Bounds on $R_0$ and final epidemic size when the next-generation matrix $M$ is only partially known

This paper derives sharp upper and lower bounds for the basic reproduction number ($R_0$) and final epidemic sizes in multitype SIR models where the next-generation matrix is only partially known through row or column sums, providing complete results for general matrices and partial results for matrices satisfying detailed balance.

The Gist

Imagine a massive, complex party where different groups of people are mingling. Some are loud and energetic (Type A), some are quiet and reserved (Type B), and some are in between. A virus is spreading through this party.

Usually, epidemiologists try to predict how big the outbreak will get by knowing exactly who talks to whom. They have a giant map (called a "Next-Generation Matrix") showing every single interaction between every group.

But here's the problem: In the real world, we rarely have that perfect map. We might only know:

• The "Row Sums": How many people each group talks to in total (e.g., "The loud group talks to 10 people a day on average").
• The "Column Sums": How many people receive attention from each group (e.g., "The quiet group is talked to by 5 people a day on average").

We don't know the specific details of who is talking to whom within those totals. It's like knowing how many apples each person in a room ate, but not knowing which specific apples they ate or who gave them to them.

This paper asks: "If we only have these partial numbers, how bad can the party get? And how good can it be?"

Here is the breakdown of their findings using simple analogies:

1. The "Basic Reproduction Number" ($R_0$)

Think of $R_0$ as the "Contagion Score."

• If the score is below 1, the virus fizzles out (like a spark hitting a wet log).
• If the score is above 1, the virus takes off (like a spark hitting a pile of dry leaves).

The Finding:
When we only know the total number of contacts (the row/column sums), we can't pinpoint the exact Contagion Score. Instead, we get a range.

• The "Worst-Case" (Upper Bound): Imagine the virus is super-efficient. Every time a person from a "high-contact" group talks, they talk to someone who is very likely to get infected. This gives us the highest possible score.
• The "Best-Case" (Lower Bound): Imagine the virus is clumsy. High-contact people mostly talk to other high-contact people who are already immune, or they talk to people who are hard to infect. This gives us the lowest possible score.

The Catch:
If the groups are very different (some talk to 100 people, some to 1), the gap between the "Best" and "Worst" case is huge. We are flying blind. However, if everyone has roughly the same number of contacts, the gap shrinks, and our prediction becomes much sharper.

2. The "Final Epidemic Size"

This is the "Party Damage Report." It asks: What percentage of the total crowd will eventually get sick?

The General Case (No Rules):
If we assume the virus can spread in any chaotic way (General Matrix), the uncertainty is massive.

• The Lower Bound: It could be zero! If the virus gets stuck in a small corner of the party and dies out, no one else gets sick.
• The Upper Bound: It could be nearly everyone. If the virus finds the perfect path to infect the most vulnerable people first, it sweeps the room.

The "Detailed Balance" Case (The Realistic Rule):
In real life, social interactions are usually reciprocal. If Alice talks to Bob, Bob is also talking to Alice. This is called "Detailed Balance."

• The Analogy: Imagine a dance floor. If Group A dances with Group B, Group B is also dancing with Group A. You can't have Group A dancing with Group B without Group B dancing with Group A.
• The Result: When we apply this "reciprocity rule," the chaos reduces. The range of possible outcomes gets narrower. We can make better predictions.
• The Surprise: The authors found a weird quirk in the math for small groups. Sometimes, if the "quiet" group starts talking to more people, the total number of infections might actually go down. It sounds counter-intuitive, but it's because spreading the virus out dilutes its power, preventing it from exploding in one specific corner.

3. Why Does This Matter?

Think of this paper as a Safety Net for Policymakers.

Imagine you are the mayor of a city. You know that "School Kids" talk to 20 people a day and "Office Workers" talk to 5. But you don't know if the kids are mostly talking to other kids or to office workers.

• Without this paper: You might panic and shut down the whole city because you fear the worst-case scenario.
• With this paper: You can say, "Okay, even in the worst-case scenario where the kids infect the office workers perfectly, the total infections will be between X and Y."

This allows leaders to make informed guesses rather than blind guesses. It tells them: "We need more data to narrow this gap, but here is the absolute worst that could happen, so let's prepare for that."

Summary

• The Problem: We often don't know exactly who infects whom, only the totals.
• The Solution: The authors created mathematical "fences" (bounds) that trap the possible outcomes.
• The Insight: If interactions are reciprocal (people talk to each other), the fences are tighter, and our predictions are more reliable. If interactions are chaotic, the fences are wide, and we must be very cautious.

In short, this paper gives us a way to draw a map of the unknown, helping us navigate the fog of an epidemic with a little more confidence.

Technical Summary

Here is a detailed technical summary of the paper "Bounds on R0 and final epidemic size when the next-generation matrix M is only partially known" by Bizzotto, Ball, and Britton.

1. Problem Statement

In multitype SIR epidemic models, the dynamics are governed by the Next-Generation Matrix (NGM), denoted as $M = \{m_{ij}\}$, where $m_{ij}$ represents the expected number of infectious contacts an individual of type $i$ makes with individuals of type $j$ during their infectious period. Two critical quantities depend on $M$:

1. The Basic Reproduction Number ($R_0$): The spectral radius (largest eigenvalue) of $M$.
2. The Final Epidemic Size ($\tau$): The vector of final infection fractions for each type, determined by a system of non-linear equations.

The Challenge: In many real-world scenarios (e.g., Social Contact Studies), the full matrix $M$ is not observable. Researchers often only have access to:

• Row sums ($r_i$): The total expected infectious contacts generated by type $i$.
• Column sums ($c_j$): The total expected infectious contacts received by type $j$.

The paper addresses the inverse problem: Given only the row sums or column sums of $M$, what are the sharp upper and lower bounds for $R_0$ and the final epidemic sizes ($\tau_i$ and total $\bar{\tau}$)?

The authors analyze two distinct scenarios:

1. General $M$: No structural constraints on the mixing patterns (non-negative entries only).
2. Detailed Balance $M$: The matrix satisfies $\pi_i m_{ij} = \pi_j m_{ji}$, which holds when mixing is symmetric and types differ only in contact rates (common in contact studies).

2. Methodology

The authors employ a combination of linear algebra, spectral theory, and non-linear optimization to derive these bounds.

• Spectral Bounds for $R_0$:
  • For general $M$, they utilize standard linear algebra results regarding the spectral radius of non-negative matrices relative to their row and column sums.
  • For detailed balance, they symmetrize the matrix using the community fractions $\pi$ (transforming $M$ into a symmetric matrix $S = D_\pi^{1/2} M D_\pi^{-1/2}$) to apply Rayleigh-Ritz variational principles.
• Final Size Bounds:
  • The final size equations are reformulated using a stochastic matrix factorization. If column sums are known, $M = P D_c$; if row sums are known, $M = D_r Q$.
  • The problem is transformed into finding the extrema of the final size function over the set of admissible stochastic matrices ($P$ or $Q$) that satisfy the fixed sum constraints.
  • Optimization: They use Lagrange multipliers and convexity arguments to solve for the extremal mixing patterns (e.g., rank-one matrices) that maximize or minimize the final size.
• Case Analysis:
  • $k=2$ (Two types): They provide a complete analytical solution for the detailed balance case, analyzing the behavior of $R_0$ and $\bar{\tau}$ as a function of a single free parameter $\theta$ (representing the mixing between types).
  • $k>2$: They provide general bounds and a conjecture based on numerical evidence regarding the structure of the optimal matrix.

3. Key Contributions and Results

A. Bounds for the Basic Reproduction Number ($R_0$)

• General $M$:
  • Sharp Bounds: $R_0$ is bounded by the minimum and maximum of the known sums.
    • If row sums $\{r_i\}$ are known: $r_{\min} \le R_0 \le r_{\max}$.
    • If column sums $\{c_j\}$ are known: $c_{\min} \le R_0 \le c_{\max}$.
  • These bounds are sharp and attained by rank-one matrices.
• Detailed Balance $M$:
  • The bounds are tighter than the general case but generally not sharp for $k > 2$.
  • Lower Bound: Derived via the Rayleigh quotient, involving weighted sums of squares of the row/column sums (e.g., $\bar{r} = \sqrt{\sum \pi_i r_i^2}$).
  • Upper Bound: Remains $r_{\max}$ (or $c_{\max}$), which is sharp.
  • Surprising Finding: For $k=2$, the lower bound for $R_0$ can decrease as the second row sum ($r_2$) increases (for small $r_2$). This occurs because increasing $r_2$ can shift contacts away from a highly infectious type to a less infectious one, effectively reducing the overall transmission potential.

B. Bounds for Final Epidemic Size ($\tau$ and $\bar{\tau}$)

• General $M$ with Known Column Sums ($c_j$):
  • Sharp bounds are derived for each type $\tau_i$ and the total size $\bar{\tau}$.
  • The bounds depend on $y^* = \min_j \{ \pi_j t_{c_j} \}$ and $y^* = \max_j \{ \pi_j t_{c_j} \}$, where $t_\alpha$ is the solution to $1-t = e^{-\alpha t}$.
  • Interpretation: The lower bound corresponds to a scenario where all infectivity comes from the type with the lowest per-capita exposure; the upper bound corresponds to the type with the highest.
• General $M$ with Known Row Sums ($r_i$):
  • Component-wise: Bounds for individual $\tau_i$ are sharp component-wise but not simultaneously sharp (a single matrix cannot achieve the upper bound for all types simultaneously).
  • Total Size ($\bar{\tau}$): A sharp upper bound is derived using a rank-one mixing matrix (separable mixing). The lower bound is simply $\min_i \{ \pi_i t_{r_i} \}$.
• Detailed Balance $M$:
  • Case $k=2$: A complete analysis is provided. The final size $\bar{\tau}$ is a function of a mixing parameter $\theta$. The paper identifies regions in the $(r_1, r_2)$ plane where $\bar{\tau}$ is monotonic or has an internal maximum.
  • Case $k>2$: The problem is analytically intractable. The authors propose a conjecture: The maximum final size is attained when the matrix $M$ is supported on a subset of types $A$ where the "force of infection" is balanced, and zero elsewhere.

4. Numerical Illustrations

The authors validate their theoretical bounds using:

1. Synthetic 2-type example: Demonstrating how the bounds tighten as $r_2$ varies and highlighting the non-monotonic behavior of the lower bound under detailed balance.
2. Belgian Social Contact Study: Re-analyzing data where age and social activity are factors. They show that knowing only row sums (average contacts per group) leads to a wide range of possible $R_0$ and final sizes. Incorporating the detailed balance assumption significantly narrows these bounds, and adding partial information about specific contact pairs narrows them further.

5. Significance and Implications

• Robustness in Data Scarcity: The paper provides a rigorous framework for epidemic forecasting when contact matrices are incomplete, a common issue in public health surveillance.
• Impact of Assumptions: It quantifies the error introduced by assuming "general mixing" versus "symmetric mixing" (detailed balance). The detailed balance assumption, while restrictive, yields significantly tighter and more realistic bounds.
• Policy Relevance: The results suggest that knowing the distribution of contacts (row/column sums) is insufficient for precise prediction without understanding the structure of mixing (assortativity). However, even with partial data, sharp bounds can be established to define the "worst-case" and "best-case" scenarios for epidemic outcomes.
• Theoretical Advancement: The derivation of sharp bounds for the final size under row-sum constraints and the analysis of the $k=2$ detailed balance case represent significant mathematical contributions to the theory of multitype epidemics.

In conclusion, this work bridges the gap between theoretical epidemic modeling and practical data limitations, offering actionable bounds for $R_0$ and final epidemic size when only aggregate contact data is available.