When a Mathematical Model Fails — Limits, Assumptions and Better Models

Students sometimes imagine that a mathematical model is either correct or useless. Real modelling is more interesting. A model may work well over one range and badly over another. It may predict average behaviour but miss extremes. It may be suitable for estimation but not for safety-critical decisions.

Learning why models fail develops mathematical judgement. It teaches students to separate calculation from interpretation and to improve a representation rather than defend it automatically.

Failure can come from a broken assumption

Suppose a tank is modelled by V=20+5t, assuming constant inflow of 5 L/min. If the pump slows as pressure changes, the constant-rate assumption fails. The algebra remains correct for the equation, but the equation no longer represents the process well.

The repair is not “do the algebra more carefully”. The repair is to change the model.

Failure can come from the wrong relationship family

A quantity growing by a constant percentage is multiplicative. Modelling it with A=A₀+kn assumes constant absolute increase instead. The two models may look similar for one or two periods but diverge later.

When errors grow systematically, ask whether the relationship should be linear, quadratic, proportional, inverse or exponential-style rather than merely adjusting a coefficient.

Failure can come from using the model outside its domain

A linear cooling model T=80−4t may work as a classroom approximation for the first five minutes. Extending it to t=30 gives −40°C, which may be absurd in the physical setting.

The model did not necessarily fail where it was designed to work. The user extended it beyond its reasonable domain.

Interpolation and extrapolation

Interpolation estimates inside the range of observed or defined data. Extrapolation extends beyond it. Extrapolation is often riskier because the relationship may change outside the known region.

If a line fits data for x from 0 to 10, using it at x=7 is interpolation. Using it at x=100 is extrapolation. The same equation is being used, but the evidential support is different.

Failure can come from missing variables

A model of travel time using only distance may fail if speed varies greatly. A model of cost using only quantity may fail if there are fixed fees. A model of crop yield using only rainfall may miss soil, temperature and nutrient effects.

Adding a variable is justified when it explains a meaningful pattern in the errors, not simply because a more complicated equation looks impressive.

Failure can come from measurement error

Even a good structural model can disagree with observations because measurements are rounded or noisy. If a length recorded as 8.3 cm is actually between 8.25 and 8.35 cm under ordinary rounding, a small discrepancy may be measurement uncertainty rather than model failure.

This connects modelling with Estimation and Bounds in Measurement.

Failure can come from random variation

A model predicting an average does not imply every case will equal that average. A class may average 70 marks while individual students score above or below it. A probability model may predict a long-run proportion without guaranteeing the next trial.

Variation around a model is not automatically evidence that the model is worthless. The pattern and size of the variation matter.

Residual patterns can diagnose failure

If predicted values are sometimes slightly high and sometimes slightly low without a clear pattern, a simple model may still be adequate. If predictions are increasingly too low as x grows, the model may be missing upward curvature.

At Secondary level, students can learn this qualitatively before formal regression techniques.

Worked failure case 1: fixed charge omitted

Observed costs are $15, $20, $25 for 1,2,3 units. A student proposes C=5n. Predictions are 5,10,15: always $10 too low. The constant residual suggests a missing fixed charge. Better model: C=10+5n.

Worked failure case 2: changing rate

Observed outputs for x=1,2,3,4 are 2,8,18,32. A straight-line model fits poorly. The values follow 2x² exactly. The increasing first differences suggest the rate is changing rather than constant.

Worked failure case 3: domain limit

A tank of capacity 200 L is modelled V=50+10t. The formula predicts 250 L at t=20. If overflow is not part of the model, the valid domain stops when V=200, which occurs at t=15.

Worked failure case 4: real-world constraint

A quadratic calculation for a rectangular design gives x=8 and x=−12. Both may satisfy the algebraic equation, but a negative physical length is outside the model’s domain. The model result must be filtered through context.

Better does not always mean more complex

A model should be as simple as possible while meeting its purpose. If y=3x+2 predicts adequately, replacing it with a high-degree polynomial may create a fragile model that behaves badly outside the observed points.

Complexity should buy something: better accuracy, better explanation, or better decisions.

Improvement strategy 1: restrict the domain

Sometimes the model is acceptable only over a stated interval. Instead of replacing the equation, say explicitly that it is intended for 0≤t≤5 or for quantities between particular limits.

Improvement strategy 2: change the rate structure

If one constant gradient does not fit, consider piecewise rates or a curved relationship. A journey may be modelled with one speed before a stop and another speed afterwards.

Improvement strategy 3: add a missing parameter

If cost data are consistently offset, add a fixed charge parameter. If a line is too steep, revise the rate parameter. If percentage growth compounds, replace an additive parameter with a multiplier.

Improvement strategy 4: improve the data

More precise or more representative measurements can distinguish model failure from noisy evidence. But collecting more data is useful only if the data address the uncertainty that matters.

Improvement strategy 5: change the purpose

A model that is poor for exact prediction may still be good for rough comparison. State what the model is suitable for instead of asking it to do more than it was designed to do.

Model limits should be communicated

A strong conclusion might say: “Under the assumption of constant flow, the tank reaches 150 L after 10 minutes. This model should not be extended beyond the tank’s capacity or used if the flow rate changes.”

That statement is mathematically stronger than presenting a bare number because it identifies the conditions supporting the result.

Common misconceptions

“A failed prediction means the mathematics was wrong.” The calculation may be correct while the assumptions are poor.

“The best model has the most variables.” More variables can add noise and complexity.

“Extrapolation is just more interpolation.” Extending beyond known data is often much less secure.

“A model must predict every case exactly.” Some models aim at averages, ranges or useful approximations.

A model-repair routine

1. Locate where the model fails.
2. Check whether the failure is random or systematic.
3. Revisit assumptions and domain.
4. Check whether the relationship family is appropriate.
5. Identify missing variables or parameters.
6. Change only what the evidence justifies.
7. Retest the revised model.
8. State the remaining limitations.

Diagnostic checkpoints

• Can the student distinguish calculation error from modelling error?
• Can they identify a broken assumption?
• Can they explain why extrapolation is risky?
• Can they use residual patterns informally?
• Can they propose a simpler or better model without overcomplicating it?
• Can they communicate a model’s valid range?

Practice prompts

1. A line underpredicts every observed value by 7. What simple repair might be considered? Add a fixed offset if context supports it.

2. A model V=30+4t is used for a 100 L tank. Find the latest time before capacity under no-overflow assumptions. 30+4t=100 gives t=17.5.

3. A growth model gives increasingly large positive residuals. What might this suggest? The model may be growing too slowly or missing curvature.

4. A model fits x=0 to 10. Is x=100 equally trustworthy? No; that is distant extrapolation unless further evidence supports it.

The larger mathematical habit

A mature mathematical model is not defended forever. It is tested, bounded, revised and sometimes replaced. The goal is not loyalty to an equation; it is a useful relationship between mathematics and the question being asked.

Build from What Is a Mathematical Model?, Variables, Assumptions and Equations and How to Test a Mathematical Model. Return to the Secondary Mathematics Master Index.