How Often Versus How Precisely: Budget Allocation Laws for Recovering Hidden Infections

Public health agencies count the cases that walk into clinics, but most infections never get counted. To estimate the true number of infections, agencies add two other data sources. One is wastewater testing, which measures viral RNA in sewage. The other is a small cohort of people who are tested on a regular schedule. Both cost money, and every program faces the same choice: sample often with low precision, or sample rarely with high precision. A sewage sample can be taken twice a week with one lab replicate or once a week with four. A cohort of 2,000 tests can be run as 20 rounds of 100 people or 4 rounds of 500. This proposal asks whether that choice follows a predictable law, so that a program could find the best schedule from a few pilot fits instead of by habit.

Video models face the same problem. A fixed context window must be split between the number of frames and the detail kept in each frame. Exploring the Design Space of Visual Context Representation in Video MLLMs fits the model's error as separate power laws in frames and in tokens per frame, then solves for the best split analytically. We carry this idea over to epidemic monitoring. The model that turns surveillance data into hidden infections is a compartmental model, meaning a set of differential equations that move people between Susceptible, Exposed, Infectious and Recovered groups. We fit it as a "universal differential equation", in which small neural networks inside the equations learn the transmission rate and the reporting rate as they change over time. We claim that this model's error in recovering hidden infections, and in recovering the time-varying reporting rate, also splits into a frequency term and a precision term. We further claim that the best split moves toward frequent, noisy sampling when transmission changes quickly. The hardest timescale should be the 2–8 week wave, where the reproduction number crosses 1.

Core Claim

Let $T$ be the number of sampling occasions over a season and $M$ the precision per occasion (lab replicates per wastewater sample, or participants per cohort round), under a budget $TM = B$. Let $E$ be the error in reconstructed latent incidence. We hypothesize

$$E(M,T) \approx C_M M^{-a} + C_T T^{-b} + E_0,$$

where $C_M, C_T, a, b$ are fitted constants and $E_0$ is error that no allocation can remove. Under the budget, the Lagrange condition gives the optimal split directly: $a\,C_M M^{-a} = b\,C_T T^{-b}$. At the optimum, the marginal gain from more precision equals the marginal gain from more frequency.

Video models already show that temporal density usually beats per-frame detail (LLaVA-Video, ScaleLong). The new content here is not that observation. It is three claims about the epidemic setting: that the law holds in an inverse problem where the model is only weakly identified, meaning several different parameter settings fit the observed data almost equally well; that its exponents can be estimated from about four pilot allocations; and that $b$ grows with how fast the reporting and transmission rates drift. The third claim is the falsifiable mechanism. If the reporting rate varies slowly, precision should win. If it shifts on the wave timescale, frequency should win.

Positioning

The closest work, and the one that supplies the method, is Wastewater-informed neural compartmental model for long-horizon case number projections. It embeds neural networks for transmission and reporting rates inside an exposed-stage compartmental model, links wastewater to infections through a viral shedding kernel, and finds that wastewater plus cases is only weakly identified until cohort tests (SentiSurv) are added. It fixes the sampling schedule and explicitly leaves open how sampling resolution affects identifiability. That open question is ours.

Accurately Estimating Unreported Infections using Information Theory (MDLInfer) chooses the reporting rate by minimizing description length, the total cost of encoding the model plus its errors, and it does so without cohort data. It assumes a constant or piecewise reporting rate. We treat it as the "do you need an anchor at all" competitor. Comparative evaluation of methodologies for estimating the effectiveness of non-pharmaceutical interventions motivates the mechanistic choice. In its simulations, integrated mechanistic fits had 0–5% bias and nominal coverage, while two-step methods (estimate the reproduction number, then regress it on policies) had up to 25% bias and often 0% coverage. The same paper also constrains us: the mechanistic fits cost 42 cluster-hours per 100 datasets. That cost is exactly why a law that needs only a few pilot fits is worth having.

The temporal-reasoning literature tells us what not to build. Foundations without Fundamentals shows that time-series foundation models pull exponential growth back toward the historical mean and largely ignore leading covariates. Chronos-2 gained 6.3% from a perfect leading indicator, while the 840-parameter linear model DLinear gained 62%. Wastewater is exactly that kind of leading indicator, about a week ahead according to Beyond clinical data. We therefore use foundation models as a comparison arm, not as the estimator. ScaleLong reports that video models do worst at intermediate timescales. That result motivates our expectation that errors concentrate at wave-scale transitions, which we test directly.

Research Program

We first build a simulator. It produces an exposed-stage compartmental epidemic with smoothly drifting transmission, an intervention, and susceptible depletion. It also produces a reporting rate $\alpha(t)$ with controllable drift speed, falling at the end toward the roughly 5% level reported in Germany. Wastewater is generated through a shedding kernel, with replicate noise and dilution events from rainfall. A cohort of tested participants follows a binomial sampling model. Cost is modeled as a fixed cost per occasion plus a per-replicate or per-test cost, so the budget is not purely multiplicative. The law is fitted under both cost models.

For each drift speed, we sweep a grid of $(T, M)$ allocations at three budget levels. We fit the universal differential equation to each simulated dataset with a fixed tuning recipe. We then measure three things: error in latent incidence, error in $\alpha(t)$, and error in the date the reproduction number crosses 1. Next we fit the power law, test whether it is separable against a model that includes a $T \times M$ interaction, and check whether exponents fitted from four pilot allocations predict the optimum found on the full grid.

The real-data stage uses the Bonn and Rhineland-Palatinate wastewater, case and SentiSurv data. Real data can only be made sparser, so we emulate lower-resolution schedules by dropping sampling days, keeping fewer replicates, and keeping fewer cohort participants. We then check whether the error ordering that the law predicts appears in 50-week out-of-sample projections. The US serology comparison from MDLInfer gives a second real check on ascertainment, the fraction of infections that are actually reported, though only for the cohort-free arm.

Baselines

The naive reference point is current practice: a uniform weekly schedule at a fixed replicate count and cohort size, matching the SentiSurv-style design, fitted with the same universal differential equation. Beating it means a predicted allocation with lower latent-incidence error at equal budget across most simulated drift regimes, not just on average.

The strongest competitor for choosing the allocation is classical optimal experimental design. It computes the Fisher information of the compartmental model's parameters, a local measure of how much each schedule would sharpen the estimates, around a pilot fit, and picks the schedule that maximizes it. Its compute budget is matched to ours: the same four pilot fits plus the information calculation. The file reports no result for this method in this setting (no reported result found on Lacuna for Fisher-information surveillance design). Our law beats it if its chosen allocation reaches lower realized error in regimes with fast-drifting $\alpha(t)$. That is where a local approximation should mislead, because the problem is weakly identified there.

For the estimation side, three references apply. The first is MDLInfer run without any cohort data, which reports 1.2× to 7.21× error reduction over curve-fitting baselines against CDC serology (MDLInfer). It sets the bar any allocation that includes cohort spending must clear. The second is a two-step reproduction-number pipeline, which in the simulation study showed up to 25% bias against 0–5% for mechanistic fits. The third is a pair of forecasters: Chronos-2 and DLinear, each given wastewater as a covariate, run on identical subsampled inputs. They can only be scored on case forecasts, since neither outputs hidden infections.

The key ablation removes the frequency–precision structure from the predictor. It replaces the fitted power law with a fixed "always maximize frequency" rule taken from the video findings. This isolates whether drift-dependent exponents add anything over the generic heuristic.

Why not just…?

Why not just use Fisher-information optimal design? This is the most serious threat, and it may succeed when the reporting rate varies slowly. Fisher information is local, and the neural rate functions make the parameter space large and poorly conditioned, which is exactly the setting the wastewater paper calls weakly identified. The simulation sweep settles the question directly. If Fisher design matches our realized error in every drift regime, the conclusion is that classical design suffices, and we will report it that way.

Why not just skip cohort testing and use MDLInfer? MDLInfer reached serology-level accuracy, but in 2020 data, with a reporting model that is constant or piecewise. Our sweep varies the drift speed of $\alpha(t)$ and includes a zero-cohort arm. That arm shows where description length alone stops being enough, and it provides the curve that budget spent on a cohort must beat.

Why not just forecast cases with DLinear or a foundation model fed wastewater? For forecasting cases, DLinear may genuinely be competitive, given its 62% covariate gain. But the question here is hidden infections and reporting rates, which these models do not represent, and foundation models flatten exponential phases (Foundations without Fundamentals). We report their case-forecast error alongside ours so the trade-off is visible.

Why not just always sample as often as possible, as video models suggest? The ablation tests exactly this. If the always-dense rule matches the fitted-law allocation everywhere, the drift-dependence claim fails.

Decisive Checks

The proposal stands or falls on four results. First, the separable law must fit the simulated error surface, with the interaction term adding little explained variance. Second, exponents from four pilot fits must pick an allocation within a small margin of the grid optimum. Third, the frequency exponent $b$ must increase with reporting-rate drift speed, and the gain from reallocating budget must concentrate in errors on the crossing date of the reproduction number. Fourth, on the Bonn and Rhineland-Palatinate data, the error ordering of emulated schedules must match the law's prediction. If the third check fails, the result reduces to the generic "denser is better" claim. That would be a modest finding, and we would say so.