Measuring discrete sensing capability for ISAC via task mutual information

Fei Shang, Haohua Du, Panlong Yang, Xin He, Jingjing Wang, Xiang-Yang LiScience China Information Sciences 2025 · DOI

TL;DR

This paper introduces a novel framework, Discrete Task Mutual Information (DTMI), to quantitatively assess the sensing capabilities of Integrated Sensing and Communication (ISAC) systems. By providing theoretical upper and lower bounds on sensing error, DTMI offers a more robust and interpretable performance metric than traditional experimental evaluations, enabling better system design and resource allocation.

What Problem Does This Paper Address?

The advent of 6G technology promises ubiquitous sensing capabilities, enabling applications from health monitoring to autonomous driving. However, evaluating the performance of these sensing systems is challenging. Traditional methods rely heavily on post-hoc experimental evaluations, which are time-consuming, costly, and often fail to capture the full range of environmental variations that can significantly degrade performance. This makes it difficult to design robust and reliable systems a priori. Existing metrics often focus on channel estimation rather than the actual sensing task, and they struggle to account for the diverse nature of sensed objects and the ambiguity in extracting relevant features. This paper aims to bridge this gap by providing a theoretical framework to measure discrete sensing capability.

Methods

The authors propose a "sensing channel encoder model" to abstract the sensing process. This model views the sensing task as a discrete channel where the target's state ($W$) is mapped to independent features ($X^n$), which are then processed by a sensing system to produce a channel embedding ($Y^n$). Finally, a decoding algorithm infers the sensed result ($\hat{W}$).

The core contribution is the definition of Discrete Task Mutual Information (DTMI), denoted as $I(X^n; Y^n)$, which quantifies the information shared between the extracted features and the system's internal representation. Using this metric, they derive theoretical lower and upper bounds for the expected sensing error ($P_e$).

  • Lower Bound: $P_e \ge \frac{H(W) - I(X^n; Y^n) - H(P_e)}{\log m}$, where $H(\cdot)$ is entropy and $m$ is the number of possible states. This bound highlights that higher DTMI leads to lower sensing error.
  • Upper Bound: A more complex expression involving error terms ($\epsilon$) and conditional mutual information, providing a bound on the achievable error.

They also establish a sufficient condition for achieving "lossless sensing" (error-free sensing) based on the DTMI.

Datasets

The paper utilizes real-world case studies for validation, including:

  • Human detection via WiFi: Data collected in a residential setting using ESP32 devices and a camera for ground truth.
  • RFID-based electrical cabinet door state monitoring: Data collected in a factory setting using RFID tags and readers.
  • Device type identification based on traffic characteristics: Utilizes an open-source dataset from previous work.
  • Direction estimation based on Music algorithm and electromagnetic signal: Simulated data based on physical models.

No new datasets are introduced; existing data and simulations are used to validate the proposed DTMI framework.

Key Insights and Ideas

  • DTMI as a Unified Metric: DTMI provides a single, theoretically grounded metric to evaluate sensing capability, unifying information from different features in a canonical form. This is crucial for comparing different sensing approaches.
  • Theoretical Bounds for Design: The derived upper and lower bounds on sensing error allow for a priori performance assessment, guiding system design and resource allocation without extensive experimentation.
  • Explaining Multimodal Sensing: The framework theoretically explains why multimodal sensing systems (using multiple types of sensors or features) tend to perform better. Adding a new modality generally increases mutual information, thereby lowering the theoretical error bound.
  • Feature Selection Guidance: DTMI can be used to compare the effectiveness of different sensing features. A feature with higher DTMI is expected to lead to lower sensing error.
  • Data Preprocessing Limitations: The paper demonstrates that while data preprocessing can improve sensing performance, it cannot guarantee lossless sensing on its own. The fundamental information captured by the features is the limiting factor.

Key Results

  • The proposed DTMI framework shows a Pearson correlation coefficient exceeding 0.9 with conventional experimental metrics like accuracy across various case studies (human identification, displacement detection, direction estimation, device recognition).
  • The theoretical bounds derived from DTMI provide accurate estimates of sensing system errors.
  • In the device type identification case study, the system achieved over 99% accuracy for KNN classification, and the DTMI values were above the theoretical lower bound, indicating the potential for achieving the desired performance.
  • The framework successfully explains the performance benefits of multimodal sensing by showing how increased modalities lead to higher mutual information and thus lower error bounds.

Limitations and Future Work

  • Model Assumptions: The model assumes independent features and a discrete sensing task. While generalized, its direct applicability to highly correlated features or continuous sensing tasks might require further extension.
  • Complexity of Bounds: While theoretically sound, calculating the exact DTMI and bounds can be computationally intensive, especially for high-dimensional feature spaces. The paper uses numerical estimation for this.
  • Future Work: The authors suggest further exploration of resource allocation strategies in ISAC systems guided by DTMI and the development of more sophisticated DTMI estimation techniques. They also highlight the potential for DTMI to unify information from different modalities more effectively.

Summary updated

BibTeX
@article{shang2025measuring,
  author = {Fei Shang and Haohua Du and Panlong Yang and Xin He and Jingjing Wang and Xiang-Yang Li},
  title = {{Measuring discrete sensing capability for ISAC via task mutual information}},
  journal = {Science China Information Sciences},
  year = {2025},
  doi = {10.1007/s11432-024-4374-y},
}

Loading related content...

Figure