Optimizing the spatio-temporal allocation of COVID-19 vaccines: Italy as a case study

While SARS-CoV-2 vaccine distribution campaigns are underway across the world, communities face the challenge of a fair and effective distribution of limited supplies. We wonder whether suitable spatial allocation strategies might significantly improve a campaign's efficacy in averting damaging outcomes. To that end, we address the problem of optimal control of COVID-19 vaccinations in a country-wide geographic and epidemiological context characterized by strong spatial heterogeneities in transmission rate and disease history. We seek the vaccine allocation strategies in space and time that minimize the number of infections in a prescribed time horizon. We examine scenarios of unfolding disease transmission across the 107 provinces of Italy, from January to April 2021, generated by a spatially explicit compartmental COVID-19 model tailored to the Italian geographic and epidemiological context. We develop a novel optimal control framework to derive optimal vaccination strategies given the epidemiological projections and constraints on vaccine supply and distribution logistic. Optimal schemes significantly outperform simple alternative allocation strategies based on incidence, population distribution, or prevalence of susceptibles in each province. Our results suggest that the complex interplay between the mobility network and the spatial heterogeneities imply highly non-trivial prioritization of local vaccination campaigns. The extent of the overall improvements in the objectives grants further inquiry aimed at refining other possibly relevant factors so far neglected. Our work thus provides a proof-of-concept of the potential of optimal control for complex and heterogeneous epidemiological contexts at country, and possibly global, scales.


Introduction 11
Supply-or deployment-limited SARS-CoV-2 vaccines [1] pose the urgent question of a 12 fair distribution of the available doses [2]. Current prioritization approaches typically 13 target groups at higher risk of severe outcomes [3,4], or their indirect protection by 14 vaccinating those with higher disease transmission [3,5,6]. Our main hypothesis is that 15 taking into account spatial heterogeneities in disease transmission when designing 16 prioritization strategies will significantly improve the effectiveness of vaccination 17 campaigns. The distribution of doses inside each country is limited by the logistic 18 capabilities of the healthcare network and the rate at which the vaccine stock is 19 replenished. Decisions concerning the best allocation strategies are to be taken under 20 these constraints. Moreover, both the complex coupling between regions due to human 21 mobility and the spatial heterogeneities in disease history and control interventions 22 make the discovery of such optimal allocation strategies an arduous task. 23 We propose an optimal control framework to explore COVID-19 vaccine distribution 24 in space and time. We study the SARS-CoV-2 epidemic in Italy, where strong spatial 25 effects arise from the geography of the disease, heterogeneous lockdown exit strategies, 26 and post-lockdown control measures [7]. The optimal control framework is applied to a 27 spatial model that has proved its reliability for Italy [8,9], updated through data 28 assimilation of a year-long epidemiological record. This allows us to unravel the best 29 possible vaccination strategy and probe the impact of vaccine allocations over the 107 30 Italian provinces. 31 The problem of vaccine allocation is of primary importance for public-health officials, 32 epidemiologists, and economists [10,11]. Roll-out strategies are conventionally based on 33 the prioritization of individuals at risk, such as health workers and elderly 34 people [12][13][14][15]. However, the heterogeneous ways in which different regions may be 35 affected by each successive wave raise questions about spatial prioritization strategies. 36 What is the best feasible spatial allocation, given supply and logistic constraints? 37 Would that differ significantly from current non geographically-optimized plans? Should 38 vaccines be distributed on the basis of demography or would it be better to prioritize 39 areas currently subject to an outbreak? How relevant are the susceptibility profile and 40 modelled future transmission in each region? of infection in each province is constituted of a local and a mobility component. The

108
The estimated vaccine efficacy and immunity duration depends on the vaccine type. 109 Because we focus on spatial patterns and differences among vaccination strategies, we 110 model the vaccination process by assuming a one-dose vaccine with instantaneous 100% 111 efficacy. Moreover, we assume that vaccines protection persists during the three months 112 of projection considered. 113 The epidemiological model, previously calibrated during the first wave of COVID-19 114 in Italy [8,9], is updated up to January 11, 2021 using an iterative particle filtering. 115 This data assimilation scheme allows us to capture the second wave of infections that 116 hit Italy in the Fall of 2020, a necessary requirement to generate model projections that 117 take into account the whole epidemic history, as shown in Figure 2. In our approach, 118 model projections are described by an ensemble of a thousand trajectories associated 119 with different parameters, whose distributions quantify the model uncertainty. 120 We consider two projection scenarios characterized by two possible rates of epidemic 121 transmission, see Figure 2. The "Optimistic" scenario assumes a constant lowering of 122 transmission from January 11, 2021 to April 11, 2021; the "Pessimistic" scenario 123 considers a gradual increase in transmission until mid February 2021, which results in a 124 third wave. 125 For each scenario, the optimal control problem is solved for one reference model 126 trajectory, whose parameters and state on January 11, 2021 are obtained as the median 127 values of the 1'000 model realizations. In this way, the reference trajectory 128 approximately represents the ensemble median in each province. Then, we assess the 129 effectiveness of the optimal allocation on the full ensemble of trajectories. 130 3) Constraints We define two types of constraints: supply constraints, which 131 determine the weekly delivery to the national stockpile; and logistic constraints, which 132 limits the maximum rate of local vaccine distribution in each province. 133 The supply constraints ensure that the model does not distribute more vaccine than 134 what is actually available in stock. We assume that the national supply of vaccine doses 135 is empty on January 11, 2021 and is replenished every Monday. We consider four 136 scenarios with weekly deliveries of 479'700 (realistic, baseline value), 1M, 1.5M and 2M 137 vaccine doses.  We control for the vaccination rate (teal arrows), aiming at minimizing incident infections (pink arrow). Individuals in compartments outside of the yellow block are able to move along the mobility network shown in B., hence the force of infection in a province is coupled to the dynamics of other connected provinces. To reduce the problem to a tractable size, we only consider the most important connections (red edges) when optimizing, but we use the full network (red and grey edges) to assess our strategies. A discussion on the effect of this simplification is provided in SI. Nodes size and color display each province's population, and edges width shows the straight of the coupling between each pair of province.
Optimal control problem formulation 151 We provide a brief methodological description of the optimal control framework. The full 152 equations are derived in the SI, along with implementation details and the source code. 153 We denote n the number of spatial nodes (n = 107 provinces in Italy) and m the 154 number of epidemic states in our model (m = 9 states). We denote as x(t) ∈ R n×m + the 155 state of the system, i.e., x(t) is a vector containing the epidemic variables  (or orange) line, representing the model trajectory obtained using the median of each ensemble parameter. A. The data on the daily hospitalizations is estimated as described in [9]; this data at the regional level is assimilated on a moving window of 14 days to update the model parameters describing the local transmission rates (see SI). B. Daily number of newly exposed individuals versus the reported positive cases. Note that, the model assumes the presence of 90% asymptomatics among the exposed individuals, who are possibly not detected by the surveillance system. C. Daily number of deaths.
all provinces: The national incidence, i.e., the sum of new infections in all provinces at time t, is 164 selected as the cost function, L(x(t), v(t)). Given our system with states x subject to 165 the dynamics (1) and controls v, the optimal control problem can be formalized as: where time is measured in days, and I b a is the set of all integers a ≤ k ≤ b. Equation (3a) 180 enforces that one can only distribute a non-negative amount of vaccine doses. Equation 181 (3b) states the logistic constraints, which limit to v max i the amount of individuals that 182 can be vaccinated each day in each node; here t d is the time at which each day starts. 183 We impose that the daily vaccination capacity of each province is proportional to its 184 population size N i , assuming a fair distribution of the sanitary infrastructure among 185 provinces, as shown in SI ( Figure S1). The constraint on the national stockpile is 186 materialized by Equation (3c), which ensures that the total vaccine allocation across all 187 nodes does not exceed the stockpile D(t). The stockpile is replenished every Monday by 188 the delivery of new vaccines, hence D(t) is a staircase function. For an overview of the 189 possible solution approaches for optimal control problems we refer the interested reader 190 to [40,41]. In particular, in this work, we use a variant of direct multiple shooting [35] 191 tailored to distributed systems [36].We solve the optimal control problem in 192 Equation (2) by a direct method, also called first discretize, then optimize, which 193 transforms the control problem into a nonlinear programming problem. We split our , and we denote as x k = x(t k ) the states at 195 time t k , and as v k the controls in interval [t k , t k+1 ]. The continuous-time dynamics F in 196 Equation (1) are transformed by numerical integration into the discrete-time operator f 197 by numerical integration. This discretization requires some care, and details are 198 provided in the SI. We thus obtain the following nonlinear programming problem: Nonlinear programming problems may be solved by readily available solvers using connections (see Figure 1), thus increasing the sparsity of the underlying spatial 215 connectivity matrix. These simplifications deliver a significant computational advantage, 216 and we verified that the impact on the model accuracy is limited. Note that, even if the 217 optimal strategy is computed using the simplified model, its impact in terms of averted 218 cases and deaths (shown in Results) is evaluated using the full epidemiological model 219 without any of these simplifications. A more detailed discussion on this subject is 220 provided in the SI.

221
The nonlinear programming problem arising from the simplified epidemiological 222 model is non-convex, and involves approximately 10 5 variables and 10 5 constraints. We 223 formulate the problem using CasADi [37] and solve it using Ipopt [38] with 224 sparsity-exploiting linear algebra solvers. In practice, solving the optimal control 225 problem takes between two to four days on a 36-cores 2.3 GHz CPU.

227
We obtain the optimal vaccination strategies for a set of eight scenarios drawn from the 228 spatial model from January 11, 2021 to April 11, 2021. These scenarios are a 229 combination of two projection scenarios (pessimistic vs optimistic) and four assumptions 230 on the weekly stockpile delivery (479'700, 1M, 1.5M or 2M doses delivered per week). In 231 each scenario, the optimal solution is a spatially explicit vaccine roll-out policy, i.e. an 232 indication of the number of vaccine doses to be deployed in each province each day.

Comparison of vaccination strategies 234
In order to measure the potential impact of the optimal allocation strategy, we compare 235 it against three non-optimized, yet reasonable alternative approaches: vaccinate 236 proportionally to the future incidence as projected by the epidemiological model;  Table 1). In the optimistic transmission scenario,  The optimal solution always outperforms all alternative strategies in terms of the 253 number of averted infections and in terms of averted infections per dose (see again Table 254 2). Incidence-based allocation comes second, while population and susceptibility-based 255 allocation are distant third and fourth respectively. The improvement between optimal 256 and incidence-based allocation is significant, ranging from 8.8% (pessimistic, 2M 257 doses/week) to 16.2% (optimistic, 479'000 doses/week). In Figure 3, the black diamonds 258 represent the percentage of adverted infections obtained using each strategy for the 259 reference trajectory, with respect to the averted infections resulting from the optimal 260 strategy. We observe that, in both the optimistic and pessimistic scenarios, the optimal 261 strategy has the largest relative benefits for the smallest stockpile.

262
In the pessimistic scenario (see Figure 3.A), when 479'700 doses are made available 263 each week, the averted infections associated with the optimal strategy in the reference 264 projection are 1.37 per dose: 25 % more compared to the strategies based on population 265 or susceptible distributions (1.08 averted infections per dose), and more than 10 % . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)  Figure 2, and Figure 3 for an assessment on a set of trajectories drawn for the posterior distribution). The first column represents the considered scenarios of weekly stockpile replenishment, i.e. the number of doses delivered to Italy every week, ranging from 479'700 to 2M.
infections per dose). These differences are smaller but still significant when increasing 268 the weekly stockpile deliveries up to 2M doses; similar results are obtained also for the 269 optimistic transmission scenario (Figure 3.B). 270 We recall that the optimal control strategy considered here is computed for a projection scenario, the optimal strategy systematically averts more infections than any 280 of the other control strategies.

281
Our results suggest that it is possible to considerably increase the impact of 282 vaccination campaigns by optimizing the vaccine allocation in space and time. For this 283 task, optimal control provides the best possible strategy and sets a benchmark for the 284 theoretical potential of a vaccination campaign.

285
Analysis of the optimal vaccine allocation 286 The optimal vaccine allocation obtained as the solution of the optimal control 287 framework is complex to analyze, and we ought to do that by unraveling the mechanism 288 behind its performance. The strategy must obey the imposed logistic and supply . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021.  infections per vaccine doses from January 11, 2021 to April 11, 2021 resulting from province-scale vaccine allocation strategies for both the pessimistic (panel A) and the optimistic (panel B) scenarios based on: the optimal solution, population, proportion of susceptible individuals, and projected incidence (see color codes in the legend). We optimize the vaccine allocation for a (median) reference trajectory (diamonds), and assess the performance of the computed vaccination strategy over the whole posterior (box plots). For each projection scenario, results are normalized by the number of averted infections in the reference solution (see Table 1 for absolute values). Results for alternative scenarios and vaccination strategies are shown in SI Figure S5.
that the number of doses distributed across the country can be of 0.5M per day at 293 maximum (more details in SI Figure S1). 294 We display the optimal vaccination course in time for the pessimistic, 479'700 295 doses/week scenario in Figure 4. We observe that the optimal allocation respects the 296 two constraints on distribution (Figure 4, A) and supply ( Figure 4, B). We observe that 297 no province is vaccinated at the maximum possible rate during the whole campaign, and 298 that provinces display a variety of vaccination patterns. We also note that the vaccines 299 received every Monday are always distributed during the following week, but that the 300 rate of delivery on a national level increases with time ( Figure 4, B). Surprisingly, the 301 optimal solution favors precise targeting over speed of delivery, in order to allocate more 302 May 6, 2021 10/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint  proportion of vaccine doses administered in the 107 provinces, (some of which are highlighted) . The local distribution rate is limited by a rate that is proportional to population. This logistic constraint is visualized here as the maximum slope, equal for every province. B. Stacked cumulative absolute number of vaccines in the 107 provinces of Italy. The national stockpile is shown in black, and is replenished every week (say on Mondays) with 479'700 doses. We display the name of the provinces with a final allocation of more than 150'000 doses.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint doses to those provinces where the impact of vaccines on the whole system is projected 303 to be higher. Hence, in order to control infections, precise targeting may trump delivery 304 speed. Visual inspection suffices to reveal that the optimal allocation strategy is more alike the incidence-based allocation, but with vaccine doses spread-out to more provinces.
Furthermore, we observe in the optimal solution that every time a province is 306 vaccinated, the rate of vaccination is equal to the maximum rate allowed by the local 307 logistic constraint. This property is common to the alternative vaccination strategies, 308 hence the difference in performance is due to the spatial allocation patterns.In Figure 5, 309 one can already see by visual inspection that the optimal allocation distributes most of 310 the available doses on a few provinces with high incidence. These provinces are neither 311 the most connected nor the most populous in Italy. The optimal strategy makes then 312 use of the information on the network connectivity to fine-tune the allocation, and 313 deploys the vaccination on more provinces than the incidence-based strategy.

314
To further investigate these patterns, in Figure 6.A we display the number of 315 administered doses vs the incidence projected without vaccines (the proxy variables 316 leading to the second-best control performance), both normalized according to the 317 resident population in each province. We observe an allocation pattern whereby 318 provinces with a higher incidence receive more vaccines. However, the allocation is 319 non-linear with respect to the projected incidence, suggesting that to better control the 320 epidemic, the optimal allocation strategy takes into account other factors such as the 321 importance of each province within the mobility network, as well as the proportion of 322 susceptibles. When the weekly stockpile delivery is increased, as shown in Figure 6.B, 323 this pattern shifts to the right while remaining qualitatively consistent. Hence, the 324 May 6, 2021 12/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint optimal allocation strategy is robust with respect to the overall vaccine availability  The goal is to distribute the vaccines where they have the strongest beneficial impact on 335 the dynamics of the epidemic. However, deriving an algorithm capable of computing 336 spatially optimal allocation strategies in real, heterogeneous settings is far from trivial 337 and our approach is, to the best of our knowledge, the first attempt in this direction. 338 We developed an novel optimal control framework that delivers the best vaccination 339 strategy under constraints on supply and logistics. This allows us to compute the 340 allocation strategy that maximizes the number of averted infections during a projection 341 of the COVID-19 epidemic in Italy from January 11, 2021 to April 11, 2021. Our results 342 show that the optimal strategy has a complex structure that mainly reflects the 343 May 6, 2021 13/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; projected incidence of each province, but also takes into account the spatial connectivity 344 provided by the mobility network and the landscape of acquired population immunity. 345 Although the reason why this strategy is optimal is not immediately intuitive, our 346 simulations clearly outline its better overall performances over other, more 347 straightforward strategies. This comparison suggests that the simplicity underlying 348 intuitive vaccination strategies may undermine their effectiveness, and calls for 349 complementing these simple approaches with rigorous and objective mathematical tools, 350 like optimal control, that allow a full account of the complexity of the problem.

351
With the present work, we show that it is possible to solve optimal control problems 352 for spatially explicit dynamical models of infectious diseases at a national scale, thus

365
Our method is obviously not devoid of limitations. The main one is that the optimal 366 vaccination strategy strongly depends on the projection of the underlying 367 epidemiological model. These projections are subject to several sources of uncertainty, 368 especially for long term horizons, for example due to model design and calibration [42], 369 the generation of baseline transmission scenarios, and unforeseen events that may Control [43], consists in compensating the performance losses expected over long 376 horizons by constantly adapting the optimal strategy. In this context, Model Predictive 377 Control might be implemented using the following steps: (a) at the beginning of each 378 week, the state of the system is estimated by using newly acquired epidemiological data; 379 (b) the optimization problem is solved over a fixed prediction horizon using the 380 estimated state as initial condition; (c) the optimal strategy for the first week is applied 381 and, as soon as the next weeks starts, these steps are repeated starting from (a). This 382 method corrects the model inaccuracies by constantly resetting the initial state to the 383 estimated one. Moreover, constraints may be updated to account for unexpected 384 deliveries or new orders. Future work will aim at further evaluating the benefits of 385 implementing this scheme for the design of optimal vaccination strategies.

386
Moreover, the epidemiological model underlying our control optimization has known 387 validity and limitations [8,9]. An additional limitation of the model for the specific 388 scopes of this work is that it does not account explicitly for risk-based classes, and thus 389 does not account for the heterogeneities that may result from the demography of the The proposed framework is constituted of two disease transmission models, one "true" 420 model and a simplified one used for control:

421
• The full model is a COVID-19 model, as designed in [8,9]. This model is  This division is necessary in order to solve the optimal control problem (OCP) in a 432 reasonable time. To adapt our framework to another model/country, one would need to 433 update the "true" model to a suitable candidate (which could be a stochastic model, a 434 Hidden Markov model, or any other kind) and design a tractable approximation of this 435 new model to be solved by optimal control.

436
In order to evaluate the effectiveness of our approach, we first compute the optimal 437 vaccination course that minimizes the objective based on the simplified model. Then, is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint It the subsections below, we first detail the full COVID-19 model, then we describe 443 the optimal control framework and the simplifications we have introduced to bring the 444 problem to a tractable form.

COVID-19 model 446
The optimal control framework may be used with any compartmental SARS-CoV-2 447 transmission model that can be approximated by ordinary differential equations. To 448 demonstrate its usefulness, we use a complex model based on previous work that was 449 aimed to describe the first wave of COVID-19 infections in Italy [8,9]. We consider the 450 107 Italian provinces and the spatial connections induced by human mobility fluxes. In 451 each province, the human population is subdivided according to infection status into the 452 epidemiological compartments of susceptible S, exposed E, pre-symptomatic P  The COVID-19 transmission dynamics are described by the following set of ordinary 459 differential equations in each node i: (S1) 461 We define N i the population of province i. Susceptible individuals get exposed to the 462 pathogen at rate λ i (t), corresponding to the force of infection for community i, thus 463 becoming latently infected (but not infectious yet). Exposed individuals transition to 464 the post-latent, infectious stage at rate δ E . Post-latent individuals progress to the next 465 infectious classes at rate δ P , developing an infection that can be either 466 symptomatic-with probability σ-or asymptomatic-with probability 1 − σ.
May 6, 2021 16/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint Vaccinated individuals are moved at rate r v i (t) from their original compartments to 479 compartment V , where they do not contribute to the infection anymore.

480
The force of infection of the full model is specified as in Gatto et al. [8], Bertuzzo et 481 al. [9]. In addition to province's local dynamics, it also considers that local susceptibles 482 may enter in contact with infected individuals that are traveling, and oppositely, 483 susceptible commuters may become infected through contact with local infected. The 484 force of infection of the OCP model slightly simplified, and detailed thereafter. 485 We split the force of infection λ i (t) as the sum of the local force of infection λ L i (t), 486 from infected in node i and a mobility-driven force of infection from the network λ N i (t), 487 hence λ i (t) = λ L i (t) + λ M i (t). We observe while running our model that λ M i (t) λ L i (t). 488 Hence this artificial separation will be exploited when simplifying our model. As 489 described below, we update λ M i (t) every day whereas λ L i (t) is updated at each 490 integration step.

491
As for the formulations of the force of infection, we recall here the formulas designed 492 by [8,9]. The local force of infection reads: and the influence of other provinces on province i is written as: where β 0 is the baseline transmission rate, while β i (t) is a spatially distributed and

502
The objective for our model is to minimize the total incidence of infections, i.e., 503 t f ti i λ i (t)S i . Note that for the present model, this is equivalent to optimizing the 504 total deaths or hospital admissions, as without risk-classes the sizes of these two 505 compartments are proportional to each other.

506
Optimal control 507 We lump the epidemiological compartments of each node i in variable 508 ) and we define v i (t) as 509 our control variable, representing the number of vaccines administered in node i at time 510 t. We express the dynamics of the epidemiological model (Equation (S1)) as an ordinary 511 differential equation in each province i: where m i (t) carries the contribution of other provinces to the force of infection of node i. 513 For simplicity, we drop the time dependence in the equations below, and we define the 514 state and control variables for the full system as is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021 for each province. This logistic constraint bounds the maximum number of vaccines to 0.5M of doses per day, with a local rate that is proportional to the node population. Here we show the maximum vaccination rate for each province (the constraint the solution has to comply with), in red, and the maximum rate prescribed by the optimal solution while simulating the pessimistic scenario with a stockpile delivery of 479'700 doses, in black. The optimal solution uses the maximal capacity of the logistic network, while respecting the constraint defined.
where n is the number of spatial node considered (n = 107). The global dynamics for all 518 provinces are denoted: The coupled force of infection in node i is denoted λ i . We define the cost function as 521 the sum of total incidence of infections (transitions S i −→ E i ) for every node i, i.e., For the sake of generality, we introduce the terminal cost M , which can be used to 524 ensure that we leave the system in a proper state instead of optimizing for short-term 525 gain. Since properly designing the terminal cost could require a long analysis, for 526 simplicity we do not use it in this work, hence M (·) = 0. 527 Given our dynamical system with states x, controls v, and dynamics F , the OCP is: where we aim at minimizing the cost function over the prediction horizon T , while 534 enforcing the modeled SARS-CoV-2 transmission dynamics (Equations (S6b) and 535 (S6c)). Moreover, constraints on vaccine availability and maximum vaccination rate are 536 lumped in function H, which reads: where time is measured in days, and I b a is the set of all integers a ≤ k ≤ b. Equation

542
(S7a) enforces that one can only distribute a non-negative amount of vaccine doses.

543
May 6, 2021 18/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint Equation (S7b) states the logistic constraints, which limit the amount of individuals 544 that can be vaccinated each day in each node to v max i ; here t d is the time at which each 545 day starts. We assume that the daily capacity of each province is proportional to the 546 population size of each node N i , because we assume a fair distribution of the sanitary 547 infrastructure among provinces with regard to population, as shown in SI Figure S1. 548 The stockpile is materialized by Equation (S7c), which ensures that the total vaccine 549 allocation across every node does not exceed the total availability D(t). The stockpile is 550 replenished every Monday by the delivery of new vaccines, hence D(t) is a staircase 551 function. 552 We convert our problem formulation (S6) to a nonlinear programming problem using 553 direct multiple shooting. Standard multiple shooting splits the time horizon [0, T ] using 554 a time grid t 0 , . . . , t N , with N + 1 points and t 0 = 0, t N = T . The control function is 555 parameterized using basis functions with local support. Common choices are a uniform 556 time grid, i.e., t k+1 = t k + δ t and a piecewise constant control function, i.e., v(t) = v k , 557 t ∈ [t k , t k+1 ]. The system dynamics are then discretized to obtain a discrete-time system 558

564
We perform the discretization using numerical integration techniques (such as a 565 fourth-order Runge-Kutta scheme, with 50 steps per days) to obtain a good 566 approximation of the true trajectory and cost. Finally, the path constraints H are 567 relaxed and imposed at a finite amount of time instants, here coinciding with the time 568 grid t 0 , . . . , t N . We ought to observe that, since in our case the constraints only involve 569 the controls, we are not introducing any approximation by enforcing these constraints In (S8), both the states x = (x 0 , . . . , x N ) and the controls v = (v 0 , . . . , v N −1 ) are 578 defined as optimization variables, which is a distinguishing trait of multiple shooting as 579 opposed to single shooting.

580
The main difficulty in solving (S8) in the context of this paper is the large dimension 581 of the system and the nonlinearity of the model, which can pose severe issues to the 582 numerical solvers. In the following, we will thus introduce a few simplifications, and we 583 will verify through numerical simulations that these simplifications do not imply large 584 errors in the solution of the OCP. 585 We discretize the OCP using a uniform grid with sampling time δt = 1 day. We . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; dynamics by introducing an auxiliary control variable z that is constrained to match the 591 force of infection due to the other nodes at the beginning of each time interval. Then, 592 the dynamics of the decoupled system in each node can be written as:

596
Discussion on Simplification (a). We ought to remark that, realistically, 597 vaccinations will occur at least eight hours per day. Our assumption, while justified as a 598 computationally convenient approximation of reality, is not a priori worse than 599 assuming that vaccine administration takes place over the whole day. More refined meaningless. The main issue in this case is that the optimizer will exploit these 607 inaccuracies in order to reduce the cost. Therefore, this issue is much more evident 608 when solving optimal control problems than when simply simulating the system 609 dynamics. We have investigated some simple approaches to tackle this issue, but no 610 technique yielded satisfactory performances. It is our impression that ad-hoc integration 611 strategies will be required in order to reliably simulate and optimize dynamics with 612 continuous vaccination rates. While this will be the subject of future research, the 613 results obtained with the current approximation have yielded sufficient accuracy.

614
Discussion on Simplification (b). This simplification has been proposed in [36] as 615 an approach to solve distributed optimal control problems by means of multiple 616 shooting. In the original version, the coupling variable z is not necessarily piecewise 617 constant, but rather piecewise polynomial. We have observed in simulations that, for 618 this problem, the piecewise constant parametrization yielded sufficient accuracy. 619 We discretize the dynamics of each node using an explicit Runge-Kutta integrator of 620 order four, with 50 integration steps per day. Alternative integrators such as explicit  Discussion on Simplification (c). We sparsify the mobility matrix by pruning 627 element below a threshold (see Figure S2). This operation reduces the number of 628 connection between nodes. Also in this case, we verified through numerical simulations 629 that the introduced simplification had a small impact on the prediction and control 630 accuracy.

631
Possible further improvements Applying optimal control in open loop, i.e., 632 solving the optimization problem once and applying the control input over the whole 633 time interval, may lead to poor performance due to model inaccuracy and external 634 perturbations. A common remedy consists in closing the loop by repeatedly solving the 635 OCP by using the most updated information on the initial states. This is the principle 636 behind Model Predictive Control (MPC) [43]. In this context, the state would be 637 May 6, 2021 20/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. Fig S2. Simplification of the mobility matrix to obtain a sparse and tractable problem. After the optimization, we assess the effectiveness of the optimal control strategy on the full model. estimated on a daily, weekly, or monthly basis so as to solve again the OCP and correct 638 the optimal strategy.

639
Implementation of the OCP 640 We implement the optimal control framework using the automatic differentiation 641 framework CasADi [44], the interior-point solver ipopt [38], and the HSL ma86 large 642 sparse symmetric indefinite solver [45]. The full framework and analysis code is 643 available here: https://github.com/jcblemai/COVID-19_italy-vaccination-oc (a 644 zenodo DOI will be added after reviews). The regional transmission rates are the main parameters governing the force of infection 652 of the model and, thus, the daily exposed individuals. To better track possible changes 653 in the transmission rates, we adopt a data assimilation strategy based on an iterative 654 particle filter [46] used on a moving window of 14 days. The filter starts considering 655 N r = 1000 model realizations at time t 0 (February 21st, 2020), whose state variables are 656 x (j) 0 , j = 1, . . . , N r , where the superscript (j) is the realization index and the subscript 657 is the temporal index. Each realization is associated with a parameter combination that 658 is randomly sampled from the posterior distribution evaluated in [9], indicated with θ (j) . 659 Possible spatial heterogeneities in regional transmission on a given day t k are obtained 660 multiplying the transmission parameter by a coefficient φ is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. (daily hospitalizations) collected in a temporal window of τ = 14 days, (t k , t k + τ . New 666 coefficients from the truncated normal distribution (mean µ k = 1, standard deviation 667 0.4, bounds 0.8µ k -1.2µ k ) are sampled at time τ = t 0 + 14 days. For each realization, we 668 run the model during the window of 14 days, assuming that the coefficients change 669 linearly for a week, from φ (j) 0,i toφ (j) 0,i , and remain constant afterwards. The regional 670 likelihood of each realization is then evaluated during these two weeks considering that 671 the daily hospitalizations follow a gamma distribution (as in [9]). A resampling step 672 (systematic resampling , see, e.g. [47]) selects and duplicates the coefficientsφ (j) 0,i 673 associated with the largest likelihood values. These coefficients are then used to update 674 the mean value µ k . Finally, the simulation is repeated on the same temporal window by 675 sampling new coefficientsφ (j) 0,i from the truncated normal distribution with the updated 676 mean µ k . This set of coefficients is used to compute state variables and parameters at 677 time t k , and then as starting condition to produce the projections used in the main text. 678 Model parameters (in the absence of vaccination) are taken from a paper [9] where 679 they were inferred in a Bayesian framework for the period February 24th -May 1st, 680 2020, on the basis of the official epidemiological bulletins released daily by Dipartimento 681 della Protezione Civile [48] (data available online at 682 https://github.com/pcm-dpc/COVID-19) and the bulletins of Epicentro, at Istituto 683 Superiore di Sanità [49,50]. All the parameters estimated for the initial phase of the 684 Italian COVID-19 epidemic, including the transmission rates, are spatially 685 homogeneous [9]. This parameterization has been used to produce all the results 686 presented in the main text.  The data to quantify nation-wide human mobility come from ISTAT (specifically, Material in [8]).

704
Alternative strategies 705 We designed alternative strategies to compare the optimal solutions. Each strategy uses 706 a decision variable, V i , as a basis for the allocation of vaccines among provinces. The 707 decision variable is one of:

708
• modelled future incidence, absolute: the modelled total future incidence in a 709 no-vaccination scenario. This is equivalent to the objective of the optimal control 710 problem with no control;

711
• modelled future incidence, per population: as above, but normalized by 712 the resident population in each node; 713 May 6, 2021 24/33 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 13, 2021. ; https://doi.org/10.1101/2021.05.06.21256732 doi: medRxiv preprint • modelled initial susceptibility, absolute: the modelled number of 714 susceptibles in each province at the start of the vaccination campaign;

715
• modelled initial susceptibility, per population: as above, but normalized 716 by the resident population in each node; 717 • province's population. 718 We define two strategies to distribute the doses:

719
• Focused Where every province is sorted (higher on top) according to its decision 720 variable V i . We then allocate the maximum local rate v max i to every province 721 going down through the list, until the stockpile is empty. In other words, 722 assuming we have an amount K of vaccines in the stockpile, we find the province 723 index i that satisfy max i V i , and we assign to vaccines. Then, we find the province j that satisfy max j,j =i V j and we assign it And so on, until no vaccine remains in the stockpile.

726
This strategy will concentrate the allocation on nodes with the highest values of 727 the considered decision variable.

728
• Proportional In this case, assuming that on a given day there is a quantity of 729 vaccine K in the stockpile, we assign to each province i an amount . This approach vaccinate each node proportionally to 731 the value of its decision variable V i .

732
In the main text, we show the results for three alternative strategies, namely 733 proportional absolute incidence, proportional population, and proportional 734 susceptibility-named respectively Incidence, Population, and Susceptibility. These show the results for all these alternative strategies.

738
Additional results for the spatial model 739 We present the results for all these strategies in Table 2, and we show them side-by-side 740 in Figure S5. The optimal solutions outperforms all the others solution. In fact, for 741 every given posterior realisation, the optimal control solution always outperforms all 742 other allocation strategies. Even if we observe some scatter when sampling the 743 posterior, the performances of optimal strategies are clearly separated from the rest of 744 the alternatives.

745
To further investigate the features of the optimal solution, we present a linear scatter 746 plot of the optimal proportion of vaccinated individuals per province (sorting variable) 747 side by side with the province population, the projected incidence without vaccination, 748 and the proportion of susceptible individuals at the start of the simulation. We present 749 these results for the optimistic scenario in Figure S6 and for the pessimistic scenario in 750 Figure S7. We find no clear visual pattern associating these covariates to the optimal 751 proportion vaccinated, highlighting again that the optimal allocation uses the  . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)   vaccine dose from January 11, 2021 to April 11, 2021 using different vaccine distribution strategies for the pessimistic (panel A) and the optimistic (panel B) scenario based on: the optimal solution, the spatial distribution of the population, the amount of susceptible individuals at the beginning of the vaccination campaign, and the projected disease incidence in the absence of control. We optimize a median realization of the modeled posterior (diamonds), and assess the performance on the whole posterior (box plots). The results are normalized by the number of averted infections in the optimized solution (see Table 2 for absolute values).