Real data rarely sits on a straight line. Populations level off, sales grow and then saturate, and the period of a pendulum grows with the square root of its length. IB Mathematics: Applications and Interpretation HL gives you the tools to model all of these: the logistic model, logarithmic linearisation of data, and non-linear regression judged with R² and residuals. Functions get 42 suggested teaching hours at AI HL and statistics 52, and this topic draws on both.
These skills also matter beyond the exam: they are the backbone of many modelling explorations. If you want structured help, our IB Maths AI HL tutoring covers the whole modelling toolkit.
The logistic model
A logistic model has the form
f(x) = L / (1 + Ce^(−kx)), with L, C and k all positive.
It describes growth that starts almost exponentially and then slows as it approaches a limit. Its key features:
| Feature | Value | Meaning in context |
|---|---|---|
| Upper horizontal asymptote | y = L | The carrying capacity or maximum |
| Lower horizontal asymptote | y = 0 | The model never goes negative |
| Initial value | f(0) = L / (1 + C) | The starting amount |
| Steepest point | where f(x) = L/2, at x = (ln C) / k | Growth is fastest here |
The steepest point is also the point of inflection: before it the curve is concave up, after it concave down. That is the difference from an exponential model, which keeps getting steeper. If you have met exponential and logarithmic functions only at SL level, our exponentials and logarithms guide is a good refresher.
Worked example 1: users of a new app
The number of users N of a new local app, t weeks after launch, is modelled by N(t) = 2400 / (1 + 39e^(−0.35t)).
- Initial users: N(0) = 2400 / (1 + 39) = 60.
- Long-term limit: as t → ∞, e^(−0.35t) → 0, so N → 2400.
- When are there 1200 users? 1 + 39e^(−0.35t) = 2, so e^(−0.35t) = 1/39, giving t = (ln 39) / 0.35 = 10.5 weeks. This is also when growth is fastest.
- When are there 2000 users? 1 + 39e^(−0.35t) = 2400/2000 = 1.2, so e^(−0.35t) = 0.2/39, giving t = (ln 195) / 0.35 = 15.1 weeks.
The GDC's numerical derivative at t = 10.5 gives about 210 users per week, the fastest growth rate. (For this model the maximum rate equals kL/4 = 0.35 × 2400 ÷ 4 = 210.)
Many GDCs have a logistic regression option, so you may be asked to fit a logistic model to data with technology and then interpret L, C and k. Always check that the fitted L makes sense in the context.
Linearising data with logarithms
Taking logarithms can turn a curve into a straight line, which makes the relationship easy to spot and its parameters easy to find.
| Suspected model | Take logs | Plot | Straight line gives |
|---|---|---|---|
| Exponential: y = Ae^(kx) | ln y = ln A + kx | ln y against x (semi-log) | gradient k, intercept ln A |
| Power: y = ax^n | ln y = ln a + n ln x | ln y against ln x (log-log) | gradient n, intercept ln a |
The test is simple: whichever graph is straight tells you the type of model. A semi-log graph that is straight means exponential; a log-log graph that is straight means a power law.
Worked example 2: bacteria growth (semi-log)
A culture is measured every 2 hours (y in thousands of bacteria):
| x (hours) | 0 | 2 | 4 | 6 | 8 |
|---|---|---|---|---|---|
| y | 3.1 | 4.6 | 7.0 | 10.2 | 15.5 |
| ln y | 1.131 | 1.526 | 1.946 | 2.322 | 2.741 |
- Enter x and ln y in the GDC and run linear regression: ln y = 0.201x + 1.13, with r = 0.99989.
- The strong straight-line fit supports an exponential model.
- Undo the logs: y = e^1.13 × e^(0.201x), so y = 3.10e^(0.201x).
- Estimate at 10 hours: y = 3.10e^(2.01) = 23.1 thousand. This is extrapolation, so say it assumes growth continues at the same rate.
Worked example 3: a pendulum (log-log)
A student times a pendulum of length L metres; T is the period in seconds:
| L | 0.2 | 0.4 | 0.6 | 0.8 | 1.0 |
|---|---|---|---|---|---|
| T | 0.90 | 1.27 | 1.55 | 1.80 | 2.01 |
| ln L | −1.609 | −0.916 | −0.511 | −0.223 | 0 |
| ln T | −0.105 | 0.239 | 0.438 | 0.588 | 0.698 |
- Linear regression of ln T on ln L gives ln T = 0.499 ln L + 0.697, with r = 0.99997.
- So n = 0.499 and a = e^0.697 = 2.01.
- Model: T = 2.01L^0.499, very close to T = 2.01√L.
- For a 1.5 m pendulum: T = 2.01 × 1.5^0.499 = 2.46 s.
The power close to 0.5 is a nice result: physics predicts a square-root relationship, and the data agrees.
Non-linear regression, residuals and R²
With technology you can fit a curve directly (quadratic, cubic, exponential, power or sine regression) rather than linearising first. You then need to judge how good the fit is.
- A residual is observed value minus predicted value: y − ŷ.
- The sum of squared residuals, SS_res, measures total error. Smaller is better, for the same data.
- The coefficient of determination R² = 1 − SS_res / SS_tot, where SS_tot is the sum of squared differences between each y and the mean of y. R² close to 1 means the model explains most of the variation.
A high R² is not the whole story. Look at the pattern of residuals too: random signs suggest a good model; a curved pattern (+, −, −, −, + for example) means the model has the wrong shape.
Worked example 4: linear or exponential?
Use the bacteria data from Example 2. Compare the least-squares line with the exponential model y = 3.10e^(0.201x). The mean of y is 8.08, and SS_tot = 97.63.
| x | y | Line ŷ = 1.52x + 2.00 | Residual | Exponential ŷ | Residual |
|---|---|---|---|---|---|
| 0 | 3.1 | 2.00 | 1.10 | 3.10 | 0.00 |
| 2 | 4.6 | 5.04 | −0.44 | 4.63 | −0.03 |
| 4 | 7.0 | 8.08 | −1.08 | 6.93 | 0.07 |
| 6 | 10.2 | 11.12 | −0.92 | 10.35 | −0.15 |
| 8 | 15.5 | 14.16 | 1.34 | 15.48 | 0.02 |
- Line: SS_res = 5.21, so R² = 1 − 5.21/97.63 = 0.947.
- Exponential: SS_res = 0.0308, so R² = 0.9997.
The line's R² of 0.947 looks respectable on its own, but its residuals run +, −, −, −, +, a clear curve. The exponential model has a far smaller SS_res and no pattern, and growth that multiplies over time makes sense for bacteria. Choose the exponential model, and say why using all three pieces of evidence: R², residuals and context.
Choosing a model: a checklist
- Context first. Is there a natural maximum (logistic)? Constant percentage growth (exponential)? Repeating behaviour (sine)?
- Shape of the scatter graph. Does it level off, keep steepening, or curve gently?
- Linearise to test. Straight on semi-log means exponential; straight on log-log means power.
- Compare fit. Lower SS_res and higher R² for the same data, with no pattern in the residuals.
- Sense-check predictions, especially outside the data range.
Common mistakes
- Plotting y against ln x when testing for an exponential. The semi-log test is ln y against x.
- Forgetting to undo the logs. The intercept is ln A, not A.
- Comparing R² values that are not like for like. Check whether your calculator's value refers to the logged data or the original data before comparing models.
- Trusting R² alone and ignoring a residual pattern.
- Extrapolating without comment, especially for exponential models that grow without limit.
- Mixing up L and L/2 in a logistic question. L is the limit; L/2 is where growth is fastest.
- Rounding parameters early, then using them in a later prediction.
Exam technique: Papers 1, 2 and 3
All AI HL papers require a GDC. Papers 1 and 2 are 120 minutes and 110 marks each, and Paper 3 is 60 minutes and 55 marks with two extended questions, where a modelling task can run through several parts: fit a model, test it, improve it, interpret it.
- Write down the regression you ran ("linear regression of ln y on x") and the full equation with parameters to 3 significant figures.
- Keep unrounded values in the GDC for later parts.
- When asked to "justify" a model, refer to R², residuals and the context, not just one of them.
The SL foundations of this topic are in our AI SL modelling with functions guide. Logistic growth also arises from a differential equation, linking to our AI HL differential equations guide, and the statistical side continues in our post on hypothesis tests and confidence intervals in AI HL.
How Maths Bodhi helps
Our tutors teach modelling as a decision process: look at the context, test with logarithms, fit with technology, then judge the fit. Students practise on fresh data sets and learn to write the short justifications that earn interpretation marks. For the exploration, we teach the methods and how the IA criteria work; the student chooses the topic and writes their own IA. See how our IB maths home tutors support AI HL students.
Home tuition is available in Gurgaon (Gurugram). Students elsewhere in India and abroad learn with us online. You can book a free demo class.
Frequently asked questions
- What is the logistic model in IB Maths AI HL?
- It is f(x) = L / (1 + Ce^(−kx)) with L, C and k positive; it grows almost exponentially at first and then levels off towards the limit L.
- How do you know whether data is exponential or a power law?
- Plot ln y against x and ln y against ln x; if the semi-log graph is straight the relationship is exponential, and if the log-log graph is straight it is a power law.
- How do you find a and n from a log-log graph?
- Fit a straight line to ln y against ln x; the gradient is n and the intercept is ln a, so a = e^(intercept).
- What is a good R² value in IB Maths AI HL?
- There is no fixed cut-off; R² closer to 1 means a better fit, but you should also check the residuals for a pattern and whether the model makes sense in context.
- When is a logistic model growing fastest?
- When the value reaches half the limit, L/2, which happens at x = (ln C) / k; this is the point of inflection.
- Do I need to linearise data if my GDC can do exponential and power regression?
- Linearising is part of the AI HL course and helps you see which model fits; the GDC regressions are then used to fit and compare models quickly.
