🔬 physics

Exact moment equivalence and structural nonidentifiability in nonlinear epidemics

This paper demonstrates that in finite-population nonlinear epidemic models, distinct group-size distributions can produce identical stochastic and deterministic dynamics if they share specific statistical moments, thereby establishing an intrinsic limit on inferring the full distribution from aggregate epidemic data.

Original authors: Roni Muslim

Published 2026-08-06

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Roni Muslim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out how a rumor spreads through a school. You know the rumor travels when people hang out in groups, but you can only see the total number of students who have heard the rumor over time. You don't know the specific sizes of the groups they were in, or how many groups of each size exist. This is the puzzle of "epidemic inference": trying to work backward from the big picture to understand the tiny, hidden details of how people interact. Scientists have long known that the size of these groups matters. If a disease spreads faster when three people are together than when two are, the "group-size distribution" (the mix of small and large groups) changes how the outbreak behaves. But a nagging question remained: If we only see the final numbers, can we ever truly know the exact mix of group sizes that caused them? Or could two completely different mixes of groups create the exact same outbreak story?

This paper, written by Roni Muslim, dives deep into that mystery using math models for diseases that spread and then either fade away or stay around (called SIS and SIR models). The author treats the spread of disease like a game where people randomly form temporary groups. The key discovery is a surprising "blind spot" in our ability to learn from data. The paper proves that if a disease spreads in a specific, non-linear way (where the risk jumps up when more infected people are in a group), the entire history of the outbreak depends on only a few specific "averages" of the group sizes, not the whole detailed list of sizes. This means two totally different groups of people—one with mostly small groups and a few huge ones, another with a mix of medium-sized groups—can generate the exact same epidemic behavior, down to the last fluctuation and the exact time the disease dies out. It's as if two different recipes, using different amounts of ingredients, could somehow bake the exact same cake, making it impossible for a food critic to tell which recipe was used just by tasting the cake.

The Great Group-Size Illusion

Imagine you are a detective trying to solve a crime, but your only clue is a blurry photo of a crowd. You know the crowd was made of groups of people, but you can't see the individuals. Now, imagine two different scenarios:

Scenario A: The crowd is made of exactly half groups of 3 people and half groups of 9 people.
Scenario B: The crowd is a chaotic mix of groups of 2, 5, and 12 people, with very specific numbers of each.

In the real world, these are two very different crowds. But in the mathematical world of this paper, if the "infection rule" is a specific type of non-linear math (specifically, a quadratic rule where the risk grows with the square of infected people), these two totally different crowds produce identical disease outbreaks.

The author calls this "structural nonidentifiability." It's not that we just don't have enough data or that our measurements are too noisy. It's that the data cannot tell the difference, even in a perfect, noise-free world. The paper shows that the disease dynamics only "see" a few specific numbers, called moments, from the group sizes. Think of these moments like the "flavor profile" of the groups.

The first moment is like the average size of the group you experience. This sets the basic speed limit for how fast the disease can start.
The second moment is like the "variability" or how much the group sizes bounce around. This controls whether the disease spreads smoothly or if it gets stuck in a weird state where it can either die out or explode, depending on how many people are sick to begin with.

The paper proves that as long as two different group-size distributions share these first few "flavor numbers," the disease behaves exactly the same. The rest of the details—the fact that one distribution has groups of size 9 and the other has groups of size 12—are completely invisible to the disease. It's like two different musical chords that sound identical to a specific instrument; the instrument just can't hear the difference.

The "First Unmatched Moment" Rule

So, how do we ever tell these two different groups apart? The paper reveals a clever rule: you have to change the game.

The author shows that the disease only "looks" at the first few moments based on how complex the infection rule is. If the infection rule is simple (linear), it only looks at the first moment. If it's quadratic (the one used in the main examples), it looks at the first two. If you want to see the third moment (the difference between the group of 9 and the group of 12), you have to upgrade the infection rule to be even more complex (cubic).

The paper demonstrates this with a "controlled experiment."

Step 1: They take the two different distributions (the 3/9 mix and the 2/5/12 mix). With a quadratic infection rule, the disease spreads exactly the same way for both. The curves on the graph overlap perfectly.
Step 2: They tweak the infection rule slightly to make it sensitive to the third moment. Suddenly, the two curves split apart! The disease now behaves differently for the two groups.

This proves that the "blindness" isn't a flaw in our observation tools; it's a fundamental limit of the transmission protocol. The paper explicitly rules out the idea that we could ever figure out the full group-size distribution from a single type of outbreak data. No matter how long we watch or how many people we count, if the infection rule stays the same, we can never distinguish between distributions that share the same first few moments.

Why This Matters (and Why It's Not a Dead End)

The paper doesn't just say "we can't know." It tells us exactly what we can know and what we need to change to know more.

The author shows that the first moment controls the "invasion threshold"—the point where the disease can start to spread. The second moment controls something called "bistability" and "hysteresis." In plain English, this means that for some diseases, having a few extra people sick at the start can push the system into a permanent "endemic" state where the disease never goes away, even if the average conditions suggest it should die out. The paper shows that the second moment is the switch that flips this behavior on or off.

The paper also clarifies that this isn't just a theoretical curiosity. They ran massive computer simulations (using something called the "master equation" and "Gillespie algorithms") to prove that the math holds up in finite populations, not just in the idealized limit of infinite people. They even checked the "survival probability" (how long the disease lasts before dying out) and the "final outbreak size" (how many people get sick in the end). In every single case, the two different distributions produced identical results until the infection rule was changed to look at a higher moment.

The takeaway is a precise limit on what we can learn. If you see an outbreak, you can figure out the "average group experience" and the "variability of group experience" that the disease cares about. But you cannot reconstruct the full list of group sizes. To do that, you would need to observe the disease spreading under different rules—perhaps a different virus that spreads differently, or a different way of measuring the groups.

In the end, the paper paints a vivid picture of a world where the details of our social lives are partially hidden from the diseases that travel through them. Two very different social structures can look exactly the same to a virus, and unless we change the rules of the game, we will never be able to tell them apart just by watching the numbers go up and down. It's a reminder that sometimes, the most important information is the stuff that the data simply cannot see.

Technical Summary: Exact Moment Equivalence and Structural Nonidentifiability in Nonlinear Epidemics

Problem Statement
Epidemic transmission frequently occurs within temporary groups (e.g., households, classrooms, public transport) where group sizes vary. While aggregate epidemic data (such as prevalence or outbreak size) are often used to infer transmission mechanisms, it remains unresolved which specific features of the underlying group-size distribution can be uniquely recovered from such data. Previous network-based approaches often reduce interactions to pairs or rely on mean-field approximations, potentially discarding information about higher-order interactions. A critical question arises: Can distinct group-size distributions generate identical stochastic epidemic dynamics, even in finite populations? This paper addresses the structural identifiability of the group-size distribution $Q(n)$ in nonlinear Susceptible-Infected-Susceptible (SIS) and Susceptible-Infected-Recovered (SIR) models.

Methodology
The author analyzes a well-mixed population of fixed size $N$ where transmission occurs through temporary groups that are independently reassembled at every infection opportunity (annealed mixing).

Model Formulation: The dynamics are defined as a continuous-time Markov process. A focal susceptible individual is exposed to a group of size $n$ drawn from a participant-view distribution $Q(n)$ . The remaining $n-1$ members are sampled uniformly without replacement from the rest of the population.
Transmission Kernel: The infection rate is governed by a polynomial kernel of order $q$ : $\lambda_{n,\ell} = \beta \ell^q / (n-1)^\eta$ , where $\ell$ is the number of infected partners, $\beta$ is the transmission scale, and $\eta$ controls group-size normalization.
Exact Averaging: Unlike mean-field closures, the author performs exact averaging over group compositions using the hypergeometric distribution. This preserves finite-population fluctuations without approximation.
Moment Reduction: By expanding the power $\ell^q$ into falling factorials using Stirling numbers of the second kind, the author derives the aggregate infection hazard. They prove that the Markov generator depends on the full distribution $Q(n)$ only through a finite set of generalized factorial moments, $M^{(\eta)}_k$ , for $k=1, \dots, q$ .
Validation: Analytical predictions are verified by solving the finite-state master equation and performing stochastic simulations (Gillespie algorithm) for both SIS and SIR models.

Key Contributions and Results

The Moment-Sufficiency Theorem:
The central theoretical result (Theorem 1) establishes that for a transmission kernel of order $q$ , two distinct group-size distributions $Q_a(n)$ and $Q_b(n)$ generate identical aggregate Markov generators if and only if their first $q$ generalized factorial moments coincide ( $M^{(\eta)}_{k,a} = M^{(\eta)}_{k,b}$ for all $k \le q$ ).
- This equivalence is structural and exact for finite $N$ .
- It implies that the entire stochastic process—including transient evolution, fluctuations, extinction-time distributions, and final outbreak sizes—is indistinguishable for any distributions within the same equivalence class, regardless of their support or shape.
Structural Nonidentifiability:
The paper demonstrates that a single aggregate dynamical protocol cannot uniquely reconstruct the full group-size distribution $Q(n)$ . Instead, it identifies only an equivalence class of distributions.
- Example: The author constructs two distributions, $Q_A$ (support on sizes 3 and 9) and $Q_B$ (support on sizes 2, 5, and 12), which have disjoint supports but identical first and second factorial moments. For a quadratic kernel ( $q=2$ ), these distributions produce numerically identical mean prevalence, survival probabilities, and final outbreak size distributions.
- The difference between them only becomes observable when the transmission mechanism activates the third moment (e.g., by increasing the kernel order to $q=3$ ).
Deterministic Phase Structure and Moment Roles:
In the deterministic limit ( $N \to \infty$ ), the hierarchy of moments dictates the epidemic phase structure:
- First Moment ( $M_1$ ): Determines the basic reproduction number $R_0$ and the linear invasion threshold.
- Second Moment ( $M_2$ ): Controls the nature of the bifurcation. Specifically, the ratio $\chi = M_2/M_1$ determines whether the transition from disease-free to endemic is continuous, tricritical, or discontinuous (saddle-node bifurcation).
- Bistability and Hysteresis: For $q \ge 2$ , sufficiently large values of $\chi$ can induce bistability, where both disease-free and endemic states are stable. The width of the hysteresis region is controlled by the second moment, while the first moment sets the linear stability boundary.
Breaking Equivalence:
The paper shows that structural nonidentifiability can be resolved by modifying the transmission protocol. Introducing a higher-order channel (e.g., adding a term proportional to $\ell^{q+1}$ ) activates the next unmatched moment. Even a weak additional channel can shift the saddle-node threshold, placing two previously equivalent systems into different dynamical regimes (e.g., one remaining disease-free while the other enters a bistable or endemic state).

Significance and Claims
The paper claims to establish an intrinsic limit on epidemic inference from aggregate data. It argues that the inability to distinguish between moment-matched distributions is not a consequence of insufficient data, measurement noise, or limited temporal resolution (practical nonidentifiability), but a fundamental property of the model's structure (structural nonidentifiability).

Inference Limit: A single nonlinear transmission protocol identifies only an equivalence class of group-size distributions, not the distribution itself.
Protocol Dependence: The equivalence classes are specific to the microscopic protocol (kernel order $q$ , normalization $\eta$ , and sampling rules). Changing the protocol (e.g., increasing $q$ ) can reveal information about higher-order moments that were previously hidden.
Distinction from Practical Issues: The author distinguishes their findings from practical nonidentifiability (caused by noise) and observable nonidentifiability (caused by limited observation types). Their result holds even for noise-free, continuous observations of the full state distribution.

The work clarifies that while group-size distributions influence epidemic thresholds and bifurcation structures, aggregate data alone cannot uniquely recover the full distribution without additional constraints or varied transmission protocols.

The Great Group-Size Illusion

The "First Unmatched Moment" Rule

Why This Matters (and Why It's Not a Dead End)

More like this