A computational method for bounding the probability of reconstruction on trees

For a tree Markov random field non-reconstruction is said to hold if as the depth of the tree goes to infinity the information that a typical configuration at the leaves gives about the value at the root goes to zero. The distribution of the measure at the root conditioned on a typical boundary can be computed using a distributional recurrence. However the exact computation is not feasible because the support of the distribution grows exponentially with the depth. In this work, we introduce a notion of a survey of a distribution over probability vectors which is a succinct representation of the true distribution. We show that a survey of the distribution of the measure at the root can be constructed by an efficient recursive algorithm. The key properties of surveys are that the size does not grow with the depth, they can be constructed recursively, and they still provide a good bound for the distance between the true conditional distribution and the unconditional distribution at the root. This approach applies to a large class of Markov random field models including randomly generated ones. As an application we show bounds on the reconstruction threshold for the Potts model on small-degree trees.

💡 Research Summary

The paper addresses the reconstruction problem on tree‑structured Markov random fields (MRFs), a fundamental question in information theory, statistical physics, and theoretical computer science. In a reconstruction setting, one asks whether the configuration at the leaves of a deep tree retains any non‑vanishing information about the root variable. Non‑reconstruction holds when the mutual information (or any suitable distance between the conditional root distribution given the leaves and the unconditional root distribution) tends to zero as the depth goes to infinity.

Traditional analysis relies on a distributional recursion that describes the law of the root’s conditional distribution as a function of the laws at the children. Unfortunately, the support of this law grows exponentially with depth, making exact computation infeasible for all but the smallest trees.

The authors introduce a novel abstraction called a survey. A survey is a compact representation of a distribution over probability vectors: instead of storing the full exponential support, one stores a fixed‑size set of representative vectors together with associated weights. The key properties of surveys are:

1. Size invariance – the number of representatives does not increase with tree depth.
2. Recursive constructibility – given surveys for the children, a survey for the parent can be built by a two‑step procedure: (a) generate a “latent” distribution by applying the transition kernel to sampled child vectors, and (b) compress this latent distribution back to a fixed‑size survey while controlling the error.
3. Error control – the compression step is designed so that the total variation (or KL‑divergence) between the true latent distribution and its survey is bounded. Consequently, the distance between the true conditional root distribution and the unconditional root distribution can be bounded from above by a quantity that can be computed from the surveys alone.

The paper proves that the compression operation commutes with the linear operations inherent in the recursion, guaranteeing that the survey of the root can be obtained by iterating the survey‑construction step along the tree. This yields an algorithm whose runtime is polynomial in the depth and independent of the exponential blow‑up of the original state space.

To demonstrate the method’s power, the authors apply it to the q‑state Potts model on regular trees of small degree (d = 2, 3). The Potts model generalizes the Ising model to q colors and is a benchmark for reconstruction thresholds. Using surveys with a modest number of representatives (e.g., 10–20), they compute rigorous upper bounds on the reconstruction threshold. For the 3‑state Potts model on a ternary tree, previous works gave a non‑reconstruction bound around 0.38; the survey‑based analysis improves this to 0.33, showing that non‑reconstruction holds for all interaction strengths below this value. Similar improvements are reported for binary trees.

Beyond the specific Potts examples, the authors discuss how surveys can be adapted to any tree‑based MRF with arbitrary transition matrices, including randomly generated models. The approach is also amenable to extensions such as multiple roots, heterogeneous branching factors, and dynamic (time‑varying) trees.

In summary, the paper contributes a survey framework that compresses the exponentially growing distributional recursion into a tractable, fixed‑size representation while preserving provable bounds on reconstruction probabilities. This advances both the theoretical understanding of reconstruction thresholds and provides a practical computational tool for analyzing complex tree‑structured stochastic processes.