# Weight-preserving simulated tempering

- 115 Downloads

## Abstract

Simulated tempering is a popular method of allowing MCMC algorithms to move between modes of a multimodal target density \(\pi \). One problem with simulated tempering for multimodal targets is that the weights of the various modes change for different inverse-temperature values, sometimes dramatically so. In this paper, we provide a fix to overcome this problem, by adjusting the mode weights to be preserved (i.e. constant) over different inverse-temperature settings. We then apply simulated tempering algorithms to multimodal targets using our mode weight correction. We present simulations in which our weight-preserving algorithm mixes between modes much more successfully than traditional tempering algorithms. We also prove a diffusion limit for an version of our algorithm, which shows that under appropriate assumptions, our algorithm mixes in time \(O(d [\log d]^2)\).

## Keywords

Simulated tempering Parallel tempering MCMC Multimodality and Monte Carlo## 1 Introduction

Consider the problem of drawing samples from a target distribution, \(\pi (x)\) on a *d*-dimensional state space \(\mathcal {X}\) where \(\pi (\cdot )\) is only known up to a scaling constant. A popular approach is to use Markov chain Monte Carlo (MCMC) which uses a Markov chain that is designed in such a way that the invariant distribution of the chain is \(\pi (\cdot )\).

However, if \(\pi (\cdot )\) exhibits multimodality, then the majority of MCMC algorithms which use tuned localised proposal mechanisms, e.g. Roberts et al. (1997) and Roberts and Rosenthal (2001), fail to explore the state space, which leads to biased samples. Two approaches to overcome this multimodality issue are the *simulated* and *parallel tempering algorithms*. These methods augment the state space with auxiliary target distributions that enable the chain to rapidly traverse the entire state space.

The major problem with these auxiliary targets is that in general they do not preserve regional mass; see Woodard et al. (2009a, b) and Bhatnagar and Randall (2016). This problem can result in the required run time of the simulated and parallel tempering algorithms growing exponentially with the dimensionality of the problem.

In this paper, we provide a fix to overcome this problem, by adjusting the mode weights to be preserved (i.e. constant) over different inverse temperatures. We apply our mode weight correction to produce new simulated and parallel tempering algorithms for multimodal target distributions. We show that assuming the chain mixes at the hottest temperature, our mode-preserving algorithm will mix well for the original target as well.

This paper is organised as follows. Section 2 reviews the simulated and parallel tempering algorithms and the existing literature for their optimal set-up. Section 3 describes the problems with modal weight preservation that are inherent with the traditional approaches to tempering, and introduces a prototype solution called the HAT algorithm that is similar to the parallel tempering algorithm but uses novel auxiliary targets. Section 4 presents some simulation studies of the new algorithms. Section 5 provides a theoretical analysis of a diffusion limit and the resulting computational complexity of the HAT algorithm in high dimensions. Section 6 concludes and provides a discussion of further work.

## 2 Tempering algorithms

### 2.1 Optimal scaling for the ST and PT algorithms

Atchadé et al. (2011) and Roberts and Rosenthal (2014) investigated the problem of selecting optimal inverse-temperature spacings for the ST and PT algorithms. Specifically, if a move between two consecutive temperature levels, \(\beta \) and \(\beta '=\beta +\epsilon \), is to be proposed, then what is the optimal choice of \(\epsilon \)? Too large, and the move will probably be rejected; too small, and the move will accomplish little (similar to the situation for the Metropolis algorithm, cf. Roberts et al. 1997 and Roberts and Rosenthal 2001).

*d*-dimensional target distributions of the i.i.d. form:

*within*each temperature, to allow them to concentrate solely on the mixing of the inverse-temperature process itself. To achieve non-degeneracy of the limiting behaviour of the inverse-temperature process as \(d \rightarrow \infty \), the spacings are scaled as \(O(d^{-1/2})\), i.e. \(\epsilon = \ell /d^{1/2}\) where \(\ell = \ell (\beta )\) a positive value to be chosen optimally.

*d*, to a specific diffusion limit, independent of dimension, which thus mixes in time

*O*(1), implying that the original ST and PT algorithms mix in time

*O*(

*d*) as \(d\rightarrow \infty \). They also prove that the mixing times of the ST and PT algorithms are optimised when the value of \(\ell \) is chosen to maximise the quantity

From a practical perspective, setting up the temperature levels to achieve optimality can be done via a stochastic approximation approach (Robbins and Monro 1951), similarly to Miasojedow et al. (2013) who use an adaptive MCMC framework (see, e.g. Roberts and Rosenthal 2009).

### 2.2 Torpid mixing of ST and PT algorithms

The above optimal scaling results suggest that the mixing time of the ST and PT algorithms through the temperature schedule is *O*(*d*), i.e. grows only linearly with the dimension of the problem, which is very promising. However, this optimal, non-degenerate scaling was derived under the assumption of immediate, infinitely fast within-temperature mixing, which is almost certainly violated in any real application. Indeed, this assumption appears to be overly strong once one considers the contrasting results regarding the scalability of the ST algorithm from Woodard et al. (2009a, b). Their approach instead relies on a detailed analysis of the spectral gap of the ST Markov chain and how it behaves asymptotically in dimension. They show that in cases where the different modal structures/scalings are distinct, this can lead to mixing times that grow exponentially in dimension, and one can only hope to attain polynomial mixing times in special cases where the modes are all symmetric.

The fundamental issue with the ST/PT approaches is that in cases where the modes are not symmetric, the tempered targets do not preserve the regional/modal weights. That motivates the current work, which is designed to preserve the modal weights even when performing tempering transformations, as we discuss next.

Interestingly, a lack of modal symmetry in the multimodal target will affect essentially all the standard multimodal-focused methods: the Equi-Energy Sampler of Kou et al. (2006), the Tempered Transitions of Neal (1996) and the Mode Jumping Proposals of Tjelmeland and Hegstad (2001), all suffer in this setting. Hence, the work in this paper is applicable beyond the immediate setting of the ST/PT approaches.

## 3 Weight-stabilised tempering

In this section, we present our modifications which preserve the weights of the different modes when performing tempering transformations. We first motivate our algorithm by considering mixtures of Gaussian distributions.

*d*-dimensional bimodal Gaussian target distribution with means, covariance matrices and weights given by \(\mu _i,~\varSigma _i,~ w_i \) for \(i=1,2\) respectively. Hence, the target density is given by:

### 3.1 Weight-stabilised Gaussian mixture targets

### Definition 1

*(Weight-Stabilised Gaussian Mixture (WSGM))*

Using these WSGM targets in the PT scheme can give substantially better performance than when using the standard power-based targets. This is very clearly illustrated in Sect. 4.1. Henceforth, when the term “WSGM ST/PT Algorithm” is used, it refers to the implementation of the standard ST/PT algorithm but now uses the WSGM targets from (10).

### 3.2 Approximating the WSGM targets

In practice, the actual target distribution would be non-Gaussian and only approximated by a Gaussian mixture target. On the other hand, due to the improved performance gained from using the WSGM over just targeting the respective power-tempered mixture, there is motivation to approximate the WSGM in the practical setting where parameters are unknown. To this end, we present a theorem establishing useful equivalent forms of the WSGM; these alternative equivalent forms give rise to a practically applicable approximation to the WSGM.

### Theorem 1

- (a)[Standard, non-weight-preserving tempering] If \(f_j(x,\beta ) = [h_j(x)]^\beta \) then$$\begin{aligned} W_{(j,\beta )} \propto w_j^\beta |\varSigma _j|^{\frac{1-\beta }{2}}. \end{aligned}$$
- (b)[Weight-preserving tempering, version #1] Denoting \(\nabla _j = \nabla \log {h_j(x)}\) and \(\nabla _j^2=\nabla ^2\log {h_j(x)}\); if \(f_j(x,\beta )\) takes the formthen \(W_{(j,\beta )} \propto w_j\).$$\begin{aligned} h_j(x)\exp \left\{ \left( \frac{1-\beta }{2}\right) (\nabla _j(x))^T \left[ \nabla ^2_j(x)\right] ^{-1} \nabla _j (x) \right\} \end{aligned}$$
- (c)[Weight-preserving tempering, version #2] Ifthen \(W_{(j,\beta )} \propto w_j\).$$\begin{aligned} f_j(x,\beta )= h_j(x)^\beta h_j(\mu _j)^{(1-\beta )} \end{aligned}$$

* Remark 1* In Theorem 1, statement (b) shows that second-order gradient information of the \(h_j\)’s can be used to preserve the component weight in this setting.

* Remark 2* Statement (c) extends statement (b) to no longer require the gradient information about the \(h_j\) but simply the mode/mean point \(\mu _j\). Essentially, this shows that by appropriately rescaling according to the height of the component as the components are “powered up,” then component weights are preserved in this setting.

* Remark 3* A simple calculation shows that statement (c) holds for a more general mixture setting when all components of the mixture share a common distribution but different location and scale parameters.

### 3.3 Hessian adjusted tempering

The results of Theorem 1 are derived under the impractical setting that the components are all known and that \(\pi (\cdot )\) is indeed a mixture target. One would like to exploit the results of (b) and (c) from Theorem 1 to aid mixing in a practical setting where the target form is unknown but may be well approximated by a mixture.

### Definition 2

*(Basic Hessian Adjusted Tempering (BHAT) Target)*For a target distribution \(\pi (\cdot )\) on \({\mathbb {R}}^d\) with a corresponding “mode point assigning function” \(\mu _{x,\beta }: \mathbb {R}^d \rightarrow \mathbb {R}^d\); the BHAT target at inverse temperature level \(\beta \in (0,\infty )\) is defined as

However, in this basic form there is an issue with this target distribution at hot temperatures when \(\beta \rightarrow 0\). The problem is that it leaves discontinuities that can grow exponentially large, and this can make the hot state temperature level mixing exponentially slow if using standard MCMC methods for the within-temperature moves.

*K*mode points \(M=\{ \mu _1,\ldots ,\mu _K \}\). This assumption seems quite strong but in general if one cannot find mode points, then this is essentially saying that one cannot find the basins of attraction and thus the desire to obtain the modal relative masses (as MCMC is trying to do) must be relatively impossible. Indeed, being able to find mode points either prior to or online in the run of the algorithm is possible, e.g. Tjelmeland and Hegstad (2001), Behrens (2008) and Tawn et al. (2018). Furthermore, assume that the target, \(\pi (\cdot )\), is \(C^2\) in a neighbourhood of the

*K*mode locations and so there is an associated collection of positive definite covariance matrices \(S=\{ \varSigma _1,\ldots ,\varSigma _K \}\) where \(\varSigma _j= -\left( \nabla ^2 \log \pi (\mu _j) \right) ^{-1}\). From this and knowing the evaluations of \(\pi (\cdot )\) at the mode points, one can approximate the weights in the regions to attain a collection \({\hat{\mathbf{W}}}=\{\hat{w}_1,\ldots ,\hat{w}_K\}\) where

### Definition 3

*(WSGM mode assignment function)*With collections

*M*,

*S*and \({\hat{\mathbf{W}}}\) specified above then for a location \(x\in \mathbb {R}^d\) and inverse temperature \(\beta \) define the WSGM mode assignment function as

Under the assumption that there are collections *M*, *S* and \({\hat{\mathbf{W}}}\) that have either been found through prior optimisation or through an adaptive online approach, we define the following.

### Definition 4

*(Hessian Adjusted Tempering (HAT) Target)*For a target distribution \(\pi (\cdot )\) on \(\mathbb {R}^d\) with collections

*M*,

*S*and \(\hat{W}\) defined above along with the associated mode assignment function given in (14), the Hessian adjusted tempering (HAT) target is defined as

*G*” specifies the target distribution when the chain’s location,

*x*, is in a part of the state space where the narrower modes expand their basins of attraction as the temperature gets hotter. Both the choice of

*G*and the mode assignment function used in Definition 4 are somewhat canonical to the Gaussian mixture setting. With the same assignment function specified in Definition 3, an alternative and seemingly robust “

*G*” that one could use is given by

*G*, then under the assumption that the target is continuous and bounded on \(\mathbb {R}^d\) and that for all \(\beta \in (0,\infty )\),

We propose to use the HAT targets in place of the power-based targets for the tempering algorithms given in Sect. 2. We thus define the following algorithms, which are explored in the following sections.

## 4 Simulation studies

### 4.1 WSGM Algorithm simulation study

We begin by comparing the performances of a ST algorithm targeting both the power-based and WSGM targets for a simple but challenging bimodal Gaussian mixture target example. The example will illustrate that the traditional ST algorithm, using power-based targets, struggles to mix effectively through the temperature levels due to a bottleneck effect caused by the lack of regional weight preservation.

This temperature schedule gave swap acceptance rates of approximately 0.23 between all levels of the power-based ST algorithm except for the coldest level swap where this degenerated to 0.17. That shows that the power-based ST algorithm was set up essentially optimally according to the results in Atchadé et al. (2011).

*x*is in mode 1 if \({\bar{x}}<0\) and in mode 2 otherwise. Then, the within-move proposal distribution for a move at inverse temperature level \(\beta \) is given by

where \(\phi _{\mu ,\varSigma }(.)\) is the density function of a Gaussian random variable with mean \(\mu \) and variance matrix \(\varSigma \).

*t*of the Markov chain, such that when \(\beta _t=1\), the magnitude of

*h*is 1 and when the temperature is at its hottest level, i.e. \(\beta _t = \beta _{\text {min}}\),

*h*is zero. Furthermore, in this example, the sign of \({\bar{x}}_t\) is a reasonable proxy to identify the mode that the chain is contained in with a negative value suggesting the chain is in the mode centred on \(\mu _1\) and \(\mu _2\) otherwise.

Figure 4 clearly illustrates that the hot state modal weight inconsistency leads the chain down a poor trajectory since at hot temperatures nearly all the mass is in modal region 1. This results in the chain never reaching the other mode in the entire (finite) run of the algorithm. Indeed, the trace plots in Fig. 4 show that the chain is effectively trapped in mode 1, which although it only has 20% of the mass in the cold state, is completely dominant at the hotter states.

### 4.2 Simulation study for HAT

As will be seen in the forthcoming simulation results, the imbalance of scales within each modal region ensures that this is a very challenging problem for the PT algorithm.

Since this target fits into the setting of Corollary 1 of Tawn and Roberts (2018), a geometric inverse-temperature schedule is approximately optimal for the HAT target in this setting. Indeed, Tawn and Roberts (2018) suggest that the geometric ratio should be tuned so that the acceptance rate for swap moves between consecutive temperatures is approximately 0.234. In this case, eight tempering levels were used to obtain effective mixing; these were geometrically spaced and given by \(\{1,0.31,0.31^2,\ldots , 0.31^7 \}\), were found to be approximately optimal and gave an average of 0.22 for the swaps between consecutive levels for the HAT algorithm.

Using this temperature schedule along with appropriately tuned RWM proposals for the within-temperature moves, ten runs of both the PT and HAT algorithms were performed. In each individual run, each temperature marginal was updated with \(m=5\) RWM proposals followed by a temperature swap move proposal and this was repeated with \(s=100{,}000\) sweeps. This results in a sample output of 600,001 of the cold state chain prior to any burn-in removal. Herein, for this example, denote \(N=600{,}001\).

As expected, the scale imbalance between the modes resulted in the PT algorithm performing poorly and with significant bias in the sample output. In contrast, the HAT approach was highly successful in converging relatively rapidly to the target distribution, exhibiting far more frequent intermodal jumps at the cold state.

*k*th iteration of the cold state chains, after removing a burn-in period of 10,000 initial iterations, for the ten runs of the PT and HAT runs, respectively. The approximation after iteration \(k \le N\) is given by

*i*th iteration. This figure indicates that the PT algorithm fails to provide a stable estimate for \(\mathbb {P}_{\pi }(-30<X^1_i<0)\) with the running weight approximations far from stable at the end of the runs; in stark contrast, the HAT algorithm exhibits very stable performance in this case. In fact, the final estimates for \(\hat{W}^N_1\) are given in Table 1.

End point estimates, \(\hat{W}^N_1\), of \(\mathbb {P}_{\pi }(-30<X^1_i<0)\) from the ten runs of the PT and HAT algorithms. The true value of 0.25 appears to be well approximated by HAT but not by PT

PT | 0.23 | 0.36 | 0.19 | 0.31 | 0.10 | 0.12 | 0.18 |
---|---|---|---|---|---|---|---|

0.39 | 0.51 | 0 | |||||

HAT | 0.27 | 0.24 | 0.26 | 0.22 | 0.22 | 0.27 | 0.23 |

0.28 | 0.25 | 0.26 |

Table 2 presents the results of using the ten runs of each algorithm in a batch-means approach to estimate the Monte Carlo variance of the estimator of \(\hat{W}^N_1\). The results in Table 2 show that the Monte Carlo error is approximately a factor of 10 higher for the PT algorithm than the HAT approach.

Using the ten runs of each algorithm in a batch-means approach to estimate the Monte Carlo variance of the pooled estimator \(\hat{W}^{10N}_1\) i.e. \(\text {SD}(\hat{W}^N_1)\). Also displayed is the average run time (RT, measured in seconds) of a single one of the ten repeated runs for both methods, respectively

\(\hat{W}^{10 N}_1\) | \(\hat{\text {SD}}(\hat{W}^N_1)\) | \(\hat{\text {SD}}\)(\(\hat{W}^{10 N}_1\)) | RT (secs) | |
---|---|---|---|---|

PT | 0.288 | 0.187 | 0.0593 | 217 |

HAT | 0.249 | 0.019 | 0.0063 | 451 |

## 5 Diffusion limit and computational complexity

In this section, we provide some theoretical analysis for our algorithm. We shall prove in Theorems 2 and 3 that as the dimension goes to infinity, a simplified and speeded-up version of our weight-preserving simulated tempering algorithm (i.e. the HAST Algorithm from Definition 5, equivalent to the ST Algorithm 1 with the adjusted target from Definition 4) converges to a certain specific diffusion limit. This limit will allow us to draw some conclusions about the computational complexity of our algorithm.

### 5.1 Assumptions

*p*and \(1-p\), respectively. Let

*I*denote the indicator of which mode the process is in, taking value 1 or 2.

We shall sometimes concentrate on the *Exponential Power Family* special case in which each of the two mixture component factors is of the form \(f_j(x) \propto e^{-\lambda _j|x|^{r_j}}\) for some \(\lambda _j,r_j>0\). This includes the Gaussian case for which \(r_1=r_2=2\) and \(\lambda _j = 1/\sigma _j^2\). (Note that the HAT target in (15) requires the existence of second derivatives about the mode points, corresponding to \(r_j \ge 2\).)

*t*for the

*d*-dimensional process. To study weak convergence, we let \( \beta ^{(\mathrm{d})}_{N(\mathrm{d}t)}\) be a continuous-time version of the \(\beta _t^{(d)}\) process, speeded up by a factor of

*d*, where \(\{N(t)\}\) is an independent standard rate 1 Poisson process. To combine the two modes into one single process, we further augment this process by multiplying it by \(-1\) when the algorithm’s state is closer to the second mode, while leaving it positive (unchanged) when state is closer to the first mode. Thus, define

### 5.2 Main results

Our first diffusion limit result (proved in Appendix), following Roberts and Rosenthal (2014), states that when we are at an inverse temperature greater than \(\beta _{\text {min}}\), the inverse temperature process behaves identically to the case where there is only one mode (i.e. \(J=1\)).

### Theorem 2

Assume the target \(\pi \) is of form (11), with \(J=2\) modes of weights \(w_1=p\) and \(w_2=1-p\), with inverse weights chosen as in (21). Then, up until the first time the process \(X^{(d)}\) hits \(\pm \beta _{\text {min}}\), as \(d\rightarrow \infty \), \(\{X_t^{(d)}\}\) converges weakly to a fixed diffusion process *X* given by (22).

*h*at \(\pm \beta _{\text {min}}\). To resolve these issues, we define

*H*continuous at 0.

### Remark 1

The process *H* leaves constant densities locally invariant, \({{\tilde{G}}}^*g(v)=0\) for all \(v\ne 0\) where \({{\tilde{G}}}^*\) is the adjoint of the infinitesimal generator of *H*, as will be shown in Appendix. This suggests that the density of the invariant distribution of *H* (if it exists) should be piecewise uniform; i.e. it should be constant for \(v>0\) and also constant for \(v<0\) though these two constants might not be equal.

To make further progress, we require a *proportionality condition*. Namely, we assume that the quantities corresponding to \(I(\beta )=\text {Var}_{\pi ^\beta }\big ( \log f(x) \big )\) are proportional to each other in the two modes. More precisely, we extend the definition of *I* to \(I(\beta ) = \text {Var}_{x\sim f_1^\beta }(\log f_1(x))\) for \(\beta >0\) (corresponding to the first mode), and \(I(\beta ) = \text {Var}_{x\sim f_2^{|\beta |}}(\log f_2(x))\) for \(\beta <0\) (corresponding to the second mode), and assume there is a fixed function \(I_0:\mathbb {R}_{+}\rightarrow \mathbb {R}_{+}\) and positive constants \(r_1\) and \(r_2\) such that we have \(I(\beta ) = I_0(\beta )/r_1\) for \(\beta >0\) (in the first mode), while \(I(\beta ) = I_0(|\beta |)/r_2\) for \(\beta <0\) (in the second mode). For example, it follows from Section 2.4 of Atchadé et al. (2011) that in the Exponential Power Family case, \(I(\beta ) = 1/r_1\beta ^2\) for \(\beta >0\) and \(I(\beta ) = 1/r_2\beta ^2\) for \(\beta <0\), so that this proportionality condition holds in that case.

*skew Brownian motion*, a generalisation of usual Brownian motion. Informally, this is a process that behaves just like a Brownian motion, except that the sign of each excursion from 0 is chosen using an independent Bernoulli random variable; for further details and constructions and discussion, see, e.g. Lejay (2006). We also require the function

### Theorem 3

*X*. Furthermore, the limit process has the property that if

*Z*is skew Brownian motion \(B^*_t\) with reflection at

### Remark 2

*a*and

*b*, respectively. This might seem surprising since the limiting weights of the modes should be equal to

*p*and \(1-p\), not proportional to

*a*and

*b*(unless \(r_1=r_2\)). The explanation is that the

*lengths*of the positive and negative parts of the domain are given by \(\left[ 2 \, \varPhi \left( {- \ell _0 \over 2 \sqrt{r_1}} \right) \right] ^{1/2}\) and \(\left[ 2 \, \varPhi \left( {- \ell _0 \over 2 \sqrt{r_2}} \right) \right] ^{1/2}\), respectively. Hence, the total stationary mass of the positive and negative parts—and hence also the limiting modes weights—is still

*p*and \(1-p\) as they should be.

### 5.3 Complexity order

Theorems 2 and 3 have implications for the computational complexity of our algorithm.

In Theorem 2, the limiting diffusion process \(H_t\) is a fixed process, not depending on dimension except through the value of \(\beta _{\text {min}}\). It follows that if \(\beta _{\text {min}}\) is kept fixed, then \(H_t\) reaches 0 (and hence mixes modes) in time *O*(1). Since \(H_t\) is derived (via \(X_t\)) from the \(\beta _t\) process speeded up by a factor of *d*, it thus follows that for fixed \(\beta _{\text {min}}\), \(\beta _t\) reaches \(\beta _{\text {min}}\) (and hence mixes modes) in time *O*(*d*). So, if \(\beta _{\text {min}}\) is kept fixed, then the mixing time of the weight-preserving tempering algorithm is *O*(*d*), which is very fast. However, this does not take into account the dependence on \(\beta _{\text {min}}\), which might also change as a function of *d*.

*d*. If the proposal scaling is optimal for within each mode at the cold temperature, then the proposal scaling is \(O(d^{-1/2})\). Then, at an inverse temperature \(\beta \), the proposal scaling is \(O((\beta d)^{-1/2})\). Hence, at an inverse temperature \(\beta \), the probability of jumping from one mode to the other (a distance \(O(\sqrt{d})\) away) is roughly of order \(e^{-\beta d^2}\). This is exponentially small unless \(\beta = O(1/d^2)\). This indicates that for our algorithm to perform well, we need to choose \(\beta _{\text {min}} = O(1/d^2)\). With this choice, the mixing time order becomes

### 5.4 More than two modes

Finally, we note that for simplicity, the above analysis was all done for just *two* modes. However, a similar analysis works more generally. Indeed, suppose now that we have *k* modes, of general weights \(p_1,p_2,\ldots ,p_k \ge 0\) with \(\sum _i p_i = 1\). Then, when \(\beta \) gets to \(\beta _{\text {min}}\), the process chooses one of the *k* modes with probability \(p_i\). This corresponds to \(\{Y_t\}\) being replaced by a Brownian motion not on \([-1,1]\), but rather on a “star” shape with *k* different length-1 line segments all meeting at the origin (corresponding, in the original scaling, to \(\beta _{\text {min}}\)), where each time the Brownian motion hits the origin it chooses one of the *k* line segments with probability \(p_i\) each. This process is called *Walsh’s Brownian motion*, see e.g. Barlow et al. (1989). (The case \(k=2\) but \(p_1\not =1/2\) corresponds to skew Brownian motion as above.) For this generalised process, a theorem similar to Theorem 2 can be then stated and proved by similar methods, leading to the same complexity bound of \(O\big (d \, [\log d]^2\big )\) iterations in the multimodal case as well.

## 6 Conclusion and further work

This article has introduced the HAT algorithm to mitigate the lack of regional weight preservation in standard power-based tempered targets. Our simulation studies show promising mixing results, and our theorems indicate the mixing times can become polynomial rather than exponential functions of the dimension *d*, and indeed of time \(O(d[\log d]^2)\) under appropriate assumptions.

Various questions remain to make our HAT approach more practically applicable. The “modal assignment function” needs to be specified in an appropriate way, and more exploration into the robustness of the current assignment mechanism is needed to understand its performance on heavier and lighter tailed distributions. The suggested HAT target assumes knowledge of the mode points which typically one will not have to begin with and one would rely on effective optimisation methods to seek these out either during or prior to the run of the algorithm. Indeed, this has been partially explored by the authors in Tawn et al. (2018). The performance of HAT is heavily reliant on the mixing at the hottest temperature level; the use of RWM here can be problematic for HAT where the mode heights of the disperse modes can be far lower than the narrower modes. As such, more advanced sampling schemes such as discretised tempered Langevin could give accelerated mixing at the hot state, the effects of which would be transferred to an improvement in the mixing at the coldest state.

In the theoretical analysis of Sect. 5, the spacing between consecutive inverse-temperature levels was taken to be \(O(d^{-1/2})\) to induce a non-trivial diffusion limit. However, this result required strong assumptions. Accompanying work in Tawn and Roberts (2018) suggests that for the HAT algorithm under more general conditions, the consecutive optimal spacing should still be \(O(d^{-1/2})\), with an associated optimal acceptance rate in the interval [0, 0.234].

## Notes

### Acknowledgements

Funding was provided by Engineering and Physical Sciences Research Council (Grant No. EP/K014463/1) and by NSERC of Canada.

## References

- Atchadé, Y.F., Liu, J.S.: The Wang–Landau algorithm for Monte Carlo computation in general state spaces. Stat. Sin.
**20**, 209–33 (2004)Google Scholar - Atchadé, Y.F., Roberts, G.O., Rosenthal, J.S.: Towards optimal scaling of Metropolis-coupled Markov chain Monte Carlo. Stat. Comput.
**21**(4), 555–568 (2011)MathSciNetCrossRefzbMATHGoogle Scholar - Barlow, M.T., Pitman, J., Yor, M.: On Walsh’s Brownian motions. Séminaire de probabilités (Strasbourg)
**23**, 275–293 (1989)MathSciNetzbMATHGoogle Scholar - Bédard, M., Rosenthal, J.S.: Optimal scaling of Metropolis algorithms: heading toward general target distributions. Canad. J. Stat.
**36**, 483–503 (2008)MathSciNetCrossRefzbMATHGoogle Scholar - Behrens, G.R.: Mode jumping in MCMC. Ph.D. thesis, University of Bath (2008)Google Scholar
- Bhatnagar, N., Randall, D.: Simulated tempering and swapping on mean-field models. J. Stat. Phys.
**164**(3), 495–530 (2016)MathSciNetCrossRefzbMATHGoogle Scholar - Ethier, S.N., Kurtz, T.G.: Markov Processes: Characterization and Convergence. Wiley, London (1986)CrossRefzbMATHGoogle Scholar
- Geyer, C.J.: Markov chain Monte Carlo maximum likelihood. Comput Sci Stat
**23**, 156–163 (1991)Google Scholar - Kone, A., Kofke, D.A.: Selection of temperature intervals for parallel-tempering simulations. J Chem Phys
**122**(20), 206101 (2005)CrossRefGoogle Scholar - Kou, S., Zhou, Q., Wong, W.H.: Equi-energy sampler with applications in statistical inference and statistical mechanics. Ann. Stat.
**34**, 1581–1619 (2006)MathSciNetCrossRefzbMATHGoogle Scholar - Lejay, A.: On the constructions of the skew Brownian motion. Probab. Surv.
**3**, 413–466 (2006)MathSciNetCrossRefzbMATHGoogle Scholar - Liggett, T.M.: Continuous Time Markov Processes: An Introduction. American Mathematical Society, Providence (2010)CrossRefzbMATHGoogle Scholar
- Marinari, E., Parisi, G.: Simulated tempering: a new Monte Carlo scheme. EPL (Europhys. Lette.)
**19**(6), 451 (1992)CrossRefGoogle Scholar - Miasojedow, B., Moulines, E., Vihola, M.: An adaptive parallel tempering algorithm. J. Comput. Graph. Stat.
**22**(3), 649–664 (2013)MathSciNetCrossRefGoogle Scholar - Neal, R.M.: Sampling from multimodal distributions using tempered transitions. Stat. Comput.
**6**(4), 353–366 (1996)CrossRefGoogle Scholar - Nemeth, C., Lindsten, F., Filippone, M., Hensman, J.: Pseudo-extended Markov Chain Monte Carlo (2017). ArXiv e-prints arXiv:1708.05239
- Predescu, C., Predescu, M., Ciobanu, C.V.: The incomplete beta function law for parallel tempering sampling of classical canonical systems. J. Chem. Phys.
**120**(9), 4119–4128 (2004)CrossRefGoogle Scholar - Revuz, D., Yor, M.: Continuous Martingales and Brownian Motion, 3rd edn. Springer, Berlin (2004)zbMATHGoogle Scholar
- Robbins, H., Monro, S.: A stochastic approximation method. Ann. Math. Stat. 400–407 (1951)Google Scholar
- Roberts, G.O., Rosenthal, J.S.: Optimal scaling of discrete approximations to Langevin diffusions. J. R. Stat. Soc. Ser. B (Stat. Methodol.)
**60**(1), 255–268 (1998)MathSciNetCrossRefzbMATHGoogle Scholar - Roberts, G.O., Rosenthal, J.S.: Optimal scaling for various Metropolis–Hastings algorithms. Stat. Sci.
**16**(4), 351–367 (2001)MathSciNetCrossRefzbMATHGoogle Scholar - Roberts, G.O., Rosenthal, J.S.: Examples of adaptive MCMC. J. Comput. Graph. Stat.
**18**(2), 349–367 (2009)MathSciNetCrossRefGoogle Scholar - Roberts, G.O., Rosenthal, J.S.: Minimising MCMC variance via diffusion limits, with an application to simulated tempering. Ann. Appl. Probab.
**24**(1), 131–149 (2014)MathSciNetCrossRefzbMATHGoogle Scholar - Roberts, G.O., Gelman, A., Gilks, W.R.: Weak convergence and optimal scaling of random walk Metropolis algorithms. Ann. Appl. Probab.
**7**(1), 110–120 (1997)MathSciNetCrossRefzbMATHGoogle Scholar - Tawn, N.: Towards optimality of the parallel tempering algorithm. Ph.D. thesis, University of Warwick (2017)Google Scholar
- Tawn, N.G., Roberts, G.O.: Optimal Temperature Spacing for Regionally Weight-preserving Tempering (2018). arxiv:1810.05845v1
- Tawn, N.G., Roberts, G.O., Moores, M., Assing, S.: Annealed leap point sampler. Manuscript in preparation (2018)Google Scholar
- Tjelmeland, H., Hegstad, B.K.: Mode jumping proposals in MCMC. Scand. J. Stat.
**28**(1), 205–223 (2001)MathSciNetCrossRefzbMATHGoogle Scholar - Wang, F., Landau, D.: Determining the density of states for classical statistical models: a random walk algorithm to produce a flat histogram. Phys. Rev. E
**64**(5), 056101 (2001)CrossRefGoogle Scholar - Wang, J.S., Swendsen, R.H.: Cluster Monte Carlo algorithms. Phys. A Stat. Mech. Appl.
**167**(3), 565–579 (1990)MathSciNetCrossRefGoogle Scholar - Woodard, D.B., Schmidler, S.C., Huber, M.: Conditions for rapid mixing of parallel and simulated tempering on multimodal distributions. Ann. Appl. Probab.
**19**, 617–640 (2009a)MathSciNetCrossRefzbMATHGoogle Scholar - Woodard, D.B., Schmidler, S.C., Huber, M.: Sufficient conditions for torpid mixing of parallel and simulated tempering. Electr. J. Probab.
**14**, 780–804 (2009b)MathSciNetCrossRefzbMATHGoogle Scholar

## Copyright information

**Open Access**This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.