# Ergodicity of Combocontinuous Adaptive MCMC algorithms

This preprint by Jeff Rosenthal and Jinyoung Yang (currently available from Jeffs webpage) might also be called “Easily verifiable adaptive MCMC”. Jeff Rosenthal gave a tutorial on adaptive MCMC during MCMSki 2016 mentioning this work.  Adaptive MCMC is based on the idea that one can use the information gathered from sampling a distribution using MCMC to improve the efficiency of the sampling process.

If two conditions, diminishing adaptation and containment are satisfied, an adaptive MCMC algorithm is valid in the sense of asymptotically consistent. Diminishing adaptation means that two consecutive Markov Kernels in the algorithm will be asymptotically equal. In other words, we either stop adaptation at some point or we know that the adaptation algorithm converges.
Containment means the number of repeated applications of all used Markov Kernels to get close to the target measure is bounded. Concretely, let $\gamma$ be a Markov kernel index, $P_\gamma^m(x,\cdot)$ be the distribution resulting from m-fold application of kernel $P_\gamma$ starting from $x$ . In other words start MCMC at point x with kernel $P_\gamma$, let it run for m iterations and consider the induced distribution for the last point. Let $\pi$ be the target distribution. Then containment requires that
$\{M_\epsilon(X_n, \Gamma_n)\}_{n=1}^\infty$  is bounded in probability for all $\epsilon > 0$. Here $M_\epsilon(x, \gamma) = \inf \{ m \geq 1 : \| P_\gamma^m(x,\cdot) - \pi(\cdot) \|_\textrm{TV}\}$ and $\|\cdot\|_\textrm{TV}\}$ is a worst case distance between distributions (total variation distance).

The paper is concerned with trying to find conditions for containment in adaptive MCMC that are more easily verified than those from earlier papers. First however it gives a kind of blueprint for adaptive algorithms that satisfy containment.

### A blueprint for consistent adaptive MCMC

Nameley, let $\mathbb{R}^d$ be the support of the target distribution and $K \subseteq \mathbb{R}^d$ some large bounded region, $D > 0$ some large constant. The blueprint, Bounded Adaptive Metropolis, is the following:

Start the algorithm at some $X_0 \in K$ and fix a $d \times d$ covariance matrix $\Sigma_*$. At iteration n generate a proposal $Y_{n+1}$ by

(1)$Y_{n+1} \sim \mathcal{N}(X_n, \Sigma_*)~\textrm{if}~X_n \notin K$
(2)$Y_{n+1} \sim \mathcal{N}(X_n, \Sigma_{n+1})~\textrm{if}~X_n \in K$

Reject if $|Y_{n+1} - X_{n}| > D$, else accept with the usual Metropolis-Hastings acceptance probability. The $\Sigma_{n+1}$ can be chosen almost arbitrarily if the diminishing adaptation condition is met, so either the mechanism of choosing is fixed asymptotically or converges.

It would seem to me that we can actually change the distribution in (2) arbitrarily if we continue to meet diminishing adaptation. So for example we could use an independent metropolis, adaptive Langevin or other sophisticated proposal inside K, so long as condition (e) in the paper is satisfied, i.e. the adaptive proposal distribution used in (2) is continuous in $X_n$. Which leads us to the actual conditions for containment.

### General conditions for containment in adaptive MCMC

Let $\mathcal{X}$ be a general state space. For example in the Bounded Metropolis we had $\mathcal{X}=\mathbb{R}^d$. The conditions the authors give are (even more simplified by me):

(a)  The probability to move more than some finite distance D > 0 is zero: $Pr(|X_{n+1} - X_n| > D) = 0$
(b) Outside of K, the algorithm uses a fixed transition kernel P that never changes (and still respects that we can at most move D far away)
(c) The fixed kernel P is bounded above by $P(x, dy) \leq M \mu_*(dy)$ for finite constant M > 0 and all x that are outside K but no farther from it than D (call that set $K_D$) and all y that are between D and 2D distance from K (call that set $K_{2D} \backslash K_D$). Here $\mu_*$ is any distribution concentrated on $K_{2D} \backslash K_D$.
(d) The fixed kernel P is bounded below by $P^{n_0}(x, A) \geq \epsilon \nu_*(A)$ for some measure $\nu_*$ on $\mathcal{X}$, some $n_0 \in \mathbb{N}$ and some event A.
(e) Let $\gamma$ be the parameter adapted by the algorithm. The overall proposal densities $q_\gamma(x,y)$ (combining the proposal in and outside of K) are continuous in $\gamma$ for fixed (x,y) and combocontinuous in x. Practically, this would be that the fixed proposal when outside  K and the adaptive proposal when inside K are both continuous.

Here, conditions (a) and (b) are very easy to ensure even when not an expert on MCMC. Conditions (c) and (d) sound harder, but as mentioned above it seems to me that they are easy to ensure by just using a (truncated, i.e. respecting (a)) gaussian random walk proposal outside of K. Finally, (e) seems to boil down to making the adaptive proposal continuous in both $\gamma$ and x.

The proofs use a generalization of piecewise continuous functions and a generalized version of Dinis theorem to prove convergence in total variation distance.

This paper seems to me to be a long way from Roberts & Rosenthal (2007, Journal of Applied Probability) which was the first paper I read on ergodicity conditions for adaptive MCMC. It truly makes checking containment much easier. My one concern is that the exposition could be clearer for people that are not MCMC researchers. Then again, this is a contribution paper rather than a tutorial.

# Operator Variational Inference

This NIPS 2016 paper by Ranganath et al. is concerned with Variational Inference using objective functions other than KL-divergence between a target density $\pi$ and a proposal density $q$. It’s called Operator VI as a fancy way to say that one is flexible in constructing how exactly the objective function uses $\pi, q$ and test functions from some family $\mathcal{F}$. I completely agree with the motivation: KL-Divergence in the form $\int q(x) \log \frac{q(x)}{\pi(x)} \mathrm{d}x$ indeed underestimates the variance of $\pi$ and approximates only one mode. Using KL the other way around, $\int \pi(x) \log \frac{pi(x)}{q(x)} \mathrm{d}x$ takes all modes into account, but still tends to underestimate variance.

As a particular case, the authors suggest an objective using what they call the Langevin-Stein Operator which does not make use of the proposal density $q$  at all but uses test functions exclusively. The only requirement is that we be able to draw samples from the proposal. The authors claim that assuming access to $q$ limits applicability of an objective/operator. This claim is not substantiated however. The example they give in equation (10) is that it is not possible to find a Jacobian correction for a certain transformation of a standard normal random variable $\epsilon \sim \mathcal{N}(0,I)$  to a bimodal distribution. However their method is not the only one to get bimodality by transforming a standard normal variable and actually the Jacobian correction can be computed even for their suggested transformation! The problem they encounter really is that they throw away one dimension of $\epsilon$, which makes the tranformation lose injectivity. However by not throwing the variable away, we keep injectivity and it is possible to compute the density of the transformed variables. The reasons for not accessing the density $q$ I thus find rather unconvincing.

To compute expectations with respect to $q$, the authors suggest Monte Carlo sums, where every summand uses an evaluation of $\pi$ or its gradient. As that is the most computationally costly part in MCMC and SMC often times, I am very curious whether the method performs any better computationally than modern adaptive Monte Carlo methods.

# Accelerating Stochastic Gradient Descent using Predictive Variance Reduction

During the super nice International Conference on Monte Carlo techniques in the beginning of July in Paris at Université Descartes (photo), which featured many outstanding talks, one by Tong Zhang particularly caught my interest. He talked about several variants of Stochastic Gradient Descent (SGD) that basically use variance reduction techniques from Monte Carlo algorithms in order to improve the convergence rate versus vanilla SGD. Even though some of the papers mentioned in the talk do not always point out the connection to Monte Carlo variance reduction techniques.

One of the first works in this line, Accelerating Stochastic Gradient Descent using Predictive Variance Reduction by Johnson and Zhang, suggests using control variates to lower the variance of the loss estimate. Let $L_j(\theta_{t-1})$ be the loss for the parameter at $t-1$ and jth data point, then the usual batch gradient descent update is $\theta_{t} =\theta_{t-1} - \frac{\eta_t}{N} \sum_{j=1}^N\nabla L_j(\theta_{t-1})$ with $\eta_t$ as step size.

In naive SGD instead one picks a data point index uniformly $j \sim \mathrm{Unif}(\{1,\dots,N\})$ and uses the update $\theta_{t} =\theta_{t-1} - \eta_t \nabla L_j(\theta_{t-1})$, usually with a decreasing step size $\eta_t$ to guarantee convergence. The expected update resulting from this Monte Carlo estimate of the batch loss is exactly the batch procedure update. However the variance of the estimate is very high, resulting in slow convergence of SGD after the first steps (even in minibatch variants).

The authors choose a well-known solution to this, namely the introduction of a control variate. Keeping a version of the parameter that is close to the optimum, say $\tilde\theta$, observe that $\nabla L_j(\tilde\theta) - \frac{1}{N} \sum_{i=1}^N \nabla L_i(\tilde\theta)$ has an expected value of 0 and is thus a possible control variate. With the possible downside that whenever $\tilde\theta$ is updated, one has to go over the complete dataset.

The contribution, apart from the novel combination of knowledge, is the proof that this improves convergence. This proof assumes smoothness and strong convexity of the overall loss function and convexity of the $L_j$ for the individual data points and then shows that the proposed procedure (termed stochastic variance reduced gradient or SVRG) enjoys geometric convergence. Even though the proof uses a slightly odd version of the algorithm, namely where $\tilde\theta \sim\mathrm{Unif}(\{\theta_0,\dots,\theta_{t-1}\})$. Rather simply setting  $\tilde\theta = \theta_{t-1}$ should intuitively improve convergence, but the authors could not report a result on that. Overall a very nice idea, and one that has been discussed in more papers quite a bit, among others by Simon Lacoste-Julien and Francis Bach.