Gaussian Process Computations¶
The main interface to the celerite solver is through the GP
class.
This main purpose of this class is to exposes methods for efficiently
computing the marginalized likelihood of a Gaussian Process (GP) model.
The covariance matrix for the GP will be specified by a kernel function as
described in the Kernel Building section.
The GP
class implements the Python: Modeling with submodels called
kernel
and mean
so the modeling language can be used to fit for
parameters for each of these models.
Below, the methods of the GP
object are described in detail but the
most important methods to look at are probably GP.compute()
,
GP.log_likelihood()
, and GP.predict()
.
-
class
celerite.
GP
(kernel, mean=0.0, fit_mean=False, log_white_noise=None, fit_white_noise=False)¶ The main interface to the celerite Gaussian Process solver
- Parameters
kernel – An instance of a subclass of
terms.Term
.mean (Optional) – A simple mean value for the process. This can either be a
float
or a subclass ofmodeling.Model
. (default:0.0
)fit_mean (optional) – If
False
, all of the parameters ofmean
will be frozen. Otherwise, the parameter states are unaffected. (default:False
)
-
apply_inverse
(y)¶ Apply the inverse of the covariance matrix to a vector or matrix
Solve
K.x = y
forx
whereK
is the covariance matrix of the GP with the white noise andyerr
components included on the diagonal.- Parameters
y (array[n] or array[n, nrhs]) – The vector or matrix
y
above. (described) –
- Returns
The solution to the linear system. This will have the same shape as
y
.- Return type
array[n] or array[n, nrhs]
- Raises
ValueError – For mismatched dimensions.
-
compute
(t, yerr=1.123e-12, check_sorted=True, A=None, U=None, V=None)¶ Compute the extended form of the covariance matrix and factorize
- Parameters
x (array[n]) – The independent coordinates of the data points. This array must be _sorted_ in ascending order.
yerr (Optional[float or array[n]]) – The measurement uncertainties for the data points at coordinates
x
. These values will be added in quadrature to the diagonal of the covariance matrix. (default:1.123e-12
)check_sorted (bool) – If
True
,x
will be checked to make sure that it is properly sorted. IfFalse
, the coordinates will be assumed to be in the correct order.
- Raises
ValueError – For un-sorted data or mismatched dimensions.
solver.LinAlgError – For non-positive definite matrices.
-
dot
(y, t=None, A=None, U=None, V=None, kernel=None, check_sorted=True)¶ Dot the covariance matrix into a vector or matrix
Compute
K.y
whereK
is the covariance matrix of the GP without the white noise oryerr
values on the diagonal.- Parameters
y (array[n] or array[n, nrhs]) – The vector or matrix
y
described above.kernel (Optional[terms.Term]) – A different kernel can optionally be provided to compute the matrix
K
from a different kernel than thekernel
property on this object.
- Returns
The dot product
K.y
as described above. This will have the same shape asy
.- Return type
array[n] or array[n, nrhs]
- Raises
ValueError – For mismatched dimensions.
-
get_matrix
(x1=None, x2=None, include_diagonal=None, include_general=None)¶ Get the covariance matrix at given independent coordinates
- Parameters
x1 (Optional[array[n1]]) – The first set of independent coordinates. If this is omitted,
x1
will be assumed to be equal tox
from a previous call toGP.compute()
.x2 (Optional[array[n2]]) – The second set of independent coordinates. If this is omitted,
x2
will be assumed to bex1
.include_diagonal (Optional[bool]) – Should the white noise and
yerr
terms be included on the diagonal? (default:False
)
-
grad_log_likelihood
(y, quiet=False)¶ Compute the gradient of the marginalized likelihood
The factorized matrix from the previous call to
GP.compute()
is used socompute
must be called first. The gradient is taken with respect to the parameters returned byGP.get_parameter_vector()
. This function requires the autograd package.- Parameters
y (array[n]) – The observations at coordinates
x
fromGP.compute()
.quiet (bool) – If true, return
-numpy.inf
and a gradient vector of zeros for non-positive definite matrices instead of throwing an error.
- Returns
The gradient of marginalized likelihood with respect to the parameter vector.
- Raises
ValueError – For mismatched dimensions.
solver.LinAlgError – For non-positive definite matrices.
-
log_likelihood
(y, _const=1.8378770664093453, quiet=False)¶ Compute the marginalized likelihood of the GP model
The factorized matrix from the previous call to
GP.compute()
is used socompute
must be called first.- Parameters
y (array[n]) – The observations at coordinates
x
fromGP.compute()
.quiet (bool) – If true, return
-numpy.inf
for non-positive definite matrices instead of throwing an error.
- Returns
The marginalized likelihood of the GP model.
- Return type
- Raises
ValueError – For mismatched dimensions.
solver.LinAlgError – For non-positive definite matrices.
-
property
mean
¶ The mean
modeling.Model
-
predict
(y, t=None, return_cov=True, return_var=False)¶ Compute the conditional predictive distribution of the model
You must call
GP.compute()
before this method.- Parameters
y (array[n]) – The observations at coordinates
x
fromGP.compute()
.t (Optional[array[ntest]]) – The independent coordinates where the prediction should be made. If this is omitted the coordinates will be assumed to be
x
fromGP.compute()
and an efficient method will be used to compute the prediction.return_cov (Optional[bool]) – If
True
, the full covariance matrix is computed and returned. Otherwise, only the mean prediction is computed. (default:True
)return_var (Optional[bool]) – If
True
, only return the diagonal of the predictive covariance; this will be faster to compute than the full covariance matrix. This overridesreturn_cov
so, if both are set toTrue
, only the diagonal is computed. (default:False
)
- Returns
mu
,(mu, cov)
, or(mu, var)
depending on the values ofreturn_cov
andreturn_var
. These output values are: (a) mu(ntest,)
: mean of the predictive distribution, (b) cov(ntest, ntest)
: the predictive covariance matrix, and (c) var(ntest,)
: the diagonal elements ofcov
.- Raises
ValueError – For mismatched dimensions.
-
sample
(size=None)¶ Sample from the prior distribution over datasets
- Parameters
size (Optional[int]) – The number of samples to draw.
- Returns
The samples from the prior distribution over datasets.
- Return type
array[n] or array[size, n]
-
sample_conditional
(y, t=None, size=None, regularize=None)¶ Sample from the conditional (predictive) distribution
Note: this method scales as
O(M^3)
for largeM
, whereM == len(t)
.- Parameters
y (array[n]) – The observations at coordinates
x
fromGP.compute()
.t (Optional[array[ntest]]) – The independent coordinates where the prediction should be made. If this is omitted the coordinates will be assumed to be
x
fromGP.compute()
and an efficient method will be used to compute the prediction.size (Optional[int]) – The number of samples to draw.
regularize (Optional[float]) – For poorly conditioned systems, you can provide a small number here to regularize the predictive covariance. This number will be added to the diagonal.
- Returns
The samples from the conditional distribution over datasets.
- Return type
array[n] or array[size, n]