emukit.quadrature.kernels package

Module contents

Kernel embeddings for Bayesian quadrature.

class emukit.quadrature.kernels.QuadratureKernel(kern, measure, variable_names)

Bases: object

Abstract class for a kernel augmented with integrability.

Note

Each specific implementation of this class must go with a specific standard kernel as input which inherits from IStandardKernel. This is because we both want the QuadratureKernel to be backend agnostic but at the same time QuadratureKernel needs access to specifics of the standard kernel. For example a specific pair of QuadratureKernel and IStandardKernel is QuadratureRBF and IRBF. The kernel embeddings are implemented w.r.t. a specific integration measure, for example the LebesgueMeasure.

Parameters:

kern (IStandardKernel) – Standard EmuKit kernel.
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

K(x1, x2)

The kernel k(x1, x2) evaluated at x1 and x2.

Parameters:

x1 (ndarray) – First argument of the kernel.
x2 (ndarray) – Second argument of the kernel.

Return type:

ndarray

Returns:

The kernel at x1, x2.

qK(x2)

The kernel with the first argument integrated out (kernel mean) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the kernel mean is evaluated, shape (n_points, input_dim).
Return type:: ndarray
Returns:: The kernel mean at x2, shape (1, n_points).

Kq(x1)

The kernel with the second argument integrated out (kernel mean) evaluated at x1.

Parameters:: x1 (ndarray) – The locations where the kernel mean is evaluated, shape (n_points, input_dim).
Return type:: ndarray
Returns:: The kernel mean at x1, shape (n_points, 1).

qKq()

The kernel integrated over both arguments x1 and x2.

Return type:: float
Returns:: Double integrated kernel.

dK_dx1(x1, x2)

The gradient of the kernel wrt x1 evaluated at pair x1, x2.

Parameters:

x1 (ndarray) – First argument of the kernel, shape (n_points N, input_dim).
x2 (ndarray) – Second argument of the kernel, shape (n_points M, input_dim).

Return type:

ndarray

Returns:

The gradient at (x1, x2), shape (input_dim, N, M).

dK_dx2(x1, x2)

The gradient of the kernel wrt x2 evaluated at pair x1, x2.

Note that this is equal to the transposed gradient of the kernel wrt x1 evaluated at x2 and x1 (swapped arguments).

Parameters:

x1 (ndarray) – First argument of the kernel, shape (n_points N, N, input_dim).
x2 (ndarray) – Second argument of the kernel, shape (n_points N, M, input_dim).

Return type:

ndarray

Returns:

The gradient at (x1, x2), shape (input_dim, N, M).

dKdiag_dx(x)

The gradient of the diagonal of the kernel \(v(x):=k(x, x)\) evaluated at x.

Parameters:: x (ndarray) – The locations where the gradient is evaluated, shape = (n_points M, input_dim).
Return type:: ndarray
Returns:: The gradient at x, shape (input_dim, M).

dqK_dx(x2)

The gradient of the kernel mean (integrated in first argument) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the gradient is evaluated, shape (n_points N, N, input_dim).
Return type:: ndarray
Returns:: The gradient at x2, shape (input_dim, N).

dKq_dx(x1)

The gradient of the kernel mean (integrated in second argument) evaluated at x1.

Parameters:: x1 (ndarray) – The locations where the gradient is evaluated, shape (n_points N, N, input_dim).
Return type:: ndarray
Returns:: The gradient with shape (N, input_dim).

class emukit.quadrature.kernels.QuadratureProductKernel(kern, measure, variable_names)

Bases: QuadratureKernel

Abstract class for a product kernel augmented with integrability.

The product kernel is of the form \(k(x, x') = \sigma^2 \prod_{i=1}^d k_i(x, x')\) where \(k_i(x, x')\) is a univariate kernel acting on dimension \(i\).

Parameters:

kern (IStandardKernel) – Standard EmuKit kernel (must be a product kernel).
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

qK(x2, skip=None)

The kernel with the first argument integrated out (kernel mean) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the kernel mean is evaluated, shape (n_points, input_dim).
Return type:: ndarray
Returns:: The kernel mean at x2, shape (1, n_points).

qKq()

The kernel integrated over both arguments x1 and x2.

Return type:: float
Returns:: Double integrated kernel.

dqK_dx(x2)

The gradient of the kernel mean (integrated in first argument) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the gradient is evaluated, shape (n_points N, N, input_dim).
Return type:: ndarray
Returns:: The gradient at x2, shape (input_dim, N).

class emukit.quadrature.kernels.LebesgueEmbedding

Bases: object

Mixin for quadrature kernel w.r.t. Lebesgue measure.

See also

emukit.quadrature.measures.LebesgueMeasure

classmethod from_integral_bounds(kern, integral_bounds, normalized=False, variable_names='')

Create the quadrature kernel w.r.t. a Lebesgue measure from the integral bounds.

Parameters:

kern (IStandardKernel) – Standard EmuKit kernel.
integral_bounds (List[Tuple[float, float]]) – The integral bounds. List of \(d\) tuples \([(a_1, b_1), (a_2, b_2), \dots, (a_d, b_d)]\), where \(d\) is the dimensionality of the integral and the tuple \((a_i, b_i)\) contains the lower and upper integration bound of dimension \(i\).
normalized (bool) – Weather the Lebesgue measure is normalized.
variable_names (str) – The (variable) name(s) of the integral

Returns:

An instance of a quadrature kernel w.r.t. Lebesgue measure.

class emukit.quadrature.kernels.GaussianEmbedding

Bases: object

Mixin for quadrature kernel w.r.t. Gaussian measure.

See also

emukit.quadrature.measures.GaussianMeasure

classmethod from_measure_params(kern, mean, variance, variable_names='')

Create the quadrature kernel w.r.t. a Gaussian measure from the measure parameters.

Parameters:

kern (IStandardKernel) – Standard EmuKit kernel.
mean (ndarray) – The mean of the Gaussian measure, shape (input_dim, ).
variance (Union[float, ndarray]) – The variances of the Gaussian measure, shape (input_dim, ). If a scalar value is given, all dimensions will have same variance.
variable_names (str) – The (variable) name(s) of the integral

Returns:

An instance of a quadrature kernel w.r.t. Gaussian measure.

class emukit.quadrature.kernels.QuadratureBrownian(brownian_kernel, measure, variable_names)

Bases: QuadratureKernel

Base class for a Brownian motion kernel augmented with integrability.

\[k(x, x') = \sigma^2 \operatorname{min}(x, x')\quad\text{with}\quad x, x' \geq 0,\]

where \(\sigma^2\) is the variance property.

Note

This class is compatible with the standard kernel IBrownian. Each child of this class implements an embedding w.r.t. a specific integration measure.

See also

Parameters:

brownian_kernel (IBrownian) – The standard EmuKit Brownian motion kernel.
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

Raises:

ValueError – If measure has wrong dimensionality.
ValueError – If the reasonable box of the measure allows for negative values.

property variance: float: The scale \(\sigma^2\) of the kernel.

class emukit.quadrature.kernels.QuadratureBrownianLebesgueMeasure(brownian_kernel, measure, variable_names='')

Bases: QuadratureBrownian, LebesgueEmbedding

A Brownian motion kernel augmented with integrability w.r.t. the standard Lebesgue measure.

See also

Parameters:

brownian_kernel (IBrownian) – The standard EmuKit Brownian motion kernel.
measure (LebesgueMeasure) – The Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.

qK(x2)

The kernel with the first argument integrated out (kernel mean) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the kernel mean is evaluated, shape (n_points, input_dim).
Return type:: ndarray
Returns:: The kernel mean at x2, shape (1, n_points).

qKq()

The kernel integrated over both arguments x1 and x2.

Return type:: float
Returns:: Double integrated kernel.

dqK_dx(x2)

The gradient of the kernel mean (integrated in first argument) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the gradient is evaluated, shape (n_points N, N, input_dim).
Return type:: ndarray
Returns:: The gradient at x2, shape (input_dim, N).

class emukit.quadrature.kernels.QuadratureRBF(rbf_kernel, measure, variable_names)

Bases: QuadratureKernel

Base class for an RBF kernel augmented with integrability.

\[k(x, x') = \sigma^2 e^{-\frac{1}{2}\sum_{i=1}^{d}r_i^2},\]

where \(d\) is the input dimensionality, \(r_i = \frac{x_i-x_i'}{\lambda_i}\) is the scaled vector difference of dimension \(i\), \(\lambda_i\) is the \(i\) th element of the lengthscales property and \(\sigma^2\) is the variance property.

Note

This class is compatible with the standard kernel IRBF. Each child of this class implements an embedding w.r.t. a specific integration measure.

See also

Parameters:

rbf_kernel (IRBF) – The standard EmuKit rbf-kernel.
measure (IntegrationMeasure) – The integration measure. None implies the standard Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.

property lengthscales: ndarray: The lengthscales \(\lambda\) of the kernel.

property variance: float: The scale \(\sigma^2\) of the kernel.

class emukit.quadrature.kernels.QuadratureRBFLebesgueMeasure(rbf_kernel, measure, variable_names='')

Bases: QuadratureRBF, LebesgueEmbedding

An RBF kernel augmented with integrability w.r.t. the standard Lebesgue measure.

See also

Parameters:

rbf_kernel (IRBF) – The standard EmuKit rbf-kernel.
measure (LebesgueMeasure) – The Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.

qK(x2)

The kernel with the first argument integrated out (kernel mean) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the kernel mean is evaluated, shape (n_points, input_dim).
Return type:: ndarray
Returns:: The kernel mean at x2, shape (1, n_points).

qKq()

The kernel integrated over both arguments x1 and x2.

Return type:: float
Returns:: Double integrated kernel.

dqK_dx(x2)

The gradient of the kernel mean (integrated in first argument) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the gradient is evaluated, shape (n_points N, N, input_dim).
Return type:: ndarray
Returns:: The gradient at x2, shape (input_dim, N).

class emukit.quadrature.kernels.QuadratureRBFGaussianMeasure(rbf_kernel, measure, variable_names='')

Bases: QuadratureRBF, GaussianEmbedding

An RBF kernel augmented with integrability w.r.t. a Gaussian measure.

See also

Parameters:

rbf_kernel (IRBF) – The standard EmuKit rbf-kernel.
measure (GaussianMeasure) – A Gaussian measure.
variable_names (str) – The (variable) name(s) of the integral.

qK(x2, scale_factor=1.0)

The kernel with the first argument integrated out (kernel mean) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the kernel mean is evaluated, shape (n_points, input_dim).
Return type:: ndarray
Returns:: The kernel mean at x2, shape (1, n_points).

qKq()

The kernel integrated over both arguments x1 and x2.

Return type:: float
Returns:: Double integrated kernel.

dqK_dx(x2)

The gradient of the kernel mean (integrated in first argument) evaluated at x2.

Parameters:: x2 (ndarray) – The locations where the gradient is evaluated, shape (n_points N, N, input_dim).
Return type:: ndarray
Returns:: The gradient at x2, shape (input_dim, N).

class emukit.quadrature.kernels.QuadratureProductMatern52(matern_kernel, measure, variable_names)

Bases: QuadratureProductKernel

Base class for a product Matern52 kernel augmented with integrability.

The kernel is of the form \(k(x, x') = \sigma^2 \prod_{i=1}^d k_i(x, x')\) where

\[k_i(x, x') = (1 + \sqrt{5} r_i + \frac{5}{3} r_i^2) \exp(- \sqrt{5} r_i).\]

Above, \(d\) is the input dimensionality, \(r_i =\frac{|x_i - x'_i|}{\lambda_i}\), is the scaled distance, \(\sigma^2\) is the variance property and \(\lambda_i\) is the \(i\) th element of the lengthscales property.

Note

This class is compatible with the standard kernel IProductMatern52. Each subclass of this class implements an embedding w.r.t. a specific integration measure.

See also

Parameters:

matern_kernel (IProductMatern52) – The standard EmuKit product Matern52 kernel.
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

property nu: float: The smoothness parameter of the kernel.

property lengthscales: ndarray: The lengthscales \(\lambda\) of the kernel.

property variance: float: The scale \(\sigma^2\) of the kernel.

class emukit.quadrature.kernels.QuadratureProductMatern52LebesgueMeasure(matern_kernel, measure, variable_names='')

Bases: QuadratureProductMatern52, LebesgueEmbedding

A product Matern52 kernel augmented with integrability w.r.t. the standard Lebesgue measure.

See also

Parameters:

matern_kernel (IProductMatern52) – The standard EmuKit product Matern52 kernel.
measure (LebesgueMeasure) – The Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.

class emukit.quadrature.kernels.QuadratureProductMatern32(matern_kernel, measure, variable_names)

Bases: QuadratureProductKernel

Base class for a product Matern32 kernel augmented with integrability.

The kernel is of the form \(k(x, x') = \sigma^2 \prod_{i=1}^d k_i(x, x')\) where

\[k_i(x, x') = (1 + \sqrt{3}r_i ) e^{-\sqrt{3} r_i}.\]

Above, \(d\) is the input dimensionality, \(r_i =\frac{|x_i - x'_i|}{\lambda_i}\), is the scaled distance, \(\sigma^2\) is the variance property and \(\lambda_i\) is the \(i\) th element of the lengthscales property.

Note

This class is compatible with the standard kernel IProductMatern32. Each subclass of this class implements an embedding w.r.t. a specific integration measure.

See also

Parameters:

matern_kernel (IProductMatern32) – The standard EmuKit product Matern32 kernel.
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

property nu: float: The smoothness parameter of the kernel.

property lengthscales: ndarray: The lengthscales \(\lambda\) of the kernel.

property variance: float: The scale \(\sigma^2\) of the kernel.

class emukit.quadrature.kernels.QuadratureProductMatern32LebesgueMeasure(matern_kernel, measure, variable_names='')

Bases: QuadratureProductMatern32, LebesgueEmbedding

A product Matern32 kernel augmented with integrability w.r.t. the standard Lebesgue measure.

See also

Parameters:

matern_kernel (IProductMatern32) – The standard EmuKit product Matern32 kernel.
measure (LebesgueMeasure) – The Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.

class emukit.quadrature.kernels.QuadratureProductMatern12(matern_kernel, measure, variable_names)

Bases: QuadratureProductKernel

Base class for a product Matern12 (a.k.a. Exponential) kernel augmented with integrability.

The kernel is of the form \(k(x, x') = \sigma^2 \prod_{i=1}^d k_i(x, x')\) where

\[k_i(x, x') = e^{-r_i}.\]

Above, \(d\) is the input dimensionality, \(r_i =\frac{|x_i - x'_i|}{\lambda_i}\), is the scaled distance, \(\sigma^2\) is the variance property and \(\lambda_i\) is the \(i\) th element of the lengthscales property.

Note

This class is compatible with the standard kernel IProductMatern12. Each subclass of this class implements an embedding w.r.t. a specific integration measure.

See also

Parameters:

matern_kernel (IProductMatern12) – The standard EmuKit product Matern12 kernel.
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

property nu: float: The smoothness parameter of the kernel.

property lengthscales: ndarray: The lengthscales \(\lambda\) of the kernel.

property variance: float: The scale \(\sigma^2\) of the kernel.

class emukit.quadrature.kernels.QuadratureProductMatern12LebesgueMeasure(matern_kernel, measure, variable_names='')

Bases: QuadratureProductMatern12, LebesgueEmbedding

A product Matern12 kernel augmented with integrability w.r.t. the standard Lebesgue measure.

See also

Parameters:

matern_kernel (IProductMatern12) – The standard EmuKit product Matern12 kernel.
measure (LebesgueMeasure) – The Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.

class emukit.quadrature.kernels.QuadratureProductBrownian(brownian_kernel, measure, variable_names)

Bases: QuadratureProductKernel

Base class for a product Brownian kernel augmented with integrability.

The kernel is of the form \(k(x, x') = \sigma^2 \prod_{i=1}^d k_i(x, x')\) where

\[k_i(x, x') = \operatorname{min}(x_i-c, x_i'-c)\quad\text{with}\quad x_i, x_i' \geq c,\]

\(d\) is the input dimensionality, \(\sigma^2\) is the variance property and \(c\) is the offset property.

Note

This class is compatible with the standard kernel IProductBrownian. Each subclass of this class implements an embedding w.r.t. a specific integration measure.

See also

Parameters:

brownian_kernel (IProductBrownian) – The standard EmuKit product Brownian kernel.
measure (IntegrationMeasure) – The integration measure.
variable_names (str) – The (variable) name(s) of the integral.

Raises:

ValueError – If the reasonable box of the measure allows for values smaller than the offset \(c\).

property variance: float: The scale \(\sigma^2\) of the kernel.

property offset: float: The offset \(c\) of the kernel.

class emukit.quadrature.kernels.QuadratureProductBrownianLebesgueMeasure(brownian_kernel, measure, variable_names='')

Bases: QuadratureProductBrownian, LebesgueEmbedding

A product Brownian kernel augmented with integrability w.r.t. the standard Lebesgue measure.

See also

Parameters:

brownian_kernel (IProductBrownian) – The standard EmuKit product Brownian kernel.
measure (LebesgueMeasure) – The Lebesgue measure.
variable_names (str) – The (variable) name(s) of the integral.