Closes the chapter by fitting curves to scattered data. Uses linearisation — taking a logarithm of the outputs — to turn exponential data into linear data that Chapter 2's regression already handles, then converts the fitted line back into an exponential model. Distinguishes exponential, logarithmic and power models by which transformation straightens the plot, and carries every caution from §2.4 forward.
Subject: Precalculus · 65 slides · symbolic lesson
Open the interactive version of this deck
Title
Precalculus · Chapter 4 — Exponential and Logarithmic Functions
§4.8 Fitting Exponential Models to Data, pp. 590-608
Objectives
Five things, each one you can check yourself on paper.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 590-608 — the pages these objectives are drawn from
Warm-up
Section 2.4 fitted a line by minimising squared residuals. A curve is harder — unless you can avoid the problem.
Discussion prompt
You have data that clearly curves upward. You have machinery for fitting lines and none for fitting exponentials. What could you do?
Hint: Could you change the data so that a line becomes appropriate?
Answer:
You could transform the data so that the relationship becomes linear, fit a line to the transformed version, and then convert the answer back.
Taking a logarithm of the outputs does exactly that for exponential data. Multiplying by a fixed factor each step becomes adding a fixed amount each step, which is what a line does.
So a hard problem becomes an easy one you have already solved. This is a recurring strategy in mathematics — transform into a case you can handle — and it is the whole content of this section.
Concept
Taking a logarithm of the outputs converts an exponential relationship into a linear one. Fitting a line to the transformed data and converting back gives the exponential model.
\[ y=ab^x \;\Longrightarrow\; \log y=\log a+x\log b \]
The two logarithm properties from §4.5 do all the work: the product property splits the coefficient from the power, and the power property brings the exponent down. What is left is a linear equation in the original input with the transformed output.
Figure (svg): The same data plotted twice: on ordinary axes it curves upward, and after taking logarithms of the outputs it falls on a straight line
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 590-595
Section
Section 1
Concept
Exponential data multiplies by a fixed factor at each step. Taking a logarithm converts that into adding a fixed amount, which is exactly what makes a relationship linear.
This is exactly why a plot with a logarithmic vertical axis straightens exponential data, as §4.4's transfer noted. The axis does the transformation for you, so the same idea appears both as a computational technique and as a way of looking.
Figure (svg): An algebraic derivation showing that taking a logarithm of an exponential model produces a linear equation in the transformed variable
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 590-596
Picture it
Two properties from §4.5 do the whole job.
Figure (svg): An algebraic derivation showing that taking a logarithm of an exponential model produces a linear equation in the transformed variable
The final line is a linear equation in x, with the transformed output as the dependent variable. Everything Chapter 2 knows about lines now applies.
Worked example
Take logarithms of the outputs and look at the differences.
\[ \begin{array}{c|cccc} x & 0 & 1 & 2 & 3 \\ \hline y & 5 & 15 & 45 & 135 \end{array} \]
Check the raw data's ratios
Why: Each output over the previous.
\[ \text{all } 3 \]
Take logarithms of the outputs
Why: Base ten.
\[ 0.699, 1.176, 1.653, 2.130 \]
Compute the differences
Why: Consecutive transformed outputs.
\[ \text{all } 0.477 \]
Recognise the constant
Why: It is the logarithm of 3.
\[ \log 3 = 0.477 \]
Figure (svg): The same data plotted twice: on ordinary axes it curves upward, and after taking logarithms of the outputs it falls on a straight line
\[ \text{slope } =\log 3 \approx 0.477 \]
Verify: check the intercept too
Why: The transformed output at input zero is the logarithm of 5, about 0.699 — which is the logarithm of the coefficient, as the derivation predicted. So the line's slope encodes the base and its intercept encodes the coefficient, and both are recovered by exponentiating.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 591-593
Faded example
A regression on the logarithms of the outputs gives an intercept of 1 and a slope of 0.301, in base ten.
Fill in the blanks
a = 10^1 = 10, \qquad b = 10^2 = ___
Why: Exponentiating the intercept gives the coefficient and exponentiating the slope gives the base. The model is 10 times 2 to the x. Both constants must be exponentiated, and leaving either as a logarithm produces a model that is wrong by a large factor.
Worked example
Exponentiate the slope and the intercept.
\[ \text{A fit gives } \log y=0.477x+0.699. \text{ Recover the exponential model.} \]
Identify the intercept
Why: It is the logarithm of the coefficient.
\[ \log a = 0.699 \]
Exponentiate
Why: Ten to that power.
\[ a = 5 \]
Identify the slope
Why: It is the logarithm of the base.
\[ \log b = 0.477 \]
Exponentiate
Why: Ten to that power.
\[ b = 3 \]
Figure (svg): The solution to Worked example convert the line back shown as a ladder of expressions, one row per legal move
\[ y=5\cdot 3^x \]
Verify: check against the original data
Why: At x equal to 2 the model gives 5 times 9, which is 45 — matching the table. The conversion back is just exponentiating both constants, since the transformation took logarithms of them. Forgetting to exponentiate leaves a model whose constants are logarithms rather than the values themselves.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 593-596
Trap
\[ \text{fit gives } \log y=0.477x+0.699 \;\Longrightarrow\; y=0.477x+0.699 \]
Take the fitted line as the answer
Why: The regression produced a line, so the line is reported as the model.
The linear equation is given as the relationship between x and y.
The line describes the LOGARITHM of y, not y itself. The transformation has to be undone before the answer is about the original data.
Exponentiating gives y equal to 5 times 3 to the x, which is a curve. Testing at x equal to 2: the line gives about 1.65 and the true value is 45.
Convert back before reporting. The linearisation was a device for making the fitting possible, and leaving the answer in transformed coordinates answers a question nobody asked.
Prediction
Exponential data is transformed by taking logarithms of the outputs and a line is fitted.
Predict first
What does the line's slope represent?
Correct: The logarithm of the base.
Why: The derivation gives log y equal to log a plus x times log b, so the coefficient of x is the logarithm of the base. Exponentiating recovers the base itself. A larger slope means faster growth, but the slope is not the base until it has been exponentiated.
Sorting
The transformation converts multiplicative structure into additive.
Sort into buckets
Sort each feature of the raw data by what it becomes.
Socratic
The strategy is to transform a hard problem into an easy one.
Discussion prompt
What property of the logarithm makes it the right transformation for exponential data?
Hint: What does a logarithm do to a product?
Answer:
A logarithm turns multiplication into addition, by §4.5's product property. Exponential data is defined by repeated multiplication, so taking a logarithm converts it into repeated addition.
Repeated addition of a constant amount is exactly a constant rate of change, which §2.1 says produces a straight line. So the transformation is precisely matched to the structure it is straightening.
This also explains why the same transformation does not straighten a polynomial. Only relationships built from multiplication are converted into linear ones, which is why the choice of transformation is itself diagnostic of the model family.
Section
Section 2
Concept
The full procedure is three steps: take logarithms of the outputs, fit a line by least squares, and exponentiate the two constants to recover the exponential model.
The final check has to be against the original data, because that is what the model claims to describe. A fit that looks excellent in transformed coordinates can look worse in the original ones, since the transformation compresses large values and therefore shrinks their apparent residuals.
Figure (svg): The same data plotted twice: on ordinary axes it curves upward, and after taking logarithms of the outputs it falls on a straight line
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 596-601
Picture it
The same data, plotted raw and transformed.
Figure (svg): The same data plotted twice: on ordinary axes it curves upward, and after taking logarithms of the outputs it falls on a straight line
The right-hand plot is where the fitting happens, because a line can be fitted there. The left-hand plot is where the answer has to be judged, because that is the data the model is about.
Worked example
Three steps, with the conversion last.
\[ \text{Data gives a fitted line } \ln y=0.25x+2.30 \text{ on the transformed values.} \]
Identify what was fitted
Why: The natural logarithm of y against x.
\[ \ln y = 0.25 x + 2.30 \]
Exponentiate both sides
Why: Using base e, since ln was used.
\[ y = e ^{0.25 x + 2.30} \]
Split the exponent
Why: Using the exponent rule.
\[ y = e ^{2.30} e ^{0.25 x} \]
Evaluate the coefficient
Why: e to the 2.30.
\[ \text{about } 10 \]
Figure (svg): The solution to Worked example fit an exponential shown as a ladder of expressions, one row per legal move
\[ y\approx 10e^{0.25x} \]
Verify: check one point
Why: At x equal to 4 the model gives 10 times e to the power 1, which is about 27.2. And the fitted line gives a natural logarithm of 3.30, whose exponential is also about 27.1. The two agree, confirming the conversion. Note the natural logarithm was used throughout, so the base is e rather than ten.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 597-599
Ranking
Three steps and a check.
Put in order
Why: Transforming makes a line appropriate, fitting produces the line, exponentiating converts it back into the model's own terms, and checking is done on the original data because that is what the model describes. Checking on the transformed data instead would confirm only the middle step.
Worked example
The check belongs on the original data.
\[ \text{A transformed fit has } r=0.998. \text{ Is the exponential model good?} \]
Note what r describes
Why: The linearity of the TRANSFORMED data.
Translate that
Why: Strong linearity there means strong exponential fit.
Note the caveat
Why: The transformation compresses large values.
Check on the original data
Why: Compare predictions with actual values.
Figure (svg): A card restating the cautions from Section 2.4 as they apply to fitted exponential models, covering extrapolation, model choice, and the effect of the transformation on outliers
\[ \text{Strong evidence; check the residuals on the ORIGINAL scale too.} \]
Verify: explain the compression
Why: A logarithm turns a factor of ten into one unit, so a prediction off by a factor of 2 at a large value contributes only about 0.3 to the transformed residual. That makes the transformed fit look better than the original-scale fit at the large end, which is worth knowing before quoting a correlation as evidence of accuracy.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 599-601
Error analysis
A student converts a fitted line back into an exponential model.
Annotate
On: \( \log y=0.477x+0.699 \;\Longrightarrow\; y=0.699\cdot 3^x \)
Both constants are logarithms in the transformed equation, so both need exponentiating. Checking the model at one data point catches the omission immediately.
Faded example
A fit gives the natural logarithm of y equal to 0.4x plus 1.6.
Fill in the blanks
y = e^4.95e^0.4, \quad \text___ ___ \text___ ___e^___x}
Why: Exponentiating the whole right-hand side splits into the constant e to the 1.6, about 4.95, times e to the 0.4x. The slope stays in the exponent rather than being exponentiated separately, because it is multiplied by x — this is the form the natural logarithm produces, and it matches §4.7's continuous-rate models directly.
Prediction
An exponential model has been fitted by linearisation.
Predict first
On which data should its residuals be examined?
Correct: The original data, since that is what the model describes.
Why: The transformation compresses large values, so residuals at the high end look artificially small in transformed coordinates. Judging the fit where the model will actually be used means examining the residuals on the original scale, where a proportional error at a large value is visible at full size.
Explain it to yourself
A logarithm compresses large values more than small ones.
Discussion prompt
Explain why that makes a transformed fit look better than it is at the high end.
Hint: What does a factor-of-two error become after taking a logarithm?
Answer:
A logarithm turns ratios into differences. An error of a factor of 2 becomes a difference of about 0.3 in base ten, whatever the size of the values involved.
So a prediction off by a factor of two at a value of 10 and one off by a factor of two at a value of 10000 contribute the same transformed residual — even though the second is off by thousands in absolute terms.
Least squares on the transformed data therefore treats proportional errors as equally important everywhere, which is sometimes exactly what you want and sometimes hides large absolute errors at the top end. Knowing which you want is part of choosing the method, and checking on the original scale is what reveals the difference.
Section
Section 3
Concept
Different model families are straightened by different transformations, so which plot comes out straight is evidence about which model the data follows.
This makes the transformation a diagnostic as well as a technique. Plotting the data three ways and seeing which comes out straight is a genuine method of model identification, and it is far more informative than comparing correlation coefficients on untransformed fits.
Figure (svg): A table showing which variable to take logarithms of for each model family: outputs only for exponential, inputs only for logarithmic, and both for power models
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 601-606
Picture it
Each family is straightened by taking logarithms of a different variable.
Figure (svg): A table showing which variable to take logarithms of for each model family: outputs only for exponential, inputs only for logarithmic, and both for power models
The last row is the one people miss: a log-log plot straightening means a power law rather than an exponential, and the two are often confused because both plots look logarithmic.
Worked example
Which plot straightens tells you the family.
\[ \text{Data is straight on a log-log plot but curved on a semi-log plot. Which model?} \]
Recall what semi-log straightens
Why: Exponential models.
Recall what log-log straightens
Why: Power models.
Write the form
Why: A constant times a power of the input.
\[ y = a x ^{b} \]
Read the exponent
Why: It is the log-log plot's slope.
Figure (svg): A table showing which variable to take logarithms of for each model family: outputs only for exponential, inputs only for logarithmic, and both for power models
\[ y=ax^b, \text{ with } b \text{ the log-log slope} \]
Verify: confirm the derivation
Why: Taking logarithms of a power model gives log y equal to log a plus b times log x, which is linear in log x — so a log-log plot straightens it and its slope is the exponent b. The semi-log plot does not straighten it because the input was not transformed, which is exactly what the data showed.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 602-604
Matching
Which variable gets transformed decides.
Match the pairs
Why: Each transformation is matched to where the model's nonlinearity lives. An exponential has it in the output's dependence on x, a power law has it in both, a logarithmic model has it in the input, and a line has none at all.
Worked example
The plots decide where correlation coefficients cannot.
\[ \text{Data could be exponential or a power law. How do you decide?} \]
Make a semi-log plot
Why: Logarithms of the outputs only.
Make a log-log plot
Why: Logarithms of both.
Compare
Why: Whichever is straighter names the family.
Fit in that coordinate system
Why: Then convert back.
Figure (svg): The solution to Worked example distinguish two candidate models shown as a ladder of expressions, one row per legal move
\[ \text{semi-log straight } \Rightarrow \text{ exponential}; \text{ log-log straight } \Rightarrow \text{ power} \]
Verify: note why this beats comparing fits
Why: Fitting both models to the raw data and comparing correlations is unreliable, since both may fit acceptably over a limited range. The straightening test asks a sharper question — which structure the data actually has — and it answers it visually, which is exactly §2.4's argument for plotting before computing.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 604-606
Trap
\[ \text{data straightens on a log-log plot} \;\Longrightarrow\; \text{exponential growth} \]
Note that a logarithmic plot straightened the data
Why: A logarithmic axis was involved, so exponential growth is inferred.
The relationship is described as exponential.
A log-log plot straightens a POWER law, not an exponential. An exponential is straightened by a semi-log plot, with only the outputs transformed.
The two behave completely differently: a power law's growth rate slows relative to itself, while an exponential's does not, and §4.1 showed the exponential eventually wins by any margin.
Check which axes were transformed. One logarithmic axis means exponential; two means a power law, and the distinction changes every extrapolation you make.
Sorting
Identify the family first, then the transformation.
Sort into buckets
Sort each relationship.
Prediction
Data straightens on a log-log plot with a slope of 3.
Predict first
What does that say about the relationship?
Correct: The output varies as the cube of the input.
Why: A log-log plot straightens a power law and its slope is the exponent, so a slope of 3 means a cube relationship — exactly §3.9's variation language. The second option describes an exponential, which would straighten on a semi-log plot instead and whose slope would be a logarithm rather than the base.
Real world
Scientists reach for a log-log plot to test for a power law.
Discussion prompt
Why is a straight log-log plot considered strong evidence of a power law?
Hint: How many relationships produce a straight line under that transformation?
Answer:
Because only a power law straightens under that transformation. The algebra shows log y equal to log a plus b times log x, and no other common family produces a linear relationship between the two logarithms.
So straightness on a log-log plot is a structural finding rather than a good fit — it says the relationship has the form of a fixed power, and the slope reads off which power.
This is why the plot is a standard tool in physics and biology, where power laws with meaningful exponents are common: metabolic rate against body mass, or earthquake frequency against magnitude. The exponent is the discovery, and the plot delivers it as a slope.
Section
Section 4
Concept
Fitting a curve does not escape any of the cautions about fitting a line. Extrapolation is more dangerous rather than less, and a good fit still says nothing about causation.
The first point deserves emphasis. Extending a linear model too far produces an error proportional to the distance; extending an exponential produces an error that grows by a factor for every step. A modest extrapolation of an exponential can be wrong by orders of magnitude.
Figure (svg): A card restating the cautions from Section 2.4 as they apply to fitted exponential models, covering extrapolation, model choice, and the effect of the transformation on outliers
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 606-608
Picture it
Three cautions, all inherited from §2.4 and sharpened here.
Figure (svg): A card restating the cautions from Section 2.4 as they apply to fitted exponential models, covering extrapolation, model choice, and the effect of the transformation on outliers
The third is specific to this section: the transformation changes what the fitting procedure is minimising, which is worth knowing before quoting a correlation.
Worked example
The same distance out, two very different errors.
\[ \text{A linear and an exponential model both fit data on } [0,10]. \text{ Compare their forecasts at } 30. \]
Consider the linear model
Why: Errors grow in proportion to the distance.
Consider the exponential
Why: Each step multiplies by the base.
Quantify roughly
Why: A base of 2 over 20 extra steps.
Conclude
Why: The exponential extrapolation is far more sensitive.
Figure (svg): A card restating the cautions from Section 2.4 as they apply to fitted exponential models, covering extrapolation, model choice, and the effect of the transformation on outliers
\[ \text{exponential error compounds; linear error accumulates} \]
Verify: make the sensitivity concrete
Why: If the fitted base is 2 percent too high, then over 20 further steps the forecast is out by a factor of 1.02 to the twentieth, about 1.49 — nearly 50 percent too high, from a 2 percent error in one constant. A linear model with a 2 percent slope error would be out by 2 percent. That is the difference between compounding and accumulating.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 606-607
Sorting
Some facts help and some are irrelevant.
Sort into buckets
Sort each fact about a fitted exponential model.
Worked example
An unbounded model applied where a limit exists.
\[ \text{An exponential fits a population well for } 20 \text{ years. Forecast } 200 \text{ years out?} \]
Note what the model assumes
Why: Unlimited growth at a fixed rate.
Ask what the situation permits
Why: Food, space and resources are finite.
Predict the failure
Why: The model exceeds any capacity eventually.
Name the better model
Why: One with a capacity built in.
Figure (svg): The solution to Worked example notice a model breaking down shown as a ladder of expressions, one row per legal move
\[ \text{Model breakdown is certain; a logistic model is appropriate.} \]
Verify: connect to the earlier sections
Why: This is §2.3's model breakdown and §4.7's model choice arriving together. The exponential fits the data and is wrong about the future, and the data cannot reveal that — only knowing the situation can. A good fit over 20 years is not evidence about 200.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 607-608
Error analysis
A student justifies a distant forecast.
Annotate
On: \( r=0.997 \text{ on } [0,10] \;\Longrightarrow\; \text{the forecast at } x=50 \text{ is reliable} \)
This is §2.4's warning unchanged, and it bites harder for exponentials. Fit quality and extrapolation range are independent, and a high correlation never licenses reaching far beyond the evidence.
Prediction
A linear and an exponential model are each extended the same distance beyond their data.
Predict first
Which forecast is more sensitive to a small error in the fitted constants?
Correct: The exponential, because errors compound.
Why: A small error in an exponential's base is multiplied at every step, so it grows as a power of the distance. A linear model's slope error grows only in proportion. Over twenty steps a 2 percent base error becomes about a 49 percent error, while a 2 percent slope error stays 2 percent.
Two truths and a lie
Two of these are true of fitted exponential models and one is false.
Eliminate the wrong options
One of these claims is wrong.
Survives elimination: B
Why: B is the false claim. A correlation describes the fit over the measured range and carries no information about regions with no data. For an exponential this matters more than for a line, since the errors compound rather than accumulating.
Explain it
Both linear and exponential extrapolation are risky, and one is riskier.
Discussion prompt
Explain to a classmate why extending an exponential model is more dangerous than extending a line.
Hint: What happens to an error in the fitted constant at each further step?
Answer:
A line's error accumulates: a slightly wrong slope adds a slightly wrong amount at each step, so the total error grows in proportion to the distance.
An exponential's error compounds: a slightly wrong base multiplies the error at each step, so the total error grows as a power of the distance. Twenty steps of a 2 percent error becomes nearly 50 percent.
So the same relative uncertainty in a fitted constant produces a vastly larger forecast error for an exponential. A good explanation adds the practical consequence: exponential forecasts should be quoted with ranges and short horizons, and treated with more suspicion than linear ones at the same distance.
Section
Section 5
Concept
The full procedure combines this chapter's models with Chapter 2's fitting, and every step has a reason.
The last clause is what makes the answer honest. A fitted model without a stated range invites exactly the extrapolation the previous idea warned about, and stating it costs one sentence.
Figure (svg): A table showing which variable to take logarithms of for each model family: outputs only for exponential, inputs only for logarithmic, and both for power models
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 590-608
Picture it
Which transformation straightens the data names the family.
Figure (svg): A table showing which variable to take logarithms of for each model family: outputs only for exponential, inputs only for logarithmic, and both for power models
Trying all three is cheap and settles the question. It is also more informative than fitting several models and comparing their correlations.
Worked example
Six steps, in order.
\[ \text{Data curves upward. Fit a model.} \]
Plot and look
Why: Confirm the shape is a smooth curve.
Try a semi-log plot
Why: Logarithms of the outputs.
Fit a line to the transformed data
Why: Ordinary least squares.
Exponentiate both constants and check
Why: Convert back and test on the raw data.
Figure (svg): A table showing which variable to take logarithms of for each model family: outputs only for exponential, inputs only for logarithmic, and both for power models
\[ \text{semi-log straight } \Rightarrow \text{ fit } \Rightarrow \text{ exponentiate } \Rightarrow \text{ check} \]
Verify: state what would have changed the route
Why: If the semi-log plot had curved and the log-log plot had straightened, the family would be a power law and the transformation would have included the inputs. Trying both is quick, and skipping the identification step is how an exponential gets fitted to power-law data — which fits acceptably over a short range and extrapolates completely wrongly.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 601-606
Ranking
Six steps, each with a reason.
Put in order
Why: Identification comes first because it decides which transformation to apply. Fitting happens in transformed coordinates, converting brings the answer into the model's own terms, and checking happens on the original data because that is what the model describes.
Worked example
The range matters as much as the constants.
\[ \text{A model } y=5(1.4)^x \text{ was fitted to data with } x \text{ from } 0 \text{ to } 12. \]
State the model
Why: With its constants.
\[ y = 5(1.4) ^{x} \]
State the data range
Why: Where the evidence is.
\[ \text{fitted on } 0\text{ to } 12 \]
State the fit quality
Why: From the transformed correlation.
Warn about extrapolation
Why: Errors compound beyond the range.
\[ \text{unreliable far beyond } 12 \]
Figure (svg): The solution to Worked example report the model honestly shown as a ladder of expressions, one row per legal move
\[ y=5(1.4)^x \text{ for } 0\le x\le 12; \text{ extrapolate with caution} \]
Verify: consider what a reader can do with this
Why: With the range stated, a reader can see immediately whether a value of interest is interpolation or extrapolation. Without it, the model looks equally authoritative everywhere — which is precisely the impression §2.4 warned against, and it costs one clause to avoid.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 606-608
Trap
\[ \text{data curves upward} \;\Longrightarrow\; \text{fit an exponential} \]
Notice the upward curve and choose a model
Why: Upward curvature is taken as evidence of exponential growth.
An exponential model is fitted without checking which transformation straightens the data.
Upward curvature is consistent with several families: exponential, a power law with exponent above 1, and others. The curve alone does not identify it.
Trying the semi-log and log-log plots takes a minute and settles it. Fitting the wrong family gives a model that works over the data range and extrapolates completely wrongly.
Identify before fitting. The two families diverge enormously outside the data, so getting the identification wrong is a far bigger error than any imprecision in the constants.
Prediction
An exponential is fitted to data that is really a power law.
Predict first
What happens?
Correct: It fits acceptably over the data range and extrapolates badly.
Why: Over a limited range both families can produce similar upward curves, so the fit will look reasonable. Outside that range they diverge enormously, since an exponential eventually beats every power law. That is why identification matters more than fit quality, and why it must be done before fitting.
Faded example
A fit gives 8 times 1.25 to the x, on data from 0 to 15.
Fill in the blanks
y = 8(1.25)^x \text0 15 \le x \le ___
Why: The range is part of the model's statement, not an optional footnote. It lets a reader see immediately whether any particular prediction is interpolation or extrapolation, which is information the constants alone do not carry.
Real world
Fitting exponentials to data is routine across the sciences.
Discussion prompt
A biologist measures bacterial counts hourly and wants a growth model. Walk through what they would do.
Hint: What would they plot, and in what order?
Answer:
Plot the raw counts against time and see a curve. Then plot the logarithms of the counts against time and see whether it straightens — which for bacteria in unlimited nutrient it will.
Fit a line to the transformed data, then exponentiate to get the growth rate and the initial count. The slope gives the doubling time by §4.7's relationship.
Then check on the original counts and state the range. And critically, note that the model is exponential only while nutrients last — beyond that a logistic model applies, which §4.7 established and which the data will eventually show as the semi-log plot begins to bend.
Comparison
Fill the blanks from memory. Which plot straightens identifies the family.
Comparison matrix
| model | transform | slope means | |
|---|---|---|---|
| exponential | y = ab^x | log the outputs only | log of the base |
| power | y = ax^b | log both | the exponent itself |
| logarithmic | y = a + b ln x | log the inputs only | the coefficient b directly |
| linear | y = mx + c | none needed | the slope itself |
The second row's slope is the most directly useful: a log-log plot hands you the exponent as a slope, which is why power laws are discovered that way.
Pattern
Six steps, and the first two decide the rest.
Step 2 is worth the extra minute. Choosing the wrong family produces a model that fits the data and extrapolates wrongly, which is a far more expensive error than a slightly imprecise constant.
OpenStax Algebra and Trigonometry 2e, §6.8 Fitting Exponential Models to Data §6.8
Check
Which variable gets transformed?
Check your understanding
To linearise exponential data, what do you take logarithms of?
Answer: A
Why: Taking a logarithm of both sides of the exponential model gives log y equal to log a plus x times log b, which is linear in the untransformed x. So only the outputs need transforming, which is what a semi-log plot does.
Check
Both constants are logarithms.
Check your understanding
A fit gives log y equal to 0.6x plus 1. What is the exponential model, in base ten?
Answer: A
Why: Exponentiating gives ten to the power of the whole right-hand side, which splits into ten to the first, times ten to the 0.6x. So the coefficient is 10 and the base is ten to the 0.6, about 3.98.
Check
Two logarithmic axes.
Check your understanding
Data straightens on a log-log plot. Which model does it follow?
Answer: A
Why: Taking logarithms of both variables straightens a power law, and the resulting slope is the exponent. An exponential is straightened by transforming only the outputs, which is a semi-log plot.
Real world
The semi-log plot is one of the most-used diagnostics in science.
Discussion prompt
During an epidemic, case counts are routinely shown on a logarithmic vertical axis. What does that plot reveal that a linear one does not?
Hint: What does a straight line on that plot mean?
Answer:
A straight line means exponential growth, and its slope is the growth rate. That is far easier to judge by eye than exponential curvature on a linear axis, where everything looks flat and then vertical.
More usefully, a bend is immediately visible: when the line starts to flatten, growth is slowing, which is the first sign of a logistic shape emerging. On a linear plot that change is invisible until much later.
It also compresses an enormous range onto one picture, so early and late data are both readable. The transformation that makes fitting possible is the same one that makes the trend visible, which is why the technique and the diagnostic are the same idea.
Commit first
State your confidence along with your answer.
Predict first
A fitted exponential has an excellent correlation on the transformed data. What does that establish?
Correct: The model fits well over the range of the data.
Why: A correlation measures fit where the data is and says nothing about anywhere else, exactly as §2.4 established. It also establishes no mechanism, and a power law often fits comparably over a limited range — which is why the identification step matters more than the correlation.
Explain it
The test of understanding is being able to say why, not just what.
Discussion prompt
Explain to a classmate why taking logarithms lets you fit an exponential using only the tools from Chapter 2.
Hint: What does a logarithm do to repeated multiplication?
Answer:
Exponential data multiplies by a fixed factor at each step. A logarithm turns multiplication into addition, so the transformed data adds a fixed amount at each step.
Adding a fixed amount per step is a constant rate of change, which is exactly §2.1's definition of linear. So the transformed data lies on a line, and §2.4's regression fits lines.
Then the answer is converted back by exponentiating. A good explanation stresses that no new fitting machinery was needed — the whole technique is turning an unfamiliar problem into one that was already solved, which is a strategy worth recognising because it recurs constantly.
Exit ticket
One honest answer, so the next chapter can start in the right place.
Predict first
Which idea from this lesson would you most want to see again?
Correct: Any of these is a legitimate answer; the useful one is the honest one.
Why: There is no correct choice here. The third is the one that prevents the most expensive error, since fitting the wrong family produces a model that looks fine and extrapolates wrongly. The second is where the mechanical marks are, and the conversion step is the one most often left undone.
Connect it up
One page, drawn from memory, closes the chapter.
Draw it
Sketch the same curved data twice: once raw and once with the outputs transformed, showing the second one straight. Beside them write the four-line derivation showing why the logarithm straightens it, marking which constant becomes the slope and which the intercept. Then list the three families with the transformation each needs, and note one caution that carries over from Section 2.4.
If your derivation shows the base becoming the slope and the coefficient becoming the intercept, you have the fact that makes the conversion back possible.
Recap
Five things, closing a chapter that began with a new family and ends by fitting it to data.
| if you remember one thing | it should be this |
|---|---|
| about the technique | transform, fit a line, and convert back by exponentiating |
| about identification | which plot straightens names the family |
| about converting back | both constants are logarithms and both need exponentiating |
| about extrapolation | exponential errors compound rather than accumulate |
Chapter 5 leaves growth behind for a family that does neither: the trigonometric functions, which repeat forever and describe everything that oscillates.
OpenStax, Precalculus, §4.8 Fitting Exponential Models to Data §4.8, pp. 590-608 — everything on these slides traces back here
Want this taught 1-on-1? Alexander tutors Precalculus — $55/session, free consultation.