this optimization, and in fact go even further and provide optimizations for no radiation, and for radiation with massless neutrinos coded in cython. Consult the FLRW subclasses and scalar_inv_efuncs for the details.

However, the important point is that it is not necessary to do this.

Astropy Interoperability: I/O and your Cosmology Package#

If you are developing a package and want to be able to interoperate with Cosmology, you’re in the right place! Here we will discuss how to enable Astropy to read and write your file formats, and convert your cosmology objects to and from Astropy’s Cosmology.

The following presumes knowledge of how Astropy structures I/O functions. For a quick tutorial see Cosmology I/O & Convert.

Now that we know how to build and register functions into read(), write(), from_format(), to_format(), we can do this in your package.

Consider a package – since this is mine, it’s cleverly named mypackage – with the following file structure: a module for cosmology codes and a module for defining related input/output functions. In the cosmology module are defined cosmology classes and a file format – myformat – and everything should interoperate with astropy. The tests are done with pytest and are integrated within the code structure.


Converting Objects Between Packages#

We want to enable conversion between cosmology objects from mypackage to/from Cosmology. All the Astropy interface code is defined in mypackage/io/ The following is a rough outline of the necessary functions and how to register them with astropy’s unified I/O to be automatically available to from_format() and to_format().

Reading and Writing#

Everything Astropy read/write related is defined in mypackage/io/ The following is a rough outline of the read, write, and identify functions and how to register them with astropy’s unified IO to be automatically available to read() and write().

If Astropy is an optional dependency#

The astropy_io and astropy_convert modules are written assuming Astropy is installed. If in mypackage it is an optional dependency then it is important to detect if Astropy is installed (and the correct version) before importing astropy_io and astropy_convert. We do this in mypackage/io/

Astropy Interoperability Tests#

Lastly, it’s important to test that everything works. In this example package all such tests are contained in mypackage/io/tests/ These tests require Astropy and will be skipped if it is not installed (and not the correct version), so at least one test in the test matrix should include astropy >= 5.0.