The Hidden Flaw in Pandemic Models—and the Fix That Could Save Lives
When epidemiologists fit disease models to data, they often get the wrong answer without knowing it. A mathematical reformulation ensures they find the truth.
Standard disease models swap two critical numbers without anyone noticing—giving wrong predictions 80% of the time.
In the spring of 2020, as COVID-19 cases climbed and governments scrambled to respond, epidemiological models offered one of the few tools for glimpsing the future. Policymakers pored over curves and projections, hoping to understand when hospitals would be overwhelmed and when normal life might resume. But hidden inside those models was a problem that most officials never saw: the mathematical machinery driving the forecasts was, in a fundamental sense, broken. Not broken like a broken window—visible, dramatic, obviously wrong. Broken like a clock with identical hour and minute hands. The model looked fine. It ran without errors. But it was mathematically incapable of telling two crucial numbers apart, and this hidden flaw was quietly corrupting predictions in ways that nobody could see.
The two numbers were the incubation rate and the recovery rate—the speed at which people become infectious after exposure, and the speed at which they recover. Get these wrong, and everything downstream—projected deaths, hospital demand, the timing of peak infections—becomes unreliable. Yet as Eduard Campillo-Funollet and James Van Yperen demonstrate in a new paper published on arXiv, standard epidemiological models cannot distinguish between these two rates from observed data alone. Swapping them produces identical predictions. An optimization algorithm that thinks it's found the true values might instead have found their mirror image.
The implications ripple outward. If the models underpinning pandemic responses are structurally incapable of unique answers, how much confidence should we place in their forecasts? What happens when policymakers, journalists, and exhausted citizens try to interpret results that may be quietly wrong? Campillo-Funollet and Van Yperen's work doesn't just diagnose this problem—it provides a cure. Through a mathematical technique they call the "observational system approach," they transform the broken clock into a precise instrument, one that reliably finds the correct answer rather than one of two equally plausible wrong ones. Their reformulated model converges to accurate parameters nearly 100 percent of the time, compared to the original's 20 percent success rate. And it runs faster, too.
The Science of Modeling Disease Spread
Before understanding why standard models fail, it helps to understand how they work. The Susceptible-Exposed-Infectious-Removed model—universally known as SEIR—is one of the foundational tools of mathematical epidemiology. Picture a population divided into four compartments. Susceptible individuals (S) are those who haven't been infected and could become sick. Exposed individuals (E) have encountered the pathogen but aren't yet infectious themselves—there's a潜伏 period, a gap between initial infection and the ability to spread it to others. Infectious individuals (I) are actively shedding the pathogen and can infect others. Removed individuals (R) have recovered or died and are no longer part of the transmission chain.
The model describes how people flow between these compartments over time using a system of differential equations. When a susceptible person encounters an infectious person, there's a probability of transmission. Exposed individuals progress to infectiousness at a rate determined by the average incubation period. Infectious individuals recover—or are isolated, or die—at a rate determined by the average infectious period. The math captures the intuition that diseases spread through contact, that there's often a delay before newly infected people can infect others, and that eventually people stop being infectious.
The SEIR model gained enormous prominence during the COVID-19 pandemic. It formed the backbone of forecasts produced by universities, consulting firms, and government agencies. Researchers layered additional complexity onto it—age structure, regional variation, vaccination status, behavior changes—creating increasingly elaborate versions that aimed to capture real-world nuance. But beneath all that complexity, the core model retained a hidden flaw.
The three fundamental parameters of SEIR are β (the transmission rate), α (the incubation rate, governing how fast exposed people become infectious), and γ (the recovery rate, governing how fast infectious people recover). To use a SEIR model for prediction, you need to know these numbers. You can't measure them directly—no epidemiologist walks around with a device that directly reads the average incubation period of a novel coronavirus. Instead, you infer them by fitting the model to observed data: daily case counts, hospitalizations, deaths. The fitting process is an optimization problem. You propose a set of parameters, run the model forward in time, see how far the model's predictions diverge from what actually happened, then adjust the parameters to reduce that gap. Repeat millions of times with clever algorithms, and you hope to converge on the true values.
This is where the trouble begins.
The Identifiability Problem
Structural identifiability is a property that asks a deceptively simple question: if you knew the exact, perfect output of a system, could you recover the exact parameters that produced it? For many physical systems, the answer is yes. If you drop a ball from a known height and measure how long it takes to hit the ground, you can uniquely solve for gravitational acceleration. The parameters are identifiable. But for the SEIR model, the answer is no—and understanding why requires a short journey into the mathematics of observability.
Campillo-Funollet and Van Yperen conducted a comprehensive identifiability analysis using what they call an observational system approach. The core idea is elegant: rather than studying the model directly, you derive a new system of equations that describes only the observable quantities—those things you can actually measure, like the number of infectious individuals at each point in time. By algebraic manipulation of the original SEIR equations, they showed how the observable output depends on the parameters.
The result is striking. The observable data uniquely determines three quantities: β (the transmission rate), α+γ (the sum of incubation and recovery rates), and α×γ (their product). But it cannot determine α and γ individually. The reason is mathematical: the equations that relate the data to the parameters are symmetric in α and γ. Swap them—use 0.1 for α and 0.3 for γ instead of the other way around—and the observable output looks identical. This isn't a bug in the computer code or a limitation of the data. It's a property of the mathematics itself.
To see this concretely, consider the following: suppose you're watching an epidemic unfold. You can count how many people are infectious each day. You can calculate the rate of change of that number—how fast it's increasing or decreasing. But that rate of change depends on a combination of incubation and recovery, and there's no way to decompose that single measurement into its two parts. A high rate of change could mean fast incubation and slow recovery, or slow incubation and fast recovery. The data doesn't distinguish.
The authors proved this formally and also showed its practical consequences. They generated synthetic data using specific parameters—α=0.3 and γ=0.1—and then ran an optimization algorithm one million times, each time starting from different initial guesses. The algorithm's job was to recover the true values. In the standard formulation, about 80 percent of the runs converged to the wrong answer: the mirror-image solution α=0.1, γ=0.3. Only about 20 percent found the correct values. And a significant fraction of runs failed to converge at all, either getting stuck in numerical difficulties or timing out before finding any answer.
The figure shows this visually. Each point represents a starting guess for the optimization algorithm. Blue regions show where the algorithm correctly converged to the true values. Red regions show where it incorrectly converged to the swapped solution. Gray regions show where the algorithm failed to converge. The blue region hugging the true answer is small and isolated, surrounded by a vast sea of red. Starting from almost anywhere else, the algorithm falls into the wrong basin of attraction and stays there.
This has profound implications for epidemiological practice. When researchers fit a SEIR model to COVID-19 data and report that the incubation period is, say, 5 days and the infectious period is 10 days, they may have actually found the mirror image: 10 days and 5 days, respectively. The model doesn't know the difference. Neither might the researchers. And because downstream predictions—the shape of the epidemic curve, the timing of peak hospitalizations, the effectiveness of interventions—depend on these parameters, getting them swapped quietly corrupts everything that follows.
The Cure: A Globally Identifiable Reformulation
The standard SEIR model treats α and γ as the fundamental parameters to estimate. They're intuitive—they correspond to real biological processes—and they appear directly in the differential equations. But because they're not individually identifiable, this intuitive choice creates the problem. Campillo-Funollet and Van Yperen's solution is to abandon the intuitive parameters and work with what the data actually determines.
Their reformulation—called the UVY model, after its three state variables u, v, and y—replaces α and γ with the sum c₁ = α+γ and the product c₃ = α×γ. These new parameters are globally identifiable. The observable data determines them uniquely. Optimization algorithms can find them reliably. And here's the crucial part: once you have c₁ and c₃, you can recover the original α and γ through a simple formula, up to the ordering ambiguity. The mathematics tells you both numbers, just not which one is which.
The transformation is not just mathematical sleight of hand. The UVY model describes the same underlying dynamics as SEIR, just expressed in a different coordinate system. If you simulate both models with the same true parameter values, they produce identical predictions for everything you can observe. The only difference is what you're optimizing over. SEIR optimizes over parameters that the data can't distinguish. UVY optimizes over parameters that the data can distinguish. The first task is fundamentally impossible. The second is tractable.
To demonstrate the improvement, the authors repeated their million-iteration experiment with the UVY model. Where the original formulation converged to the correct parameters only about 20 percent of the time, the reformulation converged nearly 100 percent of the time. The wrong answer—α=0.1, γ=0.3—was no longer an option. The algorithm had only one basin of attraction to fall into, and it was centered on the truth.
The figure illustrates this dramatic improvement. The left panel shows convergence behavior for the reformulated parameters. Every trajectory heads toward the same solution. The right panel shows that from the recovered values of c₁ and c₃, you can compute x₊ and x₋, which correspond to α and γ. The algorithm recovers the true values (0.3 and 0.1), though it can't tell which is which. This ambiguity is unavoidable—the data genuinely doesn't distinguish the ordering—but at least the values themselves are correct.
The bottom panels compare convergence frequency across different optimization algorithms. The UVY model (shown in blue) consistently achieves near-perfect convergence rates, while the SEIR model (shown in red) struggles. The difference is stark. For practical purposes, if you're fitting a SEIR model to data, you might as well flip a coin to decide whether your answer is correct. The UVY reformulation eliminates the coin flip.
Numerically Stable and Computationally Faster
The identifiability fix addresses the theoretical problem, but Campillo-Funollet and Van Yperen went further. They also tackled numerical issues that plague standard implementations.
Differential equations like SEIR are solved numerically, typically by stepping forward in small increments and approximating derivatives. For some parameter regimes, the equations become "stiff"—small changes in parameters cause the solution to blow up or oscillate wildly. This creates practical difficulties: numerical solvers may take tiny steps, run slowly, or fail entirely with overflow errors. The authors found that the UVY reformulation sidesteps many of these difficulties. The change of variables transforms the dynamics into a more numerically tractable form.
When fitting all six parameters simultaneously (β, α, γ, and three initial conditions for SEIR; c₁, c₂, c₃, and the corresponding transformed initial conditions for UVY), the performance gap widens further. The UVY model converges in roughly 80 percent of trials, even with varied optimization algorithms and starting points. The original SEIR model struggles to converge at all—it frequently hits numerical overflow errors or exhausts the iteration limit without finding any solution.
Correct Convergence Rate by Model Type
Percentage of optimization trials that converged to the correct parameter values when starting from 1 million randomized initial guesses. The standard SEIR model converges to the correct solution only about 20% of the time, often finding the mirror-image wrong answer instead. The UVY reformulation achieves near-perfect convergence.
| Label | Value |
|---|---|
| UVY Model (Reformulated) | 99.8 |
| SEIR Model (Standard) | 20.3 |
The convergence disparity is stark. Across one million randomized starting points, the reformulated model finds the correct solution four times as often as the original.
Full Parameter Estimation Convergence
When fitting all six parameters simultaneously (three model parameters plus three initial conditions), the gap between the UVY reformulation and standard SEIR widens. The original model struggles to converge at all, frequently hitting numerical errors before finding any solution.
| Label | Value |
|---|---|
| UVY Model (Reformulated) | 79.4 |
| SEIR Model (Standard) | 18.6 |
Perhaps more surprisingly, when the original model does converge, it takes much longer to do so. The UVY model is not just more reliable—it's faster. Part of this is mechanical: if you're running an optimization loop that often fails, you waste time on failures. But the authors also found that the UVY dynamics are intrinsically easier to solve numerically, avoiding the stiffness that makes SEIR computationally expensive.
Effect of Sensitivity Equations on Convergence
Combining the UVY reformulation with first-order sensitivity equations (which provide exact gradients to the optimizer) yields the highest convergence rates. Even with SEIR, sensitivity equations improve reliability, but the UVY model remains substantially better.
| Label | Value |
|---|---|
| UVY Model with Sensitivity Equations | 95.2 |
| UVY Model Standard | 79.4 |
| SEIR Model with Sensitivity Equations | 44.3 |
| SEIR Model Standard | 18.6 |
The timing analysis shows that even when both models successfully converge, the UVY model requires substantially less compute time. For real-time forecasting applications where speed matters—this forecast needs to be ready before this afternoon's policy meeting—this advantage could be significant.
A Surprising Discovery: Some Things You Can Still Know
Among the mathematical results, one finding stands out for its practical significance. The effective reproduction number—often denoted ℛₜ, the average number of secondary infections generated by a typical infected person at time t—is globally identifiable even when the underlying α and γ are not. You might expect that if you can't tell α and γ apart, you can't compute anything that depends on them. But ℛₜ depends on the ratio β/γ, and while γ itself is ambiguous, the observable data constrains this ratio sufficiently.
This matters because ℛₜ is one of the most important quantities in pandemic management. When ℛₜ exceeds 1, the epidemic is growing; when it falls below 1, the epidemic is shrinking. Governments have used ℛₜ estimates to justify lockdowns, relaxations, and everything in between. If these estimates were unreliable due to identifiability issues, policy decisions based on them might have been misguided.
Campillo-Funollet and Van Yperen derived an explicit formula for ℛₜ in terms of observable quantities. Their expression shows that the effective reproduction number can be computed directly from the time series of observed infectious individuals and its derivatives—no need to first recover α and γ, with their attendant ambiguity. This provides a robust, identifiable target for estimation that sidesteps the parameter-swapping problem entirely.
The basic reproduction number ℛ₀—the expected number of secondary infections from a typical case in a fully susceptible population—is not identifiable. But during an ongoing epidemic, when population immunity is building and behavior is changing, ℛₜ is what policymakers actually need to track. The authors' result means that this crucial quantity can be estimated reliably even with a model that can't uniquely identify its component parameters.
What This Changes for Epidemiological Practice
Epidemiological modeling has a reproducibility problem that is rarely discussed in public. When two research groups fit SEIR models to the same data, they may report different parameter estimates—and not because the data is noisy or one group is more skilled than another. They may be reporting mirror-image solutions, two equally valid answers that the model cannot distinguish. A model that always returns an answer, even a wrong one, can feel like it's working. The wrongness is invisible.
During the COVID-19 pandemic, this invisibility had real consequences. Forecasts of hospital demand, which guided resource allocation and triage decisions, depended on parameter estimates that may have been systematically wrong in ways nobody could detect. Models that looked scientific and rigorous, complete with confidence intervals and sensitivity analyses, were built on a foundation that couldn't deliver unique answers.
The UVY reformulation doesn't just improve individual analyses. It changes the epistemic situation. When you fit a UVY model, you know you're getting the correct parameters (up to the unavoidable ordering ambiguity). The answer is unique. When you fit a standard SEIR model, you don't know—and can't know—whether you've found the truth or its mirror image. This is the difference between science and numerology.
The authors also explored incorporating first-order sensitivity equations into the optimization process. Sensitivity equations describe how the model output changes with respect to parameter changes—the gradient, in optimization terms. Most general-purpose optimizers approximate gradients numerically, which is slow and imprecise. Providing exact gradients, computed via the sensitivity equations, dramatically improves both convergence speed and reliability. In tests with all six parameters free, models using sensitivity equations converged more often and faster than those using standard gradient approximations.
Caveats and Open Questions
No single paper resolves all outstanding issues, and Campillo-Funollet and Van Yperen acknowledge several limitations.
First, the observable in their main analysis is the number of infectious individuals. Real data is rarely so clean. Case counts are under-reported due to asymptomatic infections, limited testing, and reporting delays. Deaths are more reliably counted but arrive weeks after infections occur. If the observable is cumulative infections or incident cases rather than current prevalence, the identifiability structure changes. The authors show that cumulative infections are globally identifiable, but other common observables require further analysis.
Second, the analysis assumes noise-free, continuous observations. Real data is noisy (measurement error, reporting variation) and discrete (you're not watching the epidemic in continuous time—you have daily or weekly summaries). Noise and discretization can further degrade identifiability. The structural identifiability results tell you what's possible in principle; practical estimation must handle the messiness of reality.
Third, the UVY formulation requires knowing the constraint that c₁² - 4c₃ ≥ 0. This inequality ensures that the recovered α and γ are real numbers rather than complex ones. Enforcing this constraint during optimization adds complexity. The authors used an optimizer that handles nonlinear constraints, which may not be available in all software packages.
Finally, while the UVY model eliminates the α-γ ambiguity, other parameters may still be problematic. β (the transmission rate) is globally identifiable, but the initial conditions S₀ and E₀ are only locally identifiable. This means there may be multiple correct starting states that produce identical trajectories, though the authors show how to characterize them.
Several directions for future work emerge. Extending the analysis to age-structured models, which were widely used during COVID-19, would be valuable. Combining the observational system approach with Bayesian inference methods could provide principled uncertainty quantification even with the residual ambiguities. And applying these techniques to real pandemic datasets—checking whether the improved convergence translates to better retrospective forecasts—would ground the theoretical results in empirical validation.
The Path Forward
The SEIR model has been in use since the 1970s. It has been extended, modified, and applied to influenza, measles, Ebola, and COVID-19. In all that time, the identifiability problem lurked unrecognized. Campillo-Funollet and Van Yperen's contribution is not just a mathematical fix but a diagnostic insight: the model wasn't wrong, but it was asking the wrong question. It asked for parameters that the data couldn't provide.
The reformulation asks a different question—one the data can answer. This is the essence of good modeling: not just finding a mathematical representation that fits, but finding one that fits the right things. The UVY model does exactly that.
For public health practitioners, the message is both cautionary and hopeful. Cautionary because existing forecasts may be unreliable in ways that can't be detected. Hopeful because better tools now exist. Implementing the UVY formulation is a matter of software and education, not new data collection or fundamental research. The path from this paper to improved pandemic forecasting is direct.
The mathematical machinery of epidemiology has always been invisible to the people whose lives it shapes. Most people will never know that their lockdown was informed by a SEIR model, or that a model they never heard of helped determine when schools closed and opened. But the decisions that reshape society rest on these foundations—and those foundations can now be made more solid.
As the world confronts the next pandemic threat, whether it arrives as a novel influenza, an unknown pathogen, or a resurgence of something familiar, the models will run again. Policymakers will ask what happens next, and the models will answer. With the observational system approach in hand, those answers can be trusted a little more. The clock, at last, can tell time.
The model doesn't know the difference. Neither might the researchers. And because downstream predictions depend on these parameters, getting them swapped quietly corrupts everything that follows.
Sign in to join the conversation.
Comments (0)
No comments yet. Be the first to share your thoughts.