Chapter 39 — Quiz

The capstone's real assessment is the portfolio project itself (see exercises.md). This quiz is different from the computational quizzes of earlier chapters: every question tests modeling judgment — which tool fits which situation, how to read a model's output, and what its assumptions cost. There is little arithmetic here and a great deal of "why." Answer each, then expand the solution.


1. A public-health team wants to know whether a new disease will cause an outbreak or fizzle out, given a transmission rate $\beta$ and a recovery rate $\gamma$. Which single quantity answers the question, and what is the threshold?

Answer The **basic reproduction number** $R_0 = \beta/\gamma$ (§39.3.3, anchor from Chapter 19). The disease takes off if $R_0 > 1$ and dies out if $R_0 < 1$, because at the start $\frac{dI}{dt}\approx(\beta-\gamma)I$, which is positive exactly when $\beta>\gamma$. $R_0$ is the average number of new infections one infectious person causes in a fully susceptible population.

2. A modeler reports that for an epidemic with $R_0 = 3$, the herd-immunity threshold is about 67%, and concludes "so about 67% of people will eventually be infected." What is wrong with that conclusion?

Answer The threshold $1 - 1/R_0$ is where the epidemic *peaks and begins to decline* — not where it ends (§39.3.5, Common Pitfall). Because of **overshoot**, the *final attack rate* is larger; it is given by the final-size equation $s_\infty = e^{-R_0(1-s_\infty)}$, which for $R_0=3$ yields an attack rate near 94%. Infections continue after the turnaround because a large pool of infectious people is still working through the remaining susceptibles. Confusing the turnaround point with the end point is the most common SIR interpretation error.

3. You must choose a calculus tool for each of these tasks. Match the tool to the job: (a) find how a system evolves second-by-second; (b) find the best input mix under a budget; (c) recover a total from a rate; (d) find the direction to adjust parameters to reduce error.

Answer (a) a **differential equation** (Chapter 19) — dynamics over time; (b) **constrained optimization / Lagrange multipliers** (Chapter 31) — the best choice under a constraint; (c) the **integral** (Chapters 13–14, the FTC) — accumulation; (d) the **gradient** (Chapter 30) — the negative gradient is the steepest-descent direction. This is the §39.7 synthesis table read as a quiz: one tool per job.

4. An SIR simulation reports that $S + I + R$ slowly drifts from 1000 to 1006 over the run. Is this a real epidemiological effect or a problem? What should you do?

Answer It is a **bug or numerical-tolerance problem**, not an effect (§39.3.5, validation). The three SIR rates sum to zero, so $N = S+I+R$ is conserved exactly; the model *cannot* create population. A drift means the solver tolerances are too loose (tighten `rtol`/`atol`) or the equations were typed wrong (a sign error, a missing term). Conservation is the first validation check precisely because the model must obey it.

5. The SIR system has "no elementary closed-form solution," yet we still solve it. How — and what earlier idea is the numerical solver descended from?

Answer We integrate it **numerically** with a solver like `scipy.integrate.solve_ivp` (§39.3.4). Under the hood it is an adaptive Runge–Kutta method, a sophisticated descendant of the **Euler stepping** introduced in Chapter 19, which is itself rooted in the **Riemann-sum** idea from Chapter 13 (build a total from small increments). "Approximation is the soul of calculus" (Theme 6): close enough, made rigorous.

6. In the Cobb–Douglas model $Q = AL^aK^b$, you derive that the firm spends the fraction $\frac{a}{a+b}$ of its budget on labor. Why is this a good feature of the model, not just a tidy formula?

Answer Because it is **testable** (§39.4.2). The exponents $a, b$ are abstract elasticities, but the budget shares are observable in real expenditure data. A model that turns an unobservable parameter into a checkable prediction can be confronted with reality — which is what separates a model from a story. It also makes the comparative static $\partial L^*/\partial w < 0$ (higher wages reduce labor demand, Chapter 6) directly interpretable.

7. A data scientist trains a line by gradient descent and sees the loss values go 2.1, 5.8, 41, nan. Diagnose the problem and name the fix.

Answer The **learning rate $\eta$ is too large** (§39.6.2, Common Pitfall). Steps overshoot the minimum, the loss grows, and the parameters explode to `nan`. The fixes: reduce $\eta$ until the loss decreases monotonically, and **scale the features** so all partial derivatives have comparable magnitude (which widens the range of usable $\eta$). $\eta$ is the first hyperparameter you tune; there is no universally correct value.

8. A model fits its training data with zero error but predicts new data terribly. Name the failure, and name the calculus-flavored remedy and the validation technique.

Answer **Overfitting** (§39.8). A degree-$(n-1)$ polynomial can pass through any $n$ points exactly, but wiggles wildly between them and memorizes noise. The remedy is Occam's razor made quantitative — prefer the simplest model that fits — and the validation technique is **cross-validation**: hold out data the model never saw and measure error there. Training error measures memorization; held-out error measures understanding.

9. The Hohmann-transfer model predicts a total $\Delta v$ of about 3.89 km/s for a LEO-to-GEO trip. A mission planner uses this number but adds small corrections. Which assumptions of the model force those corrections, and which calculus tool supplies the transfer time?

Answer The model assumes **instantaneous burns** and a **two-body universe** (§39.5, §39.8) — real burns take finite time, and Earth's oblateness ($J_2$) and other bodies perturb the orbit, so the planner adds corrections. The transfer time comes from **Kepler's third law**, $t = \pi\sqrt{a_t^3/\mu}$ (half the transfer ellipse's period) — a conic-section fact from Chapter 27, giving about 5.27 hours.

10. Across the four tracks, the "integral" appears as the SIR final-size relation, as consumer surplus, as orbital work/energy, and as accumulated loss. What single sentence (the thesis of the synthesis in §39.7) does this illustrate, and why does it matter for your portfolio?

Answer That **calculus appears in every quantitative field, and it is the same calculus** (Theme 5, §39.7): the integral means *accumulation* whether of immune individuals, of consumer welfare, of energy, or of error. It matters because your portfolio's job is to demonstrate exactly this fluency — that you can pick up one toolkit (derivative, integral, ODE, optimization, series, gradient) and deploy it on a problem nobody handed you pre-digested, tracing every tool back to the chapter that forged it.

Scoring Guide

Score Interpretation
9–10 Excellent. You think like a modeler — you choose tools by judgment, read outputs critically, and respect assumptions. You are ready to build the portfolio.
7–8 Strong. Revisit any missed item's referenced section, especially the overshoot (Q2) and overfitting (Q8) pitfalls.
5–6 Developing. Re-read §39.2 (the modeling cycle) and §39.7–39.8 (synthesis and validation) before starting your portfolio.
0–4 Re-read the full chapter. These questions are about judgment, not arithmetic; the chapter's worked models (§39.3–39.6) demonstrate every judgment tested here.

The quiz checks whether you can reason about models. The portfolio (exercises.md) checks whether you can build one. Do both.