Branching Correlated Sampling Approach

• Branching correlated sampling approach is a statistical framework that leverages exact and MCMC methods to estimate offspring distributions from partially observed branching processes.
• It integrates random node sampling with preserved ancestral paths to enable rigorous inference even with highly incomplete data.
• The methodology applies to diverse fields such as biology, epidemiology, and network science, offering computational efficiency for moderate sample sizes.

The branching correlated sampling approach refers to a suite of statistical and algorithmic techniques designed to estimate properties of branching processes—specifically, the offspring distribution of a Galton–Watson process—from partial, sampled observations. This methodology leverages both the random selection of a subset of nodes in the population and the crucial retention of their ancestral identity, thereby enabling rigorous inference even with highly incomplete data. The approach integrates explicit probabilistic modeling, combinatorial enumeration, and advanced Monte Carlo techniques to attain both theoretical guarantees and practical feasibility.

1. Methodological Foundation

The branching correlated sampling approach comprises two primary inferential strategies:

• Exact Inference Method: Computes the marginal likelihood of the observed sample $S$ by performing a complete sum over all non-isomorphic trees of bounded height $L$ and maximum offspring $W$ that are consistent with the sample. The procedure acknowledges the combinatorial redundancy by weighting each non-isomorphic tree according to its number of automorphisms (multiplicities) and computes, for parameter vector $\theta$:

$$P(S|\theta) = \sum_{G \in \mathcal{G}_{L,W}} P(S|G)\,P(G|\theta)$$

where $P(G|\theta)$ is the tree probability under $\theta$, and $P(S|G)$ is the likelihood of observing $S$ under $G$, computed as:

$$P(S|G) = C_{G,S} \cdot p^{|V'|}(1-p)^{|V\setminus V'|}$$

with $C_{G,S}$ accounting for the number of root-preserving embeddings of $S$ into $G$, $p$ the sampling probability, $V'$ the set of observed nodes, and $V$ the set of all nodes in $G$.

• Approximate Inference (MCMC/Metropolis–Hastings): When the exact enumeration is intractable due to the exponential growth of the tree space, a Metropolis–Hastings chain is used to sample over possible trees, targeting a proposal distribution $g(G)$ that emphasizes compatibility with $S$. Importance sampling corrections are made via the ratio $P(G|\theta)/P(G|\theta_0)$, where $\theta_0$ is typically a tractable, empirically estimated distribution. Transitions are made by randomly selecting nodes to add or prune subtrees, with acceptance probability:

$$r = \min\left\{1,\, \frac{P(S|X_{i+1})\,P(X_{i+1}|\theta_0)\,q(X_{i+1}\to X_i)}{P(S|X_i)\,P(X_i|\theta_0)\,q(X_i\to X_{i+1})} \right\}$$

This method efficiently explores the subset of trees with high posterior probability conditional on the observed sample, thus sidestepping the need for full enumeration.

2. Sampling Scheme and Data Structure

The sampling protocol adopts random node inclusion with probability $p$, ensuring that each observed node contributes its ancestral trajectory up to the root. Formally, for each sampled node, the union of paths to the root forms the observed subgraph $S$. This structure is critical since in branching processes, much of the information about offspring distribution is encoded in the manner in which lineages coalesce and split. Empirically, sampling as little as $p = 0.14$ (14% of the population) suffices for highly accurate estimation, as demonstrated in the experiments.

3. Computational Characteristics, Accuracy, and Efficiency

Method	Accuracy	Scalability	Typical Use Case
Exact Enumeration	Highest (small $L,W$)	Poor (state explosion)	Small sample trees
MCMC / MH Sampling	Near-Exact (large $n$)	Good (up to 2000 nodes)	Medium-size samples

• Exact enumeration is precise but scales exponentially with $L$ and $W$: only practical for small trees. For example, the set $\mathcal{G}_{L,W}$ grows rapidly, rendering full enumeration infeasible beyond several hundred nodes.
• Approximate inference (MCMC) achieves comparable mean squared error (MSE) and Kullback–Leibler divergence on moderate sample sizes, while remaining tractable.

Empirical boxplots of MSE per parameter indicate sharply diminishing returns in accuracy as $p$ increases beyond intermediate sampling rates; most information is concentrated in early tree levels.

4. Mathematical Formulation and Algorithmic Core

Galton–Watson recursion:

$$X_0 = 1,\quad X_{n+1} = \sum_{i=1}^{X_n} Y^{(n)}_i,\quad Y^{(n)}_i \sim \theta$$

Tree likelihood:

$P(G|\theta) = \prod_{j=1}^{W} \theta_j^{c_j}$

where $c_j$ is the count of nodes with $j$ offspring.

Sample likelihood:

$$P(S|G) = C_{G,S}\cdot p^{|V'|}(1-p)^{|V\setminus V'|}$$

MCMC transition acceptance:

$$r = \min \left\{ 1,\, \frac{ P(S|X_{i+1}) P(X_{i+1}|\theta_{0}) q(X_{i+1}\to X_i) }{ P(S|X_{i}) P(X_{i}|\theta_{0}) q(X_i\to X_{i+1}) } \right\}$$

This mathematically rigorous approach enables computation of the maximum likelihood or Bayesian posterior of the offspring parameter $\theta$ subject to the observed, partially sampled data.

5. Applications Across Domains

The methodology is immediately relevant for:

• Population genetics and evolutionary biology: Reconstruction of lineages under severe observation limitations, such as ancient DNA with partial survival or sampling bottlenecks.
• Epidemiology and social diffusion: Tracing the spread of contagion or information where only a sample of cases' ancestral paths is available (e.g., sampled contact tracing).
• Internet topology and traceroute studies: Inferring global network properties from sampled subtrees, where complete observation is infeasible due to resource constraints or privacy.
• Statistical network science: Generalizable to any process with branching or recursive generative structure and partially observed, ancestry-tagged samples.

6. Comparative Analysis With Naïve and Heuristic Methods

Naïve estimators based on observed degree distributions, especially those tallying only frontier nodes or top-level degrees, systematically underestimate higher offspring counts due to their failure in accounting for the missing branches. Both the exact and MCMC approaches are designed to adjust for this "missingness" by directly modeling the full branching process and the sampling mechanism, thus correcting for structural biases.

Traceroute-based empirical methods share similar observational limitations, and the likelihood-based framework presented here, with ancestry information, generalizes the correction strategies and extends inference to richer models.

7. Limitations, Extensions, and Theoretical Underpinnings

• State space explosion is the principal limitation of the exact approach.
• MCMC methods depend on efficient proposal designs and correct handling of non-isomorphic trees; mixing diagnostics are critical for validity.
• Extensions: The core methodology is compatible with extensions to multi-type branching, non-homogeneous offspring distributions, and dynamic or time-varying processes.
• Theoretical guarantees: The methodology provides bounds on the number of samples required and proves that accurate offspring estimation is possible even for relatively small sampled fractions.

In conclusion, the branching correlated sampling approach formalizes the estimation of the offspring distribution in a Galton–Watson process from sampled data with ancestry information via dual exact and approximate (MCMC) methods. The framework is widely applicable, mathematically rigorous, computationally efficient for moderate-size problems, and demonstrably outperforms empirical degree-based heuristics, especially in accounting for the information contained in ancestral paths. This situates the approach as a foundational tool for inference in partially observed branching processes across biology, sociology, technology, and beyond (Murai et al., 2013).