Mode Collapse in Nested Sampling

Technical Summary: Mode Collapse in Nested Sampling

Problem Statement
Nested sampling is a Monte Carlo algorithm widely used for Bayesian model comparison and posterior inference, particularly valued for its robustness in handling multi-modal posteriors. The algorithm maintains a population of $K$ "live points" sampled from the prior. In each iteration, the point with the lowest likelihood is discarded and replaced by a new sample drawn from the prior, constrained to have a likelihood higher than the discarded point.

While the algorithm is designed to naturally eliminate modes ruled out by the data, this paper identifies a specific failure mode: accidental mode die-out (or mode collapse). This occurs when the stochastic nature of point replenishment causes a valid mode to be unintentionally abandoned, even if the data supports it. This risk is particularly acute in high-dimensional, multi-modal settings where the Markov chains used for replacement (e.g., slice sampling, hit-and-run) struggle to traverse disconnected modes. If a mode loses all its live points, the sampler cannot rediscover it, leading to a biased posterior and incorrect evidence estimation.

Methodology
The author develops a probabilistic model to quantify the likelihood of this accidental collapse, drawing an analogy to the neutral Moran process in population genetics.

1. Assumptions and State Model:

   • The system is modeled with two modes (0 and 1) having fractional constrained prior masses $p$ and $1-p$.
   • The population consists of $K$ live points. The state is defined by $i$, the number of live points in mode 1.
   • The analysis assumes the likelihood-restricted prior sampler (LRPS) is faithful (samples proportional to remaining mass) and that mode rediscovery is negligible once a mode is empty.
   • To provide a conservative estimate, the model assumes equal likelihood gradients in both modes. This simplifies the removal probability to be proportional to the current occupancy ($i/K$), ignoring the fact that steeper gradients might actually protect a mode from removal.

2. Transition Dynamics:

   • In each iteration, a point is removed and a replacement is generated based on a seed drawn uniformly from the current $K$ live points.
   • Under the equal-gradient assumption, the probability of the state $i$ changing to $i+1$, $i-1$, or remaining $i$ follows a symmetric random walk on the set $\{0, 1, \dots, K\}$.
   • States $0$ and $K$ are absorbing (once a mode is empty or fully occupied, it stays that way).

3. Numerical Evaluation:

   • The probability distribution of states over time is computed by iteratively multiplying the initial binomial distribution (Eq. 1) by the transition matrix defined by the random walk probabilities (Eqs. 2–4).
   • The "evolved collapse probability" $P_{evol}(t)$ is calculated as the conditional probability of reaching an absorbing state by iteration $t$, given that both modes were initially populated.

Key Results
The numerical evaluation reveals that the probability of mode collapse follows a sigmoidal curve when plotted against the number of iterations. Crucially, the curves for different values of $K$ and $p$ align when the iteration count is normalized by the factor $K^2 \times \min(p, 1-p)$.

• Scaling Law: The probability of evolved mode collapse remains below 50% as long as the number of iterations $t$ satisfies $t \lesssim K^2 \times \min(p, 1-p)$.
• Heuristic Rule: Since the total number of iterations required to update from prior to posterior is approximately $K \times H$ (where $H$ is the information gain or KL divergence), the paper derives a rule of thumb for the minimum number of live points required to make accidental collapse unlikely:
  $$K \gtrsim \frac{H}{\min(p, 1-p)}$$
  To achieve a collapse probability well below 50%, $K$ should comfortably exceed this lower bound by a factor of a few.

Significance and Claims
The paper claims to provide a general, dimension-independent framework for understanding the risk of mode collapse in nested sampling. Its primary contribution is the derivation of a simple heuristic (Eq. 10) that links the necessary population size $K$ to the information gain $H$ and the fractional mass of the smallest mode $p$.

The author notes that for typical information gains ($H \sim 10\text{--}1000$ nats) and a modest smallest-mode fraction ($p=0.1$), the rule suggests $K$ values in the range of 100–10,000. However, the paper modestly concludes that under commonly used configurations in the literature, accidental mode die-out is likely "exceedingly rare." The analysis serves as a diagnostic tool: after an initial run, Eq. 10 can be used to estimate the probability that unintentional mode die-out occurred.

The paper also contextualizes its findings against existing dynamic prevention methods (e.g., MultiNest's clustering or UltraNest's reactive nested sampling), noting that while those methods rely on detecting modes to allocate resources, the derived scaling law offers a fundamental baseline for setting $K$ to ensure retention without explicit clustering.