## [TM] Stein’s lemma

This post is about Stein’s lemma, which first appeared in the landmark paper of Stein in 1981. This lemma is leads to Stein’s unbiased risk estimator and is useful for proving central limit theorems. It has also been called Gaussian integration by parts’, which is in fact a high level description of the proof.

Lemma 1 (Stein’s lemma) If ${y}$ follows the standard normal distribution, then

$\displaystyle \mathbb{E}\left[x f(x) \right] = \mathbb{E}\left[f'(x)\right],$

if the expectations are well-defined.

Proof: If ${f}$ is nice’, then

$\displaystyle \begin{array}{rcl} \mathbb{E}\left[f'(x)\right] &=& \int_{-\infty}^{\infty} f'(x) \frac{\exp(-x^2/2)}{\sqrt{2\pi}} dx\\ &=& 0 - \int_{-\infty}^{\infty} f(x) \left(-x \frac{\exp(-x^2/2)}{\sqrt{2\pi}} \right) dx \\ &=& \mathbb{E}\left[x f(x)\right]. \end{array}$

$\Box$

It is also convinient to denote ${\phi(x)}$ as the standard Normal density and remember that

$\displaystyle \phi'(x) = -x \phi(x).$

Stein’s lemma can be generalized to exponential family distributions. In particular, for multivariate normals, if ${y \sim \mathrm{NVM}_{p}(\mu, \sigma^2 I)}$ and ${\hat{\mu} = r(y)}$ is any differentiable estimator, then we have

$\displaystyle \textrm{cov}\left(\hat{\mu}_i, y_i\right) = \sigma^2 \mathbb{E}\left[\frac{\partial \hat{\mu}_i}{\partial y_i}\right].$

This is Equation (12.56) in Computer Age Statistical Inference’ by Efron and Hastie.

This post is the second in the category trivial matters’, where I formally write down some notes to myself about identities, tricks, and facts that I repeatedly use but (unfortunately) need to re-derive everytime I use them. Although these posts are short, they discuss important topics. The term trivial matters’ is used as a sarcasm, because my blood boils everytime I see terms like obviously’ or it is trivial to show that …’ when I grade students’ homeworks.

## [TM] Change of Variables in MCMC

This post is about change of variables in Markov chain Monte Carlo (MCMC), which is used quite often when the target distribution is supported on a subset of ${\mathbb{R}^n}$. For example, the Exponential distribution and the Log-Normal distribution are only supported on positive reals.

Consider a target distribution ${\pi(x)}$ that is supported on a subset ${S \subset \mathbb{R}^n}$. If we use a random walk proposal ${q_X(x' \mid x) = \mathrm{MVN}(x' ; x,\Sigma)}$, then we might end up with a proposal ${x'}$ such that ${\pi(x') = 0}$ and, this might cause too few acceptance in the MCMC chain. If we can find a transformation ${h:D \to \mathbb{R}^n}$ that is one-to-one, differentiable and spans ${\mathbb{R}^n}$, then we can consider a proposal ${x' = h^{-1}(y')}$ where ${y' \sim q_Y(\cdot \mid y = h(x))}$. This proposal always yields a proposal ${x'}$ such that ${\pi(x') > 0.}$

Of course, when we employ such a transformation in the proposal kernel, we need to be careful about evaluating the proposal densities. We know that the acceptance probability is ${\alpha(x,x') = 1 \wedge \frac{\pi(x') q(x \mid x')}{\pi(x) q(x' \mid x)}}$, and it should be no surprise that ${q_X(x' \mid x) \not= q_Y(y' \mid y)}$ unless ${h}$ is the identity map.

Let’s work out the acceptance ratio together carefully. Recall that change of variables proceeds as follows: when ${Y \sim f_Y(y) }$ and we consider the transformation ${ y = h^{-1}(x)}$, the pdf of ${X = h^{-1}(Y)}$ is

$\displaystyle f_X(x) = f_Y(h(x))|J_{h}(x)|.$

When we apply this to the kernels ${q_Y}$ and ${q_X}$ we get that

$\displaystyle q_X(x' \mid x ) = q_Y(h(x') \mid h(x)) \cdot |J_{h}(x')|.$

Example 1 {Symmetric proposal on transformed space} If ${q_Y(y' \mid y)}$ is a symmetric proposal, then the acceptance probability becomes

$\displaystyle \alpha(x,x') = 1 \wedge \frac{\pi(x') |J_{h}(x)|}{\pi(x)|J_{h}(x')|} .$

Here are two common transformations.

Example 2 (Log-transformation for ${x}$ supported on ${\mathbb{R}_+}$)

If ${h = \log}$, then ${|J_{h}(x)| = 1 / x}$ and acceptance probability is

$\displaystyle \alpha(x,x') = 1 \wedge \frac{\pi(x')x'}{\pi(x)x}.$

Example 3 (Logit transformation for ${x}$ supported on ${(0,1)}$) If ${h(x) = \log(\frac{x}{1 - x})}$,then the inverse transformation is ${h^{-1}(y) = \frac{exp(y)}{1 + \exp(y)}.}$ The acceptance probability is

$\displaystyle \alpha(x,x') = 1 \wedge \frac{\pi(x')x'(1-x')}{\pi(x)x(1-x)}.$

This post is the first one in the category trivial matters’, where I formally write down some notes to myself about tricks and facts that I repeatedly use but (unfortunately) need to re-derive everytime I use them.

## SRN – “Scalable Metropolis-Hastings for Exact Bayesian Inference with Large Datasets” by Cornish et al.

When faced with large datasets, standard MCMC methods have a cost of O(n) per step due to evaluating the log-likelihood for n data points. Can we make it faster?

My COVID-19 self-isolation plan includes trying to finish as many to-read articles as possible and to blog my progress on this blog as much as possible. This weekend, I read the paper “Scalable Metropolis-Hastings for Exact Bayesian Inference with Large Datasets” by Robert Cornish, Paul Vanetti, Alexandre Bouchard-Cȏté, George Deligiannidis, and Arnaud Doucet. This is an ICML 2019 paper and it is a dense one. It attempts to solve one problem. When faced with large datasets,  standard MCMC methods have a cost of O(n) per step due to evaluating the log-likelihood for n data points. Can we make it faster? In a typical MH step (with symmetric proposals), we accept the proposal with probability $\alpha(\theta,\theta') = \min(1, \frac{\pi(\theta')}{\pi(\theta)})$. This is typically implemented with

log_alpha <- log_post(theta_proposal) - log_post(theta_current);
if (log(runif(1)) < log_alpha ){
theta_current <- theta_proposal;
}

The authors realize that it is wasteful having to evaluate the log-likelihood at all data points before rejecting the proposal. Their solution, Salable Metropolis-Hastings (SMH), is based on a combination of three ideas: (1) factorization of posterior density (trivial factorization is 1 + n factors), (2) fast simulation of rejection using thinning of a Poisson point process, and (3) posterior concentration around mode when n is large.

I first get excited about this paper seeing the fast simulation of Bernoulli random variables (in Algorithm 1) because there is a very similar problem on simulating a Poisson process with inhomogenous rates in Stat 210, for which I was a Teaching Fellow (TF). I am very happy to see this result being useful in cutting-edge research. The fast simulation relies on access to a lower bound for each factor in the posterior distribution and the tighter this lower bound the more efficient this implementation. Suppose the target distribution has m factors, then once we pay a once-off O(m) cost, each acceptance step can be implemented in O(upper bound of negative-log-posterior) number of steps on average.

Another highlight of the paper is the smart application of the Bernstein-von Mises (BvM) phenomenon, that which states that posterior distribution concentrates around the mode with rate $1/\sqrt{n}$ as n goes to infinity. BvM was an important topic in Stat 213, which is another class that I TFed. I have seen BvM used (1) to select step-size in random-walk MH, (2) to initialize an MH chain, and (3) to motivate Laplace approximation of the posterior distribution. In this paper, it motivates a Taylor expansion of the potential function near posterior mode, which subsequently leads to the factorization behind the SMH kernel. The SMH kernel leaves the target distribution invariant, and, if not mistaken by me, this property explains the use of “exact” in the title. More importantly, since the acceptance probability in SMH is factorizable, we can implement the acceptance step with the aforementioned fast simulation, provided that log-density can be differentiated enough times. Intuitively, the higher the differentiability, the closer the Taylor expansion, the tighter the bound for acceptance rate and the fewer the number of evaluations we need. In fact, as later justified with Theorem 3.1, the average computational cost decays geometrically with respect to the order of Taylor expansion. With a second-order Taylor expansion, which requires 3rd-order differentiability, the expected computational cost behaves like $O(1/\sqrt{n})$. Hooray! I am always a big fan of “the bigger the problem, the easier the solution”.

I learn a lot from this paper, especially from the opening paragraph of Section 3 and Section 3.2. But there is a question that I keep asking myself: if n is very large and we have an estimate of posterior mode that is $O(1/\sqrt{n})$ distance from the true mode, how long does it take for SMH to be better than that declaring the posterior is Gaussian by invoking BvM and sleeping on the result?

## SRN – my second attempt to understand ‘Controlled SMC’ by Heng et al.

It has been two years since I last tried to understand the paper ‘Controlled Sequential Monte Carlo’ by Jeremy Heng et al. Over the past two years, I picked this paper up many times, and every time I read it, I understood it a little better. I am finally ready to share my understanding of this paper on this blog.

It has been two years since I last tried to understand the paper ‘Controlled Sequential Monte Carlo’ by Jeremy Heng et al. Over the past two years, I picked this paper up many times, and every time I read it, I understood it a little better. I am finally ready to share my understanding of this paper on this blog.

This paper opens with an introduction to the Feynman-Kac path measure. A triple of (1) initial distribution, (2) a collection of Markov kernels, and (3) potential functions determines a Feynman-Kac path measure. For readers familiar with state-space models, this is not something new: the triple can correspond to (1) the initial distribution of hidden process, (2) Markov kernels of the hidden process and (3) observation densities.

For each Feynman-Kac path measure, there is a sequential Monte Carlo algorithm (aka particle filters) that provides a particle approximation of the path measure. Furthermore, Feynman-Kac path measures can be twisted. Although all the kernels and the potentials are changed, the path measure is the same at the terminal time and the normalizing constant at the terminal time $Z_T$ stays the same.

To infer state-space models, we are interested in the smoothing distribution $x_{0:T}$ given $y_{0:T}$ or the marginal likelihood $p(Y_{0:T})$. The smoothing distribution has Feynman-Kac measure representations, with the marginal likelihood being the normalizing constant.  The most natural one is described above, and it corresponds to the bootstrap particle filter. But it is not unique. For example, the path measure associated with a 1-step look ahead kernel corresponds to the popular ‘fully-adapted auxiliary particle filter’. As mentioned above, each of these path measures can be twisted. Proposition 1 defines an optimal policy with respect to each path measure. Optimality is defined in the sense of approximating the normalizing constant almost surely. The optimal policy with respect to the path measure corresponding to the bootstrap particle filter is the backward information filter $\phi^{\star}(x_t) = p(y_{t:T}|x_t)$

The optimal policies $\phi^{\star}$ are usually intractable, but they can be approximated. After running any SMC algorithm, we have a particle approximation of the path measure, which gives us a particle approximation of the optimal policy. We can use regression to learn a simple policy $\phi$ that minimized the difference between $\log\phi$ and particle approximation of $\log\phi^{\star}$. This regression is done recursively starting at $t = T$ and ending at $t = 0$, and it is called approximate dynamic programming in the paper.  Given this policy, we have a new path measure, which means that we can run another particle filter according to it and find the optimal policy with respect to the new measure. This process is called policy refinement. According to the authors, in practice, only a few iterations of policy refinement is necessary. They also offer performance monitoring criteria based on effective sample size and regression residuals.

What fascinates (and puzzles) me is that errors do not accumulate over either iteration of policy refinement or approximate dynamic programming. I hope to read this paper is greater detail and be able to report back to my readers soon.

## SRN – Further and stronger analogy between sampling and optimization by Arnak Dalalyan

Lately, I have been pondering the connection between Monte Carlo based sampling methods and optimation methods. It brings me to read a paper that I have been wanting to read in detail long ago. In this week’s Sunday Reading Notes series on Phyllis with Data, I discuss the paper “Further and stronger analogy between sampling and optimization: Langevin Monte Carlo and gradient descent” by Arnak Dalalyan.

Lately, I have been pondering the connection between Monte Carlo based sampling methods and optimation methods. It brings me to read a paper that I have been wanting to read in detail long ago. In this week’s Sunday Reading Notes series on Phyllis with Data, I discuss the paper “Further and stronger analogy between sampling and optimization: Langevin Monte Carlo and gradient descent” by Arnak Dalalyan.

This paper focuses on the Langevin Monte Carlo algorithm, which I believe is also called the Unadjusted Langevin algorithm (ULA) in contrast to the Metropolis adjusted Langevin algorithm (MALA). Using the Wasserstein distance as the metric, this paper establishes an upper bound of the Wasserstein distance, more precisely it shows that one needs $O(\frac{d}{\epsilon^2}\log(\frac{d}{\epsilon}))$ number of iterations to reach the precision level . Considering the close connection between ULA and gradient descent algorithm for optimization, the author proves an upper bound for the L2 distance for gradient descent by cleverly utilizing a tempering argument (details in section 3 of the paper). The well-known result for optimization is that it takes $O(\log(\frac{1}{\epsilon}))$ iterations to reach precision in gradient descent. Intuitively sampling should be a more difficult task than optimization because it requires an exploration of the whole parameter space, which can be of high-dimension. The bounds in this paper confirm this intuition. More importantly, it points out how these two problems are “continuously” connected (as the temperature converges to 0) and that this connection naturally leads to the deduction of the optimization bound from the sampling bound.

Admittedly this is not a new paper as it was first arxived in 2016, there have been many papers tackling similar problems in sampling, for example using the KL-divergence (Cheng and Bartleet 2019) or extending the results to MALA (Dwivedi et al. 2019). When I first picked up this paper two years ago as a second-year Ph.D. student, I was intimidated by the proof, which in fact was only 6-pages long. This time, as I sat down and went through the proof line by line, I found it quite accessible and elegant. I have been equipped with the mathematical knowledge required to understand the proof in college. This time, I was defeated by my own fear. For me, this paper is not only a delightful read, but also a wake-up call to trust myself.

## SRN – Distilling Importance Sampling

n this week’s Sunday Reading Notes, I discuss the paper Distilling Importance Sampling by Dennis Prangle. It aims to find the normalising flow $latex q$ that minimizes the distance to the target distribution $latex p$.

In this week’s Sunday Reading Notes, I discuss the paper Distilling Importance Sampling by Dennis Prangle. It aims to find the normalising flow $q$ that minimizes the distance to the target distribution $p$. DIS uses the inclusive Kulbleck-Leiber divergence (KL divergence from the target distribution $p$ to the approximating distribution $q$), in order to avoid over-concentration. The process of distilling refers to utilizing tempering distributions $p_{\epsilon}$ that converges to $p$. In each step, DIS using stochastic gradient descent to minimize KL from $p_{\epsilon}$ to $q$ where the gradients are estimated with self-normalized importance sampling.

Recently I came across many papers involving both Monte Carlo methods and neural networks. I learned many new concepts, among them, was normalising flow. A normalising flow is a bijective transformation from simple distributions such as normal or uniform. This paper considers non-volume preserving (NVP) normalising flow, particularly the coupling layer. It transforms input vector $u$ into output vector $v$ by fixing $d$ coordinates and shifting and scaling the remaining coordinates by neural network outputs. The resulting family of densities can be sampled rapidly and their gradients are computable.

One question I have while reading this paper is about convergence guarantees. What does the distribution $q^{\star}$ converge to? Numerical experiments suggest that DIS output is very close to the targeting distribution by comparing it to MCMC output. But as Prangle acknowledges, DIS output has less accurate tails than those of MCMC. Intuitively speaking DIS output should converge to the argmin of inclusive KL from target distribution to normalising flows. But does the distilling procedure guarantee convergence, despite the gradient estimates and tempering steps? I also keep thinking about what would be the SMC counterpart of DIS.

Reference:

Prangle, D. (2019). Distilling importance sampling. arXiv preprint arXiv:1910.03632. https://arxiv.org/abs/1910.03632

## SRN – A Framework for Adaptive MCMC Targeting Multimodal Distribution by Pompe et al.

After a very long break, I come back to the Sunday Reading Notes and write about the paper “A Frame for Adaptive MCMC Targeting Multimodal Distributions” by Emilia Pompe et al. This paper proposes a Monte Carlo method to sample from multimodal distributions.

After a very long break, I come back to the Sunday Reading Notes and write about the paper “A Frame for Adaptive MCMC Targeting Multimodal Distributions” by Emilia Pompe et al. This paper proposes a Monte Carlo method to sample from multimodal distributions. The “divide and conquer” idea separates the sampling task into (1) identifying the modes, (2) sample efficiently within modes and (3) moving between the modes. While the paper emphasizes more heavily on addressing task (2) and (3) and establishing ergodicity , it provides a framework that unites all three tasks.

The main algorithm in this paper is Jumping Adaptive Multimodal Sampler (JAMS). It has a burn-in algorithm that uses optimization-based procedure to identify modes, to merge them and to estimate covariance matrices around the modes. After task (1) is solved by the burn-in algorithm, JAMS alternates between the later two tasks with local moves that samples around a mode and jump moves that facilitate crossing the low density regions. The local moves are adaptive in the sense that it relies on local kernels with parameters that are learned on the fly. The jump moves propose a new mode and choose a new point around the new mode.

This paper proves ergodicity of JAMS by considering it as a case of a general class of Auxiliary Variable Adaptive MCMC. It also provides extensive simulation studies on multivariate Gaussian (d = 200) , mixtures of t-distributions and bananas and a Bayesian model for sensor network localisation. The results are compared against Adaptive Parallel Tempering (APT). In all the cases tested, JAMS out performs APT in terms of root mean squared error scaled by dimension of the problem. JAMS is also more robust to initialization in term of recovering all the modes.

I usually think of Monte Carlo integration and optimization as very similar tasks and even neglect the area of optimization. This paper is a wake up call for me, as I believe the current JAMS algorithm relies on a successful mode identifying stage. The discussions in Section 6 also suggest mode finding as the bottle neck.

## SRN – Controlled Sequential Monte Carlo by Heng et al. (Part 1)

Returning to my Sunday Reading Notes series from Thanksgiving recess, I want to discuss the paper ‘Controlled Sequential Monte Carlo’ by Jeremy Heng, Adrian N. Bishop, George Deligiannidis and Arnaud Doucet.

Returning to my Sunday Reading Notes series from Thanksgiving recess, I want to discuss the paper ‘Controlled Sequential Monte Carlo’ by Jeremy Heng, Adrian N. Bishop, George Deligiannidis and Arnaud Doucet. You can read Xi’an’s blog post discussing this paper.

As pointed out in Xi’an’s blog, this is a massive paper, and I have only managed to go through Section 3. This paper starts with introducing Feynman-Kac models and twisted Feynman-Kac models, which is motivated by designing observations dependent kernels for sequential Monte Carlo (SMC) methods. At the terminal time $T$, the twisted model and original model are the same and at all times $t = 0,1,2,\cdots, T$ the two models have the same normalizing constant which we try to approximate with SMC. With a current policy $\psi$, there is a policy refinement $\psi^\star = \psi \cdot \phi^\star$ that is a admissible policy and $\psi^star$ is referred to as the optimal policy by the authors because of its connection with Kullback-Leibler optimal control problem.

The policy refinements can be found using backward recursion $\psi^\star_T = G_T^\phi$ and $\psi^\star_t = Q_t^\phi \phi^{\star}_{t+1}$ for $t = 0,\cdots, T-1$ where $Q$ are Bellman operators. Because the policies $\psi^\star$ are intractable, the authors introduce logarithmic projection as a means of approximation. This is the Approximate Dynamic Programming (ADP) algorithm that the authors outlined in Algorithm 2. The final algorithm (Algorithm 3) alternates between performing a twisted SMC and ADP to find the next policy refinement.

References:

Heng, J., Bishop, A. N., Deligiannidis, G., & Doucet, A. (2017). Controlled Sequential Monte Carlo. arXiv preprint arXiv:1708.08396.

## SRN – The Iterated Auxiliary Particle Filter by Guarniero et al.

In this week’s Sunday Reading Post, I want to discuss the paper ‘The Iterated Auxiliary Particle Filter’ by Pieralberto Guarniero, Adam Johansen and Antony Lee.  The algorithm proposed, iAPF, approximates an idealized particle filter, which would have zero variance in terms of estimating the marginal likelihood of a hidden Markov model. It is worth mentioning that finding if we were to use an idealized particle filter, we would only need one particle instead of a system of particles for marginal likelihood estimations and that the idealized particle filter would be the backward smoothing provided a perfect forward filtering step.

In this week’s Sunday Reading Post, I want to discuss the paper ‘The Iterated Auxiliary Particle Filter’ (iAPF) by Pieralberto Guarniero, Adam Johansen and Antony Lee.

The algorithm proposed, iAPF, approximates an idealized particle filter, which would have zero variance in terms of estimating the marginal likelihood of a hidden Markov model. The motivation is that if we were to use an idealized particle filter, we would only need one particle instead of a system of particles for marginal likelihood estimations. The idealized particle filter would be the backward-smoothing provided a perfect forward-filtering step.

This paper builds upon previous studies of particle filters, including twisted particle filter and the Auxiliary Particle Filter (APF). The idea is to introduce a family of ‘twisted’ models through functions $\psi$ in order to define a new model that would have the same marginal likelihood as the original model. Because the optimal policy $\psi^\star$,which corresponds to the idealized particle filter, is intractable in general cases, the authors use a series of functional approximations $\psi^0,\psi^1, \cdots, \psi^l$ to approximate $\psi^\star.$

The key the question to me is how to and how well can we approximate $\psi^\star$. The ‘how to’ qustions is addressed in Section 3.3. The series of backward recursions that are used to approximate $\psi^\star$ relies on the iterative relationship $\psi^\star(x_t) = p(y_t|x_t) f(x_t,\psi^\star_{t+1}) = p(y_t|x_t)\int_X f(x_{t},x_{t+1})\psi^{\star}_{t+1}(x_{t+1}) dx_{t+1}.$ In Algorithms 3, the authors suggest that we ‘choose $\psi^{t}$ on the basis of $\psi_{t}^{1:N}$ and $\psi_t^{1:N}.$ Later in Section 5, they mention that this step is implemented with a parametric optimization step for their experiments. In the ‘Discussions’ section, the authors mention alternatives such as nonparametric estimates which would require much higher cost. More importantly is the ‘how well’ part, because it is also possible that the class of functions $\Psi$ that we consider does not contain the optimal function $\psi$. I think this is a very interesting and importantly case and the authors report ‘fairly well’ performance of iAPF in this case.

As demonstrated in the ‘Applications and Examples’ section, “iAPF can provide substantially better estimates of the marginal likelihood than the Bootstrap Particle Filter (BPS) at the same computational cost.

References:

• Guarniero, P., Johansen, A. M., & Lee, A. (2017). The iterated auxiliary particle filter. Journal of the American Statistical Association112(520), 1636-1647.
• Pitt, M. K., & Shephard, N. (1999). Filtering via simulation: Auxiliary particle filters. Journal of the American statistical association94(446), 590-599
• Whiteley, N., & Lee, A. (2014). Twisted particle filters. The Annals of Statistics42(1), 115-141.

## SRN – Racing Thompson by Zhou et al.

While browsing the accepted papers list of ICML 2018, I discovered this paper ‘Racing Thompson: an Efficient Algorithm for Thompson Sampling with Non-conjugate Priors‘ by Zhou, Zhu, and Zhuo. Thompson sampling is a popular algorithm for exploration-exploitation tradeoff problems and is also known as Bayesian bandits. I decided to write my Sunday Reading Notes post on this paper because have been interested in the exploration-exploitation tradeoff for a while and explored this topic through Bayesian optimization and my WSDM’19 paper on sequential A/B testing.

Suppose we want to identify the best arm among $K$ arms and we have some prior knowledge about their rewards $\mu \sim \pi.$ Thompson sampling (TS) balances exploring unexplored arms getting rewards from arms already yielding high rewards by choosing the $k$th arm according to its the posterior probability of being the optimal arm $P_{it} = \pi\left( \mu_i = \max_j \mu_j \right).$ The computational challenge is to compute the probabilities $P_{it}$. Because TS is often used as an online algorithm, efficient calculation of the posterior probabilities is very important. In the conjugate prior case, this calculation is done in $O(K)$. With non-conjugate priors, I have seen in the literature that people using Markov Chain Monte Carlo (MCMC) and sequential Monte Carlo (SMC). The authors recognize this probability as an expectation and propose to use an Importance Sampling (IS) step combined with Gumbel-Max trick, which transforms the sampling problem to an optimization problem, to sample $k$ at time $k$ according to probabilities $\pi\left( \mu_i = \max_j \mu_j \right) = \mathbb{E}_{\mu\sim \pi(\cdot|X(1:t))}\left[\mathbb{I}[\mu_k = \max_j {\mu_j}] \right] = \mathbb{E}_{\mu\sim B_t}\left[\mathbb{I}[\mu_k = \max_j {\mu_j}] \frac{\pi(\mu|X(1:t)}{B_t(\mu}\right].$

The benefits of this IS step comes from flexibility to choose $B_t$ at each time step and also the authors leveraged the stopping rule of racing algorithms to deterime the number of IS samples needed to approximate the expectation.

The resulting algorithm, which combines benefits from Importance Sampling, Gumble-Max trick, and the racing algorithm, is proved to be $(\delta,\sigma)-$PAC, which is asymptotic good in the sense the total variance distance between the true value $P_{it}$ and its estimate converges to zero.

What I find very interesting from the regret analysis section is the fact that the racing TS in this paper can provide much lower regret compared to Thompson sampling and prior-swapping (PS) even though it uses much few particles than SMC and PS. It is not intuitive to me why this should happen. But upon a little further investigation, I found that the priors used for TS and PS & Racing are different in both plots. For ease of implementation, the authors have chosen a conjugate prior for TS. This leaves me wondering what the results would be if we were to use MCMC or SMC with more particles as the baseline for the regret analysis.

References:

• Zhou, Y., Zhu, J. & Zhuo, J.. (2018). Racing Thompson: an Efficient Algorithm for Thompson Sampling with Non-conjugate Priors. Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:6000-6008