Bang–Bang Optimal Control of Vaccination in Metapopulation Epidemics with Linear Cost Structures

Lucas M. Moschen, M. Soledad Aronna · 2025

Read Bang–Bang Optimal Control of Vaccination in Metapopulation Epidemics with Linear Cost Structures by Lucas M. Moschen, M. Soledad Aronna (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15154.txt.

Introduction

When a disease spreads between linked towns, health officials must choose when and where to vaccinate. Lucas Machado Moschen and Maria Soledad Aronna studied that choice in 2025 for a model with several local populations. They used a linear cost, which counts things like total vaccines or hospital stays, rather than the more common squared cost.

The authors derive the rules that an optimal plan must obey, using a classic result called Pontryagin's Maximum Principle. Those rules rule out a tricky middle case called a singular arc. The result is a full picture of the best control, which takes a bang-bang form: at any time, vaccinate at full rate or not at all.

Catalog record

CREATORLucas M. Moschen, M. Soledad Aronna
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-15154

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bang–Bang Optimal Control of Vaccination in Metapopulation Epidemics with Linear Cost Structures Published in IEEE Control Systems Letters (L-CSS) at https://ieeexplore.ieee.org/document/11018532 Lucas M. Moschen, M. Soledad Aronna Abstract This paper investigates optimal vaccination strategies in a metapopulation epidemic model. We consider a linear cost to better capture operational considerations, such as the total number of vaccines or hospitalizations, in contrast to the standard quadratic cost assumption on the control. The model incorporates state and mixed control-state constraints, and we derive necessary optimality conditions based on Pontryagin’s Maximum Principle. We use Pontryagin’s result to rule out the possibility of the occurrence of singular arcs and to provide a full characterization of the optimal control. I Introduction Recent outbreaks of Ebola, Influenza, and COVID-19 have underscored the vulnerability of humanity to infectious diseases. Mathematical models provide a framework for quantifying epidemic damage, forecasting, preventing, and mitigating future outbreaks. In particular, vaccination has been shown to reduce the number of new cases and alleviate public health burdens. However, optimal vaccination policies require tractable models that capture real-world distribution and administration constraints. This paper studies optimal vaccination strategies in a general epidemic model defined over a metapopulation of K distinct groups, capturing heterogeneous transmission dynamics and resource constraints. Metapopulation epidemic models, rooted in ecological patch models [ 1 ] , have been adapted for epidemiological applications [ 2 ] , with a focus on direct graph structures and infections occurring at the nodes [ 3 , 4 , 5 ] . Recent work has enriched these models with time-varying parameters, demographic effects, and control strategies [ 6 , 7 , 8 ] . Motivated by these models, several authors have formulated optimal control problems using mainly quadratic costs for the control function [ 9 , 10 , 11 ] . The single-patch linear-cost model has been completely characterized by [ 12 ] . Recent studies have benchmarked numerical schemes for metapopulation models [ 13 ] and developed computational frameworks for COVID-19 control [ 14 , 15 ] . Alongside these computational advances, the mathematical treatment of constrained optimal control in epidemiology has been considered with integral, state, and mixed control-state constraints [ 16 ] . Theoretical results on constrained SEIR models are presented in [ 17 ] , while a control-affine SEIR model with time-specific vaccine limits is introduced in [ 18 ] . Recent contributions [ 19 , 20 ] further refined these approaches by incorporating hospital capacity and supply constraints. While most optimal vaccination models use quadratic cost functions for their convexity properties and ease of analysis [ 21 , 22 ] , real-world vaccination costs often scale linearly or sublinearly due to economies of scale [ 23 , 24 ] . This is because, as the number of vaccines administered increases, fixed costs (e.g., cold-chain infrastructure, training, logistics setup) are spread over more units. We therefore investigate how this linear cost structure influences optimal vaccination strategies on metapopulations, demonstrating a bang–bang control behavior. The main contributions of this paper are as follows. • Proof of the existence of optimal solutions for the constrained vaccination control problem in a heterogeneous, multi-group epidemic model ( 7 ); • Analysis of the optimal control problem with linear vaccination cost, including a formal proof that the optimal policy is bang–bang across K regions; • Complete characterization of the optimal control structure, with numerical experiments illustrating the theoretical results. The remainder of the paper is organized as follows. Section II formulates the metapopulation epidemic vaccination model as an optimal control problem, and Section III presents necessary optimality conditions for it. Section IV is dedicated to the main results, namely the absence of singular arcs, the bang-bang structure, and other technical lemmas. Section V presents numerical experiments confirming our theoretical results, and Section VI concludes with possible extensions. Notation. We write \mathbb{R}_{\geq 0}=[0,\infty) and \|\cdot\|_{\infty} for the essential-supremum norm. The space L^{\infty}([0,T];\mathbb{R}^{k}_{\geq 0}) consists of (equivalence classes of) essentially bounded measurable maps u:[0,T]\to\mathbb{R}^{k}_{\geq 0} . If g:\mathbb{R}^{n}\to\mathbb{R}^{m} , then D_{x}g or \bigl[\partial g_{j}/\partial x_{i}\bigr]_{j,i} denotes its m\times n Jacobian; for any matrix M , M_{ij} is its entry in row i , column j . Given a set S , \bar{S} is its closure, \inf S its infimum, and \sup S its supremum. Finally, “a.e.” abbreviates “almost everywhere.” II General Epidemic Model Formulation We consider a heterogeneous population partitioned into K compartments (e.g., by spatial position or age), each one of size n_{i}>0 , for i\in\{1,\dots,K\} . The epidemic evolves over a finite time horizon [0,T] and, for each group i\in\{1,\dots,K\} , we define at time t : • the proportion of susceptible individuals: s^{(i)}(t)\in\mathbb{R} ; • the proportions of individuals in compartments of other stages of the disease (e.g. exposed, infectious): x^{(i)}(t)=(x_{1}^{(i)}(t),\dots,x_{d}^{(i)}(t))^{\top}\in\mathbb{R}^{d}; and • the proportion of vaccinated in group i: V^{(i)}(t). Set \boldsymbol{s}=(s^{(1)},\dots,s^{(K)}) , \boldsymbol{x}=(x^{(1)}|\dots|x^{(K)})\in\mathbb{R}^{d\times K} and the total vaccinated population V(t)\coloneqq\sum_{i=1}^{K}n_{i}V^{(i)}(t) . The epidemic dynamics, subject to vaccination rates u_{i}\colon[0,T]\to\mathbb{R}_{\geq 0} , is governed by \displaystyle\dot{s}^{(i)} \displaystyle=-s^{(i)}f_{i}(\boldsymbol{x})-u_{i}s^{(i)}, (1) \displaystyle\dot{x}^{(i)} \displaystyle=s^{(i)}f_{i}(\boldsymbol{x})e_{1}+g(x^{(i)}), \displaystyle\dot{V} \displaystyle=\sum_{i=1}^{K}u_{i}n_{i}s^{(i)}, where e_{1}=(1,0,\dots,0)^{\top}\in\mathbb{R}^{d} (meaning that people that exit the susceptible compartment s^{(i)} enter in the group corresponding to the first component of x^{(i)} ), and the functions f_{i}:\mathbb{R}^{d\times K}\to\mathbb{R} and g:\mathbb{R}^{d}\to\mathbb{R}^{d} are twice continuously differentiable. The function g models within-group disease progression (for instance, passage from the exposed to infectious compartments), while f_{i} defines the force of infection per susceptible in group i . The nonlinear term s^{(i)}f_{i}(\boldsymbol{x}) captures the infection process due to temporary interregional interactions through which susceptible people from region i become infected. Examples of models with this structure include classical epidemiological models such as SIR and SEIR with vaccination (both single-group and stratified versions), as well as more complex variants, discussed in [ 14 , 15 , 25 ] . We assume that the effect of new births and deaths during the time interval of study is negligible, hence we suppose the total population is constant, a condition enforced by \sum_{j=1}^{d}g_{j}(x)=0,\text{ for all }x\in\mathbb{R}^{d}, (2) where g_{j} refers to the j th component of g . Additionally, g(0)=0 guarantees that the disease-free equilibrium is invariant. For positivity, we assume f_{i}(\boldsymbol{x})\geq 0 if \boldsymbol{x}\geq 0 and g_{i} is essentially non-negative , this is, if x\in\mathbb{R}^{d} , x\geq 0 , and x_{j}=0 , then g_{j}(x)\geq 0 . Remark 1 Propositions 1 , 2 and 3 remain valid if the susceptible dynamics include any Lipschitz-continuous migration term q_{i}(\mathbf{s}) to the s^{(i)} -dynamics. The bang–bang proof, however, relies on monotonicity, so extending Theorems 1 and 2 to include q_{i} is left for future work. Numerical tests with linear q_{i} (see Figure 2 ) also show one switch per week, suggesting the qualitative behavior is robust. Proposition 1 Under the standing assumptions, the system ( 1 ), with initial data s^{(i)}(0)>0 , x^{(i)}_{j}(0)\geq 0 , V^{(i)}(0)\geq 0, for all i\in\{1,\dots,K\} and j\in\{1,\dots,d\} , admits a unique bounded, absolutely continuous solution for each u\in L^{\infty}([0,T];\mathbb{R}^{K}_{\geq 0}) that remains in \mathbb{R}^{K(d+1)+1}_{\geq 0} . Moreover, s^{(i)}(t)\geq l_{i}>0, for all t\in[0,T] , i\in\{1,\dots,K\} . Proof: For a given u\in L^{\infty}([0,T];\mathbb{R}^{m}_{\geq 0}) , define the state vector w(t)\coloneqq(\boldsymbol{s}(t),\boldsymbol{x}(t),V(t)) so that system ( 1 ) can be written in compact form as w^{\prime}(t)=F(t,w(t)),\quad w(0)=w_{0}. The existence and uniqueness of a solution for the latter system follow from Carathéodory’s theorem [

Read the complete text