Choosing Models — Choosing a Model: Linear vs Exponential...

SECTION 1

Historical Context & Motivation

The practice of fitting mathematical functions to observed phenomena is as old as modern science itself. From the moment merchants began tracking prices and populations grew beyond what simple counting could manage, analysts have sought compact expressions that capture the essential behavior of complex systems. The question of model selection — choosing the right functional form for a given data set — sits at the heart of quantitative business analysis and has evolved significantly over the past four centuries.

1637

Descartes & Coordinate Geometry

René Descartes published La Géométrie, introducing the Cartesian plane and making it possible to represent algebraic relationships as geometric curves. This laid the groundwork for graphing functions and visually comparing model types.

1798

Malthus & Exponential Growth

Thomas Malthus argued that populations grow exponentially while food supplies grow linearly, introducing one of the earliest explicit comparisons between linear and exponential models in a policy context.

1614–1800s

Napier's Logarithms & Industrial Applications

John Napier introduced logarithms in 1614 as computational aids. By the 19th century, logarithmic scales were standard tools in engineering, acoustics, and economics — wherever diminishing-return phenomena appeared.

1805

Legendre & Least Squares

Adrien-Marie Legendre published the method of least squares, giving analysts a formal criterion for choosing among competing models by minimizing the sum of squared residuals.

20th–21st Century

Computational Model Selection

With spreadsheets, statistical software, and machine learning, model selection became both more powerful and more critical. Businesses now routinely compare linear, exponential, and logarithmic fits to forecast revenues, costs, and market penetration.

The central question this lesson addresses is deceptively simple: given a set of business data — sales figures, cost projections, market saturation curves — how do you decide whether a linear, exponential, or logarithmic function best describes the underlying relationship? Making the wrong choice leads to models that either underpredict growth, overestimate returns, or miss critical inflection points entirely.

SECTION 2

Core Principles & Definitions

Before comparing the three model families, it is essential to understand what each one describes in terms of the rate and pattern of change. A model is not merely a curve drawn through data points; it embodies an assumption about the mechanism driving change. Selecting a model therefore requires matching the structural behavior of the function to the structural behavior of the real-world process.

Linear Model

A linear function has the form y = mx + b. It assumes a constant rate of change — every unit increase in x produces the same change in y. Use it when growth or decline is steady and uniform.

Exponential Model

An exponential function has the form y = a·b^x (b > 0, b ≠ 1). It assumes a constant percentage rate of change. Each unit step multiplies y by the same factor.

Logarithmic Model

A logarithmic function has the form y = a + b·ln(x). It assumes that gains diminish as the input grows. Early changes are dramatic; later changes taper off.

First Differences & Ratios

For equally spaced data, compute first differences (Δy = y_n+1 − y_n). If constant → linear. Compute ratios (y_n+1/y_n). If constant → exponential.

Residual Analysis

After fitting a model, examine residuals (observed − predicted). If residuals show a pattern (curved, fanning), the model form is likely wrong, regardless of a high R² value.

✦ KEY TAKEAWAY

Think of model selection like choosing a vehicle for a trip. A linear model is a train on straight tracks — constant speed, predictable, but it cannot handle curves. An exponential model is a rocket — accelerating faster and faster, ideal for capturing compounding growth or decay. A logarithmic model is a car climbing a steep hill that gradually flattens — initial progress is rapid, but each additional effort yields less distance. Choosing the wrong vehicle doesn't just slow you down; it sends you in the wrong direction entirely.

SECTION 3

Visual Comparison of the Three Models

The most immediate way to distinguish linear, exponential, and logarithmic behavior is to see all three curves on the same coordinate plane. The following diagram plots representative functions from each family over a common domain, making it visually clear how they diverge as x increases.

The blue linear curve rises at a constant slope. The violet exponential curve starts slowly but accelerates dramatically. The cyan logarithmic curve rises steeply at first then flattens, exhibiting diminishing returns.

Notice the critical divergence as x grows large. The linear function climbs steadily, adding the same absolute amount per unit step. The exponential function overtakes it because each step adds a fixed percentage of the current value — a small difference early on that compounds into an enormous gap. The logarithmic function, by contrast, effectively "runs out of steam": it never stops increasing, but the rate of increase approaches zero. In a business context, these three behaviors correspond to steady-state operations (linear), compounding growth or viral adoption (exponential), and market saturation or learning curves (logarithmic).

SECTION 4

Mathematical Framework

Each model family is characterized by a distinctive equation, derivative structure, and diagnostic test. Understanding the calculus behind each function clarifies why the curves look the way they do and provides rigorous criteria for model selection.

LINEAR MODEL

y = mx + b dy/dx = m (constant)

m = slope (constant rate of change per unit x); b = y-intercept (value when x = 0). The derivative is a constant, confirming that the function adds the same amount with every step.

EXPONENTIAL MODEL

y = a · bˣ dy/dx = a · bˣ · ln(b) = y · ln(b)

a = initial value (y-intercept when x = 0); b = growth factor (b > 1 for growth, 0 < b < 1 for decay). The derivative is proportional to y itself, so the rate of change accelerates (or decelerates) along with the function's current value.

LOGARITHMIC MODEL

y = a + b · ln(x) dy/dx = b / x

a = vertical shift; b = scaling constant; ln(x) = natural logarithm. The derivative b/x decreases as x grows, confirming diminishing returns. The function is only defined for x > 0.

DIAGNOSTIC: FIRST DIFFERENCES vs RATIOS

Δyₙ = yₙ₊₁ − yₙ rₙ = yₙ₊₁ / yₙ

For equally spaced x-values: if all Δyₙ are approximately equal, choose linear. If all rₙ are approximately equal, choose exponential. If Δyₙ decreases and the differences of the x-values in the original data grow proportionally, consider logarithmic.

📐 Connection to Calculus

The derivative provides the most elegant diagnostic: if dy/dx is constant, the model is linear. If dy/dx is proportional to y, the model is exponential. If dy/dx is proportional to 1/x, the model is logarithmic. In business calculus, analyzing marginal quantities (marginal cost, marginal revenue) is equivalent to examining the derivative, so model selection connects directly to marginal analysis.

SECTION 5

Decision Framework & Data Diagnostics

Choosing the right model involves a systematic process: inspect the scatter plot, compute first differences and ratios, consider the real-world context, and verify the choice with residual analysis. The flowchart below synthesizes this process into a repeatable decision tree that you can apply to any business data set.

Follow the flowchart from top to bottom. Begin by plotting the scatter plot, then compute first differences and ratios. Constant differences point to linear; constant ratios point to exponential; decreasing differences with a concave-down shape suggest logarithmic. Always verify your choice with a residual check.

Summary of diagnostic signals for each model type

Diagnostic	Linear Signal	Exponential Signal	Logarithmic Signal
Scatter plot shape	Points cluster along a straight line	Points curve upward (growth) or downward (decay) with increasing steepness	Points rise steeply at first, then level off (concave down)
First differences (Δy)	Approximately constant	Increasing (growth) or decreasing in magnitude (decay)	Decreasing and approaching zero
Consecutive ratios (yₙ₊₁/yₙ)	Varying (not a useful diagnostic)	Approximately constant	Approaching 1
Derivative behavior	dy/dx = constant	dy/dx ∝ y	dy/dx ∝ 1/x

Context also matters enormously. Even if your data looks roughly linear over a narrow range, knowledge that the underlying process involves compounding (interest rates, viral sharing, bacterial growth) should push you toward an exponential model. Similarly, phenomena known to exhibit diminishing marginal returns — such as advertising effectiveness, learning curves, or utility from consumption — are natural candidates for logarithmic fits, even before you crunch the numbers.

SECTION 6

Worked Example: Selecting a Model for Sales Data

A startup tracks monthly subscribers (in thousands) over the first six months of a product launch. The data are: Month 1: 2, Month 2: 4.1, Month 3: 8.0, Month 4: 16.2, Month 5: 32.1, Month 6: 64.5. Determine whether a linear, exponential, or logarithmic model best fits this data, find the model equation, and project Month 8 subscribers.

Choosing and Fitting a Model to Subscriber Data

Step 1 — Compute First DifferencesCalculate Δy for each consecutive pair: Δy₁ = 4.1 − 2 = 2.1, Δy₂ = 8.0 − 4.1 = 3.9, Δy₃ = 16.2 − 8.0 = 8.2, Δy₄ = 32.1 − 16.2 = 15.9, Δy₅ = 64.5 − 32.1 = 32.4. These differences are clearly not constant — they roughly double each period. This rules out a linear model.

First differences are not constant → not linear

Step 2 — Compute Consecutive RatiosCalculate rₙ = yₙ₊₁/yₙ: r₁ = 4.1/2 = 2.05, r₂ = 8.0/4.1 ≈ 1.95, r₃ = 16.2/8.0 ≈ 2.025, r₄ = 32.1/16.2 ≈ 1.981, r₅ = 64.5/32.1 ≈ 2.009. These ratios are approximately constant at about 2.0 — strong evidence for an exponential model.

Consecutive ratios ≈ 2.0 → exponential model confirmed

Step 3 — Determine the Exponential EquationThe general form is y = a · b^x. The average ratio b ≈ 2.0. Using the point (1, 2): 2 = a · 2¹, so a = 1. Our model is y = 1 · 2^x or equivalently y = 2^x. Verification: at x = 6, y = 2⁶ = 64, which closely matches the observed 64.5 (within 0.8% error).

Model: y = 2ˣ (thousands of subscribers)

Step 4 — Project Month 8 SubscribersSubstituting x = 8: y = 2⁸ = 256. The model projects approximately 256,000 subscribers by Month 8. Note the important caveat: exponential growth rarely persists indefinitely in business settings. Market saturation, competition, and resource constraints will eventually slow growth, and a logistic or logarithmic model may become more appropriate at that stage.

Month 8 projection: ≈ 256,000 subscribers

SECTION 7

Strengths & Limitations of Each Model

No single model is universally "best." Each excels in particular scenarios and fails in others. The table below summarizes the strengths, limitations, and typical business applications for each model family, helping you develop intuition for when to deploy each one.

Comparative strengths and limitations of the three model families

Feature	Linear	Exponential	Logarithmic
Strengths	Simple, interpretable, easy to fit. Excellent for short-term forecasting of steady processes.	Captures compounding phenomena accurately. Models growth/decay where rate depends on current value.	Models diminishing returns naturally. Handles situations where early gains are large but taper off.
Limitations	Cannot capture accelerating or decelerating trends. Extrapolation predicts negative values incorrectly.	Predicts infinite growth, which is unrealistic long-term. Very sensitive to parameter estimation.	Cannot model accelerating growth. Undefined at x = 0. Poor for early-stage explosive growth.
Business examples	Fixed monthly rent, straight-line depreciation, hourly wages, uniform production costs	Compound interest, viral marketing, population/user growth, radioactive decay in manufacturing	Diminishing ad returns, learning curves, customer satisfaction vs. features, Richter-scale phenomena
Long-run behavior	y → ±∞ at a constant rate	y → ∞ explosively (growth) or y → 0 asymptotically (decay)	y → ∞ but extremely slowly (quasi-flat)

✦ KEY TAKEAWAY

A model is a simplification, not a prediction of truth. In practice, many business phenomena transition between model types over time: a startup's user base may grow exponentially at launch, settle into linear growth during maturity, and eventually follow a logarithmic pattern as the market saturates. The best analysts reassess their model choice as new data arrives, rather than forcing all data into a single function.

SECTION 8

Connection to Advanced Modeling Techniques

The linear, exponential, and logarithmic models are foundational building blocks that extend into more sophisticated modeling frameworks. Understanding their behavior prepares you for the powerful tools used in advanced business analytics, operations research, and data science.

How basic models connect to advanced techniques

Basic Model	Advanced Extension	Key Idea
Linear: y = mx + b	Multiple linear regression: y = b₀ + b₁x₁ + b₂x₂ + ⋯	Add more predictor variables while retaining the constant-rate-of-change assumption for each.
Exponential: y = a·bˣ	Logistic model: y = L / (1 + e^{−k(x−x₀)})	Adds a carrying capacity L so that exponential growth transitions to saturation — ideal for market adoption curves.
Logarithmic: y = a + b·ln(x)	Power model: y = a·x^b	Generalizes the logarithmic relationship; when b < 1, the function shows diminishing returns but with more flexibility than pure log.
All three	Nonlinear regression & AIC/BIC	Fit arbitrary nonlinear functions and use information criteria (Akaike, Bayesian) to formally compare model quality while penalizing complexity.

An especially important extension is the logistic growth model, which behaves exponentially at first but asymptotically approaches a maximum carrying capacity. This is the standard model for product adoption in marketing (the S-curve), diffusion of innovation, and epidemic modeling. If you master the distinction between linear, exponential, and logarithmic behavior in this lesson, you will find the logistic model intuitive: it is essentially an exponential model that "switches" to logarithmic behavior once saturation effects dominate.

🔗 Linearization Trick

Many advanced models can be reduced to linear form through transformations. Taking the natural log of both sides of y = a·b^x yields ln(y) = ln(a) + x·ln(b), which is linear in x. Similarly, y = a·x^b becomes ln(y) = ln(a) + b·ln(x), linear in ln(x). This is why semi-log and log-log plots are staples of business data analysis — they reveal whether an exponential or power model is appropriate by converting curves into straight lines.

SECTION 9

Practice Problems

PROBLEM 1 — CONCEPTUAL

A company's total fixed costs increase by exactly $5,000 for every additional unit of factory space leased, regardless of how many units have already been leased. Which model type — linear, exponential, or logarithmic — should be used to represent total fixed cost as a function of leased units? Explain your reasoning in terms of the derivative.

PROBLEM 2 — BASIC CALCULATION

An investment of $1,000 earns 6% interest compounded annually. Write the exponential model for the account balance A after t years, and compute the balance at t = 10 years. Round to the nearest cent.

PROBLEM 3 — INTERMEDIATE

A marketing team collects the following data on ad spend (in $1,000s) vs. new customer acquisitions: (1, 50), (2, 85), (4, 120), (8, 155), (16, 188). Determine whether a linear or logarithmic model fits this data better by computing first differences and examining the pattern relative to the x-values.

PROBLEM 4 — APPLIED

A SaaS company reports quarterly revenue (in millions) for two years: Q1: $2.0, Q2: $2.5, Q3: $3.1, Q4: $3.9, Q5: $4.9, Q6: $6.1, Q7: $7.6, Q8: $9.5. (a) Use the first differences and ratios diagnostic to determine the best model. (b) Fit the model. (c) Project Q12 revenue. (d) Discuss one reason the projection might be unreliable.

PROBLEM 5 — CRITICAL THINKING

A dataset appears to fit both an exponential model (R² = 0.94) and a logarithmic model (R² = 0.91) reasonably well over the observed range x ∈ [1, 10]. However, you need to extrapolate to x = 50. (a) Explain why the two models will give vastly different predictions at x = 50. (b) Describe a mathematical and a contextual strategy for deciding which model to trust for extrapolation. (c) Under what circumstances might neither model be appropriate?

SUMMARY

Lesson Summary

Choosing the right model is a foundational skill in business calculus. A linear model (y = mx + b) is appropriate when data exhibits a constant rate of change, identified by constant first differences in equally spaced data. An exponential model (y = a·bˣ) applies when the percentage rate of change is constant, signaled by constant consecutive ratios — typical of compound interest, viral growth, and depreciation. A logarithmic model (y = a + b·ln(x)) captures diminishing returns, where initial inputs produce large gains but subsequent increases yield progressively less.

The derivative provides the most elegant diagnostic: constant dy/dx → linear; dy/dx ∝ y → exponential; dy/dx ∝ 1/x → logarithmic. Always combine numerical diagnostics with contextual knowledge of the business process and verify your choice through residual analysis. Remember that real-world phenomena often transition between model types over time, and the best analysts revisit their model selection as new data arrives. Mastering these three fundamental model families prepares you for advanced techniques including logistic models, power models, and nonlinear regression.