2.1. Reversibility and Metropolis kernel

Before analysing the non-reversible Markov kernel, we first recall the definition of reversibility. Reversibility is important throughout the paper since our construction of a non-reversible Markov kernel is based on classes of reversible Markov kernels. A Markov kernel Q on a measurable space $(E,\mathcal{E})$ is $\mu$ -reversible for a $\sigma$ -finite measure $\mu$ if

(3) \begin{equation}\int_A\mu(\mathrm{d} x)Q(x,B)=\int_B\mu(\mathrm{d} x)Q(x,A)\end{equation}

for any $A, B\in\mathcal{E}$ . If Q is $\mu$ -reversible, then Q is $\mu$ -invariant. There is a strong connection between ergodicity and $\mu$ -reversibility. See [Reference Kipnis and Varadhan30, Reference Kontoyiannis and Meyn31, Reference Roberts and Rosenthal50, Reference Roberts and Tweedie53].

As we mentioned above, our non-reversible Markov kernel is based on a class of reversible kernels. Suppose that $\mu$ is a probability measure on $(E,\mathcal{E})$ where E is closed by a summation operator. A simple approach to constructing a reversible kernel is to first describe $\mu$ as an image measure of a convolution of probability measures $\mu_Y, \mu_Z$ under a measurable map f, i.e., $\mu=(\mu_Y*\mu_Z)\circ f^{-1}$ . Here, an image measure of a measure $\mu$ under a map $f\;:\;E\rightarrow E$ is defined by

\begin{align*}\mu\circ f^{-1}(A)=\mu(\{x\in E\;:\;f(x)\in A\}), \end{align*}

and a convolution of $\mu_1$ and $\mu_2$ is defined by

\begin{align*}(\mu_1*\mu_2)(A)=\int_E\mu_1(A-x)\mu_2(\mathrm{d} x)\end{align*}

where $A-x=\{y\in E\;:\;x+y\in A\}$ . Then define independent random variables $Y_1, Y_2\sim\mu_Y$ and $Z\sim \mu_Z$ . Finally, construct Q as the conditional distribution of $X_2=f(Y_2+Z)$ given $X_1=f(Y_1+Z)$ . Then the probabilities in (3) are $\mathbb{P}(X_1\in A, X_2\in B)$ and $\mathbb{P}(X_1\in B, X_2\in A)$ , which are the same by construction. We refer to this as the convolution-type construction. A probability distribution can be written by convolution if it is infinitely divisible [Reference Sato55]. Therefore, many popular probability distributions, such as the normal distribution, the Student distribution and the gamma distribution, can be written by convolution. Three specific examples are given below to illustrate the convolutional design. The illustration of the Haar mixture kernel and the $\Delta$ -guided Metropolis kernel, described later, will also be based on these three kernels.

Let $\mathbb{R}_+=(0,\infty)$ . Let $I_d$ be the $d\times d$ identity matrix.

Example 1. (Autoregressive kernel.) We first describe the well-known autoregressive kernel resulting from the above convolution-type construction. Let $\rho\in (0,1]$ , let M be a $d\times d$ positive definite symmetric matrix, and let $x_0\in\mathbb{R}^d$ .

Further, let $\mathcal{N}_d(x, M)$ be the normal distribution with mean $x\in\mathbb{R}^d$ and covariance matrix M. By the reproductive property of the normal distribution, $\mu=\mathcal{N}_d(x_0, M)$ is a convolution of probability measures $\mu_Y=\mathcal{N}_d(0,\rho M)$ and $\mu_Z=\mathcal{N}_d(0,(1-\rho) M)$ with $f(x)=x_0+x$ in the notation above. Then the random variables $X_1$ and $X_2$ in the above notation follow $\mu$ with covariance

\begin{align*}\operatorname{Cov}(X_1,X_2)=\operatorname{Var}(Z)=(1-\rho)M. \end{align*}

By the change-of-variables formula, the conditional distribution $Q(x,\cdot)=\mathbb{P}(X_2\in\cdot\ |X_1=x)$ is the autoregressive kernel, which is defined as

\begin{align*}Q(x,\cdot)=\mathcal{N}_d(x_0+(1-\rho)^{1/2}\;(x-x_0), \rho M). \end{align*}

By the nature of convolution, it is $\mu=\mathcal{N}_d(x_0, M)$ -reversible.

Example 2. (Beta--gamma kernel.) Let $\mathcal{G}(\nu,\alpha)$ be the gamma distribution with shape parameter $\nu$ and rate parameter $\alpha$ . Let $\mu=\mathcal{G}(k,1)$ , $\mu_Y=\mathcal{G}(k(1-\rho), 1)$ , $\mu_Z=\mathcal{G}(k\rho,1)$ , and $f(x)=x$ , where $k\in\mathbb{R}_+$ and $\rho\in (0,1)$ . The conditional distribution of $b\;:\!=\;Z/X_1$ given $X_1$ , in the notation above, is $\mathcal{B}e(k\rho,k(1-\rho))$ , where $\mathcal{B}e(\alpha,\beta)$ is the beta distribution with shape parameters $\alpha$ and $\beta$ . Therefore, the conditional distribution $Q(x,\mathrm{d} y)=\mathbb{P}(X_2\in\mathrm{d} y|X_1=x)$ on $E=\mathbb{R}_+$ , called the beta–gamma (autoregressive) kernel in this paper, is given by

\begin{align*}y = b x+ c,\ b\sim\mathcal{B}e(k\rho,k(1-\rho)),\ c\sim\mathcal{G}(k(1-\rho),1),\end{align*}

where b, c are independent, and c corresponds to $Y_2$ in the above notation. The kernel is $\mu=\mathcal{G}(k,1)$ -reversible by construction. See [Reference Lewis, McKenzie and Hugus33].

Example 3. (Chi-squared kernel.) We construct a $\mu=\mathcal{G}(L/2,1/2)$ -reversible kernel for $L\in\mathbb{N}$ . Let $\mu_Y=\mathcal{N}_L(0,\rho I_L)$ , $\mu_Z=\mathcal{N}_L(0,(1-\rho) I_L)$ , and $f(x_1,\ldots, x_L)=\sum_{l=1}^Lx_l^2$ . By the reproductive property, if $Y_1, Y_2\sim\mu_Y$ and $Z\sim\mu_Z$ , then $X_i'\;:\!=\;Y_i+Z\sim\mathcal{N}_L(0, I_L)$ . Therefore, $X_i=f(X_i')\sim\mu$ since $\mu$ is the chi-squared distribution with L degrees of freedom. The conditional distribution $Q(x,\mathrm{d}y)=\mathbb{P}(X_2\in\mathrm{d}y|X_1=x)$ is $\mu$ -reversible by construction. We show that the conditional distribution is given by

(4) \begin{equation}y = \left[\left\{(1-\rho)\;x\right\}^{1/2}+\rho^{1/2}\;w_1\right]^2+\sum_{l=2}^L\rho\;w_l^2,\end{equation}

where $w_1,\ldots, w_L$ are independent and follow the standard normal distribution. To see this, first note that the law of $\rho^{-1/2}X_2'$ given $X_1'=x'$ is $\mathcal{N}_L(\rho^{-1/2}(1-\rho)^{1/2}x',I_L)$ . Then the law of $\rho^{-1}X_2=f(\rho^{-1/2}X_2')$ given $X_1'=x'$ is the non-central chi-squared distribution with L degrees of freedom and the non-central parameter $f(\rho^{-1/2}(1-\rho)^{1/2}x')=\rho^{-1}(1-\rho)x$ . The expression (4) follows from the properties of the non-central chi-squared distribution.

The Metropolis algorithm is a clever way to construct a reversible Markov kernel with respect to a given probability measure $\Pi$ . The following definition is somewhat broader than the usual one. It even includes the independent Metropolis–Hastings kernel, which is usually classified as a Metropolis–Hastings kernel and not a Metropolis kernel. An important feature of this kernel compared to the more general Metropolis–Hastings kernel is that we do not need to know the explicit density function of the proposed Markov kernel $Q(x,\cdot)$ . The idea behind this naming is that if the acceptance probability can be written explicitly in $\mu$ and $\Pi$ , it is called a Metropolis kernel, and if it also depends on Q, it is called a Metropolis–Hastings kernel.

Definition 1. (Metropolis kernel.) Let $\mu$ be a measure, and let $\Pi$ be a probability measure with probability density function $\pi(x)$ with respect to $\mu$ . Let Q be a $\mu$ -reversible Markov kernel. A Markov kernel P is called a Metropolis kernel of $(Q,\Pi)$ if

\begin{equation}\begin{split}P(x,\mathrm{d} y)&=Q(x,\mathrm{d} y)\alpha(x,y)+\delta_x(\mathrm{d} y)\left\{1-\int_EQ(x,\mathrm{d} y)\alpha(x,y)\right\}\end{split}\nonumber\end{equation}

for

(5) \begin{equation} \alpha(x,y)=\min\left\{1, \frac{\pi(y)}{\pi(x)}\right\}. \end{equation}

The function $\alpha$ is called the acceptance probability, and the Markov kernel Q is called the proposal kernel.

A Metropolis kernel P is $\Pi$ -reversible. It is easy to create a Metropolis version of the proposal kernels presented in Examples 1–3.

2.2. Haar mixture kernel

We introduce Markov kernels using the Haar measure. The Haar measure enables us to construct a random walk on a locally compact topological group, which is a crucial step towards obtaining non-reversible Markov kernels in this paper. The connection between the Markov kernels and the random walk will be made clear in Section 3, and the connection with non-reversible Markov kernels will be clear in Section 4.

The idea of constructing Haar mixture kernels is to introduce an auxiliary variable g corresponding to the scaling parameter or the shift parameter of the state space. We set a prior distribution on g. In each Markov chain Monte Carlo iteration, before the transition kernel generates a new proposal for the state space, we generate the parameter g based on its prior distribution and the conditional distribution given the state space. The Haar mixture kernel uses the Haar measure as the prior distribution of g. As remarked above, the reason for using the Haar measure will be made clear in later sections.

Let $(G,\times)$ be a locally compact topological group equipped with the Borel $\sigma$ -algebra. Let E be a left G-set. We assume that E is equipped with a $\sigma$ -algebra $\mathcal{E}$ and the left group action is jointly measurable. Let Q be a $\mu$ -reversible Markov kernel on $(E,\mathcal{E})$ , where $\mu$ is a $\sigma$ -finite measure. Let

\begin{align*}Q_g(x,A)=Q(gx, gA)\qquad (x\in E,\ A\in\mathcal{E},\ g\in G),\end{align*}

where $gA=\{gx\;:\;x\in A\}\in\mathcal{E}$ . Then $Q_g$ is $\mu_g$ -reversible, where

\begin{align*}\mu_g(A)=\mu(gA). \end{align*}

Let $\nu$ be the right Haar measure on G. It satisfies $\nu(Hg)=\nu(H)$ , where $Hg=\{hg\;:\;h\in H\} \subset G$ . Set

(6) \begin{equation}\mu_*(A)=\int_{g\in G}\mu_g(A)\nu(\mathrm{d} g)\ (A\in\mathcal{E}). \end{equation}

Assume that $\mu_*$ is a $\sigma$ -finite measure. Then $\mu_*$ is a left-invariant measure. Indeed, by the definitions of $\mu_*$ and $\mu_b$ , we have

\begin{align*} \mu_*(aA)=\int_{b\in G}\mu_b(aA)\nu(\mathrm{d} b)=\int_{b\in G}\mu(baA)\nu(\mathrm{d} b),\end{align*}

and by right-invariance of $\nu$ , we have

\begin{align*}\mu_*(aA)&=\int_{b\in G}\mu(bA)\nu(\mathrm{d} b)\\[5pt] &=\int_{b\in G}\mu_b(A)\nu(\mathrm{d} b)\\[5pt] &=\mu_*(A). \end{align*}

Suppose that $\mu$ is absolutely continuous with respect to $\mu_*$ . Then $(g,x)\mapsto \mathrm{d}\mu_g/\mathrm{d}\mu_*(x)$ is jointly measurable. This is because $\mathrm{d}\mu_g/\mathrm{d}\mu_*(x)=\mathrm{d}\mu/\mathrm{d}\mu_*(gx)$ by the left-invariance of $\mu_*$ , and $(g,x)\mapsto gx$ is assumed to be jointly measurable. Let

(7) \begin{equation}K(x,\mathrm{d} g)=\frac{\mathrm{d}\mu_{g}}{\mathrm{d}\mu_*}(x)\nu(\mathrm{d} g).\end{equation}

Define

(8) \begin{equation}Q_*(x, A)=\int_{g\in G}K(x,\mathrm{d} g)Q_g(x, A). \end{equation}

Example 4. (Autoregressive mixture kernel.) Consider the autoregressive kernel in Example 1. Let $E=\mathbb{R}^d$ , $G=(\mathbb{R}_+,\times)$ , and $\mu=\mathcal{N}(x_0,M)$ , and set $(g,x)\mapsto x_0+g^{1/2}(x-x_0)$ . Then the Haar measure is $\nu(\mathrm{d} g)\propto g^{-1}\mathrm{d} g$ . A simple calculation yields $\mu_g=\mathcal{N}_d(x_0, g^{-1}M)$ and $Q_g(x,\cdot)=\mathcal{N}_d(x_0+(1-\rho)^{1/2}\;(x-x_0), g^{-1}\rho M)$ . Also, $\mu_*(\mathrm{d} x)\propto (\Delta x)^{-d/2}\mathrm{d} x$ and $K(x,\mathrm{d} g)=\mathcal{G}(d/2, \Delta x/2)$ , where $\Delta x =(x-x_0)^{\top}M^{-1}(x-x_0)$ . We have a closed-form (up to a constant) expression for $Q_*(x,\cdot)$ as follows:

\begin{align*} Q_*(x,\mathrm{d}y)\propto \left[1+\frac{\Delta(y-(1-\rho)^{1/2}(x-x_0))}{\rho\Delta x}\right]^{-d}\mathrm{d} x. \end{align*}

Proof. Let $A, B\in\mathcal{E}$ . By the definitions of $Q_*$ and K, we have

\begin{align*} \int_{A}\mu_*(\mathrm{d} x)Q_*(x, B)&= \int_{g\in G}\int_{x\in A}\mu_*(\mathrm{d} x)K(x,\mathrm{d} g)Q_g(x, B)\\[5pt] &= \int_{g\in G}\int_{x\in A}\mu_g(\mathrm{d} x)Q_g(x, B)\nu(\mathrm{d} g), \end{align*}

and by $\mu_g$ -reversibility of $Q_g$ , we have

\begin{align*} \int_{A}\mu_*(\mathrm{d} x)Q_*(x, B)&=\int_{g\in G}\int_{x\in B}\mu_g(\mathrm{d} x)Q_g(x, A)\nu(\mathrm{d} g)\\[5pt] &=\int_{B}\mu_*(\mathrm{d} x)Q_*(x, A). \end{align*}

From this, we can define the following Metropolis kernel.

The Metropolis–Haar kernel is implemented through the following algorithm, where $\pi(x)=(\mathrm{d}\Pi/\mathrm{d}\mu_*)(x)$ . In the algorithm, $\mathcal{U}[0,1]$ is the uniform distribution on [0, 1].

Algorithm 1: Metropolis–Haar kernel

The Metropolis–Haar kernel is reversible, but important in its own right. The underlying reference measure $\mu_*$ is heavier than $\mu_g$ , which is expected to lead to a robust algorithm. Examples of Metropolis–Haar kernels will be described in Section 4.3.

3.1. Unbiasedness

In this section, we introduce the unbiasedness property for efficient construction of the non-reversible kernel. Any measurable map $\Delta\;:\;E\rightarrow G$ , where $G=(G,\le\!)$ is a totally ordered set, is called a statistic in this paper. In Section 4, a statistic $\Delta$ will guide a Markov kernel $Q(x,\mathrm{d} y)$ according to the auxiliary directional variable $i\in\{-,+\}$ as in [Reference Gustafson23]. When the positive direction $i=+$ is selected, y is sampled according to $Q(x,\mathrm{d} y)$ unless $\Delta x\le\Delta y$ by rejection sampling. If the negative direction $i=-$ is selected, y is sampled unless $\Delta y\le \Delta x$ . It is typical that one of the rejection sampling directions has high rejection probability (see Example 7). To avoid this inefficiency, we consider a class of Markov kernels Q such that the probabilities of the events $\Delta x\le\Delta y$ and $\Delta y\le \Delta x$ measured by $Q(x,\cdot)$ are the same. We say Q is unbiased if this property is satisfied. If unbiasedness is violated, the rejection sampling may be inefficient, because it takes a long time to exit the while loop of the rejection sampling. Therefore, the unbiasedness property is necessary for efficient construction of the non-reversible kernel in our approach.

Definition 4. ( $\Delta$ -unbiasedness.) Let $\Delta\;:\;E\rightarrow G$ be a statistic. We say a Markov kernel Q on E is $\Delta$ -unbiased if

\begin{align*}Q(x,\{y\in E\;:\;\Delta x\le\Delta y\})=Q(x,\{y\in E\;:\;\Delta y\le\Delta x\})\end{align*}

for any $x\in E$ . Also, we say that two statistics $\Delta$ and $\Delta'$ from E to possibly different totally ordered sets are equivalent if

\begin{equation}\begin{split}Q(x,\{y\in E\;:\;\Delta x\le \Delta y\}\ominus\{y\in E\;:\;\Delta' x\le \Delta' y\})&=0,\\[5pt] Q(x,\{y\in E\;:\;\Delta y\le \Delta x\}\ominus\{y\in E\;:\;\Delta' y\le \Delta' x\})&=0\end{split}\nonumber\end{equation}

for $x\in E$ , where $A\ominus B=(A\cap B^c)\cup (A^c\cap B)$ .

If $\Delta$ and $\Delta'$ are equivalent, then $\Delta$ -unbiasedness implies $\Delta'$ -unbiasedness.

Example 7. (Random-walk kernel.) Let $v^{\top}$ be the transpose of $v\in\mathbb{R}^d$ , and let $\Gamma$ be a probability measure on $\mathbb{R}^d$ which is symmetric about the origin; that is, $\Gamma(A)=\Gamma(\!-A)$ for $-A=\{x\in E\;:\; -\!x\in A\}$ . Let $Q(x,A)=\Gamma(A-x)$ . Then Q is $\Delta$ -unbiased for $\Delta x=v^{\top}x$ for some $v\in\mathbb{R}^d$ , since

\begin{equation*}\begin{split}Q(x,\{y\;:\;\Delta x\le \Delta y\})& =\Gamma(\{z\;:\;0\le v^{\top}z\})\\[5pt] & =\Gamma(\{z\;:\;v^{\top}z\le 0\}). \end{split}\end{equation*}

On the other hand, Q is not $\Delta'$ -unbiased for $\Delta' x=x_1^2+\cdots +x_d^2$ , where $x=(x_1,\ldots, x_d)$ , if $\Gamma$ is not the Dirac measure centred on $(0,\ldots,0)$ . In particular, if $\Gamma(\{(0,\ldots, 0)\})=0$ , then $Q(x, \{\Delta' y\le \Delta' x\})=0$ for $x=(0,\ldots, 0)$ .

3.2. Random-walk property

Constructing a $\Delta$ -unbiased Markov kernel is a crucial step for our approach. However, determining how to construct a $\Delta$ -unbiased Markov kernel is non-trivial. The random-walk property is the key for this construction.

Let G be a topological group.

Definition 5. ( $(\Delta,\Gamma)$ -random-walk.) A Markov kernel $Q(x,\mathrm{d} y)$ has the $(\Delta,\Gamma)$ -random-walk property if there is a function $\Delta\;:\;E\rightarrow G$ with a probability measure $\Gamma$ on a topological group G such that $\Gamma(H)=\Gamma(H^{-1})$ for any Borel set H of G and

(9) \begin{equation} Q(x, \{y\in E\;:\;\Delta y\in H\})=\Gamma((\Delta x)^{-1}H). \end{equation}

Here, $H^{-1}=\{g\in G\;:\;g^{-1}\in H\}$ .

A typical example of a Markov kernel with the $(\Delta,\Gamma)$ -random-walk property is given in Example 7. We assume that $(G,\le\!)$ is an ordered group.

Proof. Let $H=[\Delta x,+\infty)=\{g\in H\;:\;\Delta x\le g\}$ . Then for the unit element e,

\begin{align*}Q(x,\{y\in E\;:\;\Delta x\le\Delta y\})&=Q(x,\{y\in E\;:\;\Delta y\in H\})\\[5pt] &=\Gamma((\Delta x)^{-1}H)=\Gamma([e,+\infty)).\end{align*}

Similarly, $Q(x,\{y\in E\;:\;\Delta y\le\Delta x\})=\Gamma((\!-\!\infty, e])$ . Since $[e,+\infty)^{-1}=(\!-\!\infty,e]$ , and since $\Gamma(H)=\Gamma(H^{-1})$ , Q is $\Delta$ -unbiased.

3.3. Sufficiency

So far in this section, we have introduced $\Delta$ -unbiasedness, which is the important property for the $\Delta$ -guided Metropolis kernel in Section 3.1. In Section 3.2, we showed that the $(\Delta,\Gamma)$ -random-walk property is sufficient for $\Delta$ -unbiasedness. In this section we will show that for the Haar mixture kernel, the sufficiency property introduced below is sufficient for the $(\Delta,\Gamma)$ -random-walk property, and thus for the $\Delta$ -unbiasedness property.

We would like to mention the intuition behind the sufficiency property. In general, the conditional law of $\Delta y$ given x is not completely determined by $\Delta x$ . If it is completely determined by $\Delta x$ , we call $\Delta$ sufficient. If $\Delta$ is sufficient, the equation (9) is satisfied, although $\Gamma$ is not symmetric in general. When Q is the Haar mixture kernel with some additional technical conditions, we will show that $\Gamma$ is symmetric thanks to the Haar measure property.

Let $(G,\times)$ be a unimodular locally compact topological group. Also, let $(G,\le\!)$ be an ordered group, and let E be a left G-set. In this paper, a statistic $\Delta\;:\;E\rightarrow G$ is called a G-statistic if $\Delta gx=g\Delta x$ for $g\in G$ and $x\in E$ . For a $\sigma$ -finite measure $\Pi$ on E and a G-statistic $\Delta\;:\;E\rightarrow G$ , let $\widehat{\Pi}=\Pi\circ\Delta^{-1}$ , that is, the image measure of $\Pi$ under $\Delta$ . Let $\widehat{\mu}_*$ be the image measure of $\mu_*$ under $\Delta$ . Then it is a left Haar measure, since

\begin{align*}\widehat{\mu}_*(gH)&=\mu_*(\{y\in E\;:\;\Delta y\in gH\})\\[5pt] &=\mu_*(\{y\in E\;:\;\Delta( g^{-1}y)\in H\})\end{align*}

by the property $\Delta (g^{-1}y)=g^{-1}\Delta y$ , and

\begin{align*}\widehat{\mu}_*(gH)&=\mu_*(\{y\in E\;:\;\Delta y\in H\})=\widehat{\mu}_*(H)\end{align*}

by the left-invariance of $\mu_*$ . Since G is unimodular, the left Haar measure $\widehat{\mu}_*$ and the right Haar measure $\nu$ coincide up to a multiplicative constant. From this fact, we can assume

\begin{align*}\widehat{\mu}_*=\nu\end{align*}

without loss of generality. By construction, $\mu$ and $\mu_*$ are measures on E, and $\hat{\mu}$ , $\hat{\mu}_*$ , and $\nu$ are measures on G. Let Q be a $\mu$ -reversible kernel.

Definition 6. (Sufficiency.) Let $\mu$ be a $\sigma$ -finite measure. We call a G-statistic $\Delta$ sufficient for $(\nu, \mu, Q)$ if there is a Markov kernel $\widehat{Q}$ and a measurable function $h_1$ on G such that

\begin{align*}Q(x, \{y\in E\;:\;\Delta y\in H\})=\widehat{Q}(\Delta x, H)\end{align*}

and

\begin{align*}\frac{\mathrm{d} \mu}{\mathrm{d}\mu_*}(x)=h_1(\Delta x)\end{align*}

$\mu_*$ -almost surely.

Suppose that $\Delta$ is sufficient for $(\nu, \mu, Q)$ , and let $X_n$ be a Markov chain with transition kernel Q. Then the law of $\Delta X_n$ given $X_{n-1}=x$ depends on x only through $\Delta x$ . Moreover, $K(x,\mathrm{d} g)$ depends on x only through $\Delta x$ . More precisely, by the left-invariance of $\mu_*$ , we have

(10) \begin{equation} \frac{\mathrm{d}\mu_g}{\mathrm{d}\mu_*}(x)=h_1(g\Delta x)\end{equation}

since

\begin{align*} \mu_g(A)&=\mu(gA)=\int_{gA} h_1(\Delta x)\mu_*(\mathrm{d} x)=\int_{A} h_1(g\Delta x)\mu_*(\mathrm{d} x). \end{align*}

Let $\widehat{\mu}$ be the image measure of $\mu$ under $\Delta$ . Then $\widehat{Q}$ is $\widehat{\mu}$ -reversible and

(11) \begin{equation} \frac{\mathrm{d}\widehat{\mu}}{\mathrm{d}\nu}(a)=\frac{\mathrm{d}\widehat{\mu}}{\mathrm{d}\widehat{\mu}_*}(a)=h_1(a). \end{equation}

Example 8. (Sufficiency of the autoregressive mixture kernel.) Consider the autoregressive kernel Q and the measure $\mu$ in Example 1 and the statistic $\Delta$ defined in Example 4. We show that $\Delta $ is sufficient for $(\nu,\mu,Q)$ . The Markov kernel $Q(x,\mathrm{d} y)$ corresponds to the update

\begin{align*}y \leftarrow x_0+(1-\rho)^{1/2}(x-x_0)+\rho^{1/2}\;M^{1/2}\;w\end{align*}

where $w\sim\mathcal{N}_d(0, I_d)$ . For $\xi=(1-\rho)^{1/2}\rho^{-1/2}M^{-1/2}(x-x_0)$ ,

\begin{align*}\Delta y=\rho\left\|\xi+w\right\|^2,\end{align*}

where $\|\cdot\|$ is the Euclidean norm. Therefore, $\rho^{-1}\Delta y$ conditioned on x follows the non-central chi-squared distribution with d degrees of freedom and non-central parameter $\|\xi\|^2=(1-\rho)\rho^{-1}\Delta x$ . Hence, the law of $\Delta y$ depends on x only through $\Delta x$ , and thus there exists a Markov kernel $\widehat{Q}$ as in Definition 6. Also, a simple calculation yields $h_1(g)\propto g^{d/2}\exp(\!-\!g/2)$ . Therefore, $\Delta$ is sufficient.

For a measure $\nu$ , we write $\nu^{\otimes k}$ for the kth product of $\nu$ , defined by

\begin{align*}\nu^{\otimes k}(\mathrm{d} x_1\cdots\mathrm{d} x_k)=\nu(\mathrm{d} x_1)\cdots\nu(\mathrm{d} x_k),\end{align*}

for $k\in\mathbb{N}$ .

Proof. Let h(a, b) be the Radon–Nikodým derivative:

\begin{equation}h(a,b)\nu(\mathrm{d} a)\nu(\mathrm{d} b)=\widehat{\mu}(\mathrm{d} a)\widehat{Q}(a,\mathrm{d} b). \nonumber\end{equation}

By the $\widehat{\mu}$ -reversibility of $\widehat{Q}$ , $h(a,b)=h(b,a)$ almost surely. From the sufficiency property, we can rewrite $h_1$ and $\widehat{Q}$ in terms of h(a, b) and $\nu$ :

(12) \begin{equation}\begin{cases}&h_1(a)=\int_{b\in G}h(a,b)\nu(\mathrm{d} b),\\[5pt] &h_1(a)\widehat{Q}(a,\mathrm{d} b)=h(a,b)\nu(\mathrm{d} b),\end{cases}\end{equation}

$\nu$ -almost surely. By the definition of $Q_*$ , K, and $\widehat{Q}$ , we have

\begin{align*}Q_*(x,\{y\;:\;\Delta y\in H\})=& \int_{a\in G}K(x,\mathrm{d} a)Q(ax, \{y\;:\;\Delta y\in aH\})\\[5pt] =& \int_{a\in G}\frac{\mathrm{d}\mu_a}{\mathrm{d}\mu_*}(x)\nu(\mathrm{d} a)\widehat{Q}(a \Delta x, aH), \end{align*}

and also, by (10) and (12), we have

\begin{align*}Q_*(x,\{y\;:\;\Delta y\in H\})=& \int_{a\in G}h_1(a\Delta x)\widehat{Q}(a \Delta x, aH)\nu(\mathrm{d} a)\\[5pt] =& \int_{a\in G}\int_{b\in H}h(a\Delta x, ab)\nu(\mathrm{d} a)\nu(\mathrm{d} b)\\[5pt] =& \int_{a\in G}\int_{b\in H}h(a, a(\Delta x)^{-1}b)\nu(\mathrm{d} a)\nu(\mathrm{d} b),\end{align*}

where the last equality follows from the right-invariance of $\nu$ . Let

\begin{align*} \widehat{h}(b)= \int_{a\in G}h(a,ab)\nu(\mathrm{d} a). \end{align*}

From $h(a,b)=h(b,a)$ ,

\begin{align*}\widehat{h}(b^{-1})&= \int_{a\in G}h(a,ab^{-1})\nu(\mathrm{d} a)\\[5pt] &= \int_{a\in G}h(ab,a)\nu(\mathrm{d} a)=\widehat{h}(b). \end{align*}

By using $\widehat{h}$ , we can write

\begin{align*}Q_*(x,\{y\;:\;\Delta y\in H\})&= \int_{b\in H}\widehat{h}((\Delta x)^{-1}b)\nu(\mathrm{d} b)\\[5pt] & = \int_{b\in (\Delta x)^{-1}H}\widehat{h}(b)\nu(\mathrm{d} b). \end{align*}

The above is guaranteed to have the $(\Delta,\Gamma)$ -random-walk property, where we introduce $\Gamma(H)=\int_{a\in H}\widehat{h}(a)\nu(\mathrm{d} a)$ , because

\begin{align*}Q(x,\{y\;:\;\Delta y\in H\})=\Gamma((\Delta x)^{-1}H)\end{align*}

and

\begin{align*}\Gamma(H^{-1})&=\int_{a^{-1}\in H}\widehat{h}(a)\nu(\mathrm{d} a)=\int_{a\in H}\widehat{h}(a)\nu(\mathrm{d} a)=\Gamma(H). \end{align*}

Hence, it is $\Delta$ -unbiased by Proposition 2.

3.4. Multivariate versions of one-dimensional kernels

Essentially, we have introduced three Markov kernels: the autoregressive kernel, the chi-squared kernel, and the beta–gamma kernel. The state space of the first kernel is a general Euclidean space, and that of the last two kernels is a subspace of the one-dimensional Euclidean space. In this subsection, we consider the multivariate versions of the latter two kernels.

We present different strategies for the two kernels. For the chi-squared kernel, there is a sophisticated structure that allows for a multivariate version of the state space. For the beta–gamma kernel, there does not seem to be a special structure, and so we apply a general approach which does not require any structure. First we show how to construct a multivariate extension for the chi-squared kernel.

Example 10. (Multivariate chi-squared mixture kernel.) For the chi-squared kernel (Examples 3 and 6), we use the operation $(g,x)\mapsto (gx_1,\ldots, gx_d)$ with $G=\mathbb{R}_+$ and $E=\mathbb{R}_+^d$ . Let Q be the Markov kernel defined in Example 3. Let

\begin{align*}\mathcal{Q}(x,\mathrm{d} y)=Q(x_1,\mathrm{d} y_1)\cdots Q(x_d,\mathrm{d} y_d)\end{align*}

and $\mu(\mathrm{d} x)=\mathcal{G}(L/2,1/2)^{\otimes d}$ . Let $\Delta x=x_1+\cdots +x_d$ . In this case, $\nu(\mathrm{d} g)\propto g^{-1}\mathrm{d} g$ and $\mu_g(\mathrm{d} x)=\mathcal{G}(L/2,g/2)^{\otimes d}$ , and $\mathcal{Q}_g$ on $\mathbb{R}^d_+$ is the product of $Q_g$ on $\mathbb{R}_+$ defined in Example 6; that is,

\begin{align*}\mathcal{Q}_g(x,\mathrm{d} y)=Q_g(x_1,\mathrm{d} y_1)\cdots Q_g(x_d,\mathrm{d} y_d). \end{align*}

Then

\begin{align*}\mu_*(\mathrm{d} x)\propto (x_1\cdots x_d)^{L/2-1}(\Delta x)^{-dL/2}\mathrm{d} x_1\cdots\mathrm{d} x_d\end{align*}

and $K(x,\mathrm{d} g)=\mathcal{G}(Ld/2, \Delta x/2)$ . From this expression, $h_1(g)\propto g^{dL/2}\exp(\!-\!g/2)$ . Moreover, by the property of the non-central chi-squared distribution, the law of $\rho^{-1}\Delta y$ where $y\sim \mathcal{Q}(x,\mathrm{d} y)$ is the non-central chi-squared distribution with dL degrees of freedom and with the non-central parameter $(1-\rho)\rho^{-1}\Delta x$ . Therefore there exists a Markov kernel $\widehat{\mathcal{Q}}(g,\cdot)$ which is the scaled non-central chi-squared distribution for each g. It is not difficult to check that $\widehat{Q}$ has a density function with respect to $\nu$ . The statistic $\Delta$ is sufficient, and the multivariate version of chi-squared mixture kernel $\mathcal{Q}_*$ is $\Delta$ -unbiased from this fact.

Example 11. (Multivariate beta--gamma mixture kernel.) For the beta–gamma kernel (Examples 2 and 5), we use the operation $(g,x)\mapsto (g_1x_1,\ldots, g_dx_d)$ , with $G=(\mathbb{R}_+^d,\times)$ and $E=\mathbb{R}_+^d$ , where $g=(g_1,\ldots, g_d)$ and $x=(x_1,\ldots, x_d)$ . We define the binary operation of G by $(x,y)\mapsto (x_1y_1,\cdots, x_dy_d)$ and the identity element by $e=(1,\ldots, 1)$ . In this case, the Markov kernel $\mathcal{Q}_g$ on $\mathbb{R}^d$ is the product of $Q_g$ on $\mathbb{R}$ defined in Example 5; that is,

\begin{align*}\mathcal{Q}_g(x,\mathrm{d} y)=Q_{g_1}(x_1,\mathrm{d} y_1)\cdots Q_{g_d}(x_d,\mathrm{d} y_d). \end{align*}

Also, we have $K(x,\mathrm{d} g)=\mathcal{G}(k, x_1)\cdots\mathcal{G}(k, x_d)$ and $\mu_*(\mathrm{d} x)=(x_1\cdots x_d)^{-1}\mathrm{d} x_1\cdots \mathrm{d} x_d $ . The G-statistic $\Delta x=x$ is sufficient, and hence the multivariate version of the beta–gamma mixture kernel $\mathcal{Q}_*$ is $\Delta$ -unbiased by Proposition 3.

For $G=\mathbb{R}_+^d$ in Example 11, several types of order relations are possible. Any ordering will do as long as (2) is satisfied. The popular lexicographic order depends on how we index the coordinates. To avoid this unfavourable property, we consider the modified lexicographic order defined below.

Example 12. (Modified lexicographic order.) Let $G=(\mathbb{R}_+^d,\times)$ . For $x=(x_1,\ldots, x_d)\in G$ , let

\begin{align*}s(x)_i=x_i\times\cdots\times x_d\end{align*}

be a partial product of the vector x from the ith element to the dth element. A version of lexicographic order $\le $ can be defined as follows. Counting from $i=1,\ldots, d$ ,

• if $s(x)_i=s(y)_i$ for all i or

• if the first index i such that $s(x)_i\neq s(y)_i$ satisfies $s(x)_i<s(y)_i$ ,

then we write $x\le y$ . It is not difficult to check that this ordering satisfies (2).

Since (2) is satisfied, the multivariate beta–gamma mixture kernel is $\Delta$ -unbiased with this order for G, where $\Delta$ is the modified lexicographic order. Note that the modified lexicographic order $\Delta$ still has the same problem as the (unmodified) lexicographic order; that is, it depends how we index the coordinates. However, the problem occurs with probability 0. This is because the first step of the sort (i.e. $s(x)_1<s(y)_1$ or $s(y)_1<s(x)_1$ ) does not depend on the order of the indices, and the first step determines the order with probability 1. Indeed, by construction,

\begin{align*} Q(\{y\in E\;:\;\Delta x\le \Delta y\})&= Q(\{y\in E\;:\;s(x)_1\le s(y)_1\})\end{align*}

since $s(x)_1=s(y)_1$ occurs with probability 0. More precisely,

\begin{align*}\Delta(x_1,\ldots, x_d)=(x_1,\ldots, x_d)\end{align*}

with the modified lexicographic ordering and

\begin{align*}\Delta'(x_1,\ldots,x_d)=x_1\times\cdots\times x_d\end{align*}

in $\mathbb{R}_+$ with the usual ordering are equivalent in the sense of Definition 4, because $\Delta'(x)=s(x)_1$ . In particular, the multivariate beta–gamma mixture kernel is $\Delta'$ -unbiased since the kernel is $\Delta$ -unbiased. Note that $\Delta'$ is not a G-statistic, since it does not satisfy $\Delta'gx=g\Delta'x$ .

4.1. $\Delta$ -guided Metropolis kernel

Definition 7. ( $\Delta$ -guided Metropolis kernel.) For $\Delta$ -unbiased Markov kernel Q, a probability measure $\Pi$ , and a measurable function $\alpha\;:\;E\times E\rightarrow [0,1]$ defined in (5), we say a Markov kernel $P_G$ on $E\times \{-,+\}$ is the $\Delta$ -guided Metropolis kernel of $(Q,\Pi)$ if

\begin{align*}P_G(x,+,\mathrm{d} y,+)&=Q_+(x,\mathrm{d} y)\alpha(x,y)\\[5pt] P_G(x,+,\mathrm{d} y,-)&=\delta_x(\mathrm{d} y)\left\{1-\int_EQ_+(x,\mathrm{d} y)\alpha(x,y)\right\}\\[5pt] P_G(x,-,\mathrm{d} y,-)&=Q_-(x,\mathrm{d} y)\alpha(x,y)\\[5pt] P_G(x,-,\mathrm{d} y,+)&=\delta_x(\mathrm{d} y)\left\{1-\int_EQ_-(x,\mathrm{d} y)\alpha(x,y)\right\},\end{align*}

where

\begin{align*}Q_+(x,\mathrm{d} y)&=2Q(x,\mathrm{d} y)1_{\{\Delta x <\Delta y\}}+Q(x,\mathrm{d} y)1_{\{\Delta x=\Delta y\}},\\[5pt] Q_-(x,\mathrm{d} y)&=2Q(x,\mathrm{d} y)1_{\{\Delta y<\Delta x\}}+Q(x,\mathrm{d} y)1_{\{\Delta x=\Delta y\}}. \end{align*}

The Markov kernel $P_G$ satisfies the so-called $\Pi_G$ -skew-reversible property

\begin{equation}\begin{split}\Pi_G(\mathrm{d} x,+)P_G(x,+,\mathrm{d} y,+)&=\Pi_G(\mathrm{d} y,-)P_G(y,-,\mathrm{d} x,-),\\[5pt] \Pi_G(\mathrm{d} x,+)P_G(x,+,\mathrm{d} y,-)&=\Pi_G(\mathrm{d} y,-)P_G(y,-,\mathrm{d} x,+),\end{split}\nonumber\end{equation}

where

\begin{align*}\Pi_G=\Pi\otimes(\delta_{-}+\delta_{+})/2. \end{align*}

Here, for probability measures $\nu$ and $\mu$ , $(\nu\otimes\mu)(\mathrm{d} x\mathrm{d} y)=\nu(\mathrm{d} x)\mu(\mathrm{d} y)$ . With this property, it is straightforward to check that $P_G$ is $\Pi_G$ -invariant.

As described in Proposition 3, a Haar mixture kernel $Q_*$ is $\Delta$ -unbiased if $\Delta$ is sufficient and some other technical conditions are satisfied. Therefore, we can construct a $\Delta$ -guided Metropolis kernel $(Q_*,\Pi)$ using the Haar mixture kernel $Q_*$ .

The $\Delta$ -guided Metropolis–Haar kernel is given as Algorithm 2, where we let $\pi(x)=\mathrm{d}\Pi/\mathrm{d}\mu_*(x)$ . This Metropolis–Haar kernel is further discussed in detail in Sections 4.2 and 4.3.

Algorithm 2: $\Delta$ -guided Metropolis–Haar kernel

Let P be the Metropolis kernel of $(Q,\Pi)$ . We now see that $P_G$ is always expected to be better than P in the sense of the asymptotic variance corresponding to the central limit theorem. The inner product $\langle f,g\rangle=\int f(x)g(x)\Pi(\mathrm{d} x)$ and the norm $\|f\|=(\langle f,f\rangle)^{1/2}$ can be defined on the space of $\Pi$ -square integrable functions. Let $(X_0, X_1,\ldots\!)$ be a Markov chain with Markov kernel P and $X_0\sim \Pi$ . Then we define the asymptotic variance

\begin{align*}\operatorname{Var}(f,P)=\lim_{N\rightarrow\infty}\operatorname{Var}\left(N^{-1/2}\sum_{n=1}^Nf(X_n)\right)\end{align*}

if the right-hand side exists. The existence of the right-hand side limit is a kernel-specific problem and is not addressed here. Let $\lambda \in [0,1)$ . As in [Reference Andrieu1], to avoid a kernel-specific argument, we consider a pseudo-asymptotic variance

\begin{align*}\operatorname{Var}_\lambda(f,P)=\|f_0\|^2+2\sum_{n=1}^\infty\lambda^n\langle f_0,P^n f_0\rangle,\end{align*}

where $f_0=f-\Pi(f)$ , which always exists. Under some conditions, $\lim_{\lambda\uparrow 1-}\operatorname{Var}_\lambda(f,P)=\operatorname{Var}(f,P)$ . We can also define $\operatorname{Var}_\lambda(f,P_G)$ for a $\Pi$ -square integrable function f on E by considering $f((x,i))=f(x)$ .

By taking $\lambda\uparrow 1$ , we can expect that the non-reversible kernel $P_G$ is better than P in the sense of having smaller asymptotic variance.

4.2. Step-by-step instructions for creating a $\Delta$ -guided Metropolis–Haar kernel

Below is a set of necessary conditions to build a Haar mixture kernel $Q_*$ and a Metropolis–Haar kernel $(Q_*,\Pi)$ :

1. 1. $G=(G,\times)$ is a locally compact topological group equipped with the Borel $\sigma$ -algebra and the right Haar measure $\nu$ .

2. 2. The state space E is a left G-set.

3. 3. The measure $\mu$ is a $\sigma$ -finite measure and Q is a $\mu$ -reversible Markov kernel on $(E,\mathcal{E})$ .

4. 4. There exists a Markov kernel $K(x,\mathrm{d} g)$ as in (7).

Then we can construct a Haar–mixture kernel $Q_*$ as in (8). Below is an additional set of necessary conditions to build a $\Delta$ -guided Metropolis–Haar kernel:

1. 1. $G=(G,\le\!)$ is an ordered group, and $G=(G,\times)$ is a unimodular locally compact topological group.

2. 2. $\Delta$ is a G-statistic.

3. 3. $\Delta$ is sufficient for Q.

Based on the above necessary conditions, we think it is not difficult to construct the Haar mixture kernel $Q_*$ as in Algorithm 2. In practice, we also need to think about the efficiency of sampling from $K(x,\mathrm{d} g)$ and the cost of evaluating $\Delta x$ , and we will present the details with some concrete examples in the next section.

4.3. Examples of $\Delta$ -guided Metropolis–Haar kernels

Here we present some of the $\Delta$ -guided Metropolis–Haar kernels.

5.1. $\Delta$ -guided Metropolis–Haar kernel on $\mathbb{R}^d$

In this simulation, we consider the autoregressive-based kernel considered in Example 14. More precisely, we study the preconditioned Crank–Nicolson kernel, the mixed-preconditioned Crank–Nicolson kernel, and the $\Delta$ -guided mixed preconditioned Crank–Nicolson kernel. The random-walk Metropolis kernel is also compared for reference. All these methods are gradient-free methods, in the sense that the proposal kernel does not use the derivative of $\log\pi(x)$ . Although this may sound daunting, sometimes a simple structure leads to robustness and efficiency, as shown through simulation experiments. Moreover, parameter tuning for these Markov kernels based on a reversible proposal kernel is relatively straightforward. We can learn the parameters of the reference measure $\mu$ or $\mu_*$ using the standard technique, by treating the Markov chain Monte Carlo outputs as if they came from identical and independent observations of $\mu$ or $\mu_*$ , even though $\mu_*$ is generally an improper distribution. Other parameters, such as step size, can be tuned by monitoring the acceptance probability. Since parameter tuning is not our main focus, we do not elaborate on this point in this paper.

We also compare these methods with gradient-based, informed algorithms. The Metropolis-adjusted Langevin algorithm [Reference Roberts and Tweedie52, Reference Rossky, Doll and Friedman54] and the Hamiltonian Monte Carlo algorithm [Reference Duane, Kennedy, Pendleton and Roweth17, Reference Neal40] are popular gradient-based algorithms. Furthermore, we consider methods that use both gradient-based and autoregressive-kernel-based ideas. This class includes, for example, the infinite-dimensional Metropolis-adjusted Langevin algorithm [Reference Beskos, Roberts, Stuart and Voss8, Reference Cotter, Roberts, Stuart and White13], a marginal sampler proposed in [Reference Titsias and Papaspiliopoulos60] which we will refer to as marginal gradient-based sampling, and the infinite-dimensional Hamiltonian Monte Carlo [Reference Beskos4, Reference Neal40, Reference Ottobre, Pillai, Pinski and Stuart43].

We performed all experiments using a desktop computer with 6 Intel i7-5930K (3.50GHz) CPU cores. All algorithms other than the Hamiltonian Monte Carlo algorithm were coded in R, Version 3.6.3 [47], using the RcppArmadillo package, Version 0.9.850.1.0 [Reference Eddelbuettel and Sanderson18]. The results for the Hamiltonian Monte Carlo algorithm were obtained using RStan, Version 2.19.3 [59]. For a fair comparison, we used a single core and chain for RStan. The code for all experiments is available in the online repository at https://github.com/Xiaolin-Song/Non-reversible-guided-Metropolis-kernel.

5.1.1. Discrete observation of stochastic diffusion process.

First we consider a problem in which it is difficult to apply gradient-based Markov chain Monte Carlo methodologies because of the high cost of derivative calculation. Let $\alpha\in\mathbb{R}^k$ . Suppose that $(X_t)_{t\in [0,T]}$ is a solution process of a stochastic differential equation

\begin{align*}\mathrm{d} X_t=a(X_t,\alpha)\mathrm{d} t+b(X_t)\mathrm{d} W_t;\; X_0=x_0,\end{align*}

where $(W_t)_{t\in [0,T]}$ is the d-dimensional standard Wiener process, and $a\;:\;\mathbb{R}^d\times\mathbb{R}^k\rightarrow \mathbb{R}^d$ and $b\;:\;\mathbb{R}^d\rightarrow \mathbb{R}^{d\times d}$ are the drift and diffusion coefficient, respectively. We only observe $X_0, X_h, X_{2h},\ldots, X_{Nh}$ , where $N\in\mathbb{N}$ and $h=T/N$ .

We consider a Bayesian inference based on the local Gaussian approximated likelihood, since an explicit form of the probability density function is not available in general. The local Gaussian approximation approach, including the simple least-squares estimate approach, has been studied in, for example, [Reference Florens-zmirou19, Reference Prakasa Rao45, Reference Prakasa Rao46, Reference Yoshida66]. See also [Reference Beskos, Papaspiliopoulos and Roberts5, Reference Beskos, Papaspiliopoulos, Roberts and Fearnhead6] for a non-local Gaussian approach based on unbiased estimation of the likelihood.

We consider a Bayesian inference for $\alpha\in \mathbb{R}^{50}$ using the local Gaussian approximated likelihood. We set the diffusion coefficient to be $b\equiv 1$ and the drift coefficient to be

\begin{align*}a(x,\alpha)=\frac{1}{2}\nabla\log\pi(x-\alpha),\end{align*}

with $\pi(x)\propto 1/(1+x^{\top}\Sigma^{-1}x/20)^{35}$ , where $\pi(x)$ is the probability density function with respect to the Lebesgue measure. See [Reference Kotz and Nadarajah32]. Here $\Sigma$ is generated from a Wishart distribution with 50 degrees of freedom and the identity matrix as the scale matrix. The terminal time is $T=10$ and the number of observations is $N=10^3$ . The prior distribution is a normal distribution $\mathcal{N}_{50}(0,10\;I_{50})$ .

The Markov kernels used in this simulation are listed in Table 1. The first four kernels in the table are gradient-free kernels. The last five kernels are gradient-based, informed kernels. All kernels other than the first, fifth, and eighth algorithms in Table 1 use the prior distribution as the reference distribution. ‘Reference measure’ here means that either the proposal kernel itself is reversible with respect to the measure, or the proposal kernel approximates another Markov kernel that is reversible with respect to the measure.

Table 1. Markov kernels in Section 5.1. The first four algorithms are gradient-free algorithms. The last five algorithms are gradient-based, informed algorithms.

We apply the Markov chain Monte Carlo algorithms via a two-step procedure. In the first step, we run the random-walk Metropolis algorithm as a burn-in stage. For Gaussian reference kernels, $x_0$ is estimated by the empirical mean in the burn-in stage. After the burn-in, we run each algorithm. The results are presented in Table 1 and Figure 2. In this example, the covariance matrix is not preconditioned; we use the prior’s covariance matrix instead.

Figure 1. Effective sample sizes of log-likelihood per second of the stochastic diffusion process in Section 5.1.1 for the nine Markov kernels listed in Table 1. The y-axis is on a logarithmic scale.

Figure 2. Trace plots of log-likelihood of the stochastic diffusion process in Section 5.1.1 for the nine Markov kernels listed in Table 1.

The acceptance rates for the first two algorithms in Table 1 were set at 25%. For the third and fourth algorithms, acceptance rates were set to 30% to 50%. As suggested by [Reference Roberts and Rosenthal51, Reference Titsias and Papaspiliopoulos60], for the fifth, sixth, and seventh algorithms, the acceptance probabilities were set to approximately 60%. The eighth algorithm was tuned in two steps. First, we set the number of leapfrog steps to 1 and tuned the leapfrog step size so that the acceptance rate would be between 60% and 80% according to [Reference Beskos, Pillai, Roberts, Sanz-Serna and Stuart7]. Then we increased the number of leapfrog steps until the time-normalised effective sample size decreased. The tuning parameters of the Hamiltonian Monte Carlo algorithm were controlled using the RStan package. As a quantitative measure of efficiency, we used the effective sample size of log-likelihood per second. It was estimated using the R package coda [Reference Plummer, Best, Cowles and Vines44].

The effective log-likelihood sample sizes per second are shown in Figure 1. The box plot is constructed from 50 independent simulations for each algorithm. The fifth, sixth, and seventh algorithms, which are Langevin-diffusion-based algorithms, show the worst performance. Since we have to evaluate the derivatives several times per step of the Markov chain, the Hamiltonian Monte Carlo and the infinite-dimensional Hamiltonian Monte Carlo are still worse than the random-walk Metropolis kernel. The random-walk Metropolis kernel and the preconditioned Crank–Nicolson kernel are better than gradient-based kernels, but the mixed preconditioned Crank–Nicolson kernel is much better. The $\Delta$ -guided version is even better than the non- $\Delta$ -guided version thanks to the non-reversible property. A trace plot is also shown in Figure 4; it illustrates that the Hamiltonian Monte Carlo method has good performance per iteration, but the cost is high compared to other algorithms.

5.1.2. Logistic regression.

Next we apply the algorithms to a logistic regression model with the Sonar data set from the University of California, Irvine, repository [Reference Dua and Graff16]. The data set contains 208 observations and 60 explanatory variables. The prior distribution is $\mathcal{N}(0,10^2)$ for each set of parameters. We used a relatively large variance of the normal distribution, because we did not have enough prior information at this stage.

Estimation of the preconditioning matrix is necessary for this problem because of the existence of a strong correlation between the variables. We performed $2.0\times 10^5$ iterations to estimate $\mu_0$ and estimated the preconditioning matrix $\Sigma_0$ using empirical means. Then we ran $10^5$ iterations for each algorithm, discarding the first $2\times 10^4$ iterations as burn-in. Furthermore, we ran each experiment 50 times using different seeds. We evaluated the effective sample size of log-likelihood per second; the results of all the algorithms are presented in box plots (Figure 3). The algorithms based on the Lebesgue measure (the first, fifth, and eighth algorithms in Table 1) are worse than other algorithms based on the Gaussian reference measure. The performances of the gradient-based algorithms are divergent, which might reflect the sensitivity of the gradient-based algorithms, which is well described in [Reference Chopin and Ridgway12]. In particular, the infinite-dimensional Hamiltonian Monte Carlo algorithm shows better performance in this case, although it shows poor performance in the previous simulation. The $\Delta$ -guided mixed preconditioned Crank–Nicolson kernel was slightly worse than infinite-dimensional Hamiltonian Monte Carlo algorithm and better than all the other algorithms. The Metropolis–Haar and $\Delta$ -guided Metropolis–Haar kernels show good and robust results for the two simulation experiments.

Figure 3. Effective sample sizes of log-likelihood per second in logistic regression example in Section 5.1.2 for the nine Markov kernels listed in Table 1.

Figure 4. Sampling paths of the logistic regression example illustrated in Section 5.1.2. The Hamiltonian Monte Carlo algorithm is excluded from this simulation because the initial values of the algorithm are automatically selected in the RStan package.

We also investigate the sensitivity of the gradient-based algorithms for the same model as displayed in Figure 4. In this example, 10 initial values are randomly generated from a multivariate normal distribution for each algorithm. The number of iterations of each algorithm is $5 \times 10^3$ . The paths of the gradient-based algorithms depend strongly on the initial values, except in the case of the infinite-dimensional Hamiltonian Monte Carlo algorithm.

5.1.3. Sensitivity of the choice of $x_0$ .

To illustrate the importance of $x_0$ , we additionally run a numerical experiment on a 50-dimensional multivariate central t-distribution with degrees of freedom $\nu=3$ and identity covariance matrix [Reference Kotz and Nadarajah32]. The first element of $x_0$ is $\xi\ge 0$ , and all the other elements are set to be zero. When $\xi$ is large, the direction is less important for increasing or decreasing the likelihood. We run the algorithms on the target distribution for $10^5$ iterations. The experiment shows that the benefit of non-reversibility diminishes as the importance of the direction shrinks (Table 2).

Table 2. Effective sample sizes of log-likelihood per second target on a 50-dimensional Student distribution as in Section 5.1.3.

Table 3. Description of Markov kernels in Figure 5 in Section 5.2.

Figure 5. Trace plots of the Metropolis kernels in Section 5.2. The guided kernels (in the rightmost panels) are more variable than their non-guided counterparts. The solid lines correspond to the negative direction and the dashed lines to the positive direction.

5.2. $\Delta$ -guided Metropolis–Haar kernels on $\mathbb{R}_+^d$

Next, we consider the beta–gamma-based kernels considered in Example 15 and the chi-squared-based kernels considered in Example 16 with $L=1$ . Thus, we consider a total of six Markov kernels. These are the Metropolis kernel, the Metropolis–Haar kernel, and the $\Delta$ -guided Metropolis–Haar kernel for each of the beta–gamma-based and chi-squared-based kernels.

Our goal is not to compare the beta–gamma-based kernels with the chi-squared-based kernels, but to compare the guided kernels with the non-guided kernels. In this simulation, we illustrate the difference in behaviour between the guided Metropolis kernel and other kernels by plotting trajectories in two dimensions.

We consider a Poisson hierarchical model of the form

\begin{align*}x_{m,n}|\theta_m \sim \mathrm{Poisson}(\theta_m),\quad n=1,\ldots, N, \end{align*}

\begin{align*}\theta_m \sim \mathcal{G}(\alpha,\beta),\quad m=1,\ldots, M, \end{align*}

\begin{align*}\alpha \sim \mathcal{G}(1/20,1/20), \quad \beta \sim \mathcal{G}(1/20,1/20),\end{align*}

where $x=\{x_{m,n}\;:\;m=1,\dots,M,\ n=1,\dots N\}$ is the observation. In our simulations we set $M=25$ and $N=5$ . The number of unknown parameters is $M+2=27$ in this case. The parameter $\theta=(\theta_1,\ldots,\theta_M)$ has a closed-form conditional distribution

\begin{align*}\theta_m|\alpha,\beta,x \sim \mathcal{G}\left(\sum_{n=1}^N x_{m,n}+\alpha,N+\beta\right) \quad m=1,\ldots,M.\end{align*}

Therefore we can use the Gibbs sampler for generating the parameter $\theta$ . On the other hand, since the conditional distribution of $\alpha, \beta$ is complicated, we apply the Monte Carlo algorithms mentioned above.

We created two-dimensional trajectory plots to illustrate the difference in behaviour between the Metropolis–Haar kernel and the $\Delta$ -guided version of it. The tuning parameters are chosen so that the average acceptance probabilities are 30–40% in $5 \times 10^4$ iterations. Figure 5 shows the trace plots of the last 300 iterations for the kernels. One can clearly see the larger variation for the guided kernels. Thanks to the incident variables, the guided kernel maintains its direction when the proposed value is accepted. The property of maintaining direction greatly contributes to the increase in variability.