Optimality of Deterministic Thresholds in Decentralized Inference

17 concepts11 papers6 proposals

Imagine a network of distributed sensors tasked with detecting a cyberattack, identifying a sparse signal, or estimating a statistical parameter. Each sensor collects high-dimensional local data, but the communication bandwidth to the central fusion center is severely constrained—perhaps to just a single bit per sample. Intuitively, one might assume that compressing continuous data into such a restrictive format requires sophisticated, lossy compression algorithms or randomized quantization strategies to avoid introducing systematic bias.

However, a powerful body of research demonstrates that the mathematically optimal solution is surprisingly simple: local sensors should use deterministic, threshold-based rules to partition their observation space. By drawing sharp boundaries rather than employing complex or randomized encodings, decentralized networks can achieve near-optimal inference performance with minimal communication overhead.

The Sufficiency of Likelihood Ratio Quantization

Why do simple thresholds work so well? The answer begins with the classical statistical concept of a sufficient statistic. In centralized hypothesis testing, the Neyman-Pearson lemma establishes that the likelihood ratio (LR)—the ratio of the probability of the data under two competing hypotheses—contains all the information necessary to make an optimal decision.

When moving to a decentralized setting with communication constraints, the authors of Decentralized Data Reduction With Quantization Constraints demonstrated that this principle generalizes beautifully. When sensors collect independent observations, quantizing the likelihood ratio itself, rather than the raw data, is a structurally optimal approach for preserving information under bit constraints. Instead of trying to compress high-dimensional raw observations, a sensor only needs to compute the scalar likelihood ratio and compress that.

As shown by Kien C. Nguyen et al. in A Decentralized Bayesian Attack Detection Algorithm for Network Security, mapping high-dimensional observations into binary signals via threshold-based rules derived from likelihood ratios provides a computationally tractable and highly effective strategy, even when data dependencies complicate the global fusion rule. Similarly, Chengxi Li 0001 et al. leveraged this concept in Distributed Detection of Sparse Stochastic Signals via Fusion of 1-bit Local Likelihood Ratios, showing that carefully designing the quantization thresholds of local likelihood ratios significantly reduces the number of sensors required to achieve high detection accuracy.

Why Determinism Wins: The Geometry of Quasiconcavity

Knowing that we should quantize the likelihood ratio leaves open another question: should the quantization thresholds be deterministic or randomized? In many sequential decision problems, one might expect that randomizing a decision boundary could smooth out errors over time or prevent a system from getting stuck in suboptimal states.

However, foundational work by XuanLong Nguyen et al. in On optimal quantization rules for sequential decision problems and XuanLong Nguyen et al. in On Optimal Quantization Rules for Some Problems in Sequential Decentralized Detection disproved long-standing conjectures regarding the necessity of randomized strategies. By analyzing the asymptotic cost functions of sequential decentralized detection systems, the authors revealed a fundamental geometric property: the underlying cost coefficient function for binary outputs is quasiconcave.

In optimization theory, minimizing a quasiconcave function over a convex space guarantees that the optimal solution lies at the extreme points of the decision space. Because randomized strategies represent interior points (mixtures) of this space, they are inherently suboptimal compared to the extreme points, which correspond to deterministic threshold rules. Thus, when systems are restricted to stationary or blockwise stationary rules to bound computational complexity, the optimal local decision functions are guaranteed to be deterministic likelihood-ratio thresholds that cleanly partition the observation space based on the ratio of hypothesis probabilities.

Partitioning the Empirical Distribution Space

At ultra-low bit rates—such as sending a single bit or a "trit" (three states)—this deterministic thresholding manifests as a geometric partitioning of the empirical distribution (or "type") space.

The authors of On Distributed Learning with Constant Communication Bits proved that for distributed detection under constant bit constraints, sensors only need to evaluate the empirical distribution of their data. The optimal communication strategy divides this distribution space into contiguous regions separated by sharp boundaries. A sensor simply transmits a message indicating which side of the boundary its observed data distribution falls. This aligns with the theoretical results of Pierre Escamilla et al. in Distributed Hypothesis Testing with Concurrent Detections, where mapping data to a fixed-size index via distribution partitioning allows multiple detectors to reliably identify their respective target hypotheses without scaling communication overhead.

Even when dealing with more complex composite hypotheses, the structural simplicity of thresholding remains intact. Yajun Mei investigated this in Decentralized two-sided sequential tests for A normal mean and demonstrated that the optimal quantizers—termed Unambiguous Likelihood Quantizers (ULQs)—surprisingly maintain a basic interval-based thresholding form, such as checking if an observation falls within a specific range ($\lambda_1 \le X_n \le \lambda_2$). This challenges the intuition that composite hypotheses demand highly intricate, multi-level quantization schemes.

Generalization to Secrecy and Estimation

The power of deterministic thresholding extends beyond standard hypothesis testing into adversarial and estimation domains. For instance, in decentralized networks vulnerable to passive eavesdroppers, one might assume that sensors should inject randomness into their quantization rules to confuse the adversary. However, a survey by Bhavya Kailkhura et al. in Distributed Inference in the Presence of Eavesdroppers: A Survey highlights that even under strict secrecy constraints, the optimal local sensor quantizers remain deterministic Likelihood-Ratio Tests. The authors note that sensors maximize security not by degrading their own local accuracy through randomization, but by passing sharp, deterministic decisions to the fusion center, leaving security-oriented obfuscation to the final fusion stage.

Furthermore, this principle bridges the gap between hypothesis testing and parameter estimation. Ankit Pensia et al. established in The Sample Complexity of Distributed Simple Binary Hypothesis Testing under Information Constraints that interval-based reporting achieves near-optimal sample complexity in distributed statistical testing, proving that complex data compression techniques are theoretically unnecessary. Taking this a step further, Ivan Lau et al. in Sequential 1-bit Mean Estimation with Near-Optimal Sample Complexity designed a sequential 1-bit mean estimation algorithm using simple binary interval queries. By adaptively shifting the threshold based on a "localization and refinement" strategy, the authors achieved a sample complexity that nearly matches full-precision estimation.

Ultimately, the success of deterministic thresholding across these diverse settings stems from a singular truth: when information must be squeezed through a narrow communication bottleneck, the most efficient strategy is not to blur the data with randomness or complicate it with intricate compression, but to draw a sharp, mathematically optimized line in the space of evidence.

Key Foundational Work

On Optimal Quantization Rules for Some Problems in Sequential Decentralized Detection (2008) — contributed 3 concept elements to this direction.
On optimal quantization rules for sequential decision problems (2006) — contributed 2 concept elements to this direction.
A Decentralized Bayesian Attack Detection Algorithm for Network Security (2008) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_likelihood_ratio_thresholding_3a82e20e,
  title = {likelihood-ratio thresholding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/likelihood-ratio-thresholding/rdr_3a82e20e0bbd4d05a9cefce87b6b42a1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_likelihood_ratio_thresholding_3a82e20e_v1,
  title = {likelihood-ratio thresholding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/likelihood-ratio-thresholding/rdr_3a82e20e0bbd4d05a9cefce87b6b42a1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake