2. Generic kernel classes

All kernels in lsqfitgp are subclasses of Kernel, StationaryKernel or IsotropicKernel (the latter two are themselves subclasses of Kernel).

Kernel itself is a subclass of CrossKernel, which represents the covariance function between two different processes. In symbols, the kernel of a process \(f\) is

\[k_f(x, y) = \operatorname{Cov}[f(x), f(y)],\]

while the cross kernel between processes \(f\) and \(g\) is

\[k_{fg}(x, y) = \operatorname{Cov}[f(x), g(y)].\]

However you will probably never need to deal directly with a cross kernel, they are generated automatically behind the scenes by GP.

The three general classes can be used directly by instantiating them with a callable which will do the actual computation. However, this can be done in a simpler and more functional way using the decorators kernel(), stationarykernel() and isotropickernel().

class lsqfitgp.Kernel(kernel, *, dim=None, loc=None, scale=None, forcekron=False, derivable=None, saveargs=False, maxdim=sys.maxsize, batchbytes=None, **kw)

Base class for objects representing covariance kernels.

A Kernel object is callable, the signature is obj(x, y). Kernel objects can be summed and multiplied between them and with scalars, or raised to power with a scalar exponent.

Parameters:

kernelcallable

A function with signature kernel(x, y), where x and y are two broadcastable numpy arrays, which computes the covariance of f(x) with f(y) where f is the Gaussian process.

dimNone or str

When the input arrays are structured arrays, if dim is None the kernel will operate on all fields, i.e., it will be passed the whole arrays. If dim is a string, kernel will see only the arrays for the field named dim. If dim is a string and the array is not structured, an exception is raised. If the field for name dim has a nontrivial shape, the array passed to kernel is still structured but has only field dim.

loc, scalescalar

The inputs to kernel are transformed as (x - loc) / scale.

forcekronbool

If True, when calling kernel, if x and y are structured arrays, i.e., if they represent multidimensional input, kernel is invoked separately for each dimension, and the result is the product. Default False. The selection indicated by dim limits the dimensions over which forcekron applies.

derivablebool, int, None, or callable

Specifies how many times the kernel can be derived, only for error checking purposes. True means infinitely many times derivable. If callable, it is called with the same keyword arguments as kernel. If None (default) it means that the degree of derivability is unknown.

saveargsbool

If True, save the all the initialization arguments in a dictionary under the attribute initargs. Default False.

maxdimint, callable or None

The maximum input dimensionality accepted by the kernel. If callable, it is called with the same keyword arguments as the kernel. Default sys.maxsize.

batchbytesnumber, optional

If specified, apply batch(batchbytes) to the kernel.

**kw

Additional keyword arguments are passed to kernel.

Attributes:

derivableint or None

How many times the process represented by the kernel is derivable. sys.maxsize if it is smooth. None if the derivability is unknown.

maxdimint or None

Maximum input dimensionality. None means unknown.

Methods

rescale(xfun, yfun)	Multiply the kernel by functions of its arguments.
diff(xderiv, yderiv)	Return a Kernel-like object that computes the derivatives of this kernel.
xtransf(xfun, yfun)	Transform the inputs of the kernel.
fourier(dox, doy)	Compute the Fourier series of the kernel.
taylor(dox, doy)	Compute the Taylor series of the kernel.
batch(maxnbytes)	Return a batched version of the kernel.

batch(maxnbytes)

Return a batched version of the kernel.

The batched kernel processes its inputs in chunks to try to limit memory usage.

Parameters:

maxnbytesnumber

The maximum number of input bytes per chunk, counted after broadcasting the inputs (actual broadcasting may not occur if not induced by the operations in the kernel).

Returns:

batched_kernelCrossKernel

The same kernel but with batched computations.

diff(xderiv, yderiv)

Return a Kernel-like object that computes the derivatives of this kernel. The derivatives are computed automatically with JAX. If xderiv and yderiv are trivial, this is a no-op.

\[h(x, y) = \frac{\partial^n}{\partial x^n} \frac{\partial^m}{\partial y^m} k(x, y)\]

Parameters:

xderiv, yderivDeriv-like

A Deriv object or something that can be converted to a Deriv object.

Returns:

hKernel-like

An object representing the derivatives of this one. If xderiv == yderiv, it is actually another Kernel.

Raises:

RuntimeError

The derivative orders are greater than the derivative attribute.

fourier(dox, doy)

Compute the Fourier series of the kernel.

\[\begin{split}h(k, y) = \begin{cases} \frac2T \int_0^T \mathrm dx\, k(x, y) \cos\left(\frac{2\pi}T \frac k2 x\right) & \text{if $k$ is even} \\ \frac2T \int_0^T \mathrm dx\, k(x, y) \sin\left(\frac{2\pi}T \frac{k+1}2 x\right) & \text{if $k$ is odd} \end{cases}\end{split}\]

The period T is implicit in the definition of the kernel.

Parameters:

dox, doybool

Specify if to compute the series w.r.t. x, y or both. If both are False, this is a no-op.

Returns:

hKernel-like

A Kernel-like object computing the Fourier series. If dox and doy are equal, it is a Kernel.

rescale(xfun, yfun)

Multiply the kernel by functions of its arguments.

\[h(x, y) = f(x) k(x, y) g(y)\]

Parameters:

xfun, yfuncallable or None

Functions from the type of the arguments of the kernel to scalar. If both are None, this is a no-op.

Returns:

hKernel-like

The rescaled kernel. If xfun is yfun, it is a Kernel object, otherwise a Kernel-like one.

taylor(dox, doy)

Compute the Taylor series of the kernel.

\[h(k, y) = \left. \frac{\partial^k}{\partial x^k} k(x, y) \right|_{x_0}\]

The expansion point x0 is implicit in the definition of the kernel.

Parameters:

dox, doybool

Specify if to compute the series w.r.t. x, y or both. If both are False, this is a no-op.

Returns:

hKernel-like

A Kernel-like object computing the Taylor series. If dox and doy are equal, it is a Kernel.

xtransf(xfun, yfun)

Transform the inputs of the kernel.

\[h(x, y) = k(f(x), g(y))\]

Parameters:

xfun, yfuncallable or None

Functions mapping a new kind of input to the kind of input accepted by the kernel. If both are None, this is a no-op.

Returns:

hKernel-like

The transformed kernel. If xfun is yfun, it is a Kernel object, otherwise a Kernel-like one.

class lsqfitgp.StationaryKernel(kernel, *, input='signed', scale=None, **kw)

Subclass of Kernel for isotropic kernels.

Parameters:

kernelcallable

A function taking one argument delta which is the difference between x and y, plus optionally keyword arguments.

input{‘signed’, ‘soft’, ‘hard’}

If ‘signed’ (default), kernel is passed the bare difference. If ‘soft’, kernel is passed the absolute value of the difference, and the difference of equal points is a small number instead of zero. If ‘hard’, the absolute value.

scalescalar

The difference is divided by scale.

**kw

Additional keyword arguments are passed to the Kernel init.

class lsqfitgp.IsotropicKernel(kernel, *, input='squared', scale=None, **kw)

Subclass of Kernel for isotropic kernels.

Parameters:

kernelcallable

A function taking one argument r2 which is the squared distance between x and y, plus optionally keyword arguments.

input{‘squared’, ‘hard’, ‘soft’, ‘raw’}

If ‘squared’ (default), kernel is passed the squared distance. If ‘hard’, it is passed the distance (not squared). If ‘soft’, it is passed the distance, and the distance of equal points is a small number instead of zero. If ‘raw’, the kernel is passed both points separately like non-stationary kernels.

scalescalar

The distance is divided by scale.

**kw

Additional keyword arguments are passed to the Kernel init.

Notes

The ‘soft’ option will cause problems with second derivatives in more than one dimension.