Critiquing & Refining Models — I can critique the reasonableness of a result and refine a model or method when assumptions fail. - Math 3

SECTION 1

Historical Context & Motivation

Mathematics has always been about more than just getting an answer — it is about getting a reasonable answer. Throughout history, scientists and mathematicians have built mathematical models — simplified representations of real-world situations — and then had to go back and fix them when the predictions didn't match reality. A model is only as good as the assumptions behind it, and questioning those assumptions is one of the most powerful skills in mathematics.

~240 BC • Eratosthenes Measures Earth

Eratosthenes assumed Earth was a perfect sphere and used shadow angles to estimate its circumference. His result was remarkably close, but later scientists refined the model by accounting for Earth's slight oblateness — it bulges at the equator.

1687 • Newton's Gravitational Model

Newton's law of universal gravitation modeled planetary orbits with extraordinary accuracy. However, it assumed instantaneous gravitational interaction and couldn't explain Mercury's orbital precession, revealing a limitation in its assumptions.

1915 • Einstein Refines the Model

Einstein's general relativity replaced Newton's model with one that accounted for curved spacetime. The refined model correctly predicted Mercury's orbit, demonstrating how critiquing and revising assumptions leads to better mathematics.

2020 • Pandemic Modeling

Early COVID-19 models assumed uniform mixing of populations and constant transmission rates. As data revealed the importance of age groups, regional behavior, and vaccination, modelers continually refined their assumptions to improve predictions.

The common thread across all these examples is a cycle: build a model, test its predictions, identify where it fails, question the assumptions, and refine. This lesson focuses on developing your ability to carry out that cycle yourself. How do you know when a mathematical result is unreasonable? What do you do when your model breaks down? These are the questions we will answer.

SECTION 2

Core Principles of Model Critique & Refinement

Before you can refine a model, you need a clear framework for evaluating one. Every mathematical model rests on a set of assumptions — simplifications that make the math manageable. The process of critique involves checking whether those assumptions hold, whether the results make sense in context, and whether the model can be improved without becoming unnecessarily complex.

Identify Assumptions

Every model simplifies reality. List what the model takes for granted — constant rates, linear relationships, independence of variables — so you know exactly what could fail.

Check Reasonableness

Does the answer make sense in context? Use estimation, unit analysis, boundary checks, and common sense. A model predicting a person runs 500 mph is clearly unreasonable.

Compare with Data

When real-world data is available, compare the model's predictions to observed values. Systematic errors suggest a flawed assumption, not just random noise.

Diagnose the Failure

Pinpoint which assumption breaks down. Does a linear model curve away from the data at high values? Does a constant-rate model ignore acceleration? Diagnosis guides the fix.

Refine & Re-evaluate

Adjust the model by relaxing the broken assumption — switch from linear to quadratic, add a decay factor, or incorporate a new variable. Then repeat the critique cycle.

✦

KEY TAKEAWAY

Think of a mathematical model like a map. A road map is great for driving, but terrible for hiking through a forest — not because the map is "wrong," but because its assumptions (roads exist, terrain is flat) don't match the new situation. Critiquing a model means asking: Is this the right kind of map for the terrain I'm actually on? Refining means drawing a better map.

SECTION 3

The Model Critique Cycle — Visual Explanation

The five-stage critique cycle: Build a model, Predict outcomes, Critique reasonableness, Diagnose broken assumptions, and Refine the model. Then repeat.

The diagram above shows the iterative nature of model critique. Notice that the cycle doesn't end — even a refined model should be tested again. In practice, you might go through this loop several times on a single problem. The key insight is that finding an unreasonable answer is not failure; it is valuable diagnostic information that tells you something about the situation your model missed.

SECTION 4

Mathematical Framework — Reasonableness Checks

How do you actually decide whether a result is reasonable? There are several mathematical strategies you can use. These aren't separate formulas to memorize — they're thinking tools that apply to any model, from a simple linear equation to a complex exponential growth function.

Strategy 1: Boundary & Extreme-Value Checks

Plug in extreme values — zero, very large numbers, or negative numbers — and see if the model behaves sensibly. For example, if a model for profit is P(x) = 50x − 200, where x is the number of items sold, then P(0) = −200, which makes sense (you lose money if you sell nothing). But if your model gave P(0) = 5000, that would be a red flag.

BOUNDARY CHECK

Does f(0), f(∞), or f(−x) produce a result that makes real-world sense?

f(0) checks the starting condition; f(∞) checks long-term behavior; f(−x) checks whether negative inputs are meaningful in context.

Strategy 2: Unit & Dimensional Analysis

Check whether the units of your answer match what's expected. If you're solving for time and your answer comes out in dollars, something went wrong. Even in pure math, you can track what each variable represents to make sure the output quantity is logical.

UNIT CHECK

Units of output = Units implied by the context of the problem

For example, if distance = rate × time, then miles = (miles/hour) × hours. If the units don't simplify correctly, revisit the model.

Strategy 3: Order-of-Magnitude Estimation

Before doing a detailed calculation, round the numbers and estimate the answer. If your estimate says the answer should be around 500 and your exact calculation gives 50,000, you likely made an error. This technique catches arithmetic mistakes and modeling errors alike.

ESTIMATION

Round inputs → compute rough answer → compare to exact result

If the exact result differs from the estimate by more than one order of magnitude (a factor of 10), investigate.

Strategy 4: Residual Analysis

When you have data, compute the residuals — the differences between each observed data point and the model's prediction. If the residuals show a pattern (for instance, they're all positive at high x-values), the model is systematically off, which signals a flawed assumption.

RESIDUAL

Residual = Observed value − Predicted value

Random scatter in residuals suggests a good fit. A curved pattern in residuals suggests the model type (e.g., linear) is wrong.

SECTION 5

When Models Fail — Common Assumption Breakdowns

Understanding the most common ways assumptions fail helps you diagnose problems faster. The diagram below shows a concrete example: a linear model is fit to data that actually follows a quadratic pattern. On the left, the linear fit looks reasonable over a small range, but as you zoom out, the model diverges dramatically from reality.

A linear model may approximate data well over a small range (left panel), but as the range increases, the actual quadratic trend diverges from the prediction. The growing residuals in the right panel signal that the linearity assumption has broken down.

Broken Assumption	What You See	How to Refine
Constant rate of change (linear model used for nonlinear data)	Residuals curve — positive at extremes, negative in the middle (or vice versa)	Switch to a quadratic, exponential, or logarithmic model
Unlimited growth (exponential model with no cap)	Predictions become absurdly large (e.g., more bacteria than atoms in the universe)	Use a logistic model with a carrying capacity
Independence of variables	Changing one input unexpectedly changes another; predictions are inconsistent	Introduce interaction terms or use a multivariable model
No external influences (closed-system assumption)	Model works in lab conditions but fails in real-world application	Add terms to account for friction, air resistance, competition, etc.

Common assumption breakdowns and their refinement strategies

SECTION 6

Worked Example — Critiquing a Population Model

A biologist models a rabbit population in a meadow using the exponential function P(t) = 20 × 1.5^t, where P is the population and t is time in years. After collecting data for five years, she notices the model's predictions increasingly overshoot the actual population. Let's walk through the critique and refinement process.

Critiquing and Refining an Exponential Population Model

1
Step 1 — State the Model and Its Assumptions

The model P(t) = 20 × 1.5^t assumes: (a) the population grows by 50% each year, (b) there is no limit on resources, (c) no predators or disease affect the population. These are the assumptions to scrutinize.
2
Step 2 — Generate Predictions and Compare to Data

Year 1: P(1) = 30, Observed = 29. Year 2: P(2) = 45, Observed = 41. Year 3: P(3) = 67.5, Observed = 52. Year 4: P(4) = 101.25, Observed = 58. Year 5: P(5) = 151.88, Observed = 61. The gap between predictions and observations widens significantly.

Residuals grow from −1 to +90.88, showing systematic over-prediction.
3
Step 3 — Diagnose the Broken Assumption

The actual population appears to level off near 60 rabbits. This suggests the meadow has a limited carrying capacity — it can only sustain a certain number of rabbits. The "unlimited growth" assumption fails because food, space, and other resources are finite.

Broken assumption identified: unlimited resources.
4
Step 4 — Refine the Model

Replace the exponential model with a logistic model: P(t) = K / (1 + Ae^−rt), where K = 62 (carrying capacity estimated from data), A = (K − P₀)/P₀ = (62 − 20)/20 = 2.1, and r is a growth rate. Using trial with the data, r ≈ 1.2 produces a much closer fit.

Refined model: P(t) = 62 / (1 + 2.1 × $e^{−1.2t}$ )
5
Step 5 — Re-evaluate the Refined Model

Testing: P(3) ≈ 53.4 vs. observed 52, P(5) ≈ 60.7 vs. observed 61. The logistic model's residuals are much smaller and show no systematic pattern. The boundary check also passes: as t → ∞, P(t) → 62, which is a reasonable long-term population for a meadow.

Refined model passes reasonableness checks ✓

SECTION 7

Strengths & Limitations of Different Model Types

Every model type has situations where it shines and situations where it struggles. When you critique a model, one of the first questions to ask is whether the type of model (linear, quadratic, exponential, logistic, etc.) is appropriate for the situation. The table below compares the most common model families you encounter in Math 3.

Model Type	Best For	Key Limitation
Linear (y = mx + b)	Constant rate of change; short-range approximations of curved data	Cannot capture acceleration, decay, or leveling-off behavior
Quadratic (y = ax² + bx + c)	Projectile motion, area problems, situations with a single maximum or minimum	Assumes symmetric rise and fall; doesn't model sustained growth or decay
Exponential (y = a × $b^x$ )	Unrestricted growth or decay — compound interest, radioactive decay, early-stage epidemics	Predicts infinite growth, which is unrealistic in resource-limited systems
Logistic (y = K / (1 + $Ae^{−rx}$ ))	Growth with a natural limit — populations, adoption of technology, spread of information	Requires estimating the carrying capacity K, which may be unknown or changing
Piecewise (different rules for different intervals)	Situations where the rules change — tax brackets, shipping rates, tiered pricing	Can be discontinuous at breakpoints; harder to analyze algebraically

Comparison of common model types in Math 3

✦

KEY TAKEAWAY

Choosing a model is like choosing a tool. A hammer is perfect for nails, but terrible for screws. When a model gives unreasonable results, sometimes the fix isn't to tweak the numbers — it's to switch to an entirely different type of model that matches the situation's behavior.

SECTION 8

Connection to Advanced Mathematical Reasoning

The skills you're building in this lesson — questioning assumptions, testing reasonableness, and iterating on models — are the foundation of advanced mathematical practice in college and beyond. In statistics, this process becomes hypothesis testing and regression analysis. In engineering, it's called design iteration. In data science, it drives model validation and cross-validation. The thinking is the same — only the tools get more sophisticated.

What You Learn Now	What It Becomes Later
Checking if an answer makes sense in context	Validating statistical models with test data, checking p-values and confidence intervals
Looking at residual patterns	Formal residual analysis, R² values, and goodness-of-fit tests in regression
Switching model type when assumptions fail	Model selection criteria (AIC, BIC) in statistics and machine learning
Testing boundary and extreme values	Asymptotic analysis, limit behavior, and edge-case testing in computer science

How model critique skills connect to advanced coursework

The key difference between a student who "just does the math" and one who truly thinks mathematically is this critical habit: always asking, Does this answer actually make sense, and if not, why not? Developing this instinct now will serve you across every quantitative discipline you encounter in college and your career.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A student uses a linear model to predict the temperature of a cup of coffee over time. The model gives T(t) = 90 − 5t, where T is temperature in °C and t is time in minutes. According to this model, what temperature will the coffee be at t = 20 minutes? Explain why this result is unreasonable and identify the assumption that fails.

Show answer

At t = 20, T(20) = 90 − 5(20) = 90 − 100 = −10°C. This is unreasonable because a cup of coffee sitting in a room cannot cool below room temperature (approximately 20–25°C), and it certainly cannot reach −10°C. The broken assumption is that the coffee cools at a constant rate of 5°C per minute indefinitely. In reality, the rate of cooling slows down as the coffee approaches room temperature (Newton's Law of Cooling), making an exponential decay model more appropriate.

Question 2: BASIC CALCULATION

A model predicts that a plant's height (in cm) after t weeks is H(t) = 3t + 5. The actual measured heights are: Week 1: 7.5, Week 2: 10, Week 3: 14, Week 4: 19, Week 5: 26. Calculate the residual for each week. Do the residuals suggest the model is reasonable?

Show answer

The predicted heights are: H(1) = 8, H(2) = 11, H(3) = 14, H(4) = 17, H(5) = 20. The residuals (observed − predicted) are: Week 1: 7.5 − 8 = −0.5, Week 2: 10 − 11 = −1, Week 3: 14 − 14 = 0, Week 4: 19 − 17 = +2, Week 5: 26 − 20 = +6. The residuals show a clear increasing pattern — they go from negative to increasingly positive. This systematic trend indicates the linear model underestimates growth at later weeks. The plant is likely growing at an increasing rate (perhaps exponentially), not a constant rate.

Question 3: INTERMEDIATE

A city's population is modeled by P(t) = 50000 ×

1.03^t

, where t is years after 2020. (a) What does this model predict for the year 2120 (t = 100)? (b) The city is on a small island with a maximum capacity of 200,000 people. Explain why the exponential model fails for long-term prediction. (c) Suggest a refined model and estimate a reasonable value for its parameters.

Show answer

(a) P(100) = 50000 × 1.03¹⁰⁰ ≈ 50000 × 19.22 ≈ 961,000. (b) The island can only support 200,000 people, so a prediction of 961,000 is physically impossible. The exponential model assumes unlimited resources and space, which fails for an island with finite capacity. (c) A logistic model P(t) = K / (1 +

Ae^{−rt}

) would be more appropriate. Using K = 200,000 (carrying capacity), P₀ = 50,000, we get A = (200000 − 50000)/50000 = 3. Keeping the initial growth rate consistent with 3% annually gives r ≈ 0.03. So the refined model is P(t) = 200000 / (1 +

3e^{−0.03t}

), which levels off near 200,000.

Question 4: APPLIED

A business models its monthly revenue R (in thousands of dollars) as a function of advertising spending A (in thousands of dollars) using R(A) = 12A. After collecting data, they find: A = 1, R = 11; A = 2, R = 20; A = 5, R = 40; A = 10, R = 55; A = 20, R = 62. (a) Calculate the residuals. (b) What assumption in the linear model breaks down? (c) What type of model would better capture the diminishing returns evident in the data?

Show answer

(a) Predictions: R(1) = 12, R(2) = 24, R(5) = 60, R(10) = 120, R(20) = 240. Residuals: −1, −4, −20, −65, −178. The residuals are increasingly negative, meaning the model wildly over-predicts revenue at higher spending levels. (b) The linear model assumes each additional dollar of advertising generates the same revenue (constant returns). In reality, the data shows diminishing returns — the first dollars of advertising are very effective, but additional spending yields less and less additional revenue. (c) A logarithmic model like R(A) = a × ln(A) + b or a square root model like R(A) = a√A + b would capture diminishing returns. Fitting R(A) = 20ln(A) + 12 gives approximate values: R(1) ≈ 12, R(2) ≈ 25.9, R(5) ≈ 44.2, R(10) ≈ 58.1, R(20) ≈ 71.9, which is a much closer fit to the observed data.

Question 5: CRITICAL THINKING

Two students are modeling the height of a ball thrown upward from a 10-meter platform. Student A uses h(t) = −4.9t² + 15t + 10 (a quadratic model accounting for gravity). Student B uses h(t) = −4.9t² + 15t + 10 − 0.3t³ (adding a cubic term to account for air resistance). (a) At what time does each model predict the ball hits the ground? (b) For what values of t do the two models give nearly identical results? When do they diverge? (c) Under what conditions would Student B's refinement be justified, and when would it be unnecessary over-complication? Use specific numerical comparisons to support your argument.

Show answer

(a) For Student A: setting −4.9t² + 15t + 10 = 0 and solving with the quadratic formula gives t ≈ 3.62 seconds (taking the positive root). For Student B: solving −0.3t³ − 4.9t² + 15t + 10 = 0 numerically gives t ≈ 3.38 seconds. (b) At t = 0.5: A gives 16.275 m, B gives 16.2375 m — a difference of only 0.04 m. At t = 1: A gives 20.1 m, B gives 19.8 m — difference of 0.3 m. At t = 3: A gives 10.9 m, B gives 2.8 m — difference of 8.1 m. The models agree closely for small t (under about 1 second) and diverge significantly after t ≈ 2 seconds. (c) Student B's cubic term is justified for fast-moving projectiles, long flight times, or high-precision applications where air resistance measurably affects the trajectory (e.g., a lightweight ball in windy conditions). However, for a dense ball over a short trajectory, the quadratic model is sufficiently accurate and more practical. Adding the cubic term without physical justification risks overfitting — making the model more complex without actually improving its accuracy for the specific situation. The principle of parsimony suggests using the simpler model unless data clearly supports the need for the extra term.

SUMMARY

Lesson Summary

Summary

Every mathematical model is built on assumptions that simplify reality. Critiquing a model means using reasonableness checks — boundary values, unit analysis, estimation, and residual analysis — to determine whether the model's predictions match reality. When they don't, you diagnose which assumption failed and refine the model by relaxing that assumption — for example, switching from a linear to a quadratic model, or from exponential growth to a logistic model with a carrying capacity.

This process is not a one-time fix — it is an iterative cycle of building, predicting, critiquing, diagnosing, and refining. The goal is not a "perfect" model (no model is perfect) but a model that is accurate enough for the situation while remaining as simple as possible. Developing the habit of questioning your results and revising your approach is one of the most important skills in all of mathematics.