SDE Pandemic Simulator

• SDE pandemic simulator is a computational framework that models epidemic dynamics via continuous-time stochastic processes across various compartments.
• It employs advanced numerical schemes, Monte Carlo simulations, and data assimilation techniques to address uncertainty, heterogeneity, and non-linear epidemic behavior.
• The framework supports policy analysis through counterfactual intervention design and optimal control, enabling robust public health strategy evaluation.

Searching arXiv for recent and relevant papers on SDE-based epidemic and pandemic simulators. A stochastic differential equation (SDE) pandemic simulator is a computational epidemiological framework in which epidemic states, transmission parameters, or both evolve as stochastic processes in continuous time. In the contemporary literature, such simulators range from direct compartmental SDE systems for SIR-, SIS-, and SEIHRD-type dynamics to hybrid constructions in which a stochastic process drives one or more epidemiological coefficients, and to spatial or mean-field formulations whose large-population limits are PDEs or SPDEs. Their central purpose is to represent uncertainty, path variability, early-phase fadeout or amplification, intervention effects, and structured heterogeneity that deterministic ODE models suppress (Pasquali et al., 2021, Leitao et al., 2020, Chen et al., 2 Oct 2025).

1. Scope and model families

The core object in an SDE pandemic simulator is a stochastic state vector over epidemiological compartments. Representative constructions include stochastic SIR and SIS systems, delayed SIRD and SAI(L)RD processes simulated by Monte Carlo, stochastic $\theta$-SEIHRD models driven by a random contact rate, compliance–noncompliance epidemic systems with stochastic disease and social-contagion transmission rates, and structured or spatial models in which the state is indexed by household, age, or location (Calleri et al., 2021, Kiouvrekis et al., 2023, Parkinson et al., 24 Jun 2025, Yaqine et al., 16 Jan 2026, Emakoua, 2023).

Formulation	State representation	Distinguishing feature
Direct compartmental SDE	$S,I,R$ or $S,I$ with multiplicative noise	Brownian perturbation enters infection or recovery flows
Stochastic-parameter hybrid	Deterministic compartments plus stochastic $\beta$ process	Positivity-preserving parameter dynamics, e.g. CIR
Spatial or structured stochastic system	Empirical measures or stratified subpopulations	Movement, household mixing, or mean-field interaction

A significant boundary in the literature separates genuinely Markovian SDEs from continuous-time stochastic epidemic processes with delays. In the delayed SIRD and SAI(L)RD constructions, the state is still continuous-time and stochastic, but incubation, recovery, isolation, and immunity-loss delays make the dynamics non-Markovian when those delays are retained explicitly (Calleri et al., 2021). This is one reason the phrase “SDE pandemic simulator” is best understood as an umbrella notion centered on continuous-time stochastic epidemic simulation rather than as a single canonical equation class.

2. State variables, noise structure, and epidemiological semantics

Most SDE pandemic simulators begin from a compartmental decomposition and then specify how stochasticity enters infection, recovery, death, quarantine, or behavioral transitions. In the delayed COVID-19 models of Ferrante and collaborators, the SIRD state is $(S,I,R,D)$ and the extended SAI(L)RD state is $(S,A,I,L,R,D)$, where $A$ denotes asymptomatic individuals and $L$ isolated symptomatic infected people. Infection occurs at rate proportional to $S\times(I/N)$ with contact frequency $\beta$, while incubation and subsequent recovery or death are controlled by delays $S,I,R$0, asymptomatic branching probability $S,I,R$1, lethality $S,I,R$2, and immunity-loss probability $S,I,R$3 (Calleri et al., 2021).

In direct Itô SDE formulations, multiplicative noise is typically attached to bilinear infection terms. For the SIS model with constant total population $S,I,R$4, one formulation is

$S,I,R$5

with $S,I,R$6; the diffusion vanishes at the epidemiological boundaries, which is crucial for physical admissibility (Kiouvrekis et al., 2023). In the Italian COVID-19 SIR model, stochasticity is introduced by allowing $S,I,R$7 and $S,I,R$8 to evolve as Brownian motions with drift, which yields a two-dimensional SDE for $S,I,R$9 after mass conservation is enforced (Pasquali et al., 2021). In the compliance model, uncertainty is injected into both disease transmission $S,I$0 and social contagion transmission $S,I$1, with a single scalar Wiener process modulating multiplicative terms such as $S,I$2, $S,I$3, and $S,I$4; the authors prove global existence and nonnegativity for strictly positive initial data (Parkinson et al., 24 Jun 2025).

Another widely used construction randomizes the transmission parameter itself rather than every compartmental flow. In the stochastic $S,I$5-SEIHRD model, the only explicit stochastic term is the contact rate $S,I$6, modeled as a Cox–Ingersoll–Ross process

$S,I$7

while the compartmental evolution is conditionally deterministic given a realization of $S,I$8 (Leitao et al., 2020). This design preserves non-negativity of the contact rate and separates epidemiological bookkeeping from transmission-rate uncertainty.

Noise can also alter threshold structure. In the compliance–noncompliance model, the stochastic basic reproduction numbers for fully compliant and fully noncompliant disease-free states contain additional $S,I$9 terms, and the analysis shows that multiplicative noise raises the extinction threshold relative to the deterministic model (Parkinson et al., 24 Jun 2025). A common misconception is therefore that stochasticity merely adds dispersion around a deterministic threshold; in some models it changes the threshold itself.

3. Simulation methodology and numerical schemes

Because closed-form solutions are rare, SDE pandemic simulators are primarily numerical objects. The delayed COVID-19 SIRD and SAI(L)RD systems are solved by Monte Carlo simulation. Individuals are labeled by compartment, each infected individual generates future contact times from an exponential law $\beta$0, susceptible contacts become infected after incubation, and Bernoulli draws assign recovery versus death and symptomatic versus asymptomatic trajectories. Simulation is repeated until the average total epidemic time changes by less than a specified tolerance, and trajectories are averaged to reduce statistical noise (Calleri et al., 2021).

Discretization quality is not a minor implementation detail but a structural issue. For the stochastic SIS SDE, the naive Euler–Maruyama scheme does not preserve the epidemiological domain $\beta$1 and may generate negative or otherwise unphysical values. To avoid this, an explicit semi-discrete method is constructed after the transformation $\beta$2. The resulting update

$\beta$3

is domain preserving by construction, strongly convergent of order $\beta$4, and inherits the exponential stability of the original SDE (Kiouvrekis et al., 2023). Domain preservation is therefore a defining numerical requirement for epidemic SDE simulation, not merely a desirable property.

Other simulators adopt different schemes according to the noise structure. The Italian stochastic SIR model uses Euler–Maruyama discretization inside a Rao–Blackwellized particle filter for joint state and parameter learning (Pasquali et al., 2021). The compliance–noncompliance model uses Milstein’s method, explicitly to handle multiplicative noise more accurately than first-order Euler discretization (Parkinson et al., 24 Jun 2025). The stochastic $\beta$5-SEIHRD simulator generates exact stepwise samples of the CIR process via a non-central chi-square distribution and then integrates the compartmental ODEs with Runge–Kutta 5(4) (Dormand–Prince); Monte Carlo aggregation yields means, quartiles, confidence intervals, and worst-case scenarios (Leitao et al., 2020).

At larger scales, spatial models motivate a multilevel viewpoint. Individual-based SIR systems with Brownian movement are microscopically simulated through SDEs for positions and jump transitions for infection and recovery, but their empirical measures converge as $\beta$6 to deterministic reaction–diffusion PDEs, while $\beta$7-scaled fluctuations converge to Gaussian distribution-valued limits or linear SPDEs (Tsakou et al., 2020, Emakoua, 2023). This places SDE pandemic simulators on a continuum from particle systems to macroscopic stochastic field models.

4. Calibration, inference, and data assimilation

A pandemic simulator becomes operational only after calibration and an observation model are specified. In the stochastic SIR analysis of the Italian COVID-19 epidemic, under-detection is modeled directly: observed infected and removed counts are random fractions of the latent true state, with the proportional factors taken as Beta-distributed random variables. The Beta parameters are informed by external epidemiological indicators such as IFR/CFR ratios and, for the first wave, serological survey data. Because a single global SIR model failed over the full epidemic, the timeline is partitioned at government decrees, with each interval assigned its own SDE-SIR model and posterior initialization inherited from the previous phase (Pasquali et al., 2021).

The same work couples the discretized SDE with a Rao–Blackwellized particle filter: particles propagate epidemic states, while Kalman-type updates exploit linearity in the drift to learn parameters. This supports continuously updated estimation of $\beta$8 through posterior averages of $\beta$9 and produces satisfactory short-term forecasts on horizons of roughly one to two weeks, whereas longer-term projections remain substantially more uncertain (Pasquali et al., 2021). The observation model is thus as important as the state dynamics; a visually good fit under misspecified under-detection can still yield biased parameters and unrealistic latent states.

A later reinforcement-learning-oriented SDE simulator advances calibration and validation to national scale. That simulator uses five compartments, $(S,I,R,D)$0, $(S,I,R,D)$1, $(S,I,R,D)$2, $(S,I,R,D)$3, and $(S,I,R,D)$4, with diffusion terms of the form $(S,I,R,D)$5. The transmission parameter $(S,I,R,D)$6 is calibrated against zero-intervention COVID-19 growth curves using the Kolmogorov–Smirnov test, with $(S,I,R,D)$7 reported as the closest fit and $(S,I,R,D)$8. Validation is performed against data from 176 countries using relative AUC errors on new daily cases; reported errors include UK $(S,I,R,D)$9, USA $(S,A,I,L,R,D)$0, Italy $(S,A,I,L,R,D)$1, Argentina $(S,A,I,L,R,D)$2, New Zealand $(S,A,I,L,R,D)$3, and Vietnam $(S,A,I,L,R,D)$4, while agent-based and SIR baselines are described as having errors two to three orders of magnitude higher in the reported comparisons (Chen et al., 2 Oct 2025).

For latent SDE formulations of epidemic time series, scalable training has become an additional concern. SDE Matching replaces simulation-based adjoint sensitivity training with a simulation-free objective based on direct sampling from time marginals, achieving $(S,A,I,L,R,D)$5 time and memory per optimization step and reporting more than $(S,A,I,L,R,D)$6 speedup per iteration and roughly $(S,A,I,L,R,D)$7 faster convergence in optimization steps in its experiments. This suggests a scalable route for fitting latent stochastic epidemic models to high-dimensional or irregular public-health time series, although that paper is methodological rather than disease-specific (Bartosh et al., 4 Feb 2025).

5. Intervention design, control, and structured policy simulation

One of the most consequential uses of SDE pandemic simulators is counterfactual policy analysis. In the delayed COVID-19 Monte Carlo framework, social distancing is implemented by reducing the contact rate $(S,A,I,L,R,D)$8 once the infected fraction exceeds a threshold. The reported effects are the standard flatten-the-curve pattern: lower epidemic peak, longer duration, reduced acute health-care burden, and the possibility of second waves when immunity is temporary or restrictions are relaxed. The same simulations suggest that a pure herd-immunity strategy is too risky, with herd immunity reached only at infected fractions of roughly $(S,A,I,L,R,D)$9–$A$0 and mortality around $A$1 of the total population in the stated scenario (Calleri et al., 2021).

A more formal control-theoretic strand treats intervention as a stochastic optimal control problem. In the stochastic SIR control model, vaccination $A$2 and isolation $A$3 enter the state equations, the stochastic Hamilton–Jacobi–Bellman equation is reformulated through the stochastic minimum principle, and the resulting system of forward-backward stochastic differential equations is solved numerically by Monte Carlo simulation. Conditional expectations in the BSDE solver are approximated through Longstaff–Schwartz regression with Hermite polynomials, and the experiments use 2000 paths with daily steps over a one-year horizon under isolation-only, vaccination-only, and combined policies with high- and low-cost settings (Lesniewski, 2020).

Recent work integrates SDE simulation directly with multi-objective reinforcement learning. The five-compartment $A$4–$A$5–$A$6–$A$7–$A$8 simulator is embedded in an MORL framework with discrete action levels $A$9–$L$0 for closures, vaccination, and quarantine, and a three-component reward vector for new infections, new deaths, and economic cost. A Pareto-Conditioned Network is trained so that multiple Pareto-optimal intervention strategies can be generated without retraining for new preference vectors. The same framework is reparameterized for polio, influenza, and measles, and the measles experiments report that a modest $L$1 drop in vaccination coverage necessitates significantly more stringent and costly interventions (Chen et al., 2 Oct 2025).

Structured contact models sharpen these intervention results. In the household SDE extension of the SIR framework, the population is stratified by age and household size, public and within-household contacts are encoded by distinct matrices, and the basic reproduction number is the spectral radius of a mean offspring matrix derived from a branching approximation near the disease-free equilibrium. Simulations calibrated to equal final epidemic size show that models with household structure peak earlier, at $L$2–$L$3 days, than models without households, at $L$4–$L$5 days. At epidemic peak, household transmission accounts for about $L$6 of transmission overall, and Sobol sensitivity analysis identifies the public activity level of seniors in 4-person households as highly influential (Yaqine et al., 16 Jan 2026). This indicates that structured stochastic simulators do not merely add variance; they can reverse qualitative conclusions about timing and intervention leverage.

6. Theoretical guarantees, interpretation, and limitations

The mathematical literature on epidemic SDEs places strong emphasis on well-posedness and long-run behavior. For the SIS SDE with non-linear incidence, the state space $L$7 is invariant almost surely, positivity is guaranteed, sufficient conditions are given for asymptotic stability of the disease-free equilibrium and for almost-sure extinction, and separate conditions ensure stochastic persistence and the existence of a unique stationary measure (Builes et al., 2024). In the McKean–Vlasov extension, the drift depends on the law of $L$8 and the diffusion may be a sum of power functions; under stated coefficient conditions there is a unique strong solution, extinction and persistence regimes can be analyzed, and an interacting-particle Euler–Maruyama approximation admits an explicit strong error estimate in $L$9th moment for $S\times(I/N)$0 (Kalinin et al., 21 Apr 2026).

Spatial models provide an analogous hierarchy of guarantees. Continuous-space SIR systems with Brownian movement yield law-of-large-numbers limits given by reaction–diffusion PDEs and central-limit corrections given by Gaussian distribution-valued processes or Ornstein–Uhlenbeck SPDEs (Tsakou et al., 2020, Emakoua, 2023). These results justify replacing very large stochastic simulators by deterministic or Gaussian-corrected continuum models when the modeling objective is average behavior plus fluctuation quantification rather than exact particle-level trajectories.

Interpretation remains delicate. Adjacent stochastic epidemic frameworks based on birth–death–immigration processes and event-driven simulation, while not themselves standard Itô compartmental SDEs, show that epidemic observables can follow negative binomial laws with small $S\times(I/N)$1, long right tails, and substantial separation between mean and median. Their simulation studies emphasize that one observed epidemic curve is only one sample path among many and that early random events can determine later magnitude by factors of tens or hundreds (Kobayashi, 2020, Kobayashi, 2021). The practical implication for SDE pandemic simulators is similar: reporting only the mean trajectory is inadequate when the state distribution is skewed, multimodal, or threshold-sensitive.

A second misconception is that any stochastic discretization is automatically epidemiologically faithful. The SIS literature shows that explicit schemes can violate positivity unless they are designed to preserve the admissible domain (Kiouvrekis et al., 2023). A third is that stochasticity can be treated as an afterthought layered onto a deterministic compartment model. The compliance model, the Italian under-detection analysis, and the household-structured SDE framework all show that stochastic terms interact with observation error, social behavior, and contact heterogeneity at the level of threshold location, identifiability, and policy ranking (Parkinson et al., 24 Jun 2025, Pasquali et al., 2021, Yaqine et al., 16 Jan 2026).

In this sense, the modern SDE pandemic simulator is less a single model than a mathematically controlled simulation architecture: it combines a stochastic state equation, a numerically admissible discretization, an observation or calibration layer, and an intervention interface. Its distinguishing contribution is not merely random perturbation of a deterministic curve, but rigorous representation of uncertainty, structured heterogeneity, and policy-dependent path variability in epidemic dynamics.