Parameter Sensitivity - Differential Equations

SECTION 1

Historical Context & Motivation

Every mathematical model of the real world relies on numbers called parameters — values such as growth rates, friction coefficients, or decay constants that we plug into our equations. But how confident can we be in those numbers? Even a tiny measurement error or rounding choice can ripple through a differential equation and change the predicted outcome. The study of parameter sensitivity grew directly out of scientists' need to know which parameters matter most and how much "wiggle room" a model can tolerate before its predictions fall apart.

1890s • Poincaré and the Three-Body Problem

Henri Poincaré showed that tiny changes in initial conditions of gravitational systems could produce wildly different orbits, foreshadowing modern sensitivity analysis.

1920s • Engineering Control Theory

Engineers designing feedback controllers for ships and aircraft began systematically studying how component tolerances (parameters) affected system stability.

1960s • Lorenz and the Butterfly Effect

Meteorologist Edward Lorenz discovered that rounding a parameter from 0.506127 to 0.506 completely changed a weather simulation's output, dramatically illustrating parameter sensitivity in differential equations.

1990s–Today • Computational Sensitivity Analysis

With modern computers, researchers now routinely perform sensitivity analysis on models in medicine, climate science, economics, and engineering to identify which parameters require the most precise measurement.

The core question that parameter sensitivity addresses is this: If I change a parameter by a small amount, how much does the solution of my differential equation change? Answering this question helps us decide which measurements need to be precise, which assumptions are safe, and where a model might break down.

SECTION 2

Core Principles & Definitions

Before diving into calculations, let's establish the foundational ideas behind parameter sensitivity. These principles apply whether you're modeling population growth, chemical reactions, or the motion of a pendulum.

Parameters vs. Variables

A variable is the unknown quantity that changes over time (like population). A parameter is a fixed constant that shapes the equation's behavior (like the growth rate).

Sensitivity Defined

A model is sensitive to a parameter when a small change in that parameter causes a large change in the solution. It is insensitive (or robust) when changes in the parameter barely affect the outcome.

Local vs. Global Sensitivity

Local sensitivity examines the effect of a tiny perturbation around a specific parameter value. Global sensitivity explores the effect across a wide range of possible parameter values.

The Sensitivity Coefficient

The sensitivity coefficient is a number that tells you the ratio: how much the output changes divided by how much the parameter changed. Larger absolute values mean greater sensitivity.

✦

KEY TAKEAWAY

Think of parameter sensitivity like the volume knob on a speaker. Some knobs are extremely touchy — a tiny twist makes the music blast from quiet to deafening (high sensitivity). Other knobs barely change the sound no matter how far you turn them (low sensitivity). In a differential equation, sensitivity analysis tells you which "knobs" in your model are the touchy ones, so you know where to measure carefully.

SECTION 3

Visualizing Parameter Sensitivity

The diagram below shows the exponential growth model dy/dt = ky, where three slightly different values of the growth-rate parameter k are used. Notice how solutions that start at the same initial value fan apart over time — a hallmark of parameter sensitivity.

All three curves start at the same initial condition y₀ = 1, but the green curve (k = 0.20) grows modestly while the pink curve (k = 0.40) shoots upward much faster. The spreading fan between curves demonstrates that the exponential model is highly sensitive to the parameter k, especially at later times.

The key observation here is that the curves don't just shift up or down — they diverge from each other as time goes on. Early in the simulation (small t), all three solutions look similar. But by t = 10, the pink curve (k = 0.40) has reached a value far larger than the green curve (k = 0.20). This amplification of small parameter differences is exactly what makes sensitivity analysis so important in real-world modeling.

SECTION 4

Mathematical Framework

Let's formalize what we mean by parameter sensitivity. Suppose we have a differential equation whose solution y(t) depends on a parameter p. We want to quantify how y changes when p changes by a small amount.

GENERAL FORM

dy/dt = f(t, y, p)

Here y is the dependent variable, t is time, p is the parameter of interest, and f is a function that defines the rate of change.

SENSITIVITY COEFFICIENT

S(t) = ∂y/∂p

The sensitivity coefficient S(t) is the partial derivative of the solution with respect to the parameter p. It tells you: for each unit change in p, how many units does y change at time t?

For the exponential model dy/dt = ky with y(0) = y₀, the exact solution is y(t) = y₀e^kt. If we treat k as the parameter, we can compute the sensitivity coefficient by taking the partial derivative with respect to k.

EXPONENTIAL SENSITIVITY

S(t) = ∂y/∂k = y₀ · t · e^(kt)

Notice the factor of t — sensitivity grows over time. This confirms what we saw in the graph: the curves spread apart more at larger values of t.

RELATIVE SENSITIVITY

S_rel = (p / y) × (∂y/∂p)

The relative (or normalized) sensitivity removes units and lets you compare sensitivity across different parameters. If S_rel = 2, a 1% change in p causes approximately a 2% change in y.

💡

Practical Shortcut

When you can't compute the derivative analytically, you can estimate sensitivity numerically by running your model twice — once with the original parameter value p and once with a slightly changed value p + Δp — then computing S ≈ [y(t; p + Δp) − y(t; p)] / Δp.

SECTION 5

Types of Sensitivity & When They Matter

Not all parameters affect a model equally, and the degree of sensitivity can vary depending on the context. The diagram below classifies sensitivity behavior into three categories and shows how the solution curve responds in each case.

Each panel shows a base solution (solid line) alongside two perturbed solutions (dashed lines) from slightly different parameter values. In the low sensitivity panel, the curves are nearly indistinguishable. In the high sensitivity panel, even a small parameter shift causes the curves to diverge dramatically.

Category	\|S_rel\| Range	Interpretation	Action Required
Low	< 0.5	A 1% change in p causes less than a 0.5% change in y	Rough estimates of the parameter are sufficient
Moderate	0.5 – 2	A 1% change in p causes up to a 2% change in y	Measure the parameter with reasonable care
High	> 2	A 1% change in p causes more than a 2% change in y	High-precision measurement essential; consider model redesign

Sensitivity classification guide

SECTION 6

Worked Example: Population Growth Sensitivity

A biologist models a bacteria colony with the exponential growth equation dy/dt = ky, where the initial population is y₀ = 100 cells and the estimated growth rate is k = 0.50 per hour. She wants to know: if the true growth rate is actually k = 0.55 (a 10% error), how much will the predicted population at t = 4 hours differ from the true population?

Bacteria Colony — Sensitivity to Growth Rate k

1
Step 1 — Write the solution

The general solution of dy/dt = ky is y(t) = y₀ × e^kt. We will evaluate this at t = 4 hours for both k values.

y(t) = 100 × e^k×4
2
Step 2 — Compute y with k = 0.50

Substituting k = 0.50 and t = 4: y(4) = 100 × e^{0.50 × 4} = 100 × e^2.0 ≈ 100 × 7.389 = 738.9 cells.

y(4) ≈ 738.9 cells (estimated)
3
Step 3 — Compute y with k = 0.55

Now with k = 0.55: y(4) = 100 × e^{0.55 × 4} = 100 × e^2.2 ≈ 100 × 9.025 = 902.5 cells.

y(4) ≈ 902.5 cells (true)
4
Step 4 — Find the absolute and percent error

The absolute difference is 902.5 − 738.9 = 163.6 cells. The percent error relative to the true value is (163.6 / 902.5) × 100% ≈ 18.1%. So a 10% error in k produced roughly an 18% error in the predicted population.

≈ 18.1% prediction error from a 10% parameter error
5
Step 5 — Compute the sensitivity coefficient

The analytical sensitivity coefficient is S(t) = ∂y/∂k = y₀ × t × e^kt. At k = 0.50 and t = 4: S(4) = 100 × 4 × e^2.0 ≈ 2955.6. The relative sensitivity is S_rel = (k/y) × S = (0.50 / 738.9) × 2955.6 ≈ 2.0. This means the model has moderate-to-high sensitivity to k at t = 4: a 1% change in k causes roughly a 2% change in the output.

S_rel ≈ 2.0 (moderate-to-high sensitivity)

SECTION 7

Strengths, Limitations & Common Pitfalls

Parameter sensitivity analysis is a powerful tool, but like any technique, it has strengths and limitations that you should understand before applying it.

Strengths	Limitations
Identifies which parameters matter most, focusing measurement effort where it counts	Local sensitivity only applies near one parameter value — the conclusion may differ at other values
Can be computed analytically for many standard differential equations	For complex models with many parameters, computing all sensitivity coefficients becomes tedious or computationally expensive
Helps validate models — if a model is highly sensitive to a poorly measured parameter, you know to distrust its predictions	Sensitivity analysis tells you how much the output changes, not whether the model itself is correct
Works for linear and nonlinear systems alike	Interactions between parameters are not captured by simple one-at-a-time analysis

Strengths and limitations of parameter sensitivity analysis

✦

KEY TAKEAWAY

Sensitivity analysis is like checking a recipe before cooking for a crowd. If the dish tastes great whether you add 1 teaspoon or 1½ teaspoons of salt (low sensitivity), you can eyeball it. But if the difference between 1 and 1½ teaspoons turns the dish from perfect to inedible (high sensitivity), you'd better use a precise measuring spoon. Sensitivity analysis tells you which ingredients in your model need the measuring spoon.

SECTION 8

Connections to Advanced Theory

The local sensitivity analysis you learned in this lesson is a starting point. As you continue studying differential equations and modeling, you'll encounter more sophisticated approaches. The table below compares what we've covered with two advanced techniques.

Feature	Local Sensitivity (This Lesson)	Global Sensitivity (Advanced)	Bifurcation Analysis (Advanced)
What it measures	How the output changes for a tiny perturbation in one parameter	How the output changes across the full range of all parameter values simultaneously	Critical parameter values where the qualitative behavior of the system changes entirely
Math required	Partial derivatives or simple numerical estimation	Monte Carlo simulation, variance decomposition	Eigenvalue analysis, phase-plane diagrams
Key advantage	Quick and intuitive	Handles parameter interactions and large uncertainties	Reveals when a system "tips" into completely new behavior (e.g., stable → oscillating)
Typical course level	High school / introductory college	Upper-level undergraduate	Upper-level undergraduate / graduate

Local sensitivity vs. advanced techniques

In future courses, you may also encounter the concept of structural sensitivity, which asks not just how parameter values affect the solution, but how the choice of the model equation itself changes outcomes. For now, the local sensitivity coefficient gives you a practical, powerful tool for evaluating any model you build with differential equations.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A student builds two models of the same population. In Model A, changing the birth rate by 5% changes the predicted population at t = 10 by 3%. In Model B, the same 5% change causes a 25% change in the prediction. Which model is more sensitive to the birth rate, and what does that imply about the importance of accurately measuring the birth rate in each model?

Show answer

Model B is far more sensitive to the birth rate than Model A. In Model A, a 5% change in the parameter produces only a 3% change in the output (relative sensitivity ≈ 0.6), indicating low sensitivity. In Model B, the same 5% change produces a 25% change in the output (relative sensitivity ≈ 5), indicating very high sensitivity. This means that for Model B, measuring the birth rate with high precision is critical — even a small measurement error could make the model's predictions unreliable. For Model A, a rough estimate of the birth rate would be sufficient.

Question 2: BASIC CALCULATION

Consider the decay model dy/dt = −ky with y(0) = 500 and k = 0.10 per minute. Compute y(5) for k = 0.10 and k = 0.12. Then find the percent change in y(5) and the approximate relative sensitivity S_rel.

Show answer

The solution is y(t) =

500e^{−kt}

. For k = 0.10: y(5) =

500e^{−0.50}

≈ 500 × 0.6065 ≈ 303.3. For k = 0.12: y(5) =

500e^{−0.60}

≈ 500 × 0.5488 ≈ 274.4. The percent change in y is (303.3 − 274.4)/303.3 × 100% ≈ 9.5%. The percent change in k is (0.12 − 0.10)/0.10 × 100% = 20%. So the approximate relative sensitivity is 9.5/20 ≈ 0.48. This is low sensitivity — a 20% error in k produces less than a 10% error in y(5).

Question 3: INTERMEDIATE

For the exponential growth model y(t) = y₀

e^{kt}

, the analytical sensitivity coefficient is S(t) = ∂y/∂k = y₀ × t ×

e^{kt}

. Compute S(t) at t = 2, 5, and 10 with y₀ = 50 and k = 0.30. What trend do you notice, and what does it tell you about making predictions far into the future?

Show answer

At t = 2: S(2) = 50 × 2 ×

e^{0.60}

≈ 100 × 1.822 ≈ 182.2. At t = 5: S(5) = 50 × 5 ×

e^{1.50}

≈ 250 × 4.482 ≈ 1120.4. At t = 10: S(10) = 50 × 10 ×

e^{3.0}

≈ 500 × 20.086 ≈ 10,043. The sensitivity coefficient grows rapidly with time — from about 182 at t = 2 to over 10,000 at t = 10. This tells us that exponential models become increasingly unreliable for long-term predictions if the growth rate k has any uncertainty. A small error in k that barely matters at t = 2 could produce enormous errors at t = 10.

Question 4: APPLIED

A pharmaceutical company models the concentration C(t) of a drug in the bloodstream as C(t) = C₀

e^{−λt}

, where C₀ = 200 mg/L and the elimination rate λ is estimated at 0.15 per hour. The drug is toxic above 100 mg/L and ineffective below 20 mg/L. If λ could actually be anywhere from 0.12 to 0.18, at t = 8 hours, what is the range of possible concentrations? Does sensitivity to λ pose a safety concern?

Show answer

For λ = 0.12: C(8) =

200e^{−0.96}

≈ 200 × 0.383 ≈ 76.6 mg/L. For λ = 0.15 (estimated): C(8) =

200e^{−1.20}

≈ 200 × 0.301 ≈ 60.2 mg/L. For λ = 0.18: C(8) =

200e^{−1.44}

≈ 200 × 0.237 ≈ 47.4 mg/L. The range is approximately 47 to 77 mg/L. All three values are within the therapeutic window (between 20 and 100 mg/L), so at t = 8 hours the uncertainty in λ does not pose a safety concern. However, the spread of about 30 mg/L shows moderate sensitivity. At earlier times (closer to t = 0), the upper estimate could approach the toxic threshold, so sensitivity analysis at multiple time points would be important for establishing safe dosing protocols.

Question 5: CRITICAL THINKING

Consider two models: (A) dy/dt = ky with solution y = y₀

e^{kt}

, and (B) dy/dt = a, a constant, with solution y = y₀ + at. Both models have one parameter (k in Model A, a in Model B). Compute the relative sensitivity S_rel for each model and explain why exponential models are inherently more sensitive than linear models. Under what circumstances might you prefer a linear approximation even if an exponential model is more realistic?

Show answer

For Model A: ∂y/∂k = y₀

te^{kt}

and y = y₀

e^{kt}

, so S_rel = (k/y)(∂y/∂k) = k × t. This grows without bound as t increases. For Model B: ∂y/∂a = t and y = y₀ + at, so S_rel = (a/(y₀ + at)) × t = at/(y₀ + at). As t → ∞, this approaches 1. So for the exponential model, relative sensitivity equals kt and grows linearly in time, while for the linear model, relative sensitivity is bounded and never exceeds 1. Exponential models amplify parameter errors over time because the parameter k appears in the exponent, meaning its effect compounds. You might prefer a linear approximation when making long-range predictions with uncertain parameters, because the linear model's bounded sensitivity means errors stay controlled. The trade-off is accuracy: the linear model may be less realistic but more trustworthy when parameter uncertainty is large.

SUMMARY

Lesson Summary

Summary

Parameter sensitivity measures how much the solution of a differential equation changes when one of its parameters is slightly altered. The sensitivity coefficient S = ∂y/∂p quantifies this relationship, while the relative sensitivity S_rel = (p/y)(∂y/∂p) gives a unitless ratio that lets you compare sensitivity across different parameters. Models with |S_rel| > 2 are classified as highly sensitive and require precise parameter measurements.

In exponential models like dy/dt = ky, the relative sensitivity equals kt, which grows over time — meaning long-term predictions are especially vulnerable to parameter uncertainty. By contrast, linear models have bounded relative sensitivity. Whether you're modeling populations, drug concentrations, or physical systems, sensitivity analysis helps you identify which "knobs" in your model matter most and where to invest your measurement effort.