In Bayesian inference, we often rely on Markov chain Monte Carlo (MCMC) or Variational Inference (VI) to explore complex probability landscapes. The goal is to obtain a representative sample of the "posterior distribution"—the mathematical landscape representing our updated beliefs. However, a recurring problem across the literature is that our tools for checking if these algorithms have "finished" are often blind to what they haven't seen yet.
This is the problem of missing modes. When a distribution has multiple "peaks" of high probability separated by vast "valleys" of low probability, an algorithm can become trapped in one peak, reporting total success while remaining completely oblivious to the rest of the landscape.
The Illusion of Convergence
The most common way to check if an MCMC chain has converged is to look at local stability. Metrics like Effective Sample Size (ESS) or Integrated Autocorrelation Time measure how "independent" the samples are. As Arjumand Masood, Finale Doshi-Velez, and Weiwei Pan demonstrate in An Empirical Comparison of Sampling Quality Metrics: A Case Study for Bayesian Nonnegative Matrix Factorization, these metrics primarily assess local mixing. A sampler can be perfectly "mixed" within a single local mode, yielding a high ESS, while failing to explore the global posterior.
This creates a "false sense of statistical confidence," according to Jiansi Gao and colleagues in Biological causes and impacts of rugged tree landscapes in phylodynamic inference. In fields like phylogenetics, the landscape of evolutionary trees is notoriously "rugged." Chris Whidden shows in Quantifying MCMC Exploration of Phylogenetic Tree Space that standard diagnostics can drop below acceptable thresholds long before the algorithm discovers all high-probability topological peaks.
The Failure of Multi-Chain Agreement
A standard safeguard against local trapping is to run multiple independent chains starting from different points. If they all end up in the same place, we assume they have found the truth.
However, Andrew Gelman points out in For how many iterations should we run Markov chain Monte Carlo? that this agreement is a necessary but insufficient condition. If all chains are initialized within the same "basin of attraction," or if the "warmup" period isn't diverse enough, they will all sequester in the same local mode.
This is particularly problematic for Variational Inference. Nando de Freitas, Michael Jordan, and others in Variational MCMC note that if an initial variational approximation misses a distant peak, any subsequent sampling guided by that approximation is unlikely to ever find it. Aki Vehtari and Daniel Simpson emphasize in Yes, but Did It Work?: Evaluating Variational Inference that diagnostics like the $\hat{k}$ diagnostic can report a high-quality fit for the discovered mode while providing no hint that other significant modes exist.
New Frontiers in Global Diagnostics
To move beyond local mixing, researchers are developing "global" diagnostics that are sensitive to missing components:
- Nested $\hat{R}$: In Nested Rˆ: Assessing the Convergence of Markov Chain Monte Carlo When Running Many Short Chains, Matthew D. Hoffman and Pavel Sountsov propose grouping chains into "superchains." This helps isolate initialization bias from sampling variance, making it easier to see if chains are stuck in different regions.
- $L$-Lag Couplings: Niloy Biswas and Pierre E. Jacob use Estimating Convergence of Markov chains with L-Lag Couplings to provide upper bounds on the distance to the true distribution. While these can still be optimistic if modes are missed, they offer a more rigorous mathematical footing than empirical variance ratios.
- Perturbation Strategies: Even sophisticated tools like Kernelized Stein Discrepancy (KSD) have "blind spots." Xing Liu and Andrew Duncan argue in Using Perturbation to Improve Goodness-of-Fit Tests based on Kernelized Stein Discrepancy that perturbing the target distribution can reveal when a sampler is failing to capture all modes.
The shift in the field, as summarized by Christopher Nemeth in Scalable Monte Carlo for Bayesian Learning, is away from simple "random-walk" thinking and toward diagnostics that treat the posterior as a complex, potentially fragmented landscape. As Marco Francis Cusumano-Towner warns in Quantifying the probable approximation error of probabilistic inference programs, without these global checks, a "buggy" or incomplete sampler can often appear more stable—and thus more "converged"—than a correct one.