Kernel zoo

Stationary

A popular assumption, more or less implying implies that no region of the process is special. In this case the kernel may be written as a function purely of the distance between \[K(\mathbf{x},t)=K(\|\mathbf{x}-t\|)\] for some distance \(\|\cdot\|\) between the observation coordinates. This kind of translation-invariant kernel is the default. They are convenient analysed in terms of the Wienere-Khinchine theorem.

A general stationary kernel is the Hida-Matérn kernel (Dowling, Sokół, and Park 2021).

Dot-product

The kernel is a function of the inner product/dot product of the input coordinates, and we may overload notation to write \[K(\mathbf{x},t\mathbf{y})=K(\mathbf{x}\cdot \mathbf{y}).\] Hard to google for because there is the confounding fact that kernels already define inner products in some other space so it’s a dot product defined in terms of a dot product.

Such kernels are rotation invariant but not stationary. Instead of the Fourier relationships that stationary kernels have, these can be described in Legendre bases for radial functions Smola, Óvári, and Williamson (2000) manufacture some theorems which tell you whether your choice of inner product can in fact define an inner product kernel. Unfortunately the basis so defined are long and ugly and not especially tractable to work with except maybe by computational algebra systems.

NN kernels

Infinite-width random NN kernels are nearly dot product kernels. They depend on the several dot products, \(\mathbf{x}\cdot \mathbf{x}\), \(\mathbf{x}\cdot \mathbf{y}\) and \(\mathbf{y}\cdot \mathbf{y}\). See NN kernels.

Causal kernels

Time-indexed processes are more general than a standard Wiener process. 🏗

What constraints make a covariance kernel causal? This is not always easily expressed in terms of the covariance kernel; you want something like the inverse covariance/precision.

Squared exponential

A.k.a. exponentiated quadratic. Often radial basis functions mean this also.

The classic, default, analytically convenient, because it is proportional to the Gaussian density and therefore cancels out with it at opportune times.

\[K_{\textrm{SE}}(\mathbf{x}, \mathbf{x}') = \sigma^2\exp\left(-\frac{(\mathbf{x} - \mathbf{x}')^2}{2\ell^2}\right)\]

Rational Quadratic

Duvenaud reckons this is everywhere but TBH I have not seen it. Included for completeness.

\[K_{\textrm{RQ}}(\mathbf{x}, \mathbf{x}') = \sigma^2 \left( 1 + \frac{(\mathbf{x} - \mathbf{x}')^2}{2 \alpha \ell^2} \right)^{-\alpha}\]

Note that \(\lim_{\alpha\to\infty} K_{\textrm{RQ}}= K_{\textrm{SE}}\).

Matérn

The Matérn stationary (and in the Euclidean case, isotropic) covariance function is one a surprisingly convenient model for covariance. See Carl Edward Rasmussen’s Gaussian Process lecture notes for a readable explanation, or chapter 4 of his textbook (Rasmussen and Williams 2006).

\[ K_{\textrm{Mat}}(\mathbf{x}, \mathbf{x}')=\sigma^{2} \frac{2^{1-\nu}}{\Gamma(\nu)}\left(\sqrt{2 \nu} \frac{\mathbf{x} - \mathbf{x}'}{\rho}\right)^{\nu} K_{\nu}\left(\sqrt{2 \nu} \frac{\mathbf{x} - \mathbf{x}'}{\rho}\right) \]

where \(\Gamma\) is the gamma function, \(\ K_{\nu }\) is the modified Bessel function of the second kind, and \(\rho,\nu\geq 0\). We many then read the differentiability of the solution directly off the parameterization.

Periodic

\[ K_{\textrm{Per}}(\mathbf{x}, \mathbf{x}') = \sigma^2\exp\left(-\frac{2\sin^2(\pi|\mathbf{x} - \mathbf{x}'|/p)}{\ell^2}\right) \]

Locally periodic

This is an example of a composed kernel.

\[\begin{aligned} K_{\textrm{LocPer}}(\mathbf{x}, \mathbf{x}') &= K_{\textrm{Per}}(\mathbf{x}, \mathbf{x}')K_{\textrm{SE}}(\mathbf{x}, \mathbf{x}') \\ &= \sigma^2\exp\left(-\frac{2\sin^2(\pi|\mathbf{x} - \mathbf{x}'|/p)}{\ell^2}\right) \exp\left(-\frac{(\mathbf{x} - \mathbf{x}')^2}{2\ell^2}\right) \end{aligned}\]

Obviously there are other possible localisations of a periodic kernel. This is a locally periodic kernel. NB it is not local in the sense of Genton’s local stationarity, just local in the sense that one kernel is ‘enveloped’ by another.

“Integral” kernel

I just noticed the ambiguously named Integral kernel:

I’ve called the kernel the ‘integral kernel’ as we use it when we know observations of the integrals of a function, and want to estimate the function itself.

Examples include:

• Knowing how far a robot has travelled after 2, 4, 6 and 8 seconds, but wanting an estimate of its speed after 5 seconds…
• Wanting to know an estimate of the density of people aged 23, when we only have the total count for binned age ranges…

I argue that all kernels are naturally defined in terms of integrals, but the author seems to mean something particular. I suspect I would call this a sampling kernel, but that name is also overloaded. Anyway, what is actually going on here? Where is it introduced? Possibly one of (Smith, Alvarez, and Lawrence 2018; O’Callaghan and Ramos 2011; Murray-Smith and Pearlmutter 2005).

Composed kernels

See composing kernels.

Stationary spectral kernels

(Sun et al. 2018; Bochner 1959; Kom Samo and Roberts 2015; Yaglom 1987) construct spectral kernels in the sense that they use the spectral representation to design the kernel and guarantee it is positive definite and stationary. You could think of this as a kind of limiting case of composing kernels with a Fourier basis. See Bochner’s theorem.

Nonstationary spectral kernels

(Sun et al. 2018; Remes, Heinonen, and Kaski 2017; Kom Samo and Roberts 2015) use a generalised Bochner Theorem (Yaglom 1987) often called Yaglom’s Theorem, which does not presume stationarity. See Yaglom’s theorem.

It is not immediately clear how to use this; spectral representations are not an intuitive way of constructing things.

Compactly supported

We usually think about compactly supported kernels in the stationary isotropic case, where we mean kernels that vanish whenever the distance between two observation \(\mathbf{x},\mathbf{y}\) is larger than a certain cut-off distance \(L,\) i.e. \(\|\mathbf{x}-\mathbf{y}\|>L\Rightarrow K(\mathbf{x},\mathbf{y})=0\). These are great because they make the Gram matrix sparse (for example, if the cut-off is much smaller than the diameter of the observations and most observations have few covariance neighbours) and so can lead to computational efficiency even for exact inference without any special tricks. They don’t seem to be popular? Statisticians are generally nervous around inferring the support of a parameter, or assigning zero weight to any region of a prior without good reason, so maybe it is that?

Genton (2001) mentions \[ \max \left\{\left(1-\frac{\|\mathbf{x}-\mathbf{y}\|}{\tilde{\theta}}\right)^{\tilde{\nu}}, 0\right\} \] and handballs us to Gneiting (2002b) for a bigger smörgåsbord of stationary compactly supported kernels. Gneiting (2002b) has a couple of methods designed to produce certain smoothness properties at boundary and origin, but mostly concerns producing compactly supported kernels via clever integral transforms.

For inner product kernels, this can be diabolical. The Schaback and Wu method discusses some operations that preserve positive-definiteness.

NB if you are trying specifically to enforce sparsity here, it might be worth considering the kernel induced by a stochaastic convolution, which is kind of a precision parameterisation.

Markov kernels

How can we know from inspecting a kernel whether it implies an independence structure of some kind? The Wiener process and causal kernels clearly imply certain independences. Any kernel \(K(s,t)=K(s\wedge t)\) is clearly Markov. Are there more general ones? TODO: relate to kernels of bounded support. 🏗

Genton kernels

That’s my name for them because they seem to originate in (Genton 2001).

For any non-negative function \(h:\mathcal{T}\to\mathbb{R}^+\) with \(h(\mathbf{0})=0,\) the following is a kernel:

\[ K(\mathbf{x}, \mathbf{y})=\frac{[h(\mathbf{x}+\mathbf{y})-h(\mathbf{x}-\mathbf{y})}{4} \] Genton gives the example of \(h:\mathbf{x}\mapsto \|\mathbf{x}\|_2^2.\) instance, consider the function \(h(\mathbf{x})=\mathbf{x}^{\top} \mathbf{x} .\) From this we obtain the kernel: \[ K(\mathbf{x}, \mathbf{z})=\frac{1}{4}\left[(\mathbf{x}+\mathbf{z})^{\top}(\mathbf{x}+\mathbf{z})-(\mathbf{x}-\mathbf{z})^{\top}(\mathbf{x}-\mathbf{z})\right]=\mathbf{x}^{\top} \mathbf{z} \] The motivation is the identity

\[ \operatorname { Covariance }\left(Y_{1}, Y_{2}\right)= \frac{\operatorname { Variance }\left(Y_{1}+Y_{2}\right)-\operatorname { Variance }\left(Y_{1}-Y_{2}\right)}{ 4}. \]

Kernels with desired symmetry

(D. Duvenaud 2014, chap. 2) summarises Ginsbourger et al’s work on kernels with desired symmetries / invariances. 🏗 This produces for example, the periodic kernel above, but also such cute tricks as priors over Möbius strips.

Stationary reducible kernels

See kernel warping.

Other nonstationary kernels

See nonstationary kernels.

References