Abstract
This article provides essential corrections to the mathematical framework established in "Analysis, Identification, and Validation of Discrete-Time Epidemic Processes" [1]. The corrections address two typographical errors in Assumption 3 and Corollary 2, alongside three significant misstatements in Corollary 1 and Theorems 3 and 4. These revisions are critical for ensuring the theoretical validity of parameter identification in discrete-time networked epidemic models.
Introduction
Accurately modeling the spread of viruses—whether biological or digital—requires robust mathematical frameworks. The original study [1] introduced a discrete-time model designed to analyze epidemic processes and identify key parameters, specifically infection rates ($\beta$) and recovery rates ($\delta$), from observed state data.
Central to this analysis is the concept of identifiability: the ability to uniquely determine a system's underlying parameters from its observable behavior. The original publication contained errors in the conditions required for identifiability. This erratum provides the necessary and sufficient conditions to ensure that researchers can uniquely recover the "true" parameters of an epidemic from data, preventing the risk of multiple parameter sets yielding the same observations.
Mathematical Framework for Identification
The identification process is framed as a linear system of equations where the observed change in the system's state is a function of the parameters and a system matrix $\Phi$. Specifically: $$ \text{Observed Change} = \Phi \times \text{Parameters} $$
For the parameters to be uniquely identifiable, the matrix $\Phi$ (or $\Phi_i$ for node-specific models) must have full column rank. The erratum clarifies the specific requirements for the observed states ($x^k$) and the network adjacency matrix ($A$) to satisfy this rank condition.
Key Corrections and Insights
The erratum introduces several critical updates to the original theoretical results:
1. Data Richness and Non-Zero Conditions
Parameters cannot be learned from arbitrary observations. The erratum specifies "non-zero" conditions (detailed in the corrected Equations 4 and 6) ensuring the system exhibits sufficient dynamical variety. This variety prevents the infection and recovery rates from masking each other's effects in the data.
2. Time Horizon Requirements
The duration of observation required for identification depends on the model's complexity:
- Homogeneous Spread: If infection and recovery rates are uniform across the network, a single time step of change ($T > 0$) is sufficient, provided the network structure is sufficiently diverse.
- Heterogeneous Spread: When each node $i$ has unique parameters ($\beta_i, \delta_i$), at least two time steps of change ($T > 1$) are required to separate and uniquely identify the two unknown parameters for each node.
3. Endemic State Ratios
The erratum provides a corrected formula for Corollary 2, which defines the relationship between parameters when a disease reaches a steady endemic state. It precisely relates the ratio $\delta_i / \beta_i$ to the infection levels of a node and its immediate neighbors.
Summary of Revised Theorems
- Assumption 3: Corrects a summation index typo to ensure the model maintains probability bounds (the sum of infection probabilities from neighbors must not exceed 1).
- Theorem 3 & Corollary 1: Establishes that for homogeneous spread, parameters are uniquely learnable if and only if there exists a pair of nodes and time steps where the relative infection dynamics are not perfectly proportional.
- Theorem 4: For node-specific parameters, identification is possible if and only if the local neighborhood dynamics provide at least two linearly independent observations over time.
Conclusion
While this erratum does not introduce new experimental data, it solidifies the mathematical foundation for epidemic identification. The corrected conditions highlight a fundamental limitation: if an epidemic is observed in a highly synchronized or "quiet" state, unique parameter identification is mathematically impossible. These results are foundational for developing real-time observers and filters capable of estimating disease spread from noisy or incomplete datasets.
References
[1] P. E. Paré, J. Liu, C. L. Beck, B. E. Kirwan, and T. Başar, "Analysis, Identification, and Validation of Discrete-Time Epidemic Processes," IEEE Transactions on Control Systems Technology, vol. 28, no. 1, pp. 79–93, Jan. 2020.
Summary updated